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->both.word,
389 (sizeof(key->both.word)+sizeof(key->both.ptr))/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 ihold(key->shared.inode); /* implies 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)) {
466 iput(key->shared.inode);
468 case FUT_OFF_MMSHARED:
469 mmdrop(key->private.mm);
480 * futex_setup_timer - set up the sleeping hrtimer.
481 * @time: ptr to the given timeout value
482 * @timeout: the hrtimer_sleeper structure to be set up
483 * @flags: futex flags
484 * @range_ns: optional range in ns
486 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
489 static inline struct hrtimer_sleeper *
490 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
491 int flags, u64 range_ns)
496 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
497 CLOCK_REALTIME : CLOCK_MONOTONIC,
500 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
501 * effectively the same as calling hrtimer_set_expires().
503 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
509 * get_futex_key() - Get parameters which are the keys for a futex
510 * @uaddr: virtual address of the futex
511 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
512 * @key: address where result is stored.
513 * @rw: mapping needs to be read/write (values: FUTEX_READ,
516 * Return: a negative error code or 0
518 * The key words are stored in @key on success.
520 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
521 * offset_within_page). For private mappings, it's (uaddr, current->mm).
522 * We can usually work out the index without swapping in the page.
524 * lock_page() might sleep, the caller should not hold a spinlock.
527 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
529 unsigned long address = (unsigned long)uaddr;
530 struct mm_struct *mm = current->mm;
531 struct page *page, *tail;
532 struct address_space *mapping;
536 * The futex address must be "naturally" aligned.
538 key->both.offset = address % PAGE_SIZE;
539 if (unlikely((address % sizeof(u32)) != 0))
541 address -= key->both.offset;
543 if (unlikely(!access_ok(uaddr, sizeof(u32))))
546 if (unlikely(should_fail_futex(fshared)))
550 * PROCESS_PRIVATE futexes are fast.
551 * As the mm cannot disappear under us and the 'key' only needs
552 * virtual address, we dont even have to find the underlying vma.
553 * Note : We do have to check 'uaddr' is a valid user address,
554 * but access_ok() should be faster than find_vma()
557 key->private.mm = mm;
558 key->private.address = address;
559 get_futex_key_refs(key); /* implies smp_mb(); (B) */
564 /* Ignore any VERIFY_READ mapping (futex common case) */
565 if (unlikely(should_fail_futex(fshared)))
568 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
570 * If write access is not required (eg. FUTEX_WAIT), try
571 * and get read-only access.
573 if (err == -EFAULT && rw == FUTEX_READ) {
574 err = get_user_pages_fast(address, 1, 0, &page);
583 * The treatment of mapping from this point on is critical. The page
584 * lock protects many things but in this context the page lock
585 * stabilizes mapping, prevents inode freeing in the shared
586 * file-backed region case and guards against movement to swap cache.
588 * Strictly speaking the page lock is not needed in all cases being
589 * considered here and page lock forces unnecessarily serialization
590 * From this point on, mapping will be re-verified if necessary and
591 * page lock will be acquired only if it is unavoidable
593 * Mapping checks require the head page for any compound page so the
594 * head page and mapping is looked up now. For anonymous pages, it
595 * does not matter if the page splits in the future as the key is
596 * based on the address. For filesystem-backed pages, the tail is
597 * required as the index of the page determines the key. For
598 * base pages, there is no tail page and tail == page.
601 page = compound_head(page);
602 mapping = READ_ONCE(page->mapping);
605 * If page->mapping is NULL, then it cannot be a PageAnon
606 * page; but it might be the ZERO_PAGE or in the gate area or
607 * in a special mapping (all cases which we are happy to fail);
608 * or it may have been a good file page when get_user_pages_fast
609 * found it, but truncated or holepunched or subjected to
610 * invalidate_complete_page2 before we got the page lock (also
611 * cases which we are happy to fail). And we hold a reference,
612 * so refcount care in invalidate_complete_page's remove_mapping
613 * prevents drop_caches from setting mapping to NULL beneath us.
615 * The case we do have to guard against is when memory pressure made
616 * shmem_writepage move it from filecache to swapcache beneath us:
617 * an unlikely race, but we do need to retry for page->mapping.
619 if (unlikely(!mapping)) {
623 * Page lock is required to identify which special case above
624 * applies. If this is really a shmem page then the page lock
625 * will prevent unexpected transitions.
628 shmem_swizzled = PageSwapCache(page) || page->mapping;
639 * Private mappings are handled in a simple way.
641 * If the futex key is stored on an anonymous page, then the associated
642 * object is the mm which is implicitly pinned by the calling process.
644 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
645 * it's a read-only handle, it's expected that futexes attach to
646 * the object not the particular process.
648 if (PageAnon(page)) {
650 * A RO anonymous page will never change and thus doesn't make
651 * sense for futex operations.
653 if (unlikely(should_fail_futex(fshared)) || ro) {
658 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
659 key->private.mm = mm;
660 key->private.address = address;
662 get_futex_key_refs(key); /* implies smp_mb(); (B) */
668 * The associated futex object in this case is the inode and
669 * the page->mapping must be traversed. Ordinarily this should
670 * be stabilised under page lock but it's not strictly
671 * necessary in this case as we just want to pin the inode, not
672 * update the radix tree or anything like that.
674 * The RCU read lock is taken as the inode is finally freed
675 * under RCU. If the mapping still matches expectations then the
676 * mapping->host can be safely accessed as being a valid inode.
680 if (READ_ONCE(page->mapping) != mapping) {
687 inode = READ_ONCE(mapping->host);
696 * Take a reference unless it is about to be freed. Previously
697 * this reference was taken by ihold under the page lock
698 * pinning the inode in place so i_lock was unnecessary. The
699 * only way for this check to fail is if the inode was
700 * truncated in parallel which is almost certainly an
701 * application bug. In such a case, just retry.
703 * We are not calling into get_futex_key_refs() in file-backed
704 * cases, therefore a successful atomic_inc return below will
705 * guarantee that get_futex_key() will still imply smp_mb(); (B).
707 if (!atomic_inc_not_zero(&inode->i_count)) {
714 /* Should be impossible but lets be paranoid for now */
715 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
723 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
724 key->shared.inode = inode;
725 key->shared.pgoff = basepage_index(tail);
734 static inline void put_futex_key(union futex_key *key)
736 drop_futex_key_refs(key);
740 * fault_in_user_writeable() - Fault in user address and verify RW access
741 * @uaddr: pointer to faulting user space address
743 * Slow path to fixup the fault we just took in the atomic write
746 * We have no generic implementation of a non-destructive write to the
747 * user address. We know that we faulted in the atomic pagefault
748 * disabled section so we can as well avoid the #PF overhead by
749 * calling get_user_pages() right away.
751 static int fault_in_user_writeable(u32 __user *uaddr)
753 struct mm_struct *mm = current->mm;
756 down_read(&mm->mmap_sem);
757 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
758 FAULT_FLAG_WRITE, NULL);
759 up_read(&mm->mmap_sem);
761 return ret < 0 ? ret : 0;
765 * futex_top_waiter() - Return the highest priority waiter on a futex
766 * @hb: the hash bucket the futex_q's reside in
767 * @key: the futex key (to distinguish it from other futex futex_q's)
769 * Must be called with the hb lock held.
771 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
772 union futex_key *key)
774 struct futex_q *this;
776 plist_for_each_entry(this, &hb->chain, list) {
777 if (match_futex(&this->key, key))
783 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
784 u32 uval, u32 newval)
789 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
795 static int get_futex_value_locked(u32 *dest, u32 __user *from)
800 ret = __get_user(*dest, from);
803 return ret ? -EFAULT : 0;
810 static int refill_pi_state_cache(void)
812 struct futex_pi_state *pi_state;
814 if (likely(current->pi_state_cache))
817 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
822 INIT_LIST_HEAD(&pi_state->list);
823 /* pi_mutex gets initialized later */
824 pi_state->owner = NULL;
825 refcount_set(&pi_state->refcount, 1);
826 pi_state->key = FUTEX_KEY_INIT;
828 current->pi_state_cache = pi_state;
833 static struct futex_pi_state *alloc_pi_state(void)
835 struct futex_pi_state *pi_state = current->pi_state_cache;
838 current->pi_state_cache = NULL;
843 static void get_pi_state(struct futex_pi_state *pi_state)
845 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
849 * Drops a reference to the pi_state object and frees or caches it
850 * when the last reference is gone.
852 static void put_pi_state(struct futex_pi_state *pi_state)
857 if (!refcount_dec_and_test(&pi_state->refcount))
861 * If pi_state->owner is NULL, the owner is most probably dying
862 * and has cleaned up the pi_state already
864 if (pi_state->owner) {
865 struct task_struct *owner;
867 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
868 owner = pi_state->owner;
870 raw_spin_lock(&owner->pi_lock);
871 list_del_init(&pi_state->list);
872 raw_spin_unlock(&owner->pi_lock);
874 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
875 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
878 if (current->pi_state_cache) {
882 * pi_state->list is already empty.
883 * clear pi_state->owner.
884 * refcount is at 0 - put it back to 1.
886 pi_state->owner = NULL;
887 refcount_set(&pi_state->refcount, 1);
888 current->pi_state_cache = pi_state;
892 #ifdef CONFIG_FUTEX_PI
895 * This task is holding PI mutexes at exit time => bad.
896 * Kernel cleans up PI-state, but userspace is likely hosed.
897 * (Robust-futex cleanup is separate and might save the day for userspace.)
899 static void exit_pi_state_list(struct task_struct *curr)
901 struct list_head *next, *head = &curr->pi_state_list;
902 struct futex_pi_state *pi_state;
903 struct futex_hash_bucket *hb;
904 union futex_key key = FUTEX_KEY_INIT;
906 if (!futex_cmpxchg_enabled)
909 * We are a ZOMBIE and nobody can enqueue itself on
910 * pi_state_list anymore, but we have to be careful
911 * versus waiters unqueueing themselves:
913 raw_spin_lock_irq(&curr->pi_lock);
914 while (!list_empty(head)) {
916 pi_state = list_entry(next, struct futex_pi_state, list);
918 hb = hash_futex(&key);
921 * We can race against put_pi_state() removing itself from the
922 * list (a waiter going away). put_pi_state() will first
923 * decrement the reference count and then modify the list, so
924 * its possible to see the list entry but fail this reference
927 * In that case; drop the locks to let put_pi_state() make
928 * progress and retry the loop.
930 if (!refcount_inc_not_zero(&pi_state->refcount)) {
931 raw_spin_unlock_irq(&curr->pi_lock);
933 raw_spin_lock_irq(&curr->pi_lock);
936 raw_spin_unlock_irq(&curr->pi_lock);
938 spin_lock(&hb->lock);
939 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
940 raw_spin_lock(&curr->pi_lock);
942 * We dropped the pi-lock, so re-check whether this
943 * task still owns the PI-state:
945 if (head->next != next) {
946 /* retain curr->pi_lock for the loop invariant */
947 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
948 spin_unlock(&hb->lock);
949 put_pi_state(pi_state);
953 WARN_ON(pi_state->owner != curr);
954 WARN_ON(list_empty(&pi_state->list));
955 list_del_init(&pi_state->list);
956 pi_state->owner = NULL;
958 raw_spin_unlock(&curr->pi_lock);
959 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
960 spin_unlock(&hb->lock);
962 rt_mutex_futex_unlock(&pi_state->pi_mutex);
963 put_pi_state(pi_state);
965 raw_spin_lock_irq(&curr->pi_lock);
967 raw_spin_unlock_irq(&curr->pi_lock);
970 static inline void exit_pi_state_list(struct task_struct *curr) { }
974 * We need to check the following states:
976 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
978 * [1] NULL | --- | --- | 0 | 0/1 | Valid
979 * [2] NULL | --- | --- | >0 | 0/1 | Valid
981 * [3] Found | NULL | -- | Any | 0/1 | Invalid
983 * [4] Found | Found | NULL | 0 | 1 | Valid
984 * [5] Found | Found | NULL | >0 | 1 | Invalid
986 * [6] Found | Found | task | 0 | 1 | Valid
988 * [7] Found | Found | NULL | Any | 0 | Invalid
990 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
991 * [9] Found | Found | task | 0 | 0 | Invalid
992 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
994 * [1] Indicates that the kernel can acquire the futex atomically. We
995 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
997 * [2] Valid, if TID does not belong to a kernel thread. If no matching
998 * thread is found then it indicates that the owner TID has died.
1000 * [3] Invalid. The waiter is queued on a non PI futex
1002 * [4] Valid state after exit_robust_list(), which sets the user space
1003 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1005 * [5] The user space value got manipulated between exit_robust_list()
1006 * and exit_pi_state_list()
1008 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1009 * the pi_state but cannot access the user space value.
1011 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1013 * [8] Owner and user space value match
1015 * [9] There is no transient state which sets the user space TID to 0
1016 * except exit_robust_list(), but this is indicated by the
1017 * FUTEX_OWNER_DIED bit. See [4]
1019 * [10] There is no transient state which leaves owner and user space
1023 * Serialization and lifetime rules:
1027 * hb -> futex_q, relation
1028 * futex_q -> pi_state, relation
1030 * (cannot be raw because hb can contain arbitrary amount
1033 * pi_mutex->wait_lock:
1037 * (and pi_mutex 'obviously')
1041 * p->pi_state_list -> pi_state->list, relation
1043 * pi_state->refcount:
1051 * pi_mutex->wait_lock
1057 * Validate that the existing waiter has a pi_state and sanity check
1058 * the pi_state against the user space value. If correct, attach to
1061 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1062 struct futex_pi_state *pi_state,
1063 struct futex_pi_state **ps)
1065 pid_t pid = uval & FUTEX_TID_MASK;
1070 * Userspace might have messed up non-PI and PI futexes [3]
1072 if (unlikely(!pi_state))
1076 * We get here with hb->lock held, and having found a
1077 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1078 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1079 * which in turn means that futex_lock_pi() still has a reference on
1082 * The waiter holding a reference on @pi_state also protects against
1083 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1084 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1085 * free pi_state before we can take a reference ourselves.
1087 WARN_ON(!refcount_read(&pi_state->refcount));
1090 * Now that we have a pi_state, we can acquire wait_lock
1091 * and do the state validation.
1093 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1096 * Since {uval, pi_state} is serialized by wait_lock, and our current
1097 * uval was read without holding it, it can have changed. Verify it
1098 * still is what we expect it to be, otherwise retry the entire
1101 if (get_futex_value_locked(&uval2, uaddr))
1108 * Handle the owner died case:
1110 if (uval & FUTEX_OWNER_DIED) {
1112 * exit_pi_state_list sets owner to NULL and wakes the
1113 * topmost waiter. The task which acquires the
1114 * pi_state->rt_mutex will fixup owner.
1116 if (!pi_state->owner) {
1118 * No pi state owner, but the user space TID
1119 * is not 0. Inconsistent state. [5]
1124 * Take a ref on the state and return success. [4]
1130 * If TID is 0, then either the dying owner has not
1131 * yet executed exit_pi_state_list() or some waiter
1132 * acquired the rtmutex in the pi state, but did not
1133 * yet fixup the TID in user space.
1135 * Take a ref on the state and return success. [6]
1141 * If the owner died bit is not set, then the pi_state
1142 * must have an owner. [7]
1144 if (!pi_state->owner)
1149 * Bail out if user space manipulated the futex value. If pi
1150 * state exists then the owner TID must be the same as the
1151 * user space TID. [9/10]
1153 if (pid != task_pid_vnr(pi_state->owner))
1157 get_pi_state(pi_state);
1158 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1175 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1179 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1180 struct task_struct *tsk)
1185 * If the futex exit state is not yet FUTEX_STATE_DEAD, wait
1188 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1192 * Reread the user space value to handle the following situation:
1196 * sys_exit() sys_futex()
1197 * do_exit() futex_lock_pi()
1198 * futex_lock_pi_atomic()
1199 * exit_signals(tsk) No waiters:
1200 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1201 * mm_release(tsk) Set waiter bit
1202 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1203 * Set owner died attach_to_pi_owner() {
1204 * *uaddr = 0xC0000000; tsk = get_task(PID);
1205 * } if (!tsk->flags & PF_EXITING) {
1207 * tsk->futex_state = } else {
1208 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1211 * return -ESRCH; <--- FAIL
1214 * Returning ESRCH unconditionally is wrong here because the
1215 * user space value has been changed by the exiting task.
1217 * The same logic applies to the case where the exiting task is
1220 if (get_futex_value_locked(&uval2, uaddr))
1223 /* If the user space value has changed, try again. */
1228 * The exiting task did not have a robust list, the robust list was
1229 * corrupted or the user space value in *uaddr is simply bogus.
1230 * Give up and tell user space.
1236 * Lookup the task for the TID provided from user space and attach to
1237 * it after doing proper sanity checks.
1239 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1240 struct futex_pi_state **ps)
1242 pid_t pid = uval & FUTEX_TID_MASK;
1243 struct futex_pi_state *pi_state;
1244 struct task_struct *p;
1247 * We are the first waiter - try to look up the real owner and attach
1248 * the new pi_state to it, but bail out when TID = 0 [1]
1250 * The !pid check is paranoid. None of the call sites should end up
1251 * with pid == 0, but better safe than sorry. Let the caller retry
1255 p = find_get_task_by_vpid(pid);
1257 return handle_exit_race(uaddr, uval, NULL);
1259 if (unlikely(p->flags & PF_KTHREAD)) {
1265 * We need to look at the task state to figure out, whether the
1266 * task is exiting. To protect against the change of the task state
1267 * in futex_exit_release(), we do this protected by p->pi_lock:
1269 raw_spin_lock_irq(&p->pi_lock);
1270 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1272 * The task is on the way out. When the futex state is
1273 * FUTEX_STATE_DEAD, we know that the task has finished
1276 int ret = handle_exit_race(uaddr, uval, p);
1278 raw_spin_unlock_irq(&p->pi_lock);
1284 * No existing pi state. First waiter. [2]
1286 * This creates pi_state, we have hb->lock held, this means nothing can
1287 * observe this state, wait_lock is irrelevant.
1289 pi_state = alloc_pi_state();
1292 * Initialize the pi_mutex in locked state and make @p
1295 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1297 /* Store the key for possible exit cleanups: */
1298 pi_state->key = *key;
1300 WARN_ON(!list_empty(&pi_state->list));
1301 list_add(&pi_state->list, &p->pi_state_list);
1303 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1304 * because there is no concurrency as the object is not published yet.
1306 pi_state->owner = p;
1307 raw_spin_unlock_irq(&p->pi_lock);
1316 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1317 struct futex_hash_bucket *hb,
1318 union futex_key *key, struct futex_pi_state **ps)
1320 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1323 * If there is a waiter on that futex, validate it and
1324 * attach to the pi_state when the validation succeeds.
1327 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1330 * We are the first waiter - try to look up the owner based on
1331 * @uval and attach to it.
1333 return attach_to_pi_owner(uaddr, uval, key, ps);
1336 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1339 u32 uninitialized_var(curval);
1341 if (unlikely(should_fail_futex(true)))
1344 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1348 /* If user space value changed, let the caller retry */
1349 return curval != uval ? -EAGAIN : 0;
1353 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1354 * @uaddr: the pi futex user address
1355 * @hb: the pi futex hash bucket
1356 * @key: the futex key associated with uaddr and hb
1357 * @ps: the pi_state pointer where we store the result of the
1359 * @task: the task to perform the atomic lock work for. This will
1360 * be "current" except in the case of requeue pi.
1361 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1364 * - 0 - ready to wait;
1365 * - 1 - acquired the lock;
1368 * The hb->lock and futex_key refs shall be held by the caller.
1370 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1371 union futex_key *key,
1372 struct futex_pi_state **ps,
1373 struct task_struct *task, int set_waiters)
1375 u32 uval, newval, vpid = task_pid_vnr(task);
1376 struct futex_q *top_waiter;
1380 * Read the user space value first so we can validate a few
1381 * things before proceeding further.
1383 if (get_futex_value_locked(&uval, uaddr))
1386 if (unlikely(should_fail_futex(true)))
1392 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1395 if ((unlikely(should_fail_futex(true))))
1399 * Lookup existing state first. If it exists, try to attach to
1402 top_waiter = futex_top_waiter(hb, key);
1404 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1407 * No waiter and user TID is 0. We are here because the
1408 * waiters or the owner died bit is set or called from
1409 * requeue_cmp_pi or for whatever reason something took the
1412 if (!(uval & FUTEX_TID_MASK)) {
1414 * We take over the futex. No other waiters and the user space
1415 * TID is 0. We preserve the owner died bit.
1417 newval = uval & FUTEX_OWNER_DIED;
1420 /* The futex requeue_pi code can enforce the waiters bit */
1422 newval |= FUTEX_WAITERS;
1424 ret = lock_pi_update_atomic(uaddr, uval, newval);
1425 /* If the take over worked, return 1 */
1426 return ret < 0 ? ret : 1;
1430 * First waiter. Set the waiters bit before attaching ourself to
1431 * the owner. If owner tries to unlock, it will be forced into
1432 * the kernel and blocked on hb->lock.
1434 newval = uval | FUTEX_WAITERS;
1435 ret = lock_pi_update_atomic(uaddr, uval, newval);
1439 * If the update of the user space value succeeded, we try to
1440 * attach to the owner. If that fails, no harm done, we only
1441 * set the FUTEX_WAITERS bit in the user space variable.
1443 return attach_to_pi_owner(uaddr, newval, key, ps);
1447 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1448 * @q: The futex_q to unqueue
1450 * The q->lock_ptr must not be NULL and must be held by the caller.
1452 static void __unqueue_futex(struct futex_q *q)
1454 struct futex_hash_bucket *hb;
1456 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1458 lockdep_assert_held(q->lock_ptr);
1460 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1461 plist_del(&q->list, &hb->chain);
1466 * The hash bucket lock must be held when this is called.
1467 * Afterwards, the futex_q must not be accessed. Callers
1468 * must ensure to later call wake_up_q() for the actual
1471 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1473 struct task_struct *p = q->task;
1475 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1481 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1482 * is written, without taking any locks. This is possible in the event
1483 * of a spurious wakeup, for example. A memory barrier is required here
1484 * to prevent the following store to lock_ptr from getting ahead of the
1485 * plist_del in __unqueue_futex().
1487 smp_store_release(&q->lock_ptr, NULL);
1490 * Queue the task for later wakeup for after we've released
1493 wake_q_add_safe(wake_q, p);
1497 * Caller must hold a reference on @pi_state.
1499 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1501 u32 uninitialized_var(curval), newval;
1502 struct task_struct *new_owner;
1503 bool postunlock = false;
1504 DEFINE_WAKE_Q(wake_q);
1507 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1508 if (WARN_ON_ONCE(!new_owner)) {
1510 * As per the comment in futex_unlock_pi() this should not happen.
1512 * When this happens, give up our locks and try again, giving
1513 * the futex_lock_pi() instance time to complete, either by
1514 * waiting on the rtmutex or removing itself from the futex
1522 * We pass it to the next owner. The WAITERS bit is always kept
1523 * enabled while there is PI state around. We cleanup the owner
1524 * died bit, because we are the owner.
1526 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1528 if (unlikely(should_fail_futex(true)))
1531 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1532 if (!ret && (curval != uval)) {
1534 * If a unconditional UNLOCK_PI operation (user space did not
1535 * try the TID->0 transition) raced with a waiter setting the
1536 * FUTEX_WAITERS flag between get_user() and locking the hash
1537 * bucket lock, retry the operation.
1539 if ((FUTEX_TID_MASK & curval) == uval)
1549 * This is a point of no return; once we modify the uval there is no
1550 * going back and subsequent operations must not fail.
1553 raw_spin_lock(&pi_state->owner->pi_lock);
1554 WARN_ON(list_empty(&pi_state->list));
1555 list_del_init(&pi_state->list);
1556 raw_spin_unlock(&pi_state->owner->pi_lock);
1558 raw_spin_lock(&new_owner->pi_lock);
1559 WARN_ON(!list_empty(&pi_state->list));
1560 list_add(&pi_state->list, &new_owner->pi_state_list);
1561 pi_state->owner = new_owner;
1562 raw_spin_unlock(&new_owner->pi_lock);
1564 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1567 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1570 rt_mutex_postunlock(&wake_q);
1576 * Express the locking dependencies for lockdep:
1579 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1582 spin_lock(&hb1->lock);
1584 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1585 } else { /* hb1 > hb2 */
1586 spin_lock(&hb2->lock);
1587 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1592 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1594 spin_unlock(&hb1->lock);
1596 spin_unlock(&hb2->lock);
1600 * Wake up waiters matching bitset queued on this futex (uaddr).
1603 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1605 struct futex_hash_bucket *hb;
1606 struct futex_q *this, *next;
1607 union futex_key key = FUTEX_KEY_INIT;
1609 DEFINE_WAKE_Q(wake_q);
1614 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1615 if (unlikely(ret != 0))
1618 hb = hash_futex(&key);
1620 /* Make sure we really have tasks to wakeup */
1621 if (!hb_waiters_pending(hb))
1624 spin_lock(&hb->lock);
1626 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1627 if (match_futex (&this->key, &key)) {
1628 if (this->pi_state || this->rt_waiter) {
1633 /* Check if one of the bits is set in both bitsets */
1634 if (!(this->bitset & bitset))
1637 mark_wake_futex(&wake_q, this);
1638 if (++ret >= nr_wake)
1643 spin_unlock(&hb->lock);
1646 put_futex_key(&key);
1651 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1653 unsigned int op = (encoded_op & 0x70000000) >> 28;
1654 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1655 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1656 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1659 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1660 if (oparg < 0 || oparg > 31) {
1661 char comm[sizeof(current->comm)];
1663 * kill this print and return -EINVAL when userspace
1666 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1667 get_task_comm(comm, current), oparg);
1673 if (!access_ok(uaddr, sizeof(u32)))
1676 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1681 case FUTEX_OP_CMP_EQ:
1682 return oldval == cmparg;
1683 case FUTEX_OP_CMP_NE:
1684 return oldval != cmparg;
1685 case FUTEX_OP_CMP_LT:
1686 return oldval < cmparg;
1687 case FUTEX_OP_CMP_GE:
1688 return oldval >= cmparg;
1689 case FUTEX_OP_CMP_LE:
1690 return oldval <= cmparg;
1691 case FUTEX_OP_CMP_GT:
1692 return oldval > cmparg;
1699 * Wake up all waiters hashed on the physical page that is mapped
1700 * to this virtual address:
1703 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1704 int nr_wake, int nr_wake2, int op)
1706 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1707 struct futex_hash_bucket *hb1, *hb2;
1708 struct futex_q *this, *next;
1710 DEFINE_WAKE_Q(wake_q);
1713 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1714 if (unlikely(ret != 0))
1716 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1717 if (unlikely(ret != 0))
1720 hb1 = hash_futex(&key1);
1721 hb2 = hash_futex(&key2);
1724 double_lock_hb(hb1, hb2);
1725 op_ret = futex_atomic_op_inuser(op, uaddr2);
1726 if (unlikely(op_ret < 0)) {
1727 double_unlock_hb(hb1, hb2);
1729 if (!IS_ENABLED(CONFIG_MMU) ||
1730 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1732 * we don't get EFAULT from MMU faults if we don't have
1733 * an MMU, but we might get them from range checking
1739 if (op_ret == -EFAULT) {
1740 ret = fault_in_user_writeable(uaddr2);
1745 if (!(flags & FLAGS_SHARED)) {
1750 put_futex_key(&key2);
1751 put_futex_key(&key1);
1756 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1757 if (match_futex (&this->key, &key1)) {
1758 if (this->pi_state || this->rt_waiter) {
1762 mark_wake_futex(&wake_q, this);
1763 if (++ret >= nr_wake)
1770 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1771 if (match_futex (&this->key, &key2)) {
1772 if (this->pi_state || this->rt_waiter) {
1776 mark_wake_futex(&wake_q, this);
1777 if (++op_ret >= nr_wake2)
1785 double_unlock_hb(hb1, hb2);
1788 put_futex_key(&key2);
1790 put_futex_key(&key1);
1796 * requeue_futex() - Requeue a futex_q from one hb to another
1797 * @q: the futex_q to requeue
1798 * @hb1: the source hash_bucket
1799 * @hb2: the target hash_bucket
1800 * @key2: the new key for the requeued futex_q
1803 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1804 struct futex_hash_bucket *hb2, union futex_key *key2)
1808 * If key1 and key2 hash to the same bucket, no need to
1811 if (likely(&hb1->chain != &hb2->chain)) {
1812 plist_del(&q->list, &hb1->chain);
1813 hb_waiters_dec(hb1);
1814 hb_waiters_inc(hb2);
1815 plist_add(&q->list, &hb2->chain);
1816 q->lock_ptr = &hb2->lock;
1818 get_futex_key_refs(key2);
1823 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1825 * @key: the key of the requeue target futex
1826 * @hb: the hash_bucket of the requeue target futex
1828 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1829 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1830 * to the requeue target futex so the waiter can detect the wakeup on the right
1831 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1832 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1833 * to protect access to the pi_state to fixup the owner later. Must be called
1834 * with both q->lock_ptr and hb->lock held.
1837 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1838 struct futex_hash_bucket *hb)
1840 get_futex_key_refs(key);
1845 WARN_ON(!q->rt_waiter);
1846 q->rt_waiter = NULL;
1848 q->lock_ptr = &hb->lock;
1850 wake_up_state(q->task, TASK_NORMAL);
1854 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1855 * @pifutex: the user address of the to futex
1856 * @hb1: the from futex hash bucket, must be locked by the caller
1857 * @hb2: the to futex hash bucket, must be locked by the caller
1858 * @key1: the from futex key
1859 * @key2: the to futex key
1860 * @ps: address to store the pi_state pointer
1861 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1863 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1864 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1865 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1866 * hb1 and hb2 must be held by the caller.
1869 * - 0 - failed to acquire the lock atomically;
1870 * - >0 - acquired the lock, return value is vpid of the top_waiter
1873 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1874 struct futex_hash_bucket *hb1,
1875 struct futex_hash_bucket *hb2,
1876 union futex_key *key1, union futex_key *key2,
1877 struct futex_pi_state **ps, int set_waiters)
1879 struct futex_q *top_waiter = NULL;
1883 if (get_futex_value_locked(&curval, pifutex))
1886 if (unlikely(should_fail_futex(true)))
1890 * Find the top_waiter and determine if there are additional waiters.
1891 * If the caller intends to requeue more than 1 waiter to pifutex,
1892 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1893 * as we have means to handle the possible fault. If not, don't set
1894 * the bit unecessarily as it will force the subsequent unlock to enter
1897 top_waiter = futex_top_waiter(hb1, key1);
1899 /* There are no waiters, nothing for us to do. */
1903 /* Ensure we requeue to the expected futex. */
1904 if (!match_futex(top_waiter->requeue_pi_key, key2))
1908 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1909 * the contended case or if set_waiters is 1. The pi_state is returned
1910 * in ps in contended cases.
1912 vpid = task_pid_vnr(top_waiter->task);
1913 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1916 requeue_pi_wake_futex(top_waiter, key2, hb2);
1923 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1924 * @uaddr1: source futex user address
1925 * @flags: futex flags (FLAGS_SHARED, etc.)
1926 * @uaddr2: target futex user address
1927 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1928 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1929 * @cmpval: @uaddr1 expected value (or %NULL)
1930 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1931 * pi futex (pi to pi requeue is not supported)
1933 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1934 * uaddr2 atomically on behalf of the top waiter.
1937 * - >=0 - on success, the number of tasks requeued or woken;
1940 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1941 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1942 u32 *cmpval, int requeue_pi)
1944 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1945 int drop_count = 0, task_count = 0, ret;
1946 struct futex_pi_state *pi_state = NULL;
1947 struct futex_hash_bucket *hb1, *hb2;
1948 struct futex_q *this, *next;
1949 DEFINE_WAKE_Q(wake_q);
1951 if (nr_wake < 0 || nr_requeue < 0)
1955 * When PI not supported: return -ENOSYS if requeue_pi is true,
1956 * consequently the compiler knows requeue_pi is always false past
1957 * this point which will optimize away all the conditional code
1960 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1965 * Requeue PI only works on two distinct uaddrs. This
1966 * check is only valid for private futexes. See below.
1968 if (uaddr1 == uaddr2)
1972 * requeue_pi requires a pi_state, try to allocate it now
1973 * without any locks in case it fails.
1975 if (refill_pi_state_cache())
1978 * requeue_pi must wake as many tasks as it can, up to nr_wake
1979 * + nr_requeue, since it acquires the rt_mutex prior to
1980 * returning to userspace, so as to not leave the rt_mutex with
1981 * waiters and no owner. However, second and third wake-ups
1982 * cannot be predicted as they involve race conditions with the
1983 * first wake and a fault while looking up the pi_state. Both
1984 * pthread_cond_signal() and pthread_cond_broadcast() should
1992 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1993 if (unlikely(ret != 0))
1995 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1996 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1997 if (unlikely(ret != 0))
2001 * The check above which compares uaddrs is not sufficient for
2002 * shared futexes. We need to compare the keys:
2004 if (requeue_pi && match_futex(&key1, &key2)) {
2009 hb1 = hash_futex(&key1);
2010 hb2 = hash_futex(&key2);
2013 hb_waiters_inc(hb2);
2014 double_lock_hb(hb1, hb2);
2016 if (likely(cmpval != NULL)) {
2019 ret = get_futex_value_locked(&curval, uaddr1);
2021 if (unlikely(ret)) {
2022 double_unlock_hb(hb1, hb2);
2023 hb_waiters_dec(hb2);
2025 ret = get_user(curval, uaddr1);
2029 if (!(flags & FLAGS_SHARED))
2032 put_futex_key(&key2);
2033 put_futex_key(&key1);
2036 if (curval != *cmpval) {
2042 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2044 * Attempt to acquire uaddr2 and wake the top waiter. If we
2045 * intend to requeue waiters, force setting the FUTEX_WAITERS
2046 * bit. We force this here where we are able to easily handle
2047 * faults rather in the requeue loop below.
2049 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2050 &key2, &pi_state, nr_requeue);
2053 * At this point the top_waiter has either taken uaddr2 or is
2054 * waiting on it. If the former, then the pi_state will not
2055 * exist yet, look it up one more time to ensure we have a
2056 * reference to it. If the lock was taken, ret contains the
2057 * vpid of the top waiter task.
2058 * If the lock was not taken, we have pi_state and an initial
2059 * refcount on it. In case of an error we have nothing.
2066 * If we acquired the lock, then the user space value
2067 * of uaddr2 should be vpid. It cannot be changed by
2068 * the top waiter as it is blocked on hb2 lock if it
2069 * tries to do so. If something fiddled with it behind
2070 * our back the pi state lookup might unearth it. So
2071 * we rather use the known value than rereading and
2072 * handing potential crap to lookup_pi_state.
2074 * If that call succeeds then we have pi_state and an
2075 * initial refcount on it.
2077 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2082 /* We hold a reference on the pi state. */
2085 /* If the above failed, then pi_state is NULL */
2087 double_unlock_hb(hb1, hb2);
2088 hb_waiters_dec(hb2);
2089 put_futex_key(&key2);
2090 put_futex_key(&key1);
2091 ret = fault_in_user_writeable(uaddr2);
2097 * Two reasons for this:
2098 * - Owner is exiting and we just wait for the
2100 * - The user space value changed.
2102 double_unlock_hb(hb1, hb2);
2103 hb_waiters_dec(hb2);
2104 put_futex_key(&key2);
2105 put_futex_key(&key1);
2113 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2114 if (task_count - nr_wake >= nr_requeue)
2117 if (!match_futex(&this->key, &key1))
2121 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2122 * be paired with each other and no other futex ops.
2124 * We should never be requeueing a futex_q with a pi_state,
2125 * which is awaiting a futex_unlock_pi().
2127 if ((requeue_pi && !this->rt_waiter) ||
2128 (!requeue_pi && this->rt_waiter) ||
2135 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2136 * lock, we already woke the top_waiter. If not, it will be
2137 * woken by futex_unlock_pi().
2139 if (++task_count <= nr_wake && !requeue_pi) {
2140 mark_wake_futex(&wake_q, this);
2144 /* Ensure we requeue to the expected futex for requeue_pi. */
2145 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2151 * Requeue nr_requeue waiters and possibly one more in the case
2152 * of requeue_pi if we couldn't acquire the lock atomically.
2156 * Prepare the waiter to take the rt_mutex. Take a
2157 * refcount on the pi_state and store the pointer in
2158 * the futex_q object of the waiter.
2160 get_pi_state(pi_state);
2161 this->pi_state = pi_state;
2162 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2167 * We got the lock. We do neither drop the
2168 * refcount on pi_state nor clear
2169 * this->pi_state because the waiter needs the
2170 * pi_state for cleaning up the user space
2171 * value. It will drop the refcount after
2174 requeue_pi_wake_futex(this, &key2, hb2);
2179 * rt_mutex_start_proxy_lock() detected a
2180 * potential deadlock when we tried to queue
2181 * that waiter. Drop the pi_state reference
2182 * which we took above and remove the pointer
2183 * to the state from the waiters futex_q
2186 this->pi_state = NULL;
2187 put_pi_state(pi_state);
2189 * We stop queueing more waiters and let user
2190 * space deal with the mess.
2195 requeue_futex(this, hb1, hb2, &key2);
2200 * We took an extra initial reference to the pi_state either
2201 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2202 * need to drop it here again.
2204 put_pi_state(pi_state);
2207 double_unlock_hb(hb1, hb2);
2209 hb_waiters_dec(hb2);
2212 * drop_futex_key_refs() must be called outside the spinlocks. During
2213 * the requeue we moved futex_q's from the hash bucket at key1 to the
2214 * one at key2 and updated their key pointer. We no longer need to
2215 * hold the references to key1.
2217 while (--drop_count >= 0)
2218 drop_futex_key_refs(&key1);
2221 put_futex_key(&key2);
2223 put_futex_key(&key1);
2225 return ret ? ret : task_count;
2228 /* The key must be already stored in q->key. */
2229 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2230 __acquires(&hb->lock)
2232 struct futex_hash_bucket *hb;
2234 hb = hash_futex(&q->key);
2237 * Increment the counter before taking the lock so that
2238 * a potential waker won't miss a to-be-slept task that is
2239 * waiting for the spinlock. This is safe as all queue_lock()
2240 * users end up calling queue_me(). Similarly, for housekeeping,
2241 * decrement the counter at queue_unlock() when some error has
2242 * occurred and we don't end up adding the task to the list.
2244 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2246 q->lock_ptr = &hb->lock;
2248 spin_lock(&hb->lock);
2253 queue_unlock(struct futex_hash_bucket *hb)
2254 __releases(&hb->lock)
2256 spin_unlock(&hb->lock);
2260 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2265 * The priority used to register this element is
2266 * - either the real thread-priority for the real-time threads
2267 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2268 * - or MAX_RT_PRIO for non-RT threads.
2269 * Thus, all RT-threads are woken first in priority order, and
2270 * the others are woken last, in FIFO order.
2272 prio = min(current->normal_prio, MAX_RT_PRIO);
2274 plist_node_init(&q->list, prio);
2275 plist_add(&q->list, &hb->chain);
2280 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2281 * @q: The futex_q to enqueue
2282 * @hb: The destination hash bucket
2284 * The hb->lock must be held by the caller, and is released here. A call to
2285 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2286 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2287 * or nothing if the unqueue is done as part of the wake process and the unqueue
2288 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2291 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2292 __releases(&hb->lock)
2295 spin_unlock(&hb->lock);
2299 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2300 * @q: The futex_q to unqueue
2302 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2303 * be paired with exactly one earlier call to queue_me().
2306 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2307 * - 0 - if the futex_q was already removed by the waking thread
2309 static int unqueue_me(struct futex_q *q)
2311 spinlock_t *lock_ptr;
2314 /* In the common case we don't take the spinlock, which is nice. */
2317 * q->lock_ptr can change between this read and the following spin_lock.
2318 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2319 * optimizing lock_ptr out of the logic below.
2321 lock_ptr = READ_ONCE(q->lock_ptr);
2322 if (lock_ptr != NULL) {
2323 spin_lock(lock_ptr);
2325 * q->lock_ptr can change between reading it and
2326 * spin_lock(), causing us to take the wrong lock. This
2327 * corrects the race condition.
2329 * Reasoning goes like this: if we have the wrong lock,
2330 * q->lock_ptr must have changed (maybe several times)
2331 * between reading it and the spin_lock(). It can
2332 * change again after the spin_lock() but only if it was
2333 * already changed before the spin_lock(). It cannot,
2334 * however, change back to the original value. Therefore
2335 * we can detect whether we acquired the correct lock.
2337 if (unlikely(lock_ptr != q->lock_ptr)) {
2338 spin_unlock(lock_ptr);
2343 BUG_ON(q->pi_state);
2345 spin_unlock(lock_ptr);
2349 drop_futex_key_refs(&q->key);
2354 * PI futexes can not be requeued and must remove themself from the
2355 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2358 static void unqueue_me_pi(struct futex_q *q)
2359 __releases(q->lock_ptr)
2363 BUG_ON(!q->pi_state);
2364 put_pi_state(q->pi_state);
2367 spin_unlock(q->lock_ptr);
2370 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2371 struct task_struct *argowner)
2373 struct futex_pi_state *pi_state = q->pi_state;
2374 u32 uval, uninitialized_var(curval), newval;
2375 struct task_struct *oldowner, *newowner;
2379 lockdep_assert_held(q->lock_ptr);
2381 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2383 oldowner = pi_state->owner;
2386 * We are here because either:
2388 * - we stole the lock and pi_state->owner needs updating to reflect
2389 * that (@argowner == current),
2393 * - someone stole our lock and we need to fix things to point to the
2394 * new owner (@argowner == NULL).
2396 * Either way, we have to replace the TID in the user space variable.
2397 * This must be atomic as we have to preserve the owner died bit here.
2399 * Note: We write the user space value _before_ changing the pi_state
2400 * because we can fault here. Imagine swapped out pages or a fork
2401 * that marked all the anonymous memory readonly for cow.
2403 * Modifying pi_state _before_ the user space value would leave the
2404 * pi_state in an inconsistent state when we fault here, because we
2405 * need to drop the locks to handle the fault. This might be observed
2406 * in the PID check in lookup_pi_state.
2410 if (oldowner != current) {
2412 * We raced against a concurrent self; things are
2413 * already fixed up. Nothing to do.
2419 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2420 /* We got the lock after all, nothing to fix. */
2426 * Since we just failed the trylock; there must be an owner.
2428 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2431 WARN_ON_ONCE(argowner != current);
2432 if (oldowner == current) {
2434 * We raced against a concurrent self; things are
2435 * already fixed up. Nothing to do.
2440 newowner = argowner;
2443 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2445 if (!pi_state->owner)
2446 newtid |= FUTEX_OWNER_DIED;
2448 err = get_futex_value_locked(&uval, uaddr);
2453 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2455 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2465 * We fixed up user space. Now we need to fix the pi_state
2468 if (pi_state->owner != NULL) {
2469 raw_spin_lock(&pi_state->owner->pi_lock);
2470 WARN_ON(list_empty(&pi_state->list));
2471 list_del_init(&pi_state->list);
2472 raw_spin_unlock(&pi_state->owner->pi_lock);
2475 pi_state->owner = newowner;
2477 raw_spin_lock(&newowner->pi_lock);
2478 WARN_ON(!list_empty(&pi_state->list));
2479 list_add(&pi_state->list, &newowner->pi_state_list);
2480 raw_spin_unlock(&newowner->pi_lock);
2481 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2486 * In order to reschedule or handle a page fault, we need to drop the
2487 * locks here. In the case of a fault, this gives the other task
2488 * (either the highest priority waiter itself or the task which stole
2489 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2490 * are back from handling the fault we need to check the pi_state after
2491 * reacquiring the locks and before trying to do another fixup. When
2492 * the fixup has been done already we simply return.
2494 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2495 * drop hb->lock since the caller owns the hb -> futex_q relation.
2496 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2499 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2500 spin_unlock(q->lock_ptr);
2504 ret = fault_in_user_writeable(uaddr);
2518 spin_lock(q->lock_ptr);
2519 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2522 * Check if someone else fixed it for us:
2524 if (pi_state->owner != oldowner) {
2535 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2539 static long futex_wait_restart(struct restart_block *restart);
2542 * fixup_owner() - Post lock pi_state and corner case management
2543 * @uaddr: user address of the futex
2544 * @q: futex_q (contains pi_state and access to the rt_mutex)
2545 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2547 * After attempting to lock an rt_mutex, this function is called to cleanup
2548 * the pi_state owner as well as handle race conditions that may allow us to
2549 * acquire the lock. Must be called with the hb lock held.
2552 * - 1 - success, lock taken;
2553 * - 0 - success, lock not taken;
2554 * - <0 - on error (-EFAULT)
2556 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2562 * Got the lock. We might not be the anticipated owner if we
2563 * did a lock-steal - fix up the PI-state in that case:
2565 * Speculative pi_state->owner read (we don't hold wait_lock);
2566 * since we own the lock pi_state->owner == current is the
2567 * stable state, anything else needs more attention.
2569 if (q->pi_state->owner != current)
2570 ret = fixup_pi_state_owner(uaddr, q, current);
2575 * If we didn't get the lock; check if anybody stole it from us. In
2576 * that case, we need to fix up the uval to point to them instead of
2577 * us, otherwise bad things happen. [10]
2579 * Another speculative read; pi_state->owner == current is unstable
2580 * but needs our attention.
2582 if (q->pi_state->owner == current) {
2583 ret = fixup_pi_state_owner(uaddr, q, NULL);
2588 * Paranoia check. If we did not take the lock, then we should not be
2589 * the owner of the rt_mutex.
2591 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2592 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2593 "pi-state %p\n", ret,
2594 q->pi_state->pi_mutex.owner,
2595 q->pi_state->owner);
2599 return ret ? ret : locked;
2603 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2604 * @hb: the futex hash bucket, must be locked by the caller
2605 * @q: the futex_q to queue up on
2606 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2608 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2609 struct hrtimer_sleeper *timeout)
2612 * The task state is guaranteed to be set before another task can
2613 * wake it. set_current_state() is implemented using smp_store_mb() and
2614 * queue_me() calls spin_unlock() upon completion, both serializing
2615 * access to the hash list and forcing another memory barrier.
2617 set_current_state(TASK_INTERRUPTIBLE);
2622 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2625 * If we have been removed from the hash list, then another task
2626 * has tried to wake us, and we can skip the call to schedule().
2628 if (likely(!plist_node_empty(&q->list))) {
2630 * If the timer has already expired, current will already be
2631 * flagged for rescheduling. Only call schedule if there
2632 * is no timeout, or if it has yet to expire.
2634 if (!timeout || timeout->task)
2635 freezable_schedule();
2637 __set_current_state(TASK_RUNNING);
2641 * futex_wait_setup() - Prepare to wait on a futex
2642 * @uaddr: the futex userspace address
2643 * @val: the expected value
2644 * @flags: futex flags (FLAGS_SHARED, etc.)
2645 * @q: the associated futex_q
2646 * @hb: storage for hash_bucket pointer to be returned to caller
2648 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2649 * compare it with the expected value. Handle atomic faults internally.
2650 * Return with the hb lock held and a q.key reference on success, and unlocked
2651 * with no q.key reference on failure.
2654 * - 0 - uaddr contains val and hb has been locked;
2655 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2657 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2658 struct futex_q *q, struct futex_hash_bucket **hb)
2664 * Access the page AFTER the hash-bucket is locked.
2665 * Order is important:
2667 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2668 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2670 * The basic logical guarantee of a futex is that it blocks ONLY
2671 * if cond(var) is known to be true at the time of blocking, for
2672 * any cond. If we locked the hash-bucket after testing *uaddr, that
2673 * would open a race condition where we could block indefinitely with
2674 * cond(var) false, which would violate the guarantee.
2676 * On the other hand, we insert q and release the hash-bucket only
2677 * after testing *uaddr. This guarantees that futex_wait() will NOT
2678 * absorb a wakeup if *uaddr does not match the desired values
2679 * while the syscall executes.
2682 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2683 if (unlikely(ret != 0))
2687 *hb = queue_lock(q);
2689 ret = get_futex_value_locked(&uval, uaddr);
2694 ret = get_user(uval, uaddr);
2698 if (!(flags & FLAGS_SHARED))
2701 put_futex_key(&q->key);
2712 put_futex_key(&q->key);
2716 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2717 ktime_t *abs_time, u32 bitset)
2719 struct hrtimer_sleeper timeout, *to;
2720 struct restart_block *restart;
2721 struct futex_hash_bucket *hb;
2722 struct futex_q q = futex_q_init;
2729 to = futex_setup_timer(abs_time, &timeout, flags,
2730 current->timer_slack_ns);
2733 * Prepare to wait on uaddr. On success, holds hb lock and increments
2736 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2740 /* queue_me and wait for wakeup, timeout, or a signal. */
2741 futex_wait_queue_me(hb, &q, to);
2743 /* If we were woken (and unqueued), we succeeded, whatever. */
2745 /* unqueue_me() drops q.key ref */
2746 if (!unqueue_me(&q))
2749 if (to && !to->task)
2753 * We expect signal_pending(current), but we might be the
2754 * victim of a spurious wakeup as well.
2756 if (!signal_pending(current))
2763 restart = ¤t->restart_block;
2764 restart->fn = futex_wait_restart;
2765 restart->futex.uaddr = uaddr;
2766 restart->futex.val = val;
2767 restart->futex.time = *abs_time;
2768 restart->futex.bitset = bitset;
2769 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2771 ret = -ERESTART_RESTARTBLOCK;
2775 hrtimer_cancel(&to->timer);
2776 destroy_hrtimer_on_stack(&to->timer);
2782 static long futex_wait_restart(struct restart_block *restart)
2784 u32 __user *uaddr = restart->futex.uaddr;
2785 ktime_t t, *tp = NULL;
2787 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2788 t = restart->futex.time;
2791 restart->fn = do_no_restart_syscall;
2793 return (long)futex_wait(uaddr, restart->futex.flags,
2794 restart->futex.val, tp, restart->futex.bitset);
2799 * Userspace tried a 0 -> TID atomic transition of the futex value
2800 * and failed. The kernel side here does the whole locking operation:
2801 * if there are waiters then it will block as a consequence of relying
2802 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2803 * a 0 value of the futex too.).
2805 * Also serves as futex trylock_pi()'ing, and due semantics.
2807 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2808 ktime_t *time, int trylock)
2810 struct hrtimer_sleeper timeout, *to;
2811 struct futex_pi_state *pi_state = NULL;
2812 struct rt_mutex_waiter rt_waiter;
2813 struct futex_hash_bucket *hb;
2814 struct futex_q q = futex_q_init;
2817 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2820 if (refill_pi_state_cache())
2823 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2826 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2827 if (unlikely(ret != 0))
2831 hb = queue_lock(&q);
2833 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2834 if (unlikely(ret)) {
2836 * Atomic work succeeded and we got the lock,
2837 * or failed. Either way, we do _not_ block.
2841 /* We got the lock. */
2843 goto out_unlock_put_key;
2848 * Two reasons for this:
2849 * - Task is exiting and we just wait for the
2851 * - The user space value changed.
2854 put_futex_key(&q.key);
2858 goto out_unlock_put_key;
2862 WARN_ON(!q.pi_state);
2865 * Only actually queue now that the atomic ops are done:
2870 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2871 /* Fixup the trylock return value: */
2872 ret = ret ? 0 : -EWOULDBLOCK;
2876 rt_mutex_init_waiter(&rt_waiter);
2879 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2880 * hold it while doing rt_mutex_start_proxy(), because then it will
2881 * include hb->lock in the blocking chain, even through we'll not in
2882 * fact hold it while blocking. This will lead it to report -EDEADLK
2883 * and BUG when futex_unlock_pi() interleaves with this.
2885 * Therefore acquire wait_lock while holding hb->lock, but drop the
2886 * latter before calling __rt_mutex_start_proxy_lock(). This
2887 * interleaves with futex_unlock_pi() -- which does a similar lock
2888 * handoff -- such that the latter can observe the futex_q::pi_state
2889 * before __rt_mutex_start_proxy_lock() is done.
2891 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2892 spin_unlock(q.lock_ptr);
2894 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2895 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2896 * it sees the futex_q::pi_state.
2898 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2899 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2908 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2910 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2913 spin_lock(q.lock_ptr);
2915 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2916 * first acquire the hb->lock before removing the lock from the
2917 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2920 * In particular; it is important that futex_unlock_pi() can not
2921 * observe this inconsistency.
2923 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2928 * Fixup the pi_state owner and possibly acquire the lock if we
2931 res = fixup_owner(uaddr, &q, !ret);
2933 * If fixup_owner() returned an error, proprogate that. If it acquired
2934 * the lock, clear our -ETIMEDOUT or -EINTR.
2937 ret = (res < 0) ? res : 0;
2940 * If fixup_owner() faulted and was unable to handle the fault, unlock
2941 * it and return the fault to userspace.
2943 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2944 pi_state = q.pi_state;
2945 get_pi_state(pi_state);
2948 /* Unqueue and drop the lock */
2952 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2953 put_pi_state(pi_state);
2962 put_futex_key(&q.key);
2965 hrtimer_cancel(&to->timer);
2966 destroy_hrtimer_on_stack(&to->timer);
2968 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2973 ret = fault_in_user_writeable(uaddr);
2977 if (!(flags & FLAGS_SHARED))
2980 put_futex_key(&q.key);
2985 * Userspace attempted a TID -> 0 atomic transition, and failed.
2986 * This is the in-kernel slowpath: we look up the PI state (if any),
2987 * and do the rt-mutex unlock.
2989 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2991 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2992 union futex_key key = FUTEX_KEY_INIT;
2993 struct futex_hash_bucket *hb;
2994 struct futex_q *top_waiter;
2997 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3001 if (get_user(uval, uaddr))
3004 * We release only a lock we actually own:
3006 if ((uval & FUTEX_TID_MASK) != vpid)
3009 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3013 hb = hash_futex(&key);
3014 spin_lock(&hb->lock);
3017 * Check waiters first. We do not trust user space values at
3018 * all and we at least want to know if user space fiddled
3019 * with the futex value instead of blindly unlocking.
3021 top_waiter = futex_top_waiter(hb, &key);
3023 struct futex_pi_state *pi_state = top_waiter->pi_state;
3030 * If current does not own the pi_state then the futex is
3031 * inconsistent and user space fiddled with the futex value.
3033 if (pi_state->owner != current)
3036 get_pi_state(pi_state);
3038 * By taking wait_lock while still holding hb->lock, we ensure
3039 * there is no point where we hold neither; and therefore
3040 * wake_futex_pi() must observe a state consistent with what we
3043 * In particular; this forces __rt_mutex_start_proxy() to
3044 * complete such that we're guaranteed to observe the
3045 * rt_waiter. Also see the WARN in wake_futex_pi().
3047 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3048 spin_unlock(&hb->lock);
3050 /* drops pi_state->pi_mutex.wait_lock */
3051 ret = wake_futex_pi(uaddr, uval, pi_state);
3053 put_pi_state(pi_state);
3056 * Success, we're done! No tricky corner cases.
3061 * The atomic access to the futex value generated a
3062 * pagefault, so retry the user-access and the wakeup:
3067 * A unconditional UNLOCK_PI op raced against a waiter
3068 * setting the FUTEX_WAITERS bit. Try again.
3073 * wake_futex_pi has detected invalid state. Tell user
3080 * We have no kernel internal state, i.e. no waiters in the
3081 * kernel. Waiters which are about to queue themselves are stuck
3082 * on hb->lock. So we can safely ignore them. We do neither
3083 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3086 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3087 spin_unlock(&hb->lock);
3102 * If uval has changed, let user space handle it.
3104 ret = (curval == uval) ? 0 : -EAGAIN;
3107 spin_unlock(&hb->lock);
3109 put_futex_key(&key);
3113 put_futex_key(&key);
3118 put_futex_key(&key);
3120 ret = fault_in_user_writeable(uaddr);
3128 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3129 * @hb: the hash_bucket futex_q was original enqueued on
3130 * @q: the futex_q woken while waiting to be requeued
3131 * @key2: the futex_key of the requeue target futex
3132 * @timeout: the timeout associated with the wait (NULL if none)
3134 * Detect if the task was woken on the initial futex as opposed to the requeue
3135 * target futex. If so, determine if it was a timeout or a signal that caused
3136 * the wakeup and return the appropriate error code to the caller. Must be
3137 * called with the hb lock held.
3140 * - 0 = no early wakeup detected;
3141 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3144 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3145 struct futex_q *q, union futex_key *key2,
3146 struct hrtimer_sleeper *timeout)
3151 * With the hb lock held, we avoid races while we process the wakeup.
3152 * We only need to hold hb (and not hb2) to ensure atomicity as the
3153 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3154 * It can't be requeued from uaddr2 to something else since we don't
3155 * support a PI aware source futex for requeue.
3157 if (!match_futex(&q->key, key2)) {
3158 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3160 * We were woken prior to requeue by a timeout or a signal.
3161 * Unqueue the futex_q and determine which it was.
3163 plist_del(&q->list, &hb->chain);
3166 /* Handle spurious wakeups gracefully */
3168 if (timeout && !timeout->task)
3170 else if (signal_pending(current))
3171 ret = -ERESTARTNOINTR;
3177 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3178 * @uaddr: the futex we initially wait on (non-pi)
3179 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3180 * the same type, no requeueing from private to shared, etc.
3181 * @val: the expected value of uaddr
3182 * @abs_time: absolute timeout
3183 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3184 * @uaddr2: the pi futex we will take prior to returning to user-space
3186 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3187 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3188 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3189 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3190 * without one, the pi logic would not know which task to boost/deboost, if
3191 * there was a need to.
3193 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3194 * via the following--
3195 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3196 * 2) wakeup on uaddr2 after a requeue
3200 * If 3, cleanup and return -ERESTARTNOINTR.
3202 * If 2, we may then block on trying to take the rt_mutex and return via:
3203 * 5) successful lock
3206 * 8) other lock acquisition failure
3208 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3210 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3216 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3217 u32 val, ktime_t *abs_time, u32 bitset,
3220 struct hrtimer_sleeper timeout, *to;
3221 struct futex_pi_state *pi_state = NULL;
3222 struct rt_mutex_waiter rt_waiter;
3223 struct futex_hash_bucket *hb;
3224 union futex_key key2 = FUTEX_KEY_INIT;
3225 struct futex_q q = futex_q_init;
3228 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3231 if (uaddr == uaddr2)
3237 to = futex_setup_timer(abs_time, &timeout, flags,
3238 current->timer_slack_ns);
3241 * The waiter is allocated on our stack, manipulated by the requeue
3242 * code while we sleep on uaddr.
3244 rt_mutex_init_waiter(&rt_waiter);
3246 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3247 if (unlikely(ret != 0))
3251 q.rt_waiter = &rt_waiter;
3252 q.requeue_pi_key = &key2;
3255 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3258 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3263 * The check above which compares uaddrs is not sufficient for
3264 * shared futexes. We need to compare the keys:
3266 if (match_futex(&q.key, &key2)) {
3272 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3273 futex_wait_queue_me(hb, &q, to);
3275 spin_lock(&hb->lock);
3276 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3277 spin_unlock(&hb->lock);
3282 * In order for us to be here, we know our q.key == key2, and since
3283 * we took the hb->lock above, we also know that futex_requeue() has
3284 * completed and we no longer have to concern ourselves with a wakeup
3285 * race with the atomic proxy lock acquisition by the requeue code. The
3286 * futex_requeue dropped our key1 reference and incremented our key2
3290 /* Check if the requeue code acquired the second futex for us. */
3293 * Got the lock. We might not be the anticipated owner if we
3294 * did a lock-steal - fix up the PI-state in that case.
3296 if (q.pi_state && (q.pi_state->owner != current)) {
3297 spin_lock(q.lock_ptr);
3298 ret = fixup_pi_state_owner(uaddr2, &q, current);
3299 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3300 pi_state = q.pi_state;
3301 get_pi_state(pi_state);
3304 * Drop the reference to the pi state which
3305 * the requeue_pi() code acquired for us.
3307 put_pi_state(q.pi_state);
3308 spin_unlock(q.lock_ptr);
3311 struct rt_mutex *pi_mutex;
3314 * We have been woken up by futex_unlock_pi(), a timeout, or a
3315 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3318 WARN_ON(!q.pi_state);
3319 pi_mutex = &q.pi_state->pi_mutex;
3320 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3322 spin_lock(q.lock_ptr);
3323 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3326 debug_rt_mutex_free_waiter(&rt_waiter);
3328 * Fixup the pi_state owner and possibly acquire the lock if we
3331 res = fixup_owner(uaddr2, &q, !ret);
3333 * If fixup_owner() returned an error, proprogate that. If it
3334 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3337 ret = (res < 0) ? res : 0;
3340 * If fixup_pi_state_owner() faulted and was unable to handle
3341 * the fault, unlock the rt_mutex and return the fault to
3344 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3345 pi_state = q.pi_state;
3346 get_pi_state(pi_state);
3349 /* Unqueue and drop the lock. */
3354 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3355 put_pi_state(pi_state);
3358 if (ret == -EINTR) {
3360 * We've already been requeued, but cannot restart by calling
3361 * futex_lock_pi() directly. We could restart this syscall, but
3362 * it would detect that the user space "val" changed and return
3363 * -EWOULDBLOCK. Save the overhead of the restart and return
3364 * -EWOULDBLOCK directly.
3370 put_futex_key(&q.key);
3372 put_futex_key(&key2);
3376 hrtimer_cancel(&to->timer);
3377 destroy_hrtimer_on_stack(&to->timer);
3383 * Support for robust futexes: the kernel cleans up held futexes at
3386 * Implementation: user-space maintains a per-thread list of locks it
3387 * is holding. Upon do_exit(), the kernel carefully walks this list,
3388 * and marks all locks that are owned by this thread with the
3389 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3390 * always manipulated with the lock held, so the list is private and
3391 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3392 * field, to allow the kernel to clean up if the thread dies after
3393 * acquiring the lock, but just before it could have added itself to
3394 * the list. There can only be one such pending lock.
3398 * sys_set_robust_list() - Set the robust-futex list head of a task
3399 * @head: pointer to the list-head
3400 * @len: length of the list-head, as userspace expects
3402 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3405 if (!futex_cmpxchg_enabled)
3408 * The kernel knows only one size for now:
3410 if (unlikely(len != sizeof(*head)))
3413 current->robust_list = head;
3419 * sys_get_robust_list() - Get the robust-futex list head of a task
3420 * @pid: pid of the process [zero for current task]
3421 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3422 * @len_ptr: pointer to a length field, the kernel fills in the header size
3424 SYSCALL_DEFINE3(get_robust_list, int, pid,
3425 struct robust_list_head __user * __user *, head_ptr,
3426 size_t __user *, len_ptr)
3428 struct robust_list_head __user *head;
3430 struct task_struct *p;
3432 if (!futex_cmpxchg_enabled)
3441 p = find_task_by_vpid(pid);
3447 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3450 head = p->robust_list;
3453 if (put_user(sizeof(*head), len_ptr))
3455 return put_user(head, head_ptr);
3463 /* Constants for the pending_op argument of handle_futex_death */
3464 #define HANDLE_DEATH_PENDING true
3465 #define HANDLE_DEATH_LIST false
3468 * Process a futex-list entry, check whether it's owned by the
3469 * dying task, and do notification if so:
3471 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3472 bool pi, bool pending_op)
3474 u32 uval, uninitialized_var(nval), mval;
3477 /* Futex address must be 32bit aligned */
3478 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3482 if (get_user(uval, uaddr))
3486 * Special case for regular (non PI) futexes. The unlock path in
3487 * user space has two race scenarios:
3489 * 1. The unlock path releases the user space futex value and
3490 * before it can execute the futex() syscall to wake up
3491 * waiters it is killed.
3493 * 2. A woken up waiter is killed before it can acquire the
3494 * futex in user space.
3496 * In both cases the TID validation below prevents a wakeup of
3497 * potential waiters which can cause these waiters to block
3500 * In both cases the following conditions are met:
3502 * 1) task->robust_list->list_op_pending != NULL
3503 * @pending_op == true
3504 * 2) User space futex value == 0
3505 * 3) Regular futex: @pi == false
3507 * If these conditions are met, it is safe to attempt waking up a
3508 * potential waiter without touching the user space futex value and
3509 * trying to set the OWNER_DIED bit. The user space futex value is
3510 * uncontended and the rest of the user space mutex state is
3511 * consistent, so a woken waiter will just take over the
3512 * uncontended futex. Setting the OWNER_DIED bit would create
3513 * inconsistent state and malfunction of the user space owner died
3516 if (pending_op && !pi && !uval) {
3517 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3521 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3525 * Ok, this dying thread is truly holding a futex
3526 * of interest. Set the OWNER_DIED bit atomically
3527 * via cmpxchg, and if the value had FUTEX_WAITERS
3528 * set, wake up a waiter (if any). (We have to do a
3529 * futex_wake() even if OWNER_DIED is already set -
3530 * to handle the rare but possible case of recursive
3531 * thread-death.) The rest of the cleanup is done in
3534 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3537 * We are not holding a lock here, but we want to have
3538 * the pagefault_disable/enable() protection because
3539 * we want to handle the fault gracefully. If the
3540 * access fails we try to fault in the futex with R/W
3541 * verification via get_user_pages. get_user() above
3542 * does not guarantee R/W access. If that fails we
3543 * give up and leave the futex locked.
3545 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3548 if (fault_in_user_writeable(uaddr))
3566 * Wake robust non-PI futexes here. The wakeup of
3567 * PI futexes happens in exit_pi_state():
3569 if (!pi && (uval & FUTEX_WAITERS))
3570 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3576 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3578 static inline int fetch_robust_entry(struct robust_list __user **entry,
3579 struct robust_list __user * __user *head,
3582 unsigned long uentry;
3584 if (get_user(uentry, (unsigned long __user *)head))
3587 *entry = (void __user *)(uentry & ~1UL);
3594 * Walk curr->robust_list (very carefully, it's a userspace list!)
3595 * and mark any locks found there dead, and notify any waiters.
3597 * We silently return on any sign of list-walking problem.
3599 static void exit_robust_list(struct task_struct *curr)
3601 struct robust_list_head __user *head = curr->robust_list;
3602 struct robust_list __user *entry, *next_entry, *pending;
3603 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3604 unsigned int uninitialized_var(next_pi);
3605 unsigned long futex_offset;
3608 if (!futex_cmpxchg_enabled)
3612 * Fetch the list head (which was registered earlier, via
3613 * sys_set_robust_list()):
3615 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3618 * Fetch the relative futex offset:
3620 if (get_user(futex_offset, &head->futex_offset))
3623 * Fetch any possibly pending lock-add first, and handle it
3626 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3629 next_entry = NULL; /* avoid warning with gcc */
3630 while (entry != &head->list) {
3632 * Fetch the next entry in the list before calling
3633 * handle_futex_death:
3635 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3637 * A pending lock might already be on the list, so
3638 * don't process it twice:
3640 if (entry != pending) {
3641 if (handle_futex_death((void __user *)entry + futex_offset,
3642 curr, pi, HANDLE_DEATH_LIST))
3650 * Avoid excessively long or circular lists:
3659 handle_futex_death((void __user *)pending + futex_offset,
3660 curr, pip, HANDLE_DEATH_PENDING);
3664 void futex_mm_release(struct task_struct *tsk)
3666 if (unlikely(tsk->robust_list)) {
3667 exit_robust_list(tsk);
3668 tsk->robust_list = NULL;
3671 #ifdef CONFIG_COMPAT
3672 if (unlikely(tsk->compat_robust_list)) {
3673 compat_exit_robust_list(tsk);
3674 tsk->compat_robust_list = NULL;
3678 if (unlikely(!list_empty(&tsk->pi_state_list)))
3679 exit_pi_state_list(tsk);
3682 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3683 u32 __user *uaddr2, u32 val2, u32 val3)
3685 int cmd = op & FUTEX_CMD_MASK;
3686 unsigned int flags = 0;
3688 if (!(op & FUTEX_PRIVATE_FLAG))
3689 flags |= FLAGS_SHARED;
3691 if (op & FUTEX_CLOCK_REALTIME) {
3692 flags |= FLAGS_CLOCKRT;
3693 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3694 cmd != FUTEX_WAIT_REQUEUE_PI)
3700 case FUTEX_UNLOCK_PI:
3701 case FUTEX_TRYLOCK_PI:
3702 case FUTEX_WAIT_REQUEUE_PI:
3703 case FUTEX_CMP_REQUEUE_PI:
3704 if (!futex_cmpxchg_enabled)
3710 val3 = FUTEX_BITSET_MATCH_ANY;
3712 case FUTEX_WAIT_BITSET:
3713 return futex_wait(uaddr, flags, val, timeout, val3);
3715 val3 = FUTEX_BITSET_MATCH_ANY;
3717 case FUTEX_WAKE_BITSET:
3718 return futex_wake(uaddr, flags, val, val3);
3720 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3721 case FUTEX_CMP_REQUEUE:
3722 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3724 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3726 return futex_lock_pi(uaddr, flags, timeout, 0);
3727 case FUTEX_UNLOCK_PI:
3728 return futex_unlock_pi(uaddr, flags);
3729 case FUTEX_TRYLOCK_PI:
3730 return futex_lock_pi(uaddr, flags, NULL, 1);
3731 case FUTEX_WAIT_REQUEUE_PI:
3732 val3 = FUTEX_BITSET_MATCH_ANY;
3733 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3735 case FUTEX_CMP_REQUEUE_PI:
3736 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3742 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3743 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3746 struct timespec64 ts;
3747 ktime_t t, *tp = NULL;
3749 int cmd = op & FUTEX_CMD_MASK;
3751 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3752 cmd == FUTEX_WAIT_BITSET ||
3753 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3754 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3756 if (get_timespec64(&ts, utime))
3758 if (!timespec64_valid(&ts))
3761 t = timespec64_to_ktime(ts);
3762 if (cmd == FUTEX_WAIT)
3763 t = ktime_add_safe(ktime_get(), t);
3767 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3768 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3770 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3771 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3772 val2 = (u32) (unsigned long) utime;
3774 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3777 #ifdef CONFIG_COMPAT
3779 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3782 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3783 compat_uptr_t __user *head, unsigned int *pi)
3785 if (get_user(*uentry, head))
3788 *entry = compat_ptr((*uentry) & ~1);
3789 *pi = (unsigned int)(*uentry) & 1;
3794 static void __user *futex_uaddr(struct robust_list __user *entry,
3795 compat_long_t futex_offset)
3797 compat_uptr_t base = ptr_to_compat(entry);
3798 void __user *uaddr = compat_ptr(base + futex_offset);
3804 * Walk curr->robust_list (very carefully, it's a userspace list!)
3805 * and mark any locks found there dead, and notify any waiters.
3807 * We silently return on any sign of list-walking problem.
3809 static void compat_exit_robust_list(struct task_struct *curr)
3811 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3812 struct robust_list __user *entry, *next_entry, *pending;
3813 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3814 unsigned int uninitialized_var(next_pi);
3815 compat_uptr_t uentry, next_uentry, upending;
3816 compat_long_t futex_offset;
3819 if (!futex_cmpxchg_enabled)
3823 * Fetch the list head (which was registered earlier, via
3824 * sys_set_robust_list()):
3826 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3829 * Fetch the relative futex offset:
3831 if (get_user(futex_offset, &head->futex_offset))
3834 * Fetch any possibly pending lock-add first, and handle it
3837 if (compat_fetch_robust_entry(&upending, &pending,
3838 &head->list_op_pending, &pip))
3841 next_entry = NULL; /* avoid warning with gcc */
3842 while (entry != (struct robust_list __user *) &head->list) {
3844 * Fetch the next entry in the list before calling
3845 * handle_futex_death:
3847 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3848 (compat_uptr_t __user *)&entry->next, &next_pi);
3850 * A pending lock might already be on the list, so
3851 * dont process it twice:
3853 if (entry != pending) {
3854 void __user *uaddr = futex_uaddr(entry, futex_offset);
3856 if (handle_futex_death(uaddr, curr, pi,
3862 uentry = next_uentry;
3866 * Avoid excessively long or circular lists:
3874 void __user *uaddr = futex_uaddr(pending, futex_offset);
3876 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3880 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3881 struct compat_robust_list_head __user *, head,
3884 if (!futex_cmpxchg_enabled)
3887 if (unlikely(len != sizeof(*head)))
3890 current->compat_robust_list = head;
3895 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3896 compat_uptr_t __user *, head_ptr,
3897 compat_size_t __user *, len_ptr)
3899 struct compat_robust_list_head __user *head;
3901 struct task_struct *p;
3903 if (!futex_cmpxchg_enabled)
3912 p = find_task_by_vpid(pid);
3918 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3921 head = p->compat_robust_list;
3924 if (put_user(sizeof(*head), len_ptr))
3926 return put_user(ptr_to_compat(head), head_ptr);
3933 #endif /* CONFIG_COMPAT */
3935 #ifdef CONFIG_COMPAT_32BIT_TIME
3936 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3937 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3940 struct timespec64 ts;
3941 ktime_t t, *tp = NULL;
3943 int cmd = op & FUTEX_CMD_MASK;
3945 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3946 cmd == FUTEX_WAIT_BITSET ||
3947 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3948 if (get_old_timespec32(&ts, utime))
3950 if (!timespec64_valid(&ts))
3953 t = timespec64_to_ktime(ts);
3954 if (cmd == FUTEX_WAIT)
3955 t = ktime_add_safe(ktime_get(), t);
3958 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3959 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3960 val2 = (int) (unsigned long) utime;
3962 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3964 #endif /* CONFIG_COMPAT_32BIT_TIME */
3966 static void __init futex_detect_cmpxchg(void)
3968 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3972 * This will fail and we want it. Some arch implementations do
3973 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3974 * functionality. We want to know that before we call in any
3975 * of the complex code paths. Also we want to prevent
3976 * registration of robust lists in that case. NULL is
3977 * guaranteed to fault and we get -EFAULT on functional
3978 * implementation, the non-functional ones will return
3981 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3982 futex_cmpxchg_enabled = 1;
3986 static int __init futex_init(void)
3988 unsigned int futex_shift;
3991 #if CONFIG_BASE_SMALL
3992 futex_hashsize = 16;
3994 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3997 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3999 futex_hashsize < 256 ? HASH_SMALL : 0,
4001 futex_hashsize, futex_hashsize);
4002 futex_hashsize = 1UL << futex_shift;
4004 futex_detect_cmpxchg();
4006 for (i = 0; i < futex_hashsize; i++) {
4007 atomic_set(&futex_queues[i].waiters, 0);
4008 plist_head_init(&futex_queues[i].chain);
4009 spin_lock_init(&futex_queues[i].lock);
4014 core_initcall(futex_init);