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/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/syscalls.h>
38 #include <linux/freezer.h>
39 #include <linux/memblock.h>
40 #include <linux/fault-inject.h>
41 #include <linux/time_namespace.h>
43 #include <asm/futex.h>
45 #include "locking/rtmutex_common.h"
48 * READ this before attempting to hack on futexes!
50 * Basic futex operation and ordering guarantees
51 * =============================================
53 * The waiter reads the futex value in user space and calls
54 * futex_wait(). This function computes the hash bucket and acquires
55 * the hash bucket lock. After that it reads the futex user space value
56 * again and verifies that the data has not changed. If it has not changed
57 * it enqueues itself into the hash bucket, releases the hash bucket lock
60 * The waker side modifies the user space value of the futex and calls
61 * futex_wake(). This function computes the hash bucket and acquires the
62 * hash bucket lock. Then it looks for waiters on that futex in the hash
63 * bucket and wakes them.
65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
66 * the hb spinlock can be avoided and simply return. In order for this
67 * optimization to work, ordering guarantees must exist so that the waiter
68 * being added to the list is acknowledged when the list is concurrently being
69 * checked by the waker, avoiding scenarios like the following:
73 * sys_futex(WAIT, futex, val);
74 * futex_wait(futex, val);
77 * sys_futex(WAKE, futex);
82 * lock(hash_bucket(futex));
84 * unlock(hash_bucket(futex));
87 * This would cause the waiter on CPU 0 to wait forever because it
88 * missed the transition of the user space value from val to newval
89 * and the waker did not find the waiter in the hash bucket queue.
91 * The correct serialization ensures that a waiter either observes
92 * the changed user space value before blocking or is woken by a
97 * sys_futex(WAIT, futex, val);
98 * futex_wait(futex, val);
101 * smp_mb(); (A) <-- paired with -.
103 * lock(hash_bucket(futex)); |
107 * | sys_futex(WAKE, futex);
108 * | futex_wake(futex);
110 * `--------> smp_mb(); (B)
113 * unlock(hash_bucket(futex));
114 * schedule(); if (waiters)
115 * lock(hash_bucket(futex));
116 * else wake_waiters(futex);
117 * waiters--; (b) unlock(hash_bucket(futex));
119 * Where (A) orders the waiters increment and the futex value read through
120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
121 * to futex and the waiters read (see hb_waiters_pending()).
123 * This yields the following case (where X:=waiters, Y:=futex):
131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
132 * the guarantee that we cannot both miss the futex variable change and the
135 * Note that a new waiter is accounted for in (a) even when it is possible that
136 * the wait call can return error, in which case we backtrack from it in (b).
137 * Refer to the comment in queue_lock().
139 * Similarly, in order to account for waiters being requeued on another
140 * address we always increment the waiters for the destination bucket before
141 * acquiring the lock. It then decrements them again after releasing it -
142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
143 * will do the additional required waiter count housekeeping. This is done for
144 * double_lock_hb() and double_unlock_hb(), respectively.
147 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
148 #define futex_cmpxchg_enabled 1
150 static int __read_mostly futex_cmpxchg_enabled;
154 * Futex flags used to encode options to functions and preserve them across
158 # define FLAGS_SHARED 0x01
161 * NOMMU does not have per process address space. Let the compiler optimize
164 # define FLAGS_SHARED 0x00
166 #define FLAGS_CLOCKRT 0x02
167 #define FLAGS_HAS_TIMEOUT 0x04
170 * Priority Inheritance state:
172 struct futex_pi_state {
174 * list of 'owned' pi_state instances - these have to be
175 * cleaned up in do_exit() if the task exits prematurely:
177 struct list_head list;
182 struct rt_mutex_base pi_mutex;
184 struct task_struct *owner;
188 } __randomize_layout;
191 * struct futex_q - The hashed futex queue entry, one per waiting task
192 * @list: priority-sorted list of tasks waiting on this futex
193 * @task: the task waiting on the futex
194 * @lock_ptr: the hash bucket lock
195 * @key: the key the futex is hashed on
196 * @pi_state: optional priority inheritance state
197 * @rt_waiter: rt_waiter storage for use with requeue_pi
198 * @requeue_pi_key: the requeue_pi target futex key
199 * @bitset: bitset for the optional bitmasked wakeup
201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
202 * we can wake only the relevant ones (hashed queues may be shared).
204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
206 * The order of wakeup is always to make the first condition true, then
209 * PI futexes are typically woken before they are removed from the hash list via
210 * the rt_mutex code. See unqueue_me_pi().
213 struct plist_node list;
215 struct task_struct *task;
216 spinlock_t *lock_ptr;
218 struct futex_pi_state *pi_state;
219 struct rt_mutex_waiter *rt_waiter;
220 union futex_key *requeue_pi_key;
222 } __randomize_layout;
224 static const struct futex_q futex_q_init = {
225 /* list gets initialized in queue_me()*/
226 .key = FUTEX_KEY_INIT,
227 .bitset = FUTEX_BITSET_MATCH_ANY
231 * Hash buckets are shared by all the futex_keys that hash to the same
232 * location. Each key may have multiple futex_q structures, one for each task
233 * waiting on a futex.
235 struct futex_hash_bucket {
238 struct plist_head chain;
239 } ____cacheline_aligned_in_smp;
242 * The base of the bucket array and its size are always used together
243 * (after initialization only in hash_futex()), so ensure that they
244 * reside in the same cacheline.
247 struct futex_hash_bucket *queues;
248 unsigned long hashsize;
249 } __futex_data __read_mostly __aligned(2*sizeof(long));
250 #define futex_queues (__futex_data.queues)
251 #define futex_hashsize (__futex_data.hashsize)
255 * Fault injections for futexes.
257 #ifdef CONFIG_FAIL_FUTEX
260 struct fault_attr attr;
264 .attr = FAULT_ATTR_INITIALIZER,
265 .ignore_private = false,
268 static int __init setup_fail_futex(char *str)
270 return setup_fault_attr(&fail_futex.attr, str);
272 __setup("fail_futex=", setup_fail_futex);
274 static bool should_fail_futex(bool fshared)
276 if (fail_futex.ignore_private && !fshared)
279 return should_fail(&fail_futex.attr, 1);
282 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
284 static int __init fail_futex_debugfs(void)
286 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
289 dir = fault_create_debugfs_attr("fail_futex", NULL,
294 debugfs_create_bool("ignore-private", mode, dir,
295 &fail_futex.ignore_private);
299 late_initcall(fail_futex_debugfs);
301 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
304 static inline bool should_fail_futex(bool fshared)
308 #endif /* CONFIG_FAIL_FUTEX */
311 static void compat_exit_robust_list(struct task_struct *curr);
315 * Reflects a new waiter being added to the waitqueue.
317 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
320 atomic_inc(&hb->waiters);
322 * Full barrier (A), see the ordering comment above.
324 smp_mb__after_atomic();
329 * Reflects a waiter being removed from the waitqueue by wakeup
332 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
335 atomic_dec(&hb->waiters);
339 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
343 * Full barrier (B), see the ordering comment above.
346 return atomic_read(&hb->waiters);
353 * hash_futex - Return the hash bucket in the global hash
354 * @key: Pointer to the futex key for which the hash is calculated
356 * We hash on the keys returned from get_futex_key (see below) and return the
357 * corresponding hash bucket in the global hash.
359 static struct futex_hash_bucket *hash_futex(union futex_key *key)
361 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
364 return &futex_queues[hash & (futex_hashsize - 1)];
369 * match_futex - Check whether two futex keys are equal
370 * @key1: Pointer to key1
371 * @key2: Pointer to key2
373 * Return 1 if two futex_keys are equal, 0 otherwise.
375 static inline int match_futex(union futex_key *key1, union futex_key *key2)
378 && key1->both.word == key2->both.word
379 && key1->both.ptr == key2->both.ptr
380 && key1->both.offset == key2->both.offset);
389 * futex_setup_timer - set up the sleeping hrtimer.
390 * @time: ptr to the given timeout value
391 * @timeout: the hrtimer_sleeper structure to be set up
392 * @flags: futex flags
393 * @range_ns: optional range in ns
395 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
398 static inline struct hrtimer_sleeper *
399 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
400 int flags, u64 range_ns)
405 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
406 CLOCK_REALTIME : CLOCK_MONOTONIC,
409 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
410 * effectively the same as calling hrtimer_set_expires().
412 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
418 * Generate a machine wide unique identifier for this inode.
420 * This relies on u64 not wrapping in the life-time of the machine; which with
421 * 1ns resolution means almost 585 years.
423 * This further relies on the fact that a well formed program will not unmap
424 * the file while it has a (shared) futex waiting on it. This mapping will have
425 * a file reference which pins the mount and inode.
427 * If for some reason an inode gets evicted and read back in again, it will get
428 * a new sequence number and will _NOT_ match, even though it is the exact same
431 * It is important that match_futex() will never have a false-positive, esp.
432 * for PI futexes that can mess up the state. The above argues that false-negatives
433 * are only possible for malformed programs.
435 static u64 get_inode_sequence_number(struct inode *inode)
437 static atomic64_t i_seq;
440 /* Does the inode already have a sequence number? */
441 old = atomic64_read(&inode->i_sequence);
446 u64 new = atomic64_add_return(1, &i_seq);
447 if (WARN_ON_ONCE(!new))
450 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
458 * get_futex_key() - Get parameters which are the keys for a futex
459 * @uaddr: virtual address of the futex
460 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
461 * @key: address where result is stored.
462 * @rw: mapping needs to be read/write (values: FUTEX_READ,
465 * Return: a negative error code or 0
467 * The key words are stored in @key on success.
469 * For shared mappings (when @fshared), the key is:
471 * ( inode->i_sequence, page->index, offset_within_page )
473 * [ also see get_inode_sequence_number() ]
475 * For private mappings (or when !@fshared), the key is:
477 * ( current->mm, address, 0 )
479 * This allows (cross process, where applicable) identification of the futex
480 * without keeping the page pinned for the duration of the FUTEX_WAIT.
482 * lock_page() might sleep, the caller should not hold a spinlock.
484 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
485 enum futex_access rw)
487 unsigned long address = (unsigned long)uaddr;
488 struct mm_struct *mm = current->mm;
489 struct page *page, *tail;
490 struct address_space *mapping;
494 * The futex address must be "naturally" aligned.
496 key->both.offset = address % PAGE_SIZE;
497 if (unlikely((address % sizeof(u32)) != 0))
499 address -= key->both.offset;
501 if (unlikely(!access_ok(uaddr, sizeof(u32))))
504 if (unlikely(should_fail_futex(fshared)))
508 * PROCESS_PRIVATE futexes are fast.
509 * As the mm cannot disappear under us and the 'key' only needs
510 * virtual address, we dont even have to find the underlying vma.
511 * Note : We do have to check 'uaddr' is a valid user address,
512 * but access_ok() should be faster than find_vma()
515 key->private.mm = mm;
516 key->private.address = address;
521 /* Ignore any VERIFY_READ mapping (futex common case) */
522 if (unlikely(should_fail_futex(true)))
525 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
527 * If write access is not required (eg. FUTEX_WAIT), try
528 * and get read-only access.
530 if (err == -EFAULT && rw == FUTEX_READ) {
531 err = get_user_pages_fast(address, 1, 0, &page);
540 * The treatment of mapping from this point on is critical. The page
541 * lock protects many things but in this context the page lock
542 * stabilizes mapping, prevents inode freeing in the shared
543 * file-backed region case and guards against movement to swap cache.
545 * Strictly speaking the page lock is not needed in all cases being
546 * considered here and page lock forces unnecessarily serialization
547 * From this point on, mapping will be re-verified if necessary and
548 * page lock will be acquired only if it is unavoidable
550 * Mapping checks require the head page for any compound page so the
551 * head page and mapping is looked up now. For anonymous pages, it
552 * does not matter if the page splits in the future as the key is
553 * based on the address. For filesystem-backed pages, the tail is
554 * required as the index of the page determines the key. For
555 * base pages, there is no tail page and tail == page.
558 page = compound_head(page);
559 mapping = READ_ONCE(page->mapping);
562 * If page->mapping is NULL, then it cannot be a PageAnon
563 * page; but it might be the ZERO_PAGE or in the gate area or
564 * in a special mapping (all cases which we are happy to fail);
565 * or it may have been a good file page when get_user_pages_fast
566 * found it, but truncated or holepunched or subjected to
567 * invalidate_complete_page2 before we got the page lock (also
568 * cases which we are happy to fail). And we hold a reference,
569 * so refcount care in invalidate_complete_page's remove_mapping
570 * prevents drop_caches from setting mapping to NULL beneath us.
572 * The case we do have to guard against is when memory pressure made
573 * shmem_writepage move it from filecache to swapcache beneath us:
574 * an unlikely race, but we do need to retry for page->mapping.
576 if (unlikely(!mapping)) {
580 * Page lock is required to identify which special case above
581 * applies. If this is really a shmem page then the page lock
582 * will prevent unexpected transitions.
585 shmem_swizzled = PageSwapCache(page) || page->mapping;
596 * Private mappings are handled in a simple way.
598 * If the futex key is stored on an anonymous page, then the associated
599 * object is the mm which is implicitly pinned by the calling process.
601 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
602 * it's a read-only handle, it's expected that futexes attach to
603 * the object not the particular process.
605 if (PageAnon(page)) {
607 * A RO anonymous page will never change and thus doesn't make
608 * sense for futex operations.
610 if (unlikely(should_fail_futex(true)) || ro) {
615 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
616 key->private.mm = mm;
617 key->private.address = address;
623 * The associated futex object in this case is the inode and
624 * the page->mapping must be traversed. Ordinarily this should
625 * be stabilised under page lock but it's not strictly
626 * necessary in this case as we just want to pin the inode, not
627 * update the radix tree or anything like that.
629 * The RCU read lock is taken as the inode is finally freed
630 * under RCU. If the mapping still matches expectations then the
631 * mapping->host can be safely accessed as being a valid inode.
635 if (READ_ONCE(page->mapping) != mapping) {
642 inode = READ_ONCE(mapping->host);
650 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
651 key->shared.i_seq = get_inode_sequence_number(inode);
652 key->shared.pgoff = page_to_pgoff(tail);
662 * fault_in_user_writeable() - Fault in user address and verify RW access
663 * @uaddr: pointer to faulting user space address
665 * Slow path to fixup the fault we just took in the atomic write
668 * We have no generic implementation of a non-destructive write to the
669 * user address. We know that we faulted in the atomic pagefault
670 * disabled section so we can as well avoid the #PF overhead by
671 * calling get_user_pages() right away.
673 static int fault_in_user_writeable(u32 __user *uaddr)
675 struct mm_struct *mm = current->mm;
679 ret = fixup_user_fault(mm, (unsigned long)uaddr,
680 FAULT_FLAG_WRITE, NULL);
681 mmap_read_unlock(mm);
683 return ret < 0 ? ret : 0;
687 * futex_top_waiter() - Return the highest priority waiter on a futex
688 * @hb: the hash bucket the futex_q's reside in
689 * @key: the futex key (to distinguish it from other futex futex_q's)
691 * Must be called with the hb lock held.
693 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
694 union futex_key *key)
696 struct futex_q *this;
698 plist_for_each_entry(this, &hb->chain, list) {
699 if (match_futex(&this->key, key))
705 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
706 u32 uval, u32 newval)
711 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
717 static int get_futex_value_locked(u32 *dest, u32 __user *from)
722 ret = __get_user(*dest, from);
725 return ret ? -EFAULT : 0;
732 static int refill_pi_state_cache(void)
734 struct futex_pi_state *pi_state;
736 if (likely(current->pi_state_cache))
739 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
744 INIT_LIST_HEAD(&pi_state->list);
745 /* pi_mutex gets initialized later */
746 pi_state->owner = NULL;
747 refcount_set(&pi_state->refcount, 1);
748 pi_state->key = FUTEX_KEY_INIT;
750 current->pi_state_cache = pi_state;
755 static struct futex_pi_state *alloc_pi_state(void)
757 struct futex_pi_state *pi_state = current->pi_state_cache;
760 current->pi_state_cache = NULL;
765 static void pi_state_update_owner(struct futex_pi_state *pi_state,
766 struct task_struct *new_owner)
768 struct task_struct *old_owner = pi_state->owner;
770 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
773 raw_spin_lock(&old_owner->pi_lock);
774 WARN_ON(list_empty(&pi_state->list));
775 list_del_init(&pi_state->list);
776 raw_spin_unlock(&old_owner->pi_lock);
780 raw_spin_lock(&new_owner->pi_lock);
781 WARN_ON(!list_empty(&pi_state->list));
782 list_add(&pi_state->list, &new_owner->pi_state_list);
783 pi_state->owner = new_owner;
784 raw_spin_unlock(&new_owner->pi_lock);
788 static void get_pi_state(struct futex_pi_state *pi_state)
790 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
794 * Drops a reference to the pi_state object and frees or caches it
795 * when the last reference is gone.
797 static void put_pi_state(struct futex_pi_state *pi_state)
802 if (!refcount_dec_and_test(&pi_state->refcount))
806 * If pi_state->owner is NULL, the owner is most probably dying
807 * and has cleaned up the pi_state already
809 if (pi_state->owner) {
812 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
813 pi_state_update_owner(pi_state, NULL);
814 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
815 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
818 if (current->pi_state_cache) {
822 * pi_state->list is already empty.
823 * clear pi_state->owner.
824 * refcount is at 0 - put it back to 1.
826 pi_state->owner = NULL;
827 refcount_set(&pi_state->refcount, 1);
828 current->pi_state_cache = pi_state;
832 #ifdef CONFIG_FUTEX_PI
835 * This task is holding PI mutexes at exit time => bad.
836 * Kernel cleans up PI-state, but userspace is likely hosed.
837 * (Robust-futex cleanup is separate and might save the day for userspace.)
839 static void exit_pi_state_list(struct task_struct *curr)
841 struct list_head *next, *head = &curr->pi_state_list;
842 struct futex_pi_state *pi_state;
843 struct futex_hash_bucket *hb;
844 union futex_key key = FUTEX_KEY_INIT;
846 if (!futex_cmpxchg_enabled)
849 * We are a ZOMBIE and nobody can enqueue itself on
850 * pi_state_list anymore, but we have to be careful
851 * versus waiters unqueueing themselves:
853 raw_spin_lock_irq(&curr->pi_lock);
854 while (!list_empty(head)) {
856 pi_state = list_entry(next, struct futex_pi_state, list);
858 hb = hash_futex(&key);
861 * We can race against put_pi_state() removing itself from the
862 * list (a waiter going away). put_pi_state() will first
863 * decrement the reference count and then modify the list, so
864 * its possible to see the list entry but fail this reference
867 * In that case; drop the locks to let put_pi_state() make
868 * progress and retry the loop.
870 if (!refcount_inc_not_zero(&pi_state->refcount)) {
871 raw_spin_unlock_irq(&curr->pi_lock);
873 raw_spin_lock_irq(&curr->pi_lock);
876 raw_spin_unlock_irq(&curr->pi_lock);
878 spin_lock(&hb->lock);
879 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
880 raw_spin_lock(&curr->pi_lock);
882 * We dropped the pi-lock, so re-check whether this
883 * task still owns the PI-state:
885 if (head->next != next) {
886 /* retain curr->pi_lock for the loop invariant */
887 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
888 spin_unlock(&hb->lock);
889 put_pi_state(pi_state);
893 WARN_ON(pi_state->owner != curr);
894 WARN_ON(list_empty(&pi_state->list));
895 list_del_init(&pi_state->list);
896 pi_state->owner = NULL;
898 raw_spin_unlock(&curr->pi_lock);
899 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
900 spin_unlock(&hb->lock);
902 rt_mutex_futex_unlock(&pi_state->pi_mutex);
903 put_pi_state(pi_state);
905 raw_spin_lock_irq(&curr->pi_lock);
907 raw_spin_unlock_irq(&curr->pi_lock);
910 static inline void exit_pi_state_list(struct task_struct *curr) { }
914 * We need to check the following states:
916 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
918 * [1] NULL | --- | --- | 0 | 0/1 | Valid
919 * [2] NULL | --- | --- | >0 | 0/1 | Valid
921 * [3] Found | NULL | -- | Any | 0/1 | Invalid
923 * [4] Found | Found | NULL | 0 | 1 | Valid
924 * [5] Found | Found | NULL | >0 | 1 | Invalid
926 * [6] Found | Found | task | 0 | 1 | Valid
928 * [7] Found | Found | NULL | Any | 0 | Invalid
930 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
931 * [9] Found | Found | task | 0 | 0 | Invalid
932 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
934 * [1] Indicates that the kernel can acquire the futex atomically. We
935 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
937 * [2] Valid, if TID does not belong to a kernel thread. If no matching
938 * thread is found then it indicates that the owner TID has died.
940 * [3] Invalid. The waiter is queued on a non PI futex
942 * [4] Valid state after exit_robust_list(), which sets the user space
943 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
945 * [5] The user space value got manipulated between exit_robust_list()
946 * and exit_pi_state_list()
948 * [6] Valid state after exit_pi_state_list() which sets the new owner in
949 * the pi_state but cannot access the user space value.
951 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
953 * [8] Owner and user space value match
955 * [9] There is no transient state which sets the user space TID to 0
956 * except exit_robust_list(), but this is indicated by the
957 * FUTEX_OWNER_DIED bit. See [4]
959 * [10] There is no transient state which leaves owner and user space
960 * TID out of sync. Except one error case where the kernel is denied
961 * write access to the user address, see fixup_pi_state_owner().
964 * Serialization and lifetime rules:
968 * hb -> futex_q, relation
969 * futex_q -> pi_state, relation
971 * (cannot be raw because hb can contain arbitrary amount
974 * pi_mutex->wait_lock:
978 * (and pi_mutex 'obviously')
982 * p->pi_state_list -> pi_state->list, relation
983 * pi_mutex->owner -> pi_state->owner, relation
985 * pi_state->refcount:
993 * pi_mutex->wait_lock
999 * Validate that the existing waiter has a pi_state and sanity check
1000 * the pi_state against the user space value. If correct, attach to
1003 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1004 struct futex_pi_state *pi_state,
1005 struct futex_pi_state **ps)
1007 pid_t pid = uval & FUTEX_TID_MASK;
1012 * Userspace might have messed up non-PI and PI futexes [3]
1014 if (unlikely(!pi_state))
1018 * We get here with hb->lock held, and having found a
1019 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1020 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1021 * which in turn means that futex_lock_pi() still has a reference on
1024 * The waiter holding a reference on @pi_state also protects against
1025 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1026 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1027 * free pi_state before we can take a reference ourselves.
1029 WARN_ON(!refcount_read(&pi_state->refcount));
1032 * Now that we have a pi_state, we can acquire wait_lock
1033 * and do the state validation.
1035 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1038 * Since {uval, pi_state} is serialized by wait_lock, and our current
1039 * uval was read without holding it, it can have changed. Verify it
1040 * still is what we expect it to be, otherwise retry the entire
1043 if (get_futex_value_locked(&uval2, uaddr))
1050 * Handle the owner died case:
1052 if (uval & FUTEX_OWNER_DIED) {
1054 * exit_pi_state_list sets owner to NULL and wakes the
1055 * topmost waiter. The task which acquires the
1056 * pi_state->rt_mutex will fixup owner.
1058 if (!pi_state->owner) {
1060 * No pi state owner, but the user space TID
1061 * is not 0. Inconsistent state. [5]
1066 * Take a ref on the state and return success. [4]
1072 * If TID is 0, then either the dying owner has not
1073 * yet executed exit_pi_state_list() or some waiter
1074 * acquired the rtmutex in the pi state, but did not
1075 * yet fixup the TID in user space.
1077 * Take a ref on the state and return success. [6]
1083 * If the owner died bit is not set, then the pi_state
1084 * must have an owner. [7]
1086 if (!pi_state->owner)
1091 * Bail out if user space manipulated the futex value. If pi
1092 * state exists then the owner TID must be the same as the
1093 * user space TID. [9/10]
1095 if (pid != task_pid_vnr(pi_state->owner))
1099 get_pi_state(pi_state);
1100 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1117 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1122 * wait_for_owner_exiting - Block until the owner has exited
1123 * @ret: owner's current futex lock status
1124 * @exiting: Pointer to the exiting task
1126 * Caller must hold a refcount on @exiting.
1128 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1130 if (ret != -EBUSY) {
1131 WARN_ON_ONCE(exiting);
1135 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1138 mutex_lock(&exiting->futex_exit_mutex);
1140 * No point in doing state checking here. If the waiter got here
1141 * while the task was in exec()->exec_futex_release() then it can
1142 * have any FUTEX_STATE_* value when the waiter has acquired the
1143 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1144 * already. Highly unlikely and not a problem. Just one more round
1145 * through the futex maze.
1147 mutex_unlock(&exiting->futex_exit_mutex);
1149 put_task_struct(exiting);
1152 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1153 struct task_struct *tsk)
1158 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1159 * caller that the alleged owner is busy.
1161 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1165 * Reread the user space value to handle the following situation:
1169 * sys_exit() sys_futex()
1170 * do_exit() futex_lock_pi()
1171 * futex_lock_pi_atomic()
1172 * exit_signals(tsk) No waiters:
1173 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1174 * mm_release(tsk) Set waiter bit
1175 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1176 * Set owner died attach_to_pi_owner() {
1177 * *uaddr = 0xC0000000; tsk = get_task(PID);
1178 * } if (!tsk->flags & PF_EXITING) {
1180 * tsk->futex_state = } else {
1181 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1184 * return -ESRCH; <--- FAIL
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1190 * The same logic applies to the case where the exiting task is
1193 if (get_futex_value_locked(&uval2, uaddr))
1196 /* If the user space value has changed, try again. */
1201 * The exiting task did not have a robust list, the robust list was
1202 * corrupted or the user space value in *uaddr is simply bogus.
1203 * Give up and tell user space.
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1212 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213 struct futex_pi_state **ps,
1214 struct task_struct **exiting)
1216 pid_t pid = uval & FUTEX_TID_MASK;
1217 struct futex_pi_state *pi_state;
1218 struct task_struct *p;
1221 * We are the first waiter - try to look up the real owner and attach
1222 * the new pi_state to it, but bail out when TID = 0 [1]
1224 * The !pid check is paranoid. None of the call sites should end up
1225 * with pid == 0, but better safe than sorry. Let the caller retry
1229 p = find_get_task_by_vpid(pid);
1231 return handle_exit_race(uaddr, uval, NULL);
1233 if (unlikely(p->flags & PF_KTHREAD)) {
1239 * We need to look at the task state to figure out, whether the
1240 * task is exiting. To protect against the change of the task state
1241 * in futex_exit_release(), we do this protected by p->pi_lock:
1243 raw_spin_lock_irq(&p->pi_lock);
1244 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1246 * The task is on the way out. When the futex state is
1247 * FUTEX_STATE_DEAD, we know that the task has finished
1250 int ret = handle_exit_race(uaddr, uval, p);
1252 raw_spin_unlock_irq(&p->pi_lock);
1254 * If the owner task is between FUTEX_STATE_EXITING and
1255 * FUTEX_STATE_DEAD then store the task pointer and keep
1256 * the reference on the task struct. The calling code will
1257 * drop all locks, wait for the task to reach
1258 * FUTEX_STATE_DEAD and then drop the refcount. This is
1259 * required to prevent a live lock when the current task
1260 * preempted the exiting task between the two states.
1270 * No existing pi state. First waiter. [2]
1272 * This creates pi_state, we have hb->lock held, this means nothing can
1273 * observe this state, wait_lock is irrelevant.
1275 pi_state = alloc_pi_state();
1278 * Initialize the pi_mutex in locked state and make @p
1281 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1283 /* Store the key for possible exit cleanups: */
1284 pi_state->key = *key;
1286 WARN_ON(!list_empty(&pi_state->list));
1287 list_add(&pi_state->list, &p->pi_state_list);
1289 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1290 * because there is no concurrency as the object is not published yet.
1292 pi_state->owner = p;
1293 raw_spin_unlock_irq(&p->pi_lock);
1302 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1307 if (unlikely(should_fail_futex(true)))
1310 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1314 /* If user space value changed, let the caller retry */
1315 return curval != uval ? -EAGAIN : 0;
1319 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1320 * @uaddr: the pi futex user address
1321 * @hb: the pi futex hash bucket
1322 * @key: the futex key associated with uaddr and hb
1323 * @ps: the pi_state pointer where we store the result of the
1325 * @task: the task to perform the atomic lock work for. This will
1326 * be "current" except in the case of requeue pi.
1327 * @exiting: Pointer to store the task pointer of the owner task
1328 * which is in the middle of exiting
1329 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1332 * - 0 - ready to wait;
1333 * - 1 - acquired the lock;
1336 * The hb->lock must be held by the caller.
1338 * @exiting is only set when the return value is -EBUSY. If so, this holds
1339 * a refcount on the exiting task on return and the caller needs to drop it
1340 * after waiting for the exit to complete.
1342 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1343 union futex_key *key,
1344 struct futex_pi_state **ps,
1345 struct task_struct *task,
1346 struct task_struct **exiting,
1349 u32 uval, newval, vpid = task_pid_vnr(task);
1350 struct futex_q *top_waiter;
1354 * Read the user space value first so we can validate a few
1355 * things before proceeding further.
1357 if (get_futex_value_locked(&uval, uaddr))
1360 if (unlikely(should_fail_futex(true)))
1366 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1369 if ((unlikely(should_fail_futex(true))))
1373 * Lookup existing state first. If it exists, try to attach to
1376 top_waiter = futex_top_waiter(hb, key);
1378 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1381 * No waiter and user TID is 0. We are here because the
1382 * waiters or the owner died bit is set or called from
1383 * requeue_cmp_pi or for whatever reason something took the
1386 if (!(uval & FUTEX_TID_MASK)) {
1388 * We take over the futex. No other waiters and the user space
1389 * TID is 0. We preserve the owner died bit.
1391 newval = uval & FUTEX_OWNER_DIED;
1394 /* The futex requeue_pi code can enforce the waiters bit */
1396 newval |= FUTEX_WAITERS;
1398 ret = lock_pi_update_atomic(uaddr, uval, newval);
1399 /* If the take over worked, return 1 */
1400 return ret < 0 ? ret : 1;
1404 * First waiter. Set the waiters bit before attaching ourself to
1405 * the owner. If owner tries to unlock, it will be forced into
1406 * the kernel and blocked on hb->lock.
1408 newval = uval | FUTEX_WAITERS;
1409 ret = lock_pi_update_atomic(uaddr, uval, newval);
1413 * If the update of the user space value succeeded, we try to
1414 * attach to the owner. If that fails, no harm done, we only
1415 * set the FUTEX_WAITERS bit in the user space variable.
1417 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1421 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1422 * @q: The futex_q to unqueue
1424 * The q->lock_ptr must not be NULL and must be held by the caller.
1426 static void __unqueue_futex(struct futex_q *q)
1428 struct futex_hash_bucket *hb;
1430 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1432 lockdep_assert_held(q->lock_ptr);
1434 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1435 plist_del(&q->list, &hb->chain);
1440 * The hash bucket lock must be held when this is called.
1441 * Afterwards, the futex_q must not be accessed. Callers
1442 * must ensure to later call wake_up_q() for the actual
1445 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1447 struct task_struct *p = q->task;
1449 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1455 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1456 * is written, without taking any locks. This is possible in the event
1457 * of a spurious wakeup, for example. A memory barrier is required here
1458 * to prevent the following store to lock_ptr from getting ahead of the
1459 * plist_del in __unqueue_futex().
1461 smp_store_release(&q->lock_ptr, NULL);
1464 * Queue the task for later wakeup for after we've released
1467 wake_q_add_safe(wake_q, p);
1471 * Caller must hold a reference on @pi_state.
1473 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1475 struct rt_mutex_waiter *top_waiter;
1476 struct task_struct *new_owner;
1477 bool postunlock = false;
1478 DEFINE_RT_WAKE_Q(wqh);
1482 top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1483 if (WARN_ON_ONCE(!top_waiter)) {
1485 * As per the comment in futex_unlock_pi() this should not happen.
1487 * When this happens, give up our locks and try again, giving
1488 * the futex_lock_pi() instance time to complete, either by
1489 * waiting on the rtmutex or removing itself from the futex
1496 new_owner = top_waiter->task;
1499 * We pass it to the next owner. The WAITERS bit is always kept
1500 * enabled while there is PI state around. We cleanup the owner
1501 * died bit, because we are the owner.
1503 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1505 if (unlikely(should_fail_futex(true))) {
1510 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1511 if (!ret && (curval != uval)) {
1513 * If a unconditional UNLOCK_PI operation (user space did not
1514 * try the TID->0 transition) raced with a waiter setting the
1515 * FUTEX_WAITERS flag between get_user() and locking the hash
1516 * bucket lock, retry the operation.
1518 if ((FUTEX_TID_MASK & curval) == uval)
1526 * This is a point of no return; once we modified the uval
1527 * there is no going back and subsequent operations must
1530 pi_state_update_owner(pi_state, new_owner);
1531 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wqh);
1535 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1538 rt_mutex_postunlock(&wqh);
1544 * Express the locking dependencies for lockdep:
1547 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1550 spin_lock(&hb1->lock);
1552 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1553 } else { /* hb1 > hb2 */
1554 spin_lock(&hb2->lock);
1555 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1560 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1562 spin_unlock(&hb1->lock);
1564 spin_unlock(&hb2->lock);
1568 * Wake up waiters matching bitset queued on this futex (uaddr).
1571 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1573 struct futex_hash_bucket *hb;
1574 struct futex_q *this, *next;
1575 union futex_key key = FUTEX_KEY_INIT;
1577 DEFINE_WAKE_Q(wake_q);
1582 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1583 if (unlikely(ret != 0))
1586 hb = hash_futex(&key);
1588 /* Make sure we really have tasks to wakeup */
1589 if (!hb_waiters_pending(hb))
1592 spin_lock(&hb->lock);
1594 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1595 if (match_futex (&this->key, &key)) {
1596 if (this->pi_state || this->rt_waiter) {
1601 /* Check if one of the bits is set in both bitsets */
1602 if (!(this->bitset & bitset))
1605 mark_wake_futex(&wake_q, this);
1606 if (++ret >= nr_wake)
1611 spin_unlock(&hb->lock);
1616 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1618 unsigned int op = (encoded_op & 0x70000000) >> 28;
1619 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1620 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1621 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1624 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1625 if (oparg < 0 || oparg > 31) {
1626 char comm[sizeof(current->comm)];
1628 * kill this print and return -EINVAL when userspace
1631 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1632 get_task_comm(comm, current), oparg);
1638 pagefault_disable();
1639 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1645 case FUTEX_OP_CMP_EQ:
1646 return oldval == cmparg;
1647 case FUTEX_OP_CMP_NE:
1648 return oldval != cmparg;
1649 case FUTEX_OP_CMP_LT:
1650 return oldval < cmparg;
1651 case FUTEX_OP_CMP_GE:
1652 return oldval >= cmparg;
1653 case FUTEX_OP_CMP_LE:
1654 return oldval <= cmparg;
1655 case FUTEX_OP_CMP_GT:
1656 return oldval > cmparg;
1663 * Wake up all waiters hashed on the physical page that is mapped
1664 * to this virtual address:
1667 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1668 int nr_wake, int nr_wake2, int op)
1670 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1671 struct futex_hash_bucket *hb1, *hb2;
1672 struct futex_q *this, *next;
1674 DEFINE_WAKE_Q(wake_q);
1677 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1678 if (unlikely(ret != 0))
1680 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1681 if (unlikely(ret != 0))
1684 hb1 = hash_futex(&key1);
1685 hb2 = hash_futex(&key2);
1688 double_lock_hb(hb1, hb2);
1689 op_ret = futex_atomic_op_inuser(op, uaddr2);
1690 if (unlikely(op_ret < 0)) {
1691 double_unlock_hb(hb1, hb2);
1693 if (!IS_ENABLED(CONFIG_MMU) ||
1694 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1696 * we don't get EFAULT from MMU faults if we don't have
1697 * an MMU, but we might get them from range checking
1703 if (op_ret == -EFAULT) {
1704 ret = fault_in_user_writeable(uaddr2);
1710 if (!(flags & FLAGS_SHARED))
1715 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1716 if (match_futex (&this->key, &key1)) {
1717 if (this->pi_state || this->rt_waiter) {
1721 mark_wake_futex(&wake_q, this);
1722 if (++ret >= nr_wake)
1729 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1730 if (match_futex (&this->key, &key2)) {
1731 if (this->pi_state || this->rt_waiter) {
1735 mark_wake_futex(&wake_q, this);
1736 if (++op_ret >= nr_wake2)
1744 double_unlock_hb(hb1, hb2);
1750 * requeue_futex() - Requeue a futex_q from one hb to another
1751 * @q: the futex_q to requeue
1752 * @hb1: the source hash_bucket
1753 * @hb2: the target hash_bucket
1754 * @key2: the new key for the requeued futex_q
1757 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1758 struct futex_hash_bucket *hb2, union futex_key *key2)
1762 * If key1 and key2 hash to the same bucket, no need to
1765 if (likely(&hb1->chain != &hb2->chain)) {
1766 plist_del(&q->list, &hb1->chain);
1767 hb_waiters_dec(hb1);
1768 hb_waiters_inc(hb2);
1769 plist_add(&q->list, &hb2->chain);
1770 q->lock_ptr = &hb2->lock;
1776 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1778 * @key: the key of the requeue target futex
1779 * @hb: the hash_bucket of the requeue target futex
1781 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1782 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1783 * to the requeue target futex so the waiter can detect the wakeup on the right
1784 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1785 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1786 * to protect access to the pi_state to fixup the owner later. Must be called
1787 * with both q->lock_ptr and hb->lock held.
1790 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1791 struct futex_hash_bucket *hb)
1797 WARN_ON(!q->rt_waiter);
1798 q->rt_waiter = NULL;
1800 q->lock_ptr = &hb->lock;
1802 wake_up_state(q->task, TASK_NORMAL);
1806 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1807 * @pifutex: the user address of the to futex
1808 * @hb1: the from futex hash bucket, must be locked by the caller
1809 * @hb2: the to futex hash bucket, must be locked by the caller
1810 * @key1: the from futex key
1811 * @key2: the to futex key
1812 * @ps: address to store the pi_state pointer
1813 * @exiting: Pointer to store the task pointer of the owner task
1814 * which is in the middle of exiting
1815 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1817 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1818 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1819 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1820 * hb1 and hb2 must be held by the caller.
1822 * @exiting is only set when the return value is -EBUSY. If so, this holds
1823 * a refcount on the exiting task on return and the caller needs to drop it
1824 * after waiting for the exit to complete.
1827 * - 0 - failed to acquire the lock atomically;
1828 * - >0 - acquired the lock, return value is vpid of the top_waiter
1832 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1833 struct futex_hash_bucket *hb2, union futex_key *key1,
1834 union futex_key *key2, struct futex_pi_state **ps,
1835 struct task_struct **exiting, int set_waiters)
1837 struct futex_q *top_waiter = NULL;
1841 if (get_futex_value_locked(&curval, pifutex))
1844 if (unlikely(should_fail_futex(true)))
1848 * Find the top_waiter and determine if there are additional waiters.
1849 * If the caller intends to requeue more than 1 waiter to pifutex,
1850 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1851 * as we have means to handle the possible fault. If not, don't set
1852 * the bit unnecessarily as it will force the subsequent unlock to enter
1855 top_waiter = futex_top_waiter(hb1, key1);
1857 /* There are no waiters, nothing for us to do. */
1862 * Ensure that this is a waiter sitting in futex_wait_requeue_pi()
1863 * and waiting on the 'waitqueue' futex which is always !PI.
1865 if (!top_waiter->rt_waiter || top_waiter->pi_state)
1868 /* Ensure we requeue to the expected futex. */
1869 if (!match_futex(top_waiter->requeue_pi_key, key2))
1873 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1874 * the contended case or if set_waiters is 1. The pi_state is returned
1875 * in ps in contended cases.
1877 vpid = task_pid_vnr(top_waiter->task);
1878 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1879 exiting, set_waiters);
1881 requeue_pi_wake_futex(top_waiter, key2, hb2);
1888 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1889 * @uaddr1: source futex user address
1890 * @flags: futex flags (FLAGS_SHARED, etc.)
1891 * @uaddr2: target futex user address
1892 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1893 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1894 * @cmpval: @uaddr1 expected value (or %NULL)
1895 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1896 * pi futex (pi to pi requeue is not supported)
1898 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1899 * uaddr2 atomically on behalf of the top waiter.
1902 * - >=0 - on success, the number of tasks requeued or woken;
1905 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1906 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1907 u32 *cmpval, int requeue_pi)
1909 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1910 int task_count = 0, ret;
1911 struct futex_pi_state *pi_state = NULL;
1912 struct futex_hash_bucket *hb1, *hb2;
1913 struct futex_q *this, *next;
1914 DEFINE_WAKE_Q(wake_q);
1916 if (nr_wake < 0 || nr_requeue < 0)
1920 * When PI not supported: return -ENOSYS if requeue_pi is true,
1921 * consequently the compiler knows requeue_pi is always false past
1922 * this point which will optimize away all the conditional code
1925 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1930 * Requeue PI only works on two distinct uaddrs. This
1931 * check is only valid for private futexes. See below.
1933 if (uaddr1 == uaddr2)
1937 * futex_requeue() allows the caller to define the number
1938 * of waiters to wake up via the @nr_wake argument. With
1939 * REQUEUE_PI, waking up more than one waiter is creating
1940 * more problems than it solves. Waking up a waiter makes
1941 * only sense if the PI futex @uaddr2 is uncontended as
1942 * this allows the requeue code to acquire the futex
1943 * @uaddr2 before waking the waiter. The waiter can then
1944 * return to user space without further action. A secondary
1945 * wakeup would just make the futex_wait_requeue_pi()
1946 * handling more complex, because that code would have to
1947 * look up pi_state and do more or less all the handling
1948 * which the requeue code has to do for the to be requeued
1949 * waiters. So restrict the number of waiters to wake to
1950 * one, and only wake it up when the PI futex is
1951 * uncontended. Otherwise requeue it and let the unlock of
1952 * the PI futex handle the wakeup.
1954 * All REQUEUE_PI users, e.g. pthread_cond_signal() and
1955 * pthread_cond_broadcast() must use nr_wake=1.
1961 * requeue_pi requires a pi_state, try to allocate it now
1962 * without any locks in case it fails.
1964 if (refill_pi_state_cache())
1969 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1970 if (unlikely(ret != 0))
1972 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1973 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1974 if (unlikely(ret != 0))
1978 * The check above which compares uaddrs is not sufficient for
1979 * shared futexes. We need to compare the keys:
1981 if (requeue_pi && match_futex(&key1, &key2))
1984 hb1 = hash_futex(&key1);
1985 hb2 = hash_futex(&key2);
1988 hb_waiters_inc(hb2);
1989 double_lock_hb(hb1, hb2);
1991 if (likely(cmpval != NULL)) {
1994 ret = get_futex_value_locked(&curval, uaddr1);
1996 if (unlikely(ret)) {
1997 double_unlock_hb(hb1, hb2);
1998 hb_waiters_dec(hb2);
2000 ret = get_user(curval, uaddr1);
2004 if (!(flags & FLAGS_SHARED))
2009 if (curval != *cmpval) {
2016 struct task_struct *exiting = NULL;
2019 * Attempt to acquire uaddr2 and wake the top waiter. If we
2020 * intend to requeue waiters, force setting the FUTEX_WAITERS
2021 * bit. We force this here where we are able to easily handle
2022 * faults rather in the requeue loop below.
2024 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2026 &exiting, nr_requeue);
2029 * At this point the top_waiter has either taken uaddr2 or is
2030 * waiting on it. If the former, then the pi_state will not
2031 * exist yet, look it up one more time to ensure we have a
2032 * reference to it. If the lock was taken, @ret contains the
2033 * VPID of the top waiter task.
2034 * If the lock was not taken, we have pi_state and an initial
2035 * refcount on it. In case of an error we have nothing.
2041 * If futex_proxy_trylock_atomic() acquired the
2042 * user space futex, then the user space value
2043 * @uaddr2 has been set to the @hb1's top waiter
2044 * task VPID. This task is guaranteed to be alive
2045 * and cannot be exiting because it is either
2046 * sleeping or blocked on @hb2 lock.
2048 * The @uaddr2 futex cannot have waiters either as
2049 * otherwise futex_proxy_trylock_atomic() would not
2052 * In order to requeue waiters to @hb2, pi state is
2053 * required. Hand in the VPID value (@ret) and
2054 * allocate PI state with an initial refcount on
2057 ret = attach_to_pi_owner(uaddr2, ret, &key2, &pi_state,
2064 /* We hold a reference on the pi state. */
2067 /* If the above failed, then pi_state is NULL */
2069 double_unlock_hb(hb1, hb2);
2070 hb_waiters_dec(hb2);
2071 ret = fault_in_user_writeable(uaddr2);
2078 * Two reasons for this:
2079 * - EBUSY: Owner is exiting and we just wait for the
2081 * - EAGAIN: The user space value changed.
2083 double_unlock_hb(hb1, hb2);
2084 hb_waiters_dec(hb2);
2086 * Handle the case where the owner is in the middle of
2087 * exiting. Wait for the exit to complete otherwise
2088 * this task might loop forever, aka. live lock.
2090 wait_for_owner_exiting(ret, exiting);
2098 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2099 if (task_count - nr_wake >= nr_requeue)
2102 if (!match_futex(&this->key, &key1))
2106 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2107 * be paired with each other and no other futex ops.
2109 * We should never be requeueing a futex_q with a pi_state,
2110 * which is awaiting a futex_unlock_pi().
2112 if ((requeue_pi && !this->rt_waiter) ||
2113 (!requeue_pi && this->rt_waiter) ||
2119 /* Plain futexes just wake or requeue and are done */
2121 if (++task_count <= nr_wake)
2122 mark_wake_futex(&wake_q, this);
2124 requeue_futex(this, hb1, hb2, &key2);
2128 /* Ensure we requeue to the expected futex for requeue_pi. */
2129 if (!match_futex(this->requeue_pi_key, &key2)) {
2135 * Requeue nr_requeue waiters and possibly one more in the case
2136 * of requeue_pi if we couldn't acquire the lock atomically.
2138 * Prepare the waiter to take the rt_mutex. Take a refcount
2139 * on the pi_state and store the pointer in the futex_q
2140 * object of the waiter.
2142 get_pi_state(pi_state);
2143 this->pi_state = pi_state;
2144 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2145 this->rt_waiter, this->task);
2148 * We got the lock. We do neither drop the refcount
2149 * on pi_state nor clear this->pi_state because the
2150 * waiter needs the pi_state for cleaning up the
2151 * user space value. It will drop the refcount
2154 requeue_pi_wake_futex(this, &key2, hb2);
2159 * rt_mutex_start_proxy_lock() detected a potential
2160 * deadlock when we tried to queue that waiter.
2161 * Drop the pi_state reference which we took above
2162 * and remove the pointer to the state from the
2163 * waiters futex_q object.
2165 this->pi_state = NULL;
2166 put_pi_state(pi_state);
2168 * We stop queueing more waiters and let user space
2169 * deal with the mess.
2173 /* Waiter is queued, move it to hb2 */
2174 requeue_futex(this, hb1, hb2, &key2);
2179 * We took an extra initial reference to the pi_state either in
2180 * futex_proxy_trylock_atomic() or in attach_to_pi_owner(). We need
2181 * to drop it here again.
2183 put_pi_state(pi_state);
2186 double_unlock_hb(hb1, hb2);
2188 hb_waiters_dec(hb2);
2189 return ret ? ret : task_count;
2192 /* The key must be already stored in q->key. */
2193 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2194 __acquires(&hb->lock)
2196 struct futex_hash_bucket *hb;
2198 hb = hash_futex(&q->key);
2201 * Increment the counter before taking the lock so that
2202 * a potential waker won't miss a to-be-slept task that is
2203 * waiting for the spinlock. This is safe as all queue_lock()
2204 * users end up calling queue_me(). Similarly, for housekeeping,
2205 * decrement the counter at queue_unlock() when some error has
2206 * occurred and we don't end up adding the task to the list.
2208 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2210 q->lock_ptr = &hb->lock;
2212 spin_lock(&hb->lock);
2217 queue_unlock(struct futex_hash_bucket *hb)
2218 __releases(&hb->lock)
2220 spin_unlock(&hb->lock);
2224 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2229 * The priority used to register this element is
2230 * - either the real thread-priority for the real-time threads
2231 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2232 * - or MAX_RT_PRIO for non-RT threads.
2233 * Thus, all RT-threads are woken first in priority order, and
2234 * the others are woken last, in FIFO order.
2236 prio = min(current->normal_prio, MAX_RT_PRIO);
2238 plist_node_init(&q->list, prio);
2239 plist_add(&q->list, &hb->chain);
2244 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2245 * @q: The futex_q to enqueue
2246 * @hb: The destination hash bucket
2248 * The hb->lock must be held by the caller, and is released here. A call to
2249 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2250 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2251 * or nothing if the unqueue is done as part of the wake process and the unqueue
2252 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2255 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2256 __releases(&hb->lock)
2259 spin_unlock(&hb->lock);
2263 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2264 * @q: The futex_q to unqueue
2266 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2267 * be paired with exactly one earlier call to queue_me().
2270 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2271 * - 0 - if the futex_q was already removed by the waking thread
2273 static int unqueue_me(struct futex_q *q)
2275 spinlock_t *lock_ptr;
2278 /* In the common case we don't take the spinlock, which is nice. */
2281 * q->lock_ptr can change between this read and the following spin_lock.
2282 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2283 * optimizing lock_ptr out of the logic below.
2285 lock_ptr = READ_ONCE(q->lock_ptr);
2286 if (lock_ptr != NULL) {
2287 spin_lock(lock_ptr);
2289 * q->lock_ptr can change between reading it and
2290 * spin_lock(), causing us to take the wrong lock. This
2291 * corrects the race condition.
2293 * Reasoning goes like this: if we have the wrong lock,
2294 * q->lock_ptr must have changed (maybe several times)
2295 * between reading it and the spin_lock(). It can
2296 * change again after the spin_lock() but only if it was
2297 * already changed before the spin_lock(). It cannot,
2298 * however, change back to the original value. Therefore
2299 * we can detect whether we acquired the correct lock.
2301 if (unlikely(lock_ptr != q->lock_ptr)) {
2302 spin_unlock(lock_ptr);
2307 BUG_ON(q->pi_state);
2309 spin_unlock(lock_ptr);
2317 * PI futexes can not be requeued and must remove themselves from the
2318 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
2320 static void unqueue_me_pi(struct futex_q *q)
2324 BUG_ON(!q->pi_state);
2325 put_pi_state(q->pi_state);
2329 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2330 struct task_struct *argowner)
2332 struct futex_pi_state *pi_state = q->pi_state;
2333 struct task_struct *oldowner, *newowner;
2334 u32 uval, curval, newval, newtid;
2337 oldowner = pi_state->owner;
2340 * We are here because either:
2342 * - we stole the lock and pi_state->owner needs updating to reflect
2343 * that (@argowner == current),
2347 * - someone stole our lock and we need to fix things to point to the
2348 * new owner (@argowner == NULL).
2350 * Either way, we have to replace the TID in the user space variable.
2351 * This must be atomic as we have to preserve the owner died bit here.
2353 * Note: We write the user space value _before_ changing the pi_state
2354 * because we can fault here. Imagine swapped out pages or a fork
2355 * that marked all the anonymous memory readonly for cow.
2357 * Modifying pi_state _before_ the user space value would leave the
2358 * pi_state in an inconsistent state when we fault here, because we
2359 * need to drop the locks to handle the fault. This might be observed
2360 * in the PID checks when attaching to PI state .
2364 if (oldowner != current) {
2366 * We raced against a concurrent self; things are
2367 * already fixed up. Nothing to do.
2372 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2373 /* We got the lock. pi_state is correct. Tell caller. */
2378 * The trylock just failed, so either there is an owner or
2379 * there is a higher priority waiter than this one.
2381 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2383 * If the higher priority waiter has not yet taken over the
2384 * rtmutex then newowner is NULL. We can't return here with
2385 * that state because it's inconsistent vs. the user space
2386 * state. So drop the locks and try again. It's a valid
2387 * situation and not any different from the other retry
2390 if (unlikely(!newowner)) {
2395 WARN_ON_ONCE(argowner != current);
2396 if (oldowner == current) {
2398 * We raced against a concurrent self; things are
2399 * already fixed up. Nothing to do.
2403 newowner = argowner;
2406 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2408 if (!pi_state->owner)
2409 newtid |= FUTEX_OWNER_DIED;
2411 err = get_futex_value_locked(&uval, uaddr);
2416 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2418 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2428 * We fixed up user space. Now we need to fix the pi_state
2431 pi_state_update_owner(pi_state, newowner);
2433 return argowner == current;
2436 * In order to reschedule or handle a page fault, we need to drop the
2437 * locks here. In the case of a fault, this gives the other task
2438 * (either the highest priority waiter itself or the task which stole
2439 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2440 * are back from handling the fault we need to check the pi_state after
2441 * reacquiring the locks and before trying to do another fixup. When
2442 * the fixup has been done already we simply return.
2444 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2445 * drop hb->lock since the caller owns the hb -> futex_q relation.
2446 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2449 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2450 spin_unlock(q->lock_ptr);
2454 err = fault_in_user_writeable(uaddr);
2467 spin_lock(q->lock_ptr);
2468 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2471 * Check if someone else fixed it for us:
2473 if (pi_state->owner != oldowner)
2474 return argowner == current;
2476 /* Retry if err was -EAGAIN or the fault in succeeded */
2481 * fault_in_user_writeable() failed so user state is immutable. At
2482 * best we can make the kernel state consistent but user state will
2483 * be most likely hosed and any subsequent unlock operation will be
2484 * rejected due to PI futex rule [10].
2486 * Ensure that the rtmutex owner is also the pi_state owner despite
2487 * the user space value claiming something different. There is no
2488 * point in unlocking the rtmutex if current is the owner as it
2489 * would need to wait until the next waiter has taken the rtmutex
2490 * to guarantee consistent state. Keep it simple. Userspace asked
2491 * for this wreckaged state.
2493 * The rtmutex has an owner - either current or some other
2494 * task. See the EAGAIN loop above.
2496 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2501 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2502 struct task_struct *argowner)
2504 struct futex_pi_state *pi_state = q->pi_state;
2507 lockdep_assert_held(q->lock_ptr);
2509 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2510 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2511 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2515 static long futex_wait_restart(struct restart_block *restart);
2518 * fixup_owner() - Post lock pi_state and corner case management
2519 * @uaddr: user address of the futex
2520 * @q: futex_q (contains pi_state and access to the rt_mutex)
2521 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2523 * After attempting to lock an rt_mutex, this function is called to cleanup
2524 * the pi_state owner as well as handle race conditions that may allow us to
2525 * acquire the lock. Must be called with the hb lock held.
2528 * - 1 - success, lock taken;
2529 * - 0 - success, lock not taken;
2530 * - <0 - on error (-EFAULT)
2532 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2536 * Got the lock. We might not be the anticipated owner if we
2537 * did a lock-steal - fix up the PI-state in that case:
2539 * Speculative pi_state->owner read (we don't hold wait_lock);
2540 * since we own the lock pi_state->owner == current is the
2541 * stable state, anything else needs more attention.
2543 if (q->pi_state->owner != current)
2544 return fixup_pi_state_owner(uaddr, q, current);
2549 * If we didn't get the lock; check if anybody stole it from us. In
2550 * that case, we need to fix up the uval to point to them instead of
2551 * us, otherwise bad things happen. [10]
2553 * Another speculative read; pi_state->owner == current is unstable
2554 * but needs our attention.
2556 if (q->pi_state->owner == current)
2557 return fixup_pi_state_owner(uaddr, q, NULL);
2560 * Paranoia check. If we did not take the lock, then we should not be
2561 * the owner of the rt_mutex. Warn and establish consistent state.
2563 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2564 return fixup_pi_state_owner(uaddr, q, current);
2570 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571 * @hb: the futex hash bucket, must be locked by the caller
2572 * @q: the futex_q to queue up on
2573 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2575 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2576 struct hrtimer_sleeper *timeout)
2579 * The task state is guaranteed to be set before another task can
2580 * wake it. set_current_state() is implemented using smp_store_mb() and
2581 * queue_me() calls spin_unlock() upon completion, both serializing
2582 * access to the hash list and forcing another memory barrier.
2584 set_current_state(TASK_INTERRUPTIBLE);
2589 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2592 * If we have been removed from the hash list, then another task
2593 * has tried to wake us, and we can skip the call to schedule().
2595 if (likely(!plist_node_empty(&q->list))) {
2597 * If the timer has already expired, current will already be
2598 * flagged for rescheduling. Only call schedule if there
2599 * is no timeout, or if it has yet to expire.
2601 if (!timeout || timeout->task)
2602 freezable_schedule();
2604 __set_current_state(TASK_RUNNING);
2608 * futex_wait_setup() - Prepare to wait on a futex
2609 * @uaddr: the futex userspace address
2610 * @val: the expected value
2611 * @flags: futex flags (FLAGS_SHARED, etc.)
2612 * @q: the associated futex_q
2613 * @hb: storage for hash_bucket pointer to be returned to caller
2615 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2616 * compare it with the expected value. Handle atomic faults internally.
2617 * Return with the hb lock held on success, and unlocked on failure.
2620 * - 0 - uaddr contains val and hb has been locked;
2621 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2623 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2624 struct futex_q *q, struct futex_hash_bucket **hb)
2630 * Access the page AFTER the hash-bucket is locked.
2631 * Order is important:
2633 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2634 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2636 * The basic logical guarantee of a futex is that it blocks ONLY
2637 * if cond(var) is known to be true at the time of blocking, for
2638 * any cond. If we locked the hash-bucket after testing *uaddr, that
2639 * would open a race condition where we could block indefinitely with
2640 * cond(var) false, which would violate the guarantee.
2642 * On the other hand, we insert q and release the hash-bucket only
2643 * after testing *uaddr. This guarantees that futex_wait() will NOT
2644 * absorb a wakeup if *uaddr does not match the desired values
2645 * while the syscall executes.
2648 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2649 if (unlikely(ret != 0))
2653 *hb = queue_lock(q);
2655 ret = get_futex_value_locked(&uval, uaddr);
2660 ret = get_user(uval, uaddr);
2664 if (!(flags & FLAGS_SHARED))
2678 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2679 ktime_t *abs_time, u32 bitset)
2681 struct hrtimer_sleeper timeout, *to;
2682 struct restart_block *restart;
2683 struct futex_hash_bucket *hb;
2684 struct futex_q q = futex_q_init;
2691 to = futex_setup_timer(abs_time, &timeout, flags,
2692 current->timer_slack_ns);
2695 * Prepare to wait on uaddr. On success, it holds hb->lock and q
2698 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2702 /* queue_me and wait for wakeup, timeout, or a signal. */
2703 futex_wait_queue_me(hb, &q, to);
2705 /* If we were woken (and unqueued), we succeeded, whatever. */
2707 if (!unqueue_me(&q))
2710 if (to && !to->task)
2714 * We expect signal_pending(current), but we might be the
2715 * victim of a spurious wakeup as well.
2717 if (!signal_pending(current))
2724 restart = ¤t->restart_block;
2725 restart->futex.uaddr = uaddr;
2726 restart->futex.val = val;
2727 restart->futex.time = *abs_time;
2728 restart->futex.bitset = bitset;
2729 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2731 ret = set_restart_fn(restart, futex_wait_restart);
2735 hrtimer_cancel(&to->timer);
2736 destroy_hrtimer_on_stack(&to->timer);
2742 static long futex_wait_restart(struct restart_block *restart)
2744 u32 __user *uaddr = restart->futex.uaddr;
2745 ktime_t t, *tp = NULL;
2747 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2748 t = restart->futex.time;
2751 restart->fn = do_no_restart_syscall;
2753 return (long)futex_wait(uaddr, restart->futex.flags,
2754 restart->futex.val, tp, restart->futex.bitset);
2759 * Userspace tried a 0 -> TID atomic transition of the futex value
2760 * and failed. The kernel side here does the whole locking operation:
2761 * if there are waiters then it will block as a consequence of relying
2762 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2763 * a 0 value of the futex too.).
2765 * Also serves as futex trylock_pi()'ing, and due semantics.
2767 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2768 ktime_t *time, int trylock)
2770 struct hrtimer_sleeper timeout, *to;
2771 struct task_struct *exiting = NULL;
2772 struct rt_mutex_waiter rt_waiter;
2773 struct futex_hash_bucket *hb;
2774 struct futex_q q = futex_q_init;
2777 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2780 if (refill_pi_state_cache())
2783 to = futex_setup_timer(time, &timeout, flags, 0);
2786 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2787 if (unlikely(ret != 0))
2791 hb = queue_lock(&q);
2793 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2795 if (unlikely(ret)) {
2797 * Atomic work succeeded and we got the lock,
2798 * or failed. Either way, we do _not_ block.
2802 /* We got the lock. */
2804 goto out_unlock_put_key;
2810 * Two reasons for this:
2811 * - EBUSY: Task is exiting and we just wait for the
2813 * - EAGAIN: The user space value changed.
2817 * Handle the case where the owner is in the middle of
2818 * exiting. Wait for the exit to complete otherwise
2819 * this task might loop forever, aka. live lock.
2821 wait_for_owner_exiting(ret, exiting);
2825 goto out_unlock_put_key;
2829 WARN_ON(!q.pi_state);
2832 * Only actually queue now that the atomic ops are done:
2837 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2838 /* Fixup the trylock return value: */
2839 ret = ret ? 0 : -EWOULDBLOCK;
2843 rt_mutex_init_waiter(&rt_waiter);
2846 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2847 * hold it while doing rt_mutex_start_proxy(), because then it will
2848 * include hb->lock in the blocking chain, even through we'll not in
2849 * fact hold it while blocking. This will lead it to report -EDEADLK
2850 * and BUG when futex_unlock_pi() interleaves with this.
2852 * Therefore acquire wait_lock while holding hb->lock, but drop the
2853 * latter before calling __rt_mutex_start_proxy_lock(). This
2854 * interleaves with futex_unlock_pi() -- which does a similar lock
2855 * handoff -- such that the latter can observe the futex_q::pi_state
2856 * before __rt_mutex_start_proxy_lock() is done.
2858 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2859 spin_unlock(q.lock_ptr);
2861 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2862 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2863 * it sees the futex_q::pi_state.
2865 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2866 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2875 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2877 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2880 spin_lock(q.lock_ptr);
2882 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2883 * first acquire the hb->lock before removing the lock from the
2884 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2887 * In particular; it is important that futex_unlock_pi() can not
2888 * observe this inconsistency.
2890 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2895 * Fixup the pi_state owner and possibly acquire the lock if we
2898 res = fixup_owner(uaddr, &q, !ret);
2900 * If fixup_owner() returned an error, propagate that. If it acquired
2901 * the lock, clear our -ETIMEDOUT or -EINTR.
2904 ret = (res < 0) ? res : 0;
2907 spin_unlock(q.lock_ptr);
2915 hrtimer_cancel(&to->timer);
2916 destroy_hrtimer_on_stack(&to->timer);
2918 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2923 ret = fault_in_user_writeable(uaddr);
2927 if (!(flags & FLAGS_SHARED))
2934 * Userspace attempted a TID -> 0 atomic transition, and failed.
2935 * This is the in-kernel slowpath: we look up the PI state (if any),
2936 * and do the rt-mutex unlock.
2938 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2940 u32 curval, uval, vpid = task_pid_vnr(current);
2941 union futex_key key = FUTEX_KEY_INIT;
2942 struct futex_hash_bucket *hb;
2943 struct futex_q *top_waiter;
2946 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2950 if (get_user(uval, uaddr))
2953 * We release only a lock we actually own:
2955 if ((uval & FUTEX_TID_MASK) != vpid)
2958 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2962 hb = hash_futex(&key);
2963 spin_lock(&hb->lock);
2966 * Check waiters first. We do not trust user space values at
2967 * all and we at least want to know if user space fiddled
2968 * with the futex value instead of blindly unlocking.
2970 top_waiter = futex_top_waiter(hb, &key);
2972 struct futex_pi_state *pi_state = top_waiter->pi_state;
2979 * If current does not own the pi_state then the futex is
2980 * inconsistent and user space fiddled with the futex value.
2982 if (pi_state->owner != current)
2985 get_pi_state(pi_state);
2987 * By taking wait_lock while still holding hb->lock, we ensure
2988 * there is no point where we hold neither; and therefore
2989 * wake_futex_pi() must observe a state consistent with what we
2992 * In particular; this forces __rt_mutex_start_proxy() to
2993 * complete such that we're guaranteed to observe the
2994 * rt_waiter. Also see the WARN in wake_futex_pi().
2996 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2997 spin_unlock(&hb->lock);
2999 /* drops pi_state->pi_mutex.wait_lock */
3000 ret = wake_futex_pi(uaddr, uval, pi_state);
3002 put_pi_state(pi_state);
3005 * Success, we're done! No tricky corner cases.
3010 * The atomic access to the futex value generated a
3011 * pagefault, so retry the user-access and the wakeup:
3016 * A unconditional UNLOCK_PI op raced against a waiter
3017 * setting the FUTEX_WAITERS bit. Try again.
3022 * wake_futex_pi has detected invalid state. Tell user
3029 * We have no kernel internal state, i.e. no waiters in the
3030 * kernel. Waiters which are about to queue themselves are stuck
3031 * on hb->lock. So we can safely ignore them. We do neither
3032 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3035 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3036 spin_unlock(&hb->lock);
3051 * If uval has changed, let user space handle it.
3053 ret = (curval == uval) ? 0 : -EAGAIN;
3056 spin_unlock(&hb->lock);
3065 ret = fault_in_user_writeable(uaddr);
3073 * handle_early_requeue_pi_wakeup() - Handle early wakeup on the initial futex
3074 * @hb: the hash_bucket futex_q was original enqueued on
3075 * @q: the futex_q woken while waiting to be requeued
3076 * @timeout: the timeout associated with the wait (NULL if none)
3078 * Determine the cause for the early wakeup.
3081 * -EWOULDBLOCK or -ETIMEDOUT or -ERESTARTNOINTR
3084 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3086 struct hrtimer_sleeper *timeout)
3091 * With the hb lock held, we avoid races while we process the wakeup.
3092 * We only need to hold hb (and not hb2) to ensure atomicity as the
3093 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3094 * It can't be requeued from uaddr2 to something else since we don't
3095 * support a PI aware source futex for requeue.
3097 WARN_ON_ONCE(&hb->lock != q->lock_ptr);
3100 * We were woken prior to requeue by a timeout or a signal.
3101 * Unqueue the futex_q and determine which it was.
3103 plist_del(&q->list, &hb->chain);
3106 /* Handle spurious wakeups gracefully */
3108 if (timeout && !timeout->task)
3110 else if (signal_pending(current))
3111 ret = -ERESTARTNOINTR;
3116 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3117 * @uaddr: the futex we initially wait on (non-pi)
3118 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3119 * the same type, no requeueing from private to shared, etc.
3120 * @val: the expected value of uaddr
3121 * @abs_time: absolute timeout
3122 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3123 * @uaddr2: the pi futex we will take prior to returning to user-space
3125 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3126 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3127 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3128 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3129 * without one, the pi logic would not know which task to boost/deboost, if
3130 * there was a need to.
3132 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3133 * via the following--
3134 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3135 * 2) wakeup on uaddr2 after a requeue
3139 * If 3, cleanup and return -ERESTARTNOINTR.
3141 * If 2, we may then block on trying to take the rt_mutex and return via:
3142 * 5) successful lock
3145 * 8) other lock acquisition failure
3147 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3149 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3155 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3156 u32 val, ktime_t *abs_time, u32 bitset,
3159 struct hrtimer_sleeper timeout, *to;
3160 struct rt_mutex_waiter rt_waiter;
3161 struct futex_hash_bucket *hb;
3162 union futex_key key2 = FUTEX_KEY_INIT;
3163 struct futex_q q = futex_q_init;
3166 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3169 if (uaddr == uaddr2)
3175 to = futex_setup_timer(abs_time, &timeout, flags,
3176 current->timer_slack_ns);
3179 * The waiter is allocated on our stack, manipulated by the requeue
3180 * code while we sleep on uaddr.
3182 rt_mutex_init_waiter(&rt_waiter);
3184 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3185 if (unlikely(ret != 0))
3189 q.rt_waiter = &rt_waiter;
3190 q.requeue_pi_key = &key2;
3193 * Prepare to wait on uaddr. On success, it holds hb->lock and q
3196 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3201 * The check above which compares uaddrs is not sufficient for
3202 * shared futexes. We need to compare the keys:
3204 if (match_futex(&q.key, &key2)) {
3210 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3211 futex_wait_queue_me(hb, &q, to);
3213 spin_lock(&hb->lock);
3214 /* Is @q still queued on uaddr1? */
3215 if (!match_futex(&q->key, key2))
3216 ret = handle_early_requeue_pi_wakeup(hb, &q, to);
3217 spin_unlock(&hb->lock);
3222 * In order for us to be here, we know our q.key == key2, and since
3223 * we took the hb->lock above, we also know that futex_requeue() has
3224 * completed and we no longer have to concern ourselves with a wakeup
3225 * race with the atomic proxy lock acquisition by the requeue code.
3229 * Check if the requeue code acquired the second futex for us and do
3230 * any pertinent fixup.
3233 if (q.pi_state && (q.pi_state->owner != current)) {
3234 spin_lock(q.lock_ptr);
3235 ret = fixup_owner(uaddr2, &q, true);
3237 * Drop the reference to the pi state which
3238 * the requeue_pi() code acquired for us.
3240 put_pi_state(q.pi_state);
3241 spin_unlock(q.lock_ptr);
3243 * Adjust the return value. It's either -EFAULT or
3244 * success (1) but the caller expects 0 for success.
3246 ret = ret < 0 ? ret : 0;
3249 struct rt_mutex_base *pi_mutex;
3252 * We have been woken up by futex_unlock_pi(), a timeout, or a
3253 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3256 WARN_ON(!q.pi_state);
3257 pi_mutex = &q.pi_state->pi_mutex;
3258 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3260 spin_lock(q.lock_ptr);
3261 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3264 debug_rt_mutex_free_waiter(&rt_waiter);
3266 * Fixup the pi_state owner and possibly acquire the lock if we
3269 res = fixup_owner(uaddr2, &q, !ret);
3271 * If fixup_owner() returned an error, propagate that. If it
3272 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3275 ret = (res < 0) ? res : 0;
3278 spin_unlock(q.lock_ptr);
3281 if (ret == -EINTR) {
3283 * We've already been requeued, but cannot restart by calling
3284 * futex_lock_pi() directly. We could restart this syscall, but
3285 * it would detect that the user space "val" changed and return
3286 * -EWOULDBLOCK. Save the overhead of the restart and return
3287 * -EWOULDBLOCK directly.
3294 hrtimer_cancel(&to->timer);
3295 destroy_hrtimer_on_stack(&to->timer);
3301 * Support for robust futexes: the kernel cleans up held futexes at
3304 * Implementation: user-space maintains a per-thread list of locks it
3305 * is holding. Upon do_exit(), the kernel carefully walks this list,
3306 * and marks all locks that are owned by this thread with the
3307 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3308 * always manipulated with the lock held, so the list is private and
3309 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3310 * field, to allow the kernel to clean up if the thread dies after
3311 * acquiring the lock, but just before it could have added itself to
3312 * the list. There can only be one such pending lock.
3316 * sys_set_robust_list() - Set the robust-futex list head of a task
3317 * @head: pointer to the list-head
3318 * @len: length of the list-head, as userspace expects
3320 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3323 if (!futex_cmpxchg_enabled)
3326 * The kernel knows only one size for now:
3328 if (unlikely(len != sizeof(*head)))
3331 current->robust_list = head;
3337 * sys_get_robust_list() - Get the robust-futex list head of a task
3338 * @pid: pid of the process [zero for current task]
3339 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3340 * @len_ptr: pointer to a length field, the kernel fills in the header size
3342 SYSCALL_DEFINE3(get_robust_list, int, pid,
3343 struct robust_list_head __user * __user *, head_ptr,
3344 size_t __user *, len_ptr)
3346 struct robust_list_head __user *head;
3348 struct task_struct *p;
3350 if (!futex_cmpxchg_enabled)
3359 p = find_task_by_vpid(pid);
3365 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3368 head = p->robust_list;
3371 if (put_user(sizeof(*head), len_ptr))
3373 return put_user(head, head_ptr);
3381 /* Constants for the pending_op argument of handle_futex_death */
3382 #define HANDLE_DEATH_PENDING true
3383 #define HANDLE_DEATH_LIST false
3386 * Process a futex-list entry, check whether it's owned by the
3387 * dying task, and do notification if so:
3389 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3390 bool pi, bool pending_op)
3392 u32 uval, nval, mval;
3395 /* Futex address must be 32bit aligned */
3396 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3400 if (get_user(uval, uaddr))
3404 * Special case for regular (non PI) futexes. The unlock path in
3405 * user space has two race scenarios:
3407 * 1. The unlock path releases the user space futex value and
3408 * before it can execute the futex() syscall to wake up
3409 * waiters it is killed.
3411 * 2. A woken up waiter is killed before it can acquire the
3412 * futex in user space.
3414 * In both cases the TID validation below prevents a wakeup of
3415 * potential waiters which can cause these waiters to block
3418 * In both cases the following conditions are met:
3420 * 1) task->robust_list->list_op_pending != NULL
3421 * @pending_op == true
3422 * 2) User space futex value == 0
3423 * 3) Regular futex: @pi == false
3425 * If these conditions are met, it is safe to attempt waking up a
3426 * potential waiter without touching the user space futex value and
3427 * trying to set the OWNER_DIED bit. The user space futex value is
3428 * uncontended and the rest of the user space mutex state is
3429 * consistent, so a woken waiter will just take over the
3430 * uncontended futex. Setting the OWNER_DIED bit would create
3431 * inconsistent state and malfunction of the user space owner died
3434 if (pending_op && !pi && !uval) {
3435 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3439 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3443 * Ok, this dying thread is truly holding a futex
3444 * of interest. Set the OWNER_DIED bit atomically
3445 * via cmpxchg, and if the value had FUTEX_WAITERS
3446 * set, wake up a waiter (if any). (We have to do a
3447 * futex_wake() even if OWNER_DIED is already set -
3448 * to handle the rare but possible case of recursive
3449 * thread-death.) The rest of the cleanup is done in
3452 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3455 * We are not holding a lock here, but we want to have
3456 * the pagefault_disable/enable() protection because
3457 * we want to handle the fault gracefully. If the
3458 * access fails we try to fault in the futex with R/W
3459 * verification via get_user_pages. get_user() above
3460 * does not guarantee R/W access. If that fails we
3461 * give up and leave the futex locked.
3463 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3466 if (fault_in_user_writeable(uaddr))
3484 * Wake robust non-PI futexes here. The wakeup of
3485 * PI futexes happens in exit_pi_state():
3487 if (!pi && (uval & FUTEX_WAITERS))
3488 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3494 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3496 static inline int fetch_robust_entry(struct robust_list __user **entry,
3497 struct robust_list __user * __user *head,
3500 unsigned long uentry;
3502 if (get_user(uentry, (unsigned long __user *)head))
3505 *entry = (void __user *)(uentry & ~1UL);
3512 * Walk curr->robust_list (very carefully, it's a userspace list!)
3513 * and mark any locks found there dead, and notify any waiters.
3515 * We silently return on any sign of list-walking problem.
3517 static void exit_robust_list(struct task_struct *curr)
3519 struct robust_list_head __user *head = curr->robust_list;
3520 struct robust_list __user *entry, *next_entry, *pending;
3521 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3522 unsigned int next_pi;
3523 unsigned long futex_offset;
3526 if (!futex_cmpxchg_enabled)
3530 * Fetch the list head (which was registered earlier, via
3531 * sys_set_robust_list()):
3533 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3536 * Fetch the relative futex offset:
3538 if (get_user(futex_offset, &head->futex_offset))
3541 * Fetch any possibly pending lock-add first, and handle it
3544 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3547 next_entry = NULL; /* avoid warning with gcc */
3548 while (entry != &head->list) {
3550 * Fetch the next entry in the list before calling
3551 * handle_futex_death:
3553 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3555 * A pending lock might already be on the list, so
3556 * don't process it twice:
3558 if (entry != pending) {
3559 if (handle_futex_death((void __user *)entry + futex_offset,
3560 curr, pi, HANDLE_DEATH_LIST))
3568 * Avoid excessively long or circular lists:
3577 handle_futex_death((void __user *)pending + futex_offset,
3578 curr, pip, HANDLE_DEATH_PENDING);
3582 static void futex_cleanup(struct task_struct *tsk)
3584 if (unlikely(tsk->robust_list)) {
3585 exit_robust_list(tsk);
3586 tsk->robust_list = NULL;
3589 #ifdef CONFIG_COMPAT
3590 if (unlikely(tsk->compat_robust_list)) {
3591 compat_exit_robust_list(tsk);
3592 tsk->compat_robust_list = NULL;
3596 if (unlikely(!list_empty(&tsk->pi_state_list)))
3597 exit_pi_state_list(tsk);
3601 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3602 * @tsk: task to set the state on
3604 * Set the futex exit state of the task lockless. The futex waiter code
3605 * observes that state when a task is exiting and loops until the task has
3606 * actually finished the futex cleanup. The worst case for this is that the
3607 * waiter runs through the wait loop until the state becomes visible.
3609 * This is called from the recursive fault handling path in do_exit().
3611 * This is best effort. Either the futex exit code has run already or
3612 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3613 * take it over. If not, the problem is pushed back to user space. If the
3614 * futex exit code did not run yet, then an already queued waiter might
3615 * block forever, but there is nothing which can be done about that.
3617 void futex_exit_recursive(struct task_struct *tsk)
3619 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3620 if (tsk->futex_state == FUTEX_STATE_EXITING)
3621 mutex_unlock(&tsk->futex_exit_mutex);
3622 tsk->futex_state = FUTEX_STATE_DEAD;
3625 static void futex_cleanup_begin(struct task_struct *tsk)
3628 * Prevent various race issues against a concurrent incoming waiter
3629 * including live locks by forcing the waiter to block on
3630 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3631 * attach_to_pi_owner().
3633 mutex_lock(&tsk->futex_exit_mutex);
3636 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3638 * This ensures that all subsequent checks of tsk->futex_state in
3639 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3640 * tsk->pi_lock held.
3642 * It guarantees also that a pi_state which was queued right before
3643 * the state change under tsk->pi_lock by a concurrent waiter must
3644 * be observed in exit_pi_state_list().
3646 raw_spin_lock_irq(&tsk->pi_lock);
3647 tsk->futex_state = FUTEX_STATE_EXITING;
3648 raw_spin_unlock_irq(&tsk->pi_lock);
3651 static void futex_cleanup_end(struct task_struct *tsk, int state)
3654 * Lockless store. The only side effect is that an observer might
3655 * take another loop until it becomes visible.
3657 tsk->futex_state = state;
3659 * Drop the exit protection. This unblocks waiters which observed
3660 * FUTEX_STATE_EXITING to reevaluate the state.
3662 mutex_unlock(&tsk->futex_exit_mutex);
3665 void futex_exec_release(struct task_struct *tsk)
3668 * The state handling is done for consistency, but in the case of
3669 * exec() there is no way to prevent further damage as the PID stays
3670 * the same. But for the unlikely and arguably buggy case that a
3671 * futex is held on exec(), this provides at least as much state
3672 * consistency protection which is possible.
3674 futex_cleanup_begin(tsk);
3677 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3678 * exec a new binary.
3680 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3683 void futex_exit_release(struct task_struct *tsk)
3685 futex_cleanup_begin(tsk);
3687 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3690 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3691 u32 __user *uaddr2, u32 val2, u32 val3)
3693 int cmd = op & FUTEX_CMD_MASK;
3694 unsigned int flags = 0;
3696 if (!(op & FUTEX_PRIVATE_FLAG))
3697 flags |= FLAGS_SHARED;
3699 if (op & FUTEX_CLOCK_REALTIME) {
3700 flags |= FLAGS_CLOCKRT;
3701 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI &&
3702 cmd != FUTEX_LOCK_PI2)
3708 case FUTEX_LOCK_PI2:
3709 case FUTEX_UNLOCK_PI:
3710 case FUTEX_TRYLOCK_PI:
3711 case FUTEX_WAIT_REQUEUE_PI:
3712 case FUTEX_CMP_REQUEUE_PI:
3713 if (!futex_cmpxchg_enabled)
3719 val3 = FUTEX_BITSET_MATCH_ANY;
3721 case FUTEX_WAIT_BITSET:
3722 return futex_wait(uaddr, flags, val, timeout, val3);
3724 val3 = FUTEX_BITSET_MATCH_ANY;
3726 case FUTEX_WAKE_BITSET:
3727 return futex_wake(uaddr, flags, val, val3);
3729 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3730 case FUTEX_CMP_REQUEUE:
3731 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3733 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3735 flags |= FLAGS_CLOCKRT;
3737 case FUTEX_LOCK_PI2:
3738 return futex_lock_pi(uaddr, flags, timeout, 0);
3739 case FUTEX_UNLOCK_PI:
3740 return futex_unlock_pi(uaddr, flags);
3741 case FUTEX_TRYLOCK_PI:
3742 return futex_lock_pi(uaddr, flags, NULL, 1);
3743 case FUTEX_WAIT_REQUEUE_PI:
3744 val3 = FUTEX_BITSET_MATCH_ANY;
3745 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3747 case FUTEX_CMP_REQUEUE_PI:
3748 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3753 static __always_inline bool futex_cmd_has_timeout(u32 cmd)
3758 case FUTEX_LOCK_PI2:
3759 case FUTEX_WAIT_BITSET:
3760 case FUTEX_WAIT_REQUEUE_PI:
3766 static __always_inline int
3767 futex_init_timeout(u32 cmd, u32 op, struct timespec64 *ts, ktime_t *t)
3769 if (!timespec64_valid(ts))
3772 *t = timespec64_to_ktime(*ts);
3773 if (cmd == FUTEX_WAIT)
3774 *t = ktime_add_safe(ktime_get(), *t);
3775 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3776 *t = timens_ktime_to_host(CLOCK_MONOTONIC, *t);
3780 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3781 const struct __kernel_timespec __user *, utime,
3782 u32 __user *, uaddr2, u32, val3)
3784 int ret, cmd = op & FUTEX_CMD_MASK;
3785 ktime_t t, *tp = NULL;
3786 struct timespec64 ts;
3788 if (utime && futex_cmd_has_timeout(cmd)) {
3789 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3791 if (get_timespec64(&ts, utime))
3793 ret = futex_init_timeout(cmd, op, &ts, &t);
3799 return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3802 #ifdef CONFIG_COMPAT
3804 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3807 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3808 compat_uptr_t __user *head, unsigned int *pi)
3810 if (get_user(*uentry, head))
3813 *entry = compat_ptr((*uentry) & ~1);
3814 *pi = (unsigned int)(*uentry) & 1;
3819 static void __user *futex_uaddr(struct robust_list __user *entry,
3820 compat_long_t futex_offset)
3822 compat_uptr_t base = ptr_to_compat(entry);
3823 void __user *uaddr = compat_ptr(base + futex_offset);
3829 * Walk curr->robust_list (very carefully, it's a userspace list!)
3830 * and mark any locks found there dead, and notify any waiters.
3832 * We silently return on any sign of list-walking problem.
3834 static void compat_exit_robust_list(struct task_struct *curr)
3836 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3837 struct robust_list __user *entry, *next_entry, *pending;
3838 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3839 unsigned int next_pi;
3840 compat_uptr_t uentry, next_uentry, upending;
3841 compat_long_t futex_offset;
3844 if (!futex_cmpxchg_enabled)
3848 * Fetch the list head (which was registered earlier, via
3849 * sys_set_robust_list()):
3851 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3854 * Fetch the relative futex offset:
3856 if (get_user(futex_offset, &head->futex_offset))
3859 * Fetch any possibly pending lock-add first, and handle it
3862 if (compat_fetch_robust_entry(&upending, &pending,
3863 &head->list_op_pending, &pip))
3866 next_entry = NULL; /* avoid warning with gcc */
3867 while (entry != (struct robust_list __user *) &head->list) {
3869 * Fetch the next entry in the list before calling
3870 * handle_futex_death:
3872 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3873 (compat_uptr_t __user *)&entry->next, &next_pi);
3875 * A pending lock might already be on the list, so
3876 * dont process it twice:
3878 if (entry != pending) {
3879 void __user *uaddr = futex_uaddr(entry, futex_offset);
3881 if (handle_futex_death(uaddr, curr, pi,
3887 uentry = next_uentry;
3891 * Avoid excessively long or circular lists:
3899 void __user *uaddr = futex_uaddr(pending, futex_offset);
3901 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3905 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3906 struct compat_robust_list_head __user *, head,
3909 if (!futex_cmpxchg_enabled)
3912 if (unlikely(len != sizeof(*head)))
3915 current->compat_robust_list = head;
3920 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3921 compat_uptr_t __user *, head_ptr,
3922 compat_size_t __user *, len_ptr)
3924 struct compat_robust_list_head __user *head;
3926 struct task_struct *p;
3928 if (!futex_cmpxchg_enabled)
3937 p = find_task_by_vpid(pid);
3943 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3946 head = p->compat_robust_list;
3949 if (put_user(sizeof(*head), len_ptr))
3951 return put_user(ptr_to_compat(head), head_ptr);
3958 #endif /* CONFIG_COMPAT */
3960 #ifdef CONFIG_COMPAT_32BIT_TIME
3961 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3962 const struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3965 int ret, cmd = op & FUTEX_CMD_MASK;
3966 ktime_t t, *tp = NULL;
3967 struct timespec64 ts;
3969 if (utime && futex_cmd_has_timeout(cmd)) {
3970 if (get_old_timespec32(&ts, utime))
3972 ret = futex_init_timeout(cmd, op, &ts, &t);
3978 return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3980 #endif /* CONFIG_COMPAT_32BIT_TIME */
3982 static void __init futex_detect_cmpxchg(void)
3984 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3988 * This will fail and we want it. Some arch implementations do
3989 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3990 * functionality. We want to know that before we call in any
3991 * of the complex code paths. Also we want to prevent
3992 * registration of robust lists in that case. NULL is
3993 * guaranteed to fault and we get -EFAULT on functional
3994 * implementation, the non-functional ones will return
3997 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3998 futex_cmpxchg_enabled = 1;
4002 static int __init futex_init(void)
4004 unsigned int futex_shift;
4007 #if CONFIG_BASE_SMALL
4008 futex_hashsize = 16;
4010 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4013 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4015 futex_hashsize < 256 ? HASH_SMALL : 0,
4017 futex_hashsize, futex_hashsize);
4018 futex_hashsize = 1UL << futex_shift;
4020 futex_detect_cmpxchg();
4022 for (i = 0; i < futex_hashsize; i++) {
4023 atomic_set(&futex_queues[i].waiters, 0);
4024 plist_head_init(&futex_queues[i].chain);
4025 spin_lock_init(&futex_queues[i].lock);
4030 core_initcall(futex_init);