2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/bootmem.h>
70 #include <linux/fault-inject.h>
72 #include <asm/futex.h>
74 #include "locking/rtmutex_common.h"
77 * READ this before attempting to hack on futexes!
79 * Basic futex operation and ordering guarantees
80 * =============================================
82 * The waiter reads the futex value in user space and calls
83 * futex_wait(). This function computes the hash bucket and acquires
84 * the hash bucket lock. After that it reads the futex user space value
85 * again and verifies that the data has not changed. If it has not changed
86 * it enqueues itself into the hash bucket, releases the hash bucket lock
89 * The waker side modifies the user space value of the futex and calls
90 * futex_wake(). This function computes the hash bucket and acquires the
91 * hash bucket lock. Then it looks for waiters on that futex in the hash
92 * bucket and wakes them.
94 * In futex wake up scenarios where no tasks are blocked on a futex, taking
95 * the hb spinlock can be avoided and simply return. In order for this
96 * optimization to work, ordering guarantees must exist so that the waiter
97 * being added to the list is acknowledged when the list is concurrently being
98 * checked by the waker, avoiding scenarios like the following:
102 * sys_futex(WAIT, futex, val);
103 * futex_wait(futex, val);
106 * sys_futex(WAKE, futex);
111 * lock(hash_bucket(futex));
113 * unlock(hash_bucket(futex));
116 * This would cause the waiter on CPU 0 to wait forever because it
117 * missed the transition of the user space value from val to newval
118 * and the waker did not find the waiter in the hash bucket queue.
120 * The correct serialization ensures that a waiter either observes
121 * the changed user space value before blocking or is woken by a
126 * sys_futex(WAIT, futex, val);
127 * futex_wait(futex, val);
130 * smp_mb(); (A) <-- paired with -.
132 * lock(hash_bucket(futex)); |
136 * | sys_futex(WAKE, futex);
137 * | futex_wake(futex);
139 * `--------> smp_mb(); (B)
142 * unlock(hash_bucket(futex));
143 * schedule(); if (waiters)
144 * lock(hash_bucket(futex));
145 * else wake_waiters(futex);
146 * waiters--; (b) unlock(hash_bucket(futex));
148 * Where (A) orders the waiters increment and the futex value read through
149 * atomic operations (see hb_waiters_inc) and where (B) orders the write
150 * to futex and the waiters read -- this is done by the barriers for both
151 * shared and private futexes in get_futex_key_refs().
153 * This yields the following case (where X:=waiters, Y:=futex):
161 * Which guarantees that x==0 && y==0 is impossible; which translates back into
162 * the guarantee that we cannot both miss the futex variable change and the
165 * Note that a new waiter is accounted for in (a) even when it is possible that
166 * the wait call can return error, in which case we backtrack from it in (b).
167 * Refer to the comment in queue_lock().
169 * Similarly, in order to account for waiters being requeued on another
170 * address we always increment the waiters for the destination bucket before
171 * acquiring the lock. It then decrements them again after releasing it -
172 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173 * will do the additional required waiter count housekeeping. This is done for
174 * double_lock_hb() and double_unlock_hb(), respectively.
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
180 static int __read_mostly futex_cmpxchg_enabled;
184 * Futex flags used to encode options to functions and preserve them across
188 # define FLAGS_SHARED 0x01
191 * NOMMU does not have per process address space. Let the compiler optimize
194 # define FLAGS_SHARED 0x00
196 #define FLAGS_CLOCKRT 0x02
197 #define FLAGS_HAS_TIMEOUT 0x04
200 * Priority Inheritance state:
202 struct futex_pi_state {
204 * list of 'owned' pi_state instances - these have to be
205 * cleaned up in do_exit() if the task exits prematurely:
207 struct list_head list;
212 struct rt_mutex pi_mutex;
214 struct task_struct *owner;
218 } __randomize_layout;
221 * struct futex_q - The hashed futex queue entry, one per waiting task
222 * @list: priority-sorted list of tasks waiting on this futex
223 * @task: the task waiting on the futex
224 * @lock_ptr: the hash bucket lock
225 * @key: the key the futex is hashed on
226 * @pi_state: optional priority inheritance state
227 * @rt_waiter: rt_waiter storage for use with requeue_pi
228 * @requeue_pi_key: the requeue_pi target futex key
229 * @bitset: bitset for the optional bitmasked wakeup
231 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232 * we can wake only the relevant ones (hashed queues may be shared).
234 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236 * The order of wakeup is always to make the first condition true, then
239 * PI futexes are typically woken before they are removed from the hash list via
240 * the rt_mutex code. See unqueue_me_pi().
243 struct plist_node list;
245 struct task_struct *task;
246 spinlock_t *lock_ptr;
248 struct futex_pi_state *pi_state;
249 struct rt_mutex_waiter *rt_waiter;
250 union futex_key *requeue_pi_key;
252 } __randomize_layout;
254 static const struct futex_q futex_q_init = {
255 /* list gets initialized in queue_me()*/
256 .key = FUTEX_KEY_INIT,
257 .bitset = FUTEX_BITSET_MATCH_ANY
261 * Hash buckets are shared by all the futex_keys that hash to the same
262 * location. Each key may have multiple futex_q structures, one for each task
263 * waiting on a futex.
265 struct futex_hash_bucket {
268 struct plist_head chain;
269 } ____cacheline_aligned_in_smp;
272 * The base of the bucket array and its size are always used together
273 * (after initialization only in hash_futex()), so ensure that they
274 * reside in the same cacheline.
277 struct futex_hash_bucket *queues;
278 unsigned long hashsize;
279 } __futex_data __read_mostly __aligned(2*sizeof(long));
280 #define futex_queues (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
285 * Fault injections for futexes.
287 #ifdef CONFIG_FAIL_FUTEX
290 struct fault_attr attr;
294 .attr = FAULT_ATTR_INITIALIZER,
295 .ignore_private = false,
298 static int __init setup_fail_futex(char *str)
300 return setup_fault_attr(&fail_futex.attr, str);
302 __setup("fail_futex=", setup_fail_futex);
304 static bool should_fail_futex(bool fshared)
306 if (fail_futex.ignore_private && !fshared)
309 return should_fail(&fail_futex.attr, 1);
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314 static int __init fail_futex_debugfs(void)
316 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
319 dir = fault_create_debugfs_attr("fail_futex", NULL,
324 if (!debugfs_create_bool("ignore-private", mode, dir,
325 &fail_futex.ignore_private)) {
326 debugfs_remove_recursive(dir);
333 late_initcall(fail_futex_debugfs);
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
338 static inline bool should_fail_futex(bool fshared)
342 #endif /* CONFIG_FAIL_FUTEX */
345 static void compat_exit_robust_list(struct task_struct *curr);
347 static inline void compat_exit_robust_list(struct task_struct *curr) { }
350 static inline void futex_get_mm(union futex_key *key)
352 mmgrab(key->private.mm);
354 * Ensure futex_get_mm() implies a full barrier such that
355 * get_futex_key() implies a full barrier. This is relied upon
356 * as smp_mb(); (B), see the ordering comment above.
358 smp_mb__after_atomic();
362 * Reflects a new waiter being added to the waitqueue.
364 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
367 atomic_inc(&hb->waiters);
369 * Full barrier (A), see the ordering comment above.
371 smp_mb__after_atomic();
376 * Reflects a waiter being removed from the waitqueue by wakeup
379 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
382 atomic_dec(&hb->waiters);
386 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
389 return atomic_read(&hb->waiters);
396 * hash_futex - Return the hash bucket in the global hash
397 * @key: Pointer to the futex key for which the hash is calculated
399 * We hash on the keys returned from get_futex_key (see below) and return the
400 * corresponding hash bucket in the global hash.
402 static struct futex_hash_bucket *hash_futex(union futex_key *key)
404 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
407 return &futex_queues[hash & (futex_hashsize - 1)];
412 * match_futex - Check whether two futex keys are equal
413 * @key1: Pointer to key1
414 * @key2: Pointer to key2
416 * Return 1 if two futex_keys are equal, 0 otherwise.
418 static inline int match_futex(union futex_key *key1, union futex_key *key2)
421 && key1->both.word == key2->both.word
422 && key1->both.ptr == key2->both.ptr
423 && key1->both.offset == key2->both.offset);
427 * Take a reference to the resource addressed by a key.
428 * Can be called while holding spinlocks.
431 static void get_futex_key_refs(union futex_key *key)
437 * On MMU less systems futexes are always "private" as there is no per
438 * process address space. We need the smp wmb nevertheless - yes,
439 * arch/blackfin has MMU less SMP ...
441 if (!IS_ENABLED(CONFIG_MMU)) {
442 smp_mb(); /* explicit smp_mb(); (B) */
446 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
448 smp_mb(); /* explicit smp_mb(); (B) */
450 case FUT_OFF_MMSHARED:
451 futex_get_mm(key); /* implies smp_mb(); (B) */
455 * Private futexes do not hold reference on an inode or
456 * mm, therefore the only purpose of calling get_futex_key_refs
457 * is because we need the barrier for the lockless waiter check.
459 smp_mb(); /* explicit smp_mb(); (B) */
464 * Drop a reference to the resource addressed by a key.
465 * The hash bucket spinlock must not be held. This is
466 * a no-op for private futexes, see comment in the get
469 static void drop_futex_key_refs(union futex_key *key)
471 if (!key->both.ptr) {
472 /* If we're here then we tried to put a key we failed to get */
477 if (!IS_ENABLED(CONFIG_MMU))
480 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
483 case FUT_OFF_MMSHARED:
484 mmdrop(key->private.mm);
490 * Generate a machine wide unique identifier for this inode.
492 * This relies on u64 not wrapping in the life-time of the machine; which with
493 * 1ns resolution means almost 585 years.
495 * This further relies on the fact that a well formed program will not unmap
496 * the file while it has a (shared) futex waiting on it. This mapping will have
497 * a file reference which pins the mount and inode.
499 * If for some reason an inode gets evicted and read back in again, it will get
500 * a new sequence number and will _NOT_ match, even though it is the exact same
503 * It is important that match_futex() will never have a false-positive, esp.
504 * for PI futexes that can mess up the state. The above argues that false-negatives
505 * are only possible for malformed programs.
507 static u64 get_inode_sequence_number(struct inode *inode)
509 static atomic64_t i_seq;
512 /* Does the inode already have a sequence number? */
513 old = atomic64_read(&inode->i_sequence);
518 u64 new = atomic64_add_return(1, &i_seq);
519 if (WARN_ON_ONCE(!new))
522 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
530 * get_futex_key() - Get parameters which are the keys for a futex
531 * @uaddr: virtual address of the futex
532 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
533 * @key: address where result is stored.
534 * @rw: mapping needs to be read/write (values: VERIFY_READ,
537 * Return: a negative error code or 0
539 * The key words are stored in @key on success.
541 * For shared mappings (when @fshared), the key is:
542 * ( inode->i_sequence, page->index, offset_within_page )
543 * [ also see get_inode_sequence_number() ]
545 * For private mappings (or when !@fshared), the key is:
546 * ( current->mm, address, 0 )
548 * This allows (cross process, where applicable) identification of the futex
549 * without keeping the page pinned for the duration of the FUTEX_WAIT.
551 * lock_page() might sleep, the caller should not hold a spinlock.
554 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
556 unsigned long address = (unsigned long)uaddr;
557 struct mm_struct *mm = current->mm;
558 struct page *page, *tail;
559 struct address_space *mapping;
563 * The futex address must be "naturally" aligned.
565 key->both.offset = address % PAGE_SIZE;
566 if (unlikely((address % sizeof(u32)) != 0))
568 address -= key->both.offset;
570 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
573 if (unlikely(should_fail_futex(fshared)))
577 * PROCESS_PRIVATE futexes are fast.
578 * As the mm cannot disappear under us and the 'key' only needs
579 * virtual address, we dont even have to find the underlying vma.
580 * Note : We do have to check 'uaddr' is a valid user address,
581 * but access_ok() should be faster than find_vma()
584 key->private.mm = mm;
585 key->private.address = address;
586 get_futex_key_refs(key); /* implies smp_mb(); (B) */
591 /* Ignore any VERIFY_READ mapping (futex common case) */
592 if (unlikely(should_fail_futex(fshared)))
595 err = get_user_pages_fast(address, 1, 1, &page);
597 * If write access is not required (eg. FUTEX_WAIT), try
598 * and get read-only access.
600 if (err == -EFAULT && rw == VERIFY_READ) {
601 err = get_user_pages_fast(address, 1, 0, &page);
610 * The treatment of mapping from this point on is critical. The page
611 * lock protects many things but in this context the page lock
612 * stabilizes mapping, prevents inode freeing in the shared
613 * file-backed region case and guards against movement to swap cache.
615 * Strictly speaking the page lock is not needed in all cases being
616 * considered here and page lock forces unnecessarily serialization
617 * From this point on, mapping will be re-verified if necessary and
618 * page lock will be acquired only if it is unavoidable
620 * Mapping checks require the head page for any compound page so the
621 * head page and mapping is looked up now. For anonymous pages, it
622 * does not matter if the page splits in the future as the key is
623 * based on the address. For filesystem-backed pages, the tail is
624 * required as the index of the page determines the key. For
625 * base pages, there is no tail page and tail == page.
628 page = compound_head(page);
629 mapping = READ_ONCE(page->mapping);
632 * If page->mapping is NULL, then it cannot be a PageAnon
633 * page; but it might be the ZERO_PAGE or in the gate area or
634 * in a special mapping (all cases which we are happy to fail);
635 * or it may have been a good file page when get_user_pages_fast
636 * found it, but truncated or holepunched or subjected to
637 * invalidate_complete_page2 before we got the page lock (also
638 * cases which we are happy to fail). And we hold a reference,
639 * so refcount care in invalidate_complete_page's remove_mapping
640 * prevents drop_caches from setting mapping to NULL beneath us.
642 * The case we do have to guard against is when memory pressure made
643 * shmem_writepage move it from filecache to swapcache beneath us:
644 * an unlikely race, but we do need to retry for page->mapping.
646 if (unlikely(!mapping)) {
650 * Page lock is required to identify which special case above
651 * applies. If this is really a shmem page then the page lock
652 * will prevent unexpected transitions.
655 shmem_swizzled = PageSwapCache(page) || page->mapping;
666 * Private mappings are handled in a simple way.
668 * If the futex key is stored on an anonymous page, then the associated
669 * object is the mm which is implicitly pinned by the calling process.
671 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
672 * it's a read-only handle, it's expected that futexes attach to
673 * the object not the particular process.
675 if (PageAnon(page)) {
677 * A RO anonymous page will never change and thus doesn't make
678 * sense for futex operations.
680 if (unlikely(should_fail_futex(fshared)) || ro) {
685 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
686 key->private.mm = mm;
687 key->private.address = address;
693 * The associated futex object in this case is the inode and
694 * the page->mapping must be traversed. Ordinarily this should
695 * be stabilised under page lock but it's not strictly
696 * necessary in this case as we just want to pin the inode, not
697 * update the radix tree or anything like that.
699 * The RCU read lock is taken as the inode is finally freed
700 * under RCU. If the mapping still matches expectations then the
701 * mapping->host can be safely accessed as being a valid inode.
705 if (READ_ONCE(page->mapping) != mapping) {
712 inode = READ_ONCE(mapping->host);
720 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
721 key->shared.i_seq = get_inode_sequence_number(inode);
722 key->shared.pgoff = page_to_pgoff(tail);
726 get_futex_key_refs(key); /* implies smp_mb(); (B) */
733 static inline void put_futex_key(union futex_key *key)
735 drop_futex_key_refs(key);
739 * fault_in_user_writeable() - Fault in user address and verify RW access
740 * @uaddr: pointer to faulting user space address
742 * Slow path to fixup the fault we just took in the atomic write
745 * We have no generic implementation of a non-destructive write to the
746 * user address. We know that we faulted in the atomic pagefault
747 * disabled section so we can as well avoid the #PF overhead by
748 * calling get_user_pages() right away.
750 static int fault_in_user_writeable(u32 __user *uaddr)
752 struct mm_struct *mm = current->mm;
755 down_read(&mm->mmap_sem);
756 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
757 FAULT_FLAG_WRITE, NULL);
758 up_read(&mm->mmap_sem);
760 return ret < 0 ? ret : 0;
764 * futex_top_waiter() - Return the highest priority waiter on a futex
765 * @hb: the hash bucket the futex_q's reside in
766 * @key: the futex key (to distinguish it from other futex futex_q's)
768 * Must be called with the hb lock held.
770 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
771 union futex_key *key)
773 struct futex_q *this;
775 plist_for_each_entry(this, &hb->chain, list) {
776 if (match_futex(&this->key, key))
782 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
783 u32 uval, u32 newval)
788 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
794 static int get_futex_value_locked(u32 *dest, u32 __user *from)
799 ret = __get_user(*dest, from);
802 return ret ? -EFAULT : 0;
809 static int refill_pi_state_cache(void)
811 struct futex_pi_state *pi_state;
813 if (likely(current->pi_state_cache))
816 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
821 INIT_LIST_HEAD(&pi_state->list);
822 /* pi_mutex gets initialized later */
823 pi_state->owner = NULL;
824 atomic_set(&pi_state->refcount, 1);
825 pi_state->key = FUTEX_KEY_INIT;
827 current->pi_state_cache = pi_state;
832 static struct futex_pi_state *alloc_pi_state(void)
834 struct futex_pi_state *pi_state = current->pi_state_cache;
837 current->pi_state_cache = NULL;
842 static void pi_state_update_owner(struct futex_pi_state *pi_state,
843 struct task_struct *new_owner)
845 struct task_struct *old_owner = pi_state->owner;
847 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
850 raw_spin_lock(&old_owner->pi_lock);
851 WARN_ON(list_empty(&pi_state->list));
852 list_del_init(&pi_state->list);
853 raw_spin_unlock(&old_owner->pi_lock);
857 raw_spin_lock(&new_owner->pi_lock);
858 WARN_ON(!list_empty(&pi_state->list));
859 list_add(&pi_state->list, &new_owner->pi_state_list);
860 pi_state->owner = new_owner;
861 raw_spin_unlock(&new_owner->pi_lock);
865 static void get_pi_state(struct futex_pi_state *pi_state)
867 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
871 * Drops a reference to the pi_state object and frees or caches it
872 * when the last reference is gone.
874 static void put_pi_state(struct futex_pi_state *pi_state)
879 if (!atomic_dec_and_test(&pi_state->refcount))
883 * If pi_state->owner is NULL, the owner is most probably dying
884 * and has cleaned up the pi_state already
886 if (pi_state->owner) {
889 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
890 pi_state_update_owner(pi_state, NULL);
891 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
892 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
895 if (current->pi_state_cache) {
899 * pi_state->list is already empty.
900 * clear pi_state->owner.
901 * refcount is at 0 - put it back to 1.
903 pi_state->owner = NULL;
904 atomic_set(&pi_state->refcount, 1);
905 current->pi_state_cache = pi_state;
909 #ifdef CONFIG_FUTEX_PI
912 * This task is holding PI mutexes at exit time => bad.
913 * Kernel cleans up PI-state, but userspace is likely hosed.
914 * (Robust-futex cleanup is separate and might save the day for userspace.)
916 static void exit_pi_state_list(struct task_struct *curr)
918 struct list_head *next, *head = &curr->pi_state_list;
919 struct futex_pi_state *pi_state;
920 struct futex_hash_bucket *hb;
921 union futex_key key = FUTEX_KEY_INIT;
923 if (!futex_cmpxchg_enabled)
926 * We are a ZOMBIE and nobody can enqueue itself on
927 * pi_state_list anymore, but we have to be careful
928 * versus waiters unqueueing themselves:
930 raw_spin_lock_irq(&curr->pi_lock);
931 while (!list_empty(head)) {
933 pi_state = list_entry(next, struct futex_pi_state, list);
935 hb = hash_futex(&key);
938 * We can race against put_pi_state() removing itself from the
939 * list (a waiter going away). put_pi_state() will first
940 * decrement the reference count and then modify the list, so
941 * its possible to see the list entry but fail this reference
944 * In that case; drop the locks to let put_pi_state() make
945 * progress and retry the loop.
947 if (!atomic_inc_not_zero(&pi_state->refcount)) {
948 raw_spin_unlock_irq(&curr->pi_lock);
950 raw_spin_lock_irq(&curr->pi_lock);
953 raw_spin_unlock_irq(&curr->pi_lock);
955 spin_lock(&hb->lock);
956 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
957 raw_spin_lock(&curr->pi_lock);
959 * We dropped the pi-lock, so re-check whether this
960 * task still owns the PI-state:
962 if (head->next != next) {
963 /* retain curr->pi_lock for the loop invariant */
964 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
965 spin_unlock(&hb->lock);
966 put_pi_state(pi_state);
970 WARN_ON(pi_state->owner != curr);
971 WARN_ON(list_empty(&pi_state->list));
972 list_del_init(&pi_state->list);
973 pi_state->owner = NULL;
975 raw_spin_unlock(&curr->pi_lock);
976 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
977 spin_unlock(&hb->lock);
979 rt_mutex_futex_unlock(&pi_state->pi_mutex);
980 put_pi_state(pi_state);
982 raw_spin_lock_irq(&curr->pi_lock);
984 raw_spin_unlock_irq(&curr->pi_lock);
987 static inline void exit_pi_state_list(struct task_struct *curr) { }
991 * We need to check the following states:
993 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
995 * [1] NULL | --- | --- | 0 | 0/1 | Valid
996 * [2] NULL | --- | --- | >0 | 0/1 | Valid
998 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1000 * [4] Found | Found | NULL | 0 | 1 | Valid
1001 * [5] Found | Found | NULL | >0 | 1 | Invalid
1003 * [6] Found | Found | task | 0 | 1 | Valid
1005 * [7] Found | Found | NULL | Any | 0 | Invalid
1007 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1008 * [9] Found | Found | task | 0 | 0 | Invalid
1009 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1011 * [1] Indicates that the kernel can acquire the futex atomically. We
1012 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1014 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1015 * thread is found then it indicates that the owner TID has died.
1017 * [3] Invalid. The waiter is queued on a non PI futex
1019 * [4] Valid state after exit_robust_list(), which sets the user space
1020 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1022 * [5] The user space value got manipulated between exit_robust_list()
1023 * and exit_pi_state_list()
1025 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1026 * the pi_state but cannot access the user space value.
1028 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1030 * [8] Owner and user space value match
1032 * [9] There is no transient state which sets the user space TID to 0
1033 * except exit_robust_list(), but this is indicated by the
1034 * FUTEX_OWNER_DIED bit. See [4]
1036 * [10] There is no transient state which leaves owner and user space
1037 * TID out of sync. Except one error case where the kernel is denied
1038 * write access to the user address, see fixup_pi_state_owner().
1041 * Serialization and lifetime rules:
1045 * hb -> futex_q, relation
1046 * futex_q -> pi_state, relation
1048 * (cannot be raw because hb can contain arbitrary amount
1051 * pi_mutex->wait_lock:
1055 * (and pi_mutex 'obviously')
1059 * p->pi_state_list -> pi_state->list, relation
1061 * pi_state->refcount:
1069 * pi_mutex->wait_lock
1075 * Validate that the existing waiter has a pi_state and sanity check
1076 * the pi_state against the user space value. If correct, attach to
1079 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1080 struct futex_pi_state *pi_state,
1081 struct futex_pi_state **ps)
1083 pid_t pid = uval & FUTEX_TID_MASK;
1088 * Userspace might have messed up non-PI and PI futexes [3]
1090 if (unlikely(!pi_state))
1094 * We get here with hb->lock held, and having found a
1095 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1096 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1097 * which in turn means that futex_lock_pi() still has a reference on
1100 * The waiter holding a reference on @pi_state also protects against
1101 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1102 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1103 * free pi_state before we can take a reference ourselves.
1105 WARN_ON(!atomic_read(&pi_state->refcount));
1108 * Now that we have a pi_state, we can acquire wait_lock
1109 * and do the state validation.
1111 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1114 * Since {uval, pi_state} is serialized by wait_lock, and our current
1115 * uval was read without holding it, it can have changed. Verify it
1116 * still is what we expect it to be, otherwise retry the entire
1119 if (get_futex_value_locked(&uval2, uaddr))
1126 * Handle the owner died case:
1128 if (uval & FUTEX_OWNER_DIED) {
1130 * exit_pi_state_list sets owner to NULL and wakes the
1131 * topmost waiter. The task which acquires the
1132 * pi_state->rt_mutex will fixup owner.
1134 if (!pi_state->owner) {
1136 * No pi state owner, but the user space TID
1137 * is not 0. Inconsistent state. [5]
1142 * Take a ref on the state and return success. [4]
1148 * If TID is 0, then either the dying owner has not
1149 * yet executed exit_pi_state_list() or some waiter
1150 * acquired the rtmutex in the pi state, but did not
1151 * yet fixup the TID in user space.
1153 * Take a ref on the state and return success. [6]
1159 * If the owner died bit is not set, then the pi_state
1160 * must have an owner. [7]
1162 if (!pi_state->owner)
1167 * Bail out if user space manipulated the futex value. If pi
1168 * state exists then the owner TID must be the same as the
1169 * user space TID. [9/10]
1171 if (pid != task_pid_vnr(pi_state->owner))
1175 get_pi_state(pi_state);
1176 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1193 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1198 * wait_for_owner_exiting - Block until the owner has exited
1199 * @exiting: Pointer to the exiting task
1201 * Caller must hold a refcount on @exiting.
1203 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1205 if (ret != -EBUSY) {
1206 WARN_ON_ONCE(exiting);
1210 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1213 mutex_lock(&exiting->futex_exit_mutex);
1215 * No point in doing state checking here. If the waiter got here
1216 * while the task was in exec()->exec_futex_release() then it can
1217 * have any FUTEX_STATE_* value when the waiter has acquired the
1218 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1219 * already. Highly unlikely and not a problem. Just one more round
1220 * through the futex maze.
1222 mutex_unlock(&exiting->futex_exit_mutex);
1224 put_task_struct(exiting);
1227 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1228 struct task_struct *tsk)
1233 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1234 * caller that the alleged owner is busy.
1236 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1240 * Reread the user space value to handle the following situation:
1244 * sys_exit() sys_futex()
1245 * do_exit() futex_lock_pi()
1246 * futex_lock_pi_atomic()
1247 * exit_signals(tsk) No waiters:
1248 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1249 * mm_release(tsk) Set waiter bit
1250 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1251 * Set owner died attach_to_pi_owner() {
1252 * *uaddr = 0xC0000000; tsk = get_task(PID);
1253 * } if (!tsk->flags & PF_EXITING) {
1255 * tsk->futex_state = } else {
1256 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1259 * return -ESRCH; <--- FAIL
1262 * Returning ESRCH unconditionally is wrong here because the
1263 * user space value has been changed by the exiting task.
1265 * The same logic applies to the case where the exiting task is
1268 if (get_futex_value_locked(&uval2, uaddr))
1271 /* If the user space value has changed, try again. */
1276 * The exiting task did not have a robust list, the robust list was
1277 * corrupted or the user space value in *uaddr is simply bogus.
1278 * Give up and tell user space.
1284 * Lookup the task for the TID provided from user space and attach to
1285 * it after doing proper sanity checks.
1287 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1288 struct futex_pi_state **ps,
1289 struct task_struct **exiting)
1291 pid_t pid = uval & FUTEX_TID_MASK;
1292 struct futex_pi_state *pi_state;
1293 struct task_struct *p;
1296 * We are the first waiter - try to look up the real owner and attach
1297 * the new pi_state to it, but bail out when TID = 0 [1]
1299 * The !pid check is paranoid. None of the call sites should end up
1300 * with pid == 0, but better safe than sorry. Let the caller retry
1304 p = find_get_task_by_vpid(pid);
1306 return handle_exit_race(uaddr, uval, NULL);
1308 if (unlikely(p->flags & PF_KTHREAD)) {
1314 * We need to look at the task state to figure out, whether the
1315 * task is exiting. To protect against the change of the task state
1316 * in futex_exit_release(), we do this protected by p->pi_lock:
1318 raw_spin_lock_irq(&p->pi_lock);
1319 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1321 * The task is on the way out. When the futex state is
1322 * FUTEX_STATE_DEAD, we know that the task has finished
1325 int ret = handle_exit_race(uaddr, uval, p);
1327 raw_spin_unlock_irq(&p->pi_lock);
1329 * If the owner task is between FUTEX_STATE_EXITING and
1330 * FUTEX_STATE_DEAD then store the task pointer and keep
1331 * the reference on the task struct. The calling code will
1332 * drop all locks, wait for the task to reach
1333 * FUTEX_STATE_DEAD and then drop the refcount. This is
1334 * required to prevent a live lock when the current task
1335 * preempted the exiting task between the two states.
1345 * No existing pi state. First waiter. [2]
1347 * This creates pi_state, we have hb->lock held, this means nothing can
1348 * observe this state, wait_lock is irrelevant.
1350 pi_state = alloc_pi_state();
1353 * Initialize the pi_mutex in locked state and make @p
1356 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1358 /* Store the key for possible exit cleanups: */
1359 pi_state->key = *key;
1361 WARN_ON(!list_empty(&pi_state->list));
1362 list_add(&pi_state->list, &p->pi_state_list);
1364 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1365 * because there is no concurrency as the object is not published yet.
1367 pi_state->owner = p;
1368 raw_spin_unlock_irq(&p->pi_lock);
1377 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1378 struct futex_hash_bucket *hb,
1379 union futex_key *key, struct futex_pi_state **ps,
1380 struct task_struct **exiting)
1382 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1385 * If there is a waiter on that futex, validate it and
1386 * attach to the pi_state when the validation succeeds.
1389 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1392 * We are the first waiter - try to look up the owner based on
1393 * @uval and attach to it.
1395 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1398 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1401 u32 uninitialized_var(curval);
1403 if (unlikely(should_fail_futex(true)))
1406 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1410 /* If user space value changed, let the caller retry */
1411 return curval != uval ? -EAGAIN : 0;
1415 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1416 * @uaddr: the pi futex user address
1417 * @hb: the pi futex hash bucket
1418 * @key: the futex key associated with uaddr and hb
1419 * @ps: the pi_state pointer where we store the result of the
1421 * @task: the task to perform the atomic lock work for. This will
1422 * be "current" except in the case of requeue pi.
1423 * @exiting: Pointer to store the task pointer of the owner task
1424 * which is in the middle of exiting
1425 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1428 * - 0 - ready to wait;
1429 * - 1 - acquired the lock;
1432 * The hb->lock and futex_key refs shall be held by the caller.
1434 * @exiting is only set when the return value is -EBUSY. If so, this holds
1435 * a refcount on the exiting task on return and the caller needs to drop it
1436 * after waiting for the exit to complete.
1438 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1439 union futex_key *key,
1440 struct futex_pi_state **ps,
1441 struct task_struct *task,
1442 struct task_struct **exiting,
1445 u32 uval, newval, vpid = task_pid_vnr(task);
1446 struct futex_q *top_waiter;
1450 * Read the user space value first so we can validate a few
1451 * things before proceeding further.
1453 if (get_futex_value_locked(&uval, uaddr))
1456 if (unlikely(should_fail_futex(true)))
1462 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1465 if ((unlikely(should_fail_futex(true))))
1469 * Lookup existing state first. If it exists, try to attach to
1472 top_waiter = futex_top_waiter(hb, key);
1474 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1477 * No waiter and user TID is 0. We are here because the
1478 * waiters or the owner died bit is set or called from
1479 * requeue_cmp_pi or for whatever reason something took the
1482 if (!(uval & FUTEX_TID_MASK)) {
1484 * We take over the futex. No other waiters and the user space
1485 * TID is 0. We preserve the owner died bit.
1487 newval = uval & FUTEX_OWNER_DIED;
1490 /* The futex requeue_pi code can enforce the waiters bit */
1492 newval |= FUTEX_WAITERS;
1494 ret = lock_pi_update_atomic(uaddr, uval, newval);
1495 /* If the take over worked, return 1 */
1496 return ret < 0 ? ret : 1;
1500 * First waiter. Set the waiters bit before attaching ourself to
1501 * the owner. If owner tries to unlock, it will be forced into
1502 * the kernel and blocked on hb->lock.
1504 newval = uval | FUTEX_WAITERS;
1505 ret = lock_pi_update_atomic(uaddr, uval, newval);
1509 * If the update of the user space value succeeded, we try to
1510 * attach to the owner. If that fails, no harm done, we only
1511 * set the FUTEX_WAITERS bit in the user space variable.
1513 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1517 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1518 * @q: The futex_q to unqueue
1520 * The q->lock_ptr must not be NULL and must be held by the caller.
1522 static void __unqueue_futex(struct futex_q *q)
1524 struct futex_hash_bucket *hb;
1526 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1527 || WARN_ON(plist_node_empty(&q->list)))
1530 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1531 plist_del(&q->list, &hb->chain);
1536 * The hash bucket lock must be held when this is called.
1537 * Afterwards, the futex_q must not be accessed. Callers
1538 * must ensure to later call wake_up_q() for the actual
1541 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1543 struct task_struct *p = q->task;
1545 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1551 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1552 * is written, without taking any locks. This is possible in the event
1553 * of a spurious wakeup, for example. A memory barrier is required here
1554 * to prevent the following store to lock_ptr from getting ahead of the
1555 * plist_del in __unqueue_futex().
1557 smp_store_release(&q->lock_ptr, NULL);
1560 * Queue the task for later wakeup for after we've released
1561 * the hb->lock. wake_q_add() grabs reference to p.
1563 wake_q_add(wake_q, p);
1568 * Caller must hold a reference on @pi_state.
1570 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1572 u32 uninitialized_var(curval), newval;
1573 struct task_struct *new_owner;
1574 bool postunlock = false;
1575 DEFINE_WAKE_Q(wake_q);
1578 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1579 if (WARN_ON_ONCE(!new_owner)) {
1581 * As per the comment in futex_unlock_pi() this should not happen.
1583 * When this happens, give up our locks and try again, giving
1584 * the futex_lock_pi() instance time to complete, either by
1585 * waiting on the rtmutex or removing itself from the futex
1593 * We pass it to the next owner. The WAITERS bit is always kept
1594 * enabled while there is PI state around. We cleanup the owner
1595 * died bit, because we are the owner.
1597 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1599 if (unlikely(should_fail_futex(true))) {
1604 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1605 if (!ret && (curval != uval)) {
1607 * If a unconditional UNLOCK_PI operation (user space did not
1608 * try the TID->0 transition) raced with a waiter setting the
1609 * FUTEX_WAITERS flag between get_user() and locking the hash
1610 * bucket lock, retry the operation.
1612 if ((FUTEX_TID_MASK & curval) == uval)
1620 * This is a point of no return; once we modified the uval
1621 * there is no going back and subsequent operations must
1624 pi_state_update_owner(pi_state, new_owner);
1625 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1629 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1632 rt_mutex_postunlock(&wake_q);
1638 * Express the locking dependencies for lockdep:
1641 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1644 spin_lock(&hb1->lock);
1646 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1647 } else { /* hb1 > hb2 */
1648 spin_lock(&hb2->lock);
1649 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1654 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1656 spin_unlock(&hb1->lock);
1658 spin_unlock(&hb2->lock);
1662 * Wake up waiters matching bitset queued on this futex (uaddr).
1665 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1667 struct futex_hash_bucket *hb;
1668 struct futex_q *this, *next;
1669 union futex_key key = FUTEX_KEY_INIT;
1671 DEFINE_WAKE_Q(wake_q);
1676 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1677 if (unlikely(ret != 0))
1680 hb = hash_futex(&key);
1682 /* Make sure we really have tasks to wakeup */
1683 if (!hb_waiters_pending(hb))
1686 spin_lock(&hb->lock);
1688 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1689 if (match_futex (&this->key, &key)) {
1690 if (this->pi_state || this->rt_waiter) {
1695 /* Check if one of the bits is set in both bitsets */
1696 if (!(this->bitset & bitset))
1699 mark_wake_futex(&wake_q, this);
1700 if (++ret >= nr_wake)
1705 spin_unlock(&hb->lock);
1708 put_futex_key(&key);
1713 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1715 unsigned int op = (encoded_op & 0x70000000) >> 28;
1716 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1717 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1718 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1721 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1722 if (oparg < 0 || oparg > 31) {
1723 char comm[sizeof(current->comm)];
1725 * kill this print and return -EINVAL when userspace
1728 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1729 get_task_comm(comm, current), oparg);
1735 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1738 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1743 case FUTEX_OP_CMP_EQ:
1744 return oldval == cmparg;
1745 case FUTEX_OP_CMP_NE:
1746 return oldval != cmparg;
1747 case FUTEX_OP_CMP_LT:
1748 return oldval < cmparg;
1749 case FUTEX_OP_CMP_GE:
1750 return oldval >= cmparg;
1751 case FUTEX_OP_CMP_LE:
1752 return oldval <= cmparg;
1753 case FUTEX_OP_CMP_GT:
1754 return oldval > cmparg;
1761 * Wake up all waiters hashed on the physical page that is mapped
1762 * to this virtual address:
1765 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1766 int nr_wake, int nr_wake2, int op)
1768 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1769 struct futex_hash_bucket *hb1, *hb2;
1770 struct futex_q *this, *next;
1772 DEFINE_WAKE_Q(wake_q);
1775 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1776 if (unlikely(ret != 0))
1778 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1779 if (unlikely(ret != 0))
1782 hb1 = hash_futex(&key1);
1783 hb2 = hash_futex(&key2);
1786 double_lock_hb(hb1, hb2);
1787 op_ret = futex_atomic_op_inuser(op, uaddr2);
1788 if (unlikely(op_ret < 0)) {
1789 double_unlock_hb(hb1, hb2);
1791 if (!IS_ENABLED(CONFIG_MMU) ||
1792 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1794 * we don't get EFAULT from MMU faults if we don't have
1795 * an MMU, but we might get them from range checking
1801 if (op_ret == -EFAULT) {
1802 ret = fault_in_user_writeable(uaddr2);
1807 if (!(flags & FLAGS_SHARED)) {
1812 put_futex_key(&key2);
1813 put_futex_key(&key1);
1818 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1819 if (match_futex (&this->key, &key1)) {
1820 if (this->pi_state || this->rt_waiter) {
1824 mark_wake_futex(&wake_q, this);
1825 if (++ret >= nr_wake)
1832 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1833 if (match_futex (&this->key, &key2)) {
1834 if (this->pi_state || this->rt_waiter) {
1838 mark_wake_futex(&wake_q, this);
1839 if (++op_ret >= nr_wake2)
1847 double_unlock_hb(hb1, hb2);
1850 put_futex_key(&key2);
1852 put_futex_key(&key1);
1858 * requeue_futex() - Requeue a futex_q from one hb to another
1859 * @q: the futex_q to requeue
1860 * @hb1: the source hash_bucket
1861 * @hb2: the target hash_bucket
1862 * @key2: the new key for the requeued futex_q
1865 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1866 struct futex_hash_bucket *hb2, union futex_key *key2)
1870 * If key1 and key2 hash to the same bucket, no need to
1873 if (likely(&hb1->chain != &hb2->chain)) {
1874 plist_del(&q->list, &hb1->chain);
1875 hb_waiters_dec(hb1);
1876 hb_waiters_inc(hb2);
1877 plist_add(&q->list, &hb2->chain);
1878 q->lock_ptr = &hb2->lock;
1880 get_futex_key_refs(key2);
1885 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1887 * @key: the key of the requeue target futex
1888 * @hb: the hash_bucket of the requeue target futex
1890 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1891 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1892 * to the requeue target futex so the waiter can detect the wakeup on the right
1893 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1894 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1895 * to protect access to the pi_state to fixup the owner later. Must be called
1896 * with both q->lock_ptr and hb->lock held.
1899 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1900 struct futex_hash_bucket *hb)
1902 get_futex_key_refs(key);
1907 WARN_ON(!q->rt_waiter);
1908 q->rt_waiter = NULL;
1910 q->lock_ptr = &hb->lock;
1912 wake_up_state(q->task, TASK_NORMAL);
1916 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1917 * @pifutex: the user address of the to futex
1918 * @hb1: the from futex hash bucket, must be locked by the caller
1919 * @hb2: the to futex hash bucket, must be locked by the caller
1920 * @key1: the from futex key
1921 * @key2: the to futex key
1922 * @ps: address to store the pi_state pointer
1923 * @exiting: Pointer to store the task pointer of the owner task
1924 * which is in the middle of exiting
1925 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1927 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1928 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1929 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1930 * hb1 and hb2 must be held by the caller.
1932 * @exiting is only set when the return value is -EBUSY. If so, this holds
1933 * a refcount on the exiting task on return and the caller needs to drop it
1934 * after waiting for the exit to complete.
1937 * - 0 - failed to acquire the lock atomically;
1938 * - >0 - acquired the lock, return value is vpid of the top_waiter
1942 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1943 struct futex_hash_bucket *hb2, union futex_key *key1,
1944 union futex_key *key2, struct futex_pi_state **ps,
1945 struct task_struct **exiting, int set_waiters)
1947 struct futex_q *top_waiter = NULL;
1951 if (get_futex_value_locked(&curval, pifutex))
1954 if (unlikely(should_fail_futex(true)))
1958 * Find the top_waiter and determine if there are additional waiters.
1959 * If the caller intends to requeue more than 1 waiter to pifutex,
1960 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1961 * as we have means to handle the possible fault. If not, don't set
1962 * the bit unecessarily as it will force the subsequent unlock to enter
1965 top_waiter = futex_top_waiter(hb1, key1);
1967 /* There are no waiters, nothing for us to do. */
1971 /* Ensure we requeue to the expected futex. */
1972 if (!match_futex(top_waiter->requeue_pi_key, key2))
1976 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1977 * the contended case or if set_waiters is 1. The pi_state is returned
1978 * in ps in contended cases.
1980 vpid = task_pid_vnr(top_waiter->task);
1981 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1982 exiting, set_waiters);
1984 requeue_pi_wake_futex(top_waiter, key2, hb2);
1991 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1992 * @uaddr1: source futex user address
1993 * @flags: futex flags (FLAGS_SHARED, etc.)
1994 * @uaddr2: target futex user address
1995 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1996 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1997 * @cmpval: @uaddr1 expected value (or %NULL)
1998 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1999 * pi futex (pi to pi requeue is not supported)
2001 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
2002 * uaddr2 atomically on behalf of the top waiter.
2005 * - >=0 - on success, the number of tasks requeued or woken;
2008 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
2009 u32 __user *uaddr2, int nr_wake, int nr_requeue,
2010 u32 *cmpval, int requeue_pi)
2012 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
2013 int drop_count = 0, task_count = 0, ret;
2014 struct futex_pi_state *pi_state = NULL;
2015 struct futex_hash_bucket *hb1, *hb2;
2016 struct futex_q *this, *next;
2017 DEFINE_WAKE_Q(wake_q);
2019 if (nr_wake < 0 || nr_requeue < 0)
2023 * When PI not supported: return -ENOSYS if requeue_pi is true,
2024 * consequently the compiler knows requeue_pi is always false past
2025 * this point which will optimize away all the conditional code
2028 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
2033 * Requeue PI only works on two distinct uaddrs. This
2034 * check is only valid for private futexes. See below.
2036 if (uaddr1 == uaddr2)
2040 * requeue_pi requires a pi_state, try to allocate it now
2041 * without any locks in case it fails.
2043 if (refill_pi_state_cache())
2046 * requeue_pi must wake as many tasks as it can, up to nr_wake
2047 * + nr_requeue, since it acquires the rt_mutex prior to
2048 * returning to userspace, so as to not leave the rt_mutex with
2049 * waiters and no owner. However, second and third wake-ups
2050 * cannot be predicted as they involve race conditions with the
2051 * first wake and a fault while looking up the pi_state. Both
2052 * pthread_cond_signal() and pthread_cond_broadcast() should
2060 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
2061 if (unlikely(ret != 0))
2063 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2064 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
2065 if (unlikely(ret != 0))
2069 * The check above which compares uaddrs is not sufficient for
2070 * shared futexes. We need to compare the keys:
2072 if (requeue_pi && match_futex(&key1, &key2)) {
2077 hb1 = hash_futex(&key1);
2078 hb2 = hash_futex(&key2);
2081 hb_waiters_inc(hb2);
2082 double_lock_hb(hb1, hb2);
2084 if (likely(cmpval != NULL)) {
2087 ret = get_futex_value_locked(&curval, uaddr1);
2089 if (unlikely(ret)) {
2090 double_unlock_hb(hb1, hb2);
2091 hb_waiters_dec(hb2);
2093 ret = get_user(curval, uaddr1);
2097 if (!(flags & FLAGS_SHARED))
2100 put_futex_key(&key2);
2101 put_futex_key(&key1);
2104 if (curval != *cmpval) {
2110 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2111 struct task_struct *exiting = NULL;
2114 * Attempt to acquire uaddr2 and wake the top waiter. If we
2115 * intend to requeue waiters, force setting the FUTEX_WAITERS
2116 * bit. We force this here where we are able to easily handle
2117 * faults rather in the requeue loop below.
2119 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2121 &exiting, nr_requeue);
2124 * At this point the top_waiter has either taken uaddr2 or is
2125 * waiting on it. If the former, then the pi_state will not
2126 * exist yet, look it up one more time to ensure we have a
2127 * reference to it. If the lock was taken, ret contains the
2128 * vpid of the top waiter task.
2129 * If the lock was not taken, we have pi_state and an initial
2130 * refcount on it. In case of an error we have nothing.
2137 * If we acquired the lock, then the user space value
2138 * of uaddr2 should be vpid. It cannot be changed by
2139 * the top waiter as it is blocked on hb2 lock if it
2140 * tries to do so. If something fiddled with it behind
2141 * our back the pi state lookup might unearth it. So
2142 * we rather use the known value than rereading and
2143 * handing potential crap to lookup_pi_state.
2145 * If that call succeeds then we have pi_state and an
2146 * initial refcount on it.
2148 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2149 &pi_state, &exiting);
2154 /* We hold a reference on the pi state. */
2157 /* If the above failed, then pi_state is NULL */
2159 double_unlock_hb(hb1, hb2);
2160 hb_waiters_dec(hb2);
2161 put_futex_key(&key2);
2162 put_futex_key(&key1);
2163 ret = fault_in_user_writeable(uaddr2);
2170 * Two reasons for this:
2171 * - EBUSY: Owner is exiting and we just wait for the
2173 * - EAGAIN: The user space value changed.
2175 double_unlock_hb(hb1, hb2);
2176 hb_waiters_dec(hb2);
2177 put_futex_key(&key2);
2178 put_futex_key(&key1);
2180 * Handle the case where the owner is in the middle of
2181 * exiting. Wait for the exit to complete otherwise
2182 * this task might loop forever, aka. live lock.
2184 wait_for_owner_exiting(ret, exiting);
2192 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2193 if (task_count - nr_wake >= nr_requeue)
2196 if (!match_futex(&this->key, &key1))
2200 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2201 * be paired with each other and no other futex ops.
2203 * We should never be requeueing a futex_q with a pi_state,
2204 * which is awaiting a futex_unlock_pi().
2206 if ((requeue_pi && !this->rt_waiter) ||
2207 (!requeue_pi && this->rt_waiter) ||
2214 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2215 * lock, we already woke the top_waiter. If not, it will be
2216 * woken by futex_unlock_pi().
2218 if (++task_count <= nr_wake && !requeue_pi) {
2219 mark_wake_futex(&wake_q, this);
2223 /* Ensure we requeue to the expected futex for requeue_pi. */
2224 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2230 * Requeue nr_requeue waiters and possibly one more in the case
2231 * of requeue_pi if we couldn't acquire the lock atomically.
2235 * Prepare the waiter to take the rt_mutex. Take a
2236 * refcount on the pi_state and store the pointer in
2237 * the futex_q object of the waiter.
2239 get_pi_state(pi_state);
2240 this->pi_state = pi_state;
2241 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2246 * We got the lock. We do neither drop the
2247 * refcount on pi_state nor clear
2248 * this->pi_state because the waiter needs the
2249 * pi_state for cleaning up the user space
2250 * value. It will drop the refcount after
2253 requeue_pi_wake_futex(this, &key2, hb2);
2258 * rt_mutex_start_proxy_lock() detected a
2259 * potential deadlock when we tried to queue
2260 * that waiter. Drop the pi_state reference
2261 * which we took above and remove the pointer
2262 * to the state from the waiters futex_q
2265 this->pi_state = NULL;
2266 put_pi_state(pi_state);
2268 * We stop queueing more waiters and let user
2269 * space deal with the mess.
2274 requeue_futex(this, hb1, hb2, &key2);
2279 * We took an extra initial reference to the pi_state either
2280 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2281 * need to drop it here again.
2283 put_pi_state(pi_state);
2286 double_unlock_hb(hb1, hb2);
2288 hb_waiters_dec(hb2);
2291 * drop_futex_key_refs() must be called outside the spinlocks. During
2292 * the requeue we moved futex_q's from the hash bucket at key1 to the
2293 * one at key2 and updated their key pointer. We no longer need to
2294 * hold the references to key1.
2296 while (--drop_count >= 0)
2297 drop_futex_key_refs(&key1);
2300 put_futex_key(&key2);
2302 put_futex_key(&key1);
2304 return ret ? ret : task_count;
2307 /* The key must be already stored in q->key. */
2308 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2309 __acquires(&hb->lock)
2311 struct futex_hash_bucket *hb;
2313 hb = hash_futex(&q->key);
2316 * Increment the counter before taking the lock so that
2317 * a potential waker won't miss a to-be-slept task that is
2318 * waiting for the spinlock. This is safe as all queue_lock()
2319 * users end up calling queue_me(). Similarly, for housekeeping,
2320 * decrement the counter at queue_unlock() when some error has
2321 * occurred and we don't end up adding the task to the list.
2325 q->lock_ptr = &hb->lock;
2327 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2332 queue_unlock(struct futex_hash_bucket *hb)
2333 __releases(&hb->lock)
2335 spin_unlock(&hb->lock);
2339 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2344 * The priority used to register this element is
2345 * - either the real thread-priority for the real-time threads
2346 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2347 * - or MAX_RT_PRIO for non-RT threads.
2348 * Thus, all RT-threads are woken first in priority order, and
2349 * the others are woken last, in FIFO order.
2351 prio = min(current->normal_prio, MAX_RT_PRIO);
2353 plist_node_init(&q->list, prio);
2354 plist_add(&q->list, &hb->chain);
2359 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2360 * @q: The futex_q to enqueue
2361 * @hb: The destination hash bucket
2363 * The hb->lock must be held by the caller, and is released here. A call to
2364 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2365 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2366 * or nothing if the unqueue is done as part of the wake process and the unqueue
2367 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2370 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2371 __releases(&hb->lock)
2374 spin_unlock(&hb->lock);
2378 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2379 * @q: The futex_q to unqueue
2381 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2382 * be paired with exactly one earlier call to queue_me().
2385 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2386 * - 0 - if the futex_q was already removed by the waking thread
2388 static int unqueue_me(struct futex_q *q)
2390 spinlock_t *lock_ptr;
2393 /* In the common case we don't take the spinlock, which is nice. */
2396 * q->lock_ptr can change between this read and the following spin_lock.
2397 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2398 * optimizing lock_ptr out of the logic below.
2400 lock_ptr = READ_ONCE(q->lock_ptr);
2401 if (lock_ptr != NULL) {
2402 spin_lock(lock_ptr);
2404 * q->lock_ptr can change between reading it and
2405 * spin_lock(), causing us to take the wrong lock. This
2406 * corrects the race condition.
2408 * Reasoning goes like this: if we have the wrong lock,
2409 * q->lock_ptr must have changed (maybe several times)
2410 * between reading it and the spin_lock(). It can
2411 * change again after the spin_lock() but only if it was
2412 * already changed before the spin_lock(). It cannot,
2413 * however, change back to the original value. Therefore
2414 * we can detect whether we acquired the correct lock.
2416 if (unlikely(lock_ptr != q->lock_ptr)) {
2417 spin_unlock(lock_ptr);
2422 BUG_ON(q->pi_state);
2424 spin_unlock(lock_ptr);
2428 drop_futex_key_refs(&q->key);
2433 * PI futexes can not be requeued and must remove themself from the
2434 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2437 static void unqueue_me_pi(struct futex_q *q)
2438 __releases(q->lock_ptr)
2442 BUG_ON(!q->pi_state);
2443 put_pi_state(q->pi_state);
2446 spin_unlock(q->lock_ptr);
2449 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2450 struct task_struct *argowner)
2452 u32 uval, uninitialized_var(curval), newval, newtid;
2453 struct futex_pi_state *pi_state = q->pi_state;
2454 struct task_struct *oldowner, *newowner;
2457 oldowner = pi_state->owner;
2460 * We are here because either:
2462 * - we stole the lock and pi_state->owner needs updating to reflect
2463 * that (@argowner == current),
2467 * - someone stole our lock and we need to fix things to point to the
2468 * new owner (@argowner == NULL).
2470 * Either way, we have to replace the TID in the user space variable.
2471 * This must be atomic as we have to preserve the owner died bit here.
2473 * Note: We write the user space value _before_ changing the pi_state
2474 * because we can fault here. Imagine swapped out pages or a fork
2475 * that marked all the anonymous memory readonly for cow.
2477 * Modifying pi_state _before_ the user space value would leave the
2478 * pi_state in an inconsistent state when we fault here, because we
2479 * need to drop the locks to handle the fault. This might be observed
2480 * in the PID check in lookup_pi_state.
2484 if (oldowner != current) {
2486 * We raced against a concurrent self; things are
2487 * already fixed up. Nothing to do.
2492 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2493 /* We got the lock. pi_state is correct. Tell caller. */
2498 * The trylock just failed, so either there is an owner or
2499 * there is a higher priority waiter than this one.
2501 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2503 * If the higher priority waiter has not yet taken over the
2504 * rtmutex then newowner is NULL. We can't return here with
2505 * that state because it's inconsistent vs. the user space
2506 * state. So drop the locks and try again. It's a valid
2507 * situation and not any different from the other retry
2510 if (unlikely(!newowner)) {
2515 WARN_ON_ONCE(argowner != current);
2516 if (oldowner == current) {
2518 * We raced against a concurrent self; things are
2519 * already fixed up. Nothing to do.
2523 newowner = argowner;
2526 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2528 if (!pi_state->owner)
2529 newtid |= FUTEX_OWNER_DIED;
2531 err = get_futex_value_locked(&uval, uaddr);
2536 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2538 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2548 * We fixed up user space. Now we need to fix the pi_state
2551 pi_state_update_owner(pi_state, newowner);
2553 return argowner == current;
2556 * In order to reschedule or handle a page fault, we need to drop the
2557 * locks here. In the case of a fault, this gives the other task
2558 * (either the highest priority waiter itself or the task which stole
2559 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2560 * are back from handling the fault we need to check the pi_state after
2561 * reacquiring the locks and before trying to do another fixup. When
2562 * the fixup has been done already we simply return.
2564 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2565 * drop hb->lock since the caller owns the hb -> futex_q relation.
2566 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2569 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2570 spin_unlock(q->lock_ptr);
2574 err = fault_in_user_writeable(uaddr);
2587 spin_lock(q->lock_ptr);
2588 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2591 * Check if someone else fixed it for us:
2593 if (pi_state->owner != oldowner)
2594 return argowner == current;
2596 /* Retry if err was -EAGAIN or the fault in succeeded */
2601 * fault_in_user_writeable() failed so user state is immutable. At
2602 * best we can make the kernel state consistent but user state will
2603 * be most likely hosed and any subsequent unlock operation will be
2604 * rejected due to PI futex rule [10].
2606 * Ensure that the rtmutex owner is also the pi_state owner despite
2607 * the user space value claiming something different. There is no
2608 * point in unlocking the rtmutex if current is the owner as it
2609 * would need to wait until the next waiter has taken the rtmutex
2610 * to guarantee consistent state. Keep it simple. Userspace asked
2611 * for this wreckaged state.
2613 * The rtmutex has an owner - either current or some other
2614 * task. See the EAGAIN loop above.
2616 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2621 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2622 struct task_struct *argowner)
2624 struct futex_pi_state *pi_state = q->pi_state;
2627 lockdep_assert_held(q->lock_ptr);
2629 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2630 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2631 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2635 static long futex_wait_restart(struct restart_block *restart);
2638 * fixup_owner() - Post lock pi_state and corner case management
2639 * @uaddr: user address of the futex
2640 * @q: futex_q (contains pi_state and access to the rt_mutex)
2641 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2643 * After attempting to lock an rt_mutex, this function is called to cleanup
2644 * the pi_state owner as well as handle race conditions that may allow us to
2645 * acquire the lock. Must be called with the hb lock held.
2648 * - 1 - success, lock taken;
2649 * - 0 - success, lock not taken;
2650 * - <0 - on error (-EFAULT)
2652 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2656 * Got the lock. We might not be the anticipated owner if we
2657 * did a lock-steal - fix up the PI-state in that case:
2659 * Speculative pi_state->owner read (we don't hold wait_lock);
2660 * since we own the lock pi_state->owner == current is the
2661 * stable state, anything else needs more attention.
2663 if (q->pi_state->owner != current)
2664 return fixup_pi_state_owner(uaddr, q, current);
2669 * If we didn't get the lock; check if anybody stole it from us. In
2670 * that case, we need to fix up the uval to point to them instead of
2671 * us, otherwise bad things happen. [10]
2673 * Another speculative read; pi_state->owner == current is unstable
2674 * but needs our attention.
2676 if (q->pi_state->owner == current)
2677 return fixup_pi_state_owner(uaddr, q, NULL);
2680 * Paranoia check. If we did not take the lock, then we should not be
2681 * the owner of the rt_mutex. Warn and establish consistent state.
2683 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2684 return fixup_pi_state_owner(uaddr, q, current);
2690 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2691 * @hb: the futex hash bucket, must be locked by the caller
2692 * @q: the futex_q to queue up on
2693 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2695 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2696 struct hrtimer_sleeper *timeout)
2699 * The task state is guaranteed to be set before another task can
2700 * wake it. set_current_state() is implemented using smp_store_mb() and
2701 * queue_me() calls spin_unlock() upon completion, both serializing
2702 * access to the hash list and forcing another memory barrier.
2704 set_current_state(TASK_INTERRUPTIBLE);
2709 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2712 * If we have been removed from the hash list, then another task
2713 * has tried to wake us, and we can skip the call to schedule().
2715 if (likely(!plist_node_empty(&q->list))) {
2717 * If the timer has already expired, current will already be
2718 * flagged for rescheduling. Only call schedule if there
2719 * is no timeout, or if it has yet to expire.
2721 if (!timeout || timeout->task)
2722 freezable_schedule();
2724 __set_current_state(TASK_RUNNING);
2728 * futex_wait_setup() - Prepare to wait on a futex
2729 * @uaddr: the futex userspace address
2730 * @val: the expected value
2731 * @flags: futex flags (FLAGS_SHARED, etc.)
2732 * @q: the associated futex_q
2733 * @hb: storage for hash_bucket pointer to be returned to caller
2735 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2736 * compare it with the expected value. Handle atomic faults internally.
2737 * Return with the hb lock held and a q.key reference on success, and unlocked
2738 * with no q.key reference on failure.
2741 * - 0 - uaddr contains val and hb has been locked;
2742 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2744 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2745 struct futex_q *q, struct futex_hash_bucket **hb)
2751 * Access the page AFTER the hash-bucket is locked.
2752 * Order is important:
2754 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2755 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2757 * The basic logical guarantee of a futex is that it blocks ONLY
2758 * if cond(var) is known to be true at the time of blocking, for
2759 * any cond. If we locked the hash-bucket after testing *uaddr, that
2760 * would open a race condition where we could block indefinitely with
2761 * cond(var) false, which would violate the guarantee.
2763 * On the other hand, we insert q and release the hash-bucket only
2764 * after testing *uaddr. This guarantees that futex_wait() will NOT
2765 * absorb a wakeup if *uaddr does not match the desired values
2766 * while the syscall executes.
2769 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2770 if (unlikely(ret != 0))
2774 *hb = queue_lock(q);
2776 ret = get_futex_value_locked(&uval, uaddr);
2781 ret = get_user(uval, uaddr);
2785 if (!(flags & FLAGS_SHARED))
2788 put_futex_key(&q->key);
2799 put_futex_key(&q->key);
2803 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2804 ktime_t *abs_time, u32 bitset)
2806 struct hrtimer_sleeper timeout, *to = NULL;
2807 struct restart_block *restart;
2808 struct futex_hash_bucket *hb;
2809 struct futex_q q = futex_q_init;
2819 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2820 CLOCK_REALTIME : CLOCK_MONOTONIC,
2822 hrtimer_init_sleeper(to, current);
2823 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2824 current->timer_slack_ns);
2829 * Prepare to wait on uaddr. On success, holds hb lock and increments
2832 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2836 /* queue_me and wait for wakeup, timeout, or a signal. */
2837 futex_wait_queue_me(hb, &q, to);
2839 /* If we were woken (and unqueued), we succeeded, whatever. */
2841 /* unqueue_me() drops q.key ref */
2842 if (!unqueue_me(&q))
2845 if (to && !to->task)
2849 * We expect signal_pending(current), but we might be the
2850 * victim of a spurious wakeup as well.
2852 if (!signal_pending(current))
2859 restart = ¤t->restart_block;
2860 restart->futex.uaddr = uaddr;
2861 restart->futex.val = val;
2862 restart->futex.time = *abs_time;
2863 restart->futex.bitset = bitset;
2864 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2866 ret = set_restart_fn(restart, futex_wait_restart);
2870 hrtimer_cancel(&to->timer);
2871 destroy_hrtimer_on_stack(&to->timer);
2877 static long futex_wait_restart(struct restart_block *restart)
2879 u32 __user *uaddr = restart->futex.uaddr;
2880 ktime_t t, *tp = NULL;
2882 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2883 t = restart->futex.time;
2886 restart->fn = do_no_restart_syscall;
2888 return (long)futex_wait(uaddr, restart->futex.flags,
2889 restart->futex.val, tp, restart->futex.bitset);
2894 * Userspace tried a 0 -> TID atomic transition of the futex value
2895 * and failed. The kernel side here does the whole locking operation:
2896 * if there are waiters then it will block as a consequence of relying
2897 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2898 * a 0 value of the futex too.).
2900 * Also serves as futex trylock_pi()'ing, and due semantics.
2902 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2903 ktime_t *time, int trylock)
2905 struct hrtimer_sleeper timeout, *to = NULL;
2906 struct task_struct *exiting = NULL;
2907 struct rt_mutex_waiter rt_waiter;
2908 struct futex_hash_bucket *hb;
2909 struct futex_q q = futex_q_init;
2912 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2915 if (refill_pi_state_cache())
2920 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2922 hrtimer_init_sleeper(to, current);
2923 hrtimer_set_expires(&to->timer, *time);
2927 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2928 if (unlikely(ret != 0))
2932 hb = queue_lock(&q);
2934 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2936 if (unlikely(ret)) {
2938 * Atomic work succeeded and we got the lock,
2939 * or failed. Either way, we do _not_ block.
2943 /* We got the lock. */
2945 goto out_unlock_put_key;
2951 * Two reasons for this:
2952 * - EBUSY: Task is exiting and we just wait for the
2954 * - EAGAIN: The user space value changed.
2957 put_futex_key(&q.key);
2959 * Handle the case where the owner is in the middle of
2960 * exiting. Wait for the exit to complete otherwise
2961 * this task might loop forever, aka. live lock.
2963 wait_for_owner_exiting(ret, exiting);
2967 goto out_unlock_put_key;
2971 WARN_ON(!q.pi_state);
2974 * Only actually queue now that the atomic ops are done:
2979 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2980 /* Fixup the trylock return value: */
2981 ret = ret ? 0 : -EWOULDBLOCK;
2985 rt_mutex_init_waiter(&rt_waiter);
2988 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2989 * hold it while doing rt_mutex_start_proxy(), because then it will
2990 * include hb->lock in the blocking chain, even through we'll not in
2991 * fact hold it while blocking. This will lead it to report -EDEADLK
2992 * and BUG when futex_unlock_pi() interleaves with this.
2994 * Therefore acquire wait_lock while holding hb->lock, but drop the
2995 * latter before calling __rt_mutex_start_proxy_lock(). This
2996 * interleaves with futex_unlock_pi() -- which does a similar lock
2997 * handoff -- such that the latter can observe the futex_q::pi_state
2998 * before __rt_mutex_start_proxy_lock() is done.
3000 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
3001 spin_unlock(q.lock_ptr);
3003 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
3004 * such that futex_unlock_pi() is guaranteed to observe the waiter when
3005 * it sees the futex_q::pi_state.
3007 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
3008 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3017 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
3019 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3022 spin_lock(q.lock_ptr);
3024 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3025 * first acquire the hb->lock before removing the lock from the
3026 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3029 * In particular; it is important that futex_unlock_pi() can not
3030 * observe this inconsistency.
3032 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3037 * Fixup the pi_state owner and possibly acquire the lock if we
3040 res = fixup_owner(uaddr, &q, !ret);
3042 * If fixup_owner() returned an error, proprogate that. If it acquired
3043 * the lock, clear our -ETIMEDOUT or -EINTR.
3046 ret = (res < 0) ? res : 0;
3048 /* Unqueue and drop the lock */
3057 put_futex_key(&q.key);
3060 hrtimer_cancel(&to->timer);
3061 destroy_hrtimer_on_stack(&to->timer);
3063 return ret != -EINTR ? ret : -ERESTARTNOINTR;
3068 ret = fault_in_user_writeable(uaddr);
3072 if (!(flags & FLAGS_SHARED))
3075 put_futex_key(&q.key);
3080 * Userspace attempted a TID -> 0 atomic transition, and failed.
3081 * This is the in-kernel slowpath: we look up the PI state (if any),
3082 * and do the rt-mutex unlock.
3084 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3086 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
3087 union futex_key key = FUTEX_KEY_INIT;
3088 struct futex_hash_bucket *hb;
3089 struct futex_q *top_waiter;
3092 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3096 if (get_user(uval, uaddr))
3099 * We release only a lock we actually own:
3101 if ((uval & FUTEX_TID_MASK) != vpid)
3104 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
3108 hb = hash_futex(&key);
3109 spin_lock(&hb->lock);
3112 * Check waiters first. We do not trust user space values at
3113 * all and we at least want to know if user space fiddled
3114 * with the futex value instead of blindly unlocking.
3116 top_waiter = futex_top_waiter(hb, &key);
3118 struct futex_pi_state *pi_state = top_waiter->pi_state;
3125 * If current does not own the pi_state then the futex is
3126 * inconsistent and user space fiddled with the futex value.
3128 if (pi_state->owner != current)
3131 get_pi_state(pi_state);
3133 * By taking wait_lock while still holding hb->lock, we ensure
3134 * there is no point where we hold neither; and therefore
3135 * wake_futex_pi() must observe a state consistent with what we
3138 * In particular; this forces __rt_mutex_start_proxy() to
3139 * complete such that we're guaranteed to observe the
3140 * rt_waiter. Also see the WARN in wake_futex_pi().
3142 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3143 spin_unlock(&hb->lock);
3145 /* drops pi_state->pi_mutex.wait_lock */
3146 ret = wake_futex_pi(uaddr, uval, pi_state);
3148 put_pi_state(pi_state);
3151 * Success, we're done! No tricky corner cases.
3156 * The atomic access to the futex value generated a
3157 * pagefault, so retry the user-access and the wakeup:
3162 * A unconditional UNLOCK_PI op raced against a waiter
3163 * setting the FUTEX_WAITERS bit. Try again.
3168 * wake_futex_pi has detected invalid state. Tell user
3175 * We have no kernel internal state, i.e. no waiters in the
3176 * kernel. Waiters which are about to queue themselves are stuck
3177 * on hb->lock. So we can safely ignore them. We do neither
3178 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3181 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3182 spin_unlock(&hb->lock);
3197 * If uval has changed, let user space handle it.
3199 ret = (curval == uval) ? 0 : -EAGAIN;
3202 spin_unlock(&hb->lock);
3204 put_futex_key(&key);
3208 put_futex_key(&key);
3213 put_futex_key(&key);
3215 ret = fault_in_user_writeable(uaddr);
3223 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3224 * @hb: the hash_bucket futex_q was original enqueued on
3225 * @q: the futex_q woken while waiting to be requeued
3226 * @key2: the futex_key of the requeue target futex
3227 * @timeout: the timeout associated with the wait (NULL if none)
3229 * Detect if the task was woken on the initial futex as opposed to the requeue
3230 * target futex. If so, determine if it was a timeout or a signal that caused
3231 * the wakeup and return the appropriate error code to the caller. Must be
3232 * called with the hb lock held.
3235 * - 0 = no early wakeup detected;
3236 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3239 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3240 struct futex_q *q, union futex_key *key2,
3241 struct hrtimer_sleeper *timeout)
3246 * With the hb lock held, we avoid races while we process the wakeup.
3247 * We only need to hold hb (and not hb2) to ensure atomicity as the
3248 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3249 * It can't be requeued from uaddr2 to something else since we don't
3250 * support a PI aware source futex for requeue.
3252 if (!match_futex(&q->key, key2)) {
3253 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3255 * We were woken prior to requeue by a timeout or a signal.
3256 * Unqueue the futex_q and determine which it was.
3258 plist_del(&q->list, &hb->chain);
3261 /* Handle spurious wakeups gracefully */
3263 if (timeout && !timeout->task)
3265 else if (signal_pending(current))
3266 ret = -ERESTARTNOINTR;
3272 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3273 * @uaddr: the futex we initially wait on (non-pi)
3274 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3275 * the same type, no requeueing from private to shared, etc.
3276 * @val: the expected value of uaddr
3277 * @abs_time: absolute timeout
3278 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3279 * @uaddr2: the pi futex we will take prior to returning to user-space
3281 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3282 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3283 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3284 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3285 * without one, the pi logic would not know which task to boost/deboost, if
3286 * there was a need to.
3288 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3289 * via the following--
3290 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3291 * 2) wakeup on uaddr2 after a requeue
3295 * If 3, cleanup and return -ERESTARTNOINTR.
3297 * If 2, we may then block on trying to take the rt_mutex and return via:
3298 * 5) successful lock
3301 * 8) other lock acquisition failure
3303 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3305 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3311 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3312 u32 val, ktime_t *abs_time, u32 bitset,
3315 struct hrtimer_sleeper timeout, *to = NULL;
3316 struct rt_mutex_waiter rt_waiter;
3317 struct futex_hash_bucket *hb;
3318 union futex_key key2 = FUTEX_KEY_INIT;
3319 struct futex_q q = futex_q_init;
3322 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3325 if (uaddr == uaddr2)
3333 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3334 CLOCK_REALTIME : CLOCK_MONOTONIC,
3336 hrtimer_init_sleeper(to, current);
3337 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3338 current->timer_slack_ns);
3342 * The waiter is allocated on our stack, manipulated by the requeue
3343 * code while we sleep on uaddr.
3345 rt_mutex_init_waiter(&rt_waiter);
3347 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3348 if (unlikely(ret != 0))
3352 q.rt_waiter = &rt_waiter;
3353 q.requeue_pi_key = &key2;
3356 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3359 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3364 * The check above which compares uaddrs is not sufficient for
3365 * shared futexes. We need to compare the keys:
3367 if (match_futex(&q.key, &key2)) {
3373 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3374 futex_wait_queue_me(hb, &q, to);
3376 spin_lock(&hb->lock);
3377 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3378 spin_unlock(&hb->lock);
3383 * In order for us to be here, we know our q.key == key2, and since
3384 * we took the hb->lock above, we also know that futex_requeue() has
3385 * completed and we no longer have to concern ourselves with a wakeup
3386 * race with the atomic proxy lock acquisition by the requeue code. The
3387 * futex_requeue dropped our key1 reference and incremented our key2
3391 /* Check if the requeue code acquired the second futex for us. */
3394 * Got the lock. We might not be the anticipated owner if we
3395 * did a lock-steal - fix up the PI-state in that case.
3397 if (q.pi_state && (q.pi_state->owner != current)) {
3398 spin_lock(q.lock_ptr);
3399 ret = fixup_pi_state_owner(uaddr2, &q, current);
3401 * Drop the reference to the pi state which
3402 * the requeue_pi() code acquired for us.
3404 put_pi_state(q.pi_state);
3405 spin_unlock(q.lock_ptr);
3407 * Adjust the return value. It's either -EFAULT or
3408 * success (1) but the caller expects 0 for success.
3410 ret = ret < 0 ? ret : 0;
3413 struct rt_mutex *pi_mutex;
3416 * We have been woken up by futex_unlock_pi(), a timeout, or a
3417 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3420 WARN_ON(!q.pi_state);
3421 pi_mutex = &q.pi_state->pi_mutex;
3422 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3424 spin_lock(q.lock_ptr);
3425 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3428 debug_rt_mutex_free_waiter(&rt_waiter);
3430 * Fixup the pi_state owner and possibly acquire the lock if we
3433 res = fixup_owner(uaddr2, &q, !ret);
3435 * If fixup_owner() returned an error, proprogate that. If it
3436 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3439 ret = (res < 0) ? res : 0;
3441 /* Unqueue and drop the lock. */
3445 if (ret == -EINTR) {
3447 * We've already been requeued, but cannot restart by calling
3448 * futex_lock_pi() directly. We could restart this syscall, but
3449 * it would detect that the user space "val" changed and return
3450 * -EWOULDBLOCK. Save the overhead of the restart and return
3451 * -EWOULDBLOCK directly.
3457 put_futex_key(&q.key);
3459 put_futex_key(&key2);
3463 hrtimer_cancel(&to->timer);
3464 destroy_hrtimer_on_stack(&to->timer);
3470 * Support for robust futexes: the kernel cleans up held futexes at
3473 * Implementation: user-space maintains a per-thread list of locks it
3474 * is holding. Upon do_exit(), the kernel carefully walks this list,
3475 * and marks all locks that are owned by this thread with the
3476 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3477 * always manipulated with the lock held, so the list is private and
3478 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3479 * field, to allow the kernel to clean up if the thread dies after
3480 * acquiring the lock, but just before it could have added itself to
3481 * the list. There can only be one such pending lock.
3485 * sys_set_robust_list() - Set the robust-futex list head of a task
3486 * @head: pointer to the list-head
3487 * @len: length of the list-head, as userspace expects
3489 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3492 if (!futex_cmpxchg_enabled)
3495 * The kernel knows only one size for now:
3497 if (unlikely(len != sizeof(*head)))
3500 current->robust_list = head;
3506 * sys_get_robust_list() - Get the robust-futex list head of a task
3507 * @pid: pid of the process [zero for current task]
3508 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3509 * @len_ptr: pointer to a length field, the kernel fills in the header size
3511 SYSCALL_DEFINE3(get_robust_list, int, pid,
3512 struct robust_list_head __user * __user *, head_ptr,
3513 size_t __user *, len_ptr)
3515 struct robust_list_head __user *head;
3517 struct task_struct *p;
3519 if (!futex_cmpxchg_enabled)
3528 p = find_task_by_vpid(pid);
3534 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3537 head = p->robust_list;
3540 if (put_user(sizeof(*head), len_ptr))
3542 return put_user(head, head_ptr);
3550 /* Constants for the pending_op argument of handle_futex_death */
3551 #define HANDLE_DEATH_PENDING true
3552 #define HANDLE_DEATH_LIST false
3555 * Process a futex-list entry, check whether it's owned by the
3556 * dying task, and do notification if so:
3558 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3559 bool pi, bool pending_op)
3561 u32 uval, uninitialized_var(nval), mval;
3564 /* Futex address must be 32bit aligned */
3565 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3569 if (get_user(uval, uaddr))
3573 * Special case for regular (non PI) futexes. The unlock path in
3574 * user space has two race scenarios:
3576 * 1. The unlock path releases the user space futex value and
3577 * before it can execute the futex() syscall to wake up
3578 * waiters it is killed.
3580 * 2. A woken up waiter is killed before it can acquire the
3581 * futex in user space.
3583 * In both cases the TID validation below prevents a wakeup of
3584 * potential waiters which can cause these waiters to block
3587 * In both cases the following conditions are met:
3589 * 1) task->robust_list->list_op_pending != NULL
3590 * @pending_op == true
3591 * 2) User space futex value == 0
3592 * 3) Regular futex: @pi == false
3594 * If these conditions are met, it is safe to attempt waking up a
3595 * potential waiter without touching the user space futex value and
3596 * trying to set the OWNER_DIED bit. The user space futex value is
3597 * uncontended and the rest of the user space mutex state is
3598 * consistent, so a woken waiter will just take over the
3599 * uncontended futex. Setting the OWNER_DIED bit would create
3600 * inconsistent state and malfunction of the user space owner died
3603 if (pending_op && !pi && !uval) {
3604 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3608 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3612 * Ok, this dying thread is truly holding a futex
3613 * of interest. Set the OWNER_DIED bit atomically
3614 * via cmpxchg, and if the value had FUTEX_WAITERS
3615 * set, wake up a waiter (if any). (We have to do a
3616 * futex_wake() even if OWNER_DIED is already set -
3617 * to handle the rare but possible case of recursive
3618 * thread-death.) The rest of the cleanup is done in
3621 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3624 * We are not holding a lock here, but we want to have
3625 * the pagefault_disable/enable() protection because
3626 * we want to handle the fault gracefully. If the
3627 * access fails we try to fault in the futex with R/W
3628 * verification via get_user_pages. get_user() above
3629 * does not guarantee R/W access. If that fails we
3630 * give up and leave the futex locked.
3632 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3635 if (fault_in_user_writeable(uaddr))
3653 * Wake robust non-PI futexes here. The wakeup of
3654 * PI futexes happens in exit_pi_state():
3656 if (!pi && (uval & FUTEX_WAITERS))
3657 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3663 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3665 static inline int fetch_robust_entry(struct robust_list __user **entry,
3666 struct robust_list __user * __user *head,
3669 unsigned long uentry;
3671 if (get_user(uentry, (unsigned long __user *)head))
3674 *entry = (void __user *)(uentry & ~1UL);
3681 * Walk curr->robust_list (very carefully, it's a userspace list!)
3682 * and mark any locks found there dead, and notify any waiters.
3684 * We silently return on any sign of list-walking problem.
3686 static void exit_robust_list(struct task_struct *curr)
3688 struct robust_list_head __user *head = curr->robust_list;
3689 struct robust_list __user *entry, *next_entry, *pending;
3690 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3691 unsigned int uninitialized_var(next_pi);
3692 unsigned long futex_offset;
3695 if (!futex_cmpxchg_enabled)
3699 * Fetch the list head (which was registered earlier, via
3700 * sys_set_robust_list()):
3702 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3705 * Fetch the relative futex offset:
3707 if (get_user(futex_offset, &head->futex_offset))
3710 * Fetch any possibly pending lock-add first, and handle it
3713 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3716 next_entry = NULL; /* avoid warning with gcc */
3717 while (entry != &head->list) {
3719 * Fetch the next entry in the list before calling
3720 * handle_futex_death:
3722 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3724 * A pending lock might already be on the list, so
3725 * don't process it twice:
3727 if (entry != pending) {
3728 if (handle_futex_death((void __user *)entry + futex_offset,
3729 curr, pi, HANDLE_DEATH_LIST))
3737 * Avoid excessively long or circular lists:
3746 handle_futex_death((void __user *)pending + futex_offset,
3747 curr, pip, HANDLE_DEATH_PENDING);
3751 static void futex_cleanup(struct task_struct *tsk)
3753 if (unlikely(tsk->robust_list)) {
3754 exit_robust_list(tsk);
3755 tsk->robust_list = NULL;
3758 #ifdef CONFIG_COMPAT
3759 if (unlikely(tsk->compat_robust_list)) {
3760 compat_exit_robust_list(tsk);
3761 tsk->compat_robust_list = NULL;
3765 if (unlikely(!list_empty(&tsk->pi_state_list)))
3766 exit_pi_state_list(tsk);
3770 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3771 * @tsk: task to set the state on
3773 * Set the futex exit state of the task lockless. The futex waiter code
3774 * observes that state when a task is exiting and loops until the task has
3775 * actually finished the futex cleanup. The worst case for this is that the
3776 * waiter runs through the wait loop until the state becomes visible.
3778 * This is called from the recursive fault handling path in do_exit().
3780 * This is best effort. Either the futex exit code has run already or
3781 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3782 * take it over. If not, the problem is pushed back to user space. If the
3783 * futex exit code did not run yet, then an already queued waiter might
3784 * block forever, but there is nothing which can be done about that.
3786 void futex_exit_recursive(struct task_struct *tsk)
3788 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3789 if (tsk->futex_state == FUTEX_STATE_EXITING)
3790 mutex_unlock(&tsk->futex_exit_mutex);
3791 tsk->futex_state = FUTEX_STATE_DEAD;
3794 static void futex_cleanup_begin(struct task_struct *tsk)
3797 * Prevent various race issues against a concurrent incoming waiter
3798 * including live locks by forcing the waiter to block on
3799 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3800 * attach_to_pi_owner().
3802 mutex_lock(&tsk->futex_exit_mutex);
3805 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3807 * This ensures that all subsequent checks of tsk->futex_state in
3808 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3809 * tsk->pi_lock held.
3811 * It guarantees also that a pi_state which was queued right before
3812 * the state change under tsk->pi_lock by a concurrent waiter must
3813 * be observed in exit_pi_state_list().
3815 raw_spin_lock_irq(&tsk->pi_lock);
3816 tsk->futex_state = FUTEX_STATE_EXITING;
3817 raw_spin_unlock_irq(&tsk->pi_lock);
3820 static void futex_cleanup_end(struct task_struct *tsk, int state)
3823 * Lockless store. The only side effect is that an observer might
3824 * take another loop until it becomes visible.
3826 tsk->futex_state = state;
3828 * Drop the exit protection. This unblocks waiters which observed
3829 * FUTEX_STATE_EXITING to reevaluate the state.
3831 mutex_unlock(&tsk->futex_exit_mutex);
3834 void futex_exec_release(struct task_struct *tsk)
3837 * The state handling is done for consistency, but in the case of
3838 * exec() there is no way to prevent futher damage as the PID stays
3839 * the same. But for the unlikely and arguably buggy case that a
3840 * futex is held on exec(), this provides at least as much state
3841 * consistency protection which is possible.
3843 futex_cleanup_begin(tsk);
3846 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3847 * exec a new binary.
3849 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3852 void futex_exit_release(struct task_struct *tsk)
3854 futex_cleanup_begin(tsk);
3856 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3859 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3860 u32 __user *uaddr2, u32 val2, u32 val3)
3862 int cmd = op & FUTEX_CMD_MASK;
3863 unsigned int flags = 0;
3865 if (!(op & FUTEX_PRIVATE_FLAG))
3866 flags |= FLAGS_SHARED;
3868 if (op & FUTEX_CLOCK_REALTIME) {
3869 flags |= FLAGS_CLOCKRT;
3870 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3876 case FUTEX_UNLOCK_PI:
3877 case FUTEX_TRYLOCK_PI:
3878 case FUTEX_WAIT_REQUEUE_PI:
3879 case FUTEX_CMP_REQUEUE_PI:
3880 if (!futex_cmpxchg_enabled)
3886 val3 = FUTEX_BITSET_MATCH_ANY;
3888 case FUTEX_WAIT_BITSET:
3889 return futex_wait(uaddr, flags, val, timeout, val3);
3891 val3 = FUTEX_BITSET_MATCH_ANY;
3893 case FUTEX_WAKE_BITSET:
3894 return futex_wake(uaddr, flags, val, val3);
3896 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3897 case FUTEX_CMP_REQUEUE:
3898 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3900 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3902 return futex_lock_pi(uaddr, flags, timeout, 0);
3903 case FUTEX_UNLOCK_PI:
3904 return futex_unlock_pi(uaddr, flags);
3905 case FUTEX_TRYLOCK_PI:
3906 return futex_lock_pi(uaddr, flags, NULL, 1);
3907 case FUTEX_WAIT_REQUEUE_PI:
3908 val3 = FUTEX_BITSET_MATCH_ANY;
3909 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3911 case FUTEX_CMP_REQUEUE_PI:
3912 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3918 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3919 struct timespec __user *, utime, u32 __user *, uaddr2,
3923 ktime_t t, *tp = NULL;
3925 int cmd = op & FUTEX_CMD_MASK;
3927 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3928 cmd == FUTEX_WAIT_BITSET ||
3929 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3930 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3932 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3934 if (!timespec_valid(&ts))
3937 t = timespec_to_ktime(ts);
3938 if (cmd == FUTEX_WAIT)
3939 t = ktime_add_safe(ktime_get(), t);
3943 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3944 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3946 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3947 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3948 val2 = (u32) (unsigned long) utime;
3950 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3953 #ifdef CONFIG_COMPAT
3955 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3958 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3959 compat_uptr_t __user *head, unsigned int *pi)
3961 if (get_user(*uentry, head))
3964 *entry = compat_ptr((*uentry) & ~1);
3965 *pi = (unsigned int)(*uentry) & 1;
3970 static void __user *futex_uaddr(struct robust_list __user *entry,
3971 compat_long_t futex_offset)
3973 compat_uptr_t base = ptr_to_compat(entry);
3974 void __user *uaddr = compat_ptr(base + futex_offset);
3980 * Walk curr->robust_list (very carefully, it's a userspace list!)
3981 * and mark any locks found there dead, and notify any waiters.
3983 * We silently return on any sign of list-walking problem.
3985 static void compat_exit_robust_list(struct task_struct *curr)
3987 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3988 struct robust_list __user *entry, *next_entry, *pending;
3989 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3990 unsigned int uninitialized_var(next_pi);
3991 compat_uptr_t uentry, next_uentry, upending;
3992 compat_long_t futex_offset;
3995 if (!futex_cmpxchg_enabled)
3999 * Fetch the list head (which was registered earlier, via
4000 * sys_set_robust_list()):
4002 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
4005 * Fetch the relative futex offset:
4007 if (get_user(futex_offset, &head->futex_offset))
4010 * Fetch any possibly pending lock-add first, and handle it
4013 if (compat_fetch_robust_entry(&upending, &pending,
4014 &head->list_op_pending, &pip))
4017 next_entry = NULL; /* avoid warning with gcc */
4018 while (entry != (struct robust_list __user *) &head->list) {
4020 * Fetch the next entry in the list before calling
4021 * handle_futex_death:
4023 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
4024 (compat_uptr_t __user *)&entry->next, &next_pi);
4026 * A pending lock might already be on the list, so
4027 * dont process it twice:
4029 if (entry != pending) {
4030 void __user *uaddr = futex_uaddr(entry, futex_offset);
4032 if (handle_futex_death(uaddr, curr, pi,
4038 uentry = next_uentry;
4042 * Avoid excessively long or circular lists:
4050 void __user *uaddr = futex_uaddr(pending, futex_offset);
4052 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
4056 COMPAT_SYSCALL_DEFINE2(set_robust_list,
4057 struct compat_robust_list_head __user *, head,
4060 if (!futex_cmpxchg_enabled)
4063 if (unlikely(len != sizeof(*head)))
4066 current->compat_robust_list = head;
4071 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
4072 compat_uptr_t __user *, head_ptr,
4073 compat_size_t __user *, len_ptr)
4075 struct compat_robust_list_head __user *head;
4077 struct task_struct *p;
4079 if (!futex_cmpxchg_enabled)
4088 p = find_task_by_vpid(pid);
4094 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
4097 head = p->compat_robust_list;
4100 if (put_user(sizeof(*head), len_ptr))
4102 return put_user(ptr_to_compat(head), head_ptr);
4110 COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
4111 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
4115 ktime_t t, *tp = NULL;
4117 int cmd = op & FUTEX_CMD_MASK;
4119 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4120 cmd == FUTEX_WAIT_BITSET ||
4121 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4122 if (compat_get_timespec(&ts, utime))
4124 if (!timespec_valid(&ts))
4127 t = timespec_to_ktime(ts);
4128 if (cmd == FUTEX_WAIT)
4129 t = ktime_add_safe(ktime_get(), t);
4132 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4133 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4134 val2 = (int) (unsigned long) utime;
4136 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4138 #endif /* CONFIG_COMPAT */
4140 static void __init futex_detect_cmpxchg(void)
4142 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4146 * This will fail and we want it. Some arch implementations do
4147 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4148 * functionality. We want to know that before we call in any
4149 * of the complex code paths. Also we want to prevent
4150 * registration of robust lists in that case. NULL is
4151 * guaranteed to fault and we get -EFAULT on functional
4152 * implementation, the non-functional ones will return
4155 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4156 futex_cmpxchg_enabled = 1;
4160 static int __init futex_init(void)
4162 unsigned int futex_shift;
4165 #if CONFIG_BASE_SMALL
4166 futex_hashsize = 16;
4168 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4171 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4173 futex_hashsize < 256 ? HASH_SMALL : 0,
4175 futex_hashsize, futex_hashsize);
4176 futex_hashsize = 1UL << futex_shift;
4178 futex_detect_cmpxchg();
4180 for (i = 0; i < futex_hashsize; i++) {
4181 atomic_set(&futex_queues[i].waiters, 0);
4182 plist_head_init(&futex_queues[i].chain);
4183 spin_lock_init(&futex_queues[i].lock);
4188 core_initcall(futex_init);