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
;
792 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
793 owner
= pi_state
->owner
;
795 raw_spin_lock(&owner
->pi_lock
);
796 list_del_init(&pi_state
->list
);
797 raw_spin_unlock(&owner
->pi_lock
);
799 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
800 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
803 if (current
->pi_state_cache
) {
807 * pi_state->list is already empty.
808 * clear pi_state->owner.
809 * refcount is at 0 - put it back to 1.
811 pi_state
->owner
= NULL
;
812 refcount_set(&pi_state
->refcount
, 1);
813 current
->pi_state_cache
= pi_state
;
817 #ifdef CONFIG_FUTEX_PI
820 * This task is holding PI mutexes at exit time => bad.
821 * Kernel cleans up PI-state, but userspace is likely hosed.
822 * (Robust-futex cleanup is separate and might save the day for userspace.)
824 static void exit_pi_state_list(struct task_struct
*curr
)
826 struct list_head
*next
, *head
= &curr
->pi_state_list
;
827 struct futex_pi_state
*pi_state
;
828 struct futex_hash_bucket
*hb
;
829 union futex_key key
= FUTEX_KEY_INIT
;
831 if (!futex_cmpxchg_enabled
)
834 * We are a ZOMBIE and nobody can enqueue itself on
835 * pi_state_list anymore, but we have to be careful
836 * versus waiters unqueueing themselves:
838 raw_spin_lock_irq(&curr
->pi_lock
);
839 while (!list_empty(head
)) {
841 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
843 hb
= hash_futex(&key
);
846 * We can race against put_pi_state() removing itself from the
847 * list (a waiter going away). put_pi_state() will first
848 * decrement the reference count and then modify the list, so
849 * its possible to see the list entry but fail this reference
852 * In that case; drop the locks to let put_pi_state() make
853 * progress and retry the loop.
855 if (!refcount_inc_not_zero(&pi_state
->refcount
)) {
856 raw_spin_unlock_irq(&curr
->pi_lock
);
858 raw_spin_lock_irq(&curr
->pi_lock
);
861 raw_spin_unlock_irq(&curr
->pi_lock
);
863 spin_lock(&hb
->lock
);
864 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
865 raw_spin_lock(&curr
->pi_lock
);
867 * We dropped the pi-lock, so re-check whether this
868 * task still owns the PI-state:
870 if (head
->next
!= next
) {
871 /* retain curr->pi_lock for the loop invariant */
872 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
873 spin_unlock(&hb
->lock
);
874 put_pi_state(pi_state
);
878 WARN_ON(pi_state
->owner
!= curr
);
879 WARN_ON(list_empty(&pi_state
->list
));
880 list_del_init(&pi_state
->list
);
881 pi_state
->owner
= NULL
;
883 raw_spin_unlock(&curr
->pi_lock
);
884 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
885 spin_unlock(&hb
->lock
);
887 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
888 put_pi_state(pi_state
);
890 raw_spin_lock_irq(&curr
->pi_lock
);
892 raw_spin_unlock_irq(&curr
->pi_lock
);
895 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
899 * We need to check the following states:
901 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
903 * [1] NULL | --- | --- | 0 | 0/1 | Valid
904 * [2] NULL | --- | --- | >0 | 0/1 | Valid
906 * [3] Found | NULL | -- | Any | 0/1 | Invalid
908 * [4] Found | Found | NULL | 0 | 1 | Valid
909 * [5] Found | Found | NULL | >0 | 1 | Invalid
911 * [6] Found | Found | task | 0 | 1 | Valid
913 * [7] Found | Found | NULL | Any | 0 | Invalid
915 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
916 * [9] Found | Found | task | 0 | 0 | Invalid
917 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
919 * [1] Indicates that the kernel can acquire the futex atomically. We
920 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
922 * [2] Valid, if TID does not belong to a kernel thread. If no matching
923 * thread is found then it indicates that the owner TID has died.
925 * [3] Invalid. The waiter is queued on a non PI futex
927 * [4] Valid state after exit_robust_list(), which sets the user space
928 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
930 * [5] The user space value got manipulated between exit_robust_list()
931 * and exit_pi_state_list()
933 * [6] Valid state after exit_pi_state_list() which sets the new owner in
934 * the pi_state but cannot access the user space value.
936 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
938 * [8] Owner and user space value match
940 * [9] There is no transient state which sets the user space TID to 0
941 * except exit_robust_list(), but this is indicated by the
942 * FUTEX_OWNER_DIED bit. See [4]
944 * [10] There is no transient state which leaves owner and user space
948 * Serialization and lifetime rules:
952 * hb -> futex_q, relation
953 * futex_q -> pi_state, relation
955 * (cannot be raw because hb can contain arbitrary amount
958 * pi_mutex->wait_lock:
962 * (and pi_mutex 'obviously')
966 * p->pi_state_list -> pi_state->list, relation
968 * pi_state->refcount:
976 * pi_mutex->wait_lock
982 * Validate that the existing waiter has a pi_state and sanity check
983 * the pi_state against the user space value. If correct, attach to
986 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
987 struct futex_pi_state
*pi_state
,
988 struct futex_pi_state
**ps
)
990 pid_t pid
= uval
& FUTEX_TID_MASK
;
995 * Userspace might have messed up non-PI and PI futexes [3]
997 if (unlikely(!pi_state
))
1001 * We get here with hb->lock held, and having found a
1002 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1003 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1004 * which in turn means that futex_lock_pi() still has a reference on
1007 * The waiter holding a reference on @pi_state also protects against
1008 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1009 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1010 * free pi_state before we can take a reference ourselves.
1012 WARN_ON(!refcount_read(&pi_state
->refcount
));
1015 * Now that we have a pi_state, we can acquire wait_lock
1016 * and do the state validation.
1018 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1021 * Since {uval, pi_state} is serialized by wait_lock, and our current
1022 * uval was read without holding it, it can have changed. Verify it
1023 * still is what we expect it to be, otherwise retry the entire
1026 if (get_futex_value_locked(&uval2
, uaddr
))
1033 * Handle the owner died case:
1035 if (uval
& FUTEX_OWNER_DIED
) {
1037 * exit_pi_state_list sets owner to NULL and wakes the
1038 * topmost waiter. The task which acquires the
1039 * pi_state->rt_mutex will fixup owner.
1041 if (!pi_state
->owner
) {
1043 * No pi state owner, but the user space TID
1044 * is not 0. Inconsistent state. [5]
1049 * Take a ref on the state and return success. [4]
1055 * If TID is 0, then either the dying owner has not
1056 * yet executed exit_pi_state_list() or some waiter
1057 * acquired the rtmutex in the pi state, but did not
1058 * yet fixup the TID in user space.
1060 * Take a ref on the state and return success. [6]
1066 * If the owner died bit is not set, then the pi_state
1067 * must have an owner. [7]
1069 if (!pi_state
->owner
)
1074 * Bail out if user space manipulated the futex value. If pi
1075 * state exists then the owner TID must be the same as the
1076 * user space TID. [9/10]
1078 if (pid
!= task_pid_vnr(pi_state
->owner
))
1082 get_pi_state(pi_state
);
1083 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1100 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1105 * wait_for_owner_exiting - Block until the owner has exited
1106 * @ret: owner's current futex lock status
1107 * @exiting: Pointer to the exiting task
1109 * Caller must hold a refcount on @exiting.
1111 static void wait_for_owner_exiting(int ret
, struct task_struct
*exiting
)
1113 if (ret
!= -EBUSY
) {
1114 WARN_ON_ONCE(exiting
);
1118 if (WARN_ON_ONCE(ret
== -EBUSY
&& !exiting
))
1121 mutex_lock(&exiting
->futex_exit_mutex
);
1123 * No point in doing state checking here. If the waiter got here
1124 * while the task was in exec()->exec_futex_release() then it can
1125 * have any FUTEX_STATE_* value when the waiter has acquired the
1126 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1127 * already. Highly unlikely and not a problem. Just one more round
1128 * through the futex maze.
1130 mutex_unlock(&exiting
->futex_exit_mutex
);
1132 put_task_struct(exiting
);
1135 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1136 struct task_struct
*tsk
)
1141 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1142 * caller that the alleged owner is busy.
1144 if (tsk
&& tsk
->futex_state
!= FUTEX_STATE_DEAD
)
1148 * Reread the user space value to handle the following situation:
1152 * sys_exit() sys_futex()
1153 * do_exit() futex_lock_pi()
1154 * futex_lock_pi_atomic()
1155 * exit_signals(tsk) No waiters:
1156 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1157 * mm_release(tsk) Set waiter bit
1158 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1159 * Set owner died attach_to_pi_owner() {
1160 * *uaddr = 0xC0000000; tsk = get_task(PID);
1161 * } if (!tsk->flags & PF_EXITING) {
1163 * tsk->futex_state = } else {
1164 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1167 * return -ESRCH; <--- FAIL
1170 * Returning ESRCH unconditionally is wrong here because the
1171 * user space value has been changed by the exiting task.
1173 * The same logic applies to the case where the exiting task is
1176 if (get_futex_value_locked(&uval2
, uaddr
))
1179 /* If the user space value has changed, try again. */
1184 * The exiting task did not have a robust list, the robust list was
1185 * corrupted or the user space value in *uaddr is simply bogus.
1186 * Give up and tell user space.
1192 * Lookup the task for the TID provided from user space and attach to
1193 * it after doing proper sanity checks.
1195 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1196 struct futex_pi_state
**ps
,
1197 struct task_struct
**exiting
)
1199 pid_t pid
= uval
& FUTEX_TID_MASK
;
1200 struct futex_pi_state
*pi_state
;
1201 struct task_struct
*p
;
1204 * We are the first waiter - try to look up the real owner and attach
1205 * the new pi_state to it, but bail out when TID = 0 [1]
1207 * The !pid check is paranoid. None of the call sites should end up
1208 * with pid == 0, but better safe than sorry. Let the caller retry
1212 p
= find_get_task_by_vpid(pid
);
1214 return handle_exit_race(uaddr
, uval
, NULL
);
1216 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1222 * We need to look at the task state to figure out, whether the
1223 * task is exiting. To protect against the change of the task state
1224 * in futex_exit_release(), we do this protected by p->pi_lock:
1226 raw_spin_lock_irq(&p
->pi_lock
);
1227 if (unlikely(p
->futex_state
!= FUTEX_STATE_OK
)) {
1229 * The task is on the way out. When the futex state is
1230 * FUTEX_STATE_DEAD, we know that the task has finished
1233 int ret
= handle_exit_race(uaddr
, uval
, p
);
1235 raw_spin_unlock_irq(&p
->pi_lock
);
1237 * If the owner task is between FUTEX_STATE_EXITING and
1238 * FUTEX_STATE_DEAD then store the task pointer and keep
1239 * the reference on the task struct. The calling code will
1240 * drop all locks, wait for the task to reach
1241 * FUTEX_STATE_DEAD and then drop the refcount. This is
1242 * required to prevent a live lock when the current task
1243 * preempted the exiting task between the two states.
1253 * No existing pi state. First waiter. [2]
1255 * This creates pi_state, we have hb->lock held, this means nothing can
1256 * observe this state, wait_lock is irrelevant.
1258 pi_state
= alloc_pi_state();
1261 * Initialize the pi_mutex in locked state and make @p
1264 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1266 /* Store the key for possible exit cleanups: */
1267 pi_state
->key
= *key
;
1269 WARN_ON(!list_empty(&pi_state
->list
));
1270 list_add(&pi_state
->list
, &p
->pi_state_list
);
1272 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1273 * because there is no concurrency as the object is not published yet.
1275 pi_state
->owner
= p
;
1276 raw_spin_unlock_irq(&p
->pi_lock
);
1285 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1286 struct futex_hash_bucket
*hb
,
1287 union futex_key
*key
, struct futex_pi_state
**ps
,
1288 struct task_struct
**exiting
)
1290 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1293 * If there is a waiter on that futex, validate it and
1294 * attach to the pi_state when the validation succeeds.
1297 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1300 * We are the first waiter - try to look up the owner based on
1301 * @uval and attach to it.
1303 return attach_to_pi_owner(uaddr
, uval
, key
, ps
, exiting
);
1306 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1311 if (unlikely(should_fail_futex(true)))
1314 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1318 /* If user space value changed, let the caller retry */
1319 return curval
!= uval
? -EAGAIN
: 0;
1323 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1324 * @uaddr: the pi futex user address
1325 * @hb: the pi futex hash bucket
1326 * @key: the futex key associated with uaddr and hb
1327 * @ps: the pi_state pointer where we store the result of the
1329 * @task: the task to perform the atomic lock work for. This will
1330 * be "current" except in the case of requeue pi.
1331 * @exiting: Pointer to store the task pointer of the owner task
1332 * which is in the middle of exiting
1333 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1336 * - 0 - ready to wait;
1337 * - 1 - acquired the lock;
1340 * The hb->lock and futex_key refs shall be held by the caller.
1342 * @exiting is only set when the return value is -EBUSY. If so, this holds
1343 * a refcount on the exiting task on return and the caller needs to drop it
1344 * after waiting for the exit to complete.
1346 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1347 union futex_key
*key
,
1348 struct futex_pi_state
**ps
,
1349 struct task_struct
*task
,
1350 struct task_struct
**exiting
,
1353 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1354 struct futex_q
*top_waiter
;
1358 * Read the user space value first so we can validate a few
1359 * things before proceeding further.
1361 if (get_futex_value_locked(&uval
, uaddr
))
1364 if (unlikely(should_fail_futex(true)))
1370 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1373 if ((unlikely(should_fail_futex(true))))
1377 * Lookup existing state first. If it exists, try to attach to
1380 top_waiter
= futex_top_waiter(hb
, key
);
1382 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1385 * No waiter and user TID is 0. We are here because the
1386 * waiters or the owner died bit is set or called from
1387 * requeue_cmp_pi or for whatever reason something took the
1390 if (!(uval
& FUTEX_TID_MASK
)) {
1392 * We take over the futex. No other waiters and the user space
1393 * TID is 0. We preserve the owner died bit.
1395 newval
= uval
& FUTEX_OWNER_DIED
;
1398 /* The futex requeue_pi code can enforce the waiters bit */
1400 newval
|= FUTEX_WAITERS
;
1402 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1403 /* If the take over worked, return 1 */
1404 return ret
< 0 ? ret
: 1;
1408 * First waiter. Set the waiters bit before attaching ourself to
1409 * the owner. If owner tries to unlock, it will be forced into
1410 * the kernel and blocked on hb->lock.
1412 newval
= uval
| FUTEX_WAITERS
;
1413 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1417 * If the update of the user space value succeeded, we try to
1418 * attach to the owner. If that fails, no harm done, we only
1419 * set the FUTEX_WAITERS bit in the user space variable.
1421 return attach_to_pi_owner(uaddr
, newval
, key
, ps
, exiting
);
1425 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1426 * @q: The futex_q to unqueue
1428 * The q->lock_ptr must not be NULL and must be held by the caller.
1430 static void __unqueue_futex(struct futex_q
*q
)
1432 struct futex_hash_bucket
*hb
;
1434 if (WARN_ON_SMP(!q
->lock_ptr
) || WARN_ON(plist_node_empty(&q
->list
)))
1436 lockdep_assert_held(q
->lock_ptr
);
1438 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1439 plist_del(&q
->list
, &hb
->chain
);
1444 * The hash bucket lock must be held when this is called.
1445 * Afterwards, the futex_q must not be accessed. Callers
1446 * must ensure to later call wake_up_q() for the actual
1449 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1451 struct task_struct
*p
= q
->task
;
1453 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1459 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1460 * is written, without taking any locks. This is possible in the event
1461 * of a spurious wakeup, for example. A memory barrier is required here
1462 * to prevent the following store to lock_ptr from getting ahead of the
1463 * plist_del in __unqueue_futex().
1465 smp_store_release(&q
->lock_ptr
, NULL
);
1468 * Queue the task for later wakeup for after we've released
1471 wake_q_add_safe(wake_q
, p
);
1475 * Caller must hold a reference on @pi_state.
1477 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1480 struct task_struct
*new_owner
;
1481 bool postunlock
= false;
1482 DEFINE_WAKE_Q(wake_q
);
1485 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1486 if (WARN_ON_ONCE(!new_owner
)) {
1488 * As per the comment in futex_unlock_pi() this should not happen.
1490 * When this happens, give up our locks and try again, giving
1491 * the futex_lock_pi() instance time to complete, either by
1492 * waiting on the rtmutex or removing itself from the futex
1500 * We pass it to the next owner. The WAITERS bit is always kept
1501 * enabled while there is PI state around. We cleanup the owner
1502 * died bit, because we are the owner.
1504 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1506 if (unlikely(should_fail_futex(true))) {
1511 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1512 if (!ret
&& (curval
!= uval
)) {
1514 * If a unconditional UNLOCK_PI operation (user space did not
1515 * try the TID->0 transition) raced with a waiter setting the
1516 * FUTEX_WAITERS flag between get_user() and locking the hash
1517 * bucket lock, retry the operation.
1519 if ((FUTEX_TID_MASK
& curval
) == uval
)
1529 * This is a point of no return; once we modify the uval there is no
1530 * going back and subsequent operations must not fail.
1533 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1534 WARN_ON(list_empty(&pi_state
->list
));
1535 list_del_init(&pi_state
->list
);
1536 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1538 raw_spin_lock(&new_owner
->pi_lock
);
1539 WARN_ON(!list_empty(&pi_state
->list
));
1540 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1541 pi_state
->owner
= new_owner
;
1542 raw_spin_unlock(&new_owner
->pi_lock
);
1544 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1547 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1550 rt_mutex_postunlock(&wake_q
);
1556 * Express the locking dependencies for lockdep:
1559 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1562 spin_lock(&hb1
->lock
);
1564 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1565 } else { /* hb1 > hb2 */
1566 spin_lock(&hb2
->lock
);
1567 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1572 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1574 spin_unlock(&hb1
->lock
);
1576 spin_unlock(&hb2
->lock
);
1580 * Wake up waiters matching bitset queued on this futex (uaddr).
1583 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1585 struct futex_hash_bucket
*hb
;
1586 struct futex_q
*this, *next
;
1587 union futex_key key
= FUTEX_KEY_INIT
;
1589 DEFINE_WAKE_Q(wake_q
);
1594 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_READ
);
1595 if (unlikely(ret
!= 0))
1598 hb
= hash_futex(&key
);
1600 /* Make sure we really have tasks to wakeup */
1601 if (!hb_waiters_pending(hb
))
1604 spin_lock(&hb
->lock
);
1606 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1607 if (match_futex (&this->key
, &key
)) {
1608 if (this->pi_state
|| this->rt_waiter
) {
1613 /* Check if one of the bits is set in both bitsets */
1614 if (!(this->bitset
& bitset
))
1617 mark_wake_futex(&wake_q
, this);
1618 if (++ret
>= nr_wake
)
1623 spin_unlock(&hb
->lock
);
1628 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1630 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1631 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1632 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1633 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1636 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1637 if (oparg
< 0 || oparg
> 31) {
1638 char comm
[sizeof(current
->comm
)];
1640 * kill this print and return -EINVAL when userspace
1643 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1644 get_task_comm(comm
, current
), oparg
);
1650 pagefault_disable();
1651 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1657 case FUTEX_OP_CMP_EQ
:
1658 return oldval
== cmparg
;
1659 case FUTEX_OP_CMP_NE
:
1660 return oldval
!= cmparg
;
1661 case FUTEX_OP_CMP_LT
:
1662 return oldval
< cmparg
;
1663 case FUTEX_OP_CMP_GE
:
1664 return oldval
>= cmparg
;
1665 case FUTEX_OP_CMP_LE
:
1666 return oldval
<= cmparg
;
1667 case FUTEX_OP_CMP_GT
:
1668 return oldval
> cmparg
;
1675 * Wake up all waiters hashed on the physical page that is mapped
1676 * to this virtual address:
1679 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1680 int nr_wake
, int nr_wake2
, int op
)
1682 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1683 struct futex_hash_bucket
*hb1
, *hb2
;
1684 struct futex_q
*this, *next
;
1686 DEFINE_WAKE_Q(wake_q
);
1689 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1690 if (unlikely(ret
!= 0))
1692 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
1693 if (unlikely(ret
!= 0))
1696 hb1
= hash_futex(&key1
);
1697 hb2
= hash_futex(&key2
);
1700 double_lock_hb(hb1
, hb2
);
1701 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1702 if (unlikely(op_ret
< 0)) {
1703 double_unlock_hb(hb1
, hb2
);
1705 if (!IS_ENABLED(CONFIG_MMU
) ||
1706 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1708 * we don't get EFAULT from MMU faults if we don't have
1709 * an MMU, but we might get them from range checking
1715 if (op_ret
== -EFAULT
) {
1716 ret
= fault_in_user_writeable(uaddr2
);
1721 if (!(flags
& FLAGS_SHARED
)) {
1730 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1731 if (match_futex (&this->key
, &key1
)) {
1732 if (this->pi_state
|| this->rt_waiter
) {
1736 mark_wake_futex(&wake_q
, this);
1737 if (++ret
>= nr_wake
)
1744 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1745 if (match_futex (&this->key
, &key2
)) {
1746 if (this->pi_state
|| this->rt_waiter
) {
1750 mark_wake_futex(&wake_q
, this);
1751 if (++op_ret
>= nr_wake2
)
1759 double_unlock_hb(hb1
, hb2
);
1765 * requeue_futex() - Requeue a futex_q from one hb to another
1766 * @q: the futex_q to requeue
1767 * @hb1: the source hash_bucket
1768 * @hb2: the target hash_bucket
1769 * @key2: the new key for the requeued futex_q
1772 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1773 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1777 * If key1 and key2 hash to the same bucket, no need to
1780 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1781 plist_del(&q
->list
, &hb1
->chain
);
1782 hb_waiters_dec(hb1
);
1783 hb_waiters_inc(hb2
);
1784 plist_add(&q
->list
, &hb2
->chain
);
1785 q
->lock_ptr
= &hb2
->lock
;
1791 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1793 * @key: the key of the requeue target futex
1794 * @hb: the hash_bucket of the requeue target futex
1796 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1797 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1798 * to the requeue target futex so the waiter can detect the wakeup on the right
1799 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1800 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1801 * to protect access to the pi_state to fixup the owner later. Must be called
1802 * with both q->lock_ptr and hb->lock held.
1805 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1806 struct futex_hash_bucket
*hb
)
1812 WARN_ON(!q
->rt_waiter
);
1813 q
->rt_waiter
= NULL
;
1815 q
->lock_ptr
= &hb
->lock
;
1817 wake_up_state(q
->task
, TASK_NORMAL
);
1821 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1822 * @pifutex: the user address of the to futex
1823 * @hb1: the from futex hash bucket, must be locked by the caller
1824 * @hb2: the to futex hash bucket, must be locked by the caller
1825 * @key1: the from futex key
1826 * @key2: the to futex key
1827 * @ps: address to store the pi_state pointer
1828 * @exiting: Pointer to store the task pointer of the owner task
1829 * which is in the middle of exiting
1830 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1832 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1833 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1834 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1835 * hb1 and hb2 must be held by the caller.
1837 * @exiting is only set when the return value is -EBUSY. If so, this holds
1838 * a refcount on the exiting task on return and the caller needs to drop it
1839 * after waiting for the exit to complete.
1842 * - 0 - failed to acquire the lock atomically;
1843 * - >0 - acquired the lock, return value is vpid of the top_waiter
1847 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1848 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1849 union futex_key
*key2
, struct futex_pi_state
**ps
,
1850 struct task_struct
**exiting
, int set_waiters
)
1852 struct futex_q
*top_waiter
= NULL
;
1856 if (get_futex_value_locked(&curval
, pifutex
))
1859 if (unlikely(should_fail_futex(true)))
1863 * Find the top_waiter and determine if there are additional waiters.
1864 * If the caller intends to requeue more than 1 waiter to pifutex,
1865 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1866 * as we have means to handle the possible fault. If not, don't set
1867 * the bit unecessarily as it will force the subsequent unlock to enter
1870 top_waiter
= futex_top_waiter(hb1
, key1
);
1872 /* There are no waiters, nothing for us to do. */
1876 /* Ensure we requeue to the expected futex. */
1877 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1881 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1882 * the contended case or if set_waiters is 1. The pi_state is returned
1883 * in ps in contended cases.
1885 vpid
= task_pid_vnr(top_waiter
->task
);
1886 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1887 exiting
, set_waiters
);
1889 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1896 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1897 * @uaddr1: source futex user address
1898 * @flags: futex flags (FLAGS_SHARED, etc.)
1899 * @uaddr2: target futex user address
1900 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1901 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1902 * @cmpval: @uaddr1 expected value (or %NULL)
1903 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1904 * pi futex (pi to pi requeue is not supported)
1906 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1907 * uaddr2 atomically on behalf of the top waiter.
1910 * - >=0 - on success, the number of tasks requeued or woken;
1913 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1914 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1915 u32
*cmpval
, int requeue_pi
)
1917 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1918 int task_count
= 0, ret
;
1919 struct futex_pi_state
*pi_state
= NULL
;
1920 struct futex_hash_bucket
*hb1
, *hb2
;
1921 struct futex_q
*this, *next
;
1922 DEFINE_WAKE_Q(wake_q
);
1924 if (nr_wake
< 0 || nr_requeue
< 0)
1928 * When PI not supported: return -ENOSYS if requeue_pi is true,
1929 * consequently the compiler knows requeue_pi is always false past
1930 * this point which will optimize away all the conditional code
1933 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1938 * Requeue PI only works on two distinct uaddrs. This
1939 * check is only valid for private futexes. See below.
1941 if (uaddr1
== uaddr2
)
1945 * requeue_pi requires a pi_state, try to allocate it now
1946 * without any locks in case it fails.
1948 if (refill_pi_state_cache())
1951 * requeue_pi must wake as many tasks as it can, up to nr_wake
1952 * + nr_requeue, since it acquires the rt_mutex prior to
1953 * returning to userspace, so as to not leave the rt_mutex with
1954 * waiters and no owner. However, second and third wake-ups
1955 * cannot be predicted as they involve race conditions with the
1956 * first wake and a fault while looking up the pi_state. Both
1957 * pthread_cond_signal() and pthread_cond_broadcast() should
1965 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1966 if (unlikely(ret
!= 0))
1968 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1969 requeue_pi
? FUTEX_WRITE
: FUTEX_READ
);
1970 if (unlikely(ret
!= 0))
1974 * The check above which compares uaddrs is not sufficient for
1975 * shared futexes. We need to compare the keys:
1977 if (requeue_pi
&& match_futex(&key1
, &key2
))
1980 hb1
= hash_futex(&key1
);
1981 hb2
= hash_futex(&key2
);
1984 hb_waiters_inc(hb2
);
1985 double_lock_hb(hb1
, hb2
);
1987 if (likely(cmpval
!= NULL
)) {
1990 ret
= get_futex_value_locked(&curval
, uaddr1
);
1992 if (unlikely(ret
)) {
1993 double_unlock_hb(hb1
, hb2
);
1994 hb_waiters_dec(hb2
);
1996 ret
= get_user(curval
, uaddr1
);
2000 if (!(flags
& FLAGS_SHARED
))
2005 if (curval
!= *cmpval
) {
2011 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2012 struct task_struct
*exiting
= NULL
;
2015 * Attempt to acquire uaddr2 and wake the top waiter. If we
2016 * intend to requeue waiters, force setting the FUTEX_WAITERS
2017 * bit. We force this here where we are able to easily handle
2018 * faults rather in the requeue loop below.
2020 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2022 &exiting
, nr_requeue
);
2025 * At this point the top_waiter has either taken uaddr2 or is
2026 * waiting on it. If the former, then the pi_state will not
2027 * exist yet, look it up one more time to ensure we have a
2028 * reference to it. If the lock was taken, ret contains the
2029 * vpid of the top waiter task.
2030 * If the lock was not taken, we have pi_state and an initial
2031 * refcount on it. In case of an error we have nothing.
2037 * If we acquired the lock, then the user space value
2038 * of uaddr2 should be vpid. It cannot be changed by
2039 * the top waiter as it is blocked on hb2 lock if it
2040 * tries to do so. If something fiddled with it behind
2041 * our back the pi state lookup might unearth it. So
2042 * we rather use the known value than rereading and
2043 * handing potential crap to lookup_pi_state.
2045 * If that call succeeds then we have pi_state and an
2046 * initial refcount on it.
2048 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2049 &pi_state
, &exiting
);
2054 /* We hold a reference on the pi state. */
2057 /* If the above failed, then pi_state is NULL */
2059 double_unlock_hb(hb1
, hb2
);
2060 hb_waiters_dec(hb2
);
2061 ret
= fault_in_user_writeable(uaddr2
);
2068 * Two reasons for this:
2069 * - EBUSY: Owner is exiting and we just wait for the
2071 * - EAGAIN: The user space value changed.
2073 double_unlock_hb(hb1
, hb2
);
2074 hb_waiters_dec(hb2
);
2076 * Handle the case where the owner is in the middle of
2077 * exiting. Wait for the exit to complete otherwise
2078 * this task might loop forever, aka. live lock.
2080 wait_for_owner_exiting(ret
, exiting
);
2088 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2089 if (task_count
- nr_wake
>= nr_requeue
)
2092 if (!match_futex(&this->key
, &key1
))
2096 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2097 * be paired with each other and no other futex ops.
2099 * We should never be requeueing a futex_q with a pi_state,
2100 * which is awaiting a futex_unlock_pi().
2102 if ((requeue_pi
&& !this->rt_waiter
) ||
2103 (!requeue_pi
&& this->rt_waiter
) ||
2110 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2111 * lock, we already woke the top_waiter. If not, it will be
2112 * woken by futex_unlock_pi().
2114 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2115 mark_wake_futex(&wake_q
, this);
2119 /* Ensure we requeue to the expected futex for requeue_pi. */
2120 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2126 * Requeue nr_requeue waiters and possibly one more in the case
2127 * of requeue_pi if we couldn't acquire the lock atomically.
2131 * Prepare the waiter to take the rt_mutex. Take a
2132 * refcount on the pi_state and store the pointer in
2133 * the futex_q object of the waiter.
2135 get_pi_state(pi_state
);
2136 this->pi_state
= pi_state
;
2137 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2142 * We got the lock. We do neither drop the
2143 * refcount on pi_state nor clear
2144 * this->pi_state because the waiter needs the
2145 * pi_state for cleaning up the user space
2146 * value. It will drop the refcount after
2149 requeue_pi_wake_futex(this, &key2
, hb2
);
2153 * rt_mutex_start_proxy_lock() detected a
2154 * potential deadlock when we tried to queue
2155 * that waiter. Drop the pi_state reference
2156 * which we took above and remove the pointer
2157 * to the state from the waiters futex_q
2160 this->pi_state
= NULL
;
2161 put_pi_state(pi_state
);
2163 * We stop queueing more waiters and let user
2164 * space deal with the mess.
2169 requeue_futex(this, hb1
, hb2
, &key2
);
2173 * We took an extra initial reference to the pi_state either
2174 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2175 * need to drop it here again.
2177 put_pi_state(pi_state
);
2180 double_unlock_hb(hb1
, hb2
);
2182 hb_waiters_dec(hb2
);
2183 return ret
? ret
: task_count
;
2186 /* The key must be already stored in q->key. */
2187 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2188 __acquires(&hb
->lock
)
2190 struct futex_hash_bucket
*hb
;
2192 hb
= hash_futex(&q
->key
);
2195 * Increment the counter before taking the lock so that
2196 * a potential waker won't miss a to-be-slept task that is
2197 * waiting for the spinlock. This is safe as all queue_lock()
2198 * users end up calling queue_me(). Similarly, for housekeeping,
2199 * decrement the counter at queue_unlock() when some error has
2200 * occurred and we don't end up adding the task to the list.
2202 hb_waiters_inc(hb
); /* implies smp_mb(); (A) */
2204 q
->lock_ptr
= &hb
->lock
;
2206 spin_lock(&hb
->lock
);
2211 queue_unlock(struct futex_hash_bucket
*hb
)
2212 __releases(&hb
->lock
)
2214 spin_unlock(&hb
->lock
);
2218 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2223 * The priority used to register this element is
2224 * - either the real thread-priority for the real-time threads
2225 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2226 * - or MAX_RT_PRIO for non-RT threads.
2227 * Thus, all RT-threads are woken first in priority order, and
2228 * the others are woken last, in FIFO order.
2230 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2232 plist_node_init(&q
->list
, prio
);
2233 plist_add(&q
->list
, &hb
->chain
);
2238 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2239 * @q: The futex_q to enqueue
2240 * @hb: The destination hash bucket
2242 * The hb->lock must be held by the caller, and is released here. A call to
2243 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2244 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2245 * or nothing if the unqueue is done as part of the wake process and the unqueue
2246 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2249 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2250 __releases(&hb
->lock
)
2253 spin_unlock(&hb
->lock
);
2257 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2258 * @q: The futex_q to unqueue
2260 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2261 * be paired with exactly one earlier call to queue_me().
2264 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2265 * - 0 - if the futex_q was already removed by the waking thread
2267 static int unqueue_me(struct futex_q
*q
)
2269 spinlock_t
*lock_ptr
;
2272 /* In the common case we don't take the spinlock, which is nice. */
2275 * q->lock_ptr can change between this read and the following spin_lock.
2276 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2277 * optimizing lock_ptr out of the logic below.
2279 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2280 if (lock_ptr
!= NULL
) {
2281 spin_lock(lock_ptr
);
2283 * q->lock_ptr can change between reading it and
2284 * spin_lock(), causing us to take the wrong lock. This
2285 * corrects the race condition.
2287 * Reasoning goes like this: if we have the wrong lock,
2288 * q->lock_ptr must have changed (maybe several times)
2289 * between reading it and the spin_lock(). It can
2290 * change again after the spin_lock() but only if it was
2291 * already changed before the spin_lock(). It cannot,
2292 * however, change back to the original value. Therefore
2293 * we can detect whether we acquired the correct lock.
2295 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2296 spin_unlock(lock_ptr
);
2301 BUG_ON(q
->pi_state
);
2303 spin_unlock(lock_ptr
);
2311 * PI futexes can not be requeued and must remove themself from the
2312 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2315 static void unqueue_me_pi(struct futex_q
*q
)
2316 __releases(q
->lock_ptr
)
2320 BUG_ON(!q
->pi_state
);
2321 put_pi_state(q
->pi_state
);
2324 spin_unlock(q
->lock_ptr
);
2327 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2328 struct task_struct
*argowner
)
2330 struct futex_pi_state
*pi_state
= q
->pi_state
;
2331 u32 uval
, curval
, newval
;
2332 struct task_struct
*oldowner
, *newowner
;
2336 lockdep_assert_held(q
->lock_ptr
);
2338 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2340 oldowner
= pi_state
->owner
;
2343 * We are here because either:
2345 * - we stole the lock and pi_state->owner needs updating to reflect
2346 * that (@argowner == current),
2350 * - someone stole our lock and we need to fix things to point to the
2351 * new owner (@argowner == NULL).
2353 * Either way, we have to replace the TID in the user space variable.
2354 * This must be atomic as we have to preserve the owner died bit here.
2356 * Note: We write the user space value _before_ changing the pi_state
2357 * because we can fault here. Imagine swapped out pages or a fork
2358 * that marked all the anonymous memory readonly for cow.
2360 * Modifying pi_state _before_ the user space value would leave the
2361 * pi_state in an inconsistent state when we fault here, because we
2362 * need to drop the locks to handle the fault. This might be observed
2363 * in the PID check in lookup_pi_state.
2367 if (oldowner
!= current
) {
2369 * We raced against a concurrent self; things are
2370 * already fixed up. Nothing to do.
2376 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2377 /* We got the lock after all, nothing to fix. */
2383 * Since we just failed the trylock; there must be an owner.
2385 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2388 WARN_ON_ONCE(argowner
!= current
);
2389 if (oldowner
== current
) {
2391 * We raced against a concurrent self; things are
2392 * already fixed up. Nothing to do.
2397 newowner
= argowner
;
2400 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2402 if (!pi_state
->owner
)
2403 newtid
|= FUTEX_OWNER_DIED
;
2405 err
= get_futex_value_locked(&uval
, uaddr
);
2410 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2412 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2422 * We fixed up user space. Now we need to fix the pi_state
2425 if (pi_state
->owner
!= NULL
) {
2426 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2427 WARN_ON(list_empty(&pi_state
->list
));
2428 list_del_init(&pi_state
->list
);
2429 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2432 pi_state
->owner
= newowner
;
2434 raw_spin_lock(&newowner
->pi_lock
);
2435 WARN_ON(!list_empty(&pi_state
->list
));
2436 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2437 raw_spin_unlock(&newowner
->pi_lock
);
2438 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2443 * In order to reschedule or handle a page fault, we need to drop the
2444 * locks here. In the case of a fault, this gives the other task
2445 * (either the highest priority waiter itself or the task which stole
2446 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2447 * are back from handling the fault we need to check the pi_state after
2448 * reacquiring the locks and before trying to do another fixup. When
2449 * the fixup has been done already we simply return.
2451 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2452 * drop hb->lock since the caller owns the hb -> futex_q relation.
2453 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2456 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2457 spin_unlock(q
->lock_ptr
);
2461 ret
= fault_in_user_writeable(uaddr
);
2475 spin_lock(q
->lock_ptr
);
2476 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2479 * Check if someone else fixed it for us:
2481 if (pi_state
->owner
!= oldowner
) {
2492 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2496 static long futex_wait_restart(struct restart_block
*restart
);
2499 * fixup_owner() - Post lock pi_state and corner case management
2500 * @uaddr: user address of the futex
2501 * @q: futex_q (contains pi_state and access to the rt_mutex)
2502 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2504 * After attempting to lock an rt_mutex, this function is called to cleanup
2505 * the pi_state owner as well as handle race conditions that may allow us to
2506 * acquire the lock. Must be called with the hb lock held.
2509 * - 1 - success, lock taken;
2510 * - 0 - success, lock not taken;
2511 * - <0 - on error (-EFAULT)
2513 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2519 * Got the lock. We might not be the anticipated owner if we
2520 * did a lock-steal - fix up the PI-state in that case:
2522 * Speculative pi_state->owner read (we don't hold wait_lock);
2523 * since we own the lock pi_state->owner == current is the
2524 * stable state, anything else needs more attention.
2526 if (q
->pi_state
->owner
!= current
)
2527 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2528 return ret
? ret
: locked
;
2532 * If we didn't get the lock; check if anybody stole it from us. In
2533 * that case, we need to fix up the uval to point to them instead of
2534 * us, otherwise bad things happen. [10]
2536 * Another speculative read; pi_state->owner == current is unstable
2537 * but needs our attention.
2539 if (q
->pi_state
->owner
== current
) {
2540 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2545 * Paranoia check. If we did not take the lock, then we should not be
2546 * the owner of the rt_mutex.
2548 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2549 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2550 "pi-state %p\n", ret
,
2551 q
->pi_state
->pi_mutex
.owner
,
2552 q
->pi_state
->owner
);
2559 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2560 * @hb: the futex hash bucket, must be locked by the caller
2561 * @q: the futex_q to queue up on
2562 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2564 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2565 struct hrtimer_sleeper
*timeout
)
2568 * The task state is guaranteed to be set before another task can
2569 * wake it. set_current_state() is implemented using smp_store_mb() and
2570 * queue_me() calls spin_unlock() upon completion, both serializing
2571 * access to the hash list and forcing another memory barrier.
2573 set_current_state(TASK_INTERRUPTIBLE
);
2578 hrtimer_sleeper_start_expires(timeout
, HRTIMER_MODE_ABS
);
2581 * If we have been removed from the hash list, then another task
2582 * has tried to wake us, and we can skip the call to schedule().
2584 if (likely(!plist_node_empty(&q
->list
))) {
2586 * If the timer has already expired, current will already be
2587 * flagged for rescheduling. Only call schedule if there
2588 * is no timeout, or if it has yet to expire.
2590 if (!timeout
|| timeout
->task
)
2591 freezable_schedule();
2593 __set_current_state(TASK_RUNNING
);
2597 * futex_wait_setup() - Prepare to wait on a futex
2598 * @uaddr: the futex userspace address
2599 * @val: the expected value
2600 * @flags: futex flags (FLAGS_SHARED, etc.)
2601 * @q: the associated futex_q
2602 * @hb: storage for hash_bucket pointer to be returned to caller
2604 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2605 * compare it with the expected value. Handle atomic faults internally.
2606 * Return with the hb lock held and a q.key reference on success, and unlocked
2607 * with no q.key reference on failure.
2610 * - 0 - uaddr contains val and hb has been locked;
2611 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2613 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2614 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2620 * Access the page AFTER the hash-bucket is locked.
2621 * Order is important:
2623 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2624 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2626 * The basic logical guarantee of a futex is that it blocks ONLY
2627 * if cond(var) is known to be true at the time of blocking, for
2628 * any cond. If we locked the hash-bucket after testing *uaddr, that
2629 * would open a race condition where we could block indefinitely with
2630 * cond(var) false, which would violate the guarantee.
2632 * On the other hand, we insert q and release the hash-bucket only
2633 * after testing *uaddr. This guarantees that futex_wait() will NOT
2634 * absorb a wakeup if *uaddr does not match the desired values
2635 * while the syscall executes.
2638 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, FUTEX_READ
);
2639 if (unlikely(ret
!= 0))
2643 *hb
= queue_lock(q
);
2645 ret
= get_futex_value_locked(&uval
, uaddr
);
2650 ret
= get_user(uval
, uaddr
);
2654 if (!(flags
& FLAGS_SHARED
))
2668 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2669 ktime_t
*abs_time
, u32 bitset
)
2671 struct hrtimer_sleeper timeout
, *to
;
2672 struct restart_block
*restart
;
2673 struct futex_hash_bucket
*hb
;
2674 struct futex_q q
= futex_q_init
;
2681 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
2682 current
->timer_slack_ns
);
2685 * Prepare to wait on uaddr. On success, holds hb lock and increments
2688 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2692 /* queue_me and wait for wakeup, timeout, or a signal. */
2693 futex_wait_queue_me(hb
, &q
, to
);
2695 /* If we were woken (and unqueued), we succeeded, whatever. */
2697 /* unqueue_me() drops q.key ref */
2698 if (!unqueue_me(&q
))
2701 if (to
&& !to
->task
)
2705 * We expect signal_pending(current), but we might be the
2706 * victim of a spurious wakeup as well.
2708 if (!signal_pending(current
))
2715 restart
= ¤t
->restart_block
;
2716 restart
->fn
= futex_wait_restart
;
2717 restart
->futex
.uaddr
= uaddr
;
2718 restart
->futex
.val
= val
;
2719 restart
->futex
.time
= *abs_time
;
2720 restart
->futex
.bitset
= bitset
;
2721 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2723 ret
= -ERESTART_RESTARTBLOCK
;
2727 hrtimer_cancel(&to
->timer
);
2728 destroy_hrtimer_on_stack(&to
->timer
);
2734 static long futex_wait_restart(struct restart_block
*restart
)
2736 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2737 ktime_t t
, *tp
= NULL
;
2739 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2740 t
= restart
->futex
.time
;
2743 restart
->fn
= do_no_restart_syscall
;
2745 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2746 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2751 * Userspace tried a 0 -> TID atomic transition of the futex value
2752 * and failed. The kernel side here does the whole locking operation:
2753 * if there are waiters then it will block as a consequence of relying
2754 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2755 * a 0 value of the futex too.).
2757 * Also serves as futex trylock_pi()'ing, and due semantics.
2759 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2760 ktime_t
*time
, int trylock
)
2762 struct hrtimer_sleeper timeout
, *to
;
2763 struct futex_pi_state
*pi_state
= NULL
;
2764 struct task_struct
*exiting
= NULL
;
2765 struct rt_mutex_waiter rt_waiter
;
2766 struct futex_hash_bucket
*hb
;
2767 struct futex_q q
= futex_q_init
;
2770 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2773 if (refill_pi_state_cache())
2776 to
= futex_setup_timer(time
, &timeout
, FLAGS_CLOCKRT
, 0);
2779 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, FUTEX_WRITE
);
2780 if (unlikely(ret
!= 0))
2784 hb
= queue_lock(&q
);
2786 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2788 if (unlikely(ret
)) {
2790 * Atomic work succeeded and we got the lock,
2791 * or failed. Either way, we do _not_ block.
2795 /* We got the lock. */
2797 goto out_unlock_put_key
;
2803 * Two reasons for this:
2804 * - EBUSY: Task is exiting and we just wait for the
2806 * - EAGAIN: The user space value changed.
2810 * Handle the case where the owner is in the middle of
2811 * exiting. Wait for the exit to complete otherwise
2812 * this task might loop forever, aka. live lock.
2814 wait_for_owner_exiting(ret
, exiting
);
2818 goto out_unlock_put_key
;
2822 WARN_ON(!q
.pi_state
);
2825 * Only actually queue now that the atomic ops are done:
2830 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2831 /* Fixup the trylock return value: */
2832 ret
= ret
? 0 : -EWOULDBLOCK
;
2836 rt_mutex_init_waiter(&rt_waiter
);
2839 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2840 * hold it while doing rt_mutex_start_proxy(), because then it will
2841 * include hb->lock in the blocking chain, even through we'll not in
2842 * fact hold it while blocking. This will lead it to report -EDEADLK
2843 * and BUG when futex_unlock_pi() interleaves with this.
2845 * Therefore acquire wait_lock while holding hb->lock, but drop the
2846 * latter before calling __rt_mutex_start_proxy_lock(). This
2847 * interleaves with futex_unlock_pi() -- which does a similar lock
2848 * handoff -- such that the latter can observe the futex_q::pi_state
2849 * before __rt_mutex_start_proxy_lock() is done.
2851 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2852 spin_unlock(q
.lock_ptr
);
2854 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2855 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2856 * it sees the futex_q::pi_state.
2858 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2859 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2868 hrtimer_sleeper_start_expires(to
, HRTIMER_MODE_ABS
);
2870 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2873 spin_lock(q
.lock_ptr
);
2875 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2876 * first acquire the hb->lock before removing the lock from the
2877 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2880 * In particular; it is important that futex_unlock_pi() can not
2881 * observe this inconsistency.
2883 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2888 * Fixup the pi_state owner and possibly acquire the lock if we
2891 res
= fixup_owner(uaddr
, &q
, !ret
);
2893 * If fixup_owner() returned an error, proprogate that. If it acquired
2894 * the lock, clear our -ETIMEDOUT or -EINTR.
2897 ret
= (res
< 0) ? res
: 0;
2900 * If fixup_owner() faulted and was unable to handle the fault, unlock
2901 * it and return the fault to userspace.
2903 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2904 pi_state
= q
.pi_state
;
2905 get_pi_state(pi_state
);
2908 /* Unqueue and drop the lock */
2912 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2913 put_pi_state(pi_state
);
2923 hrtimer_cancel(&to
->timer
);
2924 destroy_hrtimer_on_stack(&to
->timer
);
2926 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2931 ret
= fault_in_user_writeable(uaddr
);
2935 if (!(flags
& FLAGS_SHARED
))
2942 * Userspace attempted a TID -> 0 atomic transition, and failed.
2943 * This is the in-kernel slowpath: we look up the PI state (if any),
2944 * and do the rt-mutex unlock.
2946 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2948 u32 curval
, uval
, vpid
= task_pid_vnr(current
);
2949 union futex_key key
= FUTEX_KEY_INIT
;
2950 struct futex_hash_bucket
*hb
;
2951 struct futex_q
*top_waiter
;
2954 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2958 if (get_user(uval
, uaddr
))
2961 * We release only a lock we actually own:
2963 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2966 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_WRITE
);
2970 hb
= hash_futex(&key
);
2971 spin_lock(&hb
->lock
);
2974 * Check waiters first. We do not trust user space values at
2975 * all and we at least want to know if user space fiddled
2976 * with the futex value instead of blindly unlocking.
2978 top_waiter
= futex_top_waiter(hb
, &key
);
2980 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
2987 * If current does not own the pi_state then the futex is
2988 * inconsistent and user space fiddled with the futex value.
2990 if (pi_state
->owner
!= current
)
2993 get_pi_state(pi_state
);
2995 * By taking wait_lock while still holding hb->lock, we ensure
2996 * there is no point where we hold neither; and therefore
2997 * wake_futex_pi() must observe a state consistent with what we
3000 * In particular; this forces __rt_mutex_start_proxy() to
3001 * complete such that we're guaranteed to observe the
3002 * rt_waiter. Also see the WARN in wake_futex_pi().
3004 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3005 spin_unlock(&hb
->lock
);
3007 /* drops pi_state->pi_mutex.wait_lock */
3008 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3010 put_pi_state(pi_state
);
3013 * Success, we're done! No tricky corner cases.
3018 * The atomic access to the futex value generated a
3019 * pagefault, so retry the user-access and the wakeup:
3024 * A unconditional UNLOCK_PI op raced against a waiter
3025 * setting the FUTEX_WAITERS bit. Try again.
3030 * wake_futex_pi has detected invalid state. Tell user
3037 * We have no kernel internal state, i.e. no waiters in the
3038 * kernel. Waiters which are about to queue themselves are stuck
3039 * on hb->lock. So we can safely ignore them. We do neither
3040 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3043 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3044 spin_unlock(&hb
->lock
);
3059 * If uval has changed, let user space handle it.
3061 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3064 spin_unlock(&hb
->lock
);
3074 ret
= fault_in_user_writeable(uaddr
);
3082 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3083 * @hb: the hash_bucket futex_q was original enqueued on
3084 * @q: the futex_q woken while waiting to be requeued
3085 * @key2: the futex_key of the requeue target futex
3086 * @timeout: the timeout associated with the wait (NULL if none)
3088 * Detect if the task was woken on the initial futex as opposed to the requeue
3089 * target futex. If so, determine if it was a timeout or a signal that caused
3090 * the wakeup and return the appropriate error code to the caller. Must be
3091 * called with the hb lock held.
3094 * - 0 = no early wakeup detected;
3095 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3098 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3099 struct futex_q
*q
, union futex_key
*key2
,
3100 struct hrtimer_sleeper
*timeout
)
3105 * With the hb lock held, we avoid races while we process the wakeup.
3106 * We only need to hold hb (and not hb2) to ensure atomicity as the
3107 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3108 * It can't be requeued from uaddr2 to something else since we don't
3109 * support a PI aware source futex for requeue.
3111 if (!match_futex(&q
->key
, key2
)) {
3112 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3114 * We were woken prior to requeue by a timeout or a signal.
3115 * Unqueue the futex_q and determine which it was.
3117 plist_del(&q
->list
, &hb
->chain
);
3120 /* Handle spurious wakeups gracefully */
3122 if (timeout
&& !timeout
->task
)
3124 else if (signal_pending(current
))
3125 ret
= -ERESTARTNOINTR
;
3131 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3132 * @uaddr: the futex we initially wait on (non-pi)
3133 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3134 * the same type, no requeueing from private to shared, etc.
3135 * @val: the expected value of uaddr
3136 * @abs_time: absolute timeout
3137 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3138 * @uaddr2: the pi futex we will take prior to returning to user-space
3140 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3141 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3142 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3143 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3144 * without one, the pi logic would not know which task to boost/deboost, if
3145 * there was a need to.
3147 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3148 * via the following--
3149 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3150 * 2) wakeup on uaddr2 after a requeue
3154 * If 3, cleanup and return -ERESTARTNOINTR.
3156 * If 2, we may then block on trying to take the rt_mutex and return via:
3157 * 5) successful lock
3160 * 8) other lock acquisition failure
3162 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3164 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3170 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3171 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3174 struct hrtimer_sleeper timeout
, *to
;
3175 struct futex_pi_state
*pi_state
= NULL
;
3176 struct rt_mutex_waiter rt_waiter
;
3177 struct futex_hash_bucket
*hb
;
3178 union futex_key key2
= FUTEX_KEY_INIT
;
3179 struct futex_q q
= futex_q_init
;
3182 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3185 if (uaddr
== uaddr2
)
3191 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
3192 current
->timer_slack_ns
);
3195 * The waiter is allocated on our stack, manipulated by the requeue
3196 * code while we sleep on uaddr.
3198 rt_mutex_init_waiter(&rt_waiter
);
3200 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
3201 if (unlikely(ret
!= 0))
3205 q
.rt_waiter
= &rt_waiter
;
3206 q
.requeue_pi_key
= &key2
;
3209 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3212 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3217 * The check above which compares uaddrs is not sufficient for
3218 * shared futexes. We need to compare the keys:
3220 if (match_futex(&q
.key
, &key2
)) {
3226 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3227 futex_wait_queue_me(hb
, &q
, to
);
3229 spin_lock(&hb
->lock
);
3230 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3231 spin_unlock(&hb
->lock
);
3236 * In order for us to be here, we know our q.key == key2, and since
3237 * we took the hb->lock above, we also know that futex_requeue() has
3238 * completed and we no longer have to concern ourselves with a wakeup
3239 * race with the atomic proxy lock acquisition by the requeue code. The
3240 * futex_requeue dropped our key1 reference and incremented our key2
3244 /* Check if the requeue code acquired the second futex for us. */
3247 * Got the lock. We might not be the anticipated owner if we
3248 * did a lock-steal - fix up the PI-state in that case.
3250 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3251 spin_lock(q
.lock_ptr
);
3252 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3253 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3254 pi_state
= q
.pi_state
;
3255 get_pi_state(pi_state
);
3258 * Drop the reference to the pi state which
3259 * the requeue_pi() code acquired for us.
3261 put_pi_state(q
.pi_state
);
3262 spin_unlock(q
.lock_ptr
);
3265 struct rt_mutex
*pi_mutex
;
3268 * We have been woken up by futex_unlock_pi(), a timeout, or a
3269 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3272 WARN_ON(!q
.pi_state
);
3273 pi_mutex
= &q
.pi_state
->pi_mutex
;
3274 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3276 spin_lock(q
.lock_ptr
);
3277 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3280 debug_rt_mutex_free_waiter(&rt_waiter
);
3282 * Fixup the pi_state owner and possibly acquire the lock if we
3285 res
= fixup_owner(uaddr2
, &q
, !ret
);
3287 * If fixup_owner() returned an error, proprogate that. If it
3288 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3291 ret
= (res
< 0) ? res
: 0;
3294 * If fixup_pi_state_owner() faulted and was unable to handle
3295 * the fault, unlock the rt_mutex and return the fault to
3298 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3299 pi_state
= q
.pi_state
;
3300 get_pi_state(pi_state
);
3303 /* Unqueue and drop the lock. */
3308 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3309 put_pi_state(pi_state
);
3312 if (ret
== -EINTR
) {
3314 * We've already been requeued, but cannot restart by calling
3315 * futex_lock_pi() directly. We could restart this syscall, but
3316 * it would detect that the user space "val" changed and return
3317 * -EWOULDBLOCK. Save the overhead of the restart and return
3318 * -EWOULDBLOCK directly.
3325 hrtimer_cancel(&to
->timer
);
3326 destroy_hrtimer_on_stack(&to
->timer
);
3332 * Support for robust futexes: the kernel cleans up held futexes at
3335 * Implementation: user-space maintains a per-thread list of locks it
3336 * is holding. Upon do_exit(), the kernel carefully walks this list,
3337 * and marks all locks that are owned by this thread with the
3338 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3339 * always manipulated with the lock held, so the list is private and
3340 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3341 * field, to allow the kernel to clean up if the thread dies after
3342 * acquiring the lock, but just before it could have added itself to
3343 * the list. There can only be one such pending lock.
3347 * sys_set_robust_list() - Set the robust-futex list head of a task
3348 * @head: pointer to the list-head
3349 * @len: length of the list-head, as userspace expects
3351 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3354 if (!futex_cmpxchg_enabled
)
3357 * The kernel knows only one size for now:
3359 if (unlikely(len
!= sizeof(*head
)))
3362 current
->robust_list
= head
;
3368 * sys_get_robust_list() - Get the robust-futex list head of a task
3369 * @pid: pid of the process [zero for current task]
3370 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3371 * @len_ptr: pointer to a length field, the kernel fills in the header size
3373 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3374 struct robust_list_head __user
* __user
*, head_ptr
,
3375 size_t __user
*, len_ptr
)
3377 struct robust_list_head __user
*head
;
3379 struct task_struct
*p
;
3381 if (!futex_cmpxchg_enabled
)
3390 p
= find_task_by_vpid(pid
);
3396 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3399 head
= p
->robust_list
;
3402 if (put_user(sizeof(*head
), len_ptr
))
3404 return put_user(head
, head_ptr
);
3412 /* Constants for the pending_op argument of handle_futex_death */
3413 #define HANDLE_DEATH_PENDING true
3414 #define HANDLE_DEATH_LIST false
3417 * Process a futex-list entry, check whether it's owned by the
3418 * dying task, and do notification if so:
3420 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3421 bool pi
, bool pending_op
)
3423 u32 uval
, nval
, mval
;
3426 /* Futex address must be 32bit aligned */
3427 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3431 if (get_user(uval
, uaddr
))
3435 * Special case for regular (non PI) futexes. The unlock path in
3436 * user space has two race scenarios:
3438 * 1. The unlock path releases the user space futex value and
3439 * before it can execute the futex() syscall to wake up
3440 * waiters it is killed.
3442 * 2. A woken up waiter is killed before it can acquire the
3443 * futex in user space.
3445 * In both cases the TID validation below prevents a wakeup of
3446 * potential waiters which can cause these waiters to block
3449 * In both cases the following conditions are met:
3451 * 1) task->robust_list->list_op_pending != NULL
3452 * @pending_op == true
3453 * 2) User space futex value == 0
3454 * 3) Regular futex: @pi == false
3456 * If these conditions are met, it is safe to attempt waking up a
3457 * potential waiter without touching the user space futex value and
3458 * trying to set the OWNER_DIED bit. The user space futex value is
3459 * uncontended and the rest of the user space mutex state is
3460 * consistent, so a woken waiter will just take over the
3461 * uncontended futex. Setting the OWNER_DIED bit would create
3462 * inconsistent state and malfunction of the user space owner died
3465 if (pending_op
&& !pi
&& !uval
) {
3466 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3470 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3474 * Ok, this dying thread is truly holding a futex
3475 * of interest. Set the OWNER_DIED bit atomically
3476 * via cmpxchg, and if the value had FUTEX_WAITERS
3477 * set, wake up a waiter (if any). (We have to do a
3478 * futex_wake() even if OWNER_DIED is already set -
3479 * to handle the rare but possible case of recursive
3480 * thread-death.) The rest of the cleanup is done in
3483 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3486 * We are not holding a lock here, but we want to have
3487 * the pagefault_disable/enable() protection because
3488 * we want to handle the fault gracefully. If the
3489 * access fails we try to fault in the futex with R/W
3490 * verification via get_user_pages. get_user() above
3491 * does not guarantee R/W access. If that fails we
3492 * give up and leave the futex locked.
3494 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3497 if (fault_in_user_writeable(uaddr
))
3515 * Wake robust non-PI futexes here. The wakeup of
3516 * PI futexes happens in exit_pi_state():
3518 if (!pi
&& (uval
& FUTEX_WAITERS
))
3519 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3525 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3527 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3528 struct robust_list __user
* __user
*head
,
3531 unsigned long uentry
;
3533 if (get_user(uentry
, (unsigned long __user
*)head
))
3536 *entry
= (void __user
*)(uentry
& ~1UL);
3543 * Walk curr->robust_list (very carefully, it's a userspace list!)
3544 * and mark any locks found there dead, and notify any waiters.
3546 * We silently return on any sign of list-walking problem.
3548 static void exit_robust_list(struct task_struct
*curr
)
3550 struct robust_list_head __user
*head
= curr
->robust_list
;
3551 struct robust_list __user
*entry
, *next_entry
, *pending
;
3552 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3553 unsigned int next_pi
;
3554 unsigned long futex_offset
;
3557 if (!futex_cmpxchg_enabled
)
3561 * Fetch the list head (which was registered earlier, via
3562 * sys_set_robust_list()):
3564 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3567 * Fetch the relative futex offset:
3569 if (get_user(futex_offset
, &head
->futex_offset
))
3572 * Fetch any possibly pending lock-add first, and handle it
3575 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3578 next_entry
= NULL
; /* avoid warning with gcc */
3579 while (entry
!= &head
->list
) {
3581 * Fetch the next entry in the list before calling
3582 * handle_futex_death:
3584 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3586 * A pending lock might already be on the list, so
3587 * don't process it twice:
3589 if (entry
!= pending
) {
3590 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3591 curr
, pi
, HANDLE_DEATH_LIST
))
3599 * Avoid excessively long or circular lists:
3608 handle_futex_death((void __user
*)pending
+ futex_offset
,
3609 curr
, pip
, HANDLE_DEATH_PENDING
);
3613 static void futex_cleanup(struct task_struct
*tsk
)
3615 if (unlikely(tsk
->robust_list
)) {
3616 exit_robust_list(tsk
);
3617 tsk
->robust_list
= NULL
;
3620 #ifdef CONFIG_COMPAT
3621 if (unlikely(tsk
->compat_robust_list
)) {
3622 compat_exit_robust_list(tsk
);
3623 tsk
->compat_robust_list
= NULL
;
3627 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3628 exit_pi_state_list(tsk
);
3632 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3633 * @tsk: task to set the state on
3635 * Set the futex exit state of the task lockless. The futex waiter code
3636 * observes that state when a task is exiting and loops until the task has
3637 * actually finished the futex cleanup. The worst case for this is that the
3638 * waiter runs through the wait loop until the state becomes visible.
3640 * This is called from the recursive fault handling path in do_exit().
3642 * This is best effort. Either the futex exit code has run already or
3643 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3644 * take it over. If not, the problem is pushed back to user space. If the
3645 * futex exit code did not run yet, then an already queued waiter might
3646 * block forever, but there is nothing which can be done about that.
3648 void futex_exit_recursive(struct task_struct
*tsk
)
3650 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3651 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3652 mutex_unlock(&tsk
->futex_exit_mutex
);
3653 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3656 static void futex_cleanup_begin(struct task_struct
*tsk
)
3659 * Prevent various race issues against a concurrent incoming waiter
3660 * including live locks by forcing the waiter to block on
3661 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3662 * attach_to_pi_owner().
3664 mutex_lock(&tsk
->futex_exit_mutex
);
3667 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3669 * This ensures that all subsequent checks of tsk->futex_state in
3670 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3671 * tsk->pi_lock held.
3673 * It guarantees also that a pi_state which was queued right before
3674 * the state change under tsk->pi_lock by a concurrent waiter must
3675 * be observed in exit_pi_state_list().
3677 raw_spin_lock_irq(&tsk
->pi_lock
);
3678 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3679 raw_spin_unlock_irq(&tsk
->pi_lock
);
3682 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3685 * Lockless store. The only side effect is that an observer might
3686 * take another loop until it becomes visible.
3688 tsk
->futex_state
= state
;
3690 * Drop the exit protection. This unblocks waiters which observed
3691 * FUTEX_STATE_EXITING to reevaluate the state.
3693 mutex_unlock(&tsk
->futex_exit_mutex
);
3696 void futex_exec_release(struct task_struct
*tsk
)
3699 * The state handling is done for consistency, but in the case of
3700 * exec() there is no way to prevent futher damage as the PID stays
3701 * the same. But for the unlikely and arguably buggy case that a
3702 * futex is held on exec(), this provides at least as much state
3703 * consistency protection which is possible.
3705 futex_cleanup_begin(tsk
);
3708 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3709 * exec a new binary.
3711 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3714 void futex_exit_release(struct task_struct
*tsk
)
3716 futex_cleanup_begin(tsk
);
3718 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3721 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3722 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3724 int cmd
= op
& FUTEX_CMD_MASK
;
3725 unsigned int flags
= 0;
3727 if (!(op
& FUTEX_PRIVATE_FLAG
))
3728 flags
|= FLAGS_SHARED
;
3730 if (op
& FUTEX_CLOCK_REALTIME
) {
3731 flags
|= FLAGS_CLOCKRT
;
3732 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3733 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3739 case FUTEX_UNLOCK_PI
:
3740 case FUTEX_TRYLOCK_PI
:
3741 case FUTEX_WAIT_REQUEUE_PI
:
3742 case FUTEX_CMP_REQUEUE_PI
:
3743 if (!futex_cmpxchg_enabled
)
3749 val3
= FUTEX_BITSET_MATCH_ANY
;
3751 case FUTEX_WAIT_BITSET
:
3752 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3754 val3
= FUTEX_BITSET_MATCH_ANY
;
3756 case FUTEX_WAKE_BITSET
:
3757 return futex_wake(uaddr
, flags
, val
, val3
);
3759 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3760 case FUTEX_CMP_REQUEUE
:
3761 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3763 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3765 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3766 case FUTEX_UNLOCK_PI
:
3767 return futex_unlock_pi(uaddr
, flags
);
3768 case FUTEX_TRYLOCK_PI
:
3769 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3770 case FUTEX_WAIT_REQUEUE_PI
:
3771 val3
= FUTEX_BITSET_MATCH_ANY
;
3772 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3774 case FUTEX_CMP_REQUEUE_PI
:
3775 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3781 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3782 struct __kernel_timespec __user
*, utime
, u32 __user
*, uaddr2
,
3785 struct timespec64 ts
;
3786 ktime_t t
, *tp
= NULL
;
3788 int cmd
= op
& FUTEX_CMD_MASK
;
3790 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3791 cmd
== FUTEX_WAIT_BITSET
||
3792 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3793 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3795 if (get_timespec64(&ts
, utime
))
3797 if (!timespec64_valid(&ts
))
3800 t
= timespec64_to_ktime(ts
);
3801 if (cmd
== FUTEX_WAIT
)
3802 t
= ktime_add_safe(ktime_get(), t
);
3803 else if (!(op
& FUTEX_CLOCK_REALTIME
))
3804 t
= timens_ktime_to_host(CLOCK_MONOTONIC
, t
);
3808 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3809 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3811 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3812 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3813 val2
= (u32
) (unsigned long) utime
;
3815 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3818 #ifdef CONFIG_COMPAT
3820 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3823 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3824 compat_uptr_t __user
*head
, unsigned int *pi
)
3826 if (get_user(*uentry
, head
))
3829 *entry
= compat_ptr((*uentry
) & ~1);
3830 *pi
= (unsigned int)(*uentry
) & 1;
3835 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3836 compat_long_t futex_offset
)
3838 compat_uptr_t base
= ptr_to_compat(entry
);
3839 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3845 * Walk curr->robust_list (very carefully, it's a userspace list!)
3846 * and mark any locks found there dead, and notify any waiters.
3848 * We silently return on any sign of list-walking problem.
3850 static void compat_exit_robust_list(struct task_struct
*curr
)
3852 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3853 struct robust_list __user
*entry
, *next_entry
, *pending
;
3854 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3855 unsigned int next_pi
;
3856 compat_uptr_t uentry
, next_uentry
, upending
;
3857 compat_long_t futex_offset
;
3860 if (!futex_cmpxchg_enabled
)
3864 * Fetch the list head (which was registered earlier, via
3865 * sys_set_robust_list()):
3867 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3870 * Fetch the relative futex offset:
3872 if (get_user(futex_offset
, &head
->futex_offset
))
3875 * Fetch any possibly pending lock-add first, and handle it
3878 if (compat_fetch_robust_entry(&upending
, &pending
,
3879 &head
->list_op_pending
, &pip
))
3882 next_entry
= NULL
; /* avoid warning with gcc */
3883 while (entry
!= (struct robust_list __user
*) &head
->list
) {
3885 * Fetch the next entry in the list before calling
3886 * handle_futex_death:
3888 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
3889 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
3891 * A pending lock might already be on the list, so
3892 * dont process it twice:
3894 if (entry
!= pending
) {
3895 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
3897 if (handle_futex_death(uaddr
, curr
, pi
,
3903 uentry
= next_uentry
;
3907 * Avoid excessively long or circular lists:
3915 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
3917 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
3921 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
3922 struct compat_robust_list_head __user
*, head
,
3925 if (!futex_cmpxchg_enabled
)
3928 if (unlikely(len
!= sizeof(*head
)))
3931 current
->compat_robust_list
= head
;
3936 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3937 compat_uptr_t __user
*, head_ptr
,
3938 compat_size_t __user
*, len_ptr
)
3940 struct compat_robust_list_head __user
*head
;
3942 struct task_struct
*p
;
3944 if (!futex_cmpxchg_enabled
)
3953 p
= find_task_by_vpid(pid
);
3959 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3962 head
= p
->compat_robust_list
;
3965 if (put_user(sizeof(*head
), len_ptr
))
3967 return put_user(ptr_to_compat(head
), head_ptr
);
3974 #endif /* CONFIG_COMPAT */
3976 #ifdef CONFIG_COMPAT_32BIT_TIME
3977 SYSCALL_DEFINE6(futex_time32
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3978 struct old_timespec32 __user
*, utime
, u32 __user
*, uaddr2
,
3981 struct timespec64 ts
;
3982 ktime_t t
, *tp
= NULL
;
3984 int cmd
= op
& FUTEX_CMD_MASK
;
3986 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3987 cmd
== FUTEX_WAIT_BITSET
||
3988 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3989 if (get_old_timespec32(&ts
, utime
))
3991 if (!timespec64_valid(&ts
))
3994 t
= timespec64_to_ktime(ts
);
3995 if (cmd
== FUTEX_WAIT
)
3996 t
= ktime_add_safe(ktime_get(), t
);
3997 else if (!(op
& FUTEX_CLOCK_REALTIME
))
3998 t
= timens_ktime_to_host(CLOCK_MONOTONIC
, t
);
4001 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
4002 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
4003 val2
= (int) (unsigned long) utime
;
4005 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
4007 #endif /* CONFIG_COMPAT_32BIT_TIME */
4009 static void __init
futex_detect_cmpxchg(void)
4011 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4015 * This will fail and we want it. Some arch implementations do
4016 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4017 * functionality. We want to know that before we call in any
4018 * of the complex code paths. Also we want to prevent
4019 * registration of robust lists in that case. NULL is
4020 * guaranteed to fault and we get -EFAULT on functional
4021 * implementation, the non-functional ones will return
4024 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4025 futex_cmpxchg_enabled
= 1;
4029 static int __init
futex_init(void)
4031 unsigned int futex_shift
;
4034 #if CONFIG_BASE_SMALL
4035 futex_hashsize
= 16;
4037 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4040 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4042 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4044 futex_hashsize
, futex_hashsize
);
4045 futex_hashsize
= 1UL << futex_shift
;
4047 futex_detect_cmpxchg();
4049 for (i
= 0; i
< futex_hashsize
; i
++) {
4050 atomic_set(&futex_queues
[i
].waiters
, 0);
4051 plist_head_init(&futex_queues
[i
].chain
);
4052 spin_lock_init(&futex_queues
[i
].lock
);
4057 core_initcall(futex_init
);