2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/wake_q.h>
65 #include <linux/sched/mm.h>
66 #include <linux/hugetlb.h>
67 #include <linux/freezer.h>
68 #include <linux/bootmem.h>
69 #include <linux/fault-inject.h>
71 #include <asm/futex.h>
73 #include "locking/rtmutex_common.h"
76 * READ this before attempting to hack on futexes!
78 * Basic futex operation and ordering guarantees
79 * =============================================
81 * The waiter reads the futex value in user space and calls
82 * futex_wait(). This function computes the hash bucket and acquires
83 * the hash bucket lock. After that it reads the futex user space value
84 * again and verifies that the data has not changed. If it has not changed
85 * it enqueues itself into the hash bucket, releases the hash bucket lock
88 * The waker side modifies the user space value of the futex and calls
89 * futex_wake(). This function computes the hash bucket and acquires the
90 * hash bucket lock. Then it looks for waiters on that futex in the hash
91 * bucket and wakes them.
93 * In futex wake up scenarios where no tasks are blocked on a futex, taking
94 * the hb spinlock can be avoided and simply return. In order for this
95 * optimization to work, ordering guarantees must exist so that the waiter
96 * being added to the list is acknowledged when the list is concurrently being
97 * checked by the waker, avoiding scenarios like the following:
101 * sys_futex(WAIT, futex, val);
102 * futex_wait(futex, val);
105 * sys_futex(WAKE, futex);
110 * lock(hash_bucket(futex));
112 * unlock(hash_bucket(futex));
115 * This would cause the waiter on CPU 0 to wait forever because it
116 * missed the transition of the user space value from val to newval
117 * and the waker did not find the waiter in the hash bucket queue.
119 * The correct serialization ensures that a waiter either observes
120 * the changed user space value before blocking or is woken by a
125 * sys_futex(WAIT, futex, val);
126 * futex_wait(futex, val);
129 * smp_mb(); (A) <-- paired with -.
131 * lock(hash_bucket(futex)); |
135 * | sys_futex(WAKE, futex);
136 * | futex_wake(futex);
138 * `--------> smp_mb(); (B)
141 * unlock(hash_bucket(futex));
142 * schedule(); if (waiters)
143 * lock(hash_bucket(futex));
144 * else wake_waiters(futex);
145 * waiters--; (b) unlock(hash_bucket(futex));
147 * Where (A) orders the waiters increment and the futex value read through
148 * atomic operations (see hb_waiters_inc) and where (B) orders the write
149 * to futex and the waiters read -- this is done by the barriers for both
150 * shared and private futexes in get_futex_key_refs().
152 * This yields the following case (where X:=waiters, Y:=futex):
160 * Which guarantees that x==0 && y==0 is impossible; which translates back into
161 * the guarantee that we cannot both miss the futex variable change and the
164 * Note that a new waiter is accounted for in (a) even when it is possible that
165 * the wait call can return error, in which case we backtrack from it in (b).
166 * Refer to the comment in queue_lock().
168 * Similarly, in order to account for waiters being requeued on another
169 * address we always increment the waiters for the destination bucket before
170 * acquiring the lock. It then decrements them again after releasing it -
171 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172 * will do the additional required waiter count housekeeping. This is done for
173 * double_lock_hb() and double_unlock_hb(), respectively.
176 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
177 int __read_mostly futex_cmpxchg_enabled
;
181 * Futex flags used to encode options to functions and preserve them across
185 # define FLAGS_SHARED 0x01
188 * NOMMU does not have per process address space. Let the compiler optimize
191 # define FLAGS_SHARED 0x00
193 #define FLAGS_CLOCKRT 0x02
194 #define FLAGS_HAS_TIMEOUT 0x04
197 * Priority Inheritance state:
199 struct futex_pi_state
{
201 * list of 'owned' pi_state instances - these have to be
202 * cleaned up in do_exit() if the task exits prematurely:
204 struct list_head list
;
209 struct rt_mutex pi_mutex
;
211 struct task_struct
*owner
;
215 } __randomize_layout
;
218 * struct futex_q - The hashed futex queue entry, one per waiting task
219 * @list: priority-sorted list of tasks waiting on this futex
220 * @task: the task waiting on the futex
221 * @lock_ptr: the hash bucket lock
222 * @key: the key the futex is hashed on
223 * @pi_state: optional priority inheritance state
224 * @rt_waiter: rt_waiter storage for use with requeue_pi
225 * @requeue_pi_key: the requeue_pi target futex key
226 * @bitset: bitset for the optional bitmasked wakeup
228 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
229 * we can wake only the relevant ones (hashed queues may be shared).
231 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
232 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
233 * The order of wakeup is always to make the first condition true, then
236 * PI futexes are typically woken before they are removed from the hash list via
237 * the rt_mutex code. See unqueue_me_pi().
240 struct plist_node list
;
242 struct task_struct
*task
;
243 spinlock_t
*lock_ptr
;
245 struct futex_pi_state
*pi_state
;
246 struct rt_mutex_waiter
*rt_waiter
;
247 union futex_key
*requeue_pi_key
;
249 } __randomize_layout
;
251 static const struct futex_q futex_q_init
= {
252 /* list gets initialized in queue_me()*/
253 .key
= FUTEX_KEY_INIT
,
254 .bitset
= FUTEX_BITSET_MATCH_ANY
258 * Hash buckets are shared by all the futex_keys that hash to the same
259 * location. Each key may have multiple futex_q structures, one for each task
260 * waiting on a futex.
262 struct futex_hash_bucket
{
265 struct plist_head chain
;
266 } ____cacheline_aligned_in_smp
;
269 * The base of the bucket array and its size are always used together
270 * (after initialization only in hash_futex()), so ensure that they
271 * reside in the same cacheline.
274 struct futex_hash_bucket
*queues
;
275 unsigned long hashsize
;
276 } __futex_data __read_mostly
__aligned(2*sizeof(long));
277 #define futex_queues (__futex_data.queues)
278 #define futex_hashsize (__futex_data.hashsize)
282 * Fault injections for futexes.
284 #ifdef CONFIG_FAIL_FUTEX
287 struct fault_attr attr
;
291 .attr
= FAULT_ATTR_INITIALIZER
,
292 .ignore_private
= false,
295 static int __init
setup_fail_futex(char *str
)
297 return setup_fault_attr(&fail_futex
.attr
, str
);
299 __setup("fail_futex=", setup_fail_futex
);
301 static bool should_fail_futex(bool fshared
)
303 if (fail_futex
.ignore_private
&& !fshared
)
306 return should_fail(&fail_futex
.attr
, 1);
309 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
311 static int __init
fail_futex_debugfs(void)
313 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
316 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
321 if (!debugfs_create_bool("ignore-private", mode
, dir
,
322 &fail_futex
.ignore_private
)) {
323 debugfs_remove_recursive(dir
);
330 late_initcall(fail_futex_debugfs
);
332 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
335 static inline bool should_fail_futex(bool fshared
)
339 #endif /* CONFIG_FAIL_FUTEX */
341 static inline void futex_get_mm(union futex_key
*key
)
343 mmgrab(key
->private.mm
);
345 * Ensure futex_get_mm() implies a full barrier such that
346 * get_futex_key() implies a full barrier. This is relied upon
347 * as smp_mb(); (B), see the ordering comment above.
349 smp_mb__after_atomic();
353 * Reflects a new waiter being added to the waitqueue.
355 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
358 atomic_inc(&hb
->waiters
);
360 * Full barrier (A), see the ordering comment above.
362 smp_mb__after_atomic();
367 * Reflects a waiter being removed from the waitqueue by wakeup
370 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
373 atomic_dec(&hb
->waiters
);
377 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
380 return atomic_read(&hb
->waiters
);
387 * hash_futex - Return the hash bucket in the global hash
388 * @key: Pointer to the futex key for which the hash is calculated
390 * We hash on the keys returned from get_futex_key (see below) and return the
391 * corresponding hash bucket in the global hash.
393 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
395 u32 hash
= jhash2((u32
*)&key
->both
.word
,
396 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
398 return &futex_queues
[hash
& (futex_hashsize
- 1)];
403 * match_futex - Check whether two futex keys are equal
404 * @key1: Pointer to key1
405 * @key2: Pointer to key2
407 * Return 1 if two futex_keys are equal, 0 otherwise.
409 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
412 && key1
->both
.word
== key2
->both
.word
413 && key1
->both
.ptr
== key2
->both
.ptr
414 && key1
->both
.offset
== key2
->both
.offset
);
418 * Take a reference to the resource addressed by a key.
419 * Can be called while holding spinlocks.
422 static void get_futex_key_refs(union futex_key
*key
)
428 * On MMU less systems futexes are always "private" as there is no per
429 * process address space. We need the smp wmb nevertheless - yes,
430 * arch/blackfin has MMU less SMP ...
432 if (!IS_ENABLED(CONFIG_MMU
)) {
433 smp_mb(); /* explicit smp_mb(); (B) */
437 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
439 smp_mb(); /* explicit smp_mb(); (B) */
441 case FUT_OFF_MMSHARED
:
442 futex_get_mm(key
); /* implies smp_mb(); (B) */
446 * Private futexes do not hold reference on an inode or
447 * mm, therefore the only purpose of calling get_futex_key_refs
448 * is because we need the barrier for the lockless waiter check.
450 smp_mb(); /* explicit smp_mb(); (B) */
455 * Drop a reference to the resource addressed by a key.
456 * The hash bucket spinlock must not be held. This is
457 * a no-op for private futexes, see comment in the get
460 static void drop_futex_key_refs(union futex_key
*key
)
462 if (!key
->both
.ptr
) {
463 /* If we're here then we tried to put a key we failed to get */
468 if (!IS_ENABLED(CONFIG_MMU
))
471 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
474 case FUT_OFF_MMSHARED
:
475 mmdrop(key
->private.mm
);
481 * Generate a machine wide unique identifier for this inode.
483 * This relies on u64 not wrapping in the life-time of the machine; which with
484 * 1ns resolution means almost 585 years.
486 * This further relies on the fact that a well formed program will not unmap
487 * the file while it has a (shared) futex waiting on it. This mapping will have
488 * a file reference which pins the mount and inode.
490 * If for some reason an inode gets evicted and read back in again, it will get
491 * a new sequence number and will _NOT_ match, even though it is the exact same
494 * It is important that match_futex() will never have a false-positive, esp.
495 * for PI futexes that can mess up the state. The above argues that false-negatives
496 * are only possible for malformed programs.
498 static u64
get_inode_sequence_number(struct inode
*inode
)
500 static atomic64_t i_seq
;
503 /* Does the inode already have a sequence number? */
504 old
= atomic64_read(&inode
->i_sequence
);
509 u64
new = atomic64_add_return(1, &i_seq
);
510 if (WARN_ON_ONCE(!new))
513 old
= atomic64_cmpxchg_relaxed(&inode
->i_sequence
, 0, new);
521 * get_futex_key() - Get parameters which are the keys for a futex
522 * @uaddr: virtual address of the futex
523 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
524 * @key: address where result is stored.
525 * @rw: mapping needs to be read/write (values: VERIFY_READ,
528 * Return: a negative error code or 0
530 * The key words are stored in @key on success.
532 * For shared mappings (when @fshared), the key is:
533 * ( inode->i_sequence, page->index, offset_within_page )
534 * [ also see get_inode_sequence_number() ]
536 * For private mappings (or when !@fshared), the key is:
537 * ( current->mm, address, 0 )
539 * This allows (cross process, where applicable) identification of the futex
540 * without keeping the page pinned for the duration of the FUTEX_WAIT.
542 * lock_page() might sleep, the caller should not hold a spinlock.
545 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
547 unsigned long address
= (unsigned long)uaddr
;
548 struct mm_struct
*mm
= current
->mm
;
549 struct page
*page
, *tail
;
550 struct address_space
*mapping
;
554 * The futex address must be "naturally" aligned.
556 key
->both
.offset
= address
% PAGE_SIZE
;
557 if (unlikely((address
% sizeof(u32
)) != 0))
559 address
-= key
->both
.offset
;
561 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
564 if (unlikely(should_fail_futex(fshared
)))
568 * PROCESS_PRIVATE futexes are fast.
569 * As the mm cannot disappear under us and the 'key' only needs
570 * virtual address, we dont even have to find the underlying vma.
571 * Note : We do have to check 'uaddr' is a valid user address,
572 * but access_ok() should be faster than find_vma()
575 key
->private.mm
= mm
;
576 key
->private.address
= address
;
577 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
582 /* Ignore any VERIFY_READ mapping (futex common case) */
583 if (unlikely(should_fail_futex(fshared
)))
586 err
= get_user_pages_fast(address
, 1, 1, &page
);
588 * If write access is not required (eg. FUTEX_WAIT), try
589 * and get read-only access.
591 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
592 err
= get_user_pages_fast(address
, 1, 0, &page
);
601 * The treatment of mapping from this point on is critical. The page
602 * lock protects many things but in this context the page lock
603 * stabilizes mapping, prevents inode freeing in the shared
604 * file-backed region case and guards against movement to swap cache.
606 * Strictly speaking the page lock is not needed in all cases being
607 * considered here and page lock forces unnecessarily serialization
608 * From this point on, mapping will be re-verified if necessary and
609 * page lock will be acquired only if it is unavoidable
611 * Mapping checks require the head page for any compound page so the
612 * head page and mapping is looked up now. For anonymous pages, it
613 * does not matter if the page splits in the future as the key is
614 * based on the address. For filesystem-backed pages, the tail is
615 * required as the index of the page determines the key. For
616 * base pages, there is no tail page and tail == page.
619 page
= compound_head(page
);
620 mapping
= READ_ONCE(page
->mapping
);
623 * If page->mapping is NULL, then it cannot be a PageAnon
624 * page; but it might be the ZERO_PAGE or in the gate area or
625 * in a special mapping (all cases which we are happy to fail);
626 * or it may have been a good file page when get_user_pages_fast
627 * found it, but truncated or holepunched or subjected to
628 * invalidate_complete_page2 before we got the page lock (also
629 * cases which we are happy to fail). And we hold a reference,
630 * so refcount care in invalidate_complete_page's remove_mapping
631 * prevents drop_caches from setting mapping to NULL beneath us.
633 * The case we do have to guard against is when memory pressure made
634 * shmem_writepage move it from filecache to swapcache beneath us:
635 * an unlikely race, but we do need to retry for page->mapping.
637 if (unlikely(!mapping
)) {
641 * Page lock is required to identify which special case above
642 * applies. If this is really a shmem page then the page lock
643 * will prevent unexpected transitions.
646 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
657 * Private mappings are handled in a simple way.
659 * If the futex key is stored on an anonymous page, then the associated
660 * object is the mm which is implicitly pinned by the calling process.
662 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
663 * it's a read-only handle, it's expected that futexes attach to
664 * the object not the particular process.
666 if (PageAnon(page
)) {
668 * A RO anonymous page will never change and thus doesn't make
669 * sense for futex operations.
671 if (unlikely(should_fail_futex(fshared
)) || ro
) {
676 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
677 key
->private.mm
= mm
;
678 key
->private.address
= address
;
684 * The associated futex object in this case is the inode and
685 * the page->mapping must be traversed. Ordinarily this should
686 * be stabilised under page lock but it's not strictly
687 * necessary in this case as we just want to pin the inode, not
688 * update the radix tree or anything like that.
690 * The RCU read lock is taken as the inode is finally freed
691 * under RCU. If the mapping still matches expectations then the
692 * mapping->host can be safely accessed as being a valid inode.
696 if (READ_ONCE(page
->mapping
) != mapping
) {
703 inode
= READ_ONCE(mapping
->host
);
711 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
712 key
->shared
.i_seq
= get_inode_sequence_number(inode
);
713 key
->shared
.pgoff
= basepage_index(tail
);
717 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
724 static inline void put_futex_key(union futex_key
*key
)
726 drop_futex_key_refs(key
);
730 * fault_in_user_writeable() - Fault in user address and verify RW access
731 * @uaddr: pointer to faulting user space address
733 * Slow path to fixup the fault we just took in the atomic write
736 * We have no generic implementation of a non-destructive write to the
737 * user address. We know that we faulted in the atomic pagefault
738 * disabled section so we can as well avoid the #PF overhead by
739 * calling get_user_pages() right away.
741 static int fault_in_user_writeable(u32 __user
*uaddr
)
743 struct mm_struct
*mm
= current
->mm
;
746 down_read(&mm
->mmap_sem
);
747 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
748 FAULT_FLAG_WRITE
, NULL
);
749 up_read(&mm
->mmap_sem
);
751 return ret
< 0 ? ret
: 0;
755 * futex_top_waiter() - Return the highest priority waiter on a futex
756 * @hb: the hash bucket the futex_q's reside in
757 * @key: the futex key (to distinguish it from other futex futex_q's)
759 * Must be called with the hb lock held.
761 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
762 union futex_key
*key
)
764 struct futex_q
*this;
766 plist_for_each_entry(this, &hb
->chain
, list
) {
767 if (match_futex(&this->key
, key
))
773 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
774 u32 uval
, u32 newval
)
779 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
785 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
790 ret
= __get_user(*dest
, from
);
793 return ret
? -EFAULT
: 0;
800 static int refill_pi_state_cache(void)
802 struct futex_pi_state
*pi_state
;
804 if (likely(current
->pi_state_cache
))
807 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
812 INIT_LIST_HEAD(&pi_state
->list
);
813 /* pi_mutex gets initialized later */
814 pi_state
->owner
= NULL
;
815 atomic_set(&pi_state
->refcount
, 1);
816 pi_state
->key
= FUTEX_KEY_INIT
;
818 current
->pi_state_cache
= pi_state
;
823 static struct futex_pi_state
*alloc_pi_state(void)
825 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
828 current
->pi_state_cache
= NULL
;
833 static void pi_state_update_owner(struct futex_pi_state
*pi_state
,
834 struct task_struct
*new_owner
)
836 struct task_struct
*old_owner
= pi_state
->owner
;
838 lockdep_assert_held(&pi_state
->pi_mutex
.wait_lock
);
841 raw_spin_lock(&old_owner
->pi_lock
);
842 WARN_ON(list_empty(&pi_state
->list
));
843 list_del_init(&pi_state
->list
);
844 raw_spin_unlock(&old_owner
->pi_lock
);
848 raw_spin_lock(&new_owner
->pi_lock
);
849 WARN_ON(!list_empty(&pi_state
->list
));
850 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
851 pi_state
->owner
= new_owner
;
852 raw_spin_unlock(&new_owner
->pi_lock
);
856 static void get_pi_state(struct futex_pi_state
*pi_state
)
858 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state
->refcount
));
862 * Drops a reference to the pi_state object and frees or caches it
863 * when the last reference is gone.
865 static void put_pi_state(struct futex_pi_state
*pi_state
)
870 if (!atomic_dec_and_test(&pi_state
->refcount
))
874 * If pi_state->owner is NULL, the owner is most probably dying
875 * and has cleaned up the pi_state already
877 if (pi_state
->owner
) {
878 struct task_struct
*owner
;
880 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
881 owner
= pi_state
->owner
;
883 raw_spin_lock(&owner
->pi_lock
);
884 list_del_init(&pi_state
->list
);
885 raw_spin_unlock(&owner
->pi_lock
);
887 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
888 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
891 if (current
->pi_state_cache
) {
895 * pi_state->list is already empty.
896 * clear pi_state->owner.
897 * refcount is at 0 - put it back to 1.
899 pi_state
->owner
= NULL
;
900 atomic_set(&pi_state
->refcount
, 1);
901 current
->pi_state_cache
= pi_state
;
906 * Look up the task based on what TID userspace gave us.
909 static struct task_struct
*futex_find_get_task(pid_t pid
)
911 struct task_struct
*p
;
914 p
= find_task_by_vpid(pid
);
923 #ifdef CONFIG_FUTEX_PI
926 * This task is holding PI mutexes at exit time => bad.
927 * Kernel cleans up PI-state, but userspace is likely hosed.
928 * (Robust-futex cleanup is separate and might save the day for userspace.)
930 void exit_pi_state_list(struct task_struct
*curr
)
932 struct list_head
*next
, *head
= &curr
->pi_state_list
;
933 struct futex_pi_state
*pi_state
;
934 struct futex_hash_bucket
*hb
;
935 union futex_key key
= FUTEX_KEY_INIT
;
937 if (!futex_cmpxchg_enabled
)
940 * We are a ZOMBIE and nobody can enqueue itself on
941 * pi_state_list anymore, but we have to be careful
942 * versus waiters unqueueing themselves:
944 raw_spin_lock_irq(&curr
->pi_lock
);
945 while (!list_empty(head
)) {
947 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
949 hb
= hash_futex(&key
);
952 * We can race against put_pi_state() removing itself from the
953 * list (a waiter going away). put_pi_state() will first
954 * decrement the reference count and then modify the list, so
955 * its possible to see the list entry but fail this reference
958 * In that case; drop the locks to let put_pi_state() make
959 * progress and retry the loop.
961 if (!atomic_inc_not_zero(&pi_state
->refcount
)) {
962 raw_spin_unlock_irq(&curr
->pi_lock
);
964 raw_spin_lock_irq(&curr
->pi_lock
);
967 raw_spin_unlock_irq(&curr
->pi_lock
);
969 spin_lock(&hb
->lock
);
970 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
971 raw_spin_lock(&curr
->pi_lock
);
973 * We dropped the pi-lock, so re-check whether this
974 * task still owns the PI-state:
976 if (head
->next
!= next
) {
977 /* retain curr->pi_lock for the loop invariant */
978 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
979 spin_unlock(&hb
->lock
);
980 put_pi_state(pi_state
);
984 WARN_ON(pi_state
->owner
!= curr
);
985 WARN_ON(list_empty(&pi_state
->list
));
986 list_del_init(&pi_state
->list
);
987 pi_state
->owner
= NULL
;
989 raw_spin_unlock(&curr
->pi_lock
);
990 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
991 spin_unlock(&hb
->lock
);
993 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
994 put_pi_state(pi_state
);
996 raw_spin_lock_irq(&curr
->pi_lock
);
998 raw_spin_unlock_irq(&curr
->pi_lock
);
1004 * We need to check the following states:
1006 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
1008 * [1] NULL | --- | --- | 0 | 0/1 | Valid
1009 * [2] NULL | --- | --- | >0 | 0/1 | Valid
1011 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1013 * [4] Found | Found | NULL | 0 | 1 | Valid
1014 * [5] Found | Found | NULL | >0 | 1 | Invalid
1016 * [6] Found | Found | task | 0 | 1 | Valid
1018 * [7] Found | Found | NULL | Any | 0 | Invalid
1020 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1021 * [9] Found | Found | task | 0 | 0 | Invalid
1022 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1024 * [1] Indicates that the kernel can acquire the futex atomically. We
1025 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1027 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1028 * thread is found then it indicates that the owner TID has died.
1030 * [3] Invalid. The waiter is queued on a non PI futex
1032 * [4] Valid state after exit_robust_list(), which sets the user space
1033 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1035 * [5] The user space value got manipulated between exit_robust_list()
1036 * and exit_pi_state_list()
1038 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1039 * the pi_state but cannot access the user space value.
1041 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1043 * [8] Owner and user space value match
1045 * [9] There is no transient state which sets the user space TID to 0
1046 * except exit_robust_list(), but this is indicated by the
1047 * FUTEX_OWNER_DIED bit. See [4]
1049 * [10] There is no transient state which leaves owner and user space
1053 * Serialization and lifetime rules:
1057 * hb -> futex_q, relation
1058 * futex_q -> pi_state, relation
1060 * (cannot be raw because hb can contain arbitrary amount
1063 * pi_mutex->wait_lock:
1067 * (and pi_mutex 'obviously')
1071 * p->pi_state_list -> pi_state->list, relation
1073 * pi_state->refcount:
1081 * pi_mutex->wait_lock
1087 * Validate that the existing waiter has a pi_state and sanity check
1088 * the pi_state against the user space value. If correct, attach to
1091 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1092 struct futex_pi_state
*pi_state
,
1093 struct futex_pi_state
**ps
)
1095 pid_t pid
= uval
& FUTEX_TID_MASK
;
1100 * Userspace might have messed up non-PI and PI futexes [3]
1102 if (unlikely(!pi_state
))
1106 * We get here with hb->lock held, and having found a
1107 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1108 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1109 * which in turn means that futex_lock_pi() still has a reference on
1112 * The waiter holding a reference on @pi_state also protects against
1113 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1114 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1115 * free pi_state before we can take a reference ourselves.
1117 WARN_ON(!atomic_read(&pi_state
->refcount
));
1120 * Now that we have a pi_state, we can acquire wait_lock
1121 * and do the state validation.
1123 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1126 * Since {uval, pi_state} is serialized by wait_lock, and our current
1127 * uval was read without holding it, it can have changed. Verify it
1128 * still is what we expect it to be, otherwise retry the entire
1131 if (get_futex_value_locked(&uval2
, uaddr
))
1138 * Handle the owner died case:
1140 if (uval
& FUTEX_OWNER_DIED
) {
1142 * exit_pi_state_list sets owner to NULL and wakes the
1143 * topmost waiter. The task which acquires the
1144 * pi_state->rt_mutex will fixup owner.
1146 if (!pi_state
->owner
) {
1148 * No pi state owner, but the user space TID
1149 * is not 0. Inconsistent state. [5]
1154 * Take a ref on the state and return success. [4]
1160 * If TID is 0, then either the dying owner has not
1161 * yet executed exit_pi_state_list() or some waiter
1162 * acquired the rtmutex in the pi state, but did not
1163 * yet fixup the TID in user space.
1165 * Take a ref on the state and return success. [6]
1171 * If the owner died bit is not set, then the pi_state
1172 * must have an owner. [7]
1174 if (!pi_state
->owner
)
1179 * Bail out if user space manipulated the futex value. If pi
1180 * state exists then the owner TID must be the same as the
1181 * user space TID. [9/10]
1183 if (pid
!= task_pid_vnr(pi_state
->owner
))
1187 get_pi_state(pi_state
);
1188 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1205 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1209 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1210 struct task_struct
*tsk
)
1215 * If PF_EXITPIDONE is not yet set, then try again.
1217 if (tsk
&& !(tsk
->flags
& PF_EXITPIDONE
))
1221 * Reread the user space value to handle the following situation:
1225 * sys_exit() sys_futex()
1226 * do_exit() futex_lock_pi()
1227 * futex_lock_pi_atomic()
1228 * exit_signals(tsk) No waiters:
1229 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1230 * mm_release(tsk) Set waiter bit
1231 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1232 * Set owner died attach_to_pi_owner() {
1233 * *uaddr = 0xC0000000; tsk = get_task(PID);
1234 * } if (!tsk->flags & PF_EXITING) {
1236 * tsk->flags |= PF_EXITPIDONE; } else {
1237 * if (!(tsk->flags & PF_EXITPIDONE))
1239 * return -ESRCH; <--- FAIL
1242 * Returning ESRCH unconditionally is wrong here because the
1243 * user space value has been changed by the exiting task.
1245 * The same logic applies to the case where the exiting task is
1248 if (get_futex_value_locked(&uval2
, uaddr
))
1251 /* If the user space value has changed, try again. */
1256 * The exiting task did not have a robust list, the robust list was
1257 * corrupted or the user space value in *uaddr is simply bogus.
1258 * Give up and tell user space.
1264 * Lookup the task for the TID provided from user space and attach to
1265 * it after doing proper sanity checks.
1267 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1268 struct futex_pi_state
**ps
)
1270 pid_t pid
= uval
& FUTEX_TID_MASK
;
1271 struct futex_pi_state
*pi_state
;
1272 struct task_struct
*p
;
1275 * We are the first waiter - try to look up the real owner and attach
1276 * the new pi_state to it, but bail out when TID = 0 [1]
1278 * The !pid check is paranoid. None of the call sites should end up
1279 * with pid == 0, but better safe than sorry. Let the caller retry
1283 p
= futex_find_get_task(pid
);
1285 return handle_exit_race(uaddr
, uval
, NULL
);
1287 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1293 * We need to look at the task state flags to figure out,
1294 * whether the task is exiting. To protect against the do_exit
1295 * change of the task flags, we do this protected by
1298 raw_spin_lock_irq(&p
->pi_lock
);
1299 if (unlikely(p
->flags
& PF_EXITING
)) {
1301 * The task is on the way out. When PF_EXITPIDONE is
1302 * set, we know that the task has finished the
1305 int ret
= handle_exit_race(uaddr
, uval
, p
);
1307 raw_spin_unlock_irq(&p
->pi_lock
);
1313 * No existing pi state. First waiter. [2]
1315 * This creates pi_state, we have hb->lock held, this means nothing can
1316 * observe this state, wait_lock is irrelevant.
1318 pi_state
= alloc_pi_state();
1321 * Initialize the pi_mutex in locked state and make @p
1324 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1326 /* Store the key for possible exit cleanups: */
1327 pi_state
->key
= *key
;
1329 WARN_ON(!list_empty(&pi_state
->list
));
1330 list_add(&pi_state
->list
, &p
->pi_state_list
);
1332 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1333 * because there is no concurrency as the object is not published yet.
1335 pi_state
->owner
= p
;
1336 raw_spin_unlock_irq(&p
->pi_lock
);
1345 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1346 struct futex_hash_bucket
*hb
,
1347 union futex_key
*key
, struct futex_pi_state
**ps
)
1349 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1352 * If there is a waiter on that futex, validate it and
1353 * attach to the pi_state when the validation succeeds.
1356 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1359 * We are the first waiter - try to look up the owner based on
1360 * @uval and attach to it.
1362 return attach_to_pi_owner(uaddr
, uval
, key
, ps
);
1365 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1367 u32
uninitialized_var(curval
);
1369 if (unlikely(should_fail_futex(true)))
1372 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1375 /* If user space value changed, let the caller retry */
1376 return curval
!= uval
? -EAGAIN
: 0;
1380 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1381 * @uaddr: the pi futex user address
1382 * @hb: the pi futex hash bucket
1383 * @key: the futex key associated with uaddr and hb
1384 * @ps: the pi_state pointer where we store the result of the
1386 * @task: the task to perform the atomic lock work for. This will
1387 * be "current" except in the case of requeue pi.
1388 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1391 * - 0 - ready to wait;
1392 * - 1 - acquired the lock;
1395 * The hb->lock and futex_key refs shall be held by the caller.
1397 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1398 union futex_key
*key
,
1399 struct futex_pi_state
**ps
,
1400 struct task_struct
*task
, int set_waiters
)
1402 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1403 struct futex_q
*top_waiter
;
1407 * Read the user space value first so we can validate a few
1408 * things before proceeding further.
1410 if (get_futex_value_locked(&uval
, uaddr
))
1413 if (unlikely(should_fail_futex(true)))
1419 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1422 if ((unlikely(should_fail_futex(true))))
1426 * Lookup existing state first. If it exists, try to attach to
1429 top_waiter
= futex_top_waiter(hb
, key
);
1431 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1434 * No waiter and user TID is 0. We are here because the
1435 * waiters or the owner died bit is set or called from
1436 * requeue_cmp_pi or for whatever reason something took the
1439 if (!(uval
& FUTEX_TID_MASK
)) {
1441 * We take over the futex. No other waiters and the user space
1442 * TID is 0. We preserve the owner died bit.
1444 newval
= uval
& FUTEX_OWNER_DIED
;
1447 /* The futex requeue_pi code can enforce the waiters bit */
1449 newval
|= FUTEX_WAITERS
;
1451 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1452 /* If the take over worked, return 1 */
1453 return ret
< 0 ? ret
: 1;
1457 * First waiter. Set the waiters bit before attaching ourself to
1458 * the owner. If owner tries to unlock, it will be forced into
1459 * the kernel and blocked on hb->lock.
1461 newval
= uval
| FUTEX_WAITERS
;
1462 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1466 * If the update of the user space value succeeded, we try to
1467 * attach to the owner. If that fails, no harm done, we only
1468 * set the FUTEX_WAITERS bit in the user space variable.
1470 return attach_to_pi_owner(uaddr
, newval
, key
, ps
);
1474 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1475 * @q: The futex_q to unqueue
1477 * The q->lock_ptr must not be NULL and must be held by the caller.
1479 static void __unqueue_futex(struct futex_q
*q
)
1481 struct futex_hash_bucket
*hb
;
1483 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1484 || WARN_ON(plist_node_empty(&q
->list
)))
1487 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1488 plist_del(&q
->list
, &hb
->chain
);
1493 * The hash bucket lock must be held when this is called.
1494 * Afterwards, the futex_q must not be accessed. Callers
1495 * must ensure to later call wake_up_q() for the actual
1498 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1500 struct task_struct
*p
= q
->task
;
1502 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1508 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1509 * is written, without taking any locks. This is possible in the event
1510 * of a spurious wakeup, for example. A memory barrier is required here
1511 * to prevent the following store to lock_ptr from getting ahead of the
1512 * plist_del in __unqueue_futex().
1514 smp_store_release(&q
->lock_ptr
, NULL
);
1517 * Queue the task for later wakeup for after we've released
1518 * the hb->lock. wake_q_add() grabs reference to p.
1520 wake_q_add(wake_q
, p
);
1525 * Caller must hold a reference on @pi_state.
1527 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1529 u32
uninitialized_var(curval
), newval
;
1530 struct task_struct
*new_owner
;
1531 bool postunlock
= false;
1532 DEFINE_WAKE_Q(wake_q
);
1535 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1536 if (WARN_ON_ONCE(!new_owner
)) {
1538 * As per the comment in futex_unlock_pi() this should not happen.
1540 * When this happens, give up our locks and try again, giving
1541 * the futex_lock_pi() instance time to complete, either by
1542 * waiting on the rtmutex or removing itself from the futex
1550 * We pass it to the next owner. The WAITERS bit is always kept
1551 * enabled while there is PI state around. We cleanup the owner
1552 * died bit, because we are the owner.
1554 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1556 if (unlikely(should_fail_futex(true)))
1559 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)) {
1562 } else if (curval
!= uval
) {
1564 * If a unconditional UNLOCK_PI operation (user space did not
1565 * try the TID->0 transition) raced with a waiter setting the
1566 * FUTEX_WAITERS flag between get_user() and locking the hash
1567 * bucket lock, retry the operation.
1569 if ((FUTEX_TID_MASK
& curval
) == uval
)
1577 * This is a point of no return; once we modified the uval
1578 * there is no going back and subsequent operations must
1581 pi_state_update_owner(pi_state
, new_owner
);
1582 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1586 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1589 rt_mutex_postunlock(&wake_q
);
1595 * Express the locking dependencies for lockdep:
1598 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1601 spin_lock(&hb1
->lock
);
1603 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1604 } else { /* hb1 > hb2 */
1605 spin_lock(&hb2
->lock
);
1606 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1611 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1613 spin_unlock(&hb1
->lock
);
1615 spin_unlock(&hb2
->lock
);
1619 * Wake up waiters matching bitset queued on this futex (uaddr).
1622 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1624 struct futex_hash_bucket
*hb
;
1625 struct futex_q
*this, *next
;
1626 union futex_key key
= FUTEX_KEY_INIT
;
1628 DEFINE_WAKE_Q(wake_q
);
1633 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1634 if (unlikely(ret
!= 0))
1637 hb
= hash_futex(&key
);
1639 /* Make sure we really have tasks to wakeup */
1640 if (!hb_waiters_pending(hb
))
1643 spin_lock(&hb
->lock
);
1645 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1646 if (match_futex (&this->key
, &key
)) {
1647 if (this->pi_state
|| this->rt_waiter
) {
1652 /* Check if one of the bits is set in both bitsets */
1653 if (!(this->bitset
& bitset
))
1656 mark_wake_futex(&wake_q
, this);
1657 if (++ret
>= nr_wake
)
1662 spin_unlock(&hb
->lock
);
1665 put_futex_key(&key
);
1670 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1672 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1673 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1674 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 12);
1675 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 12);
1678 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1679 if (oparg
< 0 || oparg
> 31) {
1680 char comm
[sizeof(current
->comm
)];
1682 * kill this print and return -EINVAL when userspace
1685 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1686 get_task_comm(comm
, current
), oparg
);
1692 if (!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
)))
1695 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1700 case FUTEX_OP_CMP_EQ
:
1701 return oldval
== cmparg
;
1702 case FUTEX_OP_CMP_NE
:
1703 return oldval
!= cmparg
;
1704 case FUTEX_OP_CMP_LT
:
1705 return oldval
< cmparg
;
1706 case FUTEX_OP_CMP_GE
:
1707 return oldval
>= cmparg
;
1708 case FUTEX_OP_CMP_LE
:
1709 return oldval
<= cmparg
;
1710 case FUTEX_OP_CMP_GT
:
1711 return oldval
> cmparg
;
1718 * Wake up all waiters hashed on the physical page that is mapped
1719 * to this virtual address:
1722 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1723 int nr_wake
, int nr_wake2
, int op
)
1725 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1726 struct futex_hash_bucket
*hb1
, *hb2
;
1727 struct futex_q
*this, *next
;
1729 DEFINE_WAKE_Q(wake_q
);
1732 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1733 if (unlikely(ret
!= 0))
1735 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1736 if (unlikely(ret
!= 0))
1739 hb1
= hash_futex(&key1
);
1740 hb2
= hash_futex(&key2
);
1743 double_lock_hb(hb1
, hb2
);
1744 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1745 if (unlikely(op_ret
< 0)) {
1747 double_unlock_hb(hb1
, hb2
);
1751 * we don't get EFAULT from MMU faults if we don't have an MMU,
1752 * but we might get them from range checking
1758 if (unlikely(op_ret
!= -EFAULT
)) {
1763 ret
= fault_in_user_writeable(uaddr2
);
1767 if (!(flags
& FLAGS_SHARED
))
1770 put_futex_key(&key2
);
1771 put_futex_key(&key1
);
1775 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1776 if (match_futex (&this->key
, &key1
)) {
1777 if (this->pi_state
|| this->rt_waiter
) {
1781 mark_wake_futex(&wake_q
, this);
1782 if (++ret
>= nr_wake
)
1789 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1790 if (match_futex (&this->key
, &key2
)) {
1791 if (this->pi_state
|| this->rt_waiter
) {
1795 mark_wake_futex(&wake_q
, this);
1796 if (++op_ret
>= nr_wake2
)
1804 double_unlock_hb(hb1
, hb2
);
1807 put_futex_key(&key2
);
1809 put_futex_key(&key1
);
1815 * requeue_futex() - Requeue a futex_q from one hb to another
1816 * @q: the futex_q to requeue
1817 * @hb1: the source hash_bucket
1818 * @hb2: the target hash_bucket
1819 * @key2: the new key for the requeued futex_q
1822 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1823 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1827 * If key1 and key2 hash to the same bucket, no need to
1830 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1831 plist_del(&q
->list
, &hb1
->chain
);
1832 hb_waiters_dec(hb1
);
1833 hb_waiters_inc(hb2
);
1834 plist_add(&q
->list
, &hb2
->chain
);
1835 q
->lock_ptr
= &hb2
->lock
;
1837 get_futex_key_refs(key2
);
1842 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1844 * @key: the key of the requeue target futex
1845 * @hb: the hash_bucket of the requeue target futex
1847 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1848 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1849 * to the requeue target futex so the waiter can detect the wakeup on the right
1850 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1851 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1852 * to protect access to the pi_state to fixup the owner later. Must be called
1853 * with both q->lock_ptr and hb->lock held.
1856 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1857 struct futex_hash_bucket
*hb
)
1859 get_futex_key_refs(key
);
1864 WARN_ON(!q
->rt_waiter
);
1865 q
->rt_waiter
= NULL
;
1867 q
->lock_ptr
= &hb
->lock
;
1869 wake_up_state(q
->task
, TASK_NORMAL
);
1873 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1874 * @pifutex: the user address of the to futex
1875 * @hb1: the from futex hash bucket, must be locked by the caller
1876 * @hb2: the to futex hash bucket, must be locked by the caller
1877 * @key1: the from futex key
1878 * @key2: the to futex key
1879 * @ps: address to store the pi_state pointer
1880 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1882 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1883 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1884 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1885 * hb1 and hb2 must be held by the caller.
1888 * - 0 - failed to acquire the lock atomically;
1889 * - >0 - acquired the lock, return value is vpid of the top_waiter
1892 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1893 struct futex_hash_bucket
*hb1
,
1894 struct futex_hash_bucket
*hb2
,
1895 union futex_key
*key1
, union futex_key
*key2
,
1896 struct futex_pi_state
**ps
, int set_waiters
)
1898 struct futex_q
*top_waiter
= NULL
;
1902 if (get_futex_value_locked(&curval
, pifutex
))
1905 if (unlikely(should_fail_futex(true)))
1909 * Find the top_waiter and determine if there are additional waiters.
1910 * If the caller intends to requeue more than 1 waiter to pifutex,
1911 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1912 * as we have means to handle the possible fault. If not, don't set
1913 * the bit unecessarily as it will force the subsequent unlock to enter
1916 top_waiter
= futex_top_waiter(hb1
, key1
);
1918 /* There are no waiters, nothing for us to do. */
1922 /* Ensure we requeue to the expected futex. */
1923 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1927 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1928 * the contended case or if set_waiters is 1. The pi_state is returned
1929 * in ps in contended cases.
1931 vpid
= task_pid_vnr(top_waiter
->task
);
1932 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1935 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1942 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1943 * @uaddr1: source futex user address
1944 * @flags: futex flags (FLAGS_SHARED, etc.)
1945 * @uaddr2: target futex user address
1946 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1947 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1948 * @cmpval: @uaddr1 expected value (or %NULL)
1949 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1950 * pi futex (pi to pi requeue is not supported)
1952 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1953 * uaddr2 atomically on behalf of the top waiter.
1956 * - >=0 - on success, the number of tasks requeued or woken;
1959 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1960 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1961 u32
*cmpval
, int requeue_pi
)
1963 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1964 int drop_count
= 0, task_count
= 0, ret
;
1965 struct futex_pi_state
*pi_state
= NULL
;
1966 struct futex_hash_bucket
*hb1
, *hb2
;
1967 struct futex_q
*this, *next
;
1968 DEFINE_WAKE_Q(wake_q
);
1970 if (nr_wake
< 0 || nr_requeue
< 0)
1974 * When PI not supported: return -ENOSYS if requeue_pi is true,
1975 * consequently the compiler knows requeue_pi is always false past
1976 * this point which will optimize away all the conditional code
1979 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1984 * Requeue PI only works on two distinct uaddrs. This
1985 * check is only valid for private futexes. See below.
1987 if (uaddr1
== uaddr2
)
1991 * requeue_pi requires a pi_state, try to allocate it now
1992 * without any locks in case it fails.
1994 if (refill_pi_state_cache())
1997 * requeue_pi must wake as many tasks as it can, up to nr_wake
1998 * + nr_requeue, since it acquires the rt_mutex prior to
1999 * returning to userspace, so as to not leave the rt_mutex with
2000 * waiters and no owner. However, second and third wake-ups
2001 * cannot be predicted as they involve race conditions with the
2002 * first wake and a fault while looking up the pi_state. Both
2003 * pthread_cond_signal() and pthread_cond_broadcast() should
2011 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
2012 if (unlikely(ret
!= 0))
2014 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
2015 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
2016 if (unlikely(ret
!= 0))
2020 * The check above which compares uaddrs is not sufficient for
2021 * shared futexes. We need to compare the keys:
2023 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
2028 hb1
= hash_futex(&key1
);
2029 hb2
= hash_futex(&key2
);
2032 hb_waiters_inc(hb2
);
2033 double_lock_hb(hb1
, hb2
);
2035 if (likely(cmpval
!= NULL
)) {
2038 ret
= get_futex_value_locked(&curval
, uaddr1
);
2040 if (unlikely(ret
)) {
2041 double_unlock_hb(hb1
, hb2
);
2042 hb_waiters_dec(hb2
);
2044 ret
= get_user(curval
, uaddr1
);
2048 if (!(flags
& FLAGS_SHARED
))
2051 put_futex_key(&key2
);
2052 put_futex_key(&key1
);
2055 if (curval
!= *cmpval
) {
2061 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2063 * Attempt to acquire uaddr2 and wake the top waiter. If we
2064 * intend to requeue waiters, force setting the FUTEX_WAITERS
2065 * bit. We force this here where we are able to easily handle
2066 * faults rather in the requeue loop below.
2068 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2069 &key2
, &pi_state
, nr_requeue
);
2072 * At this point the top_waiter has either taken uaddr2 or is
2073 * waiting on it. If the former, then the pi_state will not
2074 * exist yet, look it up one more time to ensure we have a
2075 * reference to it. If the lock was taken, ret contains the
2076 * vpid of the top waiter task.
2077 * If the lock was not taken, we have pi_state and an initial
2078 * refcount on it. In case of an error we have nothing.
2085 * If we acquired the lock, then the user space value
2086 * of uaddr2 should be vpid. It cannot be changed by
2087 * the top waiter as it is blocked on hb2 lock if it
2088 * tries to do so. If something fiddled with it behind
2089 * our back the pi state lookup might unearth it. So
2090 * we rather use the known value than rereading and
2091 * handing potential crap to lookup_pi_state.
2093 * If that call succeeds then we have pi_state and an
2094 * initial refcount on it.
2096 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
, &pi_state
);
2101 /* We hold a reference on the pi state. */
2104 /* If the above failed, then pi_state is NULL */
2106 double_unlock_hb(hb1
, hb2
);
2107 hb_waiters_dec(hb2
);
2108 put_futex_key(&key2
);
2109 put_futex_key(&key1
);
2110 ret
= fault_in_user_writeable(uaddr2
);
2116 * Two reasons for this:
2117 * - Owner is exiting and we just wait for the
2119 * - The user space value changed.
2121 double_unlock_hb(hb1
, hb2
);
2122 hb_waiters_dec(hb2
);
2123 put_futex_key(&key2
);
2124 put_futex_key(&key1
);
2132 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2133 if (task_count
- nr_wake
>= nr_requeue
)
2136 if (!match_futex(&this->key
, &key1
))
2140 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2141 * be paired with each other and no other futex ops.
2143 * We should never be requeueing a futex_q with a pi_state,
2144 * which is awaiting a futex_unlock_pi().
2146 if ((requeue_pi
&& !this->rt_waiter
) ||
2147 (!requeue_pi
&& this->rt_waiter
) ||
2154 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2155 * lock, we already woke the top_waiter. If not, it will be
2156 * woken by futex_unlock_pi().
2158 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2159 mark_wake_futex(&wake_q
, this);
2163 /* Ensure we requeue to the expected futex for requeue_pi. */
2164 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2170 * Requeue nr_requeue waiters and possibly one more in the case
2171 * of requeue_pi if we couldn't acquire the lock atomically.
2175 * Prepare the waiter to take the rt_mutex. Take a
2176 * refcount on the pi_state and store the pointer in
2177 * the futex_q object of the waiter.
2179 get_pi_state(pi_state
);
2180 this->pi_state
= pi_state
;
2181 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2186 * We got the lock. We do neither drop the
2187 * refcount on pi_state nor clear
2188 * this->pi_state because the waiter needs the
2189 * pi_state for cleaning up the user space
2190 * value. It will drop the refcount after
2193 requeue_pi_wake_futex(this, &key2
, hb2
);
2198 * rt_mutex_start_proxy_lock() detected a
2199 * potential deadlock when we tried to queue
2200 * that waiter. Drop the pi_state reference
2201 * which we took above and remove the pointer
2202 * to the state from the waiters futex_q
2205 this->pi_state
= NULL
;
2206 put_pi_state(pi_state
);
2208 * We stop queueing more waiters and let user
2209 * space deal with the mess.
2214 requeue_futex(this, hb1
, hb2
, &key2
);
2219 * We took an extra initial reference to the pi_state either
2220 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2221 * need to drop it here again.
2223 put_pi_state(pi_state
);
2226 double_unlock_hb(hb1
, hb2
);
2228 hb_waiters_dec(hb2
);
2231 * drop_futex_key_refs() must be called outside the spinlocks. During
2232 * the requeue we moved futex_q's from the hash bucket at key1 to the
2233 * one at key2 and updated their key pointer. We no longer need to
2234 * hold the references to key1.
2236 while (--drop_count
>= 0)
2237 drop_futex_key_refs(&key1
);
2240 put_futex_key(&key2
);
2242 put_futex_key(&key1
);
2244 return ret
? ret
: task_count
;
2247 /* The key must be already stored in q->key. */
2248 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2249 __acquires(&hb
->lock
)
2251 struct futex_hash_bucket
*hb
;
2253 hb
= hash_futex(&q
->key
);
2256 * Increment the counter before taking the lock so that
2257 * a potential waker won't miss a to-be-slept task that is
2258 * waiting for the spinlock. This is safe as all queue_lock()
2259 * users end up calling queue_me(). Similarly, for housekeeping,
2260 * decrement the counter at queue_unlock() when some error has
2261 * occurred and we don't end up adding the task to the list.
2265 q
->lock_ptr
= &hb
->lock
;
2267 spin_lock(&hb
->lock
); /* implies smp_mb(); (A) */
2272 queue_unlock(struct futex_hash_bucket
*hb
)
2273 __releases(&hb
->lock
)
2275 spin_unlock(&hb
->lock
);
2279 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2284 * The priority used to register this element is
2285 * - either the real thread-priority for the real-time threads
2286 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2287 * - or MAX_RT_PRIO for non-RT threads.
2288 * Thus, all RT-threads are woken first in priority order, and
2289 * the others are woken last, in FIFO order.
2291 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2293 plist_node_init(&q
->list
, prio
);
2294 plist_add(&q
->list
, &hb
->chain
);
2299 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2300 * @q: The futex_q to enqueue
2301 * @hb: The destination hash bucket
2303 * The hb->lock must be held by the caller, and is released here. A call to
2304 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2305 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2306 * or nothing if the unqueue is done as part of the wake process and the unqueue
2307 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2310 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2311 __releases(&hb
->lock
)
2314 spin_unlock(&hb
->lock
);
2318 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2319 * @q: The futex_q to unqueue
2321 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2322 * be paired with exactly one earlier call to queue_me().
2325 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2326 * - 0 - if the futex_q was already removed by the waking thread
2328 static int unqueue_me(struct futex_q
*q
)
2330 spinlock_t
*lock_ptr
;
2333 /* In the common case we don't take the spinlock, which is nice. */
2336 * q->lock_ptr can change between this read and the following spin_lock.
2337 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2338 * optimizing lock_ptr out of the logic below.
2340 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2341 if (lock_ptr
!= NULL
) {
2342 spin_lock(lock_ptr
);
2344 * q->lock_ptr can change between reading it and
2345 * spin_lock(), causing us to take the wrong lock. This
2346 * corrects the race condition.
2348 * Reasoning goes like this: if we have the wrong lock,
2349 * q->lock_ptr must have changed (maybe several times)
2350 * between reading it and the spin_lock(). It can
2351 * change again after the spin_lock() but only if it was
2352 * already changed before the spin_lock(). It cannot,
2353 * however, change back to the original value. Therefore
2354 * we can detect whether we acquired the correct lock.
2356 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2357 spin_unlock(lock_ptr
);
2362 BUG_ON(q
->pi_state
);
2364 spin_unlock(lock_ptr
);
2368 drop_futex_key_refs(&q
->key
);
2373 * PI futexes can not be requeued and must remove themself from the
2374 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2377 static void unqueue_me_pi(struct futex_q
*q
)
2378 __releases(q
->lock_ptr
)
2382 BUG_ON(!q
->pi_state
);
2383 put_pi_state(q
->pi_state
);
2386 spin_unlock(q
->lock_ptr
);
2389 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2390 struct task_struct
*argowner
)
2392 struct futex_pi_state
*pi_state
= q
->pi_state
;
2393 u32 uval
, uninitialized_var(curval
), newval
;
2394 struct task_struct
*oldowner
, *newowner
;
2398 lockdep_assert_held(q
->lock_ptr
);
2400 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2402 oldowner
= pi_state
->owner
;
2405 * We are here because either:
2407 * - we stole the lock and pi_state->owner needs updating to reflect
2408 * that (@argowner == current),
2412 * - someone stole our lock and we need to fix things to point to the
2413 * new owner (@argowner == NULL).
2415 * Either way, we have to replace the TID in the user space variable.
2416 * This must be atomic as we have to preserve the owner died bit here.
2418 * Note: We write the user space value _before_ changing the pi_state
2419 * because we can fault here. Imagine swapped out pages or a fork
2420 * that marked all the anonymous memory readonly for cow.
2422 * Modifying pi_state _before_ the user space value would leave the
2423 * pi_state in an inconsistent state when we fault here, because we
2424 * need to drop the locks to handle the fault. This might be observed
2425 * in the PID check in lookup_pi_state.
2429 if (oldowner
!= current
) {
2431 * We raced against a concurrent self; things are
2432 * already fixed up. Nothing to do.
2438 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2439 /* We got the lock after all, nothing to fix. */
2445 * Since we just failed the trylock; there must be an owner.
2447 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2450 WARN_ON_ONCE(argowner
!= current
);
2451 if (oldowner
== current
) {
2453 * We raced against a concurrent self; things are
2454 * already fixed up. Nothing to do.
2459 newowner
= argowner
;
2462 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2464 if (!pi_state
->owner
)
2465 newtid
|= FUTEX_OWNER_DIED
;
2467 if (get_futex_value_locked(&uval
, uaddr
))
2471 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2473 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
2481 * We fixed up user space. Now we need to fix the pi_state
2484 pi_state_update_owner(pi_state
, newowner
);
2485 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2490 * To handle the page fault we need to drop the locks here. That gives
2491 * the other task (either the highest priority waiter itself or the
2492 * task which stole the rtmutex) the chance to try the fixup of the
2493 * pi_state. So once we are back from handling the fault we need to
2494 * check the pi_state after reacquiring the locks and before trying to
2495 * do another fixup. When the fixup has been done already we simply
2498 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2499 * drop hb->lock since the caller owns the hb -> futex_q relation.
2500 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2503 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2504 spin_unlock(q
->lock_ptr
);
2506 ret
= fault_in_user_writeable(uaddr
);
2508 spin_lock(q
->lock_ptr
);
2509 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2512 * Check if someone else fixed it for us:
2514 if (pi_state
->owner
!= oldowner
) {
2525 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2529 static long futex_wait_restart(struct restart_block
*restart
);
2532 * fixup_owner() - Post lock pi_state and corner case management
2533 * @uaddr: user address of the futex
2534 * @q: futex_q (contains pi_state and access to the rt_mutex)
2535 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2537 * After attempting to lock an rt_mutex, this function is called to cleanup
2538 * the pi_state owner as well as handle race conditions that may allow us to
2539 * acquire the lock. Must be called with the hb lock held.
2542 * - 1 - success, lock taken;
2543 * - 0 - success, lock not taken;
2544 * - <0 - on error (-EFAULT)
2546 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2552 * Got the lock. We might not be the anticipated owner if we
2553 * did a lock-steal - fix up the PI-state in that case:
2555 * Speculative pi_state->owner read (we don't hold wait_lock);
2556 * since we own the lock pi_state->owner == current is the
2557 * stable state, anything else needs more attention.
2559 if (q
->pi_state
->owner
!= current
)
2560 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2565 * If we didn't get the lock; check if anybody stole it from us. In
2566 * that case, we need to fix up the uval to point to them instead of
2567 * us, otherwise bad things happen. [10]
2569 * Another speculative read; pi_state->owner == current is unstable
2570 * but needs our attention.
2572 if (q
->pi_state
->owner
== current
) {
2573 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2578 * Paranoia check. If we did not take the lock, then we should not be
2579 * the owner of the rt_mutex.
2581 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2582 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2583 "pi-state %p\n", ret
,
2584 q
->pi_state
->pi_mutex
.owner
,
2585 q
->pi_state
->owner
);
2589 return ret
? ret
: locked
;
2593 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2594 * @hb: the futex hash bucket, must be locked by the caller
2595 * @q: the futex_q to queue up on
2596 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2598 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2599 struct hrtimer_sleeper
*timeout
)
2602 * The task state is guaranteed to be set before another task can
2603 * wake it. set_current_state() is implemented using smp_store_mb() and
2604 * queue_me() calls spin_unlock() upon completion, both serializing
2605 * access to the hash list and forcing another memory barrier.
2607 set_current_state(TASK_INTERRUPTIBLE
);
2612 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2615 * If we have been removed from the hash list, then another task
2616 * has tried to wake us, and we can skip the call to schedule().
2618 if (likely(!plist_node_empty(&q
->list
))) {
2620 * If the timer has already expired, current will already be
2621 * flagged for rescheduling. Only call schedule if there
2622 * is no timeout, or if it has yet to expire.
2624 if (!timeout
|| timeout
->task
)
2625 freezable_schedule();
2627 __set_current_state(TASK_RUNNING
);
2631 * futex_wait_setup() - Prepare to wait on a futex
2632 * @uaddr: the futex userspace address
2633 * @val: the expected value
2634 * @flags: futex flags (FLAGS_SHARED, etc.)
2635 * @q: the associated futex_q
2636 * @hb: storage for hash_bucket pointer to be returned to caller
2638 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2639 * compare it with the expected value. Handle atomic faults internally.
2640 * Return with the hb lock held and a q.key reference on success, and unlocked
2641 * with no q.key reference on failure.
2644 * - 0 - uaddr contains val and hb has been locked;
2645 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2647 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2648 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2654 * Access the page AFTER the hash-bucket is locked.
2655 * Order is important:
2657 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2658 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2660 * The basic logical guarantee of a futex is that it blocks ONLY
2661 * if cond(var) is known to be true at the time of blocking, for
2662 * any cond. If we locked the hash-bucket after testing *uaddr, that
2663 * would open a race condition where we could block indefinitely with
2664 * cond(var) false, which would violate the guarantee.
2666 * On the other hand, we insert q and release the hash-bucket only
2667 * after testing *uaddr. This guarantees that futex_wait() will NOT
2668 * absorb a wakeup if *uaddr does not match the desired values
2669 * while the syscall executes.
2672 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2673 if (unlikely(ret
!= 0))
2677 *hb
= queue_lock(q
);
2679 ret
= get_futex_value_locked(&uval
, uaddr
);
2684 ret
= get_user(uval
, uaddr
);
2688 if (!(flags
& FLAGS_SHARED
))
2691 put_futex_key(&q
->key
);
2702 put_futex_key(&q
->key
);
2706 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2707 ktime_t
*abs_time
, u32 bitset
)
2709 struct hrtimer_sleeper timeout
, *to
= NULL
;
2710 struct restart_block
*restart
;
2711 struct futex_hash_bucket
*hb
;
2712 struct futex_q q
= futex_q_init
;
2722 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2723 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2725 hrtimer_init_sleeper(to
, current
);
2726 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2727 current
->timer_slack_ns
);
2732 * Prepare to wait on uaddr. On success, holds hb lock and increments
2735 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2739 /* queue_me and wait for wakeup, timeout, or a signal. */
2740 futex_wait_queue_me(hb
, &q
, to
);
2742 /* If we were woken (and unqueued), we succeeded, whatever. */
2744 /* unqueue_me() drops q.key ref */
2745 if (!unqueue_me(&q
))
2748 if (to
&& !to
->task
)
2752 * We expect signal_pending(current), but we might be the
2753 * victim of a spurious wakeup as well.
2755 if (!signal_pending(current
))
2762 restart
= ¤t
->restart_block
;
2763 restart
->fn
= futex_wait_restart
;
2764 restart
->futex
.uaddr
= uaddr
;
2765 restart
->futex
.val
= val
;
2766 restart
->futex
.time
= *abs_time
;
2767 restart
->futex
.bitset
= bitset
;
2768 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2770 ret
= -ERESTART_RESTARTBLOCK
;
2774 hrtimer_cancel(&to
->timer
);
2775 destroy_hrtimer_on_stack(&to
->timer
);
2781 static long futex_wait_restart(struct restart_block
*restart
)
2783 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2784 ktime_t t
, *tp
= NULL
;
2786 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2787 t
= restart
->futex
.time
;
2790 restart
->fn
= do_no_restart_syscall
;
2792 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2793 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2798 * Userspace tried a 0 -> TID atomic transition of the futex value
2799 * and failed. The kernel side here does the whole locking operation:
2800 * if there are waiters then it will block as a consequence of relying
2801 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2802 * a 0 value of the futex too.).
2804 * Also serves as futex trylock_pi()'ing, and due semantics.
2806 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2807 ktime_t
*time
, int trylock
)
2809 struct hrtimer_sleeper timeout
, *to
= NULL
;
2810 struct futex_pi_state
*pi_state
= NULL
;
2811 struct rt_mutex_waiter rt_waiter
;
2812 struct futex_hash_bucket
*hb
;
2813 struct futex_q q
= futex_q_init
;
2816 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2819 if (refill_pi_state_cache())
2824 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2826 hrtimer_init_sleeper(to
, current
);
2827 hrtimer_set_expires(&to
->timer
, *time
);
2831 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2832 if (unlikely(ret
!= 0))
2836 hb
= queue_lock(&q
);
2838 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2839 if (unlikely(ret
)) {
2841 * Atomic work succeeded and we got the lock,
2842 * or failed. Either way, we do _not_ block.
2846 /* We got the lock. */
2848 goto out_unlock_put_key
;
2853 * Two reasons for this:
2854 * - Task is exiting and we just wait for the
2856 * - The user space value changed.
2859 put_futex_key(&q
.key
);
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_start_expires(&to
->timer
, 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
, VERIFY_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.
3075 if (ret
== -EAGAIN
) {
3076 put_futex_key(&key
);
3080 * wake_futex_pi has detected invalid state. Tell user
3087 * We have no kernel internal state, i.e. no waiters in the
3088 * kernel. Waiters which are about to queue themselves are stuck
3089 * on hb->lock. So we can safely ignore them. We do neither
3090 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3093 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0)) {
3094 spin_unlock(&hb
->lock
);
3099 * If uval has changed, let user space handle it.
3101 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3104 spin_unlock(&hb
->lock
);
3106 put_futex_key(&key
);
3110 put_futex_key(&key
);
3112 ret
= fault_in_user_writeable(uaddr
);
3120 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3121 * @hb: the hash_bucket futex_q was original enqueued on
3122 * @q: the futex_q woken while waiting to be requeued
3123 * @key2: the futex_key of the requeue target futex
3124 * @timeout: the timeout associated with the wait (NULL if none)
3126 * Detect if the task was woken on the initial futex as opposed to the requeue
3127 * target futex. If so, determine if it was a timeout or a signal that caused
3128 * the wakeup and return the appropriate error code to the caller. Must be
3129 * called with the hb lock held.
3132 * - 0 = no early wakeup detected;
3133 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3136 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3137 struct futex_q
*q
, union futex_key
*key2
,
3138 struct hrtimer_sleeper
*timeout
)
3143 * With the hb lock held, we avoid races while we process the wakeup.
3144 * We only need to hold hb (and not hb2) to ensure atomicity as the
3145 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3146 * It can't be requeued from uaddr2 to something else since we don't
3147 * support a PI aware source futex for requeue.
3149 if (!match_futex(&q
->key
, key2
)) {
3150 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3152 * We were woken prior to requeue by a timeout or a signal.
3153 * Unqueue the futex_q and determine which it was.
3155 plist_del(&q
->list
, &hb
->chain
);
3158 /* Handle spurious wakeups gracefully */
3160 if (timeout
&& !timeout
->task
)
3162 else if (signal_pending(current
))
3163 ret
= -ERESTARTNOINTR
;
3169 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3170 * @uaddr: the futex we initially wait on (non-pi)
3171 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3172 * the same type, no requeueing from private to shared, etc.
3173 * @val: the expected value of uaddr
3174 * @abs_time: absolute timeout
3175 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3176 * @uaddr2: the pi futex we will take prior to returning to user-space
3178 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3179 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3180 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3181 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3182 * without one, the pi logic would not know which task to boost/deboost, if
3183 * there was a need to.
3185 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3186 * via the following--
3187 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3188 * 2) wakeup on uaddr2 after a requeue
3192 * If 3, cleanup and return -ERESTARTNOINTR.
3194 * If 2, we may then block on trying to take the rt_mutex and return via:
3195 * 5) successful lock
3198 * 8) other lock acquisition failure
3200 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3202 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3208 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3209 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3212 struct hrtimer_sleeper timeout
, *to
= NULL
;
3213 struct futex_pi_state
*pi_state
= NULL
;
3214 struct rt_mutex_waiter rt_waiter
;
3215 struct futex_hash_bucket
*hb
;
3216 union futex_key key2
= FUTEX_KEY_INIT
;
3217 struct futex_q q
= futex_q_init
;
3220 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3223 if (uaddr
== uaddr2
)
3231 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
3232 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
3234 hrtimer_init_sleeper(to
, current
);
3235 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
3236 current
->timer_slack_ns
);
3240 * The waiter is allocated on our stack, manipulated by the requeue
3241 * code while we sleep on uaddr.
3243 rt_mutex_init_waiter(&rt_waiter
);
3245 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
3246 if (unlikely(ret
!= 0))
3250 q
.rt_waiter
= &rt_waiter
;
3251 q
.requeue_pi_key
= &key2
;
3254 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3257 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3262 * The check above which compares uaddrs is not sufficient for
3263 * shared futexes. We need to compare the keys:
3265 if (match_futex(&q
.key
, &key2
)) {
3271 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3272 futex_wait_queue_me(hb
, &q
, to
);
3274 spin_lock(&hb
->lock
);
3275 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3276 spin_unlock(&hb
->lock
);
3281 * In order for us to be here, we know our q.key == key2, and since
3282 * we took the hb->lock above, we also know that futex_requeue() has
3283 * completed and we no longer have to concern ourselves with a wakeup
3284 * race with the atomic proxy lock acquisition by the requeue code. The
3285 * futex_requeue dropped our key1 reference and incremented our key2
3289 /* Check if the requeue code acquired the second futex for us. */
3292 * Got the lock. We might not be the anticipated owner if we
3293 * did a lock-steal - fix up the PI-state in that case.
3295 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3296 spin_lock(q
.lock_ptr
);
3297 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3298 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3299 pi_state
= q
.pi_state
;
3300 get_pi_state(pi_state
);
3303 * Drop the reference to the pi state which
3304 * the requeue_pi() code acquired for us.
3306 put_pi_state(q
.pi_state
);
3307 spin_unlock(q
.lock_ptr
);
3310 struct rt_mutex
*pi_mutex
;
3313 * We have been woken up by futex_unlock_pi(), a timeout, or a
3314 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3317 WARN_ON(!q
.pi_state
);
3318 pi_mutex
= &q
.pi_state
->pi_mutex
;
3319 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3321 spin_lock(q
.lock_ptr
);
3322 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3325 debug_rt_mutex_free_waiter(&rt_waiter
);
3327 * Fixup the pi_state owner and possibly acquire the lock if we
3330 res
= fixup_owner(uaddr2
, &q
, !ret
);
3332 * If fixup_owner() returned an error, proprogate that. If it
3333 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3336 ret
= (res
< 0) ? res
: 0;
3339 * If fixup_pi_state_owner() faulted and was unable to handle
3340 * the fault, unlock the rt_mutex and return the fault to
3343 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3344 pi_state
= q
.pi_state
;
3345 get_pi_state(pi_state
);
3348 /* Unqueue and drop the lock. */
3353 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3354 put_pi_state(pi_state
);
3357 if (ret
== -EINTR
) {
3359 * We've already been requeued, but cannot restart by calling
3360 * futex_lock_pi() directly. We could restart this syscall, but
3361 * it would detect that the user space "val" changed and return
3362 * -EWOULDBLOCK. Save the overhead of the restart and return
3363 * -EWOULDBLOCK directly.
3369 put_futex_key(&q
.key
);
3371 put_futex_key(&key2
);
3375 hrtimer_cancel(&to
->timer
);
3376 destroy_hrtimer_on_stack(&to
->timer
);
3382 * Support for robust futexes: the kernel cleans up held futexes at
3385 * Implementation: user-space maintains a per-thread list of locks it
3386 * is holding. Upon do_exit(), the kernel carefully walks this list,
3387 * and marks all locks that are owned by this thread with the
3388 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3389 * always manipulated with the lock held, so the list is private and
3390 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3391 * field, to allow the kernel to clean up if the thread dies after
3392 * acquiring the lock, but just before it could have added itself to
3393 * the list. There can only be one such pending lock.
3397 * sys_set_robust_list() - Set the robust-futex list head of a task
3398 * @head: pointer to the list-head
3399 * @len: length of the list-head, as userspace expects
3401 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3404 if (!futex_cmpxchg_enabled
)
3407 * The kernel knows only one size for now:
3409 if (unlikely(len
!= sizeof(*head
)))
3412 current
->robust_list
= head
;
3418 * sys_get_robust_list() - Get the robust-futex list head of a task
3419 * @pid: pid of the process [zero for current task]
3420 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3421 * @len_ptr: pointer to a length field, the kernel fills in the header size
3423 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3424 struct robust_list_head __user
* __user
*, head_ptr
,
3425 size_t __user
*, len_ptr
)
3427 struct robust_list_head __user
*head
;
3429 struct task_struct
*p
;
3431 if (!futex_cmpxchg_enabled
)
3440 p
= find_task_by_vpid(pid
);
3446 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3449 head
= p
->robust_list
;
3452 if (put_user(sizeof(*head
), len_ptr
))
3454 return put_user(head
, head_ptr
);
3463 * Process a futex-list entry, check whether it's owned by the
3464 * dying task, and do notification if so:
3466 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
3468 u32 uval
, uninitialized_var(nval
), mval
;
3470 /* Futex address must be 32bit aligned */
3471 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3475 if (get_user(uval
, uaddr
))
3478 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
3480 * Ok, this dying thread is truly holding a futex
3481 * of interest. Set the OWNER_DIED bit atomically
3482 * via cmpxchg, and if the value had FUTEX_WAITERS
3483 * set, wake up a waiter (if any). (We have to do a
3484 * futex_wake() even if OWNER_DIED is already set -
3485 * to handle the rare but possible case of recursive
3486 * thread-death.) The rest of the cleanup is done in
3489 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3491 * We are not holding a lock here, but we want to have
3492 * the pagefault_disable/enable() protection because
3493 * we want to handle the fault gracefully. If the
3494 * access fails we try to fault in the futex with R/W
3495 * verification via get_user_pages. get_user() above
3496 * does not guarantee R/W access. If that fails we
3497 * give up and leave the futex locked.
3499 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
3500 if (fault_in_user_writeable(uaddr
))
3508 * Wake robust non-PI futexes here. The wakeup of
3509 * PI futexes happens in exit_pi_state():
3511 if (!pi
&& (uval
& FUTEX_WAITERS
))
3512 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3518 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3520 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3521 struct robust_list __user
* __user
*head
,
3524 unsigned long uentry
;
3526 if (get_user(uentry
, (unsigned long __user
*)head
))
3529 *entry
= (void __user
*)(uentry
& ~1UL);
3536 * Walk curr->robust_list (very carefully, it's a userspace list!)
3537 * and mark any locks found there dead, and notify any waiters.
3539 * We silently return on any sign of list-walking problem.
3541 void exit_robust_list(struct task_struct
*curr
)
3543 struct robust_list_head __user
*head
= curr
->robust_list
;
3544 struct robust_list __user
*entry
, *next_entry
, *pending
;
3545 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3546 unsigned int uninitialized_var(next_pi
);
3547 unsigned long futex_offset
;
3550 if (!futex_cmpxchg_enabled
)
3554 * Fetch the list head (which was registered earlier, via
3555 * sys_set_robust_list()):
3557 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3560 * Fetch the relative futex offset:
3562 if (get_user(futex_offset
, &head
->futex_offset
))
3565 * Fetch any possibly pending lock-add first, and handle it
3568 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3571 next_entry
= NULL
; /* avoid warning with gcc */
3572 while (entry
!= &head
->list
) {
3574 * Fetch the next entry in the list before calling
3575 * handle_futex_death:
3577 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3579 * A pending lock might already be on the list, so
3580 * don't process it twice:
3582 if (entry
!= pending
)
3583 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3591 * Avoid excessively long or circular lists:
3600 handle_futex_death((void __user
*)pending
+ futex_offset
,
3604 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3605 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3607 int cmd
= op
& FUTEX_CMD_MASK
;
3608 unsigned int flags
= 0;
3610 if (!(op
& FUTEX_PRIVATE_FLAG
))
3611 flags
|= FLAGS_SHARED
;
3613 if (op
& FUTEX_CLOCK_REALTIME
) {
3614 flags
|= FLAGS_CLOCKRT
;
3615 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3616 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3622 case FUTEX_UNLOCK_PI
:
3623 case FUTEX_TRYLOCK_PI
:
3624 case FUTEX_WAIT_REQUEUE_PI
:
3625 case FUTEX_CMP_REQUEUE_PI
:
3626 if (!futex_cmpxchg_enabled
)
3632 val3
= FUTEX_BITSET_MATCH_ANY
;
3633 case FUTEX_WAIT_BITSET
:
3634 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3636 val3
= FUTEX_BITSET_MATCH_ANY
;
3637 case FUTEX_WAKE_BITSET
:
3638 return futex_wake(uaddr
, flags
, val
, val3
);
3640 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3641 case FUTEX_CMP_REQUEUE
:
3642 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3644 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3646 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3647 case FUTEX_UNLOCK_PI
:
3648 return futex_unlock_pi(uaddr
, flags
);
3649 case FUTEX_TRYLOCK_PI
:
3650 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3651 case FUTEX_WAIT_REQUEUE_PI
:
3652 val3
= FUTEX_BITSET_MATCH_ANY
;
3653 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3655 case FUTEX_CMP_REQUEUE_PI
:
3656 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3662 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3663 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3667 ktime_t t
, *tp
= NULL
;
3669 int cmd
= op
& FUTEX_CMD_MASK
;
3671 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3672 cmd
== FUTEX_WAIT_BITSET
||
3673 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3674 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3676 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3678 if (!timespec_valid(&ts
))
3681 t
= timespec_to_ktime(ts
);
3682 if (cmd
== FUTEX_WAIT
)
3683 t
= ktime_add_safe(ktime_get(), t
);
3687 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3688 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3690 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3691 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3692 val2
= (u32
) (unsigned long) utime
;
3694 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3697 static void __init
futex_detect_cmpxchg(void)
3699 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3703 * This will fail and we want it. Some arch implementations do
3704 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3705 * functionality. We want to know that before we call in any
3706 * of the complex code paths. Also we want to prevent
3707 * registration of robust lists in that case. NULL is
3708 * guaranteed to fault and we get -EFAULT on functional
3709 * implementation, the non-functional ones will return
3712 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3713 futex_cmpxchg_enabled
= 1;
3717 static int __init
futex_init(void)
3719 unsigned int futex_shift
;
3722 #if CONFIG_BASE_SMALL
3723 futex_hashsize
= 16;
3725 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3728 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3730 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3732 futex_hashsize
, futex_hashsize
);
3733 futex_hashsize
= 1UL << futex_shift
;
3735 futex_detect_cmpxchg();
3737 for (i
= 0; i
< futex_hashsize
; i
++) {
3738 atomic_set(&futex_queues
[i
].waiters
, 0);
3739 plist_head_init(&futex_queues
[i
].chain
);
3740 spin_lock_init(&futex_queues
[i
].lock
);
3745 core_initcall(futex_init
);