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;
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 page = compound_head(page);
270 if (!page->mapping) {
277 * Private mappings are handled in a simple way.
279 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
280 * it's a read-only handle, it's expected that futexes attach to
281 * the object not the particular process.
283 if (PageAnon(page)) {
284 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
285 key->private.mm = mm;
286 key->private.address = address;
288 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
289 key->shared.inode = page->mapping->host;
290 key->shared.pgoff = page->index;
293 get_futex_key_refs(key);
300 static inline void put_futex_key(union futex_key *key)
302 drop_futex_key_refs(key);
306 * fault_in_user_writeable() - Fault in user address and verify RW access
307 * @uaddr: pointer to faulting user space address
309 * Slow path to fixup the fault we just took in the atomic write
312 * We have no generic implementation of a non-destructive write to the
313 * user address. We know that we faulted in the atomic pagefault
314 * disabled section so we can as well avoid the #PF overhead by
315 * calling get_user_pages() right away.
317 static int fault_in_user_writeable(u32 __user *uaddr)
319 struct mm_struct *mm = current->mm;
322 down_read(&mm->mmap_sem);
323 ret = get_user_pages(current, mm, (unsigned long)uaddr,
324 1, 1, 0, NULL, NULL);
325 up_read(&mm->mmap_sem);
327 return ret < 0 ? ret : 0;
331 * futex_top_waiter() - Return the highest priority waiter on a futex
332 * @hb: the hash bucket the futex_q's reside in
333 * @key: the futex key (to distinguish it from other futex futex_q's)
335 * Must be called with the hb lock held.
337 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
338 union futex_key *key)
340 struct futex_q *this;
342 plist_for_each_entry(this, &hb->chain, list) {
343 if (match_futex(&this->key, key))
349 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
354 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
360 static int get_futex_value_locked(u32 *dest, u32 __user *from)
365 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
368 return ret ? -EFAULT : 0;
375 static int refill_pi_state_cache(void)
377 struct futex_pi_state *pi_state;
379 if (likely(current->pi_state_cache))
382 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
387 INIT_LIST_HEAD(&pi_state->list);
388 /* pi_mutex gets initialized later */
389 pi_state->owner = NULL;
390 atomic_set(&pi_state->refcount, 1);
391 pi_state->key = FUTEX_KEY_INIT;
393 current->pi_state_cache = pi_state;
398 static struct futex_pi_state * alloc_pi_state(void)
400 struct futex_pi_state *pi_state = current->pi_state_cache;
403 current->pi_state_cache = NULL;
408 static void free_pi_state(struct futex_pi_state *pi_state)
410 if (!atomic_dec_and_test(&pi_state->refcount))
414 * If pi_state->owner is NULL, the owner is most probably dying
415 * and has cleaned up the pi_state already
417 if (pi_state->owner) {
418 raw_spin_lock_irq(&pi_state->owner->pi_lock);
419 list_del_init(&pi_state->list);
420 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
422 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
425 if (current->pi_state_cache)
429 * pi_state->list is already empty.
430 * clear pi_state->owner.
431 * refcount is at 0 - put it back to 1.
433 pi_state->owner = NULL;
434 atomic_set(&pi_state->refcount, 1);
435 current->pi_state_cache = pi_state;
440 * Look up the task based on what TID userspace gave us.
443 static struct task_struct * futex_find_get_task(pid_t pid)
445 struct task_struct *p;
448 p = find_task_by_vpid(pid);
458 * This task is holding PI mutexes at exit time => bad.
459 * Kernel cleans up PI-state, but userspace is likely hosed.
460 * (Robust-futex cleanup is separate and might save the day for userspace.)
462 void exit_pi_state_list(struct task_struct *curr)
464 struct list_head *next, *head = &curr->pi_state_list;
465 struct futex_pi_state *pi_state;
466 struct futex_hash_bucket *hb;
467 union futex_key key = FUTEX_KEY_INIT;
469 if (!futex_cmpxchg_enabled)
472 * We are a ZOMBIE and nobody can enqueue itself on
473 * pi_state_list anymore, but we have to be careful
474 * versus waiters unqueueing themselves:
476 raw_spin_lock_irq(&curr->pi_lock);
477 while (!list_empty(head)) {
480 pi_state = list_entry(next, struct futex_pi_state, list);
482 hb = hash_futex(&key);
483 raw_spin_unlock_irq(&curr->pi_lock);
485 spin_lock(&hb->lock);
487 raw_spin_lock_irq(&curr->pi_lock);
489 * We dropped the pi-lock, so re-check whether this
490 * task still owns the PI-state:
492 if (head->next != next) {
493 spin_unlock(&hb->lock);
497 WARN_ON(pi_state->owner != curr);
498 WARN_ON(list_empty(&pi_state->list));
499 list_del_init(&pi_state->list);
500 pi_state->owner = NULL;
501 raw_spin_unlock_irq(&curr->pi_lock);
503 rt_mutex_unlock(&pi_state->pi_mutex);
505 spin_unlock(&hb->lock);
507 raw_spin_lock_irq(&curr->pi_lock);
509 raw_spin_unlock_irq(&curr->pi_lock);
513 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
514 union futex_key *key, struct futex_pi_state **ps)
516 struct futex_pi_state *pi_state = NULL;
517 struct futex_q *this, *next;
518 struct plist_head *head;
519 struct task_struct *p;
520 pid_t pid = uval & FUTEX_TID_MASK;
524 plist_for_each_entry_safe(this, next, head, list) {
525 if (match_futex(&this->key, key)) {
527 * Another waiter already exists - bump up
528 * the refcount and return its pi_state:
530 pi_state = this->pi_state;
532 * Userspace might have messed up non-PI and PI futexes
534 if (unlikely(!pi_state))
537 WARN_ON(!atomic_read(&pi_state->refcount));
540 * When pi_state->owner is NULL then the owner died
541 * and another waiter is on the fly. pi_state->owner
542 * is fixed up by the task which acquires
543 * pi_state->rt_mutex.
545 * We do not check for pid == 0 which can happen when
546 * the owner died and robust_list_exit() cleared the
549 if (pid && pi_state->owner) {
551 * Bail out if user space manipulated the
554 if (pid != task_pid_vnr(pi_state->owner))
558 atomic_inc(&pi_state->refcount);
566 * We are the first waiter - try to look up the real owner and attach
567 * the new pi_state to it, but bail out when TID = 0
571 p = futex_find_get_task(pid);
576 * We need to look at the task state flags to figure out,
577 * whether the task is exiting. To protect against the do_exit
578 * change of the task flags, we do this protected by
581 raw_spin_lock_irq(&p->pi_lock);
582 if (unlikely(p->flags & PF_EXITING)) {
584 * The task is on the way out. When PF_EXITPIDONE is
585 * set, we know that the task has finished the
588 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
590 raw_spin_unlock_irq(&p->pi_lock);
595 pi_state = alloc_pi_state();
598 * Initialize the pi_mutex in locked state and make 'p'
601 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
603 /* Store the key for possible exit cleanups: */
604 pi_state->key = *key;
606 WARN_ON(!list_empty(&pi_state->list));
607 list_add(&pi_state->list, &p->pi_state_list);
609 raw_spin_unlock_irq(&p->pi_lock);
619 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
620 * @uaddr: the pi futex user address
621 * @hb: the pi futex hash bucket
622 * @key: the futex key associated with uaddr and hb
623 * @ps: the pi_state pointer where we store the result of the
625 * @task: the task to perform the atomic lock work for. This will
626 * be "current" except in the case of requeue pi.
627 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
631 * 1 - acquired the lock
634 * The hb->lock and futex_key refs shall be held by the caller.
636 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
637 union futex_key *key,
638 struct futex_pi_state **ps,
639 struct task_struct *task, int set_waiters)
641 int lock_taken, ret, ownerdied = 0;
642 u32 uval, newval, curval;
645 ret = lock_taken = 0;
648 * To avoid races, we attempt to take the lock here again
649 * (by doing a 0 -> TID atomic cmpxchg), while holding all
650 * the locks. It will most likely not succeed.
652 newval = task_pid_vnr(task);
654 newval |= FUTEX_WAITERS;
656 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
658 if (unlikely(curval == -EFAULT))
664 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
668 * Surprise - we got the lock. Just return to userspace:
670 if (unlikely(!curval))
676 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
677 * to wake at the next unlock.
679 newval = curval | FUTEX_WAITERS;
682 * There are two cases, where a futex might have no owner (the
683 * owner TID is 0): OWNER_DIED. We take over the futex in this
684 * case. We also do an unconditional take over, when the owner
687 * This is safe as we are protected by the hash bucket lock !
689 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
690 /* Keep the OWNER_DIED bit */
691 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
696 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
698 if (unlikely(curval == -EFAULT))
700 if (unlikely(curval != uval))
704 * We took the lock due to owner died take over.
706 if (unlikely(lock_taken))
710 * We dont have the lock. Look up the PI state (or create it if
711 * we are the first waiter):
713 ret = lookup_pi_state(uval, hb, key, ps);
719 * No owner found for this futex. Check if the
720 * OWNER_DIED bit is set to figure out whether
721 * this is a robust futex or not.
723 if (get_futex_value_locked(&curval, uaddr))
727 * We simply start over in case of a robust
728 * futex. The code above will take the futex
731 if (curval & FUTEX_OWNER_DIED) {
744 * The hash bucket lock must be held when this is called.
745 * Afterwards, the futex_q must not be accessed.
747 static void wake_futex(struct futex_q *q)
749 struct task_struct *p = q->task;
752 * We set q->lock_ptr = NULL _before_ we wake up the task. If
753 * a non-futex wake up happens on another CPU then the task
754 * might exit and p would dereference a non-existing task
755 * struct. Prevent this by holding a reference on p across the
760 plist_del(&q->list, &q->list.plist);
762 * The waiting task can free the futex_q as soon as
763 * q->lock_ptr = NULL is written, without taking any locks. A
764 * memory barrier is required here to prevent the following
765 * store to lock_ptr from getting ahead of the plist_del.
770 wake_up_state(p, TASK_NORMAL);
774 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
776 struct task_struct *new_owner;
777 struct futex_pi_state *pi_state = this->pi_state;
784 * If current does not own the pi_state then the futex is
785 * inconsistent and user space fiddled with the futex value.
787 if (pi_state->owner != current)
790 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
791 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
794 * It is possible that the next waiter (the one that brought
795 * this owner to the kernel) timed out and is no longer
796 * waiting on the lock.
799 new_owner = this->task;
802 * We pass it to the next owner. (The WAITERS bit is always
803 * kept enabled while there is PI state around. We must also
804 * preserve the owner died bit.)
806 if (!(uval & FUTEX_OWNER_DIED)) {
809 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
811 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
813 if (curval == -EFAULT)
815 else if (curval != uval)
818 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
823 raw_spin_lock_irq(&pi_state->owner->pi_lock);
824 WARN_ON(list_empty(&pi_state->list));
825 list_del_init(&pi_state->list);
826 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
828 raw_spin_lock_irq(&new_owner->pi_lock);
829 WARN_ON(!list_empty(&pi_state->list));
830 list_add(&pi_state->list, &new_owner->pi_state_list);
831 pi_state->owner = new_owner;
832 raw_spin_unlock_irq(&new_owner->pi_lock);
834 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
835 rt_mutex_unlock(&pi_state->pi_mutex);
840 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
845 * There is no waiter, so we unlock the futex. The owner died
846 * bit has not to be preserved here. We are the owner:
848 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
850 if (oldval == -EFAULT)
859 * Express the locking dependencies for lockdep:
862 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
865 spin_lock(&hb1->lock);
867 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
868 } else { /* hb1 > hb2 */
869 spin_lock(&hb2->lock);
870 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
875 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
877 spin_unlock(&hb1->lock);
879 spin_unlock(&hb2->lock);
883 * Wake up waiters matching bitset queued on this futex (uaddr).
886 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
888 struct futex_hash_bucket *hb;
889 struct futex_q *this, *next;
890 struct plist_head *head;
891 union futex_key key = FUTEX_KEY_INIT;
897 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
898 if (unlikely(ret != 0))
901 hb = hash_futex(&key);
902 spin_lock(&hb->lock);
905 plist_for_each_entry_safe(this, next, head, list) {
906 if (match_futex (&this->key, &key)) {
907 if (this->pi_state || this->rt_waiter) {
912 /* Check if one of the bits is set in both bitsets */
913 if (!(this->bitset & bitset))
917 if (++ret >= nr_wake)
922 spin_unlock(&hb->lock);
929 * Wake up all waiters hashed on the physical page that is mapped
930 * to this virtual address:
933 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
934 int nr_wake, int nr_wake2, int op)
936 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
937 struct futex_hash_bucket *hb1, *hb2;
938 struct plist_head *head;
939 struct futex_q *this, *next;
943 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
944 if (unlikely(ret != 0))
946 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
947 if (unlikely(ret != 0))
950 hb1 = hash_futex(&key1);
951 hb2 = hash_futex(&key2);
954 double_lock_hb(hb1, hb2);
955 op_ret = futex_atomic_op_inuser(op, uaddr2);
956 if (unlikely(op_ret < 0)) {
958 double_unlock_hb(hb1, hb2);
962 * we don't get EFAULT from MMU faults if we don't have an MMU,
963 * but we might get them from range checking
969 if (unlikely(op_ret != -EFAULT)) {
974 ret = fault_in_user_writeable(uaddr2);
978 if (!(flags & FLAGS_SHARED))
981 put_futex_key(&key2);
982 put_futex_key(&key1);
988 plist_for_each_entry_safe(this, next, head, list) {
989 if (match_futex (&this->key, &key1)) {
991 if (++ret >= nr_wake)
1000 plist_for_each_entry_safe(this, next, head, list) {
1001 if (match_futex (&this->key, &key2)) {
1003 if (++op_ret >= nr_wake2)
1010 double_unlock_hb(hb1, hb2);
1012 put_futex_key(&key2);
1014 put_futex_key(&key1);
1020 * requeue_futex() - Requeue a futex_q from one hb to another
1021 * @q: the futex_q to requeue
1022 * @hb1: the source hash_bucket
1023 * @hb2: the target hash_bucket
1024 * @key2: the new key for the requeued futex_q
1027 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1028 struct futex_hash_bucket *hb2, union futex_key *key2)
1032 * If key1 and key2 hash to the same bucket, no need to
1035 if (likely(&hb1->chain != &hb2->chain)) {
1036 plist_del(&q->list, &hb1->chain);
1037 plist_add(&q->list, &hb2->chain);
1038 q->lock_ptr = &hb2->lock;
1039 #ifdef CONFIG_DEBUG_PI_LIST
1040 q->list.plist.spinlock = &hb2->lock;
1043 get_futex_key_refs(key2);
1048 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1050 * @key: the key of the requeue target futex
1051 * @hb: the hash_bucket of the requeue target futex
1053 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1054 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1055 * to the requeue target futex so the waiter can detect the wakeup on the right
1056 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1057 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1058 * to protect access to the pi_state to fixup the owner later. Must be called
1059 * with both q->lock_ptr and hb->lock held.
1062 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1063 struct futex_hash_bucket *hb)
1065 get_futex_key_refs(key);
1068 WARN_ON(plist_node_empty(&q->list));
1069 plist_del(&q->list, &q->list.plist);
1071 WARN_ON(!q->rt_waiter);
1072 q->rt_waiter = NULL;
1074 q->lock_ptr = &hb->lock;
1075 #ifdef CONFIG_DEBUG_PI_LIST
1076 q->list.plist.spinlock = &hb->lock;
1079 wake_up_state(q->task, TASK_NORMAL);
1083 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1084 * @pifutex: the user address of the to futex
1085 * @hb1: the from futex hash bucket, must be locked by the caller
1086 * @hb2: the to futex hash bucket, must be locked by the caller
1087 * @key1: the from futex key
1088 * @key2: the to futex key
1089 * @ps: address to store the pi_state pointer
1090 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1092 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1093 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1094 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1095 * hb1 and hb2 must be held by the caller.
1098 * 0 - failed to acquire the lock atomicly
1099 * 1 - acquired the lock
1102 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1103 struct futex_hash_bucket *hb1,
1104 struct futex_hash_bucket *hb2,
1105 union futex_key *key1, union futex_key *key2,
1106 struct futex_pi_state **ps, int set_waiters)
1108 struct futex_q *top_waiter = NULL;
1112 if (get_futex_value_locked(&curval, pifutex))
1116 * Find the top_waiter and determine if there are additional waiters.
1117 * If the caller intends to requeue more than 1 waiter to pifutex,
1118 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1119 * as we have means to handle the possible fault. If not, don't set
1120 * the bit unecessarily as it will force the subsequent unlock to enter
1123 top_waiter = futex_top_waiter(hb1, key1);
1125 /* There are no waiters, nothing for us to do. */
1129 /* Ensure we requeue to the expected futex. */
1130 if (!match_futex(top_waiter->requeue_pi_key, key2))
1134 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1135 * the contended case or if set_waiters is 1. The pi_state is returned
1136 * in ps in contended cases.
1138 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1141 requeue_pi_wake_futex(top_waiter, key2, hb2);
1147 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1148 * @uaddr1: source futex user address
1149 * @flags: futex flags (FLAGS_SHARED, etc.)
1150 * @uaddr2: target futex user address
1151 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1152 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1153 * @cmpval: @uaddr1 expected value (or %NULL)
1154 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1155 * pi futex (pi to pi requeue is not supported)
1157 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1158 * uaddr2 atomically on behalf of the top waiter.
1161 * >=0 - on success, the number of tasks requeued or woken
1164 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1165 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1166 u32 *cmpval, int requeue_pi)
1168 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1169 int drop_count = 0, task_count = 0, ret;
1170 struct futex_pi_state *pi_state = NULL;
1171 struct futex_hash_bucket *hb1, *hb2;
1172 struct plist_head *head1;
1173 struct futex_q *this, *next;
1178 * requeue_pi requires a pi_state, try to allocate it now
1179 * without any locks in case it fails.
1181 if (refill_pi_state_cache())
1184 * requeue_pi must wake as many tasks as it can, up to nr_wake
1185 * + nr_requeue, since it acquires the rt_mutex prior to
1186 * returning to userspace, so as to not leave the rt_mutex with
1187 * waiters and no owner. However, second and third wake-ups
1188 * cannot be predicted as they involve race conditions with the
1189 * first wake and a fault while looking up the pi_state. Both
1190 * pthread_cond_signal() and pthread_cond_broadcast() should
1198 if (pi_state != NULL) {
1200 * We will have to lookup the pi_state again, so free this one
1201 * to keep the accounting correct.
1203 free_pi_state(pi_state);
1207 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
1208 if (unlikely(ret != 0))
1210 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
1211 if (unlikely(ret != 0))
1214 hb1 = hash_futex(&key1);
1215 hb2 = hash_futex(&key2);
1218 double_lock_hb(hb1, hb2);
1220 if (likely(cmpval != NULL)) {
1223 ret = get_futex_value_locked(&curval, uaddr1);
1225 if (unlikely(ret)) {
1226 double_unlock_hb(hb1, hb2);
1228 ret = get_user(curval, uaddr1);
1232 if (!(flags & FLAGS_SHARED))
1235 put_futex_key(&key2);
1236 put_futex_key(&key1);
1239 if (curval != *cmpval) {
1245 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1247 * Attempt to acquire uaddr2 and wake the top waiter. If we
1248 * intend to requeue waiters, force setting the FUTEX_WAITERS
1249 * bit. We force this here where we are able to easily handle
1250 * faults rather in the requeue loop below.
1252 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1253 &key2, &pi_state, nr_requeue);
1256 * At this point the top_waiter has either taken uaddr2 or is
1257 * waiting on it. If the former, then the pi_state will not
1258 * exist yet, look it up one more time to ensure we have a
1265 ret = get_futex_value_locked(&curval2, uaddr2);
1267 ret = lookup_pi_state(curval2, hb2, &key2,
1275 double_unlock_hb(hb1, hb2);
1276 put_futex_key(&key2);
1277 put_futex_key(&key1);
1278 ret = fault_in_user_writeable(uaddr2);
1283 /* The owner was exiting, try again. */
1284 double_unlock_hb(hb1, hb2);
1285 put_futex_key(&key2);
1286 put_futex_key(&key1);
1294 head1 = &hb1->chain;
1295 plist_for_each_entry_safe(this, next, head1, list) {
1296 if (task_count - nr_wake >= nr_requeue)
1299 if (!match_futex(&this->key, &key1))
1303 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1304 * be paired with each other and no other futex ops.
1306 if ((requeue_pi && !this->rt_waiter) ||
1307 (!requeue_pi && this->rt_waiter)) {
1313 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1314 * lock, we already woke the top_waiter. If not, it will be
1315 * woken by futex_unlock_pi().
1317 if (++task_count <= nr_wake && !requeue_pi) {
1322 /* Ensure we requeue to the expected futex for requeue_pi. */
1323 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1329 * Requeue nr_requeue waiters and possibly one more in the case
1330 * of requeue_pi if we couldn't acquire the lock atomically.
1333 /* Prepare the waiter to take the rt_mutex. */
1334 atomic_inc(&pi_state->refcount);
1335 this->pi_state = pi_state;
1336 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1340 /* We got the lock. */
1341 requeue_pi_wake_futex(this, &key2, hb2);
1346 this->pi_state = NULL;
1347 free_pi_state(pi_state);
1351 requeue_futex(this, hb1, hb2, &key2);
1356 double_unlock_hb(hb1, hb2);
1359 * drop_futex_key_refs() must be called outside the spinlocks. During
1360 * the requeue we moved futex_q's from the hash bucket at key1 to the
1361 * one at key2 and updated their key pointer. We no longer need to
1362 * hold the references to key1.
1364 while (--drop_count >= 0)
1365 drop_futex_key_refs(&key1);
1368 put_futex_key(&key2);
1370 put_futex_key(&key1);
1372 if (pi_state != NULL)
1373 free_pi_state(pi_state);
1374 return ret ? ret : task_count;
1377 /* The key must be already stored in q->key. */
1378 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1379 __acquires(&hb->lock)
1381 struct futex_hash_bucket *hb;
1383 hb = hash_futex(&q->key);
1384 q->lock_ptr = &hb->lock;
1386 spin_lock(&hb->lock);
1391 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1392 __releases(&hb->lock)
1394 spin_unlock(&hb->lock);
1398 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1399 * @q: The futex_q to enqueue
1400 * @hb: The destination hash bucket
1402 * The hb->lock must be held by the caller, and is released here. A call to
1403 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1404 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1405 * or nothing if the unqueue is done as part of the wake process and the unqueue
1406 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1409 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1410 __releases(&hb->lock)
1415 * The priority used to register this element is
1416 * - either the real thread-priority for the real-time threads
1417 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1418 * - or MAX_RT_PRIO for non-RT threads.
1419 * Thus, all RT-threads are woken first in priority order, and
1420 * the others are woken last, in FIFO order.
1422 prio = min(current->normal_prio, MAX_RT_PRIO);
1424 plist_node_init(&q->list, prio);
1425 #ifdef CONFIG_DEBUG_PI_LIST
1426 q->list.plist.spinlock = &hb->lock;
1428 plist_add(&q->list, &hb->chain);
1430 spin_unlock(&hb->lock);
1434 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1435 * @q: The futex_q to unqueue
1437 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1438 * be paired with exactly one earlier call to queue_me().
1441 * 1 - if the futex_q was still queued (and we removed unqueued it)
1442 * 0 - if the futex_q was already removed by the waking thread
1444 static int unqueue_me(struct futex_q *q)
1446 spinlock_t *lock_ptr;
1449 /* In the common case we don't take the spinlock, which is nice. */
1451 lock_ptr = q->lock_ptr;
1453 if (lock_ptr != NULL) {
1454 spin_lock(lock_ptr);
1456 * q->lock_ptr can change between reading it and
1457 * spin_lock(), causing us to take the wrong lock. This
1458 * corrects the race condition.
1460 * Reasoning goes like this: if we have the wrong lock,
1461 * q->lock_ptr must have changed (maybe several times)
1462 * between reading it and the spin_lock(). It can
1463 * change again after the spin_lock() but only if it was
1464 * already changed before the spin_lock(). It cannot,
1465 * however, change back to the original value. Therefore
1466 * we can detect whether we acquired the correct lock.
1468 if (unlikely(lock_ptr != q->lock_ptr)) {
1469 spin_unlock(lock_ptr);
1472 WARN_ON(plist_node_empty(&q->list));
1473 plist_del(&q->list, &q->list.plist);
1475 BUG_ON(q->pi_state);
1477 spin_unlock(lock_ptr);
1481 drop_futex_key_refs(&q->key);
1486 * PI futexes can not be requeued and must remove themself from the
1487 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1490 static void unqueue_me_pi(struct futex_q *q)
1491 __releases(q->lock_ptr)
1493 WARN_ON(plist_node_empty(&q->list));
1494 plist_del(&q->list, &q->list.plist);
1496 BUG_ON(!q->pi_state);
1497 free_pi_state(q->pi_state);
1500 spin_unlock(q->lock_ptr);
1504 * Fixup the pi_state owner with the new owner.
1506 * Must be called with hash bucket lock held and mm->sem held for non
1509 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1510 struct task_struct *newowner)
1512 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1513 struct futex_pi_state *pi_state = q->pi_state;
1514 struct task_struct *oldowner = pi_state->owner;
1515 u32 uval, curval, newval;
1519 if (!pi_state->owner)
1520 newtid |= FUTEX_OWNER_DIED;
1523 * We are here either because we stole the rtmutex from the
1524 * pending owner or we are the pending owner which failed to
1525 * get the rtmutex. We have to replace the pending owner TID
1526 * in the user space variable. This must be atomic as we have
1527 * to preserve the owner died bit here.
1529 * Note: We write the user space value _before_ changing the pi_state
1530 * because we can fault here. Imagine swapped out pages or a fork
1531 * that marked all the anonymous memory readonly for cow.
1533 * Modifying pi_state _before_ the user space value would
1534 * leave the pi_state in an inconsistent state when we fault
1535 * here, because we need to drop the hash bucket lock to
1536 * handle the fault. This might be observed in the PID check
1537 * in lookup_pi_state.
1540 if (get_futex_value_locked(&uval, uaddr))
1544 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1546 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1548 if (curval == -EFAULT)
1556 * We fixed up user space. Now we need to fix the pi_state
1559 if (pi_state->owner != NULL) {
1560 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1561 WARN_ON(list_empty(&pi_state->list));
1562 list_del_init(&pi_state->list);
1563 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1566 pi_state->owner = newowner;
1568 raw_spin_lock_irq(&newowner->pi_lock);
1569 WARN_ON(!list_empty(&pi_state->list));
1570 list_add(&pi_state->list, &newowner->pi_state_list);
1571 raw_spin_unlock_irq(&newowner->pi_lock);
1575 * To handle the page fault we need to drop the hash bucket
1576 * lock here. That gives the other task (either the pending
1577 * owner itself or the task which stole the rtmutex) the
1578 * chance to try the fixup of the pi_state. So once we are
1579 * back from handling the fault we need to check the pi_state
1580 * after reacquiring the hash bucket lock and before trying to
1581 * do another fixup. When the fixup has been done already we
1585 spin_unlock(q->lock_ptr);
1587 ret = fault_in_user_writeable(uaddr);
1589 spin_lock(q->lock_ptr);
1592 * Check if someone else fixed it for us:
1594 if (pi_state->owner != oldowner)
1603 static long futex_wait_restart(struct restart_block *restart);
1606 * fixup_owner() - Post lock pi_state and corner case management
1607 * @uaddr: user address of the futex
1608 * @q: futex_q (contains pi_state and access to the rt_mutex)
1609 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1611 * After attempting to lock an rt_mutex, this function is called to cleanup
1612 * the pi_state owner as well as handle race conditions that may allow us to
1613 * acquire the lock. Must be called with the hb lock held.
1616 * 1 - success, lock taken
1617 * 0 - success, lock not taken
1618 * <0 - on error (-EFAULT)
1620 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1622 struct task_struct *owner;
1627 * Got the lock. We might not be the anticipated owner if we
1628 * did a lock-steal - fix up the PI-state in that case:
1630 if (q->pi_state->owner != current)
1631 ret = fixup_pi_state_owner(uaddr, q, current);
1636 * Catch the rare case, where the lock was released when we were on the
1637 * way back before we locked the hash bucket.
1639 if (q->pi_state->owner == current) {
1641 * Try to get the rt_mutex now. This might fail as some other
1642 * task acquired the rt_mutex after we removed ourself from the
1643 * rt_mutex waiters list.
1645 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1651 * pi_state is incorrect, some other task did a lock steal and
1652 * we returned due to timeout or signal without taking the
1653 * rt_mutex. Too late. We can access the rt_mutex_owner without
1654 * locking, as the other task is now blocked on the hash bucket
1655 * lock. Fix the state up.
1657 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1658 ret = fixup_pi_state_owner(uaddr, q, owner);
1663 * Paranoia check. If we did not take the lock, then we should not be
1664 * the owner, nor the pending owner, of the rt_mutex.
1666 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1667 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1668 "pi-state %p\n", ret,
1669 q->pi_state->pi_mutex.owner,
1670 q->pi_state->owner);
1673 return ret ? ret : locked;
1677 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1678 * @hb: the futex hash bucket, must be locked by the caller
1679 * @q: the futex_q to queue up on
1680 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1682 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1683 struct hrtimer_sleeper *timeout)
1686 * The task state is guaranteed to be set before another task can
1687 * wake it. set_current_state() is implemented using set_mb() and
1688 * queue_me() calls spin_unlock() upon completion, both serializing
1689 * access to the hash list and forcing another memory barrier.
1691 set_current_state(TASK_INTERRUPTIBLE);
1696 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1697 if (!hrtimer_active(&timeout->timer))
1698 timeout->task = NULL;
1702 * If we have been removed from the hash list, then another task
1703 * has tried to wake us, and we can skip the call to schedule().
1705 if (likely(!plist_node_empty(&q->list))) {
1707 * If the timer has already expired, current will already be
1708 * flagged for rescheduling. Only call schedule if there
1709 * is no timeout, or if it has yet to expire.
1711 if (!timeout || timeout->task)
1714 __set_current_state(TASK_RUNNING);
1718 * futex_wait_setup() - Prepare to wait on a futex
1719 * @uaddr: the futex userspace address
1720 * @val: the expected value
1721 * @flags: futex flags (FLAGS_SHARED, etc.)
1722 * @q: the associated futex_q
1723 * @hb: storage for hash_bucket pointer to be returned to caller
1725 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1726 * compare it with the expected value. Handle atomic faults internally.
1727 * Return with the hb lock held and a q.key reference on success, and unlocked
1728 * with no q.key reference on failure.
1731 * 0 - uaddr contains val and hb has been locked
1732 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1734 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1735 struct futex_q *q, struct futex_hash_bucket **hb)
1741 * Access the page AFTER the hash-bucket is locked.
1742 * Order is important:
1744 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1745 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1747 * The basic logical guarantee of a futex is that it blocks ONLY
1748 * if cond(var) is known to be true at the time of blocking, for
1749 * any cond. If we queued after testing *uaddr, that would open
1750 * a race condition where we could block indefinitely with
1751 * cond(var) false, which would violate the guarantee.
1753 * A consequence is that futex_wait() can return zero and absorb
1754 * a wakeup when *uaddr != val on entry to the syscall. This is
1758 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key);
1759 if (unlikely(ret != 0))
1763 *hb = queue_lock(q);
1765 ret = get_futex_value_locked(&uval, uaddr);
1768 queue_unlock(q, *hb);
1770 ret = get_user(uval, uaddr);
1774 if (!(flags & FLAGS_SHARED))
1777 put_futex_key(&q->key);
1782 queue_unlock(q, *hb);
1788 put_futex_key(&q->key);
1792 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1793 ktime_t *abs_time, u32 bitset)
1795 struct hrtimer_sleeper timeout, *to = NULL;
1796 struct restart_block *restart;
1797 struct futex_hash_bucket *hb;
1798 struct futex_q q = futex_q_init;
1808 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1809 CLOCK_REALTIME : CLOCK_MONOTONIC,
1811 hrtimer_init_sleeper(to, current);
1812 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1813 current->timer_slack_ns);
1818 * Prepare to wait on uaddr. On success, holds hb lock and increments
1821 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1825 /* queue_me and wait for wakeup, timeout, or a signal. */
1826 futex_wait_queue_me(hb, &q, to);
1828 /* If we were woken (and unqueued), we succeeded, whatever. */
1830 /* unqueue_me() drops q.key ref */
1831 if (!unqueue_me(&q))
1834 if (to && !to->task)
1838 * We expect signal_pending(current), but we might be the
1839 * victim of a spurious wakeup as well.
1841 if (!signal_pending(current))
1848 restart = ¤t_thread_info()->restart_block;
1849 restart->fn = futex_wait_restart;
1850 restart->futex.uaddr = uaddr;
1851 restart->futex.val = val;
1852 restart->futex.time = abs_time->tv64;
1853 restart->futex.bitset = bitset;
1854 restart->futex.flags = flags;
1856 ret = -ERESTART_RESTARTBLOCK;
1860 hrtimer_cancel(&to->timer);
1861 destroy_hrtimer_on_stack(&to->timer);
1867 static long futex_wait_restart(struct restart_block *restart)
1869 u32 __user *uaddr = restart->futex.uaddr;
1870 ktime_t t, *tp = NULL;
1872 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1873 t.tv64 = restart->futex.time;
1876 restart->fn = do_no_restart_syscall;
1878 return (long)futex_wait(uaddr, restart->futex.flags,
1879 restart->futex.val, tp, restart->futex.bitset);
1884 * Userspace tried a 0 -> TID atomic transition of the futex value
1885 * and failed. The kernel side here does the whole locking operation:
1886 * if there are waiters then it will block, it does PI, etc. (Due to
1887 * races the kernel might see a 0 value of the futex too.)
1889 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1890 ktime_t *time, int trylock)
1892 struct hrtimer_sleeper timeout, *to = NULL;
1893 struct futex_hash_bucket *hb;
1894 struct futex_q q = futex_q_init;
1897 if (refill_pi_state_cache())
1902 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1904 hrtimer_init_sleeper(to, current);
1905 hrtimer_set_expires(&to->timer, *time);
1909 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key);
1910 if (unlikely(ret != 0))
1914 hb = queue_lock(&q);
1916 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1917 if (unlikely(ret)) {
1920 /* We got the lock. */
1922 goto out_unlock_put_key;
1927 * Task is exiting and we just wait for the
1930 queue_unlock(&q, hb);
1931 put_futex_key(&q.key);
1935 goto out_unlock_put_key;
1940 * Only actually queue now that the atomic ops are done:
1944 WARN_ON(!q.pi_state);
1946 * Block on the PI mutex:
1949 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1951 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1952 /* Fixup the trylock return value: */
1953 ret = ret ? 0 : -EWOULDBLOCK;
1956 spin_lock(q.lock_ptr);
1958 * Fixup the pi_state owner and possibly acquire the lock if we
1961 res = fixup_owner(uaddr, &q, !ret);
1963 * If fixup_owner() returned an error, proprogate that. If it acquired
1964 * the lock, clear our -ETIMEDOUT or -EINTR.
1967 ret = (res < 0) ? res : 0;
1970 * If fixup_owner() faulted and was unable to handle the fault, unlock
1971 * it and return the fault to userspace.
1973 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1974 rt_mutex_unlock(&q.pi_state->pi_mutex);
1976 /* Unqueue and drop the lock */
1982 queue_unlock(&q, hb);
1985 put_futex_key(&q.key);
1988 destroy_hrtimer_on_stack(&to->timer);
1989 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1992 queue_unlock(&q, hb);
1994 ret = fault_in_user_writeable(uaddr);
1998 if (!(flags & FLAGS_SHARED))
2001 put_futex_key(&q.key);
2006 * Userspace attempted a TID -> 0 atomic transition, and failed.
2007 * This is the in-kernel slowpath: we look up the PI state (if any),
2008 * and do the rt-mutex unlock.
2010 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2012 struct futex_hash_bucket *hb;
2013 struct futex_q *this, *next;
2015 struct plist_head *head;
2016 union futex_key key = FUTEX_KEY_INIT;
2020 if (get_user(uval, uaddr))
2023 * We release only a lock we actually own:
2025 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2028 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
2029 if (unlikely(ret != 0))
2032 hb = hash_futex(&key);
2033 spin_lock(&hb->lock);
2036 * To avoid races, try to do the TID -> 0 atomic transition
2037 * again. If it succeeds then we can return without waking
2040 if (!(uval & FUTEX_OWNER_DIED))
2041 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2044 if (unlikely(uval == -EFAULT))
2047 * Rare case: we managed to release the lock atomically,
2048 * no need to wake anyone else up:
2050 if (unlikely(uval == task_pid_vnr(current)))
2054 * Ok, other tasks may need to be woken up - check waiters
2055 * and do the wakeup if necessary:
2059 plist_for_each_entry_safe(this, next, head, list) {
2060 if (!match_futex (&this->key, &key))
2062 ret = wake_futex_pi(uaddr, uval, this);
2064 * The atomic access to the futex value
2065 * generated a pagefault, so retry the
2066 * user-access and the wakeup:
2073 * No waiters - kernel unlocks the futex:
2075 if (!(uval & FUTEX_OWNER_DIED)) {
2076 ret = unlock_futex_pi(uaddr, uval);
2082 spin_unlock(&hb->lock);
2083 put_futex_key(&key);
2089 spin_unlock(&hb->lock);
2090 put_futex_key(&key);
2092 ret = fault_in_user_writeable(uaddr);
2100 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2101 * @hb: the hash_bucket futex_q was original enqueued on
2102 * @q: the futex_q woken while waiting to be requeued
2103 * @key2: the futex_key of the requeue target futex
2104 * @timeout: the timeout associated with the wait (NULL if none)
2106 * Detect if the task was woken on the initial futex as opposed to the requeue
2107 * target futex. If so, determine if it was a timeout or a signal that caused
2108 * the wakeup and return the appropriate error code to the caller. Must be
2109 * called with the hb lock held.
2112 * 0 - no early wakeup detected
2113 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2116 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2117 struct futex_q *q, union futex_key *key2,
2118 struct hrtimer_sleeper *timeout)
2123 * With the hb lock held, we avoid races while we process the wakeup.
2124 * We only need to hold hb (and not hb2) to ensure atomicity as the
2125 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2126 * It can't be requeued from uaddr2 to something else since we don't
2127 * support a PI aware source futex for requeue.
2129 if (!match_futex(&q->key, key2)) {
2130 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2132 * We were woken prior to requeue by a timeout or a signal.
2133 * Unqueue the futex_q and determine which it was.
2135 plist_del(&q->list, &q->list.plist);
2137 /* Handle spurious wakeups gracefully */
2139 if (timeout && !timeout->task)
2141 else if (signal_pending(current))
2142 ret = -ERESTARTNOINTR;
2148 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2149 * @uaddr: the futex we initially wait on (non-pi)
2150 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2151 * the same type, no requeueing from private to shared, etc.
2152 * @val: the expected value of uaddr
2153 * @abs_time: absolute timeout
2154 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2155 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2156 * @uaddr2: the pi futex we will take prior to returning to user-space
2158 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2159 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2160 * complete the acquisition of the rt_mutex prior to returning to userspace.
2161 * This ensures the rt_mutex maintains an owner when it has waiters; without
2162 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2165 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2166 * via the following:
2167 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2168 * 2) wakeup on uaddr2 after a requeue
2172 * If 3, cleanup and return -ERESTARTNOINTR.
2174 * If 2, we may then block on trying to take the rt_mutex and return via:
2175 * 5) successful lock
2178 * 8) other lock acquisition failure
2180 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2182 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2188 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2189 u32 val, ktime_t *abs_time, u32 bitset,
2192 struct hrtimer_sleeper timeout, *to = NULL;
2193 struct rt_mutex_waiter rt_waiter;
2194 struct rt_mutex *pi_mutex = NULL;
2195 struct futex_hash_bucket *hb;
2196 union futex_key key2 = FUTEX_KEY_INIT;
2197 struct futex_q q = futex_q_init;
2205 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2206 CLOCK_REALTIME : CLOCK_MONOTONIC,
2208 hrtimer_init_sleeper(to, current);
2209 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2210 current->timer_slack_ns);
2214 * The waiter is allocated on our stack, manipulated by the requeue
2215 * code while we sleep on uaddr.
2217 debug_rt_mutex_init_waiter(&rt_waiter);
2218 rt_waiter.task = NULL;
2220 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
2221 if (unlikely(ret != 0))
2225 q.rt_waiter = &rt_waiter;
2226 q.requeue_pi_key = &key2;
2229 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2232 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2236 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2237 futex_wait_queue_me(hb, &q, to);
2239 spin_lock(&hb->lock);
2240 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2241 spin_unlock(&hb->lock);
2246 * In order for us to be here, we know our q.key == key2, and since
2247 * we took the hb->lock above, we also know that futex_requeue() has
2248 * completed and we no longer have to concern ourselves with a wakeup
2249 * race with the atomic proxy lock acquisition by the requeue code. The
2250 * futex_requeue dropped our key1 reference and incremented our key2
2254 /* Check if the requeue code acquired the second futex for us. */
2257 * Got the lock. We might not be the anticipated owner if we
2258 * did a lock-steal - fix up the PI-state in that case.
2260 if (q.pi_state && (q.pi_state->owner != current)) {
2261 spin_lock(q.lock_ptr);
2262 ret = fixup_pi_state_owner(uaddr2, &q, current);
2263 spin_unlock(q.lock_ptr);
2267 * We have been woken up by futex_unlock_pi(), a timeout, or a
2268 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2271 WARN_ON(!&q.pi_state);
2272 pi_mutex = &q.pi_state->pi_mutex;
2273 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2274 debug_rt_mutex_free_waiter(&rt_waiter);
2276 spin_lock(q.lock_ptr);
2278 * Fixup the pi_state owner and possibly acquire the lock if we
2281 res = fixup_owner(uaddr2, &q, !ret);
2283 * If fixup_owner() returned an error, proprogate that. If it
2284 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2287 ret = (res < 0) ? res : 0;
2289 /* Unqueue and drop the lock. */
2294 * If fixup_pi_state_owner() faulted and was unable to handle the
2295 * fault, unlock the rt_mutex and return the fault to userspace.
2297 if (ret == -EFAULT) {
2298 if (rt_mutex_owner(pi_mutex) == current)
2299 rt_mutex_unlock(pi_mutex);
2300 } else if (ret == -EINTR) {
2302 * We've already been requeued, but cannot restart by calling
2303 * futex_lock_pi() directly. We could restart this syscall, but
2304 * it would detect that the user space "val" changed and return
2305 * -EWOULDBLOCK. Save the overhead of the restart and return
2306 * -EWOULDBLOCK directly.
2312 put_futex_key(&q.key);
2314 put_futex_key(&key2);
2318 hrtimer_cancel(&to->timer);
2319 destroy_hrtimer_on_stack(&to->timer);
2325 * Support for robust futexes: the kernel cleans up held futexes at
2328 * Implementation: user-space maintains a per-thread list of locks it
2329 * is holding. Upon do_exit(), the kernel carefully walks this list,
2330 * and marks all locks that are owned by this thread with the
2331 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2332 * always manipulated with the lock held, so the list is private and
2333 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2334 * field, to allow the kernel to clean up if the thread dies after
2335 * acquiring the lock, but just before it could have added itself to
2336 * the list. There can only be one such pending lock.
2340 * sys_set_robust_list() - Set the robust-futex list head of a task
2341 * @head: pointer to the list-head
2342 * @len: length of the list-head, as userspace expects
2344 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2347 if (!futex_cmpxchg_enabled)
2350 * The kernel knows only one size for now:
2352 if (unlikely(len != sizeof(*head)))
2355 current->robust_list = head;
2361 * sys_get_robust_list() - Get the robust-futex list head of a task
2362 * @pid: pid of the process [zero for current task]
2363 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2364 * @len_ptr: pointer to a length field, the kernel fills in the header size
2366 SYSCALL_DEFINE3(get_robust_list, int, pid,
2367 struct robust_list_head __user * __user *, head_ptr,
2368 size_t __user *, len_ptr)
2370 struct robust_list_head __user *head;
2372 const struct cred *cred = current_cred(), *pcred;
2374 if (!futex_cmpxchg_enabled)
2378 head = current->robust_list;
2380 struct task_struct *p;
2384 p = find_task_by_vpid(pid);
2388 pcred = __task_cred(p);
2389 if (cred->euid != pcred->euid &&
2390 cred->euid != pcred->uid &&
2391 !capable(CAP_SYS_PTRACE))
2393 head = p->robust_list;
2397 if (put_user(sizeof(*head), len_ptr))
2399 return put_user(head, head_ptr);
2408 * Process a futex-list entry, check whether it's owned by the
2409 * dying task, and do notification if so:
2411 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2413 u32 uval, nval, mval;
2416 if (get_user(uval, uaddr))
2419 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2421 * Ok, this dying thread is truly holding a futex
2422 * of interest. Set the OWNER_DIED bit atomically
2423 * via cmpxchg, and if the value had FUTEX_WAITERS
2424 * set, wake up a waiter (if any). (We have to do a
2425 * futex_wake() even if OWNER_DIED is already set -
2426 * to handle the rare but possible case of recursive
2427 * thread-death.) The rest of the cleanup is done in
2430 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2431 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2433 if (nval == -EFAULT)
2440 * Wake robust non-PI futexes here. The wakeup of
2441 * PI futexes happens in exit_pi_state():
2443 if (!pi && (uval & FUTEX_WAITERS))
2444 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2450 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2452 static inline int fetch_robust_entry(struct robust_list __user **entry,
2453 struct robust_list __user * __user *head,
2456 unsigned long uentry;
2458 if (get_user(uentry, (unsigned long __user *)head))
2461 *entry = (void __user *)(uentry & ~1UL);
2468 * Walk curr->robust_list (very carefully, it's a userspace list!)
2469 * and mark any locks found there dead, and notify any waiters.
2471 * We silently return on any sign of list-walking problem.
2473 void exit_robust_list(struct task_struct *curr)
2475 struct robust_list_head __user *head = curr->robust_list;
2476 struct robust_list __user *entry, *next_entry, *pending;
2477 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2478 unsigned int uninitialized_var(next_pi);
2479 unsigned long futex_offset;
2482 if (!futex_cmpxchg_enabled)
2486 * Fetch the list head (which was registered earlier, via
2487 * sys_set_robust_list()):
2489 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2492 * Fetch the relative futex offset:
2494 if (get_user(futex_offset, &head->futex_offset))
2497 * Fetch any possibly pending lock-add first, and handle it
2500 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2503 next_entry = NULL; /* avoid warning with gcc */
2504 while (entry != &head->list) {
2506 * Fetch the next entry in the list before calling
2507 * handle_futex_death:
2509 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2511 * A pending lock might already be on the list, so
2512 * don't process it twice:
2514 if (entry != pending)
2515 if (handle_futex_death((void __user *)entry + futex_offset,
2523 * Avoid excessively long or circular lists:
2532 handle_futex_death((void __user *)pending + futex_offset,
2536 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2537 u32 __user *uaddr2, u32 val2, u32 val3)
2539 int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2540 unsigned int flags = 0;
2542 if (!(op & FUTEX_PRIVATE_FLAG))
2543 flags |= FLAGS_SHARED;
2545 if (op & FUTEX_CLOCK_REALTIME) {
2546 flags |= FLAGS_CLOCKRT;
2547 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2553 val3 = FUTEX_BITSET_MATCH_ANY;
2554 case FUTEX_WAIT_BITSET:
2555 ret = futex_wait(uaddr, flags, val, timeout, val3);
2558 val3 = FUTEX_BITSET_MATCH_ANY;
2559 case FUTEX_WAKE_BITSET:
2560 ret = futex_wake(uaddr, flags, val, val3);
2563 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2565 case FUTEX_CMP_REQUEUE:
2566 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2569 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2572 if (futex_cmpxchg_enabled)
2573 ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2575 case FUTEX_UNLOCK_PI:
2576 if (futex_cmpxchg_enabled)
2577 ret = futex_unlock_pi(uaddr, flags);
2579 case FUTEX_TRYLOCK_PI:
2580 if (futex_cmpxchg_enabled)
2581 ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2583 case FUTEX_WAIT_REQUEUE_PI:
2584 val3 = FUTEX_BITSET_MATCH_ANY;
2585 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2588 case FUTEX_CMP_REQUEUE_PI:
2589 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2598 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2599 struct timespec __user *, utime, u32 __user *, uaddr2,
2603 ktime_t t, *tp = NULL;
2605 int cmd = op & FUTEX_CMD_MASK;
2607 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2608 cmd == FUTEX_WAIT_BITSET ||
2609 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2610 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2612 if (!timespec_valid(&ts))
2615 t = timespec_to_ktime(ts);
2616 if (cmd == FUTEX_WAIT)
2617 t = ktime_add_safe(ktime_get(), t);
2621 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2622 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2624 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2625 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2626 val2 = (u32) (unsigned long) utime;
2628 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2631 static int __init futex_init(void)
2637 * This will fail and we want it. Some arch implementations do
2638 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2639 * functionality. We want to know that before we call in any
2640 * of the complex code paths. Also we want to prevent
2641 * registration of robust lists in that case. NULL is
2642 * guaranteed to fault and we get -EFAULT on functional
2643 * implementation, the non-functional ones will return
2646 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2647 if (curval == -EFAULT)
2648 futex_cmpxchg_enabled = 1;
2650 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2651 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2652 spin_lock_init(&futex_queues[i].lock);
2657 __initcall(futex_init);