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 get_pi_state(struct futex_pi_state *pi_state)
768 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
772 * Drops a reference to the pi_state object and frees or caches it
773 * when the last reference is gone.
775 static void put_pi_state(struct futex_pi_state *pi_state)
780 if (!refcount_dec_and_test(&pi_state->refcount))
784 * If pi_state->owner is NULL, the owner is most probably dying
785 * and has cleaned up the pi_state already
787 if (pi_state->owner) {
788 struct task_struct *owner;
791 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
792 owner = pi_state->owner;
794 raw_spin_lock(&owner->pi_lock);
795 list_del_init(&pi_state->list);
796 raw_spin_unlock(&owner->pi_lock);
798 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
799 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
802 if (current->pi_state_cache) {
806 * pi_state->list is already empty.
807 * clear pi_state->owner.
808 * refcount is at 0 - put it back to 1.
810 pi_state->owner = NULL;
811 refcount_set(&pi_state->refcount, 1);
812 current->pi_state_cache = pi_state;
816 #ifdef CONFIG_FUTEX_PI
819 * This task is holding PI mutexes at exit time => bad.
820 * Kernel cleans up PI-state, but userspace is likely hosed.
821 * (Robust-futex cleanup is separate and might save the day for userspace.)
823 static void exit_pi_state_list(struct task_struct *curr)
825 struct list_head *next, *head = &curr->pi_state_list;
826 struct futex_pi_state *pi_state;
827 struct futex_hash_bucket *hb;
828 union futex_key key = FUTEX_KEY_INIT;
830 if (!futex_cmpxchg_enabled)
833 * We are a ZOMBIE and nobody can enqueue itself on
834 * pi_state_list anymore, but we have to be careful
835 * versus waiters unqueueing themselves:
837 raw_spin_lock_irq(&curr->pi_lock);
838 while (!list_empty(head)) {
840 pi_state = list_entry(next, struct futex_pi_state, list);
842 hb = hash_futex(&key);
845 * We can race against put_pi_state() removing itself from the
846 * list (a waiter going away). put_pi_state() will first
847 * decrement the reference count and then modify the list, so
848 * its possible to see the list entry but fail this reference
851 * In that case; drop the locks to let put_pi_state() make
852 * progress and retry the loop.
854 if (!refcount_inc_not_zero(&pi_state->refcount)) {
855 raw_spin_unlock_irq(&curr->pi_lock);
857 raw_spin_lock_irq(&curr->pi_lock);
860 raw_spin_unlock_irq(&curr->pi_lock);
862 spin_lock(&hb->lock);
863 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
864 raw_spin_lock(&curr->pi_lock);
866 * We dropped the pi-lock, so re-check whether this
867 * task still owns the PI-state:
869 if (head->next != next) {
870 /* retain curr->pi_lock for the loop invariant */
871 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
872 spin_unlock(&hb->lock);
873 put_pi_state(pi_state);
877 WARN_ON(pi_state->owner != curr);
878 WARN_ON(list_empty(&pi_state->list));
879 list_del_init(&pi_state->list);
880 pi_state->owner = NULL;
882 raw_spin_unlock(&curr->pi_lock);
883 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
884 spin_unlock(&hb->lock);
886 rt_mutex_futex_unlock(&pi_state->pi_mutex);
887 put_pi_state(pi_state);
889 raw_spin_lock_irq(&curr->pi_lock);
891 raw_spin_unlock_irq(&curr->pi_lock);
894 static inline void exit_pi_state_list(struct task_struct *curr) { }
898 * We need to check the following states:
900 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
902 * [1] NULL | --- | --- | 0 | 0/1 | Valid
903 * [2] NULL | --- | --- | >0 | 0/1 | Valid
905 * [3] Found | NULL | -- | Any | 0/1 | Invalid
907 * [4] Found | Found | NULL | 0 | 1 | Valid
908 * [5] Found | Found | NULL | >0 | 1 | Invalid
910 * [6] Found | Found | task | 0 | 1 | Valid
912 * [7] Found | Found | NULL | Any | 0 | Invalid
914 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
915 * [9] Found | Found | task | 0 | 0 | Invalid
916 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
918 * [1] Indicates that the kernel can acquire the futex atomically. We
919 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
921 * [2] Valid, if TID does not belong to a kernel thread. If no matching
922 * thread is found then it indicates that the owner TID has died.
924 * [3] Invalid. The waiter is queued on a non PI futex
926 * [4] Valid state after exit_robust_list(), which sets the user space
927 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
929 * [5] The user space value got manipulated between exit_robust_list()
930 * and exit_pi_state_list()
932 * [6] Valid state after exit_pi_state_list() which sets the new owner in
933 * the pi_state but cannot access the user space value.
935 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
937 * [8] Owner and user space value match
939 * [9] There is no transient state which sets the user space TID to 0
940 * except exit_robust_list(), but this is indicated by the
941 * FUTEX_OWNER_DIED bit. See [4]
943 * [10] There is no transient state which leaves owner and user space
947 * Serialization and lifetime rules:
951 * hb -> futex_q, relation
952 * futex_q -> pi_state, relation
954 * (cannot be raw because hb can contain arbitrary amount
957 * pi_mutex->wait_lock:
961 * (and pi_mutex 'obviously')
965 * p->pi_state_list -> pi_state->list, relation
967 * pi_state->refcount:
975 * pi_mutex->wait_lock
981 * Validate that the existing waiter has a pi_state and sanity check
982 * the pi_state against the user space value. If correct, attach to
985 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
986 struct futex_pi_state *pi_state,
987 struct futex_pi_state **ps)
989 pid_t pid = uval & FUTEX_TID_MASK;
994 * Userspace might have messed up non-PI and PI futexes [3]
996 if (unlikely(!pi_state))
1000 * We get here with hb->lock held, and having found a
1001 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1002 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1003 * which in turn means that futex_lock_pi() still has a reference on
1006 * The waiter holding a reference on @pi_state also protects against
1007 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1008 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1009 * free pi_state before we can take a reference ourselves.
1011 WARN_ON(!refcount_read(&pi_state->refcount));
1014 * Now that we have a pi_state, we can acquire wait_lock
1015 * and do the state validation.
1017 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1020 * Since {uval, pi_state} is serialized by wait_lock, and our current
1021 * uval was read without holding it, it can have changed. Verify it
1022 * still is what we expect it to be, otherwise retry the entire
1025 if (get_futex_value_locked(&uval2, uaddr))
1032 * Handle the owner died case:
1034 if (uval & FUTEX_OWNER_DIED) {
1036 * exit_pi_state_list sets owner to NULL and wakes the
1037 * topmost waiter. The task which acquires the
1038 * pi_state->rt_mutex will fixup owner.
1040 if (!pi_state->owner) {
1042 * No pi state owner, but the user space TID
1043 * is not 0. Inconsistent state. [5]
1048 * Take a ref on the state and return success. [4]
1054 * If TID is 0, then either the dying owner has not
1055 * yet executed exit_pi_state_list() or some waiter
1056 * acquired the rtmutex in the pi state, but did not
1057 * yet fixup the TID in user space.
1059 * Take a ref on the state and return success. [6]
1065 * If the owner died bit is not set, then the pi_state
1066 * must have an owner. [7]
1068 if (!pi_state->owner)
1073 * Bail out if user space manipulated the futex value. If pi
1074 * state exists then the owner TID must be the same as the
1075 * user space TID. [9/10]
1077 if (pid != task_pid_vnr(pi_state->owner))
1081 get_pi_state(pi_state);
1082 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1099 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1104 * wait_for_owner_exiting - Block until the owner has exited
1105 * @ret: owner's current futex lock status
1106 * @exiting: Pointer to the exiting task
1108 * Caller must hold a refcount on @exiting.
1110 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1112 if (ret != -EBUSY) {
1113 WARN_ON_ONCE(exiting);
1117 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1120 mutex_lock(&exiting->futex_exit_mutex);
1122 * No point in doing state checking here. If the waiter got here
1123 * while the task was in exec()->exec_futex_release() then it can
1124 * have any FUTEX_STATE_* value when the waiter has acquired the
1125 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1126 * already. Highly unlikely and not a problem. Just one more round
1127 * through the futex maze.
1129 mutex_unlock(&exiting->futex_exit_mutex);
1131 put_task_struct(exiting);
1134 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1135 struct task_struct *tsk)
1140 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1141 * caller that the alleged owner is busy.
1143 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1147 * Reread the user space value to handle the following situation:
1151 * sys_exit() sys_futex()
1152 * do_exit() futex_lock_pi()
1153 * futex_lock_pi_atomic()
1154 * exit_signals(tsk) No waiters:
1155 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1156 * mm_release(tsk) Set waiter bit
1157 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1158 * Set owner died attach_to_pi_owner() {
1159 * *uaddr = 0xC0000000; tsk = get_task(PID);
1160 * } if (!tsk->flags & PF_EXITING) {
1162 * tsk->futex_state = } else {
1163 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1166 * return -ESRCH; <--- FAIL
1169 * Returning ESRCH unconditionally is wrong here because the
1170 * user space value has been changed by the exiting task.
1172 * The same logic applies to the case where the exiting task is
1175 if (get_futex_value_locked(&uval2, uaddr))
1178 /* If the user space value has changed, try again. */
1183 * The exiting task did not have a robust list, the robust list was
1184 * corrupted or the user space value in *uaddr is simply bogus.
1185 * Give up and tell user space.
1191 * Lookup the task for the TID provided from user space and attach to
1192 * it after doing proper sanity checks.
1194 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1195 struct futex_pi_state **ps,
1196 struct task_struct **exiting)
1198 pid_t pid = uval & FUTEX_TID_MASK;
1199 struct futex_pi_state *pi_state;
1200 struct task_struct *p;
1203 * We are the first waiter - try to look up the real owner and attach
1204 * the new pi_state to it, but bail out when TID = 0 [1]
1206 * The !pid check is paranoid. None of the call sites should end up
1207 * with pid == 0, but better safe than sorry. Let the caller retry
1211 p = find_get_task_by_vpid(pid);
1213 return handle_exit_race(uaddr, uval, NULL);
1215 if (unlikely(p->flags & PF_KTHREAD)) {
1221 * We need to look at the task state to figure out, whether the
1222 * task is exiting. To protect against the change of the task state
1223 * in futex_exit_release(), we do this protected by p->pi_lock:
1225 raw_spin_lock_irq(&p->pi_lock);
1226 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1228 * The task is on the way out. When the futex state is
1229 * FUTEX_STATE_DEAD, we know that the task has finished
1232 int ret = handle_exit_race(uaddr, uval, p);
1234 raw_spin_unlock_irq(&p->pi_lock);
1236 * If the owner task is between FUTEX_STATE_EXITING and
1237 * FUTEX_STATE_DEAD then store the task pointer and keep
1238 * the reference on the task struct. The calling code will
1239 * drop all locks, wait for the task to reach
1240 * FUTEX_STATE_DEAD and then drop the refcount. This is
1241 * required to prevent a live lock when the current task
1242 * preempted the exiting task between the two states.
1252 * No existing pi state. First waiter. [2]
1254 * This creates pi_state, we have hb->lock held, this means nothing can
1255 * observe this state, wait_lock is irrelevant.
1257 pi_state = alloc_pi_state();
1260 * Initialize the pi_mutex in locked state and make @p
1263 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1265 /* Store the key for possible exit cleanups: */
1266 pi_state->key = *key;
1268 WARN_ON(!list_empty(&pi_state->list));
1269 list_add(&pi_state->list, &p->pi_state_list);
1271 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1272 * because there is no concurrency as the object is not published yet.
1274 pi_state->owner = p;
1275 raw_spin_unlock_irq(&p->pi_lock);
1284 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1285 struct futex_hash_bucket *hb,
1286 union futex_key *key, struct futex_pi_state **ps,
1287 struct task_struct **exiting)
1289 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1292 * If there is a waiter on that futex, validate it and
1293 * attach to the pi_state when the validation succeeds.
1296 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1299 * We are the first waiter - try to look up the owner based on
1300 * @uval and attach to it.
1302 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1305 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1310 if (unlikely(should_fail_futex(true)))
1313 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1317 /* If user space value changed, let the caller retry */
1318 return curval != uval ? -EAGAIN : 0;
1322 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1323 * @uaddr: the pi futex user address
1324 * @hb: the pi futex hash bucket
1325 * @key: the futex key associated with uaddr and hb
1326 * @ps: the pi_state pointer where we store the result of the
1328 * @task: the task to perform the atomic lock work for. This will
1329 * be "current" except in the case of requeue pi.
1330 * @exiting: Pointer to store the task pointer of the owner task
1331 * which is in the middle of exiting
1332 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1335 * - 0 - ready to wait;
1336 * - 1 - acquired the lock;
1339 * The hb->lock and futex_key refs shall be held by the caller.
1341 * @exiting is only set when the return value is -EBUSY. If so, this holds
1342 * a refcount on the exiting task on return and the caller needs to drop it
1343 * after waiting for the exit to complete.
1345 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1346 union futex_key *key,
1347 struct futex_pi_state **ps,
1348 struct task_struct *task,
1349 struct task_struct **exiting,
1352 u32 uval, newval, vpid = task_pid_vnr(task);
1353 struct futex_q *top_waiter;
1357 * Read the user space value first so we can validate a few
1358 * things before proceeding further.
1360 if (get_futex_value_locked(&uval, uaddr))
1363 if (unlikely(should_fail_futex(true)))
1369 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1372 if ((unlikely(should_fail_futex(true))))
1376 * Lookup existing state first. If it exists, try to attach to
1379 top_waiter = futex_top_waiter(hb, key);
1381 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1384 * No waiter and user TID is 0. We are here because the
1385 * waiters or the owner died bit is set or called from
1386 * requeue_cmp_pi or for whatever reason something took the
1389 if (!(uval & FUTEX_TID_MASK)) {
1391 * We take over the futex. No other waiters and the user space
1392 * TID is 0. We preserve the owner died bit.
1394 newval = uval & FUTEX_OWNER_DIED;
1397 /* The futex requeue_pi code can enforce the waiters bit */
1399 newval |= FUTEX_WAITERS;
1401 ret = lock_pi_update_atomic(uaddr, uval, newval);
1402 /* If the take over worked, return 1 */
1403 return ret < 0 ? ret : 1;
1407 * First waiter. Set the waiters bit before attaching ourself to
1408 * the owner. If owner tries to unlock, it will be forced into
1409 * the kernel and blocked on hb->lock.
1411 newval = uval | FUTEX_WAITERS;
1412 ret = lock_pi_update_atomic(uaddr, uval, newval);
1416 * If the update of the user space value succeeded, we try to
1417 * attach to the owner. If that fails, no harm done, we only
1418 * set the FUTEX_WAITERS bit in the user space variable.
1420 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1424 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1425 * @q: The futex_q to unqueue
1427 * The q->lock_ptr must not be NULL and must be held by the caller.
1429 static void __unqueue_futex(struct futex_q *q)
1431 struct futex_hash_bucket *hb;
1433 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1435 lockdep_assert_held(q->lock_ptr);
1437 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1438 plist_del(&q->list, &hb->chain);
1443 * The hash bucket lock must be held when this is called.
1444 * Afterwards, the futex_q must not be accessed. Callers
1445 * must ensure to later call wake_up_q() for the actual
1448 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1450 struct task_struct *p = q->task;
1452 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1458 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1459 * is written, without taking any locks. This is possible in the event
1460 * of a spurious wakeup, for example. A memory barrier is required here
1461 * to prevent the following store to lock_ptr from getting ahead of the
1462 * plist_del in __unqueue_futex().
1464 smp_store_release(&q->lock_ptr, NULL);
1467 * Queue the task for later wakeup for after we've released
1470 wake_q_add_safe(wake_q, p);
1474 * Caller must hold a reference on @pi_state.
1476 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1479 struct task_struct *new_owner;
1480 bool postunlock = false;
1481 DEFINE_WAKE_Q(wake_q);
1484 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1485 if (WARN_ON_ONCE(!new_owner)) {
1487 * As per the comment in futex_unlock_pi() this should not happen.
1489 * When this happens, give up our locks and try again, giving
1490 * the futex_lock_pi() instance time to complete, either by
1491 * waiting on the rtmutex or removing itself from the futex
1499 * We pass it to the next owner. The WAITERS bit is always kept
1500 * enabled while there is PI state around. We cleanup the owner
1501 * died bit, because we are the owner.
1503 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1505 if (unlikely(should_fail_futex(true))) {
1510 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1511 if (!ret && (curval != uval)) {
1513 * If a unconditional UNLOCK_PI operation (user space did not
1514 * try the TID->0 transition) raced with a waiter setting the
1515 * FUTEX_WAITERS flag between get_user() and locking the hash
1516 * bucket lock, retry the operation.
1518 if ((FUTEX_TID_MASK & curval) == uval)
1528 * This is a point of no return; once we modify the uval there is no
1529 * going back and subsequent operations must not fail.
1532 raw_spin_lock(&pi_state->owner->pi_lock);
1533 WARN_ON(list_empty(&pi_state->list));
1534 list_del_init(&pi_state->list);
1535 raw_spin_unlock(&pi_state->owner->pi_lock);
1537 raw_spin_lock(&new_owner->pi_lock);
1538 WARN_ON(!list_empty(&pi_state->list));
1539 list_add(&pi_state->list, &new_owner->pi_state_list);
1540 pi_state->owner = new_owner;
1541 raw_spin_unlock(&new_owner->pi_lock);
1543 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1546 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1549 rt_mutex_postunlock(&wake_q);
1555 * Express the locking dependencies for lockdep:
1558 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1561 spin_lock(&hb1->lock);
1563 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1564 } else { /* hb1 > hb2 */
1565 spin_lock(&hb2->lock);
1566 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1571 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1573 spin_unlock(&hb1->lock);
1575 spin_unlock(&hb2->lock);
1579 * Wake up waiters matching bitset queued on this futex (uaddr).
1582 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1584 struct futex_hash_bucket *hb;
1585 struct futex_q *this, *next;
1586 union futex_key key = FUTEX_KEY_INIT;
1588 DEFINE_WAKE_Q(wake_q);
1593 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1594 if (unlikely(ret != 0))
1597 hb = hash_futex(&key);
1599 /* Make sure we really have tasks to wakeup */
1600 if (!hb_waiters_pending(hb))
1603 spin_lock(&hb->lock);
1605 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1606 if (match_futex (&this->key, &key)) {
1607 if (this->pi_state || this->rt_waiter) {
1612 /* Check if one of the bits is set in both bitsets */
1613 if (!(this->bitset & bitset))
1616 mark_wake_futex(&wake_q, this);
1617 if (++ret >= nr_wake)
1622 spin_unlock(&hb->lock);
1627 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1629 unsigned int op = (encoded_op & 0x70000000) >> 28;
1630 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1631 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1632 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1635 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1636 if (oparg < 0 || oparg > 31) {
1637 char comm[sizeof(current->comm)];
1639 * kill this print and return -EINVAL when userspace
1642 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1643 get_task_comm(comm, current), oparg);
1649 pagefault_disable();
1650 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1656 case FUTEX_OP_CMP_EQ:
1657 return oldval == cmparg;
1658 case FUTEX_OP_CMP_NE:
1659 return oldval != cmparg;
1660 case FUTEX_OP_CMP_LT:
1661 return oldval < cmparg;
1662 case FUTEX_OP_CMP_GE:
1663 return oldval >= cmparg;
1664 case FUTEX_OP_CMP_LE:
1665 return oldval <= cmparg;
1666 case FUTEX_OP_CMP_GT:
1667 return oldval > cmparg;
1674 * Wake up all waiters hashed on the physical page that is mapped
1675 * to this virtual address:
1678 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1679 int nr_wake, int nr_wake2, int op)
1681 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1682 struct futex_hash_bucket *hb1, *hb2;
1683 struct futex_q *this, *next;
1685 DEFINE_WAKE_Q(wake_q);
1688 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1689 if (unlikely(ret != 0))
1691 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1692 if (unlikely(ret != 0))
1695 hb1 = hash_futex(&key1);
1696 hb2 = hash_futex(&key2);
1699 double_lock_hb(hb1, hb2);
1700 op_ret = futex_atomic_op_inuser(op, uaddr2);
1701 if (unlikely(op_ret < 0)) {
1702 double_unlock_hb(hb1, hb2);
1704 if (!IS_ENABLED(CONFIG_MMU) ||
1705 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1707 * we don't get EFAULT from MMU faults if we don't have
1708 * an MMU, but we might get them from range checking
1714 if (op_ret == -EFAULT) {
1715 ret = fault_in_user_writeable(uaddr2);
1720 if (!(flags & FLAGS_SHARED)) {
1729 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1730 if (match_futex (&this->key, &key1)) {
1731 if (this->pi_state || this->rt_waiter) {
1735 mark_wake_futex(&wake_q, this);
1736 if (++ret >= nr_wake)
1743 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1744 if (match_futex (&this->key, &key2)) {
1745 if (this->pi_state || this->rt_waiter) {
1749 mark_wake_futex(&wake_q, this);
1750 if (++op_ret >= nr_wake2)
1758 double_unlock_hb(hb1, hb2);
1764 * requeue_futex() - Requeue a futex_q from one hb to another
1765 * @q: the futex_q to requeue
1766 * @hb1: the source hash_bucket
1767 * @hb2: the target hash_bucket
1768 * @key2: the new key for the requeued futex_q
1771 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1772 struct futex_hash_bucket *hb2, union futex_key *key2)
1776 * If key1 and key2 hash to the same bucket, no need to
1779 if (likely(&hb1->chain != &hb2->chain)) {
1780 plist_del(&q->list, &hb1->chain);
1781 hb_waiters_dec(hb1);
1782 hb_waiters_inc(hb2);
1783 plist_add(&q->list, &hb2->chain);
1784 q->lock_ptr = &hb2->lock;
1790 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1792 * @key: the key of the requeue target futex
1793 * @hb: the hash_bucket of the requeue target futex
1795 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1796 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1797 * to the requeue target futex so the waiter can detect the wakeup on the right
1798 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1799 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1800 * to protect access to the pi_state to fixup the owner later. Must be called
1801 * with both q->lock_ptr and hb->lock held.
1804 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1805 struct futex_hash_bucket *hb)
1811 WARN_ON(!q->rt_waiter);
1812 q->rt_waiter = NULL;
1814 q->lock_ptr = &hb->lock;
1816 wake_up_state(q->task, TASK_NORMAL);
1820 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1821 * @pifutex: the user address of the to futex
1822 * @hb1: the from futex hash bucket, must be locked by the caller
1823 * @hb2: the to futex hash bucket, must be locked by the caller
1824 * @key1: the from futex key
1825 * @key2: the to futex key
1826 * @ps: address to store the pi_state pointer
1827 * @exiting: Pointer to store the task pointer of the owner task
1828 * which is in the middle of exiting
1829 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1831 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1832 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1833 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1834 * hb1 and hb2 must be held by the caller.
1836 * @exiting is only set when the return value is -EBUSY. If so, this holds
1837 * a refcount on the exiting task on return and the caller needs to drop it
1838 * after waiting for the exit to complete.
1841 * - 0 - failed to acquire the lock atomically;
1842 * - >0 - acquired the lock, return value is vpid of the top_waiter
1846 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1847 struct futex_hash_bucket *hb2, union futex_key *key1,
1848 union futex_key *key2, struct futex_pi_state **ps,
1849 struct task_struct **exiting, int set_waiters)
1851 struct futex_q *top_waiter = NULL;
1855 if (get_futex_value_locked(&curval, pifutex))
1858 if (unlikely(should_fail_futex(true)))
1862 * Find the top_waiter and determine if there are additional waiters.
1863 * If the caller intends to requeue more than 1 waiter to pifutex,
1864 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1865 * as we have means to handle the possible fault. If not, don't set
1866 * the bit unecessarily as it will force the subsequent unlock to enter
1869 top_waiter = futex_top_waiter(hb1, key1);
1871 /* There are no waiters, nothing for us to do. */
1875 /* Ensure we requeue to the expected futex. */
1876 if (!match_futex(top_waiter->requeue_pi_key, key2))
1880 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1881 * the contended case or if set_waiters is 1. The pi_state is returned
1882 * in ps in contended cases.
1884 vpid = task_pid_vnr(top_waiter->task);
1885 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1886 exiting, set_waiters);
1888 requeue_pi_wake_futex(top_waiter, key2, hb2);
1895 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1896 * @uaddr1: source futex user address
1897 * @flags: futex flags (FLAGS_SHARED, etc.)
1898 * @uaddr2: target futex user address
1899 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1900 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1901 * @cmpval: @uaddr1 expected value (or %NULL)
1902 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1903 * pi futex (pi to pi requeue is not supported)
1905 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1906 * uaddr2 atomically on behalf of the top waiter.
1909 * - >=0 - on success, the number of tasks requeued or woken;
1912 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1913 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1914 u32 *cmpval, int requeue_pi)
1916 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1917 int task_count = 0, ret;
1918 struct futex_pi_state *pi_state = NULL;
1919 struct futex_hash_bucket *hb1, *hb2;
1920 struct futex_q *this, *next;
1921 DEFINE_WAKE_Q(wake_q);
1923 if (nr_wake < 0 || nr_requeue < 0)
1927 * When PI not supported: return -ENOSYS if requeue_pi is true,
1928 * consequently the compiler knows requeue_pi is always false past
1929 * this point which will optimize away all the conditional code
1932 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1937 * Requeue PI only works on two distinct uaddrs. This
1938 * check is only valid for private futexes. See below.
1940 if (uaddr1 == uaddr2)
1944 * requeue_pi requires a pi_state, try to allocate it now
1945 * without any locks in case it fails.
1947 if (refill_pi_state_cache())
1950 * requeue_pi must wake as many tasks as it can, up to nr_wake
1951 * + nr_requeue, since it acquires the rt_mutex prior to
1952 * returning to userspace, so as to not leave the rt_mutex with
1953 * waiters and no owner. However, second and third wake-ups
1954 * cannot be predicted as they involve race conditions with the
1955 * first wake and a fault while looking up the pi_state. Both
1956 * pthread_cond_signal() and pthread_cond_broadcast() should
1964 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1965 if (unlikely(ret != 0))
1967 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1968 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1969 if (unlikely(ret != 0))
1973 * The check above which compares uaddrs is not sufficient for
1974 * shared futexes. We need to compare the keys:
1976 if (requeue_pi && match_futex(&key1, &key2))
1979 hb1 = hash_futex(&key1);
1980 hb2 = hash_futex(&key2);
1983 hb_waiters_inc(hb2);
1984 double_lock_hb(hb1, hb2);
1986 if (likely(cmpval != NULL)) {
1989 ret = get_futex_value_locked(&curval, uaddr1);
1991 if (unlikely(ret)) {
1992 double_unlock_hb(hb1, hb2);
1993 hb_waiters_dec(hb2);
1995 ret = get_user(curval, uaddr1);
1999 if (!(flags & FLAGS_SHARED))
2004 if (curval != *cmpval) {
2010 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2011 struct task_struct *exiting = NULL;
2014 * Attempt to acquire uaddr2 and wake the top waiter. If we
2015 * intend to requeue waiters, force setting the FUTEX_WAITERS
2016 * bit. We force this here where we are able to easily handle
2017 * faults rather in the requeue loop below.
2019 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2021 &exiting, nr_requeue);
2024 * At this point the top_waiter has either taken uaddr2 or is
2025 * waiting on it. If the former, then the pi_state will not
2026 * exist yet, look it up one more time to ensure we have a
2027 * reference to it. If the lock was taken, ret contains the
2028 * vpid of the top waiter task.
2029 * If the lock was not taken, we have pi_state and an initial
2030 * refcount on it. In case of an error we have nothing.
2036 * If we acquired the lock, then the user space value
2037 * of uaddr2 should be vpid. It cannot be changed by
2038 * the top waiter as it is blocked on hb2 lock if it
2039 * tries to do so. If something fiddled with it behind
2040 * our back the pi state lookup might unearth it. So
2041 * we rather use the known value than rereading and
2042 * handing potential crap to lookup_pi_state.
2044 * If that call succeeds then we have pi_state and an
2045 * initial refcount on it.
2047 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2048 &pi_state, &exiting);
2053 /* We hold a reference on the pi state. */
2056 /* If the above failed, then pi_state is NULL */
2058 double_unlock_hb(hb1, hb2);
2059 hb_waiters_dec(hb2);
2060 ret = fault_in_user_writeable(uaddr2);
2067 * Two reasons for this:
2068 * - EBUSY: Owner is exiting and we just wait for the
2070 * - EAGAIN: The user space value changed.
2072 double_unlock_hb(hb1, hb2);
2073 hb_waiters_dec(hb2);
2075 * Handle the case where the owner is in the middle of
2076 * exiting. Wait for the exit to complete otherwise
2077 * this task might loop forever, aka. live lock.
2079 wait_for_owner_exiting(ret, exiting);
2087 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2088 if (task_count - nr_wake >= nr_requeue)
2091 if (!match_futex(&this->key, &key1))
2095 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2096 * be paired with each other and no other futex ops.
2098 * We should never be requeueing a futex_q with a pi_state,
2099 * which is awaiting a futex_unlock_pi().
2101 if ((requeue_pi && !this->rt_waiter) ||
2102 (!requeue_pi && this->rt_waiter) ||
2109 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2110 * lock, we already woke the top_waiter. If not, it will be
2111 * woken by futex_unlock_pi().
2113 if (++task_count <= nr_wake && !requeue_pi) {
2114 mark_wake_futex(&wake_q, this);
2118 /* Ensure we requeue to the expected futex for requeue_pi. */
2119 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2125 * Requeue nr_requeue waiters and possibly one more in the case
2126 * of requeue_pi if we couldn't acquire the lock atomically.
2130 * Prepare the waiter to take the rt_mutex. Take a
2131 * refcount on the pi_state and store the pointer in
2132 * the futex_q object of the waiter.
2134 get_pi_state(pi_state);
2135 this->pi_state = pi_state;
2136 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2141 * We got the lock. We do neither drop the
2142 * refcount on pi_state nor clear
2143 * this->pi_state because the waiter needs the
2144 * pi_state for cleaning up the user space
2145 * value. It will drop the refcount after
2148 requeue_pi_wake_futex(this, &key2, hb2);
2152 * rt_mutex_start_proxy_lock() detected a
2153 * potential deadlock when we tried to queue
2154 * that waiter. Drop the pi_state reference
2155 * which we took above and remove the pointer
2156 * to the state from the waiters futex_q
2159 this->pi_state = NULL;
2160 put_pi_state(pi_state);
2162 * We stop queueing more waiters and let user
2163 * space deal with the mess.
2168 requeue_futex(this, hb1, hb2, &key2);
2172 * We took an extra initial reference to the pi_state either
2173 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2174 * need to drop it here again.
2176 put_pi_state(pi_state);
2179 double_unlock_hb(hb1, hb2);
2181 hb_waiters_dec(hb2);
2182 return ret ? ret : task_count;
2185 /* The key must be already stored in q->key. */
2186 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2187 __acquires(&hb->lock)
2189 struct futex_hash_bucket *hb;
2191 hb = hash_futex(&q->key);
2194 * Increment the counter before taking the lock so that
2195 * a potential waker won't miss a to-be-slept task that is
2196 * waiting for the spinlock. This is safe as all queue_lock()
2197 * users end up calling queue_me(). Similarly, for housekeeping,
2198 * decrement the counter at queue_unlock() when some error has
2199 * occurred and we don't end up adding the task to the list.
2201 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2203 q->lock_ptr = &hb->lock;
2205 spin_lock(&hb->lock);
2210 queue_unlock(struct futex_hash_bucket *hb)
2211 __releases(&hb->lock)
2213 spin_unlock(&hb->lock);
2217 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2222 * The priority used to register this element is
2223 * - either the real thread-priority for the real-time threads
2224 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2225 * - or MAX_RT_PRIO for non-RT threads.
2226 * Thus, all RT-threads are woken first in priority order, and
2227 * the others are woken last, in FIFO order.
2229 prio = min(current->normal_prio, MAX_RT_PRIO);
2231 plist_node_init(&q->list, prio);
2232 plist_add(&q->list, &hb->chain);
2237 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2238 * @q: The futex_q to enqueue
2239 * @hb: The destination hash bucket
2241 * The hb->lock must be held by the caller, and is released here. A call to
2242 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2243 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2244 * or nothing if the unqueue is done as part of the wake process and the unqueue
2245 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2248 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2249 __releases(&hb->lock)
2252 spin_unlock(&hb->lock);
2256 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2257 * @q: The futex_q to unqueue
2259 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2260 * be paired with exactly one earlier call to queue_me().
2263 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2264 * - 0 - if the futex_q was already removed by the waking thread
2266 static int unqueue_me(struct futex_q *q)
2268 spinlock_t *lock_ptr;
2271 /* In the common case we don't take the spinlock, which is nice. */
2274 * q->lock_ptr can change between this read and the following spin_lock.
2275 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2276 * optimizing lock_ptr out of the logic below.
2278 lock_ptr = READ_ONCE(q->lock_ptr);
2279 if (lock_ptr != NULL) {
2280 spin_lock(lock_ptr);
2282 * q->lock_ptr can change between reading it and
2283 * spin_lock(), causing us to take the wrong lock. This
2284 * corrects the race condition.
2286 * Reasoning goes like this: if we have the wrong lock,
2287 * q->lock_ptr must have changed (maybe several times)
2288 * between reading it and the spin_lock(). It can
2289 * change again after the spin_lock() but only if it was
2290 * already changed before the spin_lock(). It cannot,
2291 * however, change back to the original value. Therefore
2292 * we can detect whether we acquired the correct lock.
2294 if (unlikely(lock_ptr != q->lock_ptr)) {
2295 spin_unlock(lock_ptr);
2300 BUG_ON(q->pi_state);
2302 spin_unlock(lock_ptr);
2310 * PI futexes can not be requeued and must remove themself from the
2311 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2314 static void unqueue_me_pi(struct futex_q *q)
2315 __releases(q->lock_ptr)
2319 BUG_ON(!q->pi_state);
2320 put_pi_state(q->pi_state);
2323 spin_unlock(q->lock_ptr);
2326 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2327 struct task_struct *argowner)
2329 struct futex_pi_state *pi_state = q->pi_state;
2330 u32 uval, curval, newval;
2331 struct task_struct *oldowner, *newowner;
2335 lockdep_assert_held(q->lock_ptr);
2337 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2339 oldowner = pi_state->owner;
2342 * We are here because either:
2344 * - we stole the lock and pi_state->owner needs updating to reflect
2345 * that (@argowner == current),
2349 * - someone stole our lock and we need to fix things to point to the
2350 * new owner (@argowner == NULL).
2352 * Either way, we have to replace the TID in the user space variable.
2353 * This must be atomic as we have to preserve the owner died bit here.
2355 * Note: We write the user space value _before_ changing the pi_state
2356 * because we can fault here. Imagine swapped out pages or a fork
2357 * that marked all the anonymous memory readonly for cow.
2359 * Modifying pi_state _before_ the user space value would leave the
2360 * pi_state in an inconsistent state when we fault here, because we
2361 * need to drop the locks to handle the fault. This might be observed
2362 * in the PID check in lookup_pi_state.
2366 if (oldowner != current) {
2368 * We raced against a concurrent self; things are
2369 * already fixed up. Nothing to do.
2375 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2376 /* We got the lock after all, nothing to fix. */
2382 * The trylock just failed, so either there is an owner or
2383 * there is a higher priority waiter than this one.
2385 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2387 * If the higher priority waiter has not yet taken over the
2388 * rtmutex then newowner is NULL. We can't return here with
2389 * that state because it's inconsistent vs. the user space
2390 * state. So drop the locks and try again. It's a valid
2391 * situation and not any different from the other retry
2394 if (unlikely(!newowner)) {
2399 WARN_ON_ONCE(argowner != current);
2400 if (oldowner == current) {
2402 * We raced against a concurrent self; things are
2403 * already fixed up. Nothing to do.
2408 newowner = argowner;
2411 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2413 if (!pi_state->owner)
2414 newtid |= FUTEX_OWNER_DIED;
2416 err = get_futex_value_locked(&uval, uaddr);
2421 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2423 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2433 * We fixed up user space. Now we need to fix the pi_state
2436 if (pi_state->owner != NULL) {
2437 raw_spin_lock(&pi_state->owner->pi_lock);
2438 WARN_ON(list_empty(&pi_state->list));
2439 list_del_init(&pi_state->list);
2440 raw_spin_unlock(&pi_state->owner->pi_lock);
2443 pi_state->owner = newowner;
2445 raw_spin_lock(&newowner->pi_lock);
2446 WARN_ON(!list_empty(&pi_state->list));
2447 list_add(&pi_state->list, &newowner->pi_state_list);
2448 raw_spin_unlock(&newowner->pi_lock);
2449 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2454 * In order to reschedule or handle a page fault, we need to drop the
2455 * locks here. In the case of a fault, this gives the other task
2456 * (either the highest priority waiter itself or the task which stole
2457 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2458 * are back from handling the fault we need to check the pi_state after
2459 * reacquiring the locks and before trying to do another fixup. When
2460 * the fixup has been done already we simply return.
2462 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2463 * drop hb->lock since the caller owns the hb -> futex_q relation.
2464 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2467 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2468 spin_unlock(q->lock_ptr);
2472 ret = fault_in_user_writeable(uaddr);
2486 spin_lock(q->lock_ptr);
2487 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2490 * Check if someone else fixed it for us:
2492 if (pi_state->owner != oldowner) {
2503 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2507 static long futex_wait_restart(struct restart_block *restart);
2510 * fixup_owner() - Post lock pi_state and corner case management
2511 * @uaddr: user address of the futex
2512 * @q: futex_q (contains pi_state and access to the rt_mutex)
2513 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2515 * After attempting to lock an rt_mutex, this function is called to cleanup
2516 * the pi_state owner as well as handle race conditions that may allow us to
2517 * acquire the lock. Must be called with the hb lock held.
2520 * - 1 - success, lock taken;
2521 * - 0 - success, lock not taken;
2522 * - <0 - on error (-EFAULT)
2524 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2530 * Got the lock. We might not be the anticipated owner if we
2531 * did a lock-steal - fix up the PI-state in that case:
2533 * Speculative pi_state->owner read (we don't hold wait_lock);
2534 * since we own the lock pi_state->owner == current is the
2535 * stable state, anything else needs more attention.
2537 if (q->pi_state->owner != current)
2538 ret = fixup_pi_state_owner(uaddr, q, current);
2539 return ret ? ret : locked;
2543 * If we didn't get the lock; check if anybody stole it from us. In
2544 * that case, we need to fix up the uval to point to them instead of
2545 * us, otherwise bad things happen. [10]
2547 * Another speculative read; pi_state->owner == current is unstable
2548 * but needs our attention.
2550 if (q->pi_state->owner == current) {
2551 ret = fixup_pi_state_owner(uaddr, q, NULL);
2556 * Paranoia check. If we did not take the lock, then we should not be
2557 * the owner of the rt_mutex.
2559 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2560 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2561 "pi-state %p\n", ret,
2562 q->pi_state->pi_mutex.owner,
2563 q->pi_state->owner);
2570 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571 * @hb: the futex hash bucket, must be locked by the caller
2572 * @q: the futex_q to queue up on
2573 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2575 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2576 struct hrtimer_sleeper *timeout)
2579 * The task state is guaranteed to be set before another task can
2580 * wake it. set_current_state() is implemented using smp_store_mb() and
2581 * queue_me() calls spin_unlock() upon completion, both serializing
2582 * access to the hash list and forcing another memory barrier.
2584 set_current_state(TASK_INTERRUPTIBLE);
2589 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2592 * If we have been removed from the hash list, then another task
2593 * has tried to wake us, and we can skip the call to schedule().
2595 if (likely(!plist_node_empty(&q->list))) {
2597 * If the timer has already expired, current will already be
2598 * flagged for rescheduling. Only call schedule if there
2599 * is no timeout, or if it has yet to expire.
2601 if (!timeout || timeout->task)
2602 freezable_schedule();
2604 __set_current_state(TASK_RUNNING);
2608 * futex_wait_setup() - Prepare to wait on a futex
2609 * @uaddr: the futex userspace address
2610 * @val: the expected value
2611 * @flags: futex flags (FLAGS_SHARED, etc.)
2612 * @q: the associated futex_q
2613 * @hb: storage for hash_bucket pointer to be returned to caller
2615 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2616 * compare it with the expected value. Handle atomic faults internally.
2617 * Return with the hb lock held and a q.key reference on success, and unlocked
2618 * with no q.key reference on failure.
2621 * - 0 - uaddr contains val and hb has been locked;
2622 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2624 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2625 struct futex_q *q, struct futex_hash_bucket **hb)
2631 * Access the page AFTER the hash-bucket is locked.
2632 * Order is important:
2634 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2635 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2637 * The basic logical guarantee of a futex is that it blocks ONLY
2638 * if cond(var) is known to be true at the time of blocking, for
2639 * any cond. If we locked the hash-bucket after testing *uaddr, that
2640 * would open a race condition where we could block indefinitely with
2641 * cond(var) false, which would violate the guarantee.
2643 * On the other hand, we insert q and release the hash-bucket only
2644 * after testing *uaddr. This guarantees that futex_wait() will NOT
2645 * absorb a wakeup if *uaddr does not match the desired values
2646 * while the syscall executes.
2649 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2650 if (unlikely(ret != 0))
2654 *hb = queue_lock(q);
2656 ret = get_futex_value_locked(&uval, uaddr);
2661 ret = get_user(uval, uaddr);
2665 if (!(flags & FLAGS_SHARED))
2679 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2680 ktime_t *abs_time, u32 bitset)
2682 struct hrtimer_sleeper timeout, *to;
2683 struct restart_block *restart;
2684 struct futex_hash_bucket *hb;
2685 struct futex_q q = futex_q_init;
2692 to = futex_setup_timer(abs_time, &timeout, flags,
2693 current->timer_slack_ns);
2696 * Prepare to wait on uaddr. On success, holds hb lock and increments
2699 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2703 /* queue_me and wait for wakeup, timeout, or a signal. */
2704 futex_wait_queue_me(hb, &q, to);
2706 /* If we were woken (and unqueued), we succeeded, whatever. */
2708 /* unqueue_me() drops q.key ref */
2709 if (!unqueue_me(&q))
2712 if (to && !to->task)
2716 * We expect signal_pending(current), but we might be the
2717 * victim of a spurious wakeup as well.
2719 if (!signal_pending(current))
2726 restart = ¤t->restart_block;
2727 restart->fn = futex_wait_restart;
2728 restart->futex.uaddr = uaddr;
2729 restart->futex.val = val;
2730 restart->futex.time = *abs_time;
2731 restart->futex.bitset = bitset;
2732 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2734 ret = -ERESTART_RESTARTBLOCK;
2738 hrtimer_cancel(&to->timer);
2739 destroy_hrtimer_on_stack(&to->timer);
2745 static long futex_wait_restart(struct restart_block *restart)
2747 u32 __user *uaddr = restart->futex.uaddr;
2748 ktime_t t, *tp = NULL;
2750 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2751 t = restart->futex.time;
2754 restart->fn = do_no_restart_syscall;
2756 return (long)futex_wait(uaddr, restart->futex.flags,
2757 restart->futex.val, tp, restart->futex.bitset);
2762 * Userspace tried a 0 -> TID atomic transition of the futex value
2763 * and failed. The kernel side here does the whole locking operation:
2764 * if there are waiters then it will block as a consequence of relying
2765 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2766 * a 0 value of the futex too.).
2768 * Also serves as futex trylock_pi()'ing, and due semantics.
2770 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2771 ktime_t *time, int trylock)
2773 struct hrtimer_sleeper timeout, *to;
2774 struct futex_pi_state *pi_state = NULL;
2775 struct task_struct *exiting = NULL;
2776 struct rt_mutex_waiter rt_waiter;
2777 struct futex_hash_bucket *hb;
2778 struct futex_q q = futex_q_init;
2781 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2784 if (refill_pi_state_cache())
2787 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2790 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2791 if (unlikely(ret != 0))
2795 hb = queue_lock(&q);
2797 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2799 if (unlikely(ret)) {
2801 * Atomic work succeeded and we got the lock,
2802 * or failed. Either way, we do _not_ block.
2806 /* We got the lock. */
2808 goto out_unlock_put_key;
2814 * Two reasons for this:
2815 * - EBUSY: Task is exiting and we just wait for the
2817 * - EAGAIN: The user space value changed.
2821 * Handle the case where the owner is in the middle of
2822 * exiting. Wait for the exit to complete otherwise
2823 * this task might loop forever, aka. live lock.
2825 wait_for_owner_exiting(ret, exiting);
2829 goto out_unlock_put_key;
2833 WARN_ON(!q.pi_state);
2836 * Only actually queue now that the atomic ops are done:
2841 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2842 /* Fixup the trylock return value: */
2843 ret = ret ? 0 : -EWOULDBLOCK;
2847 rt_mutex_init_waiter(&rt_waiter);
2850 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2851 * hold it while doing rt_mutex_start_proxy(), because then it will
2852 * include hb->lock in the blocking chain, even through we'll not in
2853 * fact hold it while blocking. This will lead it to report -EDEADLK
2854 * and BUG when futex_unlock_pi() interleaves with this.
2856 * Therefore acquire wait_lock while holding hb->lock, but drop the
2857 * latter before calling __rt_mutex_start_proxy_lock(). This
2858 * interleaves with futex_unlock_pi() -- which does a similar lock
2859 * handoff -- such that the latter can observe the futex_q::pi_state
2860 * before __rt_mutex_start_proxy_lock() is done.
2862 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2863 spin_unlock(q.lock_ptr);
2865 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2866 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2867 * it sees the futex_q::pi_state.
2869 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2870 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2879 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2881 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2884 spin_lock(q.lock_ptr);
2886 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2887 * first acquire the hb->lock before removing the lock from the
2888 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2891 * In particular; it is important that futex_unlock_pi() can not
2892 * observe this inconsistency.
2894 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2899 * Fixup the pi_state owner and possibly acquire the lock if we
2902 res = fixup_owner(uaddr, &q, !ret);
2904 * If fixup_owner() returned an error, proprogate that. If it acquired
2905 * the lock, clear our -ETIMEDOUT or -EINTR.
2908 ret = (res < 0) ? res : 0;
2911 * If fixup_owner() faulted and was unable to handle the fault, unlock
2912 * it and return the fault to userspace.
2914 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2915 pi_state = q.pi_state;
2916 get_pi_state(pi_state);
2919 /* Unqueue and drop the lock */
2923 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2924 put_pi_state(pi_state);
2934 hrtimer_cancel(&to->timer);
2935 destroy_hrtimer_on_stack(&to->timer);
2937 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2942 ret = fault_in_user_writeable(uaddr);
2946 if (!(flags & FLAGS_SHARED))
2953 * Userspace attempted a TID -> 0 atomic transition, and failed.
2954 * This is the in-kernel slowpath: we look up the PI state (if any),
2955 * and do the rt-mutex unlock.
2957 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2959 u32 curval, uval, vpid = task_pid_vnr(current);
2960 union futex_key key = FUTEX_KEY_INIT;
2961 struct futex_hash_bucket *hb;
2962 struct futex_q *top_waiter;
2965 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2969 if (get_user(uval, uaddr))
2972 * We release only a lock we actually own:
2974 if ((uval & FUTEX_TID_MASK) != vpid)
2977 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2981 hb = hash_futex(&key);
2982 spin_lock(&hb->lock);
2985 * Check waiters first. We do not trust user space values at
2986 * all and we at least want to know if user space fiddled
2987 * with the futex value instead of blindly unlocking.
2989 top_waiter = futex_top_waiter(hb, &key);
2991 struct futex_pi_state *pi_state = top_waiter->pi_state;
2998 * If current does not own the pi_state then the futex is
2999 * inconsistent and user space fiddled with the futex value.
3001 if (pi_state->owner != current)
3004 get_pi_state(pi_state);
3006 * By taking wait_lock while still holding hb->lock, we ensure
3007 * there is no point where we hold neither; and therefore
3008 * wake_futex_pi() must observe a state consistent with what we
3011 * In particular; this forces __rt_mutex_start_proxy() to
3012 * complete such that we're guaranteed to observe the
3013 * rt_waiter. Also see the WARN in wake_futex_pi().
3015 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3016 spin_unlock(&hb->lock);
3018 /* drops pi_state->pi_mutex.wait_lock */
3019 ret = wake_futex_pi(uaddr, uval, pi_state);
3021 put_pi_state(pi_state);
3024 * Success, we're done! No tricky corner cases.
3029 * The atomic access to the futex value generated a
3030 * pagefault, so retry the user-access and the wakeup:
3035 * A unconditional UNLOCK_PI op raced against a waiter
3036 * setting the FUTEX_WAITERS bit. Try again.
3041 * wake_futex_pi has detected invalid state. Tell user
3048 * We have no kernel internal state, i.e. no waiters in the
3049 * kernel. Waiters which are about to queue themselves are stuck
3050 * on hb->lock. So we can safely ignore them. We do neither
3051 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3054 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3055 spin_unlock(&hb->lock);
3070 * If uval has changed, let user space handle it.
3072 ret = (curval == uval) ? 0 : -EAGAIN;
3075 spin_unlock(&hb->lock);
3085 ret = fault_in_user_writeable(uaddr);
3093 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3094 * @hb: the hash_bucket futex_q was original enqueued on
3095 * @q: the futex_q woken while waiting to be requeued
3096 * @key2: the futex_key of the requeue target futex
3097 * @timeout: the timeout associated with the wait (NULL if none)
3099 * Detect if the task was woken on the initial futex as opposed to the requeue
3100 * target futex. If so, determine if it was a timeout or a signal that caused
3101 * the wakeup and return the appropriate error code to the caller. Must be
3102 * called with the hb lock held.
3105 * - 0 = no early wakeup detected;
3106 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3109 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3110 struct futex_q *q, union futex_key *key2,
3111 struct hrtimer_sleeper *timeout)
3116 * With the hb lock held, we avoid races while we process the wakeup.
3117 * We only need to hold hb (and not hb2) to ensure atomicity as the
3118 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3119 * It can't be requeued from uaddr2 to something else since we don't
3120 * support a PI aware source futex for requeue.
3122 if (!match_futex(&q->key, key2)) {
3123 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3125 * We were woken prior to requeue by a timeout or a signal.
3126 * Unqueue the futex_q and determine which it was.
3128 plist_del(&q->list, &hb->chain);
3131 /* Handle spurious wakeups gracefully */
3133 if (timeout && !timeout->task)
3135 else if (signal_pending(current))
3136 ret = -ERESTARTNOINTR;
3142 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3143 * @uaddr: the futex we initially wait on (non-pi)
3144 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3145 * the same type, no requeueing from private to shared, etc.
3146 * @val: the expected value of uaddr
3147 * @abs_time: absolute timeout
3148 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3149 * @uaddr2: the pi futex we will take prior to returning to user-space
3151 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3152 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3153 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3154 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3155 * without one, the pi logic would not know which task to boost/deboost, if
3156 * there was a need to.
3158 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3159 * via the following--
3160 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3161 * 2) wakeup on uaddr2 after a requeue
3165 * If 3, cleanup and return -ERESTARTNOINTR.
3167 * If 2, we may then block on trying to take the rt_mutex and return via:
3168 * 5) successful lock
3171 * 8) other lock acquisition failure
3173 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3175 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3181 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3182 u32 val, ktime_t *abs_time, u32 bitset,
3185 struct hrtimer_sleeper timeout, *to;
3186 struct futex_pi_state *pi_state = NULL;
3187 struct rt_mutex_waiter rt_waiter;
3188 struct futex_hash_bucket *hb;
3189 union futex_key key2 = FUTEX_KEY_INIT;
3190 struct futex_q q = futex_q_init;
3193 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3196 if (uaddr == uaddr2)
3202 to = futex_setup_timer(abs_time, &timeout, flags,
3203 current->timer_slack_ns);
3206 * The waiter is allocated on our stack, manipulated by the requeue
3207 * code while we sleep on uaddr.
3209 rt_mutex_init_waiter(&rt_waiter);
3211 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3212 if (unlikely(ret != 0))
3216 q.rt_waiter = &rt_waiter;
3217 q.requeue_pi_key = &key2;
3220 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3223 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3228 * The check above which compares uaddrs is not sufficient for
3229 * shared futexes. We need to compare the keys:
3231 if (match_futex(&q.key, &key2)) {
3237 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3238 futex_wait_queue_me(hb, &q, to);
3240 spin_lock(&hb->lock);
3241 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3242 spin_unlock(&hb->lock);
3247 * In order for us to be here, we know our q.key == key2, and since
3248 * we took the hb->lock above, we also know that futex_requeue() has
3249 * completed and we no longer have to concern ourselves with a wakeup
3250 * race with the atomic proxy lock acquisition by the requeue code. The
3251 * futex_requeue dropped our key1 reference and incremented our key2
3255 /* Check if the requeue code acquired the second futex for us. */
3258 * Got the lock. We might not be the anticipated owner if we
3259 * did a lock-steal - fix up the PI-state in that case.
3261 if (q.pi_state && (q.pi_state->owner != current)) {
3262 spin_lock(q.lock_ptr);
3263 ret = fixup_pi_state_owner(uaddr2, &q, current);
3264 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3265 pi_state = q.pi_state;
3266 get_pi_state(pi_state);
3269 * Drop the reference to the pi state which
3270 * the requeue_pi() code acquired for us.
3272 put_pi_state(q.pi_state);
3273 spin_unlock(q.lock_ptr);
3276 struct rt_mutex *pi_mutex;
3279 * We have been woken up by futex_unlock_pi(), a timeout, or a
3280 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3283 WARN_ON(!q.pi_state);
3284 pi_mutex = &q.pi_state->pi_mutex;
3285 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3287 spin_lock(q.lock_ptr);
3288 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3291 debug_rt_mutex_free_waiter(&rt_waiter);
3293 * Fixup the pi_state owner and possibly acquire the lock if we
3296 res = fixup_owner(uaddr2, &q, !ret);
3298 * If fixup_owner() returned an error, proprogate that. If it
3299 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3302 ret = (res < 0) ? res : 0;
3305 * If fixup_pi_state_owner() faulted and was unable to handle
3306 * the fault, unlock the rt_mutex and return the fault to
3309 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3310 pi_state = q.pi_state;
3311 get_pi_state(pi_state);
3314 /* Unqueue and drop the lock. */
3319 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3320 put_pi_state(pi_state);
3323 if (ret == -EINTR) {
3325 * We've already been requeued, but cannot restart by calling
3326 * futex_lock_pi() directly. We could restart this syscall, but
3327 * it would detect that the user space "val" changed and return
3328 * -EWOULDBLOCK. Save the overhead of the restart and return
3329 * -EWOULDBLOCK directly.
3336 hrtimer_cancel(&to->timer);
3337 destroy_hrtimer_on_stack(&to->timer);
3343 * Support for robust futexes: the kernel cleans up held futexes at
3346 * Implementation: user-space maintains a per-thread list of locks it
3347 * is holding. Upon do_exit(), the kernel carefully walks this list,
3348 * and marks all locks that are owned by this thread with the
3349 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3350 * always manipulated with the lock held, so the list is private and
3351 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3352 * field, to allow the kernel to clean up if the thread dies after
3353 * acquiring the lock, but just before it could have added itself to
3354 * the list. There can only be one such pending lock.
3358 * sys_set_robust_list() - Set the robust-futex list head of a task
3359 * @head: pointer to the list-head
3360 * @len: length of the list-head, as userspace expects
3362 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3365 if (!futex_cmpxchg_enabled)
3368 * The kernel knows only one size for now:
3370 if (unlikely(len != sizeof(*head)))
3373 current->robust_list = head;
3379 * sys_get_robust_list() - Get the robust-futex list head of a task
3380 * @pid: pid of the process [zero for current task]
3381 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3382 * @len_ptr: pointer to a length field, the kernel fills in the header size
3384 SYSCALL_DEFINE3(get_robust_list, int, pid,
3385 struct robust_list_head __user * __user *, head_ptr,
3386 size_t __user *, len_ptr)
3388 struct robust_list_head __user *head;
3390 struct task_struct *p;
3392 if (!futex_cmpxchg_enabled)
3401 p = find_task_by_vpid(pid);
3407 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3410 head = p->robust_list;
3413 if (put_user(sizeof(*head), len_ptr))
3415 return put_user(head, head_ptr);
3423 /* Constants for the pending_op argument of handle_futex_death */
3424 #define HANDLE_DEATH_PENDING true
3425 #define HANDLE_DEATH_LIST false
3428 * Process a futex-list entry, check whether it's owned by the
3429 * dying task, and do notification if so:
3431 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3432 bool pi, bool pending_op)
3434 u32 uval, nval, mval;
3437 /* Futex address must be 32bit aligned */
3438 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3442 if (get_user(uval, uaddr))
3446 * Special case for regular (non PI) futexes. The unlock path in
3447 * user space has two race scenarios:
3449 * 1. The unlock path releases the user space futex value and
3450 * before it can execute the futex() syscall to wake up
3451 * waiters it is killed.
3453 * 2. A woken up waiter is killed before it can acquire the
3454 * futex in user space.
3456 * In both cases the TID validation below prevents a wakeup of
3457 * potential waiters which can cause these waiters to block
3460 * In both cases the following conditions are met:
3462 * 1) task->robust_list->list_op_pending != NULL
3463 * @pending_op == true
3464 * 2) User space futex value == 0
3465 * 3) Regular futex: @pi == false
3467 * If these conditions are met, it is safe to attempt waking up a
3468 * potential waiter without touching the user space futex value and
3469 * trying to set the OWNER_DIED bit. The user space futex value is
3470 * uncontended and the rest of the user space mutex state is
3471 * consistent, so a woken waiter will just take over the
3472 * uncontended futex. Setting the OWNER_DIED bit would create
3473 * inconsistent state and malfunction of the user space owner died
3476 if (pending_op && !pi && !uval) {
3477 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3481 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3485 * Ok, this dying thread is truly holding a futex
3486 * of interest. Set the OWNER_DIED bit atomically
3487 * via cmpxchg, and if the value had FUTEX_WAITERS
3488 * set, wake up a waiter (if any). (We have to do a
3489 * futex_wake() even if OWNER_DIED is already set -
3490 * to handle the rare but possible case of recursive
3491 * thread-death.) The rest of the cleanup is done in
3494 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3497 * We are not holding a lock here, but we want to have
3498 * the pagefault_disable/enable() protection because
3499 * we want to handle the fault gracefully. If the
3500 * access fails we try to fault in the futex with R/W
3501 * verification via get_user_pages. get_user() above
3502 * does not guarantee R/W access. If that fails we
3503 * give up and leave the futex locked.
3505 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3508 if (fault_in_user_writeable(uaddr))
3526 * Wake robust non-PI futexes here. The wakeup of
3527 * PI futexes happens in exit_pi_state():
3529 if (!pi && (uval & FUTEX_WAITERS))
3530 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3536 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3538 static inline int fetch_robust_entry(struct robust_list __user **entry,
3539 struct robust_list __user * __user *head,
3542 unsigned long uentry;
3544 if (get_user(uentry, (unsigned long __user *)head))
3547 *entry = (void __user *)(uentry & ~1UL);
3554 * Walk curr->robust_list (very carefully, it's a userspace list!)
3555 * and mark any locks found there dead, and notify any waiters.
3557 * We silently return on any sign of list-walking problem.
3559 static void exit_robust_list(struct task_struct *curr)
3561 struct robust_list_head __user *head = curr->robust_list;
3562 struct robust_list __user *entry, *next_entry, *pending;
3563 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3564 unsigned int next_pi;
3565 unsigned long futex_offset;
3568 if (!futex_cmpxchg_enabled)
3572 * Fetch the list head (which was registered earlier, via
3573 * sys_set_robust_list()):
3575 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3578 * Fetch the relative futex offset:
3580 if (get_user(futex_offset, &head->futex_offset))
3583 * Fetch any possibly pending lock-add first, and handle it
3586 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3589 next_entry = NULL; /* avoid warning with gcc */
3590 while (entry != &head->list) {
3592 * Fetch the next entry in the list before calling
3593 * handle_futex_death:
3595 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3597 * A pending lock might already be on the list, so
3598 * don't process it twice:
3600 if (entry != pending) {
3601 if (handle_futex_death((void __user *)entry + futex_offset,
3602 curr, pi, HANDLE_DEATH_LIST))
3610 * Avoid excessively long or circular lists:
3619 handle_futex_death((void __user *)pending + futex_offset,
3620 curr, pip, HANDLE_DEATH_PENDING);
3624 static void futex_cleanup(struct task_struct *tsk)
3626 if (unlikely(tsk->robust_list)) {
3627 exit_robust_list(tsk);
3628 tsk->robust_list = NULL;
3631 #ifdef CONFIG_COMPAT
3632 if (unlikely(tsk->compat_robust_list)) {
3633 compat_exit_robust_list(tsk);
3634 tsk->compat_robust_list = NULL;
3638 if (unlikely(!list_empty(&tsk->pi_state_list)))
3639 exit_pi_state_list(tsk);
3643 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3644 * @tsk: task to set the state on
3646 * Set the futex exit state of the task lockless. The futex waiter code
3647 * observes that state when a task is exiting and loops until the task has
3648 * actually finished the futex cleanup. The worst case for this is that the
3649 * waiter runs through the wait loop until the state becomes visible.
3651 * This is called from the recursive fault handling path in do_exit().
3653 * This is best effort. Either the futex exit code has run already or
3654 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3655 * take it over. If not, the problem is pushed back to user space. If the
3656 * futex exit code did not run yet, then an already queued waiter might
3657 * block forever, but there is nothing which can be done about that.
3659 void futex_exit_recursive(struct task_struct *tsk)
3661 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3662 if (tsk->futex_state == FUTEX_STATE_EXITING)
3663 mutex_unlock(&tsk->futex_exit_mutex);
3664 tsk->futex_state = FUTEX_STATE_DEAD;
3667 static void futex_cleanup_begin(struct task_struct *tsk)
3670 * Prevent various race issues against a concurrent incoming waiter
3671 * including live locks by forcing the waiter to block on
3672 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3673 * attach_to_pi_owner().
3675 mutex_lock(&tsk->futex_exit_mutex);
3678 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3680 * This ensures that all subsequent checks of tsk->futex_state in
3681 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3682 * tsk->pi_lock held.
3684 * It guarantees also that a pi_state which was queued right before
3685 * the state change under tsk->pi_lock by a concurrent waiter must
3686 * be observed in exit_pi_state_list().
3688 raw_spin_lock_irq(&tsk->pi_lock);
3689 tsk->futex_state = FUTEX_STATE_EXITING;
3690 raw_spin_unlock_irq(&tsk->pi_lock);
3693 static void futex_cleanup_end(struct task_struct *tsk, int state)
3696 * Lockless store. The only side effect is that an observer might
3697 * take another loop until it becomes visible.
3699 tsk->futex_state = state;
3701 * Drop the exit protection. This unblocks waiters which observed
3702 * FUTEX_STATE_EXITING to reevaluate the state.
3704 mutex_unlock(&tsk->futex_exit_mutex);
3707 void futex_exec_release(struct task_struct *tsk)
3710 * The state handling is done for consistency, but in the case of
3711 * exec() there is no way to prevent futher damage as the PID stays
3712 * the same. But for the unlikely and arguably buggy case that a
3713 * futex is held on exec(), this provides at least as much state
3714 * consistency protection which is possible.
3716 futex_cleanup_begin(tsk);
3719 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3720 * exec a new binary.
3722 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3725 void futex_exit_release(struct task_struct *tsk)
3727 futex_cleanup_begin(tsk);
3729 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3732 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3733 u32 __user *uaddr2, u32 val2, u32 val3)
3735 int cmd = op & FUTEX_CMD_MASK;
3736 unsigned int flags = 0;
3738 if (!(op & FUTEX_PRIVATE_FLAG))
3739 flags |= FLAGS_SHARED;
3741 if (op & FUTEX_CLOCK_REALTIME) {
3742 flags |= FLAGS_CLOCKRT;
3743 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3744 cmd != FUTEX_WAIT_REQUEUE_PI)
3750 case FUTEX_UNLOCK_PI:
3751 case FUTEX_TRYLOCK_PI:
3752 case FUTEX_WAIT_REQUEUE_PI:
3753 case FUTEX_CMP_REQUEUE_PI:
3754 if (!futex_cmpxchg_enabled)
3760 val3 = FUTEX_BITSET_MATCH_ANY;
3762 case FUTEX_WAIT_BITSET:
3763 return futex_wait(uaddr, flags, val, timeout, val3);
3765 val3 = FUTEX_BITSET_MATCH_ANY;
3767 case FUTEX_WAKE_BITSET:
3768 return futex_wake(uaddr, flags, val, val3);
3770 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3771 case FUTEX_CMP_REQUEUE:
3772 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3774 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3776 return futex_lock_pi(uaddr, flags, timeout, 0);
3777 case FUTEX_UNLOCK_PI:
3778 return futex_unlock_pi(uaddr, flags);
3779 case FUTEX_TRYLOCK_PI:
3780 return futex_lock_pi(uaddr, flags, NULL, 1);
3781 case FUTEX_WAIT_REQUEUE_PI:
3782 val3 = FUTEX_BITSET_MATCH_ANY;
3783 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3785 case FUTEX_CMP_REQUEUE_PI:
3786 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3792 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3793 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3796 struct timespec64 ts;
3797 ktime_t t, *tp = NULL;
3799 int cmd = op & FUTEX_CMD_MASK;
3801 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3802 cmd == FUTEX_WAIT_BITSET ||
3803 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3804 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3806 if (get_timespec64(&ts, utime))
3808 if (!timespec64_valid(&ts))
3811 t = timespec64_to_ktime(ts);
3812 if (cmd == FUTEX_WAIT)
3813 t = ktime_add_safe(ktime_get(), t);
3814 else if (!(op & FUTEX_CLOCK_REALTIME))
3815 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3819 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3820 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3822 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3823 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3824 val2 = (u32) (unsigned long) utime;
3826 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3829 #ifdef CONFIG_COMPAT
3831 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3834 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3835 compat_uptr_t __user *head, unsigned int *pi)
3837 if (get_user(*uentry, head))
3840 *entry = compat_ptr((*uentry) & ~1);
3841 *pi = (unsigned int)(*uentry) & 1;
3846 static void __user *futex_uaddr(struct robust_list __user *entry,
3847 compat_long_t futex_offset)
3849 compat_uptr_t base = ptr_to_compat(entry);
3850 void __user *uaddr = compat_ptr(base + futex_offset);
3856 * Walk curr->robust_list (very carefully, it's a userspace list!)
3857 * and mark any locks found there dead, and notify any waiters.
3859 * We silently return on any sign of list-walking problem.
3861 static void compat_exit_robust_list(struct task_struct *curr)
3863 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3864 struct robust_list __user *entry, *next_entry, *pending;
3865 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3866 unsigned int next_pi;
3867 compat_uptr_t uentry, next_uentry, upending;
3868 compat_long_t futex_offset;
3871 if (!futex_cmpxchg_enabled)
3875 * Fetch the list head (which was registered earlier, via
3876 * sys_set_robust_list()):
3878 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3881 * Fetch the relative futex offset:
3883 if (get_user(futex_offset, &head->futex_offset))
3886 * Fetch any possibly pending lock-add first, and handle it
3889 if (compat_fetch_robust_entry(&upending, &pending,
3890 &head->list_op_pending, &pip))
3893 next_entry = NULL; /* avoid warning with gcc */
3894 while (entry != (struct robust_list __user *) &head->list) {
3896 * Fetch the next entry in the list before calling
3897 * handle_futex_death:
3899 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3900 (compat_uptr_t __user *)&entry->next, &next_pi);
3902 * A pending lock might already be on the list, so
3903 * dont process it twice:
3905 if (entry != pending) {
3906 void __user *uaddr = futex_uaddr(entry, futex_offset);
3908 if (handle_futex_death(uaddr, curr, pi,
3914 uentry = next_uentry;
3918 * Avoid excessively long or circular lists:
3926 void __user *uaddr = futex_uaddr(pending, futex_offset);
3928 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3932 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3933 struct compat_robust_list_head __user *, head,
3936 if (!futex_cmpxchg_enabled)
3939 if (unlikely(len != sizeof(*head)))
3942 current->compat_robust_list = head;
3947 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3948 compat_uptr_t __user *, head_ptr,
3949 compat_size_t __user *, len_ptr)
3951 struct compat_robust_list_head __user *head;
3953 struct task_struct *p;
3955 if (!futex_cmpxchg_enabled)
3964 p = find_task_by_vpid(pid);
3970 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3973 head = p->compat_robust_list;
3976 if (put_user(sizeof(*head), len_ptr))
3978 return put_user(ptr_to_compat(head), head_ptr);
3985 #endif /* CONFIG_COMPAT */
3987 #ifdef CONFIG_COMPAT_32BIT_TIME
3988 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3989 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3992 struct timespec64 ts;
3993 ktime_t t, *tp = NULL;
3995 int cmd = op & FUTEX_CMD_MASK;
3997 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3998 cmd == FUTEX_WAIT_BITSET ||
3999 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4000 if (get_old_timespec32(&ts, utime))
4002 if (!timespec64_valid(&ts))
4005 t = timespec64_to_ktime(ts);
4006 if (cmd == FUTEX_WAIT)
4007 t = ktime_add_safe(ktime_get(), t);
4008 else if (!(op & FUTEX_CLOCK_REALTIME))
4009 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
4012 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4013 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4014 val2 = (int) (unsigned long) utime;
4016 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4018 #endif /* CONFIG_COMPAT_32BIT_TIME */
4020 static void __init futex_detect_cmpxchg(void)
4022 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4026 * This will fail and we want it. Some arch implementations do
4027 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4028 * functionality. We want to know that before we call in any
4029 * of the complex code paths. Also we want to prevent
4030 * registration of robust lists in that case. NULL is
4031 * guaranteed to fault and we get -EFAULT on functional
4032 * implementation, the non-functional ones will return
4035 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4036 futex_cmpxchg_enabled = 1;
4040 static int __init futex_init(void)
4042 unsigned int futex_shift;
4045 #if CONFIG_BASE_SMALL
4046 futex_hashsize = 16;
4048 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4051 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4053 futex_hashsize < 256 ? HASH_SMALL : 0,
4055 futex_hashsize, futex_hashsize);
4056 futex_hashsize = 1UL << futex_shift;
4058 futex_detect_cmpxchg();
4060 for (i = 0; i < futex_hashsize; i++) {
4061 atomic_set(&futex_queues[i].waiters, 0);
4062 plist_head_init(&futex_queues[i].chain);
4063 spin_lock_init(&futex_queues[i].lock);
4068 core_initcall(futex_init);