1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/slab.h>
36 #include <linux/poll.h>
38 #include <linux/file.h>
39 #include <linux/jhash.h>
40 #include <linux/init.h>
41 #include <linux/futex.h>
42 #include <linux/mount.h>
43 #include <linux/pagemap.h>
44 #include <linux/syscalls.h>
45 #include <linux/signal.h>
46 #include <linux/export.h>
47 #include <linux/magic.h>
48 #include <linux/pid.h>
49 #include <linux/nsproxy.h>
50 #include <linux/ptrace.h>
51 #include <linux/sched/rt.h>
52 #include <linux/sched/wake_q.h>
53 #include <linux/sched/mm.h>
54 #include <linux/hugetlb.h>
55 #include <linux/freezer.h>
56 #include <linux/memblock.h>
57 #include <linux/fault-inject.h>
58 #include <linux/refcount.h>
60 #include <asm/futex.h>
62 #include "locking/rtmutex_common.h"
65 * READ this before attempting to hack on futexes!
67 * Basic futex operation and ordering guarantees
68 * =============================================
70 * The waiter reads the futex value in user space and calls
71 * futex_wait(). This function computes the hash bucket and acquires
72 * the hash bucket lock. After that it reads the futex user space value
73 * again and verifies that the data has not changed. If it has not changed
74 * it enqueues itself into the hash bucket, releases the hash bucket lock
77 * The waker side modifies the user space value of the futex and calls
78 * futex_wake(). This function computes the hash bucket and acquires the
79 * hash bucket lock. Then it looks for waiters on that futex in the hash
80 * bucket and wakes them.
82 * In futex wake up scenarios where no tasks are blocked on a futex, taking
83 * the hb spinlock can be avoided and simply return. In order for this
84 * optimization to work, ordering guarantees must exist so that the waiter
85 * being added to the list is acknowledged when the list is concurrently being
86 * checked by the waker, avoiding scenarios like the following:
90 * sys_futex(WAIT, futex, val);
91 * futex_wait(futex, val);
94 * sys_futex(WAKE, futex);
99 * lock(hash_bucket(futex));
101 * unlock(hash_bucket(futex));
104 * This would cause the waiter on CPU 0 to wait forever because it
105 * missed the transition of the user space value from val to newval
106 * and the waker did not find the waiter in the hash bucket queue.
108 * The correct serialization ensures that a waiter either observes
109 * the changed user space value before blocking or is woken by a
114 * sys_futex(WAIT, futex, val);
115 * futex_wait(futex, val);
118 * smp_mb(); (A) <-- paired with -.
120 * lock(hash_bucket(futex)); |
124 * | sys_futex(WAKE, futex);
125 * | futex_wake(futex);
127 * `--------> smp_mb(); (B)
130 * unlock(hash_bucket(futex));
131 * schedule(); if (waiters)
132 * lock(hash_bucket(futex));
133 * else wake_waiters(futex);
134 * waiters--; (b) unlock(hash_bucket(futex));
136 * Where (A) orders the waiters increment and the futex value read through
137 * atomic operations (see hb_waiters_inc) and where (B) orders the write
138 * to futex and the waiters read -- this is done by the barriers for both
139 * shared and private futexes in get_futex_key_refs().
141 * This yields the following case (where X:=waiters, Y:=futex):
149 * Which guarantees that x==0 && y==0 is impossible; which translates back into
150 * the guarantee that we cannot both miss the futex variable change and the
153 * Note that a new waiter is accounted for in (a) even when it is possible that
154 * the wait call can return error, in which case we backtrack from it in (b).
155 * Refer to the comment in queue_lock().
157 * Similarly, in order to account for waiters being requeued on another
158 * address we always increment the waiters for the destination bucket before
159 * acquiring the lock. It then decrements them again after releasing it -
160 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
161 * will do the additional required waiter count housekeeping. This is done for
162 * double_lock_hb() and double_unlock_hb(), respectively.
165 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
166 #define futex_cmpxchg_enabled 1
168 static int __read_mostly futex_cmpxchg_enabled;
172 * Futex flags used to encode options to functions and preserve them across
176 # define FLAGS_SHARED 0x01
179 * NOMMU does not have per process address space. Let the compiler optimize
182 # define FLAGS_SHARED 0x00
184 #define FLAGS_CLOCKRT 0x02
185 #define FLAGS_HAS_TIMEOUT 0x04
188 * Priority Inheritance state:
190 struct futex_pi_state {
192 * list of 'owned' pi_state instances - these have to be
193 * cleaned up in do_exit() if the task exits prematurely:
195 struct list_head list;
200 struct rt_mutex pi_mutex;
202 struct task_struct *owner;
206 } __randomize_layout;
209 * struct futex_q - The hashed futex queue entry, one per waiting task
210 * @list: priority-sorted list of tasks waiting on this futex
211 * @task: the task waiting on the futex
212 * @lock_ptr: the hash bucket lock
213 * @key: the key the futex is hashed on
214 * @pi_state: optional priority inheritance state
215 * @rt_waiter: rt_waiter storage for use with requeue_pi
216 * @requeue_pi_key: the requeue_pi target futex key
217 * @bitset: bitset for the optional bitmasked wakeup
219 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
220 * we can wake only the relevant ones (hashed queues may be shared).
222 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
223 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
224 * The order of wakeup is always to make the first condition true, then
227 * PI futexes are typically woken before they are removed from the hash list via
228 * the rt_mutex code. See unqueue_me_pi().
231 struct plist_node list;
233 struct task_struct *task;
234 spinlock_t *lock_ptr;
236 struct futex_pi_state *pi_state;
237 struct rt_mutex_waiter *rt_waiter;
238 union futex_key *requeue_pi_key;
240 } __randomize_layout;
242 static const struct futex_q futex_q_init = {
243 /* list gets initialized in queue_me()*/
244 .key = FUTEX_KEY_INIT,
245 .bitset = FUTEX_BITSET_MATCH_ANY
249 * Hash buckets are shared by all the futex_keys that hash to the same
250 * location. Each key may have multiple futex_q structures, one for each task
251 * waiting on a futex.
253 struct futex_hash_bucket {
256 struct plist_head chain;
257 } ____cacheline_aligned_in_smp;
260 * The base of the bucket array and its size are always used together
261 * (after initialization only in hash_futex()), so ensure that they
262 * reside in the same cacheline.
265 struct futex_hash_bucket *queues;
266 unsigned long hashsize;
267 } __futex_data __read_mostly __aligned(2*sizeof(long));
268 #define futex_queues (__futex_data.queues)
269 #define futex_hashsize (__futex_data.hashsize)
273 * Fault injections for futexes.
275 #ifdef CONFIG_FAIL_FUTEX
278 struct fault_attr attr;
282 .attr = FAULT_ATTR_INITIALIZER,
283 .ignore_private = false,
286 static int __init setup_fail_futex(char *str)
288 return setup_fault_attr(&fail_futex.attr, str);
290 __setup("fail_futex=", setup_fail_futex);
292 static bool should_fail_futex(bool fshared)
294 if (fail_futex.ignore_private && !fshared)
297 return should_fail(&fail_futex.attr, 1);
300 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
302 static int __init fail_futex_debugfs(void)
304 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
307 dir = fault_create_debugfs_attr("fail_futex", NULL,
312 debugfs_create_bool("ignore-private", mode, dir,
313 &fail_futex.ignore_private);
317 late_initcall(fail_futex_debugfs);
319 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
322 static inline bool should_fail_futex(bool fshared)
326 #endif /* CONFIG_FAIL_FUTEX */
329 static void compat_exit_robust_list(struct task_struct *curr);
331 static inline void compat_exit_robust_list(struct task_struct *curr) { }
334 static inline void futex_get_mm(union futex_key *key)
336 mmgrab(key->private.mm);
338 * Ensure futex_get_mm() implies a full barrier such that
339 * get_futex_key() implies a full barrier. This is relied upon
340 * as smp_mb(); (B), see the ordering comment above.
342 smp_mb__after_atomic();
346 * Reflects a new waiter being added to the waitqueue.
348 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
351 atomic_inc(&hb->waiters);
353 * Full barrier (A), see the ordering comment above.
355 smp_mb__after_atomic();
360 * Reflects a waiter being removed from the waitqueue by wakeup
363 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
366 atomic_dec(&hb->waiters);
370 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
373 return atomic_read(&hb->waiters);
380 * hash_futex - Return the hash bucket in the global hash
381 * @key: Pointer to the futex key for which the hash is calculated
383 * We hash on the keys returned from get_futex_key (see below) and return the
384 * corresponding hash bucket in the global hash.
386 static struct futex_hash_bucket *hash_futex(union futex_key *key)
388 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
391 return &futex_queues[hash & (futex_hashsize - 1)];
396 * match_futex - Check whether two futex keys are equal
397 * @key1: Pointer to key1
398 * @key2: Pointer to key2
400 * Return 1 if two futex_keys are equal, 0 otherwise.
402 static inline int match_futex(union futex_key *key1, union futex_key *key2)
405 && key1->both.word == key2->both.word
406 && key1->both.ptr == key2->both.ptr
407 && key1->both.offset == key2->both.offset);
411 * Take a reference to the resource addressed by a key.
412 * Can be called while holding spinlocks.
415 static void get_futex_key_refs(union futex_key *key)
421 * On MMU less systems futexes are always "private" as there is no per
422 * process address space. We need the smp wmb nevertheless - yes,
423 * arch/blackfin has MMU less SMP ...
425 if (!IS_ENABLED(CONFIG_MMU)) {
426 smp_mb(); /* explicit smp_mb(); (B) */
430 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
432 smp_mb(); /* explicit smp_mb(); (B) */
434 case FUT_OFF_MMSHARED:
435 futex_get_mm(key); /* implies smp_mb(); (B) */
439 * Private futexes do not hold reference on an inode or
440 * mm, therefore the only purpose of calling get_futex_key_refs
441 * is because we need the barrier for the lockless waiter check.
443 smp_mb(); /* explicit smp_mb(); (B) */
448 * Drop a reference to the resource addressed by a key.
449 * The hash bucket spinlock must not be held. This is
450 * a no-op for private futexes, see comment in the get
453 static void drop_futex_key_refs(union futex_key *key)
455 if (!key->both.ptr) {
456 /* If we're here then we tried to put a key we failed to get */
461 if (!IS_ENABLED(CONFIG_MMU))
464 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
467 case FUT_OFF_MMSHARED:
468 mmdrop(key->private.mm);
479 * futex_setup_timer - set up the sleeping hrtimer.
480 * @time: ptr to the given timeout value
481 * @timeout: the hrtimer_sleeper structure to be set up
482 * @flags: futex flags
483 * @range_ns: optional range in ns
485 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
488 static inline struct hrtimer_sleeper *
489 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
490 int flags, u64 range_ns)
495 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
496 CLOCK_REALTIME : CLOCK_MONOTONIC,
499 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
500 * effectively the same as calling hrtimer_set_expires().
502 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
508 * Generate a machine wide unique identifier for this inode.
510 * This relies on u64 not wrapping in the life-time of the machine; which with
511 * 1ns resolution means almost 585 years.
513 * This further relies on the fact that a well formed program will not unmap
514 * the file while it has a (shared) futex waiting on it. This mapping will have
515 * a file reference which pins the mount and inode.
517 * If for some reason an inode gets evicted and read back in again, it will get
518 * a new sequence number and will _NOT_ match, even though it is the exact same
521 * It is important that match_futex() will never have a false-positive, esp.
522 * for PI futexes that can mess up the state. The above argues that false-negatives
523 * are only possible for malformed programs.
525 static u64 get_inode_sequence_number(struct inode *inode)
527 static atomic64_t i_seq;
530 /* Does the inode already have a sequence number? */
531 old = atomic64_read(&inode->i_sequence);
536 u64 new = atomic64_add_return(1, &i_seq);
537 if (WARN_ON_ONCE(!new))
540 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
548 * get_futex_key() - Get parameters which are the keys for a futex
549 * @uaddr: virtual address of the futex
550 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
551 * @key: address where result is stored.
552 * @rw: mapping needs to be read/write (values: FUTEX_READ,
555 * Return: a negative error code or 0
557 * The key words are stored in @key on success.
559 * For shared mappings (when @fshared), the key is:
560 * ( inode->i_sequence, page->index, offset_within_page )
561 * [ also see get_inode_sequence_number() ]
563 * For private mappings (or when !@fshared), the key is:
564 * ( current->mm, address, 0 )
566 * This allows (cross process, where applicable) identification of the futex
567 * without keeping the page pinned for the duration of the FUTEX_WAIT.
569 * lock_page() might sleep, the caller should not hold a spinlock.
572 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
574 unsigned long address = (unsigned long)uaddr;
575 struct mm_struct *mm = current->mm;
576 struct page *page, *tail;
577 struct address_space *mapping;
581 * The futex address must be "naturally" aligned.
583 key->both.offset = address % PAGE_SIZE;
584 if (unlikely((address % sizeof(u32)) != 0))
586 address -= key->both.offset;
588 if (unlikely(!access_ok(uaddr, sizeof(u32))))
591 if (unlikely(should_fail_futex(fshared)))
595 * PROCESS_PRIVATE futexes are fast.
596 * As the mm cannot disappear under us and the 'key' only needs
597 * virtual address, we dont even have to find the underlying vma.
598 * Note : We do have to check 'uaddr' is a valid user address,
599 * but access_ok() should be faster than find_vma()
602 key->private.mm = mm;
603 key->private.address = address;
604 get_futex_key_refs(key); /* implies smp_mb(); (B) */
609 /* Ignore any VERIFY_READ mapping (futex common case) */
610 if (unlikely(should_fail_futex(fshared)))
613 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
615 * If write access is not required (eg. FUTEX_WAIT), try
616 * and get read-only access.
618 if (err == -EFAULT && rw == FUTEX_READ) {
619 err = get_user_pages_fast(address, 1, 0, &page);
628 * The treatment of mapping from this point on is critical. The page
629 * lock protects many things but in this context the page lock
630 * stabilizes mapping, prevents inode freeing in the shared
631 * file-backed region case and guards against movement to swap cache.
633 * Strictly speaking the page lock is not needed in all cases being
634 * considered here and page lock forces unnecessarily serialization
635 * From this point on, mapping will be re-verified if necessary and
636 * page lock will be acquired only if it is unavoidable
638 * Mapping checks require the head page for any compound page so the
639 * head page and mapping is looked up now. For anonymous pages, it
640 * does not matter if the page splits in the future as the key is
641 * based on the address. For filesystem-backed pages, the tail is
642 * required as the index of the page determines the key. For
643 * base pages, there is no tail page and tail == page.
646 page = compound_head(page);
647 mapping = READ_ONCE(page->mapping);
650 * If page->mapping is NULL, then it cannot be a PageAnon
651 * page; but it might be the ZERO_PAGE or in the gate area or
652 * in a special mapping (all cases which we are happy to fail);
653 * or it may have been a good file page when get_user_pages_fast
654 * found it, but truncated or holepunched or subjected to
655 * invalidate_complete_page2 before we got the page lock (also
656 * cases which we are happy to fail). And we hold a reference,
657 * so refcount care in invalidate_complete_page's remove_mapping
658 * prevents drop_caches from setting mapping to NULL beneath us.
660 * The case we do have to guard against is when memory pressure made
661 * shmem_writepage move it from filecache to swapcache beneath us:
662 * an unlikely race, but we do need to retry for page->mapping.
664 if (unlikely(!mapping)) {
668 * Page lock is required to identify which special case above
669 * applies. If this is really a shmem page then the page lock
670 * will prevent unexpected transitions.
673 shmem_swizzled = PageSwapCache(page) || page->mapping;
684 * Private mappings are handled in a simple way.
686 * If the futex key is stored on an anonymous page, then the associated
687 * object is the mm which is implicitly pinned by the calling process.
689 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
690 * it's a read-only handle, it's expected that futexes attach to
691 * the object not the particular process.
693 if (PageAnon(page)) {
695 * A RO anonymous page will never change and thus doesn't make
696 * sense for futex operations.
698 if (unlikely(should_fail_futex(fshared)) || ro) {
703 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
704 key->private.mm = mm;
705 key->private.address = address;
711 * The associated futex object in this case is the inode and
712 * the page->mapping must be traversed. Ordinarily this should
713 * be stabilised under page lock but it's not strictly
714 * necessary in this case as we just want to pin the inode, not
715 * update the radix tree or anything like that.
717 * The RCU read lock is taken as the inode is finally freed
718 * under RCU. If the mapping still matches expectations then the
719 * mapping->host can be safely accessed as being a valid inode.
723 if (READ_ONCE(page->mapping) != mapping) {
730 inode = READ_ONCE(mapping->host);
738 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
739 key->shared.i_seq = get_inode_sequence_number(inode);
740 key->shared.pgoff = basepage_index(tail);
744 get_futex_key_refs(key); /* implies smp_mb(); (B) */
751 static inline void put_futex_key(union futex_key *key)
753 drop_futex_key_refs(key);
757 * fault_in_user_writeable() - Fault in user address and verify RW access
758 * @uaddr: pointer to faulting user space address
760 * Slow path to fixup the fault we just took in the atomic write
763 * We have no generic implementation of a non-destructive write to the
764 * user address. We know that we faulted in the atomic pagefault
765 * disabled section so we can as well avoid the #PF overhead by
766 * calling get_user_pages() right away.
768 static int fault_in_user_writeable(u32 __user *uaddr)
770 struct mm_struct *mm = current->mm;
773 down_read(&mm->mmap_sem);
774 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
775 FAULT_FLAG_WRITE, NULL);
776 up_read(&mm->mmap_sem);
778 return ret < 0 ? ret : 0;
782 * futex_top_waiter() - Return the highest priority waiter on a futex
783 * @hb: the hash bucket the futex_q's reside in
784 * @key: the futex key (to distinguish it from other futex futex_q's)
786 * Must be called with the hb lock held.
788 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
789 union futex_key *key)
791 struct futex_q *this;
793 plist_for_each_entry(this, &hb->chain, list) {
794 if (match_futex(&this->key, key))
800 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
801 u32 uval, u32 newval)
806 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
812 static int get_futex_value_locked(u32 *dest, u32 __user *from)
817 ret = __get_user(*dest, from);
820 return ret ? -EFAULT : 0;
827 static int refill_pi_state_cache(void)
829 struct futex_pi_state *pi_state;
831 if (likely(current->pi_state_cache))
834 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
839 INIT_LIST_HEAD(&pi_state->list);
840 /* pi_mutex gets initialized later */
841 pi_state->owner = NULL;
842 refcount_set(&pi_state->refcount, 1);
843 pi_state->key = FUTEX_KEY_INIT;
845 current->pi_state_cache = pi_state;
850 static struct futex_pi_state *alloc_pi_state(void)
852 struct futex_pi_state *pi_state = current->pi_state_cache;
855 current->pi_state_cache = NULL;
860 static void get_pi_state(struct futex_pi_state *pi_state)
862 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
866 * Drops a reference to the pi_state object and frees or caches it
867 * when the last reference is gone.
869 static void put_pi_state(struct futex_pi_state *pi_state)
874 if (!refcount_dec_and_test(&pi_state->refcount))
878 * If pi_state->owner is NULL, the owner is most probably dying
879 * and has cleaned up the pi_state already
881 if (pi_state->owner) {
882 struct task_struct *owner;
884 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
885 owner = pi_state->owner;
887 raw_spin_lock(&owner->pi_lock);
888 list_del_init(&pi_state->list);
889 raw_spin_unlock(&owner->pi_lock);
891 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
892 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
895 if (current->pi_state_cache) {
899 * pi_state->list is already empty.
900 * clear pi_state->owner.
901 * refcount is at 0 - put it back to 1.
903 pi_state->owner = NULL;
904 refcount_set(&pi_state->refcount, 1);
905 current->pi_state_cache = pi_state;
909 #ifdef CONFIG_FUTEX_PI
912 * This task is holding PI mutexes at exit time => bad.
913 * Kernel cleans up PI-state, but userspace is likely hosed.
914 * (Robust-futex cleanup is separate and might save the day for userspace.)
916 static void exit_pi_state_list(struct task_struct *curr)
918 struct list_head *next, *head = &curr->pi_state_list;
919 struct futex_pi_state *pi_state;
920 struct futex_hash_bucket *hb;
921 union futex_key key = FUTEX_KEY_INIT;
923 if (!futex_cmpxchg_enabled)
926 * We are a ZOMBIE and nobody can enqueue itself on
927 * pi_state_list anymore, but we have to be careful
928 * versus waiters unqueueing themselves:
930 raw_spin_lock_irq(&curr->pi_lock);
931 while (!list_empty(head)) {
933 pi_state = list_entry(next, struct futex_pi_state, list);
935 hb = hash_futex(&key);
938 * We can race against put_pi_state() removing itself from the
939 * list (a waiter going away). put_pi_state() will first
940 * decrement the reference count and then modify the list, so
941 * its possible to see the list entry but fail this reference
944 * In that case; drop the locks to let put_pi_state() make
945 * progress and retry the loop.
947 if (!refcount_inc_not_zero(&pi_state->refcount)) {
948 raw_spin_unlock_irq(&curr->pi_lock);
950 raw_spin_lock_irq(&curr->pi_lock);
953 raw_spin_unlock_irq(&curr->pi_lock);
955 spin_lock(&hb->lock);
956 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
957 raw_spin_lock(&curr->pi_lock);
959 * We dropped the pi-lock, so re-check whether this
960 * task still owns the PI-state:
962 if (head->next != next) {
963 /* retain curr->pi_lock for the loop invariant */
964 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
965 spin_unlock(&hb->lock);
966 put_pi_state(pi_state);
970 WARN_ON(pi_state->owner != curr);
971 WARN_ON(list_empty(&pi_state->list));
972 list_del_init(&pi_state->list);
973 pi_state->owner = NULL;
975 raw_spin_unlock(&curr->pi_lock);
976 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
977 spin_unlock(&hb->lock);
979 rt_mutex_futex_unlock(&pi_state->pi_mutex);
980 put_pi_state(pi_state);
982 raw_spin_lock_irq(&curr->pi_lock);
984 raw_spin_unlock_irq(&curr->pi_lock);
987 static inline void exit_pi_state_list(struct task_struct *curr) { }
991 * We need to check the following states:
993 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
995 * [1] NULL | --- | --- | 0 | 0/1 | Valid
996 * [2] NULL | --- | --- | >0 | 0/1 | Valid
998 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1000 * [4] Found | Found | NULL | 0 | 1 | Valid
1001 * [5] Found | Found | NULL | >0 | 1 | Invalid
1003 * [6] Found | Found | task | 0 | 1 | Valid
1005 * [7] Found | Found | NULL | Any | 0 | Invalid
1007 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1008 * [9] Found | Found | task | 0 | 0 | Invalid
1009 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1011 * [1] Indicates that the kernel can acquire the futex atomically. We
1012 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1014 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1015 * thread is found then it indicates that the owner TID has died.
1017 * [3] Invalid. The waiter is queued on a non PI futex
1019 * [4] Valid state after exit_robust_list(), which sets the user space
1020 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1022 * [5] The user space value got manipulated between exit_robust_list()
1023 * and exit_pi_state_list()
1025 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1026 * the pi_state but cannot access the user space value.
1028 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1030 * [8] Owner and user space value match
1032 * [9] There is no transient state which sets the user space TID to 0
1033 * except exit_robust_list(), but this is indicated by the
1034 * FUTEX_OWNER_DIED bit. See [4]
1036 * [10] There is no transient state which leaves owner and user space
1040 * Serialization and lifetime rules:
1044 * hb -> futex_q, relation
1045 * futex_q -> pi_state, relation
1047 * (cannot be raw because hb can contain arbitrary amount
1050 * pi_mutex->wait_lock:
1054 * (and pi_mutex 'obviously')
1058 * p->pi_state_list -> pi_state->list, relation
1060 * pi_state->refcount:
1068 * pi_mutex->wait_lock
1074 * Validate that the existing waiter has a pi_state and sanity check
1075 * the pi_state against the user space value. If correct, attach to
1078 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1079 struct futex_pi_state *pi_state,
1080 struct futex_pi_state **ps)
1082 pid_t pid = uval & FUTEX_TID_MASK;
1087 * Userspace might have messed up non-PI and PI futexes [3]
1089 if (unlikely(!pi_state))
1093 * We get here with hb->lock held, and having found a
1094 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1095 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1096 * which in turn means that futex_lock_pi() still has a reference on
1099 * The waiter holding a reference on @pi_state also protects against
1100 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1101 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1102 * free pi_state before we can take a reference ourselves.
1104 WARN_ON(!refcount_read(&pi_state->refcount));
1107 * Now that we have a pi_state, we can acquire wait_lock
1108 * and do the state validation.
1110 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1113 * Since {uval, pi_state} is serialized by wait_lock, and our current
1114 * uval was read without holding it, it can have changed. Verify it
1115 * still is what we expect it to be, otherwise retry the entire
1118 if (get_futex_value_locked(&uval2, uaddr))
1125 * Handle the owner died case:
1127 if (uval & FUTEX_OWNER_DIED) {
1129 * exit_pi_state_list sets owner to NULL and wakes the
1130 * topmost waiter. The task which acquires the
1131 * pi_state->rt_mutex will fixup owner.
1133 if (!pi_state->owner) {
1135 * No pi state owner, but the user space TID
1136 * is not 0. Inconsistent state. [5]
1141 * Take a ref on the state and return success. [4]
1147 * If TID is 0, then either the dying owner has not
1148 * yet executed exit_pi_state_list() or some waiter
1149 * acquired the rtmutex in the pi state, but did not
1150 * yet fixup the TID in user space.
1152 * Take a ref on the state and return success. [6]
1158 * If the owner died bit is not set, then the pi_state
1159 * must have an owner. [7]
1161 if (!pi_state->owner)
1166 * Bail out if user space manipulated the futex value. If pi
1167 * state exists then the owner TID must be the same as the
1168 * user space TID. [9/10]
1170 if (pid != task_pid_vnr(pi_state->owner))
1174 get_pi_state(pi_state);
1175 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1192 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1197 * wait_for_owner_exiting - Block until the owner has exited
1198 * @ret: owner's current futex lock status
1199 * @exiting: Pointer to the exiting task
1201 * Caller must hold a refcount on @exiting.
1203 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1205 if (ret != -EBUSY) {
1206 WARN_ON_ONCE(exiting);
1210 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1213 mutex_lock(&exiting->futex_exit_mutex);
1215 * No point in doing state checking here. If the waiter got here
1216 * while the task was in exec()->exec_futex_release() then it can
1217 * have any FUTEX_STATE_* value when the waiter has acquired the
1218 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1219 * already. Highly unlikely and not a problem. Just one more round
1220 * through the futex maze.
1222 mutex_unlock(&exiting->futex_exit_mutex);
1224 put_task_struct(exiting);
1227 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1228 struct task_struct *tsk)
1233 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1234 * caller that the alleged owner is busy.
1236 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1240 * Reread the user space value to handle the following situation:
1244 * sys_exit() sys_futex()
1245 * do_exit() futex_lock_pi()
1246 * futex_lock_pi_atomic()
1247 * exit_signals(tsk) No waiters:
1248 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1249 * mm_release(tsk) Set waiter bit
1250 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1251 * Set owner died attach_to_pi_owner() {
1252 * *uaddr = 0xC0000000; tsk = get_task(PID);
1253 * } if (!tsk->flags & PF_EXITING) {
1255 * tsk->futex_state = } else {
1256 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1259 * return -ESRCH; <--- FAIL
1262 * Returning ESRCH unconditionally is wrong here because the
1263 * user space value has been changed by the exiting task.
1265 * The same logic applies to the case where the exiting task is
1268 if (get_futex_value_locked(&uval2, uaddr))
1271 /* If the user space value has changed, try again. */
1276 * The exiting task did not have a robust list, the robust list was
1277 * corrupted or the user space value in *uaddr is simply bogus.
1278 * Give up and tell user space.
1284 * Lookup the task for the TID provided from user space and attach to
1285 * it after doing proper sanity checks.
1287 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1288 struct futex_pi_state **ps,
1289 struct task_struct **exiting)
1291 pid_t pid = uval & FUTEX_TID_MASK;
1292 struct futex_pi_state *pi_state;
1293 struct task_struct *p;
1296 * We are the first waiter - try to look up the real owner and attach
1297 * the new pi_state to it, but bail out when TID = 0 [1]
1299 * The !pid check is paranoid. None of the call sites should end up
1300 * with pid == 0, but better safe than sorry. Let the caller retry
1304 p = find_get_task_by_vpid(pid);
1306 return handle_exit_race(uaddr, uval, NULL);
1308 if (unlikely(p->flags & PF_KTHREAD)) {
1314 * We need to look at the task state to figure out, whether the
1315 * task is exiting. To protect against the change of the task state
1316 * in futex_exit_release(), we do this protected by p->pi_lock:
1318 raw_spin_lock_irq(&p->pi_lock);
1319 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1321 * The task is on the way out. When the futex state is
1322 * FUTEX_STATE_DEAD, we know that the task has finished
1325 int ret = handle_exit_race(uaddr, uval, p);
1327 raw_spin_unlock_irq(&p->pi_lock);
1329 * If the owner task is between FUTEX_STATE_EXITING and
1330 * FUTEX_STATE_DEAD then store the task pointer and keep
1331 * the reference on the task struct. The calling code will
1332 * drop all locks, wait for the task to reach
1333 * FUTEX_STATE_DEAD and then drop the refcount. This is
1334 * required to prevent a live lock when the current task
1335 * preempted the exiting task between the two states.
1345 * No existing pi state. First waiter. [2]
1347 * This creates pi_state, we have hb->lock held, this means nothing can
1348 * observe this state, wait_lock is irrelevant.
1350 pi_state = alloc_pi_state();
1353 * Initialize the pi_mutex in locked state and make @p
1356 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1358 /* Store the key for possible exit cleanups: */
1359 pi_state->key = *key;
1361 WARN_ON(!list_empty(&pi_state->list));
1362 list_add(&pi_state->list, &p->pi_state_list);
1364 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1365 * because there is no concurrency as the object is not published yet.
1367 pi_state->owner = p;
1368 raw_spin_unlock_irq(&p->pi_lock);
1377 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1378 struct futex_hash_bucket *hb,
1379 union futex_key *key, struct futex_pi_state **ps,
1380 struct task_struct **exiting)
1382 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1385 * If there is a waiter on that futex, validate it and
1386 * attach to the pi_state when the validation succeeds.
1389 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1392 * We are the first waiter - try to look up the owner based on
1393 * @uval and attach to it.
1395 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1398 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1401 u32 uninitialized_var(curval);
1403 if (unlikely(should_fail_futex(true)))
1406 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1410 /* If user space value changed, let the caller retry */
1411 return curval != uval ? -EAGAIN : 0;
1415 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1416 * @uaddr: the pi futex user address
1417 * @hb: the pi futex hash bucket
1418 * @key: the futex key associated with uaddr and hb
1419 * @ps: the pi_state pointer where we store the result of the
1421 * @task: the task to perform the atomic lock work for. This will
1422 * be "current" except in the case of requeue pi.
1423 * @exiting: Pointer to store the task pointer of the owner task
1424 * which is in the middle of exiting
1425 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1428 * - 0 - ready to wait;
1429 * - 1 - acquired the lock;
1432 * The hb->lock and futex_key refs shall be held by the caller.
1434 * @exiting is only set when the return value is -EBUSY. If so, this holds
1435 * a refcount on the exiting task on return and the caller needs to drop it
1436 * after waiting for the exit to complete.
1438 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1439 union futex_key *key,
1440 struct futex_pi_state **ps,
1441 struct task_struct *task,
1442 struct task_struct **exiting,
1445 u32 uval, newval, vpid = task_pid_vnr(task);
1446 struct futex_q *top_waiter;
1450 * Read the user space value first so we can validate a few
1451 * things before proceeding further.
1453 if (get_futex_value_locked(&uval, uaddr))
1456 if (unlikely(should_fail_futex(true)))
1462 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1465 if ((unlikely(should_fail_futex(true))))
1469 * Lookup existing state first. If it exists, try to attach to
1472 top_waiter = futex_top_waiter(hb, key);
1474 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1477 * No waiter and user TID is 0. We are here because the
1478 * waiters or the owner died bit is set or called from
1479 * requeue_cmp_pi or for whatever reason something took the
1482 if (!(uval & FUTEX_TID_MASK)) {
1484 * We take over the futex. No other waiters and the user space
1485 * TID is 0. We preserve the owner died bit.
1487 newval = uval & FUTEX_OWNER_DIED;
1490 /* The futex requeue_pi code can enforce the waiters bit */
1492 newval |= FUTEX_WAITERS;
1494 ret = lock_pi_update_atomic(uaddr, uval, newval);
1495 /* If the take over worked, return 1 */
1496 return ret < 0 ? ret : 1;
1500 * First waiter. Set the waiters bit before attaching ourself to
1501 * the owner. If owner tries to unlock, it will be forced into
1502 * the kernel and blocked on hb->lock.
1504 newval = uval | FUTEX_WAITERS;
1505 ret = lock_pi_update_atomic(uaddr, uval, newval);
1509 * If the update of the user space value succeeded, we try to
1510 * attach to the owner. If that fails, no harm done, we only
1511 * set the FUTEX_WAITERS bit in the user space variable.
1513 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1517 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1518 * @q: The futex_q to unqueue
1520 * The q->lock_ptr must not be NULL and must be held by the caller.
1522 static void __unqueue_futex(struct futex_q *q)
1524 struct futex_hash_bucket *hb;
1526 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1528 lockdep_assert_held(q->lock_ptr);
1530 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1531 plist_del(&q->list, &hb->chain);
1536 * The hash bucket lock must be held when this is called.
1537 * Afterwards, the futex_q must not be accessed. Callers
1538 * must ensure to later call wake_up_q() for the actual
1541 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1543 struct task_struct *p = q->task;
1545 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1551 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1552 * is written, without taking any locks. This is possible in the event
1553 * of a spurious wakeup, for example. A memory barrier is required here
1554 * to prevent the following store to lock_ptr from getting ahead of the
1555 * plist_del in __unqueue_futex().
1557 smp_store_release(&q->lock_ptr, NULL);
1560 * Queue the task for later wakeup for after we've released
1563 wake_q_add_safe(wake_q, p);
1567 * Caller must hold a reference on @pi_state.
1569 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1571 u32 uninitialized_var(curval), newval;
1572 struct task_struct *new_owner;
1573 bool postunlock = false;
1574 DEFINE_WAKE_Q(wake_q);
1577 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1578 if (WARN_ON_ONCE(!new_owner)) {
1580 * As per the comment in futex_unlock_pi() this should not happen.
1582 * When this happens, give up our locks and try again, giving
1583 * the futex_lock_pi() instance time to complete, either by
1584 * waiting on the rtmutex or removing itself from the futex
1592 * We pass it to the next owner. The WAITERS bit is always kept
1593 * enabled while there is PI state around. We cleanup the owner
1594 * died bit, because we are the owner.
1596 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1598 if (unlikely(should_fail_futex(true)))
1601 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1602 if (!ret && (curval != uval)) {
1604 * If a unconditional UNLOCK_PI operation (user space did not
1605 * try the TID->0 transition) raced with a waiter setting the
1606 * FUTEX_WAITERS flag between get_user() and locking the hash
1607 * bucket lock, retry the operation.
1609 if ((FUTEX_TID_MASK & curval) == uval)
1619 * This is a point of no return; once we modify the uval there is no
1620 * going back and subsequent operations must not fail.
1623 raw_spin_lock(&pi_state->owner->pi_lock);
1624 WARN_ON(list_empty(&pi_state->list));
1625 list_del_init(&pi_state->list);
1626 raw_spin_unlock(&pi_state->owner->pi_lock);
1628 raw_spin_lock(&new_owner->pi_lock);
1629 WARN_ON(!list_empty(&pi_state->list));
1630 list_add(&pi_state->list, &new_owner->pi_state_list);
1631 pi_state->owner = new_owner;
1632 raw_spin_unlock(&new_owner->pi_lock);
1634 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1637 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1640 rt_mutex_postunlock(&wake_q);
1646 * Express the locking dependencies for lockdep:
1649 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1652 spin_lock(&hb1->lock);
1654 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1655 } else { /* hb1 > hb2 */
1656 spin_lock(&hb2->lock);
1657 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1662 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1664 spin_unlock(&hb1->lock);
1666 spin_unlock(&hb2->lock);
1670 * Wake up waiters matching bitset queued on this futex (uaddr).
1673 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1675 struct futex_hash_bucket *hb;
1676 struct futex_q *this, *next;
1677 union futex_key key = FUTEX_KEY_INIT;
1679 DEFINE_WAKE_Q(wake_q);
1684 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1685 if (unlikely(ret != 0))
1688 hb = hash_futex(&key);
1690 /* Make sure we really have tasks to wakeup */
1691 if (!hb_waiters_pending(hb))
1694 spin_lock(&hb->lock);
1696 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1697 if (match_futex (&this->key, &key)) {
1698 if (this->pi_state || this->rt_waiter) {
1703 /* Check if one of the bits is set in both bitsets */
1704 if (!(this->bitset & bitset))
1707 mark_wake_futex(&wake_q, this);
1708 if (++ret >= nr_wake)
1713 spin_unlock(&hb->lock);
1716 put_futex_key(&key);
1721 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1723 unsigned int op = (encoded_op & 0x70000000) >> 28;
1724 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1725 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1726 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1729 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1730 if (oparg < 0 || oparg > 31) {
1731 char comm[sizeof(current->comm)];
1733 * kill this print and return -EINVAL when userspace
1736 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1737 get_task_comm(comm, current), oparg);
1743 if (!access_ok(uaddr, sizeof(u32)))
1746 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1751 case FUTEX_OP_CMP_EQ:
1752 return oldval == cmparg;
1753 case FUTEX_OP_CMP_NE:
1754 return oldval != cmparg;
1755 case FUTEX_OP_CMP_LT:
1756 return oldval < cmparg;
1757 case FUTEX_OP_CMP_GE:
1758 return oldval >= cmparg;
1759 case FUTEX_OP_CMP_LE:
1760 return oldval <= cmparg;
1761 case FUTEX_OP_CMP_GT:
1762 return oldval > cmparg;
1769 * Wake up all waiters hashed on the physical page that is mapped
1770 * to this virtual address:
1773 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1774 int nr_wake, int nr_wake2, int op)
1776 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1777 struct futex_hash_bucket *hb1, *hb2;
1778 struct futex_q *this, *next;
1780 DEFINE_WAKE_Q(wake_q);
1783 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1784 if (unlikely(ret != 0))
1786 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1787 if (unlikely(ret != 0))
1790 hb1 = hash_futex(&key1);
1791 hb2 = hash_futex(&key2);
1794 double_lock_hb(hb1, hb2);
1795 op_ret = futex_atomic_op_inuser(op, uaddr2);
1796 if (unlikely(op_ret < 0)) {
1797 double_unlock_hb(hb1, hb2);
1799 if (!IS_ENABLED(CONFIG_MMU) ||
1800 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1802 * we don't get EFAULT from MMU faults if we don't have
1803 * an MMU, but we might get them from range checking
1809 if (op_ret == -EFAULT) {
1810 ret = fault_in_user_writeable(uaddr2);
1815 if (!(flags & FLAGS_SHARED)) {
1820 put_futex_key(&key2);
1821 put_futex_key(&key1);
1826 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1827 if (match_futex (&this->key, &key1)) {
1828 if (this->pi_state || this->rt_waiter) {
1832 mark_wake_futex(&wake_q, this);
1833 if (++ret >= nr_wake)
1840 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1841 if (match_futex (&this->key, &key2)) {
1842 if (this->pi_state || this->rt_waiter) {
1846 mark_wake_futex(&wake_q, this);
1847 if (++op_ret >= nr_wake2)
1855 double_unlock_hb(hb1, hb2);
1858 put_futex_key(&key2);
1860 put_futex_key(&key1);
1866 * requeue_futex() - Requeue a futex_q from one hb to another
1867 * @q: the futex_q to requeue
1868 * @hb1: the source hash_bucket
1869 * @hb2: the target hash_bucket
1870 * @key2: the new key for the requeued futex_q
1873 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1874 struct futex_hash_bucket *hb2, union futex_key *key2)
1878 * If key1 and key2 hash to the same bucket, no need to
1881 if (likely(&hb1->chain != &hb2->chain)) {
1882 plist_del(&q->list, &hb1->chain);
1883 hb_waiters_dec(hb1);
1884 hb_waiters_inc(hb2);
1885 plist_add(&q->list, &hb2->chain);
1886 q->lock_ptr = &hb2->lock;
1888 get_futex_key_refs(key2);
1893 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1895 * @key: the key of the requeue target futex
1896 * @hb: the hash_bucket of the requeue target futex
1898 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1899 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1900 * to the requeue target futex so the waiter can detect the wakeup on the right
1901 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1902 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1903 * to protect access to the pi_state to fixup the owner later. Must be called
1904 * with both q->lock_ptr and hb->lock held.
1907 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1908 struct futex_hash_bucket *hb)
1910 get_futex_key_refs(key);
1915 WARN_ON(!q->rt_waiter);
1916 q->rt_waiter = NULL;
1918 q->lock_ptr = &hb->lock;
1920 wake_up_state(q->task, TASK_NORMAL);
1924 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1925 * @pifutex: the user address of the to futex
1926 * @hb1: the from futex hash bucket, must be locked by the caller
1927 * @hb2: the to futex hash bucket, must be locked by the caller
1928 * @key1: the from futex key
1929 * @key2: the to futex key
1930 * @ps: address to store the pi_state pointer
1931 * @exiting: Pointer to store the task pointer of the owner task
1932 * which is in the middle of exiting
1933 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1935 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1936 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1937 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1938 * hb1 and hb2 must be held by the caller.
1940 * @exiting is only set when the return value is -EBUSY. If so, this holds
1941 * a refcount on the exiting task on return and the caller needs to drop it
1942 * after waiting for the exit to complete.
1945 * - 0 - failed to acquire the lock atomically;
1946 * - >0 - acquired the lock, return value is vpid of the top_waiter
1950 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1951 struct futex_hash_bucket *hb2, union futex_key *key1,
1952 union futex_key *key2, struct futex_pi_state **ps,
1953 struct task_struct **exiting, int set_waiters)
1955 struct futex_q *top_waiter = NULL;
1959 if (get_futex_value_locked(&curval, pifutex))
1962 if (unlikely(should_fail_futex(true)))
1966 * Find the top_waiter and determine if there are additional waiters.
1967 * If the caller intends to requeue more than 1 waiter to pifutex,
1968 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1969 * as we have means to handle the possible fault. If not, don't set
1970 * the bit unecessarily as it will force the subsequent unlock to enter
1973 top_waiter = futex_top_waiter(hb1, key1);
1975 /* There are no waiters, nothing for us to do. */
1979 /* Ensure we requeue to the expected futex. */
1980 if (!match_futex(top_waiter->requeue_pi_key, key2))
1984 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1985 * the contended case or if set_waiters is 1. The pi_state is returned
1986 * in ps in contended cases.
1988 vpid = task_pid_vnr(top_waiter->task);
1989 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1990 exiting, set_waiters);
1992 requeue_pi_wake_futex(top_waiter, key2, hb2);
1999 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
2000 * @uaddr1: source futex user address
2001 * @flags: futex flags (FLAGS_SHARED, etc.)
2002 * @uaddr2: target futex user address
2003 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
2004 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
2005 * @cmpval: @uaddr1 expected value (or %NULL)
2006 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
2007 * pi futex (pi to pi requeue is not supported)
2009 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
2010 * uaddr2 atomically on behalf of the top waiter.
2013 * - >=0 - on success, the number of tasks requeued or woken;
2016 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
2017 u32 __user *uaddr2, int nr_wake, int nr_requeue,
2018 u32 *cmpval, int requeue_pi)
2020 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
2021 int drop_count = 0, task_count = 0, ret;
2022 struct futex_pi_state *pi_state = NULL;
2023 struct futex_hash_bucket *hb1, *hb2;
2024 struct futex_q *this, *next;
2025 DEFINE_WAKE_Q(wake_q);
2027 if (nr_wake < 0 || nr_requeue < 0)
2031 * When PI not supported: return -ENOSYS if requeue_pi is true,
2032 * consequently the compiler knows requeue_pi is always false past
2033 * this point which will optimize away all the conditional code
2036 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
2041 * Requeue PI only works on two distinct uaddrs. This
2042 * check is only valid for private futexes. See below.
2044 if (uaddr1 == uaddr2)
2048 * requeue_pi requires a pi_state, try to allocate it now
2049 * without any locks in case it fails.
2051 if (refill_pi_state_cache())
2054 * requeue_pi must wake as many tasks as it can, up to nr_wake
2055 * + nr_requeue, since it acquires the rt_mutex prior to
2056 * returning to userspace, so as to not leave the rt_mutex with
2057 * waiters and no owner. However, second and third wake-ups
2058 * cannot be predicted as they involve race conditions with the
2059 * first wake and a fault while looking up the pi_state. Both
2060 * pthread_cond_signal() and pthread_cond_broadcast() should
2068 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
2069 if (unlikely(ret != 0))
2071 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2072 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
2073 if (unlikely(ret != 0))
2077 * The check above which compares uaddrs is not sufficient for
2078 * shared futexes. We need to compare the keys:
2080 if (requeue_pi && match_futex(&key1, &key2)) {
2085 hb1 = hash_futex(&key1);
2086 hb2 = hash_futex(&key2);
2089 hb_waiters_inc(hb2);
2090 double_lock_hb(hb1, hb2);
2092 if (likely(cmpval != NULL)) {
2095 ret = get_futex_value_locked(&curval, uaddr1);
2097 if (unlikely(ret)) {
2098 double_unlock_hb(hb1, hb2);
2099 hb_waiters_dec(hb2);
2101 ret = get_user(curval, uaddr1);
2105 if (!(flags & FLAGS_SHARED))
2108 put_futex_key(&key2);
2109 put_futex_key(&key1);
2112 if (curval != *cmpval) {
2118 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2119 struct task_struct *exiting = NULL;
2122 * Attempt to acquire uaddr2 and wake the top waiter. If we
2123 * intend to requeue waiters, force setting the FUTEX_WAITERS
2124 * bit. We force this here where we are able to easily handle
2125 * faults rather in the requeue loop below.
2127 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2129 &exiting, nr_requeue);
2132 * At this point the top_waiter has either taken uaddr2 or is
2133 * waiting on it. If the former, then the pi_state will not
2134 * exist yet, look it up one more time to ensure we have a
2135 * reference to it. If the lock was taken, ret contains the
2136 * vpid of the top waiter task.
2137 * If the lock was not taken, we have pi_state and an initial
2138 * refcount on it. In case of an error we have nothing.
2145 * If we acquired the lock, then the user space value
2146 * of uaddr2 should be vpid. It cannot be changed by
2147 * the top waiter as it is blocked on hb2 lock if it
2148 * tries to do so. If something fiddled with it behind
2149 * our back the pi state lookup might unearth it. So
2150 * we rather use the known value than rereading and
2151 * handing potential crap to lookup_pi_state.
2153 * If that call succeeds then we have pi_state and an
2154 * initial refcount on it.
2156 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2157 &pi_state, &exiting);
2162 /* We hold a reference on the pi state. */
2165 /* If the above failed, then pi_state is NULL */
2167 double_unlock_hb(hb1, hb2);
2168 hb_waiters_dec(hb2);
2169 put_futex_key(&key2);
2170 put_futex_key(&key1);
2171 ret = fault_in_user_writeable(uaddr2);
2178 * Two reasons for this:
2179 * - EBUSY: Owner is exiting and we just wait for the
2181 * - EAGAIN: The user space value changed.
2183 double_unlock_hb(hb1, hb2);
2184 hb_waiters_dec(hb2);
2185 put_futex_key(&key2);
2186 put_futex_key(&key1);
2188 * Handle the case where the owner is in the middle of
2189 * exiting. Wait for the exit to complete otherwise
2190 * this task might loop forever, aka. live lock.
2192 wait_for_owner_exiting(ret, exiting);
2200 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2201 if (task_count - nr_wake >= nr_requeue)
2204 if (!match_futex(&this->key, &key1))
2208 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2209 * be paired with each other and no other futex ops.
2211 * We should never be requeueing a futex_q with a pi_state,
2212 * which is awaiting a futex_unlock_pi().
2214 if ((requeue_pi && !this->rt_waiter) ||
2215 (!requeue_pi && this->rt_waiter) ||
2222 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2223 * lock, we already woke the top_waiter. If not, it will be
2224 * woken by futex_unlock_pi().
2226 if (++task_count <= nr_wake && !requeue_pi) {
2227 mark_wake_futex(&wake_q, this);
2231 /* Ensure we requeue to the expected futex for requeue_pi. */
2232 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2238 * Requeue nr_requeue waiters and possibly one more in the case
2239 * of requeue_pi if we couldn't acquire the lock atomically.
2243 * Prepare the waiter to take the rt_mutex. Take a
2244 * refcount on the pi_state and store the pointer in
2245 * the futex_q object of the waiter.
2247 get_pi_state(pi_state);
2248 this->pi_state = pi_state;
2249 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2254 * We got the lock. We do neither drop the
2255 * refcount on pi_state nor clear
2256 * this->pi_state because the waiter needs the
2257 * pi_state for cleaning up the user space
2258 * value. It will drop the refcount after
2261 requeue_pi_wake_futex(this, &key2, hb2);
2266 * rt_mutex_start_proxy_lock() detected a
2267 * potential deadlock when we tried to queue
2268 * that waiter. Drop the pi_state reference
2269 * which we took above and remove the pointer
2270 * to the state from the waiters futex_q
2273 this->pi_state = NULL;
2274 put_pi_state(pi_state);
2276 * We stop queueing more waiters and let user
2277 * space deal with the mess.
2282 requeue_futex(this, hb1, hb2, &key2);
2287 * We took an extra initial reference to the pi_state either
2288 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2289 * need to drop it here again.
2291 put_pi_state(pi_state);
2294 double_unlock_hb(hb1, hb2);
2296 hb_waiters_dec(hb2);
2299 * drop_futex_key_refs() must be called outside the spinlocks. During
2300 * the requeue we moved futex_q's from the hash bucket at key1 to the
2301 * one at key2 and updated their key pointer. We no longer need to
2302 * hold the references to key1.
2304 while (--drop_count >= 0)
2305 drop_futex_key_refs(&key1);
2308 put_futex_key(&key2);
2310 put_futex_key(&key1);
2312 return ret ? ret : task_count;
2315 /* The key must be already stored in q->key. */
2316 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2317 __acquires(&hb->lock)
2319 struct futex_hash_bucket *hb;
2321 hb = hash_futex(&q->key);
2324 * Increment the counter before taking the lock so that
2325 * a potential waker won't miss a to-be-slept task that is
2326 * waiting for the spinlock. This is safe as all queue_lock()
2327 * users end up calling queue_me(). Similarly, for housekeeping,
2328 * decrement the counter at queue_unlock() when some error has
2329 * occurred and we don't end up adding the task to the list.
2331 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2333 q->lock_ptr = &hb->lock;
2335 spin_lock(&hb->lock);
2340 queue_unlock(struct futex_hash_bucket *hb)
2341 __releases(&hb->lock)
2343 spin_unlock(&hb->lock);
2347 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2352 * The priority used to register this element is
2353 * - either the real thread-priority for the real-time threads
2354 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2355 * - or MAX_RT_PRIO for non-RT threads.
2356 * Thus, all RT-threads are woken first in priority order, and
2357 * the others are woken last, in FIFO order.
2359 prio = min(current->normal_prio, MAX_RT_PRIO);
2361 plist_node_init(&q->list, prio);
2362 plist_add(&q->list, &hb->chain);
2367 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2368 * @q: The futex_q to enqueue
2369 * @hb: The destination hash bucket
2371 * The hb->lock must be held by the caller, and is released here. A call to
2372 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2373 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2374 * or nothing if the unqueue is done as part of the wake process and the unqueue
2375 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2378 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2379 __releases(&hb->lock)
2382 spin_unlock(&hb->lock);
2386 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2387 * @q: The futex_q to unqueue
2389 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2390 * be paired with exactly one earlier call to queue_me().
2393 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2394 * - 0 - if the futex_q was already removed by the waking thread
2396 static int unqueue_me(struct futex_q *q)
2398 spinlock_t *lock_ptr;
2401 /* In the common case we don't take the spinlock, which is nice. */
2404 * q->lock_ptr can change between this read and the following spin_lock.
2405 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2406 * optimizing lock_ptr out of the logic below.
2408 lock_ptr = READ_ONCE(q->lock_ptr);
2409 if (lock_ptr != NULL) {
2410 spin_lock(lock_ptr);
2412 * q->lock_ptr can change between reading it and
2413 * spin_lock(), causing us to take the wrong lock. This
2414 * corrects the race condition.
2416 * Reasoning goes like this: if we have the wrong lock,
2417 * q->lock_ptr must have changed (maybe several times)
2418 * between reading it and the spin_lock(). It can
2419 * change again after the spin_lock() but only if it was
2420 * already changed before the spin_lock(). It cannot,
2421 * however, change back to the original value. Therefore
2422 * we can detect whether we acquired the correct lock.
2424 if (unlikely(lock_ptr != q->lock_ptr)) {
2425 spin_unlock(lock_ptr);
2430 BUG_ON(q->pi_state);
2432 spin_unlock(lock_ptr);
2436 drop_futex_key_refs(&q->key);
2441 * PI futexes can not be requeued and must remove themself from the
2442 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2445 static void unqueue_me_pi(struct futex_q *q)
2446 __releases(q->lock_ptr)
2450 BUG_ON(!q->pi_state);
2451 put_pi_state(q->pi_state);
2454 spin_unlock(q->lock_ptr);
2457 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2458 struct task_struct *argowner)
2460 struct futex_pi_state *pi_state = q->pi_state;
2461 u32 uval, uninitialized_var(curval), newval;
2462 struct task_struct *oldowner, *newowner;
2466 lockdep_assert_held(q->lock_ptr);
2468 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2470 oldowner = pi_state->owner;
2473 * We are here because either:
2475 * - we stole the lock and pi_state->owner needs updating to reflect
2476 * that (@argowner == current),
2480 * - someone stole our lock and we need to fix things to point to the
2481 * new owner (@argowner == NULL).
2483 * Either way, we have to replace the TID in the user space variable.
2484 * This must be atomic as we have to preserve the owner died bit here.
2486 * Note: We write the user space value _before_ changing the pi_state
2487 * because we can fault here. Imagine swapped out pages or a fork
2488 * that marked all the anonymous memory readonly for cow.
2490 * Modifying pi_state _before_ the user space value would leave the
2491 * pi_state in an inconsistent state when we fault here, because we
2492 * need to drop the locks to handle the fault. This might be observed
2493 * in the PID check in lookup_pi_state.
2497 if (oldowner != current) {
2499 * We raced against a concurrent self; things are
2500 * already fixed up. Nothing to do.
2506 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2507 /* We got the lock after all, nothing to fix. */
2513 * Since we just failed the trylock; there must be an owner.
2515 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2518 WARN_ON_ONCE(argowner != current);
2519 if (oldowner == current) {
2521 * We raced against a concurrent self; things are
2522 * already fixed up. Nothing to do.
2527 newowner = argowner;
2530 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2532 if (!pi_state->owner)
2533 newtid |= FUTEX_OWNER_DIED;
2535 err = get_futex_value_locked(&uval, uaddr);
2540 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2542 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2552 * We fixed up user space. Now we need to fix the pi_state
2555 if (pi_state->owner != NULL) {
2556 raw_spin_lock(&pi_state->owner->pi_lock);
2557 WARN_ON(list_empty(&pi_state->list));
2558 list_del_init(&pi_state->list);
2559 raw_spin_unlock(&pi_state->owner->pi_lock);
2562 pi_state->owner = newowner;
2564 raw_spin_lock(&newowner->pi_lock);
2565 WARN_ON(!list_empty(&pi_state->list));
2566 list_add(&pi_state->list, &newowner->pi_state_list);
2567 raw_spin_unlock(&newowner->pi_lock);
2568 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2573 * In order to reschedule or handle a page fault, we need to drop the
2574 * locks here. In the case of a fault, this gives the other task
2575 * (either the highest priority waiter itself or the task which stole
2576 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2577 * are back from handling the fault we need to check the pi_state after
2578 * reacquiring the locks and before trying to do another fixup. When
2579 * the fixup has been done already we simply return.
2581 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2582 * drop hb->lock since the caller owns the hb -> futex_q relation.
2583 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2586 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2587 spin_unlock(q->lock_ptr);
2591 ret = fault_in_user_writeable(uaddr);
2605 spin_lock(q->lock_ptr);
2606 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2609 * Check if someone else fixed it for us:
2611 if (pi_state->owner != oldowner) {
2622 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2626 static long futex_wait_restart(struct restart_block *restart);
2629 * fixup_owner() - Post lock pi_state and corner case management
2630 * @uaddr: user address of the futex
2631 * @q: futex_q (contains pi_state and access to the rt_mutex)
2632 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2634 * After attempting to lock an rt_mutex, this function is called to cleanup
2635 * the pi_state owner as well as handle race conditions that may allow us to
2636 * acquire the lock. Must be called with the hb lock held.
2639 * - 1 - success, lock taken;
2640 * - 0 - success, lock not taken;
2641 * - <0 - on error (-EFAULT)
2643 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2649 * Got the lock. We might not be the anticipated owner if we
2650 * did a lock-steal - fix up the PI-state in that case:
2652 * Speculative pi_state->owner read (we don't hold wait_lock);
2653 * since we own the lock pi_state->owner == current is the
2654 * stable state, anything else needs more attention.
2656 if (q->pi_state->owner != current)
2657 ret = fixup_pi_state_owner(uaddr, q, current);
2662 * If we didn't get the lock; check if anybody stole it from us. In
2663 * that case, we need to fix up the uval to point to them instead of
2664 * us, otherwise bad things happen. [10]
2666 * Another speculative read; pi_state->owner == current is unstable
2667 * but needs our attention.
2669 if (q->pi_state->owner == current) {
2670 ret = fixup_pi_state_owner(uaddr, q, NULL);
2675 * Paranoia check. If we did not take the lock, then we should not be
2676 * the owner of the rt_mutex.
2678 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2679 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2680 "pi-state %p\n", ret,
2681 q->pi_state->pi_mutex.owner,
2682 q->pi_state->owner);
2686 return ret ? ret : locked;
2690 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2691 * @hb: the futex hash bucket, must be locked by the caller
2692 * @q: the futex_q to queue up on
2693 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2695 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2696 struct hrtimer_sleeper *timeout)
2699 * The task state is guaranteed to be set before another task can
2700 * wake it. set_current_state() is implemented using smp_store_mb() and
2701 * queue_me() calls spin_unlock() upon completion, both serializing
2702 * access to the hash list and forcing another memory barrier.
2704 set_current_state(TASK_INTERRUPTIBLE);
2709 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2712 * If we have been removed from the hash list, then another task
2713 * has tried to wake us, and we can skip the call to schedule().
2715 if (likely(!plist_node_empty(&q->list))) {
2717 * If the timer has already expired, current will already be
2718 * flagged for rescheduling. Only call schedule if there
2719 * is no timeout, or if it has yet to expire.
2721 if (!timeout || timeout->task)
2722 freezable_schedule();
2724 __set_current_state(TASK_RUNNING);
2728 * futex_wait_setup() - Prepare to wait on a futex
2729 * @uaddr: the futex userspace address
2730 * @val: the expected value
2731 * @flags: futex flags (FLAGS_SHARED, etc.)
2732 * @q: the associated futex_q
2733 * @hb: storage for hash_bucket pointer to be returned to caller
2735 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2736 * compare it with the expected value. Handle atomic faults internally.
2737 * Return with the hb lock held and a q.key reference on success, and unlocked
2738 * with no q.key reference on failure.
2741 * - 0 - uaddr contains val and hb has been locked;
2742 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2744 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2745 struct futex_q *q, struct futex_hash_bucket **hb)
2751 * Access the page AFTER the hash-bucket is locked.
2752 * Order is important:
2754 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2755 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2757 * The basic logical guarantee of a futex is that it blocks ONLY
2758 * if cond(var) is known to be true at the time of blocking, for
2759 * any cond. If we locked the hash-bucket after testing *uaddr, that
2760 * would open a race condition where we could block indefinitely with
2761 * cond(var) false, which would violate the guarantee.
2763 * On the other hand, we insert q and release the hash-bucket only
2764 * after testing *uaddr. This guarantees that futex_wait() will NOT
2765 * absorb a wakeup if *uaddr does not match the desired values
2766 * while the syscall executes.
2769 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2770 if (unlikely(ret != 0))
2774 *hb = queue_lock(q);
2776 ret = get_futex_value_locked(&uval, uaddr);
2781 ret = get_user(uval, uaddr);
2785 if (!(flags & FLAGS_SHARED))
2788 put_futex_key(&q->key);
2799 put_futex_key(&q->key);
2803 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2804 ktime_t *abs_time, u32 bitset)
2806 struct hrtimer_sleeper timeout, *to;
2807 struct restart_block *restart;
2808 struct futex_hash_bucket *hb;
2809 struct futex_q q = futex_q_init;
2816 to = futex_setup_timer(abs_time, &timeout, flags,
2817 current->timer_slack_ns);
2820 * Prepare to wait on uaddr. On success, holds hb lock and increments
2823 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2827 /* queue_me and wait for wakeup, timeout, or a signal. */
2828 futex_wait_queue_me(hb, &q, to);
2830 /* If we were woken (and unqueued), we succeeded, whatever. */
2832 /* unqueue_me() drops q.key ref */
2833 if (!unqueue_me(&q))
2836 if (to && !to->task)
2840 * We expect signal_pending(current), but we might be the
2841 * victim of a spurious wakeup as well.
2843 if (!signal_pending(current))
2850 restart = ¤t->restart_block;
2851 restart->fn = futex_wait_restart;
2852 restart->futex.uaddr = uaddr;
2853 restart->futex.val = val;
2854 restart->futex.time = *abs_time;
2855 restart->futex.bitset = bitset;
2856 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2858 ret = -ERESTART_RESTARTBLOCK;
2862 hrtimer_cancel(&to->timer);
2863 destroy_hrtimer_on_stack(&to->timer);
2869 static long futex_wait_restart(struct restart_block *restart)
2871 u32 __user *uaddr = restart->futex.uaddr;
2872 ktime_t t, *tp = NULL;
2874 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2875 t = restart->futex.time;
2878 restart->fn = do_no_restart_syscall;
2880 return (long)futex_wait(uaddr, restart->futex.flags,
2881 restart->futex.val, tp, restart->futex.bitset);
2886 * Userspace tried a 0 -> TID atomic transition of the futex value
2887 * and failed. The kernel side here does the whole locking operation:
2888 * if there are waiters then it will block as a consequence of relying
2889 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2890 * a 0 value of the futex too.).
2892 * Also serves as futex trylock_pi()'ing, and due semantics.
2894 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2895 ktime_t *time, int trylock)
2897 struct hrtimer_sleeper timeout, *to;
2898 struct futex_pi_state *pi_state = NULL;
2899 struct task_struct *exiting = NULL;
2900 struct rt_mutex_waiter rt_waiter;
2901 struct futex_hash_bucket *hb;
2902 struct futex_q q = futex_q_init;
2905 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2908 if (refill_pi_state_cache())
2911 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2914 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2915 if (unlikely(ret != 0))
2919 hb = queue_lock(&q);
2921 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2923 if (unlikely(ret)) {
2925 * Atomic work succeeded and we got the lock,
2926 * or failed. Either way, we do _not_ block.
2930 /* We got the lock. */
2932 goto out_unlock_put_key;
2938 * Two reasons for this:
2939 * - EBUSY: Task is exiting and we just wait for the
2941 * - EAGAIN: The user space value changed.
2944 put_futex_key(&q.key);
2946 * Handle the case where the owner is in the middle of
2947 * exiting. Wait for the exit to complete otherwise
2948 * this task might loop forever, aka. live lock.
2950 wait_for_owner_exiting(ret, exiting);
2954 goto out_unlock_put_key;
2958 WARN_ON(!q.pi_state);
2961 * Only actually queue now that the atomic ops are done:
2966 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2967 /* Fixup the trylock return value: */
2968 ret = ret ? 0 : -EWOULDBLOCK;
2972 rt_mutex_init_waiter(&rt_waiter);
2975 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2976 * hold it while doing rt_mutex_start_proxy(), because then it will
2977 * include hb->lock in the blocking chain, even through we'll not in
2978 * fact hold it while blocking. This will lead it to report -EDEADLK
2979 * and BUG when futex_unlock_pi() interleaves with this.
2981 * Therefore acquire wait_lock while holding hb->lock, but drop the
2982 * latter before calling __rt_mutex_start_proxy_lock(). This
2983 * interleaves with futex_unlock_pi() -- which does a similar lock
2984 * handoff -- such that the latter can observe the futex_q::pi_state
2985 * before __rt_mutex_start_proxy_lock() is done.
2987 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2988 spin_unlock(q.lock_ptr);
2990 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2991 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2992 * it sees the futex_q::pi_state.
2994 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2995 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3004 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
3006 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3009 spin_lock(q.lock_ptr);
3011 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3012 * first acquire the hb->lock before removing the lock from the
3013 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3016 * In particular; it is important that futex_unlock_pi() can not
3017 * observe this inconsistency.
3019 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3024 * Fixup the pi_state owner and possibly acquire the lock if we
3027 res = fixup_owner(uaddr, &q, !ret);
3029 * If fixup_owner() returned an error, proprogate that. If it acquired
3030 * the lock, clear our -ETIMEDOUT or -EINTR.
3033 ret = (res < 0) ? res : 0;
3036 * If fixup_owner() faulted and was unable to handle the fault, unlock
3037 * it and return the fault to userspace.
3039 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
3040 pi_state = q.pi_state;
3041 get_pi_state(pi_state);
3044 /* Unqueue and drop the lock */
3048 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3049 put_pi_state(pi_state);
3058 put_futex_key(&q.key);
3061 hrtimer_cancel(&to->timer);
3062 destroy_hrtimer_on_stack(&to->timer);
3064 return ret != -EINTR ? ret : -ERESTARTNOINTR;
3069 ret = fault_in_user_writeable(uaddr);
3073 if (!(flags & FLAGS_SHARED))
3076 put_futex_key(&q.key);
3081 * Userspace attempted a TID -> 0 atomic transition, and failed.
3082 * This is the in-kernel slowpath: we look up the PI state (if any),
3083 * and do the rt-mutex unlock.
3085 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3087 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
3088 union futex_key key = FUTEX_KEY_INIT;
3089 struct futex_hash_bucket *hb;
3090 struct futex_q *top_waiter;
3093 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3097 if (get_user(uval, uaddr))
3100 * We release only a lock we actually own:
3102 if ((uval & FUTEX_TID_MASK) != vpid)
3105 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3109 hb = hash_futex(&key);
3110 spin_lock(&hb->lock);
3113 * Check waiters first. We do not trust user space values at
3114 * all and we at least want to know if user space fiddled
3115 * with the futex value instead of blindly unlocking.
3117 top_waiter = futex_top_waiter(hb, &key);
3119 struct futex_pi_state *pi_state = top_waiter->pi_state;
3126 * If current does not own the pi_state then the futex is
3127 * inconsistent and user space fiddled with the futex value.
3129 if (pi_state->owner != current)
3132 get_pi_state(pi_state);
3134 * By taking wait_lock while still holding hb->lock, we ensure
3135 * there is no point where we hold neither; and therefore
3136 * wake_futex_pi() must observe a state consistent with what we
3139 * In particular; this forces __rt_mutex_start_proxy() to
3140 * complete such that we're guaranteed to observe the
3141 * rt_waiter. Also see the WARN in wake_futex_pi().
3143 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3144 spin_unlock(&hb->lock);
3146 /* drops pi_state->pi_mutex.wait_lock */
3147 ret = wake_futex_pi(uaddr, uval, pi_state);
3149 put_pi_state(pi_state);
3152 * Success, we're done! No tricky corner cases.
3157 * The atomic access to the futex value generated a
3158 * pagefault, so retry the user-access and the wakeup:
3163 * A unconditional UNLOCK_PI op raced against a waiter
3164 * setting the FUTEX_WAITERS bit. Try again.
3169 * wake_futex_pi has detected invalid state. Tell user
3176 * We have no kernel internal state, i.e. no waiters in the
3177 * kernel. Waiters which are about to queue themselves are stuck
3178 * on hb->lock. So we can safely ignore them. We do neither
3179 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3182 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3183 spin_unlock(&hb->lock);
3198 * If uval has changed, let user space handle it.
3200 ret = (curval == uval) ? 0 : -EAGAIN;
3203 spin_unlock(&hb->lock);
3205 put_futex_key(&key);
3209 put_futex_key(&key);
3214 put_futex_key(&key);
3216 ret = fault_in_user_writeable(uaddr);
3224 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3225 * @hb: the hash_bucket futex_q was original enqueued on
3226 * @q: the futex_q woken while waiting to be requeued
3227 * @key2: the futex_key of the requeue target futex
3228 * @timeout: the timeout associated with the wait (NULL if none)
3230 * Detect if the task was woken on the initial futex as opposed to the requeue
3231 * target futex. If so, determine if it was a timeout or a signal that caused
3232 * the wakeup and return the appropriate error code to the caller. Must be
3233 * called with the hb lock held.
3236 * - 0 = no early wakeup detected;
3237 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3240 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3241 struct futex_q *q, union futex_key *key2,
3242 struct hrtimer_sleeper *timeout)
3247 * With the hb lock held, we avoid races while we process the wakeup.
3248 * We only need to hold hb (and not hb2) to ensure atomicity as the
3249 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3250 * It can't be requeued from uaddr2 to something else since we don't
3251 * support a PI aware source futex for requeue.
3253 if (!match_futex(&q->key, key2)) {
3254 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3256 * We were woken prior to requeue by a timeout or a signal.
3257 * Unqueue the futex_q and determine which it was.
3259 plist_del(&q->list, &hb->chain);
3262 /* Handle spurious wakeups gracefully */
3264 if (timeout && !timeout->task)
3266 else if (signal_pending(current))
3267 ret = -ERESTARTNOINTR;
3273 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3274 * @uaddr: the futex we initially wait on (non-pi)
3275 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3276 * the same type, no requeueing from private to shared, etc.
3277 * @val: the expected value of uaddr
3278 * @abs_time: absolute timeout
3279 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3280 * @uaddr2: the pi futex we will take prior to returning to user-space
3282 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3283 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3284 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3285 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3286 * without one, the pi logic would not know which task to boost/deboost, if
3287 * there was a need to.
3289 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3290 * via the following--
3291 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3292 * 2) wakeup on uaddr2 after a requeue
3296 * If 3, cleanup and return -ERESTARTNOINTR.
3298 * If 2, we may then block on trying to take the rt_mutex and return via:
3299 * 5) successful lock
3302 * 8) other lock acquisition failure
3304 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3306 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3312 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3313 u32 val, ktime_t *abs_time, u32 bitset,
3316 struct hrtimer_sleeper timeout, *to;
3317 struct futex_pi_state *pi_state = NULL;
3318 struct rt_mutex_waiter rt_waiter;
3319 struct futex_hash_bucket *hb;
3320 union futex_key key2 = FUTEX_KEY_INIT;
3321 struct futex_q q = futex_q_init;
3324 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3327 if (uaddr == uaddr2)
3333 to = futex_setup_timer(abs_time, &timeout, flags,
3334 current->timer_slack_ns);
3337 * The waiter is allocated on our stack, manipulated by the requeue
3338 * code while we sleep on uaddr.
3340 rt_mutex_init_waiter(&rt_waiter);
3342 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3343 if (unlikely(ret != 0))
3347 q.rt_waiter = &rt_waiter;
3348 q.requeue_pi_key = &key2;
3351 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3354 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3359 * The check above which compares uaddrs is not sufficient for
3360 * shared futexes. We need to compare the keys:
3362 if (match_futex(&q.key, &key2)) {
3368 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3369 futex_wait_queue_me(hb, &q, to);
3371 spin_lock(&hb->lock);
3372 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3373 spin_unlock(&hb->lock);
3378 * In order for us to be here, we know our q.key == key2, and since
3379 * we took the hb->lock above, we also know that futex_requeue() has
3380 * completed and we no longer have to concern ourselves with a wakeup
3381 * race with the atomic proxy lock acquisition by the requeue code. The
3382 * futex_requeue dropped our key1 reference and incremented our key2
3386 /* Check if the requeue code acquired the second futex for us. */
3389 * Got the lock. We might not be the anticipated owner if we
3390 * did a lock-steal - fix up the PI-state in that case.
3392 if (q.pi_state && (q.pi_state->owner != current)) {
3393 spin_lock(q.lock_ptr);
3394 ret = fixup_pi_state_owner(uaddr2, &q, current);
3395 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3396 pi_state = q.pi_state;
3397 get_pi_state(pi_state);
3400 * Drop the reference to the pi state which
3401 * the requeue_pi() code acquired for us.
3403 put_pi_state(q.pi_state);
3404 spin_unlock(q.lock_ptr);
3407 struct rt_mutex *pi_mutex;
3410 * We have been woken up by futex_unlock_pi(), a timeout, or a
3411 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3414 WARN_ON(!q.pi_state);
3415 pi_mutex = &q.pi_state->pi_mutex;
3416 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3418 spin_lock(q.lock_ptr);
3419 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3422 debug_rt_mutex_free_waiter(&rt_waiter);
3424 * Fixup the pi_state owner and possibly acquire the lock if we
3427 res = fixup_owner(uaddr2, &q, !ret);
3429 * If fixup_owner() returned an error, proprogate that. If it
3430 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3433 ret = (res < 0) ? res : 0;
3436 * If fixup_pi_state_owner() faulted and was unable to handle
3437 * the fault, unlock the rt_mutex and return the fault to
3440 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3441 pi_state = q.pi_state;
3442 get_pi_state(pi_state);
3445 /* Unqueue and drop the lock. */
3450 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3451 put_pi_state(pi_state);
3454 if (ret == -EINTR) {
3456 * We've already been requeued, but cannot restart by calling
3457 * futex_lock_pi() directly. We could restart this syscall, but
3458 * it would detect that the user space "val" changed and return
3459 * -EWOULDBLOCK. Save the overhead of the restart and return
3460 * -EWOULDBLOCK directly.
3466 put_futex_key(&q.key);
3468 put_futex_key(&key2);
3472 hrtimer_cancel(&to->timer);
3473 destroy_hrtimer_on_stack(&to->timer);
3479 * Support for robust futexes: the kernel cleans up held futexes at
3482 * Implementation: user-space maintains a per-thread list of locks it
3483 * is holding. Upon do_exit(), the kernel carefully walks this list,
3484 * and marks all locks that are owned by this thread with the
3485 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3486 * always manipulated with the lock held, so the list is private and
3487 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3488 * field, to allow the kernel to clean up if the thread dies after
3489 * acquiring the lock, but just before it could have added itself to
3490 * the list. There can only be one such pending lock.
3494 * sys_set_robust_list() - Set the robust-futex list head of a task
3495 * @head: pointer to the list-head
3496 * @len: length of the list-head, as userspace expects
3498 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3501 if (!futex_cmpxchg_enabled)
3504 * The kernel knows only one size for now:
3506 if (unlikely(len != sizeof(*head)))
3509 current->robust_list = head;
3515 * sys_get_robust_list() - Get the robust-futex list head of a task
3516 * @pid: pid of the process [zero for current task]
3517 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3518 * @len_ptr: pointer to a length field, the kernel fills in the header size
3520 SYSCALL_DEFINE3(get_robust_list, int, pid,
3521 struct robust_list_head __user * __user *, head_ptr,
3522 size_t __user *, len_ptr)
3524 struct robust_list_head __user *head;
3526 struct task_struct *p;
3528 if (!futex_cmpxchg_enabled)
3537 p = find_task_by_vpid(pid);
3543 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3546 head = p->robust_list;
3549 if (put_user(sizeof(*head), len_ptr))
3551 return put_user(head, head_ptr);
3559 /* Constants for the pending_op argument of handle_futex_death */
3560 #define HANDLE_DEATH_PENDING true
3561 #define HANDLE_DEATH_LIST false
3564 * Process a futex-list entry, check whether it's owned by the
3565 * dying task, and do notification if so:
3567 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3568 bool pi, bool pending_op)
3570 u32 uval, uninitialized_var(nval), mval;
3573 /* Futex address must be 32bit aligned */
3574 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3578 if (get_user(uval, uaddr))
3582 * Special case for regular (non PI) futexes. The unlock path in
3583 * user space has two race scenarios:
3585 * 1. The unlock path releases the user space futex value and
3586 * before it can execute the futex() syscall to wake up
3587 * waiters it is killed.
3589 * 2. A woken up waiter is killed before it can acquire the
3590 * futex in user space.
3592 * In both cases the TID validation below prevents a wakeup of
3593 * potential waiters which can cause these waiters to block
3596 * In both cases the following conditions are met:
3598 * 1) task->robust_list->list_op_pending != NULL
3599 * @pending_op == true
3600 * 2) User space futex value == 0
3601 * 3) Regular futex: @pi == false
3603 * If these conditions are met, it is safe to attempt waking up a
3604 * potential waiter without touching the user space futex value and
3605 * trying to set the OWNER_DIED bit. The user space futex value is
3606 * uncontended and the rest of the user space mutex state is
3607 * consistent, so a woken waiter will just take over the
3608 * uncontended futex. Setting the OWNER_DIED bit would create
3609 * inconsistent state and malfunction of the user space owner died
3612 if (pending_op && !pi && !uval) {
3613 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3617 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3621 * Ok, this dying thread is truly holding a futex
3622 * of interest. Set the OWNER_DIED bit atomically
3623 * via cmpxchg, and if the value had FUTEX_WAITERS
3624 * set, wake up a waiter (if any). (We have to do a
3625 * futex_wake() even if OWNER_DIED is already set -
3626 * to handle the rare but possible case of recursive
3627 * thread-death.) The rest of the cleanup is done in
3630 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3633 * We are not holding a lock here, but we want to have
3634 * the pagefault_disable/enable() protection because
3635 * we want to handle the fault gracefully. If the
3636 * access fails we try to fault in the futex with R/W
3637 * verification via get_user_pages. get_user() above
3638 * does not guarantee R/W access. If that fails we
3639 * give up and leave the futex locked.
3641 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3644 if (fault_in_user_writeable(uaddr))
3662 * Wake robust non-PI futexes here. The wakeup of
3663 * PI futexes happens in exit_pi_state():
3665 if (!pi && (uval & FUTEX_WAITERS))
3666 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3672 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3674 static inline int fetch_robust_entry(struct robust_list __user **entry,
3675 struct robust_list __user * __user *head,
3678 unsigned long uentry;
3680 if (get_user(uentry, (unsigned long __user *)head))
3683 *entry = (void __user *)(uentry & ~1UL);
3690 * Walk curr->robust_list (very carefully, it's a userspace list!)
3691 * and mark any locks found there dead, and notify any waiters.
3693 * We silently return on any sign of list-walking problem.
3695 static void exit_robust_list(struct task_struct *curr)
3697 struct robust_list_head __user *head = curr->robust_list;
3698 struct robust_list __user *entry, *next_entry, *pending;
3699 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3700 unsigned int uninitialized_var(next_pi);
3701 unsigned long futex_offset;
3704 if (!futex_cmpxchg_enabled)
3708 * Fetch the list head (which was registered earlier, via
3709 * sys_set_robust_list()):
3711 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3714 * Fetch the relative futex offset:
3716 if (get_user(futex_offset, &head->futex_offset))
3719 * Fetch any possibly pending lock-add first, and handle it
3722 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3725 next_entry = NULL; /* avoid warning with gcc */
3726 while (entry != &head->list) {
3728 * Fetch the next entry in the list before calling
3729 * handle_futex_death:
3731 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3733 * A pending lock might already be on the list, so
3734 * don't process it twice:
3736 if (entry != pending) {
3737 if (handle_futex_death((void __user *)entry + futex_offset,
3738 curr, pi, HANDLE_DEATH_LIST))
3746 * Avoid excessively long or circular lists:
3755 handle_futex_death((void __user *)pending + futex_offset,
3756 curr, pip, HANDLE_DEATH_PENDING);
3760 static void futex_cleanup(struct task_struct *tsk)
3762 if (unlikely(tsk->robust_list)) {
3763 exit_robust_list(tsk);
3764 tsk->robust_list = NULL;
3767 #ifdef CONFIG_COMPAT
3768 if (unlikely(tsk->compat_robust_list)) {
3769 compat_exit_robust_list(tsk);
3770 tsk->compat_robust_list = NULL;
3774 if (unlikely(!list_empty(&tsk->pi_state_list)))
3775 exit_pi_state_list(tsk);
3779 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3780 * @tsk: task to set the state on
3782 * Set the futex exit state of the task lockless. The futex waiter code
3783 * observes that state when a task is exiting and loops until the task has
3784 * actually finished the futex cleanup. The worst case for this is that the
3785 * waiter runs through the wait loop until the state becomes visible.
3787 * This is called from the recursive fault handling path in do_exit().
3789 * This is best effort. Either the futex exit code has run already or
3790 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3791 * take it over. If not, the problem is pushed back to user space. If the
3792 * futex exit code did not run yet, then an already queued waiter might
3793 * block forever, but there is nothing which can be done about that.
3795 void futex_exit_recursive(struct task_struct *tsk)
3797 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3798 if (tsk->futex_state == FUTEX_STATE_EXITING)
3799 mutex_unlock(&tsk->futex_exit_mutex);
3800 tsk->futex_state = FUTEX_STATE_DEAD;
3803 static void futex_cleanup_begin(struct task_struct *tsk)
3806 * Prevent various race issues against a concurrent incoming waiter
3807 * including live locks by forcing the waiter to block on
3808 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3809 * attach_to_pi_owner().
3811 mutex_lock(&tsk->futex_exit_mutex);
3814 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3816 * This ensures that all subsequent checks of tsk->futex_state in
3817 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3818 * tsk->pi_lock held.
3820 * It guarantees also that a pi_state which was queued right before
3821 * the state change under tsk->pi_lock by a concurrent waiter must
3822 * be observed in exit_pi_state_list().
3824 raw_spin_lock_irq(&tsk->pi_lock);
3825 tsk->futex_state = FUTEX_STATE_EXITING;
3826 raw_spin_unlock_irq(&tsk->pi_lock);
3829 static void futex_cleanup_end(struct task_struct *tsk, int state)
3832 * Lockless store. The only side effect is that an observer might
3833 * take another loop until it becomes visible.
3835 tsk->futex_state = state;
3837 * Drop the exit protection. This unblocks waiters which observed
3838 * FUTEX_STATE_EXITING to reevaluate the state.
3840 mutex_unlock(&tsk->futex_exit_mutex);
3843 void futex_exec_release(struct task_struct *tsk)
3846 * The state handling is done for consistency, but in the case of
3847 * exec() there is no way to prevent futher damage as the PID stays
3848 * the same. But for the unlikely and arguably buggy case that a
3849 * futex is held on exec(), this provides at least as much state
3850 * consistency protection which is possible.
3852 futex_cleanup_begin(tsk);
3855 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3856 * exec a new binary.
3858 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3861 void futex_exit_release(struct task_struct *tsk)
3863 futex_cleanup_begin(tsk);
3865 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3868 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3869 u32 __user *uaddr2, u32 val2, u32 val3)
3871 int cmd = op & FUTEX_CMD_MASK;
3872 unsigned int flags = 0;
3874 if (!(op & FUTEX_PRIVATE_FLAG))
3875 flags |= FLAGS_SHARED;
3877 if (op & FUTEX_CLOCK_REALTIME) {
3878 flags |= FLAGS_CLOCKRT;
3879 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3880 cmd != FUTEX_WAIT_REQUEUE_PI)
3886 case FUTEX_UNLOCK_PI:
3887 case FUTEX_TRYLOCK_PI:
3888 case FUTEX_WAIT_REQUEUE_PI:
3889 case FUTEX_CMP_REQUEUE_PI:
3890 if (!futex_cmpxchg_enabled)
3896 val3 = FUTEX_BITSET_MATCH_ANY;
3898 case FUTEX_WAIT_BITSET:
3899 return futex_wait(uaddr, flags, val, timeout, val3);
3901 val3 = FUTEX_BITSET_MATCH_ANY;
3903 case FUTEX_WAKE_BITSET:
3904 return futex_wake(uaddr, flags, val, val3);
3906 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3907 case FUTEX_CMP_REQUEUE:
3908 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3910 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3912 return futex_lock_pi(uaddr, flags, timeout, 0);
3913 case FUTEX_UNLOCK_PI:
3914 return futex_unlock_pi(uaddr, flags);
3915 case FUTEX_TRYLOCK_PI:
3916 return futex_lock_pi(uaddr, flags, NULL, 1);
3917 case FUTEX_WAIT_REQUEUE_PI:
3918 val3 = FUTEX_BITSET_MATCH_ANY;
3919 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3921 case FUTEX_CMP_REQUEUE_PI:
3922 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3928 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3929 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3932 struct timespec64 ts;
3933 ktime_t t, *tp = NULL;
3935 int cmd = op & FUTEX_CMD_MASK;
3937 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3938 cmd == FUTEX_WAIT_BITSET ||
3939 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3940 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3942 if (get_timespec64(&ts, utime))
3944 if (!timespec64_valid(&ts))
3947 t = timespec64_to_ktime(ts);
3948 if (cmd == FUTEX_WAIT)
3949 t = ktime_add_safe(ktime_get(), t);
3953 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3954 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3956 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3957 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3958 val2 = (u32) (unsigned long) utime;
3960 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3963 #ifdef CONFIG_COMPAT
3965 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3968 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3969 compat_uptr_t __user *head, unsigned int *pi)
3971 if (get_user(*uentry, head))
3974 *entry = compat_ptr((*uentry) & ~1);
3975 *pi = (unsigned int)(*uentry) & 1;
3980 static void __user *futex_uaddr(struct robust_list __user *entry,
3981 compat_long_t futex_offset)
3983 compat_uptr_t base = ptr_to_compat(entry);
3984 void __user *uaddr = compat_ptr(base + futex_offset);
3990 * Walk curr->robust_list (very carefully, it's a userspace list!)
3991 * and mark any locks found there dead, and notify any waiters.
3993 * We silently return on any sign of list-walking problem.
3995 static void compat_exit_robust_list(struct task_struct *curr)
3997 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3998 struct robust_list __user *entry, *next_entry, *pending;
3999 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
4000 unsigned int uninitialized_var(next_pi);
4001 compat_uptr_t uentry, next_uentry, upending;
4002 compat_long_t futex_offset;
4005 if (!futex_cmpxchg_enabled)
4009 * Fetch the list head (which was registered earlier, via
4010 * sys_set_robust_list()):
4012 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
4015 * Fetch the relative futex offset:
4017 if (get_user(futex_offset, &head->futex_offset))
4020 * Fetch any possibly pending lock-add first, and handle it
4023 if (compat_fetch_robust_entry(&upending, &pending,
4024 &head->list_op_pending, &pip))
4027 next_entry = NULL; /* avoid warning with gcc */
4028 while (entry != (struct robust_list __user *) &head->list) {
4030 * Fetch the next entry in the list before calling
4031 * handle_futex_death:
4033 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
4034 (compat_uptr_t __user *)&entry->next, &next_pi);
4036 * A pending lock might already be on the list, so
4037 * dont process it twice:
4039 if (entry != pending) {
4040 void __user *uaddr = futex_uaddr(entry, futex_offset);
4042 if (handle_futex_death(uaddr, curr, pi,
4048 uentry = next_uentry;
4052 * Avoid excessively long or circular lists:
4060 void __user *uaddr = futex_uaddr(pending, futex_offset);
4062 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
4066 COMPAT_SYSCALL_DEFINE2(set_robust_list,
4067 struct compat_robust_list_head __user *, head,
4070 if (!futex_cmpxchg_enabled)
4073 if (unlikely(len != sizeof(*head)))
4076 current->compat_robust_list = head;
4081 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
4082 compat_uptr_t __user *, head_ptr,
4083 compat_size_t __user *, len_ptr)
4085 struct compat_robust_list_head __user *head;
4087 struct task_struct *p;
4089 if (!futex_cmpxchg_enabled)
4098 p = find_task_by_vpid(pid);
4104 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
4107 head = p->compat_robust_list;
4110 if (put_user(sizeof(*head), len_ptr))
4112 return put_user(ptr_to_compat(head), head_ptr);
4119 #endif /* CONFIG_COMPAT */
4121 #ifdef CONFIG_COMPAT_32BIT_TIME
4122 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
4123 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
4126 struct timespec64 ts;
4127 ktime_t t, *tp = NULL;
4129 int cmd = op & FUTEX_CMD_MASK;
4131 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4132 cmd == FUTEX_WAIT_BITSET ||
4133 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4134 if (get_old_timespec32(&ts, utime))
4136 if (!timespec64_valid(&ts))
4139 t = timespec64_to_ktime(ts);
4140 if (cmd == FUTEX_WAIT)
4141 t = ktime_add_safe(ktime_get(), t);
4144 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4145 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4146 val2 = (int) (unsigned long) utime;
4148 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4150 #endif /* CONFIG_COMPAT_32BIT_TIME */
4152 static void __init futex_detect_cmpxchg(void)
4154 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4158 * This will fail and we want it. Some arch implementations do
4159 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4160 * functionality. We want to know that before we call in any
4161 * of the complex code paths. Also we want to prevent
4162 * registration of robust lists in that case. NULL is
4163 * guaranteed to fault and we get -EFAULT on functional
4164 * implementation, the non-functional ones will return
4167 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4168 futex_cmpxchg_enabled = 1;
4172 static int __init futex_init(void)
4174 unsigned int futex_shift;
4177 #if CONFIG_BASE_SMALL
4178 futex_hashsize = 16;
4180 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4183 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4185 futex_hashsize < 256 ? HASH_SMALL : 0,
4187 futex_hashsize, futex_hashsize);
4188 futex_hashsize = 1UL << futex_shift;
4190 futex_detect_cmpxchg();
4192 for (i = 0; i < futex_hashsize; i++) {
4193 atomic_set(&futex_queues[i].waiters, 0);
4194 plist_head_init(&futex_queues[i].chain);
4195 spin_lock_init(&futex_queues[i].lock);
4200 core_initcall(futex_init);