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);
316 * Reflects a new waiter being added to the waitqueue.
318 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
321 atomic_inc(&hb->waiters);
323 * Full barrier (A), see the ordering comment above.
325 smp_mb__after_atomic();
330 * Reflects a waiter being removed from the waitqueue by wakeup
333 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
336 atomic_dec(&hb->waiters);
340 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
344 * Full barrier (B), see the ordering comment above.
347 return atomic_read(&hb->waiters);
354 * hash_futex - Return the hash bucket in the global hash
355 * @key: Pointer to the futex key for which the hash is calculated
357 * We hash on the keys returned from get_futex_key (see below) and return the
358 * corresponding hash bucket in the global hash.
360 static struct futex_hash_bucket *hash_futex(union futex_key *key)
362 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
365 return &futex_queues[hash & (futex_hashsize - 1)];
370 * match_futex - Check whether two futex keys are equal
371 * @key1: Pointer to key1
372 * @key2: Pointer to key2
374 * Return 1 if two futex_keys are equal, 0 otherwise.
376 static inline int match_futex(union futex_key *key1, union futex_key *key2)
379 && key1->both.word == key2->both.word
380 && key1->both.ptr == key2->both.ptr
381 && key1->both.offset == key2->both.offset);
390 * futex_setup_timer - set up the sleeping hrtimer.
391 * @time: ptr to the given timeout value
392 * @timeout: the hrtimer_sleeper structure to be set up
393 * @flags: futex flags
394 * @range_ns: optional range in ns
396 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
399 static inline struct hrtimer_sleeper *
400 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
401 int flags, u64 range_ns)
406 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
407 CLOCK_REALTIME : CLOCK_MONOTONIC,
410 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
411 * effectively the same as calling hrtimer_set_expires().
413 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
419 * Generate a machine wide unique identifier for this inode.
421 * This relies on u64 not wrapping in the life-time of the machine; which with
422 * 1ns resolution means almost 585 years.
424 * This further relies on the fact that a well formed program will not unmap
425 * the file while it has a (shared) futex waiting on it. This mapping will have
426 * a file reference which pins the mount and inode.
428 * If for some reason an inode gets evicted and read back in again, it will get
429 * a new sequence number and will _NOT_ match, even though it is the exact same
432 * It is important that match_futex() will never have a false-positive, esp.
433 * for PI futexes that can mess up the state. The above argues that false-negatives
434 * are only possible for malformed programs.
436 static u64 get_inode_sequence_number(struct inode *inode)
438 static atomic64_t i_seq;
441 /* Does the inode already have a sequence number? */
442 old = atomic64_read(&inode->i_sequence);
447 u64 new = atomic64_add_return(1, &i_seq);
448 if (WARN_ON_ONCE(!new))
451 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
459 * get_futex_key() - Get parameters which are the keys for a futex
460 * @uaddr: virtual address of the futex
461 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
462 * @key: address where result is stored.
463 * @rw: mapping needs to be read/write (values: FUTEX_READ,
466 * Return: a negative error code or 0
468 * The key words are stored in @key on success.
470 * For shared mappings (when @fshared), the key is:
472 * ( inode->i_sequence, page->index, offset_within_page )
474 * [ also see get_inode_sequence_number() ]
476 * For private mappings (or when !@fshared), the key is:
478 * ( current->mm, address, 0 )
480 * This allows (cross process, where applicable) identification of the futex
481 * without keeping the page pinned for the duration of the FUTEX_WAIT.
483 * lock_page() might sleep, the caller should not hold a spinlock.
485 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
486 enum futex_access rw)
488 unsigned long address = (unsigned long)uaddr;
489 struct mm_struct *mm = current->mm;
490 struct page *page, *tail;
491 struct address_space *mapping;
495 * The futex address must be "naturally" aligned.
497 key->both.offset = address % PAGE_SIZE;
498 if (unlikely((address % sizeof(u32)) != 0))
500 address -= key->both.offset;
502 if (unlikely(!access_ok(uaddr, sizeof(u32))))
505 if (unlikely(should_fail_futex(fshared)))
509 * PROCESS_PRIVATE futexes are fast.
510 * As the mm cannot disappear under us and the 'key' only needs
511 * virtual address, we dont even have to find the underlying vma.
512 * Note : We do have to check 'uaddr' is a valid user address,
513 * but access_ok() should be faster than find_vma()
516 key->private.mm = mm;
517 key->private.address = address;
522 /* Ignore any VERIFY_READ mapping (futex common case) */
523 if (unlikely(should_fail_futex(true)))
526 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
528 * If write access is not required (eg. FUTEX_WAIT), try
529 * and get read-only access.
531 if (err == -EFAULT && rw == FUTEX_READ) {
532 err = get_user_pages_fast(address, 1, 0, &page);
541 * The treatment of mapping from this point on is critical. The page
542 * lock protects many things but in this context the page lock
543 * stabilizes mapping, prevents inode freeing in the shared
544 * file-backed region case and guards against movement to swap cache.
546 * Strictly speaking the page lock is not needed in all cases being
547 * considered here and page lock forces unnecessarily serialization
548 * From this point on, mapping will be re-verified if necessary and
549 * page lock will be acquired only if it is unavoidable
551 * Mapping checks require the head page for any compound page so the
552 * head page and mapping is looked up now. For anonymous pages, it
553 * does not matter if the page splits in the future as the key is
554 * based on the address. For filesystem-backed pages, the tail is
555 * required as the index of the page determines the key. For
556 * base pages, there is no tail page and tail == page.
559 page = compound_head(page);
560 mapping = READ_ONCE(page->mapping);
563 * If page->mapping is NULL, then it cannot be a PageAnon
564 * page; but it might be the ZERO_PAGE or in the gate area or
565 * in a special mapping (all cases which we are happy to fail);
566 * or it may have been a good file page when get_user_pages_fast
567 * found it, but truncated or holepunched or subjected to
568 * invalidate_complete_page2 before we got the page lock (also
569 * cases which we are happy to fail). And we hold a reference,
570 * so refcount care in invalidate_complete_page's remove_mapping
571 * prevents drop_caches from setting mapping to NULL beneath us.
573 * The case we do have to guard against is when memory pressure made
574 * shmem_writepage move it from filecache to swapcache beneath us:
575 * an unlikely race, but we do need to retry for page->mapping.
577 if (unlikely(!mapping)) {
581 * Page lock is required to identify which special case above
582 * applies. If this is really a shmem page then the page lock
583 * will prevent unexpected transitions.
586 shmem_swizzled = PageSwapCache(page) || page->mapping;
597 * Private mappings are handled in a simple way.
599 * If the futex key is stored on an anonymous page, then the associated
600 * object is the mm which is implicitly pinned by the calling process.
602 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
603 * it's a read-only handle, it's expected that futexes attach to
604 * the object not the particular process.
606 if (PageAnon(page)) {
608 * A RO anonymous page will never change and thus doesn't make
609 * sense for futex operations.
611 if (unlikely(should_fail_futex(true)) || ro) {
616 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
617 key->private.mm = mm;
618 key->private.address = address;
624 * The associated futex object in this case is the inode and
625 * the page->mapping must be traversed. Ordinarily this should
626 * be stabilised under page lock but it's not strictly
627 * necessary in this case as we just want to pin the inode, not
628 * update the radix tree or anything like that.
630 * The RCU read lock is taken as the inode is finally freed
631 * under RCU. If the mapping still matches expectations then the
632 * mapping->host can be safely accessed as being a valid inode.
636 if (READ_ONCE(page->mapping) != mapping) {
643 inode = READ_ONCE(mapping->host);
651 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
652 key->shared.i_seq = get_inode_sequence_number(inode);
653 key->shared.pgoff = basepage_index(tail);
663 * fault_in_user_writeable() - Fault in user address and verify RW access
664 * @uaddr: pointer to faulting user space address
666 * Slow path to fixup the fault we just took in the atomic write
669 * We have no generic implementation of a non-destructive write to the
670 * user address. We know that we faulted in the atomic pagefault
671 * disabled section so we can as well avoid the #PF overhead by
672 * calling get_user_pages() right away.
674 static int fault_in_user_writeable(u32 __user *uaddr)
676 struct mm_struct *mm = current->mm;
680 ret = fixup_user_fault(mm, (unsigned long)uaddr,
681 FAULT_FLAG_WRITE, NULL);
682 mmap_read_unlock(mm);
684 return ret < 0 ? ret : 0;
688 * futex_top_waiter() - Return the highest priority waiter on a futex
689 * @hb: the hash bucket the futex_q's reside in
690 * @key: the futex key (to distinguish it from other futex futex_q's)
692 * Must be called with the hb lock held.
694 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
695 union futex_key *key)
697 struct futex_q *this;
699 plist_for_each_entry(this, &hb->chain, list) {
700 if (match_futex(&this->key, key))
706 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
707 u32 uval, u32 newval)
712 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
718 static int get_futex_value_locked(u32 *dest, u32 __user *from)
723 ret = __get_user(*dest, from);
726 return ret ? -EFAULT : 0;
733 static int refill_pi_state_cache(void)
735 struct futex_pi_state *pi_state;
737 if (likely(current->pi_state_cache))
740 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
745 INIT_LIST_HEAD(&pi_state->list);
746 /* pi_mutex gets initialized later */
747 pi_state->owner = NULL;
748 refcount_set(&pi_state->refcount, 1);
749 pi_state->key = FUTEX_KEY_INIT;
751 current->pi_state_cache = pi_state;
756 static struct futex_pi_state *alloc_pi_state(void)
758 struct futex_pi_state *pi_state = current->pi_state_cache;
761 current->pi_state_cache = NULL;
766 static void pi_state_update_owner(struct futex_pi_state *pi_state,
767 struct task_struct *new_owner)
769 struct task_struct *old_owner = pi_state->owner;
771 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
774 raw_spin_lock(&old_owner->pi_lock);
775 WARN_ON(list_empty(&pi_state->list));
776 list_del_init(&pi_state->list);
777 raw_spin_unlock(&old_owner->pi_lock);
781 raw_spin_lock(&new_owner->pi_lock);
782 WARN_ON(!list_empty(&pi_state->list));
783 list_add(&pi_state->list, &new_owner->pi_state_list);
784 pi_state->owner = new_owner;
785 raw_spin_unlock(&new_owner->pi_lock);
789 static void get_pi_state(struct futex_pi_state *pi_state)
791 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
795 * Drops a reference to the pi_state object and frees or caches it
796 * when the last reference is gone.
798 static void put_pi_state(struct futex_pi_state *pi_state)
803 if (!refcount_dec_and_test(&pi_state->refcount))
807 * If pi_state->owner is NULL, the owner is most probably dying
808 * and has cleaned up the pi_state already
810 if (pi_state->owner) {
813 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
814 pi_state_update_owner(pi_state, NULL);
815 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
816 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
819 if (current->pi_state_cache) {
823 * pi_state->list is already empty.
824 * clear pi_state->owner.
825 * refcount is at 0 - put it back to 1.
827 pi_state->owner = NULL;
828 refcount_set(&pi_state->refcount, 1);
829 current->pi_state_cache = pi_state;
833 #ifdef CONFIG_FUTEX_PI
836 * This task is holding PI mutexes at exit time => bad.
837 * Kernel cleans up PI-state, but userspace is likely hosed.
838 * (Robust-futex cleanup is separate and might save the day for userspace.)
840 static void exit_pi_state_list(struct task_struct *curr)
842 struct list_head *next, *head = &curr->pi_state_list;
843 struct futex_pi_state *pi_state;
844 struct futex_hash_bucket *hb;
845 union futex_key key = FUTEX_KEY_INIT;
847 if (!futex_cmpxchg_enabled)
850 * We are a ZOMBIE and nobody can enqueue itself on
851 * pi_state_list anymore, but we have to be careful
852 * versus waiters unqueueing themselves:
854 raw_spin_lock_irq(&curr->pi_lock);
855 while (!list_empty(head)) {
857 pi_state = list_entry(next, struct futex_pi_state, list);
859 hb = hash_futex(&key);
862 * We can race against put_pi_state() removing itself from the
863 * list (a waiter going away). put_pi_state() will first
864 * decrement the reference count and then modify the list, so
865 * its possible to see the list entry but fail this reference
868 * In that case; drop the locks to let put_pi_state() make
869 * progress and retry the loop.
871 if (!refcount_inc_not_zero(&pi_state->refcount)) {
872 raw_spin_unlock_irq(&curr->pi_lock);
874 raw_spin_lock_irq(&curr->pi_lock);
877 raw_spin_unlock_irq(&curr->pi_lock);
879 spin_lock(&hb->lock);
880 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
881 raw_spin_lock(&curr->pi_lock);
883 * We dropped the pi-lock, so re-check whether this
884 * task still owns the PI-state:
886 if (head->next != next) {
887 /* retain curr->pi_lock for the loop invariant */
888 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
889 spin_unlock(&hb->lock);
890 put_pi_state(pi_state);
894 WARN_ON(pi_state->owner != curr);
895 WARN_ON(list_empty(&pi_state->list));
896 list_del_init(&pi_state->list);
897 pi_state->owner = NULL;
899 raw_spin_unlock(&curr->pi_lock);
900 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
901 spin_unlock(&hb->lock);
903 rt_mutex_futex_unlock(&pi_state->pi_mutex);
904 put_pi_state(pi_state);
906 raw_spin_lock_irq(&curr->pi_lock);
908 raw_spin_unlock_irq(&curr->pi_lock);
911 static inline void exit_pi_state_list(struct task_struct *curr) { }
915 * We need to check the following states:
917 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
919 * [1] NULL | --- | --- | 0 | 0/1 | Valid
920 * [2] NULL | --- | --- | >0 | 0/1 | Valid
922 * [3] Found | NULL | -- | Any | 0/1 | Invalid
924 * [4] Found | Found | NULL | 0 | 1 | Valid
925 * [5] Found | Found | NULL | >0 | 1 | Invalid
927 * [6] Found | Found | task | 0 | 1 | Valid
929 * [7] Found | Found | NULL | Any | 0 | Invalid
931 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
932 * [9] Found | Found | task | 0 | 0 | Invalid
933 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
935 * [1] Indicates that the kernel can acquire the futex atomically. We
936 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
938 * [2] Valid, if TID does not belong to a kernel thread. If no matching
939 * thread is found then it indicates that the owner TID has died.
941 * [3] Invalid. The waiter is queued on a non PI futex
943 * [4] Valid state after exit_robust_list(), which sets the user space
944 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
946 * [5] The user space value got manipulated between exit_robust_list()
947 * and exit_pi_state_list()
949 * [6] Valid state after exit_pi_state_list() which sets the new owner in
950 * the pi_state but cannot access the user space value.
952 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
954 * [8] Owner and user space value match
956 * [9] There is no transient state which sets the user space TID to 0
957 * except exit_robust_list(), but this is indicated by the
958 * FUTEX_OWNER_DIED bit. See [4]
960 * [10] There is no transient state which leaves owner and user space
961 * TID out of sync. Except one error case where the kernel is denied
962 * write access to the user address, see fixup_pi_state_owner().
965 * Serialization and lifetime rules:
969 * hb -> futex_q, relation
970 * futex_q -> pi_state, relation
972 * (cannot be raw because hb can contain arbitrary amount
975 * pi_mutex->wait_lock:
979 * (and pi_mutex 'obviously')
983 * p->pi_state_list -> pi_state->list, relation
985 * pi_state->refcount:
993 * pi_mutex->wait_lock
999 * Validate that the existing waiter has a pi_state and sanity check
1000 * the pi_state against the user space value. If correct, attach to
1003 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1004 struct futex_pi_state *pi_state,
1005 struct futex_pi_state **ps)
1007 pid_t pid = uval & FUTEX_TID_MASK;
1012 * Userspace might have messed up non-PI and PI futexes [3]
1014 if (unlikely(!pi_state))
1018 * We get here with hb->lock held, and having found a
1019 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1020 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1021 * which in turn means that futex_lock_pi() still has a reference on
1024 * The waiter holding a reference on @pi_state also protects against
1025 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1026 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1027 * free pi_state before we can take a reference ourselves.
1029 WARN_ON(!refcount_read(&pi_state->refcount));
1032 * Now that we have a pi_state, we can acquire wait_lock
1033 * and do the state validation.
1035 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1038 * Since {uval, pi_state} is serialized by wait_lock, and our current
1039 * uval was read without holding it, it can have changed. Verify it
1040 * still is what we expect it to be, otherwise retry the entire
1043 if (get_futex_value_locked(&uval2, uaddr))
1050 * Handle the owner died case:
1052 if (uval & FUTEX_OWNER_DIED) {
1054 * exit_pi_state_list sets owner to NULL and wakes the
1055 * topmost waiter. The task which acquires the
1056 * pi_state->rt_mutex will fixup owner.
1058 if (!pi_state->owner) {
1060 * No pi state owner, but the user space TID
1061 * is not 0. Inconsistent state. [5]
1066 * Take a ref on the state and return success. [4]
1072 * If TID is 0, then either the dying owner has not
1073 * yet executed exit_pi_state_list() or some waiter
1074 * acquired the rtmutex in the pi state, but did not
1075 * yet fixup the TID in user space.
1077 * Take a ref on the state and return success. [6]
1083 * If the owner died bit is not set, then the pi_state
1084 * must have an owner. [7]
1086 if (!pi_state->owner)
1091 * Bail out if user space manipulated the futex value. If pi
1092 * state exists then the owner TID must be the same as the
1093 * user space TID. [9/10]
1095 if (pid != task_pid_vnr(pi_state->owner))
1099 get_pi_state(pi_state);
1100 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1117 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1122 * wait_for_owner_exiting - Block until the owner has exited
1123 * @ret: owner's current futex lock status
1124 * @exiting: Pointer to the exiting task
1126 * Caller must hold a refcount on @exiting.
1128 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1130 if (ret != -EBUSY) {
1131 WARN_ON_ONCE(exiting);
1135 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1138 mutex_lock(&exiting->futex_exit_mutex);
1140 * No point in doing state checking here. If the waiter got here
1141 * while the task was in exec()->exec_futex_release() then it can
1142 * have any FUTEX_STATE_* value when the waiter has acquired the
1143 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1144 * already. Highly unlikely and not a problem. Just one more round
1145 * through the futex maze.
1147 mutex_unlock(&exiting->futex_exit_mutex);
1149 put_task_struct(exiting);
1152 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1153 struct task_struct *tsk)
1158 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1159 * caller that the alleged owner is busy.
1161 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1165 * Reread the user space value to handle the following situation:
1169 * sys_exit() sys_futex()
1170 * do_exit() futex_lock_pi()
1171 * futex_lock_pi_atomic()
1172 * exit_signals(tsk) No waiters:
1173 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1174 * mm_release(tsk) Set waiter bit
1175 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1176 * Set owner died attach_to_pi_owner() {
1177 * *uaddr = 0xC0000000; tsk = get_task(PID);
1178 * } if (!tsk->flags & PF_EXITING) {
1180 * tsk->futex_state = } else {
1181 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1184 * return -ESRCH; <--- FAIL
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1190 * The same logic applies to the case where the exiting task is
1193 if (get_futex_value_locked(&uval2, uaddr))
1196 /* If the user space value has changed, try again. */
1201 * The exiting task did not have a robust list, the robust list was
1202 * corrupted or the user space value in *uaddr is simply bogus.
1203 * Give up and tell user space.
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1212 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213 struct futex_pi_state **ps,
1214 struct task_struct **exiting)
1216 pid_t pid = uval & FUTEX_TID_MASK;
1217 struct futex_pi_state *pi_state;
1218 struct task_struct *p;
1221 * We are the first waiter - try to look up the real owner and attach
1222 * the new pi_state to it, but bail out when TID = 0 [1]
1224 * The !pid check is paranoid. None of the call sites should end up
1225 * with pid == 0, but better safe than sorry. Let the caller retry
1229 p = find_get_task_by_vpid(pid);
1231 return handle_exit_race(uaddr, uval, NULL);
1233 if (unlikely(p->flags & PF_KTHREAD)) {
1239 * We need to look at the task state to figure out, whether the
1240 * task is exiting. To protect against the change of the task state
1241 * in futex_exit_release(), we do this protected by p->pi_lock:
1243 raw_spin_lock_irq(&p->pi_lock);
1244 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1246 * The task is on the way out. When the futex state is
1247 * FUTEX_STATE_DEAD, we know that the task has finished
1250 int ret = handle_exit_race(uaddr, uval, p);
1252 raw_spin_unlock_irq(&p->pi_lock);
1254 * If the owner task is between FUTEX_STATE_EXITING and
1255 * FUTEX_STATE_DEAD then store the task pointer and keep
1256 * the reference on the task struct. The calling code will
1257 * drop all locks, wait for the task to reach
1258 * FUTEX_STATE_DEAD and then drop the refcount. This is
1259 * required to prevent a live lock when the current task
1260 * preempted the exiting task between the two states.
1270 * No existing pi state. First waiter. [2]
1272 * This creates pi_state, we have hb->lock held, this means nothing can
1273 * observe this state, wait_lock is irrelevant.
1275 pi_state = alloc_pi_state();
1278 * Initialize the pi_mutex in locked state and make @p
1281 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1283 /* Store the key for possible exit cleanups: */
1284 pi_state->key = *key;
1286 WARN_ON(!list_empty(&pi_state->list));
1287 list_add(&pi_state->list, &p->pi_state_list);
1289 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1290 * because there is no concurrency as the object is not published yet.
1292 pi_state->owner = p;
1293 raw_spin_unlock_irq(&p->pi_lock);
1302 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1303 struct futex_hash_bucket *hb,
1304 union futex_key *key, struct futex_pi_state **ps,
1305 struct task_struct **exiting)
1307 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1310 * If there is a waiter on that futex, validate it and
1311 * attach to the pi_state when the validation succeeds.
1314 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1317 * We are the first waiter - try to look up the owner based on
1318 * @uval and attach to it.
1320 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1323 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1328 if (unlikely(should_fail_futex(true)))
1331 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1335 /* If user space value changed, let the caller retry */
1336 return curval != uval ? -EAGAIN : 0;
1340 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1341 * @uaddr: the pi futex user address
1342 * @hb: the pi futex hash bucket
1343 * @key: the futex key associated with uaddr and hb
1344 * @ps: the pi_state pointer where we store the result of the
1346 * @task: the task to perform the atomic lock work for. This will
1347 * be "current" except in the case of requeue pi.
1348 * @exiting: Pointer to store the task pointer of the owner task
1349 * which is in the middle of exiting
1350 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1353 * - 0 - ready to wait;
1354 * - 1 - acquired the lock;
1357 * The hb->lock and futex_key refs shall be held by the caller.
1359 * @exiting is only set when the return value is -EBUSY. If so, this holds
1360 * a refcount on the exiting task on return and the caller needs to drop it
1361 * after waiting for the exit to complete.
1363 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1364 union futex_key *key,
1365 struct futex_pi_state **ps,
1366 struct task_struct *task,
1367 struct task_struct **exiting,
1370 u32 uval, newval, vpid = task_pid_vnr(task);
1371 struct futex_q *top_waiter;
1375 * Read the user space value first so we can validate a few
1376 * things before proceeding further.
1378 if (get_futex_value_locked(&uval, uaddr))
1381 if (unlikely(should_fail_futex(true)))
1387 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1390 if ((unlikely(should_fail_futex(true))))
1394 * Lookup existing state first. If it exists, try to attach to
1397 top_waiter = futex_top_waiter(hb, key);
1399 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1402 * No waiter and user TID is 0. We are here because the
1403 * waiters or the owner died bit is set or called from
1404 * requeue_cmp_pi or for whatever reason something took the
1407 if (!(uval & FUTEX_TID_MASK)) {
1409 * We take over the futex. No other waiters and the user space
1410 * TID is 0. We preserve the owner died bit.
1412 newval = uval & FUTEX_OWNER_DIED;
1415 /* The futex requeue_pi code can enforce the waiters bit */
1417 newval |= FUTEX_WAITERS;
1419 ret = lock_pi_update_atomic(uaddr, uval, newval);
1420 /* If the take over worked, return 1 */
1421 return ret < 0 ? ret : 1;
1425 * First waiter. Set the waiters bit before attaching ourself to
1426 * the owner. If owner tries to unlock, it will be forced into
1427 * the kernel and blocked on hb->lock.
1429 newval = uval | FUTEX_WAITERS;
1430 ret = lock_pi_update_atomic(uaddr, uval, newval);
1434 * If the update of the user space value succeeded, we try to
1435 * attach to the owner. If that fails, no harm done, we only
1436 * set the FUTEX_WAITERS bit in the user space variable.
1438 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1442 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1443 * @q: The futex_q to unqueue
1445 * The q->lock_ptr must not be NULL and must be held by the caller.
1447 static void __unqueue_futex(struct futex_q *q)
1449 struct futex_hash_bucket *hb;
1451 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1453 lockdep_assert_held(q->lock_ptr);
1455 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1456 plist_del(&q->list, &hb->chain);
1461 * The hash bucket lock must be held when this is called.
1462 * Afterwards, the futex_q must not be accessed. Callers
1463 * must ensure to later call wake_up_q() for the actual
1466 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1468 struct task_struct *p = q->task;
1470 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1476 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1477 * is written, without taking any locks. This is possible in the event
1478 * of a spurious wakeup, for example. A memory barrier is required here
1479 * to prevent the following store to lock_ptr from getting ahead of the
1480 * plist_del in __unqueue_futex().
1482 smp_store_release(&q->lock_ptr, NULL);
1485 * Queue the task for later wakeup for after we've released
1488 wake_q_add_safe(wake_q, p);
1492 * Caller must hold a reference on @pi_state.
1494 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1497 struct rt_mutex_waiter *top_waiter;
1498 struct task_struct *new_owner;
1499 bool postunlock = false;
1500 DEFINE_WAKE_Q(wake_q);
1503 top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1504 if (WARN_ON_ONCE(!top_waiter)) {
1506 * As per the comment in futex_unlock_pi() this should not happen.
1508 * When this happens, give up our locks and try again, giving
1509 * the futex_lock_pi() instance time to complete, either by
1510 * waiting on the rtmutex or removing itself from the futex
1517 new_owner = top_waiter->task;
1520 * We pass it to the next owner. The WAITERS bit is always kept
1521 * enabled while there is PI state around. We cleanup the owner
1522 * died bit, because we are the owner.
1524 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1526 if (unlikely(should_fail_futex(true))) {
1531 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1532 if (!ret && (curval != uval)) {
1534 * If a unconditional UNLOCK_PI operation (user space did not
1535 * try the TID->0 transition) raced with a waiter setting the
1536 * FUTEX_WAITERS flag between get_user() and locking the hash
1537 * bucket lock, retry the operation.
1539 if ((FUTEX_TID_MASK & curval) == uval)
1547 * This is a point of no return; once we modified the uval
1548 * there is no going back and subsequent operations must
1551 pi_state_update_owner(pi_state, new_owner);
1552 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1556 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1559 rt_mutex_postunlock(&wake_q);
1565 * Express the locking dependencies for lockdep:
1568 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1571 spin_lock(&hb1->lock);
1573 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1574 } else { /* hb1 > hb2 */
1575 spin_lock(&hb2->lock);
1576 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1581 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1583 spin_unlock(&hb1->lock);
1585 spin_unlock(&hb2->lock);
1589 * Wake up waiters matching bitset queued on this futex (uaddr).
1592 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1594 struct futex_hash_bucket *hb;
1595 struct futex_q *this, *next;
1596 union futex_key key = FUTEX_KEY_INIT;
1598 DEFINE_WAKE_Q(wake_q);
1603 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1604 if (unlikely(ret != 0))
1607 hb = hash_futex(&key);
1609 /* Make sure we really have tasks to wakeup */
1610 if (!hb_waiters_pending(hb))
1613 spin_lock(&hb->lock);
1615 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1616 if (match_futex (&this->key, &key)) {
1617 if (this->pi_state || this->rt_waiter) {
1622 /* Check if one of the bits is set in both bitsets */
1623 if (!(this->bitset & bitset))
1626 mark_wake_futex(&wake_q, this);
1627 if (++ret >= nr_wake)
1632 spin_unlock(&hb->lock);
1637 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1639 unsigned int op = (encoded_op & 0x70000000) >> 28;
1640 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1641 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1642 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1645 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1646 if (oparg < 0 || oparg > 31) {
1647 char comm[sizeof(current->comm)];
1649 * kill this print and return -EINVAL when userspace
1652 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1653 get_task_comm(comm, current), oparg);
1659 pagefault_disable();
1660 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1666 case FUTEX_OP_CMP_EQ:
1667 return oldval == cmparg;
1668 case FUTEX_OP_CMP_NE:
1669 return oldval != cmparg;
1670 case FUTEX_OP_CMP_LT:
1671 return oldval < cmparg;
1672 case FUTEX_OP_CMP_GE:
1673 return oldval >= cmparg;
1674 case FUTEX_OP_CMP_LE:
1675 return oldval <= cmparg;
1676 case FUTEX_OP_CMP_GT:
1677 return oldval > cmparg;
1684 * Wake up all waiters hashed on the physical page that is mapped
1685 * to this virtual address:
1688 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1689 int nr_wake, int nr_wake2, int op)
1691 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1692 struct futex_hash_bucket *hb1, *hb2;
1693 struct futex_q *this, *next;
1695 DEFINE_WAKE_Q(wake_q);
1698 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1699 if (unlikely(ret != 0))
1701 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1702 if (unlikely(ret != 0))
1705 hb1 = hash_futex(&key1);
1706 hb2 = hash_futex(&key2);
1709 double_lock_hb(hb1, hb2);
1710 op_ret = futex_atomic_op_inuser(op, uaddr2);
1711 if (unlikely(op_ret < 0)) {
1712 double_unlock_hb(hb1, hb2);
1714 if (!IS_ENABLED(CONFIG_MMU) ||
1715 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1717 * we don't get EFAULT from MMU faults if we don't have
1718 * an MMU, but we might get them from range checking
1724 if (op_ret == -EFAULT) {
1725 ret = fault_in_user_writeable(uaddr2);
1730 if (!(flags & FLAGS_SHARED)) {
1739 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1740 if (match_futex (&this->key, &key1)) {
1741 if (this->pi_state || this->rt_waiter) {
1745 mark_wake_futex(&wake_q, this);
1746 if (++ret >= nr_wake)
1753 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1754 if (match_futex (&this->key, &key2)) {
1755 if (this->pi_state || this->rt_waiter) {
1759 mark_wake_futex(&wake_q, this);
1760 if (++op_ret >= nr_wake2)
1768 double_unlock_hb(hb1, hb2);
1774 * requeue_futex() - Requeue a futex_q from one hb to another
1775 * @q: the futex_q to requeue
1776 * @hb1: the source hash_bucket
1777 * @hb2: the target hash_bucket
1778 * @key2: the new key for the requeued futex_q
1781 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1782 struct futex_hash_bucket *hb2, union futex_key *key2)
1786 * If key1 and key2 hash to the same bucket, no need to
1789 if (likely(&hb1->chain != &hb2->chain)) {
1790 plist_del(&q->list, &hb1->chain);
1791 hb_waiters_dec(hb1);
1792 hb_waiters_inc(hb2);
1793 plist_add(&q->list, &hb2->chain);
1794 q->lock_ptr = &hb2->lock;
1800 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1802 * @key: the key of the requeue target futex
1803 * @hb: the hash_bucket of the requeue target futex
1805 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1806 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1807 * to the requeue target futex so the waiter can detect the wakeup on the right
1808 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1809 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1810 * to protect access to the pi_state to fixup the owner later. Must be called
1811 * with both q->lock_ptr and hb->lock held.
1814 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1815 struct futex_hash_bucket *hb)
1821 WARN_ON(!q->rt_waiter);
1822 q->rt_waiter = NULL;
1824 q->lock_ptr = &hb->lock;
1826 wake_up_state(q->task, TASK_NORMAL);
1830 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1831 * @pifutex: the user address of the to futex
1832 * @hb1: the from futex hash bucket, must be locked by the caller
1833 * @hb2: the to futex hash bucket, must be locked by the caller
1834 * @key1: the from futex key
1835 * @key2: the to futex key
1836 * @ps: address to store the pi_state pointer
1837 * @exiting: Pointer to store the task pointer of the owner task
1838 * which is in the middle of exiting
1839 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1841 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1842 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1843 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1844 * hb1 and hb2 must be held by the caller.
1846 * @exiting is only set when the return value is -EBUSY. If so, this holds
1847 * a refcount on the exiting task on return and the caller needs to drop it
1848 * after waiting for the exit to complete.
1851 * - 0 - failed to acquire the lock atomically;
1852 * - >0 - acquired the lock, return value is vpid of the top_waiter
1856 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1857 struct futex_hash_bucket *hb2, union futex_key *key1,
1858 union futex_key *key2, struct futex_pi_state **ps,
1859 struct task_struct **exiting, int set_waiters)
1861 struct futex_q *top_waiter = NULL;
1865 if (get_futex_value_locked(&curval, pifutex))
1868 if (unlikely(should_fail_futex(true)))
1872 * Find the top_waiter and determine if there are additional waiters.
1873 * If the caller intends to requeue more than 1 waiter to pifutex,
1874 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1875 * as we have means to handle the possible fault. If not, don't set
1876 * the bit unecessarily as it will force the subsequent unlock to enter
1879 top_waiter = futex_top_waiter(hb1, key1);
1881 /* There are no waiters, nothing for us to do. */
1885 /* Ensure we requeue to the expected futex. */
1886 if (!match_futex(top_waiter->requeue_pi_key, key2))
1890 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1891 * the contended case or if set_waiters is 1. The pi_state is returned
1892 * in ps in contended cases.
1894 vpid = task_pid_vnr(top_waiter->task);
1895 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1896 exiting, set_waiters);
1898 requeue_pi_wake_futex(top_waiter, key2, hb2);
1905 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1906 * @uaddr1: source futex user address
1907 * @flags: futex flags (FLAGS_SHARED, etc.)
1908 * @uaddr2: target futex user address
1909 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1910 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1911 * @cmpval: @uaddr1 expected value (or %NULL)
1912 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1913 * pi futex (pi to pi requeue is not supported)
1915 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1916 * uaddr2 atomically on behalf of the top waiter.
1919 * - >=0 - on success, the number of tasks requeued or woken;
1922 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1923 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1924 u32 *cmpval, int requeue_pi)
1926 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1927 int task_count = 0, ret;
1928 struct futex_pi_state *pi_state = NULL;
1929 struct futex_hash_bucket *hb1, *hb2;
1930 struct futex_q *this, *next;
1931 DEFINE_WAKE_Q(wake_q);
1933 if (nr_wake < 0 || nr_requeue < 0)
1937 * When PI not supported: return -ENOSYS if requeue_pi is true,
1938 * consequently the compiler knows requeue_pi is always false past
1939 * this point which will optimize away all the conditional code
1942 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1947 * Requeue PI only works on two distinct uaddrs. This
1948 * check is only valid for private futexes. See below.
1950 if (uaddr1 == uaddr2)
1954 * requeue_pi requires a pi_state, try to allocate it now
1955 * without any locks in case it fails.
1957 if (refill_pi_state_cache())
1960 * requeue_pi must wake as many tasks as it can, up to nr_wake
1961 * + nr_requeue, since it acquires the rt_mutex prior to
1962 * returning to userspace, so as to not leave the rt_mutex with
1963 * waiters and no owner. However, second and third wake-ups
1964 * cannot be predicted as they involve race conditions with the
1965 * first wake and a fault while looking up the pi_state. Both
1966 * pthread_cond_signal() and pthread_cond_broadcast() should
1974 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1975 if (unlikely(ret != 0))
1977 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1978 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1979 if (unlikely(ret != 0))
1983 * The check above which compares uaddrs is not sufficient for
1984 * shared futexes. We need to compare the keys:
1986 if (requeue_pi && match_futex(&key1, &key2))
1989 hb1 = hash_futex(&key1);
1990 hb2 = hash_futex(&key2);
1993 hb_waiters_inc(hb2);
1994 double_lock_hb(hb1, hb2);
1996 if (likely(cmpval != NULL)) {
1999 ret = get_futex_value_locked(&curval, uaddr1);
2001 if (unlikely(ret)) {
2002 double_unlock_hb(hb1, hb2);
2003 hb_waiters_dec(hb2);
2005 ret = get_user(curval, uaddr1);
2009 if (!(flags & FLAGS_SHARED))
2014 if (curval != *cmpval) {
2020 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2021 struct task_struct *exiting = NULL;
2024 * Attempt to acquire uaddr2 and wake the top waiter. If we
2025 * intend to requeue waiters, force setting the FUTEX_WAITERS
2026 * bit. We force this here where we are able to easily handle
2027 * faults rather in the requeue loop below.
2029 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2031 &exiting, nr_requeue);
2034 * At this point the top_waiter has either taken uaddr2 or is
2035 * waiting on it. If the former, then the pi_state will not
2036 * exist yet, look it up one more time to ensure we have a
2037 * reference to it. If the lock was taken, ret contains the
2038 * vpid of the top waiter task.
2039 * If the lock was not taken, we have pi_state and an initial
2040 * refcount on it. In case of an error we have nothing.
2046 * If we acquired the lock, then the user space value
2047 * of uaddr2 should be vpid. It cannot be changed by
2048 * the top waiter as it is blocked on hb2 lock if it
2049 * tries to do so. If something fiddled with it behind
2050 * our back the pi state lookup might unearth it. So
2051 * we rather use the known value than rereading and
2052 * handing potential crap to lookup_pi_state.
2054 * If that call succeeds then we have pi_state and an
2055 * initial refcount on it.
2057 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2058 &pi_state, &exiting);
2063 /* We hold a reference on the pi state. */
2066 /* If the above failed, then pi_state is NULL */
2068 double_unlock_hb(hb1, hb2);
2069 hb_waiters_dec(hb2);
2070 ret = fault_in_user_writeable(uaddr2);
2077 * Two reasons for this:
2078 * - EBUSY: Owner is exiting and we just wait for the
2080 * - EAGAIN: The user space value changed.
2082 double_unlock_hb(hb1, hb2);
2083 hb_waiters_dec(hb2);
2085 * Handle the case where the owner is in the middle of
2086 * exiting. Wait for the exit to complete otherwise
2087 * this task might loop forever, aka. live lock.
2089 wait_for_owner_exiting(ret, exiting);
2097 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2098 if (task_count - nr_wake >= nr_requeue)
2101 if (!match_futex(&this->key, &key1))
2105 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2106 * be paired with each other and no other futex ops.
2108 * We should never be requeueing a futex_q with a pi_state,
2109 * which is awaiting a futex_unlock_pi().
2111 if ((requeue_pi && !this->rt_waiter) ||
2112 (!requeue_pi && this->rt_waiter) ||
2119 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2120 * lock, we already woke the top_waiter. If not, it will be
2121 * woken by futex_unlock_pi().
2123 if (++task_count <= nr_wake && !requeue_pi) {
2124 mark_wake_futex(&wake_q, this);
2128 /* Ensure we requeue to the expected futex for requeue_pi. */
2129 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2135 * Requeue nr_requeue waiters and possibly one more in the case
2136 * of requeue_pi if we couldn't acquire the lock atomically.
2140 * Prepare the waiter to take the rt_mutex. Take a
2141 * refcount on the pi_state and store the pointer in
2142 * the futex_q object of the waiter.
2144 get_pi_state(pi_state);
2145 this->pi_state = pi_state;
2146 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2151 * We got the lock. We do neither drop the
2152 * refcount on pi_state nor clear
2153 * this->pi_state because the waiter needs the
2154 * pi_state for cleaning up the user space
2155 * value. It will drop the refcount after
2158 requeue_pi_wake_futex(this, &key2, hb2);
2162 * rt_mutex_start_proxy_lock() detected a
2163 * potential deadlock when we tried to queue
2164 * that waiter. Drop the pi_state reference
2165 * which we took above and remove the pointer
2166 * to the state from the waiters futex_q
2169 this->pi_state = NULL;
2170 put_pi_state(pi_state);
2172 * We stop queueing more waiters and let user
2173 * space deal with the mess.
2178 requeue_futex(this, hb1, hb2, &key2);
2182 * We took an extra initial reference to the pi_state either
2183 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2184 * need to drop it here again.
2186 put_pi_state(pi_state);
2189 double_unlock_hb(hb1, hb2);
2191 hb_waiters_dec(hb2);
2192 return ret ? ret : task_count;
2195 /* The key must be already stored in q->key. */
2196 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2197 __acquires(&hb->lock)
2199 struct futex_hash_bucket *hb;
2201 hb = hash_futex(&q->key);
2204 * Increment the counter before taking the lock so that
2205 * a potential waker won't miss a to-be-slept task that is
2206 * waiting for the spinlock. This is safe as all queue_lock()
2207 * users end up calling queue_me(). Similarly, for housekeeping,
2208 * decrement the counter at queue_unlock() when some error has
2209 * occurred and we don't end up adding the task to the list.
2211 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2213 q->lock_ptr = &hb->lock;
2215 spin_lock(&hb->lock);
2220 queue_unlock(struct futex_hash_bucket *hb)
2221 __releases(&hb->lock)
2223 spin_unlock(&hb->lock);
2227 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2232 * The priority used to register this element is
2233 * - either the real thread-priority for the real-time threads
2234 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2235 * - or MAX_RT_PRIO for non-RT threads.
2236 * Thus, all RT-threads are woken first in priority order, and
2237 * the others are woken last, in FIFO order.
2239 prio = min(current->normal_prio, MAX_RT_PRIO);
2241 plist_node_init(&q->list, prio);
2242 plist_add(&q->list, &hb->chain);
2247 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2248 * @q: The futex_q to enqueue
2249 * @hb: The destination hash bucket
2251 * The hb->lock must be held by the caller, and is released here. A call to
2252 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2253 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2254 * or nothing if the unqueue is done as part of the wake process and the unqueue
2255 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2258 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2259 __releases(&hb->lock)
2262 spin_unlock(&hb->lock);
2266 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2267 * @q: The futex_q to unqueue
2269 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2270 * be paired with exactly one earlier call to queue_me().
2273 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2274 * - 0 - if the futex_q was already removed by the waking thread
2276 static int unqueue_me(struct futex_q *q)
2278 spinlock_t *lock_ptr;
2281 /* In the common case we don't take the spinlock, which is nice. */
2284 * q->lock_ptr can change between this read and the following spin_lock.
2285 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2286 * optimizing lock_ptr out of the logic below.
2288 lock_ptr = READ_ONCE(q->lock_ptr);
2289 if (lock_ptr != NULL) {
2290 spin_lock(lock_ptr);
2292 * q->lock_ptr can change between reading it and
2293 * spin_lock(), causing us to take the wrong lock. This
2294 * corrects the race condition.
2296 * Reasoning goes like this: if we have the wrong lock,
2297 * q->lock_ptr must have changed (maybe several times)
2298 * between reading it and the spin_lock(). It can
2299 * change again after the spin_lock() but only if it was
2300 * already changed before the spin_lock(). It cannot,
2301 * however, change back to the original value. Therefore
2302 * we can detect whether we acquired the correct lock.
2304 if (unlikely(lock_ptr != q->lock_ptr)) {
2305 spin_unlock(lock_ptr);
2310 BUG_ON(q->pi_state);
2312 spin_unlock(lock_ptr);
2320 * PI futexes can not be requeued and must remove themself from the
2321 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2324 static void unqueue_me_pi(struct futex_q *q)
2325 __releases(q->lock_ptr)
2329 BUG_ON(!q->pi_state);
2330 put_pi_state(q->pi_state);
2333 spin_unlock(q->lock_ptr);
2336 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2337 struct task_struct *argowner)
2339 struct futex_pi_state *pi_state = q->pi_state;
2340 struct task_struct *oldowner, *newowner;
2341 u32 uval, curval, newval, newtid;
2344 oldowner = pi_state->owner;
2347 * We are here because either:
2349 * - we stole the lock and pi_state->owner needs updating to reflect
2350 * that (@argowner == current),
2354 * - someone stole our lock and we need to fix things to point to the
2355 * new owner (@argowner == NULL).
2357 * Either way, we have to replace the TID in the user space variable.
2358 * This must be atomic as we have to preserve the owner died bit here.
2360 * Note: We write the user space value _before_ changing the pi_state
2361 * because we can fault here. Imagine swapped out pages or a fork
2362 * that marked all the anonymous memory readonly for cow.
2364 * Modifying pi_state _before_ the user space value would leave the
2365 * pi_state in an inconsistent state when we fault here, because we
2366 * need to drop the locks to handle the fault. This might be observed
2367 * in the PID check in lookup_pi_state.
2371 if (oldowner != current) {
2373 * We raced against a concurrent self; things are
2374 * already fixed up. Nothing to do.
2379 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2380 /* We got the lock. pi_state is correct. Tell caller. */
2385 * The trylock just failed, so either there is an owner or
2386 * there is a higher priority waiter than this one.
2388 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2390 * If the higher priority waiter has not yet taken over the
2391 * rtmutex then newowner is NULL. We can't return here with
2392 * that state because it's inconsistent vs. the user space
2393 * state. So drop the locks and try again. It's a valid
2394 * situation and not any different from the other retry
2397 if (unlikely(!newowner)) {
2402 WARN_ON_ONCE(argowner != current);
2403 if (oldowner == current) {
2405 * We raced against a concurrent self; things are
2406 * already fixed up. Nothing to do.
2410 newowner = argowner;
2413 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2415 if (!pi_state->owner)
2416 newtid |= FUTEX_OWNER_DIED;
2418 err = get_futex_value_locked(&uval, uaddr);
2423 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2425 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2435 * We fixed up user space. Now we need to fix the pi_state
2438 pi_state_update_owner(pi_state, newowner);
2440 return argowner == current;
2443 * In order to reschedule or handle a page fault, we need to drop the
2444 * locks here. In the case of a fault, this gives the other task
2445 * (either the highest priority waiter itself or the task which stole
2446 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2447 * are back from handling the fault we need to check the pi_state after
2448 * reacquiring the locks and before trying to do another fixup. When
2449 * the fixup has been done already we simply return.
2451 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2452 * drop hb->lock since the caller owns the hb -> futex_q relation.
2453 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2456 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2457 spin_unlock(q->lock_ptr);
2461 err = fault_in_user_writeable(uaddr);
2474 spin_lock(q->lock_ptr);
2475 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2478 * Check if someone else fixed it for us:
2480 if (pi_state->owner != oldowner)
2481 return argowner == current;
2483 /* Retry if err was -EAGAIN or the fault in succeeded */
2488 * fault_in_user_writeable() failed so user state is immutable. At
2489 * best we can make the kernel state consistent but user state will
2490 * be most likely hosed and any subsequent unlock operation will be
2491 * rejected due to PI futex rule [10].
2493 * Ensure that the rtmutex owner is also the pi_state owner despite
2494 * the user space value claiming something different. There is no
2495 * point in unlocking the rtmutex if current is the owner as it
2496 * would need to wait until the next waiter has taken the rtmutex
2497 * to guarantee consistent state. Keep it simple. Userspace asked
2498 * for this wreckaged state.
2500 * The rtmutex has an owner - either current or some other
2501 * task. See the EAGAIN loop above.
2503 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2508 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2509 struct task_struct *argowner)
2511 struct futex_pi_state *pi_state = q->pi_state;
2514 lockdep_assert_held(q->lock_ptr);
2516 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2517 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2518 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2522 static long futex_wait_restart(struct restart_block *restart);
2525 * fixup_owner() - Post lock pi_state and corner case management
2526 * @uaddr: user address of the futex
2527 * @q: futex_q (contains pi_state and access to the rt_mutex)
2528 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2530 * After attempting to lock an rt_mutex, this function is called to cleanup
2531 * the pi_state owner as well as handle race conditions that may allow us to
2532 * acquire the lock. Must be called with the hb lock held.
2535 * - 1 - success, lock taken;
2536 * - 0 - success, lock not taken;
2537 * - <0 - on error (-EFAULT)
2539 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2543 * Got the lock. We might not be the anticipated owner if we
2544 * did a lock-steal - fix up the PI-state in that case:
2546 * Speculative pi_state->owner read (we don't hold wait_lock);
2547 * since we own the lock pi_state->owner == current is the
2548 * stable state, anything else needs more attention.
2550 if (q->pi_state->owner != current)
2551 return fixup_pi_state_owner(uaddr, q, current);
2556 * If we didn't get the lock; check if anybody stole it from us. In
2557 * that case, we need to fix up the uval to point to them instead of
2558 * us, otherwise bad things happen. [10]
2560 * Another speculative read; pi_state->owner == current is unstable
2561 * but needs our attention.
2563 if (q->pi_state->owner == current)
2564 return fixup_pi_state_owner(uaddr, q, NULL);
2567 * Paranoia check. If we did not take the lock, then we should not be
2568 * the owner of the rt_mutex. Warn and establish consistent state.
2570 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2571 return fixup_pi_state_owner(uaddr, q, current);
2577 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2578 * @hb: the futex hash bucket, must be locked by the caller
2579 * @q: the futex_q to queue up on
2580 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2582 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2583 struct hrtimer_sleeper *timeout)
2586 * The task state is guaranteed to be set before another task can
2587 * wake it. set_current_state() is implemented using smp_store_mb() and
2588 * queue_me() calls spin_unlock() upon completion, both serializing
2589 * access to the hash list and forcing another memory barrier.
2591 set_current_state(TASK_INTERRUPTIBLE);
2596 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2599 * If we have been removed from the hash list, then another task
2600 * has tried to wake us, and we can skip the call to schedule().
2602 if (likely(!plist_node_empty(&q->list))) {
2604 * If the timer has already expired, current will already be
2605 * flagged for rescheduling. Only call schedule if there
2606 * is no timeout, or if it has yet to expire.
2608 if (!timeout || timeout->task)
2609 freezable_schedule();
2611 __set_current_state(TASK_RUNNING);
2615 * futex_wait_setup() - Prepare to wait on a futex
2616 * @uaddr: the futex userspace address
2617 * @val: the expected value
2618 * @flags: futex flags (FLAGS_SHARED, etc.)
2619 * @q: the associated futex_q
2620 * @hb: storage for hash_bucket pointer to be returned to caller
2622 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2623 * compare it with the expected value. Handle atomic faults internally.
2624 * Return with the hb lock held and a q.key reference on success, and unlocked
2625 * with no q.key reference on failure.
2628 * - 0 - uaddr contains val and hb has been locked;
2629 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2631 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2632 struct futex_q *q, struct futex_hash_bucket **hb)
2638 * Access the page AFTER the hash-bucket is locked.
2639 * Order is important:
2641 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2642 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2644 * The basic logical guarantee of a futex is that it blocks ONLY
2645 * if cond(var) is known to be true at the time of blocking, for
2646 * any cond. If we locked the hash-bucket after testing *uaddr, that
2647 * would open a race condition where we could block indefinitely with
2648 * cond(var) false, which would violate the guarantee.
2650 * On the other hand, we insert q and release the hash-bucket only
2651 * after testing *uaddr. This guarantees that futex_wait() will NOT
2652 * absorb a wakeup if *uaddr does not match the desired values
2653 * while the syscall executes.
2656 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2657 if (unlikely(ret != 0))
2661 *hb = queue_lock(q);
2663 ret = get_futex_value_locked(&uval, uaddr);
2668 ret = get_user(uval, uaddr);
2672 if (!(flags & FLAGS_SHARED))
2686 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2687 ktime_t *abs_time, u32 bitset)
2689 struct hrtimer_sleeper timeout, *to;
2690 struct restart_block *restart;
2691 struct futex_hash_bucket *hb;
2692 struct futex_q q = futex_q_init;
2699 to = futex_setup_timer(abs_time, &timeout, flags,
2700 current->timer_slack_ns);
2703 * Prepare to wait on uaddr. On success, holds hb lock and increments
2706 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2710 /* queue_me and wait for wakeup, timeout, or a signal. */
2711 futex_wait_queue_me(hb, &q, to);
2713 /* If we were woken (and unqueued), we succeeded, whatever. */
2715 /* unqueue_me() drops q.key ref */
2716 if (!unqueue_me(&q))
2719 if (to && !to->task)
2723 * We expect signal_pending(current), but we might be the
2724 * victim of a spurious wakeup as well.
2726 if (!signal_pending(current))
2733 restart = ¤t->restart_block;
2734 restart->fn = futex_wait_restart;
2735 restart->futex.uaddr = uaddr;
2736 restart->futex.val = val;
2737 restart->futex.time = *abs_time;
2738 restart->futex.bitset = bitset;
2739 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2741 ret = -ERESTART_RESTARTBLOCK;
2745 hrtimer_cancel(&to->timer);
2746 destroy_hrtimer_on_stack(&to->timer);
2752 static long futex_wait_restart(struct restart_block *restart)
2754 u32 __user *uaddr = restart->futex.uaddr;
2755 ktime_t t, *tp = NULL;
2757 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2758 t = restart->futex.time;
2761 restart->fn = do_no_restart_syscall;
2763 return (long)futex_wait(uaddr, restart->futex.flags,
2764 restart->futex.val, tp, restart->futex.bitset);
2769 * Userspace tried a 0 -> TID atomic transition of the futex value
2770 * and failed. The kernel side here does the whole locking operation:
2771 * if there are waiters then it will block as a consequence of relying
2772 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2773 * a 0 value of the futex too.).
2775 * Also serves as futex trylock_pi()'ing, and due semantics.
2777 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2778 ktime_t *time, int trylock)
2780 struct hrtimer_sleeper timeout, *to;
2781 struct task_struct *exiting = NULL;
2782 struct rt_mutex_waiter rt_waiter;
2783 struct futex_hash_bucket *hb;
2784 struct futex_q q = futex_q_init;
2787 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2790 if (refill_pi_state_cache())
2793 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2796 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2797 if (unlikely(ret != 0))
2801 hb = queue_lock(&q);
2803 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2805 if (unlikely(ret)) {
2807 * Atomic work succeeded and we got the lock,
2808 * or failed. Either way, we do _not_ block.
2812 /* We got the lock. */
2814 goto out_unlock_put_key;
2820 * Two reasons for this:
2821 * - EBUSY: Task is exiting and we just wait for the
2823 * - EAGAIN: The user space value changed.
2827 * Handle the case where the owner is in the middle of
2828 * exiting. Wait for the exit to complete otherwise
2829 * this task might loop forever, aka. live lock.
2831 wait_for_owner_exiting(ret, exiting);
2835 goto out_unlock_put_key;
2839 WARN_ON(!q.pi_state);
2842 * Only actually queue now that the atomic ops are done:
2847 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2848 /* Fixup the trylock return value: */
2849 ret = ret ? 0 : -EWOULDBLOCK;
2853 rt_mutex_init_waiter(&rt_waiter);
2856 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2857 * hold it while doing rt_mutex_start_proxy(), because then it will
2858 * include hb->lock in the blocking chain, even through we'll not in
2859 * fact hold it while blocking. This will lead it to report -EDEADLK
2860 * and BUG when futex_unlock_pi() interleaves with this.
2862 * Therefore acquire wait_lock while holding hb->lock, but drop the
2863 * latter before calling __rt_mutex_start_proxy_lock(). This
2864 * interleaves with futex_unlock_pi() -- which does a similar lock
2865 * handoff -- such that the latter can observe the futex_q::pi_state
2866 * before __rt_mutex_start_proxy_lock() is done.
2868 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2869 spin_unlock(q.lock_ptr);
2871 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2872 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2873 * it sees the futex_q::pi_state.
2875 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2876 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2885 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2887 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2890 spin_lock(q.lock_ptr);
2892 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2893 * first acquire the hb->lock before removing the lock from the
2894 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2897 * In particular; it is important that futex_unlock_pi() can not
2898 * observe this inconsistency.
2900 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2905 * Fixup the pi_state owner and possibly acquire the lock if we
2908 res = fixup_owner(uaddr, &q, !ret);
2910 * If fixup_owner() returned an error, proprogate that. If it acquired
2911 * the lock, clear our -ETIMEDOUT or -EINTR.
2914 ret = (res < 0) ? res : 0;
2916 /* Unqueue and drop the lock */
2925 hrtimer_cancel(&to->timer);
2926 destroy_hrtimer_on_stack(&to->timer);
2928 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2933 ret = fault_in_user_writeable(uaddr);
2937 if (!(flags & FLAGS_SHARED))
2944 * Userspace attempted a TID -> 0 atomic transition, and failed.
2945 * This is the in-kernel slowpath: we look up the PI state (if any),
2946 * and do the rt-mutex unlock.
2948 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2950 u32 curval, uval, vpid = task_pid_vnr(current);
2951 union futex_key key = FUTEX_KEY_INIT;
2952 struct futex_hash_bucket *hb;
2953 struct futex_q *top_waiter;
2956 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2960 if (get_user(uval, uaddr))
2963 * We release only a lock we actually own:
2965 if ((uval & FUTEX_TID_MASK) != vpid)
2968 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2972 hb = hash_futex(&key);
2973 spin_lock(&hb->lock);
2976 * Check waiters first. We do not trust user space values at
2977 * all and we at least want to know if user space fiddled
2978 * with the futex value instead of blindly unlocking.
2980 top_waiter = futex_top_waiter(hb, &key);
2982 struct futex_pi_state *pi_state = top_waiter->pi_state;
2989 * If current does not own the pi_state then the futex is
2990 * inconsistent and user space fiddled with the futex value.
2992 if (pi_state->owner != current)
2995 get_pi_state(pi_state);
2997 * By taking wait_lock while still holding hb->lock, we ensure
2998 * there is no point where we hold neither; and therefore
2999 * wake_futex_pi() must observe a state consistent with what we
3002 * In particular; this forces __rt_mutex_start_proxy() to
3003 * complete such that we're guaranteed to observe the
3004 * rt_waiter. Also see the WARN in wake_futex_pi().
3006 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3007 spin_unlock(&hb->lock);
3009 /* drops pi_state->pi_mutex.wait_lock */
3010 ret = wake_futex_pi(uaddr, uval, pi_state);
3012 put_pi_state(pi_state);
3015 * Success, we're done! No tricky corner cases.
3020 * The atomic access to the futex value generated a
3021 * pagefault, so retry the user-access and the wakeup:
3026 * A unconditional UNLOCK_PI op raced against a waiter
3027 * setting the FUTEX_WAITERS bit. Try again.
3032 * wake_futex_pi has detected invalid state. Tell user
3039 * We have no kernel internal state, i.e. no waiters in the
3040 * kernel. Waiters which are about to queue themselves are stuck
3041 * on hb->lock. So we can safely ignore them. We do neither
3042 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3045 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3046 spin_unlock(&hb->lock);
3061 * If uval has changed, let user space handle it.
3063 ret = (curval == uval) ? 0 : -EAGAIN;
3066 spin_unlock(&hb->lock);
3075 ret = fault_in_user_writeable(uaddr);
3083 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3084 * @hb: the hash_bucket futex_q was original enqueued on
3085 * @q: the futex_q woken while waiting to be requeued
3086 * @key2: the futex_key of the requeue target futex
3087 * @timeout: the timeout associated with the wait (NULL if none)
3089 * Detect if the task was woken on the initial futex as opposed to the requeue
3090 * target futex. If so, determine if it was a timeout or a signal that caused
3091 * the wakeup and return the appropriate error code to the caller. Must be
3092 * called with the hb lock held.
3095 * - 0 = no early wakeup detected;
3096 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3099 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3100 struct futex_q *q, union futex_key *key2,
3101 struct hrtimer_sleeper *timeout)
3106 * With the hb lock held, we avoid races while we process the wakeup.
3107 * We only need to hold hb (and not hb2) to ensure atomicity as the
3108 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3109 * It can't be requeued from uaddr2 to something else since we don't
3110 * support a PI aware source futex for requeue.
3112 if (!match_futex(&q->key, key2)) {
3113 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3115 * We were woken prior to requeue by a timeout or a signal.
3116 * Unqueue the futex_q and determine which it was.
3118 plist_del(&q->list, &hb->chain);
3121 /* Handle spurious wakeups gracefully */
3123 if (timeout && !timeout->task)
3125 else if (signal_pending(current))
3126 ret = -ERESTARTNOINTR;
3132 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3133 * @uaddr: the futex we initially wait on (non-pi)
3134 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3135 * the same type, no requeueing from private to shared, etc.
3136 * @val: the expected value of uaddr
3137 * @abs_time: absolute timeout
3138 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3139 * @uaddr2: the pi futex we will take prior to returning to user-space
3141 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3142 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3143 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3144 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3145 * without one, the pi logic would not know which task to boost/deboost, if
3146 * there was a need to.
3148 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3149 * via the following--
3150 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3151 * 2) wakeup on uaddr2 after a requeue
3155 * If 3, cleanup and return -ERESTARTNOINTR.
3157 * If 2, we may then block on trying to take the rt_mutex and return via:
3158 * 5) successful lock
3161 * 8) other lock acquisition failure
3163 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3165 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3171 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3172 u32 val, ktime_t *abs_time, u32 bitset,
3175 struct hrtimer_sleeper timeout, *to;
3176 struct rt_mutex_waiter rt_waiter;
3177 struct futex_hash_bucket *hb;
3178 union futex_key key2 = FUTEX_KEY_INIT;
3179 struct futex_q q = futex_q_init;
3182 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3185 if (uaddr == uaddr2)
3191 to = futex_setup_timer(abs_time, &timeout, flags,
3192 current->timer_slack_ns);
3195 * The waiter is allocated on our stack, manipulated by the requeue
3196 * code while we sleep on uaddr.
3198 rt_mutex_init_waiter(&rt_waiter);
3200 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3201 if (unlikely(ret != 0))
3205 q.rt_waiter = &rt_waiter;
3206 q.requeue_pi_key = &key2;
3209 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3212 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3217 * The check above which compares uaddrs is not sufficient for
3218 * shared futexes. We need to compare the keys:
3220 if (match_futex(&q.key, &key2)) {
3226 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3227 futex_wait_queue_me(hb, &q, to);
3229 spin_lock(&hb->lock);
3230 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3231 spin_unlock(&hb->lock);
3236 * In order for us to be here, we know our q.key == key2, and since
3237 * we took the hb->lock above, we also know that futex_requeue() has
3238 * completed and we no longer have to concern ourselves with a wakeup
3239 * race with the atomic proxy lock acquisition by the requeue code. The
3240 * futex_requeue dropped our key1 reference and incremented our key2
3245 * Check if the requeue code acquired the second futex for us and do
3246 * any pertinent fixup.
3249 if (q.pi_state && (q.pi_state->owner != current)) {
3250 spin_lock(q.lock_ptr);
3251 ret = fixup_owner(uaddr2, &q, true);
3253 * Drop the reference to the pi state which
3254 * the requeue_pi() code acquired for us.
3256 put_pi_state(q.pi_state);
3257 spin_unlock(q.lock_ptr);
3259 * Adjust the return value. It's either -EFAULT or
3260 * success (1) but the caller expects 0 for success.
3262 ret = ret < 0 ? ret : 0;
3265 struct rt_mutex *pi_mutex;
3268 * We have been woken up by futex_unlock_pi(), a timeout, or a
3269 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3272 WARN_ON(!q.pi_state);
3273 pi_mutex = &q.pi_state->pi_mutex;
3274 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3276 spin_lock(q.lock_ptr);
3277 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3280 debug_rt_mutex_free_waiter(&rt_waiter);
3282 * Fixup the pi_state owner and possibly acquire the lock if we
3285 res = fixup_owner(uaddr2, &q, !ret);
3287 * If fixup_owner() returned an error, proprogate that. If it
3288 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3291 ret = (res < 0) ? res : 0;
3293 /* Unqueue and drop the lock. */
3297 if (ret == -EINTR) {
3299 * We've already been requeued, but cannot restart by calling
3300 * futex_lock_pi() directly. We could restart this syscall, but
3301 * it would detect that the user space "val" changed and return
3302 * -EWOULDBLOCK. Save the overhead of the restart and return
3303 * -EWOULDBLOCK directly.
3310 hrtimer_cancel(&to->timer);
3311 destroy_hrtimer_on_stack(&to->timer);
3317 * Support for robust futexes: the kernel cleans up held futexes at
3320 * Implementation: user-space maintains a per-thread list of locks it
3321 * is holding. Upon do_exit(), the kernel carefully walks this list,
3322 * and marks all locks that are owned by this thread with the
3323 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3324 * always manipulated with the lock held, so the list is private and
3325 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3326 * field, to allow the kernel to clean up if the thread dies after
3327 * acquiring the lock, but just before it could have added itself to
3328 * the list. There can only be one such pending lock.
3332 * sys_set_robust_list() - Set the robust-futex list head of a task
3333 * @head: pointer to the list-head
3334 * @len: length of the list-head, as userspace expects
3336 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3339 if (!futex_cmpxchg_enabled)
3342 * The kernel knows only one size for now:
3344 if (unlikely(len != sizeof(*head)))
3347 current->robust_list = head;
3353 * sys_get_robust_list() - Get the robust-futex list head of a task
3354 * @pid: pid of the process [zero for current task]
3355 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3356 * @len_ptr: pointer to a length field, the kernel fills in the header size
3358 SYSCALL_DEFINE3(get_robust_list, int, pid,
3359 struct robust_list_head __user * __user *, head_ptr,
3360 size_t __user *, len_ptr)
3362 struct robust_list_head __user *head;
3364 struct task_struct *p;
3366 if (!futex_cmpxchg_enabled)
3375 p = find_task_by_vpid(pid);
3381 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3384 head = p->robust_list;
3387 if (put_user(sizeof(*head), len_ptr))
3389 return put_user(head, head_ptr);
3397 /* Constants for the pending_op argument of handle_futex_death */
3398 #define HANDLE_DEATH_PENDING true
3399 #define HANDLE_DEATH_LIST false
3402 * Process a futex-list entry, check whether it's owned by the
3403 * dying task, and do notification if so:
3405 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3406 bool pi, bool pending_op)
3408 u32 uval, nval, mval;
3411 /* Futex address must be 32bit aligned */
3412 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3416 if (get_user(uval, uaddr))
3420 * Special case for regular (non PI) futexes. The unlock path in
3421 * user space has two race scenarios:
3423 * 1. The unlock path releases the user space futex value and
3424 * before it can execute the futex() syscall to wake up
3425 * waiters it is killed.
3427 * 2. A woken up waiter is killed before it can acquire the
3428 * futex in user space.
3430 * In both cases the TID validation below prevents a wakeup of
3431 * potential waiters which can cause these waiters to block
3434 * In both cases the following conditions are met:
3436 * 1) task->robust_list->list_op_pending != NULL
3437 * @pending_op == true
3438 * 2) User space futex value == 0
3439 * 3) Regular futex: @pi == false
3441 * If these conditions are met, it is safe to attempt waking up a
3442 * potential waiter without touching the user space futex value and
3443 * trying to set the OWNER_DIED bit. The user space futex value is
3444 * uncontended and the rest of the user space mutex state is
3445 * consistent, so a woken waiter will just take over the
3446 * uncontended futex. Setting the OWNER_DIED bit would create
3447 * inconsistent state and malfunction of the user space owner died
3450 if (pending_op && !pi && !uval) {
3451 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3455 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3459 * Ok, this dying thread is truly holding a futex
3460 * of interest. Set the OWNER_DIED bit atomically
3461 * via cmpxchg, and if the value had FUTEX_WAITERS
3462 * set, wake up a waiter (if any). (We have to do a
3463 * futex_wake() even if OWNER_DIED is already set -
3464 * to handle the rare but possible case of recursive
3465 * thread-death.) The rest of the cleanup is done in
3468 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3471 * We are not holding a lock here, but we want to have
3472 * the pagefault_disable/enable() protection because
3473 * we want to handle the fault gracefully. If the
3474 * access fails we try to fault in the futex with R/W
3475 * verification via get_user_pages. get_user() above
3476 * does not guarantee R/W access. If that fails we
3477 * give up and leave the futex locked.
3479 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3482 if (fault_in_user_writeable(uaddr))
3500 * Wake robust non-PI futexes here. The wakeup of
3501 * PI futexes happens in exit_pi_state():
3503 if (!pi && (uval & FUTEX_WAITERS))
3504 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3510 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3512 static inline int fetch_robust_entry(struct robust_list __user **entry,
3513 struct robust_list __user * __user *head,
3516 unsigned long uentry;
3518 if (get_user(uentry, (unsigned long __user *)head))
3521 *entry = (void __user *)(uentry & ~1UL);
3528 * Walk curr->robust_list (very carefully, it's a userspace list!)
3529 * and mark any locks found there dead, and notify any waiters.
3531 * We silently return on any sign of list-walking problem.
3533 static void exit_robust_list(struct task_struct *curr)
3535 struct robust_list_head __user *head = curr->robust_list;
3536 struct robust_list __user *entry, *next_entry, *pending;
3537 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3538 unsigned int next_pi;
3539 unsigned long futex_offset;
3542 if (!futex_cmpxchg_enabled)
3546 * Fetch the list head (which was registered earlier, via
3547 * sys_set_robust_list()):
3549 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3552 * Fetch the relative futex offset:
3554 if (get_user(futex_offset, &head->futex_offset))
3557 * Fetch any possibly pending lock-add first, and handle it
3560 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3563 next_entry = NULL; /* avoid warning with gcc */
3564 while (entry != &head->list) {
3566 * Fetch the next entry in the list before calling
3567 * handle_futex_death:
3569 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3571 * A pending lock might already be on the list, so
3572 * don't process it twice:
3574 if (entry != pending) {
3575 if (handle_futex_death((void __user *)entry + futex_offset,
3576 curr, pi, HANDLE_DEATH_LIST))
3584 * Avoid excessively long or circular lists:
3593 handle_futex_death((void __user *)pending + futex_offset,
3594 curr, pip, HANDLE_DEATH_PENDING);
3598 static void futex_cleanup(struct task_struct *tsk)
3600 if (unlikely(tsk->robust_list)) {
3601 exit_robust_list(tsk);
3602 tsk->robust_list = NULL;
3605 #ifdef CONFIG_COMPAT
3606 if (unlikely(tsk->compat_robust_list)) {
3607 compat_exit_robust_list(tsk);
3608 tsk->compat_robust_list = NULL;
3612 if (unlikely(!list_empty(&tsk->pi_state_list)))
3613 exit_pi_state_list(tsk);
3617 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3618 * @tsk: task to set the state on
3620 * Set the futex exit state of the task lockless. The futex waiter code
3621 * observes that state when a task is exiting and loops until the task has
3622 * actually finished the futex cleanup. The worst case for this is that the
3623 * waiter runs through the wait loop until the state becomes visible.
3625 * This is called from the recursive fault handling path in do_exit().
3627 * This is best effort. Either the futex exit code has run already or
3628 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3629 * take it over. If not, the problem is pushed back to user space. If the
3630 * futex exit code did not run yet, then an already queued waiter might
3631 * block forever, but there is nothing which can be done about that.
3633 void futex_exit_recursive(struct task_struct *tsk)
3635 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3636 if (tsk->futex_state == FUTEX_STATE_EXITING)
3637 mutex_unlock(&tsk->futex_exit_mutex);
3638 tsk->futex_state = FUTEX_STATE_DEAD;
3641 static void futex_cleanup_begin(struct task_struct *tsk)
3644 * Prevent various race issues against a concurrent incoming waiter
3645 * including live locks by forcing the waiter to block on
3646 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3647 * attach_to_pi_owner().
3649 mutex_lock(&tsk->futex_exit_mutex);
3652 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3654 * This ensures that all subsequent checks of tsk->futex_state in
3655 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3656 * tsk->pi_lock held.
3658 * It guarantees also that a pi_state which was queued right before
3659 * the state change under tsk->pi_lock by a concurrent waiter must
3660 * be observed in exit_pi_state_list().
3662 raw_spin_lock_irq(&tsk->pi_lock);
3663 tsk->futex_state = FUTEX_STATE_EXITING;
3664 raw_spin_unlock_irq(&tsk->pi_lock);
3667 static void futex_cleanup_end(struct task_struct *tsk, int state)
3670 * Lockless store. The only side effect is that an observer might
3671 * take another loop until it becomes visible.
3673 tsk->futex_state = state;
3675 * Drop the exit protection. This unblocks waiters which observed
3676 * FUTEX_STATE_EXITING to reevaluate the state.
3678 mutex_unlock(&tsk->futex_exit_mutex);
3681 void futex_exec_release(struct task_struct *tsk)
3684 * The state handling is done for consistency, but in the case of
3685 * exec() there is no way to prevent futher damage as the PID stays
3686 * the same. But for the unlikely and arguably buggy case that a
3687 * futex is held on exec(), this provides at least as much state
3688 * consistency protection which is possible.
3690 futex_cleanup_begin(tsk);
3693 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3694 * exec a new binary.
3696 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3699 void futex_exit_release(struct task_struct *tsk)
3701 futex_cleanup_begin(tsk);
3703 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3706 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3707 u32 __user *uaddr2, u32 val2, u32 val3)
3709 int cmd = op & FUTEX_CMD_MASK;
3710 unsigned int flags = 0;
3712 if (!(op & FUTEX_PRIVATE_FLAG))
3713 flags |= FLAGS_SHARED;
3715 if (op & FUTEX_CLOCK_REALTIME) {
3716 flags |= FLAGS_CLOCKRT;
3717 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3718 cmd != FUTEX_WAIT_REQUEUE_PI)
3724 case FUTEX_UNLOCK_PI:
3725 case FUTEX_TRYLOCK_PI:
3726 case FUTEX_WAIT_REQUEUE_PI:
3727 case FUTEX_CMP_REQUEUE_PI:
3728 if (!futex_cmpxchg_enabled)
3734 val3 = FUTEX_BITSET_MATCH_ANY;
3736 case FUTEX_WAIT_BITSET:
3737 return futex_wait(uaddr, flags, val, timeout, val3);
3739 val3 = FUTEX_BITSET_MATCH_ANY;
3741 case FUTEX_WAKE_BITSET:
3742 return futex_wake(uaddr, flags, val, val3);
3744 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3745 case FUTEX_CMP_REQUEUE:
3746 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3748 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3750 return futex_lock_pi(uaddr, flags, timeout, 0);
3751 case FUTEX_UNLOCK_PI:
3752 return futex_unlock_pi(uaddr, flags);
3753 case FUTEX_TRYLOCK_PI:
3754 return futex_lock_pi(uaddr, flags, NULL, 1);
3755 case FUTEX_WAIT_REQUEUE_PI:
3756 val3 = FUTEX_BITSET_MATCH_ANY;
3757 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3759 case FUTEX_CMP_REQUEUE_PI:
3760 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3766 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3767 const struct __kernel_timespec __user *, utime,
3768 u32 __user *, uaddr2, u32, val3)
3770 struct timespec64 ts;
3771 ktime_t t, *tp = NULL;
3773 int cmd = op & FUTEX_CMD_MASK;
3775 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3776 cmd == FUTEX_WAIT_BITSET ||
3777 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3778 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3780 if (get_timespec64(&ts, utime))
3782 if (!timespec64_valid(&ts))
3785 t = timespec64_to_ktime(ts);
3786 if (cmd == FUTEX_WAIT)
3787 t = ktime_add_safe(ktime_get(), t);
3788 else if (!(op & FUTEX_CLOCK_REALTIME))
3789 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3793 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3794 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3796 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3797 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3798 val2 = (u32) (unsigned long) utime;
3800 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3803 #ifdef CONFIG_COMPAT
3805 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3808 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3809 compat_uptr_t __user *head, unsigned int *pi)
3811 if (get_user(*uentry, head))
3814 *entry = compat_ptr((*uentry) & ~1);
3815 *pi = (unsigned int)(*uentry) & 1;
3820 static void __user *futex_uaddr(struct robust_list __user *entry,
3821 compat_long_t futex_offset)
3823 compat_uptr_t base = ptr_to_compat(entry);
3824 void __user *uaddr = compat_ptr(base + futex_offset);
3830 * Walk curr->robust_list (very carefully, it's a userspace list!)
3831 * and mark any locks found there dead, and notify any waiters.
3833 * We silently return on any sign of list-walking problem.
3835 static void compat_exit_robust_list(struct task_struct *curr)
3837 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3838 struct robust_list __user *entry, *next_entry, *pending;
3839 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3840 unsigned int next_pi;
3841 compat_uptr_t uentry, next_uentry, upending;
3842 compat_long_t futex_offset;
3845 if (!futex_cmpxchg_enabled)
3849 * Fetch the list head (which was registered earlier, via
3850 * sys_set_robust_list()):
3852 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3855 * Fetch the relative futex offset:
3857 if (get_user(futex_offset, &head->futex_offset))
3860 * Fetch any possibly pending lock-add first, and handle it
3863 if (compat_fetch_robust_entry(&upending, &pending,
3864 &head->list_op_pending, &pip))
3867 next_entry = NULL; /* avoid warning with gcc */
3868 while (entry != (struct robust_list __user *) &head->list) {
3870 * Fetch the next entry in the list before calling
3871 * handle_futex_death:
3873 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3874 (compat_uptr_t __user *)&entry->next, &next_pi);
3876 * A pending lock might already be on the list, so
3877 * dont process it twice:
3879 if (entry != pending) {
3880 void __user *uaddr = futex_uaddr(entry, futex_offset);
3882 if (handle_futex_death(uaddr, curr, pi,
3888 uentry = next_uentry;
3892 * Avoid excessively long or circular lists:
3900 void __user *uaddr = futex_uaddr(pending, futex_offset);
3902 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3906 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3907 struct compat_robust_list_head __user *, head,
3910 if (!futex_cmpxchg_enabled)
3913 if (unlikely(len != sizeof(*head)))
3916 current->compat_robust_list = head;
3921 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3922 compat_uptr_t __user *, head_ptr,
3923 compat_size_t __user *, len_ptr)
3925 struct compat_robust_list_head __user *head;
3927 struct task_struct *p;
3929 if (!futex_cmpxchg_enabled)
3938 p = find_task_by_vpid(pid);
3944 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3947 head = p->compat_robust_list;
3950 if (put_user(sizeof(*head), len_ptr))
3952 return put_user(ptr_to_compat(head), head_ptr);
3959 #endif /* CONFIG_COMPAT */
3961 #ifdef CONFIG_COMPAT_32BIT_TIME
3962 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3963 const struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3966 struct timespec64 ts;
3967 ktime_t t, *tp = NULL;
3969 int cmd = op & FUTEX_CMD_MASK;
3971 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3972 cmd == FUTEX_WAIT_BITSET ||
3973 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3974 if (get_old_timespec32(&ts, utime))
3976 if (!timespec64_valid(&ts))
3979 t = timespec64_to_ktime(ts);
3980 if (cmd == FUTEX_WAIT)
3981 t = ktime_add_safe(ktime_get(), t);
3982 else if (!(op & FUTEX_CLOCK_REALTIME))
3983 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3986 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3987 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3988 val2 = (int) (unsigned long) utime;
3990 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3992 #endif /* CONFIG_COMPAT_32BIT_TIME */
3994 static void __init futex_detect_cmpxchg(void)
3996 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4000 * This will fail and we want it. Some arch implementations do
4001 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4002 * functionality. We want to know that before we call in any
4003 * of the complex code paths. Also we want to prevent
4004 * registration of robust lists in that case. NULL is
4005 * guaranteed to fault and we get -EFAULT on functional
4006 * implementation, the non-functional ones will return
4009 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4010 futex_cmpxchg_enabled = 1;
4014 static int __init futex_init(void)
4016 unsigned int futex_shift;
4019 #if CONFIG_BASE_SMALL
4020 futex_hashsize = 16;
4022 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4025 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4027 futex_hashsize < 256 ? HASH_SMALL : 0,
4029 futex_hashsize, futex_hashsize);
4030 futex_hashsize = 1UL << futex_shift;
4032 futex_detect_cmpxchg();
4034 for (i = 0; i < futex_hashsize; i++) {
4035 atomic_set(&futex_queues[i].waiters, 0);
4036 plist_head_init(&futex_queues[i].chain);
4037 spin_lock_init(&futex_queues[i].lock);
4042 core_initcall(futex_init);