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/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
63 #include <asm/futex.h>
65 #include "rtmutex_common.h"
67 int __read_mostly futex_cmpxchg_enabled;
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
72 * Futex flags used to encode options to functions and preserve them across
75 #define FLAGS_SHARED 0x01
76 #define FLAGS_CLOCKRT 0x02
77 #define FLAGS_HAS_TIMEOUT 0x04
80 * Priority Inheritance state:
82 struct futex_pi_state {
84 * list of 'owned' pi_state instances - these have to be
85 * cleaned up in do_exit() if the task exits prematurely:
87 struct list_head list;
92 struct rt_mutex pi_mutex;
94 struct task_struct *owner;
101 * struct futex_q - The hashed futex queue entry, one per waiting task
102 * @list: priority-sorted list of tasks waiting on this futex
103 * @task: the task waiting on the futex
104 * @lock_ptr: the hash bucket lock
105 * @key: the key the futex is hashed on
106 * @pi_state: optional priority inheritance state
107 * @rt_waiter: rt_waiter storage for use with requeue_pi
108 * @requeue_pi_key: the requeue_pi target futex key
109 * @bitset: bitset for the optional bitmasked wakeup
111 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
112 * we can wake only the relevant ones (hashed queues may be shared).
114 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
115 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
116 * The order of wakeup is always to make the first condition true, then
119 * PI futexes are typically woken before they are removed from the hash list via
120 * the rt_mutex code. See unqueue_me_pi().
123 struct plist_node list;
125 struct task_struct *task;
126 spinlock_t *lock_ptr;
128 struct futex_pi_state *pi_state;
129 struct rt_mutex_waiter *rt_waiter;
130 union futex_key *requeue_pi_key;
134 static const struct futex_q futex_q_init = {
135 /* list gets initialized in queue_me()*/
136 .key = FUTEX_KEY_INIT,
137 .bitset = FUTEX_BITSET_MATCH_ANY
141 * Hash buckets are shared by all the futex_keys that hash to the same
142 * location. Each key may have multiple futex_q structures, one for each task
143 * waiting on a futex.
145 struct futex_hash_bucket {
147 struct plist_head chain;
150 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
153 * We hash on the keys returned from get_futex_key (see below).
155 static struct futex_hash_bucket *hash_futex(union futex_key *key)
157 u32 hash = jhash2((u32*)&key->both.word,
158 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
160 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
164 * Return 1 if two futex_keys are equal, 0 otherwise.
166 static inline int match_futex(union futex_key *key1, union futex_key *key2)
169 && key1->both.word == key2->both.word
170 && key1->both.ptr == key2->both.ptr
171 && key1->both.offset == key2->both.offset);
175 * Take a reference to the resource addressed by a key.
176 * Can be called while holding spinlocks.
179 static void get_futex_key_refs(union futex_key *key)
184 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
186 ihold(key->shared.inode);
188 case FUT_OFF_MMSHARED:
189 atomic_inc(&key->private.mm->mm_count);
195 * Drop a reference to the resource addressed by a key.
196 * The hash bucket spinlock must not be held.
198 static void drop_futex_key_refs(union futex_key *key)
200 if (!key->both.ptr) {
201 /* If we're here then we tried to put a key we failed to get */
206 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
208 iput(key->shared.inode);
210 case FUT_OFF_MMSHARED:
211 mmdrop(key->private.mm);
217 * get_futex_key() - Get parameters which are the keys for a futex
218 * @uaddr: virtual address of the futex
219 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
220 * @key: address where result is stored.
222 * Returns a negative error code or 0
223 * The key words are stored in *key on success.
225 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
226 * offset_within_page). For private mappings, it's (uaddr, current->mm).
227 * We can usually work out the index without swapping in the page.
229 * lock_page() might sleep, the caller should not hold a spinlock.
232 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
234 unsigned long address = (unsigned long)uaddr;
235 struct mm_struct *mm = current->mm;
236 struct page *page, *page_head;
240 * The futex address must be "naturally" aligned.
242 key->both.offset = address % PAGE_SIZE;
243 if (unlikely((address % sizeof(u32)) != 0))
245 address -= key->both.offset;
248 * PROCESS_PRIVATE futexes are fast.
249 * As the mm cannot disappear under us and the 'key' only needs
250 * virtual address, we dont even have to find the underlying vma.
251 * Note : We do have to check 'uaddr' is a valid user address,
252 * but access_ok() should be faster than find_vma()
255 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
257 key->private.mm = mm;
258 key->private.address = address;
259 get_futex_key_refs(key);
264 err = get_user_pages_fast(address, 1, 1, &page);
268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
270 if (unlikely(PageTail(page))) {
272 /* serialize against __split_huge_page_splitting() */
274 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
275 page_head = compound_head(page);
277 * page_head is valid pointer but we must pin
278 * it before taking the PG_lock and/or
279 * PG_compound_lock. The moment we re-enable
280 * irqs __split_huge_page_splitting() can
281 * return and the head page can be freed from
282 * under us. We can't take the PG_lock and/or
283 * PG_compound_lock on a page that could be
284 * freed from under us.
286 if (page != page_head) {
297 page_head = compound_head(page);
298 if (page != page_head) {
304 lock_page(page_head);
305 if (!page_head->mapping) {
306 unlock_page(page_head);
312 * Private mappings are handled in a simple way.
314 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
315 * it's a read-only handle, it's expected that futexes attach to
316 * the object not the particular process.
318 if (PageAnon(page_head)) {
319 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
320 key->private.mm = mm;
321 key->private.address = address;
323 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
324 key->shared.inode = page_head->mapping->host;
325 key->shared.pgoff = page_head->index;
328 get_futex_key_refs(key);
330 unlock_page(page_head);
335 static inline void put_futex_key(union futex_key *key)
337 drop_futex_key_refs(key);
341 * fault_in_user_writeable() - Fault in user address and verify RW access
342 * @uaddr: pointer to faulting user space address
344 * Slow path to fixup the fault we just took in the atomic write
347 * We have no generic implementation of a non-destructive write to the
348 * user address. We know that we faulted in the atomic pagefault
349 * disabled section so we can as well avoid the #PF overhead by
350 * calling get_user_pages() right away.
352 static int fault_in_user_writeable(u32 __user *uaddr)
354 struct mm_struct *mm = current->mm;
357 down_read(&mm->mmap_sem);
358 ret = get_user_pages(current, mm, (unsigned long)uaddr,
359 1, 1, 0, NULL, NULL);
360 up_read(&mm->mmap_sem);
362 return ret < 0 ? ret : 0;
366 * futex_top_waiter() - Return the highest priority waiter on a futex
367 * @hb: the hash bucket the futex_q's reside in
368 * @key: the futex key (to distinguish it from other futex futex_q's)
370 * Must be called with the hb lock held.
372 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
373 union futex_key *key)
375 struct futex_q *this;
377 plist_for_each_entry(this, &hb->chain, list) {
378 if (match_futex(&this->key, key))
384 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
385 u32 uval, u32 newval)
390 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
396 static int get_futex_value_locked(u32 *dest, u32 __user *from)
401 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
404 return ret ? -EFAULT : 0;
411 static int refill_pi_state_cache(void)
413 struct futex_pi_state *pi_state;
415 if (likely(current->pi_state_cache))
418 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
423 INIT_LIST_HEAD(&pi_state->list);
424 /* pi_mutex gets initialized later */
425 pi_state->owner = NULL;
426 atomic_set(&pi_state->refcount, 1);
427 pi_state->key = FUTEX_KEY_INIT;
429 current->pi_state_cache = pi_state;
434 static struct futex_pi_state * alloc_pi_state(void)
436 struct futex_pi_state *pi_state = current->pi_state_cache;
439 current->pi_state_cache = NULL;
444 static void free_pi_state(struct futex_pi_state *pi_state)
446 if (!atomic_dec_and_test(&pi_state->refcount))
450 * If pi_state->owner is NULL, the owner is most probably dying
451 * and has cleaned up the pi_state already
453 if (pi_state->owner) {
454 raw_spin_lock_irq(&pi_state->owner->pi_lock);
455 list_del_init(&pi_state->list);
456 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
458 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
461 if (current->pi_state_cache)
465 * pi_state->list is already empty.
466 * clear pi_state->owner.
467 * refcount is at 0 - put it back to 1.
469 pi_state->owner = NULL;
470 atomic_set(&pi_state->refcount, 1);
471 current->pi_state_cache = pi_state;
476 * Look up the task based on what TID userspace gave us.
479 static struct task_struct * futex_find_get_task(pid_t pid)
481 struct task_struct *p;
484 p = find_task_by_vpid(pid);
494 * This task is holding PI mutexes at exit time => bad.
495 * Kernel cleans up PI-state, but userspace is likely hosed.
496 * (Robust-futex cleanup is separate and might save the day for userspace.)
498 void exit_pi_state_list(struct task_struct *curr)
500 struct list_head *next, *head = &curr->pi_state_list;
501 struct futex_pi_state *pi_state;
502 struct futex_hash_bucket *hb;
503 union futex_key key = FUTEX_KEY_INIT;
505 if (!futex_cmpxchg_enabled)
508 * We are a ZOMBIE and nobody can enqueue itself on
509 * pi_state_list anymore, but we have to be careful
510 * versus waiters unqueueing themselves:
512 raw_spin_lock_irq(&curr->pi_lock);
513 while (!list_empty(head)) {
516 pi_state = list_entry(next, struct futex_pi_state, list);
518 hb = hash_futex(&key);
519 raw_spin_unlock_irq(&curr->pi_lock);
521 spin_lock(&hb->lock);
523 raw_spin_lock_irq(&curr->pi_lock);
525 * We dropped the pi-lock, so re-check whether this
526 * task still owns the PI-state:
528 if (head->next != next) {
529 spin_unlock(&hb->lock);
533 WARN_ON(pi_state->owner != curr);
534 WARN_ON(list_empty(&pi_state->list));
535 list_del_init(&pi_state->list);
536 pi_state->owner = NULL;
537 raw_spin_unlock_irq(&curr->pi_lock);
539 rt_mutex_unlock(&pi_state->pi_mutex);
541 spin_unlock(&hb->lock);
543 raw_spin_lock_irq(&curr->pi_lock);
545 raw_spin_unlock_irq(&curr->pi_lock);
549 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
550 union futex_key *key, struct futex_pi_state **ps)
552 struct futex_pi_state *pi_state = NULL;
553 struct futex_q *this, *next;
554 struct plist_head *head;
555 struct task_struct *p;
556 pid_t pid = uval & FUTEX_TID_MASK;
560 plist_for_each_entry_safe(this, next, head, list) {
561 if (match_futex(&this->key, key)) {
563 * Another waiter already exists - bump up
564 * the refcount and return its pi_state:
566 pi_state = this->pi_state;
568 * Userspace might have messed up non-PI and PI futexes
570 if (unlikely(!pi_state))
573 WARN_ON(!atomic_read(&pi_state->refcount));
576 * When pi_state->owner is NULL then the owner died
577 * and another waiter is on the fly. pi_state->owner
578 * is fixed up by the task which acquires
579 * pi_state->rt_mutex.
581 * We do not check for pid == 0 which can happen when
582 * the owner died and robust_list_exit() cleared the
585 if (pid && pi_state->owner) {
587 * Bail out if user space manipulated the
590 if (pid != task_pid_vnr(pi_state->owner))
594 atomic_inc(&pi_state->refcount);
602 * We are the first waiter - try to look up the real owner and attach
603 * the new pi_state to it, but bail out when TID = 0
607 p = futex_find_get_task(pid);
612 * We need to look at the task state flags to figure out,
613 * whether the task is exiting. To protect against the do_exit
614 * change of the task flags, we do this protected by
617 raw_spin_lock_irq(&p->pi_lock);
618 if (unlikely(p->flags & PF_EXITING)) {
620 * The task is on the way out. When PF_EXITPIDONE is
621 * set, we know that the task has finished the
624 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
626 raw_spin_unlock_irq(&p->pi_lock);
631 pi_state = alloc_pi_state();
634 * Initialize the pi_mutex in locked state and make 'p'
637 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
639 /* Store the key for possible exit cleanups: */
640 pi_state->key = *key;
642 WARN_ON(!list_empty(&pi_state->list));
643 list_add(&pi_state->list, &p->pi_state_list);
645 raw_spin_unlock_irq(&p->pi_lock);
655 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
656 * @uaddr: the pi futex user address
657 * @hb: the pi futex hash bucket
658 * @key: the futex key associated with uaddr and hb
659 * @ps: the pi_state pointer where we store the result of the
661 * @task: the task to perform the atomic lock work for. This will
662 * be "current" except in the case of requeue pi.
663 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
667 * 1 - acquired the lock
670 * The hb->lock and futex_key refs shall be held by the caller.
672 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
673 union futex_key *key,
674 struct futex_pi_state **ps,
675 struct task_struct *task, int set_waiters)
677 int lock_taken, ret, ownerdied = 0;
678 u32 uval, newval, curval, vpid = task_pid_vnr(task);
681 ret = lock_taken = 0;
684 * To avoid races, we attempt to take the lock here again
685 * (by doing a 0 -> TID atomic cmpxchg), while holding all
686 * the locks. It will most likely not succeed.
690 newval |= FUTEX_WAITERS;
692 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
698 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
702 * Surprise - we got the lock. Just return to userspace:
704 if (unlikely(!curval))
710 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
711 * to wake at the next unlock.
713 newval = curval | FUTEX_WAITERS;
716 * There are two cases, where a futex might have no owner (the
717 * owner TID is 0): OWNER_DIED. We take over the futex in this
718 * case. We also do an unconditional take over, when the owner
721 * This is safe as we are protected by the hash bucket lock !
723 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
724 /* Keep the OWNER_DIED bit */
725 newval = (curval & ~FUTEX_TID_MASK) | vpid;
730 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
732 if (unlikely(curval != uval))
736 * We took the lock due to owner died take over.
738 if (unlikely(lock_taken))
742 * We dont have the lock. Look up the PI state (or create it if
743 * we are the first waiter):
745 ret = lookup_pi_state(uval, hb, key, ps);
751 * No owner found for this futex. Check if the
752 * OWNER_DIED bit is set to figure out whether
753 * this is a robust futex or not.
755 if (get_futex_value_locked(&curval, uaddr))
759 * We simply start over in case of a robust
760 * futex. The code above will take the futex
763 if (curval & FUTEX_OWNER_DIED) {
776 * The hash bucket lock must be held when this is called.
777 * Afterwards, the futex_q must not be accessed.
779 static void wake_futex(struct futex_q *q)
781 struct task_struct *p = q->task;
784 * We set q->lock_ptr = NULL _before_ we wake up the task. If
785 * a non-futex wake up happens on another CPU then the task
786 * might exit and p would dereference a non-existing task
787 * struct. Prevent this by holding a reference on p across the
792 plist_del(&q->list, &q->list.plist);
794 * The waiting task can free the futex_q as soon as
795 * q->lock_ptr = NULL is written, without taking any locks. A
796 * memory barrier is required here to prevent the following
797 * store to lock_ptr from getting ahead of the plist_del.
802 wake_up_state(p, TASK_NORMAL);
806 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
808 struct task_struct *new_owner;
809 struct futex_pi_state *pi_state = this->pi_state;
816 * If current does not own the pi_state then the futex is
817 * inconsistent and user space fiddled with the futex value.
819 if (pi_state->owner != current)
822 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
823 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
826 * It is possible that the next waiter (the one that brought
827 * this owner to the kernel) timed out and is no longer
828 * waiting on the lock.
831 new_owner = this->task;
834 * We pass it to the next owner. (The WAITERS bit is always
835 * kept enabled while there is PI state around. We must also
836 * preserve the owner died bit.)
838 if (!(uval & FUTEX_OWNER_DIED)) {
841 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
843 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
845 else if (curval != uval)
848 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
853 raw_spin_lock_irq(&pi_state->owner->pi_lock);
854 WARN_ON(list_empty(&pi_state->list));
855 list_del_init(&pi_state->list);
856 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
858 raw_spin_lock_irq(&new_owner->pi_lock);
859 WARN_ON(!list_empty(&pi_state->list));
860 list_add(&pi_state->list, &new_owner->pi_state_list);
861 pi_state->owner = new_owner;
862 raw_spin_unlock_irq(&new_owner->pi_lock);
864 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
865 rt_mutex_unlock(&pi_state->pi_mutex);
870 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
875 * There is no waiter, so we unlock the futex. The owner died
876 * bit has not to be preserved here. We are the owner:
878 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
887 * Express the locking dependencies for lockdep:
890 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
893 spin_lock(&hb1->lock);
895 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
896 } else { /* hb1 > hb2 */
897 spin_lock(&hb2->lock);
898 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
903 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
905 spin_unlock(&hb1->lock);
907 spin_unlock(&hb2->lock);
911 * Wake up waiters matching bitset queued on this futex (uaddr).
914 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
916 struct futex_hash_bucket *hb;
917 struct futex_q *this, *next;
918 struct plist_head *head;
919 union futex_key key = FUTEX_KEY_INIT;
925 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
926 if (unlikely(ret != 0))
929 hb = hash_futex(&key);
930 spin_lock(&hb->lock);
933 plist_for_each_entry_safe(this, next, head, list) {
934 if (match_futex (&this->key, &key)) {
935 if (this->pi_state || this->rt_waiter) {
940 /* Check if one of the bits is set in both bitsets */
941 if (!(this->bitset & bitset))
945 if (++ret >= nr_wake)
950 spin_unlock(&hb->lock);
957 * Wake up all waiters hashed on the physical page that is mapped
958 * to this virtual address:
961 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
962 int nr_wake, int nr_wake2, int op)
964 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
965 struct futex_hash_bucket *hb1, *hb2;
966 struct plist_head *head;
967 struct futex_q *this, *next;
971 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
972 if (unlikely(ret != 0))
974 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
975 if (unlikely(ret != 0))
978 hb1 = hash_futex(&key1);
979 hb2 = hash_futex(&key2);
982 double_lock_hb(hb1, hb2);
983 op_ret = futex_atomic_op_inuser(op, uaddr2);
984 if (unlikely(op_ret < 0)) {
986 double_unlock_hb(hb1, hb2);
990 * we don't get EFAULT from MMU faults if we don't have an MMU,
991 * but we might get them from range checking
997 if (unlikely(op_ret != -EFAULT)) {
1002 ret = fault_in_user_writeable(uaddr2);
1006 if (!(flags & FLAGS_SHARED))
1009 put_futex_key(&key2);
1010 put_futex_key(&key1);
1016 plist_for_each_entry_safe(this, next, head, list) {
1017 if (match_futex (&this->key, &key1)) {
1019 if (++ret >= nr_wake)
1028 plist_for_each_entry_safe(this, next, head, list) {
1029 if (match_futex (&this->key, &key2)) {
1031 if (++op_ret >= nr_wake2)
1038 double_unlock_hb(hb1, hb2);
1040 put_futex_key(&key2);
1042 put_futex_key(&key1);
1048 * requeue_futex() - Requeue a futex_q from one hb to another
1049 * @q: the futex_q to requeue
1050 * @hb1: the source hash_bucket
1051 * @hb2: the target hash_bucket
1052 * @key2: the new key for the requeued futex_q
1055 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1056 struct futex_hash_bucket *hb2, union futex_key *key2)
1060 * If key1 and key2 hash to the same bucket, no need to
1063 if (likely(&hb1->chain != &hb2->chain)) {
1064 plist_del(&q->list, &hb1->chain);
1065 plist_add(&q->list, &hb2->chain);
1066 q->lock_ptr = &hb2->lock;
1067 #ifdef CONFIG_DEBUG_PI_LIST
1068 q->list.plist.spinlock = &hb2->lock;
1071 get_futex_key_refs(key2);
1076 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1078 * @key: the key of the requeue target futex
1079 * @hb: the hash_bucket of the requeue target futex
1081 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1082 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1083 * to the requeue target futex so the waiter can detect the wakeup on the right
1084 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1085 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1086 * to protect access to the pi_state to fixup the owner later. Must be called
1087 * with both q->lock_ptr and hb->lock held.
1090 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1091 struct futex_hash_bucket *hb)
1093 get_futex_key_refs(key);
1096 WARN_ON(plist_node_empty(&q->list));
1097 plist_del(&q->list, &q->list.plist);
1099 WARN_ON(!q->rt_waiter);
1100 q->rt_waiter = NULL;
1102 q->lock_ptr = &hb->lock;
1103 #ifdef CONFIG_DEBUG_PI_LIST
1104 q->list.plist.spinlock = &hb->lock;
1107 wake_up_state(q->task, TASK_NORMAL);
1111 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1112 * @pifutex: the user address of the to futex
1113 * @hb1: the from futex hash bucket, must be locked by the caller
1114 * @hb2: the to futex hash bucket, must be locked by the caller
1115 * @key1: the from futex key
1116 * @key2: the to futex key
1117 * @ps: address to store the pi_state pointer
1118 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1120 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1121 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1122 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1123 * hb1 and hb2 must be held by the caller.
1126 * 0 - failed to acquire the lock atomicly
1127 * 1 - acquired the lock
1130 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1131 struct futex_hash_bucket *hb1,
1132 struct futex_hash_bucket *hb2,
1133 union futex_key *key1, union futex_key *key2,
1134 struct futex_pi_state **ps, int set_waiters)
1136 struct futex_q *top_waiter = NULL;
1140 if (get_futex_value_locked(&curval, pifutex))
1144 * Find the top_waiter and determine if there are additional waiters.
1145 * If the caller intends to requeue more than 1 waiter to pifutex,
1146 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1147 * as we have means to handle the possible fault. If not, don't set
1148 * the bit unecessarily as it will force the subsequent unlock to enter
1151 top_waiter = futex_top_waiter(hb1, key1);
1153 /* There are no waiters, nothing for us to do. */
1157 /* Ensure we requeue to the expected futex. */
1158 if (!match_futex(top_waiter->requeue_pi_key, key2))
1162 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1163 * the contended case or if set_waiters is 1. The pi_state is returned
1164 * in ps in contended cases.
1166 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1169 requeue_pi_wake_futex(top_waiter, key2, hb2);
1175 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1176 * @uaddr1: source futex user address
1177 * @flags: futex flags (FLAGS_SHARED, etc.)
1178 * @uaddr2: target futex user address
1179 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1180 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1181 * @cmpval: @uaddr1 expected value (or %NULL)
1182 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1183 * pi futex (pi to pi requeue is not supported)
1185 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1186 * uaddr2 atomically on behalf of the top waiter.
1189 * >=0 - on success, the number of tasks requeued or woken
1192 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1193 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1194 u32 *cmpval, int requeue_pi)
1196 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1197 int drop_count = 0, task_count = 0, ret;
1198 struct futex_pi_state *pi_state = NULL;
1199 struct futex_hash_bucket *hb1, *hb2;
1200 struct plist_head *head1;
1201 struct futex_q *this, *next;
1206 * requeue_pi requires a pi_state, try to allocate it now
1207 * without any locks in case it fails.
1209 if (refill_pi_state_cache())
1212 * requeue_pi must wake as many tasks as it can, up to nr_wake
1213 * + nr_requeue, since it acquires the rt_mutex prior to
1214 * returning to userspace, so as to not leave the rt_mutex with
1215 * waiters and no owner. However, second and third wake-ups
1216 * cannot be predicted as they involve race conditions with the
1217 * first wake and a fault while looking up the pi_state. Both
1218 * pthread_cond_signal() and pthread_cond_broadcast() should
1226 if (pi_state != NULL) {
1228 * We will have to lookup the pi_state again, so free this one
1229 * to keep the accounting correct.
1231 free_pi_state(pi_state);
1235 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
1236 if (unlikely(ret != 0))
1238 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
1239 if (unlikely(ret != 0))
1242 hb1 = hash_futex(&key1);
1243 hb2 = hash_futex(&key2);
1246 double_lock_hb(hb1, hb2);
1248 if (likely(cmpval != NULL)) {
1251 ret = get_futex_value_locked(&curval, uaddr1);
1253 if (unlikely(ret)) {
1254 double_unlock_hb(hb1, hb2);
1256 ret = get_user(curval, uaddr1);
1260 if (!(flags & FLAGS_SHARED))
1263 put_futex_key(&key2);
1264 put_futex_key(&key1);
1267 if (curval != *cmpval) {
1273 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1275 * Attempt to acquire uaddr2 and wake the top waiter. If we
1276 * intend to requeue waiters, force setting the FUTEX_WAITERS
1277 * bit. We force this here where we are able to easily handle
1278 * faults rather in the requeue loop below.
1280 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1281 &key2, &pi_state, nr_requeue);
1284 * At this point the top_waiter has either taken uaddr2 or is
1285 * waiting on it. If the former, then the pi_state will not
1286 * exist yet, look it up one more time to ensure we have a
1293 ret = get_futex_value_locked(&curval2, uaddr2);
1295 ret = lookup_pi_state(curval2, hb2, &key2,
1303 double_unlock_hb(hb1, hb2);
1304 put_futex_key(&key2);
1305 put_futex_key(&key1);
1306 ret = fault_in_user_writeable(uaddr2);
1311 /* The owner was exiting, try again. */
1312 double_unlock_hb(hb1, hb2);
1313 put_futex_key(&key2);
1314 put_futex_key(&key1);
1322 head1 = &hb1->chain;
1323 plist_for_each_entry_safe(this, next, head1, list) {
1324 if (task_count - nr_wake >= nr_requeue)
1327 if (!match_futex(&this->key, &key1))
1331 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1332 * be paired with each other and no other futex ops.
1334 if ((requeue_pi && !this->rt_waiter) ||
1335 (!requeue_pi && this->rt_waiter)) {
1341 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1342 * lock, we already woke the top_waiter. If not, it will be
1343 * woken by futex_unlock_pi().
1345 if (++task_count <= nr_wake && !requeue_pi) {
1350 /* Ensure we requeue to the expected futex for requeue_pi. */
1351 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1357 * Requeue nr_requeue waiters and possibly one more in the case
1358 * of requeue_pi if we couldn't acquire the lock atomically.
1361 /* Prepare the waiter to take the rt_mutex. */
1362 atomic_inc(&pi_state->refcount);
1363 this->pi_state = pi_state;
1364 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1368 /* We got the lock. */
1369 requeue_pi_wake_futex(this, &key2, hb2);
1374 this->pi_state = NULL;
1375 free_pi_state(pi_state);
1379 requeue_futex(this, hb1, hb2, &key2);
1384 double_unlock_hb(hb1, hb2);
1387 * drop_futex_key_refs() must be called outside the spinlocks. During
1388 * the requeue we moved futex_q's from the hash bucket at key1 to the
1389 * one at key2 and updated their key pointer. We no longer need to
1390 * hold the references to key1.
1392 while (--drop_count >= 0)
1393 drop_futex_key_refs(&key1);
1396 put_futex_key(&key2);
1398 put_futex_key(&key1);
1400 if (pi_state != NULL)
1401 free_pi_state(pi_state);
1402 return ret ? ret : task_count;
1405 /* The key must be already stored in q->key. */
1406 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1407 __acquires(&hb->lock)
1409 struct futex_hash_bucket *hb;
1411 hb = hash_futex(&q->key);
1412 q->lock_ptr = &hb->lock;
1414 spin_lock(&hb->lock);
1419 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1420 __releases(&hb->lock)
1422 spin_unlock(&hb->lock);
1426 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1427 * @q: The futex_q to enqueue
1428 * @hb: The destination hash bucket
1430 * The hb->lock must be held by the caller, and is released here. A call to
1431 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1432 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1433 * or nothing if the unqueue is done as part of the wake process and the unqueue
1434 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1437 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1438 __releases(&hb->lock)
1443 * The priority used to register this element is
1444 * - either the real thread-priority for the real-time threads
1445 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1446 * - or MAX_RT_PRIO for non-RT threads.
1447 * Thus, all RT-threads are woken first in priority order, and
1448 * the others are woken last, in FIFO order.
1450 prio = min(current->normal_prio, MAX_RT_PRIO);
1452 plist_node_init(&q->list, prio);
1453 #ifdef CONFIG_DEBUG_PI_LIST
1454 q->list.plist.spinlock = &hb->lock;
1456 plist_add(&q->list, &hb->chain);
1458 spin_unlock(&hb->lock);
1462 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1463 * @q: The futex_q to unqueue
1465 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1466 * be paired with exactly one earlier call to queue_me().
1469 * 1 - if the futex_q was still queued (and we removed unqueued it)
1470 * 0 - if the futex_q was already removed by the waking thread
1472 static int unqueue_me(struct futex_q *q)
1474 spinlock_t *lock_ptr;
1477 /* In the common case we don't take the spinlock, which is nice. */
1479 lock_ptr = q->lock_ptr;
1481 if (lock_ptr != NULL) {
1482 spin_lock(lock_ptr);
1484 * q->lock_ptr can change between reading it and
1485 * spin_lock(), causing us to take the wrong lock. This
1486 * corrects the race condition.
1488 * Reasoning goes like this: if we have the wrong lock,
1489 * q->lock_ptr must have changed (maybe several times)
1490 * between reading it and the spin_lock(). It can
1491 * change again after the spin_lock() but only if it was
1492 * already changed before the spin_lock(). It cannot,
1493 * however, change back to the original value. Therefore
1494 * we can detect whether we acquired the correct lock.
1496 if (unlikely(lock_ptr != q->lock_ptr)) {
1497 spin_unlock(lock_ptr);
1500 WARN_ON(plist_node_empty(&q->list));
1501 plist_del(&q->list, &q->list.plist);
1503 BUG_ON(q->pi_state);
1505 spin_unlock(lock_ptr);
1509 drop_futex_key_refs(&q->key);
1514 * PI futexes can not be requeued and must remove themself from the
1515 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1518 static void unqueue_me_pi(struct futex_q *q)
1519 __releases(q->lock_ptr)
1521 WARN_ON(plist_node_empty(&q->list));
1522 plist_del(&q->list, &q->list.plist);
1524 BUG_ON(!q->pi_state);
1525 free_pi_state(q->pi_state);
1528 spin_unlock(q->lock_ptr);
1532 * Fixup the pi_state owner with the new owner.
1534 * Must be called with hash bucket lock held and mm->sem held for non
1537 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1538 struct task_struct *newowner)
1540 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1541 struct futex_pi_state *pi_state = q->pi_state;
1542 struct task_struct *oldowner = pi_state->owner;
1543 u32 uval, curval, newval;
1547 if (!pi_state->owner)
1548 newtid |= FUTEX_OWNER_DIED;
1551 * We are here either because we stole the rtmutex from the
1552 * pending owner or we are the pending owner which failed to
1553 * get the rtmutex. We have to replace the pending owner TID
1554 * in the user space variable. This must be atomic as we have
1555 * to preserve the owner died bit here.
1557 * Note: We write the user space value _before_ changing the pi_state
1558 * because we can fault here. Imagine swapped out pages or a fork
1559 * that marked all the anonymous memory readonly for cow.
1561 * Modifying pi_state _before_ the user space value would
1562 * leave the pi_state in an inconsistent state when we fault
1563 * here, because we need to drop the hash bucket lock to
1564 * handle the fault. This might be observed in the PID check
1565 * in lookup_pi_state.
1568 if (get_futex_value_locked(&uval, uaddr))
1572 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1574 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1582 * We fixed up user space. Now we need to fix the pi_state
1585 if (pi_state->owner != NULL) {
1586 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1587 WARN_ON(list_empty(&pi_state->list));
1588 list_del_init(&pi_state->list);
1589 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1592 pi_state->owner = newowner;
1594 raw_spin_lock_irq(&newowner->pi_lock);
1595 WARN_ON(!list_empty(&pi_state->list));
1596 list_add(&pi_state->list, &newowner->pi_state_list);
1597 raw_spin_unlock_irq(&newowner->pi_lock);
1601 * To handle the page fault we need to drop the hash bucket
1602 * lock here. That gives the other task (either the pending
1603 * owner itself or the task which stole the rtmutex) the
1604 * chance to try the fixup of the pi_state. So once we are
1605 * back from handling the fault we need to check the pi_state
1606 * after reacquiring the hash bucket lock and before trying to
1607 * do another fixup. When the fixup has been done already we
1611 spin_unlock(q->lock_ptr);
1613 ret = fault_in_user_writeable(uaddr);
1615 spin_lock(q->lock_ptr);
1618 * Check if someone else fixed it for us:
1620 if (pi_state->owner != oldowner)
1629 static long futex_wait_restart(struct restart_block *restart);
1632 * fixup_owner() - Post lock pi_state and corner case management
1633 * @uaddr: user address of the futex
1634 * @q: futex_q (contains pi_state and access to the rt_mutex)
1635 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1637 * After attempting to lock an rt_mutex, this function is called to cleanup
1638 * the pi_state owner as well as handle race conditions that may allow us to
1639 * acquire the lock. Must be called with the hb lock held.
1642 * 1 - success, lock taken
1643 * 0 - success, lock not taken
1644 * <0 - on error (-EFAULT)
1646 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1648 struct task_struct *owner;
1653 * Got the lock. We might not be the anticipated owner if we
1654 * did a lock-steal - fix up the PI-state in that case:
1656 if (q->pi_state->owner != current)
1657 ret = fixup_pi_state_owner(uaddr, q, current);
1662 * Catch the rare case, where the lock was released when we were on the
1663 * way back before we locked the hash bucket.
1665 if (q->pi_state->owner == current) {
1667 * Try to get the rt_mutex now. This might fail as some other
1668 * task acquired the rt_mutex after we removed ourself from the
1669 * rt_mutex waiters list.
1671 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1677 * pi_state is incorrect, some other task did a lock steal and
1678 * we returned due to timeout or signal without taking the
1679 * rt_mutex. Too late. We can access the rt_mutex_owner without
1680 * locking, as the other task is now blocked on the hash bucket
1681 * lock. Fix the state up.
1683 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1684 ret = fixup_pi_state_owner(uaddr, q, owner);
1689 * Paranoia check. If we did not take the lock, then we should not be
1690 * the owner, nor the pending owner, of the rt_mutex.
1692 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1693 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1694 "pi-state %p\n", ret,
1695 q->pi_state->pi_mutex.owner,
1696 q->pi_state->owner);
1699 return ret ? ret : locked;
1703 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1704 * @hb: the futex hash bucket, must be locked by the caller
1705 * @q: the futex_q to queue up on
1706 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1708 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1709 struct hrtimer_sleeper *timeout)
1712 * The task state is guaranteed to be set before another task can
1713 * wake it. set_current_state() is implemented using set_mb() and
1714 * queue_me() calls spin_unlock() upon completion, both serializing
1715 * access to the hash list and forcing another memory barrier.
1717 set_current_state(TASK_INTERRUPTIBLE);
1722 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1723 if (!hrtimer_active(&timeout->timer))
1724 timeout->task = NULL;
1728 * If we have been removed from the hash list, then another task
1729 * has tried to wake us, and we can skip the call to schedule().
1731 if (likely(!plist_node_empty(&q->list))) {
1733 * If the timer has already expired, current will already be
1734 * flagged for rescheduling. Only call schedule if there
1735 * is no timeout, or if it has yet to expire.
1737 if (!timeout || timeout->task)
1740 __set_current_state(TASK_RUNNING);
1744 * futex_wait_setup() - Prepare to wait on a futex
1745 * @uaddr: the futex userspace address
1746 * @val: the expected value
1747 * @flags: futex flags (FLAGS_SHARED, etc.)
1748 * @q: the associated futex_q
1749 * @hb: storage for hash_bucket pointer to be returned to caller
1751 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1752 * compare it with the expected value. Handle atomic faults internally.
1753 * Return with the hb lock held and a q.key reference on success, and unlocked
1754 * with no q.key reference on failure.
1757 * 0 - uaddr contains val and hb has been locked
1758 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1760 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1761 struct futex_q *q, struct futex_hash_bucket **hb)
1767 * Access the page AFTER the hash-bucket is locked.
1768 * Order is important:
1770 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1771 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1773 * The basic logical guarantee of a futex is that it blocks ONLY
1774 * if cond(var) is known to be true at the time of blocking, for
1775 * any cond. If we locked the hash-bucket after testing *uaddr, that
1776 * would open a race condition where we could block indefinitely with
1777 * cond(var) false, which would violate the guarantee.
1779 * On the other hand, we insert q and release the hash-bucket only
1780 * after testing *uaddr. This guarantees that futex_wait() will NOT
1781 * absorb a wakeup if *uaddr does not match the desired values
1782 * while the syscall executes.
1785 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key);
1786 if (unlikely(ret != 0))
1790 *hb = queue_lock(q);
1792 ret = get_futex_value_locked(&uval, uaddr);
1795 queue_unlock(q, *hb);
1797 ret = get_user(uval, uaddr);
1801 if (!(flags & FLAGS_SHARED))
1804 put_futex_key(&q->key);
1809 queue_unlock(q, *hb);
1815 put_futex_key(&q->key);
1819 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1820 ktime_t *abs_time, u32 bitset)
1822 struct hrtimer_sleeper timeout, *to = NULL;
1823 struct restart_block *restart;
1824 struct futex_hash_bucket *hb;
1825 struct futex_q q = futex_q_init;
1835 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1836 CLOCK_REALTIME : CLOCK_MONOTONIC,
1838 hrtimer_init_sleeper(to, current);
1839 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1840 current->timer_slack_ns);
1845 * Prepare to wait on uaddr. On success, holds hb lock and increments
1848 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1852 /* queue_me and wait for wakeup, timeout, or a signal. */
1853 futex_wait_queue_me(hb, &q, to);
1855 /* If we were woken (and unqueued), we succeeded, whatever. */
1857 /* unqueue_me() drops q.key ref */
1858 if (!unqueue_me(&q))
1861 if (to && !to->task)
1865 * We expect signal_pending(current), but we might be the
1866 * victim of a spurious wakeup as well.
1868 if (!signal_pending(current))
1875 restart = ¤t_thread_info()->restart_block;
1876 restart->fn = futex_wait_restart;
1877 restart->futex.uaddr = uaddr;
1878 restart->futex.val = val;
1879 restart->futex.time = abs_time->tv64;
1880 restart->futex.bitset = bitset;
1881 restart->futex.flags = flags;
1883 ret = -ERESTART_RESTARTBLOCK;
1887 hrtimer_cancel(&to->timer);
1888 destroy_hrtimer_on_stack(&to->timer);
1894 static long futex_wait_restart(struct restart_block *restart)
1896 u32 __user *uaddr = restart->futex.uaddr;
1897 ktime_t t, *tp = NULL;
1899 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1900 t.tv64 = restart->futex.time;
1903 restart->fn = do_no_restart_syscall;
1905 return (long)futex_wait(uaddr, restart->futex.flags,
1906 restart->futex.val, tp, restart->futex.bitset);
1911 * Userspace tried a 0 -> TID atomic transition of the futex value
1912 * and failed. The kernel side here does the whole locking operation:
1913 * if there are waiters then it will block, it does PI, etc. (Due to
1914 * races the kernel might see a 0 value of the futex too.)
1916 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1917 ktime_t *time, int trylock)
1919 struct hrtimer_sleeper timeout, *to = NULL;
1920 struct futex_hash_bucket *hb;
1921 struct futex_q q = futex_q_init;
1924 if (refill_pi_state_cache())
1929 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1931 hrtimer_init_sleeper(to, current);
1932 hrtimer_set_expires(&to->timer, *time);
1936 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key);
1937 if (unlikely(ret != 0))
1941 hb = queue_lock(&q);
1943 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1944 if (unlikely(ret)) {
1947 /* We got the lock. */
1949 goto out_unlock_put_key;
1954 * Task is exiting and we just wait for the
1957 queue_unlock(&q, hb);
1958 put_futex_key(&q.key);
1962 goto out_unlock_put_key;
1967 * Only actually queue now that the atomic ops are done:
1971 WARN_ON(!q.pi_state);
1973 * Block on the PI mutex:
1976 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1978 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1979 /* Fixup the trylock return value: */
1980 ret = ret ? 0 : -EWOULDBLOCK;
1983 spin_lock(q.lock_ptr);
1985 * Fixup the pi_state owner and possibly acquire the lock if we
1988 res = fixup_owner(uaddr, &q, !ret);
1990 * If fixup_owner() returned an error, proprogate that. If it acquired
1991 * the lock, clear our -ETIMEDOUT or -EINTR.
1994 ret = (res < 0) ? res : 0;
1997 * If fixup_owner() faulted and was unable to handle the fault, unlock
1998 * it and return the fault to userspace.
2000 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2001 rt_mutex_unlock(&q.pi_state->pi_mutex);
2003 /* Unqueue and drop the lock */
2009 queue_unlock(&q, hb);
2012 put_futex_key(&q.key);
2015 destroy_hrtimer_on_stack(&to->timer);
2016 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2019 queue_unlock(&q, hb);
2021 ret = fault_in_user_writeable(uaddr);
2025 if (!(flags & FLAGS_SHARED))
2028 put_futex_key(&q.key);
2033 * Userspace attempted a TID -> 0 atomic transition, and failed.
2034 * This is the in-kernel slowpath: we look up the PI state (if any),
2035 * and do the rt-mutex unlock.
2037 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2039 struct futex_hash_bucket *hb;
2040 struct futex_q *this, *next;
2041 struct plist_head *head;
2042 union futex_key key = FUTEX_KEY_INIT;
2043 u32 uval, vpid = task_pid_vnr(current);
2047 if (get_user(uval, uaddr))
2050 * We release only a lock we actually own:
2052 if ((uval & FUTEX_TID_MASK) != vpid)
2055 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
2056 if (unlikely(ret != 0))
2059 hb = hash_futex(&key);
2060 spin_lock(&hb->lock);
2063 * To avoid races, try to do the TID -> 0 atomic transition
2064 * again. If it succeeds then we can return without waking
2067 if (!(uval & FUTEX_OWNER_DIED) &&
2068 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2071 * Rare case: we managed to release the lock atomically,
2072 * no need to wake anyone else up:
2074 if (unlikely(uval == vpid))
2078 * Ok, other tasks may need to be woken up - check waiters
2079 * and do the wakeup if necessary:
2083 plist_for_each_entry_safe(this, next, head, list) {
2084 if (!match_futex (&this->key, &key))
2086 ret = wake_futex_pi(uaddr, uval, this);
2088 * The atomic access to the futex value
2089 * generated a pagefault, so retry the
2090 * user-access and the wakeup:
2097 * No waiters - kernel unlocks the futex:
2099 if (!(uval & FUTEX_OWNER_DIED)) {
2100 ret = unlock_futex_pi(uaddr, uval);
2106 spin_unlock(&hb->lock);
2107 put_futex_key(&key);
2113 spin_unlock(&hb->lock);
2114 put_futex_key(&key);
2116 ret = fault_in_user_writeable(uaddr);
2124 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2125 * @hb: the hash_bucket futex_q was original enqueued on
2126 * @q: the futex_q woken while waiting to be requeued
2127 * @key2: the futex_key of the requeue target futex
2128 * @timeout: the timeout associated with the wait (NULL if none)
2130 * Detect if the task was woken on the initial futex as opposed to the requeue
2131 * target futex. If so, determine if it was a timeout or a signal that caused
2132 * the wakeup and return the appropriate error code to the caller. Must be
2133 * called with the hb lock held.
2136 * 0 - no early wakeup detected
2137 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2140 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2141 struct futex_q *q, union futex_key *key2,
2142 struct hrtimer_sleeper *timeout)
2147 * With the hb lock held, we avoid races while we process the wakeup.
2148 * We only need to hold hb (and not hb2) to ensure atomicity as the
2149 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2150 * It can't be requeued from uaddr2 to something else since we don't
2151 * support a PI aware source futex for requeue.
2153 if (!match_futex(&q->key, key2)) {
2154 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2156 * We were woken prior to requeue by a timeout or a signal.
2157 * Unqueue the futex_q and determine which it was.
2159 plist_del(&q->list, &q->list.plist);
2161 /* Handle spurious wakeups gracefully */
2163 if (timeout && !timeout->task)
2165 else if (signal_pending(current))
2166 ret = -ERESTARTNOINTR;
2172 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2173 * @uaddr: the futex we initially wait on (non-pi)
2174 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2175 * the same type, no requeueing from private to shared, etc.
2176 * @val: the expected value of uaddr
2177 * @abs_time: absolute timeout
2178 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2179 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2180 * @uaddr2: the pi futex we will take prior to returning to user-space
2182 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2183 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2184 * complete the acquisition of the rt_mutex prior to returning to userspace.
2185 * This ensures the rt_mutex maintains an owner when it has waiters; without
2186 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2189 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2190 * via the following:
2191 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2192 * 2) wakeup on uaddr2 after a requeue
2196 * If 3, cleanup and return -ERESTARTNOINTR.
2198 * If 2, we may then block on trying to take the rt_mutex and return via:
2199 * 5) successful lock
2202 * 8) other lock acquisition failure
2204 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2206 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2212 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2213 u32 val, ktime_t *abs_time, u32 bitset,
2216 struct hrtimer_sleeper timeout, *to = NULL;
2217 struct rt_mutex_waiter rt_waiter;
2218 struct rt_mutex *pi_mutex = NULL;
2219 struct futex_hash_bucket *hb;
2220 union futex_key key2 = FUTEX_KEY_INIT;
2221 struct futex_q q = futex_q_init;
2229 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2230 CLOCK_REALTIME : CLOCK_MONOTONIC,
2232 hrtimer_init_sleeper(to, current);
2233 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2234 current->timer_slack_ns);
2238 * The waiter is allocated on our stack, manipulated by the requeue
2239 * code while we sleep on uaddr.
2241 debug_rt_mutex_init_waiter(&rt_waiter);
2242 rt_waiter.task = NULL;
2244 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
2245 if (unlikely(ret != 0))
2249 q.rt_waiter = &rt_waiter;
2250 q.requeue_pi_key = &key2;
2253 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2256 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2260 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2261 futex_wait_queue_me(hb, &q, to);
2263 spin_lock(&hb->lock);
2264 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2265 spin_unlock(&hb->lock);
2270 * In order for us to be here, we know our q.key == key2, and since
2271 * we took the hb->lock above, we also know that futex_requeue() has
2272 * completed and we no longer have to concern ourselves with a wakeup
2273 * race with the atomic proxy lock acquisition by the requeue code. The
2274 * futex_requeue dropped our key1 reference and incremented our key2
2278 /* Check if the requeue code acquired the second futex for us. */
2281 * Got the lock. We might not be the anticipated owner if we
2282 * did a lock-steal - fix up the PI-state in that case.
2284 if (q.pi_state && (q.pi_state->owner != current)) {
2285 spin_lock(q.lock_ptr);
2286 ret = fixup_pi_state_owner(uaddr2, &q, current);
2287 spin_unlock(q.lock_ptr);
2291 * We have been woken up by futex_unlock_pi(), a timeout, or a
2292 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2295 WARN_ON(!&q.pi_state);
2296 pi_mutex = &q.pi_state->pi_mutex;
2297 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2298 debug_rt_mutex_free_waiter(&rt_waiter);
2300 spin_lock(q.lock_ptr);
2302 * Fixup the pi_state owner and possibly acquire the lock if we
2305 res = fixup_owner(uaddr2, &q, !ret);
2307 * If fixup_owner() returned an error, proprogate that. If it
2308 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2311 ret = (res < 0) ? res : 0;
2313 /* Unqueue and drop the lock. */
2318 * If fixup_pi_state_owner() faulted and was unable to handle the
2319 * fault, unlock the rt_mutex and return the fault to userspace.
2321 if (ret == -EFAULT) {
2322 if (rt_mutex_owner(pi_mutex) == current)
2323 rt_mutex_unlock(pi_mutex);
2324 } else if (ret == -EINTR) {
2326 * We've already been requeued, but cannot restart by calling
2327 * futex_lock_pi() directly. We could restart this syscall, but
2328 * it would detect that the user space "val" changed and return
2329 * -EWOULDBLOCK. Save the overhead of the restart and return
2330 * -EWOULDBLOCK directly.
2336 put_futex_key(&q.key);
2338 put_futex_key(&key2);
2342 hrtimer_cancel(&to->timer);
2343 destroy_hrtimer_on_stack(&to->timer);
2349 * Support for robust futexes: the kernel cleans up held futexes at
2352 * Implementation: user-space maintains a per-thread list of locks it
2353 * is holding. Upon do_exit(), the kernel carefully walks this list,
2354 * and marks all locks that are owned by this thread with the
2355 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2356 * always manipulated with the lock held, so the list is private and
2357 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2358 * field, to allow the kernel to clean up if the thread dies after
2359 * acquiring the lock, but just before it could have added itself to
2360 * the list. There can only be one such pending lock.
2364 * sys_set_robust_list() - Set the robust-futex list head of a task
2365 * @head: pointer to the list-head
2366 * @len: length of the list-head, as userspace expects
2368 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2371 if (!futex_cmpxchg_enabled)
2374 * The kernel knows only one size for now:
2376 if (unlikely(len != sizeof(*head)))
2379 current->robust_list = head;
2385 * sys_get_robust_list() - Get the robust-futex list head of a task
2386 * @pid: pid of the process [zero for current task]
2387 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2388 * @len_ptr: pointer to a length field, the kernel fills in the header size
2390 SYSCALL_DEFINE3(get_robust_list, int, pid,
2391 struct robust_list_head __user * __user *, head_ptr,
2392 size_t __user *, len_ptr)
2394 struct robust_list_head __user *head;
2396 const struct cred *cred = current_cred(), *pcred;
2398 if (!futex_cmpxchg_enabled)
2402 head = current->robust_list;
2404 struct task_struct *p;
2408 p = find_task_by_vpid(pid);
2412 pcred = __task_cred(p);
2413 if (cred->euid != pcred->euid &&
2414 cred->euid != pcred->uid &&
2415 !capable(CAP_SYS_PTRACE))
2417 head = p->robust_list;
2421 if (put_user(sizeof(*head), len_ptr))
2423 return put_user(head, head_ptr);
2432 * Process a futex-list entry, check whether it's owned by the
2433 * dying task, and do notification if so:
2435 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2437 u32 uval, nval, mval;
2440 if (get_user(uval, uaddr))
2443 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2445 * Ok, this dying thread is truly holding a futex
2446 * of interest. Set the OWNER_DIED bit atomically
2447 * via cmpxchg, and if the value had FUTEX_WAITERS
2448 * set, wake up a waiter (if any). (We have to do a
2449 * futex_wake() even if OWNER_DIED is already set -
2450 * to handle the rare but possible case of recursive
2451 * thread-death.) The rest of the cleanup is done in
2454 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2455 if (futex_atomic_cmpxchg_inatomic(&nval, uaddr, uval, mval))
2462 * Wake robust non-PI futexes here. The wakeup of
2463 * PI futexes happens in exit_pi_state():
2465 if (!pi && (uval & FUTEX_WAITERS))
2466 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2472 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2474 static inline int fetch_robust_entry(struct robust_list __user **entry,
2475 struct robust_list __user * __user *head,
2478 unsigned long uentry;
2480 if (get_user(uentry, (unsigned long __user *)head))
2483 *entry = (void __user *)(uentry & ~1UL);
2490 * Walk curr->robust_list (very carefully, it's a userspace list!)
2491 * and mark any locks found there dead, and notify any waiters.
2493 * We silently return on any sign of list-walking problem.
2495 void exit_robust_list(struct task_struct *curr)
2497 struct robust_list_head __user *head = curr->robust_list;
2498 struct robust_list __user *entry, *next_entry, *pending;
2499 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2500 unsigned int uninitialized_var(next_pi);
2501 unsigned long futex_offset;
2504 if (!futex_cmpxchg_enabled)
2508 * Fetch the list head (which was registered earlier, via
2509 * sys_set_robust_list()):
2511 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2514 * Fetch the relative futex offset:
2516 if (get_user(futex_offset, &head->futex_offset))
2519 * Fetch any possibly pending lock-add first, and handle it
2522 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2525 next_entry = NULL; /* avoid warning with gcc */
2526 while (entry != &head->list) {
2528 * Fetch the next entry in the list before calling
2529 * handle_futex_death:
2531 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2533 * A pending lock might already be on the list, so
2534 * don't process it twice:
2536 if (entry != pending)
2537 if (handle_futex_death((void __user *)entry + futex_offset,
2545 * Avoid excessively long or circular lists:
2554 handle_futex_death((void __user *)pending + futex_offset,
2558 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2559 u32 __user *uaddr2, u32 val2, u32 val3)
2561 int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2562 unsigned int flags = 0;
2564 if (!(op & FUTEX_PRIVATE_FLAG))
2565 flags |= FLAGS_SHARED;
2567 if (op & FUTEX_CLOCK_REALTIME) {
2568 flags |= FLAGS_CLOCKRT;
2569 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2575 val3 = FUTEX_BITSET_MATCH_ANY;
2576 case FUTEX_WAIT_BITSET:
2577 ret = futex_wait(uaddr, flags, val, timeout, val3);
2580 val3 = FUTEX_BITSET_MATCH_ANY;
2581 case FUTEX_WAKE_BITSET:
2582 ret = futex_wake(uaddr, flags, val, val3);
2585 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2587 case FUTEX_CMP_REQUEUE:
2588 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2591 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2594 if (futex_cmpxchg_enabled)
2595 ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2597 case FUTEX_UNLOCK_PI:
2598 if (futex_cmpxchg_enabled)
2599 ret = futex_unlock_pi(uaddr, flags);
2601 case FUTEX_TRYLOCK_PI:
2602 if (futex_cmpxchg_enabled)
2603 ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2605 case FUTEX_WAIT_REQUEUE_PI:
2606 val3 = FUTEX_BITSET_MATCH_ANY;
2607 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2610 case FUTEX_CMP_REQUEUE_PI:
2611 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2620 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2621 struct timespec __user *, utime, u32 __user *, uaddr2,
2625 ktime_t t, *tp = NULL;
2627 int cmd = op & FUTEX_CMD_MASK;
2629 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2630 cmd == FUTEX_WAIT_BITSET ||
2631 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2632 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2634 if (!timespec_valid(&ts))
2637 t = timespec_to_ktime(ts);
2638 if (cmd == FUTEX_WAIT)
2639 t = ktime_add_safe(ktime_get(), t);
2643 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2644 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2646 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2647 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2648 val2 = (u32) (unsigned long) utime;
2650 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2653 static int __init futex_init(void)
2659 * This will fail and we want it. Some arch implementations do
2660 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2661 * functionality. We want to know that before we call in any
2662 * of the complex code paths. Also we want to prevent
2663 * registration of robust lists in that case. NULL is
2664 * guaranteed to fault and we get -EFAULT on functional
2665 * implementation, the non-functional ones will return
2668 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2669 futex_cmpxchg_enabled = 1;
2671 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2672 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2673 spin_lock_init(&futex_queues[i].lock);
2678 __initcall(futex_init);