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/hugetlb.h>
66 #include <linux/freezer.h>
67 #include <linux/bootmem.h>
68 #include <linux/fault-inject.h>
70 #include <asm/futex.h>
72 #include "locking/rtmutex_common.h"
75 * READ this before attempting to hack on futexes!
77 * Basic futex operation and ordering guarantees
78 * =============================================
80 * The waiter reads the futex value in user space and calls
81 * futex_wait(). This function computes the hash bucket and acquires
82 * the hash bucket lock. After that it reads the futex user space value
83 * again and verifies that the data has not changed. If it has not changed
84 * it enqueues itself into the hash bucket, releases the hash bucket lock
87 * The waker side modifies the user space value of the futex and calls
88 * futex_wake(). This function computes the hash bucket and acquires the
89 * hash bucket lock. Then it looks for waiters on that futex in the hash
90 * bucket and wakes them.
92 * In futex wake up scenarios where no tasks are blocked on a futex, taking
93 * the hb spinlock can be avoided and simply return. In order for this
94 * optimization to work, ordering guarantees must exist so that the waiter
95 * being added to the list is acknowledged when the list is concurrently being
96 * checked by the waker, avoiding scenarios like the following:
100 * sys_futex(WAIT, futex, val);
101 * futex_wait(futex, val);
104 * sys_futex(WAKE, futex);
109 * lock(hash_bucket(futex));
111 * unlock(hash_bucket(futex));
114 * This would cause the waiter on CPU 0 to wait forever because it
115 * missed the transition of the user space value from val to newval
116 * and the waker did not find the waiter in the hash bucket queue.
118 * The correct serialization ensures that a waiter either observes
119 * the changed user space value before blocking or is woken by a
124 * sys_futex(WAIT, futex, val);
125 * futex_wait(futex, val);
128 * smp_mb(); (A) <-- paired with -.
130 * lock(hash_bucket(futex)); |
134 * | sys_futex(WAKE, futex);
135 * | futex_wake(futex);
137 * `--------> smp_mb(); (B)
140 * unlock(hash_bucket(futex));
141 * schedule(); if (waiters)
142 * lock(hash_bucket(futex));
143 * else wake_waiters(futex);
144 * waiters--; (b) unlock(hash_bucket(futex));
146 * Where (A) orders the waiters increment and the futex value read through
147 * atomic operations (see hb_waiters_inc) and where (B) orders the write
148 * to futex and the waiters read -- this is done by the barriers for both
149 * shared and private futexes in get_futex_key_refs().
151 * This yields the following case (where X:=waiters, Y:=futex):
159 * Which guarantees that x==0 && y==0 is impossible; which translates back into
160 * the guarantee that we cannot both miss the futex variable change and the
163 * Note that a new waiter is accounted for in (a) even when it is possible that
164 * the wait call can return error, in which case we backtrack from it in (b).
165 * Refer to the comment in queue_lock().
167 * Similarly, in order to account for waiters being requeued on another
168 * address we always increment the waiters for the destination bucket before
169 * acquiring the lock. It then decrements them again after releasing it -
170 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
171 * will do the additional required waiter count housekeeping. This is done for
172 * double_lock_hb() and double_unlock_hb(), respectively.
175 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
176 #define futex_cmpxchg_enabled 1
178 static int __read_mostly futex_cmpxchg_enabled;
182 * Futex flags used to encode options to functions and preserve them across
186 # define FLAGS_SHARED 0x01
189 * NOMMU does not have per process address space. Let the compiler optimize
192 # define FLAGS_SHARED 0x00
194 #define FLAGS_CLOCKRT 0x02
195 #define FLAGS_HAS_TIMEOUT 0x04
198 * Priority Inheritance state:
200 struct futex_pi_state {
202 * list of 'owned' pi_state instances - these have to be
203 * cleaned up in do_exit() if the task exits prematurely:
205 struct list_head list;
210 struct rt_mutex pi_mutex;
212 struct task_struct *owner;
219 * struct futex_q - The hashed futex queue entry, one per waiting task
220 * @list: priority-sorted list of tasks waiting on this futex
221 * @task: the task waiting on the futex
222 * @lock_ptr: the hash bucket lock
223 * @key: the key the futex is hashed on
224 * @pi_state: optional priority inheritance state
225 * @rt_waiter: rt_waiter storage for use with requeue_pi
226 * @requeue_pi_key: the requeue_pi target futex key
227 * @bitset: bitset for the optional bitmasked wakeup
229 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
230 * we can wake only the relevant ones (hashed queues may be shared).
232 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
233 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
234 * The order of wakeup is always to make the first condition true, then
237 * PI futexes are typically woken before they are removed from the hash list via
238 * the rt_mutex code. See unqueue_me_pi().
241 struct plist_node list;
243 struct task_struct *task;
244 spinlock_t *lock_ptr;
246 struct futex_pi_state *pi_state;
247 struct rt_mutex_waiter *rt_waiter;
248 union futex_key *requeue_pi_key;
252 static const struct futex_q futex_q_init = {
253 /* list gets initialized in queue_me()*/
254 .key = FUTEX_KEY_INIT,
255 .bitset = FUTEX_BITSET_MATCH_ANY
259 * Hash buckets are shared by all the futex_keys that hash to the same
260 * location. Each key may have multiple futex_q structures, one for each task
261 * waiting on a futex.
263 struct futex_hash_bucket {
266 struct plist_head chain;
267 } ____cacheline_aligned_in_smp;
270 * The base of the bucket array and its size are always used together
271 * (after initialization only in hash_futex()), so ensure that they
272 * reside in the same cacheline.
275 struct futex_hash_bucket *queues;
276 unsigned long hashsize;
277 } __futex_data __read_mostly __aligned(2*sizeof(long));
278 #define futex_queues (__futex_data.queues)
279 #define futex_hashsize (__futex_data.hashsize)
283 * Fault injections for futexes.
285 #ifdef CONFIG_FAIL_FUTEX
288 struct fault_attr attr;
292 .attr = FAULT_ATTR_INITIALIZER,
293 .ignore_private = false,
296 static int __init setup_fail_futex(char *str)
298 return setup_fault_attr(&fail_futex.attr, str);
300 __setup("fail_futex=", setup_fail_futex);
302 static bool should_fail_futex(bool fshared)
304 if (fail_futex.ignore_private && !fshared)
307 return should_fail(&fail_futex.attr, 1);
310 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
312 static int __init fail_futex_debugfs(void)
314 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
317 dir = fault_create_debugfs_attr("fail_futex", NULL,
322 if (!debugfs_create_bool("ignore-private", mode, dir,
323 &fail_futex.ignore_private)) {
324 debugfs_remove_recursive(dir);
331 late_initcall(fail_futex_debugfs);
333 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
336 static inline bool should_fail_futex(bool fshared)
340 #endif /* CONFIG_FAIL_FUTEX */
343 static void compat_exit_robust_list(struct task_struct *curr);
345 static inline void compat_exit_robust_list(struct task_struct *curr) { }
348 static inline void futex_get_mm(union futex_key *key)
350 atomic_inc(&key->private.mm->mm_count);
352 * Ensure futex_get_mm() implies a full barrier such that
353 * get_futex_key() implies a full barrier. This is relied upon
354 * as smp_mb(); (B), see the ordering comment above.
356 smp_mb__after_atomic();
360 * Reflects a new waiter being added to the waitqueue.
362 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
365 atomic_inc(&hb->waiters);
367 * Full barrier (A), see the ordering comment above.
369 smp_mb__after_atomic();
374 * Reflects a waiter being removed from the waitqueue by wakeup
377 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
380 atomic_dec(&hb->waiters);
384 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
387 return atomic_read(&hb->waiters);
394 * hash_futex - Return the hash bucket in the global hash
395 * @key: Pointer to the futex key for which the hash is calculated
397 * We hash on the keys returned from get_futex_key (see below) and return the
398 * corresponding hash bucket in the global hash.
400 static struct futex_hash_bucket *hash_futex(union futex_key *key)
402 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
405 return &futex_queues[hash & (futex_hashsize - 1)];
410 * match_futex - Check whether two futex keys are equal
411 * @key1: Pointer to key1
412 * @key2: Pointer to key2
414 * Return 1 if two futex_keys are equal, 0 otherwise.
416 static inline int match_futex(union futex_key *key1, union futex_key *key2)
419 && key1->both.word == key2->both.word
420 && key1->both.ptr == key2->both.ptr
421 && key1->both.offset == key2->both.offset);
425 * Take a reference to the resource addressed by a key.
426 * Can be called while holding spinlocks.
429 static void get_futex_key_refs(union futex_key *key)
435 * On MMU less systems futexes are always "private" as there is no per
436 * process address space. We need the smp wmb nevertheless - yes,
437 * arch/blackfin has MMU less SMP ...
439 if (!IS_ENABLED(CONFIG_MMU)) {
440 smp_mb(); /* explicit smp_mb(); (B) */
444 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
446 smp_mb(); /* explicit smp_mb(); (B) */
448 case FUT_OFF_MMSHARED:
449 futex_get_mm(key); /* implies smp_mb(); (B) */
453 * Private futexes do not hold reference on an inode or
454 * mm, therefore the only purpose of calling get_futex_key_refs
455 * is because we need the barrier for the lockless waiter check.
457 smp_mb(); /* explicit smp_mb(); (B) */
462 * Drop a reference to the resource addressed by a key.
463 * The hash bucket spinlock must not be held. This is
464 * a no-op for private futexes, see comment in the get
467 static void drop_futex_key_refs(union futex_key *key)
469 if (!key->both.ptr) {
470 /* If we're here then we tried to put a key we failed to get */
475 if (!IS_ENABLED(CONFIG_MMU))
478 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
481 case FUT_OFF_MMSHARED:
482 mmdrop(key->private.mm);
488 * Generate a machine wide unique identifier for this inode.
490 * This relies on u64 not wrapping in the life-time of the machine; which with
491 * 1ns resolution means almost 585 years.
493 * This further relies on the fact that a well formed program will not unmap
494 * the file while it has a (shared) futex waiting on it. This mapping will have
495 * a file reference which pins the mount and inode.
497 * If for some reason an inode gets evicted and read back in again, it will get
498 * a new sequence number and will _NOT_ match, even though it is the exact same
501 * It is important that match_futex() will never have a false-positive, esp.
502 * for PI futexes that can mess up the state. The above argues that false-negatives
503 * are only possible for malformed programs.
505 static u64 get_inode_sequence_number(struct inode *inode)
507 static atomic64_t i_seq;
510 /* Does the inode already have a sequence number? */
511 old = atomic64_read(&inode->i_sequence);
516 u64 new = atomic64_add_return(1, &i_seq);
517 if (WARN_ON_ONCE(!new))
520 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
528 * get_futex_key() - Get parameters which are the keys for a futex
529 * @uaddr: virtual address of the futex
530 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
531 * @key: address where result is stored.
532 * @rw: mapping needs to be read/write (values: VERIFY_READ,
535 * Return: a negative error code or 0
537 * The key words are stored in *key on success.
539 * For shared mappings (when @fshared), the key is:
540 * ( inode->i_sequence, page->index, offset_within_page )
541 * [ also see get_inode_sequence_number() ]
543 * For private mappings (or when !@fshared), the key is:
544 * ( current->mm, address, 0 )
546 * This allows (cross process, where applicable) identification of the futex
547 * without keeping the page pinned for the duration of the FUTEX_WAIT.
549 * lock_page() might sleep, the caller should not hold a spinlock.
552 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
554 unsigned long address = (unsigned long)uaddr;
555 struct mm_struct *mm = current->mm;
556 struct page *page, *tail;
557 struct address_space *mapping;
561 * The futex address must be "naturally" aligned.
563 key->both.offset = address % PAGE_SIZE;
564 if (unlikely((address % sizeof(u32)) != 0))
566 address -= key->both.offset;
568 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
571 if (unlikely(should_fail_futex(fshared)))
575 * PROCESS_PRIVATE futexes are fast.
576 * As the mm cannot disappear under us and the 'key' only needs
577 * virtual address, we dont even have to find the underlying vma.
578 * Note : We do have to check 'uaddr' is a valid user address,
579 * but access_ok() should be faster than find_vma()
582 key->private.mm = mm;
583 key->private.address = address;
584 get_futex_key_refs(key); /* implies smp_mb(); (B) */
589 /* Ignore any VERIFY_READ mapping (futex common case) */
590 if (unlikely(should_fail_futex(fshared)))
593 err = get_user_pages_fast(address, 1, 1, &page);
595 * If write access is not required (eg. FUTEX_WAIT), try
596 * and get read-only access.
598 if (err == -EFAULT && rw == VERIFY_READ) {
599 err = get_user_pages_fast(address, 1, 0, &page);
608 * The treatment of mapping from this point on is critical. The page
609 * lock protects many things but in this context the page lock
610 * stabilizes mapping, prevents inode freeing in the shared
611 * file-backed region case and guards against movement to swap cache.
613 * Strictly speaking the page lock is not needed in all cases being
614 * considered here and page lock forces unnecessarily serialization
615 * From this point on, mapping will be re-verified if necessary and
616 * page lock will be acquired only if it is unavoidable
618 * Mapping checks require the head page for any compound page so the
619 * head page and mapping is looked up now. For anonymous pages, it
620 * does not matter if the page splits in the future as the key is
621 * based on the address. For filesystem-backed pages, the tail is
622 * required as the index of the page determines the key. For
623 * base pages, there is no tail page and tail == page.
626 page = compound_head(page);
627 mapping = READ_ONCE(page->mapping);
630 * If page->mapping is NULL, then it cannot be a PageAnon
631 * page; but it might be the ZERO_PAGE or in the gate area or
632 * in a special mapping (all cases which we are happy to fail);
633 * or it may have been a good file page when get_user_pages_fast
634 * found it, but truncated or holepunched or subjected to
635 * invalidate_complete_page2 before we got the page lock (also
636 * cases which we are happy to fail). And we hold a reference,
637 * so refcount care in invalidate_complete_page's remove_mapping
638 * prevents drop_caches from setting mapping to NULL beneath us.
640 * The case we do have to guard against is when memory pressure made
641 * shmem_writepage move it from filecache to swapcache beneath us:
642 * an unlikely race, but we do need to retry for page->mapping.
644 if (unlikely(!mapping)) {
648 * Page lock is required to identify which special case above
649 * applies. If this is really a shmem page then the page lock
650 * will prevent unexpected transitions.
653 shmem_swizzled = PageSwapCache(page) || page->mapping;
664 * Private mappings are handled in a simple way.
666 * If the futex key is stored on an anonymous page, then the associated
667 * object is the mm which is implicitly pinned by the calling process.
669 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
670 * it's a read-only handle, it's expected that futexes attach to
671 * the object not the particular process.
673 if (PageAnon(page)) {
675 * A RO anonymous page will never change and thus doesn't make
676 * sense for futex operations.
678 if (unlikely(should_fail_futex(fshared)) || ro) {
683 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
684 key->private.mm = mm;
685 key->private.address = address;
691 * The associated futex object in this case is the inode and
692 * the page->mapping must be traversed. Ordinarily this should
693 * be stabilised under page lock but it's not strictly
694 * necessary in this case as we just want to pin the inode, not
695 * update the radix tree or anything like that.
697 * The RCU read lock is taken as the inode is finally freed
698 * under RCU. If the mapping still matches expectations then the
699 * mapping->host can be safely accessed as being a valid inode.
703 if (READ_ONCE(page->mapping) != mapping) {
710 inode = READ_ONCE(mapping->host);
718 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
719 key->shared.i_seq = get_inode_sequence_number(inode);
720 key->shared.pgoff = page_to_pgoff(tail);
724 get_futex_key_refs(key); /* implies smp_mb(); (B) */
731 static inline void put_futex_key(union futex_key *key)
733 drop_futex_key_refs(key);
737 * fault_in_user_writeable() - Fault in user address and verify RW access
738 * @uaddr: pointer to faulting user space address
740 * Slow path to fixup the fault we just took in the atomic write
743 * We have no generic implementation of a non-destructive write to the
744 * user address. We know that we faulted in the atomic pagefault
745 * disabled section so we can as well avoid the #PF overhead by
746 * calling get_user_pages() right away.
748 static int fault_in_user_writeable(u32 __user *uaddr)
750 struct mm_struct *mm = current->mm;
753 down_read(&mm->mmap_sem);
754 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
755 FAULT_FLAG_WRITE, NULL);
756 up_read(&mm->mmap_sem);
758 return ret < 0 ? ret : 0;
762 * futex_top_waiter() - Return the highest priority waiter on a futex
763 * @hb: the hash bucket the futex_q's reside in
764 * @key: the futex key (to distinguish it from other futex futex_q's)
766 * Must be called with the hb lock held.
768 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
769 union futex_key *key)
771 struct futex_q *this;
773 plist_for_each_entry(this, &hb->chain, list) {
774 if (match_futex(&this->key, key))
780 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
781 u32 uval, u32 newval)
786 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
792 static int get_futex_value_locked(u32 *dest, u32 __user *from)
797 ret = __get_user(*dest, from);
800 return ret ? -EFAULT : 0;
807 static int refill_pi_state_cache(void)
809 struct futex_pi_state *pi_state;
811 if (likely(current->pi_state_cache))
814 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
819 INIT_LIST_HEAD(&pi_state->list);
820 /* pi_mutex gets initialized later */
821 pi_state->owner = NULL;
822 atomic_set(&pi_state->refcount, 1);
823 pi_state->key = FUTEX_KEY_INIT;
825 current->pi_state_cache = pi_state;
830 static struct futex_pi_state *alloc_pi_state(void)
832 struct futex_pi_state *pi_state = current->pi_state_cache;
835 current->pi_state_cache = NULL;
840 static void pi_state_update_owner(struct futex_pi_state *pi_state,
841 struct task_struct *new_owner)
843 struct task_struct *old_owner = pi_state->owner;
845 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
848 raw_spin_lock(&old_owner->pi_lock);
849 WARN_ON(list_empty(&pi_state->list));
850 list_del_init(&pi_state->list);
851 raw_spin_unlock(&old_owner->pi_lock);
855 raw_spin_lock(&new_owner->pi_lock);
856 WARN_ON(!list_empty(&pi_state->list));
857 list_add(&pi_state->list, &new_owner->pi_state_list);
858 pi_state->owner = new_owner;
859 raw_spin_unlock(&new_owner->pi_lock);
863 static void get_pi_state(struct futex_pi_state *pi_state)
865 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
869 * Drops a reference to the pi_state object and frees or caches it
870 * when the last reference is gone.
872 static void put_pi_state(struct futex_pi_state *pi_state)
877 if (!atomic_dec_and_test(&pi_state->refcount))
881 * If pi_state->owner is NULL, the owner is most probably dying
882 * and has cleaned up the pi_state already
884 if (pi_state->owner) {
887 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
888 pi_state_update_owner(pi_state, NULL);
889 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
890 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
893 if (current->pi_state_cache) {
897 * pi_state->list is already empty.
898 * clear pi_state->owner.
899 * refcount is at 0 - put it back to 1.
901 pi_state->owner = NULL;
902 atomic_set(&pi_state->refcount, 1);
903 current->pi_state_cache = pi_state;
908 * Look up the task based on what TID userspace gave us.
911 static struct task_struct *futex_find_get_task(pid_t pid)
913 struct task_struct *p;
916 p = find_task_by_vpid(pid);
926 * This task is holding PI mutexes at exit time => bad.
927 * Kernel cleans up PI-state, but userspace is likely hosed.
928 * (Robust-futex cleanup is separate and might save the day for userspace.)
930 static void exit_pi_state_list(struct task_struct *curr)
932 struct list_head *next, *head = &curr->pi_state_list;
933 struct futex_pi_state *pi_state;
934 struct futex_hash_bucket *hb;
935 union futex_key key = FUTEX_KEY_INIT;
937 if (!futex_cmpxchg_enabled)
940 * We are a ZOMBIE and nobody can enqueue itself on
941 * pi_state_list anymore, but we have to be careful
942 * versus waiters unqueueing themselves:
944 raw_spin_lock_irq(&curr->pi_lock);
945 while (!list_empty(head)) {
947 pi_state = list_entry(next, struct futex_pi_state, list);
949 hb = hash_futex(&key);
952 * We can race against put_pi_state() removing itself from the
953 * list (a waiter going away). put_pi_state() will first
954 * decrement the reference count and then modify the list, so
955 * its possible to see the list entry but fail this reference
958 * In that case; drop the locks to let put_pi_state() make
959 * progress and retry the loop.
961 if (!atomic_inc_not_zero(&pi_state->refcount)) {
962 raw_spin_unlock_irq(&curr->pi_lock);
964 raw_spin_lock_irq(&curr->pi_lock);
967 raw_spin_unlock_irq(&curr->pi_lock);
969 spin_lock(&hb->lock);
970 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
971 raw_spin_lock(&curr->pi_lock);
973 * We dropped the pi-lock, so re-check whether this
974 * task still owns the PI-state:
976 if (head->next != next) {
977 /* retain curr->pi_lock for the loop invariant */
978 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
979 spin_unlock(&hb->lock);
980 put_pi_state(pi_state);
984 WARN_ON(pi_state->owner != curr);
985 WARN_ON(list_empty(&pi_state->list));
986 list_del_init(&pi_state->list);
987 pi_state->owner = NULL;
989 raw_spin_unlock(&curr->pi_lock);
990 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
991 spin_unlock(&hb->lock);
993 rt_mutex_futex_unlock(&pi_state->pi_mutex);
994 put_pi_state(pi_state);
996 raw_spin_lock_irq(&curr->pi_lock);
998 raw_spin_unlock_irq(&curr->pi_lock);
1002 * We need to check the following states:
1004 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
1006 * [1] NULL | --- | --- | 0 | 0/1 | Valid
1007 * [2] NULL | --- | --- | >0 | 0/1 | Valid
1009 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1011 * [4] Found | Found | NULL | 0 | 1 | Valid
1012 * [5] Found | Found | NULL | >0 | 1 | Invalid
1014 * [6] Found | Found | task | 0 | 1 | Valid
1016 * [7] Found | Found | NULL | Any | 0 | Invalid
1018 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1019 * [9] Found | Found | task | 0 | 0 | Invalid
1020 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1022 * [1] Indicates that the kernel can acquire the futex atomically. We
1023 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1025 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1026 * thread is found then it indicates that the owner TID has died.
1028 * [3] Invalid. The waiter is queued on a non PI futex
1030 * [4] Valid state after exit_robust_list(), which sets the user space
1031 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1033 * [5] The user space value got manipulated between exit_robust_list()
1034 * and exit_pi_state_list()
1036 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1037 * the pi_state but cannot access the user space value.
1039 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1041 * [8] Owner and user space value match
1043 * [9] There is no transient state which sets the user space TID to 0
1044 * except exit_robust_list(), but this is indicated by the
1045 * FUTEX_OWNER_DIED bit. See [4]
1047 * [10] There is no transient state which leaves owner and user space
1048 * TID out of sync. Except one error case where the kernel is denied
1049 * write access to the user address, see fixup_pi_state_owner().
1052 * Serialization and lifetime rules:
1056 * hb -> futex_q, relation
1057 * futex_q -> pi_state, relation
1059 * (cannot be raw because hb can contain arbitrary amount
1062 * pi_mutex->wait_lock:
1066 * (and pi_mutex 'obviously')
1070 * p->pi_state_list -> pi_state->list, relation
1072 * pi_state->refcount:
1080 * pi_mutex->wait_lock
1086 * Validate that the existing waiter has a pi_state and sanity check
1087 * the pi_state against the user space value. If correct, attach to
1090 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1091 struct futex_pi_state *pi_state,
1092 struct futex_pi_state **ps)
1094 pid_t pid = uval & FUTEX_TID_MASK;
1098 * Userspace might have messed up non-PI and PI futexes [3]
1100 if (unlikely(!pi_state))
1104 * We get here with hb->lock held, and having found a
1105 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1106 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1107 * which in turn means that futex_lock_pi() still has a reference on
1110 * The waiter holding a reference on @pi_state also protects against
1111 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1112 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1113 * free pi_state before we can take a reference ourselves.
1115 WARN_ON(!atomic_read(&pi_state->refcount));
1118 * Now that we have a pi_state, we can acquire wait_lock
1119 * and do the state validation.
1121 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1124 * Since {uval, pi_state} is serialized by wait_lock, and our current
1125 * uval was read without holding it, it can have changed. Verify it
1126 * still is what we expect it to be, otherwise retry the entire
1129 if (get_futex_value_locked(&uval2, uaddr))
1136 * Handle the owner died case:
1138 if (uval & FUTEX_OWNER_DIED) {
1140 * exit_pi_state_list sets owner to NULL and wakes the
1141 * topmost waiter. The task which acquires the
1142 * pi_state->rt_mutex will fixup owner.
1144 if (!pi_state->owner) {
1146 * No pi state owner, but the user space TID
1147 * is not 0. Inconsistent state. [5]
1152 * Take a ref on the state and return success. [4]
1158 * If TID is 0, then either the dying owner has not
1159 * yet executed exit_pi_state_list() or some waiter
1160 * acquired the rtmutex in the pi state, but did not
1161 * yet fixup the TID in user space.
1163 * Take a ref on the state and return success. [6]
1169 * If the owner died bit is not set, then the pi_state
1170 * must have an owner. [7]
1172 if (!pi_state->owner)
1177 * Bail out if user space manipulated the futex value. If pi
1178 * state exists then the owner TID must be the same as the
1179 * user space TID. [9/10]
1181 if (pid != task_pid_vnr(pi_state->owner))
1185 get_pi_state(pi_state);
1186 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1203 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1208 * wait_for_owner_exiting - Block until the owner has exited
1209 * @exiting: Pointer to the exiting task
1211 * Caller must hold a refcount on @exiting.
1213 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1215 if (ret != -EBUSY) {
1216 WARN_ON_ONCE(exiting);
1220 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1223 mutex_lock(&exiting->futex_exit_mutex);
1225 * No point in doing state checking here. If the waiter got here
1226 * while the task was in exec()->exec_futex_release() then it can
1227 * have any FUTEX_STATE_* value when the waiter has acquired the
1228 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1229 * already. Highly unlikely and not a problem. Just one more round
1230 * through the futex maze.
1232 mutex_unlock(&exiting->futex_exit_mutex);
1234 put_task_struct(exiting);
1237 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1238 struct task_struct *tsk)
1243 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1244 * caller that the alleged owner is busy.
1246 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1250 * Reread the user space value to handle the following situation:
1254 * sys_exit() sys_futex()
1255 * do_exit() futex_lock_pi()
1256 * futex_lock_pi_atomic()
1257 * exit_signals(tsk) No waiters:
1258 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1259 * mm_release(tsk) Set waiter bit
1260 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1261 * Set owner died attach_to_pi_owner() {
1262 * *uaddr = 0xC0000000; tsk = get_task(PID);
1263 * } if (!tsk->flags & PF_EXITING) {
1265 * tsk->futex_state = } else {
1266 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1269 * return -ESRCH; <--- FAIL
1272 * Returning ESRCH unconditionally is wrong here because the
1273 * user space value has been changed by the exiting task.
1275 * The same logic applies to the case where the exiting task is
1278 if (get_futex_value_locked(&uval2, uaddr))
1281 /* If the user space value has changed, try again. */
1286 * The exiting task did not have a robust list, the robust list was
1287 * corrupted or the user space value in *uaddr is simply bogus.
1288 * Give up and tell user space.
1294 * Lookup the task for the TID provided from user space and attach to
1295 * it after doing proper sanity checks.
1297 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1298 struct futex_pi_state **ps,
1299 struct task_struct **exiting)
1301 pid_t pid = uval & FUTEX_TID_MASK;
1302 struct futex_pi_state *pi_state;
1303 struct task_struct *p;
1306 * We are the first waiter - try to look up the real owner and attach
1307 * the new pi_state to it, but bail out when TID = 0 [1]
1309 * The !pid check is paranoid. None of the call sites should end up
1310 * with pid == 0, but better safe than sorry. Let the caller retry
1314 p = futex_find_get_task(pid);
1316 return handle_exit_race(uaddr, uval, NULL);
1318 if (unlikely(p->flags & PF_KTHREAD)) {
1324 * We need to look at the task state to figure out, whether the
1325 * task is exiting. To protect against the change of the task state
1326 * in futex_exit_release(), we do this protected by p->pi_lock:
1328 raw_spin_lock_irq(&p->pi_lock);
1329 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1331 * The task is on the way out. When the futex state is
1332 * FUTEX_STATE_DEAD, we know that the task has finished
1335 int ret = handle_exit_race(uaddr, uval, p);
1337 raw_spin_unlock_irq(&p->pi_lock);
1339 * If the owner task is between FUTEX_STATE_EXITING and
1340 * FUTEX_STATE_DEAD then store the task pointer and keep
1341 * the reference on the task struct. The calling code will
1342 * drop all locks, wait for the task to reach
1343 * FUTEX_STATE_DEAD and then drop the refcount. This is
1344 * required to prevent a live lock when the current task
1345 * preempted the exiting task between the two states.
1355 * No existing pi state. First waiter. [2]
1357 * This creates pi_state, we have hb->lock held, this means nothing can
1358 * observe this state, wait_lock is irrelevant.
1360 pi_state = alloc_pi_state();
1363 * Initialize the pi_mutex in locked state and make @p
1366 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1368 /* Store the key for possible exit cleanups: */
1369 pi_state->key = *key;
1371 WARN_ON(!list_empty(&pi_state->list));
1372 list_add(&pi_state->list, &p->pi_state_list);
1374 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1375 * because there is no concurrency as the object is not published yet.
1377 pi_state->owner = p;
1378 raw_spin_unlock_irq(&p->pi_lock);
1387 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1388 struct futex_hash_bucket *hb,
1389 union futex_key *key, struct futex_pi_state **ps,
1390 struct task_struct **exiting)
1392 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1395 * If there is a waiter on that futex, validate it and
1396 * attach to the pi_state when the validation succeeds.
1399 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1402 * We are the first waiter - try to look up the owner based on
1403 * @uval and attach to it.
1405 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1408 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1411 u32 uninitialized_var(curval);
1413 if (unlikely(should_fail_futex(true)))
1416 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1420 /* If user space value changed, let the caller retry */
1421 return curval != uval ? -EAGAIN : 0;
1425 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1426 * @uaddr: the pi futex user address
1427 * @hb: the pi futex hash bucket
1428 * @key: the futex key associated with uaddr and hb
1429 * @ps: the pi_state pointer where we store the result of the
1431 * @task: the task to perform the atomic lock work for. This will
1432 * be "current" except in the case of requeue pi.
1433 * @exiting: Pointer to store the task pointer of the owner task
1434 * which is in the middle of exiting
1435 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1438 * 0 - ready to wait;
1439 * 1 - acquired the lock;
1442 * The hb->lock and futex_key refs shall be held by the caller.
1444 * @exiting is only set when the return value is -EBUSY. If so, this holds
1445 * a refcount on the exiting task on return and the caller needs to drop it
1446 * after waiting for the exit to complete.
1448 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1449 union futex_key *key,
1450 struct futex_pi_state **ps,
1451 struct task_struct *task,
1452 struct task_struct **exiting,
1455 u32 uval, newval, vpid = task_pid_vnr(task);
1456 struct futex_q *top_waiter;
1460 * Read the user space value first so we can validate a few
1461 * things before proceeding further.
1463 if (get_futex_value_locked(&uval, uaddr))
1466 if (unlikely(should_fail_futex(true)))
1472 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1475 if ((unlikely(should_fail_futex(true))))
1479 * Lookup existing state first. If it exists, try to attach to
1482 top_waiter = futex_top_waiter(hb, key);
1484 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1487 * No waiter and user TID is 0. We are here because the
1488 * waiters or the owner died bit is set or called from
1489 * requeue_cmp_pi or for whatever reason something took the
1492 if (!(uval & FUTEX_TID_MASK)) {
1494 * We take over the futex. No other waiters and the user space
1495 * TID is 0. We preserve the owner died bit.
1497 newval = uval & FUTEX_OWNER_DIED;
1500 /* The futex requeue_pi code can enforce the waiters bit */
1502 newval |= FUTEX_WAITERS;
1504 ret = lock_pi_update_atomic(uaddr, uval, newval);
1505 /* If the take over worked, return 1 */
1506 return ret < 0 ? ret : 1;
1510 * First waiter. Set the waiters bit before attaching ourself to
1511 * the owner. If owner tries to unlock, it will be forced into
1512 * the kernel and blocked on hb->lock.
1514 newval = uval | FUTEX_WAITERS;
1515 ret = lock_pi_update_atomic(uaddr, uval, newval);
1519 * If the update of the user space value succeeded, we try to
1520 * attach to the owner. If that fails, no harm done, we only
1521 * set the FUTEX_WAITERS bit in the user space variable.
1523 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1527 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1528 * @q: The futex_q to unqueue
1530 * The q->lock_ptr must not be NULL and must be held by the caller.
1532 static void __unqueue_futex(struct futex_q *q)
1534 struct futex_hash_bucket *hb;
1536 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1537 || WARN_ON(plist_node_empty(&q->list)))
1540 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1541 plist_del(&q->list, &hb->chain);
1546 * The hash bucket lock must be held when this is called.
1547 * Afterwards, the futex_q must not be accessed. Callers
1548 * must ensure to later call wake_up_q() for the actual
1551 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1553 struct task_struct *p = q->task;
1555 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1561 * The waiting task can free the futex_q as soon as
1562 * q->lock_ptr = NULL is written, without taking any locks. A
1563 * memory barrier is required here to prevent the following
1564 * store to lock_ptr from getting ahead of the plist_del.
1566 smp_store_release(&q->lock_ptr, NULL);
1569 * Queue the task for later wakeup for after we've released
1570 * the hb->lock. wake_q_add() grabs reference to p.
1572 wake_q_add(wake_q, p);
1577 * Caller must hold a reference on @pi_state.
1579 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1581 u32 uninitialized_var(curval), newval;
1582 struct task_struct *new_owner;
1583 bool deboost = false;
1587 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1588 if (WARN_ON_ONCE(!new_owner)) {
1590 * As per the comment in futex_unlock_pi() this should not happen.
1592 * When this happens, give up our locks and try again, giving
1593 * the futex_lock_pi() instance time to complete, either by
1594 * waiting on the rtmutex or removing itself from the futex
1602 * We pass it to the next owner. The WAITERS bit is always kept
1603 * enabled while there is PI state around. We cleanup the owner
1604 * died bit, because we are the owner.
1606 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1608 if (unlikely(should_fail_futex(true))) {
1613 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1614 if (!ret && (curval != uval)) {
1616 * If a unconditional UNLOCK_PI operation (user space did not
1617 * try the TID->0 transition) raced with a waiter setting the
1618 * FUTEX_WAITERS flag between get_user() and locking the hash
1619 * bucket lock, retry the operation.
1621 if ((FUTEX_TID_MASK & curval) == uval)
1629 * This is a point of no return; once we modified the uval
1630 * there is no going back and subsequent operations must
1633 pi_state_update_owner(pi_state, new_owner);
1634 deboost = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1638 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1642 rt_mutex_adjust_prio(current);
1649 * Express the locking dependencies for lockdep:
1652 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1655 spin_lock(&hb1->lock);
1657 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1658 } else { /* hb1 > hb2 */
1659 spin_lock(&hb2->lock);
1660 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1665 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1667 spin_unlock(&hb1->lock);
1669 spin_unlock(&hb2->lock);
1673 * Wake up waiters matching bitset queued on this futex (uaddr).
1676 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1678 struct futex_hash_bucket *hb;
1679 struct futex_q *this, *next;
1680 union futex_key key = FUTEX_KEY_INIT;
1687 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1688 if (unlikely(ret != 0))
1691 hb = hash_futex(&key);
1693 /* Make sure we really have tasks to wakeup */
1694 if (!hb_waiters_pending(hb))
1697 spin_lock(&hb->lock);
1699 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1700 if (match_futex (&this->key, &key)) {
1701 if (this->pi_state || this->rt_waiter) {
1706 /* Check if one of the bits is set in both bitsets */
1707 if (!(this->bitset & bitset))
1710 mark_wake_futex(&wake_q, this);
1711 if (++ret >= nr_wake)
1716 spin_unlock(&hb->lock);
1719 put_futex_key(&key);
1724 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1726 unsigned int op = (encoded_op & 0x70000000) >> 28;
1727 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1728 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1729 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1732 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1733 if (oparg < 0 || oparg > 31) {
1734 char comm[sizeof(current->comm)];
1736 * kill this print and return -EINVAL when userspace
1739 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1740 get_task_comm(comm, current), oparg);
1746 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1749 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1754 case FUTEX_OP_CMP_EQ:
1755 return oldval == cmparg;
1756 case FUTEX_OP_CMP_NE:
1757 return oldval != cmparg;
1758 case FUTEX_OP_CMP_LT:
1759 return oldval < cmparg;
1760 case FUTEX_OP_CMP_GE:
1761 return oldval >= cmparg;
1762 case FUTEX_OP_CMP_LE:
1763 return oldval <= cmparg;
1764 case FUTEX_OP_CMP_GT:
1765 return oldval > cmparg;
1772 * Wake up all waiters hashed on the physical page that is mapped
1773 * to this virtual address:
1776 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1777 int nr_wake, int nr_wake2, int op)
1779 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1780 struct futex_hash_bucket *hb1, *hb2;
1781 struct futex_q *this, *next;
1786 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1787 if (unlikely(ret != 0))
1789 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1790 if (unlikely(ret != 0))
1793 hb1 = hash_futex(&key1);
1794 hb2 = hash_futex(&key2);
1797 double_lock_hb(hb1, hb2);
1798 op_ret = futex_atomic_op_inuser(op, uaddr2);
1799 if (unlikely(op_ret < 0)) {
1800 double_unlock_hb(hb1, hb2);
1802 if (!IS_ENABLED(CONFIG_MMU) ||
1803 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1805 * we don't get EFAULT from MMU faults if we don't have
1806 * an MMU, but we might get them from range checking
1812 if (op_ret == -EFAULT) {
1813 ret = fault_in_user_writeable(uaddr2);
1818 if (!(flags & FLAGS_SHARED)) {
1823 put_futex_key(&key2);
1824 put_futex_key(&key1);
1829 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1830 if (match_futex (&this->key, &key1)) {
1831 if (this->pi_state || this->rt_waiter) {
1835 mark_wake_futex(&wake_q, this);
1836 if (++ret >= nr_wake)
1843 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1844 if (match_futex (&this->key, &key2)) {
1845 if (this->pi_state || this->rt_waiter) {
1849 mark_wake_futex(&wake_q, this);
1850 if (++op_ret >= nr_wake2)
1858 double_unlock_hb(hb1, hb2);
1861 put_futex_key(&key2);
1863 put_futex_key(&key1);
1869 * requeue_futex() - Requeue a futex_q from one hb to another
1870 * @q: the futex_q to requeue
1871 * @hb1: the source hash_bucket
1872 * @hb2: the target hash_bucket
1873 * @key2: the new key for the requeued futex_q
1876 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1877 struct futex_hash_bucket *hb2, union futex_key *key2)
1881 * If key1 and key2 hash to the same bucket, no need to
1884 if (likely(&hb1->chain != &hb2->chain)) {
1885 plist_del(&q->list, &hb1->chain);
1886 hb_waiters_dec(hb1);
1887 hb_waiters_inc(hb2);
1888 plist_add(&q->list, &hb2->chain);
1889 q->lock_ptr = &hb2->lock;
1891 get_futex_key_refs(key2);
1896 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1898 * @key: the key of the requeue target futex
1899 * @hb: the hash_bucket of the requeue target futex
1901 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1902 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1903 * to the requeue target futex so the waiter can detect the wakeup on the right
1904 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1905 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1906 * to protect access to the pi_state to fixup the owner later. Must be called
1907 * with both q->lock_ptr and hb->lock held.
1910 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1911 struct futex_hash_bucket *hb)
1913 get_futex_key_refs(key);
1918 WARN_ON(!q->rt_waiter);
1919 q->rt_waiter = NULL;
1921 q->lock_ptr = &hb->lock;
1923 wake_up_state(q->task, TASK_NORMAL);
1927 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1928 * @pifutex: the user address of the to futex
1929 * @hb1: the from futex hash bucket, must be locked by the caller
1930 * @hb2: the to futex hash bucket, must be locked by the caller
1931 * @key1: the from futex key
1932 * @key2: the to futex key
1933 * @ps: address to store the pi_state pointer
1934 * @exiting: Pointer to store the task pointer of the owner task
1935 * which is in the middle of exiting
1936 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1938 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1939 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1940 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1941 * hb1 and hb2 must be held by the caller.
1943 * @exiting is only set when the return value is -EBUSY. If so, this holds
1944 * a refcount on the exiting task on return and the caller needs to drop it
1945 * after waiting for the exit to complete.
1948 * 0 - failed to acquire the lock atomically;
1949 * >0 - acquired the lock, return value is vpid of the top_waiter
1953 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1954 struct futex_hash_bucket *hb2, union futex_key *key1,
1955 union futex_key *key2, struct futex_pi_state **ps,
1956 struct task_struct **exiting, int set_waiters)
1958 struct futex_q *top_waiter = NULL;
1962 if (get_futex_value_locked(&curval, pifutex))
1965 if (unlikely(should_fail_futex(true)))
1969 * Find the top_waiter and determine if there are additional waiters.
1970 * If the caller intends to requeue more than 1 waiter to pifutex,
1971 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1972 * as we have means to handle the possible fault. If not, don't set
1973 * the bit unecessarily as it will force the subsequent unlock to enter
1976 top_waiter = futex_top_waiter(hb1, key1);
1978 /* There are no waiters, nothing for us to do. */
1982 /* Ensure we requeue to the expected futex. */
1983 if (!match_futex(top_waiter->requeue_pi_key, key2))
1987 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1988 * the contended case or if set_waiters is 1. The pi_state is returned
1989 * in ps in contended cases.
1991 vpid = task_pid_vnr(top_waiter->task);
1992 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1993 exiting, set_waiters);
1995 requeue_pi_wake_futex(top_waiter, key2, hb2);
2002 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
2003 * @uaddr1: source futex user address
2004 * @flags: futex flags (FLAGS_SHARED, etc.)
2005 * @uaddr2: target futex user address
2006 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
2007 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
2008 * @cmpval: @uaddr1 expected value (or %NULL)
2009 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
2010 * pi futex (pi to pi requeue is not supported)
2012 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
2013 * uaddr2 atomically on behalf of the top waiter.
2016 * >=0 - on success, the number of tasks requeued or woken;
2019 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
2020 u32 __user *uaddr2, int nr_wake, int nr_requeue,
2021 u32 *cmpval, int requeue_pi)
2023 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
2024 int drop_count = 0, task_count = 0, ret;
2025 struct futex_pi_state *pi_state = NULL;
2026 struct futex_hash_bucket *hb1, *hb2;
2027 struct futex_q *this, *next;
2030 if (nr_wake < 0 || nr_requeue < 0)
2035 * Requeue PI only works on two distinct uaddrs. This
2036 * check is only valid for private futexes. See below.
2038 if (uaddr1 == uaddr2)
2042 * requeue_pi requires a pi_state, try to allocate it now
2043 * without any locks in case it fails.
2045 if (refill_pi_state_cache())
2048 * requeue_pi must wake as many tasks as it can, up to nr_wake
2049 * + nr_requeue, since it acquires the rt_mutex prior to
2050 * returning to userspace, so as to not leave the rt_mutex with
2051 * waiters and no owner. However, second and third wake-ups
2052 * cannot be predicted as they involve race conditions with the
2053 * first wake and a fault while looking up the pi_state. Both
2054 * pthread_cond_signal() and pthread_cond_broadcast() should
2062 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
2063 if (unlikely(ret != 0))
2065 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2066 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
2067 if (unlikely(ret != 0))
2071 * The check above which compares uaddrs is not sufficient for
2072 * shared futexes. We need to compare the keys:
2074 if (requeue_pi && match_futex(&key1, &key2)) {
2079 hb1 = hash_futex(&key1);
2080 hb2 = hash_futex(&key2);
2083 hb_waiters_inc(hb2);
2084 double_lock_hb(hb1, hb2);
2086 if (likely(cmpval != NULL)) {
2089 ret = get_futex_value_locked(&curval, uaddr1);
2091 if (unlikely(ret)) {
2092 double_unlock_hb(hb1, hb2);
2093 hb_waiters_dec(hb2);
2095 ret = get_user(curval, uaddr1);
2099 if (!(flags & FLAGS_SHARED))
2102 put_futex_key(&key2);
2103 put_futex_key(&key1);
2106 if (curval != *cmpval) {
2112 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2113 struct task_struct *exiting = NULL;
2116 * Attempt to acquire uaddr2 and wake the top waiter. If we
2117 * intend to requeue waiters, force setting the FUTEX_WAITERS
2118 * bit. We force this here where we are able to easily handle
2119 * faults rather in the requeue loop below.
2121 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2123 &exiting, nr_requeue);
2126 * At this point the top_waiter has either taken uaddr2 or is
2127 * waiting on it. If the former, then the pi_state will not
2128 * exist yet, look it up one more time to ensure we have a
2129 * reference to it. If the lock was taken, ret contains the
2130 * vpid of the top waiter task.
2131 * If the lock was not taken, we have pi_state and an initial
2132 * refcount on it. In case of an error we have nothing.
2139 * If we acquired the lock, then the user space value
2140 * of uaddr2 should be vpid. It cannot be changed by
2141 * the top waiter as it is blocked on hb2 lock if it
2142 * tries to do so. If something fiddled with it behind
2143 * our back the pi state lookup might unearth it. So
2144 * we rather use the known value than rereading and
2145 * handing potential crap to lookup_pi_state.
2147 * If that call succeeds then we have pi_state and an
2148 * initial refcount on it.
2150 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2151 &pi_state, &exiting);
2156 /* We hold a reference on the pi state. */
2159 /* If the above failed, then pi_state is NULL */
2161 double_unlock_hb(hb1, hb2);
2162 hb_waiters_dec(hb2);
2163 put_futex_key(&key2);
2164 put_futex_key(&key1);
2165 ret = fault_in_user_writeable(uaddr2);
2172 * Two reasons for this:
2173 * - EBUSY: Owner is exiting and we just wait for the
2175 * - EAGAIN: The user space value changed.
2177 double_unlock_hb(hb1, hb2);
2178 hb_waiters_dec(hb2);
2179 put_futex_key(&key2);
2180 put_futex_key(&key1);
2182 * Handle the case where the owner is in the middle of
2183 * exiting. Wait for the exit to complete otherwise
2184 * this task might loop forever, aka. live lock.
2186 wait_for_owner_exiting(ret, exiting);
2194 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2195 if (task_count - nr_wake >= nr_requeue)
2198 if (!match_futex(&this->key, &key1))
2202 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2203 * be paired with each other and no other futex ops.
2205 * We should never be requeueing a futex_q with a pi_state,
2206 * which is awaiting a futex_unlock_pi().
2208 if ((requeue_pi && !this->rt_waiter) ||
2209 (!requeue_pi && this->rt_waiter) ||
2216 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2217 * lock, we already woke the top_waiter. If not, it will be
2218 * woken by futex_unlock_pi().
2220 if (++task_count <= nr_wake && !requeue_pi) {
2221 mark_wake_futex(&wake_q, this);
2225 /* Ensure we requeue to the expected futex for requeue_pi. */
2226 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2232 * Requeue nr_requeue waiters and possibly one more in the case
2233 * of requeue_pi if we couldn't acquire the lock atomically.
2237 * Prepare the waiter to take the rt_mutex. Take a
2238 * refcount on the pi_state and store the pointer in
2239 * the futex_q object of the waiter.
2241 get_pi_state(pi_state);
2242 this->pi_state = pi_state;
2243 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2248 * We got the lock. We do neither drop the
2249 * refcount on pi_state nor clear
2250 * this->pi_state because the waiter needs the
2251 * pi_state for cleaning up the user space
2252 * value. It will drop the refcount after
2255 requeue_pi_wake_futex(this, &key2, hb2);
2260 * rt_mutex_start_proxy_lock() detected a
2261 * potential deadlock when we tried to queue
2262 * that waiter. Drop the pi_state reference
2263 * which we took above and remove the pointer
2264 * to the state from the waiters futex_q
2267 this->pi_state = NULL;
2268 put_pi_state(pi_state);
2270 * We stop queueing more waiters and let user
2271 * space deal with the mess.
2276 requeue_futex(this, hb1, hb2, &key2);
2281 * We took an extra initial reference to the pi_state either
2282 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2283 * need to drop it here again.
2285 put_pi_state(pi_state);
2288 double_unlock_hb(hb1, hb2);
2290 hb_waiters_dec(hb2);
2293 * drop_futex_key_refs() must be called outside the spinlocks. During
2294 * the requeue we moved futex_q's from the hash bucket at key1 to the
2295 * one at key2 and updated their key pointer. We no longer need to
2296 * hold the references to key1.
2298 while (--drop_count >= 0)
2299 drop_futex_key_refs(&key1);
2302 put_futex_key(&key2);
2304 put_futex_key(&key1);
2306 return ret ? ret : task_count;
2309 /* The key must be already stored in q->key. */
2310 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2311 __acquires(&hb->lock)
2313 struct futex_hash_bucket *hb;
2315 hb = hash_futex(&q->key);
2318 * Increment the counter before taking the lock so that
2319 * a potential waker won't miss a to-be-slept task that is
2320 * waiting for the spinlock. This is safe as all queue_lock()
2321 * users end up calling queue_me(). Similarly, for housekeeping,
2322 * decrement the counter at queue_unlock() when some error has
2323 * occurred and we don't end up adding the task to the list.
2327 q->lock_ptr = &hb->lock;
2329 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2334 queue_unlock(struct futex_hash_bucket *hb)
2335 __releases(&hb->lock)
2337 spin_unlock(&hb->lock);
2341 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2346 * The priority used to register this element is
2347 * - either the real thread-priority for the real-time threads
2348 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2349 * - or MAX_RT_PRIO for non-RT threads.
2350 * Thus, all RT-threads are woken first in priority order, and
2351 * the others are woken last, in FIFO order.
2353 prio = min(current->normal_prio, MAX_RT_PRIO);
2355 plist_node_init(&q->list, prio);
2356 plist_add(&q->list, &hb->chain);
2361 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2362 * @q: The futex_q to enqueue
2363 * @hb: The destination hash bucket
2365 * The hb->lock must be held by the caller, and is released here. A call to
2366 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2367 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2368 * or nothing if the unqueue is done as part of the wake process and the unqueue
2369 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2372 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2373 __releases(&hb->lock)
2376 spin_unlock(&hb->lock);
2380 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2381 * @q: The futex_q to unqueue
2383 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2384 * be paired with exactly one earlier call to queue_me().
2387 * 1 - if the futex_q was still queued (and we removed unqueued it);
2388 * 0 - if the futex_q was already removed by the waking thread
2390 static int unqueue_me(struct futex_q *q)
2392 spinlock_t *lock_ptr;
2395 /* In the common case we don't take the spinlock, which is nice. */
2398 * q->lock_ptr can change between this read and the following spin_lock.
2399 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2400 * optimizing lock_ptr out of the logic below.
2402 lock_ptr = READ_ONCE(q->lock_ptr);
2403 if (lock_ptr != NULL) {
2404 spin_lock(lock_ptr);
2406 * q->lock_ptr can change between reading it and
2407 * spin_lock(), causing us to take the wrong lock. This
2408 * corrects the race condition.
2410 * Reasoning goes like this: if we have the wrong lock,
2411 * q->lock_ptr must have changed (maybe several times)
2412 * between reading it and the spin_lock(). It can
2413 * change again after the spin_lock() but only if it was
2414 * already changed before the spin_lock(). It cannot,
2415 * however, change back to the original value. Therefore
2416 * we can detect whether we acquired the correct lock.
2418 if (unlikely(lock_ptr != q->lock_ptr)) {
2419 spin_unlock(lock_ptr);
2424 BUG_ON(q->pi_state);
2426 spin_unlock(lock_ptr);
2430 drop_futex_key_refs(&q->key);
2435 * PI futexes can not be requeued and must remove themself from the
2436 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2439 static void unqueue_me_pi(struct futex_q *q)
2440 __releases(q->lock_ptr)
2444 BUG_ON(!q->pi_state);
2445 put_pi_state(q->pi_state);
2448 spin_unlock(q->lock_ptr);
2451 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2452 struct task_struct *argowner)
2454 struct futex_pi_state *pi_state = q->pi_state;
2455 struct task_struct *oldowner, *newowner;
2456 u32 uval, curval, newval, newtid;
2459 oldowner = pi_state->owner;
2462 * We are here because either:
2464 * - we stole the lock and pi_state->owner needs updating to reflect
2465 * that (@argowner == current),
2469 * - someone stole our lock and we need to fix things to point to the
2470 * new owner (@argowner == NULL).
2472 * Either way, we have to replace the TID in the user space variable.
2473 * This must be atomic as we have to preserve the owner died bit here.
2475 * Note: We write the user space value _before_ changing the pi_state
2476 * because we can fault here. Imagine swapped out pages or a fork
2477 * that marked all the anonymous memory readonly for cow.
2479 * Modifying pi_state _before_ the user space value would leave the
2480 * pi_state in an inconsistent state when we fault here, because we
2481 * need to drop the locks to handle the fault. This might be observed
2482 * in the PID check in lookup_pi_state.
2486 if (oldowner != current) {
2488 * We raced against a concurrent self; things are
2489 * already fixed up. Nothing to do.
2494 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2495 /* We got the lock. pi_state is correct. Tell caller. */
2500 * The trylock just failed, so either there is an owner or
2501 * there is a higher priority waiter than this one.
2503 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2505 * If the higher priority waiter has not yet taken over the
2506 * rtmutex then newowner is NULL. We can't return here with
2507 * that state because it's inconsistent vs. the user space
2508 * state. So drop the locks and try again. It's a valid
2509 * situation and not any different from the other retry
2512 if (unlikely(!newowner)) {
2517 WARN_ON_ONCE(argowner != current);
2518 if (oldowner == current) {
2520 * We raced against a concurrent self; things are
2521 * already fixed up. Nothing to do.
2525 newowner = argowner;
2528 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2530 if (!pi_state->owner)
2531 newtid |= FUTEX_OWNER_DIED;
2533 err = get_futex_value_locked(&uval, uaddr);
2538 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2540 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2550 * We fixed up user space. Now we need to fix the pi_state
2553 pi_state_update_owner(pi_state, newowner);
2555 return argowner == current;
2558 * In order to reschedule or handle a page fault, we need to drop the
2559 * locks here. In the case of a fault, this gives the other task
2560 * (either the highest priority waiter itself or the task which stole
2561 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2562 * are back from handling the fault we need to check the pi_state after
2563 * reacquiring the locks and before trying to do another fixup. When
2564 * the fixup has been done already we simply return.
2566 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2567 * drop hb->lock since the caller owns the hb -> futex_q relation.
2568 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2571 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2572 spin_unlock(q->lock_ptr);
2576 err = fault_in_user_writeable(uaddr);
2589 spin_lock(q->lock_ptr);
2590 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2593 * Check if someone else fixed it for us:
2595 if (pi_state->owner != oldowner)
2596 return argowner == current;
2598 /* Retry if err was -EAGAIN or the fault in succeeded */
2603 * fault_in_user_writeable() failed so user state is immutable. At
2604 * best we can make the kernel state consistent but user state will
2605 * be most likely hosed and any subsequent unlock operation will be
2606 * rejected due to PI futex rule [10].
2608 * Ensure that the rtmutex owner is also the pi_state owner despite
2609 * the user space value claiming something different. There is no
2610 * point in unlocking the rtmutex if current is the owner as it
2611 * would need to wait until the next waiter has taken the rtmutex
2612 * to guarantee consistent state. Keep it simple. Userspace asked
2613 * for this wreckaged state.
2615 * The rtmutex has an owner - either current or some other
2616 * task. See the EAGAIN loop above.
2618 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2623 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2624 struct task_struct *argowner)
2626 struct futex_pi_state *pi_state = q->pi_state;
2629 lockdep_assert_held(q->lock_ptr);
2631 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2632 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2633 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2637 static long futex_wait_restart(struct restart_block *restart);
2640 * fixup_owner() - Post lock pi_state and corner case management
2641 * @uaddr: user address of the futex
2642 * @q: futex_q (contains pi_state and access to the rt_mutex)
2643 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2645 * After attempting to lock an rt_mutex, this function is called to cleanup
2646 * the pi_state owner as well as handle race conditions that may allow us to
2647 * acquire the lock. Must be called with the hb lock held.
2650 * 1 - success, lock taken;
2651 * 0 - success, lock not taken;
2652 * <0 - on error (-EFAULT)
2654 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2658 * Got the lock. We might not be the anticipated owner if we
2659 * did a lock-steal - fix up the PI-state in that case:
2661 * Speculative pi_state->owner read (we don't hold wait_lock);
2662 * since we own the lock pi_state->owner == current is the
2663 * stable state, anything else needs more attention.
2665 if (q->pi_state->owner != current)
2666 return fixup_pi_state_owner(uaddr, q, current);
2671 * If we didn't get the lock; check if anybody stole it from us. In
2672 * that case, we need to fix up the uval to point to them instead of
2673 * us, otherwise bad things happen. [10]
2675 * Another speculative read; pi_state->owner == current is unstable
2676 * but needs our attention.
2678 if (q->pi_state->owner == current)
2679 return fixup_pi_state_owner(uaddr, q, NULL);
2682 * Paranoia check. If we did not take the lock, then we should not be
2683 * the owner of the rt_mutex. Warn and establish consistent state.
2685 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2686 return fixup_pi_state_owner(uaddr, q, current);
2692 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2693 * @hb: the futex hash bucket, must be locked by the caller
2694 * @q: the futex_q to queue up on
2695 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2697 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2698 struct hrtimer_sleeper *timeout)
2701 * The task state is guaranteed to be set before another task can
2702 * wake it. set_current_state() is implemented using smp_store_mb() and
2703 * queue_me() calls spin_unlock() upon completion, both serializing
2704 * access to the hash list and forcing another memory barrier.
2706 set_current_state(TASK_INTERRUPTIBLE);
2711 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2714 * If we have been removed from the hash list, then another task
2715 * has tried to wake us, and we can skip the call to schedule().
2717 if (likely(!plist_node_empty(&q->list))) {
2719 * If the timer has already expired, current will already be
2720 * flagged for rescheduling. Only call schedule if there
2721 * is no timeout, or if it has yet to expire.
2723 if (!timeout || timeout->task)
2724 freezable_schedule();
2726 __set_current_state(TASK_RUNNING);
2730 * futex_wait_setup() - Prepare to wait on a futex
2731 * @uaddr: the futex userspace address
2732 * @val: the expected value
2733 * @flags: futex flags (FLAGS_SHARED, etc.)
2734 * @q: the associated futex_q
2735 * @hb: storage for hash_bucket pointer to be returned to caller
2737 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2738 * compare it with the expected value. Handle atomic faults internally.
2739 * Return with the hb lock held and a q.key reference on success, and unlocked
2740 * with no q.key reference on failure.
2743 * 0 - uaddr contains val and hb has been locked;
2744 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2746 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2747 struct futex_q *q, struct futex_hash_bucket **hb)
2753 * Access the page AFTER the hash-bucket is locked.
2754 * Order is important:
2756 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2757 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2759 * The basic logical guarantee of a futex is that it blocks ONLY
2760 * if cond(var) is known to be true at the time of blocking, for
2761 * any cond. If we locked the hash-bucket after testing *uaddr, that
2762 * would open a race condition where we could block indefinitely with
2763 * cond(var) false, which would violate the guarantee.
2765 * On the other hand, we insert q and release the hash-bucket only
2766 * after testing *uaddr. This guarantees that futex_wait() will NOT
2767 * absorb a wakeup if *uaddr does not match the desired values
2768 * while the syscall executes.
2771 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2772 if (unlikely(ret != 0))
2776 *hb = queue_lock(q);
2778 ret = get_futex_value_locked(&uval, uaddr);
2783 ret = get_user(uval, uaddr);
2787 if (!(flags & FLAGS_SHARED))
2790 put_futex_key(&q->key);
2801 put_futex_key(&q->key);
2805 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2806 ktime_t *abs_time, u32 bitset)
2808 struct hrtimer_sleeper timeout, *to = NULL;
2809 struct restart_block *restart;
2810 struct futex_hash_bucket *hb;
2811 struct futex_q q = futex_q_init;
2821 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2822 CLOCK_REALTIME : CLOCK_MONOTONIC,
2824 hrtimer_init_sleeper(to, current);
2825 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2826 current->timer_slack_ns);
2831 * Prepare to wait on uaddr. On success, holds hb lock and increments
2834 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2838 /* queue_me and wait for wakeup, timeout, or a signal. */
2839 futex_wait_queue_me(hb, &q, to);
2841 /* If we were woken (and unqueued), we succeeded, whatever. */
2843 /* unqueue_me() drops q.key ref */
2844 if (!unqueue_me(&q))
2847 if (to && !to->task)
2851 * We expect signal_pending(current), but we might be the
2852 * victim of a spurious wakeup as well.
2854 if (!signal_pending(current))
2861 restart = ¤t->restart_block;
2862 restart->futex.uaddr = uaddr;
2863 restart->futex.val = val;
2864 restart->futex.time = abs_time->tv64;
2865 restart->futex.bitset = bitset;
2866 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2868 ret = set_restart_fn(restart, futex_wait_restart);
2872 hrtimer_cancel(&to->timer);
2873 destroy_hrtimer_on_stack(&to->timer);
2879 static long futex_wait_restart(struct restart_block *restart)
2881 u32 __user *uaddr = restart->futex.uaddr;
2882 ktime_t t, *tp = NULL;
2884 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2885 t.tv64 = restart->futex.time;
2888 restart->fn = do_no_restart_syscall;
2890 return (long)futex_wait(uaddr, restart->futex.flags,
2891 restart->futex.val, tp, restart->futex.bitset);
2896 * Userspace tried a 0 -> TID atomic transition of the futex value
2897 * and failed. The kernel side here does the whole locking operation:
2898 * if there are waiters then it will block as a consequence of relying
2899 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2900 * a 0 value of the futex too.).
2902 * Also serves as futex trylock_pi()'ing, and due semantics.
2904 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2905 ktime_t *time, int trylock)
2907 struct hrtimer_sleeper timeout, *to = NULL;
2908 struct task_struct *exiting = NULL;
2909 struct rt_mutex_waiter rt_waiter;
2910 struct futex_hash_bucket *hb;
2911 struct futex_q q = futex_q_init;
2914 if (refill_pi_state_cache())
2919 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2921 hrtimer_init_sleeper(to, current);
2922 hrtimer_set_expires(&to->timer, *time);
2926 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2927 if (unlikely(ret != 0))
2931 hb = queue_lock(&q);
2933 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2935 if (unlikely(ret)) {
2937 * Atomic work succeeded and we got the lock,
2938 * or failed. Either way, we do _not_ block.
2942 /* We got the lock. */
2944 goto out_unlock_put_key;
2950 * Two reasons for this:
2951 * - EBUSY: Task is exiting and we just wait for the
2953 * - EAGAIN: The user space value changed.
2956 put_futex_key(&q.key);
2958 * Handle the case where the owner is in the middle of
2959 * exiting. Wait for the exit to complete otherwise
2960 * this task might loop forever, aka. live lock.
2962 wait_for_owner_exiting(ret, exiting);
2966 goto out_unlock_put_key;
2970 WARN_ON(!q.pi_state);
2973 * Only actually queue now that the atomic ops are done:
2978 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2979 /* Fixup the trylock return value: */
2980 ret = ret ? 0 : -EWOULDBLOCK;
2984 rt_mutex_init_waiter(&rt_waiter);
2987 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2988 * hold it while doing rt_mutex_start_proxy(), because then it will
2989 * include hb->lock in the blocking chain, even through we'll not in
2990 * fact hold it while blocking. This will lead it to report -EDEADLK
2991 * and BUG when futex_unlock_pi() interleaves with this.
2993 * Therefore acquire wait_lock while holding hb->lock, but drop the
2994 * latter before calling __rt_mutex_start_proxy_lock(). This
2995 * interleaves with futex_unlock_pi() -- which does a similar lock
2996 * handoff -- such that the latter can observe the futex_q::pi_state
2997 * before __rt_mutex_start_proxy_lock() is done.
2999 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
3000 spin_unlock(q.lock_ptr);
3002 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
3003 * such that futex_unlock_pi() is guaranteed to observe the waiter when
3004 * it sees the futex_q::pi_state.
3006 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
3007 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3016 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
3018 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3021 spin_lock(q.lock_ptr);
3023 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3024 * first acquire the hb->lock before removing the lock from the
3025 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3028 * In particular; it is important that futex_unlock_pi() can not
3029 * observe this inconsistency.
3031 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3036 * Fixup the pi_state owner and possibly acquire the lock if we
3039 res = fixup_owner(uaddr, &q, !ret);
3041 * If fixup_owner() returned an error, proprogate that. If it acquired
3042 * the lock, clear our -ETIMEDOUT or -EINTR.
3045 ret = (res < 0) ? res : 0;
3047 /* Unqueue and drop the lock */
3056 put_futex_key(&q.key);
3059 hrtimer_cancel(&to->timer);
3060 destroy_hrtimer_on_stack(&to->timer);
3062 return ret != -EINTR ? ret : -ERESTARTNOINTR;
3067 ret = fault_in_user_writeable(uaddr);
3071 if (!(flags & FLAGS_SHARED))
3074 put_futex_key(&q.key);
3079 * Userspace attempted a TID -> 0 atomic transition, and failed.
3080 * This is the in-kernel slowpath: we look up the PI state (if any),
3081 * and do the rt-mutex unlock.
3083 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3085 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
3086 union futex_key key = FUTEX_KEY_INIT;
3087 struct futex_hash_bucket *hb;
3088 struct futex_q *top_waiter;
3092 if (get_user(uval, uaddr))
3095 * We release only a lock we actually own:
3097 if ((uval & FUTEX_TID_MASK) != vpid)
3100 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
3104 hb = hash_futex(&key);
3105 spin_lock(&hb->lock);
3108 * Check waiters first. We do not trust user space values at
3109 * all and we at least want to know if user space fiddled
3110 * with the futex value instead of blindly unlocking.
3112 top_waiter = futex_top_waiter(hb, &key);
3114 struct futex_pi_state *pi_state = top_waiter->pi_state;
3121 * If current does not own the pi_state then the futex is
3122 * inconsistent and user space fiddled with the futex value.
3124 if (pi_state->owner != current)
3127 get_pi_state(pi_state);
3129 * By taking wait_lock while still holding hb->lock, we ensure
3130 * there is no point where we hold neither; and therefore
3131 * wake_futex_pi() must observe a state consistent with what we
3134 * In particular; this forces __rt_mutex_start_proxy() to
3135 * complete such that we're guaranteed to observe the
3136 * rt_waiter. Also see the WARN in wake_futex_pi().
3138 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3139 spin_unlock(&hb->lock);
3141 /* drops pi_state->pi_mutex.wait_lock */
3142 ret = wake_futex_pi(uaddr, uval, pi_state);
3144 put_pi_state(pi_state);
3147 * Success, we're done! No tricky corner cases.
3152 * The atomic access to the futex value generated a
3153 * pagefault, so retry the user-access and the wakeup:
3158 * A unconditional UNLOCK_PI op raced against a waiter
3159 * setting the FUTEX_WAITERS bit. Try again.
3164 * wake_futex_pi has detected invalid state. Tell user
3171 * We have no kernel internal state, i.e. no waiters in the
3172 * kernel. Waiters which are about to queue themselves are stuck
3173 * on hb->lock. So we can safely ignore them. We do neither
3174 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3177 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3178 spin_unlock(&hb->lock);
3193 * If uval has changed, let user space handle it.
3195 ret = (curval == uval) ? 0 : -EAGAIN;
3198 spin_unlock(&hb->lock);
3200 put_futex_key(&key);
3204 put_futex_key(&key);
3209 put_futex_key(&key);
3211 ret = fault_in_user_writeable(uaddr);
3219 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3220 * @hb: the hash_bucket futex_q was original enqueued on
3221 * @q: the futex_q woken while waiting to be requeued
3222 * @key2: the futex_key of the requeue target futex
3223 * @timeout: the timeout associated with the wait (NULL if none)
3225 * Detect if the task was woken on the initial futex as opposed to the requeue
3226 * target futex. If so, determine if it was a timeout or a signal that caused
3227 * the wakeup and return the appropriate error code to the caller. Must be
3228 * called with the hb lock held.
3231 * 0 = no early wakeup detected;
3232 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
3235 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3236 struct futex_q *q, union futex_key *key2,
3237 struct hrtimer_sleeper *timeout)
3242 * With the hb lock held, we avoid races while we process the wakeup.
3243 * We only need to hold hb (and not hb2) to ensure atomicity as the
3244 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3245 * It can't be requeued from uaddr2 to something else since we don't
3246 * support a PI aware source futex for requeue.
3248 if (!match_futex(&q->key, key2)) {
3249 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3251 * We were woken prior to requeue by a timeout or a signal.
3252 * Unqueue the futex_q and determine which it was.
3254 plist_del(&q->list, &hb->chain);
3257 /* Handle spurious wakeups gracefully */
3259 if (timeout && !timeout->task)
3261 else if (signal_pending(current))
3262 ret = -ERESTARTNOINTR;
3268 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3269 * @uaddr: the futex we initially wait on (non-pi)
3270 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3271 * the same type, no requeueing from private to shared, etc.
3272 * @val: the expected value of uaddr
3273 * @abs_time: absolute timeout
3274 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3275 * @uaddr2: the pi futex we will take prior to returning to user-space
3277 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3278 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3279 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3280 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3281 * without one, the pi logic would not know which task to boost/deboost, if
3282 * there was a need to.
3284 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3285 * via the following--
3286 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3287 * 2) wakeup on uaddr2 after a requeue
3291 * If 3, cleanup and return -ERESTARTNOINTR.
3293 * If 2, we may then block on trying to take the rt_mutex and return via:
3294 * 5) successful lock
3297 * 8) other lock acquisition failure
3299 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3301 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3307 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3308 u32 val, ktime_t *abs_time, u32 bitset,
3311 struct hrtimer_sleeper timeout, *to = NULL;
3312 struct rt_mutex_waiter rt_waiter;
3313 struct futex_hash_bucket *hb;
3314 union futex_key key2 = FUTEX_KEY_INIT;
3315 struct futex_q q = futex_q_init;
3318 if (uaddr == uaddr2)
3326 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3327 CLOCK_REALTIME : CLOCK_MONOTONIC,
3329 hrtimer_init_sleeper(to, current);
3330 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3331 current->timer_slack_ns);
3335 * The waiter is allocated on our stack, manipulated by the requeue
3336 * code while we sleep on uaddr.
3338 rt_mutex_init_waiter(&rt_waiter);
3340 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3341 if (unlikely(ret != 0))
3345 q.rt_waiter = &rt_waiter;
3346 q.requeue_pi_key = &key2;
3349 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3352 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3357 * The check above which compares uaddrs is not sufficient for
3358 * shared futexes. We need to compare the keys:
3360 if (match_futex(&q.key, &key2)) {
3366 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3367 futex_wait_queue_me(hb, &q, to);
3369 spin_lock(&hb->lock);
3370 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3371 spin_unlock(&hb->lock);
3376 * In order for us to be here, we know our q.key == key2, and since
3377 * we took the hb->lock above, we also know that futex_requeue() has
3378 * completed and we no longer have to concern ourselves with a wakeup
3379 * race with the atomic proxy lock acquisition by the requeue code. The
3380 * futex_requeue dropped our key1 reference and incremented our key2
3384 /* Check if the requeue code acquired the second futex for us. */
3387 * Got the lock. We might not be the anticipated owner if we
3388 * did a lock-steal - fix up the PI-state in that case.
3390 if (q.pi_state && (q.pi_state->owner != current)) {
3391 spin_lock(q.lock_ptr);
3392 ret = fixup_pi_state_owner(uaddr2, &q, current);
3394 * Drop the reference to the pi state which
3395 * the requeue_pi() code acquired for us.
3397 put_pi_state(q.pi_state);
3398 spin_unlock(q.lock_ptr);
3400 * Adjust the return value. It's either -EFAULT or
3401 * success (1) but the caller expects 0 for success.
3403 ret = ret < 0 ? ret : 0;
3406 struct rt_mutex *pi_mutex;
3409 * We have been woken up by futex_unlock_pi(), a timeout, or a
3410 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3413 WARN_ON(!q.pi_state);
3414 pi_mutex = &q.pi_state->pi_mutex;
3415 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3417 spin_lock(q.lock_ptr);
3418 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3421 debug_rt_mutex_free_waiter(&rt_waiter);
3423 * Fixup the pi_state owner and possibly acquire the lock if we
3426 res = fixup_owner(uaddr2, &q, !ret);
3428 * If fixup_owner() returned an error, proprogate that. If it
3429 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3432 ret = (res < 0) ? res : 0;
3434 /* Unqueue and drop the lock. */
3438 if (ret == -EINTR) {
3440 * We've already been requeued, but cannot restart by calling
3441 * futex_lock_pi() directly. We could restart this syscall, but
3442 * it would detect that the user space "val" changed and return
3443 * -EWOULDBLOCK. Save the overhead of the restart and return
3444 * -EWOULDBLOCK directly.
3450 put_futex_key(&q.key);
3452 put_futex_key(&key2);
3456 hrtimer_cancel(&to->timer);
3457 destroy_hrtimer_on_stack(&to->timer);
3463 * Support for robust futexes: the kernel cleans up held futexes at
3466 * Implementation: user-space maintains a per-thread list of locks it
3467 * is holding. Upon do_exit(), the kernel carefully walks this list,
3468 * and marks all locks that are owned by this thread with the
3469 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3470 * always manipulated with the lock held, so the list is private and
3471 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3472 * field, to allow the kernel to clean up if the thread dies after
3473 * acquiring the lock, but just before it could have added itself to
3474 * the list. There can only be one such pending lock.
3478 * sys_set_robust_list() - Set the robust-futex list head of a task
3479 * @head: pointer to the list-head
3480 * @len: length of the list-head, as userspace expects
3482 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3485 if (!futex_cmpxchg_enabled)
3488 * The kernel knows only one size for now:
3490 if (unlikely(len != sizeof(*head)))
3493 current->robust_list = head;
3499 * sys_get_robust_list() - Get the robust-futex list head of a task
3500 * @pid: pid of the process [zero for current task]
3501 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3502 * @len_ptr: pointer to a length field, the kernel fills in the header size
3504 SYSCALL_DEFINE3(get_robust_list, int, pid,
3505 struct robust_list_head __user * __user *, head_ptr,
3506 size_t __user *, len_ptr)
3508 struct robust_list_head __user *head;
3510 struct task_struct *p;
3512 if (!futex_cmpxchg_enabled)
3521 p = find_task_by_vpid(pid);
3527 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3530 head = p->robust_list;
3533 if (put_user(sizeof(*head), len_ptr))
3535 return put_user(head, head_ptr);
3543 /* Constants for the pending_op argument of handle_futex_death */
3544 #define HANDLE_DEATH_PENDING true
3545 #define HANDLE_DEATH_LIST false
3548 * Process a futex-list entry, check whether it's owned by the
3549 * dying task, and do notification if so:
3551 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3552 bool pi, bool pending_op)
3554 u32 uval, uninitialized_var(nval), mval;
3557 /* Futex address must be 32bit aligned */
3558 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3562 if (get_user(uval, uaddr))
3566 * Special case for regular (non PI) futexes. The unlock path in
3567 * user space has two race scenarios:
3569 * 1. The unlock path releases the user space futex value and
3570 * before it can execute the futex() syscall to wake up
3571 * waiters it is killed.
3573 * 2. A woken up waiter is killed before it can acquire the
3574 * futex in user space.
3576 * In both cases the TID validation below prevents a wakeup of
3577 * potential waiters which can cause these waiters to block
3580 * In both cases the following conditions are met:
3582 * 1) task->robust_list->list_op_pending != NULL
3583 * @pending_op == true
3584 * 2) User space futex value == 0
3585 * 3) Regular futex: @pi == false
3587 * If these conditions are met, it is safe to attempt waking up a
3588 * potential waiter without touching the user space futex value and
3589 * trying to set the OWNER_DIED bit. The user space futex value is
3590 * uncontended and the rest of the user space mutex state is
3591 * consistent, so a woken waiter will just take over the
3592 * uncontended futex. Setting the OWNER_DIED bit would create
3593 * inconsistent state and malfunction of the user space owner died
3596 if (pending_op && !pi && !uval) {
3597 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3601 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3605 * Ok, this dying thread is truly holding a futex
3606 * of interest. Set the OWNER_DIED bit atomically
3607 * via cmpxchg, and if the value had FUTEX_WAITERS
3608 * set, wake up a waiter (if any). (We have to do a
3609 * futex_wake() even if OWNER_DIED is already set -
3610 * to handle the rare but possible case of recursive
3611 * thread-death.) The rest of the cleanup is done in
3614 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3617 * We are not holding a lock here, but we want to have
3618 * the pagefault_disable/enable() protection because
3619 * we want to handle the fault gracefully. If the
3620 * access fails we try to fault in the futex with R/W
3621 * verification via get_user_pages. get_user() above
3622 * does not guarantee R/W access. If that fails we
3623 * give up and leave the futex locked.
3625 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3628 if (fault_in_user_writeable(uaddr))
3646 * Wake robust non-PI futexes here. The wakeup of
3647 * PI futexes happens in exit_pi_state():
3649 if (!pi && (uval & FUTEX_WAITERS))
3650 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3656 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3658 static inline int fetch_robust_entry(struct robust_list __user **entry,
3659 struct robust_list __user * __user *head,
3662 unsigned long uentry;
3664 if (get_user(uentry, (unsigned long __user *)head))
3667 *entry = (void __user *)(uentry & ~1UL);
3674 * Walk curr->robust_list (very carefully, it's a userspace list!)
3675 * and mark any locks found there dead, and notify any waiters.
3677 * We silently return on any sign of list-walking problem.
3679 static void exit_robust_list(struct task_struct *curr)
3681 struct robust_list_head __user *head = curr->robust_list;
3682 struct robust_list __user *entry, *next_entry, *pending;
3683 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3684 unsigned int uninitialized_var(next_pi);
3685 unsigned long futex_offset;
3688 if (!futex_cmpxchg_enabled)
3692 * Fetch the list head (which was registered earlier, via
3693 * sys_set_robust_list()):
3695 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3698 * Fetch the relative futex offset:
3700 if (get_user(futex_offset, &head->futex_offset))
3703 * Fetch any possibly pending lock-add first, and handle it
3706 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3709 next_entry = NULL; /* avoid warning with gcc */
3710 while (entry != &head->list) {
3712 * Fetch the next entry in the list before calling
3713 * handle_futex_death:
3715 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3717 * A pending lock might already be on the list, so
3718 * don't process it twice:
3720 if (entry != pending) {
3721 if (handle_futex_death((void __user *)entry + futex_offset,
3722 curr, pi, HANDLE_DEATH_LIST))
3730 * Avoid excessively long or circular lists:
3739 handle_futex_death((void __user *)pending + futex_offset,
3740 curr, pip, HANDLE_DEATH_PENDING);
3744 static void futex_cleanup(struct task_struct *tsk)
3746 if (unlikely(tsk->robust_list)) {
3747 exit_robust_list(tsk);
3748 tsk->robust_list = NULL;
3751 #ifdef CONFIG_COMPAT
3752 if (unlikely(tsk->compat_robust_list)) {
3753 compat_exit_robust_list(tsk);
3754 tsk->compat_robust_list = NULL;
3758 if (unlikely(!list_empty(&tsk->pi_state_list)))
3759 exit_pi_state_list(tsk);
3763 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3764 * @tsk: task to set the state on
3766 * Set the futex exit state of the task lockless. The futex waiter code
3767 * observes that state when a task is exiting and loops until the task has
3768 * actually finished the futex cleanup. The worst case for this is that the
3769 * waiter runs through the wait loop until the state becomes visible.
3771 * This is called from the recursive fault handling path in do_exit().
3773 * This is best effort. Either the futex exit code has run already or
3774 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3775 * take it over. If not, the problem is pushed back to user space. If the
3776 * futex exit code did not run yet, then an already queued waiter might
3777 * block forever, but there is nothing which can be done about that.
3779 void futex_exit_recursive(struct task_struct *tsk)
3781 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3782 if (tsk->futex_state == FUTEX_STATE_EXITING)
3783 mutex_unlock(&tsk->futex_exit_mutex);
3784 tsk->futex_state = FUTEX_STATE_DEAD;
3787 static void futex_cleanup_begin(struct task_struct *tsk)
3790 * Prevent various race issues against a concurrent incoming waiter
3791 * including live locks by forcing the waiter to block on
3792 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3793 * attach_to_pi_owner().
3795 mutex_lock(&tsk->futex_exit_mutex);
3798 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3800 * This ensures that all subsequent checks of tsk->futex_state in
3801 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3802 * tsk->pi_lock held.
3804 * It guarantees also that a pi_state which was queued right before
3805 * the state change under tsk->pi_lock by a concurrent waiter must
3806 * be observed in exit_pi_state_list().
3808 raw_spin_lock_irq(&tsk->pi_lock);
3809 tsk->futex_state = FUTEX_STATE_EXITING;
3810 raw_spin_unlock_irq(&tsk->pi_lock);
3813 static void futex_cleanup_end(struct task_struct *tsk, int state)
3816 * Lockless store. The only side effect is that an observer might
3817 * take another loop until it becomes visible.
3819 tsk->futex_state = state;
3821 * Drop the exit protection. This unblocks waiters which observed
3822 * FUTEX_STATE_EXITING to reevaluate the state.
3824 mutex_unlock(&tsk->futex_exit_mutex);
3827 void futex_exec_release(struct task_struct *tsk)
3830 * The state handling is done for consistency, but in the case of
3831 * exec() there is no way to prevent futher damage as the PID stays
3832 * the same. But for the unlikely and arguably buggy case that a
3833 * futex is held on exec(), this provides at least as much state
3834 * consistency protection which is possible.
3836 futex_cleanup_begin(tsk);
3839 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3840 * exec a new binary.
3842 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3845 void futex_exit_release(struct task_struct *tsk)
3847 futex_cleanup_begin(tsk);
3849 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3852 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3853 u32 __user *uaddr2, u32 val2, u32 val3)
3855 int cmd = op & FUTEX_CMD_MASK;
3856 unsigned int flags = 0;
3858 if (!(op & FUTEX_PRIVATE_FLAG))
3859 flags |= FLAGS_SHARED;
3861 if (op & FUTEX_CLOCK_REALTIME) {
3862 flags |= FLAGS_CLOCKRT;
3863 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3869 case FUTEX_UNLOCK_PI:
3870 case FUTEX_TRYLOCK_PI:
3871 case FUTEX_WAIT_REQUEUE_PI:
3872 case FUTEX_CMP_REQUEUE_PI:
3873 if (!futex_cmpxchg_enabled)
3879 val3 = FUTEX_BITSET_MATCH_ANY;
3880 case FUTEX_WAIT_BITSET:
3881 return futex_wait(uaddr, flags, val, timeout, val3);
3883 val3 = FUTEX_BITSET_MATCH_ANY;
3884 case FUTEX_WAKE_BITSET:
3885 return futex_wake(uaddr, flags, val, val3);
3887 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3888 case FUTEX_CMP_REQUEUE:
3889 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3891 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3893 return futex_lock_pi(uaddr, flags, timeout, 0);
3894 case FUTEX_UNLOCK_PI:
3895 return futex_unlock_pi(uaddr, flags);
3896 case FUTEX_TRYLOCK_PI:
3897 return futex_lock_pi(uaddr, flags, NULL, 1);
3898 case FUTEX_WAIT_REQUEUE_PI:
3899 val3 = FUTEX_BITSET_MATCH_ANY;
3900 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3902 case FUTEX_CMP_REQUEUE_PI:
3903 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3909 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3910 struct timespec __user *, utime, u32 __user *, uaddr2,
3914 ktime_t t, *tp = NULL;
3916 int cmd = op & FUTEX_CMD_MASK;
3918 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3919 cmd == FUTEX_WAIT_BITSET ||
3920 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3921 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3923 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3925 if (!timespec_valid(&ts))
3928 t = timespec_to_ktime(ts);
3929 if (cmd == FUTEX_WAIT)
3930 t = ktime_add_safe(ktime_get(), t);
3934 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3935 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3937 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3938 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3939 val2 = (u32) (unsigned long) utime;
3941 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3944 #ifdef CONFIG_COMPAT
3946 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3949 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3950 compat_uptr_t __user *head, unsigned int *pi)
3952 if (get_user(*uentry, head))
3955 *entry = compat_ptr((*uentry) & ~1);
3956 *pi = (unsigned int)(*uentry) & 1;
3961 static void __user *futex_uaddr(struct robust_list __user *entry,
3962 compat_long_t futex_offset)
3964 compat_uptr_t base = ptr_to_compat(entry);
3965 void __user *uaddr = compat_ptr(base + futex_offset);
3971 * Walk curr->robust_list (very carefully, it's a userspace list!)
3972 * and mark any locks found there dead, and notify any waiters.
3974 * We silently return on any sign of list-walking problem.
3976 void compat_exit_robust_list(struct task_struct *curr)
3978 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3979 struct robust_list __user *entry, *next_entry, *pending;
3980 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3981 unsigned int uninitialized_var(next_pi);
3982 compat_uptr_t uentry, next_uentry, upending;
3983 compat_long_t futex_offset;
3986 if (!futex_cmpxchg_enabled)
3990 * Fetch the list head (which was registered earlier, via
3991 * sys_set_robust_list()):
3993 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3996 * Fetch the relative futex offset:
3998 if (get_user(futex_offset, &head->futex_offset))
4001 * Fetch any possibly pending lock-add first, and handle it
4004 if (compat_fetch_robust_entry(&upending, &pending,
4005 &head->list_op_pending, &pip))
4008 next_entry = NULL; /* avoid warning with gcc */
4009 while (entry != (struct robust_list __user *) &head->list) {
4011 * Fetch the next entry in the list before calling
4012 * handle_futex_death:
4014 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
4015 (compat_uptr_t __user *)&entry->next, &next_pi);
4017 * A pending lock might already be on the list, so
4018 * dont process it twice:
4020 if (entry != pending) {
4021 void __user *uaddr = futex_uaddr(entry, futex_offset);
4023 if (handle_futex_death(uaddr, curr, pi,
4029 uentry = next_uentry;
4033 * Avoid excessively long or circular lists:
4041 void __user *uaddr = futex_uaddr(pending, futex_offset);
4043 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
4047 COMPAT_SYSCALL_DEFINE2(set_robust_list,
4048 struct compat_robust_list_head __user *, head,
4051 if (!futex_cmpxchg_enabled)
4054 if (unlikely(len != sizeof(*head)))
4057 current->compat_robust_list = head;
4062 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
4063 compat_uptr_t __user *, head_ptr,
4064 compat_size_t __user *, len_ptr)
4066 struct compat_robust_list_head __user *head;
4068 struct task_struct *p;
4070 if (!futex_cmpxchg_enabled)
4079 p = find_task_by_vpid(pid);
4085 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
4088 head = p->compat_robust_list;
4091 if (put_user(sizeof(*head), len_ptr))
4093 return put_user(ptr_to_compat(head), head_ptr);
4101 COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
4102 struct compat_timespec __user *, utime, u32 __user *, uaddr2,
4106 ktime_t t, *tp = NULL;
4108 int cmd = op & FUTEX_CMD_MASK;
4110 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4111 cmd == FUTEX_WAIT_BITSET ||
4112 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4113 if (compat_get_timespec(&ts, utime))
4115 if (!timespec_valid(&ts))
4118 t = timespec_to_ktime(ts);
4119 if (cmd == FUTEX_WAIT)
4120 t = ktime_add_safe(ktime_get(), t);
4123 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4124 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4125 val2 = (int) (unsigned long) utime;
4127 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4129 #endif /* CONFIG_COMPAT */
4131 static void __init futex_detect_cmpxchg(void)
4133 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4137 * This will fail and we want it. Some arch implementations do
4138 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4139 * functionality. We want to know that before we call in any
4140 * of the complex code paths. Also we want to prevent
4141 * registration of robust lists in that case. NULL is
4142 * guaranteed to fault and we get -EFAULT on functional
4143 * implementation, the non-functional ones will return
4146 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4147 futex_cmpxchg_enabled = 1;
4151 static int __init futex_init(void)
4153 unsigned int futex_shift;
4156 #if CONFIG_BASE_SMALL
4157 futex_hashsize = 16;
4159 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4162 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4164 futex_hashsize < 256 ? HASH_SMALL : 0,
4166 futex_hashsize, futex_hashsize);
4167 futex_hashsize = 1UL << futex_shift;
4169 futex_detect_cmpxchg();
4171 for (i = 0; i < futex_hashsize; i++) {
4172 atomic_set(&futex_queues[i].waiters, 0);
4173 plist_head_init(&futex_queues[i].chain);
4174 spin_lock_init(&futex_queues[i].lock);
4179 core_initcall(futex_init);