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 ihold(key
->shared
.inode
); /* implies 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
)) {
473 iput(key
->shared
.inode
);
475 case FUT_OFF_MMSHARED
:
476 mmdrop(key
->private.mm
);
482 * get_futex_key() - Get parameters which are the keys for a futex
483 * @uaddr: virtual address of the futex
484 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485 * @key: address where result is stored.
486 * @rw: mapping needs to be read/write (values: VERIFY_READ,
489 * Return: a negative error code or 0
491 * The key words are stored in @key on success.
493 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
494 * offset_within_page). For private mappings, it's (uaddr, current->mm).
495 * We can usually work out the index without swapping in the page.
497 * lock_page() might sleep, the caller should not hold a spinlock.
500 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
502 unsigned long address
= (unsigned long)uaddr
;
503 struct mm_struct
*mm
= current
->mm
;
504 struct page
*page
, *tail
;
505 struct address_space
*mapping
;
509 * The futex address must be "naturally" aligned.
511 key
->both
.offset
= address
% PAGE_SIZE
;
512 if (unlikely((address
% sizeof(u32
)) != 0))
514 address
-= key
->both
.offset
;
516 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
519 if (unlikely(should_fail_futex(fshared
)))
523 * PROCESS_PRIVATE futexes are fast.
524 * As the mm cannot disappear under us and the 'key' only needs
525 * virtual address, we dont even have to find the underlying vma.
526 * Note : We do have to check 'uaddr' is a valid user address,
527 * but access_ok() should be faster than find_vma()
530 key
->private.mm
= mm
;
531 key
->private.address
= address
;
532 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared
)))
541 err
= get_user_pages_fast(address
, 1, 1, &page
);
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
546 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
547 err
= get_user_pages_fast(address
, 1, 0, &page
);
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
574 page
= compound_head(page
);
575 mapping
= READ_ONCE(page
->mapping
);
578 * If page->mapping is NULL, then it cannot be a PageAnon
579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
590 * an unlikely race, but we do need to retry for page->mapping.
592 if (unlikely(!mapping
)) {
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
601 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
612 * Private mappings are handled in a simple way.
614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
619 * the object not the particular process.
621 if (PageAnon(page
)) {
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
626 if (unlikely(should_fail_futex(fshared
)) || ro
) {
631 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
632 key
->private.mm
= mm
;
633 key
->private.address
= address
;
635 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
641 * The associated futex object in this case is the inode and
642 * the page->mapping must be traversed. Ordinarily this should
643 * be stabilised under page lock but it's not strictly
644 * necessary in this case as we just want to pin the inode, not
645 * update the radix tree or anything like that.
647 * The RCU read lock is taken as the inode is finally freed
648 * under RCU. If the mapping still matches expectations then the
649 * mapping->host can be safely accessed as being a valid inode.
653 if (READ_ONCE(page
->mapping
) != mapping
) {
660 inode
= READ_ONCE(mapping
->host
);
669 * Take a reference unless it is about to be freed. Previously
670 * this reference was taken by ihold under the page lock
671 * pinning the inode in place so i_lock was unnecessary. The
672 * only way for this check to fail is if the inode was
673 * truncated in parallel which is almost certainly an
674 * application bug. In such a case, just retry.
676 * We are not calling into get_futex_key_refs() in file-backed
677 * cases, therefore a successful atomic_inc return below will
678 * guarantee that get_futex_key() will still imply smp_mb(); (B).
680 if (!atomic_inc_not_zero(&inode
->i_count
)) {
687 /* Should be impossible but lets be paranoid for now */
688 if (WARN_ON_ONCE(inode
->i_mapping
!= mapping
)) {
696 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
697 key
->shared
.inode
= inode
;
698 key
->shared
.pgoff
= basepage_index(tail
);
707 static inline void put_futex_key(union futex_key
*key
)
709 drop_futex_key_refs(key
);
713 * fault_in_user_writeable() - Fault in user address and verify RW access
714 * @uaddr: pointer to faulting user space address
716 * Slow path to fixup the fault we just took in the atomic write
719 * We have no generic implementation of a non-destructive write to the
720 * user address. We know that we faulted in the atomic pagefault
721 * disabled section so we can as well avoid the #PF overhead by
722 * calling get_user_pages() right away.
724 static int fault_in_user_writeable(u32 __user
*uaddr
)
726 struct mm_struct
*mm
= current
->mm
;
729 down_read(&mm
->mmap_sem
);
730 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
731 FAULT_FLAG_WRITE
, NULL
);
732 up_read(&mm
->mmap_sem
);
734 return ret
< 0 ? ret
: 0;
738 * futex_top_waiter() - Return the highest priority waiter on a futex
739 * @hb: the hash bucket the futex_q's reside in
740 * @key: the futex key (to distinguish it from other futex futex_q's)
742 * Must be called with the hb lock held.
744 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
745 union futex_key
*key
)
747 struct futex_q
*this;
749 plist_for_each_entry(this, &hb
->chain
, list
) {
750 if (match_futex(&this->key
, key
))
756 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
757 u32 uval
, u32 newval
)
762 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
768 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
773 ret
= __get_user(*dest
, from
);
776 return ret
? -EFAULT
: 0;
783 static int refill_pi_state_cache(void)
785 struct futex_pi_state
*pi_state
;
787 if (likely(current
->pi_state_cache
))
790 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
795 INIT_LIST_HEAD(&pi_state
->list
);
796 /* pi_mutex gets initialized later */
797 pi_state
->owner
= NULL
;
798 atomic_set(&pi_state
->refcount
, 1);
799 pi_state
->key
= FUTEX_KEY_INIT
;
801 current
->pi_state_cache
= pi_state
;
806 static struct futex_pi_state
*alloc_pi_state(void)
808 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
811 current
->pi_state_cache
= NULL
;
816 static void get_pi_state(struct futex_pi_state
*pi_state
)
818 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state
->refcount
));
822 * Drops a reference to the pi_state object and frees or caches it
823 * when the last reference is gone.
825 static void put_pi_state(struct futex_pi_state
*pi_state
)
830 if (!atomic_dec_and_test(&pi_state
->refcount
))
834 * If pi_state->owner is NULL, the owner is most probably dying
835 * and has cleaned up the pi_state already
837 if (pi_state
->owner
) {
838 struct task_struct
*owner
;
840 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
841 owner
= pi_state
->owner
;
843 raw_spin_lock(&owner
->pi_lock
);
844 list_del_init(&pi_state
->list
);
845 raw_spin_unlock(&owner
->pi_lock
);
847 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
848 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
851 if (current
->pi_state_cache
) {
855 * pi_state->list is already empty.
856 * clear pi_state->owner.
857 * refcount is at 0 - put it back to 1.
859 pi_state
->owner
= NULL
;
860 atomic_set(&pi_state
->refcount
, 1);
861 current
->pi_state_cache
= pi_state
;
866 * Look up the task based on what TID userspace gave us.
869 static struct task_struct
*futex_find_get_task(pid_t pid
)
871 struct task_struct
*p
;
874 p
= find_task_by_vpid(pid
);
883 #ifdef CONFIG_FUTEX_PI
886 * This task is holding PI mutexes at exit time => bad.
887 * Kernel cleans up PI-state, but userspace is likely hosed.
888 * (Robust-futex cleanup is separate and might save the day for userspace.)
890 void exit_pi_state_list(struct task_struct
*curr
)
892 struct list_head
*next
, *head
= &curr
->pi_state_list
;
893 struct futex_pi_state
*pi_state
;
894 struct futex_hash_bucket
*hb
;
895 union futex_key key
= FUTEX_KEY_INIT
;
897 if (!futex_cmpxchg_enabled
)
900 * We are a ZOMBIE and nobody can enqueue itself on
901 * pi_state_list anymore, but we have to be careful
902 * versus waiters unqueueing themselves:
904 raw_spin_lock_irq(&curr
->pi_lock
);
905 while (!list_empty(head
)) {
907 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
909 hb
= hash_futex(&key
);
912 * We can race against put_pi_state() removing itself from the
913 * list (a waiter going away). put_pi_state() will first
914 * decrement the reference count and then modify the list, so
915 * its possible to see the list entry but fail this reference
918 * In that case; drop the locks to let put_pi_state() make
919 * progress and retry the loop.
921 if (!atomic_inc_not_zero(&pi_state
->refcount
)) {
922 raw_spin_unlock_irq(&curr
->pi_lock
);
924 raw_spin_lock_irq(&curr
->pi_lock
);
927 raw_spin_unlock_irq(&curr
->pi_lock
);
929 spin_lock(&hb
->lock
);
930 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
931 raw_spin_lock(&curr
->pi_lock
);
933 * We dropped the pi-lock, so re-check whether this
934 * task still owns the PI-state:
936 if (head
->next
!= next
) {
937 /* retain curr->pi_lock for the loop invariant */
938 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
939 spin_unlock(&hb
->lock
);
940 put_pi_state(pi_state
);
944 WARN_ON(pi_state
->owner
!= curr
);
945 WARN_ON(list_empty(&pi_state
->list
));
946 list_del_init(&pi_state
->list
);
947 pi_state
->owner
= NULL
;
949 raw_spin_unlock(&curr
->pi_lock
);
950 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
951 spin_unlock(&hb
->lock
);
953 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
954 put_pi_state(pi_state
);
956 raw_spin_lock_irq(&curr
->pi_lock
);
958 raw_spin_unlock_irq(&curr
->pi_lock
);
964 * We need to check the following states:
966 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
968 * [1] NULL | --- | --- | 0 | 0/1 | Valid
969 * [2] NULL | --- | --- | >0 | 0/1 | Valid
971 * [3] Found | NULL | -- | Any | 0/1 | Invalid
973 * [4] Found | Found | NULL | 0 | 1 | Valid
974 * [5] Found | Found | NULL | >0 | 1 | Invalid
976 * [6] Found | Found | task | 0 | 1 | Valid
978 * [7] Found | Found | NULL | Any | 0 | Invalid
980 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
981 * [9] Found | Found | task | 0 | 0 | Invalid
982 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
984 * [1] Indicates that the kernel can acquire the futex atomically. We
985 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
987 * [2] Valid, if TID does not belong to a kernel thread. If no matching
988 * thread is found then it indicates that the owner TID has died.
990 * [3] Invalid. The waiter is queued on a non PI futex
992 * [4] Valid state after exit_robust_list(), which sets the user space
993 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
995 * [5] The user space value got manipulated between exit_robust_list()
996 * and exit_pi_state_list()
998 * [6] Valid state after exit_pi_state_list() which sets the new owner in
999 * the pi_state but cannot access the user space value.
1001 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1003 * [8] Owner and user space value match
1005 * [9] There is no transient state which sets the user space TID to 0
1006 * except exit_robust_list(), but this is indicated by the
1007 * FUTEX_OWNER_DIED bit. See [4]
1009 * [10] There is no transient state which leaves owner and user space
1013 * Serialization and lifetime rules:
1017 * hb -> futex_q, relation
1018 * futex_q -> pi_state, relation
1020 * (cannot be raw because hb can contain arbitrary amount
1023 * pi_mutex->wait_lock:
1027 * (and pi_mutex 'obviously')
1031 * p->pi_state_list -> pi_state->list, relation
1033 * pi_state->refcount:
1041 * pi_mutex->wait_lock
1047 * Validate that the existing waiter has a pi_state and sanity check
1048 * the pi_state against the user space value. If correct, attach to
1051 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1052 struct futex_pi_state
*pi_state
,
1053 struct futex_pi_state
**ps
)
1055 pid_t pid
= uval
& FUTEX_TID_MASK
;
1060 * Userspace might have messed up non-PI and PI futexes [3]
1062 if (unlikely(!pi_state
))
1066 * We get here with hb->lock held, and having found a
1067 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1068 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1069 * which in turn means that futex_lock_pi() still has a reference on
1072 * The waiter holding a reference on @pi_state also protects against
1073 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1074 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1075 * free pi_state before we can take a reference ourselves.
1077 WARN_ON(!atomic_read(&pi_state
->refcount
));
1080 * Now that we have a pi_state, we can acquire wait_lock
1081 * and do the state validation.
1083 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1086 * Since {uval, pi_state} is serialized by wait_lock, and our current
1087 * uval was read without holding it, it can have changed. Verify it
1088 * still is what we expect it to be, otherwise retry the entire
1091 if (get_futex_value_locked(&uval2
, uaddr
))
1098 * Handle the owner died case:
1100 if (uval
& FUTEX_OWNER_DIED
) {
1102 * exit_pi_state_list sets owner to NULL and wakes the
1103 * topmost waiter. The task which acquires the
1104 * pi_state->rt_mutex will fixup owner.
1106 if (!pi_state
->owner
) {
1108 * No pi state owner, but the user space TID
1109 * is not 0. Inconsistent state. [5]
1114 * Take a ref on the state and return success. [4]
1120 * If TID is 0, then either the dying owner has not
1121 * yet executed exit_pi_state_list() or some waiter
1122 * acquired the rtmutex in the pi state, but did not
1123 * yet fixup the TID in user space.
1125 * Take a ref on the state and return success. [6]
1131 * If the owner died bit is not set, then the pi_state
1132 * must have an owner. [7]
1134 if (!pi_state
->owner
)
1139 * Bail out if user space manipulated the futex value. If pi
1140 * state exists then the owner TID must be the same as the
1141 * user space TID. [9/10]
1143 if (pid
!= task_pid_vnr(pi_state
->owner
))
1147 get_pi_state(pi_state
);
1148 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1165 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1169 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1170 struct task_struct
*tsk
)
1175 * If PF_EXITPIDONE is not yet set, then try again.
1177 if (tsk
&& !(tsk
->flags
& PF_EXITPIDONE
))
1181 * Reread the user space value to handle the following situation:
1185 * sys_exit() sys_futex()
1186 * do_exit() futex_lock_pi()
1187 * futex_lock_pi_atomic()
1188 * exit_signals(tsk) No waiters:
1189 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1190 * mm_release(tsk) Set waiter bit
1191 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1192 * Set owner died attach_to_pi_owner() {
1193 * *uaddr = 0xC0000000; tsk = get_task(PID);
1194 * } if (!tsk->flags & PF_EXITING) {
1196 * tsk->flags |= PF_EXITPIDONE; } else {
1197 * if (!(tsk->flags & PF_EXITPIDONE))
1199 * return -ESRCH; <--- FAIL
1202 * Returning ESRCH unconditionally is wrong here because the
1203 * user space value has been changed by the exiting task.
1205 * The same logic applies to the case where the exiting task is
1208 if (get_futex_value_locked(&uval2
, uaddr
))
1211 /* If the user space value has changed, try again. */
1216 * The exiting task did not have a robust list, the robust list was
1217 * corrupted or the user space value in *uaddr is simply bogus.
1218 * Give up and tell user space.
1224 * Lookup the task for the TID provided from user space and attach to
1225 * it after doing proper sanity checks.
1227 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1228 struct futex_pi_state
**ps
)
1230 pid_t pid
= uval
& FUTEX_TID_MASK
;
1231 struct futex_pi_state
*pi_state
;
1232 struct task_struct
*p
;
1235 * We are the first waiter - try to look up the real owner and attach
1236 * the new pi_state to it, but bail out when TID = 0 [1]
1238 * The !pid check is paranoid. None of the call sites should end up
1239 * with pid == 0, but better safe than sorry. Let the caller retry
1243 p
= futex_find_get_task(pid
);
1245 return handle_exit_race(uaddr
, uval
, NULL
);
1247 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1253 * We need to look at the task state flags to figure out,
1254 * whether the task is exiting. To protect against the do_exit
1255 * change of the task flags, we do this protected by
1258 raw_spin_lock_irq(&p
->pi_lock
);
1259 if (unlikely(p
->flags
& PF_EXITING
)) {
1261 * The task is on the way out. When PF_EXITPIDONE is
1262 * set, we know that the task has finished the
1265 int ret
= handle_exit_race(uaddr
, uval
, p
);
1267 raw_spin_unlock_irq(&p
->pi_lock
);
1273 * No existing pi state. First waiter. [2]
1275 * This creates pi_state, we have hb->lock held, this means nothing can
1276 * observe this state, wait_lock is irrelevant.
1278 pi_state
= alloc_pi_state();
1281 * Initialize the pi_mutex in locked state and make @p
1284 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1286 /* Store the key for possible exit cleanups: */
1287 pi_state
->key
= *key
;
1289 WARN_ON(!list_empty(&pi_state
->list
));
1290 list_add(&pi_state
->list
, &p
->pi_state_list
);
1292 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1293 * because there is no concurrency as the object is not published yet.
1295 pi_state
->owner
= p
;
1296 raw_spin_unlock_irq(&p
->pi_lock
);
1305 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1306 struct futex_hash_bucket
*hb
,
1307 union futex_key
*key
, struct futex_pi_state
**ps
)
1309 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1312 * If there is a waiter on that futex, validate it and
1313 * attach to the pi_state when the validation succeeds.
1316 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1319 * We are the first waiter - try to look up the owner based on
1320 * @uval and attach to it.
1322 return attach_to_pi_owner(uaddr
, uval
, key
, ps
);
1325 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1327 u32
uninitialized_var(curval
);
1329 if (unlikely(should_fail_futex(true)))
1332 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1335 /* If user space value changed, let the caller retry */
1336 return curval
!= uval
? -EAGAIN
: 0;
1340 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1341 * @uaddr: the pi futex user address
1342 * @hb: the pi futex hash bucket
1343 * @key: the futex key associated with uaddr and hb
1344 * @ps: the pi_state pointer where we store the result of the
1346 * @task: the task to perform the atomic lock work for. This will
1347 * be "current" except in the case of requeue pi.
1348 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1351 * - 0 - ready to wait;
1352 * - 1 - acquired the lock;
1355 * The hb->lock and futex_key refs shall be held by the caller.
1357 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1358 union futex_key
*key
,
1359 struct futex_pi_state
**ps
,
1360 struct task_struct
*task
, int set_waiters
)
1362 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1363 struct futex_q
*top_waiter
;
1367 * Read the user space value first so we can validate a few
1368 * things before proceeding further.
1370 if (get_futex_value_locked(&uval
, uaddr
))
1373 if (unlikely(should_fail_futex(true)))
1379 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1382 if ((unlikely(should_fail_futex(true))))
1386 * Lookup existing state first. If it exists, try to attach to
1389 top_waiter
= futex_top_waiter(hb
, key
);
1391 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1394 * No waiter and user TID is 0. We are here because the
1395 * waiters or the owner died bit is set or called from
1396 * requeue_cmp_pi or for whatever reason something took the
1399 if (!(uval
& FUTEX_TID_MASK
)) {
1401 * We take over the futex. No other waiters and the user space
1402 * TID is 0. We preserve the owner died bit.
1404 newval
= uval
& FUTEX_OWNER_DIED
;
1407 /* The futex requeue_pi code can enforce the waiters bit */
1409 newval
|= FUTEX_WAITERS
;
1411 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1412 /* If the take over worked, return 1 */
1413 return ret
< 0 ? ret
: 1;
1417 * First waiter. Set the waiters bit before attaching ourself to
1418 * the owner. If owner tries to unlock, it will be forced into
1419 * the kernel and blocked on hb->lock.
1421 newval
= uval
| FUTEX_WAITERS
;
1422 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1426 * If the update of the user space value succeeded, we try to
1427 * attach to the owner. If that fails, no harm done, we only
1428 * set the FUTEX_WAITERS bit in the user space variable.
1430 return attach_to_pi_owner(uaddr
, newval
, key
, ps
);
1434 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1435 * @q: The futex_q to unqueue
1437 * The q->lock_ptr must not be NULL and must be held by the caller.
1439 static void __unqueue_futex(struct futex_q
*q
)
1441 struct futex_hash_bucket
*hb
;
1443 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1444 || WARN_ON(plist_node_empty(&q
->list
)))
1447 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1448 plist_del(&q
->list
, &hb
->chain
);
1453 * The hash bucket lock must be held when this is called.
1454 * Afterwards, the futex_q must not be accessed. Callers
1455 * must ensure to later call wake_up_q() for the actual
1458 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1460 struct task_struct
*p
= q
->task
;
1462 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1466 * Queue the task for later wakeup for after we've released
1467 * the hb->lock. wake_q_add() grabs reference to p.
1469 wake_q_add(wake_q
, p
);
1472 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1473 * is written, without taking any locks. This is possible in the event
1474 * of a spurious wakeup, for example. A memory barrier is required here
1475 * to prevent the following store to lock_ptr from getting ahead of the
1476 * plist_del in __unqueue_futex().
1478 smp_store_release(&q
->lock_ptr
, NULL
);
1482 * Caller must hold a reference on @pi_state.
1484 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1486 u32
uninitialized_var(curval
), newval
;
1487 struct task_struct
*new_owner
;
1488 bool postunlock
= false;
1489 DEFINE_WAKE_Q(wake_q
);
1492 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1493 if (WARN_ON_ONCE(!new_owner
)) {
1495 * As per the comment in futex_unlock_pi() this should not happen.
1497 * When this happens, give up our locks and try again, giving
1498 * the futex_lock_pi() instance time to complete, either by
1499 * waiting on the rtmutex or removing itself from the futex
1507 * We pass it to the next owner. The WAITERS bit is always kept
1508 * enabled while there is PI state around. We cleanup the owner
1509 * died bit, because we are the owner.
1511 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1513 if (unlikely(should_fail_futex(true)))
1516 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)) {
1519 } else if (curval
!= uval
) {
1521 * If a unconditional UNLOCK_PI operation (user space did not
1522 * try the TID->0 transition) raced with a waiter setting the
1523 * FUTEX_WAITERS flag between get_user() and locking the hash
1524 * bucket lock, retry the operation.
1526 if ((FUTEX_TID_MASK
& curval
) == uval
)
1536 * This is a point of no return; once we modify the uval there is no
1537 * going back and subsequent operations must not fail.
1540 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1541 WARN_ON(list_empty(&pi_state
->list
));
1542 list_del_init(&pi_state
->list
);
1543 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1545 raw_spin_lock(&new_owner
->pi_lock
);
1546 WARN_ON(!list_empty(&pi_state
->list
));
1547 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1548 pi_state
->owner
= new_owner
;
1549 raw_spin_unlock(&new_owner
->pi_lock
);
1551 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1554 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1557 rt_mutex_postunlock(&wake_q
);
1563 * Express the locking dependencies for lockdep:
1566 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1569 spin_lock(&hb1
->lock
);
1571 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1572 } else { /* hb1 > hb2 */
1573 spin_lock(&hb2
->lock
);
1574 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1579 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1581 spin_unlock(&hb1
->lock
);
1583 spin_unlock(&hb2
->lock
);
1587 * Wake up waiters matching bitset queued on this futex (uaddr).
1590 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1592 struct futex_hash_bucket
*hb
;
1593 struct futex_q
*this, *next
;
1594 union futex_key key
= FUTEX_KEY_INIT
;
1596 DEFINE_WAKE_Q(wake_q
);
1601 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1602 if (unlikely(ret
!= 0))
1605 hb
= hash_futex(&key
);
1607 /* Make sure we really have tasks to wakeup */
1608 if (!hb_waiters_pending(hb
))
1611 spin_lock(&hb
->lock
);
1613 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1614 if (match_futex (&this->key
, &key
)) {
1615 if (this->pi_state
|| this->rt_waiter
) {
1620 /* Check if one of the bits is set in both bitsets */
1621 if (!(this->bitset
& bitset
))
1624 mark_wake_futex(&wake_q
, this);
1625 if (++ret
>= nr_wake
)
1630 spin_unlock(&hb
->lock
);
1633 put_futex_key(&key
);
1638 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1640 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1641 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1642 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1643 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1646 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1647 if (oparg
< 0 || oparg
> 31) {
1648 char comm
[sizeof(current
->comm
)];
1650 * kill this print and return -EINVAL when userspace
1653 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1654 get_task_comm(comm
, current
), oparg
);
1660 if (!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
)))
1663 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1668 case FUTEX_OP_CMP_EQ
:
1669 return oldval
== cmparg
;
1670 case FUTEX_OP_CMP_NE
:
1671 return oldval
!= cmparg
;
1672 case FUTEX_OP_CMP_LT
:
1673 return oldval
< cmparg
;
1674 case FUTEX_OP_CMP_GE
:
1675 return oldval
>= cmparg
;
1676 case FUTEX_OP_CMP_LE
:
1677 return oldval
<= cmparg
;
1678 case FUTEX_OP_CMP_GT
:
1679 return oldval
> cmparg
;
1686 * Wake up all waiters hashed on the physical page that is mapped
1687 * to this virtual address:
1690 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1691 int nr_wake
, int nr_wake2
, int op
)
1693 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1694 struct futex_hash_bucket
*hb1
, *hb2
;
1695 struct futex_q
*this, *next
;
1697 DEFINE_WAKE_Q(wake_q
);
1700 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1701 if (unlikely(ret
!= 0))
1703 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1704 if (unlikely(ret
!= 0))
1707 hb1
= hash_futex(&key1
);
1708 hb2
= hash_futex(&key2
);
1711 double_lock_hb(hb1
, hb2
);
1712 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1713 if (unlikely(op_ret
< 0)) {
1715 double_unlock_hb(hb1
, hb2
);
1719 * we don't get EFAULT from MMU faults if we don't have an MMU,
1720 * but we might get them from range checking
1726 if (unlikely(op_ret
!= -EFAULT
)) {
1731 ret
= fault_in_user_writeable(uaddr2
);
1735 if (!(flags
& FLAGS_SHARED
))
1738 put_futex_key(&key2
);
1739 put_futex_key(&key1
);
1743 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1744 if (match_futex (&this->key
, &key1
)) {
1745 if (this->pi_state
|| this->rt_waiter
) {
1749 mark_wake_futex(&wake_q
, this);
1750 if (++ret
>= nr_wake
)
1757 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1758 if (match_futex (&this->key
, &key2
)) {
1759 if (this->pi_state
|| this->rt_waiter
) {
1763 mark_wake_futex(&wake_q
, this);
1764 if (++op_ret
>= nr_wake2
)
1772 double_unlock_hb(hb1
, hb2
);
1775 put_futex_key(&key2
);
1777 put_futex_key(&key1
);
1783 * requeue_futex() - Requeue a futex_q from one hb to another
1784 * @q: the futex_q to requeue
1785 * @hb1: the source hash_bucket
1786 * @hb2: the target hash_bucket
1787 * @key2: the new key for the requeued futex_q
1790 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1791 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1795 * If key1 and key2 hash to the same bucket, no need to
1798 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1799 plist_del(&q
->list
, &hb1
->chain
);
1800 hb_waiters_dec(hb1
);
1801 hb_waiters_inc(hb2
);
1802 plist_add(&q
->list
, &hb2
->chain
);
1803 q
->lock_ptr
= &hb2
->lock
;
1805 get_futex_key_refs(key2
);
1810 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1812 * @key: the key of the requeue target futex
1813 * @hb: the hash_bucket of the requeue target futex
1815 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1816 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1817 * to the requeue target futex so the waiter can detect the wakeup on the right
1818 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1819 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1820 * to protect access to the pi_state to fixup the owner later. Must be called
1821 * with both q->lock_ptr and hb->lock held.
1824 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1825 struct futex_hash_bucket
*hb
)
1827 get_futex_key_refs(key
);
1832 WARN_ON(!q
->rt_waiter
);
1833 q
->rt_waiter
= NULL
;
1835 q
->lock_ptr
= &hb
->lock
;
1837 wake_up_state(q
->task
, TASK_NORMAL
);
1841 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1842 * @pifutex: the user address of the to futex
1843 * @hb1: the from futex hash bucket, must be locked by the caller
1844 * @hb2: the to futex hash bucket, must be locked by the caller
1845 * @key1: the from futex key
1846 * @key2: the to futex key
1847 * @ps: address to store the pi_state pointer
1848 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1850 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1851 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1852 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1853 * hb1 and hb2 must be held by the caller.
1856 * - 0 - failed to acquire the lock atomically;
1857 * - >0 - acquired the lock, return value is vpid of the top_waiter
1860 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1861 struct futex_hash_bucket
*hb1
,
1862 struct futex_hash_bucket
*hb2
,
1863 union futex_key
*key1
, union futex_key
*key2
,
1864 struct futex_pi_state
**ps
, int set_waiters
)
1866 struct futex_q
*top_waiter
= NULL
;
1870 if (get_futex_value_locked(&curval
, pifutex
))
1873 if (unlikely(should_fail_futex(true)))
1877 * Find the top_waiter and determine if there are additional waiters.
1878 * If the caller intends to requeue more than 1 waiter to pifutex,
1879 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1880 * as we have means to handle the possible fault. If not, don't set
1881 * the bit unecessarily as it will force the subsequent unlock to enter
1884 top_waiter
= futex_top_waiter(hb1
, key1
);
1886 /* There are no waiters, nothing for us to do. */
1890 /* Ensure we requeue to the expected futex. */
1891 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1895 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1896 * the contended case or if set_waiters is 1. The pi_state is returned
1897 * in ps in contended cases.
1899 vpid
= task_pid_vnr(top_waiter
->task
);
1900 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1903 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1910 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1911 * @uaddr1: source futex user address
1912 * @flags: futex flags (FLAGS_SHARED, etc.)
1913 * @uaddr2: target futex user address
1914 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1915 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1916 * @cmpval: @uaddr1 expected value (or %NULL)
1917 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1918 * pi futex (pi to pi requeue is not supported)
1920 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1921 * uaddr2 atomically on behalf of the top waiter.
1924 * - >=0 - on success, the number of tasks requeued or woken;
1927 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1928 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1929 u32
*cmpval
, int requeue_pi
)
1931 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1932 int drop_count
= 0, task_count
= 0, ret
;
1933 struct futex_pi_state
*pi_state
= NULL
;
1934 struct futex_hash_bucket
*hb1
, *hb2
;
1935 struct futex_q
*this, *next
;
1936 DEFINE_WAKE_Q(wake_q
);
1938 if (nr_wake
< 0 || nr_requeue
< 0)
1942 * When PI not supported: return -ENOSYS if requeue_pi is true,
1943 * consequently the compiler knows requeue_pi is always false past
1944 * this point which will optimize away all the conditional code
1947 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1952 * Requeue PI only works on two distinct uaddrs. This
1953 * check is only valid for private futexes. See below.
1955 if (uaddr1
== uaddr2
)
1959 * requeue_pi requires a pi_state, try to allocate it now
1960 * without any locks in case it fails.
1962 if (refill_pi_state_cache())
1965 * requeue_pi must wake as many tasks as it can, up to nr_wake
1966 * + nr_requeue, since it acquires the rt_mutex prior to
1967 * returning to userspace, so as to not leave the rt_mutex with
1968 * waiters and no owner. However, second and third wake-ups
1969 * cannot be predicted as they involve race conditions with the
1970 * first wake and a fault while looking up the pi_state. Both
1971 * pthread_cond_signal() and pthread_cond_broadcast() should
1979 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1980 if (unlikely(ret
!= 0))
1982 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1983 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1984 if (unlikely(ret
!= 0))
1988 * The check above which compares uaddrs is not sufficient for
1989 * shared futexes. We need to compare the keys:
1991 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1996 hb1
= hash_futex(&key1
);
1997 hb2
= hash_futex(&key2
);
2000 hb_waiters_inc(hb2
);
2001 double_lock_hb(hb1
, hb2
);
2003 if (likely(cmpval
!= NULL
)) {
2006 ret
= get_futex_value_locked(&curval
, uaddr1
);
2008 if (unlikely(ret
)) {
2009 double_unlock_hb(hb1
, hb2
);
2010 hb_waiters_dec(hb2
);
2012 ret
= get_user(curval
, uaddr1
);
2016 if (!(flags
& FLAGS_SHARED
))
2019 put_futex_key(&key2
);
2020 put_futex_key(&key1
);
2023 if (curval
!= *cmpval
) {
2029 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2031 * Attempt to acquire uaddr2 and wake the top waiter. If we
2032 * intend to requeue waiters, force setting the FUTEX_WAITERS
2033 * bit. We force this here where we are able to easily handle
2034 * faults rather in the requeue loop below.
2036 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2037 &key2
, &pi_state
, nr_requeue
);
2040 * At this point the top_waiter has either taken uaddr2 or is
2041 * waiting on it. If the former, then the pi_state will not
2042 * exist yet, look it up one more time to ensure we have a
2043 * reference to it. If the lock was taken, ret contains the
2044 * vpid of the top waiter task.
2045 * If the lock was not taken, we have pi_state and an initial
2046 * refcount on it. In case of an error we have nothing.
2053 * If we acquired the lock, then the user space value
2054 * of uaddr2 should be vpid. It cannot be changed by
2055 * the top waiter as it is blocked on hb2 lock if it
2056 * tries to do so. If something fiddled with it behind
2057 * our back the pi state lookup might unearth it. So
2058 * we rather use the known value than rereading and
2059 * handing potential crap to lookup_pi_state.
2061 * If that call succeeds then we have pi_state and an
2062 * initial refcount on it.
2064 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
, &pi_state
);
2069 /* We hold a reference on the pi state. */
2072 /* If the above failed, then pi_state is NULL */
2074 double_unlock_hb(hb1
, hb2
);
2075 hb_waiters_dec(hb2
);
2076 put_futex_key(&key2
);
2077 put_futex_key(&key1
);
2078 ret
= fault_in_user_writeable(uaddr2
);
2084 * Two reasons for this:
2085 * - Owner is exiting and we just wait for the
2087 * - The user space value changed.
2089 double_unlock_hb(hb1
, hb2
);
2090 hb_waiters_dec(hb2
);
2091 put_futex_key(&key2
);
2092 put_futex_key(&key1
);
2100 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2101 if (task_count
- nr_wake
>= nr_requeue
)
2104 if (!match_futex(&this->key
, &key1
))
2108 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2109 * be paired with each other and no other futex ops.
2111 * We should never be requeueing a futex_q with a pi_state,
2112 * which is awaiting a futex_unlock_pi().
2114 if ((requeue_pi
&& !this->rt_waiter
) ||
2115 (!requeue_pi
&& this->rt_waiter
) ||
2122 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2123 * lock, we already woke the top_waiter. If not, it will be
2124 * woken by futex_unlock_pi().
2126 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2127 mark_wake_futex(&wake_q
, this);
2131 /* Ensure we requeue to the expected futex for requeue_pi. */
2132 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2138 * Requeue nr_requeue waiters and possibly one more in the case
2139 * of requeue_pi if we couldn't acquire the lock atomically.
2143 * Prepare the waiter to take the rt_mutex. Take a
2144 * refcount on the pi_state and store the pointer in
2145 * the futex_q object of the waiter.
2147 get_pi_state(pi_state
);
2148 this->pi_state
= pi_state
;
2149 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2154 * We got the lock. We do neither drop the
2155 * refcount on pi_state nor clear
2156 * this->pi_state because the waiter needs the
2157 * pi_state for cleaning up the user space
2158 * value. It will drop the refcount after
2161 requeue_pi_wake_futex(this, &key2
, hb2
);
2166 * rt_mutex_start_proxy_lock() detected a
2167 * potential deadlock when we tried to queue
2168 * that waiter. Drop the pi_state reference
2169 * which we took above and remove the pointer
2170 * to the state from the waiters futex_q
2173 this->pi_state
= NULL
;
2174 put_pi_state(pi_state
);
2176 * We stop queueing more waiters and let user
2177 * space deal with the mess.
2182 requeue_futex(this, hb1
, hb2
, &key2
);
2187 * We took an extra initial reference to the pi_state either
2188 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2189 * need to drop it here again.
2191 put_pi_state(pi_state
);
2194 double_unlock_hb(hb1
, hb2
);
2196 hb_waiters_dec(hb2
);
2199 * drop_futex_key_refs() must be called outside the spinlocks. During
2200 * the requeue we moved futex_q's from the hash bucket at key1 to the
2201 * one at key2 and updated their key pointer. We no longer need to
2202 * hold the references to key1.
2204 while (--drop_count
>= 0)
2205 drop_futex_key_refs(&key1
);
2208 put_futex_key(&key2
);
2210 put_futex_key(&key1
);
2212 return ret
? ret
: task_count
;
2215 /* The key must be already stored in q->key. */
2216 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2217 __acquires(&hb
->lock
)
2219 struct futex_hash_bucket
*hb
;
2221 hb
= hash_futex(&q
->key
);
2224 * Increment the counter before taking the lock so that
2225 * a potential waker won't miss a to-be-slept task that is
2226 * waiting for the spinlock. This is safe as all queue_lock()
2227 * users end up calling queue_me(). Similarly, for housekeeping,
2228 * decrement the counter at queue_unlock() when some error has
2229 * occurred and we don't end up adding the task to the list.
2233 q
->lock_ptr
= &hb
->lock
;
2235 spin_lock(&hb
->lock
); /* implies smp_mb(); (A) */
2240 queue_unlock(struct futex_hash_bucket
*hb
)
2241 __releases(&hb
->lock
)
2243 spin_unlock(&hb
->lock
);
2247 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2252 * The priority used to register this element is
2253 * - either the real thread-priority for the real-time threads
2254 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2255 * - or MAX_RT_PRIO for non-RT threads.
2256 * Thus, all RT-threads are woken first in priority order, and
2257 * the others are woken last, in FIFO order.
2259 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2261 plist_node_init(&q
->list
, prio
);
2262 plist_add(&q
->list
, &hb
->chain
);
2267 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2268 * @q: The futex_q to enqueue
2269 * @hb: The destination hash bucket
2271 * The hb->lock must be held by the caller, and is released here. A call to
2272 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2273 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2274 * or nothing if the unqueue is done as part of the wake process and the unqueue
2275 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2278 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2279 __releases(&hb
->lock
)
2282 spin_unlock(&hb
->lock
);
2286 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2287 * @q: The futex_q to unqueue
2289 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2290 * be paired with exactly one earlier call to queue_me().
2293 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2294 * - 0 - if the futex_q was already removed by the waking thread
2296 static int unqueue_me(struct futex_q
*q
)
2298 spinlock_t
*lock_ptr
;
2301 /* In the common case we don't take the spinlock, which is nice. */
2304 * q->lock_ptr can change between this read and the following spin_lock.
2305 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2306 * optimizing lock_ptr out of the logic below.
2308 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2309 if (lock_ptr
!= NULL
) {
2310 spin_lock(lock_ptr
);
2312 * q->lock_ptr can change between reading it and
2313 * spin_lock(), causing us to take the wrong lock. This
2314 * corrects the race condition.
2316 * Reasoning goes like this: if we have the wrong lock,
2317 * q->lock_ptr must have changed (maybe several times)
2318 * between reading it and the spin_lock(). It can
2319 * change again after the spin_lock() but only if it was
2320 * already changed before the spin_lock(). It cannot,
2321 * however, change back to the original value. Therefore
2322 * we can detect whether we acquired the correct lock.
2324 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2325 spin_unlock(lock_ptr
);
2330 BUG_ON(q
->pi_state
);
2332 spin_unlock(lock_ptr
);
2336 drop_futex_key_refs(&q
->key
);
2341 * PI futexes can not be requeued and must remove themself from the
2342 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2345 static void unqueue_me_pi(struct futex_q
*q
)
2346 __releases(q
->lock_ptr
)
2350 BUG_ON(!q
->pi_state
);
2351 put_pi_state(q
->pi_state
);
2354 spin_unlock(q
->lock_ptr
);
2357 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2358 struct task_struct
*argowner
)
2360 struct futex_pi_state
*pi_state
= q
->pi_state
;
2361 u32 uval
, uninitialized_var(curval
), newval
;
2362 struct task_struct
*oldowner
, *newowner
;
2366 lockdep_assert_held(q
->lock_ptr
);
2368 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2370 oldowner
= pi_state
->owner
;
2373 * We are here because either:
2375 * - we stole the lock and pi_state->owner needs updating to reflect
2376 * that (@argowner == current),
2380 * - someone stole our lock and we need to fix things to point to the
2381 * new owner (@argowner == NULL).
2383 * Either way, we have to replace the TID in the user space variable.
2384 * This must be atomic as we have to preserve the owner died bit here.
2386 * Note: We write the user space value _before_ changing the pi_state
2387 * because we can fault here. Imagine swapped out pages or a fork
2388 * that marked all the anonymous memory readonly for cow.
2390 * Modifying pi_state _before_ the user space value would leave the
2391 * pi_state in an inconsistent state when we fault here, because we
2392 * need to drop the locks to handle the fault. This might be observed
2393 * in the PID check in lookup_pi_state.
2397 if (oldowner
!= current
) {
2399 * We raced against a concurrent self; things are
2400 * already fixed up. Nothing to do.
2406 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2407 /* We got the lock after all, nothing to fix. */
2413 * Since we just failed the trylock; there must be an owner.
2415 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2418 WARN_ON_ONCE(argowner
!= current
);
2419 if (oldowner
== current
) {
2421 * We raced against a concurrent self; things are
2422 * already fixed up. Nothing to do.
2427 newowner
= argowner
;
2430 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2432 if (!pi_state
->owner
)
2433 newtid
|= FUTEX_OWNER_DIED
;
2435 if (get_futex_value_locked(&uval
, uaddr
))
2439 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2441 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
2449 * We fixed up user space. Now we need to fix the pi_state
2452 if (pi_state
->owner
!= NULL
) {
2453 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2454 WARN_ON(list_empty(&pi_state
->list
));
2455 list_del_init(&pi_state
->list
);
2456 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2459 pi_state
->owner
= newowner
;
2461 raw_spin_lock(&newowner
->pi_lock
);
2462 WARN_ON(!list_empty(&pi_state
->list
));
2463 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2464 raw_spin_unlock(&newowner
->pi_lock
);
2465 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2470 * To handle the page fault we need to drop the locks here. That gives
2471 * the other task (either the highest priority waiter itself or the
2472 * task which stole the rtmutex) the chance to try the fixup of the
2473 * pi_state. So once we are back from handling the fault we need to
2474 * check the pi_state after reacquiring the locks and before trying to
2475 * do another fixup. When the fixup has been done already we simply
2478 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2479 * drop hb->lock since the caller owns the hb -> futex_q relation.
2480 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2483 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2484 spin_unlock(q
->lock_ptr
);
2486 ret
= fault_in_user_writeable(uaddr
);
2488 spin_lock(q
->lock_ptr
);
2489 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2492 * Check if someone else fixed it for us:
2494 if (pi_state
->owner
!= oldowner
) {
2505 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2509 static long futex_wait_restart(struct restart_block
*restart
);
2512 * fixup_owner() - Post lock pi_state and corner case management
2513 * @uaddr: user address of the futex
2514 * @q: futex_q (contains pi_state and access to the rt_mutex)
2515 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2517 * After attempting to lock an rt_mutex, this function is called to cleanup
2518 * the pi_state owner as well as handle race conditions that may allow us to
2519 * acquire the lock. Must be called with the hb lock held.
2522 * - 1 - success, lock taken;
2523 * - 0 - success, lock not taken;
2524 * - <0 - on error (-EFAULT)
2526 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2532 * Got the lock. We might not be the anticipated owner if we
2533 * did a lock-steal - fix up the PI-state in that case:
2535 * Speculative pi_state->owner read (we don't hold wait_lock);
2536 * since we own the lock pi_state->owner == current is the
2537 * stable state, anything else needs more attention.
2539 if (q
->pi_state
->owner
!= current
)
2540 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2545 * If we didn't get the lock; check if anybody stole it from us. In
2546 * that case, we need to fix up the uval to point to them instead of
2547 * us, otherwise bad things happen. [10]
2549 * Another speculative read; pi_state->owner == current is unstable
2550 * but needs our attention.
2552 if (q
->pi_state
->owner
== current
) {
2553 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2558 * Paranoia check. If we did not take the lock, then we should not be
2559 * the owner of the rt_mutex.
2561 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2562 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2563 "pi-state %p\n", ret
,
2564 q
->pi_state
->pi_mutex
.owner
,
2565 q
->pi_state
->owner
);
2569 return ret
? ret
: locked
;
2573 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2574 * @hb: the futex hash bucket, must be locked by the caller
2575 * @q: the futex_q to queue up on
2576 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2578 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2579 struct hrtimer_sleeper
*timeout
)
2582 * The task state is guaranteed to be set before another task can
2583 * wake it. set_current_state() is implemented using smp_store_mb() and
2584 * queue_me() calls spin_unlock() upon completion, both serializing
2585 * access to the hash list and forcing another memory barrier.
2587 set_current_state(TASK_INTERRUPTIBLE
);
2592 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2595 * If we have been removed from the hash list, then another task
2596 * has tried to wake us, and we can skip the call to schedule().
2598 if (likely(!plist_node_empty(&q
->list
))) {
2600 * If the timer has already expired, current will already be
2601 * flagged for rescheduling. Only call schedule if there
2602 * is no timeout, or if it has yet to expire.
2604 if (!timeout
|| timeout
->task
)
2605 freezable_schedule();
2607 __set_current_state(TASK_RUNNING
);
2611 * futex_wait_setup() - Prepare to wait on a futex
2612 * @uaddr: the futex userspace address
2613 * @val: the expected value
2614 * @flags: futex flags (FLAGS_SHARED, etc.)
2615 * @q: the associated futex_q
2616 * @hb: storage for hash_bucket pointer to be returned to caller
2618 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2619 * compare it with the expected value. Handle atomic faults internally.
2620 * Return with the hb lock held and a q.key reference on success, and unlocked
2621 * with no q.key reference on failure.
2624 * - 0 - uaddr contains val and hb has been locked;
2625 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2627 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2628 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2634 * Access the page AFTER the hash-bucket is locked.
2635 * Order is important:
2637 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2638 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2640 * The basic logical guarantee of a futex is that it blocks ONLY
2641 * if cond(var) is known to be true at the time of blocking, for
2642 * any cond. If we locked the hash-bucket after testing *uaddr, that
2643 * would open a race condition where we could block indefinitely with
2644 * cond(var) false, which would violate the guarantee.
2646 * On the other hand, we insert q and release the hash-bucket only
2647 * after testing *uaddr. This guarantees that futex_wait() will NOT
2648 * absorb a wakeup if *uaddr does not match the desired values
2649 * while the syscall executes.
2652 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2653 if (unlikely(ret
!= 0))
2657 *hb
= queue_lock(q
);
2659 ret
= get_futex_value_locked(&uval
, uaddr
);
2664 ret
= get_user(uval
, uaddr
);
2668 if (!(flags
& FLAGS_SHARED
))
2671 put_futex_key(&q
->key
);
2682 put_futex_key(&q
->key
);
2686 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2687 ktime_t
*abs_time
, u32 bitset
)
2689 struct hrtimer_sleeper timeout
, *to
= NULL
;
2690 struct restart_block
*restart
;
2691 struct futex_hash_bucket
*hb
;
2692 struct futex_q q
= futex_q_init
;
2702 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2703 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2705 hrtimer_init_sleeper(to
, current
);
2706 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2707 current
->timer_slack_ns
);
2712 * Prepare to wait on uaddr. On success, holds hb lock and increments
2715 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2719 /* queue_me and wait for wakeup, timeout, or a signal. */
2720 futex_wait_queue_me(hb
, &q
, to
);
2722 /* If we were woken (and unqueued), we succeeded, whatever. */
2724 /* unqueue_me() drops q.key ref */
2725 if (!unqueue_me(&q
))
2728 if (to
&& !to
->task
)
2732 * We expect signal_pending(current), but we might be the
2733 * victim of a spurious wakeup as well.
2735 if (!signal_pending(current
))
2742 restart
= ¤t
->restart_block
;
2743 restart
->fn
= futex_wait_restart
;
2744 restart
->futex
.uaddr
= uaddr
;
2745 restart
->futex
.val
= val
;
2746 restart
->futex
.time
= *abs_time
;
2747 restart
->futex
.bitset
= bitset
;
2748 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2750 ret
= -ERESTART_RESTARTBLOCK
;
2754 hrtimer_cancel(&to
->timer
);
2755 destroy_hrtimer_on_stack(&to
->timer
);
2761 static long futex_wait_restart(struct restart_block
*restart
)
2763 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2764 ktime_t t
, *tp
= NULL
;
2766 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2767 t
= restart
->futex
.time
;
2770 restart
->fn
= do_no_restart_syscall
;
2772 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2773 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2778 * Userspace tried a 0 -> TID atomic transition of the futex value
2779 * and failed. The kernel side here does the whole locking operation:
2780 * if there are waiters then it will block as a consequence of relying
2781 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2782 * a 0 value of the futex too.).
2784 * Also serves as futex trylock_pi()'ing, and due semantics.
2786 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2787 ktime_t
*time
, int trylock
)
2789 struct hrtimer_sleeper timeout
, *to
= NULL
;
2790 struct futex_pi_state
*pi_state
= NULL
;
2791 struct rt_mutex_waiter rt_waiter
;
2792 struct futex_hash_bucket
*hb
;
2793 struct futex_q q
= futex_q_init
;
2796 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2799 if (refill_pi_state_cache())
2804 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2806 hrtimer_init_sleeper(to
, current
);
2807 hrtimer_set_expires(&to
->timer
, *time
);
2811 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2812 if (unlikely(ret
!= 0))
2816 hb
= queue_lock(&q
);
2818 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2819 if (unlikely(ret
)) {
2821 * Atomic work succeeded and we got the lock,
2822 * or failed. Either way, we do _not_ block.
2826 /* We got the lock. */
2828 goto out_unlock_put_key
;
2833 * Two reasons for this:
2834 * - Task is exiting and we just wait for the
2836 * - The user space value changed.
2839 put_futex_key(&q
.key
);
2843 goto out_unlock_put_key
;
2847 WARN_ON(!q
.pi_state
);
2850 * Only actually queue now that the atomic ops are done:
2855 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2856 /* Fixup the trylock return value: */
2857 ret
= ret
? 0 : -EWOULDBLOCK
;
2861 rt_mutex_init_waiter(&rt_waiter
);
2864 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2865 * hold it while doing rt_mutex_start_proxy(), because then it will
2866 * include hb->lock in the blocking chain, even through we'll not in
2867 * fact hold it while blocking. This will lead it to report -EDEADLK
2868 * and BUG when futex_unlock_pi() interleaves with this.
2870 * Therefore acquire wait_lock while holding hb->lock, but drop the
2871 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2872 * serializes against futex_unlock_pi() as that does the exact same
2873 * lock handoff sequence.
2875 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2876 spin_unlock(q
.lock_ptr
);
2877 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2878 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2884 spin_lock(q
.lock_ptr
);
2890 hrtimer_start_expires(&to
->timer
, HRTIMER_MODE_ABS
);
2892 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2894 spin_lock(q
.lock_ptr
);
2896 * If we failed to acquire the lock (signal/timeout), we must
2897 * first acquire the hb->lock before removing the lock from the
2898 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2899 * wait lists consistent.
2901 * In particular; it is important that futex_unlock_pi() can not
2902 * observe this inconsistency.
2904 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2909 * Fixup the pi_state owner and possibly acquire the lock if we
2912 res
= fixup_owner(uaddr
, &q
, !ret
);
2914 * If fixup_owner() returned an error, proprogate that. If it acquired
2915 * the lock, clear our -ETIMEDOUT or -EINTR.
2918 ret
= (res
< 0) ? res
: 0;
2921 * If fixup_owner() faulted and was unable to handle the fault, unlock
2922 * it and return the fault to userspace.
2924 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2925 pi_state
= q
.pi_state
;
2926 get_pi_state(pi_state
);
2929 /* Unqueue and drop the lock */
2933 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2934 put_pi_state(pi_state
);
2943 put_futex_key(&q
.key
);
2946 hrtimer_cancel(&to
->timer
);
2947 destroy_hrtimer_on_stack(&to
->timer
);
2949 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2954 ret
= fault_in_user_writeable(uaddr
);
2958 if (!(flags
& FLAGS_SHARED
))
2961 put_futex_key(&q
.key
);
2966 * Userspace attempted a TID -> 0 atomic transition, and failed.
2967 * This is the in-kernel slowpath: we look up the PI state (if any),
2968 * and do the rt-mutex unlock.
2970 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2972 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2973 union futex_key key
= FUTEX_KEY_INIT
;
2974 struct futex_hash_bucket
*hb
;
2975 struct futex_q
*top_waiter
;
2978 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2982 if (get_user(uval
, uaddr
))
2985 * We release only a lock we actually own:
2987 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2990 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2994 hb
= hash_futex(&key
);
2995 spin_lock(&hb
->lock
);
2998 * Check waiters first. We do not trust user space values at
2999 * all and we at least want to know if user space fiddled
3000 * with the futex value instead of blindly unlocking.
3002 top_waiter
= futex_top_waiter(hb
, &key
);
3004 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3011 * If current does not own the pi_state then the futex is
3012 * inconsistent and user space fiddled with the futex value.
3014 if (pi_state
->owner
!= current
)
3017 get_pi_state(pi_state
);
3019 * By taking wait_lock while still holding hb->lock, we ensure
3020 * there is no point where we hold neither; and therefore
3021 * wake_futex_pi() must observe a state consistent with what we
3024 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3025 spin_unlock(&hb
->lock
);
3027 /* drops pi_state->pi_mutex.wait_lock */
3028 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3030 put_pi_state(pi_state
);
3033 * Success, we're done! No tricky corner cases.
3038 * The atomic access to the futex value generated a
3039 * pagefault, so retry the user-access and the wakeup:
3044 * A unconditional UNLOCK_PI op raced against a waiter
3045 * setting the FUTEX_WAITERS bit. Try again.
3047 if (ret
== -EAGAIN
) {
3048 put_futex_key(&key
);
3052 * wake_futex_pi has detected invalid state. Tell user
3059 * We have no kernel internal state, i.e. no waiters in the
3060 * kernel. Waiters which are about to queue themselves are stuck
3061 * on hb->lock. So we can safely ignore them. We do neither
3062 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3065 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0)) {
3066 spin_unlock(&hb
->lock
);
3071 * If uval has changed, let user space handle it.
3073 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3076 spin_unlock(&hb
->lock
);
3078 put_futex_key(&key
);
3082 put_futex_key(&key
);
3084 ret
= fault_in_user_writeable(uaddr
);
3092 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3093 * @hb: the hash_bucket futex_q was original enqueued on
3094 * @q: the futex_q woken while waiting to be requeued
3095 * @key2: the futex_key of the requeue target futex
3096 * @timeout: the timeout associated with the wait (NULL if none)
3098 * Detect if the task was woken on the initial futex as opposed to the requeue
3099 * target futex. If so, determine if it was a timeout or a signal that caused
3100 * the wakeup and return the appropriate error code to the caller. Must be
3101 * called with the hb lock held.
3104 * - 0 = no early wakeup detected;
3105 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3108 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3109 struct futex_q
*q
, union futex_key
*key2
,
3110 struct hrtimer_sleeper
*timeout
)
3115 * With the hb lock held, we avoid races while we process the wakeup.
3116 * We only need to hold hb (and not hb2) to ensure atomicity as the
3117 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3118 * It can't be requeued from uaddr2 to something else since we don't
3119 * support a PI aware source futex for requeue.
3121 if (!match_futex(&q
->key
, key2
)) {
3122 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3124 * We were woken prior to requeue by a timeout or a signal.
3125 * Unqueue the futex_q and determine which it was.
3127 plist_del(&q
->list
, &hb
->chain
);
3130 /* Handle spurious wakeups gracefully */
3132 if (timeout
&& !timeout
->task
)
3134 else if (signal_pending(current
))
3135 ret
= -ERESTARTNOINTR
;
3141 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3142 * @uaddr: the futex we initially wait on (non-pi)
3143 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3144 * the same type, no requeueing from private to shared, etc.
3145 * @val: the expected value of uaddr
3146 * @abs_time: absolute timeout
3147 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3148 * @uaddr2: the pi futex we will take prior to returning to user-space
3150 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3151 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3152 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3153 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3154 * without one, the pi logic would not know which task to boost/deboost, if
3155 * there was a need to.
3157 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3158 * via the following--
3159 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3160 * 2) wakeup on uaddr2 after a requeue
3164 * If 3, cleanup and return -ERESTARTNOINTR.
3166 * If 2, we may then block on trying to take the rt_mutex and return via:
3167 * 5) successful lock
3170 * 8) other lock acquisition failure
3172 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3174 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3180 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3181 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3184 struct hrtimer_sleeper timeout
, *to
= NULL
;
3185 struct futex_pi_state
*pi_state
= NULL
;
3186 struct rt_mutex_waiter rt_waiter
;
3187 struct futex_hash_bucket
*hb
;
3188 union futex_key key2
= FUTEX_KEY_INIT
;
3189 struct futex_q q
= futex_q_init
;
3192 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3195 if (uaddr
== uaddr2
)
3203 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
3204 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
3206 hrtimer_init_sleeper(to
, current
);
3207 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
3208 current
->timer_slack_ns
);
3212 * The waiter is allocated on our stack, manipulated by the requeue
3213 * code while we sleep on uaddr.
3215 rt_mutex_init_waiter(&rt_waiter
);
3217 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
3218 if (unlikely(ret
!= 0))
3222 q
.rt_waiter
= &rt_waiter
;
3223 q
.requeue_pi_key
= &key2
;
3226 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3229 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3234 * The check above which compares uaddrs is not sufficient for
3235 * shared futexes. We need to compare the keys:
3237 if (match_futex(&q
.key
, &key2
)) {
3243 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3244 futex_wait_queue_me(hb
, &q
, to
);
3246 spin_lock(&hb
->lock
);
3247 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3248 spin_unlock(&hb
->lock
);
3253 * In order for us to be here, we know our q.key == key2, and since
3254 * we took the hb->lock above, we also know that futex_requeue() has
3255 * completed and we no longer have to concern ourselves with a wakeup
3256 * race with the atomic proxy lock acquisition by the requeue code. The
3257 * futex_requeue dropped our key1 reference and incremented our key2
3261 /* Check if the requeue code acquired the second futex for us. */
3264 * Got the lock. We might not be the anticipated owner if we
3265 * did a lock-steal - fix up the PI-state in that case.
3267 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3268 spin_lock(q
.lock_ptr
);
3269 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3270 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3271 pi_state
= q
.pi_state
;
3272 get_pi_state(pi_state
);
3275 * Drop the reference to the pi state which
3276 * the requeue_pi() code acquired for us.
3278 put_pi_state(q
.pi_state
);
3279 spin_unlock(q
.lock_ptr
);
3282 struct rt_mutex
*pi_mutex
;
3285 * We have been woken up by futex_unlock_pi(), a timeout, or a
3286 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3289 WARN_ON(!q
.pi_state
);
3290 pi_mutex
= &q
.pi_state
->pi_mutex
;
3291 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3293 spin_lock(q
.lock_ptr
);
3294 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3297 debug_rt_mutex_free_waiter(&rt_waiter
);
3299 * Fixup the pi_state owner and possibly acquire the lock if we
3302 res
= fixup_owner(uaddr2
, &q
, !ret
);
3304 * If fixup_owner() returned an error, proprogate that. If it
3305 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3308 ret
= (res
< 0) ? res
: 0;
3311 * If fixup_pi_state_owner() faulted and was unable to handle
3312 * the fault, unlock the rt_mutex and return the fault to
3315 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3316 pi_state
= q
.pi_state
;
3317 get_pi_state(pi_state
);
3320 /* Unqueue and drop the lock. */
3325 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3326 put_pi_state(pi_state
);
3329 if (ret
== -EINTR
) {
3331 * We've already been requeued, but cannot restart by calling
3332 * futex_lock_pi() directly. We could restart this syscall, but
3333 * it would detect that the user space "val" changed and return
3334 * -EWOULDBLOCK. Save the overhead of the restart and return
3335 * -EWOULDBLOCK directly.
3341 put_futex_key(&q
.key
);
3343 put_futex_key(&key2
);
3347 hrtimer_cancel(&to
->timer
);
3348 destroy_hrtimer_on_stack(&to
->timer
);
3354 * Support for robust futexes: the kernel cleans up held futexes at
3357 * Implementation: user-space maintains a per-thread list of locks it
3358 * is holding. Upon do_exit(), the kernel carefully walks this list,
3359 * and marks all locks that are owned by this thread with the
3360 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3361 * always manipulated with the lock held, so the list is private and
3362 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3363 * field, to allow the kernel to clean up if the thread dies after
3364 * acquiring the lock, but just before it could have added itself to
3365 * the list. There can only be one such pending lock.
3369 * sys_set_robust_list() - Set the robust-futex list head of a task
3370 * @head: pointer to the list-head
3371 * @len: length of the list-head, as userspace expects
3373 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3376 if (!futex_cmpxchg_enabled
)
3379 * The kernel knows only one size for now:
3381 if (unlikely(len
!= sizeof(*head
)))
3384 current
->robust_list
= head
;
3390 * sys_get_robust_list() - Get the robust-futex list head of a task
3391 * @pid: pid of the process [zero for current task]
3392 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3393 * @len_ptr: pointer to a length field, the kernel fills in the header size
3395 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3396 struct robust_list_head __user
* __user
*, head_ptr
,
3397 size_t __user
*, len_ptr
)
3399 struct robust_list_head __user
*head
;
3401 struct task_struct
*p
;
3403 if (!futex_cmpxchg_enabled
)
3412 p
= find_task_by_vpid(pid
);
3418 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3421 head
= p
->robust_list
;
3424 if (put_user(sizeof(*head
), len_ptr
))
3426 return put_user(head
, head_ptr
);
3435 * Process a futex-list entry, check whether it's owned by the
3436 * dying task, and do notification if so:
3438 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
3440 u32 uval
, uninitialized_var(nval
), mval
;
3443 if (get_user(uval
, uaddr
))
3446 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
3448 * Ok, this dying thread is truly holding a futex
3449 * of interest. Set the OWNER_DIED bit atomically
3450 * via cmpxchg, and if the value had FUTEX_WAITERS
3451 * set, wake up a waiter (if any). (We have to do a
3452 * futex_wake() even if OWNER_DIED is already set -
3453 * to handle the rare but possible case of recursive
3454 * thread-death.) The rest of the cleanup is done in
3457 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3459 * We are not holding a lock here, but we want to have
3460 * the pagefault_disable/enable() protection because
3461 * we want to handle the fault gracefully. If the
3462 * access fails we try to fault in the futex with R/W
3463 * verification via get_user_pages. get_user() above
3464 * does not guarantee R/W access. If that fails we
3465 * give up and leave the futex locked.
3467 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
3468 if (fault_in_user_writeable(uaddr
))
3476 * Wake robust non-PI futexes here. The wakeup of
3477 * PI futexes happens in exit_pi_state():
3479 if (!pi
&& (uval
& FUTEX_WAITERS
))
3480 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3486 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3488 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3489 struct robust_list __user
* __user
*head
,
3492 unsigned long uentry
;
3494 if (get_user(uentry
, (unsigned long __user
*)head
))
3497 *entry
= (void __user
*)(uentry
& ~1UL);
3504 * Walk curr->robust_list (very carefully, it's a userspace list!)
3505 * and mark any locks found there dead, and notify any waiters.
3507 * We silently return on any sign of list-walking problem.
3509 void exit_robust_list(struct task_struct
*curr
)
3511 struct robust_list_head __user
*head
= curr
->robust_list
;
3512 struct robust_list __user
*entry
, *next_entry
, *pending
;
3513 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3514 unsigned int uninitialized_var(next_pi
);
3515 unsigned long futex_offset
;
3518 if (!futex_cmpxchg_enabled
)
3522 * Fetch the list head (which was registered earlier, via
3523 * sys_set_robust_list()):
3525 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3528 * Fetch the relative futex offset:
3530 if (get_user(futex_offset
, &head
->futex_offset
))
3533 * Fetch any possibly pending lock-add first, and handle it
3536 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3539 next_entry
= NULL
; /* avoid warning with gcc */
3540 while (entry
!= &head
->list
) {
3542 * Fetch the next entry in the list before calling
3543 * handle_futex_death:
3545 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3547 * A pending lock might already be on the list, so
3548 * don't process it twice:
3550 if (entry
!= pending
)
3551 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3559 * Avoid excessively long or circular lists:
3568 handle_futex_death((void __user
*)pending
+ futex_offset
,
3572 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3573 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3575 int cmd
= op
& FUTEX_CMD_MASK
;
3576 unsigned int flags
= 0;
3578 if (!(op
& FUTEX_PRIVATE_FLAG
))
3579 flags
|= FLAGS_SHARED
;
3581 if (op
& FUTEX_CLOCK_REALTIME
) {
3582 flags
|= FLAGS_CLOCKRT
;
3583 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3584 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3590 case FUTEX_UNLOCK_PI
:
3591 case FUTEX_TRYLOCK_PI
:
3592 case FUTEX_WAIT_REQUEUE_PI
:
3593 case FUTEX_CMP_REQUEUE_PI
:
3594 if (!futex_cmpxchg_enabled
)
3600 val3
= FUTEX_BITSET_MATCH_ANY
;
3601 case FUTEX_WAIT_BITSET
:
3602 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3604 val3
= FUTEX_BITSET_MATCH_ANY
;
3605 case FUTEX_WAKE_BITSET
:
3606 return futex_wake(uaddr
, flags
, val
, val3
);
3608 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3609 case FUTEX_CMP_REQUEUE
:
3610 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3612 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3614 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3615 case FUTEX_UNLOCK_PI
:
3616 return futex_unlock_pi(uaddr
, flags
);
3617 case FUTEX_TRYLOCK_PI
:
3618 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3619 case FUTEX_WAIT_REQUEUE_PI
:
3620 val3
= FUTEX_BITSET_MATCH_ANY
;
3621 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3623 case FUTEX_CMP_REQUEUE_PI
:
3624 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3630 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3631 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3635 ktime_t t
, *tp
= NULL
;
3637 int cmd
= op
& FUTEX_CMD_MASK
;
3639 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3640 cmd
== FUTEX_WAIT_BITSET
||
3641 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3642 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3644 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3646 if (!timespec_valid(&ts
))
3649 t
= timespec_to_ktime(ts
);
3650 if (cmd
== FUTEX_WAIT
)
3651 t
= ktime_add_safe(ktime_get(), t
);
3655 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3656 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3658 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3659 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3660 val2
= (u32
) (unsigned long) utime
;
3662 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3665 static void __init
futex_detect_cmpxchg(void)
3667 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3671 * This will fail and we want it. Some arch implementations do
3672 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3673 * functionality. We want to know that before we call in any
3674 * of the complex code paths. Also we want to prevent
3675 * registration of robust lists in that case. NULL is
3676 * guaranteed to fault and we get -EFAULT on functional
3677 * implementation, the non-functional ones will return
3680 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3681 futex_cmpxchg_enabled
= 1;
3685 static int __init
futex_init(void)
3687 unsigned int futex_shift
;
3690 #if CONFIG_BASE_SMALL
3691 futex_hashsize
= 16;
3693 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3696 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3698 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3700 futex_hashsize
, futex_hashsize
);
3701 futex_hashsize
= 1UL << futex_shift
;
3703 futex_detect_cmpxchg();
3705 for (i
= 0; i
< futex_hashsize
; i
++) {
3706 atomic_set(&futex_queues
[i
].waiters
, 0);
3707 plist_head_init(&futex_queues
[i
].chain
);
3708 spin_lock_init(&futex_queues
[i
].lock
);
3713 core_initcall(futex_init
);