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/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/bootmem.h>
70 #include <linux/fault-inject.h>
72 #include <asm/futex.h>
74 #include "locking/rtmutex_common.h"
77 * READ this before attempting to hack on futexes!
79 * Basic futex operation and ordering guarantees
80 * =============================================
82 * The waiter reads the futex value in user space and calls
83 * futex_wait(). This function computes the hash bucket and acquires
84 * the hash bucket lock. After that it reads the futex user space value
85 * again and verifies that the data has not changed. If it has not changed
86 * it enqueues itself into the hash bucket, releases the hash bucket lock
89 * The waker side modifies the user space value of the futex and calls
90 * futex_wake(). This function computes the hash bucket and acquires the
91 * hash bucket lock. Then it looks for waiters on that futex in the hash
92 * bucket and wakes them.
94 * In futex wake up scenarios where no tasks are blocked on a futex, taking
95 * the hb spinlock can be avoided and simply return. In order for this
96 * optimization to work, ordering guarantees must exist so that the waiter
97 * being added to the list is acknowledged when the list is concurrently being
98 * checked by the waker, avoiding scenarios like the following:
102 * sys_futex(WAIT, futex, val);
103 * futex_wait(futex, val);
106 * sys_futex(WAKE, futex);
111 * lock(hash_bucket(futex));
113 * unlock(hash_bucket(futex));
116 * This would cause the waiter on CPU 0 to wait forever because it
117 * missed the transition of the user space value from val to newval
118 * and the waker did not find the waiter in the hash bucket queue.
120 * The correct serialization ensures that a waiter either observes
121 * the changed user space value before blocking or is woken by a
126 * sys_futex(WAIT, futex, val);
127 * futex_wait(futex, val);
130 * smp_mb(); (A) <-- paired with -.
132 * lock(hash_bucket(futex)); |
136 * | sys_futex(WAKE, futex);
137 * | futex_wake(futex);
139 * `--------> smp_mb(); (B)
142 * unlock(hash_bucket(futex));
143 * schedule(); if (waiters)
144 * lock(hash_bucket(futex));
145 * else wake_waiters(futex);
146 * waiters--; (b) unlock(hash_bucket(futex));
148 * Where (A) orders the waiters increment and the futex value read through
149 * atomic operations (see hb_waiters_inc) and where (B) orders the write
150 * to futex and the waiters read -- this is done by the barriers for both
151 * shared and private futexes in get_futex_key_refs().
153 * This yields the following case (where X:=waiters, Y:=futex):
161 * Which guarantees that x==0 && y==0 is impossible; which translates back into
162 * the guarantee that we cannot both miss the futex variable change and the
165 * Note that a new waiter is accounted for in (a) even when it is possible that
166 * the wait call can return error, in which case we backtrack from it in (b).
167 * Refer to the comment in queue_lock().
169 * Similarly, in order to account for waiters being requeued on another
170 * address we always increment the waiters for the destination bucket before
171 * acquiring the lock. It then decrements them again after releasing it -
172 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173 * will do the additional required waiter count housekeeping. This is done for
174 * double_lock_hb() and double_unlock_hb(), respectively.
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
180 static int __read_mostly futex_cmpxchg_enabled
;
184 * Futex flags used to encode options to functions and preserve them across
188 # define FLAGS_SHARED 0x01
191 * NOMMU does not have per process address space. Let the compiler optimize
194 # define FLAGS_SHARED 0x00
196 #define FLAGS_CLOCKRT 0x02
197 #define FLAGS_HAS_TIMEOUT 0x04
200 * Priority Inheritance state:
202 struct futex_pi_state
{
204 * list of 'owned' pi_state instances - these have to be
205 * cleaned up in do_exit() if the task exits prematurely:
207 struct list_head list
;
212 struct rt_mutex pi_mutex
;
214 struct task_struct
*owner
;
218 } __randomize_layout
;
221 * struct futex_q - The hashed futex queue entry, one per waiting task
222 * @list: priority-sorted list of tasks waiting on this futex
223 * @task: the task waiting on the futex
224 * @lock_ptr: the hash bucket lock
225 * @key: the key the futex is hashed on
226 * @pi_state: optional priority inheritance state
227 * @rt_waiter: rt_waiter storage for use with requeue_pi
228 * @requeue_pi_key: the requeue_pi target futex key
229 * @bitset: bitset for the optional bitmasked wakeup
231 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232 * we can wake only the relevant ones (hashed queues may be shared).
234 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236 * The order of wakeup is always to make the first condition true, then
239 * PI futexes are typically woken before they are removed from the hash list via
240 * the rt_mutex code. See unqueue_me_pi().
243 struct plist_node list
;
245 struct task_struct
*task
;
246 spinlock_t
*lock_ptr
;
248 struct futex_pi_state
*pi_state
;
249 struct rt_mutex_waiter
*rt_waiter
;
250 union futex_key
*requeue_pi_key
;
252 } __randomize_layout
;
254 static const struct futex_q futex_q_init
= {
255 /* list gets initialized in queue_me()*/
256 .key
= FUTEX_KEY_INIT
,
257 .bitset
= FUTEX_BITSET_MATCH_ANY
261 * Hash buckets are shared by all the futex_keys that hash to the same
262 * location. Each key may have multiple futex_q structures, one for each task
263 * waiting on a futex.
265 struct futex_hash_bucket
{
268 struct plist_head chain
;
269 } ____cacheline_aligned_in_smp
;
272 * The base of the bucket array and its size are always used together
273 * (after initialization only in hash_futex()), so ensure that they
274 * reside in the same cacheline.
277 struct futex_hash_bucket
*queues
;
278 unsigned long hashsize
;
279 } __futex_data __read_mostly
__aligned(2*sizeof(long));
280 #define futex_queues (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
285 * Fault injections for futexes.
287 #ifdef CONFIG_FAIL_FUTEX
290 struct fault_attr attr
;
294 .attr
= FAULT_ATTR_INITIALIZER
,
295 .ignore_private
= false,
298 static int __init
setup_fail_futex(char *str
)
300 return setup_fault_attr(&fail_futex
.attr
, str
);
302 __setup("fail_futex=", setup_fail_futex
);
304 static bool should_fail_futex(bool fshared
)
306 if (fail_futex
.ignore_private
&& !fshared
)
309 return should_fail(&fail_futex
.attr
, 1);
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314 static int __init
fail_futex_debugfs(void)
316 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
319 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
324 if (!debugfs_create_bool("ignore-private", mode
, dir
,
325 &fail_futex
.ignore_private
)) {
326 debugfs_remove_recursive(dir
);
333 late_initcall(fail_futex_debugfs
);
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
338 static inline bool should_fail_futex(bool fshared
)
342 #endif /* CONFIG_FAIL_FUTEX */
345 static void compat_exit_robust_list(struct task_struct
*curr
);
347 static inline void compat_exit_robust_list(struct task_struct
*curr
) { }
350 static inline void futex_get_mm(union futex_key
*key
)
352 mmgrab(key
->private.mm
);
354 * Ensure futex_get_mm() implies a full barrier such that
355 * get_futex_key() implies a full barrier. This is relied upon
356 * as smp_mb(); (B), see the ordering comment above.
358 smp_mb__after_atomic();
362 * Reflects a new waiter being added to the waitqueue.
364 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
367 atomic_inc(&hb
->waiters
);
369 * Full barrier (A), see the ordering comment above.
371 smp_mb__after_atomic();
376 * Reflects a waiter being removed from the waitqueue by wakeup
379 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
382 atomic_dec(&hb
->waiters
);
386 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
389 return atomic_read(&hb
->waiters
);
396 * hash_futex - Return the hash bucket in the global hash
397 * @key: Pointer to the futex key for which the hash is calculated
399 * We hash on the keys returned from get_futex_key (see below) and return the
400 * corresponding hash bucket in the global hash.
402 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
404 u32 hash
= jhash2((u32
*)&key
->both
.word
,
405 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
407 return &futex_queues
[hash
& (futex_hashsize
- 1)];
412 * match_futex - Check whether two futex keys are equal
413 * @key1: Pointer to key1
414 * @key2: Pointer to key2
416 * Return 1 if two futex_keys are equal, 0 otherwise.
418 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
421 && key1
->both
.word
== key2
->both
.word
422 && key1
->both
.ptr
== key2
->both
.ptr
423 && key1
->both
.offset
== key2
->both
.offset
);
427 * Take a reference to the resource addressed by a key.
428 * Can be called while holding spinlocks.
431 static void get_futex_key_refs(union futex_key
*key
)
437 * On MMU less systems futexes are always "private" as there is no per
438 * process address space. We need the smp wmb nevertheless - yes,
439 * arch/blackfin has MMU less SMP ...
441 if (!IS_ENABLED(CONFIG_MMU
)) {
442 smp_mb(); /* explicit smp_mb(); (B) */
446 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
448 ihold(key
->shared
.inode
); /* implies smp_mb(); (B) */
450 case FUT_OFF_MMSHARED
:
451 futex_get_mm(key
); /* implies smp_mb(); (B) */
455 * Private futexes do not hold reference on an inode or
456 * mm, therefore the only purpose of calling get_futex_key_refs
457 * is because we need the barrier for the lockless waiter check.
459 smp_mb(); /* explicit smp_mb(); (B) */
464 * Drop a reference to the resource addressed by a key.
465 * The hash bucket spinlock must not be held. This is
466 * a no-op for private futexes, see comment in the get
469 static void drop_futex_key_refs(union futex_key
*key
)
471 if (!key
->both
.ptr
) {
472 /* If we're here then we tried to put a key we failed to get */
477 if (!IS_ENABLED(CONFIG_MMU
))
480 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
482 iput(key
->shared
.inode
);
484 case FUT_OFF_MMSHARED
:
485 mmdrop(key
->private.mm
);
491 * get_futex_key() - Get parameters which are the keys for a futex
492 * @uaddr: virtual address of the futex
493 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
494 * @key: address where result is stored.
495 * @rw: mapping needs to be read/write (values: VERIFY_READ,
498 * Return: a negative error code or 0
500 * The key words are stored in @key on success.
502 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
503 * offset_within_page). For private mappings, it's (uaddr, current->mm).
504 * We can usually work out the index without swapping in the page.
506 * lock_page() might sleep, the caller should not hold a spinlock.
509 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
511 unsigned long address
= (unsigned long)uaddr
;
512 struct mm_struct
*mm
= current
->mm
;
513 struct page
*page
, *tail
;
514 struct address_space
*mapping
;
518 * The futex address must be "naturally" aligned.
520 key
->both
.offset
= address
% PAGE_SIZE
;
521 if (unlikely((address
% sizeof(u32
)) != 0))
523 address
-= key
->both
.offset
;
525 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
528 if (unlikely(should_fail_futex(fshared
)))
532 * PROCESS_PRIVATE futexes are fast.
533 * As the mm cannot disappear under us and the 'key' only needs
534 * virtual address, we dont even have to find the underlying vma.
535 * Note : We do have to check 'uaddr' is a valid user address,
536 * but access_ok() should be faster than find_vma()
539 key
->private.mm
= mm
;
540 key
->private.address
= address
;
541 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
546 /* Ignore any VERIFY_READ mapping (futex common case) */
547 if (unlikely(should_fail_futex(fshared
)))
550 err
= get_user_pages_fast(address
, 1, 1, &page
);
552 * If write access is not required (eg. FUTEX_WAIT), try
553 * and get read-only access.
555 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
556 err
= get_user_pages_fast(address
, 1, 0, &page
);
565 * The treatment of mapping from this point on is critical. The page
566 * lock protects many things but in this context the page lock
567 * stabilizes mapping, prevents inode freeing in the shared
568 * file-backed region case and guards against movement to swap cache.
570 * Strictly speaking the page lock is not needed in all cases being
571 * considered here and page lock forces unnecessarily serialization
572 * From this point on, mapping will be re-verified if necessary and
573 * page lock will be acquired only if it is unavoidable
575 * Mapping checks require the head page for any compound page so the
576 * head page and mapping is looked up now. For anonymous pages, it
577 * does not matter if the page splits in the future as the key is
578 * based on the address. For filesystem-backed pages, the tail is
579 * required as the index of the page determines the key. For
580 * base pages, there is no tail page and tail == page.
583 page
= compound_head(page
);
584 mapping
= READ_ONCE(page
->mapping
);
587 * If page->mapping is NULL, then it cannot be a PageAnon
588 * page; but it might be the ZERO_PAGE or in the gate area or
589 * in a special mapping (all cases which we are happy to fail);
590 * or it may have been a good file page when get_user_pages_fast
591 * found it, but truncated or holepunched or subjected to
592 * invalidate_complete_page2 before we got the page lock (also
593 * cases which we are happy to fail). And we hold a reference,
594 * so refcount care in invalidate_complete_page's remove_mapping
595 * prevents drop_caches from setting mapping to NULL beneath us.
597 * The case we do have to guard against is when memory pressure made
598 * shmem_writepage move it from filecache to swapcache beneath us:
599 * an unlikely race, but we do need to retry for page->mapping.
601 if (unlikely(!mapping
)) {
605 * Page lock is required to identify which special case above
606 * applies. If this is really a shmem page then the page lock
607 * will prevent unexpected transitions.
610 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
621 * Private mappings are handled in a simple way.
623 * If the futex key is stored on an anonymous page, then the associated
624 * object is the mm which is implicitly pinned by the calling process.
626 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
627 * it's a read-only handle, it's expected that futexes attach to
628 * the object not the particular process.
630 if (PageAnon(page
)) {
632 * A RO anonymous page will never change and thus doesn't make
633 * sense for futex operations.
635 if (unlikely(should_fail_futex(fshared
)) || ro
) {
640 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
641 key
->private.mm
= mm
;
642 key
->private.address
= address
;
644 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
650 * The associated futex object in this case is the inode and
651 * the page->mapping must be traversed. Ordinarily this should
652 * be stabilised under page lock but it's not strictly
653 * necessary in this case as we just want to pin the inode, not
654 * update the radix tree or anything like that.
656 * The RCU read lock is taken as the inode is finally freed
657 * under RCU. If the mapping still matches expectations then the
658 * mapping->host can be safely accessed as being a valid inode.
662 if (READ_ONCE(page
->mapping
) != mapping
) {
669 inode
= READ_ONCE(mapping
->host
);
678 * Take a reference unless it is about to be freed. Previously
679 * this reference was taken by ihold under the page lock
680 * pinning the inode in place so i_lock was unnecessary. The
681 * only way for this check to fail is if the inode was
682 * truncated in parallel which is almost certainly an
683 * application bug. In such a case, just retry.
685 * We are not calling into get_futex_key_refs() in file-backed
686 * cases, therefore a successful atomic_inc return below will
687 * guarantee that get_futex_key() will still imply smp_mb(); (B).
689 if (!atomic_inc_not_zero(&inode
->i_count
)) {
696 /* Should be impossible but lets be paranoid for now */
697 if (WARN_ON_ONCE(inode
->i_mapping
!= mapping
)) {
705 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
706 key
->shared
.inode
= inode
;
707 key
->shared
.pgoff
= basepage_index(tail
);
716 static inline void put_futex_key(union futex_key
*key
)
718 drop_futex_key_refs(key
);
722 * fault_in_user_writeable() - Fault in user address and verify RW access
723 * @uaddr: pointer to faulting user space address
725 * Slow path to fixup the fault we just took in the atomic write
728 * We have no generic implementation of a non-destructive write to the
729 * user address. We know that we faulted in the atomic pagefault
730 * disabled section so we can as well avoid the #PF overhead by
731 * calling get_user_pages() right away.
733 static int fault_in_user_writeable(u32 __user
*uaddr
)
735 struct mm_struct
*mm
= current
->mm
;
738 down_read(&mm
->mmap_sem
);
739 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
740 FAULT_FLAG_WRITE
, NULL
);
741 up_read(&mm
->mmap_sem
);
743 return ret
< 0 ? ret
: 0;
747 * futex_top_waiter() - Return the highest priority waiter on a futex
748 * @hb: the hash bucket the futex_q's reside in
749 * @key: the futex key (to distinguish it from other futex futex_q's)
751 * Must be called with the hb lock held.
753 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
754 union futex_key
*key
)
756 struct futex_q
*this;
758 plist_for_each_entry(this, &hb
->chain
, list
) {
759 if (match_futex(&this->key
, key
))
765 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
766 u32 uval
, u32 newval
)
771 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
777 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
782 ret
= __get_user(*dest
, from
);
785 return ret
? -EFAULT
: 0;
792 static int refill_pi_state_cache(void)
794 struct futex_pi_state
*pi_state
;
796 if (likely(current
->pi_state_cache
))
799 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
804 INIT_LIST_HEAD(&pi_state
->list
);
805 /* pi_mutex gets initialized later */
806 pi_state
->owner
= NULL
;
807 atomic_set(&pi_state
->refcount
, 1);
808 pi_state
->key
= FUTEX_KEY_INIT
;
810 current
->pi_state_cache
= pi_state
;
815 static struct futex_pi_state
*alloc_pi_state(void)
817 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
820 current
->pi_state_cache
= NULL
;
825 static void get_pi_state(struct futex_pi_state
*pi_state
)
827 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state
->refcount
));
831 * Drops a reference to the pi_state object and frees or caches it
832 * when the last reference is gone.
834 static void put_pi_state(struct futex_pi_state
*pi_state
)
839 if (!atomic_dec_and_test(&pi_state
->refcount
))
843 * If pi_state->owner is NULL, the owner is most probably dying
844 * and has cleaned up the pi_state already
846 if (pi_state
->owner
) {
847 struct task_struct
*owner
;
849 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
850 owner
= pi_state
->owner
;
852 raw_spin_lock(&owner
->pi_lock
);
853 list_del_init(&pi_state
->list
);
854 raw_spin_unlock(&owner
->pi_lock
);
856 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
857 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
860 if (current
->pi_state_cache
) {
864 * pi_state->list is already empty.
865 * clear pi_state->owner.
866 * refcount is at 0 - put it back to 1.
868 pi_state
->owner
= NULL
;
869 atomic_set(&pi_state
->refcount
, 1);
870 current
->pi_state_cache
= pi_state
;
875 * Look up the task based on what TID userspace gave us.
878 static struct task_struct
*futex_find_get_task(pid_t pid
)
880 struct task_struct
*p
;
883 p
= find_task_by_vpid(pid
);
892 #ifdef CONFIG_FUTEX_PI
895 * This task is holding PI mutexes at exit time => bad.
896 * Kernel cleans up PI-state, but userspace is likely hosed.
897 * (Robust-futex cleanup is separate and might save the day for userspace.)
899 static void exit_pi_state_list(struct task_struct
*curr
)
901 struct list_head
*next
, *head
= &curr
->pi_state_list
;
902 struct futex_pi_state
*pi_state
;
903 struct futex_hash_bucket
*hb
;
904 union futex_key key
= FUTEX_KEY_INIT
;
906 if (!futex_cmpxchg_enabled
)
909 * We are a ZOMBIE and nobody can enqueue itself on
910 * pi_state_list anymore, but we have to be careful
911 * versus waiters unqueueing themselves:
913 raw_spin_lock_irq(&curr
->pi_lock
);
914 while (!list_empty(head
)) {
916 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
918 hb
= hash_futex(&key
);
921 * We can race against put_pi_state() removing itself from the
922 * list (a waiter going away). put_pi_state() will first
923 * decrement the reference count and then modify the list, so
924 * its possible to see the list entry but fail this reference
927 * In that case; drop the locks to let put_pi_state() make
928 * progress and retry the loop.
930 if (!atomic_inc_not_zero(&pi_state
->refcount
)) {
931 raw_spin_unlock_irq(&curr
->pi_lock
);
933 raw_spin_lock_irq(&curr
->pi_lock
);
936 raw_spin_unlock_irq(&curr
->pi_lock
);
938 spin_lock(&hb
->lock
);
939 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
940 raw_spin_lock(&curr
->pi_lock
);
942 * We dropped the pi-lock, so re-check whether this
943 * task still owns the PI-state:
945 if (head
->next
!= next
) {
946 /* retain curr->pi_lock for the loop invariant */
947 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
948 spin_unlock(&hb
->lock
);
949 put_pi_state(pi_state
);
953 WARN_ON(pi_state
->owner
!= curr
);
954 WARN_ON(list_empty(&pi_state
->list
));
955 list_del_init(&pi_state
->list
);
956 pi_state
->owner
= NULL
;
958 raw_spin_unlock(&curr
->pi_lock
);
959 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
960 spin_unlock(&hb
->lock
);
962 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
963 put_pi_state(pi_state
);
965 raw_spin_lock_irq(&curr
->pi_lock
);
967 raw_spin_unlock_irq(&curr
->pi_lock
);
970 static inline void exit_pi_state_list(struct task_struct
*curr
) { }
974 * We need to check the following states:
976 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
978 * [1] NULL | --- | --- | 0 | 0/1 | Valid
979 * [2] NULL | --- | --- | >0 | 0/1 | Valid
981 * [3] Found | NULL | -- | Any | 0/1 | Invalid
983 * [4] Found | Found | NULL | 0 | 1 | Valid
984 * [5] Found | Found | NULL | >0 | 1 | Invalid
986 * [6] Found | Found | task | 0 | 1 | Valid
988 * [7] Found | Found | NULL | Any | 0 | Invalid
990 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
991 * [9] Found | Found | task | 0 | 0 | Invalid
992 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
994 * [1] Indicates that the kernel can acquire the futex atomically. We
995 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
997 * [2] Valid, if TID does not belong to a kernel thread. If no matching
998 * thread is found then it indicates that the owner TID has died.
1000 * [3] Invalid. The waiter is queued on a non PI futex
1002 * [4] Valid state after exit_robust_list(), which sets the user space
1003 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1005 * [5] The user space value got manipulated between exit_robust_list()
1006 * and exit_pi_state_list()
1008 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1009 * the pi_state but cannot access the user space value.
1011 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1013 * [8] Owner and user space value match
1015 * [9] There is no transient state which sets the user space TID to 0
1016 * except exit_robust_list(), but this is indicated by the
1017 * FUTEX_OWNER_DIED bit. See [4]
1019 * [10] There is no transient state which leaves owner and user space
1023 * Serialization and lifetime rules:
1027 * hb -> futex_q, relation
1028 * futex_q -> pi_state, relation
1030 * (cannot be raw because hb can contain arbitrary amount
1033 * pi_mutex->wait_lock:
1037 * (and pi_mutex 'obviously')
1041 * p->pi_state_list -> pi_state->list, relation
1043 * pi_state->refcount:
1051 * pi_mutex->wait_lock
1057 * Validate that the existing waiter has a pi_state and sanity check
1058 * the pi_state against the user space value. If correct, attach to
1061 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1062 struct futex_pi_state
*pi_state
,
1063 struct futex_pi_state
**ps
)
1065 pid_t pid
= uval
& FUTEX_TID_MASK
;
1070 * Userspace might have messed up non-PI and PI futexes [3]
1072 if (unlikely(!pi_state
))
1076 * We get here with hb->lock held, and having found a
1077 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1078 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1079 * which in turn means that futex_lock_pi() still has a reference on
1082 * The waiter holding a reference on @pi_state also protects against
1083 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1084 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1085 * free pi_state before we can take a reference ourselves.
1087 WARN_ON(!atomic_read(&pi_state
->refcount
));
1090 * Now that we have a pi_state, we can acquire wait_lock
1091 * and do the state validation.
1093 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1096 * Since {uval, pi_state} is serialized by wait_lock, and our current
1097 * uval was read without holding it, it can have changed. Verify it
1098 * still is what we expect it to be, otherwise retry the entire
1101 if (get_futex_value_locked(&uval2
, uaddr
))
1108 * Handle the owner died case:
1110 if (uval
& FUTEX_OWNER_DIED
) {
1112 * exit_pi_state_list sets owner to NULL and wakes the
1113 * topmost waiter. The task which acquires the
1114 * pi_state->rt_mutex will fixup owner.
1116 if (!pi_state
->owner
) {
1118 * No pi state owner, but the user space TID
1119 * is not 0. Inconsistent state. [5]
1124 * Take a ref on the state and return success. [4]
1130 * If TID is 0, then either the dying owner has not
1131 * yet executed exit_pi_state_list() or some waiter
1132 * acquired the rtmutex in the pi state, but did not
1133 * yet fixup the TID in user space.
1135 * Take a ref on the state and return success. [6]
1141 * If the owner died bit is not set, then the pi_state
1142 * must have an owner. [7]
1144 if (!pi_state
->owner
)
1149 * Bail out if user space manipulated the futex value. If pi
1150 * state exists then the owner TID must be the same as the
1151 * user space TID. [9/10]
1153 if (pid
!= task_pid_vnr(pi_state
->owner
))
1157 get_pi_state(pi_state
);
1158 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1175 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1179 static int handle_exit_race(u32 __user
*uaddr
, u32 uval
,
1180 struct task_struct
*tsk
)
1185 * If the futex exit state is not yet FUTEX_STATE_DEAD, wait
1188 if (tsk
&& tsk
->futex_state
!= FUTEX_STATE_DEAD
)
1192 * Reread the user space value to handle the following situation:
1196 * sys_exit() sys_futex()
1197 * do_exit() futex_lock_pi()
1198 * futex_lock_pi_atomic()
1199 * exit_signals(tsk) No waiters:
1200 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1201 * mm_release(tsk) Set waiter bit
1202 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1203 * Set owner died attach_to_pi_owner() {
1204 * *uaddr = 0xC0000000; tsk = get_task(PID);
1205 * } if (!tsk->flags & PF_EXITING) {
1207 * tsk->futex_state = } else {
1208 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1211 * return -ESRCH; <--- FAIL
1214 * Returning ESRCH unconditionally is wrong here because the
1215 * user space value has been changed by the exiting task.
1217 * The same logic applies to the case where the exiting task is
1220 if (get_futex_value_locked(&uval2
, uaddr
))
1223 /* If the user space value has changed, try again. */
1228 * The exiting task did not have a robust list, the robust list was
1229 * corrupted or the user space value in *uaddr is simply bogus.
1230 * Give up and tell user space.
1236 * Lookup the task for the TID provided from user space and attach to
1237 * it after doing proper sanity checks.
1239 static int attach_to_pi_owner(u32 __user
*uaddr
, u32 uval
, union futex_key
*key
,
1240 struct futex_pi_state
**ps
)
1242 pid_t pid
= uval
& FUTEX_TID_MASK
;
1243 struct futex_pi_state
*pi_state
;
1244 struct task_struct
*p
;
1247 * We are the first waiter - try to look up the real owner and attach
1248 * the new pi_state to it, but bail out when TID = 0 [1]
1250 * The !pid check is paranoid. None of the call sites should end up
1251 * with pid == 0, but better safe than sorry. Let the caller retry
1255 p
= futex_find_get_task(pid
);
1257 return handle_exit_race(uaddr
, uval
, NULL
);
1259 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1265 * We need to look at the task state to figure out, whether the
1266 * task is exiting. To protect against the change of the task state
1267 * in futex_exit_release(), we do this protected by p->pi_lock:
1269 raw_spin_lock_irq(&p
->pi_lock
);
1270 if (unlikely(p
->futex_state
!= FUTEX_STATE_OK
)) {
1272 * The task is on the way out. When the futex state is
1273 * FUTEX_STATE_DEAD, we know that the task has finished
1276 int ret
= handle_exit_race(uaddr
, uval
, p
);
1278 raw_spin_unlock_irq(&p
->pi_lock
);
1284 * No existing pi state. First waiter. [2]
1286 * This creates pi_state, we have hb->lock held, this means nothing can
1287 * observe this state, wait_lock is irrelevant.
1289 pi_state
= alloc_pi_state();
1292 * Initialize the pi_mutex in locked state and make @p
1295 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1297 /* Store the key for possible exit cleanups: */
1298 pi_state
->key
= *key
;
1300 WARN_ON(!list_empty(&pi_state
->list
));
1301 list_add(&pi_state
->list
, &p
->pi_state_list
);
1303 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1304 * because there is no concurrency as the object is not published yet.
1306 pi_state
->owner
= p
;
1307 raw_spin_unlock_irq(&p
->pi_lock
);
1316 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1317 struct futex_hash_bucket
*hb
,
1318 union futex_key
*key
, struct futex_pi_state
**ps
)
1320 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1323 * If there is a waiter on that futex, validate it and
1324 * attach to the pi_state when the validation succeeds.
1327 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1330 * We are the first waiter - try to look up the owner based on
1331 * @uval and attach to it.
1333 return attach_to_pi_owner(uaddr
, uval
, key
, ps
);
1336 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1339 u32
uninitialized_var(curval
);
1341 if (unlikely(should_fail_futex(true)))
1344 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1348 /* If user space value changed, let the caller retry */
1349 return curval
!= uval
? -EAGAIN
: 0;
1353 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1354 * @uaddr: the pi futex user address
1355 * @hb: the pi futex hash bucket
1356 * @key: the futex key associated with uaddr and hb
1357 * @ps: the pi_state pointer where we store the result of the
1359 * @task: the task to perform the atomic lock work for. This will
1360 * be "current" except in the case of requeue pi.
1361 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1364 * - 0 - ready to wait;
1365 * - 1 - acquired the lock;
1368 * The hb->lock and futex_key refs shall be held by the caller.
1370 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1371 union futex_key
*key
,
1372 struct futex_pi_state
**ps
,
1373 struct task_struct
*task
, int set_waiters
)
1375 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1376 struct futex_q
*top_waiter
;
1380 * Read the user space value first so we can validate a few
1381 * things before proceeding further.
1383 if (get_futex_value_locked(&uval
, uaddr
))
1386 if (unlikely(should_fail_futex(true)))
1392 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1395 if ((unlikely(should_fail_futex(true))))
1399 * Lookup existing state first. If it exists, try to attach to
1402 top_waiter
= futex_top_waiter(hb
, key
);
1404 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1407 * No waiter and user TID is 0. We are here because the
1408 * waiters or the owner died bit is set or called from
1409 * requeue_cmp_pi or for whatever reason something took the
1412 if (!(uval
& FUTEX_TID_MASK
)) {
1414 * We take over the futex. No other waiters and the user space
1415 * TID is 0. We preserve the owner died bit.
1417 newval
= uval
& FUTEX_OWNER_DIED
;
1420 /* The futex requeue_pi code can enforce the waiters bit */
1422 newval
|= FUTEX_WAITERS
;
1424 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1425 /* If the take over worked, return 1 */
1426 return ret
< 0 ? ret
: 1;
1430 * First waiter. Set the waiters bit before attaching ourself to
1431 * the owner. If owner tries to unlock, it will be forced into
1432 * the kernel and blocked on hb->lock.
1434 newval
= uval
| FUTEX_WAITERS
;
1435 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1439 * If the update of the user space value succeeded, we try to
1440 * attach to the owner. If that fails, no harm done, we only
1441 * set the FUTEX_WAITERS bit in the user space variable.
1443 return attach_to_pi_owner(uaddr
, newval
, key
, ps
);
1447 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1448 * @q: The futex_q to unqueue
1450 * The q->lock_ptr must not be NULL and must be held by the caller.
1452 static void __unqueue_futex(struct futex_q
*q
)
1454 struct futex_hash_bucket
*hb
;
1456 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1457 || WARN_ON(plist_node_empty(&q
->list
)))
1460 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1461 plist_del(&q
->list
, &hb
->chain
);
1466 * The hash bucket lock must be held when this is called.
1467 * Afterwards, the futex_q must not be accessed. Callers
1468 * must ensure to later call wake_up_q() for the actual
1471 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1473 struct task_struct
*p
= q
->task
;
1475 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1481 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1482 * is written, without taking any locks. This is possible in the event
1483 * of a spurious wakeup, for example. A memory barrier is required here
1484 * to prevent the following store to lock_ptr from getting ahead of the
1485 * plist_del in __unqueue_futex().
1487 smp_store_release(&q
->lock_ptr
, NULL
);
1490 * Queue the task for later wakeup for after we've released
1491 * the hb->lock. wake_q_add() grabs reference to p.
1493 wake_q_add(wake_q
, p
);
1498 * Caller must hold a reference on @pi_state.
1500 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1502 u32
uninitialized_var(curval
), newval
;
1503 struct task_struct
*new_owner
;
1504 bool postunlock
= false;
1505 DEFINE_WAKE_Q(wake_q
);
1508 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1509 if (WARN_ON_ONCE(!new_owner
)) {
1511 * As per the comment in futex_unlock_pi() this should not happen.
1513 * When this happens, give up our locks and try again, giving
1514 * the futex_lock_pi() instance time to complete, either by
1515 * waiting on the rtmutex or removing itself from the futex
1523 * We pass it to the next owner. The WAITERS bit is always kept
1524 * enabled while there is PI state around. We cleanup the owner
1525 * died bit, because we are the owner.
1527 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1529 if (unlikely(should_fail_futex(true)))
1532 ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
1533 if (!ret
&& (curval
!= uval
)) {
1535 * If a unconditional UNLOCK_PI operation (user space did not
1536 * try the TID->0 transition) raced with a waiter setting the
1537 * FUTEX_WAITERS flag between get_user() and locking the hash
1538 * bucket lock, retry the operation.
1540 if ((FUTEX_TID_MASK
& curval
) == uval
)
1550 * This is a point of no return; once we modify the uval there is no
1551 * going back and subsequent operations must not fail.
1554 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1555 WARN_ON(list_empty(&pi_state
->list
));
1556 list_del_init(&pi_state
->list
);
1557 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1559 raw_spin_lock(&new_owner
->pi_lock
);
1560 WARN_ON(!list_empty(&pi_state
->list
));
1561 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1562 pi_state
->owner
= new_owner
;
1563 raw_spin_unlock(&new_owner
->pi_lock
);
1565 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1568 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1571 rt_mutex_postunlock(&wake_q
);
1577 * Express the locking dependencies for lockdep:
1580 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1583 spin_lock(&hb1
->lock
);
1585 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1586 } else { /* hb1 > hb2 */
1587 spin_lock(&hb2
->lock
);
1588 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1593 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1595 spin_unlock(&hb1
->lock
);
1597 spin_unlock(&hb2
->lock
);
1601 * Wake up waiters matching bitset queued on this futex (uaddr).
1604 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1606 struct futex_hash_bucket
*hb
;
1607 struct futex_q
*this, *next
;
1608 union futex_key key
= FUTEX_KEY_INIT
;
1610 DEFINE_WAKE_Q(wake_q
);
1615 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1616 if (unlikely(ret
!= 0))
1619 hb
= hash_futex(&key
);
1621 /* Make sure we really have tasks to wakeup */
1622 if (!hb_waiters_pending(hb
))
1625 spin_lock(&hb
->lock
);
1627 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1628 if (match_futex (&this->key
, &key
)) {
1629 if (this->pi_state
|| this->rt_waiter
) {
1634 /* Check if one of the bits is set in both bitsets */
1635 if (!(this->bitset
& bitset
))
1638 mark_wake_futex(&wake_q
, this);
1639 if (++ret
>= nr_wake
)
1644 spin_unlock(&hb
->lock
);
1647 put_futex_key(&key
);
1652 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1654 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1655 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1656 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 11);
1657 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 11);
1660 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1661 if (oparg
< 0 || oparg
> 31) {
1662 char comm
[sizeof(current
->comm
)];
1664 * kill this print and return -EINVAL when userspace
1667 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1668 get_task_comm(comm
, current
), oparg
);
1674 if (!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
)))
1677 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1682 case FUTEX_OP_CMP_EQ
:
1683 return oldval
== cmparg
;
1684 case FUTEX_OP_CMP_NE
:
1685 return oldval
!= cmparg
;
1686 case FUTEX_OP_CMP_LT
:
1687 return oldval
< cmparg
;
1688 case FUTEX_OP_CMP_GE
:
1689 return oldval
>= cmparg
;
1690 case FUTEX_OP_CMP_LE
:
1691 return oldval
<= cmparg
;
1692 case FUTEX_OP_CMP_GT
:
1693 return oldval
> cmparg
;
1700 * Wake up all waiters hashed on the physical page that is mapped
1701 * to this virtual address:
1704 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1705 int nr_wake
, int nr_wake2
, int op
)
1707 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1708 struct futex_hash_bucket
*hb1
, *hb2
;
1709 struct futex_q
*this, *next
;
1711 DEFINE_WAKE_Q(wake_q
);
1714 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1715 if (unlikely(ret
!= 0))
1717 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1718 if (unlikely(ret
!= 0))
1721 hb1
= hash_futex(&key1
);
1722 hb2
= hash_futex(&key2
);
1725 double_lock_hb(hb1
, hb2
);
1726 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1727 if (unlikely(op_ret
< 0)) {
1728 double_unlock_hb(hb1
, hb2
);
1730 if (!IS_ENABLED(CONFIG_MMU
) ||
1731 unlikely(op_ret
!= -EFAULT
&& op_ret
!= -EAGAIN
)) {
1733 * we don't get EFAULT from MMU faults if we don't have
1734 * an MMU, but we might get them from range checking
1740 if (op_ret
== -EFAULT
) {
1741 ret
= fault_in_user_writeable(uaddr2
);
1746 if (!(flags
& FLAGS_SHARED
)) {
1751 put_futex_key(&key2
);
1752 put_futex_key(&key1
);
1757 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1758 if (match_futex (&this->key
, &key1
)) {
1759 if (this->pi_state
|| this->rt_waiter
) {
1763 mark_wake_futex(&wake_q
, this);
1764 if (++ret
>= nr_wake
)
1771 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1772 if (match_futex (&this->key
, &key2
)) {
1773 if (this->pi_state
|| this->rt_waiter
) {
1777 mark_wake_futex(&wake_q
, this);
1778 if (++op_ret
>= nr_wake2
)
1786 double_unlock_hb(hb1
, hb2
);
1789 put_futex_key(&key2
);
1791 put_futex_key(&key1
);
1797 * requeue_futex() - Requeue a futex_q from one hb to another
1798 * @q: the futex_q to requeue
1799 * @hb1: the source hash_bucket
1800 * @hb2: the target hash_bucket
1801 * @key2: the new key for the requeued futex_q
1804 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1805 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1809 * If key1 and key2 hash to the same bucket, no need to
1812 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1813 plist_del(&q
->list
, &hb1
->chain
);
1814 hb_waiters_dec(hb1
);
1815 hb_waiters_inc(hb2
);
1816 plist_add(&q
->list
, &hb2
->chain
);
1817 q
->lock_ptr
= &hb2
->lock
;
1819 get_futex_key_refs(key2
);
1824 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1826 * @key: the key of the requeue target futex
1827 * @hb: the hash_bucket of the requeue target futex
1829 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1830 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1831 * to the requeue target futex so the waiter can detect the wakeup on the right
1832 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1833 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1834 * to protect access to the pi_state to fixup the owner later. Must be called
1835 * with both q->lock_ptr and hb->lock held.
1838 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1839 struct futex_hash_bucket
*hb
)
1841 get_futex_key_refs(key
);
1846 WARN_ON(!q
->rt_waiter
);
1847 q
->rt_waiter
= NULL
;
1849 q
->lock_ptr
= &hb
->lock
;
1851 wake_up_state(q
->task
, TASK_NORMAL
);
1855 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1856 * @pifutex: the user address of the to futex
1857 * @hb1: the from futex hash bucket, must be locked by the caller
1858 * @hb2: the to futex hash bucket, must be locked by the caller
1859 * @key1: the from futex key
1860 * @key2: the to futex key
1861 * @ps: address to store the pi_state pointer
1862 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1864 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1865 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1866 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1867 * hb1 and hb2 must be held by the caller.
1870 * - 0 - failed to acquire the lock atomically;
1871 * - >0 - acquired the lock, return value is vpid of the top_waiter
1874 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1875 struct futex_hash_bucket
*hb1
,
1876 struct futex_hash_bucket
*hb2
,
1877 union futex_key
*key1
, union futex_key
*key2
,
1878 struct futex_pi_state
**ps
, int set_waiters
)
1880 struct futex_q
*top_waiter
= NULL
;
1884 if (get_futex_value_locked(&curval
, pifutex
))
1887 if (unlikely(should_fail_futex(true)))
1891 * Find the top_waiter and determine if there are additional waiters.
1892 * If the caller intends to requeue more than 1 waiter to pifutex,
1893 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1894 * as we have means to handle the possible fault. If not, don't set
1895 * the bit unecessarily as it will force the subsequent unlock to enter
1898 top_waiter
= futex_top_waiter(hb1
, key1
);
1900 /* There are no waiters, nothing for us to do. */
1904 /* Ensure we requeue to the expected futex. */
1905 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1909 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1910 * the contended case or if set_waiters is 1. The pi_state is returned
1911 * in ps in contended cases.
1913 vpid
= task_pid_vnr(top_waiter
->task
);
1914 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1917 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1924 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1925 * @uaddr1: source futex user address
1926 * @flags: futex flags (FLAGS_SHARED, etc.)
1927 * @uaddr2: target futex user address
1928 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1929 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1930 * @cmpval: @uaddr1 expected value (or %NULL)
1931 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1932 * pi futex (pi to pi requeue is not supported)
1934 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1935 * uaddr2 atomically on behalf of the top waiter.
1938 * - >=0 - on success, the number of tasks requeued or woken;
1941 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1942 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1943 u32
*cmpval
, int requeue_pi
)
1945 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1946 int drop_count
= 0, task_count
= 0, ret
;
1947 struct futex_pi_state
*pi_state
= NULL
;
1948 struct futex_hash_bucket
*hb1
, *hb2
;
1949 struct futex_q
*this, *next
;
1950 DEFINE_WAKE_Q(wake_q
);
1952 if (nr_wake
< 0 || nr_requeue
< 0)
1956 * When PI not supported: return -ENOSYS if requeue_pi is true,
1957 * consequently the compiler knows requeue_pi is always false past
1958 * this point which will optimize away all the conditional code
1961 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1966 * Requeue PI only works on two distinct uaddrs. This
1967 * check is only valid for private futexes. See below.
1969 if (uaddr1
== uaddr2
)
1973 * requeue_pi requires a pi_state, try to allocate it now
1974 * without any locks in case it fails.
1976 if (refill_pi_state_cache())
1979 * requeue_pi must wake as many tasks as it can, up to nr_wake
1980 * + nr_requeue, since it acquires the rt_mutex prior to
1981 * returning to userspace, so as to not leave the rt_mutex with
1982 * waiters and no owner. However, second and third wake-ups
1983 * cannot be predicted as they involve race conditions with the
1984 * first wake and a fault while looking up the pi_state. Both
1985 * pthread_cond_signal() and pthread_cond_broadcast() should
1993 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1994 if (unlikely(ret
!= 0))
1996 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1997 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1998 if (unlikely(ret
!= 0))
2002 * The check above which compares uaddrs is not sufficient for
2003 * shared futexes. We need to compare the keys:
2005 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
2010 hb1
= hash_futex(&key1
);
2011 hb2
= hash_futex(&key2
);
2014 hb_waiters_inc(hb2
);
2015 double_lock_hb(hb1
, hb2
);
2017 if (likely(cmpval
!= NULL
)) {
2020 ret
= get_futex_value_locked(&curval
, uaddr1
);
2022 if (unlikely(ret
)) {
2023 double_unlock_hb(hb1
, hb2
);
2024 hb_waiters_dec(hb2
);
2026 ret
= get_user(curval
, uaddr1
);
2030 if (!(flags
& FLAGS_SHARED
))
2033 put_futex_key(&key2
);
2034 put_futex_key(&key1
);
2037 if (curval
!= *cmpval
) {
2043 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
2045 * Attempt to acquire uaddr2 and wake the top waiter. If we
2046 * intend to requeue waiters, force setting the FUTEX_WAITERS
2047 * bit. We force this here where we are able to easily handle
2048 * faults rather in the requeue loop below.
2050 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
2051 &key2
, &pi_state
, nr_requeue
);
2054 * At this point the top_waiter has either taken uaddr2 or is
2055 * waiting on it. If the former, then the pi_state will not
2056 * exist yet, look it up one more time to ensure we have a
2057 * reference to it. If the lock was taken, ret contains the
2058 * vpid of the top waiter task.
2059 * If the lock was not taken, we have pi_state and an initial
2060 * refcount on it. In case of an error we have nothing.
2067 * If we acquired the lock, then the user space value
2068 * of uaddr2 should be vpid. It cannot be changed by
2069 * the top waiter as it is blocked on hb2 lock if it
2070 * tries to do so. If something fiddled with it behind
2071 * our back the pi state lookup might unearth it. So
2072 * we rather use the known value than rereading and
2073 * handing potential crap to lookup_pi_state.
2075 * If that call succeeds then we have pi_state and an
2076 * initial refcount on it.
2078 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
, &pi_state
);
2083 /* We hold a reference on the pi state. */
2086 /* If the above failed, then pi_state is NULL */
2088 double_unlock_hb(hb1
, hb2
);
2089 hb_waiters_dec(hb2
);
2090 put_futex_key(&key2
);
2091 put_futex_key(&key1
);
2092 ret
= fault_in_user_writeable(uaddr2
);
2098 * Two reasons for this:
2099 * - Owner is exiting and we just wait for the
2101 * - The user space value changed.
2103 double_unlock_hb(hb1
, hb2
);
2104 hb_waiters_dec(hb2
);
2105 put_futex_key(&key2
);
2106 put_futex_key(&key1
);
2114 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2115 if (task_count
- nr_wake
>= nr_requeue
)
2118 if (!match_futex(&this->key
, &key1
))
2122 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2123 * be paired with each other and no other futex ops.
2125 * We should never be requeueing a futex_q with a pi_state,
2126 * which is awaiting a futex_unlock_pi().
2128 if ((requeue_pi
&& !this->rt_waiter
) ||
2129 (!requeue_pi
&& this->rt_waiter
) ||
2136 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2137 * lock, we already woke the top_waiter. If not, it will be
2138 * woken by futex_unlock_pi().
2140 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2141 mark_wake_futex(&wake_q
, this);
2145 /* Ensure we requeue to the expected futex for requeue_pi. */
2146 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2152 * Requeue nr_requeue waiters and possibly one more in the case
2153 * of requeue_pi if we couldn't acquire the lock atomically.
2157 * Prepare the waiter to take the rt_mutex. Take a
2158 * refcount on the pi_state and store the pointer in
2159 * the futex_q object of the waiter.
2161 get_pi_state(pi_state
);
2162 this->pi_state
= pi_state
;
2163 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2168 * We got the lock. We do neither drop the
2169 * refcount on pi_state nor clear
2170 * this->pi_state because the waiter needs the
2171 * pi_state for cleaning up the user space
2172 * value. It will drop the refcount after
2175 requeue_pi_wake_futex(this, &key2
, hb2
);
2180 * rt_mutex_start_proxy_lock() detected a
2181 * potential deadlock when we tried to queue
2182 * that waiter. Drop the pi_state reference
2183 * which we took above and remove the pointer
2184 * to the state from the waiters futex_q
2187 this->pi_state
= NULL
;
2188 put_pi_state(pi_state
);
2190 * We stop queueing more waiters and let user
2191 * space deal with the mess.
2196 requeue_futex(this, hb1
, hb2
, &key2
);
2201 * We took an extra initial reference to the pi_state either
2202 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2203 * need to drop it here again.
2205 put_pi_state(pi_state
);
2208 double_unlock_hb(hb1
, hb2
);
2210 hb_waiters_dec(hb2
);
2213 * drop_futex_key_refs() must be called outside the spinlocks. During
2214 * the requeue we moved futex_q's from the hash bucket at key1 to the
2215 * one at key2 and updated their key pointer. We no longer need to
2216 * hold the references to key1.
2218 while (--drop_count
>= 0)
2219 drop_futex_key_refs(&key1
);
2222 put_futex_key(&key2
);
2224 put_futex_key(&key1
);
2226 return ret
? ret
: task_count
;
2229 /* The key must be already stored in q->key. */
2230 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2231 __acquires(&hb
->lock
)
2233 struct futex_hash_bucket
*hb
;
2235 hb
= hash_futex(&q
->key
);
2238 * Increment the counter before taking the lock so that
2239 * a potential waker won't miss a to-be-slept task that is
2240 * waiting for the spinlock. This is safe as all queue_lock()
2241 * users end up calling queue_me(). Similarly, for housekeeping,
2242 * decrement the counter at queue_unlock() when some error has
2243 * occurred and we don't end up adding the task to the list.
2247 q
->lock_ptr
= &hb
->lock
;
2249 spin_lock(&hb
->lock
); /* implies smp_mb(); (A) */
2254 queue_unlock(struct futex_hash_bucket
*hb
)
2255 __releases(&hb
->lock
)
2257 spin_unlock(&hb
->lock
);
2261 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2266 * The priority used to register this element is
2267 * - either the real thread-priority for the real-time threads
2268 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2269 * - or MAX_RT_PRIO for non-RT threads.
2270 * Thus, all RT-threads are woken first in priority order, and
2271 * the others are woken last, in FIFO order.
2273 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2275 plist_node_init(&q
->list
, prio
);
2276 plist_add(&q
->list
, &hb
->chain
);
2281 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2282 * @q: The futex_q to enqueue
2283 * @hb: The destination hash bucket
2285 * The hb->lock must be held by the caller, and is released here. A call to
2286 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2287 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2288 * or nothing if the unqueue is done as part of the wake process and the unqueue
2289 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2292 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2293 __releases(&hb
->lock
)
2296 spin_unlock(&hb
->lock
);
2300 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2301 * @q: The futex_q to unqueue
2303 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2304 * be paired with exactly one earlier call to queue_me().
2307 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2308 * - 0 - if the futex_q was already removed by the waking thread
2310 static int unqueue_me(struct futex_q
*q
)
2312 spinlock_t
*lock_ptr
;
2315 /* In the common case we don't take the spinlock, which is nice. */
2318 * q->lock_ptr can change between this read and the following spin_lock.
2319 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2320 * optimizing lock_ptr out of the logic below.
2322 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2323 if (lock_ptr
!= NULL
) {
2324 spin_lock(lock_ptr
);
2326 * q->lock_ptr can change between reading it and
2327 * spin_lock(), causing us to take the wrong lock. This
2328 * corrects the race condition.
2330 * Reasoning goes like this: if we have the wrong lock,
2331 * q->lock_ptr must have changed (maybe several times)
2332 * between reading it and the spin_lock(). It can
2333 * change again after the spin_lock() but only if it was
2334 * already changed before the spin_lock(). It cannot,
2335 * however, change back to the original value. Therefore
2336 * we can detect whether we acquired the correct lock.
2338 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2339 spin_unlock(lock_ptr
);
2344 BUG_ON(q
->pi_state
);
2346 spin_unlock(lock_ptr
);
2350 drop_futex_key_refs(&q
->key
);
2355 * PI futexes can not be requeued and must remove themself from the
2356 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2359 static void unqueue_me_pi(struct futex_q
*q
)
2360 __releases(q
->lock_ptr
)
2364 BUG_ON(!q
->pi_state
);
2365 put_pi_state(q
->pi_state
);
2368 spin_unlock(q
->lock_ptr
);
2371 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2372 struct task_struct
*argowner
)
2374 struct futex_pi_state
*pi_state
= q
->pi_state
;
2375 u32 uval
, uninitialized_var(curval
), newval
;
2376 struct task_struct
*oldowner
, *newowner
;
2380 lockdep_assert_held(q
->lock_ptr
);
2382 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2384 oldowner
= pi_state
->owner
;
2387 * We are here because either:
2389 * - we stole the lock and pi_state->owner needs updating to reflect
2390 * that (@argowner == current),
2394 * - someone stole our lock and we need to fix things to point to the
2395 * new owner (@argowner == NULL).
2397 * Either way, we have to replace the TID in the user space variable.
2398 * This must be atomic as we have to preserve the owner died bit here.
2400 * Note: We write the user space value _before_ changing the pi_state
2401 * because we can fault here. Imagine swapped out pages or a fork
2402 * that marked all the anonymous memory readonly for cow.
2404 * Modifying pi_state _before_ the user space value would leave the
2405 * pi_state in an inconsistent state when we fault here, because we
2406 * need to drop the locks to handle the fault. This might be observed
2407 * in the PID check in lookup_pi_state.
2411 if (oldowner
!= current
) {
2413 * We raced against a concurrent self; things are
2414 * already fixed up. Nothing to do.
2420 if (__rt_mutex_futex_trylock(&pi_state
->pi_mutex
)) {
2421 /* We got the lock after all, nothing to fix. */
2427 * Since we just failed the trylock; there must be an owner.
2429 newowner
= rt_mutex_owner(&pi_state
->pi_mutex
);
2432 WARN_ON_ONCE(argowner
!= current
);
2433 if (oldowner
== current
) {
2435 * We raced against a concurrent self; things are
2436 * already fixed up. Nothing to do.
2441 newowner
= argowner
;
2444 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2446 if (!pi_state
->owner
)
2447 newtid
|= FUTEX_OWNER_DIED
;
2449 err
= get_futex_value_locked(&uval
, uaddr
);
2454 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2456 err
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
);
2466 * We fixed up user space. Now we need to fix the pi_state
2469 if (pi_state
->owner
!= NULL
) {
2470 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2471 WARN_ON(list_empty(&pi_state
->list
));
2472 list_del_init(&pi_state
->list
);
2473 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2476 pi_state
->owner
= newowner
;
2478 raw_spin_lock(&newowner
->pi_lock
);
2479 WARN_ON(!list_empty(&pi_state
->list
));
2480 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2481 raw_spin_unlock(&newowner
->pi_lock
);
2482 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2487 * In order to reschedule or handle a page fault, we need to drop the
2488 * locks here. In the case of a fault, this gives the other task
2489 * (either the highest priority waiter itself or the task which stole
2490 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2491 * are back from handling the fault we need to check the pi_state after
2492 * reacquiring the locks and before trying to do another fixup. When
2493 * the fixup has been done already we simply return.
2495 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2496 * drop hb->lock since the caller owns the hb -> futex_q relation.
2497 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2500 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2501 spin_unlock(q
->lock_ptr
);
2505 ret
= fault_in_user_writeable(uaddr
);
2519 spin_lock(q
->lock_ptr
);
2520 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2523 * Check if someone else fixed it for us:
2525 if (pi_state
->owner
!= oldowner
) {
2536 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2540 static long futex_wait_restart(struct restart_block
*restart
);
2543 * fixup_owner() - Post lock pi_state and corner case management
2544 * @uaddr: user address of the futex
2545 * @q: futex_q (contains pi_state and access to the rt_mutex)
2546 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2548 * After attempting to lock an rt_mutex, this function is called to cleanup
2549 * the pi_state owner as well as handle race conditions that may allow us to
2550 * acquire the lock. Must be called with the hb lock held.
2553 * - 1 - success, lock taken;
2554 * - 0 - success, lock not taken;
2555 * - <0 - on error (-EFAULT)
2557 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2563 * Got the lock. We might not be the anticipated owner if we
2564 * did a lock-steal - fix up the PI-state in that case:
2566 * Speculative pi_state->owner read (we don't hold wait_lock);
2567 * since we own the lock pi_state->owner == current is the
2568 * stable state, anything else needs more attention.
2570 if (q
->pi_state
->owner
!= current
)
2571 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2576 * If we didn't get the lock; check if anybody stole it from us. In
2577 * that case, we need to fix up the uval to point to them instead of
2578 * us, otherwise bad things happen. [10]
2580 * Another speculative read; pi_state->owner == current is unstable
2581 * but needs our attention.
2583 if (q
->pi_state
->owner
== current
) {
2584 ret
= fixup_pi_state_owner(uaddr
, q
, NULL
);
2589 * Paranoia check. If we did not take the lock, then we should not be
2590 * the owner of the rt_mutex.
2592 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2593 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2594 "pi-state %p\n", ret
,
2595 q
->pi_state
->pi_mutex
.owner
,
2596 q
->pi_state
->owner
);
2600 return ret
? ret
: locked
;
2604 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2605 * @hb: the futex hash bucket, must be locked by the caller
2606 * @q: the futex_q to queue up on
2607 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2609 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2610 struct hrtimer_sleeper
*timeout
)
2613 * The task state is guaranteed to be set before another task can
2614 * wake it. set_current_state() is implemented using smp_store_mb() and
2615 * queue_me() calls spin_unlock() upon completion, both serializing
2616 * access to the hash list and forcing another memory barrier.
2618 set_current_state(TASK_INTERRUPTIBLE
);
2623 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2626 * If we have been removed from the hash list, then another task
2627 * has tried to wake us, and we can skip the call to schedule().
2629 if (likely(!plist_node_empty(&q
->list
))) {
2631 * If the timer has already expired, current will already be
2632 * flagged for rescheduling. Only call schedule if there
2633 * is no timeout, or if it has yet to expire.
2635 if (!timeout
|| timeout
->task
)
2636 freezable_schedule();
2638 __set_current_state(TASK_RUNNING
);
2642 * futex_wait_setup() - Prepare to wait on a futex
2643 * @uaddr: the futex userspace address
2644 * @val: the expected value
2645 * @flags: futex flags (FLAGS_SHARED, etc.)
2646 * @q: the associated futex_q
2647 * @hb: storage for hash_bucket pointer to be returned to caller
2649 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2650 * compare it with the expected value. Handle atomic faults internally.
2651 * Return with the hb lock held and a q.key reference on success, and unlocked
2652 * with no q.key reference on failure.
2655 * - 0 - uaddr contains val and hb has been locked;
2656 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2658 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2659 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2665 * Access the page AFTER the hash-bucket is locked.
2666 * Order is important:
2668 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2669 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2671 * The basic logical guarantee of a futex is that it blocks ONLY
2672 * if cond(var) is known to be true at the time of blocking, for
2673 * any cond. If we locked the hash-bucket after testing *uaddr, that
2674 * would open a race condition where we could block indefinitely with
2675 * cond(var) false, which would violate the guarantee.
2677 * On the other hand, we insert q and release the hash-bucket only
2678 * after testing *uaddr. This guarantees that futex_wait() will NOT
2679 * absorb a wakeup if *uaddr does not match the desired values
2680 * while the syscall executes.
2683 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2684 if (unlikely(ret
!= 0))
2688 *hb
= queue_lock(q
);
2690 ret
= get_futex_value_locked(&uval
, uaddr
);
2695 ret
= get_user(uval
, uaddr
);
2699 if (!(flags
& FLAGS_SHARED
))
2702 put_futex_key(&q
->key
);
2713 put_futex_key(&q
->key
);
2717 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2718 ktime_t
*abs_time
, u32 bitset
)
2720 struct hrtimer_sleeper timeout
, *to
= NULL
;
2721 struct restart_block
*restart
;
2722 struct futex_hash_bucket
*hb
;
2723 struct futex_q q
= futex_q_init
;
2733 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2734 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2736 hrtimer_init_sleeper(to
, current
);
2737 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2738 current
->timer_slack_ns
);
2743 * Prepare to wait on uaddr. On success, holds hb lock and increments
2746 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2750 /* queue_me and wait for wakeup, timeout, or a signal. */
2751 futex_wait_queue_me(hb
, &q
, to
);
2753 /* If we were woken (and unqueued), we succeeded, whatever. */
2755 /* unqueue_me() drops q.key ref */
2756 if (!unqueue_me(&q
))
2759 if (to
&& !to
->task
)
2763 * We expect signal_pending(current), but we might be the
2764 * victim of a spurious wakeup as well.
2766 if (!signal_pending(current
))
2773 restart
= ¤t
->restart_block
;
2774 restart
->fn
= futex_wait_restart
;
2775 restart
->futex
.uaddr
= uaddr
;
2776 restart
->futex
.val
= val
;
2777 restart
->futex
.time
= *abs_time
;
2778 restart
->futex
.bitset
= bitset
;
2779 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2781 ret
= -ERESTART_RESTARTBLOCK
;
2785 hrtimer_cancel(&to
->timer
);
2786 destroy_hrtimer_on_stack(&to
->timer
);
2792 static long futex_wait_restart(struct restart_block
*restart
)
2794 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2795 ktime_t t
, *tp
= NULL
;
2797 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2798 t
= restart
->futex
.time
;
2801 restart
->fn
= do_no_restart_syscall
;
2803 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2804 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2809 * Userspace tried a 0 -> TID atomic transition of the futex value
2810 * and failed. The kernel side here does the whole locking operation:
2811 * if there are waiters then it will block as a consequence of relying
2812 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2813 * a 0 value of the futex too.).
2815 * Also serves as futex trylock_pi()'ing, and due semantics.
2817 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2818 ktime_t
*time
, int trylock
)
2820 struct hrtimer_sleeper timeout
, *to
= NULL
;
2821 struct futex_pi_state
*pi_state
= NULL
;
2822 struct rt_mutex_waiter rt_waiter
;
2823 struct futex_hash_bucket
*hb
;
2824 struct futex_q q
= futex_q_init
;
2827 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2830 if (refill_pi_state_cache())
2835 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2837 hrtimer_init_sleeper(to
, current
);
2838 hrtimer_set_expires(&to
->timer
, *time
);
2842 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2843 if (unlikely(ret
!= 0))
2847 hb
= queue_lock(&q
);
2849 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2850 if (unlikely(ret
)) {
2852 * Atomic work succeeded and we got the lock,
2853 * or failed. Either way, we do _not_ block.
2857 /* We got the lock. */
2859 goto out_unlock_put_key
;
2864 * Two reasons for this:
2865 * - Task is exiting and we just wait for the
2867 * - The user space value changed.
2870 put_futex_key(&q
.key
);
2874 goto out_unlock_put_key
;
2878 WARN_ON(!q
.pi_state
);
2881 * Only actually queue now that the atomic ops are done:
2886 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2887 /* Fixup the trylock return value: */
2888 ret
= ret
? 0 : -EWOULDBLOCK
;
2892 rt_mutex_init_waiter(&rt_waiter
);
2895 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2896 * hold it while doing rt_mutex_start_proxy(), because then it will
2897 * include hb->lock in the blocking chain, even through we'll not in
2898 * fact hold it while blocking. This will lead it to report -EDEADLK
2899 * and BUG when futex_unlock_pi() interleaves with this.
2901 * Therefore acquire wait_lock while holding hb->lock, but drop the
2902 * latter before calling __rt_mutex_start_proxy_lock(). This
2903 * interleaves with futex_unlock_pi() -- which does a similar lock
2904 * handoff -- such that the latter can observe the futex_q::pi_state
2905 * before __rt_mutex_start_proxy_lock() is done.
2907 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2908 spin_unlock(q
.lock_ptr
);
2910 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2911 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2912 * it sees the futex_q::pi_state.
2914 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2915 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2924 hrtimer_start_expires(&to
->timer
, HRTIMER_MODE_ABS
);
2926 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2929 spin_lock(q
.lock_ptr
);
2931 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2932 * first acquire the hb->lock before removing the lock from the
2933 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2936 * In particular; it is important that futex_unlock_pi() can not
2937 * observe this inconsistency.
2939 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2944 * Fixup the pi_state owner and possibly acquire the lock if we
2947 res
= fixup_owner(uaddr
, &q
, !ret
);
2949 * If fixup_owner() returned an error, proprogate that. If it acquired
2950 * the lock, clear our -ETIMEDOUT or -EINTR.
2953 ret
= (res
< 0) ? res
: 0;
2956 * If fixup_owner() faulted and was unable to handle the fault, unlock
2957 * it and return the fault to userspace.
2959 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2960 pi_state
= q
.pi_state
;
2961 get_pi_state(pi_state
);
2964 /* Unqueue and drop the lock */
2968 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2969 put_pi_state(pi_state
);
2978 put_futex_key(&q
.key
);
2981 hrtimer_cancel(&to
->timer
);
2982 destroy_hrtimer_on_stack(&to
->timer
);
2984 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2989 ret
= fault_in_user_writeable(uaddr
);
2993 if (!(flags
& FLAGS_SHARED
))
2996 put_futex_key(&q
.key
);
3001 * Userspace attempted a TID -> 0 atomic transition, and failed.
3002 * This is the in-kernel slowpath: we look up the PI state (if any),
3003 * and do the rt-mutex unlock.
3005 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
3007 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
3008 union futex_key key
= FUTEX_KEY_INIT
;
3009 struct futex_hash_bucket
*hb
;
3010 struct futex_q
*top_waiter
;
3013 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3017 if (get_user(uval
, uaddr
))
3020 * We release only a lock we actually own:
3022 if ((uval
& FUTEX_TID_MASK
) != vpid
)
3025 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
3029 hb
= hash_futex(&key
);
3030 spin_lock(&hb
->lock
);
3033 * Check waiters first. We do not trust user space values at
3034 * all and we at least want to know if user space fiddled
3035 * with the futex value instead of blindly unlocking.
3037 top_waiter
= futex_top_waiter(hb
, &key
);
3039 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
3046 * If current does not own the pi_state then the futex is
3047 * inconsistent and user space fiddled with the futex value.
3049 if (pi_state
->owner
!= current
)
3052 get_pi_state(pi_state
);
3054 * By taking wait_lock while still holding hb->lock, we ensure
3055 * there is no point where we hold neither; and therefore
3056 * wake_futex_pi() must observe a state consistent with what we
3059 * In particular; this forces __rt_mutex_start_proxy() to
3060 * complete such that we're guaranteed to observe the
3061 * rt_waiter. Also see the WARN in wake_futex_pi().
3063 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
3064 spin_unlock(&hb
->lock
);
3066 /* drops pi_state->pi_mutex.wait_lock */
3067 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
3069 put_pi_state(pi_state
);
3072 * Success, we're done! No tricky corner cases.
3077 * The atomic access to the futex value generated a
3078 * pagefault, so retry the user-access and the wakeup:
3083 * A unconditional UNLOCK_PI op raced against a waiter
3084 * setting the FUTEX_WAITERS bit. Try again.
3089 * wake_futex_pi has detected invalid state. Tell user
3096 * We have no kernel internal state, i.e. no waiters in the
3097 * kernel. Waiters which are about to queue themselves are stuck
3098 * on hb->lock. So we can safely ignore them. We do neither
3099 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3102 if ((ret
= cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))) {
3103 spin_unlock(&hb
->lock
);
3118 * If uval has changed, let user space handle it.
3120 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
3123 spin_unlock(&hb
->lock
);
3125 put_futex_key(&key
);
3129 put_futex_key(&key
);
3134 put_futex_key(&key
);
3136 ret
= fault_in_user_writeable(uaddr
);
3144 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3145 * @hb: the hash_bucket futex_q was original enqueued on
3146 * @q: the futex_q woken while waiting to be requeued
3147 * @key2: the futex_key of the requeue target futex
3148 * @timeout: the timeout associated with the wait (NULL if none)
3150 * Detect if the task was woken on the initial futex as opposed to the requeue
3151 * target futex. If so, determine if it was a timeout or a signal that caused
3152 * the wakeup and return the appropriate error code to the caller. Must be
3153 * called with the hb lock held.
3156 * - 0 = no early wakeup detected;
3157 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3160 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
3161 struct futex_q
*q
, union futex_key
*key2
,
3162 struct hrtimer_sleeper
*timeout
)
3167 * With the hb lock held, we avoid races while we process the wakeup.
3168 * We only need to hold hb (and not hb2) to ensure atomicity as the
3169 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3170 * It can't be requeued from uaddr2 to something else since we don't
3171 * support a PI aware source futex for requeue.
3173 if (!match_futex(&q
->key
, key2
)) {
3174 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
3176 * We were woken prior to requeue by a timeout or a signal.
3177 * Unqueue the futex_q and determine which it was.
3179 plist_del(&q
->list
, &hb
->chain
);
3182 /* Handle spurious wakeups gracefully */
3184 if (timeout
&& !timeout
->task
)
3186 else if (signal_pending(current
))
3187 ret
= -ERESTARTNOINTR
;
3193 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3194 * @uaddr: the futex we initially wait on (non-pi)
3195 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3196 * the same type, no requeueing from private to shared, etc.
3197 * @val: the expected value of uaddr
3198 * @abs_time: absolute timeout
3199 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3200 * @uaddr2: the pi futex we will take prior to returning to user-space
3202 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3203 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3204 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3205 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3206 * without one, the pi logic would not know which task to boost/deboost, if
3207 * there was a need to.
3209 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3210 * via the following--
3211 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3212 * 2) wakeup on uaddr2 after a requeue
3216 * If 3, cleanup and return -ERESTARTNOINTR.
3218 * If 2, we may then block on trying to take the rt_mutex and return via:
3219 * 5) successful lock
3222 * 8) other lock acquisition failure
3224 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3226 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3232 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3233 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3236 struct hrtimer_sleeper timeout
, *to
= NULL
;
3237 struct futex_pi_state
*pi_state
= NULL
;
3238 struct rt_mutex_waiter rt_waiter
;
3239 struct futex_hash_bucket
*hb
;
3240 union futex_key key2
= FUTEX_KEY_INIT
;
3241 struct futex_q q
= futex_q_init
;
3244 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3247 if (uaddr
== uaddr2
)
3255 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
3256 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
3258 hrtimer_init_sleeper(to
, current
);
3259 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
3260 current
->timer_slack_ns
);
3264 * The waiter is allocated on our stack, manipulated by the requeue
3265 * code while we sleep on uaddr.
3267 rt_mutex_init_waiter(&rt_waiter
);
3269 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
3270 if (unlikely(ret
!= 0))
3274 q
.rt_waiter
= &rt_waiter
;
3275 q
.requeue_pi_key
= &key2
;
3278 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3281 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3286 * The check above which compares uaddrs is not sufficient for
3287 * shared futexes. We need to compare the keys:
3289 if (match_futex(&q
.key
, &key2
)) {
3295 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3296 futex_wait_queue_me(hb
, &q
, to
);
3298 spin_lock(&hb
->lock
);
3299 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3300 spin_unlock(&hb
->lock
);
3305 * In order for us to be here, we know our q.key == key2, and since
3306 * we took the hb->lock above, we also know that futex_requeue() has
3307 * completed and we no longer have to concern ourselves with a wakeup
3308 * race with the atomic proxy lock acquisition by the requeue code. The
3309 * futex_requeue dropped our key1 reference and incremented our key2
3313 /* Check if the requeue code acquired the second futex for us. */
3316 * Got the lock. We might not be the anticipated owner if we
3317 * did a lock-steal - fix up the PI-state in that case.
3319 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3320 spin_lock(q
.lock_ptr
);
3321 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3322 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3323 pi_state
= q
.pi_state
;
3324 get_pi_state(pi_state
);
3327 * Drop the reference to the pi state which
3328 * the requeue_pi() code acquired for us.
3330 put_pi_state(q
.pi_state
);
3331 spin_unlock(q
.lock_ptr
);
3334 struct rt_mutex
*pi_mutex
;
3337 * We have been woken up by futex_unlock_pi(), a timeout, or a
3338 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3341 WARN_ON(!q
.pi_state
);
3342 pi_mutex
= &q
.pi_state
->pi_mutex
;
3343 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3345 spin_lock(q
.lock_ptr
);
3346 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3349 debug_rt_mutex_free_waiter(&rt_waiter
);
3351 * Fixup the pi_state owner and possibly acquire the lock if we
3354 res
= fixup_owner(uaddr2
, &q
, !ret
);
3356 * If fixup_owner() returned an error, proprogate that. If it
3357 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3360 ret
= (res
< 0) ? res
: 0;
3363 * If fixup_pi_state_owner() faulted and was unable to handle
3364 * the fault, unlock the rt_mutex and return the fault to
3367 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3368 pi_state
= q
.pi_state
;
3369 get_pi_state(pi_state
);
3372 /* Unqueue and drop the lock. */
3377 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3378 put_pi_state(pi_state
);
3381 if (ret
== -EINTR
) {
3383 * We've already been requeued, but cannot restart by calling
3384 * futex_lock_pi() directly. We could restart this syscall, but
3385 * it would detect that the user space "val" changed and return
3386 * -EWOULDBLOCK. Save the overhead of the restart and return
3387 * -EWOULDBLOCK directly.
3393 put_futex_key(&q
.key
);
3395 put_futex_key(&key2
);
3399 hrtimer_cancel(&to
->timer
);
3400 destroy_hrtimer_on_stack(&to
->timer
);
3406 * Support for robust futexes: the kernel cleans up held futexes at
3409 * Implementation: user-space maintains a per-thread list of locks it
3410 * is holding. Upon do_exit(), the kernel carefully walks this list,
3411 * and marks all locks that are owned by this thread with the
3412 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3413 * always manipulated with the lock held, so the list is private and
3414 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3415 * field, to allow the kernel to clean up if the thread dies after
3416 * acquiring the lock, but just before it could have added itself to
3417 * the list. There can only be one such pending lock.
3421 * sys_set_robust_list() - Set the robust-futex list head of a task
3422 * @head: pointer to the list-head
3423 * @len: length of the list-head, as userspace expects
3425 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3428 if (!futex_cmpxchg_enabled
)
3431 * The kernel knows only one size for now:
3433 if (unlikely(len
!= sizeof(*head
)))
3436 current
->robust_list
= head
;
3442 * sys_get_robust_list() - Get the robust-futex list head of a task
3443 * @pid: pid of the process [zero for current task]
3444 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3445 * @len_ptr: pointer to a length field, the kernel fills in the header size
3447 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3448 struct robust_list_head __user
* __user
*, head_ptr
,
3449 size_t __user
*, len_ptr
)
3451 struct robust_list_head __user
*head
;
3453 struct task_struct
*p
;
3455 if (!futex_cmpxchg_enabled
)
3464 p
= find_task_by_vpid(pid
);
3470 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3473 head
= p
->robust_list
;
3476 if (put_user(sizeof(*head
), len_ptr
))
3478 return put_user(head
, head_ptr
);
3486 /* Constants for the pending_op argument of handle_futex_death */
3487 #define HANDLE_DEATH_PENDING true
3488 #define HANDLE_DEATH_LIST false
3491 * Process a futex-list entry, check whether it's owned by the
3492 * dying task, and do notification if so:
3494 static int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
,
3495 bool pi
, bool pending_op
)
3497 u32 uval
, uninitialized_var(nval
), mval
;
3500 /* Futex address must be 32bit aligned */
3501 if ((((unsigned long)uaddr
) % sizeof(*uaddr
)) != 0)
3505 if (get_user(uval
, uaddr
))
3509 * Special case for regular (non PI) futexes. The unlock path in
3510 * user space has two race scenarios:
3512 * 1. The unlock path releases the user space futex value and
3513 * before it can execute the futex() syscall to wake up
3514 * waiters it is killed.
3516 * 2. A woken up waiter is killed before it can acquire the
3517 * futex in user space.
3519 * In both cases the TID validation below prevents a wakeup of
3520 * potential waiters which can cause these waiters to block
3523 * In both cases the following conditions are met:
3525 * 1) task->robust_list->list_op_pending != NULL
3526 * @pending_op == true
3527 * 2) User space futex value == 0
3528 * 3) Regular futex: @pi == false
3530 * If these conditions are met, it is safe to attempt waking up a
3531 * potential waiter without touching the user space futex value and
3532 * trying to set the OWNER_DIED bit. The user space futex value is
3533 * uncontended and the rest of the user space mutex state is
3534 * consistent, so a woken waiter will just take over the
3535 * uncontended futex. Setting the OWNER_DIED bit would create
3536 * inconsistent state and malfunction of the user space owner died
3539 if (pending_op
&& !pi
&& !uval
) {
3540 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3544 if ((uval
& FUTEX_TID_MASK
) != task_pid_vnr(curr
))
3548 * Ok, this dying thread is truly holding a futex
3549 * of interest. Set the OWNER_DIED bit atomically
3550 * via cmpxchg, and if the value had FUTEX_WAITERS
3551 * set, wake up a waiter (if any). (We have to do a
3552 * futex_wake() even if OWNER_DIED is already set -
3553 * to handle the rare but possible case of recursive
3554 * thread-death.) The rest of the cleanup is done in
3557 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3560 * We are not holding a lock here, but we want to have
3561 * the pagefault_disable/enable() protection because
3562 * we want to handle the fault gracefully. If the
3563 * access fails we try to fault in the futex with R/W
3564 * verification via get_user_pages. get_user() above
3565 * does not guarantee R/W access. If that fails we
3566 * give up and leave the futex locked.
3568 if ((err
= cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
))) {
3571 if (fault_in_user_writeable(uaddr
))
3589 * Wake robust non-PI futexes here. The wakeup of
3590 * PI futexes happens in exit_pi_state():
3592 if (!pi
&& (uval
& FUTEX_WAITERS
))
3593 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3599 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3601 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3602 struct robust_list __user
* __user
*head
,
3605 unsigned long uentry
;
3607 if (get_user(uentry
, (unsigned long __user
*)head
))
3610 *entry
= (void __user
*)(uentry
& ~1UL);
3617 * Walk curr->robust_list (very carefully, it's a userspace list!)
3618 * and mark any locks found there dead, and notify any waiters.
3620 * We silently return on any sign of list-walking problem.
3622 static void exit_robust_list(struct task_struct
*curr
)
3624 struct robust_list_head __user
*head
= curr
->robust_list
;
3625 struct robust_list __user
*entry
, *next_entry
, *pending
;
3626 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3627 unsigned int uninitialized_var(next_pi
);
3628 unsigned long futex_offset
;
3631 if (!futex_cmpxchg_enabled
)
3635 * Fetch the list head (which was registered earlier, via
3636 * sys_set_robust_list()):
3638 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3641 * Fetch the relative futex offset:
3643 if (get_user(futex_offset
, &head
->futex_offset
))
3646 * Fetch any possibly pending lock-add first, and handle it
3649 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3652 next_entry
= NULL
; /* avoid warning with gcc */
3653 while (entry
!= &head
->list
) {
3655 * Fetch the next entry in the list before calling
3656 * handle_futex_death:
3658 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3660 * A pending lock might already be on the list, so
3661 * don't process it twice:
3663 if (entry
!= pending
) {
3664 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3665 curr
, pi
, HANDLE_DEATH_LIST
))
3673 * Avoid excessively long or circular lists:
3682 handle_futex_death((void __user
*)pending
+ futex_offset
,
3683 curr
, pip
, HANDLE_DEATH_PENDING
);
3687 void futex_mm_release(struct task_struct
*tsk
)
3689 if (unlikely(tsk
->robust_list
)) {
3690 exit_robust_list(tsk
);
3691 tsk
->robust_list
= NULL
;
3694 #ifdef CONFIG_COMPAT
3695 if (unlikely(tsk
->compat_robust_list
)) {
3696 compat_exit_robust_list(tsk
);
3697 tsk
->compat_robust_list
= NULL
;
3701 if (unlikely(!list_empty(&tsk
->pi_state_list
)))
3702 exit_pi_state_list(tsk
);
3705 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3706 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3708 int cmd
= op
& FUTEX_CMD_MASK
;
3709 unsigned int flags
= 0;
3711 if (!(op
& FUTEX_PRIVATE_FLAG
))
3712 flags
|= FLAGS_SHARED
;
3714 if (op
& FUTEX_CLOCK_REALTIME
) {
3715 flags
|= FLAGS_CLOCKRT
;
3716 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3717 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3723 case FUTEX_UNLOCK_PI
:
3724 case FUTEX_TRYLOCK_PI
:
3725 case FUTEX_WAIT_REQUEUE_PI
:
3726 case FUTEX_CMP_REQUEUE_PI
:
3727 if (!futex_cmpxchg_enabled
)
3733 val3
= FUTEX_BITSET_MATCH_ANY
;
3734 case FUTEX_WAIT_BITSET
:
3735 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3737 val3
= FUTEX_BITSET_MATCH_ANY
;
3738 case FUTEX_WAKE_BITSET
:
3739 return futex_wake(uaddr
, flags
, val
, val3
);
3741 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3742 case FUTEX_CMP_REQUEUE
:
3743 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3745 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3747 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3748 case FUTEX_UNLOCK_PI
:
3749 return futex_unlock_pi(uaddr
, flags
);
3750 case FUTEX_TRYLOCK_PI
:
3751 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3752 case FUTEX_WAIT_REQUEUE_PI
:
3753 val3
= FUTEX_BITSET_MATCH_ANY
;
3754 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3756 case FUTEX_CMP_REQUEUE_PI
:
3757 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3763 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3764 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3768 ktime_t t
, *tp
= NULL
;
3770 int cmd
= op
& FUTEX_CMD_MASK
;
3772 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3773 cmd
== FUTEX_WAIT_BITSET
||
3774 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3775 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3777 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3779 if (!timespec_valid(&ts
))
3782 t
= timespec_to_ktime(ts
);
3783 if (cmd
== FUTEX_WAIT
)
3784 t
= ktime_add_safe(ktime_get(), t
);
3788 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3789 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3791 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3792 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3793 val2
= (u32
) (unsigned long) utime
;
3795 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3798 #ifdef CONFIG_COMPAT
3800 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3803 compat_fetch_robust_entry(compat_uptr_t
*uentry
, struct robust_list __user
**entry
,
3804 compat_uptr_t __user
*head
, unsigned int *pi
)
3806 if (get_user(*uentry
, head
))
3809 *entry
= compat_ptr((*uentry
) & ~1);
3810 *pi
= (unsigned int)(*uentry
) & 1;
3815 static void __user
*futex_uaddr(struct robust_list __user
*entry
,
3816 compat_long_t futex_offset
)
3818 compat_uptr_t base
= ptr_to_compat(entry
);
3819 void __user
*uaddr
= compat_ptr(base
+ futex_offset
);
3825 * Walk curr->robust_list (very carefully, it's a userspace list!)
3826 * and mark any locks found there dead, and notify any waiters.
3828 * We silently return on any sign of list-walking problem.
3830 static void compat_exit_robust_list(struct task_struct
*curr
)
3832 struct compat_robust_list_head __user
*head
= curr
->compat_robust_list
;
3833 struct robust_list __user
*entry
, *next_entry
, *pending
;
3834 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3835 unsigned int uninitialized_var(next_pi
);
3836 compat_uptr_t uentry
, next_uentry
, upending
;
3837 compat_long_t futex_offset
;
3840 if (!futex_cmpxchg_enabled
)
3844 * Fetch the list head (which was registered earlier, via
3845 * sys_set_robust_list()):
3847 if (compat_fetch_robust_entry(&uentry
, &entry
, &head
->list
.next
, &pi
))
3850 * Fetch the relative futex offset:
3852 if (get_user(futex_offset
, &head
->futex_offset
))
3855 * Fetch any possibly pending lock-add first, and handle it
3858 if (compat_fetch_robust_entry(&upending
, &pending
,
3859 &head
->list_op_pending
, &pip
))
3862 next_entry
= NULL
; /* avoid warning with gcc */
3863 while (entry
!= (struct robust_list __user
*) &head
->list
) {
3865 * Fetch the next entry in the list before calling
3866 * handle_futex_death:
3868 rc
= compat_fetch_robust_entry(&next_uentry
, &next_entry
,
3869 (compat_uptr_t __user
*)&entry
->next
, &next_pi
);
3871 * A pending lock might already be on the list, so
3872 * dont process it twice:
3874 if (entry
!= pending
) {
3875 void __user
*uaddr
= futex_uaddr(entry
, futex_offset
);
3877 if (handle_futex_death(uaddr
, curr
, pi
,
3883 uentry
= next_uentry
;
3887 * Avoid excessively long or circular lists:
3895 void __user
*uaddr
= futex_uaddr(pending
, futex_offset
);
3897 handle_futex_death(uaddr
, curr
, pip
, HANDLE_DEATH_PENDING
);
3901 COMPAT_SYSCALL_DEFINE2(set_robust_list
,
3902 struct compat_robust_list_head __user
*, head
,
3905 if (!futex_cmpxchg_enabled
)
3908 if (unlikely(len
!= sizeof(*head
)))
3911 current
->compat_robust_list
= head
;
3916 COMPAT_SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3917 compat_uptr_t __user
*, head_ptr
,
3918 compat_size_t __user
*, len_ptr
)
3920 struct compat_robust_list_head __user
*head
;
3922 struct task_struct
*p
;
3924 if (!futex_cmpxchg_enabled
)
3933 p
= find_task_by_vpid(pid
);
3939 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3942 head
= p
->compat_robust_list
;
3945 if (put_user(sizeof(*head
), len_ptr
))
3947 return put_user(ptr_to_compat(head
), head_ptr
);
3955 COMPAT_SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3956 struct compat_timespec __user
*, utime
, u32 __user
*, uaddr2
,
3960 ktime_t t
, *tp
= NULL
;
3962 int cmd
= op
& FUTEX_CMD_MASK
;
3964 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3965 cmd
== FUTEX_WAIT_BITSET
||
3966 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3967 if (compat_get_timespec(&ts
, utime
))
3969 if (!timespec_valid(&ts
))
3972 t
= timespec_to_ktime(ts
);
3973 if (cmd
== FUTEX_WAIT
)
3974 t
= ktime_add_safe(ktime_get(), t
);
3977 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3978 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3979 val2
= (int) (unsigned long) utime
;
3981 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3983 #endif /* CONFIG_COMPAT */
3985 static void __init
futex_detect_cmpxchg(void)
3987 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3991 * This will fail and we want it. Some arch implementations do
3992 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3993 * functionality. We want to know that before we call in any
3994 * of the complex code paths. Also we want to prevent
3995 * registration of robust lists in that case. NULL is
3996 * guaranteed to fault and we get -EFAULT on functional
3997 * implementation, the non-functional ones will return
4000 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
4001 futex_cmpxchg_enabled
= 1;
4005 static int __init
futex_init(void)
4007 unsigned int futex_shift
;
4010 #if CONFIG_BASE_SMALL
4011 futex_hashsize
= 16;
4013 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
4016 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
4018 futex_hashsize
< 256 ? HASH_SMALL
: 0,
4020 futex_hashsize
, futex_hashsize
);
4021 futex_hashsize
= 1UL << futex_shift
;
4023 futex_detect_cmpxchg();
4025 for (i
= 0; i
< futex_hashsize
; i
++) {
4026 atomic_set(&futex_queues
[i
].waiters
, 0);
4027 plist_head_init(&futex_queues
[i
].chain
);
4028 spin_lock_init(&futex_queues
[i
].lock
);
4033 core_initcall(futex_init
);