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/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/syscalls.h>
38 #include <linux/hugetlb.h>
39 #include <linux/freezer.h>
40 #include <linux/memblock.h>
41 #include <linux/fault-inject.h>
42 #include <linux/time_namespace.h>
44 #include <asm/futex.h>
46 #include "locking/rtmutex_common.h"
49 * READ this before attempting to hack on futexes!
51 * Basic futex operation and ordering guarantees
52 * =============================================
54 * The waiter reads the futex value in user space and calls
55 * futex_wait(). This function computes the hash bucket and acquires
56 * the hash bucket lock. After that it reads the futex user space value
57 * again and verifies that the data has not changed. If it has not changed
58 * it enqueues itself into the hash bucket, releases the hash bucket lock
61 * The waker side modifies the user space value of the futex and calls
62 * futex_wake(). This function computes the hash bucket and acquires the
63 * hash bucket lock. Then it looks for waiters on that futex in the hash
64 * bucket and wakes them.
66 * In futex wake up scenarios where no tasks are blocked on a futex, taking
67 * the hb spinlock can be avoided and simply return. In order for this
68 * optimization to work, ordering guarantees must exist so that the waiter
69 * being added to the list is acknowledged when the list is concurrently being
70 * checked by the waker, avoiding scenarios like the following:
74 * sys_futex(WAIT, futex, val);
75 * futex_wait(futex, val);
78 * sys_futex(WAKE, futex);
83 * lock(hash_bucket(futex));
85 * unlock(hash_bucket(futex));
88 * This would cause the waiter on CPU 0 to wait forever because it
89 * missed the transition of the user space value from val to newval
90 * and the waker did not find the waiter in the hash bucket queue.
92 * The correct serialization ensures that a waiter either observes
93 * the changed user space value before blocking or is woken by a
98 * sys_futex(WAIT, futex, val);
99 * futex_wait(futex, val);
102 * smp_mb(); (A) <-- paired with -.
104 * lock(hash_bucket(futex)); |
108 * | sys_futex(WAKE, futex);
109 * | futex_wake(futex);
111 * `--------> smp_mb(); (B)
114 * unlock(hash_bucket(futex));
115 * schedule(); if (waiters)
116 * lock(hash_bucket(futex));
117 * else wake_waiters(futex);
118 * waiters--; (b) unlock(hash_bucket(futex));
120 * Where (A) orders the waiters increment and the futex value read through
121 * atomic operations (see hb_waiters_inc) and where (B) orders the write
122 * to futex and the waiters read (see hb_waiters_pending()).
124 * This yields the following case (where X:=waiters, Y:=futex):
132 * Which guarantees that x==0 && y==0 is impossible; which translates back into
133 * the guarantee that we cannot both miss the futex variable change and the
136 * Note that a new waiter is accounted for in (a) even when it is possible that
137 * the wait call can return error, in which case we backtrack from it in (b).
138 * Refer to the comment in queue_lock().
140 * Similarly, in order to account for waiters being requeued on another
141 * address we always increment the waiters for the destination bucket before
142 * acquiring the lock. It then decrements them again after releasing it -
143 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
144 * will do the additional required waiter count housekeeping. This is done for
145 * double_lock_hb() and double_unlock_hb(), respectively.
148 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
149 #define futex_cmpxchg_enabled 1
151 static int __read_mostly futex_cmpxchg_enabled
;
155 * Futex flags used to encode options to functions and preserve them across
159 # define FLAGS_SHARED 0x01
162 * NOMMU does not have per process address space. Let the compiler optimize
165 # define FLAGS_SHARED 0x00
167 #define FLAGS_CLOCKRT 0x02
168 #define FLAGS_HAS_TIMEOUT 0x04
171 * Priority Inheritance state:
173 struct futex_pi_state
{
175 * list of 'owned' pi_state instances - these have to be
176 * cleaned up in do_exit() if the task exits prematurely:
178 struct list_head list
;
183 struct rt_mutex pi_mutex
;
185 struct task_struct
*owner
;
189 } __randomize_layout
;
192 * struct futex_q - The hashed futex queue entry, one per waiting task
193 * @list: priority-sorted list of tasks waiting on this futex
194 * @task: the task waiting on the futex
195 * @lock_ptr: the hash bucket lock
196 * @key: the key the futex is hashed on
197 * @pi_state: optional priority inheritance state
198 * @rt_waiter: rt_waiter storage for use with requeue_pi
199 * @requeue_pi_key: the requeue_pi target futex key
200 * @bitset: bitset for the optional bitmasked wakeup
202 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
203 * we can wake only the relevant ones (hashed queues may be shared).
205 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
206 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
207 * The order of wakeup is always to make the first condition true, then
210 * PI futexes are typically woken before they are removed from the hash list via
211 * the rt_mutex code. See unqueue_me_pi().
214 struct plist_node list
;
216 struct task_struct
*task
;
217 spinlock_t
*lock_ptr
;
219 struct futex_pi_state
*pi_state
;
220 struct rt_mutex_waiter
*rt_waiter
;
221 union futex_key
*requeue_pi_key
;
223 } __randomize_layout
;
225 static const struct futex_q futex_q_init
= {
226 /* list gets initialized in queue_me()*/
227 .key
= FUTEX_KEY_INIT
,
228 .bitset
= FUTEX_BITSET_MATCH_ANY
232 * Hash buckets are shared by all the futex_keys that hash to the same
233 * location. Each key may have multiple futex_q structures, one for each task
234 * waiting on a futex.
236 struct futex_hash_bucket
{
239 struct plist_head chain
;
240 } ____cacheline_aligned_in_smp
;
243 * The base of the bucket array and its size are always used together
244 * (after initialization only in hash_futex()), so ensure that they
245 * reside in the same cacheline.
248 struct futex_hash_bucket
*queues
;
249 unsigned long hashsize
;
250 } __futex_data __read_mostly
__aligned(2*sizeof(long));
251 #define futex_queues (__futex_data.queues)
252 #define futex_hashsize (__futex_data.hashsize)
256 * Fault injections for futexes.
258 #ifdef CONFIG_FAIL_FUTEX
261 struct fault_attr attr
;
265 .attr
= FAULT_ATTR_INITIALIZER
,
266 .ignore_private
= false,
269 static int __init
setup_fail_futex(char *str
)
271 return setup_fault_attr(&fail_futex
.attr
, str
);
273 __setup("fail_futex=", setup_fail_futex
);
275 static bool should_fail_futex(bool fshared
)
277 if (fail_futex
.ignore_private
&& !fshared
)
280 return should_fail(&fail_futex
.attr
, 1);
283 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
285 static int __init
fail_futex_debugfs(void)
287 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
290 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
295 debugfs_create_bool("ignore-private", mode
, dir
,
296 &fail_futex
.ignore_private
);
300 late_initcall(fail_futex_debugfs
);
302 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
305 static inline bool should_fail_futex(bool fshared
)
309 #endif /* CONFIG_FAIL_FUTEX */
312 static void compat_exit_robust_list(struct task_struct
*curr
);
314 static inline void compat_exit_robust_list(struct task_struct
*curr
) { }
318 * Reflects a new waiter being added to the waitqueue.
320 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
323 atomic_inc(&hb
->waiters
);
325 * Full barrier (A), see the ordering comment above.
327 smp_mb__after_atomic();
332 * Reflects a waiter being removed from the waitqueue by wakeup
335 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
338 atomic_dec(&hb
->waiters
);
342 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
346 * Full barrier (B), see the ordering comment above.
349 return atomic_read(&hb
->waiters
);
356 * hash_futex - Return the hash bucket in the global hash
357 * @key: Pointer to the futex key for which the hash is calculated
359 * We hash on the keys returned from get_futex_key (see below) and return the
360 * corresponding hash bucket in the global hash.
362 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
364 u32 hash
= jhash2((u32
*)key
, offsetof(typeof(*key
), both
.offset
) / 4,
367 return &futex_queues
[hash
& (futex_hashsize
- 1)];
372 * match_futex - Check whether two futex keys are equal
373 * @key1: Pointer to key1
374 * @key2: Pointer to key2
376 * Return 1 if two futex_keys are equal, 0 otherwise.
378 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
381 && key1
->both
.word
== key2
->both
.word
382 && key1
->both
.ptr
== key2
->both
.ptr
383 && key1
->both
.offset
== key2
->both
.offset
);
392 * futex_setup_timer - set up the sleeping hrtimer.
393 * @time: ptr to the given timeout value
394 * @timeout: the hrtimer_sleeper structure to be set up
395 * @flags: futex flags
396 * @range_ns: optional range in ns
398 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
401 static inline struct hrtimer_sleeper
*
402 futex_setup_timer(ktime_t
*time
, struct hrtimer_sleeper
*timeout
,
403 int flags
, u64 range_ns
)
408 hrtimer_init_sleeper_on_stack(timeout
, (flags
& FLAGS_CLOCKRT
) ?
409 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
412 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
413 * effectively the same as calling hrtimer_set_expires().
415 hrtimer_set_expires_range_ns(&timeout
->timer
, *time
, range_ns
);
421 * Generate a machine wide unique identifier for this inode.
423 * This relies on u64 not wrapping in the life-time of the machine; which with
424 * 1ns resolution means almost 585 years.
426 * This further relies on the fact that a well formed program will not unmap
427 * the file while it has a (shared) futex waiting on it. This mapping will have
428 * a file reference which pins the mount and inode.
430 * If for some reason an inode gets evicted and read back in again, it will get
431 * a new sequence number and will _NOT_ match, even though it is the exact same
434 * It is important that match_futex() will never have a false-positive, esp.
435 * for PI futexes that can mess up the state. The above argues that false-negatives
436 * are only possible for malformed programs.
438 static u64
get_inode_sequence_number(struct inode
*inode
)
440 static atomic64_t i_seq
;
443 /* Does the inode already have a sequence number? */
444 old
= atomic64_read(&inode
->i_sequence
);
449 u64
new = atomic64_add_return(1, &i_seq
);
450 if (WARN_ON_ONCE(!new))
453 old
= atomic64_cmpxchg_relaxed(&inode
->i_sequence
, 0, new);
461 * get_futex_key() - Get parameters which are the keys for a futex
462 * @uaddr: virtual address of the futex
463 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
464 * @key: address where result is stored.
465 * @rw: mapping needs to be read/write (values: FUTEX_READ,
468 * Return: a negative error code or 0
470 * The key words are stored in @key on success.
472 * For shared mappings (when @fshared), the key is:
474 * ( inode->i_sequence, page->index, offset_within_page )
476 * [ also see get_inode_sequence_number() ]
478 * For private mappings (or when !@fshared), the key is:
480 * ( current->mm, address, 0 )
482 * This allows (cross process, where applicable) identification of the futex
483 * without keeping the page pinned for the duration of the FUTEX_WAIT.
485 * lock_page() might sleep, the caller should not hold a spinlock.
487 static int get_futex_key(u32 __user
*uaddr
, bool fshared
, union futex_key
*key
,
488 enum futex_access rw
)
490 unsigned long address
= (unsigned long)uaddr
;
491 struct mm_struct
*mm
= current
->mm
;
492 struct page
*page
, *tail
;
493 struct address_space
*mapping
;
497 * The futex address must be "naturally" aligned.
499 key
->both
.offset
= address
% PAGE_SIZE
;
500 if (unlikely((address
% sizeof(u32
)) != 0))
502 address
-= key
->both
.offset
;
504 if (unlikely(!access_ok(uaddr
, sizeof(u32
))))
507 if (unlikely(should_fail_futex(fshared
)))
511 * PROCESS_PRIVATE futexes are fast.
512 * As the mm cannot disappear under us and the 'key' only needs
513 * virtual address, we dont even have to find the underlying vma.
514 * Note : We do have to check 'uaddr' is a valid user address,
515 * but access_ok() should be faster than find_vma()
518 key
->private.mm
= mm
;
519 key
->private.address
= address
;
524 /* Ignore any VERIFY_READ mapping (futex common case) */
525 if (unlikely(should_fail_futex(true)))
528 err
= get_user_pages_fast(address
, 1, FOLL_WRITE
, &page
);
530 * If write access is not required (eg. FUTEX_WAIT), try
531 * and get read-only access.
533 if (err
== -EFAULT
&& rw
== FUTEX_READ
) {
534 err
= get_user_pages_fast(address
, 1, 0, &page
);
543 * The treatment of mapping from this point on is critical. The page
544 * lock protects many things but in this context the page lock
545 * stabilizes mapping, prevents inode freeing in the shared
546 * file-backed region case and guards against movement to swap cache.
548 * Strictly speaking the page lock is not needed in all cases being
549 * considered here and page lock forces unnecessarily serialization
550 * From this point on, mapping will be re-verified if necessary and
551 * page lock will be acquired only if it is unavoidable
553 * Mapping checks require the head page for any compound page so the
554 * head page and mapping is looked up now. For anonymous pages, it
555 * does not matter if the page splits in the future as the key is
556 * based on the address. For filesystem-backed pages, the tail is
557 * required as the index of the page determines the key. For
558 * base pages, there is no tail page and tail == page.
561 page
= compound_head(page
);
562 mapping
= READ_ONCE(page
->mapping
);
565 * If page->mapping is NULL, then it cannot be a PageAnon
566 * page; but it might be the ZERO_PAGE or in the gate area or
567 * in a special mapping (all cases which we are happy to fail);
568 * or it may have been a good file page when get_user_pages_fast
569 * found it, but truncated or holepunched or subjected to
570 * invalidate_complete_page2 before we got the page lock (also
571 * cases which we are happy to fail). And we hold a reference,
572 * so refcount care in invalidate_complete_page's remove_mapping
573 * prevents drop_caches from setting mapping to NULL beneath us.
575 * The case we do have to guard against is when memory pressure made
576 * shmem_writepage move it from filecache to swapcache beneath us:
577 * an unlikely race, but we do need to retry for page->mapping.
579 if (unlikely(!mapping
)) {
583 * Page lock is required to identify which special case above
584 * applies. If this is really a shmem page then the page lock
585 * will prevent unexpected transitions.
588 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
599 * Private mappings are handled in a simple way.
601 * If the futex key is stored on an anonymous page, then the associated
602 * object is the mm which is implicitly pinned by the calling process.
604 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
605 * it's a read-only handle, it's expected that futexes attach to
606 * the object not the particular process.
608 if (PageAnon(page
)) {
610 * A RO anonymous page will never change and thus doesn't make
611 * sense for futex operations.
613 if (unlikely(should_fail_futex(true)) || ro
) {
618 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
619 key
->private.mm
= mm
;
620 key
->private.address
= address
;
626 * The associated futex object in this case is the inode and
627 * the page->mapping must be traversed. Ordinarily this should
628 * be stabilised under page lock but it's not strictly
629 * necessary in this case as we just want to pin the inode, not
630 * update the radix tree or anything like that.
632 * The RCU read lock is taken as the inode is finally freed
633 * under RCU. If the mapping still matches expectations then the
634 * mapping->host can be safely accessed as being a valid inode.
638 if (READ_ONCE(page
->mapping
) != mapping
) {
645 inode
= READ_ONCE(mapping
->host
);
653 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
654 key
->shared
.i_seq
= get_inode_sequence_number(inode
);
655 key
->shared
.pgoff
= basepage_index(tail
);
665 * fault_in_user_writeable() - Fault in user address and verify RW access
666 * @uaddr: pointer to faulting user space address
668 * Slow path to fixup the fault we just took in the atomic write
671 * We have no generic implementation of a non-destructive write to the
672 * user address. We know that we faulted in the atomic pagefault
673 * disabled section so we can as well avoid the #PF overhead by
674 * calling get_user_pages() right away.
676 static int fault_in_user_writeable(u32 __user
*uaddr
)
678 struct mm_struct
*mm
= current
->mm
;
682 ret
= fixup_user_fault(mm
, (unsigned long)uaddr
,
683 FAULT_FLAG_WRITE
, NULL
);
684 mmap_read_unlock(mm
);
686 return ret
< 0 ? ret
: 0;
690 * futex_top_waiter() - Return the highest priority waiter on a futex
691 * @hb: the hash bucket the futex_q's reside in
692 * @key: the futex key (to distinguish it from other futex futex_q's)
694 * Must be called with the hb lock held.
696 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
697 union futex_key
*key
)
699 struct futex_q
*this;
701 plist_for_each_entry(this, &hb
->chain
, list
) {
702 if (match_futex(&this->key
, key
))
708 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
709 u32 uval
, u32 newval
)
714 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
720 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
725 ret
= __get_user(*dest
, from
);
728 return ret
? -EFAULT
: 0;
735 static int refill_pi_state_cache(void)
737 struct futex_pi_state
*pi_state
;
739 if (likely(current
->pi_state_cache
))
742 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
747 INIT_LIST_HEAD(&pi_state
->list
);
748 /* pi_mutex gets initialized later */
749 pi_state
->owner
= NULL
;
750 refcount_set(&pi_state
->refcount
, 1);
751 pi_state
->key
= FUTEX_KEY_INIT
;
753 current
->pi_state_cache
= pi_state
;
758 static struct futex_pi_state
*alloc_pi_state(void)
760 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
763 current
->pi_state_cache
= NULL
;
768 static void get_pi_state(struct futex_pi_state
*pi_state
)
770 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state
->refcount
));
774 * Drops a reference to the pi_state object and frees or caches it
775 * when the last reference is gone.
777 static void put_pi_state(struct futex_pi_state
*pi_state
)
782 if (!refcount_dec_and_test(&pi_state
->refcount
))
786 * If pi_state->owner is NULL, the owner is most probably dying
787 * and has cleaned up the pi_state already
789 if (pi_state
->owner
) {
790 struct task_struct
*owner
;
793 raw_spin_lock_irqsave(&pi_state
->pi_mutex
.wait_lock
, flags
);
794 owner
= pi_state
->owner
;
796 raw_spin_lock(&owner
->pi_lock
);
797 list_del_init(&pi_state
->list
);
798 raw_spin_unlock(&owner
->pi_lock
);
800 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
801 raw_spin_unlock_irqrestore(&pi_state
->pi_mutex
.wait_lock
, flags
);
804 if (current
->pi_state_cache
) {
808 * pi_state->list is already empty.
809 * clear pi_state->owner.
810 * refcount is at 0 - put it back to 1.
812 pi_state
->owner
= NULL
;
813 refcount_set(&pi_state
->refcount
, 1);
814 current
->pi_state_cache
= pi_state
;
818 #ifdef CONFIG_FUTEX_PI
821 * This task is holding PI mutexes at exit time => bad.
822 * Kernel cleans up PI-state, but userspace is likely hosed.
823 * (Robust-futex cleanup is separate and might save the day for userspace.)
825 static void exit_pi_state_list(struct task_struct
*curr
)
827 struct list_head
*next
, *head
= &curr
->pi_state_list
;
828 struct futex_pi_state
*pi_state
;
829 struct futex_hash_bucket
*hb
;
830 union futex_key key
= FUTEX_KEY_INIT
;
832 if (!futex_cmpxchg_enabled
)
835 * We are a ZOMBIE and nobody can enqueue itself on
836 * pi_state_list anymore, but we have to be careful
837 * versus waiters unqueueing themselves:
839 raw_spin_lock_irq(&curr
->pi_lock
);
840 while (!list_empty(head
)) {
842 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
844 hb
= hash_futex(&key
);
847 * We can race against put_pi_state() removing itself from the
848 * list (a waiter going away). put_pi_state() will first
849 * decrement the reference count and then modify the list, so
850 * its possible to see the list entry but fail this reference
853 * In that case; drop the locks to let put_pi_state() make
854 * progress and retry the loop.
856 if (!refcount_inc_not_zero(&pi_state
->refcount
)) {
857 raw_spin_unlock_irq(&curr
->pi_lock
);
859 raw_spin_lock_irq(&curr
->pi_lock
);
862 raw_spin_unlock_irq(&curr
->pi_lock
);
864 spin_lock(&hb
->lock
);
865 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
866 raw_spin_lock(&curr
->pi_lock
);
868 * We dropped the pi-lock, so re-check whether this
869 * task still owns the PI-state:
871 if (head
->next
!= next
) {
872 /* retain curr->pi_lock for the loop invariant */
873 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
874 spin_unlock(&hb
->lock
);
875 put_pi_state(pi_state
);
879 WARN_ON(pi_state
->owner
!= curr
);
880 WARN_ON(list_empty(&pi_state
->list
));
881 list_del_init(&pi_state
->list
);
882 pi_state
->owner
= NULL
;
884 raw_spin_unlock(&curr
->pi_lock
);
885 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
886 spin_unlock(&hb
->lock
);
888 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
889 put_pi_state(pi_state
);
891 raw_spin_lock_irq(&curr
->pi_lock
);
893 raw_spin_unlock_irq(&curr
->pi_lock
);
896 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
900 * We need to check the following states:
902 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
904 * [1] NULL | --- | --- | 0 | 0/1 | Valid
905 * [2] NULL | --- | --- | >0 | 0/1 | Valid
907 * [3] Found | NULL | -- | Any | 0/1 | Invalid
909 * [4] Found | Found | NULL | 0 | 1 | Valid
910 * [5] Found | Found | NULL | >0 | 1 | Invalid
912 * [6] Found | Found | task | 0 | 1 | Valid
914 * [7] Found | Found | NULL | Any | 0 | Invalid
916 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
917 * [9] Found | Found | task | 0 | 0 | Invalid
918 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
920 * [1] Indicates that the kernel can acquire the futex atomically. We
921 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
923 * [2] Valid, if TID does not belong to a kernel thread. If no matching
924 * thread is found then it indicates that the owner TID has died.
926 * [3] Invalid. The waiter is queued on a non PI futex
928 * [4] Valid state after exit_robust_list(), which sets the user space
929 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
931 * [5] The user space value got manipulated between exit_robust_list()
932 * and exit_pi_state_list()
934 * [6] Valid state after exit_pi_state_list() which sets the new owner in
935 * the pi_state but cannot access the user space value.
937 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
939 * [8] Owner and user space value match
941 * [9] There is no transient state which sets the user space TID to 0
942 * except exit_robust_list(), but this is indicated by the
943 * FUTEX_OWNER_DIED bit. See [4]
945 * [10] There is no transient state which leaves owner and user space
949 * Serialization and lifetime rules:
953 * hb -> futex_q, relation
954 * futex_q -> pi_state, relation
956 * (cannot be raw because hb can contain arbitrary amount
959 * pi_mutex->wait_lock:
963 * (and pi_mutex 'obviously')
967 * p->pi_state_list -> pi_state->list, relation
969 * pi_state->refcount:
977 * pi_mutex->wait_lock
983 * Validate that the existing waiter has a pi_state and sanity check
984 * the pi_state against the user space value. If correct, attach to
987 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
988 struct futex_pi_state
*pi_state
,
989 struct futex_pi_state
**ps
)
991 pid_t pid
= uval
& FUTEX_TID_MASK
;
996 * Userspace might have messed up non-PI and PI futexes [3]
998 if (unlikely(!pi_state
))
1002 * We get here with hb->lock held, and having found a
1003 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1004 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1005 * which in turn means that futex_lock_pi() still has a reference on
1008 * The waiter holding a reference on @pi_state also protects against
1009 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1010 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1011 * free pi_state before we can take a reference ourselves.
1013 WARN_ON(!refcount_read(&pi_state
->refcount
));
1016 * Now that we have a pi_state, we can acquire wait_lock
1017 * and do the state validation.
1019 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1022 * Since {uval, pi_state} is serialized by wait_lock, and our current
1023 * uval was read without holding it, it can have changed. Verify it
1024 * still is what we expect it to be, otherwise retry the entire
1027 if (get_futex_value_locked(&uval2
, uaddr
))
1034 * Handle the owner died case:
1036 if (uval
& FUTEX_OWNER_DIED
) {
1038 * exit_pi_state_list sets owner to NULL and wakes the
1039 * topmost waiter. The task which acquires the
1040 * pi_state->rt_mutex will fixup owner.
1042 if (!pi_state
->owner
) {
1044 * No pi state owner, but the user space TID
1045 * is not 0. Inconsistent state. [5]
1050 * Take a ref on the state and return success. [4]
1056 * If TID is 0, then either the dying owner has not
1057 * yet executed exit_pi_state_list() or some waiter
1058 * acquired the rtmutex in the pi state, but did not
1059 * yet fixup the TID in user space.
1061 * Take a ref on the state and return success. [6]
1067 * If the owner died bit is not set, then the pi_state
1068 * must have an owner. [7]
1070 if (!pi_state
->owner
)
1075 * Bail out if user space manipulated the futex value. If pi
1076 * state exists then the owner TID must be the same as the
1077 * user space TID. [9/10]
1079 if (pid
!= task_pid_vnr(pi_state
->owner
))
1083 get_pi_state(pi_state
);
1084 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1101 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1106 * wait_for_owner_exiting - Block until the owner has exited
1107 * @ret: owner's current futex lock status
1108 * @exiting: Pointer to the exiting task
1110 * Caller must hold a refcount on @exiting.
1112 static void wait_for_owner_exiting(int ret
, struct task_struct
*exiting
)
1114 if (ret
!= -EBUSY
) {
1115 WARN_ON_ONCE(exiting
);
1119 if (WARN_ON_ONCE(ret
== -EBUSY
&& !exiting
))
1122 mutex_lock(&exiting
->futex_exit_mutex
);
1124 * No point in doing state checking here. If the waiter got here
1125 * while the task was in exec()->exec_futex_release() then it can
1126 * have any FUTEX_STATE_* value when the waiter has acquired the
1127 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1128 * already. Highly unlikely and not a problem. Just one more round
1129 * through the futex maze.
1131 mutex_unlock(&exiting
->futex_exit_mutex
);
1133 put_task_struct(exiting
);
1136 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1137 struct task_struct
*tsk
)
1142 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1143 * caller that the alleged owner is busy.
1145 if (tsk
&& tsk
->futex_state
!= FUTEX_STATE_DEAD
)
1149 * Reread the user space value to handle the following situation:
1153 * sys_exit() sys_futex()
1154 * do_exit() futex_lock_pi()
1155 * futex_lock_pi_atomic()
1156 * exit_signals(tsk) No waiters:
1157 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1158 * mm_release(tsk) Set waiter bit
1159 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1160 * Set owner died attach_to_pi_owner() {
1161 * *uaddr = 0xC0000000; tsk = get_task(PID);
1162 * } if (!tsk->flags & PF_EXITING) {
1164 * tsk->futex_state = } else {
1165 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1168 * return -ESRCH; <--- FAIL
1171 * Returning ESRCH unconditionally is wrong here because the
1172 * user space value has been changed by the exiting task.
1174 * The same logic applies to the case where the exiting task is
1177 if (get_futex_value_locked(&uval2
, uaddr
))
1180 /* If the user space value has changed, try again. */
1185 * The exiting task did not have a robust list, the robust list was
1186 * corrupted or the user space value in *uaddr is simply bogus.
1187 * Give up and tell user space.
1193 * Lookup the task for the TID provided from user space and attach to
1194 * it after doing proper sanity checks.
1196 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1197 struct futex_pi_state
**ps
,
1198 struct task_struct
**exiting
)
1200 pid_t pid
= uval
& FUTEX_TID_MASK
;
1201 struct futex_pi_state
*pi_state
;
1202 struct task_struct
*p
;
1205 * We are the first waiter - try to look up the real owner and attach
1206 * the new pi_state to it, but bail out when TID = 0 [1]
1208 * The !pid check is paranoid. None of the call sites should end up
1209 * with pid == 0, but better safe than sorry. Let the caller retry
1213 p
= find_get_task_by_vpid(pid
);
1215 return handle_exit_race(uaddr
, uval
, NULL
);
1217 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1223 * We need to look at the task state to figure out, whether the
1224 * task is exiting. To protect against the change of the task state
1225 * in futex_exit_release(), we do this protected by p->pi_lock:
1227 raw_spin_lock_irq(&p
->pi_lock
);
1228 if (unlikely(p
->futex_state
!= FUTEX_STATE_OK
)) {
1230 * The task is on the way out. When the futex state is
1231 * FUTEX_STATE_DEAD, we know that the task has finished
1234 int ret
= handle_exit_race(uaddr
, uval
, p
);
1236 raw_spin_unlock_irq(&p
->pi_lock
);
1238 * If the owner task is between FUTEX_STATE_EXITING and
1239 * FUTEX_STATE_DEAD then store the task pointer and keep
1240 * the reference on the task struct. The calling code will
1241 * drop all locks, wait for the task to reach
1242 * FUTEX_STATE_DEAD and then drop the refcount. This is
1243 * required to prevent a live lock when the current task
1244 * preempted the exiting task between the two states.
1254 * No existing pi state. First waiter. [2]
1256 * This creates pi_state, we have hb->lock held, this means nothing can
1257 * observe this state, wait_lock is irrelevant.
1259 pi_state
= alloc_pi_state();
1262 * Initialize the pi_mutex in locked state and make @p
1265 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1267 /* Store the key for possible exit cleanups: */
1268 pi_state
->key
= *key
;
1270 WARN_ON(!list_empty(&pi_state
->list
));
1271 list_add(&pi_state
->list
, &p
->pi_state_list
);
1273 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1274 * because there is no concurrency as the object is not published yet.
1276 pi_state
->owner
= p
;
1277 raw_spin_unlock_irq(&p
->pi_lock
);
1286 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1287 struct futex_hash_bucket
*hb
,
1288 union futex_key
*key
, struct futex_pi_state
**ps
,
1289 struct task_struct
**exiting
)
1291 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1294 * If there is a waiter on that futex, validate it and
1295 * attach to the pi_state when the validation succeeds.
1298 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1301 * We are the first waiter - try to look up the owner based on
1302 * @uval and attach to it.
1304 return attach_to_pi_owner(uaddr
, uval
, key
, ps
, exiting
);
1307 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1312 if (unlikely(should_fail_futex(true)))
1315 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1319 /* If user space value changed, let the caller retry */
1320 return curval
!= uval
? -EAGAIN
: 0;
1324 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1325 * @uaddr: the pi futex user address
1326 * @hb: the pi futex hash bucket
1327 * @key: the futex key associated with uaddr and hb
1328 * @ps: the pi_state pointer where we store the result of the
1330 * @task: the task to perform the atomic lock work for. This will
1331 * be "current" except in the case of requeue pi.
1332 * @exiting: Pointer to store the task pointer of the owner task
1333 * which is in the middle of exiting
1334 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1337 * - 0 - ready to wait;
1338 * - 1 - acquired the lock;
1341 * The hb->lock and futex_key refs shall be held by the caller.
1343 * @exiting is only set when the return value is -EBUSY. If so, this holds
1344 * a refcount on the exiting task on return and the caller needs to drop it
1345 * after waiting for the exit to complete.
1347 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1348 union futex_key
*key
,
1349 struct futex_pi_state
**ps
,
1350 struct task_struct
*task
,
1351 struct task_struct
**exiting
,
1354 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1355 struct futex_q
*top_waiter
;
1359 * Read the user space value first so we can validate a few
1360 * things before proceeding further.
1362 if (get_futex_value_locked(&uval
, uaddr
))
1365 if (unlikely(should_fail_futex(true)))
1371 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1374 if ((unlikely(should_fail_futex(true))))
1378 * Lookup existing state first. If it exists, try to attach to
1381 top_waiter
= futex_top_waiter(hb
, key
);
1383 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1386 * No waiter and user TID is 0. We are here because the
1387 * waiters or the owner died bit is set or called from
1388 * requeue_cmp_pi or for whatever reason something took the
1391 if (!(uval
& FUTEX_TID_MASK
)) {
1393 * We take over the futex. No other waiters and the user space
1394 * TID is 0. We preserve the owner died bit.
1396 newval
= uval
& FUTEX_OWNER_DIED
;
1399 /* The futex requeue_pi code can enforce the waiters bit */
1401 newval
|= FUTEX_WAITERS
;
1403 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1404 /* If the take over worked, return 1 */
1405 return ret
< 0 ? ret
: 1;
1409 * First waiter. Set the waiters bit before attaching ourself to
1410 * the owner. If owner tries to unlock, it will be forced into
1411 * the kernel and blocked on hb->lock.
1413 newval
= uval
| FUTEX_WAITERS
;
1414 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1418 * If the update of the user space value succeeded, we try to
1419 * attach to the owner. If that fails, no harm done, we only
1420 * set the FUTEX_WAITERS bit in the user space variable.
1422 return attach_to_pi_owner(uaddr
, newval
, key
, ps
, exiting
);
1426 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1427 * @q: The futex_q to unqueue
1429 * The q->lock_ptr must not be NULL and must be held by the caller.
1431 static void __unqueue_futex(struct futex_q
*q
)
1433 struct futex_hash_bucket
*hb
;
1435 if (WARN_ON_SMP(!q
->lock_ptr
) || WARN_ON(plist_node_empty(&q
->list
)))
1437 lockdep_assert_held(q
->lock_ptr
);
1439 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1440 plist_del(&q
->list
, &hb
->chain
);
1445 * The hash bucket lock must be held when this is called.
1446 * Afterwards, the futex_q must not be accessed. Callers
1447 * must ensure to later call wake_up_q() for the actual
1450 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1452 struct task_struct
*p
= q
->task
;
1454 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1460 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1461 * is written, without taking any locks. This is possible in the event
1462 * of a spurious wakeup, for example. A memory barrier is required here
1463 * to prevent the following store to lock_ptr from getting ahead of the
1464 * plist_del in __unqueue_futex().
1466 smp_store_release(&q
->lock_ptr
, NULL
);
1469 * Queue the task for later wakeup for after we've released
1472 wake_q_add_safe(wake_q
, p
);
1476 * Caller must hold a reference on @pi_state.
1478 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1481 struct task_struct
*new_owner
;
1482 bool postunlock
= false;
1483 DEFINE_WAKE_Q(wake_q
);
1486 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1487 if (WARN_ON_ONCE(!new_owner
)) {
1489 * As per the comment in futex_unlock_pi() this should not happen.
1491 * When this happens, give up our locks and try again, giving
1492 * the futex_lock_pi() instance time to complete, either by
1493 * waiting on the rtmutex or removing itself from the futex
1501 * We pass it to the next owner. The WAITERS bit is always kept
1502 * enabled while there is PI state around. We cleanup the owner
1503 * died bit, because we are the owner.
1505 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1507 if (unlikely(should_fail_futex(true))) {
1512 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1513 if (!ret
&& (curval
!= uval
)) {
1515 * If a unconditional UNLOCK_PI operation (user space did not
1516 * try the TID->0 transition) raced with a waiter setting the
1517 * FUTEX_WAITERS flag between get_user() and locking the hash
1518 * bucket lock, retry the operation.
1520 if ((FUTEX_TID_MASK
& curval
) == uval
)
1530 * This is a point of no return; once we modify the uval there is no
1531 * going back and subsequent operations must not fail.
1534 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1535 WARN_ON(list_empty(&pi_state
->list
));
1536 list_del_init(&pi_state
->list
);
1537 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1539 raw_spin_lock(&new_owner
->pi_lock
);
1540 WARN_ON(!list_empty(&pi_state
->list
));
1541 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1542 pi_state
->owner
= new_owner
;
1543 raw_spin_unlock(&new_owner
->pi_lock
);
1545 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1548 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1551 rt_mutex_postunlock(&wake_q
);
1557 * Express the locking dependencies for lockdep:
1560 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1563 spin_lock(&hb1
->lock
);
1565 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1566 } else { /* hb1 > hb2 */
1567 spin_lock(&hb2
->lock
);
1568 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1573 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1575 spin_unlock(&hb1
->lock
);
1577 spin_unlock(&hb2
->lock
);
1581 * Wake up waiters matching bitset queued on this futex (uaddr).
1584 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1586 struct futex_hash_bucket
*hb
;
1587 struct futex_q
*this, *next
;
1588 union futex_key key
= FUTEX_KEY_INIT
;
1590 DEFINE_WAKE_Q(wake_q
);
1595 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_READ
);
1596 if (unlikely(ret
!= 0))
1599 hb
= hash_futex(&key
);
1601 /* Make sure we really have tasks to wakeup */
1602 if (!hb_waiters_pending(hb
))
1605 spin_lock(&hb
->lock
);
1607 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1608 if (match_futex (&this->key
, &key
)) {
1609 if (this->pi_state
|| this->rt_waiter
) {
1614 /* Check if one of the bits is set in both bitsets */
1615 if (!(this->bitset
& bitset
))
1618 mark_wake_futex(&wake_q
, this);
1619 if (++ret
>= nr_wake
)
1624 spin_unlock(&hb
->lock
);
1629 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1631 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1632 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1633 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1634 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1637 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1638 if (oparg
< 0 || oparg
> 31) {
1639 char comm
[sizeof(current
->comm
)];
1641 * kill this print and return -EINVAL when userspace
1644 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1645 get_task_comm(comm
, current
), oparg
);
1651 pagefault_disable();
1652 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1658 case FUTEX_OP_CMP_EQ
:
1659 return oldval
== cmparg
;
1660 case FUTEX_OP_CMP_NE
:
1661 return oldval
!= cmparg
;
1662 case FUTEX_OP_CMP_LT
:
1663 return oldval
< cmparg
;
1664 case FUTEX_OP_CMP_GE
:
1665 return oldval
>= cmparg
;
1666 case FUTEX_OP_CMP_LE
:
1667 return oldval
<= cmparg
;
1668 case FUTEX_OP_CMP_GT
:
1669 return oldval
> cmparg
;
1676 * Wake up all waiters hashed on the physical page that is mapped
1677 * to this virtual address:
1680 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1681 int nr_wake
, int nr_wake2
, int op
)
1683 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1684 struct futex_hash_bucket
*hb1
, *hb2
;
1685 struct futex_q
*this, *next
;
1687 DEFINE_WAKE_Q(wake_q
);
1690 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1691 if (unlikely(ret
!= 0))
1693 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
1694 if (unlikely(ret
!= 0))
1697 hb1
= hash_futex(&key1
);
1698 hb2
= hash_futex(&key2
);
1701 double_lock_hb(hb1
, hb2
);
1702 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1703 if (unlikely(op_ret
< 0)) {
1704 double_unlock_hb(hb1
, hb2
);
1706 if (!IS_ENABLED(CONFIG_MMU
) ||
1707 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1709 * we don't get EFAULT from MMU faults if we don't have
1710 * an MMU, but we might get them from range checking
1716 if (op_ret
== -EFAULT
) {
1717 ret
= fault_in_user_writeable(uaddr2
);
1722 if (!(flags
& FLAGS_SHARED
)) {
1731 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1732 if (match_futex (&this->key
, &key1
)) {
1733 if (this->pi_state
|| this->rt_waiter
) {
1737 mark_wake_futex(&wake_q
, this);
1738 if (++ret
>= nr_wake
)
1745 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1746 if (match_futex (&this->key
, &key2
)) {
1747 if (this->pi_state
|| this->rt_waiter
) {
1751 mark_wake_futex(&wake_q
, this);
1752 if (++op_ret
>= nr_wake2
)
1760 double_unlock_hb(hb1
, hb2
);
1766 * requeue_futex() - Requeue a futex_q from one hb to another
1767 * @q: the futex_q to requeue
1768 * @hb1: the source hash_bucket
1769 * @hb2: the target hash_bucket
1770 * @key2: the new key for the requeued futex_q
1773 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1774 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1778 * If key1 and key2 hash to the same bucket, no need to
1781 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1782 plist_del(&q
->list
, &hb1
->chain
);
1783 hb_waiters_dec(hb1
);
1784 hb_waiters_inc(hb2
);
1785 plist_add(&q
->list
, &hb2
->chain
);
1786 q
->lock_ptr
= &hb2
->lock
;
1792 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1794 * @key: the key of the requeue target futex
1795 * @hb: the hash_bucket of the requeue target futex
1797 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1798 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1799 * to the requeue target futex so the waiter can detect the wakeup on the right
1800 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1801 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1802 * to protect access to the pi_state to fixup the owner later. Must be called
1803 * with both q->lock_ptr and hb->lock held.
1806 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1807 struct futex_hash_bucket
*hb
)
1813 WARN_ON(!q
->rt_waiter
);
1814 q
->rt_waiter
= NULL
;
1816 q
->lock_ptr
= &hb
->lock
;
1818 wake_up_state(q
->task
, TASK_NORMAL
);
1822 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1823 * @pifutex: the user address of the to futex
1824 * @hb1: the from futex hash bucket, must be locked by the caller
1825 * @hb2: the to futex hash bucket, must be locked by the caller
1826 * @key1: the from futex key
1827 * @key2: the to futex key
1828 * @ps: address to store the pi_state pointer
1829 * @exiting: Pointer to store the task pointer of the owner task
1830 * which is in the middle of exiting
1831 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1833 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1834 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1835 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1836 * hb1 and hb2 must be held by the caller.
1838 * @exiting is only set when the return value is -EBUSY. If so, this holds
1839 * a refcount on the exiting task on return and the caller needs to drop it
1840 * after waiting for the exit to complete.
1843 * - 0 - failed to acquire the lock atomically;
1844 * - >0 - acquired the lock, return value is vpid of the top_waiter
1848 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1849 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1850 union futex_key
*key2
, struct futex_pi_state
**ps
,
1851 struct task_struct
**exiting
, int set_waiters
)
1853 struct futex_q
*top_waiter
= NULL
;
1857 if (get_futex_value_locked(&curval
, pifutex
))
1860 if (unlikely(should_fail_futex(true)))
1864 * Find the top_waiter and determine if there are additional waiters.
1865 * If the caller intends to requeue more than 1 waiter to pifutex,
1866 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1867 * as we have means to handle the possible fault. If not, don't set
1868 * the bit unecessarily as it will force the subsequent unlock to enter
1871 top_waiter
= futex_top_waiter(hb1
, key1
);
1873 /* There are no waiters, nothing for us to do. */
1877 /* Ensure we requeue to the expected futex. */
1878 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1882 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1883 * the contended case or if set_waiters is 1. The pi_state is returned
1884 * in ps in contended cases.
1886 vpid
= task_pid_vnr(top_waiter
->task
);
1887 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1888 exiting
, set_waiters
);
1890 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1897 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1898 * @uaddr1: source futex user address
1899 * @flags: futex flags (FLAGS_SHARED, etc.)
1900 * @uaddr2: target futex user address
1901 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1902 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1903 * @cmpval: @uaddr1 expected value (or %NULL)
1904 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1905 * pi futex (pi to pi requeue is not supported)
1907 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1908 * uaddr2 atomically on behalf of the top waiter.
1911 * - >=0 - on success, the number of tasks requeued or woken;
1914 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1915 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1916 u32
*cmpval
, int requeue_pi
)
1918 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1919 int task_count
= 0, ret
;
1920 struct futex_pi_state
*pi_state
= NULL
;
1921 struct futex_hash_bucket
*hb1
, *hb2
;
1922 struct futex_q
*this, *next
;
1923 DEFINE_WAKE_Q(wake_q
);
1925 if (nr_wake
< 0 || nr_requeue
< 0)
1929 * When PI not supported: return -ENOSYS if requeue_pi is true,
1930 * consequently the compiler knows requeue_pi is always false past
1931 * this point which will optimize away all the conditional code
1934 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1939 * Requeue PI only works on two distinct uaddrs. This
1940 * check is only valid for private futexes. See below.
1942 if (uaddr1
== uaddr2
)
1946 * requeue_pi requires a pi_state, try to allocate it now
1947 * without any locks in case it fails.
1949 if (refill_pi_state_cache())
1952 * requeue_pi must wake as many tasks as it can, up to nr_wake
1953 * + nr_requeue, since it acquires the rt_mutex prior to
1954 * returning to userspace, so as to not leave the rt_mutex with
1955 * waiters and no owner. However, second and third wake-ups
1956 * cannot be predicted as they involve race conditions with the
1957 * first wake and a fault while looking up the pi_state. Both
1958 * pthread_cond_signal() and pthread_cond_broadcast() should
1966 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1967 if (unlikely(ret
!= 0))
1969 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1970 requeue_pi
? FUTEX_WRITE
: FUTEX_READ
);
1971 if (unlikely(ret
!= 0))
1975 * The check above which compares uaddrs is not sufficient for
1976 * shared futexes. We need to compare the keys:
1978 if (requeue_pi
&& match_futex(&key1
, &key2
))
1981 hb1
= hash_futex(&key1
);
1982 hb2
= hash_futex(&key2
);
1985 hb_waiters_inc(hb2
);
1986 double_lock_hb(hb1
, hb2
);
1988 if (likely(cmpval
!= NULL
)) {
1991 ret
= get_futex_value_locked(&curval
, uaddr1
);
1993 if (unlikely(ret
)) {
1994 double_unlock_hb(hb1
, hb2
);
1995 hb_waiters_dec(hb2
);
1997 ret
= get_user(curval
, uaddr1
);
2001 if (!(flags
& FLAGS_SHARED
))
2006 if (curval
!= *cmpval
) {
2012 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2013 struct task_struct
*exiting
= NULL
;
2016 * Attempt to acquire uaddr2 and wake the top waiter. If we
2017 * intend to requeue waiters, force setting the FUTEX_WAITERS
2018 * bit. We force this here where we are able to easily handle
2019 * faults rather in the requeue loop below.
2021 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2023 &exiting
, nr_requeue
);
2026 * At this point the top_waiter has either taken uaddr2 or is
2027 * waiting on it. If the former, then the pi_state will not
2028 * exist yet, look it up one more time to ensure we have a
2029 * reference to it. If the lock was taken, ret contains the
2030 * vpid of the top waiter task.
2031 * If the lock was not taken, we have pi_state and an initial
2032 * refcount on it. In case of an error we have nothing.
2038 * If we acquired the lock, then the user space value
2039 * of uaddr2 should be vpid. It cannot be changed by
2040 * the top waiter as it is blocked on hb2 lock if it
2041 * tries to do so. If something fiddled with it behind
2042 * our back the pi state lookup might unearth it. So
2043 * we rather use the known value than rereading and
2044 * handing potential crap to lookup_pi_state.
2046 * If that call succeeds then we have pi_state and an
2047 * initial refcount on it.
2049 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2050 &pi_state
, &exiting
);
2055 /* We hold a reference on the pi state. */
2058 /* If the above failed, then pi_state is NULL */
2060 double_unlock_hb(hb1
, hb2
);
2061 hb_waiters_dec(hb2
);
2062 ret
= fault_in_user_writeable(uaddr2
);
2069 * Two reasons for this:
2070 * - EBUSY: Owner is exiting and we just wait for the
2072 * - EAGAIN: The user space value changed.
2074 double_unlock_hb(hb1
, hb2
);
2075 hb_waiters_dec(hb2
);
2077 * Handle the case where the owner is in the middle of
2078 * exiting. Wait for the exit to complete otherwise
2079 * this task might loop forever, aka. live lock.
2081 wait_for_owner_exiting(ret
, exiting
);
2089 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2090 if (task_count
- nr_wake
>= nr_requeue
)
2093 if (!match_futex(&this->key
, &key1
))
2097 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2098 * be paired with each other and no other futex ops.
2100 * We should never be requeueing a futex_q with a pi_state,
2101 * which is awaiting a futex_unlock_pi().
2103 if ((requeue_pi
&& !this->rt_waiter
) ||
2104 (!requeue_pi
&& this->rt_waiter
) ||
2111 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2112 * lock, we already woke the top_waiter. If not, it will be
2113 * woken by futex_unlock_pi().
2115 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2116 mark_wake_futex(&wake_q
, this);
2120 /* Ensure we requeue to the expected futex for requeue_pi. */
2121 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2127 * Requeue nr_requeue waiters and possibly one more in the case
2128 * of requeue_pi if we couldn't acquire the lock atomically.
2132 * Prepare the waiter to take the rt_mutex. Take a
2133 * refcount on the pi_state and store the pointer in
2134 * the futex_q object of the waiter.
2136 get_pi_state(pi_state
);
2137 this->pi_state
= pi_state
;
2138 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2143 * We got the lock. We do neither drop the
2144 * refcount on pi_state nor clear
2145 * this->pi_state because the waiter needs the
2146 * pi_state for cleaning up the user space
2147 * value. It will drop the refcount after
2150 requeue_pi_wake_futex(this, &key2
, hb2
);
2154 * rt_mutex_start_proxy_lock() detected a
2155 * potential deadlock when we tried to queue
2156 * that waiter. Drop the pi_state reference
2157 * which we took above and remove the pointer
2158 * to the state from the waiters futex_q
2161 this->pi_state
= NULL
;
2162 put_pi_state(pi_state
);
2164 * We stop queueing more waiters and let user
2165 * space deal with the mess.
2170 requeue_futex(this, hb1
, hb2
, &key2
);
2174 * We took an extra initial reference to the pi_state either
2175 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2176 * need to drop it here again.
2178 put_pi_state(pi_state
);
2181 double_unlock_hb(hb1
, hb2
);
2183 hb_waiters_dec(hb2
);
2184 return ret
? ret
: task_count
;
2187 /* The key must be already stored in q->key. */
2188 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2189 __acquires(&hb
->lock
)
2191 struct futex_hash_bucket
*hb
;
2193 hb
= hash_futex(&q
->key
);
2196 * Increment the counter before taking the lock so that
2197 * a potential waker won't miss a to-be-slept task that is
2198 * waiting for the spinlock. This is safe as all queue_lock()
2199 * users end up calling queue_me(). Similarly, for housekeeping,
2200 * decrement the counter at queue_unlock() when some error has
2201 * occurred and we don't end up adding the task to the list.
2203 hb_waiters_inc(hb
); /* implies smp_mb(); (A) */
2205 q
->lock_ptr
= &hb
->lock
;
2207 spin_lock(&hb
->lock
);
2212 queue_unlock(struct futex_hash_bucket
*hb
)
2213 __releases(&hb
->lock
)
2215 spin_unlock(&hb
->lock
);
2219 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2224 * The priority used to register this element is
2225 * - either the real thread-priority for the real-time threads
2226 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2227 * - or MAX_RT_PRIO for non-RT threads.
2228 * Thus, all RT-threads are woken first in priority order, and
2229 * the others are woken last, in FIFO order.
2231 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2233 plist_node_init(&q
->list
, prio
);
2234 plist_add(&q
->list
, &hb
->chain
);
2239 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2240 * @q: The futex_q to enqueue
2241 * @hb: The destination hash bucket
2243 * The hb->lock must be held by the caller, and is released here. A call to
2244 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2245 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2246 * or nothing if the unqueue is done as part of the wake process and the unqueue
2247 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2250 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2251 __releases(&hb
->lock
)
2254 spin_unlock(&hb
->lock
);
2258 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2259 * @q: The futex_q to unqueue
2261 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2262 * be paired with exactly one earlier call to queue_me().
2265 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2266 * - 0 - if the futex_q was already removed by the waking thread
2268 static int unqueue_me(struct futex_q
*q
)
2270 spinlock_t
*lock_ptr
;
2273 /* In the common case we don't take the spinlock, which is nice. */
2276 * q->lock_ptr can change between this read and the following spin_lock.
2277 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2278 * optimizing lock_ptr out of the logic below.
2280 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2281 if (lock_ptr
!= NULL
) {
2282 spin_lock(lock_ptr
);
2284 * q->lock_ptr can change between reading it and
2285 * spin_lock(), causing us to take the wrong lock. This
2286 * corrects the race condition.
2288 * Reasoning goes like this: if we have the wrong lock,
2289 * q->lock_ptr must have changed (maybe several times)
2290 * between reading it and the spin_lock(). It can
2291 * change again after the spin_lock() but only if it was
2292 * already changed before the spin_lock(). It cannot,
2293 * however, change back to the original value. Therefore
2294 * we can detect whether we acquired the correct lock.
2296 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2297 spin_unlock(lock_ptr
);
2302 BUG_ON(q
->pi_state
);
2304 spin_unlock(lock_ptr
);
2312 * PI futexes can not be requeued and must remove themself from the
2313 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2316 static void unqueue_me_pi(struct futex_q
*q
)
2317 __releases(q
->lock_ptr
)
2321 BUG_ON(!q
->pi_state
);
2322 put_pi_state(q
->pi_state
);
2325 spin_unlock(q
->lock_ptr
);
2328 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2329 struct task_struct
*argowner
)
2331 struct futex_pi_state
*pi_state
= q
->pi_state
;
2332 u32 uval
, curval
, newval
;
2333 struct task_struct
*oldowner
, *newowner
;
2337 lockdep_assert_held(q
->lock_ptr
);
2339 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2341 oldowner
= pi_state
->owner
;
2344 * We are here because either:
2346 * - we stole the lock and pi_state->owner needs updating to reflect
2347 * that (@argowner == current),
2351 * - someone stole our lock and we need to fix things to point to the
2352 * new owner (@argowner == NULL).
2354 * Either way, we have to replace the TID in the user space variable.
2355 * This must be atomic as we have to preserve the owner died bit here.
2357 * Note: We write the user space value _before_ changing the pi_state
2358 * because we can fault here. Imagine swapped out pages or a fork
2359 * that marked all the anonymous memory readonly for cow.
2361 * Modifying pi_state _before_ the user space value would leave the
2362 * pi_state in an inconsistent state when we fault here, because we
2363 * need to drop the locks to handle the fault. This might be observed
2364 * in the PID check in lookup_pi_state.
2368 if (oldowner
!= current
) {
2370 * We raced against a concurrent self; things are
2371 * already fixed up. Nothing to do.
2377 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2378 /* We got the lock after all, nothing to fix. */
2384 * The trylock just failed, so either there is an owner or
2385 * there is a higher priority waiter than this one.
2387 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2389 * If the higher priority waiter has not yet taken over the
2390 * rtmutex then newowner is NULL. We can't return here with
2391 * that state because it's inconsistent vs. the user space
2392 * state. So drop the locks and try again. It's a valid
2393 * situation and not any different from the other retry
2396 if (unlikely(!newowner
)) {
2401 WARN_ON_ONCE(argowner
!= current
);
2402 if (oldowner
== current
) {
2404 * We raced against a concurrent self; things are
2405 * already fixed up. Nothing to do.
2410 newowner
= argowner
;
2413 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2415 if (!pi_state
->owner
)
2416 newtid
|= FUTEX_OWNER_DIED
;
2418 err
= get_futex_value_locked(&uval
, uaddr
);
2423 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2425 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2435 * We fixed up user space. Now we need to fix the pi_state
2438 if (pi_state
->owner
!= NULL
) {
2439 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2440 WARN_ON(list_empty(&pi_state
->list
));
2441 list_del_init(&pi_state
->list
);
2442 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2445 pi_state
->owner
= newowner
;
2447 raw_spin_lock(&newowner
->pi_lock
);
2448 WARN_ON(!list_empty(&pi_state
->list
));
2449 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2450 raw_spin_unlock(&newowner
->pi_lock
);
2451 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2456 * In order to reschedule or handle a page fault, we need to drop the
2457 * locks here. In the case of a fault, this gives the other task
2458 * (either the highest priority waiter itself or the task which stole
2459 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2460 * are back from handling the fault we need to check the pi_state after
2461 * reacquiring the locks and before trying to do another fixup. When
2462 * the fixup has been done already we simply return.
2464 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2465 * drop hb->lock since the caller owns the hb -> futex_q relation.
2466 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2469 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2470 spin_unlock(q
->lock_ptr
);
2474 ret
= fault_in_user_writeable(uaddr
);
2488 spin_lock(q
->lock_ptr
);
2489 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2492 * Check if someone else fixed it for us:
2494 if (pi_state
->owner
!= oldowner
) {
2505 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2509 static long futex_wait_restart(struct restart_block
*restart
);
2512 * fixup_owner() - Post lock pi_state and corner case management
2513 * @uaddr: user address of the futex
2514 * @q: futex_q (contains pi_state and access to the rt_mutex)
2515 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2517 * After attempting to lock an rt_mutex, this function is called to cleanup
2518 * the pi_state owner as well as handle race conditions that may allow us to
2519 * acquire the lock. Must be called with the hb lock held.
2522 * - 1 - success, lock taken;
2523 * - 0 - success, lock not taken;
2524 * - <0 - on error (-EFAULT)
2526 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2532 * Got the lock. We might not be the anticipated owner if we
2533 * did a lock-steal - fix up the PI-state in that case:
2535 * Speculative pi_state->owner read (we don't hold wait_lock);
2536 * since we own the lock pi_state->owner == current is the
2537 * stable state, anything else needs more attention.
2539 if (q
->pi_state
->owner
!= current
)
2540 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2541 return ret
? ret
: locked
;
2545 * If we didn't get the lock; check if anybody stole it from us. In
2546 * that case, we need to fix up the uval to point to them instead of
2547 * us, otherwise bad things happen. [10]
2549 * Another speculative read; pi_state->owner == current is unstable
2550 * but needs our attention.
2552 if (q
->pi_state
->owner
== current
) {
2553 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2558 * Paranoia check. If we did not take the lock, then we should not be
2559 * the owner of the rt_mutex.
2561 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2562 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2563 "pi-state %p\n", ret
,
2564 q
->pi_state
->pi_mutex
.owner
,
2565 q
->pi_state
->owner
);
2572 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2573 * @hb: the futex hash bucket, must be locked by the caller
2574 * @q: the futex_q to queue up on
2575 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2577 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2578 struct hrtimer_sleeper
*timeout
)
2581 * The task state is guaranteed to be set before another task can
2582 * wake it. set_current_state() is implemented using smp_store_mb() and
2583 * queue_me() calls spin_unlock() upon completion, both serializing
2584 * access to the hash list and forcing another memory barrier.
2586 set_current_state(TASK_INTERRUPTIBLE
);
2591 hrtimer_sleeper_start_expires(timeout
, HRTIMER_MODE_ABS
);
2594 * If we have been removed from the hash list, then another task
2595 * has tried to wake us, and we can skip the call to schedule().
2597 if (likely(!plist_node_empty(&q
->list
))) {
2599 * If the timer has already expired, current will already be
2600 * flagged for rescheduling. Only call schedule if there
2601 * is no timeout, or if it has yet to expire.
2603 if (!timeout
|| timeout
->task
)
2604 freezable_schedule();
2606 __set_current_state(TASK_RUNNING
);
2610 * futex_wait_setup() - Prepare to wait on a futex
2611 * @uaddr: the futex userspace address
2612 * @val: the expected value
2613 * @flags: futex flags (FLAGS_SHARED, etc.)
2614 * @q: the associated futex_q
2615 * @hb: storage for hash_bucket pointer to be returned to caller
2617 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2618 * compare it with the expected value. Handle atomic faults internally.
2619 * Return with the hb lock held and a q.key reference on success, and unlocked
2620 * with no q.key reference on failure.
2623 * - 0 - uaddr contains val and hb has been locked;
2624 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2626 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2627 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2633 * Access the page AFTER the hash-bucket is locked.
2634 * Order is important:
2636 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2637 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2639 * The basic logical guarantee of a futex is that it blocks ONLY
2640 * if cond(var) is known to be true at the time of blocking, for
2641 * any cond. If we locked the hash-bucket after testing *uaddr, that
2642 * would open a race condition where we could block indefinitely with
2643 * cond(var) false, which would violate the guarantee.
2645 * On the other hand, we insert q and release the hash-bucket only
2646 * after testing *uaddr. This guarantees that futex_wait() will NOT
2647 * absorb a wakeup if *uaddr does not match the desired values
2648 * while the syscall executes.
2651 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, FUTEX_READ
);
2652 if (unlikely(ret
!= 0))
2656 *hb
= queue_lock(q
);
2658 ret
= get_futex_value_locked(&uval
, uaddr
);
2663 ret
= get_user(uval
, uaddr
);
2667 if (!(flags
& FLAGS_SHARED
))
2681 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2682 ktime_t
*abs_time
, u32 bitset
)
2684 struct hrtimer_sleeper timeout
, *to
;
2685 struct restart_block
*restart
;
2686 struct futex_hash_bucket
*hb
;
2687 struct futex_q q
= futex_q_init
;
2694 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
2695 current
->timer_slack_ns
);
2698 * Prepare to wait on uaddr. On success, holds hb lock and increments
2701 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2705 /* queue_me and wait for wakeup, timeout, or a signal. */
2706 futex_wait_queue_me(hb
, &q
, to
);
2708 /* If we were woken (and unqueued), we succeeded, whatever. */
2710 /* unqueue_me() drops q.key ref */
2711 if (!unqueue_me(&q
))
2714 if (to
&& !to
->task
)
2718 * We expect signal_pending(current), but we might be the
2719 * victim of a spurious wakeup as well.
2721 if (!signal_pending(current
))
2728 restart
= ¤t
->restart_block
;
2729 restart
->fn
= futex_wait_restart
;
2730 restart
->futex
.uaddr
= uaddr
;
2731 restart
->futex
.val
= val
;
2732 restart
->futex
.time
= *abs_time
;
2733 restart
->futex
.bitset
= bitset
;
2734 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2736 ret
= -ERESTART_RESTARTBLOCK
;
2740 hrtimer_cancel(&to
->timer
);
2741 destroy_hrtimer_on_stack(&to
->timer
);
2747 static long futex_wait_restart(struct restart_block
*restart
)
2749 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2750 ktime_t t
, *tp
= NULL
;
2752 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2753 t
= restart
->futex
.time
;
2756 restart
->fn
= do_no_restart_syscall
;
2758 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2759 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2764 * Userspace tried a 0 -> TID atomic transition of the futex value
2765 * and failed. The kernel side here does the whole locking operation:
2766 * if there are waiters then it will block as a consequence of relying
2767 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2768 * a 0 value of the futex too.).
2770 * Also serves as futex trylock_pi()'ing, and due semantics.
2772 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2773 ktime_t
*time
, int trylock
)
2775 struct hrtimer_sleeper timeout
, *to
;
2776 struct futex_pi_state
*pi_state
= NULL
;
2777 struct task_struct
*exiting
= NULL
;
2778 struct rt_mutex_waiter rt_waiter
;
2779 struct futex_hash_bucket
*hb
;
2780 struct futex_q q
= futex_q_init
;
2783 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2786 if (refill_pi_state_cache())
2789 to
= futex_setup_timer(time
, &timeout
, FLAGS_CLOCKRT
, 0);
2792 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, FUTEX_WRITE
);
2793 if (unlikely(ret
!= 0))
2797 hb
= queue_lock(&q
);
2799 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2801 if (unlikely(ret
)) {
2803 * Atomic work succeeded and we got the lock,
2804 * or failed. Either way, we do _not_ block.
2808 /* We got the lock. */
2810 goto out_unlock_put_key
;
2816 * Two reasons for this:
2817 * - EBUSY: Task is exiting and we just wait for the
2819 * - EAGAIN: The user space value changed.
2823 * Handle the case where the owner is in the middle of
2824 * exiting. Wait for the exit to complete otherwise
2825 * this task might loop forever, aka. live lock.
2827 wait_for_owner_exiting(ret
, exiting
);
2831 goto out_unlock_put_key
;
2835 WARN_ON(!q
.pi_state
);
2838 * Only actually queue now that the atomic ops are done:
2843 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2844 /* Fixup the trylock return value: */
2845 ret
= ret
? 0 : -EWOULDBLOCK
;
2849 rt_mutex_init_waiter(&rt_waiter
);
2852 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2853 * hold it while doing rt_mutex_start_proxy(), because then it will
2854 * include hb->lock in the blocking chain, even through we'll not in
2855 * fact hold it while blocking. This will lead it to report -EDEADLK
2856 * and BUG when futex_unlock_pi() interleaves with this.
2858 * Therefore acquire wait_lock while holding hb->lock, but drop the
2859 * latter before calling __rt_mutex_start_proxy_lock(). This
2860 * interleaves with futex_unlock_pi() -- which does a similar lock
2861 * handoff -- such that the latter can observe the futex_q::pi_state
2862 * before __rt_mutex_start_proxy_lock() is done.
2864 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2865 spin_unlock(q
.lock_ptr
);
2867 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2868 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2869 * it sees the futex_q::pi_state.
2871 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2872 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2881 hrtimer_sleeper_start_expires(to
, HRTIMER_MODE_ABS
);
2883 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2886 spin_lock(q
.lock_ptr
);
2888 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2889 * first acquire the hb->lock before removing the lock from the
2890 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2893 * In particular; it is important that futex_unlock_pi() can not
2894 * observe this inconsistency.
2896 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2901 * Fixup the pi_state owner and possibly acquire the lock if we
2904 res
= fixup_owner(uaddr
, &q
, !ret
);
2906 * If fixup_owner() returned an error, proprogate that. If it acquired
2907 * the lock, clear our -ETIMEDOUT or -EINTR.
2910 ret
= (res
< 0) ? res
: 0;
2913 * If fixup_owner() faulted and was unable to handle the fault, unlock
2914 * it and return the fault to userspace.
2916 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2917 pi_state
= q
.pi_state
;
2918 get_pi_state(pi_state
);
2921 /* Unqueue and drop the lock */
2925 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2926 put_pi_state(pi_state
);
2936 hrtimer_cancel(&to
->timer
);
2937 destroy_hrtimer_on_stack(&to
->timer
);
2939 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2944 ret
= fault_in_user_writeable(uaddr
);
2948 if (!(flags
& FLAGS_SHARED
))
2955 * Userspace attempted a TID -> 0 atomic transition, and failed.
2956 * This is the in-kernel slowpath: we look up the PI state (if any),
2957 * and do the rt-mutex unlock.
2959 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2961 u32 curval
, uval
, vpid
= task_pid_vnr(current
);
2962 union futex_key key
= FUTEX_KEY_INIT
;
2963 struct futex_hash_bucket
*hb
;
2964 struct futex_q
*top_waiter
;
2967 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2971 if (get_user(uval
, uaddr
))
2974 * We release only a lock we actually own:
2976 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2979 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_WRITE
);
2983 hb
= hash_futex(&key
);
2984 spin_lock(&hb
->lock
);
2987 * Check waiters first. We do not trust user space values at
2988 * all and we at least want to know if user space fiddled
2989 * with the futex value instead of blindly unlocking.
2991 top_waiter
= futex_top_waiter(hb
, &key
);
2993 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3000 * If current does not own the pi_state then the futex is
3001 * inconsistent and user space fiddled with the futex value.
3003 if (pi_state
->owner
!= current
)
3006 get_pi_state(pi_state
);
3008 * By taking wait_lock while still holding hb->lock, we ensure
3009 * there is no point where we hold neither; and therefore
3010 * wake_futex_pi() must observe a state consistent with what we
3013 * In particular; this forces __rt_mutex_start_proxy() to
3014 * complete such that we're guaranteed to observe the
3015 * rt_waiter. Also see the WARN in wake_futex_pi().
3017 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3018 spin_unlock(&hb
->lock
);
3020 /* drops pi_state->pi_mutex.wait_lock */
3021 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3023 put_pi_state(pi_state
);
3026 * Success, we're done! No tricky corner cases.
3031 * The atomic access to the futex value generated a
3032 * pagefault, so retry the user-access and the wakeup:
3037 * A unconditional UNLOCK_PI op raced against a waiter
3038 * setting the FUTEX_WAITERS bit. Try again.
3043 * wake_futex_pi has detected invalid state. Tell user
3050 * We have no kernel internal state, i.e. no waiters in the
3051 * kernel. Waiters which are about to queue themselves are stuck
3052 * on hb->lock. So we can safely ignore them. We do neither
3053 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3056 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3057 spin_unlock(&hb
->lock
);
3072 * If uval has changed, let user space handle it.
3074 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3077 spin_unlock(&hb
->lock
);
3087 ret
= fault_in_user_writeable(uaddr
);
3095 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3096 * @hb: the hash_bucket futex_q was original enqueued on
3097 * @q: the futex_q woken while waiting to be requeued
3098 * @key2: the futex_key of the requeue target futex
3099 * @timeout: the timeout associated with the wait (NULL if none)
3101 * Detect if the task was woken on the initial futex as opposed to the requeue
3102 * target futex. If so, determine if it was a timeout or a signal that caused
3103 * the wakeup and return the appropriate error code to the caller. Must be
3104 * called with the hb lock held.
3107 * - 0 = no early wakeup detected;
3108 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3111 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3112 struct futex_q
*q
, union futex_key
*key2
,
3113 struct hrtimer_sleeper
*timeout
)
3118 * With the hb lock held, we avoid races while we process the wakeup.
3119 * We only need to hold hb (and not hb2) to ensure atomicity as the
3120 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3121 * It can't be requeued from uaddr2 to something else since we don't
3122 * support a PI aware source futex for requeue.
3124 if (!match_futex(&q
->key
, key2
)) {
3125 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3127 * We were woken prior to requeue by a timeout or a signal.
3128 * Unqueue the futex_q and determine which it was.
3130 plist_del(&q
->list
, &hb
->chain
);
3133 /* Handle spurious wakeups gracefully */
3135 if (timeout
&& !timeout
->task
)
3137 else if (signal_pending(current
))
3138 ret
= -ERESTARTNOINTR
;
3144 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3145 * @uaddr: the futex we initially wait on (non-pi)
3146 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3147 * the same type, no requeueing from private to shared, etc.
3148 * @val: the expected value of uaddr
3149 * @abs_time: absolute timeout
3150 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3151 * @uaddr2: the pi futex we will take prior to returning to user-space
3153 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3154 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3155 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3156 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3157 * without one, the pi logic would not know which task to boost/deboost, if
3158 * there was a need to.
3160 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3161 * via the following--
3162 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3163 * 2) wakeup on uaddr2 after a requeue
3167 * If 3, cleanup and return -ERESTARTNOINTR.
3169 * If 2, we may then block on trying to take the rt_mutex and return via:
3170 * 5) successful lock
3173 * 8) other lock acquisition failure
3175 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3177 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3183 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3184 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3187 struct hrtimer_sleeper timeout
, *to
;
3188 struct futex_pi_state
*pi_state
= NULL
;
3189 struct rt_mutex_waiter rt_waiter
;
3190 struct futex_hash_bucket
*hb
;
3191 union futex_key key2
= FUTEX_KEY_INIT
;
3192 struct futex_q q
= futex_q_init
;
3195 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3198 if (uaddr
== uaddr2
)
3204 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
3205 current
->timer_slack_ns
);
3208 * The waiter is allocated on our stack, manipulated by the requeue
3209 * code while we sleep on uaddr.
3211 rt_mutex_init_waiter(&rt_waiter
);
3213 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
3214 if (unlikely(ret
!= 0))
3218 q
.rt_waiter
= &rt_waiter
;
3219 q
.requeue_pi_key
= &key2
;
3222 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3225 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3230 * The check above which compares uaddrs is not sufficient for
3231 * shared futexes. We need to compare the keys:
3233 if (match_futex(&q
.key
, &key2
)) {
3239 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3240 futex_wait_queue_me(hb
, &q
, to
);
3242 spin_lock(&hb
->lock
);
3243 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3244 spin_unlock(&hb
->lock
);
3249 * In order for us to be here, we know our q.key == key2, and since
3250 * we took the hb->lock above, we also know that futex_requeue() has
3251 * completed and we no longer have to concern ourselves with a wakeup
3252 * race with the atomic proxy lock acquisition by the requeue code. The
3253 * futex_requeue dropped our key1 reference and incremented our key2
3257 /* Check if the requeue code acquired the second futex for us. */
3260 * Got the lock. We might not be the anticipated owner if we
3261 * did a lock-steal - fix up the PI-state in that case.
3263 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3264 spin_lock(q
.lock_ptr
);
3265 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3266 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3267 pi_state
= q
.pi_state
;
3268 get_pi_state(pi_state
);
3271 * Drop the reference to the pi state which
3272 * the requeue_pi() code acquired for us.
3274 put_pi_state(q
.pi_state
);
3275 spin_unlock(q
.lock_ptr
);
3278 struct rt_mutex
*pi_mutex
;
3281 * We have been woken up by futex_unlock_pi(), a timeout, or a
3282 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3285 WARN_ON(!q
.pi_state
);
3286 pi_mutex
= &q
.pi_state
->pi_mutex
;
3287 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3289 spin_lock(q
.lock_ptr
);
3290 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3293 debug_rt_mutex_free_waiter(&rt_waiter
);
3295 * Fixup the pi_state owner and possibly acquire the lock if we
3298 res
= fixup_owner(uaddr2
, &q
, !ret
);
3300 * If fixup_owner() returned an error, proprogate that. If it
3301 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3304 ret
= (res
< 0) ? res
: 0;
3307 * If fixup_pi_state_owner() faulted and was unable to handle
3308 * the fault, unlock the rt_mutex and return the fault to
3311 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3312 pi_state
= q
.pi_state
;
3313 get_pi_state(pi_state
);
3316 /* Unqueue and drop the lock. */
3321 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3322 put_pi_state(pi_state
);
3325 if (ret
== -EINTR
) {
3327 * We've already been requeued, but cannot restart by calling
3328 * futex_lock_pi() directly. We could restart this syscall, but
3329 * it would detect that the user space "val" changed and return
3330 * -EWOULDBLOCK. Save the overhead of the restart and return
3331 * -EWOULDBLOCK directly.
3338 hrtimer_cancel(&to
->timer
);
3339 destroy_hrtimer_on_stack(&to
->timer
);
3345 * Support for robust futexes: the kernel cleans up held futexes at
3348 * Implementation: user-space maintains a per-thread list of locks it
3349 * is holding. Upon do_exit(), the kernel carefully walks this list,
3350 * and marks all locks that are owned by this thread with the
3351 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3352 * always manipulated with the lock held, so the list is private and
3353 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3354 * field, to allow the kernel to clean up if the thread dies after
3355 * acquiring the lock, but just before it could have added itself to
3356 * the list. There can only be one such pending lock.
3360 * sys_set_robust_list() - Set the robust-futex list head of a task
3361 * @head: pointer to the list-head
3362 * @len: length of the list-head, as userspace expects
3364 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3367 if (!futex_cmpxchg_enabled
)
3370 * The kernel knows only one size for now:
3372 if (unlikely(len
!= sizeof(*head
)))
3375 current
->robust_list
= head
;
3381 * sys_get_robust_list() - Get the robust-futex list head of a task
3382 * @pid: pid of the process [zero for current task]
3383 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3384 * @len_ptr: pointer to a length field, the kernel fills in the header size
3386 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3387 struct robust_list_head __user
* __user
*, head_ptr
,
3388 size_t __user
*, len_ptr
)
3390 struct robust_list_head __user
*head
;
3392 struct task_struct
*p
;
3394 if (!futex_cmpxchg_enabled
)
3403 p
= find_task_by_vpid(pid
);
3409 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3412 head
= p
->robust_list
;
3415 if (put_user(sizeof(*head
), len_ptr
))
3417 return put_user(head
, head_ptr
);
3425 /* Constants for the pending_op argument of handle_futex_death */
3426 #define HANDLE_DEATH_PENDING true
3427 #define HANDLE_DEATH_LIST false
3430 * Process a futex-list entry, check whether it's owned by the
3431 * dying task, and do notification if so:
3433 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3434 bool pi
, bool pending_op
)
3436 u32 uval
, nval
, mval
;
3439 /* Futex address must be 32bit aligned */
3440 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3444 if (get_user(uval
, uaddr
))
3448 * Special case for regular (non PI) futexes. The unlock path in
3449 * user space has two race scenarios:
3451 * 1. The unlock path releases the user space futex value and
3452 * before it can execute the futex() syscall to wake up
3453 * waiters it is killed.
3455 * 2. A woken up waiter is killed before it can acquire the
3456 * futex in user space.
3458 * In both cases the TID validation below prevents a wakeup of
3459 * potential waiters which can cause these waiters to block
3462 * In both cases the following conditions are met:
3464 * 1) task->robust_list->list_op_pending != NULL
3465 * @pending_op == true
3466 * 2) User space futex value == 0
3467 * 3) Regular futex: @pi == false
3469 * If these conditions are met, it is safe to attempt waking up a
3470 * potential waiter without touching the user space futex value and
3471 * trying to set the OWNER_DIED bit. The user space futex value is
3472 * uncontended and the rest of the user space mutex state is
3473 * consistent, so a woken waiter will just take over the
3474 * uncontended futex. Setting the OWNER_DIED bit would create
3475 * inconsistent state and malfunction of the user space owner died
3478 if (pending_op
&& !pi
&& !uval
) {
3479 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3483 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3487 * Ok, this dying thread is truly holding a futex
3488 * of interest. Set the OWNER_DIED bit atomically
3489 * via cmpxchg, and if the value had FUTEX_WAITERS
3490 * set, wake up a waiter (if any). (We have to do a
3491 * futex_wake() even if OWNER_DIED is already set -
3492 * to handle the rare but possible case of recursive
3493 * thread-death.) The rest of the cleanup is done in
3496 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3499 * We are not holding a lock here, but we want to have
3500 * the pagefault_disable/enable() protection because
3501 * we want to handle the fault gracefully. If the
3502 * access fails we try to fault in the futex with R/W
3503 * verification via get_user_pages. get_user() above
3504 * does not guarantee R/W access. If that fails we
3505 * give up and leave the futex locked.
3507 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3510 if (fault_in_user_writeable(uaddr
))
3528 * Wake robust non-PI futexes here. The wakeup of
3529 * PI futexes happens in exit_pi_state():
3531 if (!pi
&& (uval
& FUTEX_WAITERS
))
3532 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3538 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3540 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3541 struct robust_list __user
* __user
*head
,
3544 unsigned long uentry
;
3546 if (get_user(uentry
, (unsigned long __user
*)head
))
3549 *entry
= (void __user
*)(uentry
& ~1UL);
3556 * Walk curr->robust_list (very carefully, it's a userspace list!)
3557 * and mark any locks found there dead, and notify any waiters.
3559 * We silently return on any sign of list-walking problem.
3561 static void exit_robust_list(struct task_struct
*curr
)
3563 struct robust_list_head __user
*head
= curr
->robust_list
;
3564 struct robust_list __user
*entry
, *next_entry
, *pending
;
3565 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3566 unsigned int next_pi
;
3567 unsigned long futex_offset
;
3570 if (!futex_cmpxchg_enabled
)
3574 * Fetch the list head (which was registered earlier, via
3575 * sys_set_robust_list()):
3577 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3580 * Fetch the relative futex offset:
3582 if (get_user(futex_offset
, &head
->futex_offset
))
3585 * Fetch any possibly pending lock-add first, and handle it
3588 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3591 next_entry
= NULL
; /* avoid warning with gcc */
3592 while (entry
!= &head
->list
) {
3594 * Fetch the next entry in the list before calling
3595 * handle_futex_death:
3597 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3599 * A pending lock might already be on the list, so
3600 * don't process it twice:
3602 if (entry
!= pending
) {
3603 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3604 curr
, pi
, HANDLE_DEATH_LIST
))
3612 * Avoid excessively long or circular lists:
3621 handle_futex_death((void __user
*)pending
+ futex_offset
,
3622 curr
, pip
, HANDLE_DEATH_PENDING
);
3626 static void futex_cleanup(struct task_struct
*tsk
)
3628 if (unlikely(tsk
->robust_list
)) {
3629 exit_robust_list(tsk
);
3630 tsk
->robust_list
= NULL
;
3633 #ifdef CONFIG_COMPAT
3634 if (unlikely(tsk
->compat_robust_list
)) {
3635 compat_exit_robust_list(tsk
);
3636 tsk
->compat_robust_list
= NULL
;
3640 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3641 exit_pi_state_list(tsk
);
3645 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3646 * @tsk: task to set the state on
3648 * Set the futex exit state of the task lockless. The futex waiter code
3649 * observes that state when a task is exiting and loops until the task has
3650 * actually finished the futex cleanup. The worst case for this is that the
3651 * waiter runs through the wait loop until the state becomes visible.
3653 * This is called from the recursive fault handling path in do_exit().
3655 * This is best effort. Either the futex exit code has run already or
3656 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3657 * take it over. If not, the problem is pushed back to user space. If the
3658 * futex exit code did not run yet, then an already queued waiter might
3659 * block forever, but there is nothing which can be done about that.
3661 void futex_exit_recursive(struct task_struct
*tsk
)
3663 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3664 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3665 mutex_unlock(&tsk
->futex_exit_mutex
);
3666 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3669 static void futex_cleanup_begin(struct task_struct
*tsk
)
3672 * Prevent various race issues against a concurrent incoming waiter
3673 * including live locks by forcing the waiter to block on
3674 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3675 * attach_to_pi_owner().
3677 mutex_lock(&tsk
->futex_exit_mutex
);
3680 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3682 * This ensures that all subsequent checks of tsk->futex_state in
3683 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3684 * tsk->pi_lock held.
3686 * It guarantees also that a pi_state which was queued right before
3687 * the state change under tsk->pi_lock by a concurrent waiter must
3688 * be observed in exit_pi_state_list().
3690 raw_spin_lock_irq(&tsk
->pi_lock
);
3691 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3692 raw_spin_unlock_irq(&tsk
->pi_lock
);
3695 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3698 * Lockless store. The only side effect is that an observer might
3699 * take another loop until it becomes visible.
3701 tsk
->futex_state
= state
;
3703 * Drop the exit protection. This unblocks waiters which observed
3704 * FUTEX_STATE_EXITING to reevaluate the state.
3706 mutex_unlock(&tsk
->futex_exit_mutex
);
3709 void futex_exec_release(struct task_struct
*tsk
)
3712 * The state handling is done for consistency, but in the case of
3713 * exec() there is no way to prevent futher damage as the PID stays
3714 * the same. But for the unlikely and arguably buggy case that a
3715 * futex is held on exec(), this provides at least as much state
3716 * consistency protection which is possible.
3718 futex_cleanup_begin(tsk
);
3721 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3722 * exec a new binary.
3724 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3727 void futex_exit_release(struct task_struct
*tsk
)
3729 futex_cleanup_begin(tsk
);
3731 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3734 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3735 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3737 int cmd
= op
& FUTEX_CMD_MASK
;
3738 unsigned int flags
= 0;
3740 if (!(op
& FUTEX_PRIVATE_FLAG
))
3741 flags
|= FLAGS_SHARED
;
3743 if (op
& FUTEX_CLOCK_REALTIME
) {
3744 flags
|= FLAGS_CLOCKRT
;
3745 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3746 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3752 case FUTEX_UNLOCK_PI
:
3753 case FUTEX_TRYLOCK_PI
:
3754 case FUTEX_WAIT_REQUEUE_PI
:
3755 case FUTEX_CMP_REQUEUE_PI
:
3756 if (!futex_cmpxchg_enabled
)
3762 val3
= FUTEX_BITSET_MATCH_ANY
;
3764 case FUTEX_WAIT_BITSET
:
3765 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3767 val3
= FUTEX_BITSET_MATCH_ANY
;
3769 case FUTEX_WAKE_BITSET
:
3770 return futex_wake(uaddr
, flags
, val
, val3
);
3772 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3773 case FUTEX_CMP_REQUEUE
:
3774 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3776 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3778 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3779 case FUTEX_UNLOCK_PI
:
3780 return futex_unlock_pi(uaddr
, flags
);
3781 case FUTEX_TRYLOCK_PI
:
3782 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3783 case FUTEX_WAIT_REQUEUE_PI
:
3784 val3
= FUTEX_BITSET_MATCH_ANY
;
3785 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3787 case FUTEX_CMP_REQUEUE_PI
:
3788 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3794 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3795 struct __kernel_timespec __user
*, utime
, u32 __user
*, uaddr2
,
3798 struct timespec64 ts
;
3799 ktime_t t
, *tp
= NULL
;
3801 int cmd
= op
& FUTEX_CMD_MASK
;
3803 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3804 cmd
== FUTEX_WAIT_BITSET
||
3805 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3806 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3808 if (get_timespec64(&ts
, utime
))
3810 if (!timespec64_valid(&ts
))
3813 t
= timespec64_to_ktime(ts
);
3814 if (cmd
== FUTEX_WAIT
)
3815 t
= ktime_add_safe(ktime_get(), t
);
3816 else if (!(op
& FUTEX_CLOCK_REALTIME
))
3817 t
= timens_ktime_to_host(CLOCK_MONOTONIC
, t
);
3821 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3822 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3824 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3825 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3826 val2
= (u32
) (unsigned long) utime
;
3828 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3831 #ifdef CONFIG_COMPAT
3833 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3836 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3837 compat_uptr_t __user
*head
, unsigned int *pi
)
3839 if (get_user(*uentry
, head
))
3842 *entry
= compat_ptr((*uentry
) & ~1);
3843 *pi
= (unsigned int)(*uentry
) & 1;
3848 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3849 compat_long_t futex_offset
)
3851 compat_uptr_t base
= ptr_to_compat(entry
);
3852 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3858 * Walk curr->robust_list (very carefully, it's a userspace list!)
3859 * and mark any locks found there dead, and notify any waiters.
3861 * We silently return on any sign of list-walking problem.
3863 static void compat_exit_robust_list(struct task_struct
*curr
)
3865 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3866 struct robust_list __user
*entry
, *next_entry
, *pending
;
3867 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3868 unsigned int next_pi
;
3869 compat_uptr_t uentry
, next_uentry
, upending
;
3870 compat_long_t futex_offset
;
3873 if (!futex_cmpxchg_enabled
)
3877 * Fetch the list head (which was registered earlier, via
3878 * sys_set_robust_list()):
3880 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3883 * Fetch the relative futex offset:
3885 if (get_user(futex_offset
, &head
->futex_offset
))
3888 * Fetch any possibly pending lock-add first, and handle it
3891 if (compat_fetch_robust_entry(&upending
, &pending
,
3892 &head
->list_op_pending
, &pip
))
3895 next_entry
= NULL
; /* avoid warning with gcc */
3896 while (entry
!= (struct robust_list __user
*) &head
->list
) {
3898 * Fetch the next entry in the list before calling
3899 * handle_futex_death:
3901 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
3902 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
3904 * A pending lock might already be on the list, so
3905 * dont process it twice:
3907 if (entry
!= pending
) {
3908 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
3910 if (handle_futex_death(uaddr
, curr
, pi
,
3916 uentry
= next_uentry
;
3920 * Avoid excessively long or circular lists:
3928 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
3930 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
3934 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
3935 struct compat_robust_list_head __user
*, head
,
3938 if (!futex_cmpxchg_enabled
)
3941 if (unlikely(len
!= sizeof(*head
)))
3944 current
->compat_robust_list
= head
;
3949 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3950 compat_uptr_t __user
*, head_ptr
,
3951 compat_size_t __user
*, len_ptr
)
3953 struct compat_robust_list_head __user
*head
;
3955 struct task_struct
*p
;
3957 if (!futex_cmpxchg_enabled
)
3966 p
= find_task_by_vpid(pid
);
3972 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3975 head
= p
->compat_robust_list
;
3978 if (put_user(sizeof(*head
), len_ptr
))
3980 return put_user(ptr_to_compat(head
), head_ptr
);
3987 #endif /* CONFIG_COMPAT */
3989 #ifdef CONFIG_COMPAT_32BIT_TIME
3990 SYSCALL_DEFINE6(futex_time32
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3991 struct old_timespec32 __user
*, utime
, u32 __user
*, uaddr2
,
3994 struct timespec64 ts
;
3995 ktime_t t
, *tp
= NULL
;
3997 int cmd
= op
& FUTEX_CMD_MASK
;
3999 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
4000 cmd
== FUTEX_WAIT_BITSET
||
4001 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
4002 if (get_old_timespec32(&ts
, utime
))
4004 if (!timespec64_valid(&ts
))
4007 t
= timespec64_to_ktime(ts
);
4008 if (cmd
== FUTEX_WAIT
)
4009 t
= ktime_add_safe(ktime_get(), t
);
4010 else if (!(op
& FUTEX_CLOCK_REALTIME
))
4011 t
= timens_ktime_to_host(CLOCK_MONOTONIC
, t
);
4014 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
4015 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
4016 val2
= (int) (unsigned long) utime
;
4018 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
4020 #endif /* CONFIG_COMPAT_32BIT_TIME */
4022 static void __init
futex_detect_cmpxchg(void)
4024 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4028 * This will fail and we want it. Some arch implementations do
4029 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4030 * functionality. We want to know that before we call in any
4031 * of the complex code paths. Also we want to prevent
4032 * registration of robust lists in that case. NULL is
4033 * guaranteed to fault and we get -EFAULT on functional
4034 * implementation, the non-functional ones will return
4037 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4038 futex_cmpxchg_enabled
= 1;
4042 static int __init
futex_init(void)
4044 unsigned int futex_shift
;
4047 #if CONFIG_BASE_SMALL
4048 futex_hashsize
= 16;
4050 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4053 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4055 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4057 futex_hashsize
, futex_hashsize
);
4058 futex_hashsize
= 1UL << futex_shift
;
4060 futex_detect_cmpxchg();
4062 for (i
= 0; i
< futex_hashsize
; i
++) {
4063 atomic_set(&futex_queues
[i
].waiters
, 0);
4064 plist_head_init(&futex_queues
[i
].chain
);
4065 spin_lock_init(&futex_queues
[i
].lock
);
4070 core_initcall(futex_init
);