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/freezer.h>
39 #include <linux/memblock.h>
40 #include <linux/fault-inject.h>
41 #include <linux/time_namespace.h>
43 #include <asm/futex.h>
45 #include "locking/rtmutex_common.h"
48 * READ this before attempting to hack on futexes!
50 * Basic futex operation and ordering guarantees
51 * =============================================
53 * The waiter reads the futex value in user space and calls
54 * futex_wait(). This function computes the hash bucket and acquires
55 * the hash bucket lock. After that it reads the futex user space value
56 * again and verifies that the data has not changed. If it has not changed
57 * it enqueues itself into the hash bucket, releases the hash bucket lock
60 * The waker side modifies the user space value of the futex and calls
61 * futex_wake(). This function computes the hash bucket and acquires the
62 * hash bucket lock. Then it looks for waiters on that futex in the hash
63 * bucket and wakes them.
65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
66 * the hb spinlock can be avoided and simply return. In order for this
67 * optimization to work, ordering guarantees must exist so that the waiter
68 * being added to the list is acknowledged when the list is concurrently being
69 * checked by the waker, avoiding scenarios like the following:
73 * sys_futex(WAIT, futex, val);
74 * futex_wait(futex, val);
77 * sys_futex(WAKE, futex);
82 * lock(hash_bucket(futex));
84 * unlock(hash_bucket(futex));
87 * This would cause the waiter on CPU 0 to wait forever because it
88 * missed the transition of the user space value from val to newval
89 * and the waker did not find the waiter in the hash bucket queue.
91 * The correct serialization ensures that a waiter either observes
92 * the changed user space value before blocking or is woken by a
97 * sys_futex(WAIT, futex, val);
98 * futex_wait(futex, val);
101 * smp_mb(); (A) <-- paired with -.
103 * lock(hash_bucket(futex)); |
107 * | sys_futex(WAKE, futex);
108 * | futex_wake(futex);
110 * `--------> smp_mb(); (B)
113 * unlock(hash_bucket(futex));
114 * schedule(); if (waiters)
115 * lock(hash_bucket(futex));
116 * else wake_waiters(futex);
117 * waiters--; (b) unlock(hash_bucket(futex));
119 * Where (A) orders the waiters increment and the futex value read through
120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
121 * to futex and the waiters read (see hb_waiters_pending()).
123 * This yields the following case (where X:=waiters, Y:=futex):
131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
132 * the guarantee that we cannot both miss the futex variable change and the
135 * Note that a new waiter is accounted for in (a) even when it is possible that
136 * the wait call can return error, in which case we backtrack from it in (b).
137 * Refer to the comment in queue_lock().
139 * Similarly, in order to account for waiters being requeued on another
140 * address we always increment the waiters for the destination bucket before
141 * acquiring the lock. It then decrements them again after releasing it -
142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
143 * will do the additional required waiter count housekeeping. This is done for
144 * double_lock_hb() and double_unlock_hb(), respectively.
147 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
148 #define futex_cmpxchg_enabled 1
150 static int __read_mostly futex_cmpxchg_enabled
;
154 * Futex flags used to encode options to functions and preserve them across
158 # define FLAGS_SHARED 0x01
161 * NOMMU does not have per process address space. Let the compiler optimize
164 # define FLAGS_SHARED 0x00
166 #define FLAGS_CLOCKRT 0x02
167 #define FLAGS_HAS_TIMEOUT 0x04
170 * Priority Inheritance state:
172 struct futex_pi_state
{
174 * list of 'owned' pi_state instances - these have to be
175 * cleaned up in do_exit() if the task exits prematurely:
177 struct list_head list
;
182 struct rt_mutex pi_mutex
;
184 struct task_struct
*owner
;
188 } __randomize_layout
;
191 * struct futex_q - The hashed futex queue entry, one per waiting task
192 * @list: priority-sorted list of tasks waiting on this futex
193 * @task: the task waiting on the futex
194 * @lock_ptr: the hash bucket lock
195 * @key: the key the futex is hashed on
196 * @pi_state: optional priority inheritance state
197 * @rt_waiter: rt_waiter storage for use with requeue_pi
198 * @requeue_pi_key: the requeue_pi target futex key
199 * @bitset: bitset for the optional bitmasked wakeup
201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
202 * we can wake only the relevant ones (hashed queues may be shared).
204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
206 * The order of wakeup is always to make the first condition true, then
209 * PI futexes are typically woken before they are removed from the hash list via
210 * the rt_mutex code. See unqueue_me_pi().
213 struct plist_node list
;
215 struct task_struct
*task
;
216 spinlock_t
*lock_ptr
;
218 struct futex_pi_state
*pi_state
;
219 struct rt_mutex_waiter
*rt_waiter
;
220 union futex_key
*requeue_pi_key
;
222 } __randomize_layout
;
224 static const struct futex_q futex_q_init
= {
225 /* list gets initialized in queue_me()*/
226 .key
= FUTEX_KEY_INIT
,
227 .bitset
= FUTEX_BITSET_MATCH_ANY
231 * Hash buckets are shared by all the futex_keys that hash to the same
232 * location. Each key may have multiple futex_q structures, one for each task
233 * waiting on a futex.
235 struct futex_hash_bucket
{
238 struct plist_head chain
;
239 } ____cacheline_aligned_in_smp
;
242 * The base of the bucket array and its size are always used together
243 * (after initialization only in hash_futex()), so ensure that they
244 * reside in the same cacheline.
247 struct futex_hash_bucket
*queues
;
248 unsigned long hashsize
;
249 } __futex_data __read_mostly
__aligned(2*sizeof(long));
250 #define futex_queues (__futex_data.queues)
251 #define futex_hashsize (__futex_data.hashsize)
255 * Fault injections for futexes.
257 #ifdef CONFIG_FAIL_FUTEX
260 struct fault_attr attr
;
264 .attr
= FAULT_ATTR_INITIALIZER
,
265 .ignore_private
= false,
268 static int __init
setup_fail_futex(char *str
)
270 return setup_fault_attr(&fail_futex
.attr
, str
);
272 __setup("fail_futex=", setup_fail_futex
);
274 static bool should_fail_futex(bool fshared
)
276 if (fail_futex
.ignore_private
&& !fshared
)
279 return should_fail(&fail_futex
.attr
, 1);
282 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
284 static int __init
fail_futex_debugfs(void)
286 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
289 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
294 debugfs_create_bool("ignore-private", mode
, dir
,
295 &fail_futex
.ignore_private
);
299 late_initcall(fail_futex_debugfs
);
301 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
304 static inline bool should_fail_futex(bool fshared
)
308 #endif /* CONFIG_FAIL_FUTEX */
311 static void compat_exit_robust_list(struct task_struct
*curr
);
315 * Reflects a new waiter being added to the waitqueue.
317 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
320 atomic_inc(&hb
->waiters
);
322 * Full barrier (A), see the ordering comment above.
324 smp_mb__after_atomic();
329 * Reflects a waiter being removed from the waitqueue by wakeup
332 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
335 atomic_dec(&hb
->waiters
);
339 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
343 * Full barrier (B), see the ordering comment above.
346 return atomic_read(&hb
->waiters
);
353 * hash_futex - Return the hash bucket in the global hash
354 * @key: Pointer to the futex key for which the hash is calculated
356 * We hash on the keys returned from get_futex_key (see below) and return the
357 * corresponding hash bucket in the global hash.
359 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
361 u32 hash
= jhash2((u32
*)key
, offsetof(typeof(*key
), both
.offset
) / 4,
364 return &futex_queues
[hash
& (futex_hashsize
- 1)];
369 * match_futex - Check whether two futex keys are equal
370 * @key1: Pointer to key1
371 * @key2: Pointer to key2
373 * Return 1 if two futex_keys are equal, 0 otherwise.
375 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
378 && key1
->both
.word
== key2
->both
.word
379 && key1
->both
.ptr
== key2
->both
.ptr
380 && key1
->both
.offset
== key2
->both
.offset
);
389 * futex_setup_timer - set up the sleeping hrtimer.
390 * @time: ptr to the given timeout value
391 * @timeout: the hrtimer_sleeper structure to be set up
392 * @flags: futex flags
393 * @range_ns: optional range in ns
395 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
398 static inline struct hrtimer_sleeper
*
399 futex_setup_timer(ktime_t
*time
, struct hrtimer_sleeper
*timeout
,
400 int flags
, u64 range_ns
)
405 hrtimer_init_sleeper_on_stack(timeout
, (flags
& FLAGS_CLOCKRT
) ?
406 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
409 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
410 * effectively the same as calling hrtimer_set_expires().
412 hrtimer_set_expires_range_ns(&timeout
->timer
, *time
, range_ns
);
418 * Generate a machine wide unique identifier for this inode.
420 * This relies on u64 not wrapping in the life-time of the machine; which with
421 * 1ns resolution means almost 585 years.
423 * This further relies on the fact that a well formed program will not unmap
424 * the file while it has a (shared) futex waiting on it. This mapping will have
425 * a file reference which pins the mount and inode.
427 * If for some reason an inode gets evicted and read back in again, it will get
428 * a new sequence number and will _NOT_ match, even though it is the exact same
431 * It is important that match_futex() will never have a false-positive, esp.
432 * for PI futexes that can mess up the state. The above argues that false-negatives
433 * are only possible for malformed programs.
435 static u64
get_inode_sequence_number(struct inode
*inode
)
437 static atomic64_t i_seq
;
440 /* Does the inode already have a sequence number? */
441 old
= atomic64_read(&inode
->i_sequence
);
446 u64
new = atomic64_add_return(1, &i_seq
);
447 if (WARN_ON_ONCE(!new))
450 old
= atomic64_cmpxchg_relaxed(&inode
->i_sequence
, 0, new);
458 * get_futex_key() - Get parameters which are the keys for a futex
459 * @uaddr: virtual address of the futex
460 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
461 * @key: address where result is stored.
462 * @rw: mapping needs to be read/write (values: FUTEX_READ,
465 * Return: a negative error code or 0
467 * The key words are stored in @key on success.
469 * For shared mappings (when @fshared), the key is:
471 * ( inode->i_sequence, page->index, offset_within_page )
473 * [ also see get_inode_sequence_number() ]
475 * For private mappings (or when !@fshared), the key is:
477 * ( current->mm, address, 0 )
479 * This allows (cross process, where applicable) identification of the futex
480 * without keeping the page pinned for the duration of the FUTEX_WAIT.
482 * lock_page() might sleep, the caller should not hold a spinlock.
484 static int get_futex_key(u32 __user
*uaddr
, bool fshared
, union futex_key
*key
,
485 enum futex_access rw
)
487 unsigned long address
= (unsigned long)uaddr
;
488 struct mm_struct
*mm
= current
->mm
;
489 struct page
*page
, *tail
;
490 struct address_space
*mapping
;
494 * The futex address must be "naturally" aligned.
496 key
->both
.offset
= address
% PAGE_SIZE
;
497 if (unlikely((address
% sizeof(u32
)) != 0))
499 address
-= key
->both
.offset
;
501 if (unlikely(!access_ok(uaddr
, sizeof(u32
))))
504 if (unlikely(should_fail_futex(fshared
)))
508 * PROCESS_PRIVATE futexes are fast.
509 * As the mm cannot disappear under us and the 'key' only needs
510 * virtual address, we dont even have to find the underlying vma.
511 * Note : We do have to check 'uaddr' is a valid user address,
512 * but access_ok() should be faster than find_vma()
515 key
->private.mm
= mm
;
516 key
->private.address
= address
;
521 /* Ignore any VERIFY_READ mapping (futex common case) */
522 if (unlikely(should_fail_futex(true)))
525 err
= get_user_pages_fast(address
, 1, FOLL_WRITE
, &page
);
527 * If write access is not required (eg. FUTEX_WAIT), try
528 * and get read-only access.
530 if (err
== -EFAULT
&& rw
== FUTEX_READ
) {
531 err
= get_user_pages_fast(address
, 1, 0, &page
);
540 * The treatment of mapping from this point on is critical. The page
541 * lock protects many things but in this context the page lock
542 * stabilizes mapping, prevents inode freeing in the shared
543 * file-backed region case and guards against movement to swap cache.
545 * Strictly speaking the page lock is not needed in all cases being
546 * considered here and page lock forces unnecessarily serialization
547 * From this point on, mapping will be re-verified if necessary and
548 * page lock will be acquired only if it is unavoidable
550 * Mapping checks require the head page for any compound page so the
551 * head page and mapping is looked up now. For anonymous pages, it
552 * does not matter if the page splits in the future as the key is
553 * based on the address. For filesystem-backed pages, the tail is
554 * required as the index of the page determines the key. For
555 * base pages, there is no tail page and tail == page.
558 page
= compound_head(page
);
559 mapping
= READ_ONCE(page
->mapping
);
562 * If page->mapping is NULL, then it cannot be a PageAnon
563 * page; but it might be the ZERO_PAGE or in the gate area or
564 * in a special mapping (all cases which we are happy to fail);
565 * or it may have been a good file page when get_user_pages_fast
566 * found it, but truncated or holepunched or subjected to
567 * invalidate_complete_page2 before we got the page lock (also
568 * cases which we are happy to fail). And we hold a reference,
569 * so refcount care in invalidate_complete_page's remove_mapping
570 * prevents drop_caches from setting mapping to NULL beneath us.
572 * The case we do have to guard against is when memory pressure made
573 * shmem_writepage move it from filecache to swapcache beneath us:
574 * an unlikely race, but we do need to retry for page->mapping.
576 if (unlikely(!mapping
)) {
580 * Page lock is required to identify which special case above
581 * applies. If this is really a shmem page then the page lock
582 * will prevent unexpected transitions.
585 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
596 * Private mappings are handled in a simple way.
598 * If the futex key is stored on an anonymous page, then the associated
599 * object is the mm which is implicitly pinned by the calling process.
601 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
602 * it's a read-only handle, it's expected that futexes attach to
603 * the object not the particular process.
605 if (PageAnon(page
)) {
607 * A RO anonymous page will never change and thus doesn't make
608 * sense for futex operations.
610 if (unlikely(should_fail_futex(true)) || ro
) {
615 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
616 key
->private.mm
= mm
;
617 key
->private.address
= address
;
623 * The associated futex object in this case is the inode and
624 * the page->mapping must be traversed. Ordinarily this should
625 * be stabilised under page lock but it's not strictly
626 * necessary in this case as we just want to pin the inode, not
627 * update the radix tree or anything like that.
629 * The RCU read lock is taken as the inode is finally freed
630 * under RCU. If the mapping still matches expectations then the
631 * mapping->host can be safely accessed as being a valid inode.
635 if (READ_ONCE(page
->mapping
) != mapping
) {
642 inode
= READ_ONCE(mapping
->host
);
650 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
651 key
->shared
.i_seq
= get_inode_sequence_number(inode
);
652 key
->shared
.pgoff
= page_to_pgoff(tail
);
662 * fault_in_user_writeable() - Fault in user address and verify RW access
663 * @uaddr: pointer to faulting user space address
665 * Slow path to fixup the fault we just took in the atomic write
668 * We have no generic implementation of a non-destructive write to the
669 * user address. We know that we faulted in the atomic pagefault
670 * disabled section so we can as well avoid the #PF overhead by
671 * calling get_user_pages() right away.
673 static int fault_in_user_writeable(u32 __user
*uaddr
)
675 struct mm_struct
*mm
= current
->mm
;
679 ret
= fixup_user_fault(mm
, (unsigned long)uaddr
,
680 FAULT_FLAG_WRITE
, NULL
);
681 mmap_read_unlock(mm
);
683 return ret
< 0 ? ret
: 0;
687 * futex_top_waiter() - Return the highest priority waiter on a futex
688 * @hb: the hash bucket the futex_q's reside in
689 * @key: the futex key (to distinguish it from other futex futex_q's)
691 * Must be called with the hb lock held.
693 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
694 union futex_key
*key
)
696 struct futex_q
*this;
698 plist_for_each_entry(this, &hb
->chain
, list
) {
699 if (match_futex(&this->key
, key
))
705 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
706 u32 uval
, u32 newval
)
711 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
717 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
722 ret
= __get_user(*dest
, from
);
725 return ret
? -EFAULT
: 0;
732 static int refill_pi_state_cache(void)
734 struct futex_pi_state
*pi_state
;
736 if (likely(current
->pi_state_cache
))
739 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
744 INIT_LIST_HEAD(&pi_state
->list
);
745 /* pi_mutex gets initialized later */
746 pi_state
->owner
= NULL
;
747 refcount_set(&pi_state
->refcount
, 1);
748 pi_state
->key
= FUTEX_KEY_INIT
;
750 current
->pi_state_cache
= pi_state
;
755 static struct futex_pi_state
*alloc_pi_state(void)
757 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
760 current
->pi_state_cache
= NULL
;
765 static void pi_state_update_owner(struct futex_pi_state
*pi_state
,
766 struct task_struct
*new_owner
)
768 struct task_struct
*old_owner
= pi_state
->owner
;
770 lockdep_assert_held(&pi_state
->pi_mutex
.wait_lock
);
773 raw_spin_lock(&old_owner
->pi_lock
);
774 WARN_ON(list_empty(&pi_state
->list
));
775 list_del_init(&pi_state
->list
);
776 raw_spin_unlock(&old_owner
->pi_lock
);
780 raw_spin_lock(&new_owner
->pi_lock
);
781 WARN_ON(!list_empty(&pi_state
->list
));
782 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
783 pi_state
->owner
= new_owner
;
784 raw_spin_unlock(&new_owner
->pi_lock
);
788 static void get_pi_state(struct futex_pi_state
*pi_state
)
790 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state
->refcount
));
794 * Drops a reference to the pi_state object and frees or caches it
795 * when the last reference is gone.
797 static void put_pi_state(struct futex_pi_state
*pi_state
)
802 if (!refcount_dec_and_test(&pi_state
->refcount
))
806 * If pi_state->owner is NULL, the owner is most probably dying
807 * and has cleaned up the pi_state already
809 if (pi_state
->owner
) {
812 raw_spin_lock_irqsave(&pi_state
->pi_mutex
.wait_lock
, flags
);
813 pi_state_update_owner(pi_state
, NULL
);
814 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
);
815 raw_spin_unlock_irqrestore(&pi_state
->pi_mutex
.wait_lock
, flags
);
818 if (current
->pi_state_cache
) {
822 * pi_state->list is already empty.
823 * clear pi_state->owner.
824 * refcount is at 0 - put it back to 1.
826 pi_state
->owner
= NULL
;
827 refcount_set(&pi_state
->refcount
, 1);
828 current
->pi_state_cache
= pi_state
;
832 #ifdef CONFIG_FUTEX_PI
835 * This task is holding PI mutexes at exit time => bad.
836 * Kernel cleans up PI-state, but userspace is likely hosed.
837 * (Robust-futex cleanup is separate and might save the day for userspace.)
839 static void exit_pi_state_list(struct task_struct
*curr
)
841 struct list_head
*next
, *head
= &curr
->pi_state_list
;
842 struct futex_pi_state
*pi_state
;
843 struct futex_hash_bucket
*hb
;
844 union futex_key key
= FUTEX_KEY_INIT
;
846 if (!futex_cmpxchg_enabled
)
849 * We are a ZOMBIE and nobody can enqueue itself on
850 * pi_state_list anymore, but we have to be careful
851 * versus waiters unqueueing themselves:
853 raw_spin_lock_irq(&curr
->pi_lock
);
854 while (!list_empty(head
)) {
856 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
858 hb
= hash_futex(&key
);
861 * We can race against put_pi_state() removing itself from the
862 * list (a waiter going away). put_pi_state() will first
863 * decrement the reference count and then modify the list, so
864 * its possible to see the list entry but fail this reference
867 * In that case; drop the locks to let put_pi_state() make
868 * progress and retry the loop.
870 if (!refcount_inc_not_zero(&pi_state
->refcount
)) {
871 raw_spin_unlock_irq(&curr
->pi_lock
);
873 raw_spin_lock_irq(&curr
->pi_lock
);
876 raw_spin_unlock_irq(&curr
->pi_lock
);
878 spin_lock(&hb
->lock
);
879 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
880 raw_spin_lock(&curr
->pi_lock
);
882 * We dropped the pi-lock, so re-check whether this
883 * task still owns the PI-state:
885 if (head
->next
!= next
) {
886 /* retain curr->pi_lock for the loop invariant */
887 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
888 spin_unlock(&hb
->lock
);
889 put_pi_state(pi_state
);
893 WARN_ON(pi_state
->owner
!= curr
);
894 WARN_ON(list_empty(&pi_state
->list
));
895 list_del_init(&pi_state
->list
);
896 pi_state
->owner
= NULL
;
898 raw_spin_unlock(&curr
->pi_lock
);
899 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
900 spin_unlock(&hb
->lock
);
902 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
903 put_pi_state(pi_state
);
905 raw_spin_lock_irq(&curr
->pi_lock
);
907 raw_spin_unlock_irq(&curr
->pi_lock
);
910 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
914 * We need to check the following states:
916 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
918 * [1] NULL | --- | --- | 0 | 0/1 | Valid
919 * [2] NULL | --- | --- | >0 | 0/1 | Valid
921 * [3] Found | NULL | -- | Any | 0/1 | Invalid
923 * [4] Found | Found | NULL | 0 | 1 | Valid
924 * [5] Found | Found | NULL | >0 | 1 | Invalid
926 * [6] Found | Found | task | 0 | 1 | Valid
928 * [7] Found | Found | NULL | Any | 0 | Invalid
930 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
931 * [9] Found | Found | task | 0 | 0 | Invalid
932 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
934 * [1] Indicates that the kernel can acquire the futex atomically. We
935 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
937 * [2] Valid, if TID does not belong to a kernel thread. If no matching
938 * thread is found then it indicates that the owner TID has died.
940 * [3] Invalid. The waiter is queued on a non PI futex
942 * [4] Valid state after exit_robust_list(), which sets the user space
943 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
945 * [5] The user space value got manipulated between exit_robust_list()
946 * and exit_pi_state_list()
948 * [6] Valid state after exit_pi_state_list() which sets the new owner in
949 * the pi_state but cannot access the user space value.
951 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
953 * [8] Owner and user space value match
955 * [9] There is no transient state which sets the user space TID to 0
956 * except exit_robust_list(), but this is indicated by the
957 * FUTEX_OWNER_DIED bit. See [4]
959 * [10] There is no transient state which leaves owner and user space
960 * TID out of sync. Except one error case where the kernel is denied
961 * write access to the user address, see fixup_pi_state_owner().
964 * Serialization and lifetime rules:
968 * hb -> futex_q, relation
969 * futex_q -> pi_state, relation
971 * (cannot be raw because hb can contain arbitrary amount
974 * pi_mutex->wait_lock:
978 * (and pi_mutex 'obviously')
982 * p->pi_state_list -> pi_state->list, relation
983 * pi_mutex->owner -> pi_state->owner, relation
985 * pi_state->refcount:
993 * pi_mutex->wait_lock
999 * Validate that the existing waiter has a pi_state and sanity check
1000 * the pi_state against the user space value. If correct, attach to
1003 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1004 struct futex_pi_state
*pi_state
,
1005 struct futex_pi_state
**ps
)
1007 pid_t pid
= uval
& FUTEX_TID_MASK
;
1012 * Userspace might have messed up non-PI and PI futexes [3]
1014 if (unlikely(!pi_state
))
1018 * We get here with hb->lock held, and having found a
1019 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1020 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1021 * which in turn means that futex_lock_pi() still has a reference on
1024 * The waiter holding a reference on @pi_state also protects against
1025 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1026 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1027 * free pi_state before we can take a reference ourselves.
1029 WARN_ON(!refcount_read(&pi_state
->refcount
));
1032 * Now that we have a pi_state, we can acquire wait_lock
1033 * and do the state validation.
1035 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1038 * Since {uval, pi_state} is serialized by wait_lock, and our current
1039 * uval was read without holding it, it can have changed. Verify it
1040 * still is what we expect it to be, otherwise retry the entire
1043 if (get_futex_value_locked(&uval2
, uaddr
))
1050 * Handle the owner died case:
1052 if (uval
& FUTEX_OWNER_DIED
) {
1054 * exit_pi_state_list sets owner to NULL and wakes the
1055 * topmost waiter. The task which acquires the
1056 * pi_state->rt_mutex will fixup owner.
1058 if (!pi_state
->owner
) {
1060 * No pi state owner, but the user space TID
1061 * is not 0. Inconsistent state. [5]
1066 * Take a ref on the state and return success. [4]
1072 * If TID is 0, then either the dying owner has not
1073 * yet executed exit_pi_state_list() or some waiter
1074 * acquired the rtmutex in the pi state, but did not
1075 * yet fixup the TID in user space.
1077 * Take a ref on the state and return success. [6]
1083 * If the owner died bit is not set, then the pi_state
1084 * must have an owner. [7]
1086 if (!pi_state
->owner
)
1091 * Bail out if user space manipulated the futex value. If pi
1092 * state exists then the owner TID must be the same as the
1093 * user space TID. [9/10]
1095 if (pid
!= task_pid_vnr(pi_state
->owner
))
1099 get_pi_state(pi_state
);
1100 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1117 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1122 * wait_for_owner_exiting - Block until the owner has exited
1123 * @ret: owner's current futex lock status
1124 * @exiting: Pointer to the exiting task
1126 * Caller must hold a refcount on @exiting.
1128 static void wait_for_owner_exiting(int ret
, struct task_struct
*exiting
)
1130 if (ret
!= -EBUSY
) {
1131 WARN_ON_ONCE(exiting
);
1135 if (WARN_ON_ONCE(ret
== -EBUSY
&& !exiting
))
1138 mutex_lock(&exiting
->futex_exit_mutex
);
1140 * No point in doing state checking here. If the waiter got here
1141 * while the task was in exec()->exec_futex_release() then it can
1142 * have any FUTEX_STATE_* value when the waiter has acquired the
1143 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1144 * already. Highly unlikely and not a problem. Just one more round
1145 * through the futex maze.
1147 mutex_unlock(&exiting
->futex_exit_mutex
);
1149 put_task_struct(exiting
);
1152 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1153 struct task_struct
*tsk
)
1158 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1159 * caller that the alleged owner is busy.
1161 if (tsk
&& tsk
->futex_state
!= FUTEX_STATE_DEAD
)
1165 * Reread the user space value to handle the following situation:
1169 * sys_exit() sys_futex()
1170 * do_exit() futex_lock_pi()
1171 * futex_lock_pi_atomic()
1172 * exit_signals(tsk) No waiters:
1173 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1174 * mm_release(tsk) Set waiter bit
1175 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1176 * Set owner died attach_to_pi_owner() {
1177 * *uaddr = 0xC0000000; tsk = get_task(PID);
1178 * } if (!tsk->flags & PF_EXITING) {
1180 * tsk->futex_state = } else {
1181 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1184 * return -ESRCH; <--- FAIL
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1190 * The same logic applies to the case where the exiting task is
1193 if (get_futex_value_locked(&uval2
, uaddr
))
1196 /* If the user space value has changed, try again. */
1201 * The exiting task did not have a robust list, the robust list was
1202 * corrupted or the user space value in *uaddr is simply bogus.
1203 * Give up and tell user space.
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1212 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1213 struct futex_pi_state
**ps
,
1214 struct task_struct
**exiting
)
1216 pid_t pid
= uval
& FUTEX_TID_MASK
;
1217 struct futex_pi_state
*pi_state
;
1218 struct task_struct
*p
;
1221 * We are the first waiter - try to look up the real owner and attach
1222 * the new pi_state to it, but bail out when TID = 0 [1]
1224 * The !pid check is paranoid. None of the call sites should end up
1225 * with pid == 0, but better safe than sorry. Let the caller retry
1229 p
= find_get_task_by_vpid(pid
);
1231 return handle_exit_race(uaddr
, uval
, NULL
);
1233 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1239 * We need to look at the task state to figure out, whether the
1240 * task is exiting. To protect against the change of the task state
1241 * in futex_exit_release(), we do this protected by p->pi_lock:
1243 raw_spin_lock_irq(&p
->pi_lock
);
1244 if (unlikely(p
->futex_state
!= FUTEX_STATE_OK
)) {
1246 * The task is on the way out. When the futex state is
1247 * FUTEX_STATE_DEAD, we know that the task has finished
1250 int ret
= handle_exit_race(uaddr
, uval
, p
);
1252 raw_spin_unlock_irq(&p
->pi_lock
);
1254 * If the owner task is between FUTEX_STATE_EXITING and
1255 * FUTEX_STATE_DEAD then store the task pointer and keep
1256 * the reference on the task struct. The calling code will
1257 * drop all locks, wait for the task to reach
1258 * FUTEX_STATE_DEAD and then drop the refcount. This is
1259 * required to prevent a live lock when the current task
1260 * preempted the exiting task between the two states.
1270 * No existing pi state. First waiter. [2]
1272 * This creates pi_state, we have hb->lock held, this means nothing can
1273 * observe this state, wait_lock is irrelevant.
1275 pi_state
= alloc_pi_state();
1278 * Initialize the pi_mutex in locked state and make @p
1281 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1283 /* Store the key for possible exit cleanups: */
1284 pi_state
->key
= *key
;
1286 WARN_ON(!list_empty(&pi_state
->list
));
1287 list_add(&pi_state
->list
, &p
->pi_state_list
);
1289 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1290 * because there is no concurrency as the object is not published yet.
1292 pi_state
->owner
= p
;
1293 raw_spin_unlock_irq(&p
->pi_lock
);
1302 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1303 struct futex_hash_bucket
*hb
,
1304 union futex_key
*key
, struct futex_pi_state
**ps
,
1305 struct task_struct
**exiting
)
1307 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1310 * If there is a waiter on that futex, validate it and
1311 * attach to the pi_state when the validation succeeds.
1314 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1317 * We are the first waiter - try to look up the owner based on
1318 * @uval and attach to it.
1320 return attach_to_pi_owner(uaddr
, uval
, key
, ps
, exiting
);
1323 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1328 if (unlikely(should_fail_futex(true)))
1331 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1335 /* If user space value changed, let the caller retry */
1336 return curval
!= uval
? -EAGAIN
: 0;
1340 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1341 * @uaddr: the pi futex user address
1342 * @hb: the pi futex hash bucket
1343 * @key: the futex key associated with uaddr and hb
1344 * @ps: the pi_state pointer where we store the result of the
1346 * @task: the task to perform the atomic lock work for. This will
1347 * be "current" except in the case of requeue pi.
1348 * @exiting: Pointer to store the task pointer of the owner task
1349 * which is in the middle of exiting
1350 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1353 * - 0 - ready to wait;
1354 * - 1 - acquired the lock;
1357 * The hb->lock and futex_key refs shall be held by the caller.
1359 * @exiting is only set when the return value is -EBUSY. If so, this holds
1360 * a refcount on the exiting task on return and the caller needs to drop it
1361 * after waiting for the exit to complete.
1363 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1364 union futex_key
*key
,
1365 struct futex_pi_state
**ps
,
1366 struct task_struct
*task
,
1367 struct task_struct
**exiting
,
1370 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1371 struct futex_q
*top_waiter
;
1375 * Read the user space value first so we can validate a few
1376 * things before proceeding further.
1378 if (get_futex_value_locked(&uval
, uaddr
))
1381 if (unlikely(should_fail_futex(true)))
1387 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1390 if ((unlikely(should_fail_futex(true))))
1394 * Lookup existing state first. If it exists, try to attach to
1397 top_waiter
= futex_top_waiter(hb
, key
);
1399 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1402 * No waiter and user TID is 0. We are here because the
1403 * waiters or the owner died bit is set or called from
1404 * requeue_cmp_pi or for whatever reason something took the
1407 if (!(uval
& FUTEX_TID_MASK
)) {
1409 * We take over the futex. No other waiters and the user space
1410 * TID is 0. We preserve the owner died bit.
1412 newval
= uval
& FUTEX_OWNER_DIED
;
1415 /* The futex requeue_pi code can enforce the waiters bit */
1417 newval
|= FUTEX_WAITERS
;
1419 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1420 /* If the take over worked, return 1 */
1421 return ret
< 0 ? ret
: 1;
1425 * First waiter. Set the waiters bit before attaching ourself to
1426 * the owner. If owner tries to unlock, it will be forced into
1427 * the kernel and blocked on hb->lock.
1429 newval
= uval
| FUTEX_WAITERS
;
1430 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1434 * If the update of the user space value succeeded, we try to
1435 * attach to the owner. If that fails, no harm done, we only
1436 * set the FUTEX_WAITERS bit in the user space variable.
1438 return attach_to_pi_owner(uaddr
, newval
, key
, ps
, exiting
);
1442 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1443 * @q: The futex_q to unqueue
1445 * The q->lock_ptr must not be NULL and must be held by the caller.
1447 static void __unqueue_futex(struct futex_q
*q
)
1449 struct futex_hash_bucket
*hb
;
1451 if (WARN_ON_SMP(!q
->lock_ptr
) || WARN_ON(plist_node_empty(&q
->list
)))
1453 lockdep_assert_held(q
->lock_ptr
);
1455 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1456 plist_del(&q
->list
, &hb
->chain
);
1461 * The hash bucket lock must be held when this is called.
1462 * Afterwards, the futex_q must not be accessed. Callers
1463 * must ensure to later call wake_up_q() for the actual
1466 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1468 struct task_struct
*p
= q
->task
;
1470 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1476 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1477 * is written, without taking any locks. This is possible in the event
1478 * of a spurious wakeup, for example. A memory barrier is required here
1479 * to prevent the following store to lock_ptr from getting ahead of the
1480 * plist_del in __unqueue_futex().
1482 smp_store_release(&q
->lock_ptr
, NULL
);
1485 * Queue the task for later wakeup for after we've released
1488 wake_q_add_safe(wake_q
, p
);
1492 * Caller must hold a reference on @pi_state.
1494 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1497 struct rt_mutex_waiter
*top_waiter
;
1498 struct task_struct
*new_owner
;
1499 bool postunlock
= false;
1500 DEFINE_WAKE_Q(wake_q
);
1503 top_waiter
= rt_mutex_top_waiter(&pi_state
->pi_mutex
);
1504 if (WARN_ON_ONCE(!top_waiter
)) {
1506 * As per the comment in futex_unlock_pi() this should not happen.
1508 * When this happens, give up our locks and try again, giving
1509 * the futex_lock_pi() instance time to complete, either by
1510 * waiting on the rtmutex or removing itself from the futex
1517 new_owner
= top_waiter
->task
;
1520 * We pass it to the next owner. The WAITERS bit is always kept
1521 * enabled while there is PI state around. We cleanup the owner
1522 * died bit, because we are the owner.
1524 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1526 if (unlikely(should_fail_futex(true))) {
1531 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1532 if (!ret
&& (curval
!= uval
)) {
1534 * If a unconditional UNLOCK_PI operation (user space did not
1535 * try the TID->0 transition) raced with a waiter setting the
1536 * FUTEX_WAITERS flag between get_user() and locking the hash
1537 * bucket lock, retry the operation.
1539 if ((FUTEX_TID_MASK
& curval
) == uval
)
1547 * This is a point of no return; once we modified the uval
1548 * there is no going back and subsequent operations must
1551 pi_state_update_owner(pi_state
, new_owner
);
1552 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1556 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1559 rt_mutex_postunlock(&wake_q
);
1565 * Express the locking dependencies for lockdep:
1568 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1571 spin_lock(&hb1
->lock
);
1573 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1574 } else { /* hb1 > hb2 */
1575 spin_lock(&hb2
->lock
);
1576 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1581 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1583 spin_unlock(&hb1
->lock
);
1585 spin_unlock(&hb2
->lock
);
1589 * Wake up waiters matching bitset queued on this futex (uaddr).
1592 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1594 struct futex_hash_bucket
*hb
;
1595 struct futex_q
*this, *next
;
1596 union futex_key key
= FUTEX_KEY_INIT
;
1598 DEFINE_WAKE_Q(wake_q
);
1603 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_READ
);
1604 if (unlikely(ret
!= 0))
1607 hb
= hash_futex(&key
);
1609 /* Make sure we really have tasks to wakeup */
1610 if (!hb_waiters_pending(hb
))
1613 spin_lock(&hb
->lock
);
1615 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1616 if (match_futex (&this->key
, &key
)) {
1617 if (this->pi_state
|| this->rt_waiter
) {
1622 /* Check if one of the bits is set in both bitsets */
1623 if (!(this->bitset
& bitset
))
1626 mark_wake_futex(&wake_q
, this);
1627 if (++ret
>= nr_wake
)
1632 spin_unlock(&hb
->lock
);
1637 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1639 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1640 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1641 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1642 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1645 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1646 if (oparg
< 0 || oparg
> 31) {
1647 char comm
[sizeof(current
->comm
)];
1649 * kill this print and return -EINVAL when userspace
1652 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1653 get_task_comm(comm
, current
), oparg
);
1659 pagefault_disable();
1660 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1666 case FUTEX_OP_CMP_EQ
:
1667 return oldval
== cmparg
;
1668 case FUTEX_OP_CMP_NE
:
1669 return oldval
!= cmparg
;
1670 case FUTEX_OP_CMP_LT
:
1671 return oldval
< cmparg
;
1672 case FUTEX_OP_CMP_GE
:
1673 return oldval
>= cmparg
;
1674 case FUTEX_OP_CMP_LE
:
1675 return oldval
<= cmparg
;
1676 case FUTEX_OP_CMP_GT
:
1677 return oldval
> cmparg
;
1684 * Wake up all waiters hashed on the physical page that is mapped
1685 * to this virtual address:
1688 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1689 int nr_wake
, int nr_wake2
, int op
)
1691 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1692 struct futex_hash_bucket
*hb1
, *hb2
;
1693 struct futex_q
*this, *next
;
1695 DEFINE_WAKE_Q(wake_q
);
1698 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1699 if (unlikely(ret
!= 0))
1701 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
1702 if (unlikely(ret
!= 0))
1705 hb1
= hash_futex(&key1
);
1706 hb2
= hash_futex(&key2
);
1709 double_lock_hb(hb1
, hb2
);
1710 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1711 if (unlikely(op_ret
< 0)) {
1712 double_unlock_hb(hb1
, hb2
);
1714 if (!IS_ENABLED(CONFIG_MMU
) ||
1715 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1717 * we don't get EFAULT from MMU faults if we don't have
1718 * an MMU, but we might get them from range checking
1724 if (op_ret
== -EFAULT
) {
1725 ret
= fault_in_user_writeable(uaddr2
);
1731 if (!(flags
& FLAGS_SHARED
))
1736 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1737 if (match_futex (&this->key
, &key1
)) {
1738 if (this->pi_state
|| this->rt_waiter
) {
1742 mark_wake_futex(&wake_q
, this);
1743 if (++ret
>= nr_wake
)
1750 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1751 if (match_futex (&this->key
, &key2
)) {
1752 if (this->pi_state
|| this->rt_waiter
) {
1756 mark_wake_futex(&wake_q
, this);
1757 if (++op_ret
>= nr_wake2
)
1765 double_unlock_hb(hb1
, hb2
);
1771 * requeue_futex() - Requeue a futex_q from one hb to another
1772 * @q: the futex_q to requeue
1773 * @hb1: the source hash_bucket
1774 * @hb2: the target hash_bucket
1775 * @key2: the new key for the requeued futex_q
1778 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1779 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1783 * If key1 and key2 hash to the same bucket, no need to
1786 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1787 plist_del(&q
->list
, &hb1
->chain
);
1788 hb_waiters_dec(hb1
);
1789 hb_waiters_inc(hb2
);
1790 plist_add(&q
->list
, &hb2
->chain
);
1791 q
->lock_ptr
= &hb2
->lock
;
1797 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1799 * @key: the key of the requeue target futex
1800 * @hb: the hash_bucket of the requeue target futex
1802 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1803 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1804 * to the requeue target futex so the waiter can detect the wakeup on the right
1805 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1806 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1807 * to protect access to the pi_state to fixup the owner later. Must be called
1808 * with both q->lock_ptr and hb->lock held.
1811 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1812 struct futex_hash_bucket
*hb
)
1818 WARN_ON(!q
->rt_waiter
);
1819 q
->rt_waiter
= NULL
;
1821 q
->lock_ptr
= &hb
->lock
;
1823 wake_up_state(q
->task
, TASK_NORMAL
);
1827 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1828 * @pifutex: the user address of the to futex
1829 * @hb1: the from futex hash bucket, must be locked by the caller
1830 * @hb2: the to futex hash bucket, must be locked by the caller
1831 * @key1: the from futex key
1832 * @key2: the to futex key
1833 * @ps: address to store the pi_state pointer
1834 * @exiting: Pointer to store the task pointer of the owner task
1835 * which is in the middle of exiting
1836 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1838 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1839 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1840 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1841 * hb1 and hb2 must be held by the caller.
1843 * @exiting is only set when the return value is -EBUSY. If so, this holds
1844 * a refcount on the exiting task on return and the caller needs to drop it
1845 * after waiting for the exit to complete.
1848 * - 0 - failed to acquire the lock atomically;
1849 * - >0 - acquired the lock, return value is vpid of the top_waiter
1853 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1854 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1855 union futex_key
*key2
, struct futex_pi_state
**ps
,
1856 struct task_struct
**exiting
, int set_waiters
)
1858 struct futex_q
*top_waiter
= NULL
;
1862 if (get_futex_value_locked(&curval
, pifutex
))
1865 if (unlikely(should_fail_futex(true)))
1869 * Find the top_waiter and determine if there are additional waiters.
1870 * If the caller intends to requeue more than 1 waiter to pifutex,
1871 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1872 * as we have means to handle the possible fault. If not, don't set
1873 * the bit unnecessarily as it will force the subsequent unlock to enter
1876 top_waiter
= futex_top_waiter(hb1
, key1
);
1878 /* There are no waiters, nothing for us to do. */
1882 /* Ensure we requeue to the expected futex. */
1883 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1887 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1888 * the contended case or if set_waiters is 1. The pi_state is returned
1889 * in ps in contended cases.
1891 vpid
= task_pid_vnr(top_waiter
->task
);
1892 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1893 exiting
, set_waiters
);
1895 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1902 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1903 * @uaddr1: source futex user address
1904 * @flags: futex flags (FLAGS_SHARED, etc.)
1905 * @uaddr2: target futex user address
1906 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1907 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1908 * @cmpval: @uaddr1 expected value (or %NULL)
1909 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1910 * pi futex (pi to pi requeue is not supported)
1912 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1913 * uaddr2 atomically on behalf of the top waiter.
1916 * - >=0 - on success, the number of tasks requeued or woken;
1919 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1920 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1921 u32
*cmpval
, int requeue_pi
)
1923 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1924 int task_count
= 0, ret
;
1925 struct futex_pi_state
*pi_state
= NULL
;
1926 struct futex_hash_bucket
*hb1
, *hb2
;
1927 struct futex_q
*this, *next
;
1928 DEFINE_WAKE_Q(wake_q
);
1930 if (nr_wake
< 0 || nr_requeue
< 0)
1934 * When PI not supported: return -ENOSYS if requeue_pi is true,
1935 * consequently the compiler knows requeue_pi is always false past
1936 * this point which will optimize away all the conditional code
1939 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1944 * Requeue PI only works on two distinct uaddrs. This
1945 * check is only valid for private futexes. See below.
1947 if (uaddr1
== uaddr2
)
1951 * requeue_pi requires a pi_state, try to allocate it now
1952 * without any locks in case it fails.
1954 if (refill_pi_state_cache())
1957 * requeue_pi must wake as many tasks as it can, up to nr_wake
1958 * + nr_requeue, since it acquires the rt_mutex prior to
1959 * returning to userspace, so as to not leave the rt_mutex with
1960 * waiters and no owner. However, second and third wake-ups
1961 * cannot be predicted as they involve race conditions with the
1962 * first wake and a fault while looking up the pi_state. Both
1963 * pthread_cond_signal() and pthread_cond_broadcast() should
1971 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1972 if (unlikely(ret
!= 0))
1974 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1975 requeue_pi
? FUTEX_WRITE
: FUTEX_READ
);
1976 if (unlikely(ret
!= 0))
1980 * The check above which compares uaddrs is not sufficient for
1981 * shared futexes. We need to compare the keys:
1983 if (requeue_pi
&& match_futex(&key1
, &key2
))
1986 hb1
= hash_futex(&key1
);
1987 hb2
= hash_futex(&key2
);
1990 hb_waiters_inc(hb2
);
1991 double_lock_hb(hb1
, hb2
);
1993 if (likely(cmpval
!= NULL
)) {
1996 ret
= get_futex_value_locked(&curval
, uaddr1
);
1998 if (unlikely(ret
)) {
1999 double_unlock_hb(hb1
, hb2
);
2000 hb_waiters_dec(hb2
);
2002 ret
= get_user(curval
, uaddr1
);
2006 if (!(flags
& FLAGS_SHARED
))
2011 if (curval
!= *cmpval
) {
2017 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2018 struct task_struct
*exiting
= NULL
;
2021 * Attempt to acquire uaddr2 and wake the top waiter. If we
2022 * intend to requeue waiters, force setting the FUTEX_WAITERS
2023 * bit. We force this here where we are able to easily handle
2024 * faults rather in the requeue loop below.
2026 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2028 &exiting
, nr_requeue
);
2031 * At this point the top_waiter has either taken uaddr2 or is
2032 * waiting on it. If the former, then the pi_state will not
2033 * exist yet, look it up one more time to ensure we have a
2034 * reference to it. If the lock was taken, ret contains the
2035 * vpid of the top waiter task.
2036 * If the lock was not taken, we have pi_state and an initial
2037 * refcount on it. In case of an error we have nothing.
2043 * If we acquired the lock, then the user space value
2044 * of uaddr2 should be vpid. It cannot be changed by
2045 * the top waiter as it is blocked on hb2 lock if it
2046 * tries to do so. If something fiddled with it behind
2047 * our back the pi state lookup might unearth it. So
2048 * we rather use the known value than rereading and
2049 * handing potential crap to lookup_pi_state.
2051 * If that call succeeds then we have pi_state and an
2052 * initial refcount on it.
2054 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2055 &pi_state
, &exiting
);
2060 /* We hold a reference on the pi state. */
2063 /* If the above failed, then pi_state is NULL */
2065 double_unlock_hb(hb1
, hb2
);
2066 hb_waiters_dec(hb2
);
2067 ret
= fault_in_user_writeable(uaddr2
);
2074 * Two reasons for this:
2075 * - EBUSY: Owner is exiting and we just wait for the
2077 * - EAGAIN: The user space value changed.
2079 double_unlock_hb(hb1
, hb2
);
2080 hb_waiters_dec(hb2
);
2082 * Handle the case where the owner is in the middle of
2083 * exiting. Wait for the exit to complete otherwise
2084 * this task might loop forever, aka. live lock.
2086 wait_for_owner_exiting(ret
, exiting
);
2094 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2095 if (task_count
- nr_wake
>= nr_requeue
)
2098 if (!match_futex(&this->key
, &key1
))
2102 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2103 * be paired with each other and no other futex ops.
2105 * We should never be requeueing a futex_q with a pi_state,
2106 * which is awaiting a futex_unlock_pi().
2108 if ((requeue_pi
&& !this->rt_waiter
) ||
2109 (!requeue_pi
&& this->rt_waiter
) ||
2116 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2117 * lock, we already woke the top_waiter. If not, it will be
2118 * woken by futex_unlock_pi().
2120 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2121 mark_wake_futex(&wake_q
, this);
2125 /* Ensure we requeue to the expected futex for requeue_pi. */
2126 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2132 * Requeue nr_requeue waiters and possibly one more in the case
2133 * of requeue_pi if we couldn't acquire the lock atomically.
2137 * Prepare the waiter to take the rt_mutex. Take a
2138 * refcount on the pi_state and store the pointer in
2139 * the futex_q object of the waiter.
2141 get_pi_state(pi_state
);
2142 this->pi_state
= pi_state
;
2143 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2148 * We got the lock. We do neither drop the
2149 * refcount on pi_state nor clear
2150 * this->pi_state because the waiter needs the
2151 * pi_state for cleaning up the user space
2152 * value. It will drop the refcount after
2155 requeue_pi_wake_futex(this, &key2
, hb2
);
2159 * rt_mutex_start_proxy_lock() detected a
2160 * potential deadlock when we tried to queue
2161 * that waiter. Drop the pi_state reference
2162 * which we took above and remove the pointer
2163 * to the state from the waiters futex_q
2166 this->pi_state
= NULL
;
2167 put_pi_state(pi_state
);
2169 * We stop queueing more waiters and let user
2170 * space deal with the mess.
2175 requeue_futex(this, hb1
, hb2
, &key2
);
2179 * We took an extra initial reference to the pi_state either
2180 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2181 * need to drop it here again.
2183 put_pi_state(pi_state
);
2186 double_unlock_hb(hb1
, hb2
);
2188 hb_waiters_dec(hb2
);
2189 return ret
? ret
: task_count
;
2192 /* The key must be already stored in q->key. */
2193 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2194 __acquires(&hb
->lock
)
2196 struct futex_hash_bucket
*hb
;
2198 hb
= hash_futex(&q
->key
);
2201 * Increment the counter before taking the lock so that
2202 * a potential waker won't miss a to-be-slept task that is
2203 * waiting for the spinlock. This is safe as all queue_lock()
2204 * users end up calling queue_me(). Similarly, for housekeeping,
2205 * decrement the counter at queue_unlock() when some error has
2206 * occurred and we don't end up adding the task to the list.
2208 hb_waiters_inc(hb
); /* implies smp_mb(); (A) */
2210 q
->lock_ptr
= &hb
->lock
;
2212 spin_lock(&hb
->lock
);
2217 queue_unlock(struct futex_hash_bucket
*hb
)
2218 __releases(&hb
->lock
)
2220 spin_unlock(&hb
->lock
);
2224 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2229 * The priority used to register this element is
2230 * - either the real thread-priority for the real-time threads
2231 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2232 * - or MAX_RT_PRIO for non-RT threads.
2233 * Thus, all RT-threads are woken first in priority order, and
2234 * the others are woken last, in FIFO order.
2236 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2238 plist_node_init(&q
->list
, prio
);
2239 plist_add(&q
->list
, &hb
->chain
);
2244 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2245 * @q: The futex_q to enqueue
2246 * @hb: The destination hash bucket
2248 * The hb->lock must be held by the caller, and is released here. A call to
2249 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2250 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2251 * or nothing if the unqueue is done as part of the wake process and the unqueue
2252 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2255 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2256 __releases(&hb
->lock
)
2259 spin_unlock(&hb
->lock
);
2263 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2264 * @q: The futex_q to unqueue
2266 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2267 * be paired with exactly one earlier call to queue_me().
2270 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2271 * - 0 - if the futex_q was already removed by the waking thread
2273 static int unqueue_me(struct futex_q
*q
)
2275 spinlock_t
*lock_ptr
;
2278 /* In the common case we don't take the spinlock, which is nice. */
2281 * q->lock_ptr can change between this read and the following spin_lock.
2282 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2283 * optimizing lock_ptr out of the logic below.
2285 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2286 if (lock_ptr
!= NULL
) {
2287 spin_lock(lock_ptr
);
2289 * q->lock_ptr can change between reading it and
2290 * spin_lock(), causing us to take the wrong lock. This
2291 * corrects the race condition.
2293 * Reasoning goes like this: if we have the wrong lock,
2294 * q->lock_ptr must have changed (maybe several times)
2295 * between reading it and the spin_lock(). It can
2296 * change again after the spin_lock() but only if it was
2297 * already changed before the spin_lock(). It cannot,
2298 * however, change back to the original value. Therefore
2299 * we can detect whether we acquired the correct lock.
2301 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2302 spin_unlock(lock_ptr
);
2307 BUG_ON(q
->pi_state
);
2309 spin_unlock(lock_ptr
);
2317 * PI futexes can not be requeued and must remove themselves from the
2318 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
2320 static void unqueue_me_pi(struct futex_q
*q
)
2324 BUG_ON(!q
->pi_state
);
2325 put_pi_state(q
->pi_state
);
2329 static int __fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2330 struct task_struct
*argowner
)
2332 struct futex_pi_state
*pi_state
= q
->pi_state
;
2333 struct task_struct
*oldowner
, *newowner
;
2334 u32 uval
, curval
, newval
, newtid
;
2337 oldowner
= pi_state
->owner
;
2340 * We are here because either:
2342 * - we stole the lock and pi_state->owner needs updating to reflect
2343 * that (@argowner == current),
2347 * - someone stole our lock and we need to fix things to point to the
2348 * new owner (@argowner == NULL).
2350 * Either way, we have to replace the TID in the user space variable.
2351 * This must be atomic as we have to preserve the owner died bit here.
2353 * Note: We write the user space value _before_ changing the pi_state
2354 * because we can fault here. Imagine swapped out pages or a fork
2355 * that marked all the anonymous memory readonly for cow.
2357 * Modifying pi_state _before_ the user space value would leave the
2358 * pi_state in an inconsistent state when we fault here, because we
2359 * need to drop the locks to handle the fault. This might be observed
2360 * in the PID check in lookup_pi_state.
2364 if (oldowner
!= current
) {
2366 * We raced against a concurrent self; things are
2367 * already fixed up. Nothing to do.
2372 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2373 /* We got the lock. pi_state is correct. Tell caller. */
2378 * The trylock just failed, so either there is an owner or
2379 * there is a higher priority waiter than this one.
2381 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2383 * If the higher priority waiter has not yet taken over the
2384 * rtmutex then newowner is NULL. We can't return here with
2385 * that state because it's inconsistent vs. the user space
2386 * state. So drop the locks and try again. It's a valid
2387 * situation and not any different from the other retry
2390 if (unlikely(!newowner
)) {
2395 WARN_ON_ONCE(argowner
!= current
);
2396 if (oldowner
== current
) {
2398 * We raced against a concurrent self; things are
2399 * already fixed up. Nothing to do.
2403 newowner
= argowner
;
2406 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2408 if (!pi_state
->owner
)
2409 newtid
|= FUTEX_OWNER_DIED
;
2411 err
= get_futex_value_locked(&uval
, uaddr
);
2416 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2418 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2428 * We fixed up user space. Now we need to fix the pi_state
2431 pi_state_update_owner(pi_state
, newowner
);
2433 return argowner
== current
;
2436 * In order to reschedule or handle a page fault, we need to drop the
2437 * locks here. In the case of a fault, this gives the other task
2438 * (either the highest priority waiter itself or the task which stole
2439 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2440 * are back from handling the fault we need to check the pi_state after
2441 * reacquiring the locks and before trying to do another fixup. When
2442 * the fixup has been done already we simply return.
2444 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2445 * drop hb->lock since the caller owns the hb -> futex_q relation.
2446 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2449 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2450 spin_unlock(q
->lock_ptr
);
2454 err
= fault_in_user_writeable(uaddr
);
2467 spin_lock(q
->lock_ptr
);
2468 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2471 * Check if someone else fixed it for us:
2473 if (pi_state
->owner
!= oldowner
)
2474 return argowner
== current
;
2476 /* Retry if err was -EAGAIN or the fault in succeeded */
2481 * fault_in_user_writeable() failed so user state is immutable. At
2482 * best we can make the kernel state consistent but user state will
2483 * be most likely hosed and any subsequent unlock operation will be
2484 * rejected due to PI futex rule [10].
2486 * Ensure that the rtmutex owner is also the pi_state owner despite
2487 * the user space value claiming something different. There is no
2488 * point in unlocking the rtmutex if current is the owner as it
2489 * would need to wait until the next waiter has taken the rtmutex
2490 * to guarantee consistent state. Keep it simple. Userspace asked
2491 * for this wreckaged state.
2493 * The rtmutex has an owner - either current or some other
2494 * task. See the EAGAIN loop above.
2496 pi_state_update_owner(pi_state
, rt_mutex_owner(&pi_state
->pi_mutex
));
2501 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2502 struct task_struct
*argowner
)
2504 struct futex_pi_state
*pi_state
= q
->pi_state
;
2507 lockdep_assert_held(q
->lock_ptr
);
2509 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2510 ret
= __fixup_pi_state_owner(uaddr
, q
, argowner
);
2511 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2515 static long futex_wait_restart(struct restart_block
*restart
);
2518 * fixup_owner() - Post lock pi_state and corner case management
2519 * @uaddr: user address of the futex
2520 * @q: futex_q (contains pi_state and access to the rt_mutex)
2521 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2523 * After attempting to lock an rt_mutex, this function is called to cleanup
2524 * the pi_state owner as well as handle race conditions that may allow us to
2525 * acquire the lock. Must be called with the hb lock held.
2528 * - 1 - success, lock taken;
2529 * - 0 - success, lock not taken;
2530 * - <0 - on error (-EFAULT)
2532 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2536 * Got the lock. We might not be the anticipated owner if we
2537 * did a lock-steal - fix up the PI-state in that case:
2539 * Speculative pi_state->owner read (we don't hold wait_lock);
2540 * since we own the lock pi_state->owner == current is the
2541 * stable state, anything else needs more attention.
2543 if (q
->pi_state
->owner
!= current
)
2544 return fixup_pi_state_owner(uaddr
, q
, current
);
2549 * If we didn't get the lock; check if anybody stole it from us. In
2550 * that case, we need to fix up the uval to point to them instead of
2551 * us, otherwise bad things happen. [10]
2553 * Another speculative read; pi_state->owner == current is unstable
2554 * but needs our attention.
2556 if (q
->pi_state
->owner
== current
)
2557 return fixup_pi_state_owner(uaddr
, q
, NULL
);
2560 * Paranoia check. If we did not take the lock, then we should not be
2561 * the owner of the rt_mutex. Warn and establish consistent state.
2563 if (WARN_ON_ONCE(rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
))
2564 return fixup_pi_state_owner(uaddr
, q
, current
);
2570 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571 * @hb: the futex hash bucket, must be locked by the caller
2572 * @q: the futex_q to queue up on
2573 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2575 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2576 struct hrtimer_sleeper
*timeout
)
2579 * The task state is guaranteed to be set before another task can
2580 * wake it. set_current_state() is implemented using smp_store_mb() and
2581 * queue_me() calls spin_unlock() upon completion, both serializing
2582 * access to the hash list and forcing another memory barrier.
2584 set_current_state(TASK_INTERRUPTIBLE
);
2589 hrtimer_sleeper_start_expires(timeout
, HRTIMER_MODE_ABS
);
2592 * If we have been removed from the hash list, then another task
2593 * has tried to wake us, and we can skip the call to schedule().
2595 if (likely(!plist_node_empty(&q
->list
))) {
2597 * If the timer has already expired, current will already be
2598 * flagged for rescheduling. Only call schedule if there
2599 * is no timeout, or if it has yet to expire.
2601 if (!timeout
|| timeout
->task
)
2602 freezable_schedule();
2604 __set_current_state(TASK_RUNNING
);
2608 * futex_wait_setup() - Prepare to wait on a futex
2609 * @uaddr: the futex userspace address
2610 * @val: the expected value
2611 * @flags: futex flags (FLAGS_SHARED, etc.)
2612 * @q: the associated futex_q
2613 * @hb: storage for hash_bucket pointer to be returned to caller
2615 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2616 * compare it with the expected value. Handle atomic faults internally.
2617 * Return with the hb lock held and a q.key reference on success, and unlocked
2618 * with no q.key reference on failure.
2621 * - 0 - uaddr contains val and hb has been locked;
2622 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2624 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2625 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2631 * Access the page AFTER the hash-bucket is locked.
2632 * Order is important:
2634 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2635 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2637 * The basic logical guarantee of a futex is that it blocks ONLY
2638 * if cond(var) is known to be true at the time of blocking, for
2639 * any cond. If we locked the hash-bucket after testing *uaddr, that
2640 * would open a race condition where we could block indefinitely with
2641 * cond(var) false, which would violate the guarantee.
2643 * On the other hand, we insert q and release the hash-bucket only
2644 * after testing *uaddr. This guarantees that futex_wait() will NOT
2645 * absorb a wakeup if *uaddr does not match the desired values
2646 * while the syscall executes.
2649 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, FUTEX_READ
);
2650 if (unlikely(ret
!= 0))
2654 *hb
= queue_lock(q
);
2656 ret
= get_futex_value_locked(&uval
, uaddr
);
2661 ret
= get_user(uval
, uaddr
);
2665 if (!(flags
& FLAGS_SHARED
))
2679 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2680 ktime_t
*abs_time
, u32 bitset
)
2682 struct hrtimer_sleeper timeout
, *to
;
2683 struct restart_block
*restart
;
2684 struct futex_hash_bucket
*hb
;
2685 struct futex_q q
= futex_q_init
;
2692 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
2693 current
->timer_slack_ns
);
2696 * Prepare to wait on uaddr. On success, holds hb lock and increments
2699 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2703 /* queue_me and wait for wakeup, timeout, or a signal. */
2704 futex_wait_queue_me(hb
, &q
, to
);
2706 /* If we were woken (and unqueued), we succeeded, whatever. */
2708 /* unqueue_me() drops q.key ref */
2709 if (!unqueue_me(&q
))
2712 if (to
&& !to
->task
)
2716 * We expect signal_pending(current), but we might be the
2717 * victim of a spurious wakeup as well.
2719 if (!signal_pending(current
))
2726 restart
= ¤t
->restart_block
;
2727 restart
->futex
.uaddr
= uaddr
;
2728 restart
->futex
.val
= val
;
2729 restart
->futex
.time
= *abs_time
;
2730 restart
->futex
.bitset
= bitset
;
2731 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2733 ret
= set_restart_fn(restart
, futex_wait_restart
);
2737 hrtimer_cancel(&to
->timer
);
2738 destroy_hrtimer_on_stack(&to
->timer
);
2744 static long futex_wait_restart(struct restart_block
*restart
)
2746 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2747 ktime_t t
, *tp
= NULL
;
2749 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2750 t
= restart
->futex
.time
;
2753 restart
->fn
= do_no_restart_syscall
;
2755 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2756 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2761 * Userspace tried a 0 -> TID atomic transition of the futex value
2762 * and failed. The kernel side here does the whole locking operation:
2763 * if there are waiters then it will block as a consequence of relying
2764 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2765 * a 0 value of the futex too.).
2767 * Also serves as futex trylock_pi()'ing, and due semantics.
2769 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2770 ktime_t
*time
, int trylock
)
2772 struct hrtimer_sleeper timeout
, *to
;
2773 struct task_struct
*exiting
= NULL
;
2774 struct rt_mutex_waiter rt_waiter
;
2775 struct futex_hash_bucket
*hb
;
2776 struct futex_q q
= futex_q_init
;
2779 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2782 if (refill_pi_state_cache())
2785 to
= futex_setup_timer(time
, &timeout
, flags
, 0);
2788 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, FUTEX_WRITE
);
2789 if (unlikely(ret
!= 0))
2793 hb
= queue_lock(&q
);
2795 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2797 if (unlikely(ret
)) {
2799 * Atomic work succeeded and we got the lock,
2800 * or failed. Either way, we do _not_ block.
2804 /* We got the lock. */
2806 goto out_unlock_put_key
;
2812 * Two reasons for this:
2813 * - EBUSY: Task is exiting and we just wait for the
2815 * - EAGAIN: The user space value changed.
2819 * Handle the case where the owner is in the middle of
2820 * exiting. Wait for the exit to complete otherwise
2821 * this task might loop forever, aka. live lock.
2823 wait_for_owner_exiting(ret
, exiting
);
2827 goto out_unlock_put_key
;
2831 WARN_ON(!q
.pi_state
);
2834 * Only actually queue now that the atomic ops are done:
2839 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2840 /* Fixup the trylock return value: */
2841 ret
= ret
? 0 : -EWOULDBLOCK
;
2845 rt_mutex_init_waiter(&rt_waiter
);
2848 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2849 * hold it while doing rt_mutex_start_proxy(), because then it will
2850 * include hb->lock in the blocking chain, even through we'll not in
2851 * fact hold it while blocking. This will lead it to report -EDEADLK
2852 * and BUG when futex_unlock_pi() interleaves with this.
2854 * Therefore acquire wait_lock while holding hb->lock, but drop the
2855 * latter before calling __rt_mutex_start_proxy_lock(). This
2856 * interleaves with futex_unlock_pi() -- which does a similar lock
2857 * handoff -- such that the latter can observe the futex_q::pi_state
2858 * before __rt_mutex_start_proxy_lock() is done.
2860 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2861 spin_unlock(q
.lock_ptr
);
2863 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2864 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2865 * it sees the futex_q::pi_state.
2867 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2868 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2877 hrtimer_sleeper_start_expires(to
, HRTIMER_MODE_ABS
);
2879 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2882 spin_lock(q
.lock_ptr
);
2884 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2885 * first acquire the hb->lock before removing the lock from the
2886 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2889 * In particular; it is important that futex_unlock_pi() can not
2890 * observe this inconsistency.
2892 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2897 * Fixup the pi_state owner and possibly acquire the lock if we
2900 res
= fixup_owner(uaddr
, &q
, !ret
);
2902 * If fixup_owner() returned an error, propagate that. If it acquired
2903 * the lock, clear our -ETIMEDOUT or -EINTR.
2906 ret
= (res
< 0) ? res
: 0;
2909 spin_unlock(q
.lock_ptr
);
2917 hrtimer_cancel(&to
->timer
);
2918 destroy_hrtimer_on_stack(&to
->timer
);
2920 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2925 ret
= fault_in_user_writeable(uaddr
);
2929 if (!(flags
& FLAGS_SHARED
))
2936 * Userspace attempted a TID -> 0 atomic transition, and failed.
2937 * This is the in-kernel slowpath: we look up the PI state (if any),
2938 * and do the rt-mutex unlock.
2940 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2942 u32 curval
, uval
, vpid
= task_pid_vnr(current
);
2943 union futex_key key
= FUTEX_KEY_INIT
;
2944 struct futex_hash_bucket
*hb
;
2945 struct futex_q
*top_waiter
;
2948 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2952 if (get_user(uval
, uaddr
))
2955 * We release only a lock we actually own:
2957 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2960 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_WRITE
);
2964 hb
= hash_futex(&key
);
2965 spin_lock(&hb
->lock
);
2968 * Check waiters first. We do not trust user space values at
2969 * all and we at least want to know if user space fiddled
2970 * with the futex value instead of blindly unlocking.
2972 top_waiter
= futex_top_waiter(hb
, &key
);
2974 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
2981 * If current does not own the pi_state then the futex is
2982 * inconsistent and user space fiddled with the futex value.
2984 if (pi_state
->owner
!= current
)
2987 get_pi_state(pi_state
);
2989 * By taking wait_lock while still holding hb->lock, we ensure
2990 * there is no point where we hold neither; and therefore
2991 * wake_futex_pi() must observe a state consistent with what we
2994 * In particular; this forces __rt_mutex_start_proxy() to
2995 * complete such that we're guaranteed to observe the
2996 * rt_waiter. Also see the WARN in wake_futex_pi().
2998 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2999 spin_unlock(&hb
->lock
);
3001 /* drops pi_state->pi_mutex.wait_lock */
3002 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3004 put_pi_state(pi_state
);
3007 * Success, we're done! No tricky corner cases.
3012 * The atomic access to the futex value generated a
3013 * pagefault, so retry the user-access and the wakeup:
3018 * A unconditional UNLOCK_PI op raced against a waiter
3019 * setting the FUTEX_WAITERS bit. Try again.
3024 * wake_futex_pi has detected invalid state. Tell user
3031 * We have no kernel internal state, i.e. no waiters in the
3032 * kernel. Waiters which are about to queue themselves are stuck
3033 * on hb->lock. So we can safely ignore them. We do neither
3034 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3037 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3038 spin_unlock(&hb
->lock
);
3053 * If uval has changed, let user space handle it.
3055 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3058 spin_unlock(&hb
->lock
);
3067 ret
= fault_in_user_writeable(uaddr
);
3075 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3076 * @hb: the hash_bucket futex_q was original enqueued on
3077 * @q: the futex_q woken while waiting to be requeued
3078 * @key2: the futex_key of the requeue target futex
3079 * @timeout: the timeout associated with the wait (NULL if none)
3081 * Detect if the task was woken on the initial futex as opposed to the requeue
3082 * target futex. If so, determine if it was a timeout or a signal that caused
3083 * the wakeup and return the appropriate error code to the caller. Must be
3084 * called with the hb lock held.
3087 * - 0 = no early wakeup detected;
3088 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3091 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3092 struct futex_q
*q
, union futex_key
*key2
,
3093 struct hrtimer_sleeper
*timeout
)
3098 * With the hb lock held, we avoid races while we process the wakeup.
3099 * We only need to hold hb (and not hb2) to ensure atomicity as the
3100 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3101 * It can't be requeued from uaddr2 to something else since we don't
3102 * support a PI aware source futex for requeue.
3104 if (!match_futex(&q
->key
, key2
)) {
3105 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3107 * We were woken prior to requeue by a timeout or a signal.
3108 * Unqueue the futex_q and determine which it was.
3110 plist_del(&q
->list
, &hb
->chain
);
3113 /* Handle spurious wakeups gracefully */
3115 if (timeout
&& !timeout
->task
)
3117 else if (signal_pending(current
))
3118 ret
= -ERESTARTNOINTR
;
3124 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3125 * @uaddr: the futex we initially wait on (non-pi)
3126 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3127 * the same type, no requeueing from private to shared, etc.
3128 * @val: the expected value of uaddr
3129 * @abs_time: absolute timeout
3130 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3131 * @uaddr2: the pi futex we will take prior to returning to user-space
3133 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3134 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3135 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3136 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3137 * without one, the pi logic would not know which task to boost/deboost, if
3138 * there was a need to.
3140 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3141 * via the following--
3142 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3143 * 2) wakeup on uaddr2 after a requeue
3147 * If 3, cleanup and return -ERESTARTNOINTR.
3149 * If 2, we may then block on trying to take the rt_mutex and return via:
3150 * 5) successful lock
3153 * 8) other lock acquisition failure
3155 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3157 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3163 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3164 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3167 struct hrtimer_sleeper timeout
, *to
;
3168 struct rt_mutex_waiter rt_waiter
;
3169 struct futex_hash_bucket
*hb
;
3170 union futex_key key2
= FUTEX_KEY_INIT
;
3171 struct futex_q q
= futex_q_init
;
3174 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3177 if (uaddr
== uaddr2
)
3183 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
3184 current
->timer_slack_ns
);
3187 * The waiter is allocated on our stack, manipulated by the requeue
3188 * code while we sleep on uaddr.
3190 rt_mutex_init_waiter(&rt_waiter
);
3192 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
3193 if (unlikely(ret
!= 0))
3197 q
.rt_waiter
= &rt_waiter
;
3198 q
.requeue_pi_key
= &key2
;
3201 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3204 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3209 * The check above which compares uaddrs is not sufficient for
3210 * shared futexes. We need to compare the keys:
3212 if (match_futex(&q
.key
, &key2
)) {
3218 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3219 futex_wait_queue_me(hb
, &q
, to
);
3221 spin_lock(&hb
->lock
);
3222 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3223 spin_unlock(&hb
->lock
);
3228 * In order for us to be here, we know our q.key == key2, and since
3229 * we took the hb->lock above, we also know that futex_requeue() has
3230 * completed and we no longer have to concern ourselves with a wakeup
3231 * race with the atomic proxy lock acquisition by the requeue code. The
3232 * futex_requeue dropped our key1 reference and incremented our key2
3237 * Check if the requeue code acquired the second futex for us and do
3238 * any pertinent fixup.
3241 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3242 spin_lock(q
.lock_ptr
);
3243 ret
= fixup_owner(uaddr2
, &q
, true);
3245 * Drop the reference to the pi state which
3246 * the requeue_pi() code acquired for us.
3248 put_pi_state(q
.pi_state
);
3249 spin_unlock(q
.lock_ptr
);
3251 * Adjust the return value. It's either -EFAULT or
3252 * success (1) but the caller expects 0 for success.
3254 ret
= ret
< 0 ? ret
: 0;
3257 struct rt_mutex
*pi_mutex
;
3260 * We have been woken up by futex_unlock_pi(), a timeout, or a
3261 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3264 WARN_ON(!q
.pi_state
);
3265 pi_mutex
= &q
.pi_state
->pi_mutex
;
3266 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3268 spin_lock(q
.lock_ptr
);
3269 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3272 debug_rt_mutex_free_waiter(&rt_waiter
);
3274 * Fixup the pi_state owner and possibly acquire the lock if we
3277 res
= fixup_owner(uaddr2
, &q
, !ret
);
3279 * If fixup_owner() returned an error, propagate that. If it
3280 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3283 ret
= (res
< 0) ? res
: 0;
3286 spin_unlock(q
.lock_ptr
);
3289 if (ret
== -EINTR
) {
3291 * We've already been requeued, but cannot restart by calling
3292 * futex_lock_pi() directly. We could restart this syscall, but
3293 * it would detect that the user space "val" changed and return
3294 * -EWOULDBLOCK. Save the overhead of the restart and return
3295 * -EWOULDBLOCK directly.
3302 hrtimer_cancel(&to
->timer
);
3303 destroy_hrtimer_on_stack(&to
->timer
);
3309 * Support for robust futexes: the kernel cleans up held futexes at
3312 * Implementation: user-space maintains a per-thread list of locks it
3313 * is holding. Upon do_exit(), the kernel carefully walks this list,
3314 * and marks all locks that are owned by this thread with the
3315 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3316 * always manipulated with the lock held, so the list is private and
3317 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3318 * field, to allow the kernel to clean up if the thread dies after
3319 * acquiring the lock, but just before it could have added itself to
3320 * the list. There can only be one such pending lock.
3324 * sys_set_robust_list() - Set the robust-futex list head of a task
3325 * @head: pointer to the list-head
3326 * @len: length of the list-head, as userspace expects
3328 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3331 if (!futex_cmpxchg_enabled
)
3334 * The kernel knows only one size for now:
3336 if (unlikely(len
!= sizeof(*head
)))
3339 current
->robust_list
= head
;
3345 * sys_get_robust_list() - Get the robust-futex list head of a task
3346 * @pid: pid of the process [zero for current task]
3347 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3348 * @len_ptr: pointer to a length field, the kernel fills in the header size
3350 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3351 struct robust_list_head __user
* __user
*, head_ptr
,
3352 size_t __user
*, len_ptr
)
3354 struct robust_list_head __user
*head
;
3356 struct task_struct
*p
;
3358 if (!futex_cmpxchg_enabled
)
3367 p
= find_task_by_vpid(pid
);
3373 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3376 head
= p
->robust_list
;
3379 if (put_user(sizeof(*head
), len_ptr
))
3381 return put_user(head
, head_ptr
);
3389 /* Constants for the pending_op argument of handle_futex_death */
3390 #define HANDLE_DEATH_PENDING true
3391 #define HANDLE_DEATH_LIST false
3394 * Process a futex-list entry, check whether it's owned by the
3395 * dying task, and do notification if so:
3397 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3398 bool pi
, bool pending_op
)
3400 u32 uval
, nval
, mval
;
3403 /* Futex address must be 32bit aligned */
3404 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3408 if (get_user(uval
, uaddr
))
3412 * Special case for regular (non PI) futexes. The unlock path in
3413 * user space has two race scenarios:
3415 * 1. The unlock path releases the user space futex value and
3416 * before it can execute the futex() syscall to wake up
3417 * waiters it is killed.
3419 * 2. A woken up waiter is killed before it can acquire the
3420 * futex in user space.
3422 * In both cases the TID validation below prevents a wakeup of
3423 * potential waiters which can cause these waiters to block
3426 * In both cases the following conditions are met:
3428 * 1) task->robust_list->list_op_pending != NULL
3429 * @pending_op == true
3430 * 2) User space futex value == 0
3431 * 3) Regular futex: @pi == false
3433 * If these conditions are met, it is safe to attempt waking up a
3434 * potential waiter without touching the user space futex value and
3435 * trying to set the OWNER_DIED bit. The user space futex value is
3436 * uncontended and the rest of the user space mutex state is
3437 * consistent, so a woken waiter will just take over the
3438 * uncontended futex. Setting the OWNER_DIED bit would create
3439 * inconsistent state and malfunction of the user space owner died
3442 if (pending_op
&& !pi
&& !uval
) {
3443 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3447 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3451 * Ok, this dying thread is truly holding a futex
3452 * of interest. Set the OWNER_DIED bit atomically
3453 * via cmpxchg, and if the value had FUTEX_WAITERS
3454 * set, wake up a waiter (if any). (We have to do a
3455 * futex_wake() even if OWNER_DIED is already set -
3456 * to handle the rare but possible case of recursive
3457 * thread-death.) The rest of the cleanup is done in
3460 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3463 * We are not holding a lock here, but we want to have
3464 * the pagefault_disable/enable() protection because
3465 * we want to handle the fault gracefully. If the
3466 * access fails we try to fault in the futex with R/W
3467 * verification via get_user_pages. get_user() above
3468 * does not guarantee R/W access. If that fails we
3469 * give up and leave the futex locked.
3471 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3474 if (fault_in_user_writeable(uaddr
))
3492 * Wake robust non-PI futexes here. The wakeup of
3493 * PI futexes happens in exit_pi_state():
3495 if (!pi
&& (uval
& FUTEX_WAITERS
))
3496 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3502 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3504 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3505 struct robust_list __user
* __user
*head
,
3508 unsigned long uentry
;
3510 if (get_user(uentry
, (unsigned long __user
*)head
))
3513 *entry
= (void __user
*)(uentry
& ~1UL);
3520 * Walk curr->robust_list (very carefully, it's a userspace list!)
3521 * and mark any locks found there dead, and notify any waiters.
3523 * We silently return on any sign of list-walking problem.
3525 static void exit_robust_list(struct task_struct
*curr
)
3527 struct robust_list_head __user
*head
= curr
->robust_list
;
3528 struct robust_list __user
*entry
, *next_entry
, *pending
;
3529 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3530 unsigned int next_pi
;
3531 unsigned long futex_offset
;
3534 if (!futex_cmpxchg_enabled
)
3538 * Fetch the list head (which was registered earlier, via
3539 * sys_set_robust_list()):
3541 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3544 * Fetch the relative futex offset:
3546 if (get_user(futex_offset
, &head
->futex_offset
))
3549 * Fetch any possibly pending lock-add first, and handle it
3552 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3555 next_entry
= NULL
; /* avoid warning with gcc */
3556 while (entry
!= &head
->list
) {
3558 * Fetch the next entry in the list before calling
3559 * handle_futex_death:
3561 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3563 * A pending lock might already be on the list, so
3564 * don't process it twice:
3566 if (entry
!= pending
) {
3567 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3568 curr
, pi
, HANDLE_DEATH_LIST
))
3576 * Avoid excessively long or circular lists:
3585 handle_futex_death((void __user
*)pending
+ futex_offset
,
3586 curr
, pip
, HANDLE_DEATH_PENDING
);
3590 static void futex_cleanup(struct task_struct
*tsk
)
3592 if (unlikely(tsk
->robust_list
)) {
3593 exit_robust_list(tsk
);
3594 tsk
->robust_list
= NULL
;
3597 #ifdef CONFIG_COMPAT
3598 if (unlikely(tsk
->compat_robust_list
)) {
3599 compat_exit_robust_list(tsk
);
3600 tsk
->compat_robust_list
= NULL
;
3604 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3605 exit_pi_state_list(tsk
);
3609 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3610 * @tsk: task to set the state on
3612 * Set the futex exit state of the task lockless. The futex waiter code
3613 * observes that state when a task is exiting and loops until the task has
3614 * actually finished the futex cleanup. The worst case for this is that the
3615 * waiter runs through the wait loop until the state becomes visible.
3617 * This is called from the recursive fault handling path in do_exit().
3619 * This is best effort. Either the futex exit code has run already or
3620 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3621 * take it over. If not, the problem is pushed back to user space. If the
3622 * futex exit code did not run yet, then an already queued waiter might
3623 * block forever, but there is nothing which can be done about that.
3625 void futex_exit_recursive(struct task_struct
*tsk
)
3627 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3628 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3629 mutex_unlock(&tsk
->futex_exit_mutex
);
3630 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3633 static void futex_cleanup_begin(struct task_struct
*tsk
)
3636 * Prevent various race issues against a concurrent incoming waiter
3637 * including live locks by forcing the waiter to block on
3638 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3639 * attach_to_pi_owner().
3641 mutex_lock(&tsk
->futex_exit_mutex
);
3644 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3646 * This ensures that all subsequent checks of tsk->futex_state in
3647 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3648 * tsk->pi_lock held.
3650 * It guarantees also that a pi_state which was queued right before
3651 * the state change under tsk->pi_lock by a concurrent waiter must
3652 * be observed in exit_pi_state_list().
3654 raw_spin_lock_irq(&tsk
->pi_lock
);
3655 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3656 raw_spin_unlock_irq(&tsk
->pi_lock
);
3659 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3662 * Lockless store. The only side effect is that an observer might
3663 * take another loop until it becomes visible.
3665 tsk
->futex_state
= state
;
3667 * Drop the exit protection. This unblocks waiters which observed
3668 * FUTEX_STATE_EXITING to reevaluate the state.
3670 mutex_unlock(&tsk
->futex_exit_mutex
);
3673 void futex_exec_release(struct task_struct
*tsk
)
3676 * The state handling is done for consistency, but in the case of
3677 * exec() there is no way to prevent further damage as the PID stays
3678 * the same. But for the unlikely and arguably buggy case that a
3679 * futex is held on exec(), this provides at least as much state
3680 * consistency protection which is possible.
3682 futex_cleanup_begin(tsk
);
3685 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3686 * exec a new binary.
3688 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3691 void futex_exit_release(struct task_struct
*tsk
)
3693 futex_cleanup_begin(tsk
);
3695 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3698 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3699 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3701 int cmd
= op
& FUTEX_CMD_MASK
;
3702 unsigned int flags
= 0;
3704 if (!(op
& FUTEX_PRIVATE_FLAG
))
3705 flags
|= FLAGS_SHARED
;
3707 if (op
& FUTEX_CLOCK_REALTIME
) {
3708 flags
|= FLAGS_CLOCKRT
;
3709 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
&&
3710 cmd
!= FUTEX_LOCK_PI2
)
3716 case FUTEX_LOCK_PI2
:
3717 case FUTEX_UNLOCK_PI
:
3718 case FUTEX_TRYLOCK_PI
:
3719 case FUTEX_WAIT_REQUEUE_PI
:
3720 case FUTEX_CMP_REQUEUE_PI
:
3721 if (!futex_cmpxchg_enabled
)
3727 val3
= FUTEX_BITSET_MATCH_ANY
;
3729 case FUTEX_WAIT_BITSET
:
3730 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3732 val3
= FUTEX_BITSET_MATCH_ANY
;
3734 case FUTEX_WAKE_BITSET
:
3735 return futex_wake(uaddr
, flags
, val
, val3
);
3737 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3738 case FUTEX_CMP_REQUEUE
:
3739 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3741 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3743 flags
|= FLAGS_CLOCKRT
;
3745 case FUTEX_LOCK_PI2
:
3746 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3747 case FUTEX_UNLOCK_PI
:
3748 return futex_unlock_pi(uaddr
, flags
);
3749 case FUTEX_TRYLOCK_PI
:
3750 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3751 case FUTEX_WAIT_REQUEUE_PI
:
3752 val3
= FUTEX_BITSET_MATCH_ANY
;
3753 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3755 case FUTEX_CMP_REQUEUE_PI
:
3756 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3761 static __always_inline
bool futex_cmd_has_timeout(u32 cmd
)
3766 case FUTEX_LOCK_PI2
:
3767 case FUTEX_WAIT_BITSET
:
3768 case FUTEX_WAIT_REQUEUE_PI
:
3774 static __always_inline
int
3775 futex_init_timeout(u32 cmd
, u32 op
, struct timespec64
*ts
, ktime_t
*t
)
3777 if (!timespec64_valid(ts
))
3780 *t
= timespec64_to_ktime(*ts
);
3781 if (cmd
== FUTEX_WAIT
)
3782 *t
= ktime_add_safe(ktime_get(), *t
);
3783 else if (cmd
!= FUTEX_LOCK_PI
&& !(op
& FUTEX_CLOCK_REALTIME
))
3784 *t
= timens_ktime_to_host(CLOCK_MONOTONIC
, *t
);
3788 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3789 const struct __kernel_timespec __user
*, utime
,
3790 u32 __user
*, uaddr2
, u32
, val3
)
3792 int ret
, cmd
= op
& FUTEX_CMD_MASK
;
3793 ktime_t t
, *tp
= NULL
;
3794 struct timespec64 ts
;
3796 if (utime
&& futex_cmd_has_timeout(cmd
)) {
3797 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3799 if (get_timespec64(&ts
, utime
))
3801 ret
= futex_init_timeout(cmd
, op
, &ts
, &t
);
3807 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, (unsigned long)utime
, val3
);
3810 #ifdef CONFIG_COMPAT
3812 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3815 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3816 compat_uptr_t __user
*head
, unsigned int *pi
)
3818 if (get_user(*uentry
, head
))
3821 *entry
= compat_ptr((*uentry
) & ~1);
3822 *pi
= (unsigned int)(*uentry
) & 1;
3827 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3828 compat_long_t futex_offset
)
3830 compat_uptr_t base
= ptr_to_compat(entry
);
3831 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3837 * Walk curr->robust_list (very carefully, it's a userspace list!)
3838 * and mark any locks found there dead, and notify any waiters.
3840 * We silently return on any sign of list-walking problem.
3842 static void compat_exit_robust_list(struct task_struct
*curr
)
3844 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3845 struct robust_list __user
*entry
, *next_entry
, *pending
;
3846 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3847 unsigned int next_pi
;
3848 compat_uptr_t uentry
, next_uentry
, upending
;
3849 compat_long_t futex_offset
;
3852 if (!futex_cmpxchg_enabled
)
3856 * Fetch the list head (which was registered earlier, via
3857 * sys_set_robust_list()):
3859 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3862 * Fetch the relative futex offset:
3864 if (get_user(futex_offset
, &head
->futex_offset
))
3867 * Fetch any possibly pending lock-add first, and handle it
3870 if (compat_fetch_robust_entry(&upending
, &pending
,
3871 &head
->list_op_pending
, &pip
))
3874 next_entry
= NULL
; /* avoid warning with gcc */
3875 while (entry
!= (struct robust_list __user
*) &head
->list
) {
3877 * Fetch the next entry in the list before calling
3878 * handle_futex_death:
3880 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
3881 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
3883 * A pending lock might already be on the list, so
3884 * dont process it twice:
3886 if (entry
!= pending
) {
3887 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
3889 if (handle_futex_death(uaddr
, curr
, pi
,
3895 uentry
= next_uentry
;
3899 * Avoid excessively long or circular lists:
3907 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
3909 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
3913 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
3914 struct compat_robust_list_head __user
*, head
,
3917 if (!futex_cmpxchg_enabled
)
3920 if (unlikely(len
!= sizeof(*head
)))
3923 current
->compat_robust_list
= head
;
3928 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3929 compat_uptr_t __user
*, head_ptr
,
3930 compat_size_t __user
*, len_ptr
)
3932 struct compat_robust_list_head __user
*head
;
3934 struct task_struct
*p
;
3936 if (!futex_cmpxchg_enabled
)
3945 p
= find_task_by_vpid(pid
);
3951 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3954 head
= p
->compat_robust_list
;
3957 if (put_user(sizeof(*head
), len_ptr
))
3959 return put_user(ptr_to_compat(head
), head_ptr
);
3966 #endif /* CONFIG_COMPAT */
3968 #ifdef CONFIG_COMPAT_32BIT_TIME
3969 SYSCALL_DEFINE6(futex_time32
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3970 const struct old_timespec32 __user
*, utime
, u32 __user
*, uaddr2
,
3973 int ret
, cmd
= op
& FUTEX_CMD_MASK
;
3974 ktime_t t
, *tp
= NULL
;
3975 struct timespec64 ts
;
3977 if (utime
&& futex_cmd_has_timeout(cmd
)) {
3978 if (get_old_timespec32(&ts
, utime
))
3980 ret
= futex_init_timeout(cmd
, op
, &ts
, &t
);
3986 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, (unsigned long)utime
, val3
);
3988 #endif /* CONFIG_COMPAT_32BIT_TIME */
3990 static void __init
futex_detect_cmpxchg(void)
3992 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3996 * This will fail and we want it. Some arch implementations do
3997 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3998 * functionality. We want to know that before we call in any
3999 * of the complex code paths. Also we want to prevent
4000 * registration of robust lists in that case. NULL is
4001 * guaranteed to fault and we get -EFAULT on functional
4002 * implementation, the non-functional ones will return
4005 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4006 futex_cmpxchg_enabled
= 1;
4010 static int __init
futex_init(void)
4012 unsigned int futex_shift
;
4015 #if CONFIG_BASE_SMALL
4016 futex_hashsize
= 16;
4018 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4021 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4023 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4025 futex_hashsize
, futex_hashsize
);
4026 futex_hashsize
= 1UL << futex_shift
;
4028 futex_detect_cmpxchg();
4030 for (i
= 0; i
< futex_hashsize
; i
++) {
4031 atomic_set(&futex_queues
[i
].waiters
, 0);
4032 plist_head_init(&futex_queues
[i
].chain
);
4033 spin_lock_init(&futex_queues
[i
].lock
);
4038 core_initcall(futex_init
);