1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/slab.h>
36 #include <linux/poll.h>
38 #include <linux/file.h>
39 #include <linux/jhash.h>
40 #include <linux/init.h>
41 #include <linux/futex.h>
42 #include <linux/mount.h>
43 #include <linux/pagemap.h>
44 #include <linux/syscalls.h>
45 #include <linux/signal.h>
46 #include <linux/export.h>
47 #include <linux/magic.h>
48 #include <linux/pid.h>
49 #include <linux/nsproxy.h>
50 #include <linux/ptrace.h>
51 #include <linux/sched/rt.h>
52 #include <linux/sched/wake_q.h>
53 #include <linux/sched/mm.h>
54 #include <linux/hugetlb.h>
55 #include <linux/freezer.h>
56 #include <linux/memblock.h>
57 #include <linux/fault-inject.h>
58 #include <linux/refcount.h>
60 #include <asm/futex.h>
62 #include "locking/rtmutex_common.h"
65 * READ this before attempting to hack on futexes!
67 * Basic futex operation and ordering guarantees
68 * =============================================
70 * The waiter reads the futex value in user space and calls
71 * futex_wait(). This function computes the hash bucket and acquires
72 * the hash bucket lock. After that it reads the futex user space value
73 * again and verifies that the data has not changed. If it has not changed
74 * it enqueues itself into the hash bucket, releases the hash bucket lock
77 * The waker side modifies the user space value of the futex and calls
78 * futex_wake(). This function computes the hash bucket and acquires the
79 * hash bucket lock. Then it looks for waiters on that futex in the hash
80 * bucket and wakes them.
82 * In futex wake up scenarios where no tasks are blocked on a futex, taking
83 * the hb spinlock can be avoided and simply return. In order for this
84 * optimization to work, ordering guarantees must exist so that the waiter
85 * being added to the list is acknowledged when the list is concurrently being
86 * checked by the waker, avoiding scenarios like the following:
90 * sys_futex(WAIT, futex, val);
91 * futex_wait(futex, val);
94 * sys_futex(WAKE, futex);
99 * lock(hash_bucket(futex));
101 * unlock(hash_bucket(futex));
104 * This would cause the waiter on CPU 0 to wait forever because it
105 * missed the transition of the user space value from val to newval
106 * and the waker did not find the waiter in the hash bucket queue.
108 * The correct serialization ensures that a waiter either observes
109 * the changed user space value before blocking or is woken by a
114 * sys_futex(WAIT, futex, val);
115 * futex_wait(futex, val);
118 * smp_mb(); (A) <-- paired with -.
120 * lock(hash_bucket(futex)); |
124 * | sys_futex(WAKE, futex);
125 * | futex_wake(futex);
127 * `--------> smp_mb(); (B)
130 * unlock(hash_bucket(futex));
131 * schedule(); if (waiters)
132 * lock(hash_bucket(futex));
133 * else wake_waiters(futex);
134 * waiters--; (b) unlock(hash_bucket(futex));
136 * Where (A) orders the waiters increment and the futex value read through
137 * atomic operations (see hb_waiters_inc) and where (B) orders the write
138 * to futex and the waiters read -- this is done by the barriers for both
139 * shared and private futexes in get_futex_key_refs().
141 * This yields the following case (where X:=waiters, Y:=futex):
149 * Which guarantees that x==0 && y==0 is impossible; which translates back into
150 * the guarantee that we cannot both miss the futex variable change and the
153 * Note that a new waiter is accounted for in (a) even when it is possible that
154 * the wait call can return error, in which case we backtrack from it in (b).
155 * Refer to the comment in queue_lock().
157 * Similarly, in order to account for waiters being requeued on another
158 * address we always increment the waiters for the destination bucket before
159 * acquiring the lock. It then decrements them again after releasing it -
160 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
161 * will do the additional required waiter count housekeeping. This is done for
162 * double_lock_hb() and double_unlock_hb(), respectively.
165 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
166 #define futex_cmpxchg_enabled 1
168 static int __read_mostly futex_cmpxchg_enabled;
172 * Futex flags used to encode options to functions and preserve them across
176 # define FLAGS_SHARED 0x01
179 * NOMMU does not have per process address space. Let the compiler optimize
182 # define FLAGS_SHARED 0x00
184 #define FLAGS_CLOCKRT 0x02
185 #define FLAGS_HAS_TIMEOUT 0x04
188 * Priority Inheritance state:
190 struct futex_pi_state {
192 * list of 'owned' pi_state instances - these have to be
193 * cleaned up in do_exit() if the task exits prematurely:
195 struct list_head list;
200 struct rt_mutex pi_mutex;
202 struct task_struct *owner;
206 } __randomize_layout;
209 * struct futex_q - The hashed futex queue entry, one per waiting task
210 * @list: priority-sorted list of tasks waiting on this futex
211 * @task: the task waiting on the futex
212 * @lock_ptr: the hash bucket lock
213 * @key: the key the futex is hashed on
214 * @pi_state: optional priority inheritance state
215 * @rt_waiter: rt_waiter storage for use with requeue_pi
216 * @requeue_pi_key: the requeue_pi target futex key
217 * @bitset: bitset for the optional bitmasked wakeup
219 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
220 * we can wake only the relevant ones (hashed queues may be shared).
222 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
223 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
224 * The order of wakeup is always to make the first condition true, then
227 * PI futexes are typically woken before they are removed from the hash list via
228 * the rt_mutex code. See unqueue_me_pi().
231 struct plist_node list;
233 struct task_struct *task;
234 spinlock_t *lock_ptr;
236 struct futex_pi_state *pi_state;
237 struct rt_mutex_waiter *rt_waiter;
238 union futex_key *requeue_pi_key;
240 } __randomize_layout;
242 static const struct futex_q futex_q_init = {
243 /* list gets initialized in queue_me()*/
244 .key = FUTEX_KEY_INIT,
245 .bitset = FUTEX_BITSET_MATCH_ANY
249 * Hash buckets are shared by all the futex_keys that hash to the same
250 * location. Each key may have multiple futex_q structures, one for each task
251 * waiting on a futex.
253 struct futex_hash_bucket {
256 struct plist_head chain;
257 } ____cacheline_aligned_in_smp;
260 * The base of the bucket array and its size are always used together
261 * (after initialization only in hash_futex()), so ensure that they
262 * reside in the same cacheline.
265 struct futex_hash_bucket *queues;
266 unsigned long hashsize;
267 } __futex_data __read_mostly __aligned(2*sizeof(long));
268 #define futex_queues (__futex_data.queues)
269 #define futex_hashsize (__futex_data.hashsize)
273 * Fault injections for futexes.
275 #ifdef CONFIG_FAIL_FUTEX
278 struct fault_attr attr;
282 .attr = FAULT_ATTR_INITIALIZER,
283 .ignore_private = false,
286 static int __init setup_fail_futex(char *str)
288 return setup_fault_attr(&fail_futex.attr, str);
290 __setup("fail_futex=", setup_fail_futex);
292 static bool should_fail_futex(bool fshared)
294 if (fail_futex.ignore_private && !fshared)
297 return should_fail(&fail_futex.attr, 1);
300 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
302 static int __init fail_futex_debugfs(void)
304 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
307 dir = fault_create_debugfs_attr("fail_futex", NULL,
312 debugfs_create_bool("ignore-private", mode, dir,
313 &fail_futex.ignore_private);
317 late_initcall(fail_futex_debugfs);
319 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
322 static inline bool should_fail_futex(bool fshared)
326 #endif /* CONFIG_FAIL_FUTEX */
329 static void compat_exit_robust_list(struct task_struct *curr);
331 static inline void compat_exit_robust_list(struct task_struct *curr) { }
334 static inline void futex_get_mm(union futex_key *key)
336 mmgrab(key->private.mm);
338 * Ensure futex_get_mm() implies a full barrier such that
339 * get_futex_key() implies a full barrier. This is relied upon
340 * as smp_mb(); (B), see the ordering comment above.
342 smp_mb__after_atomic();
346 * Reflects a new waiter being added to the waitqueue.
348 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
351 atomic_inc(&hb->waiters);
353 * Full barrier (A), see the ordering comment above.
355 smp_mb__after_atomic();
360 * Reflects a waiter being removed from the waitqueue by wakeup
363 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
366 atomic_dec(&hb->waiters);
370 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
373 return atomic_read(&hb->waiters);
380 * hash_futex - Return the hash bucket in the global hash
381 * @key: Pointer to the futex key for which the hash is calculated
383 * We hash on the keys returned from get_futex_key (see below) and return the
384 * corresponding hash bucket in the global hash.
386 static struct futex_hash_bucket *hash_futex(union futex_key *key)
388 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
391 return &futex_queues[hash & (futex_hashsize - 1)];
396 * match_futex - Check whether two futex keys are equal
397 * @key1: Pointer to key1
398 * @key2: Pointer to key2
400 * Return 1 if two futex_keys are equal, 0 otherwise.
402 static inline int match_futex(union futex_key *key1, union futex_key *key2)
405 && key1->both.word == key2->both.word
406 && key1->both.ptr == key2->both.ptr
407 && key1->both.offset == key2->both.offset);
411 * Take a reference to the resource addressed by a key.
412 * Can be called while holding spinlocks.
415 static void get_futex_key_refs(union futex_key *key)
421 * On MMU less systems futexes are always "private" as there is no per
422 * process address space. We need the smp wmb nevertheless - yes,
423 * arch/blackfin has MMU less SMP ...
425 if (!IS_ENABLED(CONFIG_MMU)) {
426 smp_mb(); /* explicit smp_mb(); (B) */
430 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
432 smp_mb(); /* explicit smp_mb(); (B) */
434 case FUT_OFF_MMSHARED:
435 futex_get_mm(key); /* implies smp_mb(); (B) */
439 * Private futexes do not hold reference on an inode or
440 * mm, therefore the only purpose of calling get_futex_key_refs
441 * is because we need the barrier for the lockless waiter check.
443 smp_mb(); /* explicit smp_mb(); (B) */
448 * Drop a reference to the resource addressed by a key.
449 * The hash bucket spinlock must not be held. This is
450 * a no-op for private futexes, see comment in the get
453 static void drop_futex_key_refs(union futex_key *key)
455 if (!key->both.ptr) {
456 /* If we're here then we tried to put a key we failed to get */
461 if (!IS_ENABLED(CONFIG_MMU))
464 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
467 case FUT_OFF_MMSHARED:
468 mmdrop(key->private.mm);
479 * futex_setup_timer - set up the sleeping hrtimer.
480 * @time: ptr to the given timeout value
481 * @timeout: the hrtimer_sleeper structure to be set up
482 * @flags: futex flags
483 * @range_ns: optional range in ns
485 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
488 static inline struct hrtimer_sleeper *
489 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
490 int flags, u64 range_ns)
495 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
496 CLOCK_REALTIME : CLOCK_MONOTONIC,
499 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
500 * effectively the same as calling hrtimer_set_expires().
502 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
508 * Generate a machine wide unique identifier for this inode.
510 * This relies on u64 not wrapping in the life-time of the machine; which with
511 * 1ns resolution means almost 585 years.
513 * This further relies on the fact that a well formed program will not unmap
514 * the file while it has a (shared) futex waiting on it. This mapping will have
515 * a file reference which pins the mount and inode.
517 * If for some reason an inode gets evicted and read back in again, it will get
518 * a new sequence number and will _NOT_ match, even though it is the exact same
521 * It is important that match_futex() will never have a false-positive, esp.
522 * for PI futexes that can mess up the state. The above argues that false-negatives
523 * are only possible for malformed programs.
525 static u64 get_inode_sequence_number(struct inode *inode)
527 static atomic64_t i_seq;
530 /* Does the inode already have a sequence number? */
531 old = atomic64_read(&inode->i_sequence);
536 u64 new = atomic64_add_return(1, &i_seq);
537 if (WARN_ON_ONCE(!new))
540 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
548 * get_futex_key() - Get parameters which are the keys for a futex
549 * @uaddr: virtual address of the futex
550 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
551 * @key: address where result is stored.
552 * @rw: mapping needs to be read/write (values: FUTEX_READ,
555 * Return: a negative error code or 0
557 * The key words are stored in @key on success.
559 * For shared mappings (when @fshared), the key is:
560 * ( inode->i_sequence, page->index, offset_within_page )
561 * [ also see get_inode_sequence_number() ]
563 * For private mappings (or when !@fshared), the key is:
564 * ( current->mm, address, 0 )
566 * This allows (cross process, where applicable) identification of the futex
567 * without keeping the page pinned for the duration of the FUTEX_WAIT.
569 * lock_page() might sleep, the caller should not hold a spinlock.
572 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
574 unsigned long address = (unsigned long)uaddr;
575 struct mm_struct *mm = current->mm;
576 struct page *page, *tail;
577 struct address_space *mapping;
581 * The futex address must be "naturally" aligned.
583 key->both.offset = address % PAGE_SIZE;
584 if (unlikely((address % sizeof(u32)) != 0))
586 address -= key->both.offset;
588 if (unlikely(!access_ok(uaddr, sizeof(u32))))
591 if (unlikely(should_fail_futex(fshared)))
595 * PROCESS_PRIVATE futexes are fast.
596 * As the mm cannot disappear under us and the 'key' only needs
597 * virtual address, we dont even have to find the underlying vma.
598 * Note : We do have to check 'uaddr' is a valid user address,
599 * but access_ok() should be faster than find_vma()
602 key->private.mm = mm;
603 key->private.address = address;
604 get_futex_key_refs(key); /* implies smp_mb(); (B) */
609 /* Ignore any VERIFY_READ mapping (futex common case) */
610 if (unlikely(should_fail_futex(fshared)))
613 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
615 * If write access is not required (eg. FUTEX_WAIT), try
616 * and get read-only access.
618 if (err == -EFAULT && rw == FUTEX_READ) {
619 err = get_user_pages_fast(address, 1, 0, &page);
628 * The treatment of mapping from this point on is critical. The page
629 * lock protects many things but in this context the page lock
630 * stabilizes mapping, prevents inode freeing in the shared
631 * file-backed region case and guards against movement to swap cache.
633 * Strictly speaking the page lock is not needed in all cases being
634 * considered here and page lock forces unnecessarily serialization
635 * From this point on, mapping will be re-verified if necessary and
636 * page lock will be acquired only if it is unavoidable
638 * Mapping checks require the head page for any compound page so the
639 * head page and mapping is looked up now. For anonymous pages, it
640 * does not matter if the page splits in the future as the key is
641 * based on the address. For filesystem-backed pages, the tail is
642 * required as the index of the page determines the key. For
643 * base pages, there is no tail page and tail == page.
646 page = compound_head(page);
647 mapping = READ_ONCE(page->mapping);
650 * If page->mapping is NULL, then it cannot be a PageAnon
651 * page; but it might be the ZERO_PAGE or in the gate area or
652 * in a special mapping (all cases which we are happy to fail);
653 * or it may have been a good file page when get_user_pages_fast
654 * found it, but truncated or holepunched or subjected to
655 * invalidate_complete_page2 before we got the page lock (also
656 * cases which we are happy to fail). And we hold a reference,
657 * so refcount care in invalidate_complete_page's remove_mapping
658 * prevents drop_caches from setting mapping to NULL beneath us.
660 * The case we do have to guard against is when memory pressure made
661 * shmem_writepage move it from filecache to swapcache beneath us:
662 * an unlikely race, but we do need to retry for page->mapping.
664 if (unlikely(!mapping)) {
668 * Page lock is required to identify which special case above
669 * applies. If this is really a shmem page then the page lock
670 * will prevent unexpected transitions.
673 shmem_swizzled = PageSwapCache(page) || page->mapping;
684 * Private mappings are handled in a simple way.
686 * If the futex key is stored on an anonymous page, then the associated
687 * object is the mm which is implicitly pinned by the calling process.
689 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
690 * it's a read-only handle, it's expected that futexes attach to
691 * the object not the particular process.
693 if (PageAnon(page)) {
695 * A RO anonymous page will never change and thus doesn't make
696 * sense for futex operations.
698 if (unlikely(should_fail_futex(fshared)) || ro) {
703 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
704 key->private.mm = mm;
705 key->private.address = address;
711 * The associated futex object in this case is the inode and
712 * the page->mapping must be traversed. Ordinarily this should
713 * be stabilised under page lock but it's not strictly
714 * necessary in this case as we just want to pin the inode, not
715 * update the radix tree or anything like that.
717 * The RCU read lock is taken as the inode is finally freed
718 * under RCU. If the mapping still matches expectations then the
719 * mapping->host can be safely accessed as being a valid inode.
723 if (READ_ONCE(page->mapping) != mapping) {
730 inode = READ_ONCE(mapping->host);
738 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
739 key->shared.i_seq = get_inode_sequence_number(inode);
740 key->shared.pgoff = page_to_pgoff(tail);
744 get_futex_key_refs(key); /* implies smp_mb(); (B) */
751 static inline void put_futex_key(union futex_key *key)
753 drop_futex_key_refs(key);
757 * fault_in_user_writeable() - Fault in user address and verify RW access
758 * @uaddr: pointer to faulting user space address
760 * Slow path to fixup the fault we just took in the atomic write
763 * We have no generic implementation of a non-destructive write to the
764 * user address. We know that we faulted in the atomic pagefault
765 * disabled section so we can as well avoid the #PF overhead by
766 * calling get_user_pages() right away.
768 static int fault_in_user_writeable(u32 __user *uaddr)
770 struct mm_struct *mm = current->mm;
773 down_read(&mm->mmap_sem);
774 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
775 FAULT_FLAG_WRITE, NULL);
776 up_read(&mm->mmap_sem);
778 return ret < 0 ? ret : 0;
782 * futex_top_waiter() - Return the highest priority waiter on a futex
783 * @hb: the hash bucket the futex_q's reside in
784 * @key: the futex key (to distinguish it from other futex futex_q's)
786 * Must be called with the hb lock held.
788 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
789 union futex_key *key)
791 struct futex_q *this;
793 plist_for_each_entry(this, &hb->chain, list) {
794 if (match_futex(&this->key, key))
800 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
801 u32 uval, u32 newval)
806 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
812 static int get_futex_value_locked(u32 *dest, u32 __user *from)
817 ret = __get_user(*dest, from);
820 return ret ? -EFAULT : 0;
827 static int refill_pi_state_cache(void)
829 struct futex_pi_state *pi_state;
831 if (likely(current->pi_state_cache))
834 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
839 INIT_LIST_HEAD(&pi_state->list);
840 /* pi_mutex gets initialized later */
841 pi_state->owner = NULL;
842 refcount_set(&pi_state->refcount, 1);
843 pi_state->key = FUTEX_KEY_INIT;
845 current->pi_state_cache = pi_state;
850 static struct futex_pi_state *alloc_pi_state(void)
852 struct futex_pi_state *pi_state = current->pi_state_cache;
855 current->pi_state_cache = NULL;
860 static void pi_state_update_owner(struct futex_pi_state *pi_state,
861 struct task_struct *new_owner)
863 struct task_struct *old_owner = pi_state->owner;
865 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
868 raw_spin_lock(&old_owner->pi_lock);
869 WARN_ON(list_empty(&pi_state->list));
870 list_del_init(&pi_state->list);
871 raw_spin_unlock(&old_owner->pi_lock);
875 raw_spin_lock(&new_owner->pi_lock);
876 WARN_ON(!list_empty(&pi_state->list));
877 list_add(&pi_state->list, &new_owner->pi_state_list);
878 pi_state->owner = new_owner;
879 raw_spin_unlock(&new_owner->pi_lock);
883 static void get_pi_state(struct futex_pi_state *pi_state)
885 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
889 * Drops a reference to the pi_state object and frees or caches it
890 * when the last reference is gone.
892 static void put_pi_state(struct futex_pi_state *pi_state)
897 if (!refcount_dec_and_test(&pi_state->refcount))
901 * If pi_state->owner is NULL, the owner is most probably dying
902 * and has cleaned up the pi_state already
904 if (pi_state->owner) {
907 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
908 pi_state_update_owner(pi_state, NULL);
909 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
910 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
913 if (current->pi_state_cache) {
917 * pi_state->list is already empty.
918 * clear pi_state->owner.
919 * refcount is at 0 - put it back to 1.
921 pi_state->owner = NULL;
922 refcount_set(&pi_state->refcount, 1);
923 current->pi_state_cache = pi_state;
927 #ifdef CONFIG_FUTEX_PI
930 * This task is holding PI mutexes at exit time => bad.
931 * Kernel cleans up PI-state, but userspace is likely hosed.
932 * (Robust-futex cleanup is separate and might save the day for userspace.)
934 static void exit_pi_state_list(struct task_struct *curr)
936 struct list_head *next, *head = &curr->pi_state_list;
937 struct futex_pi_state *pi_state;
938 struct futex_hash_bucket *hb;
939 union futex_key key = FUTEX_KEY_INIT;
941 if (!futex_cmpxchg_enabled)
944 * We are a ZOMBIE and nobody can enqueue itself on
945 * pi_state_list anymore, but we have to be careful
946 * versus waiters unqueueing themselves:
948 raw_spin_lock_irq(&curr->pi_lock);
949 while (!list_empty(head)) {
951 pi_state = list_entry(next, struct futex_pi_state, list);
953 hb = hash_futex(&key);
956 * We can race against put_pi_state() removing itself from the
957 * list (a waiter going away). put_pi_state() will first
958 * decrement the reference count and then modify the list, so
959 * its possible to see the list entry but fail this reference
962 * In that case; drop the locks to let put_pi_state() make
963 * progress and retry the loop.
965 if (!refcount_inc_not_zero(&pi_state->refcount)) {
966 raw_spin_unlock_irq(&curr->pi_lock);
968 raw_spin_lock_irq(&curr->pi_lock);
971 raw_spin_unlock_irq(&curr->pi_lock);
973 spin_lock(&hb->lock);
974 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
975 raw_spin_lock(&curr->pi_lock);
977 * We dropped the pi-lock, so re-check whether this
978 * task still owns the PI-state:
980 if (head->next != next) {
981 /* retain curr->pi_lock for the loop invariant */
982 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
983 spin_unlock(&hb->lock);
984 put_pi_state(pi_state);
988 WARN_ON(pi_state->owner != curr);
989 WARN_ON(list_empty(&pi_state->list));
990 list_del_init(&pi_state->list);
991 pi_state->owner = NULL;
993 raw_spin_unlock(&curr->pi_lock);
994 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
995 spin_unlock(&hb->lock);
997 rt_mutex_futex_unlock(&pi_state->pi_mutex);
998 put_pi_state(pi_state);
1000 raw_spin_lock_irq(&curr->pi_lock);
1002 raw_spin_unlock_irq(&curr->pi_lock);
1005 static inline void exit_pi_state_list(struct task_struct *curr) { }
1009 * We need to check the following states:
1011 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
1013 * [1] NULL | --- | --- | 0 | 0/1 | Valid
1014 * [2] NULL | --- | --- | >0 | 0/1 | Valid
1016 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1018 * [4] Found | Found | NULL | 0 | 1 | Valid
1019 * [5] Found | Found | NULL | >0 | 1 | Invalid
1021 * [6] Found | Found | task | 0 | 1 | Valid
1023 * [7] Found | Found | NULL | Any | 0 | Invalid
1025 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1026 * [9] Found | Found | task | 0 | 0 | Invalid
1027 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1029 * [1] Indicates that the kernel can acquire the futex atomically. We
1030 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1032 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1033 * thread is found then it indicates that the owner TID has died.
1035 * [3] Invalid. The waiter is queued on a non PI futex
1037 * [4] Valid state after exit_robust_list(), which sets the user space
1038 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1040 * [5] The user space value got manipulated between exit_robust_list()
1041 * and exit_pi_state_list()
1043 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1044 * the pi_state but cannot access the user space value.
1046 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1048 * [8] Owner and user space value match
1050 * [9] There is no transient state which sets the user space TID to 0
1051 * except exit_robust_list(), but this is indicated by the
1052 * FUTEX_OWNER_DIED bit. See [4]
1054 * [10] There is no transient state which leaves owner and user space
1055 * TID out of sync. Except one error case where the kernel is denied
1056 * write access to the user address, see fixup_pi_state_owner().
1059 * Serialization and lifetime rules:
1063 * hb -> futex_q, relation
1064 * futex_q -> pi_state, relation
1066 * (cannot be raw because hb can contain arbitrary amount
1069 * pi_mutex->wait_lock:
1073 * (and pi_mutex 'obviously')
1077 * p->pi_state_list -> pi_state->list, relation
1079 * pi_state->refcount:
1087 * pi_mutex->wait_lock
1093 * Validate that the existing waiter has a pi_state and sanity check
1094 * the pi_state against the user space value. If correct, attach to
1097 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1098 struct futex_pi_state *pi_state,
1099 struct futex_pi_state **ps)
1101 pid_t pid = uval & FUTEX_TID_MASK;
1106 * Userspace might have messed up non-PI and PI futexes [3]
1108 if (unlikely(!pi_state))
1112 * We get here with hb->lock held, and having found a
1113 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1114 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1115 * which in turn means that futex_lock_pi() still has a reference on
1118 * The waiter holding a reference on @pi_state also protects against
1119 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1120 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1121 * free pi_state before we can take a reference ourselves.
1123 WARN_ON(!refcount_read(&pi_state->refcount));
1126 * Now that we have a pi_state, we can acquire wait_lock
1127 * and do the state validation.
1129 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1132 * Since {uval, pi_state} is serialized by wait_lock, and our current
1133 * uval was read without holding it, it can have changed. Verify it
1134 * still is what we expect it to be, otherwise retry the entire
1137 if (get_futex_value_locked(&uval2, uaddr))
1144 * Handle the owner died case:
1146 if (uval & FUTEX_OWNER_DIED) {
1148 * exit_pi_state_list sets owner to NULL and wakes the
1149 * topmost waiter. The task which acquires the
1150 * pi_state->rt_mutex will fixup owner.
1152 if (!pi_state->owner) {
1154 * No pi state owner, but the user space TID
1155 * is not 0. Inconsistent state. [5]
1160 * Take a ref on the state and return success. [4]
1166 * If TID is 0, then either the dying owner has not
1167 * yet executed exit_pi_state_list() or some waiter
1168 * acquired the rtmutex in the pi state, but did not
1169 * yet fixup the TID in user space.
1171 * Take a ref on the state and return success. [6]
1177 * If the owner died bit is not set, then the pi_state
1178 * must have an owner. [7]
1180 if (!pi_state->owner)
1185 * Bail out if user space manipulated the futex value. If pi
1186 * state exists then the owner TID must be the same as the
1187 * user space TID. [9/10]
1189 if (pid != task_pid_vnr(pi_state->owner))
1193 get_pi_state(pi_state);
1194 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1211 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1216 * wait_for_owner_exiting - Block until the owner has exited
1217 * @exiting: Pointer to the exiting task
1219 * Caller must hold a refcount on @exiting.
1221 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1223 if (ret != -EBUSY) {
1224 WARN_ON_ONCE(exiting);
1228 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1231 mutex_lock(&exiting->futex_exit_mutex);
1233 * No point in doing state checking here. If the waiter got here
1234 * while the task was in exec()->exec_futex_release() then it can
1235 * have any FUTEX_STATE_* value when the waiter has acquired the
1236 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1237 * already. Highly unlikely and not a problem. Just one more round
1238 * through the futex maze.
1240 mutex_unlock(&exiting->futex_exit_mutex);
1242 put_task_struct(exiting);
1245 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1246 struct task_struct *tsk)
1251 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1252 * caller that the alleged owner is busy.
1254 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1258 * Reread the user space value to handle the following situation:
1262 * sys_exit() sys_futex()
1263 * do_exit() futex_lock_pi()
1264 * futex_lock_pi_atomic()
1265 * exit_signals(tsk) No waiters:
1266 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1267 * mm_release(tsk) Set waiter bit
1268 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1269 * Set owner died attach_to_pi_owner() {
1270 * *uaddr = 0xC0000000; tsk = get_task(PID);
1271 * } if (!tsk->flags & PF_EXITING) {
1273 * tsk->futex_state = } else {
1274 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1277 * return -ESRCH; <--- FAIL
1280 * Returning ESRCH unconditionally is wrong here because the
1281 * user space value has been changed by the exiting task.
1283 * The same logic applies to the case where the exiting task is
1286 if (get_futex_value_locked(&uval2, uaddr))
1289 /* If the user space value has changed, try again. */
1294 * The exiting task did not have a robust list, the robust list was
1295 * corrupted or the user space value in *uaddr is simply bogus.
1296 * Give up and tell user space.
1302 * Lookup the task for the TID provided from user space and attach to
1303 * it after doing proper sanity checks.
1305 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1306 struct futex_pi_state **ps,
1307 struct task_struct **exiting)
1309 pid_t pid = uval & FUTEX_TID_MASK;
1310 struct futex_pi_state *pi_state;
1311 struct task_struct *p;
1314 * We are the first waiter - try to look up the real owner and attach
1315 * the new pi_state to it, but bail out when TID = 0 [1]
1317 * The !pid check is paranoid. None of the call sites should end up
1318 * with pid == 0, but better safe than sorry. Let the caller retry
1322 p = find_get_task_by_vpid(pid);
1324 return handle_exit_race(uaddr, uval, NULL);
1326 if (unlikely(p->flags & PF_KTHREAD)) {
1332 * We need to look at the task state to figure out, whether the
1333 * task is exiting. To protect against the change of the task state
1334 * in futex_exit_release(), we do this protected by p->pi_lock:
1336 raw_spin_lock_irq(&p->pi_lock);
1337 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1339 * The task is on the way out. When the futex state is
1340 * FUTEX_STATE_DEAD, we know that the task has finished
1343 int ret = handle_exit_race(uaddr, uval, p);
1345 raw_spin_unlock_irq(&p->pi_lock);
1347 * If the owner task is between FUTEX_STATE_EXITING and
1348 * FUTEX_STATE_DEAD then store the task pointer and keep
1349 * the reference on the task struct. The calling code will
1350 * drop all locks, wait for the task to reach
1351 * FUTEX_STATE_DEAD and then drop the refcount. This is
1352 * required to prevent a live lock when the current task
1353 * preempted the exiting task between the two states.
1363 * No existing pi state. First waiter. [2]
1365 * This creates pi_state, we have hb->lock held, this means nothing can
1366 * observe this state, wait_lock is irrelevant.
1368 pi_state = alloc_pi_state();
1371 * Initialize the pi_mutex in locked state and make @p
1374 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1376 /* Store the key for possible exit cleanups: */
1377 pi_state->key = *key;
1379 WARN_ON(!list_empty(&pi_state->list));
1380 list_add(&pi_state->list, &p->pi_state_list);
1382 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1383 * because there is no concurrency as the object is not published yet.
1385 pi_state->owner = p;
1386 raw_spin_unlock_irq(&p->pi_lock);
1395 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1396 struct futex_hash_bucket *hb,
1397 union futex_key *key, struct futex_pi_state **ps,
1398 struct task_struct **exiting)
1400 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1403 * If there is a waiter on that futex, validate it and
1404 * attach to the pi_state when the validation succeeds.
1407 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1410 * We are the first waiter - try to look up the owner based on
1411 * @uval and attach to it.
1413 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1416 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1419 u32 uninitialized_var(curval);
1421 if (unlikely(should_fail_futex(true)))
1424 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1428 /* If user space value changed, let the caller retry */
1429 return curval != uval ? -EAGAIN : 0;
1433 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1434 * @uaddr: the pi futex user address
1435 * @hb: the pi futex hash bucket
1436 * @key: the futex key associated with uaddr and hb
1437 * @ps: the pi_state pointer where we store the result of the
1439 * @task: the task to perform the atomic lock work for. This will
1440 * be "current" except in the case of requeue pi.
1441 * @exiting: Pointer to store the task pointer of the owner task
1442 * which is in the middle of exiting
1443 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1446 * - 0 - ready to wait;
1447 * - 1 - acquired the lock;
1450 * The hb->lock and futex_key refs shall be held by the caller.
1452 * @exiting is only set when the return value is -EBUSY. If so, this holds
1453 * a refcount on the exiting task on return and the caller needs to drop it
1454 * after waiting for the exit to complete.
1456 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1457 union futex_key *key,
1458 struct futex_pi_state **ps,
1459 struct task_struct *task,
1460 struct task_struct **exiting,
1463 u32 uval, newval, vpid = task_pid_vnr(task);
1464 struct futex_q *top_waiter;
1468 * Read the user space value first so we can validate a few
1469 * things before proceeding further.
1471 if (get_futex_value_locked(&uval, uaddr))
1474 if (unlikely(should_fail_futex(true)))
1480 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1483 if ((unlikely(should_fail_futex(true))))
1487 * Lookup existing state first. If it exists, try to attach to
1490 top_waiter = futex_top_waiter(hb, key);
1492 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1495 * No waiter and user TID is 0. We are here because the
1496 * waiters or the owner died bit is set or called from
1497 * requeue_cmp_pi or for whatever reason something took the
1500 if (!(uval & FUTEX_TID_MASK)) {
1502 * We take over the futex. No other waiters and the user space
1503 * TID is 0. We preserve the owner died bit.
1505 newval = uval & FUTEX_OWNER_DIED;
1508 /* The futex requeue_pi code can enforce the waiters bit */
1510 newval |= FUTEX_WAITERS;
1512 ret = lock_pi_update_atomic(uaddr, uval, newval);
1513 /* If the take over worked, return 1 */
1514 return ret < 0 ? ret : 1;
1518 * First waiter. Set the waiters bit before attaching ourself to
1519 * the owner. If owner tries to unlock, it will be forced into
1520 * the kernel and blocked on hb->lock.
1522 newval = uval | FUTEX_WAITERS;
1523 ret = lock_pi_update_atomic(uaddr, uval, newval);
1527 * If the update of the user space value succeeded, we try to
1528 * attach to the owner. If that fails, no harm done, we only
1529 * set the FUTEX_WAITERS bit in the user space variable.
1531 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1535 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1536 * @q: The futex_q to unqueue
1538 * The q->lock_ptr must not be NULL and must be held by the caller.
1540 static void __unqueue_futex(struct futex_q *q)
1542 struct futex_hash_bucket *hb;
1544 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1546 lockdep_assert_held(q->lock_ptr);
1548 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1549 plist_del(&q->list, &hb->chain);
1554 * The hash bucket lock must be held when this is called.
1555 * Afterwards, the futex_q must not be accessed. Callers
1556 * must ensure to later call wake_up_q() for the actual
1559 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1561 struct task_struct *p = q->task;
1563 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1569 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1570 * is written, without taking any locks. This is possible in the event
1571 * of a spurious wakeup, for example. A memory barrier is required here
1572 * to prevent the following store to lock_ptr from getting ahead of the
1573 * plist_del in __unqueue_futex().
1575 smp_store_release(&q->lock_ptr, NULL);
1578 * Queue the task for later wakeup for after we've released
1579 * the hb->lock. wake_q_add() grabs reference to p.
1581 wake_q_add_safe(wake_q, p);
1585 * Caller must hold a reference on @pi_state.
1587 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1589 u32 uninitialized_var(curval), newval;
1590 struct task_struct *new_owner;
1591 bool postunlock = false;
1592 DEFINE_WAKE_Q(wake_q);
1595 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1596 if (WARN_ON_ONCE(!new_owner)) {
1598 * As per the comment in futex_unlock_pi() this should not happen.
1600 * When this happens, give up our locks and try again, giving
1601 * the futex_lock_pi() instance time to complete, either by
1602 * waiting on the rtmutex or removing itself from the futex
1610 * We pass it to the next owner. The WAITERS bit is always kept
1611 * enabled while there is PI state around. We cleanup the owner
1612 * died bit, because we are the owner.
1614 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1616 if (unlikely(should_fail_futex(true))) {
1621 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1622 if (!ret && (curval != uval)) {
1624 * If a unconditional UNLOCK_PI operation (user space did not
1625 * try the TID->0 transition) raced with a waiter setting the
1626 * FUTEX_WAITERS flag between get_user() and locking the hash
1627 * bucket lock, retry the operation.
1629 if ((FUTEX_TID_MASK & curval) == uval)
1637 * This is a point of no return; once we modified the uval
1638 * there is no going back and subsequent operations must
1641 pi_state_update_owner(pi_state, new_owner);
1642 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1646 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1649 rt_mutex_postunlock(&wake_q);
1655 * Express the locking dependencies for lockdep:
1658 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1661 spin_lock(&hb1->lock);
1663 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1664 } else { /* hb1 > hb2 */
1665 spin_lock(&hb2->lock);
1666 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1671 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1673 spin_unlock(&hb1->lock);
1675 spin_unlock(&hb2->lock);
1679 * Wake up waiters matching bitset queued on this futex (uaddr).
1682 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1684 struct futex_hash_bucket *hb;
1685 struct futex_q *this, *next;
1686 union futex_key key = FUTEX_KEY_INIT;
1688 DEFINE_WAKE_Q(wake_q);
1693 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1694 if (unlikely(ret != 0))
1697 hb = hash_futex(&key);
1699 /* Make sure we really have tasks to wakeup */
1700 if (!hb_waiters_pending(hb))
1703 spin_lock(&hb->lock);
1705 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1706 if (match_futex (&this->key, &key)) {
1707 if (this->pi_state || this->rt_waiter) {
1712 /* Check if one of the bits is set in both bitsets */
1713 if (!(this->bitset & bitset))
1716 mark_wake_futex(&wake_q, this);
1717 if (++ret >= nr_wake)
1722 spin_unlock(&hb->lock);
1725 put_futex_key(&key);
1730 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1732 unsigned int op = (encoded_op & 0x70000000) >> 28;
1733 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1734 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1735 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1738 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1739 if (oparg < 0 || oparg > 31) {
1740 char comm[sizeof(current->comm)];
1742 * kill this print and return -EINVAL when userspace
1745 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1746 get_task_comm(comm, current), oparg);
1752 if (!access_ok(uaddr, sizeof(u32)))
1755 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1760 case FUTEX_OP_CMP_EQ:
1761 return oldval == cmparg;
1762 case FUTEX_OP_CMP_NE:
1763 return oldval != cmparg;
1764 case FUTEX_OP_CMP_LT:
1765 return oldval < cmparg;
1766 case FUTEX_OP_CMP_GE:
1767 return oldval >= cmparg;
1768 case FUTEX_OP_CMP_LE:
1769 return oldval <= cmparg;
1770 case FUTEX_OP_CMP_GT:
1771 return oldval > cmparg;
1778 * Wake up all waiters hashed on the physical page that is mapped
1779 * to this virtual address:
1782 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1783 int nr_wake, int nr_wake2, int op)
1785 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1786 struct futex_hash_bucket *hb1, *hb2;
1787 struct futex_q *this, *next;
1789 DEFINE_WAKE_Q(wake_q);
1792 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1793 if (unlikely(ret != 0))
1795 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1796 if (unlikely(ret != 0))
1799 hb1 = hash_futex(&key1);
1800 hb2 = hash_futex(&key2);
1803 double_lock_hb(hb1, hb2);
1804 op_ret = futex_atomic_op_inuser(op, uaddr2);
1805 if (unlikely(op_ret < 0)) {
1806 double_unlock_hb(hb1, hb2);
1808 if (!IS_ENABLED(CONFIG_MMU) ||
1809 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1811 * we don't get EFAULT from MMU faults if we don't have
1812 * an MMU, but we might get them from range checking
1818 if (op_ret == -EFAULT) {
1819 ret = fault_in_user_writeable(uaddr2);
1824 if (!(flags & FLAGS_SHARED)) {
1829 put_futex_key(&key2);
1830 put_futex_key(&key1);
1835 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1836 if (match_futex (&this->key, &key1)) {
1837 if (this->pi_state || this->rt_waiter) {
1841 mark_wake_futex(&wake_q, this);
1842 if (++ret >= nr_wake)
1849 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1850 if (match_futex (&this->key, &key2)) {
1851 if (this->pi_state || this->rt_waiter) {
1855 mark_wake_futex(&wake_q, this);
1856 if (++op_ret >= nr_wake2)
1864 double_unlock_hb(hb1, hb2);
1867 put_futex_key(&key2);
1869 put_futex_key(&key1);
1875 * requeue_futex() - Requeue a futex_q from one hb to another
1876 * @q: the futex_q to requeue
1877 * @hb1: the source hash_bucket
1878 * @hb2: the target hash_bucket
1879 * @key2: the new key for the requeued futex_q
1882 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1883 struct futex_hash_bucket *hb2, union futex_key *key2)
1887 * If key1 and key2 hash to the same bucket, no need to
1890 if (likely(&hb1->chain != &hb2->chain)) {
1891 plist_del(&q->list, &hb1->chain);
1892 hb_waiters_dec(hb1);
1893 hb_waiters_inc(hb2);
1894 plist_add(&q->list, &hb2->chain);
1895 q->lock_ptr = &hb2->lock;
1897 get_futex_key_refs(key2);
1902 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1904 * @key: the key of the requeue target futex
1905 * @hb: the hash_bucket of the requeue target futex
1907 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1908 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1909 * to the requeue target futex so the waiter can detect the wakeup on the right
1910 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1911 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1912 * to protect access to the pi_state to fixup the owner later. Must be called
1913 * with both q->lock_ptr and hb->lock held.
1916 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1917 struct futex_hash_bucket *hb)
1919 get_futex_key_refs(key);
1924 WARN_ON(!q->rt_waiter);
1925 q->rt_waiter = NULL;
1927 q->lock_ptr = &hb->lock;
1929 wake_up_state(q->task, TASK_NORMAL);
1933 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1934 * @pifutex: the user address of the to futex
1935 * @hb1: the from futex hash bucket, must be locked by the caller
1936 * @hb2: the to futex hash bucket, must be locked by the caller
1937 * @key1: the from futex key
1938 * @key2: the to futex key
1939 * @ps: address to store the pi_state pointer
1940 * @exiting: Pointer to store the task pointer of the owner task
1941 * which is in the middle of exiting
1942 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1944 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1945 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1946 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1947 * hb1 and hb2 must be held by the caller.
1949 * @exiting is only set when the return value is -EBUSY. If so, this holds
1950 * a refcount on the exiting task on return and the caller needs to drop it
1951 * after waiting for the exit to complete.
1954 * - 0 - failed to acquire the lock atomically;
1955 * - >0 - acquired the lock, return value is vpid of the top_waiter
1959 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1960 struct futex_hash_bucket *hb2, union futex_key *key1,
1961 union futex_key *key2, struct futex_pi_state **ps,
1962 struct task_struct **exiting, int set_waiters)
1964 struct futex_q *top_waiter = NULL;
1968 if (get_futex_value_locked(&curval, pifutex))
1971 if (unlikely(should_fail_futex(true)))
1975 * Find the top_waiter and determine if there are additional waiters.
1976 * If the caller intends to requeue more than 1 waiter to pifutex,
1977 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1978 * as we have means to handle the possible fault. If not, don't set
1979 * the bit unecessarily as it will force the subsequent unlock to enter
1982 top_waiter = futex_top_waiter(hb1, key1);
1984 /* There are no waiters, nothing for us to do. */
1988 /* Ensure we requeue to the expected futex. */
1989 if (!match_futex(top_waiter->requeue_pi_key, key2))
1993 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1994 * the contended case or if set_waiters is 1. The pi_state is returned
1995 * in ps in contended cases.
1997 vpid = task_pid_vnr(top_waiter->task);
1998 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1999 exiting, set_waiters);
2001 requeue_pi_wake_futex(top_waiter, key2, hb2);
2008 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
2009 * @uaddr1: source futex user address
2010 * @flags: futex flags (FLAGS_SHARED, etc.)
2011 * @uaddr2: target futex user address
2012 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
2013 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
2014 * @cmpval: @uaddr1 expected value (or %NULL)
2015 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
2016 * pi futex (pi to pi requeue is not supported)
2018 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
2019 * uaddr2 atomically on behalf of the top waiter.
2022 * - >=0 - on success, the number of tasks requeued or woken;
2025 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
2026 u32 __user *uaddr2, int nr_wake, int nr_requeue,
2027 u32 *cmpval, int requeue_pi)
2029 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
2030 int drop_count = 0, task_count = 0, ret;
2031 struct futex_pi_state *pi_state = NULL;
2032 struct futex_hash_bucket *hb1, *hb2;
2033 struct futex_q *this, *next;
2034 DEFINE_WAKE_Q(wake_q);
2036 if (nr_wake < 0 || nr_requeue < 0)
2040 * When PI not supported: return -ENOSYS if requeue_pi is true,
2041 * consequently the compiler knows requeue_pi is always false past
2042 * this point which will optimize away all the conditional code
2045 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
2050 * Requeue PI only works on two distinct uaddrs. This
2051 * check is only valid for private futexes. See below.
2053 if (uaddr1 == uaddr2)
2057 * requeue_pi requires a pi_state, try to allocate it now
2058 * without any locks in case it fails.
2060 if (refill_pi_state_cache())
2063 * requeue_pi must wake as many tasks as it can, up to nr_wake
2064 * + nr_requeue, since it acquires the rt_mutex prior to
2065 * returning to userspace, so as to not leave the rt_mutex with
2066 * waiters and no owner. However, second and third wake-ups
2067 * cannot be predicted as they involve race conditions with the
2068 * first wake and a fault while looking up the pi_state. Both
2069 * pthread_cond_signal() and pthread_cond_broadcast() should
2077 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
2078 if (unlikely(ret != 0))
2080 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2081 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
2082 if (unlikely(ret != 0))
2086 * The check above which compares uaddrs is not sufficient for
2087 * shared futexes. We need to compare the keys:
2089 if (requeue_pi && match_futex(&key1, &key2)) {
2094 hb1 = hash_futex(&key1);
2095 hb2 = hash_futex(&key2);
2098 hb_waiters_inc(hb2);
2099 double_lock_hb(hb1, hb2);
2101 if (likely(cmpval != NULL)) {
2104 ret = get_futex_value_locked(&curval, uaddr1);
2106 if (unlikely(ret)) {
2107 double_unlock_hb(hb1, hb2);
2108 hb_waiters_dec(hb2);
2110 ret = get_user(curval, uaddr1);
2114 if (!(flags & FLAGS_SHARED))
2117 put_futex_key(&key2);
2118 put_futex_key(&key1);
2121 if (curval != *cmpval) {
2127 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2128 struct task_struct *exiting = NULL;
2131 * Attempt to acquire uaddr2 and wake the top waiter. If we
2132 * intend to requeue waiters, force setting the FUTEX_WAITERS
2133 * bit. We force this here where we are able to easily handle
2134 * faults rather in the requeue loop below.
2136 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2138 &exiting, nr_requeue);
2141 * At this point the top_waiter has either taken uaddr2 or is
2142 * waiting on it. If the former, then the pi_state will not
2143 * exist yet, look it up one more time to ensure we have a
2144 * reference to it. If the lock was taken, ret contains the
2145 * vpid of the top waiter task.
2146 * If the lock was not taken, we have pi_state and an initial
2147 * refcount on it. In case of an error we have nothing.
2154 * If we acquired the lock, then the user space value
2155 * of uaddr2 should be vpid. It cannot be changed by
2156 * the top waiter as it is blocked on hb2 lock if it
2157 * tries to do so. If something fiddled with it behind
2158 * our back the pi state lookup might unearth it. So
2159 * we rather use the known value than rereading and
2160 * handing potential crap to lookup_pi_state.
2162 * If that call succeeds then we have pi_state and an
2163 * initial refcount on it.
2165 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2166 &pi_state, &exiting);
2171 /* We hold a reference on the pi state. */
2174 /* If the above failed, then pi_state is NULL */
2176 double_unlock_hb(hb1, hb2);
2177 hb_waiters_dec(hb2);
2178 put_futex_key(&key2);
2179 put_futex_key(&key1);
2180 ret = fault_in_user_writeable(uaddr2);
2187 * Two reasons for this:
2188 * - EBUSY: Owner is exiting and we just wait for the
2190 * - EAGAIN: The user space value changed.
2192 double_unlock_hb(hb1, hb2);
2193 hb_waiters_dec(hb2);
2194 put_futex_key(&key2);
2195 put_futex_key(&key1);
2197 * Handle the case where the owner is in the middle of
2198 * exiting. Wait for the exit to complete otherwise
2199 * this task might loop forever, aka. live lock.
2201 wait_for_owner_exiting(ret, exiting);
2209 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2210 if (task_count - nr_wake >= nr_requeue)
2213 if (!match_futex(&this->key, &key1))
2217 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2218 * be paired with each other and no other futex ops.
2220 * We should never be requeueing a futex_q with a pi_state,
2221 * which is awaiting a futex_unlock_pi().
2223 if ((requeue_pi && !this->rt_waiter) ||
2224 (!requeue_pi && this->rt_waiter) ||
2231 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2232 * lock, we already woke the top_waiter. If not, it will be
2233 * woken by futex_unlock_pi().
2235 if (++task_count <= nr_wake && !requeue_pi) {
2236 mark_wake_futex(&wake_q, this);
2240 /* Ensure we requeue to the expected futex for requeue_pi. */
2241 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2247 * Requeue nr_requeue waiters and possibly one more in the case
2248 * of requeue_pi if we couldn't acquire the lock atomically.
2252 * Prepare the waiter to take the rt_mutex. Take a
2253 * refcount on the pi_state and store the pointer in
2254 * the futex_q object of the waiter.
2256 get_pi_state(pi_state);
2257 this->pi_state = pi_state;
2258 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2263 * We got the lock. We do neither drop the
2264 * refcount on pi_state nor clear
2265 * this->pi_state because the waiter needs the
2266 * pi_state for cleaning up the user space
2267 * value. It will drop the refcount after
2270 requeue_pi_wake_futex(this, &key2, hb2);
2275 * rt_mutex_start_proxy_lock() detected a
2276 * potential deadlock when we tried to queue
2277 * that waiter. Drop the pi_state reference
2278 * which we took above and remove the pointer
2279 * to the state from the waiters futex_q
2282 this->pi_state = NULL;
2283 put_pi_state(pi_state);
2285 * We stop queueing more waiters and let user
2286 * space deal with the mess.
2291 requeue_futex(this, hb1, hb2, &key2);
2296 * We took an extra initial reference to the pi_state either
2297 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2298 * need to drop it here again.
2300 put_pi_state(pi_state);
2303 double_unlock_hb(hb1, hb2);
2305 hb_waiters_dec(hb2);
2308 * drop_futex_key_refs() must be called outside the spinlocks. During
2309 * the requeue we moved futex_q's from the hash bucket at key1 to the
2310 * one at key2 and updated their key pointer. We no longer need to
2311 * hold the references to key1.
2313 while (--drop_count >= 0)
2314 drop_futex_key_refs(&key1);
2317 put_futex_key(&key2);
2319 put_futex_key(&key1);
2321 return ret ? ret : task_count;
2324 /* The key must be already stored in q->key. */
2325 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2326 __acquires(&hb->lock)
2328 struct futex_hash_bucket *hb;
2330 hb = hash_futex(&q->key);
2333 * Increment the counter before taking the lock so that
2334 * a potential waker won't miss a to-be-slept task that is
2335 * waiting for the spinlock. This is safe as all queue_lock()
2336 * users end up calling queue_me(). Similarly, for housekeeping,
2337 * decrement the counter at queue_unlock() when some error has
2338 * occurred and we don't end up adding the task to the list.
2340 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2342 q->lock_ptr = &hb->lock;
2344 spin_lock(&hb->lock);
2349 queue_unlock(struct futex_hash_bucket *hb)
2350 __releases(&hb->lock)
2352 spin_unlock(&hb->lock);
2356 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2361 * The priority used to register this element is
2362 * - either the real thread-priority for the real-time threads
2363 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2364 * - or MAX_RT_PRIO for non-RT threads.
2365 * Thus, all RT-threads are woken first in priority order, and
2366 * the others are woken last, in FIFO order.
2368 prio = min(current->normal_prio, MAX_RT_PRIO);
2370 plist_node_init(&q->list, prio);
2371 plist_add(&q->list, &hb->chain);
2376 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2377 * @q: The futex_q to enqueue
2378 * @hb: The destination hash bucket
2380 * The hb->lock must be held by the caller, and is released here. A call to
2381 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2382 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2383 * or nothing if the unqueue is done as part of the wake process and the unqueue
2384 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2387 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2388 __releases(&hb->lock)
2391 spin_unlock(&hb->lock);
2395 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2396 * @q: The futex_q to unqueue
2398 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2399 * be paired with exactly one earlier call to queue_me().
2402 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2403 * - 0 - if the futex_q was already removed by the waking thread
2405 static int unqueue_me(struct futex_q *q)
2407 spinlock_t *lock_ptr;
2410 /* In the common case we don't take the spinlock, which is nice. */
2413 * q->lock_ptr can change between this read and the following spin_lock.
2414 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2415 * optimizing lock_ptr out of the logic below.
2417 lock_ptr = READ_ONCE(q->lock_ptr);
2418 if (lock_ptr != NULL) {
2419 spin_lock(lock_ptr);
2421 * q->lock_ptr can change between reading it and
2422 * spin_lock(), causing us to take the wrong lock. This
2423 * corrects the race condition.
2425 * Reasoning goes like this: if we have the wrong lock,
2426 * q->lock_ptr must have changed (maybe several times)
2427 * between reading it and the spin_lock(). It can
2428 * change again after the spin_lock() but only if it was
2429 * already changed before the spin_lock(). It cannot,
2430 * however, change back to the original value. Therefore
2431 * we can detect whether we acquired the correct lock.
2433 if (unlikely(lock_ptr != q->lock_ptr)) {
2434 spin_unlock(lock_ptr);
2439 BUG_ON(q->pi_state);
2441 spin_unlock(lock_ptr);
2445 drop_futex_key_refs(&q->key);
2450 * PI futexes can not be requeued and must remove themself from the
2451 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2454 static void unqueue_me_pi(struct futex_q *q)
2455 __releases(q->lock_ptr)
2459 BUG_ON(!q->pi_state);
2460 put_pi_state(q->pi_state);
2463 spin_unlock(q->lock_ptr);
2466 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2467 struct task_struct *argowner)
2469 u32 uval, uninitialized_var(curval), newval, newtid;
2470 struct futex_pi_state *pi_state = q->pi_state;
2471 struct task_struct *oldowner, *newowner;
2474 oldowner = pi_state->owner;
2477 * We are here because either:
2479 * - we stole the lock and pi_state->owner needs updating to reflect
2480 * that (@argowner == current),
2484 * - someone stole our lock and we need to fix things to point to the
2485 * new owner (@argowner == NULL).
2487 * Either way, we have to replace the TID in the user space variable.
2488 * This must be atomic as we have to preserve the owner died bit here.
2490 * Note: We write the user space value _before_ changing the pi_state
2491 * because we can fault here. Imagine swapped out pages or a fork
2492 * that marked all the anonymous memory readonly for cow.
2494 * Modifying pi_state _before_ the user space value would leave the
2495 * pi_state in an inconsistent state when we fault here, because we
2496 * need to drop the locks to handle the fault. This might be observed
2497 * in the PID check in lookup_pi_state.
2501 if (oldowner != current) {
2503 * We raced against a concurrent self; things are
2504 * already fixed up. Nothing to do.
2509 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2510 /* We got the lock. pi_state is correct. Tell caller. */
2515 * The trylock just failed, so either there is an owner or
2516 * there is a higher priority waiter than this one.
2518 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2520 * If the higher priority waiter has not yet taken over the
2521 * rtmutex then newowner is NULL. We can't return here with
2522 * that state because it's inconsistent vs. the user space
2523 * state. So drop the locks and try again. It's a valid
2524 * situation and not any different from the other retry
2527 if (unlikely(!newowner)) {
2532 WARN_ON_ONCE(argowner != current);
2533 if (oldowner == current) {
2535 * We raced against a concurrent self; things are
2536 * already fixed up. Nothing to do.
2540 newowner = argowner;
2543 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2545 if (!pi_state->owner)
2546 newtid |= FUTEX_OWNER_DIED;
2548 err = get_futex_value_locked(&uval, uaddr);
2553 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2555 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2565 * We fixed up user space. Now we need to fix the pi_state
2568 pi_state_update_owner(pi_state, newowner);
2570 return argowner == current;
2573 * In order to reschedule or handle a page fault, we need to drop the
2574 * locks here. In the case of a fault, this gives the other task
2575 * (either the highest priority waiter itself or the task which stole
2576 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2577 * are back from handling the fault we need to check the pi_state after
2578 * reacquiring the locks and before trying to do another fixup. When
2579 * the fixup has been done already we simply return.
2581 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2582 * drop hb->lock since the caller owns the hb -> futex_q relation.
2583 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2586 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2587 spin_unlock(q->lock_ptr);
2591 err = fault_in_user_writeable(uaddr);
2604 spin_lock(q->lock_ptr);
2605 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2608 * Check if someone else fixed it for us:
2610 if (pi_state->owner != oldowner)
2611 return argowner == current;
2613 /* Retry if err was -EAGAIN or the fault in succeeded */
2618 * fault_in_user_writeable() failed so user state is immutable. At
2619 * best we can make the kernel state consistent but user state will
2620 * be most likely hosed and any subsequent unlock operation will be
2621 * rejected due to PI futex rule [10].
2623 * Ensure that the rtmutex owner is also the pi_state owner despite
2624 * the user space value claiming something different. There is no
2625 * point in unlocking the rtmutex if current is the owner as it
2626 * would need to wait until the next waiter has taken the rtmutex
2627 * to guarantee consistent state. Keep it simple. Userspace asked
2628 * for this wreckaged state.
2630 * The rtmutex has an owner - either current or some other
2631 * task. See the EAGAIN loop above.
2633 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2638 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2639 struct task_struct *argowner)
2641 struct futex_pi_state *pi_state = q->pi_state;
2644 lockdep_assert_held(q->lock_ptr);
2646 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2647 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2648 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2652 static long futex_wait_restart(struct restart_block *restart);
2655 * fixup_owner() - Post lock pi_state and corner case management
2656 * @uaddr: user address of the futex
2657 * @q: futex_q (contains pi_state and access to the rt_mutex)
2658 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2660 * After attempting to lock an rt_mutex, this function is called to cleanup
2661 * the pi_state owner as well as handle race conditions that may allow us to
2662 * acquire the lock. Must be called with the hb lock held.
2665 * - 1 - success, lock taken;
2666 * - 0 - success, lock not taken;
2667 * - <0 - on error (-EFAULT)
2669 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2673 * Got the lock. We might not be the anticipated owner if we
2674 * did a lock-steal - fix up the PI-state in that case:
2676 * Speculative pi_state->owner read (we don't hold wait_lock);
2677 * since we own the lock pi_state->owner == current is the
2678 * stable state, anything else needs more attention.
2680 if (q->pi_state->owner != current)
2681 return fixup_pi_state_owner(uaddr, q, current);
2686 * If we didn't get the lock; check if anybody stole it from us. In
2687 * that case, we need to fix up the uval to point to them instead of
2688 * us, otherwise bad things happen. [10]
2690 * Another speculative read; pi_state->owner == current is unstable
2691 * but needs our attention.
2693 if (q->pi_state->owner == current)
2694 return fixup_pi_state_owner(uaddr, q, NULL);
2697 * Paranoia check. If we did not take the lock, then we should not be
2698 * the owner of the rt_mutex. Warn and establish consistent state.
2700 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2701 return fixup_pi_state_owner(uaddr, q, current);
2707 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2708 * @hb: the futex hash bucket, must be locked by the caller
2709 * @q: the futex_q to queue up on
2710 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2712 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2713 struct hrtimer_sleeper *timeout)
2716 * The task state is guaranteed to be set before another task can
2717 * wake it. set_current_state() is implemented using smp_store_mb() and
2718 * queue_me() calls spin_unlock() upon completion, both serializing
2719 * access to the hash list and forcing another memory barrier.
2721 set_current_state(TASK_INTERRUPTIBLE);
2726 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2729 * If we have been removed from the hash list, then another task
2730 * has tried to wake us, and we can skip the call to schedule().
2732 if (likely(!plist_node_empty(&q->list))) {
2734 * If the timer has already expired, current will already be
2735 * flagged for rescheduling. Only call schedule if there
2736 * is no timeout, or if it has yet to expire.
2738 if (!timeout || timeout->task)
2739 freezable_schedule();
2741 __set_current_state(TASK_RUNNING);
2745 * futex_wait_setup() - Prepare to wait on a futex
2746 * @uaddr: the futex userspace address
2747 * @val: the expected value
2748 * @flags: futex flags (FLAGS_SHARED, etc.)
2749 * @q: the associated futex_q
2750 * @hb: storage for hash_bucket pointer to be returned to caller
2752 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2753 * compare it with the expected value. Handle atomic faults internally.
2754 * Return with the hb lock held and a q.key reference on success, and unlocked
2755 * with no q.key reference on failure.
2758 * - 0 - uaddr contains val and hb has been locked;
2759 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2761 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2762 struct futex_q *q, struct futex_hash_bucket **hb)
2768 * Access the page AFTER the hash-bucket is locked.
2769 * Order is important:
2771 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2772 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2774 * The basic logical guarantee of a futex is that it blocks ONLY
2775 * if cond(var) is known to be true at the time of blocking, for
2776 * any cond. If we locked the hash-bucket after testing *uaddr, that
2777 * would open a race condition where we could block indefinitely with
2778 * cond(var) false, which would violate the guarantee.
2780 * On the other hand, we insert q and release the hash-bucket only
2781 * after testing *uaddr. This guarantees that futex_wait() will NOT
2782 * absorb a wakeup if *uaddr does not match the desired values
2783 * while the syscall executes.
2786 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2787 if (unlikely(ret != 0))
2791 *hb = queue_lock(q);
2793 ret = get_futex_value_locked(&uval, uaddr);
2798 ret = get_user(uval, uaddr);
2802 if (!(flags & FLAGS_SHARED))
2805 put_futex_key(&q->key);
2816 put_futex_key(&q->key);
2820 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2821 ktime_t *abs_time, u32 bitset)
2823 struct hrtimer_sleeper timeout, *to;
2824 struct restart_block *restart;
2825 struct futex_hash_bucket *hb;
2826 struct futex_q q = futex_q_init;
2833 to = futex_setup_timer(abs_time, &timeout, flags,
2834 current->timer_slack_ns);
2837 * Prepare to wait on uaddr. On success, holds hb lock and increments
2840 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2844 /* queue_me and wait for wakeup, timeout, or a signal. */
2845 futex_wait_queue_me(hb, &q, to);
2847 /* If we were woken (and unqueued), we succeeded, whatever. */
2849 /* unqueue_me() drops q.key ref */
2850 if (!unqueue_me(&q))
2853 if (to && !to->task)
2857 * We expect signal_pending(current), but we might be the
2858 * victim of a spurious wakeup as well.
2860 if (!signal_pending(current))
2867 restart = ¤t->restart_block;
2868 restart->futex.uaddr = uaddr;
2869 restart->futex.val = val;
2870 restart->futex.time = *abs_time;
2871 restart->futex.bitset = bitset;
2872 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2874 ret = set_restart_fn(restart, futex_wait_restart);
2878 hrtimer_cancel(&to->timer);
2879 destroy_hrtimer_on_stack(&to->timer);
2885 static long futex_wait_restart(struct restart_block *restart)
2887 u32 __user *uaddr = restart->futex.uaddr;
2888 ktime_t t, *tp = NULL;
2890 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2891 t = restart->futex.time;
2894 restart->fn = do_no_restart_syscall;
2896 return (long)futex_wait(uaddr, restart->futex.flags,
2897 restart->futex.val, tp, restart->futex.bitset);
2902 * Userspace tried a 0 -> TID atomic transition of the futex value
2903 * and failed. The kernel side here does the whole locking operation:
2904 * if there are waiters then it will block as a consequence of relying
2905 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2906 * a 0 value of the futex too.).
2908 * Also serves as futex trylock_pi()'ing, and due semantics.
2910 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2911 ktime_t *time, int trylock)
2913 struct hrtimer_sleeper timeout, *to;
2914 struct task_struct *exiting = NULL;
2915 struct rt_mutex_waiter rt_waiter;
2916 struct futex_hash_bucket *hb;
2917 struct futex_q q = futex_q_init;
2920 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2923 if (refill_pi_state_cache())
2926 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2929 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2930 if (unlikely(ret != 0))
2934 hb = queue_lock(&q);
2936 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2938 if (unlikely(ret)) {
2940 * Atomic work succeeded and we got the lock,
2941 * or failed. Either way, we do _not_ block.
2945 /* We got the lock. */
2947 goto out_unlock_put_key;
2953 * Two reasons for this:
2954 * - EBUSY: Task is exiting and we just wait for the
2956 * - EAGAIN: The user space value changed.
2959 put_futex_key(&q.key);
2961 * Handle the case where the owner is in the middle of
2962 * exiting. Wait for the exit to complete otherwise
2963 * this task might loop forever, aka. live lock.
2965 wait_for_owner_exiting(ret, exiting);
2969 goto out_unlock_put_key;
2973 WARN_ON(!q.pi_state);
2976 * Only actually queue now that the atomic ops are done:
2981 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2982 /* Fixup the trylock return value: */
2983 ret = ret ? 0 : -EWOULDBLOCK;
2987 rt_mutex_init_waiter(&rt_waiter);
2990 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2991 * hold it while doing rt_mutex_start_proxy(), because then it will
2992 * include hb->lock in the blocking chain, even through we'll not in
2993 * fact hold it while blocking. This will lead it to report -EDEADLK
2994 * and BUG when futex_unlock_pi() interleaves with this.
2996 * Therefore acquire wait_lock while holding hb->lock, but drop the
2997 * latter before calling __rt_mutex_start_proxy_lock(). This
2998 * interleaves with futex_unlock_pi() -- which does a similar lock
2999 * handoff -- such that the latter can observe the futex_q::pi_state
3000 * before __rt_mutex_start_proxy_lock() is done.
3002 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
3003 spin_unlock(q.lock_ptr);
3005 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
3006 * such that futex_unlock_pi() is guaranteed to observe the waiter when
3007 * it sees the futex_q::pi_state.
3009 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
3010 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3019 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
3021 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3024 spin_lock(q.lock_ptr);
3026 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3027 * first acquire the hb->lock before removing the lock from the
3028 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3031 * In particular; it is important that futex_unlock_pi() can not
3032 * observe this inconsistency.
3034 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3039 * Fixup the pi_state owner and possibly acquire the lock if we
3042 res = fixup_owner(uaddr, &q, !ret);
3044 * If fixup_owner() returned an error, proprogate that. If it acquired
3045 * the lock, clear our -ETIMEDOUT or -EINTR.
3048 ret = (res < 0) ? res : 0;
3050 /* Unqueue and drop the lock */
3059 put_futex_key(&q.key);
3062 hrtimer_cancel(&to->timer);
3063 destroy_hrtimer_on_stack(&to->timer);
3065 return ret != -EINTR ? ret : -ERESTARTNOINTR;
3070 ret = fault_in_user_writeable(uaddr);
3074 if (!(flags & FLAGS_SHARED))
3077 put_futex_key(&q.key);
3082 * Userspace attempted a TID -> 0 atomic transition, and failed.
3083 * This is the in-kernel slowpath: we look up the PI state (if any),
3084 * and do the rt-mutex unlock.
3086 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3088 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
3089 union futex_key key = FUTEX_KEY_INIT;
3090 struct futex_hash_bucket *hb;
3091 struct futex_q *top_waiter;
3094 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3098 if (get_user(uval, uaddr))
3101 * We release only a lock we actually own:
3103 if ((uval & FUTEX_TID_MASK) != vpid)
3106 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3110 hb = hash_futex(&key);
3111 spin_lock(&hb->lock);
3114 * Check waiters first. We do not trust user space values at
3115 * all and we at least want to know if user space fiddled
3116 * with the futex value instead of blindly unlocking.
3118 top_waiter = futex_top_waiter(hb, &key);
3120 struct futex_pi_state *pi_state = top_waiter->pi_state;
3127 * If current does not own the pi_state then the futex is
3128 * inconsistent and user space fiddled with the futex value.
3130 if (pi_state->owner != current)
3133 get_pi_state(pi_state);
3135 * By taking wait_lock while still holding hb->lock, we ensure
3136 * there is no point where we hold neither; and therefore
3137 * wake_futex_pi() must observe a state consistent with what we
3140 * In particular; this forces __rt_mutex_start_proxy() to
3141 * complete such that we're guaranteed to observe the
3142 * rt_waiter. Also see the WARN in wake_futex_pi().
3144 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3145 spin_unlock(&hb->lock);
3147 /* drops pi_state->pi_mutex.wait_lock */
3148 ret = wake_futex_pi(uaddr, uval, pi_state);
3150 put_pi_state(pi_state);
3153 * Success, we're done! No tricky corner cases.
3158 * The atomic access to the futex value generated a
3159 * pagefault, so retry the user-access and the wakeup:
3164 * A unconditional UNLOCK_PI op raced against a waiter
3165 * setting the FUTEX_WAITERS bit. Try again.
3170 * wake_futex_pi has detected invalid state. Tell user
3177 * We have no kernel internal state, i.e. no waiters in the
3178 * kernel. Waiters which are about to queue themselves are stuck
3179 * on hb->lock. So we can safely ignore them. We do neither
3180 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3183 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3184 spin_unlock(&hb->lock);
3199 * If uval has changed, let user space handle it.
3201 ret = (curval == uval) ? 0 : -EAGAIN;
3204 spin_unlock(&hb->lock);
3206 put_futex_key(&key);
3210 put_futex_key(&key);
3215 put_futex_key(&key);
3217 ret = fault_in_user_writeable(uaddr);
3225 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3226 * @hb: the hash_bucket futex_q was original enqueued on
3227 * @q: the futex_q woken while waiting to be requeued
3228 * @key2: the futex_key of the requeue target futex
3229 * @timeout: the timeout associated with the wait (NULL if none)
3231 * Detect if the task was woken on the initial futex as opposed to the requeue
3232 * target futex. If so, determine if it was a timeout or a signal that caused
3233 * the wakeup and return the appropriate error code to the caller. Must be
3234 * called with the hb lock held.
3237 * - 0 = no early wakeup detected;
3238 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3241 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3242 struct futex_q *q, union futex_key *key2,
3243 struct hrtimer_sleeper *timeout)
3248 * With the hb lock held, we avoid races while we process the wakeup.
3249 * We only need to hold hb (and not hb2) to ensure atomicity as the
3250 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3251 * It can't be requeued from uaddr2 to something else since we don't
3252 * support a PI aware source futex for requeue.
3254 if (!match_futex(&q->key, key2)) {
3255 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3257 * We were woken prior to requeue by a timeout or a signal.
3258 * Unqueue the futex_q and determine which it was.
3260 plist_del(&q->list, &hb->chain);
3263 /* Handle spurious wakeups gracefully */
3265 if (timeout && !timeout->task)
3267 else if (signal_pending(current))
3268 ret = -ERESTARTNOINTR;
3274 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3275 * @uaddr: the futex we initially wait on (non-pi)
3276 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3277 * the same type, no requeueing from private to shared, etc.
3278 * @val: the expected value of uaddr
3279 * @abs_time: absolute timeout
3280 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3281 * @uaddr2: the pi futex we will take prior to returning to user-space
3283 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3284 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3285 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3286 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3287 * without one, the pi logic would not know which task to boost/deboost, if
3288 * there was a need to.
3290 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3291 * via the following--
3292 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3293 * 2) wakeup on uaddr2 after a requeue
3297 * If 3, cleanup and return -ERESTARTNOINTR.
3299 * If 2, we may then block on trying to take the rt_mutex and return via:
3300 * 5) successful lock
3303 * 8) other lock acquisition failure
3305 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3307 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3313 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3314 u32 val, ktime_t *abs_time, u32 bitset,
3317 struct hrtimer_sleeper timeout, *to;
3318 struct rt_mutex_waiter rt_waiter;
3319 struct futex_hash_bucket *hb;
3320 union futex_key key2 = FUTEX_KEY_INIT;
3321 struct futex_q q = futex_q_init;
3324 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3327 if (uaddr == uaddr2)
3333 to = futex_setup_timer(abs_time, &timeout, flags,
3334 current->timer_slack_ns);
3337 * The waiter is allocated on our stack, manipulated by the requeue
3338 * code while we sleep on uaddr.
3340 rt_mutex_init_waiter(&rt_waiter);
3342 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3343 if (unlikely(ret != 0))
3347 q.rt_waiter = &rt_waiter;
3348 q.requeue_pi_key = &key2;
3351 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3354 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3359 * The check above which compares uaddrs is not sufficient for
3360 * shared futexes. We need to compare the keys:
3362 if (match_futex(&q.key, &key2)) {
3368 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3369 futex_wait_queue_me(hb, &q, to);
3371 spin_lock(&hb->lock);
3372 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3373 spin_unlock(&hb->lock);
3378 * In order for us to be here, we know our q.key == key2, and since
3379 * we took the hb->lock above, we also know that futex_requeue() has
3380 * completed and we no longer have to concern ourselves with a wakeup
3381 * race with the atomic proxy lock acquisition by the requeue code. The
3382 * futex_requeue dropped our key1 reference and incremented our key2
3386 /* Check if the requeue code acquired the second futex for us. */
3389 * Got the lock. We might not be the anticipated owner if we
3390 * did a lock-steal - fix up the PI-state in that case.
3392 if (q.pi_state && (q.pi_state->owner != current)) {
3393 spin_lock(q.lock_ptr);
3394 ret = fixup_pi_state_owner(uaddr2, &q, current);
3396 * Drop the reference to the pi state which
3397 * the requeue_pi() code acquired for us.
3399 put_pi_state(q.pi_state);
3400 spin_unlock(q.lock_ptr);
3402 * Adjust the return value. It's either -EFAULT or
3403 * success (1) but the caller expects 0 for success.
3405 ret = ret < 0 ? ret : 0;
3408 struct rt_mutex *pi_mutex;
3411 * We have been woken up by futex_unlock_pi(), a timeout, or a
3412 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3415 WARN_ON(!q.pi_state);
3416 pi_mutex = &q.pi_state->pi_mutex;
3417 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3419 spin_lock(q.lock_ptr);
3420 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3423 debug_rt_mutex_free_waiter(&rt_waiter);
3425 * Fixup the pi_state owner and possibly acquire the lock if we
3428 res = fixup_owner(uaddr2, &q, !ret);
3430 * If fixup_owner() returned an error, proprogate that. If it
3431 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3434 ret = (res < 0) ? res : 0;
3436 /* Unqueue and drop the lock. */
3440 if (ret == -EINTR) {
3442 * We've already been requeued, but cannot restart by calling
3443 * futex_lock_pi() directly. We could restart this syscall, but
3444 * it would detect that the user space "val" changed and return
3445 * -EWOULDBLOCK. Save the overhead of the restart and return
3446 * -EWOULDBLOCK directly.
3452 put_futex_key(&q.key);
3454 put_futex_key(&key2);
3458 hrtimer_cancel(&to->timer);
3459 destroy_hrtimer_on_stack(&to->timer);
3465 * Support for robust futexes: the kernel cleans up held futexes at
3468 * Implementation: user-space maintains a per-thread list of locks it
3469 * is holding. Upon do_exit(), the kernel carefully walks this list,
3470 * and marks all locks that are owned by this thread with the
3471 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3472 * always manipulated with the lock held, so the list is private and
3473 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3474 * field, to allow the kernel to clean up if the thread dies after
3475 * acquiring the lock, but just before it could have added itself to
3476 * the list. There can only be one such pending lock.
3480 * sys_set_robust_list() - Set the robust-futex list head of a task
3481 * @head: pointer to the list-head
3482 * @len: length of the list-head, as userspace expects
3484 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3487 if (!futex_cmpxchg_enabled)
3490 * The kernel knows only one size for now:
3492 if (unlikely(len != sizeof(*head)))
3495 current->robust_list = head;
3501 * sys_get_robust_list() - Get the robust-futex list head of a task
3502 * @pid: pid of the process [zero for current task]
3503 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3504 * @len_ptr: pointer to a length field, the kernel fills in the header size
3506 SYSCALL_DEFINE3(get_robust_list, int, pid,
3507 struct robust_list_head __user * __user *, head_ptr,
3508 size_t __user *, len_ptr)
3510 struct robust_list_head __user *head;
3512 struct task_struct *p;
3514 if (!futex_cmpxchg_enabled)
3523 p = find_task_by_vpid(pid);
3529 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3532 head = p->robust_list;
3535 if (put_user(sizeof(*head), len_ptr))
3537 return put_user(head, head_ptr);
3545 /* Constants for the pending_op argument of handle_futex_death */
3546 #define HANDLE_DEATH_PENDING true
3547 #define HANDLE_DEATH_LIST false
3550 * Process a futex-list entry, check whether it's owned by the
3551 * dying task, and do notification if so:
3553 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3554 bool pi, bool pending_op)
3556 u32 uval, uninitialized_var(nval), mval;
3559 /* Futex address must be 32bit aligned */
3560 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3564 if (get_user(uval, uaddr))
3568 * Special case for regular (non PI) futexes. The unlock path in
3569 * user space has two race scenarios:
3571 * 1. The unlock path releases the user space futex value and
3572 * before it can execute the futex() syscall to wake up
3573 * waiters it is killed.
3575 * 2. A woken up waiter is killed before it can acquire the
3576 * futex in user space.
3578 * In both cases the TID validation below prevents a wakeup of
3579 * potential waiters which can cause these waiters to block
3582 * In both cases the following conditions are met:
3584 * 1) task->robust_list->list_op_pending != NULL
3585 * @pending_op == true
3586 * 2) User space futex value == 0
3587 * 3) Regular futex: @pi == false
3589 * If these conditions are met, it is safe to attempt waking up a
3590 * potential waiter without touching the user space futex value and
3591 * trying to set the OWNER_DIED bit. The user space futex value is
3592 * uncontended and the rest of the user space mutex state is
3593 * consistent, so a woken waiter will just take over the
3594 * uncontended futex. Setting the OWNER_DIED bit would create
3595 * inconsistent state and malfunction of the user space owner died
3598 if (pending_op && !pi && !uval) {
3599 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3603 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3607 * Ok, this dying thread is truly holding a futex
3608 * of interest. Set the OWNER_DIED bit atomically
3609 * via cmpxchg, and if the value had FUTEX_WAITERS
3610 * set, wake up a waiter (if any). (We have to do a
3611 * futex_wake() even if OWNER_DIED is already set -
3612 * to handle the rare but possible case of recursive
3613 * thread-death.) The rest of the cleanup is done in
3616 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3619 * We are not holding a lock here, but we want to have
3620 * the pagefault_disable/enable() protection because
3621 * we want to handle the fault gracefully. If the
3622 * access fails we try to fault in the futex with R/W
3623 * verification via get_user_pages. get_user() above
3624 * does not guarantee R/W access. If that fails we
3625 * give up and leave the futex locked.
3627 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3630 if (fault_in_user_writeable(uaddr))
3648 * Wake robust non-PI futexes here. The wakeup of
3649 * PI futexes happens in exit_pi_state():
3651 if (!pi && (uval & FUTEX_WAITERS))
3652 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3658 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3660 static inline int fetch_robust_entry(struct robust_list __user **entry,
3661 struct robust_list __user * __user *head,
3664 unsigned long uentry;
3666 if (get_user(uentry, (unsigned long __user *)head))
3669 *entry = (void __user *)(uentry & ~1UL);
3676 * Walk curr->robust_list (very carefully, it's a userspace list!)
3677 * and mark any locks found there dead, and notify any waiters.
3679 * We silently return on any sign of list-walking problem.
3681 static void exit_robust_list(struct task_struct *curr)
3683 struct robust_list_head __user *head = curr->robust_list;
3684 struct robust_list __user *entry, *next_entry, *pending;
3685 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3686 unsigned int uninitialized_var(next_pi);
3687 unsigned long futex_offset;
3690 if (!futex_cmpxchg_enabled)
3694 * Fetch the list head (which was registered earlier, via
3695 * sys_set_robust_list()):
3697 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3700 * Fetch the relative futex offset:
3702 if (get_user(futex_offset, &head->futex_offset))
3705 * Fetch any possibly pending lock-add first, and handle it
3708 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3711 next_entry = NULL; /* avoid warning with gcc */
3712 while (entry != &head->list) {
3714 * Fetch the next entry in the list before calling
3715 * handle_futex_death:
3717 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3719 * A pending lock might already be on the list, so
3720 * don't process it twice:
3722 if (entry != pending) {
3723 if (handle_futex_death((void __user *)entry + futex_offset,
3724 curr, pi, HANDLE_DEATH_LIST))
3732 * Avoid excessively long or circular lists:
3741 handle_futex_death((void __user *)pending + futex_offset,
3742 curr, pip, HANDLE_DEATH_PENDING);
3746 static void futex_cleanup(struct task_struct *tsk)
3748 if (unlikely(tsk->robust_list)) {
3749 exit_robust_list(tsk);
3750 tsk->robust_list = NULL;
3753 #ifdef CONFIG_COMPAT
3754 if (unlikely(tsk->compat_robust_list)) {
3755 compat_exit_robust_list(tsk);
3756 tsk->compat_robust_list = NULL;
3760 if (unlikely(!list_empty(&tsk->pi_state_list)))
3761 exit_pi_state_list(tsk);
3765 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3766 * @tsk: task to set the state on
3768 * Set the futex exit state of the task lockless. The futex waiter code
3769 * observes that state when a task is exiting and loops until the task has
3770 * actually finished the futex cleanup. The worst case for this is that the
3771 * waiter runs through the wait loop until the state becomes visible.
3773 * This is called from the recursive fault handling path in do_exit().
3775 * This is best effort. Either the futex exit code has run already or
3776 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3777 * take it over. If not, the problem is pushed back to user space. If the
3778 * futex exit code did not run yet, then an already queued waiter might
3779 * block forever, but there is nothing which can be done about that.
3781 void futex_exit_recursive(struct task_struct *tsk)
3783 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3784 if (tsk->futex_state == FUTEX_STATE_EXITING)
3785 mutex_unlock(&tsk->futex_exit_mutex);
3786 tsk->futex_state = FUTEX_STATE_DEAD;
3789 static void futex_cleanup_begin(struct task_struct *tsk)
3792 * Prevent various race issues against a concurrent incoming waiter
3793 * including live locks by forcing the waiter to block on
3794 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3795 * attach_to_pi_owner().
3797 mutex_lock(&tsk->futex_exit_mutex);
3800 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3802 * This ensures that all subsequent checks of tsk->futex_state in
3803 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3804 * tsk->pi_lock held.
3806 * It guarantees also that a pi_state which was queued right before
3807 * the state change under tsk->pi_lock by a concurrent waiter must
3808 * be observed in exit_pi_state_list().
3810 raw_spin_lock_irq(&tsk->pi_lock);
3811 tsk->futex_state = FUTEX_STATE_EXITING;
3812 raw_spin_unlock_irq(&tsk->pi_lock);
3815 static void futex_cleanup_end(struct task_struct *tsk, int state)
3818 * Lockless store. The only side effect is that an observer might
3819 * take another loop until it becomes visible.
3821 tsk->futex_state = state;
3823 * Drop the exit protection. This unblocks waiters which observed
3824 * FUTEX_STATE_EXITING to reevaluate the state.
3826 mutex_unlock(&tsk->futex_exit_mutex);
3829 void futex_exec_release(struct task_struct *tsk)
3832 * The state handling is done for consistency, but in the case of
3833 * exec() there is no way to prevent futher damage as the PID stays
3834 * the same. But for the unlikely and arguably buggy case that a
3835 * futex is held on exec(), this provides at least as much state
3836 * consistency protection which is possible.
3838 futex_cleanup_begin(tsk);
3841 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3842 * exec a new binary.
3844 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3847 void futex_exit_release(struct task_struct *tsk)
3849 futex_cleanup_begin(tsk);
3851 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3854 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3855 u32 __user *uaddr2, u32 val2, u32 val3)
3857 int cmd = op & FUTEX_CMD_MASK;
3858 unsigned int flags = 0;
3860 if (!(op & FUTEX_PRIVATE_FLAG))
3861 flags |= FLAGS_SHARED;
3863 if (op & FUTEX_CLOCK_REALTIME) {
3864 flags |= FLAGS_CLOCKRT;
3865 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3871 case FUTEX_UNLOCK_PI:
3872 case FUTEX_TRYLOCK_PI:
3873 case FUTEX_WAIT_REQUEUE_PI:
3874 case FUTEX_CMP_REQUEUE_PI:
3875 if (!futex_cmpxchg_enabled)
3881 val3 = FUTEX_BITSET_MATCH_ANY;
3883 case FUTEX_WAIT_BITSET:
3884 return futex_wait(uaddr, flags, val, timeout, val3);
3886 val3 = FUTEX_BITSET_MATCH_ANY;
3888 case FUTEX_WAKE_BITSET:
3889 return futex_wake(uaddr, flags, val, val3);
3891 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3892 case FUTEX_CMP_REQUEUE:
3893 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3895 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3897 return futex_lock_pi(uaddr, flags, timeout, 0);
3898 case FUTEX_UNLOCK_PI:
3899 return futex_unlock_pi(uaddr, flags);
3900 case FUTEX_TRYLOCK_PI:
3901 return futex_lock_pi(uaddr, flags, NULL, 1);
3902 case FUTEX_WAIT_REQUEUE_PI:
3903 val3 = FUTEX_BITSET_MATCH_ANY;
3904 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3906 case FUTEX_CMP_REQUEUE_PI:
3907 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3913 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3914 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3917 struct timespec64 ts;
3918 ktime_t t, *tp = NULL;
3920 int cmd = op & FUTEX_CMD_MASK;
3922 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3923 cmd == FUTEX_WAIT_BITSET ||
3924 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3925 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3927 if (get_timespec64(&ts, utime))
3929 if (!timespec64_valid(&ts))
3932 t = timespec64_to_ktime(ts);
3933 if (cmd == FUTEX_WAIT)
3934 t = ktime_add_safe(ktime_get(), t);
3938 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3939 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3941 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3942 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3943 val2 = (u32) (unsigned long) utime;
3945 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3948 #ifdef CONFIG_COMPAT
3950 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3953 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3954 compat_uptr_t __user *head, unsigned int *pi)
3956 if (get_user(*uentry, head))
3959 *entry = compat_ptr((*uentry) & ~1);
3960 *pi = (unsigned int)(*uentry) & 1;
3965 static void __user *futex_uaddr(struct robust_list __user *entry,
3966 compat_long_t futex_offset)
3968 compat_uptr_t base = ptr_to_compat(entry);
3969 void __user *uaddr = compat_ptr(base + futex_offset);
3975 * Walk curr->robust_list (very carefully, it's a userspace list!)
3976 * and mark any locks found there dead, and notify any waiters.
3978 * We silently return on any sign of list-walking problem.
3980 static void compat_exit_robust_list(struct task_struct *curr)
3982 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3983 struct robust_list __user *entry, *next_entry, *pending;
3984 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3985 unsigned int uninitialized_var(next_pi);
3986 compat_uptr_t uentry, next_uentry, upending;
3987 compat_long_t futex_offset;
3990 if (!futex_cmpxchg_enabled)
3994 * Fetch the list head (which was registered earlier, via
3995 * sys_set_robust_list()):
3997 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
4000 * Fetch the relative futex offset:
4002 if (get_user(futex_offset, &head->futex_offset))
4005 * Fetch any possibly pending lock-add first, and handle it
4008 if (compat_fetch_robust_entry(&upending, &pending,
4009 &head->list_op_pending, &pip))
4012 next_entry = NULL; /* avoid warning with gcc */
4013 while (entry != (struct robust_list __user *) &head->list) {
4015 * Fetch the next entry in the list before calling
4016 * handle_futex_death:
4018 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
4019 (compat_uptr_t __user *)&entry->next, &next_pi);
4021 * A pending lock might already be on the list, so
4022 * dont process it twice:
4024 if (entry != pending) {
4025 void __user *uaddr = futex_uaddr(entry, futex_offset);
4027 if (handle_futex_death(uaddr, curr, pi,
4033 uentry = next_uentry;
4037 * Avoid excessively long or circular lists:
4045 void __user *uaddr = futex_uaddr(pending, futex_offset);
4047 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
4051 COMPAT_SYSCALL_DEFINE2(set_robust_list,
4052 struct compat_robust_list_head __user *, head,
4055 if (!futex_cmpxchg_enabled)
4058 if (unlikely(len != sizeof(*head)))
4061 current->compat_robust_list = head;
4066 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
4067 compat_uptr_t __user *, head_ptr,
4068 compat_size_t __user *, len_ptr)
4070 struct compat_robust_list_head __user *head;
4072 struct task_struct *p;
4074 if (!futex_cmpxchg_enabled)
4083 p = find_task_by_vpid(pid);
4089 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
4092 head = p->compat_robust_list;
4095 if (put_user(sizeof(*head), len_ptr))
4097 return put_user(ptr_to_compat(head), head_ptr);
4104 #endif /* CONFIG_COMPAT */
4106 #ifdef CONFIG_COMPAT_32BIT_TIME
4107 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
4108 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
4111 struct timespec64 ts;
4112 ktime_t t, *tp = NULL;
4114 int cmd = op & FUTEX_CMD_MASK;
4116 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4117 cmd == FUTEX_WAIT_BITSET ||
4118 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4119 if (get_old_timespec32(&ts, utime))
4121 if (!timespec64_valid(&ts))
4124 t = timespec64_to_ktime(ts);
4125 if (cmd == FUTEX_WAIT)
4126 t = ktime_add_safe(ktime_get(), t);
4129 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4130 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4131 val2 = (int) (unsigned long) utime;
4133 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4135 #endif /* CONFIG_COMPAT_32BIT_TIME */
4137 static void __init futex_detect_cmpxchg(void)
4139 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4143 * This will fail and we want it. Some arch implementations do
4144 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4145 * functionality. We want to know that before we call in any
4146 * of the complex code paths. Also we want to prevent
4147 * registration of robust lists in that case. NULL is
4148 * guaranteed to fault and we get -EFAULT on functional
4149 * implementation, the non-functional ones will return
4152 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4153 futex_cmpxchg_enabled = 1;
4157 static int __init futex_init(void)
4159 unsigned int futex_shift;
4162 #if CONFIG_BASE_SMALL
4163 futex_hashsize = 16;
4165 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4168 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4170 futex_hashsize < 256 ? HASH_SMALL : 0,
4172 futex_hashsize, futex_hashsize);
4173 futex_hashsize = 1UL << futex_shift;
4175 futex_detect_cmpxchg();
4177 for (i = 0; i < futex_hashsize; i++) {
4178 atomic_set(&futex_queues[i].waiters, 0);
4179 plist_head_init(&futex_queues[i].chain);
4180 spin_lock_init(&futex_queues[i].lock);
4185 core_initcall(futex_init);