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/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
68 #include <asm/futex.h>
70 #include "locking/rtmutex_common.h"
73 * Basic futex operation and ordering guarantees:
75 * The waiter reads the futex value in user space and calls
76 * futex_wait(). This function computes the hash bucket and acquires
77 * the hash bucket lock. After that it reads the futex user space value
78 * again and verifies that the data has not changed. If it has not changed
79 * it enqueues itself into the hash bucket, releases the hash bucket lock
82 * The waker side modifies the user space value of the futex and calls
83 * futex_wake(). This function computes the hash bucket and acquires the
84 * hash bucket lock. Then it looks for waiters on that futex in the hash
85 * bucket and wakes them.
87 * In futex wake up scenarios where no tasks are blocked on a futex, taking
88 * the hb spinlock can be avoided and simply return. In order for this
89 * optimization to work, ordering guarantees must exist so that the waiter
90 * being added to the list is acknowledged when the list is concurrently being
91 * checked by the waker, avoiding scenarios like the following:
95 * sys_futex(WAIT, futex, val);
96 * futex_wait(futex, val);
99 * sys_futex(WAKE, futex);
104 * lock(hash_bucket(futex));
106 * unlock(hash_bucket(futex));
109 * This would cause the waiter on CPU 0 to wait forever because it
110 * missed the transition of the user space value from val to newval
111 * and the waker did not find the waiter in the hash bucket queue.
113 * The correct serialization ensures that a waiter either observes
114 * the changed user space value before blocking or is woken by a
119 * sys_futex(WAIT, futex, val);
120 * futex_wait(futex, val);
123 * mb(); (A) <-- paired with -.
125 * lock(hash_bucket(futex)); |
129 * | sys_futex(WAKE, futex);
130 * | futex_wake(futex);
132 * `-------> mb(); (B)
135 * unlock(hash_bucket(futex));
136 * schedule(); if (waiters)
137 * lock(hash_bucket(futex));
138 * wake_waiters(futex);
139 * unlock(hash_bucket(futex));
141 * Where (A) orders the waiters increment and the futex value read -- this
142 * is guaranteed by the head counter in the hb spinlock; and where (B)
143 * orders the write to futex and the waiters read -- this is done by the
144 * barriers in get_futex_key_refs(), through either ihold or atomic_inc,
145 * depending on the futex type.
147 * This yields the following case (where X:=waiters, Y:=futex):
155 * Which guarantees that x==0 && y==0 is impossible; which translates back into
156 * the guarantee that we cannot both miss the futex variable change and the
160 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
161 int __read_mostly futex_cmpxchg_enabled;
165 * Futex flags used to encode options to functions and preserve them across
168 #define FLAGS_SHARED 0x01
169 #define FLAGS_CLOCKRT 0x02
170 #define FLAGS_HAS_TIMEOUT 0x04
173 * Priority Inheritance state:
175 struct futex_pi_state {
177 * list of 'owned' pi_state instances - these have to be
178 * cleaned up in do_exit() if the task exits prematurely:
180 struct list_head list;
185 struct rt_mutex pi_mutex;
187 struct task_struct *owner;
194 * struct futex_q - The hashed futex queue entry, one per waiting task
195 * @list: priority-sorted list of tasks waiting on this futex
196 * @task: the task waiting on the futex
197 * @lock_ptr: the hash bucket lock
198 * @key: the key the futex is hashed on
199 * @pi_state: optional priority inheritance state
200 * @rt_waiter: rt_waiter storage for use with requeue_pi
201 * @requeue_pi_key: the requeue_pi target futex key
202 * @bitset: bitset for the optional bitmasked wakeup
204 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
205 * we can wake only the relevant ones (hashed queues may be shared).
207 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
208 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
209 * The order of wakeup is always to make the first condition true, then
212 * PI futexes are typically woken before they are removed from the hash list via
213 * the rt_mutex code. See unqueue_me_pi().
216 struct plist_node list;
218 struct task_struct *task;
219 spinlock_t *lock_ptr;
221 struct futex_pi_state *pi_state;
222 struct rt_mutex_waiter *rt_waiter;
223 union futex_key *requeue_pi_key;
227 static const struct futex_q futex_q_init = {
228 /* list gets initialized in queue_me()*/
229 .key = FUTEX_KEY_INIT,
230 .bitset = FUTEX_BITSET_MATCH_ANY
234 * Hash buckets are shared by all the futex_keys that hash to the same
235 * location. Each key may have multiple futex_q structures, one for each task
236 * waiting on a futex.
238 struct futex_hash_bucket {
241 struct plist_head chain;
242 } ____cacheline_aligned_in_smp;
244 static unsigned long __read_mostly futex_hashsize;
246 static struct futex_hash_bucket *futex_queues;
248 static inline void futex_get_mm(union futex_key *key)
250 atomic_inc(&key->private.mm->mm_count);
252 * Ensure futex_get_mm() implies a full barrier such that
253 * get_futex_key() implies a full barrier. This is relied upon
254 * as full barrier (B), see the ordering comment above.
256 smp_mb__after_atomic_inc();
260 * Reflects a new waiter being added to the waitqueue.
262 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
265 atomic_inc(&hb->waiters);
267 * Full barrier (A), see the ordering comment above.
269 smp_mb__after_atomic_inc();
274 * Reflects a waiter being removed from the waitqueue by wakeup
277 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
280 atomic_dec(&hb->waiters);
284 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
287 return atomic_read(&hb->waiters);
294 * We hash on the keys returned from get_futex_key (see below).
296 static struct futex_hash_bucket *hash_futex(union futex_key *key)
298 u32 hash = jhash2((u32*)&key->both.word,
299 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
301 return &futex_queues[hash & (futex_hashsize - 1)];
305 * Return 1 if two futex_keys are equal, 0 otherwise.
307 static inline int match_futex(union futex_key *key1, union futex_key *key2)
310 && key1->both.word == key2->both.word
311 && key1->both.ptr == key2->both.ptr
312 && key1->both.offset == key2->both.offset);
316 * Take a reference to the resource addressed by a key.
317 * Can be called while holding spinlocks.
320 static void get_futex_key_refs(union futex_key *key)
325 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
327 ihold(key->shared.inode); /* implies MB (B) */
329 case FUT_OFF_MMSHARED:
330 futex_get_mm(key); /* implies MB (B) */
336 * Drop a reference to the resource addressed by a key.
337 * The hash bucket spinlock must not be held.
339 static void drop_futex_key_refs(union futex_key *key)
341 if (!key->both.ptr) {
342 /* If we're here then we tried to put a key we failed to get */
347 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
349 iput(key->shared.inode);
351 case FUT_OFF_MMSHARED:
352 mmdrop(key->private.mm);
358 * get_futex_key() - Get parameters which are the keys for a futex
359 * @uaddr: virtual address of the futex
360 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
361 * @key: address where result is stored.
362 * @rw: mapping needs to be read/write (values: VERIFY_READ,
365 * Return: a negative error code or 0
367 * The key words are stored in *key on success.
369 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
370 * offset_within_page). For private mappings, it's (uaddr, current->mm).
371 * We can usually work out the index without swapping in the page.
373 * lock_page() might sleep, the caller should not hold a spinlock.
376 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
378 unsigned long address = (unsigned long)uaddr;
379 struct mm_struct *mm = current->mm;
380 struct page *page, *page_head;
384 * The futex address must be "naturally" aligned.
386 key->both.offset = address % PAGE_SIZE;
387 if (unlikely((address % sizeof(u32)) != 0))
389 address -= key->both.offset;
391 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
395 * PROCESS_PRIVATE futexes are fast.
396 * As the mm cannot disappear under us and the 'key' only needs
397 * virtual address, we dont even have to find the underlying vma.
398 * Note : We do have to check 'uaddr' is a valid user address,
399 * but access_ok() should be faster than find_vma()
402 key->private.mm = mm;
403 key->private.address = address;
404 get_futex_key_refs(key); /* implies MB (B) */
409 err = get_user_pages_fast(address, 1, 1, &page);
411 * If write access is not required (eg. FUTEX_WAIT), try
412 * and get read-only access.
414 if (err == -EFAULT && rw == VERIFY_READ) {
415 err = get_user_pages_fast(address, 1, 0, &page);
423 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
425 if (unlikely(PageTail(page))) {
427 /* serialize against __split_huge_page_splitting() */
429 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
430 page_head = compound_head(page);
432 * page_head is valid pointer but we must pin
433 * it before taking the PG_lock and/or
434 * PG_compound_lock. The moment we re-enable
435 * irqs __split_huge_page_splitting() can
436 * return and the head page can be freed from
437 * under us. We can't take the PG_lock and/or
438 * PG_compound_lock on a page that could be
439 * freed from under us.
441 if (page != page_head) {
452 page_head = compound_head(page);
453 if (page != page_head) {
459 lock_page(page_head);
462 * If page_head->mapping is NULL, then it cannot be a PageAnon
463 * page; but it might be the ZERO_PAGE or in the gate area or
464 * in a special mapping (all cases which we are happy to fail);
465 * or it may have been a good file page when get_user_pages_fast
466 * found it, but truncated or holepunched or subjected to
467 * invalidate_complete_page2 before we got the page lock (also
468 * cases which we are happy to fail). And we hold a reference,
469 * so refcount care in invalidate_complete_page's remove_mapping
470 * prevents drop_caches from setting mapping to NULL beneath us.
472 * The case we do have to guard against is when memory pressure made
473 * shmem_writepage move it from filecache to swapcache beneath us:
474 * an unlikely race, but we do need to retry for page_head->mapping.
476 if (!page_head->mapping) {
477 int shmem_swizzled = PageSwapCache(page_head);
478 unlock_page(page_head);
486 * Private mappings are handled in a simple way.
488 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
489 * it's a read-only handle, it's expected that futexes attach to
490 * the object not the particular process.
492 if (PageAnon(page_head)) {
494 * A RO anonymous page will never change and thus doesn't make
495 * sense for futex operations.
502 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
503 key->private.mm = mm;
504 key->private.address = address;
506 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
507 key->shared.inode = page_head->mapping->host;
508 key->shared.pgoff = basepage_index(page);
511 get_futex_key_refs(key); /* implies MB (B) */
514 unlock_page(page_head);
519 static inline void put_futex_key(union futex_key *key)
521 drop_futex_key_refs(key);
525 * fault_in_user_writeable() - Fault in user address and verify RW access
526 * @uaddr: pointer to faulting user space address
528 * Slow path to fixup the fault we just took in the atomic write
531 * We have no generic implementation of a non-destructive write to the
532 * user address. We know that we faulted in the atomic pagefault
533 * disabled section so we can as well avoid the #PF overhead by
534 * calling get_user_pages() right away.
536 static int fault_in_user_writeable(u32 __user *uaddr)
538 struct mm_struct *mm = current->mm;
541 down_read(&mm->mmap_sem);
542 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
544 up_read(&mm->mmap_sem);
546 return ret < 0 ? ret : 0;
550 * futex_top_waiter() - Return the highest priority waiter on a futex
551 * @hb: the hash bucket the futex_q's reside in
552 * @key: the futex key (to distinguish it from other futex futex_q's)
554 * Must be called with the hb lock held.
556 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
557 union futex_key *key)
559 struct futex_q *this;
561 plist_for_each_entry(this, &hb->chain, list) {
562 if (match_futex(&this->key, key))
568 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
569 u32 uval, u32 newval)
574 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
580 static int get_futex_value_locked(u32 *dest, u32 __user *from)
585 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
588 return ret ? -EFAULT : 0;
595 static int refill_pi_state_cache(void)
597 struct futex_pi_state *pi_state;
599 if (likely(current->pi_state_cache))
602 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
607 INIT_LIST_HEAD(&pi_state->list);
608 /* pi_mutex gets initialized later */
609 pi_state->owner = NULL;
610 atomic_set(&pi_state->refcount, 1);
611 pi_state->key = FUTEX_KEY_INIT;
613 current->pi_state_cache = pi_state;
618 static struct futex_pi_state * alloc_pi_state(void)
620 struct futex_pi_state *pi_state = current->pi_state_cache;
623 current->pi_state_cache = NULL;
628 static void free_pi_state(struct futex_pi_state *pi_state)
630 if (!atomic_dec_and_test(&pi_state->refcount))
634 * If pi_state->owner is NULL, the owner is most probably dying
635 * and has cleaned up the pi_state already
637 if (pi_state->owner) {
638 raw_spin_lock_irq(&pi_state->owner->pi_lock);
639 list_del_init(&pi_state->list);
640 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
642 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
645 if (current->pi_state_cache)
649 * pi_state->list is already empty.
650 * clear pi_state->owner.
651 * refcount is at 0 - put it back to 1.
653 pi_state->owner = NULL;
654 atomic_set(&pi_state->refcount, 1);
655 current->pi_state_cache = pi_state;
660 * Look up the task based on what TID userspace gave us.
663 static struct task_struct * futex_find_get_task(pid_t pid)
665 struct task_struct *p;
668 p = find_task_by_vpid(pid);
678 * This task is holding PI mutexes at exit time => bad.
679 * Kernel cleans up PI-state, but userspace is likely hosed.
680 * (Robust-futex cleanup is separate and might save the day for userspace.)
682 void exit_pi_state_list(struct task_struct *curr)
684 struct list_head *next, *head = &curr->pi_state_list;
685 struct futex_pi_state *pi_state;
686 struct futex_hash_bucket *hb;
687 union futex_key key = FUTEX_KEY_INIT;
689 if (!futex_cmpxchg_enabled)
692 * We are a ZOMBIE and nobody can enqueue itself on
693 * pi_state_list anymore, but we have to be careful
694 * versus waiters unqueueing themselves:
696 raw_spin_lock_irq(&curr->pi_lock);
697 while (!list_empty(head)) {
700 pi_state = list_entry(next, struct futex_pi_state, list);
702 hb = hash_futex(&key);
703 raw_spin_unlock_irq(&curr->pi_lock);
705 spin_lock(&hb->lock);
707 raw_spin_lock_irq(&curr->pi_lock);
709 * We dropped the pi-lock, so re-check whether this
710 * task still owns the PI-state:
712 if (head->next != next) {
713 spin_unlock(&hb->lock);
717 WARN_ON(pi_state->owner != curr);
718 WARN_ON(list_empty(&pi_state->list));
719 list_del_init(&pi_state->list);
720 pi_state->owner = NULL;
721 raw_spin_unlock_irq(&curr->pi_lock);
723 rt_mutex_unlock(&pi_state->pi_mutex);
725 spin_unlock(&hb->lock);
727 raw_spin_lock_irq(&curr->pi_lock);
729 raw_spin_unlock_irq(&curr->pi_lock);
733 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
734 union futex_key *key, struct futex_pi_state **ps,
735 struct task_struct *task)
737 struct futex_pi_state *pi_state = NULL;
738 struct futex_q *this, *next;
739 struct task_struct *p;
740 pid_t pid = uval & FUTEX_TID_MASK;
742 plist_for_each_entry_safe(this, next, &hb->chain, list) {
743 if (match_futex(&this->key, key)) {
745 * Another waiter already exists - bump up
746 * the refcount and return its pi_state:
748 pi_state = this->pi_state;
750 * Userspace might have messed up non-PI and PI futexes
752 if (unlikely(!pi_state))
755 WARN_ON(!atomic_read(&pi_state->refcount));
758 * When pi_state->owner is NULL then the owner died
759 * and another waiter is on the fly. pi_state->owner
760 * is fixed up by the task which acquires
761 * pi_state->rt_mutex.
763 * We do not check for pid == 0 which can happen when
764 * the owner died and robust_list_exit() cleared the
767 if (pid && pi_state->owner) {
769 * Bail out if user space manipulated the
772 if (pid != task_pid_vnr(pi_state->owner))
777 * Protect against a corrupted uval. If uval
778 * is 0x80000000 then pid is 0 and the waiter
779 * bit is set. So the deadlock check in the
780 * calling code has failed and we did not fall
781 * into the check above due to !pid.
783 if (task && pi_state->owner == task)
786 atomic_inc(&pi_state->refcount);
794 * We are the first waiter - try to look up the real owner and attach
795 * the new pi_state to it, but bail out when TID = 0
799 p = futex_find_get_task(pid);
804 * We need to look at the task state flags to figure out,
805 * whether the task is exiting. To protect against the do_exit
806 * change of the task flags, we do this protected by
809 raw_spin_lock_irq(&p->pi_lock);
810 if (unlikely(p->flags & PF_EXITING)) {
812 * The task is on the way out. When PF_EXITPIDONE is
813 * set, we know that the task has finished the
816 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
818 raw_spin_unlock_irq(&p->pi_lock);
823 pi_state = alloc_pi_state();
826 * Initialize the pi_mutex in locked state and make 'p'
829 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
831 /* Store the key for possible exit cleanups: */
832 pi_state->key = *key;
834 WARN_ON(!list_empty(&pi_state->list));
835 list_add(&pi_state->list, &p->pi_state_list);
837 raw_spin_unlock_irq(&p->pi_lock);
847 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
848 * @uaddr: the pi futex user address
849 * @hb: the pi futex hash bucket
850 * @key: the futex key associated with uaddr and hb
851 * @ps: the pi_state pointer where we store the result of the
853 * @task: the task to perform the atomic lock work for. This will
854 * be "current" except in the case of requeue pi.
855 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
859 * 1 - acquired the lock;
862 * The hb->lock and futex_key refs shall be held by the caller.
864 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
865 union futex_key *key,
866 struct futex_pi_state **ps,
867 struct task_struct *task, int set_waiters)
869 int lock_taken, ret, force_take = 0;
870 u32 uval, newval, curval, vpid = task_pid_vnr(task);
873 ret = lock_taken = 0;
876 * To avoid races, we attempt to take the lock here again
877 * (by doing a 0 -> TID atomic cmpxchg), while holding all
878 * the locks. It will most likely not succeed.
882 newval |= FUTEX_WAITERS;
884 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
890 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
894 * Surprise - we got the lock. Just return to userspace:
896 if (unlikely(!curval))
902 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
903 * to wake at the next unlock.
905 newval = curval | FUTEX_WAITERS;
908 * Should we force take the futex? See below.
910 if (unlikely(force_take)) {
912 * Keep the OWNER_DIED and the WAITERS bit and set the
915 newval = (curval & ~FUTEX_TID_MASK) | vpid;
920 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
922 if (unlikely(curval != uval))
926 * We took the lock due to forced take over.
928 if (unlikely(lock_taken))
932 * We dont have the lock. Look up the PI state (or create it if
933 * we are the first waiter):
935 ret = lookup_pi_state(uval, hb, key, ps, task);
941 * We failed to find an owner for this
942 * futex. So we have no pi_state to block
943 * on. This can happen in two cases:
946 * 2) A stale FUTEX_WAITERS bit
948 * Re-read the futex value.
950 if (get_futex_value_locked(&curval, uaddr))
954 * If the owner died or we have a stale
955 * WAITERS bit the owner TID in the user space
958 if (!(curval & FUTEX_TID_MASK)) {
971 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
972 * @q: The futex_q to unqueue
974 * The q->lock_ptr must not be NULL and must be held by the caller.
976 static void __unqueue_futex(struct futex_q *q)
978 struct futex_hash_bucket *hb;
980 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
981 || WARN_ON(plist_node_empty(&q->list)))
984 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
985 plist_del(&q->list, &hb->chain);
990 * The hash bucket lock must be held when this is called.
991 * Afterwards, the futex_q must not be accessed.
993 static void wake_futex(struct futex_q *q)
995 struct task_struct *p = q->task;
997 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1001 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1002 * a non-futex wake up happens on another CPU then the task
1003 * might exit and p would dereference a non-existing task
1004 * struct. Prevent this by holding a reference on p across the
1011 * The waiting task can free the futex_q as soon as
1012 * q->lock_ptr = NULL is written, without taking any locks. A
1013 * memory barrier is required here to prevent the following
1014 * store to lock_ptr from getting ahead of the plist_del.
1019 wake_up_state(p, TASK_NORMAL);
1023 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1025 struct task_struct *new_owner;
1026 struct futex_pi_state *pi_state = this->pi_state;
1027 u32 uninitialized_var(curval), newval;
1033 * If current does not own the pi_state then the futex is
1034 * inconsistent and user space fiddled with the futex value.
1036 if (pi_state->owner != current)
1039 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1040 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1043 * It is possible that the next waiter (the one that brought
1044 * this owner to the kernel) timed out and is no longer
1045 * waiting on the lock.
1048 new_owner = this->task;
1051 * We pass it to the next owner. (The WAITERS bit is always
1052 * kept enabled while there is PI state around. We must also
1053 * preserve the owner died bit.)
1055 if (!(uval & FUTEX_OWNER_DIED)) {
1058 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1060 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1062 else if (curval != uval)
1065 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1070 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1071 WARN_ON(list_empty(&pi_state->list));
1072 list_del_init(&pi_state->list);
1073 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1075 raw_spin_lock_irq(&new_owner->pi_lock);
1076 WARN_ON(!list_empty(&pi_state->list));
1077 list_add(&pi_state->list, &new_owner->pi_state_list);
1078 pi_state->owner = new_owner;
1079 raw_spin_unlock_irq(&new_owner->pi_lock);
1081 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1082 rt_mutex_unlock(&pi_state->pi_mutex);
1087 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1089 u32 uninitialized_var(oldval);
1092 * There is no waiter, so we unlock the futex. The owner died
1093 * bit has not to be preserved here. We are the owner:
1095 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1104 * Express the locking dependencies for lockdep:
1107 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1110 spin_lock(&hb1->lock);
1112 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1113 } else { /* hb1 > hb2 */
1114 spin_lock(&hb2->lock);
1115 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1120 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1122 spin_unlock(&hb1->lock);
1124 spin_unlock(&hb2->lock);
1128 * Wake up waiters matching bitset queued on this futex (uaddr).
1131 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1133 struct futex_hash_bucket *hb;
1134 struct futex_q *this, *next;
1135 union futex_key key = FUTEX_KEY_INIT;
1141 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1142 if (unlikely(ret != 0))
1145 hb = hash_futex(&key);
1147 /* Make sure we really have tasks to wakeup */
1148 if (!hb_waiters_pending(hb))
1151 spin_lock(&hb->lock);
1153 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1154 if (match_futex (&this->key, &key)) {
1155 if (this->pi_state || this->rt_waiter) {
1160 /* Check if one of the bits is set in both bitsets */
1161 if (!(this->bitset & bitset))
1165 if (++ret >= nr_wake)
1170 spin_unlock(&hb->lock);
1172 put_futex_key(&key);
1178 * Wake up all waiters hashed on the physical page that is mapped
1179 * to this virtual address:
1182 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1183 int nr_wake, int nr_wake2, int op)
1185 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1186 struct futex_hash_bucket *hb1, *hb2;
1187 struct futex_q *this, *next;
1191 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1192 if (unlikely(ret != 0))
1194 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1195 if (unlikely(ret != 0))
1198 hb1 = hash_futex(&key1);
1199 hb2 = hash_futex(&key2);
1202 double_lock_hb(hb1, hb2);
1203 op_ret = futex_atomic_op_inuser(op, uaddr2);
1204 if (unlikely(op_ret < 0)) {
1206 double_unlock_hb(hb1, hb2);
1210 * we don't get EFAULT from MMU faults if we don't have an MMU,
1211 * but we might get them from range checking
1217 if (unlikely(op_ret != -EFAULT)) {
1222 ret = fault_in_user_writeable(uaddr2);
1226 if (!(flags & FLAGS_SHARED))
1229 put_futex_key(&key2);
1230 put_futex_key(&key1);
1234 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1235 if (match_futex (&this->key, &key1)) {
1236 if (this->pi_state || this->rt_waiter) {
1241 if (++ret >= nr_wake)
1248 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1249 if (match_futex (&this->key, &key2)) {
1250 if (this->pi_state || this->rt_waiter) {
1255 if (++op_ret >= nr_wake2)
1263 double_unlock_hb(hb1, hb2);
1265 put_futex_key(&key2);
1267 put_futex_key(&key1);
1273 * requeue_futex() - Requeue a futex_q from one hb to another
1274 * @q: the futex_q to requeue
1275 * @hb1: the source hash_bucket
1276 * @hb2: the target hash_bucket
1277 * @key2: the new key for the requeued futex_q
1280 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1281 struct futex_hash_bucket *hb2, union futex_key *key2)
1285 * If key1 and key2 hash to the same bucket, no need to
1288 if (likely(&hb1->chain != &hb2->chain)) {
1289 plist_del(&q->list, &hb1->chain);
1290 hb_waiters_dec(hb1);
1291 plist_add(&q->list, &hb2->chain);
1292 hb_waiters_inc(hb2);
1293 q->lock_ptr = &hb2->lock;
1295 get_futex_key_refs(key2);
1300 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1302 * @key: the key of the requeue target futex
1303 * @hb: the hash_bucket of the requeue target futex
1305 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1306 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1307 * to the requeue target futex so the waiter can detect the wakeup on the right
1308 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1309 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1310 * to protect access to the pi_state to fixup the owner later. Must be called
1311 * with both q->lock_ptr and hb->lock held.
1314 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1315 struct futex_hash_bucket *hb)
1317 get_futex_key_refs(key);
1322 WARN_ON(!q->rt_waiter);
1323 q->rt_waiter = NULL;
1325 q->lock_ptr = &hb->lock;
1327 wake_up_state(q->task, TASK_NORMAL);
1331 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1332 * @pifutex: the user address of the to futex
1333 * @hb1: the from futex hash bucket, must be locked by the caller
1334 * @hb2: the to futex hash bucket, must be locked by the caller
1335 * @key1: the from futex key
1336 * @key2: the to futex key
1337 * @ps: address to store the pi_state pointer
1338 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1340 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1341 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1342 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1343 * hb1 and hb2 must be held by the caller.
1346 * 0 - failed to acquire the lock atomically;
1347 * >0 - acquired the lock, return value is vpid of the top_waiter
1350 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1351 struct futex_hash_bucket *hb1,
1352 struct futex_hash_bucket *hb2,
1353 union futex_key *key1, union futex_key *key2,
1354 struct futex_pi_state **ps, int set_waiters)
1356 struct futex_q *top_waiter = NULL;
1360 if (get_futex_value_locked(&curval, pifutex))
1364 * Find the top_waiter and determine if there are additional waiters.
1365 * If the caller intends to requeue more than 1 waiter to pifutex,
1366 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1367 * as we have means to handle the possible fault. If not, don't set
1368 * the bit unecessarily as it will force the subsequent unlock to enter
1371 top_waiter = futex_top_waiter(hb1, key1);
1373 /* There are no waiters, nothing for us to do. */
1377 /* Ensure we requeue to the expected futex. */
1378 if (!match_futex(top_waiter->requeue_pi_key, key2))
1382 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1383 * the contended case or if set_waiters is 1. The pi_state is returned
1384 * in ps in contended cases.
1386 vpid = task_pid_vnr(top_waiter->task);
1387 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1390 requeue_pi_wake_futex(top_waiter, key2, hb2);
1397 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1398 * @uaddr1: source futex user address
1399 * @flags: futex flags (FLAGS_SHARED, etc.)
1400 * @uaddr2: target futex user address
1401 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1402 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1403 * @cmpval: @uaddr1 expected value (or %NULL)
1404 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1405 * pi futex (pi to pi requeue is not supported)
1407 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1408 * uaddr2 atomically on behalf of the top waiter.
1411 * >=0 - on success, the number of tasks requeued or woken;
1414 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1415 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1416 u32 *cmpval, int requeue_pi)
1418 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1419 int drop_count = 0, task_count = 0, ret;
1420 struct futex_pi_state *pi_state = NULL;
1421 struct futex_hash_bucket *hb1, *hb2;
1422 struct futex_q *this, *next;
1426 * requeue_pi requires a pi_state, try to allocate it now
1427 * without any locks in case it fails.
1429 if (refill_pi_state_cache())
1432 * requeue_pi must wake as many tasks as it can, up to nr_wake
1433 * + nr_requeue, since it acquires the rt_mutex prior to
1434 * returning to userspace, so as to not leave the rt_mutex with
1435 * waiters and no owner. However, second and third wake-ups
1436 * cannot be predicted as they involve race conditions with the
1437 * first wake and a fault while looking up the pi_state. Both
1438 * pthread_cond_signal() and pthread_cond_broadcast() should
1446 if (pi_state != NULL) {
1448 * We will have to lookup the pi_state again, so free this one
1449 * to keep the accounting correct.
1451 free_pi_state(pi_state);
1455 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1456 if (unlikely(ret != 0))
1458 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1459 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1460 if (unlikely(ret != 0))
1463 hb1 = hash_futex(&key1);
1464 hb2 = hash_futex(&key2);
1467 hb_waiters_inc(hb2);
1468 double_lock_hb(hb1, hb2);
1470 if (likely(cmpval != NULL)) {
1473 ret = get_futex_value_locked(&curval, uaddr1);
1475 if (unlikely(ret)) {
1476 double_unlock_hb(hb1, hb2);
1477 hb_waiters_dec(hb2);
1479 ret = get_user(curval, uaddr1);
1483 if (!(flags & FLAGS_SHARED))
1486 put_futex_key(&key2);
1487 put_futex_key(&key1);
1490 if (curval != *cmpval) {
1496 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1498 * Attempt to acquire uaddr2 and wake the top waiter. If we
1499 * intend to requeue waiters, force setting the FUTEX_WAITERS
1500 * bit. We force this here where we are able to easily handle
1501 * faults rather in the requeue loop below.
1503 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1504 &key2, &pi_state, nr_requeue);
1507 * At this point the top_waiter has either taken uaddr2 or is
1508 * waiting on it. If the former, then the pi_state will not
1509 * exist yet, look it up one more time to ensure we have a
1510 * reference to it. If the lock was taken, ret contains the
1511 * vpid of the top waiter task.
1518 * If we acquired the lock, then the user
1519 * space value of uaddr2 should be vpid. It
1520 * cannot be changed by the top waiter as it
1521 * is blocked on hb2 lock if it tries to do
1522 * so. If something fiddled with it behind our
1523 * back the pi state lookup might unearth
1524 * it. So we rather use the known value than
1525 * rereading and handing potential crap to
1528 ret = lookup_pi_state(ret, hb2, &key2, &pi_state, NULL);
1535 double_unlock_hb(hb1, hb2);
1536 hb_waiters_dec(hb2);
1537 put_futex_key(&key2);
1538 put_futex_key(&key1);
1539 ret = fault_in_user_writeable(uaddr2);
1544 /* The owner was exiting, try again. */
1545 double_unlock_hb(hb1, hb2);
1546 hb_waiters_dec(hb2);
1547 put_futex_key(&key2);
1548 put_futex_key(&key1);
1556 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1557 if (task_count - nr_wake >= nr_requeue)
1560 if (!match_futex(&this->key, &key1))
1564 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1565 * be paired with each other and no other futex ops.
1567 * We should never be requeueing a futex_q with a pi_state,
1568 * which is awaiting a futex_unlock_pi().
1570 if ((requeue_pi && !this->rt_waiter) ||
1571 (!requeue_pi && this->rt_waiter) ||
1578 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1579 * lock, we already woke the top_waiter. If not, it will be
1580 * woken by futex_unlock_pi().
1582 if (++task_count <= nr_wake && !requeue_pi) {
1587 /* Ensure we requeue to the expected futex for requeue_pi. */
1588 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1594 * Requeue nr_requeue waiters and possibly one more in the case
1595 * of requeue_pi if we couldn't acquire the lock atomically.
1598 /* Prepare the waiter to take the rt_mutex. */
1599 atomic_inc(&pi_state->refcount);
1600 this->pi_state = pi_state;
1601 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1605 /* We got the lock. */
1606 requeue_pi_wake_futex(this, &key2, hb2);
1611 this->pi_state = NULL;
1612 free_pi_state(pi_state);
1616 requeue_futex(this, hb1, hb2, &key2);
1621 double_unlock_hb(hb1, hb2);
1622 hb_waiters_dec(hb2);
1625 * drop_futex_key_refs() must be called outside the spinlocks. During
1626 * the requeue we moved futex_q's from the hash bucket at key1 to the
1627 * one at key2 and updated their key pointer. We no longer need to
1628 * hold the references to key1.
1630 while (--drop_count >= 0)
1631 drop_futex_key_refs(&key1);
1634 put_futex_key(&key2);
1636 put_futex_key(&key1);
1638 if (pi_state != NULL)
1639 free_pi_state(pi_state);
1640 return ret ? ret : task_count;
1643 /* The key must be already stored in q->key. */
1644 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1645 __acquires(&hb->lock)
1647 struct futex_hash_bucket *hb;
1649 hb = hash_futex(&q->key);
1652 * Increment the counter before taking the lock so that
1653 * a potential waker won't miss a to-be-slept task that is
1654 * waiting for the spinlock. This is safe as all queue_lock()
1655 * users end up calling queue_me(). Similarly, for housekeeping,
1656 * decrement the counter at queue_unlock() when some error has
1657 * occurred and we don't end up adding the task to the list.
1661 q->lock_ptr = &hb->lock;
1663 spin_lock(&hb->lock); /* implies MB (A) */
1668 queue_unlock(struct futex_hash_bucket *hb)
1669 __releases(&hb->lock)
1671 spin_unlock(&hb->lock);
1676 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1677 * @q: The futex_q to enqueue
1678 * @hb: The destination hash bucket
1680 * The hb->lock must be held by the caller, and is released here. A call to
1681 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1682 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1683 * or nothing if the unqueue is done as part of the wake process and the unqueue
1684 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1687 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1688 __releases(&hb->lock)
1693 * The priority used to register this element is
1694 * - either the real thread-priority for the real-time threads
1695 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1696 * - or MAX_RT_PRIO for non-RT threads.
1697 * Thus, all RT-threads are woken first in priority order, and
1698 * the others are woken last, in FIFO order.
1700 prio = min(current->normal_prio, MAX_RT_PRIO);
1702 plist_node_init(&q->list, prio);
1703 plist_add(&q->list, &hb->chain);
1705 spin_unlock(&hb->lock);
1709 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1710 * @q: The futex_q to unqueue
1712 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1713 * be paired with exactly one earlier call to queue_me().
1716 * 1 - if the futex_q was still queued (and we removed unqueued it);
1717 * 0 - if the futex_q was already removed by the waking thread
1719 static int unqueue_me(struct futex_q *q)
1721 spinlock_t *lock_ptr;
1724 /* In the common case we don't take the spinlock, which is nice. */
1726 lock_ptr = q->lock_ptr;
1728 if (lock_ptr != NULL) {
1729 spin_lock(lock_ptr);
1731 * q->lock_ptr can change between reading it and
1732 * spin_lock(), causing us to take the wrong lock. This
1733 * corrects the race condition.
1735 * Reasoning goes like this: if we have the wrong lock,
1736 * q->lock_ptr must have changed (maybe several times)
1737 * between reading it and the spin_lock(). It can
1738 * change again after the spin_lock() but only if it was
1739 * already changed before the spin_lock(). It cannot,
1740 * however, change back to the original value. Therefore
1741 * we can detect whether we acquired the correct lock.
1743 if (unlikely(lock_ptr != q->lock_ptr)) {
1744 spin_unlock(lock_ptr);
1749 BUG_ON(q->pi_state);
1751 spin_unlock(lock_ptr);
1755 drop_futex_key_refs(&q->key);
1760 * PI futexes can not be requeued and must remove themself from the
1761 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1764 static void unqueue_me_pi(struct futex_q *q)
1765 __releases(q->lock_ptr)
1769 BUG_ON(!q->pi_state);
1770 free_pi_state(q->pi_state);
1773 spin_unlock(q->lock_ptr);
1777 * Fixup the pi_state owner with the new owner.
1779 * Must be called with hash bucket lock held and mm->sem held for non
1782 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1783 struct task_struct *newowner)
1785 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1786 struct futex_pi_state *pi_state = q->pi_state;
1787 struct task_struct *oldowner = pi_state->owner;
1788 u32 uval, uninitialized_var(curval), newval;
1792 if (!pi_state->owner)
1793 newtid |= FUTEX_OWNER_DIED;
1796 * We are here either because we stole the rtmutex from the
1797 * previous highest priority waiter or we are the highest priority
1798 * waiter but failed to get the rtmutex the first time.
1799 * We have to replace the newowner TID in the user space variable.
1800 * This must be atomic as we have to preserve the owner died bit here.
1802 * Note: We write the user space value _before_ changing the pi_state
1803 * because we can fault here. Imagine swapped out pages or a fork
1804 * that marked all the anonymous memory readonly for cow.
1806 * Modifying pi_state _before_ the user space value would
1807 * leave the pi_state in an inconsistent state when we fault
1808 * here, because we need to drop the hash bucket lock to
1809 * handle the fault. This might be observed in the PID check
1810 * in lookup_pi_state.
1813 if (get_futex_value_locked(&uval, uaddr))
1817 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1819 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1827 * We fixed up user space. Now we need to fix the pi_state
1830 if (pi_state->owner != NULL) {
1831 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1832 WARN_ON(list_empty(&pi_state->list));
1833 list_del_init(&pi_state->list);
1834 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1837 pi_state->owner = newowner;
1839 raw_spin_lock_irq(&newowner->pi_lock);
1840 WARN_ON(!list_empty(&pi_state->list));
1841 list_add(&pi_state->list, &newowner->pi_state_list);
1842 raw_spin_unlock_irq(&newowner->pi_lock);
1846 * To handle the page fault we need to drop the hash bucket
1847 * lock here. That gives the other task (either the highest priority
1848 * waiter itself or the task which stole the rtmutex) the
1849 * chance to try the fixup of the pi_state. So once we are
1850 * back from handling the fault we need to check the pi_state
1851 * after reacquiring the hash bucket lock and before trying to
1852 * do another fixup. When the fixup has been done already we
1856 spin_unlock(q->lock_ptr);
1858 ret = fault_in_user_writeable(uaddr);
1860 spin_lock(q->lock_ptr);
1863 * Check if someone else fixed it for us:
1865 if (pi_state->owner != oldowner)
1874 static long futex_wait_restart(struct restart_block *restart);
1877 * fixup_owner() - Post lock pi_state and corner case management
1878 * @uaddr: user address of the futex
1879 * @q: futex_q (contains pi_state and access to the rt_mutex)
1880 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1882 * After attempting to lock an rt_mutex, this function is called to cleanup
1883 * the pi_state owner as well as handle race conditions that may allow us to
1884 * acquire the lock. Must be called with the hb lock held.
1887 * 1 - success, lock taken;
1888 * 0 - success, lock not taken;
1889 * <0 - on error (-EFAULT)
1891 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1893 struct task_struct *owner;
1898 * Got the lock. We might not be the anticipated owner if we
1899 * did a lock-steal - fix up the PI-state in that case:
1901 if (q->pi_state->owner != current)
1902 ret = fixup_pi_state_owner(uaddr, q, current);
1907 * Catch the rare case, where the lock was released when we were on the
1908 * way back before we locked the hash bucket.
1910 if (q->pi_state->owner == current) {
1912 * Try to get the rt_mutex now. This might fail as some other
1913 * task acquired the rt_mutex after we removed ourself from the
1914 * rt_mutex waiters list.
1916 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1922 * pi_state is incorrect, some other task did a lock steal and
1923 * we returned due to timeout or signal without taking the
1924 * rt_mutex. Too late.
1926 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1927 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1929 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1930 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1931 ret = fixup_pi_state_owner(uaddr, q, owner);
1936 * Paranoia check. If we did not take the lock, then we should not be
1937 * the owner of the rt_mutex.
1939 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1940 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1941 "pi-state %p\n", ret,
1942 q->pi_state->pi_mutex.owner,
1943 q->pi_state->owner);
1946 return ret ? ret : locked;
1950 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1951 * @hb: the futex hash bucket, must be locked by the caller
1952 * @q: the futex_q to queue up on
1953 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1955 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1956 struct hrtimer_sleeper *timeout)
1959 * The task state is guaranteed to be set before another task can
1960 * wake it. set_current_state() is implemented using set_mb() and
1961 * queue_me() calls spin_unlock() upon completion, both serializing
1962 * access to the hash list and forcing another memory barrier.
1964 set_current_state(TASK_INTERRUPTIBLE);
1969 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1970 if (!hrtimer_active(&timeout->timer))
1971 timeout->task = NULL;
1975 * If we have been removed from the hash list, then another task
1976 * has tried to wake us, and we can skip the call to schedule().
1978 if (likely(!plist_node_empty(&q->list))) {
1980 * If the timer has already expired, current will already be
1981 * flagged for rescheduling. Only call schedule if there
1982 * is no timeout, or if it has yet to expire.
1984 if (!timeout || timeout->task)
1985 freezable_schedule();
1987 __set_current_state(TASK_RUNNING);
1991 * futex_wait_setup() - Prepare to wait on a futex
1992 * @uaddr: the futex userspace address
1993 * @val: the expected value
1994 * @flags: futex flags (FLAGS_SHARED, etc.)
1995 * @q: the associated futex_q
1996 * @hb: storage for hash_bucket pointer to be returned to caller
1998 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1999 * compare it with the expected value. Handle atomic faults internally.
2000 * Return with the hb lock held and a q.key reference on success, and unlocked
2001 * with no q.key reference on failure.
2004 * 0 - uaddr contains val and hb has been locked;
2005 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2007 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2008 struct futex_q *q, struct futex_hash_bucket **hb)
2014 * Access the page AFTER the hash-bucket is locked.
2015 * Order is important:
2017 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2018 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2020 * The basic logical guarantee of a futex is that it blocks ONLY
2021 * if cond(var) is known to be true at the time of blocking, for
2022 * any cond. If we locked the hash-bucket after testing *uaddr, that
2023 * would open a race condition where we could block indefinitely with
2024 * cond(var) false, which would violate the guarantee.
2026 * On the other hand, we insert q and release the hash-bucket only
2027 * after testing *uaddr. This guarantees that futex_wait() will NOT
2028 * absorb a wakeup if *uaddr does not match the desired values
2029 * while the syscall executes.
2032 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2033 if (unlikely(ret != 0))
2037 *hb = queue_lock(q);
2039 ret = get_futex_value_locked(&uval, uaddr);
2044 ret = get_user(uval, uaddr);
2048 if (!(flags & FLAGS_SHARED))
2051 put_futex_key(&q->key);
2062 put_futex_key(&q->key);
2066 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2067 ktime_t *abs_time, u32 bitset)
2069 struct hrtimer_sleeper timeout, *to = NULL;
2070 struct restart_block *restart;
2071 struct futex_hash_bucket *hb;
2072 struct futex_q q = futex_q_init;
2082 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2083 CLOCK_REALTIME : CLOCK_MONOTONIC,
2085 hrtimer_init_sleeper(to, current);
2086 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2087 current->timer_slack_ns);
2092 * Prepare to wait on uaddr. On success, holds hb lock and increments
2095 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2099 /* queue_me and wait for wakeup, timeout, or a signal. */
2100 futex_wait_queue_me(hb, &q, to);
2102 /* If we were woken (and unqueued), we succeeded, whatever. */
2104 /* unqueue_me() drops q.key ref */
2105 if (!unqueue_me(&q))
2108 if (to && !to->task)
2112 * We expect signal_pending(current), but we might be the
2113 * victim of a spurious wakeup as well.
2115 if (!signal_pending(current))
2122 restart = ¤t_thread_info()->restart_block;
2123 restart->fn = futex_wait_restart;
2124 restart->futex.uaddr = uaddr;
2125 restart->futex.val = val;
2126 restart->futex.time = abs_time->tv64;
2127 restart->futex.bitset = bitset;
2128 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2130 ret = -ERESTART_RESTARTBLOCK;
2134 hrtimer_cancel(&to->timer);
2135 destroy_hrtimer_on_stack(&to->timer);
2141 static long futex_wait_restart(struct restart_block *restart)
2143 u32 __user *uaddr = restart->futex.uaddr;
2144 ktime_t t, *tp = NULL;
2146 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2147 t.tv64 = restart->futex.time;
2150 restart->fn = do_no_restart_syscall;
2152 return (long)futex_wait(uaddr, restart->futex.flags,
2153 restart->futex.val, tp, restart->futex.bitset);
2158 * Userspace tried a 0 -> TID atomic transition of the futex value
2159 * and failed. The kernel side here does the whole locking operation:
2160 * if there are waiters then it will block, it does PI, etc. (Due to
2161 * races the kernel might see a 0 value of the futex too.)
2163 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2164 ktime_t *time, int trylock)
2166 struct hrtimer_sleeper timeout, *to = NULL;
2167 struct futex_hash_bucket *hb;
2168 struct futex_q q = futex_q_init;
2171 if (refill_pi_state_cache())
2176 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2178 hrtimer_init_sleeper(to, current);
2179 hrtimer_set_expires(&to->timer, *time);
2183 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2184 if (unlikely(ret != 0))
2188 hb = queue_lock(&q);
2190 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2191 if (unlikely(ret)) {
2194 /* We got the lock. */
2196 goto out_unlock_put_key;
2201 * Task is exiting and we just wait for the
2205 put_futex_key(&q.key);
2209 goto out_unlock_put_key;
2214 * Only actually queue now that the atomic ops are done:
2218 WARN_ON(!q.pi_state);
2220 * Block on the PI mutex:
2223 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2225 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2226 /* Fixup the trylock return value: */
2227 ret = ret ? 0 : -EWOULDBLOCK;
2230 spin_lock(q.lock_ptr);
2232 * Fixup the pi_state owner and possibly acquire the lock if we
2235 res = fixup_owner(uaddr, &q, !ret);
2237 * If fixup_owner() returned an error, proprogate that. If it acquired
2238 * the lock, clear our -ETIMEDOUT or -EINTR.
2241 ret = (res < 0) ? res : 0;
2244 * If fixup_owner() faulted and was unable to handle the fault, unlock
2245 * it and return the fault to userspace.
2247 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2248 rt_mutex_unlock(&q.pi_state->pi_mutex);
2250 /* Unqueue and drop the lock */
2259 put_futex_key(&q.key);
2262 destroy_hrtimer_on_stack(&to->timer);
2263 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2268 ret = fault_in_user_writeable(uaddr);
2272 if (!(flags & FLAGS_SHARED))
2275 put_futex_key(&q.key);
2280 * Userspace attempted a TID -> 0 atomic transition, and failed.
2281 * This is the in-kernel slowpath: we look up the PI state (if any),
2282 * and do the rt-mutex unlock.
2284 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2286 struct futex_hash_bucket *hb;
2287 struct futex_q *this, *next;
2288 union futex_key key = FUTEX_KEY_INIT;
2289 u32 uval, vpid = task_pid_vnr(current);
2293 if (get_user(uval, uaddr))
2296 * We release only a lock we actually own:
2298 if ((uval & FUTEX_TID_MASK) != vpid)
2301 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2302 if (unlikely(ret != 0))
2305 hb = hash_futex(&key);
2306 spin_lock(&hb->lock);
2309 * To avoid races, try to do the TID -> 0 atomic transition
2310 * again. If it succeeds then we can return without waking
2313 if (!(uval & FUTEX_OWNER_DIED) &&
2314 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2317 * Rare case: we managed to release the lock atomically,
2318 * no need to wake anyone else up:
2320 if (unlikely(uval == vpid))
2324 * Ok, other tasks may need to be woken up - check waiters
2325 * and do the wakeup if necessary:
2327 plist_for_each_entry_safe(this, next, &hb->chain, list) {
2328 if (!match_futex (&this->key, &key))
2330 ret = wake_futex_pi(uaddr, uval, this);
2332 * The atomic access to the futex value
2333 * generated a pagefault, so retry the
2334 * user-access and the wakeup:
2341 * No waiters - kernel unlocks the futex:
2343 if (!(uval & FUTEX_OWNER_DIED)) {
2344 ret = unlock_futex_pi(uaddr, uval);
2350 spin_unlock(&hb->lock);
2351 put_futex_key(&key);
2357 spin_unlock(&hb->lock);
2358 put_futex_key(&key);
2360 ret = fault_in_user_writeable(uaddr);
2368 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2369 * @hb: the hash_bucket futex_q was original enqueued on
2370 * @q: the futex_q woken while waiting to be requeued
2371 * @key2: the futex_key of the requeue target futex
2372 * @timeout: the timeout associated with the wait (NULL if none)
2374 * Detect if the task was woken on the initial futex as opposed to the requeue
2375 * target futex. If so, determine if it was a timeout or a signal that caused
2376 * the wakeup and return the appropriate error code to the caller. Must be
2377 * called with the hb lock held.
2380 * 0 = no early wakeup detected;
2381 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2384 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2385 struct futex_q *q, union futex_key *key2,
2386 struct hrtimer_sleeper *timeout)
2391 * With the hb lock held, we avoid races while we process the wakeup.
2392 * We only need to hold hb (and not hb2) to ensure atomicity as the
2393 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2394 * It can't be requeued from uaddr2 to something else since we don't
2395 * support a PI aware source futex for requeue.
2397 if (!match_futex(&q->key, key2)) {
2398 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2400 * We were woken prior to requeue by a timeout or a signal.
2401 * Unqueue the futex_q and determine which it was.
2403 plist_del(&q->list, &hb->chain);
2406 /* Handle spurious wakeups gracefully */
2408 if (timeout && !timeout->task)
2410 else if (signal_pending(current))
2411 ret = -ERESTARTNOINTR;
2417 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2418 * @uaddr: the futex we initially wait on (non-pi)
2419 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2420 * the same type, no requeueing from private to shared, etc.
2421 * @val: the expected value of uaddr
2422 * @abs_time: absolute timeout
2423 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2424 * @uaddr2: the pi futex we will take prior to returning to user-space
2426 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2427 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2428 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2429 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2430 * without one, the pi logic would not know which task to boost/deboost, if
2431 * there was a need to.
2433 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2434 * via the following--
2435 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2436 * 2) wakeup on uaddr2 after a requeue
2440 * If 3, cleanup and return -ERESTARTNOINTR.
2442 * If 2, we may then block on trying to take the rt_mutex and return via:
2443 * 5) successful lock
2446 * 8) other lock acquisition failure
2448 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2450 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2456 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2457 u32 val, ktime_t *abs_time, u32 bitset,
2460 struct hrtimer_sleeper timeout, *to = NULL;
2461 struct rt_mutex_waiter rt_waiter;
2462 struct rt_mutex *pi_mutex = NULL;
2463 struct futex_hash_bucket *hb;
2464 union futex_key key2 = FUTEX_KEY_INIT;
2465 struct futex_q q = futex_q_init;
2468 if (uaddr == uaddr2)
2476 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2477 CLOCK_REALTIME : CLOCK_MONOTONIC,
2479 hrtimer_init_sleeper(to, current);
2480 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2481 current->timer_slack_ns);
2485 * The waiter is allocated on our stack, manipulated by the requeue
2486 * code while we sleep on uaddr.
2488 debug_rt_mutex_init_waiter(&rt_waiter);
2489 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2490 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2491 rt_waiter.task = NULL;
2493 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2494 if (unlikely(ret != 0))
2498 q.rt_waiter = &rt_waiter;
2499 q.requeue_pi_key = &key2;
2502 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2505 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2509 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2510 futex_wait_queue_me(hb, &q, to);
2512 spin_lock(&hb->lock);
2513 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2514 spin_unlock(&hb->lock);
2519 * In order for us to be here, we know our q.key == key2, and since
2520 * we took the hb->lock above, we also know that futex_requeue() has
2521 * completed and we no longer have to concern ourselves with a wakeup
2522 * race with the atomic proxy lock acquisition by the requeue code. The
2523 * futex_requeue dropped our key1 reference and incremented our key2
2527 /* Check if the requeue code acquired the second futex for us. */
2530 * Got the lock. We might not be the anticipated owner if we
2531 * did a lock-steal - fix up the PI-state in that case.
2533 if (q.pi_state && (q.pi_state->owner != current)) {
2534 spin_lock(q.lock_ptr);
2535 ret = fixup_pi_state_owner(uaddr2, &q, current);
2536 spin_unlock(q.lock_ptr);
2540 * We have been woken up by futex_unlock_pi(), a timeout, or a
2541 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2544 WARN_ON(!q.pi_state);
2545 pi_mutex = &q.pi_state->pi_mutex;
2546 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2547 debug_rt_mutex_free_waiter(&rt_waiter);
2549 spin_lock(q.lock_ptr);
2551 * Fixup the pi_state owner and possibly acquire the lock if we
2554 res = fixup_owner(uaddr2, &q, !ret);
2556 * If fixup_owner() returned an error, proprogate that. If it
2557 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2560 ret = (res < 0) ? res : 0;
2562 /* Unqueue and drop the lock. */
2567 * If fixup_pi_state_owner() faulted and was unable to handle the
2568 * fault, unlock the rt_mutex and return the fault to userspace.
2570 if (ret == -EFAULT) {
2571 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2572 rt_mutex_unlock(pi_mutex);
2573 } else if (ret == -EINTR) {
2575 * We've already been requeued, but cannot restart by calling
2576 * futex_lock_pi() directly. We could restart this syscall, but
2577 * it would detect that the user space "val" changed and return
2578 * -EWOULDBLOCK. Save the overhead of the restart and return
2579 * -EWOULDBLOCK directly.
2585 put_futex_key(&q.key);
2587 put_futex_key(&key2);
2591 hrtimer_cancel(&to->timer);
2592 destroy_hrtimer_on_stack(&to->timer);
2598 * Support for robust futexes: the kernel cleans up held futexes at
2601 * Implementation: user-space maintains a per-thread list of locks it
2602 * is holding. Upon do_exit(), the kernel carefully walks this list,
2603 * and marks all locks that are owned by this thread with the
2604 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2605 * always manipulated with the lock held, so the list is private and
2606 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2607 * field, to allow the kernel to clean up if the thread dies after
2608 * acquiring the lock, but just before it could have added itself to
2609 * the list. There can only be one such pending lock.
2613 * sys_set_robust_list() - Set the robust-futex list head of a task
2614 * @head: pointer to the list-head
2615 * @len: length of the list-head, as userspace expects
2617 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2620 if (!futex_cmpxchg_enabled)
2623 * The kernel knows only one size for now:
2625 if (unlikely(len != sizeof(*head)))
2628 current->robust_list = head;
2634 * sys_get_robust_list() - Get the robust-futex list head of a task
2635 * @pid: pid of the process [zero for current task]
2636 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2637 * @len_ptr: pointer to a length field, the kernel fills in the header size
2639 SYSCALL_DEFINE3(get_robust_list, int, pid,
2640 struct robust_list_head __user * __user *, head_ptr,
2641 size_t __user *, len_ptr)
2643 struct robust_list_head __user *head;
2645 struct task_struct *p;
2647 if (!futex_cmpxchg_enabled)
2656 p = find_task_by_vpid(pid);
2662 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2665 head = p->robust_list;
2668 if (put_user(sizeof(*head), len_ptr))
2670 return put_user(head, head_ptr);
2679 * Process a futex-list entry, check whether it's owned by the
2680 * dying task, and do notification if so:
2682 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2684 u32 uval, uninitialized_var(nval), mval;
2687 if (get_user(uval, uaddr))
2690 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2692 * Ok, this dying thread is truly holding a futex
2693 * of interest. Set the OWNER_DIED bit atomically
2694 * via cmpxchg, and if the value had FUTEX_WAITERS
2695 * set, wake up a waiter (if any). (We have to do a
2696 * futex_wake() even if OWNER_DIED is already set -
2697 * to handle the rare but possible case of recursive
2698 * thread-death.) The rest of the cleanup is done in
2701 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2703 * We are not holding a lock here, but we want to have
2704 * the pagefault_disable/enable() protection because
2705 * we want to handle the fault gracefully. If the
2706 * access fails we try to fault in the futex with R/W
2707 * verification via get_user_pages. get_user() above
2708 * does not guarantee R/W access. If that fails we
2709 * give up and leave the futex locked.
2711 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2712 if (fault_in_user_writeable(uaddr))
2720 * Wake robust non-PI futexes here. The wakeup of
2721 * PI futexes happens in exit_pi_state():
2723 if (!pi && (uval & FUTEX_WAITERS))
2724 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2730 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2732 static inline int fetch_robust_entry(struct robust_list __user **entry,
2733 struct robust_list __user * __user *head,
2736 unsigned long uentry;
2738 if (get_user(uentry, (unsigned long __user *)head))
2741 *entry = (void __user *)(uentry & ~1UL);
2748 * Walk curr->robust_list (very carefully, it's a userspace list!)
2749 * and mark any locks found there dead, and notify any waiters.
2751 * We silently return on any sign of list-walking problem.
2753 void exit_robust_list(struct task_struct *curr)
2755 struct robust_list_head __user *head = curr->robust_list;
2756 struct robust_list __user *entry, *next_entry, *pending;
2757 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2758 unsigned int uninitialized_var(next_pi);
2759 unsigned long futex_offset;
2762 if (!futex_cmpxchg_enabled)
2766 * Fetch the list head (which was registered earlier, via
2767 * sys_set_robust_list()):
2769 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2772 * Fetch the relative futex offset:
2774 if (get_user(futex_offset, &head->futex_offset))
2777 * Fetch any possibly pending lock-add first, and handle it
2780 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2783 next_entry = NULL; /* avoid warning with gcc */
2784 while (entry != &head->list) {
2786 * Fetch the next entry in the list before calling
2787 * handle_futex_death:
2789 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2791 * A pending lock might already be on the list, so
2792 * don't process it twice:
2794 if (entry != pending)
2795 if (handle_futex_death((void __user *)entry + futex_offset,
2803 * Avoid excessively long or circular lists:
2812 handle_futex_death((void __user *)pending + futex_offset,
2816 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2817 u32 __user *uaddr2, u32 val2, u32 val3)
2819 int cmd = op & FUTEX_CMD_MASK;
2820 unsigned int flags = 0;
2822 if (!(op & FUTEX_PRIVATE_FLAG))
2823 flags |= FLAGS_SHARED;
2825 if (op & FUTEX_CLOCK_REALTIME) {
2826 flags |= FLAGS_CLOCKRT;
2827 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2833 case FUTEX_UNLOCK_PI:
2834 case FUTEX_TRYLOCK_PI:
2835 case FUTEX_WAIT_REQUEUE_PI:
2836 case FUTEX_CMP_REQUEUE_PI:
2837 if (!futex_cmpxchg_enabled)
2843 val3 = FUTEX_BITSET_MATCH_ANY;
2844 case FUTEX_WAIT_BITSET:
2845 return futex_wait(uaddr, flags, val, timeout, val3);
2847 val3 = FUTEX_BITSET_MATCH_ANY;
2848 case FUTEX_WAKE_BITSET:
2849 return futex_wake(uaddr, flags, val, val3);
2851 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2852 case FUTEX_CMP_REQUEUE:
2853 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2855 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2857 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2858 case FUTEX_UNLOCK_PI:
2859 return futex_unlock_pi(uaddr, flags);
2860 case FUTEX_TRYLOCK_PI:
2861 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2862 case FUTEX_WAIT_REQUEUE_PI:
2863 val3 = FUTEX_BITSET_MATCH_ANY;
2864 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2866 case FUTEX_CMP_REQUEUE_PI:
2867 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2873 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2874 struct timespec __user *, utime, u32 __user *, uaddr2,
2878 ktime_t t, *tp = NULL;
2880 int cmd = op & FUTEX_CMD_MASK;
2882 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2883 cmd == FUTEX_WAIT_BITSET ||
2884 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2885 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2887 if (!timespec_valid(&ts))
2890 t = timespec_to_ktime(ts);
2891 if (cmd == FUTEX_WAIT)
2892 t = ktime_add_safe(ktime_get(), t);
2896 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2897 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2899 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2900 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2901 val2 = (u32) (unsigned long) utime;
2903 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2906 static void __init futex_detect_cmpxchg(void)
2908 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2912 * This will fail and we want it. Some arch implementations do
2913 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2914 * functionality. We want to know that before we call in any
2915 * of the complex code paths. Also we want to prevent
2916 * registration of robust lists in that case. NULL is
2917 * guaranteed to fault and we get -EFAULT on functional
2918 * implementation, the non-functional ones will return
2921 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2922 futex_cmpxchg_enabled = 1;
2926 static int __init futex_init(void)
2928 unsigned int futex_shift;
2931 #if CONFIG_BASE_SMALL
2932 futex_hashsize = 16;
2934 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
2937 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
2939 futex_hashsize < 256 ? HASH_SMALL : 0,
2941 futex_hashsize, futex_hashsize);
2942 futex_hashsize = 1UL << futex_shift;
2944 futex_detect_cmpxchg();
2946 for (i = 0; i < futex_hashsize; i++) {
2947 atomic_set(&futex_queues[i].waiters, 0);
2948 plist_head_init(&futex_queues[i].chain);
2949 spin_lock_init(&futex_queues[i].lock);
2954 __initcall(futex_init);