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/hugetlb.h>
39 #include <linux/freezer.h>
40 #include <linux/memblock.h>
41 #include <linux/fault-inject.h>
42 #include <linux/time_namespace.h>
44 #include <asm/futex.h>
46 #include "locking/rtmutex_common.h"
49 * READ this before attempting to hack on futexes!
51 * Basic futex operation and ordering guarantees
52 * =============================================
54 * The waiter reads the futex value in user space and calls
55 * futex_wait(). This function computes the hash bucket and acquires
56 * the hash bucket lock. After that it reads the futex user space value
57 * again and verifies that the data has not changed. If it has not changed
58 * it enqueues itself into the hash bucket, releases the hash bucket lock
61 * The waker side modifies the user space value of the futex and calls
62 * futex_wake(). This function computes the hash bucket and acquires the
63 * hash bucket lock. Then it looks for waiters on that futex in the hash
64 * bucket and wakes them.
66 * In futex wake up scenarios where no tasks are blocked on a futex, taking
67 * the hb spinlock can be avoided and simply return. In order for this
68 * optimization to work, ordering guarantees must exist so that the waiter
69 * being added to the list is acknowledged when the list is concurrently being
70 * checked by the waker, avoiding scenarios like the following:
74 * sys_futex(WAIT, futex, val);
75 * futex_wait(futex, val);
78 * sys_futex(WAKE, futex);
83 * lock(hash_bucket(futex));
85 * unlock(hash_bucket(futex));
88 * This would cause the waiter on CPU 0 to wait forever because it
89 * missed the transition of the user space value from val to newval
90 * and the waker did not find the waiter in the hash bucket queue.
92 * The correct serialization ensures that a waiter either observes
93 * the changed user space value before blocking or is woken by a
98 * sys_futex(WAIT, futex, val);
99 * futex_wait(futex, val);
102 * smp_mb(); (A) <-- paired with -.
104 * lock(hash_bucket(futex)); |
108 * | sys_futex(WAKE, futex);
109 * | futex_wake(futex);
111 * `--------> smp_mb(); (B)
114 * unlock(hash_bucket(futex));
115 * schedule(); if (waiters)
116 * lock(hash_bucket(futex));
117 * else wake_waiters(futex);
118 * waiters--; (b) unlock(hash_bucket(futex));
120 * Where (A) orders the waiters increment and the futex value read through
121 * atomic operations (see hb_waiters_inc) and where (B) orders the write
122 * to futex and the waiters read (see hb_waiters_pending()).
124 * This yields the following case (where X:=waiters, Y:=futex):
132 * Which guarantees that x==0 && y==0 is impossible; which translates back into
133 * the guarantee that we cannot both miss the futex variable change and the
136 * Note that a new waiter is accounted for in (a) even when it is possible that
137 * the wait call can return error, in which case we backtrack from it in (b).
138 * Refer to the comment in queue_lock().
140 * Similarly, in order to account for waiters being requeued on another
141 * address we always increment the waiters for the destination bucket before
142 * acquiring the lock. It then decrements them again after releasing it -
143 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
144 * will do the additional required waiter count housekeeping. This is done for
145 * double_lock_hb() and double_unlock_hb(), respectively.
148 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
149 #define futex_cmpxchg_enabled 1
151 static int __read_mostly futex_cmpxchg_enabled;
155 * Futex flags used to encode options to functions and preserve them across
159 # define FLAGS_SHARED 0x01
162 * NOMMU does not have per process address space. Let the compiler optimize
165 # define FLAGS_SHARED 0x00
167 #define FLAGS_CLOCKRT 0x02
168 #define FLAGS_HAS_TIMEOUT 0x04
171 * Priority Inheritance state:
173 struct futex_pi_state {
175 * list of 'owned' pi_state instances - these have to be
176 * cleaned up in do_exit() if the task exits prematurely:
178 struct list_head list;
183 struct rt_mutex pi_mutex;
185 struct task_struct *owner;
189 } __randomize_layout;
192 * struct futex_q - The hashed futex queue entry, one per waiting task
193 * @list: priority-sorted list of tasks waiting on this futex
194 * @task: the task waiting on the futex
195 * @lock_ptr: the hash bucket lock
196 * @key: the key the futex is hashed on
197 * @pi_state: optional priority inheritance state
198 * @rt_waiter: rt_waiter storage for use with requeue_pi
199 * @requeue_pi_key: the requeue_pi target futex key
200 * @bitset: bitset for the optional bitmasked wakeup
202 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
203 * we can wake only the relevant ones (hashed queues may be shared).
205 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
206 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
207 * The order of wakeup is always to make the first condition true, then
210 * PI futexes are typically woken before they are removed from the hash list via
211 * the rt_mutex code. See unqueue_me_pi().
214 struct plist_node list;
216 struct task_struct *task;
217 spinlock_t *lock_ptr;
219 struct futex_pi_state *pi_state;
220 struct rt_mutex_waiter *rt_waiter;
221 union futex_key *requeue_pi_key;
223 } __randomize_layout;
225 static const struct futex_q futex_q_init = {
226 /* list gets initialized in queue_me()*/
227 .key = FUTEX_KEY_INIT,
228 .bitset = FUTEX_BITSET_MATCH_ANY
232 * Hash buckets are shared by all the futex_keys that hash to the same
233 * location. Each key may have multiple futex_q structures, one for each task
234 * waiting on a futex.
236 struct futex_hash_bucket {
239 struct plist_head chain;
240 } ____cacheline_aligned_in_smp;
243 * The base of the bucket array and its size are always used together
244 * (after initialization only in hash_futex()), so ensure that they
245 * reside in the same cacheline.
248 struct futex_hash_bucket *queues;
249 unsigned long hashsize;
250 } __futex_data __read_mostly __aligned(2*sizeof(long));
251 #define futex_queues (__futex_data.queues)
252 #define futex_hashsize (__futex_data.hashsize)
256 * Fault injections for futexes.
258 #ifdef CONFIG_FAIL_FUTEX
261 struct fault_attr attr;
265 .attr = FAULT_ATTR_INITIALIZER,
266 .ignore_private = false,
269 static int __init setup_fail_futex(char *str)
271 return setup_fault_attr(&fail_futex.attr, str);
273 __setup("fail_futex=", setup_fail_futex);
275 static bool should_fail_futex(bool fshared)
277 if (fail_futex.ignore_private && !fshared)
280 return should_fail(&fail_futex.attr, 1);
283 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
285 static int __init fail_futex_debugfs(void)
287 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
290 dir = fault_create_debugfs_attr("fail_futex", NULL,
295 debugfs_create_bool("ignore-private", mode, dir,
296 &fail_futex.ignore_private);
300 late_initcall(fail_futex_debugfs);
302 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
305 static inline bool should_fail_futex(bool fshared)
309 #endif /* CONFIG_FAIL_FUTEX */
312 static void compat_exit_robust_list(struct task_struct *curr);
314 static inline void compat_exit_robust_list(struct task_struct *curr) { }
318 * Reflects a new waiter being added to the waitqueue.
320 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
323 atomic_inc(&hb->waiters);
325 * Full barrier (A), see the ordering comment above.
327 smp_mb__after_atomic();
332 * Reflects a waiter being removed from the waitqueue by wakeup
335 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
338 atomic_dec(&hb->waiters);
342 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
346 * Full barrier (B), see the ordering comment above.
349 return atomic_read(&hb->waiters);
356 * hash_futex - Return the hash bucket in the global hash
357 * @key: Pointer to the futex key for which the hash is calculated
359 * We hash on the keys returned from get_futex_key (see below) and return the
360 * corresponding hash bucket in the global hash.
362 static struct futex_hash_bucket *hash_futex(union futex_key *key)
364 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
367 return &futex_queues[hash & (futex_hashsize - 1)];
372 * match_futex - Check whether two futex keys are equal
373 * @key1: Pointer to key1
374 * @key2: Pointer to key2
376 * Return 1 if two futex_keys are equal, 0 otherwise.
378 static inline int match_futex(union futex_key *key1, union futex_key *key2)
381 && key1->both.word == key2->both.word
382 && key1->both.ptr == key2->both.ptr
383 && key1->both.offset == key2->both.offset);
392 * futex_setup_timer - set up the sleeping hrtimer.
393 * @time: ptr to the given timeout value
394 * @timeout: the hrtimer_sleeper structure to be set up
395 * @flags: futex flags
396 * @range_ns: optional range in ns
398 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
401 static inline struct hrtimer_sleeper *
402 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
403 int flags, u64 range_ns)
408 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
409 CLOCK_REALTIME : CLOCK_MONOTONIC,
412 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
413 * effectively the same as calling hrtimer_set_expires().
415 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
421 * Generate a machine wide unique identifier for this inode.
423 * This relies on u64 not wrapping in the life-time of the machine; which with
424 * 1ns resolution means almost 585 years.
426 * This further relies on the fact that a well formed program will not unmap
427 * the file while it has a (shared) futex waiting on it. This mapping will have
428 * a file reference which pins the mount and inode.
430 * If for some reason an inode gets evicted and read back in again, it will get
431 * a new sequence number and will _NOT_ match, even though it is the exact same
434 * It is important that match_futex() will never have a false-positive, esp.
435 * for PI futexes that can mess up the state. The above argues that false-negatives
436 * are only possible for malformed programs.
438 static u64 get_inode_sequence_number(struct inode *inode)
440 static atomic64_t i_seq;
443 /* Does the inode already have a sequence number? */
444 old = atomic64_read(&inode->i_sequence);
449 u64 new = atomic64_add_return(1, &i_seq);
450 if (WARN_ON_ONCE(!new))
453 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
461 * get_futex_key() - Get parameters which are the keys for a futex
462 * @uaddr: virtual address of the futex
463 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
464 * @key: address where result is stored.
465 * @rw: mapping needs to be read/write (values: FUTEX_READ,
468 * Return: a negative error code or 0
470 * The key words are stored in @key on success.
472 * For shared mappings (when @fshared), the key is:
474 * ( inode->i_sequence, page->index, offset_within_page )
476 * [ also see get_inode_sequence_number() ]
478 * For private mappings (or when !@fshared), the key is:
480 * ( current->mm, address, 0 )
482 * This allows (cross process, where applicable) identification of the futex
483 * without keeping the page pinned for the duration of the FUTEX_WAIT.
485 * lock_page() might sleep, the caller should not hold a spinlock.
487 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
488 enum futex_access rw)
490 unsigned long address = (unsigned long)uaddr;
491 struct mm_struct *mm = current->mm;
492 struct page *page, *tail;
493 struct address_space *mapping;
497 * The futex address must be "naturally" aligned.
499 key->both.offset = address % PAGE_SIZE;
500 if (unlikely((address % sizeof(u32)) != 0))
502 address -= key->both.offset;
504 if (unlikely(!access_ok(uaddr, sizeof(u32))))
507 if (unlikely(should_fail_futex(fshared)))
511 * PROCESS_PRIVATE futexes are fast.
512 * As the mm cannot disappear under us and the 'key' only needs
513 * virtual address, we dont even have to find the underlying vma.
514 * Note : We do have to check 'uaddr' is a valid user address,
515 * but access_ok() should be faster than find_vma()
518 key->private.mm = mm;
519 key->private.address = address;
524 /* Ignore any VERIFY_READ mapping (futex common case) */
525 if (unlikely(should_fail_futex(true)))
528 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
530 * If write access is not required (eg. FUTEX_WAIT), try
531 * and get read-only access.
533 if (err == -EFAULT && rw == FUTEX_READ) {
534 err = get_user_pages_fast(address, 1, 0, &page);
543 * The treatment of mapping from this point on is critical. The page
544 * lock protects many things but in this context the page lock
545 * stabilizes mapping, prevents inode freeing in the shared
546 * file-backed region case and guards against movement to swap cache.
548 * Strictly speaking the page lock is not needed in all cases being
549 * considered here and page lock forces unnecessarily serialization
550 * From this point on, mapping will be re-verified if necessary and
551 * page lock will be acquired only if it is unavoidable
553 * Mapping checks require the head page for any compound page so the
554 * head page and mapping is looked up now. For anonymous pages, it
555 * does not matter if the page splits in the future as the key is
556 * based on the address. For filesystem-backed pages, the tail is
557 * required as the index of the page determines the key. For
558 * base pages, there is no tail page and tail == page.
561 page = compound_head(page);
562 mapping = READ_ONCE(page->mapping);
565 * If page->mapping is NULL, then it cannot be a PageAnon
566 * page; but it might be the ZERO_PAGE or in the gate area or
567 * in a special mapping (all cases which we are happy to fail);
568 * or it may have been a good file page when get_user_pages_fast
569 * found it, but truncated or holepunched or subjected to
570 * invalidate_complete_page2 before we got the page lock (also
571 * cases which we are happy to fail). And we hold a reference,
572 * so refcount care in invalidate_complete_page's remove_mapping
573 * prevents drop_caches from setting mapping to NULL beneath us.
575 * The case we do have to guard against is when memory pressure made
576 * shmem_writepage move it from filecache to swapcache beneath us:
577 * an unlikely race, but we do need to retry for page->mapping.
579 if (unlikely(!mapping)) {
583 * Page lock is required to identify which special case above
584 * applies. If this is really a shmem page then the page lock
585 * will prevent unexpected transitions.
588 shmem_swizzled = PageSwapCache(page) || page->mapping;
599 * Private mappings are handled in a simple way.
601 * If the futex key is stored on an anonymous page, then the associated
602 * object is the mm which is implicitly pinned by the calling process.
604 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
605 * it's a read-only handle, it's expected that futexes attach to
606 * the object not the particular process.
608 if (PageAnon(page)) {
610 * A RO anonymous page will never change and thus doesn't make
611 * sense for futex operations.
613 if (unlikely(should_fail_futex(true)) || ro) {
618 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
619 key->private.mm = mm;
620 key->private.address = address;
626 * The associated futex object in this case is the inode and
627 * the page->mapping must be traversed. Ordinarily this should
628 * be stabilised under page lock but it's not strictly
629 * necessary in this case as we just want to pin the inode, not
630 * update the radix tree or anything like that.
632 * The RCU read lock is taken as the inode is finally freed
633 * under RCU. If the mapping still matches expectations then the
634 * mapping->host can be safely accessed as being a valid inode.
638 if (READ_ONCE(page->mapping) != mapping) {
645 inode = READ_ONCE(mapping->host);
653 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
654 key->shared.i_seq = get_inode_sequence_number(inode);
655 key->shared.pgoff = basepage_index(tail);
665 * fault_in_user_writeable() - Fault in user address and verify RW access
666 * @uaddr: pointer to faulting user space address
668 * Slow path to fixup the fault we just took in the atomic write
671 * We have no generic implementation of a non-destructive write to the
672 * user address. We know that we faulted in the atomic pagefault
673 * disabled section so we can as well avoid the #PF overhead by
674 * calling get_user_pages() right away.
676 static int fault_in_user_writeable(u32 __user *uaddr)
678 struct mm_struct *mm = current->mm;
682 ret = fixup_user_fault(mm, (unsigned long)uaddr,
683 FAULT_FLAG_WRITE, NULL);
684 mmap_read_unlock(mm);
686 return ret < 0 ? ret : 0;
690 * futex_top_waiter() - Return the highest priority waiter on a futex
691 * @hb: the hash bucket the futex_q's reside in
692 * @key: the futex key (to distinguish it from other futex futex_q's)
694 * Must be called with the hb lock held.
696 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
697 union futex_key *key)
699 struct futex_q *this;
701 plist_for_each_entry(this, &hb->chain, list) {
702 if (match_futex(&this->key, key))
708 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
709 u32 uval, u32 newval)
714 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
720 static int get_futex_value_locked(u32 *dest, u32 __user *from)
725 ret = __get_user(*dest, from);
728 return ret ? -EFAULT : 0;
735 static int refill_pi_state_cache(void)
737 struct futex_pi_state *pi_state;
739 if (likely(current->pi_state_cache))
742 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
747 INIT_LIST_HEAD(&pi_state->list);
748 /* pi_mutex gets initialized later */
749 pi_state->owner = NULL;
750 refcount_set(&pi_state->refcount, 1);
751 pi_state->key = FUTEX_KEY_INIT;
753 current->pi_state_cache = pi_state;
758 static struct futex_pi_state *alloc_pi_state(void)
760 struct futex_pi_state *pi_state = current->pi_state_cache;
763 current->pi_state_cache = NULL;
768 static void get_pi_state(struct futex_pi_state *pi_state)
770 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
774 * Drops a reference to the pi_state object and frees or caches it
775 * when the last reference is gone.
777 static void put_pi_state(struct futex_pi_state *pi_state)
782 if (!refcount_dec_and_test(&pi_state->refcount))
786 * If pi_state->owner is NULL, the owner is most probably dying
787 * and has cleaned up the pi_state already
789 if (pi_state->owner) {
790 struct task_struct *owner;
792 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
793 owner = pi_state->owner;
795 raw_spin_lock(&owner->pi_lock);
796 list_del_init(&pi_state->list);
797 raw_spin_unlock(&owner->pi_lock);
799 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
800 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
803 if (current->pi_state_cache) {
807 * pi_state->list is already empty.
808 * clear pi_state->owner.
809 * refcount is at 0 - put it back to 1.
811 pi_state->owner = NULL;
812 refcount_set(&pi_state->refcount, 1);
813 current->pi_state_cache = pi_state;
817 #ifdef CONFIG_FUTEX_PI
820 * This task is holding PI mutexes at exit time => bad.
821 * Kernel cleans up PI-state, but userspace is likely hosed.
822 * (Robust-futex cleanup is separate and might save the day for userspace.)
824 static void exit_pi_state_list(struct task_struct *curr)
826 struct list_head *next, *head = &curr->pi_state_list;
827 struct futex_pi_state *pi_state;
828 struct futex_hash_bucket *hb;
829 union futex_key key = FUTEX_KEY_INIT;
831 if (!futex_cmpxchg_enabled)
834 * We are a ZOMBIE and nobody can enqueue itself on
835 * pi_state_list anymore, but we have to be careful
836 * versus waiters unqueueing themselves:
838 raw_spin_lock_irq(&curr->pi_lock);
839 while (!list_empty(head)) {
841 pi_state = list_entry(next, struct futex_pi_state, list);
843 hb = hash_futex(&key);
846 * We can race against put_pi_state() removing itself from the
847 * list (a waiter going away). put_pi_state() will first
848 * decrement the reference count and then modify the list, so
849 * its possible to see the list entry but fail this reference
852 * In that case; drop the locks to let put_pi_state() make
853 * progress and retry the loop.
855 if (!refcount_inc_not_zero(&pi_state->refcount)) {
856 raw_spin_unlock_irq(&curr->pi_lock);
858 raw_spin_lock_irq(&curr->pi_lock);
861 raw_spin_unlock_irq(&curr->pi_lock);
863 spin_lock(&hb->lock);
864 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
865 raw_spin_lock(&curr->pi_lock);
867 * We dropped the pi-lock, so re-check whether this
868 * task still owns the PI-state:
870 if (head->next != next) {
871 /* retain curr->pi_lock for the loop invariant */
872 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
873 spin_unlock(&hb->lock);
874 put_pi_state(pi_state);
878 WARN_ON(pi_state->owner != curr);
879 WARN_ON(list_empty(&pi_state->list));
880 list_del_init(&pi_state->list);
881 pi_state->owner = NULL;
883 raw_spin_unlock(&curr->pi_lock);
884 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
885 spin_unlock(&hb->lock);
887 rt_mutex_futex_unlock(&pi_state->pi_mutex);
888 put_pi_state(pi_state);
890 raw_spin_lock_irq(&curr->pi_lock);
892 raw_spin_unlock_irq(&curr->pi_lock);
895 static inline void exit_pi_state_list(struct task_struct *curr) { }
899 * We need to check the following states:
901 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
903 * [1] NULL | --- | --- | 0 | 0/1 | Valid
904 * [2] NULL | --- | --- | >0 | 0/1 | Valid
906 * [3] Found | NULL | -- | Any | 0/1 | Invalid
908 * [4] Found | Found | NULL | 0 | 1 | Valid
909 * [5] Found | Found | NULL | >0 | 1 | Invalid
911 * [6] Found | Found | task | 0 | 1 | Valid
913 * [7] Found | Found | NULL | Any | 0 | Invalid
915 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
916 * [9] Found | Found | task | 0 | 0 | Invalid
917 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
919 * [1] Indicates that the kernel can acquire the futex atomically. We
920 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
922 * [2] Valid, if TID does not belong to a kernel thread. If no matching
923 * thread is found then it indicates that the owner TID has died.
925 * [3] Invalid. The waiter is queued on a non PI futex
927 * [4] Valid state after exit_robust_list(), which sets the user space
928 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
930 * [5] The user space value got manipulated between exit_robust_list()
931 * and exit_pi_state_list()
933 * [6] Valid state after exit_pi_state_list() which sets the new owner in
934 * the pi_state but cannot access the user space value.
936 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
938 * [8] Owner and user space value match
940 * [9] There is no transient state which sets the user space TID to 0
941 * except exit_robust_list(), but this is indicated by the
942 * FUTEX_OWNER_DIED bit. See [4]
944 * [10] There is no transient state which leaves owner and user space
948 * Serialization and lifetime rules:
952 * hb -> futex_q, relation
953 * futex_q -> pi_state, relation
955 * (cannot be raw because hb can contain arbitrary amount
958 * pi_mutex->wait_lock:
962 * (and pi_mutex 'obviously')
966 * p->pi_state_list -> pi_state->list, relation
968 * pi_state->refcount:
976 * pi_mutex->wait_lock
982 * Validate that the existing waiter has a pi_state and sanity check
983 * the pi_state against the user space value. If correct, attach to
986 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
987 struct futex_pi_state *pi_state,
988 struct futex_pi_state **ps)
990 pid_t pid = uval & FUTEX_TID_MASK;
995 * Userspace might have messed up non-PI and PI futexes [3]
997 if (unlikely(!pi_state))
1001 * We get here with hb->lock held, and having found a
1002 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1003 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1004 * which in turn means that futex_lock_pi() still has a reference on
1007 * The waiter holding a reference on @pi_state also protects against
1008 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1009 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1010 * free pi_state before we can take a reference ourselves.
1012 WARN_ON(!refcount_read(&pi_state->refcount));
1015 * Now that we have a pi_state, we can acquire wait_lock
1016 * and do the state validation.
1018 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1021 * Since {uval, pi_state} is serialized by wait_lock, and our current
1022 * uval was read without holding it, it can have changed. Verify it
1023 * still is what we expect it to be, otherwise retry the entire
1026 if (get_futex_value_locked(&uval2, uaddr))
1033 * Handle the owner died case:
1035 if (uval & FUTEX_OWNER_DIED) {
1037 * exit_pi_state_list sets owner to NULL and wakes the
1038 * topmost waiter. The task which acquires the
1039 * pi_state->rt_mutex will fixup owner.
1041 if (!pi_state->owner) {
1043 * No pi state owner, but the user space TID
1044 * is not 0. Inconsistent state. [5]
1049 * Take a ref on the state and return success. [4]
1055 * If TID is 0, then either the dying owner has not
1056 * yet executed exit_pi_state_list() or some waiter
1057 * acquired the rtmutex in the pi state, but did not
1058 * yet fixup the TID in user space.
1060 * Take a ref on the state and return success. [6]
1066 * If the owner died bit is not set, then the pi_state
1067 * must have an owner. [7]
1069 if (!pi_state->owner)
1074 * Bail out if user space manipulated the futex value. If pi
1075 * state exists then the owner TID must be the same as the
1076 * user space TID. [9/10]
1078 if (pid != task_pid_vnr(pi_state->owner))
1082 get_pi_state(pi_state);
1083 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1100 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1105 * wait_for_owner_exiting - Block until the owner has exited
1106 * @ret: owner's current futex lock status
1107 * @exiting: Pointer to the exiting task
1109 * Caller must hold a refcount on @exiting.
1111 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1113 if (ret != -EBUSY) {
1114 WARN_ON_ONCE(exiting);
1118 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1121 mutex_lock(&exiting->futex_exit_mutex);
1123 * No point in doing state checking here. If the waiter got here
1124 * while the task was in exec()->exec_futex_release() then it can
1125 * have any FUTEX_STATE_* value when the waiter has acquired the
1126 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1127 * already. Highly unlikely and not a problem. Just one more round
1128 * through the futex maze.
1130 mutex_unlock(&exiting->futex_exit_mutex);
1132 put_task_struct(exiting);
1135 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1136 struct task_struct *tsk)
1141 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1142 * caller that the alleged owner is busy.
1144 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1148 * Reread the user space value to handle the following situation:
1152 * sys_exit() sys_futex()
1153 * do_exit() futex_lock_pi()
1154 * futex_lock_pi_atomic()
1155 * exit_signals(tsk) No waiters:
1156 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1157 * mm_release(tsk) Set waiter bit
1158 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1159 * Set owner died attach_to_pi_owner() {
1160 * *uaddr = 0xC0000000; tsk = get_task(PID);
1161 * } if (!tsk->flags & PF_EXITING) {
1163 * tsk->futex_state = } else {
1164 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1167 * return -ESRCH; <--- FAIL
1170 * Returning ESRCH unconditionally is wrong here because the
1171 * user space value has been changed by the exiting task.
1173 * The same logic applies to the case where the exiting task is
1176 if (get_futex_value_locked(&uval2, uaddr))
1179 /* If the user space value has changed, try again. */
1184 * The exiting task did not have a robust list, the robust list was
1185 * corrupted or the user space value in *uaddr is simply bogus.
1186 * Give up and tell user space.
1192 * Lookup the task for the TID provided from user space and attach to
1193 * it after doing proper sanity checks.
1195 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1196 struct futex_pi_state **ps,
1197 struct task_struct **exiting)
1199 pid_t pid = uval & FUTEX_TID_MASK;
1200 struct futex_pi_state *pi_state;
1201 struct task_struct *p;
1204 * We are the first waiter - try to look up the real owner and attach
1205 * the new pi_state to it, but bail out when TID = 0 [1]
1207 * The !pid check is paranoid. None of the call sites should end up
1208 * with pid == 0, but better safe than sorry. Let the caller retry
1212 p = find_get_task_by_vpid(pid);
1214 return handle_exit_race(uaddr, uval, NULL);
1216 if (unlikely(p->flags & PF_KTHREAD)) {
1222 * We need to look at the task state to figure out, whether the
1223 * task is exiting. To protect against the change of the task state
1224 * in futex_exit_release(), we do this protected by p->pi_lock:
1226 raw_spin_lock_irq(&p->pi_lock);
1227 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1229 * The task is on the way out. When the futex state is
1230 * FUTEX_STATE_DEAD, we know that the task has finished
1233 int ret = handle_exit_race(uaddr, uval, p);
1235 raw_spin_unlock_irq(&p->pi_lock);
1237 * If the owner task is between FUTEX_STATE_EXITING and
1238 * FUTEX_STATE_DEAD then store the task pointer and keep
1239 * the reference on the task struct. The calling code will
1240 * drop all locks, wait for the task to reach
1241 * FUTEX_STATE_DEAD and then drop the refcount. This is
1242 * required to prevent a live lock when the current task
1243 * preempted the exiting task between the two states.
1253 * No existing pi state. First waiter. [2]
1255 * This creates pi_state, we have hb->lock held, this means nothing can
1256 * observe this state, wait_lock is irrelevant.
1258 pi_state = alloc_pi_state();
1261 * Initialize the pi_mutex in locked state and make @p
1264 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1266 /* Store the key for possible exit cleanups: */
1267 pi_state->key = *key;
1269 WARN_ON(!list_empty(&pi_state->list));
1270 list_add(&pi_state->list, &p->pi_state_list);
1272 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1273 * because there is no concurrency as the object is not published yet.
1275 pi_state->owner = p;
1276 raw_spin_unlock_irq(&p->pi_lock);
1285 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1286 struct futex_hash_bucket *hb,
1287 union futex_key *key, struct futex_pi_state **ps,
1288 struct task_struct **exiting)
1290 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1293 * If there is a waiter on that futex, validate it and
1294 * attach to the pi_state when the validation succeeds.
1297 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1300 * We are the first waiter - try to look up the owner based on
1301 * @uval and attach to it.
1303 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1306 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1311 if (unlikely(should_fail_futex(true)))
1314 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1318 /* If user space value changed, let the caller retry */
1319 return curval != uval ? -EAGAIN : 0;
1323 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1324 * @uaddr: the pi futex user address
1325 * @hb: the pi futex hash bucket
1326 * @key: the futex key associated with uaddr and hb
1327 * @ps: the pi_state pointer where we store the result of the
1329 * @task: the task to perform the atomic lock work for. This will
1330 * be "current" except in the case of requeue pi.
1331 * @exiting: Pointer to store the task pointer of the owner task
1332 * which is in the middle of exiting
1333 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1336 * - 0 - ready to wait;
1337 * - 1 - acquired the lock;
1340 * The hb->lock and futex_key refs shall be held by the caller.
1342 * @exiting is only set when the return value is -EBUSY. If so, this holds
1343 * a refcount on the exiting task on return and the caller needs to drop it
1344 * after waiting for the exit to complete.
1346 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1347 union futex_key *key,
1348 struct futex_pi_state **ps,
1349 struct task_struct *task,
1350 struct task_struct **exiting,
1353 u32 uval, newval, vpid = task_pid_vnr(task);
1354 struct futex_q *top_waiter;
1358 * Read the user space value first so we can validate a few
1359 * things before proceeding further.
1361 if (get_futex_value_locked(&uval, uaddr))
1364 if (unlikely(should_fail_futex(true)))
1370 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1373 if ((unlikely(should_fail_futex(true))))
1377 * Lookup existing state first. If it exists, try to attach to
1380 top_waiter = futex_top_waiter(hb, key);
1382 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1385 * No waiter and user TID is 0. We are here because the
1386 * waiters or the owner died bit is set or called from
1387 * requeue_cmp_pi or for whatever reason something took the
1390 if (!(uval & FUTEX_TID_MASK)) {
1392 * We take over the futex. No other waiters and the user space
1393 * TID is 0. We preserve the owner died bit.
1395 newval = uval & FUTEX_OWNER_DIED;
1398 /* The futex requeue_pi code can enforce the waiters bit */
1400 newval |= FUTEX_WAITERS;
1402 ret = lock_pi_update_atomic(uaddr, uval, newval);
1403 /* If the take over worked, return 1 */
1404 return ret < 0 ? ret : 1;
1408 * First waiter. Set the waiters bit before attaching ourself to
1409 * the owner. If owner tries to unlock, it will be forced into
1410 * the kernel and blocked on hb->lock.
1412 newval = uval | FUTEX_WAITERS;
1413 ret = lock_pi_update_atomic(uaddr, uval, newval);
1417 * If the update of the user space value succeeded, we try to
1418 * attach to the owner. If that fails, no harm done, we only
1419 * set the FUTEX_WAITERS bit in the user space variable.
1421 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1425 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1426 * @q: The futex_q to unqueue
1428 * The q->lock_ptr must not be NULL and must be held by the caller.
1430 static void __unqueue_futex(struct futex_q *q)
1432 struct futex_hash_bucket *hb;
1434 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1436 lockdep_assert_held(q->lock_ptr);
1438 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1439 plist_del(&q->list, &hb->chain);
1444 * The hash bucket lock must be held when this is called.
1445 * Afterwards, the futex_q must not be accessed. Callers
1446 * must ensure to later call wake_up_q() for the actual
1449 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1451 struct task_struct *p = q->task;
1453 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1459 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1460 * is written, without taking any locks. This is possible in the event
1461 * of a spurious wakeup, for example. A memory barrier is required here
1462 * to prevent the following store to lock_ptr from getting ahead of the
1463 * plist_del in __unqueue_futex().
1465 smp_store_release(&q->lock_ptr, NULL);
1468 * Queue the task for later wakeup for after we've released
1471 wake_q_add_safe(wake_q, p);
1475 * Caller must hold a reference on @pi_state.
1477 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1480 struct task_struct *new_owner;
1481 bool postunlock = false;
1482 DEFINE_WAKE_Q(wake_q);
1485 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1486 if (WARN_ON_ONCE(!new_owner)) {
1488 * As per the comment in futex_unlock_pi() this should not happen.
1490 * When this happens, give up our locks and try again, giving
1491 * the futex_lock_pi() instance time to complete, either by
1492 * waiting on the rtmutex or removing itself from the futex
1500 * We pass it to the next owner. The WAITERS bit is always kept
1501 * enabled while there is PI state around. We cleanup the owner
1502 * died bit, because we are the owner.
1504 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1506 if (unlikely(should_fail_futex(true)))
1509 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1510 if (!ret && (curval != uval)) {
1512 * If a unconditional UNLOCK_PI operation (user space did not
1513 * try the TID->0 transition) raced with a waiter setting the
1514 * FUTEX_WAITERS flag between get_user() and locking the hash
1515 * bucket lock, retry the operation.
1517 if ((FUTEX_TID_MASK & curval) == uval)
1527 * This is a point of no return; once we modify the uval there is no
1528 * going back and subsequent operations must not fail.
1531 raw_spin_lock(&pi_state->owner->pi_lock);
1532 WARN_ON(list_empty(&pi_state->list));
1533 list_del_init(&pi_state->list);
1534 raw_spin_unlock(&pi_state->owner->pi_lock);
1536 raw_spin_lock(&new_owner->pi_lock);
1537 WARN_ON(!list_empty(&pi_state->list));
1538 list_add(&pi_state->list, &new_owner->pi_state_list);
1539 pi_state->owner = new_owner;
1540 raw_spin_unlock(&new_owner->pi_lock);
1542 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1545 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1548 rt_mutex_postunlock(&wake_q);
1554 * Express the locking dependencies for lockdep:
1557 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1560 spin_lock(&hb1->lock);
1562 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1563 } else { /* hb1 > hb2 */
1564 spin_lock(&hb2->lock);
1565 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1570 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1572 spin_unlock(&hb1->lock);
1574 spin_unlock(&hb2->lock);
1578 * Wake up waiters matching bitset queued on this futex (uaddr).
1581 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1583 struct futex_hash_bucket *hb;
1584 struct futex_q *this, *next;
1585 union futex_key key = FUTEX_KEY_INIT;
1587 DEFINE_WAKE_Q(wake_q);
1592 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1593 if (unlikely(ret != 0))
1596 hb = hash_futex(&key);
1598 /* Make sure we really have tasks to wakeup */
1599 if (!hb_waiters_pending(hb))
1602 spin_lock(&hb->lock);
1604 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1605 if (match_futex (&this->key, &key)) {
1606 if (this->pi_state || this->rt_waiter) {
1611 /* Check if one of the bits is set in both bitsets */
1612 if (!(this->bitset & bitset))
1615 mark_wake_futex(&wake_q, this);
1616 if (++ret >= nr_wake)
1621 spin_unlock(&hb->lock);
1626 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1628 unsigned int op = (encoded_op & 0x70000000) >> 28;
1629 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1630 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1631 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1634 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1635 if (oparg < 0 || oparg > 31) {
1636 char comm[sizeof(current->comm)];
1638 * kill this print and return -EINVAL when userspace
1641 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1642 get_task_comm(comm, current), oparg);
1648 pagefault_disable();
1649 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1655 case FUTEX_OP_CMP_EQ:
1656 return oldval == cmparg;
1657 case FUTEX_OP_CMP_NE:
1658 return oldval != cmparg;
1659 case FUTEX_OP_CMP_LT:
1660 return oldval < cmparg;
1661 case FUTEX_OP_CMP_GE:
1662 return oldval >= cmparg;
1663 case FUTEX_OP_CMP_LE:
1664 return oldval <= cmparg;
1665 case FUTEX_OP_CMP_GT:
1666 return oldval > cmparg;
1673 * Wake up all waiters hashed on the physical page that is mapped
1674 * to this virtual address:
1677 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1678 int nr_wake, int nr_wake2, int op)
1680 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1681 struct futex_hash_bucket *hb1, *hb2;
1682 struct futex_q *this, *next;
1684 DEFINE_WAKE_Q(wake_q);
1687 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1688 if (unlikely(ret != 0))
1690 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1691 if (unlikely(ret != 0))
1694 hb1 = hash_futex(&key1);
1695 hb2 = hash_futex(&key2);
1698 double_lock_hb(hb1, hb2);
1699 op_ret = futex_atomic_op_inuser(op, uaddr2);
1700 if (unlikely(op_ret < 0)) {
1701 double_unlock_hb(hb1, hb2);
1703 if (!IS_ENABLED(CONFIG_MMU) ||
1704 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1706 * we don't get EFAULT from MMU faults if we don't have
1707 * an MMU, but we might get them from range checking
1713 if (op_ret == -EFAULT) {
1714 ret = fault_in_user_writeable(uaddr2);
1719 if (!(flags & FLAGS_SHARED)) {
1728 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1729 if (match_futex (&this->key, &key1)) {
1730 if (this->pi_state || this->rt_waiter) {
1734 mark_wake_futex(&wake_q, this);
1735 if (++ret >= nr_wake)
1742 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1743 if (match_futex (&this->key, &key2)) {
1744 if (this->pi_state || this->rt_waiter) {
1748 mark_wake_futex(&wake_q, this);
1749 if (++op_ret >= nr_wake2)
1757 double_unlock_hb(hb1, hb2);
1763 * requeue_futex() - Requeue a futex_q from one hb to another
1764 * @q: the futex_q to requeue
1765 * @hb1: the source hash_bucket
1766 * @hb2: the target hash_bucket
1767 * @key2: the new key for the requeued futex_q
1770 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1771 struct futex_hash_bucket *hb2, union futex_key *key2)
1775 * If key1 and key2 hash to the same bucket, no need to
1778 if (likely(&hb1->chain != &hb2->chain)) {
1779 plist_del(&q->list, &hb1->chain);
1780 hb_waiters_dec(hb1);
1781 hb_waiters_inc(hb2);
1782 plist_add(&q->list, &hb2->chain);
1783 q->lock_ptr = &hb2->lock;
1789 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1791 * @key: the key of the requeue target futex
1792 * @hb: the hash_bucket of the requeue target futex
1794 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1795 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1796 * to the requeue target futex so the waiter can detect the wakeup on the right
1797 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1798 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1799 * to protect access to the pi_state to fixup the owner later. Must be called
1800 * with both q->lock_ptr and hb->lock held.
1803 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1804 struct futex_hash_bucket *hb)
1810 WARN_ON(!q->rt_waiter);
1811 q->rt_waiter = NULL;
1813 q->lock_ptr = &hb->lock;
1815 wake_up_state(q->task, TASK_NORMAL);
1819 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1820 * @pifutex: the user address of the to futex
1821 * @hb1: the from futex hash bucket, must be locked by the caller
1822 * @hb2: the to futex hash bucket, must be locked by the caller
1823 * @key1: the from futex key
1824 * @key2: the to futex key
1825 * @ps: address to store the pi_state pointer
1826 * @exiting: Pointer to store the task pointer of the owner task
1827 * which is in the middle of exiting
1828 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1830 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1831 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1832 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1833 * hb1 and hb2 must be held by the caller.
1835 * @exiting is only set when the return value is -EBUSY. If so, this holds
1836 * a refcount on the exiting task on return and the caller needs to drop it
1837 * after waiting for the exit to complete.
1840 * - 0 - failed to acquire the lock atomically;
1841 * - >0 - acquired the lock, return value is vpid of the top_waiter
1845 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1846 struct futex_hash_bucket *hb2, union futex_key *key1,
1847 union futex_key *key2, struct futex_pi_state **ps,
1848 struct task_struct **exiting, int set_waiters)
1850 struct futex_q *top_waiter = NULL;
1854 if (get_futex_value_locked(&curval, pifutex))
1857 if (unlikely(should_fail_futex(true)))
1861 * Find the top_waiter and determine if there are additional waiters.
1862 * If the caller intends to requeue more than 1 waiter to pifutex,
1863 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1864 * as we have means to handle the possible fault. If not, don't set
1865 * the bit unecessarily as it will force the subsequent unlock to enter
1868 top_waiter = futex_top_waiter(hb1, key1);
1870 /* There are no waiters, nothing for us to do. */
1874 /* Ensure we requeue to the expected futex. */
1875 if (!match_futex(top_waiter->requeue_pi_key, key2))
1879 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1880 * the contended case or if set_waiters is 1. The pi_state is returned
1881 * in ps in contended cases.
1883 vpid = task_pid_vnr(top_waiter->task);
1884 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1885 exiting, set_waiters);
1887 requeue_pi_wake_futex(top_waiter, key2, hb2);
1894 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1895 * @uaddr1: source futex user address
1896 * @flags: futex flags (FLAGS_SHARED, etc.)
1897 * @uaddr2: target futex user address
1898 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1899 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1900 * @cmpval: @uaddr1 expected value (or %NULL)
1901 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1902 * pi futex (pi to pi requeue is not supported)
1904 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1905 * uaddr2 atomically on behalf of the top waiter.
1908 * - >=0 - on success, the number of tasks requeued or woken;
1911 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1912 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1913 u32 *cmpval, int requeue_pi)
1915 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1916 int task_count = 0, ret;
1917 struct futex_pi_state *pi_state = NULL;
1918 struct futex_hash_bucket *hb1, *hb2;
1919 struct futex_q *this, *next;
1920 DEFINE_WAKE_Q(wake_q);
1922 if (nr_wake < 0 || nr_requeue < 0)
1926 * When PI not supported: return -ENOSYS if requeue_pi is true,
1927 * consequently the compiler knows requeue_pi is always false past
1928 * this point which will optimize away all the conditional code
1931 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1936 * Requeue PI only works on two distinct uaddrs. This
1937 * check is only valid for private futexes. See below.
1939 if (uaddr1 == uaddr2)
1943 * requeue_pi requires a pi_state, try to allocate it now
1944 * without any locks in case it fails.
1946 if (refill_pi_state_cache())
1949 * requeue_pi must wake as many tasks as it can, up to nr_wake
1950 * + nr_requeue, since it acquires the rt_mutex prior to
1951 * returning to userspace, so as to not leave the rt_mutex with
1952 * waiters and no owner. However, second and third wake-ups
1953 * cannot be predicted as they involve race conditions with the
1954 * first wake and a fault while looking up the pi_state. Both
1955 * pthread_cond_signal() and pthread_cond_broadcast() should
1963 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1964 if (unlikely(ret != 0))
1966 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1967 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1968 if (unlikely(ret != 0))
1972 * The check above which compares uaddrs is not sufficient for
1973 * shared futexes. We need to compare the keys:
1975 if (requeue_pi && match_futex(&key1, &key2))
1978 hb1 = hash_futex(&key1);
1979 hb2 = hash_futex(&key2);
1982 hb_waiters_inc(hb2);
1983 double_lock_hb(hb1, hb2);
1985 if (likely(cmpval != NULL)) {
1988 ret = get_futex_value_locked(&curval, uaddr1);
1990 if (unlikely(ret)) {
1991 double_unlock_hb(hb1, hb2);
1992 hb_waiters_dec(hb2);
1994 ret = get_user(curval, uaddr1);
1998 if (!(flags & FLAGS_SHARED))
2003 if (curval != *cmpval) {
2009 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2010 struct task_struct *exiting = NULL;
2013 * Attempt to acquire uaddr2 and wake the top waiter. If we
2014 * intend to requeue waiters, force setting the FUTEX_WAITERS
2015 * bit. We force this here where we are able to easily handle
2016 * faults rather in the requeue loop below.
2018 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2020 &exiting, nr_requeue);
2023 * At this point the top_waiter has either taken uaddr2 or is
2024 * waiting on it. If the former, then the pi_state will not
2025 * exist yet, look it up one more time to ensure we have a
2026 * reference to it. If the lock was taken, ret contains the
2027 * vpid of the top waiter task.
2028 * If the lock was not taken, we have pi_state and an initial
2029 * refcount on it. In case of an error we have nothing.
2035 * If we acquired the lock, then the user space value
2036 * of uaddr2 should be vpid. It cannot be changed by
2037 * the top waiter as it is blocked on hb2 lock if it
2038 * tries to do so. If something fiddled with it behind
2039 * our back the pi state lookup might unearth it. So
2040 * we rather use the known value than rereading and
2041 * handing potential crap to lookup_pi_state.
2043 * If that call succeeds then we have pi_state and an
2044 * initial refcount on it.
2046 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2047 &pi_state, &exiting);
2052 /* We hold a reference on the pi state. */
2055 /* If the above failed, then pi_state is NULL */
2057 double_unlock_hb(hb1, hb2);
2058 hb_waiters_dec(hb2);
2059 ret = fault_in_user_writeable(uaddr2);
2066 * Two reasons for this:
2067 * - EBUSY: Owner is exiting and we just wait for the
2069 * - EAGAIN: The user space value changed.
2071 double_unlock_hb(hb1, hb2);
2072 hb_waiters_dec(hb2);
2074 * Handle the case where the owner is in the middle of
2075 * exiting. Wait for the exit to complete otherwise
2076 * this task might loop forever, aka. live lock.
2078 wait_for_owner_exiting(ret, exiting);
2086 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2087 if (task_count - nr_wake >= nr_requeue)
2090 if (!match_futex(&this->key, &key1))
2094 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2095 * be paired with each other and no other futex ops.
2097 * We should never be requeueing a futex_q with a pi_state,
2098 * which is awaiting a futex_unlock_pi().
2100 if ((requeue_pi && !this->rt_waiter) ||
2101 (!requeue_pi && this->rt_waiter) ||
2108 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2109 * lock, we already woke the top_waiter. If not, it will be
2110 * woken by futex_unlock_pi().
2112 if (++task_count <= nr_wake && !requeue_pi) {
2113 mark_wake_futex(&wake_q, this);
2117 /* Ensure we requeue to the expected futex for requeue_pi. */
2118 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2124 * Requeue nr_requeue waiters and possibly one more in the case
2125 * of requeue_pi if we couldn't acquire the lock atomically.
2129 * Prepare the waiter to take the rt_mutex. Take a
2130 * refcount on the pi_state and store the pointer in
2131 * the futex_q object of the waiter.
2133 get_pi_state(pi_state);
2134 this->pi_state = pi_state;
2135 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2140 * We got the lock. We do neither drop the
2141 * refcount on pi_state nor clear
2142 * this->pi_state because the waiter needs the
2143 * pi_state for cleaning up the user space
2144 * value. It will drop the refcount after
2147 requeue_pi_wake_futex(this, &key2, hb2);
2151 * rt_mutex_start_proxy_lock() detected a
2152 * potential deadlock when we tried to queue
2153 * that waiter. Drop the pi_state reference
2154 * which we took above and remove the pointer
2155 * to the state from the waiters futex_q
2158 this->pi_state = NULL;
2159 put_pi_state(pi_state);
2161 * We stop queueing more waiters and let user
2162 * space deal with the mess.
2167 requeue_futex(this, hb1, hb2, &key2);
2171 * We took an extra initial reference to the pi_state either
2172 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2173 * need to drop it here again.
2175 put_pi_state(pi_state);
2178 double_unlock_hb(hb1, hb2);
2180 hb_waiters_dec(hb2);
2181 return ret ? ret : task_count;
2184 /* The key must be already stored in q->key. */
2185 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2186 __acquires(&hb->lock)
2188 struct futex_hash_bucket *hb;
2190 hb = hash_futex(&q->key);
2193 * Increment the counter before taking the lock so that
2194 * a potential waker won't miss a to-be-slept task that is
2195 * waiting for the spinlock. This is safe as all queue_lock()
2196 * users end up calling queue_me(). Similarly, for housekeeping,
2197 * decrement the counter at queue_unlock() when some error has
2198 * occurred and we don't end up adding the task to the list.
2200 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2202 q->lock_ptr = &hb->lock;
2204 spin_lock(&hb->lock);
2209 queue_unlock(struct futex_hash_bucket *hb)
2210 __releases(&hb->lock)
2212 spin_unlock(&hb->lock);
2216 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2221 * The priority used to register this element is
2222 * - either the real thread-priority for the real-time threads
2223 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2224 * - or MAX_RT_PRIO for non-RT threads.
2225 * Thus, all RT-threads are woken first in priority order, and
2226 * the others are woken last, in FIFO order.
2228 prio = min(current->normal_prio, MAX_RT_PRIO);
2230 plist_node_init(&q->list, prio);
2231 plist_add(&q->list, &hb->chain);
2236 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2237 * @q: The futex_q to enqueue
2238 * @hb: The destination hash bucket
2240 * The hb->lock must be held by the caller, and is released here. A call to
2241 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2242 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2243 * or nothing if the unqueue is done as part of the wake process and the unqueue
2244 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2247 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2248 __releases(&hb->lock)
2251 spin_unlock(&hb->lock);
2255 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2256 * @q: The futex_q to unqueue
2258 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2259 * be paired with exactly one earlier call to queue_me().
2262 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2263 * - 0 - if the futex_q was already removed by the waking thread
2265 static int unqueue_me(struct futex_q *q)
2267 spinlock_t *lock_ptr;
2270 /* In the common case we don't take the spinlock, which is nice. */
2273 * q->lock_ptr can change between this read and the following spin_lock.
2274 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2275 * optimizing lock_ptr out of the logic below.
2277 lock_ptr = READ_ONCE(q->lock_ptr);
2278 if (lock_ptr != NULL) {
2279 spin_lock(lock_ptr);
2281 * q->lock_ptr can change between reading it and
2282 * spin_lock(), causing us to take the wrong lock. This
2283 * corrects the race condition.
2285 * Reasoning goes like this: if we have the wrong lock,
2286 * q->lock_ptr must have changed (maybe several times)
2287 * between reading it and the spin_lock(). It can
2288 * change again after the spin_lock() but only if it was
2289 * already changed before the spin_lock(). It cannot,
2290 * however, change back to the original value. Therefore
2291 * we can detect whether we acquired the correct lock.
2293 if (unlikely(lock_ptr != q->lock_ptr)) {
2294 spin_unlock(lock_ptr);
2299 BUG_ON(q->pi_state);
2301 spin_unlock(lock_ptr);
2309 * PI futexes can not be requeued and must remove themself from the
2310 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2313 static void unqueue_me_pi(struct futex_q *q)
2314 __releases(q->lock_ptr)
2318 BUG_ON(!q->pi_state);
2319 put_pi_state(q->pi_state);
2322 spin_unlock(q->lock_ptr);
2325 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2326 struct task_struct *argowner)
2328 struct futex_pi_state *pi_state = q->pi_state;
2329 u32 uval, curval, newval;
2330 struct task_struct *oldowner, *newowner;
2334 lockdep_assert_held(q->lock_ptr);
2336 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2338 oldowner = pi_state->owner;
2341 * We are here because either:
2343 * - we stole the lock and pi_state->owner needs updating to reflect
2344 * that (@argowner == current),
2348 * - someone stole our lock and we need to fix things to point to the
2349 * new owner (@argowner == NULL).
2351 * Either way, we have to replace the TID in the user space variable.
2352 * This must be atomic as we have to preserve the owner died bit here.
2354 * Note: We write the user space value _before_ changing the pi_state
2355 * because we can fault here. Imagine swapped out pages or a fork
2356 * that marked all the anonymous memory readonly for cow.
2358 * Modifying pi_state _before_ the user space value would leave the
2359 * pi_state in an inconsistent state when we fault here, because we
2360 * need to drop the locks to handle the fault. This might be observed
2361 * in the PID check in lookup_pi_state.
2365 if (oldowner != current) {
2367 * We raced against a concurrent self; things are
2368 * already fixed up. Nothing to do.
2374 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2375 /* We got the lock after all, nothing to fix. */
2381 * Since we just failed the trylock; there must be an owner.
2383 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2386 WARN_ON_ONCE(argowner != current);
2387 if (oldowner == current) {
2389 * We raced against a concurrent self; things are
2390 * already fixed up. Nothing to do.
2395 newowner = argowner;
2398 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2400 if (!pi_state->owner)
2401 newtid |= FUTEX_OWNER_DIED;
2403 err = get_futex_value_locked(&uval, uaddr);
2408 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2410 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2420 * We fixed up user space. Now we need to fix the pi_state
2423 if (pi_state->owner != NULL) {
2424 raw_spin_lock(&pi_state->owner->pi_lock);
2425 WARN_ON(list_empty(&pi_state->list));
2426 list_del_init(&pi_state->list);
2427 raw_spin_unlock(&pi_state->owner->pi_lock);
2430 pi_state->owner = newowner;
2432 raw_spin_lock(&newowner->pi_lock);
2433 WARN_ON(!list_empty(&pi_state->list));
2434 list_add(&pi_state->list, &newowner->pi_state_list);
2435 raw_spin_unlock(&newowner->pi_lock);
2436 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2441 * In order to reschedule or handle a page fault, we need to drop the
2442 * locks here. In the case of a fault, this gives the other task
2443 * (either the highest priority waiter itself or the task which stole
2444 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2445 * are back from handling the fault we need to check the pi_state after
2446 * reacquiring the locks and before trying to do another fixup. When
2447 * the fixup has been done already we simply return.
2449 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2450 * drop hb->lock since the caller owns the hb -> futex_q relation.
2451 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2454 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2455 spin_unlock(q->lock_ptr);
2459 ret = fault_in_user_writeable(uaddr);
2473 spin_lock(q->lock_ptr);
2474 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2477 * Check if someone else fixed it for us:
2479 if (pi_state->owner != oldowner) {
2490 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2494 static long futex_wait_restart(struct restart_block *restart);
2497 * fixup_owner() - Post lock pi_state and corner case management
2498 * @uaddr: user address of the futex
2499 * @q: futex_q (contains pi_state and access to the rt_mutex)
2500 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2502 * After attempting to lock an rt_mutex, this function is called to cleanup
2503 * the pi_state owner as well as handle race conditions that may allow us to
2504 * acquire the lock. Must be called with the hb lock held.
2507 * - 1 - success, lock taken;
2508 * - 0 - success, lock not taken;
2509 * - <0 - on error (-EFAULT)
2511 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2517 * Got the lock. We might not be the anticipated owner if we
2518 * did a lock-steal - fix up the PI-state in that case:
2520 * Speculative pi_state->owner read (we don't hold wait_lock);
2521 * since we own the lock pi_state->owner == current is the
2522 * stable state, anything else needs more attention.
2524 if (q->pi_state->owner != current)
2525 ret = fixup_pi_state_owner(uaddr, q, current);
2526 return ret ? ret : locked;
2530 * If we didn't get the lock; check if anybody stole it from us. In
2531 * that case, we need to fix up the uval to point to them instead of
2532 * us, otherwise bad things happen. [10]
2534 * Another speculative read; pi_state->owner == current is unstable
2535 * but needs our attention.
2537 if (q->pi_state->owner == current) {
2538 ret = fixup_pi_state_owner(uaddr, q, NULL);
2543 * Paranoia check. If we did not take the lock, then we should not be
2544 * the owner of the rt_mutex.
2546 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2547 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2548 "pi-state %p\n", ret,
2549 q->pi_state->pi_mutex.owner,
2550 q->pi_state->owner);
2557 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2558 * @hb: the futex hash bucket, must be locked by the caller
2559 * @q: the futex_q to queue up on
2560 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2562 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2563 struct hrtimer_sleeper *timeout)
2566 * The task state is guaranteed to be set before another task can
2567 * wake it. set_current_state() is implemented using smp_store_mb() and
2568 * queue_me() calls spin_unlock() upon completion, both serializing
2569 * access to the hash list and forcing another memory barrier.
2571 set_current_state(TASK_INTERRUPTIBLE);
2576 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2579 * If we have been removed from the hash list, then another task
2580 * has tried to wake us, and we can skip the call to schedule().
2582 if (likely(!plist_node_empty(&q->list))) {
2584 * If the timer has already expired, current will already be
2585 * flagged for rescheduling. Only call schedule if there
2586 * is no timeout, or if it has yet to expire.
2588 if (!timeout || timeout->task)
2589 freezable_schedule();
2591 __set_current_state(TASK_RUNNING);
2595 * futex_wait_setup() - Prepare to wait on a futex
2596 * @uaddr: the futex userspace address
2597 * @val: the expected value
2598 * @flags: futex flags (FLAGS_SHARED, etc.)
2599 * @q: the associated futex_q
2600 * @hb: storage for hash_bucket pointer to be returned to caller
2602 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2603 * compare it with the expected value. Handle atomic faults internally.
2604 * Return with the hb lock held and a q.key reference on success, and unlocked
2605 * with no q.key reference on failure.
2608 * - 0 - uaddr contains val and hb has been locked;
2609 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2611 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2612 struct futex_q *q, struct futex_hash_bucket **hb)
2618 * Access the page AFTER the hash-bucket is locked.
2619 * Order is important:
2621 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2622 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2624 * The basic logical guarantee of a futex is that it blocks ONLY
2625 * if cond(var) is known to be true at the time of blocking, for
2626 * any cond. If we locked the hash-bucket after testing *uaddr, that
2627 * would open a race condition where we could block indefinitely with
2628 * cond(var) false, which would violate the guarantee.
2630 * On the other hand, we insert q and release the hash-bucket only
2631 * after testing *uaddr. This guarantees that futex_wait() will NOT
2632 * absorb a wakeup if *uaddr does not match the desired values
2633 * while the syscall executes.
2636 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2637 if (unlikely(ret != 0))
2641 *hb = queue_lock(q);
2643 ret = get_futex_value_locked(&uval, uaddr);
2648 ret = get_user(uval, uaddr);
2652 if (!(flags & FLAGS_SHARED))
2666 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2667 ktime_t *abs_time, u32 bitset)
2669 struct hrtimer_sleeper timeout, *to;
2670 struct restart_block *restart;
2671 struct futex_hash_bucket *hb;
2672 struct futex_q q = futex_q_init;
2679 to = futex_setup_timer(abs_time, &timeout, flags,
2680 current->timer_slack_ns);
2683 * Prepare to wait on uaddr. On success, holds hb lock and increments
2686 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2690 /* queue_me and wait for wakeup, timeout, or a signal. */
2691 futex_wait_queue_me(hb, &q, to);
2693 /* If we were woken (and unqueued), we succeeded, whatever. */
2695 /* unqueue_me() drops q.key ref */
2696 if (!unqueue_me(&q))
2699 if (to && !to->task)
2703 * We expect signal_pending(current), but we might be the
2704 * victim of a spurious wakeup as well.
2706 if (!signal_pending(current))
2713 restart = ¤t->restart_block;
2714 restart->fn = futex_wait_restart;
2715 restart->futex.uaddr = uaddr;
2716 restart->futex.val = val;
2717 restart->futex.time = *abs_time;
2718 restart->futex.bitset = bitset;
2719 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2721 ret = -ERESTART_RESTARTBLOCK;
2725 hrtimer_cancel(&to->timer);
2726 destroy_hrtimer_on_stack(&to->timer);
2732 static long futex_wait_restart(struct restart_block *restart)
2734 u32 __user *uaddr = restart->futex.uaddr;
2735 ktime_t t, *tp = NULL;
2737 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2738 t = restart->futex.time;
2741 restart->fn = do_no_restart_syscall;
2743 return (long)futex_wait(uaddr, restart->futex.flags,
2744 restart->futex.val, tp, restart->futex.bitset);
2749 * Userspace tried a 0 -> TID atomic transition of the futex value
2750 * and failed. The kernel side here does the whole locking operation:
2751 * if there are waiters then it will block as a consequence of relying
2752 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2753 * a 0 value of the futex too.).
2755 * Also serves as futex trylock_pi()'ing, and due semantics.
2757 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2758 ktime_t *time, int trylock)
2760 struct hrtimer_sleeper timeout, *to;
2761 struct futex_pi_state *pi_state = NULL;
2762 struct task_struct *exiting = NULL;
2763 struct rt_mutex_waiter rt_waiter;
2764 struct futex_hash_bucket *hb;
2765 struct futex_q q = futex_q_init;
2768 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2771 if (refill_pi_state_cache())
2774 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2777 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2778 if (unlikely(ret != 0))
2782 hb = queue_lock(&q);
2784 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2786 if (unlikely(ret)) {
2788 * Atomic work succeeded and we got the lock,
2789 * or failed. Either way, we do _not_ block.
2793 /* We got the lock. */
2795 goto out_unlock_put_key;
2801 * Two reasons for this:
2802 * - EBUSY: Task is exiting and we just wait for the
2804 * - EAGAIN: The user space value changed.
2808 * Handle the case where the owner is in the middle of
2809 * exiting. Wait for the exit to complete otherwise
2810 * this task might loop forever, aka. live lock.
2812 wait_for_owner_exiting(ret, exiting);
2816 goto out_unlock_put_key;
2820 WARN_ON(!q.pi_state);
2823 * Only actually queue now that the atomic ops are done:
2828 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2829 /* Fixup the trylock return value: */
2830 ret = ret ? 0 : -EWOULDBLOCK;
2834 rt_mutex_init_waiter(&rt_waiter);
2837 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2838 * hold it while doing rt_mutex_start_proxy(), because then it will
2839 * include hb->lock in the blocking chain, even through we'll not in
2840 * fact hold it while blocking. This will lead it to report -EDEADLK
2841 * and BUG when futex_unlock_pi() interleaves with this.
2843 * Therefore acquire wait_lock while holding hb->lock, but drop the
2844 * latter before calling __rt_mutex_start_proxy_lock(). This
2845 * interleaves with futex_unlock_pi() -- which does a similar lock
2846 * handoff -- such that the latter can observe the futex_q::pi_state
2847 * before __rt_mutex_start_proxy_lock() is done.
2849 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2850 spin_unlock(q.lock_ptr);
2852 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2853 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2854 * it sees the futex_q::pi_state.
2856 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2857 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2866 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2868 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2871 spin_lock(q.lock_ptr);
2873 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2874 * first acquire the hb->lock before removing the lock from the
2875 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2878 * In particular; it is important that futex_unlock_pi() can not
2879 * observe this inconsistency.
2881 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2886 * Fixup the pi_state owner and possibly acquire the lock if we
2889 res = fixup_owner(uaddr, &q, !ret);
2891 * If fixup_owner() returned an error, proprogate that. If it acquired
2892 * the lock, clear our -ETIMEDOUT or -EINTR.
2895 ret = (res < 0) ? res : 0;
2898 * If fixup_owner() faulted and was unable to handle the fault, unlock
2899 * it and return the fault to userspace.
2901 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2902 pi_state = q.pi_state;
2903 get_pi_state(pi_state);
2906 /* Unqueue and drop the lock */
2910 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2911 put_pi_state(pi_state);
2921 hrtimer_cancel(&to->timer);
2922 destroy_hrtimer_on_stack(&to->timer);
2924 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2929 ret = fault_in_user_writeable(uaddr);
2933 if (!(flags & FLAGS_SHARED))
2940 * Userspace attempted a TID -> 0 atomic transition, and failed.
2941 * This is the in-kernel slowpath: we look up the PI state (if any),
2942 * and do the rt-mutex unlock.
2944 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2946 u32 curval, uval, vpid = task_pid_vnr(current);
2947 union futex_key key = FUTEX_KEY_INIT;
2948 struct futex_hash_bucket *hb;
2949 struct futex_q *top_waiter;
2952 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2956 if (get_user(uval, uaddr))
2959 * We release only a lock we actually own:
2961 if ((uval & FUTEX_TID_MASK) != vpid)
2964 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2968 hb = hash_futex(&key);
2969 spin_lock(&hb->lock);
2972 * Check waiters first. We do not trust user space values at
2973 * all and we at least want to know if user space fiddled
2974 * with the futex value instead of blindly unlocking.
2976 top_waiter = futex_top_waiter(hb, &key);
2978 struct futex_pi_state *pi_state = top_waiter->pi_state;
2985 * If current does not own the pi_state then the futex is
2986 * inconsistent and user space fiddled with the futex value.
2988 if (pi_state->owner != current)
2991 get_pi_state(pi_state);
2993 * By taking wait_lock while still holding hb->lock, we ensure
2994 * there is no point where we hold neither; and therefore
2995 * wake_futex_pi() must observe a state consistent with what we
2998 * In particular; this forces __rt_mutex_start_proxy() to
2999 * complete such that we're guaranteed to observe the
3000 * rt_waiter. Also see the WARN in wake_futex_pi().
3002 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3003 spin_unlock(&hb->lock);
3005 /* drops pi_state->pi_mutex.wait_lock */
3006 ret = wake_futex_pi(uaddr, uval, pi_state);
3008 put_pi_state(pi_state);
3011 * Success, we're done! No tricky corner cases.
3016 * The atomic access to the futex value generated a
3017 * pagefault, so retry the user-access and the wakeup:
3022 * A unconditional UNLOCK_PI op raced against a waiter
3023 * setting the FUTEX_WAITERS bit. Try again.
3028 * wake_futex_pi has detected invalid state. Tell user
3035 * We have no kernel internal state, i.e. no waiters in the
3036 * kernel. Waiters which are about to queue themselves are stuck
3037 * on hb->lock. So we can safely ignore them. We do neither
3038 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3041 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3042 spin_unlock(&hb->lock);
3057 * If uval has changed, let user space handle it.
3059 ret = (curval == uval) ? 0 : -EAGAIN;
3062 spin_unlock(&hb->lock);
3072 ret = fault_in_user_writeable(uaddr);
3080 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3081 * @hb: the hash_bucket futex_q was original enqueued on
3082 * @q: the futex_q woken while waiting to be requeued
3083 * @key2: the futex_key of the requeue target futex
3084 * @timeout: the timeout associated with the wait (NULL if none)
3086 * Detect if the task was woken on the initial futex as opposed to the requeue
3087 * target futex. If so, determine if it was a timeout or a signal that caused
3088 * the wakeup and return the appropriate error code to the caller. Must be
3089 * called with the hb lock held.
3092 * - 0 = no early wakeup detected;
3093 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3096 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3097 struct futex_q *q, union futex_key *key2,
3098 struct hrtimer_sleeper *timeout)
3103 * With the hb lock held, we avoid races while we process the wakeup.
3104 * We only need to hold hb (and not hb2) to ensure atomicity as the
3105 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3106 * It can't be requeued from uaddr2 to something else since we don't
3107 * support a PI aware source futex for requeue.
3109 if (!match_futex(&q->key, key2)) {
3110 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3112 * We were woken prior to requeue by a timeout or a signal.
3113 * Unqueue the futex_q and determine which it was.
3115 plist_del(&q->list, &hb->chain);
3118 /* Handle spurious wakeups gracefully */
3120 if (timeout && !timeout->task)
3122 else if (signal_pending(current))
3123 ret = -ERESTARTNOINTR;
3129 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3130 * @uaddr: the futex we initially wait on (non-pi)
3131 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3132 * the same type, no requeueing from private to shared, etc.
3133 * @val: the expected value of uaddr
3134 * @abs_time: absolute timeout
3135 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3136 * @uaddr2: the pi futex we will take prior to returning to user-space
3138 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3139 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3140 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3141 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3142 * without one, the pi logic would not know which task to boost/deboost, if
3143 * there was a need to.
3145 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3146 * via the following--
3147 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3148 * 2) wakeup on uaddr2 after a requeue
3152 * If 3, cleanup and return -ERESTARTNOINTR.
3154 * If 2, we may then block on trying to take the rt_mutex and return via:
3155 * 5) successful lock
3158 * 8) other lock acquisition failure
3160 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3162 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3168 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3169 u32 val, ktime_t *abs_time, u32 bitset,
3172 struct hrtimer_sleeper timeout, *to;
3173 struct futex_pi_state *pi_state = NULL;
3174 struct rt_mutex_waiter rt_waiter;
3175 struct futex_hash_bucket *hb;
3176 union futex_key key2 = FUTEX_KEY_INIT;
3177 struct futex_q q = futex_q_init;
3180 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3183 if (uaddr == uaddr2)
3189 to = futex_setup_timer(abs_time, &timeout, flags,
3190 current->timer_slack_ns);
3193 * The waiter is allocated on our stack, manipulated by the requeue
3194 * code while we sleep on uaddr.
3196 rt_mutex_init_waiter(&rt_waiter);
3198 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3199 if (unlikely(ret != 0))
3203 q.rt_waiter = &rt_waiter;
3204 q.requeue_pi_key = &key2;
3207 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3210 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3215 * The check above which compares uaddrs is not sufficient for
3216 * shared futexes. We need to compare the keys:
3218 if (match_futex(&q.key, &key2)) {
3224 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3225 futex_wait_queue_me(hb, &q, to);
3227 spin_lock(&hb->lock);
3228 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3229 spin_unlock(&hb->lock);
3234 * In order for us to be here, we know our q.key == key2, and since
3235 * we took the hb->lock above, we also know that futex_requeue() has
3236 * completed and we no longer have to concern ourselves with a wakeup
3237 * race with the atomic proxy lock acquisition by the requeue code. The
3238 * futex_requeue dropped our key1 reference and incremented our key2
3242 /* Check if the requeue code acquired the second futex for us. */
3245 * Got the lock. We might not be the anticipated owner if we
3246 * did a lock-steal - fix up the PI-state in that case.
3248 if (q.pi_state && (q.pi_state->owner != current)) {
3249 spin_lock(q.lock_ptr);
3250 ret = fixup_pi_state_owner(uaddr2, &q, current);
3251 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3252 pi_state = q.pi_state;
3253 get_pi_state(pi_state);
3256 * Drop the reference to the pi state which
3257 * the requeue_pi() code acquired for us.
3259 put_pi_state(q.pi_state);
3260 spin_unlock(q.lock_ptr);
3263 struct rt_mutex *pi_mutex;
3266 * We have been woken up by futex_unlock_pi(), a timeout, or a
3267 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3270 WARN_ON(!q.pi_state);
3271 pi_mutex = &q.pi_state->pi_mutex;
3272 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3274 spin_lock(q.lock_ptr);
3275 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3278 debug_rt_mutex_free_waiter(&rt_waiter);
3280 * Fixup the pi_state owner and possibly acquire the lock if we
3283 res = fixup_owner(uaddr2, &q, !ret);
3285 * If fixup_owner() returned an error, proprogate that. If it
3286 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3289 ret = (res < 0) ? res : 0;
3292 * If fixup_pi_state_owner() faulted and was unable to handle
3293 * the fault, unlock the rt_mutex and return the fault to
3296 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3297 pi_state = q.pi_state;
3298 get_pi_state(pi_state);
3301 /* Unqueue and drop the lock. */
3306 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3307 put_pi_state(pi_state);
3310 if (ret == -EINTR) {
3312 * We've already been requeued, but cannot restart by calling
3313 * futex_lock_pi() directly. We could restart this syscall, but
3314 * it would detect that the user space "val" changed and return
3315 * -EWOULDBLOCK. Save the overhead of the restart and return
3316 * -EWOULDBLOCK directly.
3323 hrtimer_cancel(&to->timer);
3324 destroy_hrtimer_on_stack(&to->timer);
3330 * Support for robust futexes: the kernel cleans up held futexes at
3333 * Implementation: user-space maintains a per-thread list of locks it
3334 * is holding. Upon do_exit(), the kernel carefully walks this list,
3335 * and marks all locks that are owned by this thread with the
3336 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3337 * always manipulated with the lock held, so the list is private and
3338 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3339 * field, to allow the kernel to clean up if the thread dies after
3340 * acquiring the lock, but just before it could have added itself to
3341 * the list. There can only be one such pending lock.
3345 * sys_set_robust_list() - Set the robust-futex list head of a task
3346 * @head: pointer to the list-head
3347 * @len: length of the list-head, as userspace expects
3349 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3352 if (!futex_cmpxchg_enabled)
3355 * The kernel knows only one size for now:
3357 if (unlikely(len != sizeof(*head)))
3360 current->robust_list = head;
3366 * sys_get_robust_list() - Get the robust-futex list head of a task
3367 * @pid: pid of the process [zero for current task]
3368 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3369 * @len_ptr: pointer to a length field, the kernel fills in the header size
3371 SYSCALL_DEFINE3(get_robust_list, int, pid,
3372 struct robust_list_head __user * __user *, head_ptr,
3373 size_t __user *, len_ptr)
3375 struct robust_list_head __user *head;
3377 struct task_struct *p;
3379 if (!futex_cmpxchg_enabled)
3388 p = find_task_by_vpid(pid);
3394 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3397 head = p->robust_list;
3400 if (put_user(sizeof(*head), len_ptr))
3402 return put_user(head, head_ptr);
3410 /* Constants for the pending_op argument of handle_futex_death */
3411 #define HANDLE_DEATH_PENDING true
3412 #define HANDLE_DEATH_LIST false
3415 * Process a futex-list entry, check whether it's owned by the
3416 * dying task, and do notification if so:
3418 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3419 bool pi, bool pending_op)
3421 u32 uval, nval, mval;
3424 /* Futex address must be 32bit aligned */
3425 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3429 if (get_user(uval, uaddr))
3433 * Special case for regular (non PI) futexes. The unlock path in
3434 * user space has two race scenarios:
3436 * 1. The unlock path releases the user space futex value and
3437 * before it can execute the futex() syscall to wake up
3438 * waiters it is killed.
3440 * 2. A woken up waiter is killed before it can acquire the
3441 * futex in user space.
3443 * In both cases the TID validation below prevents a wakeup of
3444 * potential waiters which can cause these waiters to block
3447 * In both cases the following conditions are met:
3449 * 1) task->robust_list->list_op_pending != NULL
3450 * @pending_op == true
3451 * 2) User space futex value == 0
3452 * 3) Regular futex: @pi == false
3454 * If these conditions are met, it is safe to attempt waking up a
3455 * potential waiter without touching the user space futex value and
3456 * trying to set the OWNER_DIED bit. The user space futex value is
3457 * uncontended and the rest of the user space mutex state is
3458 * consistent, so a woken waiter will just take over the
3459 * uncontended futex. Setting the OWNER_DIED bit would create
3460 * inconsistent state and malfunction of the user space owner died
3463 if (pending_op && !pi && !uval) {
3464 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3468 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3472 * Ok, this dying thread is truly holding a futex
3473 * of interest. Set the OWNER_DIED bit atomically
3474 * via cmpxchg, and if the value had FUTEX_WAITERS
3475 * set, wake up a waiter (if any). (We have to do a
3476 * futex_wake() even if OWNER_DIED is already set -
3477 * to handle the rare but possible case of recursive
3478 * thread-death.) The rest of the cleanup is done in
3481 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3484 * We are not holding a lock here, but we want to have
3485 * the pagefault_disable/enable() protection because
3486 * we want to handle the fault gracefully. If the
3487 * access fails we try to fault in the futex with R/W
3488 * verification via get_user_pages. get_user() above
3489 * does not guarantee R/W access. If that fails we
3490 * give up and leave the futex locked.
3492 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3495 if (fault_in_user_writeable(uaddr))
3513 * Wake robust non-PI futexes here. The wakeup of
3514 * PI futexes happens in exit_pi_state():
3516 if (!pi && (uval & FUTEX_WAITERS))
3517 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3523 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3525 static inline int fetch_robust_entry(struct robust_list __user **entry,
3526 struct robust_list __user * __user *head,
3529 unsigned long uentry;
3531 if (get_user(uentry, (unsigned long __user *)head))
3534 *entry = (void __user *)(uentry & ~1UL);
3541 * Walk curr->robust_list (very carefully, it's a userspace list!)
3542 * and mark any locks found there dead, and notify any waiters.
3544 * We silently return on any sign of list-walking problem.
3546 static void exit_robust_list(struct task_struct *curr)
3548 struct robust_list_head __user *head = curr->robust_list;
3549 struct robust_list __user *entry, *next_entry, *pending;
3550 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3551 unsigned int next_pi;
3552 unsigned long futex_offset;
3555 if (!futex_cmpxchg_enabled)
3559 * Fetch the list head (which was registered earlier, via
3560 * sys_set_robust_list()):
3562 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3565 * Fetch the relative futex offset:
3567 if (get_user(futex_offset, &head->futex_offset))
3570 * Fetch any possibly pending lock-add first, and handle it
3573 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3576 next_entry = NULL; /* avoid warning with gcc */
3577 while (entry != &head->list) {
3579 * Fetch the next entry in the list before calling
3580 * handle_futex_death:
3582 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3584 * A pending lock might already be on the list, so
3585 * don't process it twice:
3587 if (entry != pending) {
3588 if (handle_futex_death((void __user *)entry + futex_offset,
3589 curr, pi, HANDLE_DEATH_LIST))
3597 * Avoid excessively long or circular lists:
3606 handle_futex_death((void __user *)pending + futex_offset,
3607 curr, pip, HANDLE_DEATH_PENDING);
3611 static void futex_cleanup(struct task_struct *tsk)
3613 if (unlikely(tsk->robust_list)) {
3614 exit_robust_list(tsk);
3615 tsk->robust_list = NULL;
3618 #ifdef CONFIG_COMPAT
3619 if (unlikely(tsk->compat_robust_list)) {
3620 compat_exit_robust_list(tsk);
3621 tsk->compat_robust_list = NULL;
3625 if (unlikely(!list_empty(&tsk->pi_state_list)))
3626 exit_pi_state_list(tsk);
3630 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3631 * @tsk: task to set the state on
3633 * Set the futex exit state of the task lockless. The futex waiter code
3634 * observes that state when a task is exiting and loops until the task has
3635 * actually finished the futex cleanup. The worst case for this is that the
3636 * waiter runs through the wait loop until the state becomes visible.
3638 * This is called from the recursive fault handling path in do_exit().
3640 * This is best effort. Either the futex exit code has run already or
3641 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3642 * take it over. If not, the problem is pushed back to user space. If the
3643 * futex exit code did not run yet, then an already queued waiter might
3644 * block forever, but there is nothing which can be done about that.
3646 void futex_exit_recursive(struct task_struct *tsk)
3648 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3649 if (tsk->futex_state == FUTEX_STATE_EXITING)
3650 mutex_unlock(&tsk->futex_exit_mutex);
3651 tsk->futex_state = FUTEX_STATE_DEAD;
3654 static void futex_cleanup_begin(struct task_struct *tsk)
3657 * Prevent various race issues against a concurrent incoming waiter
3658 * including live locks by forcing the waiter to block on
3659 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3660 * attach_to_pi_owner().
3662 mutex_lock(&tsk->futex_exit_mutex);
3665 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3667 * This ensures that all subsequent checks of tsk->futex_state in
3668 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3669 * tsk->pi_lock held.
3671 * It guarantees also that a pi_state which was queued right before
3672 * the state change under tsk->pi_lock by a concurrent waiter must
3673 * be observed in exit_pi_state_list().
3675 raw_spin_lock_irq(&tsk->pi_lock);
3676 tsk->futex_state = FUTEX_STATE_EXITING;
3677 raw_spin_unlock_irq(&tsk->pi_lock);
3680 static void futex_cleanup_end(struct task_struct *tsk, int state)
3683 * Lockless store. The only side effect is that an observer might
3684 * take another loop until it becomes visible.
3686 tsk->futex_state = state;
3688 * Drop the exit protection. This unblocks waiters which observed
3689 * FUTEX_STATE_EXITING to reevaluate the state.
3691 mutex_unlock(&tsk->futex_exit_mutex);
3694 void futex_exec_release(struct task_struct *tsk)
3697 * The state handling is done for consistency, but in the case of
3698 * exec() there is no way to prevent futher damage as the PID stays
3699 * the same. But for the unlikely and arguably buggy case that a
3700 * futex is held on exec(), this provides at least as much state
3701 * consistency protection which is possible.
3703 futex_cleanup_begin(tsk);
3706 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3707 * exec a new binary.
3709 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3712 void futex_exit_release(struct task_struct *tsk)
3714 futex_cleanup_begin(tsk);
3716 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3719 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3720 u32 __user *uaddr2, u32 val2, u32 val3)
3722 int cmd = op & FUTEX_CMD_MASK;
3723 unsigned int flags = 0;
3725 if (!(op & FUTEX_PRIVATE_FLAG))
3726 flags |= FLAGS_SHARED;
3728 if (op & FUTEX_CLOCK_REALTIME) {
3729 flags |= FLAGS_CLOCKRT;
3730 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3731 cmd != FUTEX_WAIT_REQUEUE_PI)
3737 case FUTEX_UNLOCK_PI:
3738 case FUTEX_TRYLOCK_PI:
3739 case FUTEX_WAIT_REQUEUE_PI:
3740 case FUTEX_CMP_REQUEUE_PI:
3741 if (!futex_cmpxchg_enabled)
3747 val3 = FUTEX_BITSET_MATCH_ANY;
3749 case FUTEX_WAIT_BITSET:
3750 return futex_wait(uaddr, flags, val, timeout, val3);
3752 val3 = FUTEX_BITSET_MATCH_ANY;
3754 case FUTEX_WAKE_BITSET:
3755 return futex_wake(uaddr, flags, val, val3);
3757 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3758 case FUTEX_CMP_REQUEUE:
3759 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3761 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3763 return futex_lock_pi(uaddr, flags, timeout, 0);
3764 case FUTEX_UNLOCK_PI:
3765 return futex_unlock_pi(uaddr, flags);
3766 case FUTEX_TRYLOCK_PI:
3767 return futex_lock_pi(uaddr, flags, NULL, 1);
3768 case FUTEX_WAIT_REQUEUE_PI:
3769 val3 = FUTEX_BITSET_MATCH_ANY;
3770 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3772 case FUTEX_CMP_REQUEUE_PI:
3773 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3779 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3780 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3783 struct timespec64 ts;
3784 ktime_t t, *tp = NULL;
3786 int cmd = op & FUTEX_CMD_MASK;
3788 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3789 cmd == FUTEX_WAIT_BITSET ||
3790 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3791 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3793 if (get_timespec64(&ts, utime))
3795 if (!timespec64_valid(&ts))
3798 t = timespec64_to_ktime(ts);
3799 if (cmd == FUTEX_WAIT)
3800 t = ktime_add_safe(ktime_get(), t);
3801 else if (!(op & FUTEX_CLOCK_REALTIME))
3802 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3806 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3807 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3809 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3810 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3811 val2 = (u32) (unsigned long) utime;
3813 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3816 #ifdef CONFIG_COMPAT
3818 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3821 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3822 compat_uptr_t __user *head, unsigned int *pi)
3824 if (get_user(*uentry, head))
3827 *entry = compat_ptr((*uentry) & ~1);
3828 *pi = (unsigned int)(*uentry) & 1;
3833 static void __user *futex_uaddr(struct robust_list __user *entry,
3834 compat_long_t futex_offset)
3836 compat_uptr_t base = ptr_to_compat(entry);
3837 void __user *uaddr = compat_ptr(base + futex_offset);
3843 * Walk curr->robust_list (very carefully, it's a userspace list!)
3844 * and mark any locks found there dead, and notify any waiters.
3846 * We silently return on any sign of list-walking problem.
3848 static void compat_exit_robust_list(struct task_struct *curr)
3850 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3851 struct robust_list __user *entry, *next_entry, *pending;
3852 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3853 unsigned int next_pi;
3854 compat_uptr_t uentry, next_uentry, upending;
3855 compat_long_t futex_offset;
3858 if (!futex_cmpxchg_enabled)
3862 * Fetch the list head (which was registered earlier, via
3863 * sys_set_robust_list()):
3865 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3868 * Fetch the relative futex offset:
3870 if (get_user(futex_offset, &head->futex_offset))
3873 * Fetch any possibly pending lock-add first, and handle it
3876 if (compat_fetch_robust_entry(&upending, &pending,
3877 &head->list_op_pending, &pip))
3880 next_entry = NULL; /* avoid warning with gcc */
3881 while (entry != (struct robust_list __user *) &head->list) {
3883 * Fetch the next entry in the list before calling
3884 * handle_futex_death:
3886 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3887 (compat_uptr_t __user *)&entry->next, &next_pi);
3889 * A pending lock might already be on the list, so
3890 * dont process it twice:
3892 if (entry != pending) {
3893 void __user *uaddr = futex_uaddr(entry, futex_offset);
3895 if (handle_futex_death(uaddr, curr, pi,
3901 uentry = next_uentry;
3905 * Avoid excessively long or circular lists:
3913 void __user *uaddr = futex_uaddr(pending, futex_offset);
3915 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3919 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3920 struct compat_robust_list_head __user *, head,
3923 if (!futex_cmpxchg_enabled)
3926 if (unlikely(len != sizeof(*head)))
3929 current->compat_robust_list = head;
3934 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3935 compat_uptr_t __user *, head_ptr,
3936 compat_size_t __user *, len_ptr)
3938 struct compat_robust_list_head __user *head;
3940 struct task_struct *p;
3942 if (!futex_cmpxchg_enabled)
3951 p = find_task_by_vpid(pid);
3957 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3960 head = p->compat_robust_list;
3963 if (put_user(sizeof(*head), len_ptr))
3965 return put_user(ptr_to_compat(head), head_ptr);
3972 #endif /* CONFIG_COMPAT */
3974 #ifdef CONFIG_COMPAT_32BIT_TIME
3975 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3976 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3979 struct timespec64 ts;
3980 ktime_t t, *tp = NULL;
3982 int cmd = op & FUTEX_CMD_MASK;
3984 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3985 cmd == FUTEX_WAIT_BITSET ||
3986 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3987 if (get_old_timespec32(&ts, utime))
3989 if (!timespec64_valid(&ts))
3992 t = timespec64_to_ktime(ts);
3993 if (cmd == FUTEX_WAIT)
3994 t = ktime_add_safe(ktime_get(), t);
3995 else if (!(op & FUTEX_CLOCK_REALTIME))
3996 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3999 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4000 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4001 val2 = (int) (unsigned long) utime;
4003 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4005 #endif /* CONFIG_COMPAT_32BIT_TIME */
4007 static void __init futex_detect_cmpxchg(void)
4009 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4013 * This will fail and we want it. Some arch implementations do
4014 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4015 * functionality. We want to know that before we call in any
4016 * of the complex code paths. Also we want to prevent
4017 * registration of robust lists in that case. NULL is
4018 * guaranteed to fault and we get -EFAULT on functional
4019 * implementation, the non-functional ones will return
4022 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4023 futex_cmpxchg_enabled = 1;
4027 static int __init futex_init(void)
4029 unsigned int futex_shift;
4032 #if CONFIG_BASE_SMALL
4033 futex_hashsize = 16;
4035 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4038 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4040 futex_hashsize < 256 ? HASH_SMALL : 0,
4042 futex_hashsize, futex_hashsize);
4043 futex_hashsize = 1UL << futex_shift;
4045 futex_detect_cmpxchg();
4047 for (i = 0; i < futex_hashsize; i++) {
4048 atomic_set(&futex_queues[i].waiters, 0);
4049 plist_head_init(&futex_queues[i].chain);
4050 spin_lock_init(&futex_queues[i].lock);
4055 core_initcall(futex_init);