1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/slab.h>
36 #include <linux/poll.h>
38 #include <linux/file.h>
39 #include <linux/jhash.h>
40 #include <linux/init.h>
41 #include <linux/futex.h>
42 #include <linux/mount.h>
43 #include <linux/pagemap.h>
44 #include <linux/syscalls.h>
45 #include <linux/signal.h>
46 #include <linux/export.h>
47 #include <linux/magic.h>
48 #include <linux/pid.h>
49 #include <linux/nsproxy.h>
50 #include <linux/ptrace.h>
51 #include <linux/sched/rt.h>
52 #include <linux/sched/wake_q.h>
53 #include <linux/sched/mm.h>
54 #include <linux/hugetlb.h>
55 #include <linux/freezer.h>
56 #include <linux/memblock.h>
57 #include <linux/fault-inject.h>
58 #include <linux/refcount.h>
60 #include <asm/futex.h>
62 #include "locking/rtmutex_common.h"
65 * READ this before attempting to hack on futexes!
67 * Basic futex operation and ordering guarantees
68 * =============================================
70 * The waiter reads the futex value in user space and calls
71 * futex_wait(). This function computes the hash bucket and acquires
72 * the hash bucket lock. After that it reads the futex user space value
73 * again and verifies that the data has not changed. If it has not changed
74 * it enqueues itself into the hash bucket, releases the hash bucket lock
77 * The waker side modifies the user space value of the futex and calls
78 * futex_wake(). This function computes the hash bucket and acquires the
79 * hash bucket lock. Then it looks for waiters on that futex in the hash
80 * bucket and wakes them.
82 * In futex wake up scenarios where no tasks are blocked on a futex, taking
83 * the hb spinlock can be avoided and simply return. In order for this
84 * optimization to work, ordering guarantees must exist so that the waiter
85 * being added to the list is acknowledged when the list is concurrently being
86 * checked by the waker, avoiding scenarios like the following:
90 * sys_futex(WAIT, futex, val);
91 * futex_wait(futex, val);
94 * sys_futex(WAKE, futex);
99 * lock(hash_bucket(futex));
101 * unlock(hash_bucket(futex));
104 * This would cause the waiter on CPU 0 to wait forever because it
105 * missed the transition of the user space value from val to newval
106 * and the waker did not find the waiter in the hash bucket queue.
108 * The correct serialization ensures that a waiter either observes
109 * the changed user space value before blocking or is woken by a
114 * sys_futex(WAIT, futex, val);
115 * futex_wait(futex, val);
118 * smp_mb(); (A) <-- paired with -.
120 * lock(hash_bucket(futex)); |
124 * | sys_futex(WAKE, futex);
125 * | futex_wake(futex);
127 * `--------> smp_mb(); (B)
130 * unlock(hash_bucket(futex));
131 * schedule(); if (waiters)
132 * lock(hash_bucket(futex));
133 * else wake_waiters(futex);
134 * waiters--; (b) unlock(hash_bucket(futex));
136 * Where (A) orders the waiters increment and the futex value read through
137 * atomic operations (see hb_waiters_inc) and where (B) orders the write
138 * to futex and the waiters read (see hb_waiters_pending()).
140 * This yields the following case (where X:=waiters, Y:=futex):
148 * Which guarantees that x==0 && y==0 is impossible; which translates back into
149 * the guarantee that we cannot both miss the futex variable change and the
152 * Note that a new waiter is accounted for in (a) even when it is possible that
153 * the wait call can return error, in which case we backtrack from it in (b).
154 * Refer to the comment in queue_lock().
156 * Similarly, in order to account for waiters being requeued on another
157 * address we always increment the waiters for the destination bucket before
158 * acquiring the lock. It then decrements them again after releasing it -
159 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
160 * will do the additional required waiter count housekeeping. This is done for
161 * double_lock_hb() and double_unlock_hb(), respectively.
164 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
165 #define futex_cmpxchg_enabled 1
167 static int __read_mostly futex_cmpxchg_enabled
;
171 * Futex flags used to encode options to functions and preserve them across
175 # define FLAGS_SHARED 0x01
178 * NOMMU does not have per process address space. Let the compiler optimize
181 # define FLAGS_SHARED 0x00
183 #define FLAGS_CLOCKRT 0x02
184 #define FLAGS_HAS_TIMEOUT 0x04
187 * Priority Inheritance state:
189 struct futex_pi_state
{
191 * list of 'owned' pi_state instances - these have to be
192 * cleaned up in do_exit() if the task exits prematurely:
194 struct list_head list
;
199 struct rt_mutex pi_mutex
;
201 struct task_struct
*owner
;
205 } __randomize_layout
;
208 * struct futex_q - The hashed futex queue entry, one per waiting task
209 * @list: priority-sorted list of tasks waiting on this futex
210 * @task: the task waiting on the futex
211 * @lock_ptr: the hash bucket lock
212 * @key: the key the futex is hashed on
213 * @pi_state: optional priority inheritance state
214 * @rt_waiter: rt_waiter storage for use with requeue_pi
215 * @requeue_pi_key: the requeue_pi target futex key
216 * @bitset: bitset for the optional bitmasked wakeup
218 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
219 * we can wake only the relevant ones (hashed queues may be shared).
221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223 * The order of wakeup is always to make the first condition true, then
226 * PI futexes are typically woken before they are removed from the hash list via
227 * the rt_mutex code. See unqueue_me_pi().
230 struct plist_node list
;
232 struct task_struct
*task
;
233 spinlock_t
*lock_ptr
;
235 struct futex_pi_state
*pi_state
;
236 struct rt_mutex_waiter
*rt_waiter
;
237 union futex_key
*requeue_pi_key
;
239 } __randomize_layout
;
241 static const struct futex_q futex_q_init
= {
242 /* list gets initialized in queue_me()*/
243 .key
= FUTEX_KEY_INIT
,
244 .bitset
= FUTEX_BITSET_MATCH_ANY
248 * Hash buckets are shared by all the futex_keys that hash to the same
249 * location. Each key may have multiple futex_q structures, one for each task
250 * waiting on a futex.
252 struct futex_hash_bucket
{
255 struct plist_head chain
;
256 } ____cacheline_aligned_in_smp
;
259 * The base of the bucket array and its size are always used together
260 * (after initialization only in hash_futex()), so ensure that they
261 * reside in the same cacheline.
264 struct futex_hash_bucket
*queues
;
265 unsigned long hashsize
;
266 } __futex_data __read_mostly
__aligned(2*sizeof(long));
267 #define futex_queues (__futex_data.queues)
268 #define futex_hashsize (__futex_data.hashsize)
272 * Fault injections for futexes.
274 #ifdef CONFIG_FAIL_FUTEX
277 struct fault_attr attr
;
281 .attr
= FAULT_ATTR_INITIALIZER
,
282 .ignore_private
= false,
285 static int __init
setup_fail_futex(char *str
)
287 return setup_fault_attr(&fail_futex
.attr
, str
);
289 __setup("fail_futex=", setup_fail_futex
);
291 static bool should_fail_futex(bool fshared
)
293 if (fail_futex
.ignore_private
&& !fshared
)
296 return should_fail(&fail_futex
.attr
, 1);
299 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
301 static int __init
fail_futex_debugfs(void)
303 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
306 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
311 debugfs_create_bool("ignore-private", mode
, dir
,
312 &fail_futex
.ignore_private
);
316 late_initcall(fail_futex_debugfs
);
318 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
321 static inline bool should_fail_futex(bool fshared
)
325 #endif /* CONFIG_FAIL_FUTEX */
328 static void compat_exit_robust_list(struct task_struct
*curr
);
330 static inline void compat_exit_robust_list(struct task_struct
*curr
) { }
334 * Reflects a new waiter being added to the waitqueue.
336 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
339 atomic_inc(&hb
->waiters
);
341 * Full barrier (A), see the ordering comment above.
343 smp_mb__after_atomic();
348 * Reflects a waiter being removed from the waitqueue by wakeup
351 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
354 atomic_dec(&hb
->waiters
);
358 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
362 * Full barrier (B), see the ordering comment above.
365 return atomic_read(&hb
->waiters
);
372 * hash_futex - Return the hash bucket in the global hash
373 * @key: Pointer to the futex key for which the hash is calculated
375 * We hash on the keys returned from get_futex_key (see below) and return the
376 * corresponding hash bucket in the global hash.
378 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
380 u32 hash
= jhash2((u32
*)key
, offsetof(typeof(*key
), both
.offset
) / 4,
383 return &futex_queues
[hash
& (futex_hashsize
- 1)];
388 * match_futex - Check whether two futex keys are equal
389 * @key1: Pointer to key1
390 * @key2: Pointer to key2
392 * Return 1 if two futex_keys are equal, 0 otherwise.
394 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
397 && key1
->both
.word
== key2
->both
.word
398 && key1
->both
.ptr
== key2
->both
.ptr
399 && key1
->both
.offset
== key2
->both
.offset
);
408 * futex_setup_timer - set up the sleeping hrtimer.
409 * @time: ptr to the given timeout value
410 * @timeout: the hrtimer_sleeper structure to be set up
411 * @flags: futex flags
412 * @range_ns: optional range in ns
414 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
417 static inline struct hrtimer_sleeper
*
418 futex_setup_timer(ktime_t
*time
, struct hrtimer_sleeper
*timeout
,
419 int flags
, u64 range_ns
)
424 hrtimer_init_sleeper_on_stack(timeout
, (flags
& FLAGS_CLOCKRT
) ?
425 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
428 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
429 * effectively the same as calling hrtimer_set_expires().
431 hrtimer_set_expires_range_ns(&timeout
->timer
, *time
, range_ns
);
437 * Generate a machine wide unique identifier for this inode.
439 * This relies on u64 not wrapping in the life-time of the machine; which with
440 * 1ns resolution means almost 585 years.
442 * This further relies on the fact that a well formed program will not unmap
443 * the file while it has a (shared) futex waiting on it. This mapping will have
444 * a file reference which pins the mount and inode.
446 * If for some reason an inode gets evicted and read back in again, it will get
447 * a new sequence number and will _NOT_ match, even though it is the exact same
450 * It is important that match_futex() will never have a false-positive, esp.
451 * for PI futexes that can mess up the state. The above argues that false-negatives
452 * are only possible for malformed programs.
454 static u64
get_inode_sequence_number(struct inode
*inode
)
456 static atomic64_t i_seq
;
459 /* Does the inode already have a sequence number? */
460 old
= atomic64_read(&inode
->i_sequence
);
465 u64
new = atomic64_add_return(1, &i_seq
);
466 if (WARN_ON_ONCE(!new))
469 old
= atomic64_cmpxchg_relaxed(&inode
->i_sequence
, 0, new);
477 * get_futex_key() - Get parameters which are the keys for a futex
478 * @uaddr: virtual address of the futex
479 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
480 * @key: address where result is stored.
481 * @rw: mapping needs to be read/write (values: FUTEX_READ,
484 * Return: a negative error code or 0
486 * The key words are stored in @key on success.
488 * For shared mappings (when @fshared), the key is:
489 * ( inode->i_sequence, page->index, offset_within_page )
490 * [ also see get_inode_sequence_number() ]
492 * For private mappings (or when !@fshared), the key is:
493 * ( current->mm, address, 0 )
495 * This allows (cross process, where applicable) identification of the futex
496 * without keeping the page pinned for the duration of the FUTEX_WAIT.
498 * lock_page() might sleep, the caller should not hold a spinlock.
501 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, enum futex_access rw
)
503 unsigned long address
= (unsigned long)uaddr
;
504 struct mm_struct
*mm
= current
->mm
;
505 struct page
*page
, *tail
;
506 struct address_space
*mapping
;
510 * The futex address must be "naturally" aligned.
512 key
->both
.offset
= address
% PAGE_SIZE
;
513 if (unlikely((address
% sizeof(u32
)) != 0))
515 address
-= key
->both
.offset
;
517 if (unlikely(!access_ok(uaddr
, sizeof(u32
))))
520 if (unlikely(should_fail_futex(fshared
)))
524 * PROCESS_PRIVATE futexes are fast.
525 * As the mm cannot disappear under us and the 'key' only needs
526 * virtual address, we dont even have to find the underlying vma.
527 * Note : We do have to check 'uaddr' is a valid user address,
528 * but access_ok() should be faster than find_vma()
531 key
->private.mm
= mm
;
532 key
->private.address
= address
;
537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared
)))
541 err
= get_user_pages_fast(address
, 1, FOLL_WRITE
, &page
);
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
546 if (err
== -EFAULT
&& rw
== FUTEX_READ
) {
547 err
= get_user_pages_fast(address
, 1, 0, &page
);
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
574 page
= compound_head(page
);
575 mapping
= READ_ONCE(page
->mapping
);
578 * If page->mapping is NULL, then it cannot be a PageAnon
579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
590 * an unlikely race, but we do need to retry for page->mapping.
592 if (unlikely(!mapping
)) {
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
601 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
612 * Private mappings are handled in a simple way.
614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
619 * the object not the particular process.
621 if (PageAnon(page
)) {
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
626 if (unlikely(should_fail_futex(fshared
)) || ro
) {
631 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
632 key
->private.mm
= mm
;
633 key
->private.address
= address
;
639 * The associated futex object in this case is the inode and
640 * the page->mapping must be traversed. Ordinarily this should
641 * be stabilised under page lock but it's not strictly
642 * necessary in this case as we just want to pin the inode, not
643 * update the radix tree or anything like that.
645 * The RCU read lock is taken as the inode is finally freed
646 * under RCU. If the mapping still matches expectations then the
647 * mapping->host can be safely accessed as being a valid inode.
651 if (READ_ONCE(page
->mapping
) != mapping
) {
658 inode
= READ_ONCE(mapping
->host
);
666 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
667 key
->shared
.i_seq
= get_inode_sequence_number(inode
);
668 key
->shared
.pgoff
= basepage_index(tail
);
677 static inline void put_futex_key(union futex_key
*key
)
682 * fault_in_user_writeable() - Fault in user address and verify RW access
683 * @uaddr: pointer to faulting user space address
685 * Slow path to fixup the fault we just took in the atomic write
688 * We have no generic implementation of a non-destructive write to the
689 * user address. We know that we faulted in the atomic pagefault
690 * disabled section so we can as well avoid the #PF overhead by
691 * calling get_user_pages() right away.
693 static int fault_in_user_writeable(u32 __user
*uaddr
)
695 struct mm_struct
*mm
= current
->mm
;
698 down_read(&mm
->mmap_sem
);
699 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
700 FAULT_FLAG_WRITE
, NULL
);
701 up_read(&mm
->mmap_sem
);
703 return ret
< 0 ? ret
: 0;
707 * futex_top_waiter() - Return the highest priority waiter on a futex
708 * @hb: the hash bucket the futex_q's reside in
709 * @key: the futex key (to distinguish it from other futex futex_q's)
711 * Must be called with the hb lock held.
713 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
714 union futex_key
*key
)
716 struct futex_q
*this;
718 plist_for_each_entry(this, &hb
->chain
, list
) {
719 if (match_futex(&this->key
, key
))
725 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
726 u32 uval
, u32 newval
)
731 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
737 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
742 ret
= __get_user(*dest
, from
);
745 return ret
? -EFAULT
: 0;
752 static int refill_pi_state_cache(void)
754 struct futex_pi_state
*pi_state
;
756 if (likely(current
->pi_state_cache
))
759 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
764 INIT_LIST_HEAD(&pi_state
->list
);
765 /* pi_mutex gets initialized later */
766 pi_state
->owner
= NULL
;
767 refcount_set(&pi_state
->refcount
, 1);
768 pi_state
->key
= FUTEX_KEY_INIT
;
770 current
->pi_state_cache
= pi_state
;
775 static struct futex_pi_state
*alloc_pi_state(void)
777 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
780 current
->pi_state_cache
= NULL
;
785 static void get_pi_state(struct futex_pi_state
*pi_state
)
787 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state
->refcount
));
791 * Drops a reference to the pi_state object and frees or caches it
792 * when the last reference is gone.
794 static void put_pi_state(struct futex_pi_state
*pi_state
)
799 if (!refcount_dec_and_test(&pi_state
->refcount
))
803 * If pi_state->owner is NULL, the owner is most probably dying
804 * and has cleaned up the pi_state already
806 if (pi_state
->owner
) {
807 struct task_struct
*owner
;
809 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
810 owner
= pi_state
->owner
;
812 raw_spin_lock(&owner
->pi_lock
);
813 list_del_init(&pi_state
->list
);
814 raw_spin_unlock(&owner
->pi_lock
);
816 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
817 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
820 if (current
->pi_state_cache
) {
824 * pi_state->list is already empty.
825 * clear pi_state->owner.
826 * refcount is at 0 - put it back to 1.
828 pi_state
->owner
= NULL
;
829 refcount_set(&pi_state
->refcount
, 1);
830 current
->pi_state_cache
= pi_state
;
834 #ifdef CONFIG_FUTEX_PI
837 * This task is holding PI mutexes at exit time => bad.
838 * Kernel cleans up PI-state, but userspace is likely hosed.
839 * (Robust-futex cleanup is separate and might save the day for userspace.)
841 static void exit_pi_state_list(struct task_struct
*curr
)
843 struct list_head
*next
, *head
= &curr
->pi_state_list
;
844 struct futex_pi_state
*pi_state
;
845 struct futex_hash_bucket
*hb
;
846 union futex_key key
= FUTEX_KEY_INIT
;
848 if (!futex_cmpxchg_enabled
)
851 * We are a ZOMBIE and nobody can enqueue itself on
852 * pi_state_list anymore, but we have to be careful
853 * versus waiters unqueueing themselves:
855 raw_spin_lock_irq(&curr
->pi_lock
);
856 while (!list_empty(head
)) {
858 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
860 hb
= hash_futex(&key
);
863 * We can race against put_pi_state() removing itself from the
864 * list (a waiter going away). put_pi_state() will first
865 * decrement the reference count and then modify the list, so
866 * its possible to see the list entry but fail this reference
869 * In that case; drop the locks to let put_pi_state() make
870 * progress and retry the loop.
872 if (!refcount_inc_not_zero(&pi_state
->refcount
)) {
873 raw_spin_unlock_irq(&curr
->pi_lock
);
875 raw_spin_lock_irq(&curr
->pi_lock
);
878 raw_spin_unlock_irq(&curr
->pi_lock
);
880 spin_lock(&hb
->lock
);
881 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
882 raw_spin_lock(&curr
->pi_lock
);
884 * We dropped the pi-lock, so re-check whether this
885 * task still owns the PI-state:
887 if (head
->next
!= next
) {
888 /* retain curr->pi_lock for the loop invariant */
889 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
890 spin_unlock(&hb
->lock
);
891 put_pi_state(pi_state
);
895 WARN_ON(pi_state
->owner
!= curr
);
896 WARN_ON(list_empty(&pi_state
->list
));
897 list_del_init(&pi_state
->list
);
898 pi_state
->owner
= NULL
;
900 raw_spin_unlock(&curr
->pi_lock
);
901 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
902 spin_unlock(&hb
->lock
);
904 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
905 put_pi_state(pi_state
);
907 raw_spin_lock_irq(&curr
->pi_lock
);
909 raw_spin_unlock_irq(&curr
->pi_lock
);
912 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
916 * We need to check the following states:
918 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
920 * [1] NULL | --- | --- | 0 | 0/1 | Valid
921 * [2] NULL | --- | --- | >0 | 0/1 | Valid
923 * [3] Found | NULL | -- | Any | 0/1 | Invalid
925 * [4] Found | Found | NULL | 0 | 1 | Valid
926 * [5] Found | Found | NULL | >0 | 1 | Invalid
928 * [6] Found | Found | task | 0 | 1 | Valid
930 * [7] Found | Found | NULL | Any | 0 | Invalid
932 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
933 * [9] Found | Found | task | 0 | 0 | Invalid
934 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
936 * [1] Indicates that the kernel can acquire the futex atomically. We
937 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
939 * [2] Valid, if TID does not belong to a kernel thread. If no matching
940 * thread is found then it indicates that the owner TID has died.
942 * [3] Invalid. The waiter is queued on a non PI futex
944 * [4] Valid state after exit_robust_list(), which sets the user space
945 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
947 * [5] The user space value got manipulated between exit_robust_list()
948 * and exit_pi_state_list()
950 * [6] Valid state after exit_pi_state_list() which sets the new owner in
951 * the pi_state but cannot access the user space value.
953 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
955 * [8] Owner and user space value match
957 * [9] There is no transient state which sets the user space TID to 0
958 * except exit_robust_list(), but this is indicated by the
959 * FUTEX_OWNER_DIED bit. See [4]
961 * [10] There is no transient state which leaves owner and user space
965 * Serialization and lifetime rules:
969 * hb -> futex_q, relation
970 * futex_q -> pi_state, relation
972 * (cannot be raw because hb can contain arbitrary amount
975 * pi_mutex->wait_lock:
979 * (and pi_mutex 'obviously')
983 * p->pi_state_list -> pi_state->list, 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
)
1326 u32
uninitialized_var(curval
);
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
)
1496 u32
uninitialized_var(curval
), newval
;
1497 struct task_struct
*new_owner
;
1498 bool postunlock
= false;
1499 DEFINE_WAKE_Q(wake_q
);
1502 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1503 if (WARN_ON_ONCE(!new_owner
)) {
1505 * As per the comment in futex_unlock_pi() this should not happen.
1507 * When this happens, give up our locks and try again, giving
1508 * the futex_lock_pi() instance time to complete, either by
1509 * waiting on the rtmutex or removing itself from the futex
1517 * We pass it to the next owner. The WAITERS bit is always kept
1518 * enabled while there is PI state around. We cleanup the owner
1519 * died bit, because we are the owner.
1521 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1523 if (unlikely(should_fail_futex(true)))
1526 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1527 if (!ret
&& (curval
!= uval
)) {
1529 * If a unconditional UNLOCK_PI operation (user space did not
1530 * try the TID->0 transition) raced with a waiter setting the
1531 * FUTEX_WAITERS flag between get_user() and locking the hash
1532 * bucket lock, retry the operation.
1534 if ((FUTEX_TID_MASK
& curval
) == uval
)
1544 * This is a point of no return; once we modify the uval there is no
1545 * going back and subsequent operations must not fail.
1548 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1549 WARN_ON(list_empty(&pi_state
->list
));
1550 list_del_init(&pi_state
->list
);
1551 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1553 raw_spin_lock(&new_owner
->pi_lock
);
1554 WARN_ON(!list_empty(&pi_state
->list
));
1555 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1556 pi_state
->owner
= new_owner
;
1557 raw_spin_unlock(&new_owner
->pi_lock
);
1559 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1562 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1565 rt_mutex_postunlock(&wake_q
);
1571 * Express the locking dependencies for lockdep:
1574 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1577 spin_lock(&hb1
->lock
);
1579 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1580 } else { /* hb1 > hb2 */
1581 spin_lock(&hb2
->lock
);
1582 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1587 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1589 spin_unlock(&hb1
->lock
);
1591 spin_unlock(&hb2
->lock
);
1595 * Wake up waiters matching bitset queued on this futex (uaddr).
1598 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1600 struct futex_hash_bucket
*hb
;
1601 struct futex_q
*this, *next
;
1602 union futex_key key
= FUTEX_KEY_INIT
;
1604 DEFINE_WAKE_Q(wake_q
);
1609 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_READ
);
1610 if (unlikely(ret
!= 0))
1613 hb
= hash_futex(&key
);
1615 /* Make sure we really have tasks to wakeup */
1616 if (!hb_waiters_pending(hb
))
1619 spin_lock(&hb
->lock
);
1621 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1622 if (match_futex (&this->key
, &key
)) {
1623 if (this->pi_state
|| this->rt_waiter
) {
1628 /* Check if one of the bits is set in both bitsets */
1629 if (!(this->bitset
& bitset
))
1632 mark_wake_futex(&wake_q
, this);
1633 if (++ret
>= nr_wake
)
1638 spin_unlock(&hb
->lock
);
1641 put_futex_key(&key
);
1646 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1648 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1649 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1650 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1651 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1654 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1655 if (oparg
< 0 || oparg
> 31) {
1656 char comm
[sizeof(current
->comm
)];
1658 * kill this print and return -EINVAL when userspace
1661 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1662 get_task_comm(comm
, current
), oparg
);
1668 pagefault_disable();
1669 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1675 case FUTEX_OP_CMP_EQ
:
1676 return oldval
== cmparg
;
1677 case FUTEX_OP_CMP_NE
:
1678 return oldval
!= cmparg
;
1679 case FUTEX_OP_CMP_LT
:
1680 return oldval
< cmparg
;
1681 case FUTEX_OP_CMP_GE
:
1682 return oldval
>= cmparg
;
1683 case FUTEX_OP_CMP_LE
:
1684 return oldval
<= cmparg
;
1685 case FUTEX_OP_CMP_GT
:
1686 return oldval
> cmparg
;
1693 * Wake up all waiters hashed on the physical page that is mapped
1694 * to this virtual address:
1697 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1698 int nr_wake
, int nr_wake2
, int op
)
1700 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1701 struct futex_hash_bucket
*hb1
, *hb2
;
1702 struct futex_q
*this, *next
;
1704 DEFINE_WAKE_Q(wake_q
);
1707 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1708 if (unlikely(ret
!= 0))
1710 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
1711 if (unlikely(ret
!= 0))
1714 hb1
= hash_futex(&key1
);
1715 hb2
= hash_futex(&key2
);
1718 double_lock_hb(hb1
, hb2
);
1719 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1720 if (unlikely(op_ret
< 0)) {
1721 double_unlock_hb(hb1
, hb2
);
1723 if (!IS_ENABLED(CONFIG_MMU
) ||
1724 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1726 * we don't get EFAULT from MMU faults if we don't have
1727 * an MMU, but we might get them from range checking
1733 if (op_ret
== -EFAULT
) {
1734 ret
= fault_in_user_writeable(uaddr2
);
1739 if (!(flags
& FLAGS_SHARED
)) {
1744 put_futex_key(&key2
);
1745 put_futex_key(&key1
);
1750 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1751 if (match_futex (&this->key
, &key1
)) {
1752 if (this->pi_state
|| this->rt_waiter
) {
1756 mark_wake_futex(&wake_q
, this);
1757 if (++ret
>= nr_wake
)
1764 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1765 if (match_futex (&this->key
, &key2
)) {
1766 if (this->pi_state
|| this->rt_waiter
) {
1770 mark_wake_futex(&wake_q
, this);
1771 if (++op_ret
>= nr_wake2
)
1779 double_unlock_hb(hb1
, hb2
);
1782 put_futex_key(&key2
);
1784 put_futex_key(&key1
);
1790 * requeue_futex() - Requeue a futex_q from one hb to another
1791 * @q: the futex_q to requeue
1792 * @hb1: the source hash_bucket
1793 * @hb2: the target hash_bucket
1794 * @key2: the new key for the requeued futex_q
1797 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1798 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1802 * If key1 and key2 hash to the same bucket, no need to
1805 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1806 plist_del(&q
->list
, &hb1
->chain
);
1807 hb_waiters_dec(hb1
);
1808 hb_waiters_inc(hb2
);
1809 plist_add(&q
->list
, &hb2
->chain
);
1810 q
->lock_ptr
= &hb2
->lock
;
1816 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1818 * @key: the key of the requeue target futex
1819 * @hb: the hash_bucket of the requeue target futex
1821 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1822 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1823 * to the requeue target futex so the waiter can detect the wakeup on the right
1824 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1825 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1826 * to protect access to the pi_state to fixup the owner later. Must be called
1827 * with both q->lock_ptr and hb->lock held.
1830 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1831 struct futex_hash_bucket
*hb
)
1837 WARN_ON(!q
->rt_waiter
);
1838 q
->rt_waiter
= NULL
;
1840 q
->lock_ptr
= &hb
->lock
;
1842 wake_up_state(q
->task
, TASK_NORMAL
);
1846 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1847 * @pifutex: the user address of the to futex
1848 * @hb1: the from futex hash bucket, must be locked by the caller
1849 * @hb2: the to futex hash bucket, must be locked by the caller
1850 * @key1: the from futex key
1851 * @key2: the to futex key
1852 * @ps: address to store the pi_state pointer
1853 * @exiting: Pointer to store the task pointer of the owner task
1854 * which is in the middle of exiting
1855 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1857 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1858 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1859 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1860 * hb1 and hb2 must be held by the caller.
1862 * @exiting is only set when the return value is -EBUSY. If so, this holds
1863 * a refcount on the exiting task on return and the caller needs to drop it
1864 * after waiting for the exit to complete.
1867 * - 0 - failed to acquire the lock atomically;
1868 * - >0 - acquired the lock, return value is vpid of the top_waiter
1872 futex_proxy_trylock_atomic(u32 __user
*pifutex
, struct futex_hash_bucket
*hb1
,
1873 struct futex_hash_bucket
*hb2
, union futex_key
*key1
,
1874 union futex_key
*key2
, struct futex_pi_state
**ps
,
1875 struct task_struct
**exiting
, int set_waiters
)
1877 struct futex_q
*top_waiter
= NULL
;
1881 if (get_futex_value_locked(&curval
, pifutex
))
1884 if (unlikely(should_fail_futex(true)))
1888 * Find the top_waiter and determine if there are additional waiters.
1889 * If the caller intends to requeue more than 1 waiter to pifutex,
1890 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1891 * as we have means to handle the possible fault. If not, don't set
1892 * the bit unecessarily as it will force the subsequent unlock to enter
1895 top_waiter
= futex_top_waiter(hb1
, key1
);
1897 /* There are no waiters, nothing for us to do. */
1901 /* Ensure we requeue to the expected futex. */
1902 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1906 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1907 * the contended case or if set_waiters is 1. The pi_state is returned
1908 * in ps in contended cases.
1910 vpid
= task_pid_vnr(top_waiter
->task
);
1911 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1912 exiting
, set_waiters
);
1914 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1921 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1922 * @uaddr1: source futex user address
1923 * @flags: futex flags (FLAGS_SHARED, etc.)
1924 * @uaddr2: target futex user address
1925 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1926 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1927 * @cmpval: @uaddr1 expected value (or %NULL)
1928 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1929 * pi futex (pi to pi requeue is not supported)
1931 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1932 * uaddr2 atomically on behalf of the top waiter.
1935 * - >=0 - on success, the number of tasks requeued or woken;
1938 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1939 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1940 u32
*cmpval
, int requeue_pi
)
1942 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1943 int task_count
= 0, ret
;
1944 struct futex_pi_state
*pi_state
= NULL
;
1945 struct futex_hash_bucket
*hb1
, *hb2
;
1946 struct futex_q
*this, *next
;
1947 DEFINE_WAKE_Q(wake_q
);
1949 if (nr_wake
< 0 || nr_requeue
< 0)
1953 * When PI not supported: return -ENOSYS if requeue_pi is true,
1954 * consequently the compiler knows requeue_pi is always false past
1955 * this point which will optimize away all the conditional code
1958 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1963 * Requeue PI only works on two distinct uaddrs. This
1964 * check is only valid for private futexes. See below.
1966 if (uaddr1
== uaddr2
)
1970 * requeue_pi requires a pi_state, try to allocate it now
1971 * without any locks in case it fails.
1973 if (refill_pi_state_cache())
1976 * requeue_pi must wake as many tasks as it can, up to nr_wake
1977 * + nr_requeue, since it acquires the rt_mutex prior to
1978 * returning to userspace, so as to not leave the rt_mutex with
1979 * waiters and no owner. However, second and third wake-ups
1980 * cannot be predicted as they involve race conditions with the
1981 * first wake and a fault while looking up the pi_state. Both
1982 * pthread_cond_signal() and pthread_cond_broadcast() should
1990 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, FUTEX_READ
);
1991 if (unlikely(ret
!= 0))
1993 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1994 requeue_pi
? FUTEX_WRITE
: FUTEX_READ
);
1995 if (unlikely(ret
!= 0))
1999 * The check above which compares uaddrs is not sufficient for
2000 * shared futexes. We need to compare the keys:
2002 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
2007 hb1
= hash_futex(&key1
);
2008 hb2
= hash_futex(&key2
);
2011 hb_waiters_inc(hb2
);
2012 double_lock_hb(hb1
, hb2
);
2014 if (likely(cmpval
!= NULL
)) {
2017 ret
= get_futex_value_locked(&curval
, uaddr1
);
2019 if (unlikely(ret
)) {
2020 double_unlock_hb(hb1
, hb2
);
2021 hb_waiters_dec(hb2
);
2023 ret
= get_user(curval
, uaddr1
);
2027 if (!(flags
& FLAGS_SHARED
))
2030 put_futex_key(&key2
);
2031 put_futex_key(&key1
);
2034 if (curval
!= *cmpval
) {
2040 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2041 struct task_struct
*exiting
= NULL
;
2044 * Attempt to acquire uaddr2 and wake the top waiter. If we
2045 * intend to requeue waiters, force setting the FUTEX_WAITERS
2046 * bit. We force this here where we are able to easily handle
2047 * faults rather in the requeue loop below.
2049 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2051 &exiting
, nr_requeue
);
2054 * At this point the top_waiter has either taken uaddr2 or is
2055 * waiting on it. If the former, then the pi_state will not
2056 * exist yet, look it up one more time to ensure we have a
2057 * reference to it. If the lock was taken, ret contains the
2058 * vpid of the top waiter task.
2059 * If the lock was not taken, we have pi_state and an initial
2060 * refcount on it. In case of an error we have nothing.
2066 * If we acquired the lock, then the user space value
2067 * of uaddr2 should be vpid. It cannot be changed by
2068 * the top waiter as it is blocked on hb2 lock if it
2069 * tries to do so. If something fiddled with it behind
2070 * our back the pi state lookup might unearth it. So
2071 * we rather use the known value than rereading and
2072 * handing potential crap to lookup_pi_state.
2074 * If that call succeeds then we have pi_state and an
2075 * initial refcount on it.
2077 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
,
2078 &pi_state
, &exiting
);
2083 /* We hold a reference on the pi state. */
2086 /* If the above failed, then pi_state is NULL */
2088 double_unlock_hb(hb1
, hb2
);
2089 hb_waiters_dec(hb2
);
2090 put_futex_key(&key2
);
2091 put_futex_key(&key1
);
2092 ret
= fault_in_user_writeable(uaddr2
);
2099 * Two reasons for this:
2100 * - EBUSY: Owner is exiting and we just wait for the
2102 * - EAGAIN: The user space value changed.
2104 double_unlock_hb(hb1
, hb2
);
2105 hb_waiters_dec(hb2
);
2106 put_futex_key(&key2
);
2107 put_futex_key(&key1
);
2109 * Handle the case where the owner is in the middle of
2110 * exiting. Wait for the exit to complete otherwise
2111 * this task might loop forever, aka. live lock.
2113 wait_for_owner_exiting(ret
, exiting
);
2121 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2122 if (task_count
- nr_wake
>= nr_requeue
)
2125 if (!match_futex(&this->key
, &key1
))
2129 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2130 * be paired with each other and no other futex ops.
2132 * We should never be requeueing a futex_q with a pi_state,
2133 * which is awaiting a futex_unlock_pi().
2135 if ((requeue_pi
&& !this->rt_waiter
) ||
2136 (!requeue_pi
&& this->rt_waiter
) ||
2143 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2144 * lock, we already woke the top_waiter. If not, it will be
2145 * woken by futex_unlock_pi().
2147 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2148 mark_wake_futex(&wake_q
, this);
2152 /* Ensure we requeue to the expected futex for requeue_pi. */
2153 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2159 * Requeue nr_requeue waiters and possibly one more in the case
2160 * of requeue_pi if we couldn't acquire the lock atomically.
2164 * Prepare the waiter to take the rt_mutex. Take a
2165 * refcount on the pi_state and store the pointer in
2166 * the futex_q object of the waiter.
2168 get_pi_state(pi_state
);
2169 this->pi_state
= pi_state
;
2170 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2175 * We got the lock. We do neither drop the
2176 * refcount on pi_state nor clear
2177 * this->pi_state because the waiter needs the
2178 * pi_state for cleaning up the user space
2179 * value. It will drop the refcount after
2182 requeue_pi_wake_futex(this, &key2
, hb2
);
2186 * rt_mutex_start_proxy_lock() detected a
2187 * potential deadlock when we tried to queue
2188 * that waiter. Drop the pi_state reference
2189 * which we took above and remove the pointer
2190 * to the state from the waiters futex_q
2193 this->pi_state
= NULL
;
2194 put_pi_state(pi_state
);
2196 * We stop queueing more waiters and let user
2197 * space deal with the mess.
2202 requeue_futex(this, hb1
, hb2
, &key2
);
2206 * We took an extra initial reference to the pi_state either
2207 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2208 * need to drop it here again.
2210 put_pi_state(pi_state
);
2213 double_unlock_hb(hb1
, hb2
);
2215 hb_waiters_dec(hb2
);
2218 put_futex_key(&key2
);
2220 put_futex_key(&key1
);
2222 return ret
? ret
: task_count
;
2225 /* The key must be already stored in q->key. */
2226 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2227 __acquires(&hb
->lock
)
2229 struct futex_hash_bucket
*hb
;
2231 hb
= hash_futex(&q
->key
);
2234 * Increment the counter before taking the lock so that
2235 * a potential waker won't miss a to-be-slept task that is
2236 * waiting for the spinlock. This is safe as all queue_lock()
2237 * users end up calling queue_me(). Similarly, for housekeeping,
2238 * decrement the counter at queue_unlock() when some error has
2239 * occurred and we don't end up adding the task to the list.
2241 hb_waiters_inc(hb
); /* implies smp_mb(); (A) */
2243 q
->lock_ptr
= &hb
->lock
;
2245 spin_lock(&hb
->lock
);
2250 queue_unlock(struct futex_hash_bucket
*hb
)
2251 __releases(&hb
->lock
)
2253 spin_unlock(&hb
->lock
);
2257 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2262 * The priority used to register this element is
2263 * - either the real thread-priority for the real-time threads
2264 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2265 * - or MAX_RT_PRIO for non-RT threads.
2266 * Thus, all RT-threads are woken first in priority order, and
2267 * the others are woken last, in FIFO order.
2269 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2271 plist_node_init(&q
->list
, prio
);
2272 plist_add(&q
->list
, &hb
->chain
);
2277 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2278 * @q: The futex_q to enqueue
2279 * @hb: The destination hash bucket
2281 * The hb->lock must be held by the caller, and is released here. A call to
2282 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2283 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2284 * or nothing if the unqueue is done as part of the wake process and the unqueue
2285 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2288 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2289 __releases(&hb
->lock
)
2292 spin_unlock(&hb
->lock
);
2296 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2297 * @q: The futex_q to unqueue
2299 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2300 * be paired with exactly one earlier call to queue_me().
2303 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2304 * - 0 - if the futex_q was already removed by the waking thread
2306 static int unqueue_me(struct futex_q
*q
)
2308 spinlock_t
*lock_ptr
;
2311 /* In the common case we don't take the spinlock, which is nice. */
2314 * q->lock_ptr can change between this read and the following spin_lock.
2315 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2316 * optimizing lock_ptr out of the logic below.
2318 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2319 if (lock_ptr
!= NULL
) {
2320 spin_lock(lock_ptr
);
2322 * q->lock_ptr can change between reading it and
2323 * spin_lock(), causing us to take the wrong lock. This
2324 * corrects the race condition.
2326 * Reasoning goes like this: if we have the wrong lock,
2327 * q->lock_ptr must have changed (maybe several times)
2328 * between reading it and the spin_lock(). It can
2329 * change again after the spin_lock() but only if it was
2330 * already changed before the spin_lock(). It cannot,
2331 * however, change back to the original value. Therefore
2332 * we can detect whether we acquired the correct lock.
2334 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2335 spin_unlock(lock_ptr
);
2340 BUG_ON(q
->pi_state
);
2342 spin_unlock(lock_ptr
);
2350 * PI futexes can not be requeued and must remove themself from the
2351 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2354 static void unqueue_me_pi(struct futex_q
*q
)
2355 __releases(q
->lock_ptr
)
2359 BUG_ON(!q
->pi_state
);
2360 put_pi_state(q
->pi_state
);
2363 spin_unlock(q
->lock_ptr
);
2366 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2367 struct task_struct
*argowner
)
2369 struct futex_pi_state
*pi_state
= q
->pi_state
;
2370 u32 uval
, uninitialized_var(curval
), newval
;
2371 struct task_struct
*oldowner
, *newowner
;
2375 lockdep_assert_held(q
->lock_ptr
);
2377 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2379 oldowner
= pi_state
->owner
;
2382 * We are here because either:
2384 * - we stole the lock and pi_state->owner needs updating to reflect
2385 * that (@argowner == current),
2389 * - someone stole our lock and we need to fix things to point to the
2390 * new owner (@argowner == NULL).
2392 * Either way, we have to replace the TID in the user space variable.
2393 * This must be atomic as we have to preserve the owner died bit here.
2395 * Note: We write the user space value _before_ changing the pi_state
2396 * because we can fault here. Imagine swapped out pages or a fork
2397 * that marked all the anonymous memory readonly for cow.
2399 * Modifying pi_state _before_ the user space value would leave the
2400 * pi_state in an inconsistent state when we fault here, because we
2401 * need to drop the locks to handle the fault. This might be observed
2402 * in the PID check in lookup_pi_state.
2406 if (oldowner
!= current
) {
2408 * We raced against a concurrent self; things are
2409 * already fixed up. Nothing to do.
2415 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2416 /* We got the lock after all, nothing to fix. */
2422 * Since we just failed the trylock; there must be an owner.
2424 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2427 WARN_ON_ONCE(argowner
!= current
);
2428 if (oldowner
== current
) {
2430 * We raced against a concurrent self; things are
2431 * already fixed up. Nothing to do.
2436 newowner
= argowner
;
2439 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2441 if (!pi_state
->owner
)
2442 newtid
|= FUTEX_OWNER_DIED
;
2444 err
= get_futex_value_locked(&uval
, uaddr
);
2449 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2451 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2461 * We fixed up user space. Now we need to fix the pi_state
2464 if (pi_state
->owner
!= NULL
) {
2465 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2466 WARN_ON(list_empty(&pi_state
->list
));
2467 list_del_init(&pi_state
->list
);
2468 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2471 pi_state
->owner
= newowner
;
2473 raw_spin_lock(&newowner
->pi_lock
);
2474 WARN_ON(!list_empty(&pi_state
->list
));
2475 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2476 raw_spin_unlock(&newowner
->pi_lock
);
2477 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2482 * In order to reschedule or handle a page fault, we need to drop the
2483 * locks here. In the case of a fault, this gives the other task
2484 * (either the highest priority waiter itself or the task which stole
2485 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2486 * are back from handling the fault we need to check the pi_state after
2487 * reacquiring the locks and before trying to do another fixup. When
2488 * the fixup has been done already we simply return.
2490 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2491 * drop hb->lock since the caller owns the hb -> futex_q relation.
2492 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2495 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2496 spin_unlock(q
->lock_ptr
);
2500 ret
= fault_in_user_writeable(uaddr
);
2514 spin_lock(q
->lock_ptr
);
2515 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2518 * Check if someone else fixed it for us:
2520 if (pi_state
->owner
!= oldowner
) {
2531 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2535 static long futex_wait_restart(struct restart_block
*restart
);
2538 * fixup_owner() - Post lock pi_state and corner case management
2539 * @uaddr: user address of the futex
2540 * @q: futex_q (contains pi_state and access to the rt_mutex)
2541 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2543 * After attempting to lock an rt_mutex, this function is called to cleanup
2544 * the pi_state owner as well as handle race conditions that may allow us to
2545 * acquire the lock. Must be called with the hb lock held.
2548 * - 1 - success, lock taken;
2549 * - 0 - success, lock not taken;
2550 * - <0 - on error (-EFAULT)
2552 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2558 * Got the lock. We might not be the anticipated owner if we
2559 * did a lock-steal - fix up the PI-state in that case:
2561 * Speculative pi_state->owner read (we don't hold wait_lock);
2562 * since we own the lock pi_state->owner == current is the
2563 * stable state, anything else needs more attention.
2565 if (q
->pi_state
->owner
!= current
)
2566 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2571 * If we didn't get the lock; check if anybody stole it from us. In
2572 * that case, we need to fix up the uval to point to them instead of
2573 * us, otherwise bad things happen. [10]
2575 * Another speculative read; pi_state->owner == current is unstable
2576 * but needs our attention.
2578 if (q
->pi_state
->owner
== current
) {
2579 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2584 * Paranoia check. If we did not take the lock, then we should not be
2585 * the owner of the rt_mutex.
2587 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2588 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2589 "pi-state %p\n", ret
,
2590 q
->pi_state
->pi_mutex
.owner
,
2591 q
->pi_state
->owner
);
2595 return ret
? ret
: locked
;
2599 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2600 * @hb: the futex hash bucket, must be locked by the caller
2601 * @q: the futex_q to queue up on
2602 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2604 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2605 struct hrtimer_sleeper
*timeout
)
2608 * The task state is guaranteed to be set before another task can
2609 * wake it. set_current_state() is implemented using smp_store_mb() and
2610 * queue_me() calls spin_unlock() upon completion, both serializing
2611 * access to the hash list and forcing another memory barrier.
2613 set_current_state(TASK_INTERRUPTIBLE
);
2618 hrtimer_sleeper_start_expires(timeout
, HRTIMER_MODE_ABS
);
2621 * If we have been removed from the hash list, then another task
2622 * has tried to wake us, and we can skip the call to schedule().
2624 if (likely(!plist_node_empty(&q
->list
))) {
2626 * If the timer has already expired, current will already be
2627 * flagged for rescheduling. Only call schedule if there
2628 * is no timeout, or if it has yet to expire.
2630 if (!timeout
|| timeout
->task
)
2631 freezable_schedule();
2633 __set_current_state(TASK_RUNNING
);
2637 * futex_wait_setup() - Prepare to wait on a futex
2638 * @uaddr: the futex userspace address
2639 * @val: the expected value
2640 * @flags: futex flags (FLAGS_SHARED, etc.)
2641 * @q: the associated futex_q
2642 * @hb: storage for hash_bucket pointer to be returned to caller
2644 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2645 * compare it with the expected value. Handle atomic faults internally.
2646 * Return with the hb lock held and a q.key reference on success, and unlocked
2647 * with no q.key reference on failure.
2650 * - 0 - uaddr contains val and hb has been locked;
2651 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2653 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2654 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2660 * Access the page AFTER the hash-bucket is locked.
2661 * Order is important:
2663 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2664 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2666 * The basic logical guarantee of a futex is that it blocks ONLY
2667 * if cond(var) is known to be true at the time of blocking, for
2668 * any cond. If we locked the hash-bucket after testing *uaddr, that
2669 * would open a race condition where we could block indefinitely with
2670 * cond(var) false, which would violate the guarantee.
2672 * On the other hand, we insert q and release the hash-bucket only
2673 * after testing *uaddr. This guarantees that futex_wait() will NOT
2674 * absorb a wakeup if *uaddr does not match the desired values
2675 * while the syscall executes.
2678 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, FUTEX_READ
);
2679 if (unlikely(ret
!= 0))
2683 *hb
= queue_lock(q
);
2685 ret
= get_futex_value_locked(&uval
, uaddr
);
2690 ret
= get_user(uval
, uaddr
);
2694 if (!(flags
& FLAGS_SHARED
))
2697 put_futex_key(&q
->key
);
2708 put_futex_key(&q
->key
);
2712 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2713 ktime_t
*abs_time
, u32 bitset
)
2715 struct hrtimer_sleeper timeout
, *to
;
2716 struct restart_block
*restart
;
2717 struct futex_hash_bucket
*hb
;
2718 struct futex_q q
= futex_q_init
;
2725 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
2726 current
->timer_slack_ns
);
2729 * Prepare to wait on uaddr. On success, holds hb lock and increments
2732 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2736 /* queue_me and wait for wakeup, timeout, or a signal. */
2737 futex_wait_queue_me(hb
, &q
, to
);
2739 /* If we were woken (and unqueued), we succeeded, whatever. */
2741 /* unqueue_me() drops q.key ref */
2742 if (!unqueue_me(&q
))
2745 if (to
&& !to
->task
)
2749 * We expect signal_pending(current), but we might be the
2750 * victim of a spurious wakeup as well.
2752 if (!signal_pending(current
))
2759 restart
= ¤t
->restart_block
;
2760 restart
->fn
= futex_wait_restart
;
2761 restart
->futex
.uaddr
= uaddr
;
2762 restart
->futex
.val
= val
;
2763 restart
->futex
.time
= *abs_time
;
2764 restart
->futex
.bitset
= bitset
;
2765 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2767 ret
= -ERESTART_RESTARTBLOCK
;
2771 hrtimer_cancel(&to
->timer
);
2772 destroy_hrtimer_on_stack(&to
->timer
);
2778 static long futex_wait_restart(struct restart_block
*restart
)
2780 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2781 ktime_t t
, *tp
= NULL
;
2783 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2784 t
= restart
->futex
.time
;
2787 restart
->fn
= do_no_restart_syscall
;
2789 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2790 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2795 * Userspace tried a 0 -> TID atomic transition of the futex value
2796 * and failed. The kernel side here does the whole locking operation:
2797 * if there are waiters then it will block as a consequence of relying
2798 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2799 * a 0 value of the futex too.).
2801 * Also serves as futex trylock_pi()'ing, and due semantics.
2803 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2804 ktime_t
*time
, int trylock
)
2806 struct hrtimer_sleeper timeout
, *to
;
2807 struct futex_pi_state
*pi_state
= NULL
;
2808 struct task_struct
*exiting
= NULL
;
2809 struct rt_mutex_waiter rt_waiter
;
2810 struct futex_hash_bucket
*hb
;
2811 struct futex_q q
= futex_q_init
;
2814 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2817 if (refill_pi_state_cache())
2820 to
= futex_setup_timer(time
, &timeout
, FLAGS_CLOCKRT
, 0);
2823 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, FUTEX_WRITE
);
2824 if (unlikely(ret
!= 0))
2828 hb
= queue_lock(&q
);
2830 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
,
2832 if (unlikely(ret
)) {
2834 * Atomic work succeeded and we got the lock,
2835 * or failed. Either way, we do _not_ block.
2839 /* We got the lock. */
2841 goto out_unlock_put_key
;
2847 * Two reasons for this:
2848 * - EBUSY: Task is exiting and we just wait for the
2850 * - EAGAIN: The user space value changed.
2853 put_futex_key(&q
.key
);
2855 * Handle the case where the owner is in the middle of
2856 * exiting. Wait for the exit to complete otherwise
2857 * this task might loop forever, aka. live lock.
2859 wait_for_owner_exiting(ret
, exiting
);
2863 goto out_unlock_put_key
;
2867 WARN_ON(!q
.pi_state
);
2870 * Only actually queue now that the atomic ops are done:
2875 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2876 /* Fixup the trylock return value: */
2877 ret
= ret
? 0 : -EWOULDBLOCK
;
2881 rt_mutex_init_waiter(&rt_waiter
);
2884 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2885 * hold it while doing rt_mutex_start_proxy(), because then it will
2886 * include hb->lock in the blocking chain, even through we'll not in
2887 * fact hold it while blocking. This will lead it to report -EDEADLK
2888 * and BUG when futex_unlock_pi() interleaves with this.
2890 * Therefore acquire wait_lock while holding hb->lock, but drop the
2891 * latter before calling __rt_mutex_start_proxy_lock(). This
2892 * interleaves with futex_unlock_pi() -- which does a similar lock
2893 * handoff -- such that the latter can observe the futex_q::pi_state
2894 * before __rt_mutex_start_proxy_lock() is done.
2896 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2897 spin_unlock(q
.lock_ptr
);
2899 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2900 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2901 * it sees the futex_q::pi_state.
2903 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2904 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2913 hrtimer_sleeper_start_expires(to
, HRTIMER_MODE_ABS
);
2915 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2918 spin_lock(q
.lock_ptr
);
2920 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2921 * first acquire the hb->lock before removing the lock from the
2922 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2925 * In particular; it is important that futex_unlock_pi() can not
2926 * observe this inconsistency.
2928 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2933 * Fixup the pi_state owner and possibly acquire the lock if we
2936 res
= fixup_owner(uaddr
, &q
, !ret
);
2938 * If fixup_owner() returned an error, proprogate that. If it acquired
2939 * the lock, clear our -ETIMEDOUT or -EINTR.
2942 ret
= (res
< 0) ? res
: 0;
2945 * If fixup_owner() faulted and was unable to handle the fault, unlock
2946 * it and return the fault to userspace.
2948 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2949 pi_state
= q
.pi_state
;
2950 get_pi_state(pi_state
);
2953 /* Unqueue and drop the lock */
2957 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2958 put_pi_state(pi_state
);
2967 put_futex_key(&q
.key
);
2970 hrtimer_cancel(&to
->timer
);
2971 destroy_hrtimer_on_stack(&to
->timer
);
2973 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2978 ret
= fault_in_user_writeable(uaddr
);
2982 if (!(flags
& FLAGS_SHARED
))
2985 put_futex_key(&q
.key
);
2990 * Userspace attempted a TID -> 0 atomic transition, and failed.
2991 * This is the in-kernel slowpath: we look up the PI state (if any),
2992 * and do the rt-mutex unlock.
2994 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2996 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2997 union futex_key key
= FUTEX_KEY_INIT
;
2998 struct futex_hash_bucket
*hb
;
2999 struct futex_q
*top_waiter
;
3002 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3006 if (get_user(uval
, uaddr
))
3009 * We release only a lock we actually own:
3011 if ((uval
& FUTEX_TID_MASK
) != vpid
)
3014 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, FUTEX_WRITE
);
3018 hb
= hash_futex(&key
);
3019 spin_lock(&hb
->lock
);
3022 * Check waiters first. We do not trust user space values at
3023 * all and we at least want to know if user space fiddled
3024 * with the futex value instead of blindly unlocking.
3026 top_waiter
= futex_top_waiter(hb
, &key
);
3028 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3035 * If current does not own the pi_state then the futex is
3036 * inconsistent and user space fiddled with the futex value.
3038 if (pi_state
->owner
!= current
)
3041 get_pi_state(pi_state
);
3043 * By taking wait_lock while still holding hb->lock, we ensure
3044 * there is no point where we hold neither; and therefore
3045 * wake_futex_pi() must observe a state consistent with what we
3048 * In particular; this forces __rt_mutex_start_proxy() to
3049 * complete such that we're guaranteed to observe the
3050 * rt_waiter. Also see the WARN in wake_futex_pi().
3052 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3053 spin_unlock(&hb
->lock
);
3055 /* drops pi_state->pi_mutex.wait_lock */
3056 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3058 put_pi_state(pi_state
);
3061 * Success, we're done! No tricky corner cases.
3066 * The atomic access to the futex value generated a
3067 * pagefault, so retry the user-access and the wakeup:
3072 * A unconditional UNLOCK_PI op raced against a waiter
3073 * setting the FUTEX_WAITERS bit. Try again.
3078 * wake_futex_pi has detected invalid state. Tell user
3085 * We have no kernel internal state, i.e. no waiters in the
3086 * kernel. Waiters which are about to queue themselves are stuck
3087 * on hb->lock. So we can safely ignore them. We do neither
3088 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3091 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3092 spin_unlock(&hb
->lock
);
3107 * If uval has changed, let user space handle it.
3109 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3112 spin_unlock(&hb
->lock
);
3114 put_futex_key(&key
);
3118 put_futex_key(&key
);
3123 put_futex_key(&key
);
3125 ret
= fault_in_user_writeable(uaddr
);
3133 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3134 * @hb: the hash_bucket futex_q was original enqueued on
3135 * @q: the futex_q woken while waiting to be requeued
3136 * @key2: the futex_key of the requeue target futex
3137 * @timeout: the timeout associated with the wait (NULL if none)
3139 * Detect if the task was woken on the initial futex as opposed to the requeue
3140 * target futex. If so, determine if it was a timeout or a signal that caused
3141 * the wakeup and return the appropriate error code to the caller. Must be
3142 * called with the hb lock held.
3145 * - 0 = no early wakeup detected;
3146 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3149 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3150 struct futex_q
*q
, union futex_key
*key2
,
3151 struct hrtimer_sleeper
*timeout
)
3156 * With the hb lock held, we avoid races while we process the wakeup.
3157 * We only need to hold hb (and not hb2) to ensure atomicity as the
3158 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3159 * It can't be requeued from uaddr2 to something else since we don't
3160 * support a PI aware source futex for requeue.
3162 if (!match_futex(&q
->key
, key2
)) {
3163 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3165 * We were woken prior to requeue by a timeout or a signal.
3166 * Unqueue the futex_q and determine which it was.
3168 plist_del(&q
->list
, &hb
->chain
);
3171 /* Handle spurious wakeups gracefully */
3173 if (timeout
&& !timeout
->task
)
3175 else if (signal_pending(current
))
3176 ret
= -ERESTARTNOINTR
;
3182 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3183 * @uaddr: the futex we initially wait on (non-pi)
3184 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3185 * the same type, no requeueing from private to shared, etc.
3186 * @val: the expected value of uaddr
3187 * @abs_time: absolute timeout
3188 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3189 * @uaddr2: the pi futex we will take prior to returning to user-space
3191 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3192 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3193 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3194 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3195 * without one, the pi logic would not know which task to boost/deboost, if
3196 * there was a need to.
3198 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3199 * via the following--
3200 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3201 * 2) wakeup on uaddr2 after a requeue
3205 * If 3, cleanup and return -ERESTARTNOINTR.
3207 * If 2, we may then block on trying to take the rt_mutex and return via:
3208 * 5) successful lock
3211 * 8) other lock acquisition failure
3213 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3215 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3221 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3222 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3225 struct hrtimer_sleeper timeout
, *to
;
3226 struct futex_pi_state
*pi_state
= NULL
;
3227 struct rt_mutex_waiter rt_waiter
;
3228 struct futex_hash_bucket
*hb
;
3229 union futex_key key2
= FUTEX_KEY_INIT
;
3230 struct futex_q q
= futex_q_init
;
3233 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3236 if (uaddr
== uaddr2
)
3242 to
= futex_setup_timer(abs_time
, &timeout
, flags
,
3243 current
->timer_slack_ns
);
3246 * The waiter is allocated on our stack, manipulated by the requeue
3247 * code while we sleep on uaddr.
3249 rt_mutex_init_waiter(&rt_waiter
);
3251 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, FUTEX_WRITE
);
3252 if (unlikely(ret
!= 0))
3256 q
.rt_waiter
= &rt_waiter
;
3257 q
.requeue_pi_key
= &key2
;
3260 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3263 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3268 * The check above which compares uaddrs is not sufficient for
3269 * shared futexes. We need to compare the keys:
3271 if (match_futex(&q
.key
, &key2
)) {
3277 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3278 futex_wait_queue_me(hb
, &q
, to
);
3280 spin_lock(&hb
->lock
);
3281 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3282 spin_unlock(&hb
->lock
);
3287 * In order for us to be here, we know our q.key == key2, and since
3288 * we took the hb->lock above, we also know that futex_requeue() has
3289 * completed and we no longer have to concern ourselves with a wakeup
3290 * race with the atomic proxy lock acquisition by the requeue code. The
3291 * futex_requeue dropped our key1 reference and incremented our key2
3295 /* Check if the requeue code acquired the second futex for us. */
3298 * Got the lock. We might not be the anticipated owner if we
3299 * did a lock-steal - fix up the PI-state in that case.
3301 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3302 spin_lock(q
.lock_ptr
);
3303 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3304 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3305 pi_state
= q
.pi_state
;
3306 get_pi_state(pi_state
);
3309 * Drop the reference to the pi state which
3310 * the requeue_pi() code acquired for us.
3312 put_pi_state(q
.pi_state
);
3313 spin_unlock(q
.lock_ptr
);
3316 struct rt_mutex
*pi_mutex
;
3319 * We have been woken up by futex_unlock_pi(), a timeout, or a
3320 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3323 WARN_ON(!q
.pi_state
);
3324 pi_mutex
= &q
.pi_state
->pi_mutex
;
3325 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3327 spin_lock(q
.lock_ptr
);
3328 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3331 debug_rt_mutex_free_waiter(&rt_waiter
);
3333 * Fixup the pi_state owner and possibly acquire the lock if we
3336 res
= fixup_owner(uaddr2
, &q
, !ret
);
3338 * If fixup_owner() returned an error, proprogate that. If it
3339 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3342 ret
= (res
< 0) ? res
: 0;
3345 * If fixup_pi_state_owner() faulted and was unable to handle
3346 * the fault, unlock the rt_mutex and return the fault to
3349 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3350 pi_state
= q
.pi_state
;
3351 get_pi_state(pi_state
);
3354 /* Unqueue and drop the lock. */
3359 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3360 put_pi_state(pi_state
);
3363 if (ret
== -EINTR
) {
3365 * We've already been requeued, but cannot restart by calling
3366 * futex_lock_pi() directly. We could restart this syscall, but
3367 * it would detect that the user space "val" changed and return
3368 * -EWOULDBLOCK. Save the overhead of the restart and return
3369 * -EWOULDBLOCK directly.
3375 put_futex_key(&q
.key
);
3377 put_futex_key(&key2
);
3381 hrtimer_cancel(&to
->timer
);
3382 destroy_hrtimer_on_stack(&to
->timer
);
3388 * Support for robust futexes: the kernel cleans up held futexes at
3391 * Implementation: user-space maintains a per-thread list of locks it
3392 * is holding. Upon do_exit(), the kernel carefully walks this list,
3393 * and marks all locks that are owned by this thread with the
3394 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3395 * always manipulated with the lock held, so the list is private and
3396 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3397 * field, to allow the kernel to clean up if the thread dies after
3398 * acquiring the lock, but just before it could have added itself to
3399 * the list. There can only be one such pending lock.
3403 * sys_set_robust_list() - Set the robust-futex list head of a task
3404 * @head: pointer to the list-head
3405 * @len: length of the list-head, as userspace expects
3407 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3410 if (!futex_cmpxchg_enabled
)
3413 * The kernel knows only one size for now:
3415 if (unlikely(len
!= sizeof(*head
)))
3418 current
->robust_list
= head
;
3424 * sys_get_robust_list() - Get the robust-futex list head of a task
3425 * @pid: pid of the process [zero for current task]
3426 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3427 * @len_ptr: pointer to a length field, the kernel fills in the header size
3429 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3430 struct robust_list_head __user
* __user
*, head_ptr
,
3431 size_t __user
*, len_ptr
)
3433 struct robust_list_head __user
*head
;
3435 struct task_struct
*p
;
3437 if (!futex_cmpxchg_enabled
)
3446 p
= find_task_by_vpid(pid
);
3452 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3455 head
= p
->robust_list
;
3458 if (put_user(sizeof(*head
), len_ptr
))
3460 return put_user(head
, head_ptr
);
3468 /* Constants for the pending_op argument of handle_futex_death */
3469 #define HANDLE_DEATH_PENDING true
3470 #define HANDLE_DEATH_LIST false
3473 * Process a futex-list entry, check whether it's owned by the
3474 * dying task, and do notification if so:
3476 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3477 bool pi
, bool pending_op
)
3479 u32 uval
, uninitialized_var(nval
), mval
;
3482 /* Futex address must be 32bit aligned */
3483 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3487 if (get_user(uval
, uaddr
))
3491 * Special case for regular (non PI) futexes. The unlock path in
3492 * user space has two race scenarios:
3494 * 1. The unlock path releases the user space futex value and
3495 * before it can execute the futex() syscall to wake up
3496 * waiters it is killed.
3498 * 2. A woken up waiter is killed before it can acquire the
3499 * futex in user space.
3501 * In both cases the TID validation below prevents a wakeup of
3502 * potential waiters which can cause these waiters to block
3505 * In both cases the following conditions are met:
3507 * 1) task->robust_list->list_op_pending != NULL
3508 * @pending_op == true
3509 * 2) User space futex value == 0
3510 * 3) Regular futex: @pi == false
3512 * If these conditions are met, it is safe to attempt waking up a
3513 * potential waiter without touching the user space futex value and
3514 * trying to set the OWNER_DIED bit. The user space futex value is
3515 * uncontended and the rest of the user space mutex state is
3516 * consistent, so a woken waiter will just take over the
3517 * uncontended futex. Setting the OWNER_DIED bit would create
3518 * inconsistent state and malfunction of the user space owner died
3521 if (pending_op
&& !pi
&& !uval
) {
3522 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3526 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3530 * Ok, this dying thread is truly holding a futex
3531 * of interest. Set the OWNER_DIED bit atomically
3532 * via cmpxchg, and if the value had FUTEX_WAITERS
3533 * set, wake up a waiter (if any). (We have to do a
3534 * futex_wake() even if OWNER_DIED is already set -
3535 * to handle the rare but possible case of recursive
3536 * thread-death.) The rest of the cleanup is done in
3539 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3542 * We are not holding a lock here, but we want to have
3543 * the pagefault_disable/enable() protection because
3544 * we want to handle the fault gracefully. If the
3545 * access fails we try to fault in the futex with R/W
3546 * verification via get_user_pages. get_user() above
3547 * does not guarantee R/W access. If that fails we
3548 * give up and leave the futex locked.
3550 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3553 if (fault_in_user_writeable(uaddr
))
3571 * Wake robust non-PI futexes here. The wakeup of
3572 * PI futexes happens in exit_pi_state():
3574 if (!pi
&& (uval
& FUTEX_WAITERS
))
3575 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3581 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3583 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3584 struct robust_list __user
* __user
*head
,
3587 unsigned long uentry
;
3589 if (get_user(uentry
, (unsigned long __user
*)head
))
3592 *entry
= (void __user
*)(uentry
& ~1UL);
3599 * Walk curr->robust_list (very carefully, it's a userspace list!)
3600 * and mark any locks found there dead, and notify any waiters.
3602 * We silently return on any sign of list-walking problem.
3604 static void exit_robust_list(struct task_struct
*curr
)
3606 struct robust_list_head __user
*head
= curr
->robust_list
;
3607 struct robust_list __user
*entry
, *next_entry
, *pending
;
3608 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3609 unsigned int uninitialized_var(next_pi
);
3610 unsigned long futex_offset
;
3613 if (!futex_cmpxchg_enabled
)
3617 * Fetch the list head (which was registered earlier, via
3618 * sys_set_robust_list()):
3620 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3623 * Fetch the relative futex offset:
3625 if (get_user(futex_offset
, &head
->futex_offset
))
3628 * Fetch any possibly pending lock-add first, and handle it
3631 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3634 next_entry
= NULL
; /* avoid warning with gcc */
3635 while (entry
!= &head
->list
) {
3637 * Fetch the next entry in the list before calling
3638 * handle_futex_death:
3640 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3642 * A pending lock might already be on the list, so
3643 * don't process it twice:
3645 if (entry
!= pending
) {
3646 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3647 curr
, pi
, HANDLE_DEATH_LIST
))
3655 * Avoid excessively long or circular lists:
3664 handle_futex_death((void __user
*)pending
+ futex_offset
,
3665 curr
, pip
, HANDLE_DEATH_PENDING
);
3669 static void futex_cleanup(struct task_struct
*tsk
)
3671 if (unlikely(tsk
->robust_list
)) {
3672 exit_robust_list(tsk
);
3673 tsk
->robust_list
= NULL
;
3676 #ifdef CONFIG_COMPAT
3677 if (unlikely(tsk
->compat_robust_list
)) {
3678 compat_exit_robust_list(tsk
);
3679 tsk
->compat_robust_list
= NULL
;
3683 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3684 exit_pi_state_list(tsk
);
3688 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3689 * @tsk: task to set the state on
3691 * Set the futex exit state of the task lockless. The futex waiter code
3692 * observes that state when a task is exiting and loops until the task has
3693 * actually finished the futex cleanup. The worst case for this is that the
3694 * waiter runs through the wait loop until the state becomes visible.
3696 * This is called from the recursive fault handling path in do_exit().
3698 * This is best effort. Either the futex exit code has run already or
3699 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3700 * take it over. If not, the problem is pushed back to user space. If the
3701 * futex exit code did not run yet, then an already queued waiter might
3702 * block forever, but there is nothing which can be done about that.
3704 void futex_exit_recursive(struct task_struct
*tsk
)
3706 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3707 if (tsk
->futex_state
== FUTEX_STATE_EXITING
)
3708 mutex_unlock(&tsk
->futex_exit_mutex
);
3709 tsk
->futex_state
= FUTEX_STATE_DEAD
;
3712 static void futex_cleanup_begin(struct task_struct
*tsk
)
3715 * Prevent various race issues against a concurrent incoming waiter
3716 * including live locks by forcing the waiter to block on
3717 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3718 * attach_to_pi_owner().
3720 mutex_lock(&tsk
->futex_exit_mutex
);
3723 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3725 * This ensures that all subsequent checks of tsk->futex_state in
3726 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3727 * tsk->pi_lock held.
3729 * It guarantees also that a pi_state which was queued right before
3730 * the state change under tsk->pi_lock by a concurrent waiter must
3731 * be observed in exit_pi_state_list().
3733 raw_spin_lock_irq(&tsk
->pi_lock
);
3734 tsk
->futex_state
= FUTEX_STATE_EXITING
;
3735 raw_spin_unlock_irq(&tsk
->pi_lock
);
3738 static void futex_cleanup_end(struct task_struct
*tsk
, int state
)
3741 * Lockless store. The only side effect is that an observer might
3742 * take another loop until it becomes visible.
3744 tsk
->futex_state
= state
;
3746 * Drop the exit protection. This unblocks waiters which observed
3747 * FUTEX_STATE_EXITING to reevaluate the state.
3749 mutex_unlock(&tsk
->futex_exit_mutex
);
3752 void futex_exec_release(struct task_struct
*tsk
)
3755 * The state handling is done for consistency, but in the case of
3756 * exec() there is no way to prevent futher damage as the PID stays
3757 * the same. But for the unlikely and arguably buggy case that a
3758 * futex is held on exec(), this provides at least as much state
3759 * consistency protection which is possible.
3761 futex_cleanup_begin(tsk
);
3764 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3765 * exec a new binary.
3767 futex_cleanup_end(tsk
, FUTEX_STATE_OK
);
3770 void futex_exit_release(struct task_struct
*tsk
)
3772 futex_cleanup_begin(tsk
);
3774 futex_cleanup_end(tsk
, FUTEX_STATE_DEAD
);
3777 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3778 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3780 int cmd
= op
& FUTEX_CMD_MASK
;
3781 unsigned int flags
= 0;
3783 if (!(op
& FUTEX_PRIVATE_FLAG
))
3784 flags
|= FLAGS_SHARED
;
3786 if (op
& FUTEX_CLOCK_REALTIME
) {
3787 flags
|= FLAGS_CLOCKRT
;
3788 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3789 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3795 case FUTEX_UNLOCK_PI
:
3796 case FUTEX_TRYLOCK_PI
:
3797 case FUTEX_WAIT_REQUEUE_PI
:
3798 case FUTEX_CMP_REQUEUE_PI
:
3799 if (!futex_cmpxchg_enabled
)
3805 val3
= FUTEX_BITSET_MATCH_ANY
;
3807 case FUTEX_WAIT_BITSET
:
3808 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3810 val3
= FUTEX_BITSET_MATCH_ANY
;
3812 case FUTEX_WAKE_BITSET
:
3813 return futex_wake(uaddr
, flags
, val
, val3
);
3815 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3816 case FUTEX_CMP_REQUEUE
:
3817 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3819 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3821 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3822 case FUTEX_UNLOCK_PI
:
3823 return futex_unlock_pi(uaddr
, flags
);
3824 case FUTEX_TRYLOCK_PI
:
3825 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3826 case FUTEX_WAIT_REQUEUE_PI
:
3827 val3
= FUTEX_BITSET_MATCH_ANY
;
3828 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3830 case FUTEX_CMP_REQUEUE_PI
:
3831 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3837 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3838 struct __kernel_timespec __user
*, utime
, u32 __user
*, uaddr2
,
3841 struct timespec64 ts
;
3842 ktime_t t
, *tp
= NULL
;
3844 int cmd
= op
& FUTEX_CMD_MASK
;
3846 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3847 cmd
== FUTEX_WAIT_BITSET
||
3848 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3849 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3851 if (get_timespec64(&ts
, utime
))
3853 if (!timespec64_valid(&ts
))
3856 t
= timespec64_to_ktime(ts
);
3857 if (cmd
== FUTEX_WAIT
)
3858 t
= ktime_add_safe(ktime_get(), t
);
3862 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3863 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3865 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3866 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3867 val2
= (u32
) (unsigned long) utime
;
3869 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3872 #ifdef CONFIG_COMPAT
3874 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3877 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3878 compat_uptr_t __user
*head
, unsigned int *pi
)
3880 if (get_user(*uentry
, head
))
3883 *entry
= compat_ptr((*uentry
) & ~1);
3884 *pi
= (unsigned int)(*uentry
) & 1;
3889 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3890 compat_long_t futex_offset
)
3892 compat_uptr_t base
= ptr_to_compat(entry
);
3893 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3899 * Walk curr->robust_list (very carefully, it's a userspace list!)
3900 * and mark any locks found there dead, and notify any waiters.
3902 * We silently return on any sign of list-walking problem.
3904 static void compat_exit_robust_list(struct task_struct
*curr
)
3906 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3907 struct robust_list __user
*entry
, *next_entry
, *pending
;
3908 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3909 unsigned int uninitialized_var(next_pi
);
3910 compat_uptr_t uentry
, next_uentry
, upending
;
3911 compat_long_t futex_offset
;
3914 if (!futex_cmpxchg_enabled
)
3918 * Fetch the list head (which was registered earlier, via
3919 * sys_set_robust_list()):
3921 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3924 * Fetch the relative futex offset:
3926 if (get_user(futex_offset
, &head
->futex_offset
))
3929 * Fetch any possibly pending lock-add first, and handle it
3932 if (compat_fetch_robust_entry(&upending
, &pending
,
3933 &head
->list_op_pending
, &pip
))
3936 next_entry
= NULL
; /* avoid warning with gcc */
3937 while (entry
!= (struct robust_list __user
*) &head
->list
) {
3939 * Fetch the next entry in the list before calling
3940 * handle_futex_death:
3942 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
3943 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
3945 * A pending lock might already be on the list, so
3946 * dont process it twice:
3948 if (entry
!= pending
) {
3949 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
3951 if (handle_futex_death(uaddr
, curr
, pi
,
3957 uentry
= next_uentry
;
3961 * Avoid excessively long or circular lists:
3969 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
3971 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
3975 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
3976 struct compat_robust_list_head __user
*, head
,
3979 if (!futex_cmpxchg_enabled
)
3982 if (unlikely(len
!= sizeof(*head
)))
3985 current
->compat_robust_list
= head
;
3990 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3991 compat_uptr_t __user
*, head_ptr
,
3992 compat_size_t __user
*, len_ptr
)
3994 struct compat_robust_list_head __user
*head
;
3996 struct task_struct
*p
;
3998 if (!futex_cmpxchg_enabled
)
4007 p
= find_task_by_vpid(pid
);
4013 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
4016 head
= p
->compat_robust_list
;
4019 if (put_user(sizeof(*head
), len_ptr
))
4021 return put_user(ptr_to_compat(head
), head_ptr
);
4028 #endif /* CONFIG_COMPAT */
4030 #ifdef CONFIG_COMPAT_32BIT_TIME
4031 SYSCALL_DEFINE6(futex_time32
, u32 __user
*, uaddr
, int, op
, u32
, val
,
4032 struct old_timespec32 __user
*, utime
, u32 __user
*, uaddr2
,
4035 struct timespec64 ts
;
4036 ktime_t t
, *tp
= NULL
;
4038 int cmd
= op
& FUTEX_CMD_MASK
;
4040 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
4041 cmd
== FUTEX_WAIT_BITSET
||
4042 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
4043 if (get_old_timespec32(&ts
, utime
))
4045 if (!timespec64_valid(&ts
))
4048 t
= timespec64_to_ktime(ts
);
4049 if (cmd
== FUTEX_WAIT
)
4050 t
= ktime_add_safe(ktime_get(), t
);
4053 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
4054 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
4055 val2
= (int) (unsigned long) utime
;
4057 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
4059 #endif /* CONFIG_COMPAT_32BIT_TIME */
4061 static void __init
futex_detect_cmpxchg(void)
4063 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4067 * This will fail and we want it. Some arch implementations do
4068 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4069 * functionality. We want to know that before we call in any
4070 * of the complex code paths. Also we want to prevent
4071 * registration of robust lists in that case. NULL is
4072 * guaranteed to fault and we get -EFAULT on functional
4073 * implementation, the non-functional ones will return
4076 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4077 futex_cmpxchg_enabled
= 1;
4081 static int __init
futex_init(void)
4083 unsigned int futex_shift
;
4086 #if CONFIG_BASE_SMALL
4087 futex_hashsize
= 16;
4089 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4092 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4094 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4096 futex_hashsize
, futex_hashsize
);
4097 futex_hashsize
= 1UL << futex_shift
;
4099 futex_detect_cmpxchg();
4101 for (i
= 0; i
< futex_hashsize
; i
++) {
4102 atomic_set(&futex_queues
[i
].waiters
, 0);
4103 plist_head_init(&futex_queues
[i
].chain
);
4104 spin_lock_init(&futex_queues
[i
].lock
);
4109 core_initcall(futex_init
);