2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/wake_q.h>
65 #include <linux/sched/mm.h>
66 #include <linux/hugetlb.h>
67 #include <linux/freezer.h>
68 #include <linux/bootmem.h>
69 #include <linux/fault-inject.h>
71 #include <asm/futex.h>
73 #include "locking/rtmutex_common.h"
76 * READ this before attempting to hack on futexes!
78 * Basic futex operation and ordering guarantees
79 * =============================================
81 * The waiter reads the futex value in user space and calls
82 * futex_wait(). This function computes the hash bucket and acquires
83 * the hash bucket lock. After that it reads the futex user space value
84 * again and verifies that the data has not changed. If it has not changed
85 * it enqueues itself into the hash bucket, releases the hash bucket lock
88 * The waker side modifies the user space value of the futex and calls
89 * futex_wake(). This function computes the hash bucket and acquires the
90 * hash bucket lock. Then it looks for waiters on that futex in the hash
91 * bucket and wakes them.
93 * In futex wake up scenarios where no tasks are blocked on a futex, taking
94 * the hb spinlock can be avoided and simply return. In order for this
95 * optimization to work, ordering guarantees must exist so that the waiter
96 * being added to the list is acknowledged when the list is concurrently being
97 * checked by the waker, avoiding scenarios like the following:
101 * sys_futex(WAIT, futex, val);
102 * futex_wait(futex, val);
105 * sys_futex(WAKE, futex);
110 * lock(hash_bucket(futex));
112 * unlock(hash_bucket(futex));
115 * This would cause the waiter on CPU 0 to wait forever because it
116 * missed the transition of the user space value from val to newval
117 * and the waker did not find the waiter in the hash bucket queue.
119 * The correct serialization ensures that a waiter either observes
120 * the changed user space value before blocking or is woken by a
125 * sys_futex(WAIT, futex, val);
126 * futex_wait(futex, val);
129 * smp_mb(); (A) <-- paired with -.
131 * lock(hash_bucket(futex)); |
135 * | sys_futex(WAKE, futex);
136 * | futex_wake(futex);
138 * `--------> smp_mb(); (B)
141 * unlock(hash_bucket(futex));
142 * schedule(); if (waiters)
143 * lock(hash_bucket(futex));
144 * else wake_waiters(futex);
145 * waiters--; (b) unlock(hash_bucket(futex));
147 * Where (A) orders the waiters increment and the futex value read through
148 * atomic operations (see hb_waiters_inc) and where (B) orders the write
149 * to futex and the waiters read -- this is done by the barriers for both
150 * shared and private futexes in get_futex_key_refs().
152 * This yields the following case (where X:=waiters, Y:=futex):
160 * Which guarantees that x==0 && y==0 is impossible; which translates back into
161 * the guarantee that we cannot both miss the futex variable change and the
164 * Note that a new waiter is accounted for in (a) even when it is possible that
165 * the wait call can return error, in which case we backtrack from it in (b).
166 * Refer to the comment in queue_lock().
168 * Similarly, in order to account for waiters being requeued on another
169 * address we always increment the waiters for the destination bucket before
170 * acquiring the lock. It then decrements them again after releasing it -
171 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172 * will do the additional required waiter count housekeeping. This is done for
173 * double_lock_hb() and double_unlock_hb(), respectively.
176 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
177 int __read_mostly futex_cmpxchg_enabled
;
181 * Futex flags used to encode options to functions and preserve them across
185 # define FLAGS_SHARED 0x01
188 * NOMMU does not have per process address space. Let the compiler optimize
191 # define FLAGS_SHARED 0x00
193 #define FLAGS_CLOCKRT 0x02
194 #define FLAGS_HAS_TIMEOUT 0x04
197 * Priority Inheritance state:
199 struct futex_pi_state
{
201 * list of 'owned' pi_state instances - these have to be
202 * cleaned up in do_exit() if the task exits prematurely:
204 struct list_head list
;
209 struct rt_mutex pi_mutex
;
211 struct task_struct
*owner
;
215 } __randomize_layout
;
218 * struct futex_q - The hashed futex queue entry, one per waiting task
219 * @list: priority-sorted list of tasks waiting on this futex
220 * @task: the task waiting on the futex
221 * @lock_ptr: the hash bucket lock
222 * @key: the key the futex is hashed on
223 * @pi_state: optional priority inheritance state
224 * @rt_waiter: rt_waiter storage for use with requeue_pi
225 * @requeue_pi_key: the requeue_pi target futex key
226 * @bitset: bitset for the optional bitmasked wakeup
228 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
229 * we can wake only the relevant ones (hashed queues may be shared).
231 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
232 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
233 * The order of wakeup is always to make the first condition true, then
236 * PI futexes are typically woken before they are removed from the hash list via
237 * the rt_mutex code. See unqueue_me_pi().
240 struct plist_node list
;
242 struct task_struct
*task
;
243 spinlock_t
*lock_ptr
;
245 struct futex_pi_state
*pi_state
;
246 struct rt_mutex_waiter
*rt_waiter
;
247 union futex_key
*requeue_pi_key
;
249 } __randomize_layout
;
251 static const struct futex_q futex_q_init
= {
252 /* list gets initialized in queue_me()*/
253 .key
= FUTEX_KEY_INIT
,
254 .bitset
= FUTEX_BITSET_MATCH_ANY
258 * Hash buckets are shared by all the futex_keys that hash to the same
259 * location. Each key may have multiple futex_q structures, one for each task
260 * waiting on a futex.
262 struct futex_hash_bucket
{
265 struct plist_head chain
;
266 } ____cacheline_aligned_in_smp
;
269 * The base of the bucket array and its size are always used together
270 * (after initialization only in hash_futex()), so ensure that they
271 * reside in the same cacheline.
274 struct futex_hash_bucket
*queues
;
275 unsigned long hashsize
;
276 } __futex_data __read_mostly
__aligned(2*sizeof(long));
277 #define futex_queues (__futex_data.queues)
278 #define futex_hashsize (__futex_data.hashsize)
282 * Fault injections for futexes.
284 #ifdef CONFIG_FAIL_FUTEX
287 struct fault_attr attr
;
291 .attr
= FAULT_ATTR_INITIALIZER
,
292 .ignore_private
= false,
295 static int __init
setup_fail_futex(char *str
)
297 return setup_fault_attr(&fail_futex
.attr
, str
);
299 __setup("fail_futex=", setup_fail_futex
);
301 static bool should_fail_futex(bool fshared
)
303 if (fail_futex
.ignore_private
&& !fshared
)
306 return should_fail(&fail_futex
.attr
, 1);
309 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
311 static int __init
fail_futex_debugfs(void)
313 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
316 dir
= fault_create_debugfs_attr("fail_futex", NULL
,
321 if (!debugfs_create_bool("ignore-private", mode
, dir
,
322 &fail_futex
.ignore_private
)) {
323 debugfs_remove_recursive(dir
);
330 late_initcall(fail_futex_debugfs
);
332 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
335 static inline bool should_fail_futex(bool fshared
)
339 #endif /* CONFIG_FAIL_FUTEX */
341 static inline void futex_get_mm(union futex_key
*key
)
343 mmgrab(key
->private.mm
);
345 * Ensure futex_get_mm() implies a full barrier such that
346 * get_futex_key() implies a full barrier. This is relied upon
347 * as smp_mb(); (B), see the ordering comment above.
349 smp_mb__after_atomic();
353 * Reflects a new waiter being added to the waitqueue.
355 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
358 atomic_inc(&hb
->waiters
);
360 * Full barrier (A), see the ordering comment above.
362 smp_mb__after_atomic();
367 * Reflects a waiter being removed from the waitqueue by wakeup
370 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
373 atomic_dec(&hb
->waiters
);
377 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
380 return atomic_read(&hb
->waiters
);
387 * hash_futex - Return the hash bucket in the global hash
388 * @key: Pointer to the futex key for which the hash is calculated
390 * We hash on the keys returned from get_futex_key (see below) and return the
391 * corresponding hash bucket in the global hash.
393 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
395 u32 hash
= jhash2((u32
*)&key
->both
.word
,
396 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
398 return &futex_queues
[hash
& (futex_hashsize
- 1)];
403 * match_futex - Check whether two futex keys are equal
404 * @key1: Pointer to key1
405 * @key2: Pointer to key2
407 * Return 1 if two futex_keys are equal, 0 otherwise.
409 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
412 && key1
->both
.word
== key2
->both
.word
413 && key1
->both
.ptr
== key2
->both
.ptr
414 && key1
->both
.offset
== key2
->both
.offset
);
418 * Take a reference to the resource addressed by a key.
419 * Can be called while holding spinlocks.
422 static void get_futex_key_refs(union futex_key
*key
)
428 * On MMU less systems futexes are always "private" as there is no per
429 * process address space. We need the smp wmb nevertheless - yes,
430 * arch/blackfin has MMU less SMP ...
432 if (!IS_ENABLED(CONFIG_MMU
)) {
433 smp_mb(); /* explicit smp_mb(); (B) */
437 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
439 ihold(key
->shared
.inode
); /* implies smp_mb(); (B) */
441 case FUT_OFF_MMSHARED
:
442 futex_get_mm(key
); /* implies smp_mb(); (B) */
446 * Private futexes do not hold reference on an inode or
447 * mm, therefore the only purpose of calling get_futex_key_refs
448 * is because we need the barrier for the lockless waiter check.
450 smp_mb(); /* explicit smp_mb(); (B) */
455 * Drop a reference to the resource addressed by a key.
456 * The hash bucket spinlock must not be held. This is
457 * a no-op for private futexes, see comment in the get
460 static void drop_futex_key_refs(union futex_key
*key
)
462 if (!key
->both
.ptr
) {
463 /* If we're here then we tried to put a key we failed to get */
468 if (!IS_ENABLED(CONFIG_MMU
))
471 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
473 iput(key
->shared
.inode
);
475 case FUT_OFF_MMSHARED
:
476 mmdrop(key
->private.mm
);
482 * get_futex_key() - Get parameters which are the keys for a futex
483 * @uaddr: virtual address of the futex
484 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485 * @key: address where result is stored.
486 * @rw: mapping needs to be read/write (values: VERIFY_READ,
489 * Return: a negative error code or 0
491 * The key words are stored in @key on success.
493 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
494 * offset_within_page). For private mappings, it's (uaddr, current->mm).
495 * We can usually work out the index without swapping in the page.
497 * lock_page() might sleep, the caller should not hold a spinlock.
500 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
502 unsigned long address
= (unsigned long)uaddr
;
503 struct mm_struct
*mm
= current
->mm
;
504 struct page
*page
, *tail
;
505 struct address_space
*mapping
;
509 * The futex address must be "naturally" aligned.
511 key
->both
.offset
= address
% PAGE_SIZE
;
512 if (unlikely((address
% sizeof(u32
)) != 0))
514 address
-= key
->both
.offset
;
516 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
519 if (unlikely(should_fail_futex(fshared
)))
523 * PROCESS_PRIVATE futexes are fast.
524 * As the mm cannot disappear under us and the 'key' only needs
525 * virtual address, we dont even have to find the underlying vma.
526 * Note : We do have to check 'uaddr' is a valid user address,
527 * but access_ok() should be faster than find_vma()
530 key
->private.mm
= mm
;
531 key
->private.address
= address
;
532 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared
)))
541 err
= get_user_pages_fast(address
, 1, 1, &page
);
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
546 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
547 err
= get_user_pages_fast(address
, 1, 0, &page
);
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
574 page
= compound_head(page
);
575 mapping
= READ_ONCE(page
->mapping
);
578 * If page->mapping is NULL, then it cannot be a PageAnon
579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
590 * an unlikely race, but we do need to retry for page->mapping.
592 if (unlikely(!mapping
)) {
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
601 shmem_swizzled
= PageSwapCache(page
) || page
->mapping
;
612 * Private mappings are handled in a simple way.
614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
619 * the object not the particular process.
621 if (PageAnon(page
)) {
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
626 if (unlikely(should_fail_futex(fshared
)) || ro
) {
631 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
632 key
->private.mm
= mm
;
633 key
->private.address
= address
;
635 get_futex_key_refs(key
); /* implies smp_mb(); (B) */
641 * The associated futex object in this case is the inode and
642 * the page->mapping must be traversed. Ordinarily this should
643 * be stabilised under page lock but it's not strictly
644 * necessary in this case as we just want to pin the inode, not
645 * update the radix tree or anything like that.
647 * The RCU read lock is taken as the inode is finally freed
648 * under RCU. If the mapping still matches expectations then the
649 * mapping->host can be safely accessed as being a valid inode.
653 if (READ_ONCE(page
->mapping
) != mapping
) {
660 inode
= READ_ONCE(mapping
->host
);
669 * Take a reference unless it is about to be freed. Previously
670 * this reference was taken by ihold under the page lock
671 * pinning the inode in place so i_lock was unnecessary. The
672 * only way for this check to fail is if the inode was
673 * truncated in parallel which is almost certainly an
674 * application bug. In such a case, just retry.
676 * We are not calling into get_futex_key_refs() in file-backed
677 * cases, therefore a successful atomic_inc return below will
678 * guarantee that get_futex_key() will still imply smp_mb(); (B).
680 if (!atomic_inc_not_zero(&inode
->i_count
)) {
687 /* Should be impossible but lets be paranoid for now */
688 if (WARN_ON_ONCE(inode
->i_mapping
!= mapping
)) {
696 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
697 key
->shared
.inode
= inode
;
698 key
->shared
.pgoff
= basepage_index(tail
);
707 static inline void put_futex_key(union futex_key
*key
)
709 drop_futex_key_refs(key
);
713 * fault_in_user_writeable() - Fault in user address and verify RW access
714 * @uaddr: pointer to faulting user space address
716 * Slow path to fixup the fault we just took in the atomic write
719 * We have no generic implementation of a non-destructive write to the
720 * user address. We know that we faulted in the atomic pagefault
721 * disabled section so we can as well avoid the #PF overhead by
722 * calling get_user_pages() right away.
724 static int fault_in_user_writeable(u32 __user
*uaddr
)
726 struct mm_struct
*mm
= current
->mm
;
729 down_read(&mm
->mmap_sem
);
730 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
731 FAULT_FLAG_WRITE
, NULL
);
732 up_read(&mm
->mmap_sem
);
734 return ret
< 0 ? ret
: 0;
738 * futex_top_waiter() - Return the highest priority waiter on a futex
739 * @hb: the hash bucket the futex_q's reside in
740 * @key: the futex key (to distinguish it from other futex futex_q's)
742 * Must be called with the hb lock held.
744 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
745 union futex_key
*key
)
747 struct futex_q
*this;
749 plist_for_each_entry(this, &hb
->chain
, list
) {
750 if (match_futex(&this->key
, key
))
756 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
757 u32 uval
, u32 newval
)
762 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
768 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
773 ret
= __get_user(*dest
, from
);
776 return ret
? -EFAULT
: 0;
783 static int refill_pi_state_cache(void)
785 struct futex_pi_state
*pi_state
;
787 if (likely(current
->pi_state_cache
))
790 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
795 INIT_LIST_HEAD(&pi_state
->list
);
796 /* pi_mutex gets initialized later */
797 pi_state
->owner
= NULL
;
798 atomic_set(&pi_state
->refcount
, 1);
799 pi_state
->key
= FUTEX_KEY_INIT
;
801 current
->pi_state_cache
= pi_state
;
806 static struct futex_pi_state
*alloc_pi_state(void)
808 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
811 current
->pi_state_cache
= NULL
;
816 static void get_pi_state(struct futex_pi_state
*pi_state
)
818 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state
->refcount
));
822 * Drops a reference to the pi_state object and frees or caches it
823 * when the last reference is gone.
825 static void put_pi_state(struct futex_pi_state
*pi_state
)
830 if (!atomic_dec_and_test(&pi_state
->refcount
))
834 * If pi_state->owner is NULL, the owner is most probably dying
835 * and has cleaned up the pi_state already
837 if (pi_state
->owner
) {
838 struct task_struct
*owner
;
840 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
841 owner
= pi_state
->owner
;
843 raw_spin_lock(&owner
->pi_lock
);
844 list_del_init(&pi_state
->list
);
845 raw_spin_unlock(&owner
->pi_lock
);
847 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, owner
);
848 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
851 if (current
->pi_state_cache
) {
855 * pi_state->list is already empty.
856 * clear pi_state->owner.
857 * refcount is at 0 - put it back to 1.
859 pi_state
->owner
= NULL
;
860 atomic_set(&pi_state
->refcount
, 1);
861 current
->pi_state_cache
= pi_state
;
866 * Look up the task based on what TID userspace gave us.
869 static struct task_struct
*futex_find_get_task(pid_t pid
)
871 struct task_struct
*p
;
874 p
= find_task_by_vpid(pid
);
883 #ifdef CONFIG_FUTEX_PI
886 * This task is holding PI mutexes at exit time => bad.
887 * Kernel cleans up PI-state, but userspace is likely hosed.
888 * (Robust-futex cleanup is separate and might save the day for userspace.)
890 void exit_pi_state_list(struct task_struct
*curr
)
892 struct list_head
*next
, *head
= &curr
->pi_state_list
;
893 struct futex_pi_state
*pi_state
;
894 struct futex_hash_bucket
*hb
;
895 union futex_key key
= FUTEX_KEY_INIT
;
897 if (!futex_cmpxchg_enabled
)
900 * We are a ZOMBIE and nobody can enqueue itself on
901 * pi_state_list anymore, but we have to be careful
902 * versus waiters unqueueing themselves:
904 raw_spin_lock_irq(&curr
->pi_lock
);
905 while (!list_empty(head
)) {
908 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
910 hb
= hash_futex(&key
);
911 raw_spin_unlock_irq(&curr
->pi_lock
);
913 spin_lock(&hb
->lock
);
914 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
915 raw_spin_lock(&curr
->pi_lock
);
917 * We dropped the pi-lock, so re-check whether this
918 * task still owns the PI-state:
920 if (head
->next
!= next
) {
921 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
922 spin_unlock(&hb
->lock
);
926 WARN_ON(pi_state
->owner
!= curr
);
927 WARN_ON(list_empty(&pi_state
->list
));
928 list_del_init(&pi_state
->list
);
929 pi_state
->owner
= NULL
;
930 raw_spin_unlock(&curr
->pi_lock
);
932 get_pi_state(pi_state
);
933 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
934 spin_unlock(&hb
->lock
);
936 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
937 put_pi_state(pi_state
);
939 raw_spin_lock_irq(&curr
->pi_lock
);
941 raw_spin_unlock_irq(&curr
->pi_lock
);
947 * We need to check the following states:
949 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
951 * [1] NULL | --- | --- | 0 | 0/1 | Valid
952 * [2] NULL | --- | --- | >0 | 0/1 | Valid
954 * [3] Found | NULL | -- | Any | 0/1 | Invalid
956 * [4] Found | Found | NULL | 0 | 1 | Valid
957 * [5] Found | Found | NULL | >0 | 1 | Invalid
959 * [6] Found | Found | task | 0 | 1 | Valid
961 * [7] Found | Found | NULL | Any | 0 | Invalid
963 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
964 * [9] Found | Found | task | 0 | 0 | Invalid
965 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
967 * [1] Indicates that the kernel can acquire the futex atomically. We
968 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
970 * [2] Valid, if TID does not belong to a kernel thread. If no matching
971 * thread is found then it indicates that the owner TID has died.
973 * [3] Invalid. The waiter is queued on a non PI futex
975 * [4] Valid state after exit_robust_list(), which sets the user space
976 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
978 * [5] The user space value got manipulated between exit_robust_list()
979 * and exit_pi_state_list()
981 * [6] Valid state after exit_pi_state_list() which sets the new owner in
982 * the pi_state but cannot access the user space value.
984 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
986 * [8] Owner and user space value match
988 * [9] There is no transient state which sets the user space TID to 0
989 * except exit_robust_list(), but this is indicated by the
990 * FUTEX_OWNER_DIED bit. See [4]
992 * [10] There is no transient state which leaves owner and user space
996 * Serialization and lifetime rules:
1000 * hb -> futex_q, relation
1001 * futex_q -> pi_state, relation
1003 * (cannot be raw because hb can contain arbitrary amount
1006 * pi_mutex->wait_lock:
1010 * (and pi_mutex 'obviously')
1014 * p->pi_state_list -> pi_state->list, relation
1016 * pi_state->refcount:
1024 * pi_mutex->wait_lock
1030 * Validate that the existing waiter has a pi_state and sanity check
1031 * the pi_state against the user space value. If correct, attach to
1034 static int attach_to_pi_state(u32 __user
*uaddr
, u32 uval
,
1035 struct futex_pi_state
*pi_state
,
1036 struct futex_pi_state
**ps
)
1038 pid_t pid
= uval
& FUTEX_TID_MASK
;
1043 * Userspace might have messed up non-PI and PI futexes [3]
1045 if (unlikely(!pi_state
))
1049 * We get here with hb->lock held, and having found a
1050 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1051 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1052 * which in turn means that futex_lock_pi() still has a reference on
1055 * The waiter holding a reference on @pi_state also protects against
1056 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1057 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1058 * free pi_state before we can take a reference ourselves.
1060 WARN_ON(!atomic_read(&pi_state
->refcount
));
1063 * Now that we have a pi_state, we can acquire wait_lock
1064 * and do the state validation.
1066 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
1069 * Since {uval, pi_state} is serialized by wait_lock, and our current
1070 * uval was read without holding it, it can have changed. Verify it
1071 * still is what we expect it to be, otherwise retry the entire
1074 if (get_futex_value_locked(&uval2
, uaddr
))
1081 * Handle the owner died case:
1083 if (uval
& FUTEX_OWNER_DIED
) {
1085 * exit_pi_state_list sets owner to NULL and wakes the
1086 * topmost waiter. The task which acquires the
1087 * pi_state->rt_mutex will fixup owner.
1089 if (!pi_state
->owner
) {
1091 * No pi state owner, but the user space TID
1092 * is not 0. Inconsistent state. [5]
1097 * Take a ref on the state and return success. [4]
1103 * If TID is 0, then either the dying owner has not
1104 * yet executed exit_pi_state_list() or some waiter
1105 * acquired the rtmutex in the pi state, but did not
1106 * yet fixup the TID in user space.
1108 * Take a ref on the state and return success. [6]
1114 * If the owner died bit is not set, then the pi_state
1115 * must have an owner. [7]
1117 if (!pi_state
->owner
)
1122 * Bail out if user space manipulated the futex value. If pi
1123 * state exists then the owner TID must be the same as the
1124 * user space TID. [9/10]
1126 if (pid
!= task_pid_vnr(pi_state
->owner
))
1130 get_pi_state(pi_state
);
1131 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1148 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1153 * Lookup the task for the TID provided from user space and attach to
1154 * it after doing proper sanity checks.
1156 static int attach_to_pi_owner(u32 uval
, union futex_key
*key
,
1157 struct futex_pi_state
**ps
)
1159 pid_t pid
= uval
& FUTEX_TID_MASK
;
1160 struct futex_pi_state
*pi_state
;
1161 struct task_struct
*p
;
1164 * We are the first waiter - try to look up the real owner and attach
1165 * the new pi_state to it, but bail out when TID = 0 [1]
1169 p
= futex_find_get_task(pid
);
1173 if (unlikely(p
->flags
& PF_KTHREAD
)) {
1179 * We need to look at the task state flags to figure out,
1180 * whether the task is exiting. To protect against the do_exit
1181 * change of the task flags, we do this protected by
1184 raw_spin_lock_irq(&p
->pi_lock
);
1185 if (unlikely(p
->flags
& PF_EXITING
)) {
1187 * The task is on the way out. When PF_EXITPIDONE is
1188 * set, we know that the task has finished the
1191 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
1193 raw_spin_unlock_irq(&p
->pi_lock
);
1199 * No existing pi state. First waiter. [2]
1201 * This creates pi_state, we have hb->lock held, this means nothing can
1202 * observe this state, wait_lock is irrelevant.
1204 pi_state
= alloc_pi_state();
1207 * Initialize the pi_mutex in locked state and make @p
1210 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
1212 /* Store the key for possible exit cleanups: */
1213 pi_state
->key
= *key
;
1215 WARN_ON(!list_empty(&pi_state
->list
));
1216 list_add(&pi_state
->list
, &p
->pi_state_list
);
1218 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1219 * because there is no concurrency as the object is not published yet.
1221 pi_state
->owner
= p
;
1222 raw_spin_unlock_irq(&p
->pi_lock
);
1231 static int lookup_pi_state(u32 __user
*uaddr
, u32 uval
,
1232 struct futex_hash_bucket
*hb
,
1233 union futex_key
*key
, struct futex_pi_state
**ps
)
1235 struct futex_q
*top_waiter
= futex_top_waiter(hb
, key
);
1238 * If there is a waiter on that futex, validate it and
1239 * attach to the pi_state when the validation succeeds.
1242 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1245 * We are the first waiter - try to look up the owner based on
1246 * @uval and attach to it.
1248 return attach_to_pi_owner(uval
, key
, ps
);
1251 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
1253 u32
uninitialized_var(curval
);
1255 if (unlikely(should_fail_futex(true)))
1258 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
1261 /* If user space value changed, let the caller retry */
1262 return curval
!= uval
? -EAGAIN
: 0;
1266 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1267 * @uaddr: the pi futex user address
1268 * @hb: the pi futex hash bucket
1269 * @key: the futex key associated with uaddr and hb
1270 * @ps: the pi_state pointer where we store the result of the
1272 * @task: the task to perform the atomic lock work for. This will
1273 * be "current" except in the case of requeue pi.
1274 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1277 * - 0 - ready to wait;
1278 * - 1 - acquired the lock;
1281 * The hb->lock and futex_key refs shall be held by the caller.
1283 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
1284 union futex_key
*key
,
1285 struct futex_pi_state
**ps
,
1286 struct task_struct
*task
, int set_waiters
)
1288 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
1289 struct futex_q
*top_waiter
;
1293 * Read the user space value first so we can validate a few
1294 * things before proceeding further.
1296 if (get_futex_value_locked(&uval
, uaddr
))
1299 if (unlikely(should_fail_futex(true)))
1305 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1308 if ((unlikely(should_fail_futex(true))))
1312 * Lookup existing state first. If it exists, try to attach to
1315 top_waiter
= futex_top_waiter(hb
, key
);
1317 return attach_to_pi_state(uaddr
, uval
, top_waiter
->pi_state
, ps
);
1320 * No waiter and user TID is 0. We are here because the
1321 * waiters or the owner died bit is set or called from
1322 * requeue_cmp_pi or for whatever reason something took the
1325 if (!(uval
& FUTEX_TID_MASK
)) {
1327 * We take over the futex. No other waiters and the user space
1328 * TID is 0. We preserve the owner died bit.
1330 newval
= uval
& FUTEX_OWNER_DIED
;
1333 /* The futex requeue_pi code can enforce the waiters bit */
1335 newval
|= FUTEX_WAITERS
;
1337 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1338 /* If the take over worked, return 1 */
1339 return ret
< 0 ? ret
: 1;
1343 * First waiter. Set the waiters bit before attaching ourself to
1344 * the owner. If owner tries to unlock, it will be forced into
1345 * the kernel and blocked on hb->lock.
1347 newval
= uval
| FUTEX_WAITERS
;
1348 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1352 * If the update of the user space value succeeded, we try to
1353 * attach to the owner. If that fails, no harm done, we only
1354 * set the FUTEX_WAITERS bit in the user space variable.
1356 return attach_to_pi_owner(uval
, key
, ps
);
1360 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1361 * @q: The futex_q to unqueue
1363 * The q->lock_ptr must not be NULL and must be held by the caller.
1365 static void __unqueue_futex(struct futex_q
*q
)
1367 struct futex_hash_bucket
*hb
;
1369 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1370 || WARN_ON(plist_node_empty(&q
->list
)))
1373 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1374 plist_del(&q
->list
, &hb
->chain
);
1379 * The hash bucket lock must be held when this is called.
1380 * Afterwards, the futex_q must not be accessed. Callers
1381 * must ensure to later call wake_up_q() for the actual
1384 static void mark_wake_futex(struct wake_q_head
*wake_q
, struct futex_q
*q
)
1386 struct task_struct
*p
= q
->task
;
1388 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1392 * Queue the task for later wakeup for after we've released
1393 * the hb->lock. wake_q_add() grabs reference to p.
1395 wake_q_add(wake_q
, p
);
1398 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1399 * is written, without taking any locks. This is possible in the event
1400 * of a spurious wakeup, for example. A memory barrier is required here
1401 * to prevent the following store to lock_ptr from getting ahead of the
1402 * plist_del in __unqueue_futex().
1404 smp_store_release(&q
->lock_ptr
, NULL
);
1408 * Caller must hold a reference on @pi_state.
1410 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_pi_state
*pi_state
)
1412 u32
uninitialized_var(curval
), newval
;
1413 struct task_struct
*new_owner
;
1414 bool postunlock
= false;
1415 DEFINE_WAKE_Q(wake_q
);
1418 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1419 if (WARN_ON_ONCE(!new_owner
)) {
1421 * As per the comment in futex_unlock_pi() this should not happen.
1423 * When this happens, give up our locks and try again, giving
1424 * the futex_lock_pi() instance time to complete, either by
1425 * waiting on the rtmutex or removing itself from the futex
1433 * We pass it to the next owner. The WAITERS bit is always kept
1434 * enabled while there is PI state around. We cleanup the owner
1435 * died bit, because we are the owner.
1437 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1439 if (unlikely(should_fail_futex(true)))
1442 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)) {
1445 } else if (curval
!= uval
) {
1447 * If a unconditional UNLOCK_PI operation (user space did not
1448 * try the TID->0 transition) raced with a waiter setting the
1449 * FUTEX_WAITERS flag between get_user() and locking the hash
1450 * bucket lock, retry the operation.
1452 if ((FUTEX_TID_MASK
& curval
) == uval
)
1462 * This is a point of no return; once we modify the uval there is no
1463 * going back and subsequent operations must not fail.
1466 raw_spin_lock(&pi_state
->owner
->pi_lock
);
1467 WARN_ON(list_empty(&pi_state
->list
));
1468 list_del_init(&pi_state
->list
);
1469 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
1471 raw_spin_lock(&new_owner
->pi_lock
);
1472 WARN_ON(!list_empty(&pi_state
->list
));
1473 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1474 pi_state
->owner
= new_owner
;
1475 raw_spin_unlock(&new_owner
->pi_lock
);
1477 postunlock
= __rt_mutex_futex_unlock(&pi_state
->pi_mutex
, &wake_q
);
1480 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
1483 rt_mutex_postunlock(&wake_q
);
1489 * Express the locking dependencies for lockdep:
1492 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1495 spin_lock(&hb1
->lock
);
1497 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1498 } else { /* hb1 > hb2 */
1499 spin_lock(&hb2
->lock
);
1500 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1505 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1507 spin_unlock(&hb1
->lock
);
1509 spin_unlock(&hb2
->lock
);
1513 * Wake up waiters matching bitset queued on this futex (uaddr).
1516 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1518 struct futex_hash_bucket
*hb
;
1519 struct futex_q
*this, *next
;
1520 union futex_key key
= FUTEX_KEY_INIT
;
1522 DEFINE_WAKE_Q(wake_q
);
1527 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1528 if (unlikely(ret
!= 0))
1531 hb
= hash_futex(&key
);
1533 /* Make sure we really have tasks to wakeup */
1534 if (!hb_waiters_pending(hb
))
1537 spin_lock(&hb
->lock
);
1539 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1540 if (match_futex (&this->key
, &key
)) {
1541 if (this->pi_state
|| this->rt_waiter
) {
1546 /* Check if one of the bits is set in both bitsets */
1547 if (!(this->bitset
& bitset
))
1550 mark_wake_futex(&wake_q
, this);
1551 if (++ret
>= nr_wake
)
1556 spin_unlock(&hb
->lock
);
1559 put_futex_key(&key
);
1564 static int futex_atomic_op_inuser(unsigned int encoded_op
, u32 __user
*uaddr
)
1566 unsigned int op
= (encoded_op
& 0x70000000) >> 28;
1567 unsigned int cmp
= (encoded_op
& 0x0f000000) >> 24;
1568 int oparg
= sign_extend32((encoded_op
& 0x00fff000) >> 12, 12);
1569 int cmparg
= sign_extend32(encoded_op
& 0x00000fff, 12);
1572 if (encoded_op
& (FUTEX_OP_OPARG_SHIFT
<< 28)) {
1573 if (oparg
< 0 || oparg
> 31) {
1574 char comm
[sizeof(current
->comm
)];
1576 * kill this print and return -EINVAL when userspace
1579 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1580 get_task_comm(comm
, current
), oparg
);
1586 if (!access_ok(VERIFY_WRITE
, uaddr
, sizeof(u32
)))
1589 ret
= arch_futex_atomic_op_inuser(op
, oparg
, &oldval
, uaddr
);
1594 case FUTEX_OP_CMP_EQ
:
1595 return oldval
== cmparg
;
1596 case FUTEX_OP_CMP_NE
:
1597 return oldval
!= cmparg
;
1598 case FUTEX_OP_CMP_LT
:
1599 return oldval
< cmparg
;
1600 case FUTEX_OP_CMP_GE
:
1601 return oldval
>= cmparg
;
1602 case FUTEX_OP_CMP_LE
:
1603 return oldval
<= cmparg
;
1604 case FUTEX_OP_CMP_GT
:
1605 return oldval
> cmparg
;
1612 * Wake up all waiters hashed on the physical page that is mapped
1613 * to this virtual address:
1616 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1617 int nr_wake
, int nr_wake2
, int op
)
1619 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1620 struct futex_hash_bucket
*hb1
, *hb2
;
1621 struct futex_q
*this, *next
;
1623 DEFINE_WAKE_Q(wake_q
);
1626 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1627 if (unlikely(ret
!= 0))
1629 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1630 if (unlikely(ret
!= 0))
1633 hb1
= hash_futex(&key1
);
1634 hb2
= hash_futex(&key2
);
1637 double_lock_hb(hb1
, hb2
);
1638 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1639 if (unlikely(op_ret
< 0)) {
1641 double_unlock_hb(hb1
, hb2
);
1645 * we don't get EFAULT from MMU faults if we don't have an MMU,
1646 * but we might get them from range checking
1652 if (unlikely(op_ret
!= -EFAULT
)) {
1657 ret
= fault_in_user_writeable(uaddr2
);
1661 if (!(flags
& FLAGS_SHARED
))
1664 put_futex_key(&key2
);
1665 put_futex_key(&key1
);
1669 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1670 if (match_futex (&this->key
, &key1
)) {
1671 if (this->pi_state
|| this->rt_waiter
) {
1675 mark_wake_futex(&wake_q
, this);
1676 if (++ret
>= nr_wake
)
1683 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1684 if (match_futex (&this->key
, &key2
)) {
1685 if (this->pi_state
|| this->rt_waiter
) {
1689 mark_wake_futex(&wake_q
, this);
1690 if (++op_ret
>= nr_wake2
)
1698 double_unlock_hb(hb1
, hb2
);
1701 put_futex_key(&key2
);
1703 put_futex_key(&key1
);
1709 * requeue_futex() - Requeue a futex_q from one hb to another
1710 * @q: the futex_q to requeue
1711 * @hb1: the source hash_bucket
1712 * @hb2: the target hash_bucket
1713 * @key2: the new key for the requeued futex_q
1716 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1717 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1721 * If key1 and key2 hash to the same bucket, no need to
1724 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1725 plist_del(&q
->list
, &hb1
->chain
);
1726 hb_waiters_dec(hb1
);
1727 hb_waiters_inc(hb2
);
1728 plist_add(&q
->list
, &hb2
->chain
);
1729 q
->lock_ptr
= &hb2
->lock
;
1731 get_futex_key_refs(key2
);
1736 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1738 * @key: the key of the requeue target futex
1739 * @hb: the hash_bucket of the requeue target futex
1741 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1742 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1743 * to the requeue target futex so the waiter can detect the wakeup on the right
1744 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1745 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1746 * to protect access to the pi_state to fixup the owner later. Must be called
1747 * with both q->lock_ptr and hb->lock held.
1750 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1751 struct futex_hash_bucket
*hb
)
1753 get_futex_key_refs(key
);
1758 WARN_ON(!q
->rt_waiter
);
1759 q
->rt_waiter
= NULL
;
1761 q
->lock_ptr
= &hb
->lock
;
1763 wake_up_state(q
->task
, TASK_NORMAL
);
1767 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1768 * @pifutex: the user address of the to futex
1769 * @hb1: the from futex hash bucket, must be locked by the caller
1770 * @hb2: the to futex hash bucket, must be locked by the caller
1771 * @key1: the from futex key
1772 * @key2: the to futex key
1773 * @ps: address to store the pi_state pointer
1774 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1776 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1777 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1778 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1779 * hb1 and hb2 must be held by the caller.
1782 * - 0 - failed to acquire the lock atomically;
1783 * - >0 - acquired the lock, return value is vpid of the top_waiter
1786 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1787 struct futex_hash_bucket
*hb1
,
1788 struct futex_hash_bucket
*hb2
,
1789 union futex_key
*key1
, union futex_key
*key2
,
1790 struct futex_pi_state
**ps
, int set_waiters
)
1792 struct futex_q
*top_waiter
= NULL
;
1796 if (get_futex_value_locked(&curval
, pifutex
))
1799 if (unlikely(should_fail_futex(true)))
1803 * Find the top_waiter and determine if there are additional waiters.
1804 * If the caller intends to requeue more than 1 waiter to pifutex,
1805 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1806 * as we have means to handle the possible fault. If not, don't set
1807 * the bit unecessarily as it will force the subsequent unlock to enter
1810 top_waiter
= futex_top_waiter(hb1
, key1
);
1812 /* There are no waiters, nothing for us to do. */
1816 /* Ensure we requeue to the expected futex. */
1817 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1821 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1822 * the contended case or if set_waiters is 1. The pi_state is returned
1823 * in ps in contended cases.
1825 vpid
= task_pid_vnr(top_waiter
->task
);
1826 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1829 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1836 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1837 * @uaddr1: source futex user address
1838 * @flags: futex flags (FLAGS_SHARED, etc.)
1839 * @uaddr2: target futex user address
1840 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1841 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1842 * @cmpval: @uaddr1 expected value (or %NULL)
1843 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1844 * pi futex (pi to pi requeue is not supported)
1846 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1847 * uaddr2 atomically on behalf of the top waiter.
1850 * - >=0 - on success, the number of tasks requeued or woken;
1853 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1854 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1855 u32
*cmpval
, int requeue_pi
)
1857 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1858 int drop_count
= 0, task_count
= 0, ret
;
1859 struct futex_pi_state
*pi_state
= NULL
;
1860 struct futex_hash_bucket
*hb1
, *hb2
;
1861 struct futex_q
*this, *next
;
1862 DEFINE_WAKE_Q(wake_q
);
1865 * When PI not supported: return -ENOSYS if requeue_pi is true,
1866 * consequently the compiler knows requeue_pi is always false past
1867 * this point which will optimize away all the conditional code
1870 if (!IS_ENABLED(CONFIG_FUTEX_PI
) && requeue_pi
)
1875 * Requeue PI only works on two distinct uaddrs. This
1876 * check is only valid for private futexes. See below.
1878 if (uaddr1
== uaddr2
)
1882 * requeue_pi requires a pi_state, try to allocate it now
1883 * without any locks in case it fails.
1885 if (refill_pi_state_cache())
1888 * requeue_pi must wake as many tasks as it can, up to nr_wake
1889 * + nr_requeue, since it acquires the rt_mutex prior to
1890 * returning to userspace, so as to not leave the rt_mutex with
1891 * waiters and no owner. However, second and third wake-ups
1892 * cannot be predicted as they involve race conditions with the
1893 * first wake and a fault while looking up the pi_state. Both
1894 * pthread_cond_signal() and pthread_cond_broadcast() should
1902 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1903 if (unlikely(ret
!= 0))
1905 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1906 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1907 if (unlikely(ret
!= 0))
1911 * The check above which compares uaddrs is not sufficient for
1912 * shared futexes. We need to compare the keys:
1914 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1919 hb1
= hash_futex(&key1
);
1920 hb2
= hash_futex(&key2
);
1923 hb_waiters_inc(hb2
);
1924 double_lock_hb(hb1
, hb2
);
1926 if (likely(cmpval
!= NULL
)) {
1929 ret
= get_futex_value_locked(&curval
, uaddr1
);
1931 if (unlikely(ret
)) {
1932 double_unlock_hb(hb1
, hb2
);
1933 hb_waiters_dec(hb2
);
1935 ret
= get_user(curval
, uaddr1
);
1939 if (!(flags
& FLAGS_SHARED
))
1942 put_futex_key(&key2
);
1943 put_futex_key(&key1
);
1946 if (curval
!= *cmpval
) {
1952 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1954 * Attempt to acquire uaddr2 and wake the top waiter. If we
1955 * intend to requeue waiters, force setting the FUTEX_WAITERS
1956 * bit. We force this here where we are able to easily handle
1957 * faults rather in the requeue loop below.
1959 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1960 &key2
, &pi_state
, nr_requeue
);
1963 * At this point the top_waiter has either taken uaddr2 or is
1964 * waiting on it. If the former, then the pi_state will not
1965 * exist yet, look it up one more time to ensure we have a
1966 * reference to it. If the lock was taken, ret contains the
1967 * vpid of the top waiter task.
1968 * If the lock was not taken, we have pi_state and an initial
1969 * refcount on it. In case of an error we have nothing.
1976 * If we acquired the lock, then the user space value
1977 * of uaddr2 should be vpid. It cannot be changed by
1978 * the top waiter as it is blocked on hb2 lock if it
1979 * tries to do so. If something fiddled with it behind
1980 * our back the pi state lookup might unearth it. So
1981 * we rather use the known value than rereading and
1982 * handing potential crap to lookup_pi_state.
1984 * If that call succeeds then we have pi_state and an
1985 * initial refcount on it.
1987 ret
= lookup_pi_state(uaddr2
, ret
, hb2
, &key2
, &pi_state
);
1992 /* We hold a reference on the pi state. */
1995 /* If the above failed, then pi_state is NULL */
1997 double_unlock_hb(hb1
, hb2
);
1998 hb_waiters_dec(hb2
);
1999 put_futex_key(&key2
);
2000 put_futex_key(&key1
);
2001 ret
= fault_in_user_writeable(uaddr2
);
2007 * Two reasons for this:
2008 * - Owner is exiting and we just wait for the
2010 * - The user space value changed.
2012 double_unlock_hb(hb1
, hb2
);
2013 hb_waiters_dec(hb2
);
2014 put_futex_key(&key2
);
2015 put_futex_key(&key1
);
2023 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
2024 if (task_count
- nr_wake
>= nr_requeue
)
2027 if (!match_futex(&this->key
, &key1
))
2031 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2032 * be paired with each other and no other futex ops.
2034 * We should never be requeueing a futex_q with a pi_state,
2035 * which is awaiting a futex_unlock_pi().
2037 if ((requeue_pi
&& !this->rt_waiter
) ||
2038 (!requeue_pi
&& this->rt_waiter
) ||
2045 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2046 * lock, we already woke the top_waiter. If not, it will be
2047 * woken by futex_unlock_pi().
2049 if (++task_count
<= nr_wake
&& !requeue_pi
) {
2050 mark_wake_futex(&wake_q
, this);
2054 /* Ensure we requeue to the expected futex for requeue_pi. */
2055 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
2061 * Requeue nr_requeue waiters and possibly one more in the case
2062 * of requeue_pi if we couldn't acquire the lock atomically.
2066 * Prepare the waiter to take the rt_mutex. Take a
2067 * refcount on the pi_state and store the pointer in
2068 * the futex_q object of the waiter.
2070 get_pi_state(pi_state
);
2071 this->pi_state
= pi_state
;
2072 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
2077 * We got the lock. We do neither drop the
2078 * refcount on pi_state nor clear
2079 * this->pi_state because the waiter needs the
2080 * pi_state for cleaning up the user space
2081 * value. It will drop the refcount after
2084 requeue_pi_wake_futex(this, &key2
, hb2
);
2089 * rt_mutex_start_proxy_lock() detected a
2090 * potential deadlock when we tried to queue
2091 * that waiter. Drop the pi_state reference
2092 * which we took above and remove the pointer
2093 * to the state from the waiters futex_q
2096 this->pi_state
= NULL
;
2097 put_pi_state(pi_state
);
2099 * We stop queueing more waiters and let user
2100 * space deal with the mess.
2105 requeue_futex(this, hb1
, hb2
, &key2
);
2110 * We took an extra initial reference to the pi_state either
2111 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2112 * need to drop it here again.
2114 put_pi_state(pi_state
);
2117 double_unlock_hb(hb1
, hb2
);
2119 hb_waiters_dec(hb2
);
2122 * drop_futex_key_refs() must be called outside the spinlocks. During
2123 * the requeue we moved futex_q's from the hash bucket at key1 to the
2124 * one at key2 and updated their key pointer. We no longer need to
2125 * hold the references to key1.
2127 while (--drop_count
>= 0)
2128 drop_futex_key_refs(&key1
);
2131 put_futex_key(&key2
);
2133 put_futex_key(&key1
);
2135 return ret
? ret
: task_count
;
2138 /* The key must be already stored in q->key. */
2139 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
2140 __acquires(&hb
->lock
)
2142 struct futex_hash_bucket
*hb
;
2144 hb
= hash_futex(&q
->key
);
2147 * Increment the counter before taking the lock so that
2148 * a potential waker won't miss a to-be-slept task that is
2149 * waiting for the spinlock. This is safe as all queue_lock()
2150 * users end up calling queue_me(). Similarly, for housekeeping,
2151 * decrement the counter at queue_unlock() when some error has
2152 * occurred and we don't end up adding the task to the list.
2156 q
->lock_ptr
= &hb
->lock
;
2158 spin_lock(&hb
->lock
); /* implies smp_mb(); (A) */
2163 queue_unlock(struct futex_hash_bucket
*hb
)
2164 __releases(&hb
->lock
)
2166 spin_unlock(&hb
->lock
);
2170 static inline void __queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2175 * The priority used to register this element is
2176 * - either the real thread-priority for the real-time threads
2177 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2178 * - or MAX_RT_PRIO for non-RT threads.
2179 * Thus, all RT-threads are woken first in priority order, and
2180 * the others are woken last, in FIFO order.
2182 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
2184 plist_node_init(&q
->list
, prio
);
2185 plist_add(&q
->list
, &hb
->chain
);
2190 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2191 * @q: The futex_q to enqueue
2192 * @hb: The destination hash bucket
2194 * The hb->lock must be held by the caller, and is released here. A call to
2195 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2196 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2197 * or nothing if the unqueue is done as part of the wake process and the unqueue
2198 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2201 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
2202 __releases(&hb
->lock
)
2205 spin_unlock(&hb
->lock
);
2209 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2210 * @q: The futex_q to unqueue
2212 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2213 * be paired with exactly one earlier call to queue_me().
2216 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2217 * - 0 - if the futex_q was already removed by the waking thread
2219 static int unqueue_me(struct futex_q
*q
)
2221 spinlock_t
*lock_ptr
;
2224 /* In the common case we don't take the spinlock, which is nice. */
2227 * q->lock_ptr can change between this read and the following spin_lock.
2228 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2229 * optimizing lock_ptr out of the logic below.
2231 lock_ptr
= READ_ONCE(q
->lock_ptr
);
2232 if (lock_ptr
!= NULL
) {
2233 spin_lock(lock_ptr
);
2235 * q->lock_ptr can change between reading it and
2236 * spin_lock(), causing us to take the wrong lock. This
2237 * corrects the race condition.
2239 * Reasoning goes like this: if we have the wrong lock,
2240 * q->lock_ptr must have changed (maybe several times)
2241 * between reading it and the spin_lock(). It can
2242 * change again after the spin_lock() but only if it was
2243 * already changed before the spin_lock(). It cannot,
2244 * however, change back to the original value. Therefore
2245 * we can detect whether we acquired the correct lock.
2247 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
2248 spin_unlock(lock_ptr
);
2253 BUG_ON(q
->pi_state
);
2255 spin_unlock(lock_ptr
);
2259 drop_futex_key_refs(&q
->key
);
2264 * PI futexes can not be requeued and must remove themself from the
2265 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2268 static void unqueue_me_pi(struct futex_q
*q
)
2269 __releases(q
->lock_ptr
)
2273 BUG_ON(!q
->pi_state
);
2274 put_pi_state(q
->pi_state
);
2277 spin_unlock(q
->lock_ptr
);
2281 * Fixup the pi_state owner with the new owner.
2283 * Must be called with hash bucket lock held and mm->sem held for non
2286 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
2287 struct task_struct
*newowner
)
2289 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
2290 struct futex_pi_state
*pi_state
= q
->pi_state
;
2291 u32 uval
, uninitialized_var(curval
), newval
;
2292 struct task_struct
*oldowner
;
2295 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2297 oldowner
= pi_state
->owner
;
2299 if (!pi_state
->owner
)
2300 newtid
|= FUTEX_OWNER_DIED
;
2303 * We are here either because we stole the rtmutex from the
2304 * previous highest priority waiter or we are the highest priority
2305 * waiter but have failed to get the rtmutex the first time.
2307 * We have to replace the newowner TID in the user space variable.
2308 * This must be atomic as we have to preserve the owner died bit here.
2310 * Note: We write the user space value _before_ changing the pi_state
2311 * because we can fault here. Imagine swapped out pages or a fork
2312 * that marked all the anonymous memory readonly for cow.
2314 * Modifying pi_state _before_ the user space value would leave the
2315 * pi_state in an inconsistent state when we fault here, because we
2316 * need to drop the locks to handle the fault. This might be observed
2317 * in the PID check in lookup_pi_state.
2320 if (get_futex_value_locked(&uval
, uaddr
))
2324 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
2326 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
2334 * We fixed up user space. Now we need to fix the pi_state
2337 if (pi_state
->owner
!= NULL
) {
2338 raw_spin_lock(&pi_state
->owner
->pi_lock
);
2339 WARN_ON(list_empty(&pi_state
->list
));
2340 list_del_init(&pi_state
->list
);
2341 raw_spin_unlock(&pi_state
->owner
->pi_lock
);
2344 pi_state
->owner
= newowner
;
2346 raw_spin_lock(&newowner
->pi_lock
);
2347 WARN_ON(!list_empty(&pi_state
->list
));
2348 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
2349 raw_spin_unlock(&newowner
->pi_lock
);
2350 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2355 * To handle the page fault we need to drop the locks here. That gives
2356 * the other task (either the highest priority waiter itself or the
2357 * task which stole the rtmutex) the chance to try the fixup of the
2358 * pi_state. So once we are back from handling the fault we need to
2359 * check the pi_state after reacquiring the locks and before trying to
2360 * do another fixup. When the fixup has been done already we simply
2363 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2364 * drop hb->lock since the caller owns the hb -> futex_q relation.
2365 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2368 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2369 spin_unlock(q
->lock_ptr
);
2371 ret
= fault_in_user_writeable(uaddr
);
2373 spin_lock(q
->lock_ptr
);
2374 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2377 * Check if someone else fixed it for us:
2379 if (pi_state
->owner
!= oldowner
) {
2390 raw_spin_unlock_irq(&pi_state
->pi_mutex
.wait_lock
);
2394 static long futex_wait_restart(struct restart_block
*restart
);
2397 * fixup_owner() - Post lock pi_state and corner case management
2398 * @uaddr: user address of the futex
2399 * @q: futex_q (contains pi_state and access to the rt_mutex)
2400 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2402 * After attempting to lock an rt_mutex, this function is called to cleanup
2403 * the pi_state owner as well as handle race conditions that may allow us to
2404 * acquire the lock. Must be called with the hb lock held.
2407 * - 1 - success, lock taken;
2408 * - 0 - success, lock not taken;
2409 * - <0 - on error (-EFAULT)
2411 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
2417 * Got the lock. We might not be the anticipated owner if we
2418 * did a lock-steal - fix up the PI-state in that case:
2420 * We can safely read pi_state->owner without holding wait_lock
2421 * because we now own the rt_mutex, only the owner will attempt
2424 if (q
->pi_state
->owner
!= current
)
2425 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
2430 * Paranoia check. If we did not take the lock, then we should not be
2431 * the owner of the rt_mutex.
2433 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
) {
2434 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2435 "pi-state %p\n", ret
,
2436 q
->pi_state
->pi_mutex
.owner
,
2437 q
->pi_state
->owner
);
2441 return ret
? ret
: locked
;
2445 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2446 * @hb: the futex hash bucket, must be locked by the caller
2447 * @q: the futex_q to queue up on
2448 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2450 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2451 struct hrtimer_sleeper
*timeout
)
2454 * The task state is guaranteed to be set before another task can
2455 * wake it. set_current_state() is implemented using smp_store_mb() and
2456 * queue_me() calls spin_unlock() upon completion, both serializing
2457 * access to the hash list and forcing another memory barrier.
2459 set_current_state(TASK_INTERRUPTIBLE
);
2464 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2467 * If we have been removed from the hash list, then another task
2468 * has tried to wake us, and we can skip the call to schedule().
2470 if (likely(!plist_node_empty(&q
->list
))) {
2472 * If the timer has already expired, current will already be
2473 * flagged for rescheduling. Only call schedule if there
2474 * is no timeout, or if it has yet to expire.
2476 if (!timeout
|| timeout
->task
)
2477 freezable_schedule();
2479 __set_current_state(TASK_RUNNING
);
2483 * futex_wait_setup() - Prepare to wait on a futex
2484 * @uaddr: the futex userspace address
2485 * @val: the expected value
2486 * @flags: futex flags (FLAGS_SHARED, etc.)
2487 * @q: the associated futex_q
2488 * @hb: storage for hash_bucket pointer to be returned to caller
2490 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2491 * compare it with the expected value. Handle atomic faults internally.
2492 * Return with the hb lock held and a q.key reference on success, and unlocked
2493 * with no q.key reference on failure.
2496 * - 0 - uaddr contains val and hb has been locked;
2497 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2499 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2500 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2506 * Access the page AFTER the hash-bucket is locked.
2507 * Order is important:
2509 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2510 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2512 * The basic logical guarantee of a futex is that it blocks ONLY
2513 * if cond(var) is known to be true at the time of blocking, for
2514 * any cond. If we locked the hash-bucket after testing *uaddr, that
2515 * would open a race condition where we could block indefinitely with
2516 * cond(var) false, which would violate the guarantee.
2518 * On the other hand, we insert q and release the hash-bucket only
2519 * after testing *uaddr. This guarantees that futex_wait() will NOT
2520 * absorb a wakeup if *uaddr does not match the desired values
2521 * while the syscall executes.
2524 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2525 if (unlikely(ret
!= 0))
2529 *hb
= queue_lock(q
);
2531 ret
= get_futex_value_locked(&uval
, uaddr
);
2536 ret
= get_user(uval
, uaddr
);
2540 if (!(flags
& FLAGS_SHARED
))
2543 put_futex_key(&q
->key
);
2554 put_futex_key(&q
->key
);
2558 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2559 ktime_t
*abs_time
, u32 bitset
)
2561 struct hrtimer_sleeper timeout
, *to
= NULL
;
2562 struct restart_block
*restart
;
2563 struct futex_hash_bucket
*hb
;
2564 struct futex_q q
= futex_q_init
;
2574 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2575 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2577 hrtimer_init_sleeper(to
, current
);
2578 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2579 current
->timer_slack_ns
);
2584 * Prepare to wait on uaddr. On success, holds hb lock and increments
2587 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2591 /* queue_me and wait for wakeup, timeout, or a signal. */
2592 futex_wait_queue_me(hb
, &q
, to
);
2594 /* If we were woken (and unqueued), we succeeded, whatever. */
2596 /* unqueue_me() drops q.key ref */
2597 if (!unqueue_me(&q
))
2600 if (to
&& !to
->task
)
2604 * We expect signal_pending(current), but we might be the
2605 * victim of a spurious wakeup as well.
2607 if (!signal_pending(current
))
2614 restart
= ¤t
->restart_block
;
2615 restart
->fn
= futex_wait_restart
;
2616 restart
->futex
.uaddr
= uaddr
;
2617 restart
->futex
.val
= val
;
2618 restart
->futex
.time
= *abs_time
;
2619 restart
->futex
.bitset
= bitset
;
2620 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2622 ret
= -ERESTART_RESTARTBLOCK
;
2626 hrtimer_cancel(&to
->timer
);
2627 destroy_hrtimer_on_stack(&to
->timer
);
2633 static long futex_wait_restart(struct restart_block
*restart
)
2635 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2636 ktime_t t
, *tp
= NULL
;
2638 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2639 t
= restart
->futex
.time
;
2642 restart
->fn
= do_no_restart_syscall
;
2644 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2645 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2650 * Userspace tried a 0 -> TID atomic transition of the futex value
2651 * and failed. The kernel side here does the whole locking operation:
2652 * if there are waiters then it will block as a consequence of relying
2653 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2654 * a 0 value of the futex too.).
2656 * Also serves as futex trylock_pi()'ing, and due semantics.
2658 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
,
2659 ktime_t
*time
, int trylock
)
2661 struct hrtimer_sleeper timeout
, *to
= NULL
;
2662 struct futex_pi_state
*pi_state
= NULL
;
2663 struct rt_mutex_waiter rt_waiter
;
2664 struct futex_hash_bucket
*hb
;
2665 struct futex_q q
= futex_q_init
;
2668 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2671 if (refill_pi_state_cache())
2676 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2678 hrtimer_init_sleeper(to
, current
);
2679 hrtimer_set_expires(&to
->timer
, *time
);
2683 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2684 if (unlikely(ret
!= 0))
2688 hb
= queue_lock(&q
);
2690 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2691 if (unlikely(ret
)) {
2693 * Atomic work succeeded and we got the lock,
2694 * or failed. Either way, we do _not_ block.
2698 /* We got the lock. */
2700 goto out_unlock_put_key
;
2705 * Two reasons for this:
2706 * - Task is exiting and we just wait for the
2708 * - The user space value changed.
2711 put_futex_key(&q
.key
);
2715 goto out_unlock_put_key
;
2719 WARN_ON(!q
.pi_state
);
2722 * Only actually queue now that the atomic ops are done:
2727 ret
= rt_mutex_futex_trylock(&q
.pi_state
->pi_mutex
);
2728 /* Fixup the trylock return value: */
2729 ret
= ret
? 0 : -EWOULDBLOCK
;
2733 rt_mutex_init_waiter(&rt_waiter
);
2736 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2737 * hold it while doing rt_mutex_start_proxy(), because then it will
2738 * include hb->lock in the blocking chain, even through we'll not in
2739 * fact hold it while blocking. This will lead it to report -EDEADLK
2740 * and BUG when futex_unlock_pi() interleaves with this.
2742 * Therefore acquire wait_lock while holding hb->lock, but drop the
2743 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2744 * serializes against futex_unlock_pi() as that does the exact same
2745 * lock handoff sequence.
2747 raw_spin_lock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2748 spin_unlock(q
.lock_ptr
);
2749 ret
= __rt_mutex_start_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
, current
);
2750 raw_spin_unlock_irq(&q
.pi_state
->pi_mutex
.wait_lock
);
2756 spin_lock(q
.lock_ptr
);
2762 hrtimer_start_expires(&to
->timer
, HRTIMER_MODE_ABS
);
2764 ret
= rt_mutex_wait_proxy_lock(&q
.pi_state
->pi_mutex
, to
, &rt_waiter
);
2766 spin_lock(q
.lock_ptr
);
2768 * If we failed to acquire the lock (signal/timeout), we must
2769 * first acquire the hb->lock before removing the lock from the
2770 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2771 * wait lists consistent.
2773 * In particular; it is important that futex_unlock_pi() can not
2774 * observe this inconsistency.
2776 if (ret
&& !rt_mutex_cleanup_proxy_lock(&q
.pi_state
->pi_mutex
, &rt_waiter
))
2781 * Fixup the pi_state owner and possibly acquire the lock if we
2784 res
= fixup_owner(uaddr
, &q
, !ret
);
2786 * If fixup_owner() returned an error, proprogate that. If it acquired
2787 * the lock, clear our -ETIMEDOUT or -EINTR.
2790 ret
= (res
< 0) ? res
: 0;
2793 * If fixup_owner() faulted and was unable to handle the fault, unlock
2794 * it and return the fault to userspace.
2796 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
)) {
2797 pi_state
= q
.pi_state
;
2798 get_pi_state(pi_state
);
2801 /* Unqueue and drop the lock */
2805 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
2806 put_pi_state(pi_state
);
2815 put_futex_key(&q
.key
);
2818 hrtimer_cancel(&to
->timer
);
2819 destroy_hrtimer_on_stack(&to
->timer
);
2821 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2826 ret
= fault_in_user_writeable(uaddr
);
2830 if (!(flags
& FLAGS_SHARED
))
2833 put_futex_key(&q
.key
);
2838 * Userspace attempted a TID -> 0 atomic transition, and failed.
2839 * This is the in-kernel slowpath: we look up the PI state (if any),
2840 * and do the rt-mutex unlock.
2842 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2844 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2845 union futex_key key
= FUTEX_KEY_INIT
;
2846 struct futex_hash_bucket
*hb
;
2847 struct futex_q
*top_waiter
;
2850 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
2854 if (get_user(uval
, uaddr
))
2857 * We release only a lock we actually own:
2859 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2862 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2866 hb
= hash_futex(&key
);
2867 spin_lock(&hb
->lock
);
2870 * Check waiters first. We do not trust user space values at
2871 * all and we at least want to know if user space fiddled
2872 * with the futex value instead of blindly unlocking.
2874 top_waiter
= futex_top_waiter(hb
, &key
);
2876 struct futex_pi_state
*pi_state
= top_waiter
->pi_state
;
2883 * If current does not own the pi_state then the futex is
2884 * inconsistent and user space fiddled with the futex value.
2886 if (pi_state
->owner
!= current
)
2889 get_pi_state(pi_state
);
2891 * By taking wait_lock while still holding hb->lock, we ensure
2892 * there is no point where we hold neither; and therefore
2893 * wake_futex_pi() must observe a state consistent with what we
2896 raw_spin_lock_irq(&pi_state
->pi_mutex
.wait_lock
);
2897 spin_unlock(&hb
->lock
);
2899 /* drops pi_state->pi_mutex.wait_lock */
2900 ret
= wake_futex_pi(uaddr
, uval
, pi_state
);
2902 put_pi_state(pi_state
);
2905 * Success, we're done! No tricky corner cases.
2910 * The atomic access to the futex value generated a
2911 * pagefault, so retry the user-access and the wakeup:
2916 * A unconditional UNLOCK_PI op raced against a waiter
2917 * setting the FUTEX_WAITERS bit. Try again.
2919 if (ret
== -EAGAIN
) {
2920 put_futex_key(&key
);
2924 * wake_futex_pi has detected invalid state. Tell user
2931 * We have no kernel internal state, i.e. no waiters in the
2932 * kernel. Waiters which are about to queue themselves are stuck
2933 * on hb->lock. So we can safely ignore them. We do neither
2934 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2937 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0)) {
2938 spin_unlock(&hb
->lock
);
2943 * If uval has changed, let user space handle it.
2945 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
2948 spin_unlock(&hb
->lock
);
2950 put_futex_key(&key
);
2954 put_futex_key(&key
);
2956 ret
= fault_in_user_writeable(uaddr
);
2964 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2965 * @hb: the hash_bucket futex_q was original enqueued on
2966 * @q: the futex_q woken while waiting to be requeued
2967 * @key2: the futex_key of the requeue target futex
2968 * @timeout: the timeout associated with the wait (NULL if none)
2970 * Detect if the task was woken on the initial futex as opposed to the requeue
2971 * target futex. If so, determine if it was a timeout or a signal that caused
2972 * the wakeup and return the appropriate error code to the caller. Must be
2973 * called with the hb lock held.
2976 * - 0 = no early wakeup detected;
2977 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
2980 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2981 struct futex_q
*q
, union futex_key
*key2
,
2982 struct hrtimer_sleeper
*timeout
)
2987 * With the hb lock held, we avoid races while we process the wakeup.
2988 * We only need to hold hb (and not hb2) to ensure atomicity as the
2989 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2990 * It can't be requeued from uaddr2 to something else since we don't
2991 * support a PI aware source futex for requeue.
2993 if (!match_futex(&q
->key
, key2
)) {
2994 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2996 * We were woken prior to requeue by a timeout or a signal.
2997 * Unqueue the futex_q and determine which it was.
2999 plist_del(&q
->list
, &hb
->chain
);
3002 /* Handle spurious wakeups gracefully */
3004 if (timeout
&& !timeout
->task
)
3006 else if (signal_pending(current
))
3007 ret
= -ERESTARTNOINTR
;
3013 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3014 * @uaddr: the futex we initially wait on (non-pi)
3015 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3016 * the same type, no requeueing from private to shared, etc.
3017 * @val: the expected value of uaddr
3018 * @abs_time: absolute timeout
3019 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3020 * @uaddr2: the pi futex we will take prior to returning to user-space
3022 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3023 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3024 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3025 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3026 * without one, the pi logic would not know which task to boost/deboost, if
3027 * there was a need to.
3029 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3030 * via the following--
3031 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3032 * 2) wakeup on uaddr2 after a requeue
3036 * If 3, cleanup and return -ERESTARTNOINTR.
3038 * If 2, we may then block on trying to take the rt_mutex and return via:
3039 * 5) successful lock
3042 * 8) other lock acquisition failure
3044 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3046 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3052 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
3053 u32 val
, ktime_t
*abs_time
, u32 bitset
,
3056 struct hrtimer_sleeper timeout
, *to
= NULL
;
3057 struct futex_pi_state
*pi_state
= NULL
;
3058 struct rt_mutex_waiter rt_waiter
;
3059 struct futex_hash_bucket
*hb
;
3060 union futex_key key2
= FUTEX_KEY_INIT
;
3061 struct futex_q q
= futex_q_init
;
3064 if (!IS_ENABLED(CONFIG_FUTEX_PI
))
3067 if (uaddr
== uaddr2
)
3075 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
3076 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
3078 hrtimer_init_sleeper(to
, current
);
3079 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
3080 current
->timer_slack_ns
);
3084 * The waiter is allocated on our stack, manipulated by the requeue
3085 * code while we sleep on uaddr.
3087 rt_mutex_init_waiter(&rt_waiter
);
3089 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
3090 if (unlikely(ret
!= 0))
3094 q
.rt_waiter
= &rt_waiter
;
3095 q
.requeue_pi_key
= &key2
;
3098 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3101 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
3106 * The check above which compares uaddrs is not sufficient for
3107 * shared futexes. We need to compare the keys:
3109 if (match_futex(&q
.key
, &key2
)) {
3115 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3116 futex_wait_queue_me(hb
, &q
, to
);
3118 spin_lock(&hb
->lock
);
3119 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
3120 spin_unlock(&hb
->lock
);
3125 * In order for us to be here, we know our q.key == key2, and since
3126 * we took the hb->lock above, we also know that futex_requeue() has
3127 * completed and we no longer have to concern ourselves with a wakeup
3128 * race with the atomic proxy lock acquisition by the requeue code. The
3129 * futex_requeue dropped our key1 reference and incremented our key2
3133 /* Check if the requeue code acquired the second futex for us. */
3136 * Got the lock. We might not be the anticipated owner if we
3137 * did a lock-steal - fix up the PI-state in that case.
3139 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
3140 spin_lock(q
.lock_ptr
);
3141 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
3142 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3143 pi_state
= q
.pi_state
;
3144 get_pi_state(pi_state
);
3147 * Drop the reference to the pi state which
3148 * the requeue_pi() code acquired for us.
3150 put_pi_state(q
.pi_state
);
3151 spin_unlock(q
.lock_ptr
);
3154 struct rt_mutex
*pi_mutex
;
3157 * We have been woken up by futex_unlock_pi(), a timeout, or a
3158 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3161 WARN_ON(!q
.pi_state
);
3162 pi_mutex
= &q
.pi_state
->pi_mutex
;
3163 ret
= rt_mutex_wait_proxy_lock(pi_mutex
, to
, &rt_waiter
);
3165 spin_lock(q
.lock_ptr
);
3166 if (ret
&& !rt_mutex_cleanup_proxy_lock(pi_mutex
, &rt_waiter
))
3169 debug_rt_mutex_free_waiter(&rt_waiter
);
3171 * Fixup the pi_state owner and possibly acquire the lock if we
3174 res
= fixup_owner(uaddr2
, &q
, !ret
);
3176 * If fixup_owner() returned an error, proprogate that. If it
3177 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3180 ret
= (res
< 0) ? res
: 0;
3183 * If fixup_pi_state_owner() faulted and was unable to handle
3184 * the fault, unlock the rt_mutex and return the fault to
3187 if (ret
&& rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
) {
3188 pi_state
= q
.pi_state
;
3189 get_pi_state(pi_state
);
3192 /* Unqueue and drop the lock. */
3197 rt_mutex_futex_unlock(&pi_state
->pi_mutex
);
3198 put_pi_state(pi_state
);
3201 if (ret
== -EINTR
) {
3203 * We've already been requeued, but cannot restart by calling
3204 * futex_lock_pi() directly. We could restart this syscall, but
3205 * it would detect that the user space "val" changed and return
3206 * -EWOULDBLOCK. Save the overhead of the restart and return
3207 * -EWOULDBLOCK directly.
3213 put_futex_key(&q
.key
);
3215 put_futex_key(&key2
);
3219 hrtimer_cancel(&to
->timer
);
3220 destroy_hrtimer_on_stack(&to
->timer
);
3226 * Support for robust futexes: the kernel cleans up held futexes at
3229 * Implementation: user-space maintains a per-thread list of locks it
3230 * is holding. Upon do_exit(), the kernel carefully walks this list,
3231 * and marks all locks that are owned by this thread with the
3232 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3233 * always manipulated with the lock held, so the list is private and
3234 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3235 * field, to allow the kernel to clean up if the thread dies after
3236 * acquiring the lock, but just before it could have added itself to
3237 * the list. There can only be one such pending lock.
3241 * sys_set_robust_list() - Set the robust-futex list head of a task
3242 * @head: pointer to the list-head
3243 * @len: length of the list-head, as userspace expects
3245 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
3248 if (!futex_cmpxchg_enabled
)
3251 * The kernel knows only one size for now:
3253 if (unlikely(len
!= sizeof(*head
)))
3256 current
->robust_list
= head
;
3262 * sys_get_robust_list() - Get the robust-futex list head of a task
3263 * @pid: pid of the process [zero for current task]
3264 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3265 * @len_ptr: pointer to a length field, the kernel fills in the header size
3267 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
3268 struct robust_list_head __user
* __user
*, head_ptr
,
3269 size_t __user
*, len_ptr
)
3271 struct robust_list_head __user
*head
;
3273 struct task_struct
*p
;
3275 if (!futex_cmpxchg_enabled
)
3284 p
= find_task_by_vpid(pid
);
3290 if (!ptrace_may_access(p
, PTRACE_MODE_READ_REALCREDS
))
3293 head
= p
->robust_list
;
3296 if (put_user(sizeof(*head
), len_ptr
))
3298 return put_user(head
, head_ptr
);
3307 * Process a futex-list entry, check whether it's owned by the
3308 * dying task, and do notification if so:
3310 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
3312 u32 uval
, uninitialized_var(nval
), mval
;
3315 if (get_user(uval
, uaddr
))
3318 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
3320 * Ok, this dying thread is truly holding a futex
3321 * of interest. Set the OWNER_DIED bit atomically
3322 * via cmpxchg, and if the value had FUTEX_WAITERS
3323 * set, wake up a waiter (if any). (We have to do a
3324 * futex_wake() even if OWNER_DIED is already set -
3325 * to handle the rare but possible case of recursive
3326 * thread-death.) The rest of the cleanup is done in
3329 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
3331 * We are not holding a lock here, but we want to have
3332 * the pagefault_disable/enable() protection because
3333 * we want to handle the fault gracefully. If the
3334 * access fails we try to fault in the futex with R/W
3335 * verification via get_user_pages. get_user() above
3336 * does not guarantee R/W access. If that fails we
3337 * give up and leave the futex locked.
3339 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
3340 if (fault_in_user_writeable(uaddr
))
3348 * Wake robust non-PI futexes here. The wakeup of
3349 * PI futexes happens in exit_pi_state():
3351 if (!pi
&& (uval
& FUTEX_WAITERS
))
3352 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
3358 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3360 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
3361 struct robust_list __user
* __user
*head
,
3364 unsigned long uentry
;
3366 if (get_user(uentry
, (unsigned long __user
*)head
))
3369 *entry
= (void __user
*)(uentry
& ~1UL);
3376 * Walk curr->robust_list (very carefully, it's a userspace list!)
3377 * and mark any locks found there dead, and notify any waiters.
3379 * We silently return on any sign of list-walking problem.
3381 void exit_robust_list(struct task_struct
*curr
)
3383 struct robust_list_head __user
*head
= curr
->robust_list
;
3384 struct robust_list __user
*entry
, *next_entry
, *pending
;
3385 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
3386 unsigned int uninitialized_var(next_pi
);
3387 unsigned long futex_offset
;
3390 if (!futex_cmpxchg_enabled
)
3394 * Fetch the list head (which was registered earlier, via
3395 * sys_set_robust_list()):
3397 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
3400 * Fetch the relative futex offset:
3402 if (get_user(futex_offset
, &head
->futex_offset
))
3405 * Fetch any possibly pending lock-add first, and handle it
3408 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
3411 next_entry
= NULL
; /* avoid warning with gcc */
3412 while (entry
!= &head
->list
) {
3414 * Fetch the next entry in the list before calling
3415 * handle_futex_death:
3417 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
3419 * A pending lock might already be on the list, so
3420 * don't process it twice:
3422 if (entry
!= pending
)
3423 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
3431 * Avoid excessively long or circular lists:
3440 handle_futex_death((void __user
*)pending
+ futex_offset
,
3444 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
3445 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
3447 int cmd
= op
& FUTEX_CMD_MASK
;
3448 unsigned int flags
= 0;
3450 if (!(op
& FUTEX_PRIVATE_FLAG
))
3451 flags
|= FLAGS_SHARED
;
3453 if (op
& FUTEX_CLOCK_REALTIME
) {
3454 flags
|= FLAGS_CLOCKRT
;
3455 if (cmd
!= FUTEX_WAIT
&& cmd
!= FUTEX_WAIT_BITSET
&& \
3456 cmd
!= FUTEX_WAIT_REQUEUE_PI
)
3462 case FUTEX_UNLOCK_PI
:
3463 case FUTEX_TRYLOCK_PI
:
3464 case FUTEX_WAIT_REQUEUE_PI
:
3465 case FUTEX_CMP_REQUEUE_PI
:
3466 if (!futex_cmpxchg_enabled
)
3472 val3
= FUTEX_BITSET_MATCH_ANY
;
3473 case FUTEX_WAIT_BITSET
:
3474 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
3476 val3
= FUTEX_BITSET_MATCH_ANY
;
3477 case FUTEX_WAKE_BITSET
:
3478 return futex_wake(uaddr
, flags
, val
, val3
);
3480 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
3481 case FUTEX_CMP_REQUEUE
:
3482 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
3484 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
3486 return futex_lock_pi(uaddr
, flags
, timeout
, 0);
3487 case FUTEX_UNLOCK_PI
:
3488 return futex_unlock_pi(uaddr
, flags
);
3489 case FUTEX_TRYLOCK_PI
:
3490 return futex_lock_pi(uaddr
, flags
, NULL
, 1);
3491 case FUTEX_WAIT_REQUEUE_PI
:
3492 val3
= FUTEX_BITSET_MATCH_ANY
;
3493 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
3495 case FUTEX_CMP_REQUEUE_PI
:
3496 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
3502 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
3503 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
3507 ktime_t t
, *tp
= NULL
;
3509 int cmd
= op
& FUTEX_CMD_MASK
;
3511 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
3512 cmd
== FUTEX_WAIT_BITSET
||
3513 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
3514 if (unlikely(should_fail_futex(!(op
& FUTEX_PRIVATE_FLAG
))))
3516 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
3518 if (!timespec_valid(&ts
))
3521 t
= timespec_to_ktime(ts
);
3522 if (cmd
== FUTEX_WAIT
)
3523 t
= ktime_add_safe(ktime_get(), t
);
3527 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3528 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3530 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
3531 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
3532 val2
= (u32
) (unsigned long) utime
;
3534 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
3537 static void __init
futex_detect_cmpxchg(void)
3539 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3543 * This will fail and we want it. Some arch implementations do
3544 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3545 * functionality. We want to know that before we call in any
3546 * of the complex code paths. Also we want to prevent
3547 * registration of robust lists in that case. NULL is
3548 * guaranteed to fault and we get -EFAULT on functional
3549 * implementation, the non-functional ones will return
3552 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3553 futex_cmpxchg_enabled
= 1;
3557 static int __init
futex_init(void)
3559 unsigned int futex_shift
;
3562 #if CONFIG_BASE_SMALL
3563 futex_hashsize
= 16;
3565 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3568 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3570 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3572 futex_hashsize
, futex_hashsize
);
3573 futex_hashsize
= 1UL << futex_shift
;
3575 futex_detect_cmpxchg();
3577 for (i
= 0; i
< futex_hashsize
; i
++) {
3578 atomic_set(&futex_queues
[i
].waiters
, 0);
3579 plist_head_init(&futex_queues
[i
].chain
);
3580 spin_lock_init(&futex_queues
[i
].lock
);
3585 core_initcall(futex_init
);