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/memblock.h>
70 #include <linux/fault-inject.h>
71 #include <linux/refcount.h>
73 #include <asm/futex.h>
75 #include "locking/rtmutex_common.h"
78 * READ this before attempting to hack on futexes!
80 * Basic futex operation and ordering guarantees
81 * =============================================
83 * The waiter reads the futex value in user space and calls
84 * futex_wait(). This function computes the hash bucket and acquires
85 * the hash bucket lock. After that it reads the futex user space value
86 * again and verifies that the data has not changed. If it has not changed
87 * it enqueues itself into the hash bucket, releases the hash bucket lock
90 * The waker side modifies the user space value of the futex and calls
91 * futex_wake(). This function computes the hash bucket and acquires the
92 * hash bucket lock. Then it looks for waiters on that futex in the hash
93 * bucket and wakes them.
95 * In futex wake up scenarios where no tasks are blocked on a futex, taking
96 * the hb spinlock can be avoided and simply return. In order for this
97 * optimization to work, ordering guarantees must exist so that the waiter
98 * being added to the list is acknowledged when the list is concurrently being
99 * checked by the waker, avoiding scenarios like the following:
103 * sys_futex(WAIT, futex, val);
104 * futex_wait(futex, val);
107 * sys_futex(WAKE, futex);
112 * lock(hash_bucket(futex));
114 * unlock(hash_bucket(futex));
117 * This would cause the waiter on CPU 0 to wait forever because it
118 * missed the transition of the user space value from val to newval
119 * and the waker did not find the waiter in the hash bucket queue.
121 * The correct serialization ensures that a waiter either observes
122 * the changed user space value before blocking or is woken by a
127 * sys_futex(WAIT, futex, val);
128 * futex_wait(futex, val);
131 * smp_mb(); (A) <-- paired with -.
133 * lock(hash_bucket(futex)); |
137 * | sys_futex(WAKE, futex);
138 * | futex_wake(futex);
140 * `--------> smp_mb(); (B)
143 * unlock(hash_bucket(futex));
144 * schedule(); if (waiters)
145 * lock(hash_bucket(futex));
146 * else wake_waiters(futex);
147 * waiters--; (b) unlock(hash_bucket(futex));
149 * Where (A) orders the waiters increment and the futex value read through
150 * atomic operations (see hb_waiters_inc) and where (B) orders the write
151 * to futex and the waiters read -- this is done by the barriers for both
152 * shared and private futexes in get_futex_key_refs().
154 * This yields the following case (where X:=waiters, Y:=futex):
162 * Which guarantees that x==0 && y==0 is impossible; which translates back into
163 * the guarantee that we cannot both miss the futex variable change and the
166 * Note that a new waiter is accounted for in (a) even when it is possible that
167 * the wait call can return error, in which case we backtrack from it in (b).
168 * Refer to the comment in queue_lock().
170 * Similarly, in order to account for waiters being requeued on another
171 * address we always increment the waiters for the destination bucket before
172 * acquiring the lock. It then decrements them again after releasing it -
173 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
174 * will do the additional required waiter count housekeeping. This is done for
175 * double_lock_hb() and double_unlock_hb(), respectively.
178 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
179 #define futex_cmpxchg_enabled 1
181 static int __read_mostly futex_cmpxchg_enabled;
185 * Futex flags used to encode options to functions and preserve them across
189 # define FLAGS_SHARED 0x01
192 * NOMMU does not have per process address space. Let the compiler optimize
195 # define FLAGS_SHARED 0x00
197 #define FLAGS_CLOCKRT 0x02
198 #define FLAGS_HAS_TIMEOUT 0x04
201 * Priority Inheritance state:
203 struct futex_pi_state {
205 * list of 'owned' pi_state instances - these have to be
206 * cleaned up in do_exit() if the task exits prematurely:
208 struct list_head list;
213 struct rt_mutex pi_mutex;
215 struct task_struct *owner;
219 } __randomize_layout;
222 * struct futex_q - The hashed futex queue entry, one per waiting task
223 * @list: priority-sorted list of tasks waiting on this futex
224 * @task: the task waiting on the futex
225 * @lock_ptr: the hash bucket lock
226 * @key: the key the futex is hashed on
227 * @pi_state: optional priority inheritance state
228 * @rt_waiter: rt_waiter storage for use with requeue_pi
229 * @requeue_pi_key: the requeue_pi target futex key
230 * @bitset: bitset for the optional bitmasked wakeup
232 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
233 * we can wake only the relevant ones (hashed queues may be shared).
235 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
236 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
237 * The order of wakeup is always to make the first condition true, then
240 * PI futexes are typically woken before they are removed from the hash list via
241 * the rt_mutex code. See unqueue_me_pi().
244 struct plist_node list;
246 struct task_struct *task;
247 spinlock_t *lock_ptr;
249 struct futex_pi_state *pi_state;
250 struct rt_mutex_waiter *rt_waiter;
251 union futex_key *requeue_pi_key;
253 } __randomize_layout;
255 static const struct futex_q futex_q_init = {
256 /* list gets initialized in queue_me()*/
257 .key = FUTEX_KEY_INIT,
258 .bitset = FUTEX_BITSET_MATCH_ANY
262 * Hash buckets are shared by all the futex_keys that hash to the same
263 * location. Each key may have multiple futex_q structures, one for each task
264 * waiting on a futex.
266 struct futex_hash_bucket {
269 struct plist_head chain;
270 } ____cacheline_aligned_in_smp;
273 * The base of the bucket array and its size are always used together
274 * (after initialization only in hash_futex()), so ensure that they
275 * reside in the same cacheline.
278 struct futex_hash_bucket *queues;
279 unsigned long hashsize;
280 } __futex_data __read_mostly __aligned(2*sizeof(long));
281 #define futex_queues (__futex_data.queues)
282 #define futex_hashsize (__futex_data.hashsize)
286 * Fault injections for futexes.
288 #ifdef CONFIG_FAIL_FUTEX
291 struct fault_attr attr;
295 .attr = FAULT_ATTR_INITIALIZER,
296 .ignore_private = false,
299 static int __init setup_fail_futex(char *str)
301 return setup_fault_attr(&fail_futex.attr, str);
303 __setup("fail_futex=", setup_fail_futex);
305 static bool should_fail_futex(bool fshared)
307 if (fail_futex.ignore_private && !fshared)
310 return should_fail(&fail_futex.attr, 1);
313 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
315 static int __init fail_futex_debugfs(void)
317 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
320 dir = fault_create_debugfs_attr("fail_futex", NULL,
325 debugfs_create_bool("ignore-private", mode, dir,
326 &fail_futex.ignore_private);
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);
487 * futex_setup_timer - set up the sleeping hrtimer.
488 * @time: ptr to the given timeout value
489 * @timeout: the hrtimer_sleeper structure to be set up
490 * @flags: futex flags
491 * @range_ns: optional range in ns
493 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
496 static inline struct hrtimer_sleeper *
497 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
498 int flags, u64 range_ns)
503 hrtimer_init_on_stack(&timeout->timer, (flags & FLAGS_CLOCKRT) ?
504 CLOCK_REALTIME : CLOCK_MONOTONIC,
506 hrtimer_init_sleeper(timeout, current);
509 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
510 * effectively the same as calling hrtimer_set_expires().
512 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
518 * get_futex_key() - Get parameters which are the keys for a futex
519 * @uaddr: virtual address of the futex
520 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
521 * @key: address where result is stored.
522 * @rw: mapping needs to be read/write (values: FUTEX_READ,
525 * Return: a negative error code or 0
527 * The key words are stored in @key on success.
529 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
530 * offset_within_page). For private mappings, it's (uaddr, current->mm).
531 * We can usually work out the index without swapping in the page.
533 * lock_page() might sleep, the caller should not hold a spinlock.
536 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
538 unsigned long address = (unsigned long)uaddr;
539 struct mm_struct *mm = current->mm;
540 struct page *page, *tail;
541 struct address_space *mapping;
545 * The futex address must be "naturally" aligned.
547 key->both.offset = address % PAGE_SIZE;
548 if (unlikely((address % sizeof(u32)) != 0))
550 address -= key->both.offset;
552 if (unlikely(!access_ok(uaddr, sizeof(u32))))
555 if (unlikely(should_fail_futex(fshared)))
559 * PROCESS_PRIVATE futexes are fast.
560 * As the mm cannot disappear under us and the 'key' only needs
561 * virtual address, we dont even have to find the underlying vma.
562 * Note : We do have to check 'uaddr' is a valid user address,
563 * but access_ok() should be faster than find_vma()
566 key->private.mm = mm;
567 key->private.address = address;
568 get_futex_key_refs(key); /* implies smp_mb(); (B) */
573 /* Ignore any VERIFY_READ mapping (futex common case) */
574 if (unlikely(should_fail_futex(fshared)))
577 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
579 * If write access is not required (eg. FUTEX_WAIT), try
580 * and get read-only access.
582 if (err == -EFAULT && rw == FUTEX_READ) {
583 err = get_user_pages_fast(address, 1, 0, &page);
592 * The treatment of mapping from this point on is critical. The page
593 * lock protects many things but in this context the page lock
594 * stabilizes mapping, prevents inode freeing in the shared
595 * file-backed region case and guards against movement to swap cache.
597 * Strictly speaking the page lock is not needed in all cases being
598 * considered here and page lock forces unnecessarily serialization
599 * From this point on, mapping will be re-verified if necessary and
600 * page lock will be acquired only if it is unavoidable
602 * Mapping checks require the head page for any compound page so the
603 * head page and mapping is looked up now. For anonymous pages, it
604 * does not matter if the page splits in the future as the key is
605 * based on the address. For filesystem-backed pages, the tail is
606 * required as the index of the page determines the key. For
607 * base pages, there is no tail page and tail == page.
610 page = compound_head(page);
611 mapping = READ_ONCE(page->mapping);
614 * If page->mapping is NULL, then it cannot be a PageAnon
615 * page; but it might be the ZERO_PAGE or in the gate area or
616 * in a special mapping (all cases which we are happy to fail);
617 * or it may have been a good file page when get_user_pages_fast
618 * found it, but truncated or holepunched or subjected to
619 * invalidate_complete_page2 before we got the page lock (also
620 * cases which we are happy to fail). And we hold a reference,
621 * so refcount care in invalidate_complete_page's remove_mapping
622 * prevents drop_caches from setting mapping to NULL beneath us.
624 * The case we do have to guard against is when memory pressure made
625 * shmem_writepage move it from filecache to swapcache beneath us:
626 * an unlikely race, but we do need to retry for page->mapping.
628 if (unlikely(!mapping)) {
632 * Page lock is required to identify which special case above
633 * applies. If this is really a shmem page then the page lock
634 * will prevent unexpected transitions.
637 shmem_swizzled = PageSwapCache(page) || page->mapping;
648 * Private mappings are handled in a simple way.
650 * If the futex key is stored on an anonymous page, then the associated
651 * object is the mm which is implicitly pinned by the calling process.
653 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
654 * it's a read-only handle, it's expected that futexes attach to
655 * the object not the particular process.
657 if (PageAnon(page)) {
659 * A RO anonymous page will never change and thus doesn't make
660 * sense for futex operations.
662 if (unlikely(should_fail_futex(fshared)) || ro) {
667 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
668 key->private.mm = mm;
669 key->private.address = address;
671 get_futex_key_refs(key); /* implies smp_mb(); (B) */
677 * The associated futex object in this case is the inode and
678 * the page->mapping must be traversed. Ordinarily this should
679 * be stabilised under page lock but it's not strictly
680 * necessary in this case as we just want to pin the inode, not
681 * update the radix tree or anything like that.
683 * The RCU read lock is taken as the inode is finally freed
684 * under RCU. If the mapping still matches expectations then the
685 * mapping->host can be safely accessed as being a valid inode.
689 if (READ_ONCE(page->mapping) != mapping) {
696 inode = READ_ONCE(mapping->host);
705 * Take a reference unless it is about to be freed. Previously
706 * this reference was taken by ihold under the page lock
707 * pinning the inode in place so i_lock was unnecessary. The
708 * only way for this check to fail is if the inode was
709 * truncated in parallel which is almost certainly an
710 * application bug. In such a case, just retry.
712 * We are not calling into get_futex_key_refs() in file-backed
713 * cases, therefore a successful atomic_inc return below will
714 * guarantee that get_futex_key() will still imply smp_mb(); (B).
716 if (!atomic_inc_not_zero(&inode->i_count)) {
723 /* Should be impossible but lets be paranoid for now */
724 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
732 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
733 key->shared.inode = inode;
734 key->shared.pgoff = basepage_index(tail);
743 static inline void put_futex_key(union futex_key *key)
745 drop_futex_key_refs(key);
749 * fault_in_user_writeable() - Fault in user address and verify RW access
750 * @uaddr: pointer to faulting user space address
752 * Slow path to fixup the fault we just took in the atomic write
755 * We have no generic implementation of a non-destructive write to the
756 * user address. We know that we faulted in the atomic pagefault
757 * disabled section so we can as well avoid the #PF overhead by
758 * calling get_user_pages() right away.
760 static int fault_in_user_writeable(u32 __user *uaddr)
762 struct mm_struct *mm = current->mm;
765 down_read(&mm->mmap_sem);
766 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
767 FAULT_FLAG_WRITE, NULL);
768 up_read(&mm->mmap_sem);
770 return ret < 0 ? ret : 0;
774 * futex_top_waiter() - Return the highest priority waiter on a futex
775 * @hb: the hash bucket the futex_q's reside in
776 * @key: the futex key (to distinguish it from other futex futex_q's)
778 * Must be called with the hb lock held.
780 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
781 union futex_key *key)
783 struct futex_q *this;
785 plist_for_each_entry(this, &hb->chain, list) {
786 if (match_futex(&this->key, key))
792 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
793 u32 uval, u32 newval)
798 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
804 static int get_futex_value_locked(u32 *dest, u32 __user *from)
809 ret = __get_user(*dest, from);
812 return ret ? -EFAULT : 0;
819 static int refill_pi_state_cache(void)
821 struct futex_pi_state *pi_state;
823 if (likely(current->pi_state_cache))
826 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
831 INIT_LIST_HEAD(&pi_state->list);
832 /* pi_mutex gets initialized later */
833 pi_state->owner = NULL;
834 refcount_set(&pi_state->refcount, 1);
835 pi_state->key = FUTEX_KEY_INIT;
837 current->pi_state_cache = pi_state;
842 static struct futex_pi_state *alloc_pi_state(void)
844 struct futex_pi_state *pi_state = current->pi_state_cache;
847 current->pi_state_cache = NULL;
852 static void get_pi_state(struct futex_pi_state *pi_state)
854 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
858 * Drops a reference to the pi_state object and frees or caches it
859 * when the last reference is gone.
861 static void put_pi_state(struct futex_pi_state *pi_state)
866 if (!refcount_dec_and_test(&pi_state->refcount))
870 * If pi_state->owner is NULL, the owner is most probably dying
871 * and has cleaned up the pi_state already
873 if (pi_state->owner) {
874 struct task_struct *owner;
876 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
877 owner = pi_state->owner;
879 raw_spin_lock(&owner->pi_lock);
880 list_del_init(&pi_state->list);
881 raw_spin_unlock(&owner->pi_lock);
883 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
884 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
887 if (current->pi_state_cache) {
891 * pi_state->list is already empty.
892 * clear pi_state->owner.
893 * refcount is at 0 - put it back to 1.
895 pi_state->owner = NULL;
896 refcount_set(&pi_state->refcount, 1);
897 current->pi_state_cache = pi_state;
901 #ifdef CONFIG_FUTEX_PI
904 * This task is holding PI mutexes at exit time => bad.
905 * Kernel cleans up PI-state, but userspace is likely hosed.
906 * (Robust-futex cleanup is separate and might save the day for userspace.)
908 void exit_pi_state_list(struct task_struct *curr)
910 struct list_head *next, *head = &curr->pi_state_list;
911 struct futex_pi_state *pi_state;
912 struct futex_hash_bucket *hb;
913 union futex_key key = FUTEX_KEY_INIT;
915 if (!futex_cmpxchg_enabled)
918 * We are a ZOMBIE and nobody can enqueue itself on
919 * pi_state_list anymore, but we have to be careful
920 * versus waiters unqueueing themselves:
922 raw_spin_lock_irq(&curr->pi_lock);
923 while (!list_empty(head)) {
925 pi_state = list_entry(next, struct futex_pi_state, list);
927 hb = hash_futex(&key);
930 * We can race against put_pi_state() removing itself from the
931 * list (a waiter going away). put_pi_state() will first
932 * decrement the reference count and then modify the list, so
933 * its possible to see the list entry but fail this reference
936 * In that case; drop the locks to let put_pi_state() make
937 * progress and retry the loop.
939 if (!refcount_inc_not_zero(&pi_state->refcount)) {
940 raw_spin_unlock_irq(&curr->pi_lock);
942 raw_spin_lock_irq(&curr->pi_lock);
945 raw_spin_unlock_irq(&curr->pi_lock);
947 spin_lock(&hb->lock);
948 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
949 raw_spin_lock(&curr->pi_lock);
951 * We dropped the pi-lock, so re-check whether this
952 * task still owns the PI-state:
954 if (head->next != next) {
955 /* retain curr->pi_lock for the loop invariant */
956 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
957 spin_unlock(&hb->lock);
958 put_pi_state(pi_state);
962 WARN_ON(pi_state->owner != curr);
963 WARN_ON(list_empty(&pi_state->list));
964 list_del_init(&pi_state->list);
965 pi_state->owner = NULL;
967 raw_spin_unlock(&curr->pi_lock);
968 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
969 spin_unlock(&hb->lock);
971 rt_mutex_futex_unlock(&pi_state->pi_mutex);
972 put_pi_state(pi_state);
974 raw_spin_lock_irq(&curr->pi_lock);
976 raw_spin_unlock_irq(&curr->pi_lock);
982 * We need to check the following states:
984 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
986 * [1] NULL | --- | --- | 0 | 0/1 | Valid
987 * [2] NULL | --- | --- | >0 | 0/1 | Valid
989 * [3] Found | NULL | -- | Any | 0/1 | Invalid
991 * [4] Found | Found | NULL | 0 | 1 | Valid
992 * [5] Found | Found | NULL | >0 | 1 | Invalid
994 * [6] Found | Found | task | 0 | 1 | Valid
996 * [7] Found | Found | NULL | Any | 0 | Invalid
998 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
999 * [9] Found | Found | task | 0 | 0 | Invalid
1000 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1002 * [1] Indicates that the kernel can acquire the futex atomically. We
1003 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1005 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1006 * thread is found then it indicates that the owner TID has died.
1008 * [3] Invalid. The waiter is queued on a non PI futex
1010 * [4] Valid state after exit_robust_list(), which sets the user space
1011 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1013 * [5] The user space value got manipulated between exit_robust_list()
1014 * and exit_pi_state_list()
1016 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1017 * the pi_state but cannot access the user space value.
1019 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1021 * [8] Owner and user space value match
1023 * [9] There is no transient state which sets the user space TID to 0
1024 * except exit_robust_list(), but this is indicated by the
1025 * FUTEX_OWNER_DIED bit. See [4]
1027 * [10] There is no transient state which leaves owner and user space
1031 * Serialization and lifetime rules:
1035 * hb -> futex_q, relation
1036 * futex_q -> pi_state, relation
1038 * (cannot be raw because hb can contain arbitrary amount
1041 * pi_mutex->wait_lock:
1045 * (and pi_mutex 'obviously')
1049 * p->pi_state_list -> pi_state->list, relation
1051 * pi_state->refcount:
1059 * pi_mutex->wait_lock
1065 * Validate that the existing waiter has a pi_state and sanity check
1066 * the pi_state against the user space value. If correct, attach to
1069 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1070 struct futex_pi_state *pi_state,
1071 struct futex_pi_state **ps)
1073 pid_t pid = uval & FUTEX_TID_MASK;
1078 * Userspace might have messed up non-PI and PI futexes [3]
1080 if (unlikely(!pi_state))
1084 * We get here with hb->lock held, and having found a
1085 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1086 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1087 * which in turn means that futex_lock_pi() still has a reference on
1090 * The waiter holding a reference on @pi_state also protects against
1091 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1092 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1093 * free pi_state before we can take a reference ourselves.
1095 WARN_ON(!refcount_read(&pi_state->refcount));
1098 * Now that we have a pi_state, we can acquire wait_lock
1099 * and do the state validation.
1101 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1104 * Since {uval, pi_state} is serialized by wait_lock, and our current
1105 * uval was read without holding it, it can have changed. Verify it
1106 * still is what we expect it to be, otherwise retry the entire
1109 if (get_futex_value_locked(&uval2, uaddr))
1116 * Handle the owner died case:
1118 if (uval & FUTEX_OWNER_DIED) {
1120 * exit_pi_state_list sets owner to NULL and wakes the
1121 * topmost waiter. The task which acquires the
1122 * pi_state->rt_mutex will fixup owner.
1124 if (!pi_state->owner) {
1126 * No pi state owner, but the user space TID
1127 * is not 0. Inconsistent state. [5]
1132 * Take a ref on the state and return success. [4]
1138 * If TID is 0, then either the dying owner has not
1139 * yet executed exit_pi_state_list() or some waiter
1140 * acquired the rtmutex in the pi state, but did not
1141 * yet fixup the TID in user space.
1143 * Take a ref on the state and return success. [6]
1149 * If the owner died bit is not set, then the pi_state
1150 * must have an owner. [7]
1152 if (!pi_state->owner)
1157 * Bail out if user space manipulated the futex value. If pi
1158 * state exists then the owner TID must be the same as the
1159 * user space TID. [9/10]
1161 if (pid != task_pid_vnr(pi_state->owner))
1165 get_pi_state(pi_state);
1166 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1183 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1187 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1188 struct task_struct *tsk)
1193 * If PF_EXITPIDONE is not yet set, then try again.
1195 if (tsk && !(tsk->flags & PF_EXITPIDONE))
1199 * Reread the user space value to handle the following situation:
1203 * sys_exit() sys_futex()
1204 * do_exit() futex_lock_pi()
1205 * futex_lock_pi_atomic()
1206 * exit_signals(tsk) No waiters:
1207 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1208 * mm_release(tsk) Set waiter bit
1209 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1210 * Set owner died attach_to_pi_owner() {
1211 * *uaddr = 0xC0000000; tsk = get_task(PID);
1212 * } if (!tsk->flags & PF_EXITING) {
1214 * tsk->flags |= PF_EXITPIDONE; } else {
1215 * if (!(tsk->flags & PF_EXITPIDONE))
1217 * return -ESRCH; <--- FAIL
1220 * Returning ESRCH unconditionally is wrong here because the
1221 * user space value has been changed by the exiting task.
1223 * The same logic applies to the case where the exiting task is
1226 if (get_futex_value_locked(&uval2, uaddr))
1229 /* If the user space value has changed, try again. */
1234 * The exiting task did not have a robust list, the robust list was
1235 * corrupted or the user space value in *uaddr is simply bogus.
1236 * Give up and tell user space.
1242 * Lookup the task for the TID provided from user space and attach to
1243 * it after doing proper sanity checks.
1245 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1246 struct futex_pi_state **ps)
1248 pid_t pid = uval & FUTEX_TID_MASK;
1249 struct futex_pi_state *pi_state;
1250 struct task_struct *p;
1253 * We are the first waiter - try to look up the real owner and attach
1254 * the new pi_state to it, but bail out when TID = 0 [1]
1256 * The !pid check is paranoid. None of the call sites should end up
1257 * with pid == 0, but better safe than sorry. Let the caller retry
1261 p = find_get_task_by_vpid(pid);
1263 return handle_exit_race(uaddr, uval, NULL);
1265 if (unlikely(p->flags & PF_KTHREAD)) {
1271 * We need to look at the task state flags to figure out,
1272 * whether the task is exiting. To protect against the do_exit
1273 * change of the task flags, we do this protected by
1276 raw_spin_lock_irq(&p->pi_lock);
1277 if (unlikely(p->flags & PF_EXITING)) {
1279 * The task is on the way out. When PF_EXITPIDONE is
1280 * set, we know that the task has finished the
1283 int ret = handle_exit_race(uaddr, uval, p);
1285 raw_spin_unlock_irq(&p->pi_lock);
1291 * No existing pi state. First waiter. [2]
1293 * This creates pi_state, we have hb->lock held, this means nothing can
1294 * observe this state, wait_lock is irrelevant.
1296 pi_state = alloc_pi_state();
1299 * Initialize the pi_mutex in locked state and make @p
1302 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1304 /* Store the key for possible exit cleanups: */
1305 pi_state->key = *key;
1307 WARN_ON(!list_empty(&pi_state->list));
1308 list_add(&pi_state->list, &p->pi_state_list);
1310 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1311 * because there is no concurrency as the object is not published yet.
1313 pi_state->owner = p;
1314 raw_spin_unlock_irq(&p->pi_lock);
1323 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1324 struct futex_hash_bucket *hb,
1325 union futex_key *key, struct futex_pi_state **ps)
1327 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1330 * If there is a waiter on that futex, validate it and
1331 * attach to the pi_state when the validation succeeds.
1334 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1337 * We are the first waiter - try to look up the owner based on
1338 * @uval and attach to it.
1340 return attach_to_pi_owner(uaddr, uval, key, ps);
1343 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1346 u32 uninitialized_var(curval);
1348 if (unlikely(should_fail_futex(true)))
1351 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1355 /* If user space value changed, let the caller retry */
1356 return curval != uval ? -EAGAIN : 0;
1360 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1361 * @uaddr: the pi futex user address
1362 * @hb: the pi futex hash bucket
1363 * @key: the futex key associated with uaddr and hb
1364 * @ps: the pi_state pointer where we store the result of the
1366 * @task: the task to perform the atomic lock work for. This will
1367 * be "current" except in the case of requeue pi.
1368 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1371 * - 0 - ready to wait;
1372 * - 1 - acquired the lock;
1375 * The hb->lock and futex_key refs shall be held by the caller.
1377 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1378 union futex_key *key,
1379 struct futex_pi_state **ps,
1380 struct task_struct *task, int set_waiters)
1382 u32 uval, newval, vpid = task_pid_vnr(task);
1383 struct futex_q *top_waiter;
1387 * Read the user space value first so we can validate a few
1388 * things before proceeding further.
1390 if (get_futex_value_locked(&uval, uaddr))
1393 if (unlikely(should_fail_futex(true)))
1399 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1402 if ((unlikely(should_fail_futex(true))))
1406 * Lookup existing state first. If it exists, try to attach to
1409 top_waiter = futex_top_waiter(hb, key);
1411 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1414 * No waiter and user TID is 0. We are here because the
1415 * waiters or the owner died bit is set or called from
1416 * requeue_cmp_pi or for whatever reason something took the
1419 if (!(uval & FUTEX_TID_MASK)) {
1421 * We take over the futex. No other waiters and the user space
1422 * TID is 0. We preserve the owner died bit.
1424 newval = uval & FUTEX_OWNER_DIED;
1427 /* The futex requeue_pi code can enforce the waiters bit */
1429 newval |= FUTEX_WAITERS;
1431 ret = lock_pi_update_atomic(uaddr, uval, newval);
1432 /* If the take over worked, return 1 */
1433 return ret < 0 ? ret : 1;
1437 * First waiter. Set the waiters bit before attaching ourself to
1438 * the owner. If owner tries to unlock, it will be forced into
1439 * the kernel and blocked on hb->lock.
1441 newval = uval | FUTEX_WAITERS;
1442 ret = lock_pi_update_atomic(uaddr, uval, newval);
1446 * If the update of the user space value succeeded, we try to
1447 * attach to the owner. If that fails, no harm done, we only
1448 * set the FUTEX_WAITERS bit in the user space variable.
1450 return attach_to_pi_owner(uaddr, newval, key, ps);
1454 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1455 * @q: The futex_q to unqueue
1457 * The q->lock_ptr must not be NULL and must be held by the caller.
1459 static void __unqueue_futex(struct futex_q *q)
1461 struct futex_hash_bucket *hb;
1463 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1465 lockdep_assert_held(q->lock_ptr);
1467 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1468 plist_del(&q->list, &hb->chain);
1473 * The hash bucket lock must be held when this is called.
1474 * Afterwards, the futex_q must not be accessed. Callers
1475 * must ensure to later call wake_up_q() for the actual
1478 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1480 struct task_struct *p = q->task;
1482 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1488 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1489 * is written, without taking any locks. This is possible in the event
1490 * of a spurious wakeup, for example. A memory barrier is required here
1491 * to prevent the following store to lock_ptr from getting ahead of the
1492 * plist_del in __unqueue_futex().
1494 smp_store_release(&q->lock_ptr, NULL);
1497 * Queue the task for later wakeup for after we've released
1498 * the hb->lock. wake_q_add() grabs reference to p.
1500 wake_q_add_safe(wake_q, p);
1504 * Caller must hold a reference on @pi_state.
1506 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1508 u32 uninitialized_var(curval), newval;
1509 struct task_struct *new_owner;
1510 bool postunlock = false;
1511 DEFINE_WAKE_Q(wake_q);
1514 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1515 if (WARN_ON_ONCE(!new_owner)) {
1517 * As per the comment in futex_unlock_pi() this should not happen.
1519 * When this happens, give up our locks and try again, giving
1520 * the futex_lock_pi() instance time to complete, either by
1521 * waiting on the rtmutex or removing itself from the futex
1529 * We pass it to the next owner. The WAITERS bit is always kept
1530 * enabled while there is PI state around. We cleanup the owner
1531 * died bit, because we are the owner.
1533 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1535 if (unlikely(should_fail_futex(true)))
1538 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1539 if (!ret && (curval != uval)) {
1541 * If a unconditional UNLOCK_PI operation (user space did not
1542 * try the TID->0 transition) raced with a waiter setting the
1543 * FUTEX_WAITERS flag between get_user() and locking the hash
1544 * bucket lock, retry the operation.
1546 if ((FUTEX_TID_MASK & curval) == uval)
1556 * This is a point of no return; once we modify the uval there is no
1557 * going back and subsequent operations must not fail.
1560 raw_spin_lock(&pi_state->owner->pi_lock);
1561 WARN_ON(list_empty(&pi_state->list));
1562 list_del_init(&pi_state->list);
1563 raw_spin_unlock(&pi_state->owner->pi_lock);
1565 raw_spin_lock(&new_owner->pi_lock);
1566 WARN_ON(!list_empty(&pi_state->list));
1567 list_add(&pi_state->list, &new_owner->pi_state_list);
1568 pi_state->owner = new_owner;
1569 raw_spin_unlock(&new_owner->pi_lock);
1571 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1574 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1577 rt_mutex_postunlock(&wake_q);
1583 * Express the locking dependencies for lockdep:
1586 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1589 spin_lock(&hb1->lock);
1591 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1592 } else { /* hb1 > hb2 */
1593 spin_lock(&hb2->lock);
1594 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1599 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1601 spin_unlock(&hb1->lock);
1603 spin_unlock(&hb2->lock);
1607 * Wake up waiters matching bitset queued on this futex (uaddr).
1610 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1612 struct futex_hash_bucket *hb;
1613 struct futex_q *this, *next;
1614 union futex_key key = FUTEX_KEY_INIT;
1616 DEFINE_WAKE_Q(wake_q);
1621 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1622 if (unlikely(ret != 0))
1625 hb = hash_futex(&key);
1627 /* Make sure we really have tasks to wakeup */
1628 if (!hb_waiters_pending(hb))
1631 spin_lock(&hb->lock);
1633 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1634 if (match_futex (&this->key, &key)) {
1635 if (this->pi_state || this->rt_waiter) {
1640 /* Check if one of the bits is set in both bitsets */
1641 if (!(this->bitset & bitset))
1644 mark_wake_futex(&wake_q, this);
1645 if (++ret >= nr_wake)
1650 spin_unlock(&hb->lock);
1653 put_futex_key(&key);
1658 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1660 unsigned int op = (encoded_op & 0x70000000) >> 28;
1661 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1662 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1663 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1666 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1667 if (oparg < 0 || oparg > 31) {
1668 char comm[sizeof(current->comm)];
1670 * kill this print and return -EINVAL when userspace
1673 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1674 get_task_comm(comm, current), oparg);
1680 if (!access_ok(uaddr, sizeof(u32)))
1683 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1688 case FUTEX_OP_CMP_EQ:
1689 return oldval == cmparg;
1690 case FUTEX_OP_CMP_NE:
1691 return oldval != cmparg;
1692 case FUTEX_OP_CMP_LT:
1693 return oldval < cmparg;
1694 case FUTEX_OP_CMP_GE:
1695 return oldval >= cmparg;
1696 case FUTEX_OP_CMP_LE:
1697 return oldval <= cmparg;
1698 case FUTEX_OP_CMP_GT:
1699 return oldval > cmparg;
1706 * Wake up all waiters hashed on the physical page that is mapped
1707 * to this virtual address:
1710 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1711 int nr_wake, int nr_wake2, int op)
1713 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1714 struct futex_hash_bucket *hb1, *hb2;
1715 struct futex_q *this, *next;
1717 DEFINE_WAKE_Q(wake_q);
1720 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1721 if (unlikely(ret != 0))
1723 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1724 if (unlikely(ret != 0))
1727 hb1 = hash_futex(&key1);
1728 hb2 = hash_futex(&key2);
1731 double_lock_hb(hb1, hb2);
1732 op_ret = futex_atomic_op_inuser(op, uaddr2);
1733 if (unlikely(op_ret < 0)) {
1734 double_unlock_hb(hb1, hb2);
1736 if (!IS_ENABLED(CONFIG_MMU) ||
1737 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1739 * we don't get EFAULT from MMU faults if we don't have
1740 * an MMU, but we might get them from range checking
1746 if (op_ret == -EFAULT) {
1747 ret = fault_in_user_writeable(uaddr2);
1752 if (!(flags & FLAGS_SHARED)) {
1757 put_futex_key(&key2);
1758 put_futex_key(&key1);
1763 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1764 if (match_futex (&this->key, &key1)) {
1765 if (this->pi_state || this->rt_waiter) {
1769 mark_wake_futex(&wake_q, this);
1770 if (++ret >= nr_wake)
1777 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1778 if (match_futex (&this->key, &key2)) {
1779 if (this->pi_state || this->rt_waiter) {
1783 mark_wake_futex(&wake_q, this);
1784 if (++op_ret >= nr_wake2)
1792 double_unlock_hb(hb1, hb2);
1795 put_futex_key(&key2);
1797 put_futex_key(&key1);
1803 * requeue_futex() - Requeue a futex_q from one hb to another
1804 * @q: the futex_q to requeue
1805 * @hb1: the source hash_bucket
1806 * @hb2: the target hash_bucket
1807 * @key2: the new key for the requeued futex_q
1810 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1811 struct futex_hash_bucket *hb2, union futex_key *key2)
1815 * If key1 and key2 hash to the same bucket, no need to
1818 if (likely(&hb1->chain != &hb2->chain)) {
1819 plist_del(&q->list, &hb1->chain);
1820 hb_waiters_dec(hb1);
1821 hb_waiters_inc(hb2);
1822 plist_add(&q->list, &hb2->chain);
1823 q->lock_ptr = &hb2->lock;
1825 get_futex_key_refs(key2);
1830 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1832 * @key: the key of the requeue target futex
1833 * @hb: the hash_bucket of the requeue target futex
1835 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1836 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1837 * to the requeue target futex so the waiter can detect the wakeup on the right
1838 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1839 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1840 * to protect access to the pi_state to fixup the owner later. Must be called
1841 * with both q->lock_ptr and hb->lock held.
1844 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1845 struct futex_hash_bucket *hb)
1847 get_futex_key_refs(key);
1852 WARN_ON(!q->rt_waiter);
1853 q->rt_waiter = NULL;
1855 q->lock_ptr = &hb->lock;
1857 wake_up_state(q->task, TASK_NORMAL);
1861 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1862 * @pifutex: the user address of the to futex
1863 * @hb1: the from futex hash bucket, must be locked by the caller
1864 * @hb2: the to futex hash bucket, must be locked by the caller
1865 * @key1: the from futex key
1866 * @key2: the to futex key
1867 * @ps: address to store the pi_state pointer
1868 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1870 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1871 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1872 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1873 * hb1 and hb2 must be held by the caller.
1876 * - 0 - failed to acquire the lock atomically;
1877 * - >0 - acquired the lock, return value is vpid of the top_waiter
1880 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1881 struct futex_hash_bucket *hb1,
1882 struct futex_hash_bucket *hb2,
1883 union futex_key *key1, union futex_key *key2,
1884 struct futex_pi_state **ps, int set_waiters)
1886 struct futex_q *top_waiter = NULL;
1890 if (get_futex_value_locked(&curval, pifutex))
1893 if (unlikely(should_fail_futex(true)))
1897 * Find the top_waiter and determine if there are additional waiters.
1898 * If the caller intends to requeue more than 1 waiter to pifutex,
1899 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1900 * as we have means to handle the possible fault. If not, don't set
1901 * the bit unecessarily as it will force the subsequent unlock to enter
1904 top_waiter = futex_top_waiter(hb1, key1);
1906 /* There are no waiters, nothing for us to do. */
1910 /* Ensure we requeue to the expected futex. */
1911 if (!match_futex(top_waiter->requeue_pi_key, key2))
1915 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1916 * the contended case or if set_waiters is 1. The pi_state is returned
1917 * in ps in contended cases.
1919 vpid = task_pid_vnr(top_waiter->task);
1920 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1923 requeue_pi_wake_futex(top_waiter, key2, hb2);
1930 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1931 * @uaddr1: source futex user address
1932 * @flags: futex flags (FLAGS_SHARED, etc.)
1933 * @uaddr2: target futex user address
1934 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1935 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1936 * @cmpval: @uaddr1 expected value (or %NULL)
1937 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1938 * pi futex (pi to pi requeue is not supported)
1940 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1941 * uaddr2 atomically on behalf of the top waiter.
1944 * - >=0 - on success, the number of tasks requeued or woken;
1947 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1948 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1949 u32 *cmpval, int requeue_pi)
1951 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1952 int drop_count = 0, task_count = 0, ret;
1953 struct futex_pi_state *pi_state = NULL;
1954 struct futex_hash_bucket *hb1, *hb2;
1955 struct futex_q *this, *next;
1956 DEFINE_WAKE_Q(wake_q);
1958 if (nr_wake < 0 || nr_requeue < 0)
1962 * When PI not supported: return -ENOSYS if requeue_pi is true,
1963 * consequently the compiler knows requeue_pi is always false past
1964 * this point which will optimize away all the conditional code
1967 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1972 * Requeue PI only works on two distinct uaddrs. This
1973 * check is only valid for private futexes. See below.
1975 if (uaddr1 == uaddr2)
1979 * requeue_pi requires a pi_state, try to allocate it now
1980 * without any locks in case it fails.
1982 if (refill_pi_state_cache())
1985 * requeue_pi must wake as many tasks as it can, up to nr_wake
1986 * + nr_requeue, since it acquires the rt_mutex prior to
1987 * returning to userspace, so as to not leave the rt_mutex with
1988 * waiters and no owner. However, second and third wake-ups
1989 * cannot be predicted as they involve race conditions with the
1990 * first wake and a fault while looking up the pi_state. Both
1991 * pthread_cond_signal() and pthread_cond_broadcast() should
1999 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
2000 if (unlikely(ret != 0))
2002 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2003 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
2004 if (unlikely(ret != 0))
2008 * The check above which compares uaddrs is not sufficient for
2009 * shared futexes. We need to compare the keys:
2011 if (requeue_pi && match_futex(&key1, &key2)) {
2016 hb1 = hash_futex(&key1);
2017 hb2 = hash_futex(&key2);
2020 hb_waiters_inc(hb2);
2021 double_lock_hb(hb1, hb2);
2023 if (likely(cmpval != NULL)) {
2026 ret = get_futex_value_locked(&curval, uaddr1);
2028 if (unlikely(ret)) {
2029 double_unlock_hb(hb1, hb2);
2030 hb_waiters_dec(hb2);
2032 ret = get_user(curval, uaddr1);
2036 if (!(flags & FLAGS_SHARED))
2039 put_futex_key(&key2);
2040 put_futex_key(&key1);
2043 if (curval != *cmpval) {
2049 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2051 * Attempt to acquire uaddr2 and wake the top waiter. If we
2052 * intend to requeue waiters, force setting the FUTEX_WAITERS
2053 * bit. We force this here where we are able to easily handle
2054 * faults rather in the requeue loop below.
2056 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2057 &key2, &pi_state, nr_requeue);
2060 * At this point the top_waiter has either taken uaddr2 or is
2061 * waiting on it. If the former, then the pi_state will not
2062 * exist yet, look it up one more time to ensure we have a
2063 * reference to it. If the lock was taken, ret contains the
2064 * vpid of the top waiter task.
2065 * If the lock was not taken, we have pi_state and an initial
2066 * refcount on it. In case of an error we have nothing.
2073 * If we acquired the lock, then the user space value
2074 * of uaddr2 should be vpid. It cannot be changed by
2075 * the top waiter as it is blocked on hb2 lock if it
2076 * tries to do so. If something fiddled with it behind
2077 * our back the pi state lookup might unearth it. So
2078 * we rather use the known value than rereading and
2079 * handing potential crap to lookup_pi_state.
2081 * If that call succeeds then we have pi_state and an
2082 * initial refcount on it.
2084 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2089 /* We hold a reference on the pi state. */
2092 /* If the above failed, then pi_state is NULL */
2094 double_unlock_hb(hb1, hb2);
2095 hb_waiters_dec(hb2);
2096 put_futex_key(&key2);
2097 put_futex_key(&key1);
2098 ret = fault_in_user_writeable(uaddr2);
2104 * Two reasons for this:
2105 * - Owner is exiting and we just wait for the
2107 * - The user space value changed.
2109 double_unlock_hb(hb1, hb2);
2110 hb_waiters_dec(hb2);
2111 put_futex_key(&key2);
2112 put_futex_key(&key1);
2120 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2121 if (task_count - nr_wake >= nr_requeue)
2124 if (!match_futex(&this->key, &key1))
2128 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2129 * be paired with each other and no other futex ops.
2131 * We should never be requeueing a futex_q with a pi_state,
2132 * which is awaiting a futex_unlock_pi().
2134 if ((requeue_pi && !this->rt_waiter) ||
2135 (!requeue_pi && this->rt_waiter) ||
2142 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2143 * lock, we already woke the top_waiter. If not, it will be
2144 * woken by futex_unlock_pi().
2146 if (++task_count <= nr_wake && !requeue_pi) {
2147 mark_wake_futex(&wake_q, this);
2151 /* Ensure we requeue to the expected futex for requeue_pi. */
2152 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2158 * Requeue nr_requeue waiters and possibly one more in the case
2159 * of requeue_pi if we couldn't acquire the lock atomically.
2163 * Prepare the waiter to take the rt_mutex. Take a
2164 * refcount on the pi_state and store the pointer in
2165 * the futex_q object of the waiter.
2167 get_pi_state(pi_state);
2168 this->pi_state = pi_state;
2169 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2174 * We got the lock. We do neither drop the
2175 * refcount on pi_state nor clear
2176 * this->pi_state because the waiter needs the
2177 * pi_state for cleaning up the user space
2178 * value. It will drop the refcount after
2181 requeue_pi_wake_futex(this, &key2, hb2);
2186 * rt_mutex_start_proxy_lock() detected a
2187 * potential deadlock when we tried to queue
2188 * that waiter. Drop the pi_state reference
2189 * which we took above and remove the pointer
2190 * to the state from the waiters futex_q
2193 this->pi_state = NULL;
2194 put_pi_state(pi_state);
2196 * We stop queueing more waiters and let user
2197 * space deal with the mess.
2202 requeue_futex(this, hb1, hb2, &key2);
2207 * We took an extra initial reference to the pi_state either
2208 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2209 * need to drop it here again.
2211 put_pi_state(pi_state);
2214 double_unlock_hb(hb1, hb2);
2216 hb_waiters_dec(hb2);
2219 * drop_futex_key_refs() must be called outside the spinlocks. During
2220 * the requeue we moved futex_q's from the hash bucket at key1 to the
2221 * one at key2 and updated their key pointer. We no longer need to
2222 * hold the references to key1.
2224 while (--drop_count >= 0)
2225 drop_futex_key_refs(&key1);
2228 put_futex_key(&key2);
2230 put_futex_key(&key1);
2232 return ret ? ret : task_count;
2235 /* The key must be already stored in q->key. */
2236 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2237 __acquires(&hb->lock)
2239 struct futex_hash_bucket *hb;
2241 hb = hash_futex(&q->key);
2244 * Increment the counter before taking the lock so that
2245 * a potential waker won't miss a to-be-slept task that is
2246 * waiting for the spinlock. This is safe as all queue_lock()
2247 * users end up calling queue_me(). Similarly, for housekeeping,
2248 * decrement the counter at queue_unlock() when some error has
2249 * occurred and we don't end up adding the task to the list.
2251 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2253 q->lock_ptr = &hb->lock;
2255 spin_lock(&hb->lock);
2260 queue_unlock(struct futex_hash_bucket *hb)
2261 __releases(&hb->lock)
2263 spin_unlock(&hb->lock);
2267 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2272 * The priority used to register this element is
2273 * - either the real thread-priority for the real-time threads
2274 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2275 * - or MAX_RT_PRIO for non-RT threads.
2276 * Thus, all RT-threads are woken first in priority order, and
2277 * the others are woken last, in FIFO order.
2279 prio = min(current->normal_prio, MAX_RT_PRIO);
2281 plist_node_init(&q->list, prio);
2282 plist_add(&q->list, &hb->chain);
2287 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2288 * @q: The futex_q to enqueue
2289 * @hb: The destination hash bucket
2291 * The hb->lock must be held by the caller, and is released here. A call to
2292 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2293 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2294 * or nothing if the unqueue is done as part of the wake process and the unqueue
2295 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2298 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2299 __releases(&hb->lock)
2302 spin_unlock(&hb->lock);
2306 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2307 * @q: The futex_q to unqueue
2309 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2310 * be paired with exactly one earlier call to queue_me().
2313 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2314 * - 0 - if the futex_q was already removed by the waking thread
2316 static int unqueue_me(struct futex_q *q)
2318 spinlock_t *lock_ptr;
2321 /* In the common case we don't take the spinlock, which is nice. */
2324 * q->lock_ptr can change between this read and the following spin_lock.
2325 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2326 * optimizing lock_ptr out of the logic below.
2328 lock_ptr = READ_ONCE(q->lock_ptr);
2329 if (lock_ptr != NULL) {
2330 spin_lock(lock_ptr);
2332 * q->lock_ptr can change between reading it and
2333 * spin_lock(), causing us to take the wrong lock. This
2334 * corrects the race condition.
2336 * Reasoning goes like this: if we have the wrong lock,
2337 * q->lock_ptr must have changed (maybe several times)
2338 * between reading it and the spin_lock(). It can
2339 * change again after the spin_lock() but only if it was
2340 * already changed before the spin_lock(). It cannot,
2341 * however, change back to the original value. Therefore
2342 * we can detect whether we acquired the correct lock.
2344 if (unlikely(lock_ptr != q->lock_ptr)) {
2345 spin_unlock(lock_ptr);
2350 BUG_ON(q->pi_state);
2352 spin_unlock(lock_ptr);
2356 drop_futex_key_refs(&q->key);
2361 * PI futexes can not be requeued and must remove themself from the
2362 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2365 static void unqueue_me_pi(struct futex_q *q)
2366 __releases(q->lock_ptr)
2370 BUG_ON(!q->pi_state);
2371 put_pi_state(q->pi_state);
2374 spin_unlock(q->lock_ptr);
2377 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2378 struct task_struct *argowner)
2380 struct futex_pi_state *pi_state = q->pi_state;
2381 u32 uval, uninitialized_var(curval), newval;
2382 struct task_struct *oldowner, *newowner;
2386 lockdep_assert_held(q->lock_ptr);
2388 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2390 oldowner = pi_state->owner;
2393 * We are here because either:
2395 * - we stole the lock and pi_state->owner needs updating to reflect
2396 * that (@argowner == current),
2400 * - someone stole our lock and we need to fix things to point to the
2401 * new owner (@argowner == NULL).
2403 * Either way, we have to replace the TID in the user space variable.
2404 * This must be atomic as we have to preserve the owner died bit here.
2406 * Note: We write the user space value _before_ changing the pi_state
2407 * because we can fault here. Imagine swapped out pages or a fork
2408 * that marked all the anonymous memory readonly for cow.
2410 * Modifying pi_state _before_ the user space value would leave the
2411 * pi_state in an inconsistent state when we fault here, because we
2412 * need to drop the locks to handle the fault. This might be observed
2413 * in the PID check in lookup_pi_state.
2417 if (oldowner != current) {
2419 * We raced against a concurrent self; things are
2420 * already fixed up. Nothing to do.
2426 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2427 /* We got the lock after all, nothing to fix. */
2433 * Since we just failed the trylock; there must be an owner.
2435 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2438 WARN_ON_ONCE(argowner != current);
2439 if (oldowner == current) {
2441 * We raced against a concurrent self; things are
2442 * already fixed up. Nothing to do.
2447 newowner = argowner;
2450 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2452 if (!pi_state->owner)
2453 newtid |= FUTEX_OWNER_DIED;
2455 err = get_futex_value_locked(&uval, uaddr);
2460 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2462 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2472 * We fixed up user space. Now we need to fix the pi_state
2475 if (pi_state->owner != NULL) {
2476 raw_spin_lock(&pi_state->owner->pi_lock);
2477 WARN_ON(list_empty(&pi_state->list));
2478 list_del_init(&pi_state->list);
2479 raw_spin_unlock(&pi_state->owner->pi_lock);
2482 pi_state->owner = newowner;
2484 raw_spin_lock(&newowner->pi_lock);
2485 WARN_ON(!list_empty(&pi_state->list));
2486 list_add(&pi_state->list, &newowner->pi_state_list);
2487 raw_spin_unlock(&newowner->pi_lock);
2488 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2493 * In order to reschedule or handle a page fault, we need to drop the
2494 * locks here. In the case of a fault, this gives the other task
2495 * (either the highest priority waiter itself or the task which stole
2496 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2497 * are back from handling the fault we need to check the pi_state after
2498 * reacquiring the locks and before trying to do another fixup. When
2499 * the fixup has been done already we simply return.
2501 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2502 * drop hb->lock since the caller owns the hb -> futex_q relation.
2503 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2506 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2507 spin_unlock(q->lock_ptr);
2511 ret = fault_in_user_writeable(uaddr);
2525 spin_lock(q->lock_ptr);
2526 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2529 * Check if someone else fixed it for us:
2531 if (pi_state->owner != oldowner) {
2542 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2546 static long futex_wait_restart(struct restart_block *restart);
2549 * fixup_owner() - Post lock pi_state and corner case management
2550 * @uaddr: user address of the futex
2551 * @q: futex_q (contains pi_state and access to the rt_mutex)
2552 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2554 * After attempting to lock an rt_mutex, this function is called to cleanup
2555 * the pi_state owner as well as handle race conditions that may allow us to
2556 * acquire the lock. Must be called with the hb lock held.
2559 * - 1 - success, lock taken;
2560 * - 0 - success, lock not taken;
2561 * - <0 - on error (-EFAULT)
2563 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2569 * Got the lock. We might not be the anticipated owner if we
2570 * did a lock-steal - fix up the PI-state in that case:
2572 * Speculative pi_state->owner read (we don't hold wait_lock);
2573 * since we own the lock pi_state->owner == current is the
2574 * stable state, anything else needs more attention.
2576 if (q->pi_state->owner != current)
2577 ret = fixup_pi_state_owner(uaddr, q, current);
2582 * If we didn't get the lock; check if anybody stole it from us. In
2583 * that case, we need to fix up the uval to point to them instead of
2584 * us, otherwise bad things happen. [10]
2586 * Another speculative read; pi_state->owner == current is unstable
2587 * but needs our attention.
2589 if (q->pi_state->owner == current) {
2590 ret = fixup_pi_state_owner(uaddr, q, NULL);
2595 * Paranoia check. If we did not take the lock, then we should not be
2596 * the owner of the rt_mutex.
2598 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2599 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2600 "pi-state %p\n", ret,
2601 q->pi_state->pi_mutex.owner,
2602 q->pi_state->owner);
2606 return ret ? ret : locked;
2610 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2611 * @hb: the futex hash bucket, must be locked by the caller
2612 * @q: the futex_q to queue up on
2613 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2615 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2616 struct hrtimer_sleeper *timeout)
2619 * The task state is guaranteed to be set before another task can
2620 * wake it. set_current_state() is implemented using smp_store_mb() and
2621 * queue_me() calls spin_unlock() upon completion, both serializing
2622 * access to the hash list and forcing another memory barrier.
2624 set_current_state(TASK_INTERRUPTIBLE);
2629 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2632 * If we have been removed from the hash list, then another task
2633 * has tried to wake us, and we can skip the call to schedule().
2635 if (likely(!plist_node_empty(&q->list))) {
2637 * If the timer has already expired, current will already be
2638 * flagged for rescheduling. Only call schedule if there
2639 * is no timeout, or if it has yet to expire.
2641 if (!timeout || timeout->task)
2642 freezable_schedule();
2644 __set_current_state(TASK_RUNNING);
2648 * futex_wait_setup() - Prepare to wait on a futex
2649 * @uaddr: the futex userspace address
2650 * @val: the expected value
2651 * @flags: futex flags (FLAGS_SHARED, etc.)
2652 * @q: the associated futex_q
2653 * @hb: storage for hash_bucket pointer to be returned to caller
2655 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2656 * compare it with the expected value. Handle atomic faults internally.
2657 * Return with the hb lock held and a q.key reference on success, and unlocked
2658 * with no q.key reference on failure.
2661 * - 0 - uaddr contains val and hb has been locked;
2662 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2664 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2665 struct futex_q *q, struct futex_hash_bucket **hb)
2671 * Access the page AFTER the hash-bucket is locked.
2672 * Order is important:
2674 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2675 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2677 * The basic logical guarantee of a futex is that it blocks ONLY
2678 * if cond(var) is known to be true at the time of blocking, for
2679 * any cond. If we locked the hash-bucket after testing *uaddr, that
2680 * would open a race condition where we could block indefinitely with
2681 * cond(var) false, which would violate the guarantee.
2683 * On the other hand, we insert q and release the hash-bucket only
2684 * after testing *uaddr. This guarantees that futex_wait() will NOT
2685 * absorb a wakeup if *uaddr does not match the desired values
2686 * while the syscall executes.
2689 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2690 if (unlikely(ret != 0))
2694 *hb = queue_lock(q);
2696 ret = get_futex_value_locked(&uval, uaddr);
2701 ret = get_user(uval, uaddr);
2705 if (!(flags & FLAGS_SHARED))
2708 put_futex_key(&q->key);
2719 put_futex_key(&q->key);
2723 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2724 ktime_t *abs_time, u32 bitset)
2726 struct hrtimer_sleeper timeout, *to;
2727 struct restart_block *restart;
2728 struct futex_hash_bucket *hb;
2729 struct futex_q q = futex_q_init;
2736 to = futex_setup_timer(abs_time, &timeout, flags,
2737 current->timer_slack_ns);
2740 * Prepare to wait on uaddr. On success, holds hb lock and increments
2743 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2747 /* queue_me and wait for wakeup, timeout, or a signal. */
2748 futex_wait_queue_me(hb, &q, to);
2750 /* If we were woken (and unqueued), we succeeded, whatever. */
2752 /* unqueue_me() drops q.key ref */
2753 if (!unqueue_me(&q))
2756 if (to && !to->task)
2760 * We expect signal_pending(current), but we might be the
2761 * victim of a spurious wakeup as well.
2763 if (!signal_pending(current))
2770 restart = ¤t->restart_block;
2771 restart->fn = futex_wait_restart;
2772 restart->futex.uaddr = uaddr;
2773 restart->futex.val = val;
2774 restart->futex.time = *abs_time;
2775 restart->futex.bitset = bitset;
2776 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2778 ret = -ERESTART_RESTARTBLOCK;
2782 hrtimer_cancel(&to->timer);
2783 destroy_hrtimer_on_stack(&to->timer);
2789 static long futex_wait_restart(struct restart_block *restart)
2791 u32 __user *uaddr = restart->futex.uaddr;
2792 ktime_t t, *tp = NULL;
2794 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2795 t = restart->futex.time;
2798 restart->fn = do_no_restart_syscall;
2800 return (long)futex_wait(uaddr, restart->futex.flags,
2801 restart->futex.val, tp, restart->futex.bitset);
2806 * Userspace tried a 0 -> TID atomic transition of the futex value
2807 * and failed. The kernel side here does the whole locking operation:
2808 * if there are waiters then it will block as a consequence of relying
2809 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2810 * a 0 value of the futex too.).
2812 * Also serves as futex trylock_pi()'ing, and due semantics.
2814 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2815 ktime_t *time, int trylock)
2817 struct hrtimer_sleeper timeout, *to;
2818 struct futex_pi_state *pi_state = NULL;
2819 struct rt_mutex_waiter rt_waiter;
2820 struct futex_hash_bucket *hb;
2821 struct futex_q q = futex_q_init;
2824 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2827 if (refill_pi_state_cache())
2830 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2833 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2834 if (unlikely(ret != 0))
2838 hb = queue_lock(&q);
2840 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2841 if (unlikely(ret)) {
2843 * Atomic work succeeded and we got the lock,
2844 * or failed. Either way, we do _not_ block.
2848 /* We got the lock. */
2850 goto out_unlock_put_key;
2855 * Two reasons for this:
2856 * - Task is exiting and we just wait for the
2858 * - The user space value changed.
2861 put_futex_key(&q.key);
2865 goto out_unlock_put_key;
2869 WARN_ON(!q.pi_state);
2872 * Only actually queue now that the atomic ops are done:
2877 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2878 /* Fixup the trylock return value: */
2879 ret = ret ? 0 : -EWOULDBLOCK;
2883 rt_mutex_init_waiter(&rt_waiter);
2886 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2887 * hold it while doing rt_mutex_start_proxy(), because then it will
2888 * include hb->lock in the blocking chain, even through we'll not in
2889 * fact hold it while blocking. This will lead it to report -EDEADLK
2890 * and BUG when futex_unlock_pi() interleaves with this.
2892 * Therefore acquire wait_lock while holding hb->lock, but drop the
2893 * latter before calling __rt_mutex_start_proxy_lock(). This
2894 * interleaves with futex_unlock_pi() -- which does a similar lock
2895 * handoff -- such that the latter can observe the futex_q::pi_state
2896 * before __rt_mutex_start_proxy_lock() is done.
2898 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2899 spin_unlock(q.lock_ptr);
2901 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2902 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2903 * it sees the futex_q::pi_state.
2905 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2906 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2915 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2917 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2920 spin_lock(q.lock_ptr);
2922 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2923 * first acquire the hb->lock before removing the lock from the
2924 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2927 * In particular; it is important that futex_unlock_pi() can not
2928 * observe this inconsistency.
2930 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2935 * Fixup the pi_state owner and possibly acquire the lock if we
2938 res = fixup_owner(uaddr, &q, !ret);
2940 * If fixup_owner() returned an error, proprogate that. If it acquired
2941 * the lock, clear our -ETIMEDOUT or -EINTR.
2944 ret = (res < 0) ? res : 0;
2947 * If fixup_owner() faulted and was unable to handle the fault, unlock
2948 * it and return the fault to userspace.
2950 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2951 pi_state = q.pi_state;
2952 get_pi_state(pi_state);
2955 /* Unqueue and drop the lock */
2959 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2960 put_pi_state(pi_state);
2969 put_futex_key(&q.key);
2972 hrtimer_cancel(&to->timer);
2973 destroy_hrtimer_on_stack(&to->timer);
2975 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2980 ret = fault_in_user_writeable(uaddr);
2984 if (!(flags & FLAGS_SHARED))
2987 put_futex_key(&q.key);
2992 * Userspace attempted a TID -> 0 atomic transition, and failed.
2993 * This is the in-kernel slowpath: we look up the PI state (if any),
2994 * and do the rt-mutex unlock.
2996 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2998 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2999 union futex_key key = FUTEX_KEY_INIT;
3000 struct futex_hash_bucket *hb;
3001 struct futex_q *top_waiter;
3004 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3008 if (get_user(uval, uaddr))
3011 * We release only a lock we actually own:
3013 if ((uval & FUTEX_TID_MASK) != vpid)
3016 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3020 hb = hash_futex(&key);
3021 spin_lock(&hb->lock);
3024 * Check waiters first. We do not trust user space values at
3025 * all and we at least want to know if user space fiddled
3026 * with the futex value instead of blindly unlocking.
3028 top_waiter = futex_top_waiter(hb, &key);
3030 struct futex_pi_state *pi_state = top_waiter->pi_state;
3037 * If current does not own the pi_state then the futex is
3038 * inconsistent and user space fiddled with the futex value.
3040 if (pi_state->owner != current)
3043 get_pi_state(pi_state);
3045 * By taking wait_lock while still holding hb->lock, we ensure
3046 * there is no point where we hold neither; and therefore
3047 * wake_futex_pi() must observe a state consistent with what we
3050 * In particular; this forces __rt_mutex_start_proxy() to
3051 * complete such that we're guaranteed to observe the
3052 * rt_waiter. Also see the WARN in wake_futex_pi().
3054 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3055 spin_unlock(&hb->lock);
3057 /* drops pi_state->pi_mutex.wait_lock */
3058 ret = wake_futex_pi(uaddr, uval, pi_state);
3060 put_pi_state(pi_state);
3063 * Success, we're done! No tricky corner cases.
3068 * The atomic access to the futex value generated a
3069 * pagefault, so retry the user-access and the wakeup:
3074 * A unconditional UNLOCK_PI op raced against a waiter
3075 * setting the FUTEX_WAITERS bit. Try again.
3080 * wake_futex_pi has detected invalid state. Tell user
3087 * We have no kernel internal state, i.e. no waiters in the
3088 * kernel. Waiters which are about to queue themselves are stuck
3089 * on hb->lock. So we can safely ignore them. We do neither
3090 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3093 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3094 spin_unlock(&hb->lock);
3109 * If uval has changed, let user space handle it.
3111 ret = (curval == uval) ? 0 : -EAGAIN;
3114 spin_unlock(&hb->lock);
3116 put_futex_key(&key);
3120 put_futex_key(&key);
3125 put_futex_key(&key);
3127 ret = fault_in_user_writeable(uaddr);
3135 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3136 * @hb: the hash_bucket futex_q was original enqueued on
3137 * @q: the futex_q woken while waiting to be requeued
3138 * @key2: the futex_key of the requeue target futex
3139 * @timeout: the timeout associated with the wait (NULL if none)
3141 * Detect if the task was woken on the initial futex as opposed to the requeue
3142 * target futex. If so, determine if it was a timeout or a signal that caused
3143 * the wakeup and return the appropriate error code to the caller. Must be
3144 * called with the hb lock held.
3147 * - 0 = no early wakeup detected;
3148 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3151 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3152 struct futex_q *q, union futex_key *key2,
3153 struct hrtimer_sleeper *timeout)
3158 * With the hb lock held, we avoid races while we process the wakeup.
3159 * We only need to hold hb (and not hb2) to ensure atomicity as the
3160 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3161 * It can't be requeued from uaddr2 to something else since we don't
3162 * support a PI aware source futex for requeue.
3164 if (!match_futex(&q->key, key2)) {
3165 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3167 * We were woken prior to requeue by a timeout or a signal.
3168 * Unqueue the futex_q and determine which it was.
3170 plist_del(&q->list, &hb->chain);
3173 /* Handle spurious wakeups gracefully */
3175 if (timeout && !timeout->task)
3177 else if (signal_pending(current))
3178 ret = -ERESTARTNOINTR;
3184 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3185 * @uaddr: the futex we initially wait on (non-pi)
3186 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3187 * the same type, no requeueing from private to shared, etc.
3188 * @val: the expected value of uaddr
3189 * @abs_time: absolute timeout
3190 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3191 * @uaddr2: the pi futex we will take prior to returning to user-space
3193 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3194 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3195 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3196 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3197 * without one, the pi logic would not know which task to boost/deboost, if
3198 * there was a need to.
3200 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3201 * via the following--
3202 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3203 * 2) wakeup on uaddr2 after a requeue
3207 * If 3, cleanup and return -ERESTARTNOINTR.
3209 * If 2, we may then block on trying to take the rt_mutex and return via:
3210 * 5) successful lock
3213 * 8) other lock acquisition failure
3215 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3217 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3223 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3224 u32 val, ktime_t *abs_time, u32 bitset,
3227 struct hrtimer_sleeper timeout, *to;
3228 struct futex_pi_state *pi_state = NULL;
3229 struct rt_mutex_waiter rt_waiter;
3230 struct futex_hash_bucket *hb;
3231 union futex_key key2 = FUTEX_KEY_INIT;
3232 struct futex_q q = futex_q_init;
3235 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3238 if (uaddr == uaddr2)
3244 to = futex_setup_timer(abs_time, &timeout, flags,
3245 current->timer_slack_ns);
3248 * The waiter is allocated on our stack, manipulated by the requeue
3249 * code while we sleep on uaddr.
3251 rt_mutex_init_waiter(&rt_waiter);
3253 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3254 if (unlikely(ret != 0))
3258 q.rt_waiter = &rt_waiter;
3259 q.requeue_pi_key = &key2;
3262 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3265 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3270 * The check above which compares uaddrs is not sufficient for
3271 * shared futexes. We need to compare the keys:
3273 if (match_futex(&q.key, &key2)) {
3279 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3280 futex_wait_queue_me(hb, &q, to);
3282 spin_lock(&hb->lock);
3283 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3284 spin_unlock(&hb->lock);
3289 * In order for us to be here, we know our q.key == key2, and since
3290 * we took the hb->lock above, we also know that futex_requeue() has
3291 * completed and we no longer have to concern ourselves with a wakeup
3292 * race with the atomic proxy lock acquisition by the requeue code. The
3293 * futex_requeue dropped our key1 reference and incremented our key2
3297 /* Check if the requeue code acquired the second futex for us. */
3300 * Got the lock. We might not be the anticipated owner if we
3301 * did a lock-steal - fix up the PI-state in that case.
3303 if (q.pi_state && (q.pi_state->owner != current)) {
3304 spin_lock(q.lock_ptr);
3305 ret = fixup_pi_state_owner(uaddr2, &q, current);
3306 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3307 pi_state = q.pi_state;
3308 get_pi_state(pi_state);
3311 * Drop the reference to the pi state which
3312 * the requeue_pi() code acquired for us.
3314 put_pi_state(q.pi_state);
3315 spin_unlock(q.lock_ptr);
3318 struct rt_mutex *pi_mutex;
3321 * We have been woken up by futex_unlock_pi(), a timeout, or a
3322 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3325 WARN_ON(!q.pi_state);
3326 pi_mutex = &q.pi_state->pi_mutex;
3327 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3329 spin_lock(q.lock_ptr);
3330 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3333 debug_rt_mutex_free_waiter(&rt_waiter);
3335 * Fixup the pi_state owner and possibly acquire the lock if we
3338 res = fixup_owner(uaddr2, &q, !ret);
3340 * If fixup_owner() returned an error, proprogate that. If it
3341 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3344 ret = (res < 0) ? res : 0;
3347 * If fixup_pi_state_owner() faulted and was unable to handle
3348 * the fault, unlock the rt_mutex and return the fault to
3351 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3352 pi_state = q.pi_state;
3353 get_pi_state(pi_state);
3356 /* Unqueue and drop the lock. */
3361 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3362 put_pi_state(pi_state);
3365 if (ret == -EINTR) {
3367 * We've already been requeued, but cannot restart by calling
3368 * futex_lock_pi() directly. We could restart this syscall, but
3369 * it would detect that the user space "val" changed and return
3370 * -EWOULDBLOCK. Save the overhead of the restart and return
3371 * -EWOULDBLOCK directly.
3377 put_futex_key(&q.key);
3379 put_futex_key(&key2);
3383 hrtimer_cancel(&to->timer);
3384 destroy_hrtimer_on_stack(&to->timer);
3390 * Support for robust futexes: the kernel cleans up held futexes at
3393 * Implementation: user-space maintains a per-thread list of locks it
3394 * is holding. Upon do_exit(), the kernel carefully walks this list,
3395 * and marks all locks that are owned by this thread with the
3396 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3397 * always manipulated with the lock held, so the list is private and
3398 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3399 * field, to allow the kernel to clean up if the thread dies after
3400 * acquiring the lock, but just before it could have added itself to
3401 * the list. There can only be one such pending lock.
3405 * sys_set_robust_list() - Set the robust-futex list head of a task
3406 * @head: pointer to the list-head
3407 * @len: length of the list-head, as userspace expects
3409 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3412 if (!futex_cmpxchg_enabled)
3415 * The kernel knows only one size for now:
3417 if (unlikely(len != sizeof(*head)))
3420 current->robust_list = head;
3426 * sys_get_robust_list() - Get the robust-futex list head of a task
3427 * @pid: pid of the process [zero for current task]
3428 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3429 * @len_ptr: pointer to a length field, the kernel fills in the header size
3431 SYSCALL_DEFINE3(get_robust_list, int, pid,
3432 struct robust_list_head __user * __user *, head_ptr,
3433 size_t __user *, len_ptr)
3435 struct robust_list_head __user *head;
3437 struct task_struct *p;
3439 if (!futex_cmpxchg_enabled)
3448 p = find_task_by_vpid(pid);
3454 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3457 head = p->robust_list;
3460 if (put_user(sizeof(*head), len_ptr))
3462 return put_user(head, head_ptr);
3471 * Process a futex-list entry, check whether it's owned by the
3472 * dying task, and do notification if so:
3474 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3476 u32 uval, uninitialized_var(nval), mval;
3479 /* Futex address must be 32bit aligned */
3480 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3484 if (get_user(uval, uaddr))
3487 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3491 * Ok, this dying thread is truly holding a futex
3492 * of interest. Set the OWNER_DIED bit atomically
3493 * via cmpxchg, and if the value had FUTEX_WAITERS
3494 * set, wake up a waiter (if any). (We have to do a
3495 * futex_wake() even if OWNER_DIED is already set -
3496 * to handle the rare but possible case of recursive
3497 * thread-death.) The rest of the cleanup is done in
3500 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3503 * We are not holding a lock here, but we want to have
3504 * the pagefault_disable/enable() protection because
3505 * we want to handle the fault gracefully. If the
3506 * access fails we try to fault in the futex with R/W
3507 * verification via get_user_pages. get_user() above
3508 * does not guarantee R/W access. If that fails we
3509 * give up and leave the futex locked.
3511 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3514 if (fault_in_user_writeable(uaddr))
3532 * Wake robust non-PI futexes here. The wakeup of
3533 * PI futexes happens in exit_pi_state():
3535 if (!pi && (uval & FUTEX_WAITERS))
3536 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3542 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3544 static inline int fetch_robust_entry(struct robust_list __user **entry,
3545 struct robust_list __user * __user *head,
3548 unsigned long uentry;
3550 if (get_user(uentry, (unsigned long __user *)head))
3553 *entry = (void __user *)(uentry & ~1UL);
3560 * Walk curr->robust_list (very carefully, it's a userspace list!)
3561 * and mark any locks found there dead, and notify any waiters.
3563 * We silently return on any sign of list-walking problem.
3565 void exit_robust_list(struct task_struct *curr)
3567 struct robust_list_head __user *head = curr->robust_list;
3568 struct robust_list __user *entry, *next_entry, *pending;
3569 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3570 unsigned int uninitialized_var(next_pi);
3571 unsigned long futex_offset;
3574 if (!futex_cmpxchg_enabled)
3578 * Fetch the list head (which was registered earlier, via
3579 * sys_set_robust_list()):
3581 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3584 * Fetch the relative futex offset:
3586 if (get_user(futex_offset, &head->futex_offset))
3589 * Fetch any possibly pending lock-add first, and handle it
3592 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3595 next_entry = NULL; /* avoid warning with gcc */
3596 while (entry != &head->list) {
3598 * Fetch the next entry in the list before calling
3599 * handle_futex_death:
3601 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3603 * A pending lock might already be on the list, so
3604 * don't process it twice:
3606 if (entry != pending)
3607 if (handle_futex_death((void __user *)entry + futex_offset,
3615 * Avoid excessively long or circular lists:
3624 handle_futex_death((void __user *)pending + futex_offset,
3628 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3629 u32 __user *uaddr2, u32 val2, u32 val3)
3631 int cmd = op & FUTEX_CMD_MASK;
3632 unsigned int flags = 0;
3634 if (!(op & FUTEX_PRIVATE_FLAG))
3635 flags |= FLAGS_SHARED;
3637 if (op & FUTEX_CLOCK_REALTIME) {
3638 flags |= FLAGS_CLOCKRT;
3639 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3640 cmd != FUTEX_WAIT_REQUEUE_PI)
3646 case FUTEX_UNLOCK_PI:
3647 case FUTEX_TRYLOCK_PI:
3648 case FUTEX_WAIT_REQUEUE_PI:
3649 case FUTEX_CMP_REQUEUE_PI:
3650 if (!futex_cmpxchg_enabled)
3656 val3 = FUTEX_BITSET_MATCH_ANY;
3658 case FUTEX_WAIT_BITSET:
3659 return futex_wait(uaddr, flags, val, timeout, val3);
3661 val3 = FUTEX_BITSET_MATCH_ANY;
3663 case FUTEX_WAKE_BITSET:
3664 return futex_wake(uaddr, flags, val, val3);
3666 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3667 case FUTEX_CMP_REQUEUE:
3668 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3670 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3672 return futex_lock_pi(uaddr, flags, timeout, 0);
3673 case FUTEX_UNLOCK_PI:
3674 return futex_unlock_pi(uaddr, flags);
3675 case FUTEX_TRYLOCK_PI:
3676 return futex_lock_pi(uaddr, flags, NULL, 1);
3677 case FUTEX_WAIT_REQUEUE_PI:
3678 val3 = FUTEX_BITSET_MATCH_ANY;
3679 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3681 case FUTEX_CMP_REQUEUE_PI:
3682 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3688 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3689 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3692 struct timespec64 ts;
3693 ktime_t t, *tp = NULL;
3695 int cmd = op & FUTEX_CMD_MASK;
3697 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3698 cmd == FUTEX_WAIT_BITSET ||
3699 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3700 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3702 if (get_timespec64(&ts, utime))
3704 if (!timespec64_valid(&ts))
3707 t = timespec64_to_ktime(ts);
3708 if (cmd == FUTEX_WAIT)
3709 t = ktime_add_safe(ktime_get(), t);
3713 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3714 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3716 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3717 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3718 val2 = (u32) (unsigned long) utime;
3720 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3723 #ifdef CONFIG_COMPAT
3725 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3728 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3729 compat_uptr_t __user *head, unsigned int *pi)
3731 if (get_user(*uentry, head))
3734 *entry = compat_ptr((*uentry) & ~1);
3735 *pi = (unsigned int)(*uentry) & 1;
3740 static void __user *futex_uaddr(struct robust_list __user *entry,
3741 compat_long_t futex_offset)
3743 compat_uptr_t base = ptr_to_compat(entry);
3744 void __user *uaddr = compat_ptr(base + futex_offset);
3750 * Walk curr->robust_list (very carefully, it's a userspace list!)
3751 * and mark any locks found there dead, and notify any waiters.
3753 * We silently return on any sign of list-walking problem.
3755 void compat_exit_robust_list(struct task_struct *curr)
3757 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3758 struct robust_list __user *entry, *next_entry, *pending;
3759 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3760 unsigned int uninitialized_var(next_pi);
3761 compat_uptr_t uentry, next_uentry, upending;
3762 compat_long_t futex_offset;
3765 if (!futex_cmpxchg_enabled)
3769 * Fetch the list head (which was registered earlier, via
3770 * sys_set_robust_list()):
3772 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3775 * Fetch the relative futex offset:
3777 if (get_user(futex_offset, &head->futex_offset))
3780 * Fetch any possibly pending lock-add first, and handle it
3783 if (compat_fetch_robust_entry(&upending, &pending,
3784 &head->list_op_pending, &pip))
3787 next_entry = NULL; /* avoid warning with gcc */
3788 while (entry != (struct robust_list __user *) &head->list) {
3790 * Fetch the next entry in the list before calling
3791 * handle_futex_death:
3793 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3794 (compat_uptr_t __user *)&entry->next, &next_pi);
3796 * A pending lock might already be on the list, so
3797 * dont process it twice:
3799 if (entry != pending) {
3800 void __user *uaddr = futex_uaddr(entry, futex_offset);
3802 if (handle_futex_death(uaddr, curr, pi))
3807 uentry = next_uentry;
3811 * Avoid excessively long or circular lists:
3819 void __user *uaddr = futex_uaddr(pending, futex_offset);
3821 handle_futex_death(uaddr, curr, pip);
3825 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3826 struct compat_robust_list_head __user *, head,
3829 if (!futex_cmpxchg_enabled)
3832 if (unlikely(len != sizeof(*head)))
3835 current->compat_robust_list = head;
3840 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3841 compat_uptr_t __user *, head_ptr,
3842 compat_size_t __user *, len_ptr)
3844 struct compat_robust_list_head __user *head;
3846 struct task_struct *p;
3848 if (!futex_cmpxchg_enabled)
3857 p = find_task_by_vpid(pid);
3863 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3866 head = p->compat_robust_list;
3869 if (put_user(sizeof(*head), len_ptr))
3871 return put_user(ptr_to_compat(head), head_ptr);
3878 #endif /* CONFIG_COMPAT */
3880 #ifdef CONFIG_COMPAT_32BIT_TIME
3881 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3882 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3885 struct timespec64 ts;
3886 ktime_t t, *tp = NULL;
3888 int cmd = op & FUTEX_CMD_MASK;
3890 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3891 cmd == FUTEX_WAIT_BITSET ||
3892 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3893 if (get_old_timespec32(&ts, utime))
3895 if (!timespec64_valid(&ts))
3898 t = timespec64_to_ktime(ts);
3899 if (cmd == FUTEX_WAIT)
3900 t = ktime_add_safe(ktime_get(), t);
3903 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3904 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3905 val2 = (int) (unsigned long) utime;
3907 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3909 #endif /* CONFIG_COMPAT_32BIT_TIME */
3911 static void __init futex_detect_cmpxchg(void)
3913 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3917 * This will fail and we want it. Some arch implementations do
3918 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3919 * functionality. We want to know that before we call in any
3920 * of the complex code paths. Also we want to prevent
3921 * registration of robust lists in that case. NULL is
3922 * guaranteed to fault and we get -EFAULT on functional
3923 * implementation, the non-functional ones will return
3926 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3927 futex_cmpxchg_enabled = 1;
3931 static int __init futex_init(void)
3933 unsigned int futex_shift;
3936 #if CONFIG_BASE_SMALL
3937 futex_hashsize = 16;
3939 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3942 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3944 futex_hashsize < 256 ? HASH_SMALL : 0,
3946 futex_hashsize, futex_hashsize);
3947 futex_hashsize = 1UL << futex_shift;
3949 futex_detect_cmpxchg();
3951 for (i = 0; i < futex_hashsize; i++) {
3952 atomic_set(&futex_queues[i].waiters, 0);
3953 plist_head_init(&futex_queues[i].chain);
3954 spin_lock_init(&futex_queues[i].lock);
3959 core_initcall(futex_init);