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 * Priority Inheritance state:
74 struct futex_pi_state {
76 * list of 'owned' pi_state instances - these have to be
77 * cleaned up in do_exit() if the task exits prematurely:
79 struct list_head list;
84 struct rt_mutex pi_mutex;
86 struct task_struct *owner;
93 * We use this hashed waitqueue instead of a normal wait_queue_t, so
94 * we can wake only the relevant ones (hashed queues may be shared).
96 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
97 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
98 * The order of wakup is always to make the first condition true, then
99 * wake up q->waiter, then make the second condition true.
102 struct plist_node list;
103 /* Waiter reference */
104 struct task_struct *task;
106 /* Which hash list lock to use: */
107 spinlock_t *lock_ptr;
109 /* Key which the futex is hashed on: */
112 /* Optional priority inheritance state: */
113 struct futex_pi_state *pi_state;
115 /* rt_waiter storage for requeue_pi: */
116 struct rt_mutex_waiter *rt_waiter;
118 /* Bitset for the optional bitmasked wakeup */
123 * Hash buckets are shared by all the futex_keys that hash to the same
124 * location. Each key may have multiple futex_q structures, one for each task
125 * waiting on a futex.
127 struct futex_hash_bucket {
129 struct plist_head chain;
132 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
135 * We hash on the keys returned from get_futex_key (see below).
137 static struct futex_hash_bucket *hash_futex(union futex_key *key)
139 u32 hash = jhash2((u32*)&key->both.word,
140 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
142 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146 * Return 1 if two futex_keys are equal, 0 otherwise.
148 static inline int match_futex(union futex_key *key1, union futex_key *key2)
150 return (key1->both.word == key2->both.word
151 && key1->both.ptr == key2->both.ptr
152 && key1->both.offset == key2->both.offset);
156 * Take a reference to the resource addressed by a key.
157 * Can be called while holding spinlocks.
160 static void get_futex_key_refs(union futex_key *key)
165 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
167 atomic_inc(&key->shared.inode->i_count);
169 case FUT_OFF_MMSHARED:
170 atomic_inc(&key->private.mm->mm_count);
176 * Drop a reference to the resource addressed by a key.
177 * The hash bucket spinlock must not be held.
179 static void drop_futex_key_refs(union futex_key *key)
181 if (!key->both.ptr) {
182 /* If we're here then we tried to put a key we failed to get */
187 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
189 iput(key->shared.inode);
191 case FUT_OFF_MMSHARED:
192 mmdrop(key->private.mm);
198 * get_futex_key - Get parameters which are the keys for a futex.
199 * @uaddr: virtual address of the futex
200 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
201 * @key: address where result is stored.
203 * Returns a negative error code or 0
204 * The key words are stored in *key on success.
206 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
207 * offset_within_page). For private mappings, it's (uaddr, current->mm).
208 * We can usually work out the index without swapping in the page.
210 * lock_page() might sleep, the caller should not hold a spinlock.
212 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
214 unsigned long address = (unsigned long)uaddr;
215 struct mm_struct *mm = current->mm;
220 * The futex address must be "naturally" aligned.
222 key->both.offset = address % PAGE_SIZE;
223 if (unlikely((address % sizeof(u32)) != 0))
225 address -= key->both.offset;
228 * PROCESS_PRIVATE futexes are fast.
229 * As the mm cannot disappear under us and the 'key' only needs
230 * virtual address, we dont even have to find the underlying vma.
231 * Note : We do have to check 'uaddr' is a valid user address,
232 * but access_ok() should be faster than find_vma()
235 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
237 key->private.mm = mm;
238 key->private.address = address;
239 get_futex_key_refs(key);
244 err = get_user_pages_fast(address, 1, 0, &page);
249 if (!page->mapping) {
256 * Private mappings are handled in a simple way.
258 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
259 * it's a read-only handle, it's expected that futexes attach to
260 * the object not the particular process.
262 if (PageAnon(page)) {
263 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
264 key->private.mm = mm;
265 key->private.address = address;
267 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
268 key->shared.inode = page->mapping->host;
269 key->shared.pgoff = page->index;
272 get_futex_key_refs(key);
280 void put_futex_key(int fshared, union futex_key *key)
282 drop_futex_key_refs(key);
286 * futex_top_waiter() - Return the highest priority waiter on a futex
287 * @hb: the hash bucket the futex_q's reside in
288 * @key: the futex key (to distinguish it from other futex futex_q's)
290 * Must be called with the hb lock held.
292 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
293 union futex_key *key)
295 struct futex_q *this;
297 plist_for_each_entry(this, &hb->chain, list) {
298 if (match_futex(&this->key, key))
304 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
309 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
315 static int get_futex_value_locked(u32 *dest, u32 __user *from)
320 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
323 return ret ? -EFAULT : 0;
330 static int refill_pi_state_cache(void)
332 struct futex_pi_state *pi_state;
334 if (likely(current->pi_state_cache))
337 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
342 INIT_LIST_HEAD(&pi_state->list);
343 /* pi_mutex gets initialized later */
344 pi_state->owner = NULL;
345 atomic_set(&pi_state->refcount, 1);
346 pi_state->key = FUTEX_KEY_INIT;
348 current->pi_state_cache = pi_state;
353 static struct futex_pi_state * alloc_pi_state(void)
355 struct futex_pi_state *pi_state = current->pi_state_cache;
358 current->pi_state_cache = NULL;
363 static void free_pi_state(struct futex_pi_state *pi_state)
365 if (!atomic_dec_and_test(&pi_state->refcount))
369 * If pi_state->owner is NULL, the owner is most probably dying
370 * and has cleaned up the pi_state already
372 if (pi_state->owner) {
373 spin_lock_irq(&pi_state->owner->pi_lock);
374 list_del_init(&pi_state->list);
375 spin_unlock_irq(&pi_state->owner->pi_lock);
377 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
380 if (current->pi_state_cache)
384 * pi_state->list is already empty.
385 * clear pi_state->owner.
386 * refcount is at 0 - put it back to 1.
388 pi_state->owner = NULL;
389 atomic_set(&pi_state->refcount, 1);
390 current->pi_state_cache = pi_state;
395 * Look up the task based on what TID userspace gave us.
398 static struct task_struct * futex_find_get_task(pid_t pid)
400 struct task_struct *p;
401 const struct cred *cred = current_cred(), *pcred;
404 p = find_task_by_vpid(pid);
408 pcred = __task_cred(p);
409 if (cred->euid != pcred->euid &&
410 cred->euid != pcred->uid)
422 * This task is holding PI mutexes at exit time => bad.
423 * Kernel cleans up PI-state, but userspace is likely hosed.
424 * (Robust-futex cleanup is separate and might save the day for userspace.)
426 void exit_pi_state_list(struct task_struct *curr)
428 struct list_head *next, *head = &curr->pi_state_list;
429 struct futex_pi_state *pi_state;
430 struct futex_hash_bucket *hb;
431 union futex_key key = FUTEX_KEY_INIT;
433 if (!futex_cmpxchg_enabled)
436 * We are a ZOMBIE and nobody can enqueue itself on
437 * pi_state_list anymore, but we have to be careful
438 * versus waiters unqueueing themselves:
440 spin_lock_irq(&curr->pi_lock);
441 while (!list_empty(head)) {
444 pi_state = list_entry(next, struct futex_pi_state, list);
446 hb = hash_futex(&key);
447 spin_unlock_irq(&curr->pi_lock);
449 spin_lock(&hb->lock);
451 spin_lock_irq(&curr->pi_lock);
453 * We dropped the pi-lock, so re-check whether this
454 * task still owns the PI-state:
456 if (head->next != next) {
457 spin_unlock(&hb->lock);
461 WARN_ON(pi_state->owner != curr);
462 WARN_ON(list_empty(&pi_state->list));
463 list_del_init(&pi_state->list);
464 pi_state->owner = NULL;
465 spin_unlock_irq(&curr->pi_lock);
467 rt_mutex_unlock(&pi_state->pi_mutex);
469 spin_unlock(&hb->lock);
471 spin_lock_irq(&curr->pi_lock);
473 spin_unlock_irq(&curr->pi_lock);
477 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
478 union futex_key *key, struct futex_pi_state **ps)
480 struct futex_pi_state *pi_state = NULL;
481 struct futex_q *this, *next;
482 struct plist_head *head;
483 struct task_struct *p;
484 pid_t pid = uval & FUTEX_TID_MASK;
488 plist_for_each_entry_safe(this, next, head, list) {
489 if (match_futex(&this->key, key)) {
491 * Another waiter already exists - bump up
492 * the refcount and return its pi_state:
494 pi_state = this->pi_state;
496 * Userspace might have messed up non PI and PI futexes
498 if (unlikely(!pi_state))
501 WARN_ON(!atomic_read(&pi_state->refcount));
502 WARN_ON(pid && pi_state->owner &&
503 pi_state->owner->pid != pid);
505 atomic_inc(&pi_state->refcount);
513 * We are the first waiter - try to look up the real owner and attach
514 * the new pi_state to it, but bail out when TID = 0
518 p = futex_find_get_task(pid);
523 * We need to look at the task state flags to figure out,
524 * whether the task is exiting. To protect against the do_exit
525 * change of the task flags, we do this protected by
528 spin_lock_irq(&p->pi_lock);
529 if (unlikely(p->flags & PF_EXITING)) {
531 * The task is on the way out. When PF_EXITPIDONE is
532 * set, we know that the task has finished the
535 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
537 spin_unlock_irq(&p->pi_lock);
542 pi_state = alloc_pi_state();
545 * Initialize the pi_mutex in locked state and make 'p'
548 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
550 /* Store the key for possible exit cleanups: */
551 pi_state->key = *key;
553 WARN_ON(!list_empty(&pi_state->list));
554 list_add(&pi_state->list, &p->pi_state_list);
556 spin_unlock_irq(&p->pi_lock);
566 * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
567 * @uaddr: the pi futex user address
568 * @hb: the pi futex hash bucket
569 * @key: the futex key associated with uaddr and hb
570 * @ps: the pi_state pointer where we store the result of the
572 * @task: the task to perform the atomic lock work for. This will
573 * be "current" except in the case of requeue pi.
574 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
578 * 1 - acquired the lock
581 * The hb->lock and futex_key refs shall be held by the caller.
583 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
584 union futex_key *key,
585 struct futex_pi_state **ps,
586 struct task_struct *task, int set_waiters)
588 int lock_taken, ret, ownerdied = 0;
589 u32 uval, newval, curval;
592 ret = lock_taken = 0;
595 * To avoid races, we attempt to take the lock here again
596 * (by doing a 0 -> TID atomic cmpxchg), while holding all
597 * the locks. It will most likely not succeed.
599 newval = task_pid_vnr(task);
601 newval |= FUTEX_WAITERS;
603 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
605 if (unlikely(curval == -EFAULT))
611 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
615 * Surprise - we got the lock. Just return to userspace:
617 if (unlikely(!curval))
623 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
624 * to wake at the next unlock.
626 newval = curval | FUTEX_WAITERS;
629 * There are two cases, where a futex might have no owner (the
630 * owner TID is 0): OWNER_DIED. We take over the futex in this
631 * case. We also do an unconditional take over, when the owner
634 * This is safe as we are protected by the hash bucket lock !
636 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
637 /* Keep the OWNER_DIED bit */
638 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
643 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
645 if (unlikely(curval == -EFAULT))
647 if (unlikely(curval != uval))
651 * We took the lock due to owner died take over.
653 if (unlikely(lock_taken))
657 * We dont have the lock. Look up the PI state (or create it if
658 * we are the first waiter):
660 ret = lookup_pi_state(uval, hb, key, ps);
666 * No owner found for this futex. Check if the
667 * OWNER_DIED bit is set to figure out whether
668 * this is a robust futex or not.
670 if (get_futex_value_locked(&curval, uaddr))
674 * We simply start over in case of a robust
675 * futex. The code above will take the futex
678 if (curval & FUTEX_OWNER_DIED) {
691 * The hash bucket lock must be held when this is called.
692 * Afterwards, the futex_q must not be accessed.
694 static void wake_futex(struct futex_q *q)
696 struct task_struct *p = q->task;
699 * We set q->lock_ptr = NULL _before_ we wake up the task. If
700 * a non futex wake up happens on another CPU then the task
701 * might exit and p would dereference a non existing task
702 * struct. Prevent this by holding a reference on p across the
707 plist_del(&q->list, &q->list.plist);
709 * The waiting task can free the futex_q as soon as
710 * q->lock_ptr = NULL is written, without taking any locks. A
711 * memory barrier is required here to prevent the following
712 * store to lock_ptr from getting ahead of the plist_del.
717 wake_up_state(p, TASK_NORMAL);
721 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
723 struct task_struct *new_owner;
724 struct futex_pi_state *pi_state = this->pi_state;
730 spin_lock(&pi_state->pi_mutex.wait_lock);
731 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
734 * This happens when we have stolen the lock and the original
735 * pending owner did not enqueue itself back on the rt_mutex.
736 * Thats not a tragedy. We know that way, that a lock waiter
737 * is on the fly. We make the futex_q waiter the pending owner.
740 new_owner = this->task;
743 * We pass it to the next owner. (The WAITERS bit is always
744 * kept enabled while there is PI state around. We must also
745 * preserve the owner died bit.)
747 if (!(uval & FUTEX_OWNER_DIED)) {
750 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
752 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
754 if (curval == -EFAULT)
756 else if (curval != uval)
759 spin_unlock(&pi_state->pi_mutex.wait_lock);
764 spin_lock_irq(&pi_state->owner->pi_lock);
765 WARN_ON(list_empty(&pi_state->list));
766 list_del_init(&pi_state->list);
767 spin_unlock_irq(&pi_state->owner->pi_lock);
769 spin_lock_irq(&new_owner->pi_lock);
770 WARN_ON(!list_empty(&pi_state->list));
771 list_add(&pi_state->list, &new_owner->pi_state_list);
772 pi_state->owner = new_owner;
773 spin_unlock_irq(&new_owner->pi_lock);
775 spin_unlock(&pi_state->pi_mutex.wait_lock);
776 rt_mutex_unlock(&pi_state->pi_mutex);
781 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
786 * There is no waiter, so we unlock the futex. The owner died
787 * bit has not to be preserved here. We are the owner:
789 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
791 if (oldval == -EFAULT)
800 * Express the locking dependencies for lockdep:
803 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
806 spin_lock(&hb1->lock);
808 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
809 } else { /* hb1 > hb2 */
810 spin_lock(&hb2->lock);
811 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
816 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
818 spin_unlock(&hb1->lock);
820 spin_unlock(&hb2->lock);
824 * Wake up waiters matching bitset queued on this futex (uaddr).
826 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
828 struct futex_hash_bucket *hb;
829 struct futex_q *this, *next;
830 struct plist_head *head;
831 union futex_key key = FUTEX_KEY_INIT;
837 ret = get_futex_key(uaddr, fshared, &key);
838 if (unlikely(ret != 0))
841 hb = hash_futex(&key);
842 spin_lock(&hb->lock);
845 plist_for_each_entry_safe(this, next, head, list) {
846 if (match_futex (&this->key, &key)) {
847 if (this->pi_state || this->rt_waiter) {
852 /* Check if one of the bits is set in both bitsets */
853 if (!(this->bitset & bitset))
857 if (++ret >= nr_wake)
862 spin_unlock(&hb->lock);
863 put_futex_key(fshared, &key);
869 * Wake up all waiters hashed on the physical page that is mapped
870 * to this virtual address:
873 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
874 int nr_wake, int nr_wake2, int op)
876 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
877 struct futex_hash_bucket *hb1, *hb2;
878 struct plist_head *head;
879 struct futex_q *this, *next;
883 ret = get_futex_key(uaddr1, fshared, &key1);
884 if (unlikely(ret != 0))
886 ret = get_futex_key(uaddr2, fshared, &key2);
887 if (unlikely(ret != 0))
890 hb1 = hash_futex(&key1);
891 hb2 = hash_futex(&key2);
893 double_lock_hb(hb1, hb2);
895 op_ret = futex_atomic_op_inuser(op, uaddr2);
896 if (unlikely(op_ret < 0)) {
899 double_unlock_hb(hb1, hb2);
903 * we don't get EFAULT from MMU faults if we don't have an MMU,
904 * but we might get them from range checking
910 if (unlikely(op_ret != -EFAULT)) {
915 ret = get_user(dummy, uaddr2);
922 put_futex_key(fshared, &key2);
923 put_futex_key(fshared, &key1);
929 plist_for_each_entry_safe(this, next, head, list) {
930 if (match_futex (&this->key, &key1)) {
932 if (++ret >= nr_wake)
941 plist_for_each_entry_safe(this, next, head, list) {
942 if (match_futex (&this->key, &key2)) {
944 if (++op_ret >= nr_wake2)
951 double_unlock_hb(hb1, hb2);
953 put_futex_key(fshared, &key2);
955 put_futex_key(fshared, &key1);
961 * requeue_futex() - Requeue a futex_q from one hb to another
962 * @q: the futex_q to requeue
963 * @hb1: the source hash_bucket
964 * @hb2: the target hash_bucket
965 * @key2: the new key for the requeued futex_q
968 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
969 struct futex_hash_bucket *hb2, union futex_key *key2)
973 * If key1 and key2 hash to the same bucket, no need to
976 if (likely(&hb1->chain != &hb2->chain)) {
977 plist_del(&q->list, &hb1->chain);
978 plist_add(&q->list, &hb2->chain);
979 q->lock_ptr = &hb2->lock;
980 #ifdef CONFIG_DEBUG_PI_LIST
981 q->list.plist.lock = &hb2->lock;
984 get_futex_key_refs(key2);
989 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
991 * key: the key of the requeue target futex
993 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
994 * target futex if it is uncontended or via a lock steal. Set the futex_q key
995 * to the requeue target futex so the waiter can detect the wakeup on the right
996 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
997 * atomic lock acquisition. Must be called with the q->lock_ptr held.
1000 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key)
1002 drop_futex_key_refs(&q->key);
1003 get_futex_key_refs(key);
1006 WARN_ON(plist_node_empty(&q->list));
1007 plist_del(&q->list, &q->list.plist);
1009 WARN_ON(!q->rt_waiter);
1010 q->rt_waiter = NULL;
1012 wake_up_state(q->task, TASK_NORMAL);
1016 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1017 * @pifutex: the user address of the to futex
1018 * @hb1: the from futex hash bucket, must be locked by the caller
1019 * @hb2: the to futex hash bucket, must be locked by the caller
1020 * @key1: the from futex key
1021 * @key2: the to futex key
1022 * @ps: address to store the pi_state pointer
1023 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1025 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1026 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1027 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1028 * hb1 and hb2 must be held by the caller.
1031 * 0 - failed to acquire the lock atomicly
1032 * 1 - acquired the lock
1035 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1036 struct futex_hash_bucket *hb1,
1037 struct futex_hash_bucket *hb2,
1038 union futex_key *key1, union futex_key *key2,
1039 struct futex_pi_state **ps, int set_waiters)
1041 struct futex_q *top_waiter = NULL;
1045 if (get_futex_value_locked(&curval, pifutex))
1049 * Find the top_waiter and determine if there are additional waiters.
1050 * If the caller intends to requeue more than 1 waiter to pifutex,
1051 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1052 * as we have means to handle the possible fault. If not, don't set
1053 * the bit unecessarily as it will force the subsequent unlock to enter
1056 top_waiter = futex_top_waiter(hb1, key1);
1058 /* There are no waiters, nothing for us to do. */
1063 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1064 * the contended case or if set_waiters is 1. The pi_state is returned
1065 * in ps in contended cases.
1067 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1070 requeue_pi_wake_futex(top_waiter, key2);
1076 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1077 * uaddr1: source futex user address
1078 * uaddr2: target futex user address
1079 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1080 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1081 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1082 * pi futex (pi to pi requeue is not supported)
1084 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1085 * uaddr2 atomically on behalf of the top waiter.
1088 * >=0 - on success, the number of tasks requeued or woken
1091 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1092 int nr_wake, int nr_requeue, u32 *cmpval,
1095 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1096 int drop_count = 0, task_count = 0, ret;
1097 struct futex_pi_state *pi_state = NULL;
1098 struct futex_hash_bucket *hb1, *hb2;
1099 struct plist_head *head1;
1100 struct futex_q *this, *next;
1105 * requeue_pi requires a pi_state, try to allocate it now
1106 * without any locks in case it fails.
1108 if (refill_pi_state_cache())
1111 * requeue_pi must wake as many tasks as it can, up to nr_wake
1112 * + nr_requeue, since it acquires the rt_mutex prior to
1113 * returning to userspace, so as to not leave the rt_mutex with
1114 * waiters and no owner. However, second and third wake-ups
1115 * cannot be predicted as they involve race conditions with the
1116 * first wake and a fault while looking up the pi_state. Both
1117 * pthread_cond_signal() and pthread_cond_broadcast() should
1125 if (pi_state != NULL) {
1127 * We will have to lookup the pi_state again, so free this one
1128 * to keep the accounting correct.
1130 free_pi_state(pi_state);
1134 ret = get_futex_key(uaddr1, fshared, &key1);
1135 if (unlikely(ret != 0))
1137 ret = get_futex_key(uaddr2, fshared, &key2);
1138 if (unlikely(ret != 0))
1141 hb1 = hash_futex(&key1);
1142 hb2 = hash_futex(&key2);
1145 double_lock_hb(hb1, hb2);
1147 if (likely(cmpval != NULL)) {
1150 ret = get_futex_value_locked(&curval, uaddr1);
1152 if (unlikely(ret)) {
1153 double_unlock_hb(hb1, hb2);
1155 ret = get_user(curval, uaddr1);
1162 put_futex_key(fshared, &key2);
1163 put_futex_key(fshared, &key1);
1166 if (curval != *cmpval) {
1172 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1174 * Attempt to acquire uaddr2 and wake the top waiter. If we
1175 * intend to requeue waiters, force setting the FUTEX_WAITERS
1176 * bit. We force this here where we are able to easily handle
1177 * faults rather in the requeue loop below.
1179 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1180 &key2, &pi_state, nr_requeue);
1183 * At this point the top_waiter has either taken uaddr2 or is
1184 * waiting on it. If the former, then the pi_state will not
1185 * exist yet, look it up one more time to ensure we have a
1191 ret = get_futex_value_locked(&curval2, uaddr2);
1193 ret = lookup_pi_state(curval2, hb2, &key2,
1201 double_unlock_hb(hb1, hb2);
1202 put_futex_key(fshared, &key2);
1203 put_futex_key(fshared, &key1);
1204 ret = get_user(curval2, uaddr2);
1209 /* The owner was exiting, try again. */
1210 double_unlock_hb(hb1, hb2);
1211 put_futex_key(fshared, &key2);
1212 put_futex_key(fshared, &key1);
1220 head1 = &hb1->chain;
1221 plist_for_each_entry_safe(this, next, head1, list) {
1222 if (task_count - nr_wake >= nr_requeue)
1225 if (!match_futex(&this->key, &key1))
1228 WARN_ON(!requeue_pi && this->rt_waiter);
1229 WARN_ON(requeue_pi && !this->rt_waiter);
1232 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1233 * lock, we already woke the top_waiter. If not, it will be
1234 * woken by futex_unlock_pi().
1236 if (++task_count <= nr_wake && !requeue_pi) {
1242 * Requeue nr_requeue waiters and possibly one more in the case
1243 * of requeue_pi if we couldn't acquire the lock atomically.
1246 /* Prepare the waiter to take the rt_mutex. */
1247 atomic_inc(&pi_state->refcount);
1248 this->pi_state = pi_state;
1249 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1253 /* We got the lock. */
1254 requeue_pi_wake_futex(this, &key2);
1258 this->pi_state = NULL;
1259 free_pi_state(pi_state);
1263 requeue_futex(this, hb1, hb2, &key2);
1268 double_unlock_hb(hb1, hb2);
1270 /* drop_futex_key_refs() must be called outside the spinlocks. */
1271 while (--drop_count >= 0)
1272 drop_futex_key_refs(&key1);
1275 put_futex_key(fshared, &key2);
1277 put_futex_key(fshared, &key1);
1279 if (pi_state != NULL)
1280 free_pi_state(pi_state);
1281 return ret ? ret : task_count;
1284 /* The key must be already stored in q->key. */
1285 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1287 struct futex_hash_bucket *hb;
1289 get_futex_key_refs(&q->key);
1290 hb = hash_futex(&q->key);
1291 q->lock_ptr = &hb->lock;
1293 spin_lock(&hb->lock);
1297 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1302 * The priority used to register this element is
1303 * - either the real thread-priority for the real-time threads
1304 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1305 * - or MAX_RT_PRIO for non-RT threads.
1306 * Thus, all RT-threads are woken first in priority order, and
1307 * the others are woken last, in FIFO order.
1309 prio = min(current->normal_prio, MAX_RT_PRIO);
1311 plist_node_init(&q->list, prio);
1312 #ifdef CONFIG_DEBUG_PI_LIST
1313 q->list.plist.lock = &hb->lock;
1315 plist_add(&q->list, &hb->chain);
1317 spin_unlock(&hb->lock);
1321 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1323 spin_unlock(&hb->lock);
1324 drop_futex_key_refs(&q->key);
1328 * queue_me and unqueue_me must be called as a pair, each
1329 * exactly once. They are called with the hashed spinlock held.
1332 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1333 static int unqueue_me(struct futex_q *q)
1335 spinlock_t *lock_ptr;
1338 /* In the common case we don't take the spinlock, which is nice. */
1340 lock_ptr = q->lock_ptr;
1342 if (lock_ptr != NULL) {
1343 spin_lock(lock_ptr);
1345 * q->lock_ptr can change between reading it and
1346 * spin_lock(), causing us to take the wrong lock. This
1347 * corrects the race condition.
1349 * Reasoning goes like this: if we have the wrong lock,
1350 * q->lock_ptr must have changed (maybe several times)
1351 * between reading it and the spin_lock(). It can
1352 * change again after the spin_lock() but only if it was
1353 * already changed before the spin_lock(). It cannot,
1354 * however, change back to the original value. Therefore
1355 * we can detect whether we acquired the correct lock.
1357 if (unlikely(lock_ptr != q->lock_ptr)) {
1358 spin_unlock(lock_ptr);
1361 WARN_ON(plist_node_empty(&q->list));
1362 plist_del(&q->list, &q->list.plist);
1364 BUG_ON(q->pi_state);
1366 spin_unlock(lock_ptr);
1370 drop_futex_key_refs(&q->key);
1375 * PI futexes can not be requeued and must remove themself from the
1376 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1379 static void unqueue_me_pi(struct futex_q *q)
1381 WARN_ON(plist_node_empty(&q->list));
1382 plist_del(&q->list, &q->list.plist);
1384 BUG_ON(!q->pi_state);
1385 free_pi_state(q->pi_state);
1388 spin_unlock(q->lock_ptr);
1390 drop_futex_key_refs(&q->key);
1394 * Fixup the pi_state owner with the new owner.
1396 * Must be called with hash bucket lock held and mm->sem held for non
1399 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1400 struct task_struct *newowner, int fshared)
1402 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1403 struct futex_pi_state *pi_state = q->pi_state;
1404 struct task_struct *oldowner = pi_state->owner;
1405 u32 uval, curval, newval;
1409 if (!pi_state->owner)
1410 newtid |= FUTEX_OWNER_DIED;
1413 * We are here either because we stole the rtmutex from the
1414 * pending owner or we are the pending owner which failed to
1415 * get the rtmutex. We have to replace the pending owner TID
1416 * in the user space variable. This must be atomic as we have
1417 * to preserve the owner died bit here.
1419 * Note: We write the user space value _before_ changing the pi_state
1420 * because we can fault here. Imagine swapped out pages or a fork
1421 * that marked all the anonymous memory readonly for cow.
1423 * Modifying pi_state _before_ the user space value would
1424 * leave the pi_state in an inconsistent state when we fault
1425 * here, because we need to drop the hash bucket lock to
1426 * handle the fault. This might be observed in the PID check
1427 * in lookup_pi_state.
1430 if (get_futex_value_locked(&uval, uaddr))
1434 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1436 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1438 if (curval == -EFAULT)
1446 * We fixed up user space. Now we need to fix the pi_state
1449 if (pi_state->owner != NULL) {
1450 spin_lock_irq(&pi_state->owner->pi_lock);
1451 WARN_ON(list_empty(&pi_state->list));
1452 list_del_init(&pi_state->list);
1453 spin_unlock_irq(&pi_state->owner->pi_lock);
1456 pi_state->owner = newowner;
1458 spin_lock_irq(&newowner->pi_lock);
1459 WARN_ON(!list_empty(&pi_state->list));
1460 list_add(&pi_state->list, &newowner->pi_state_list);
1461 spin_unlock_irq(&newowner->pi_lock);
1465 * To handle the page fault we need to drop the hash bucket
1466 * lock here. That gives the other task (either the pending
1467 * owner itself or the task which stole the rtmutex) the
1468 * chance to try the fixup of the pi_state. So once we are
1469 * back from handling the fault we need to check the pi_state
1470 * after reacquiring the hash bucket lock and before trying to
1471 * do another fixup. When the fixup has been done already we
1475 spin_unlock(q->lock_ptr);
1477 ret = get_user(uval, uaddr);
1479 spin_lock(q->lock_ptr);
1482 * Check if someone else fixed it for us:
1484 if (pi_state->owner != oldowner)
1494 * In case we must use restart_block to restart a futex_wait,
1495 * we encode in the 'flags' shared capability
1497 #define FLAGS_SHARED 0x01
1498 #define FLAGS_CLOCKRT 0x02
1499 #define FLAGS_HAS_TIMEOUT 0x04
1501 static long futex_wait_restart(struct restart_block *restart);
1502 static long futex_lock_pi_restart(struct restart_block *restart);
1505 * fixup_owner() - Post lock pi_state and corner case management
1506 * @uaddr: user address of the futex
1507 * @fshared: whether the futex is shared (1) or not (0)
1508 * @q: futex_q (contains pi_state and access to the rt_mutex)
1509 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1511 * After attempting to lock an rt_mutex, this function is called to cleanup
1512 * the pi_state owner as well as handle race conditions that may allow us to
1513 * acquire the lock. Must be called with the hb lock held.
1516 * 1 - success, lock taken
1517 * 0 - success, lock not taken
1518 * <0 - on error (-EFAULT)
1520 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1523 struct task_struct *owner;
1528 * Got the lock. We might not be the anticipated owner if we
1529 * did a lock-steal - fix up the PI-state in that case:
1531 if (q->pi_state->owner != current)
1532 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1537 * Catch the rare case, where the lock was released when we were on the
1538 * way back before we locked the hash bucket.
1540 if (q->pi_state->owner == current) {
1542 * Try to get the rt_mutex now. This might fail as some other
1543 * task acquired the rt_mutex after we removed ourself from the
1544 * rt_mutex waiters list.
1546 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1552 * pi_state is incorrect, some other task did a lock steal and
1553 * we returned due to timeout or signal without taking the
1554 * rt_mutex. Too late. We can access the rt_mutex_owner without
1555 * locking, as the other task is now blocked on the hash bucket
1556 * lock. Fix the state up.
1558 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1559 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1564 * Paranoia check. If we did not take the lock, then we should not be
1565 * the owner, nor the pending owner, of the rt_mutex.
1567 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1568 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1569 "pi-state %p\n", ret,
1570 q->pi_state->pi_mutex.owner,
1571 q->pi_state->owner);
1574 return ret ? ret : locked;
1578 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1579 * @hb: the futex hash bucket, must be locked by the caller
1580 * @q: the futex_q to queue up on
1581 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1583 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1584 struct hrtimer_sleeper *timeout)
1589 * There might have been scheduling since the queue_me(), as we
1590 * cannot hold a spinlock across the get_user() in case it
1591 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1592 * queueing ourselves into the futex hash. This code thus has to
1593 * rely on the futex_wake() code removing us from hash when it
1596 set_current_state(TASK_INTERRUPTIBLE);
1600 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1601 if (!hrtimer_active(&timeout->timer))
1602 timeout->task = NULL;
1606 * !plist_node_empty() is safe here without any lock.
1607 * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1609 if (likely(!plist_node_empty(&q->list))) {
1611 * If the timer has already expired, current will already be
1612 * flagged for rescheduling. Only call schedule if there
1613 * is no timeout, or if it has yet to expire.
1615 if (!timeout || timeout->task)
1618 __set_current_state(TASK_RUNNING);
1622 * futex_wait_setup() - Prepare to wait on a futex
1623 * @uaddr: the futex userspace address
1624 * @val: the expected value
1625 * @fshared: whether the futex is shared (1) or not (0)
1626 * @q: the associated futex_q
1627 * @hb: storage for hash_bucket pointer to be returned to caller
1629 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1630 * compare it with the expected value. Handle atomic faults internally.
1631 * Return with the hb lock held and a q.key reference on success, and unlocked
1632 * with no q.key reference on failure.
1635 * 0 - uaddr contains val and hb has been locked
1636 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1638 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1639 struct futex_q *q, struct futex_hash_bucket **hb)
1645 * Access the page AFTER the hash-bucket is locked.
1646 * Order is important:
1648 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1649 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1651 * The basic logical guarantee of a futex is that it blocks ONLY
1652 * if cond(var) is known to be true at the time of blocking, for
1653 * any cond. If we queued after testing *uaddr, that would open
1654 * a race condition where we could block indefinitely with
1655 * cond(var) false, which would violate the guarantee.
1657 * A consequence is that futex_wait() can return zero and absorb
1658 * a wakeup when *uaddr != val on entry to the syscall. This is
1662 q->key = FUTEX_KEY_INIT;
1663 ret = get_futex_key(uaddr, fshared, &q->key);
1664 if (unlikely(ret != 0))
1668 *hb = queue_lock(q);
1670 ret = get_futex_value_locked(&uval, uaddr);
1673 queue_unlock(q, *hb);
1675 ret = get_user(uval, uaddr);
1682 put_futex_key(fshared, &q->key);
1687 queue_unlock(q, *hb);
1693 put_futex_key(fshared, &q->key);
1697 static int futex_wait(u32 __user *uaddr, int fshared,
1698 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1700 struct hrtimer_sleeper timeout, *to = NULL;
1701 struct restart_block *restart;
1702 struct futex_hash_bucket *hb;
1716 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1717 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1718 hrtimer_init_sleeper(to, current);
1719 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1720 current->timer_slack_ns);
1723 /* Prepare to wait on uaddr. */
1724 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1728 /* queue_me and wait for wakeup, timeout, or a signal. */
1729 futex_wait_queue_me(hb, &q, to);
1731 /* If we were woken (and unqueued), we succeeded, whatever. */
1733 if (!unqueue_me(&q))
1736 if (to && !to->task)
1740 * We expect signal_pending(current), but another thread may
1741 * have handled it for us already.
1747 restart = ¤t_thread_info()->restart_block;
1748 restart->fn = futex_wait_restart;
1749 restart->futex.uaddr = (u32 *)uaddr;
1750 restart->futex.val = val;
1751 restart->futex.time = abs_time->tv64;
1752 restart->futex.bitset = bitset;
1753 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1756 restart->futex.flags |= FLAGS_SHARED;
1758 restart->futex.flags |= FLAGS_CLOCKRT;
1760 ret = -ERESTART_RESTARTBLOCK;
1763 put_futex_key(fshared, &q.key);
1766 hrtimer_cancel(&to->timer);
1767 destroy_hrtimer_on_stack(&to->timer);
1773 static long futex_wait_restart(struct restart_block *restart)
1775 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1777 ktime_t t, *tp = NULL;
1779 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1780 t.tv64 = restart->futex.time;
1783 restart->fn = do_no_restart_syscall;
1784 if (restart->futex.flags & FLAGS_SHARED)
1786 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1787 restart->futex.bitset,
1788 restart->futex.flags & FLAGS_CLOCKRT);
1793 * Userspace tried a 0 -> TID atomic transition of the futex value
1794 * and failed. The kernel side here does the whole locking operation:
1795 * if there are waiters then it will block, it does PI, etc. (Due to
1796 * races the kernel might see a 0 value of the futex too.)
1798 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1799 int detect, ktime_t *time, int trylock)
1801 struct hrtimer_sleeper timeout, *to = NULL;
1802 struct futex_hash_bucket *hb;
1807 if (refill_pi_state_cache())
1812 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1814 hrtimer_init_sleeper(to, current);
1815 hrtimer_set_expires(&to->timer, *time);
1821 q.key = FUTEX_KEY_INIT;
1822 ret = get_futex_key(uaddr, fshared, &q.key);
1823 if (unlikely(ret != 0))
1827 hb = queue_lock(&q);
1829 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1830 if (unlikely(ret)) {
1833 /* We got the lock. */
1835 goto out_unlock_put_key;
1840 * Task is exiting and we just wait for the
1843 queue_unlock(&q, hb);
1844 put_futex_key(fshared, &q.key);
1848 goto out_unlock_put_key;
1853 * Only actually queue now that the atomic ops are done:
1857 WARN_ON(!q.pi_state);
1859 * Block on the PI mutex:
1862 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1864 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1865 /* Fixup the trylock return value: */
1866 ret = ret ? 0 : -EWOULDBLOCK;
1869 spin_lock(q.lock_ptr);
1871 * Fixup the pi_state owner and possibly acquire the lock if we
1874 res = fixup_owner(uaddr, fshared, &q, !ret);
1876 * If fixup_owner() returned an error, proprogate that. If it acquired
1877 * the lock, clear our -ETIMEDOUT or -EINTR.
1880 ret = (res < 0) ? res : 0;
1883 * If fixup_owner() faulted and was unable to handle the fault, unlock
1884 * it and return the fault to userspace.
1886 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1887 rt_mutex_unlock(&q.pi_state->pi_mutex);
1889 /* Unqueue and drop the lock */
1895 queue_unlock(&q, hb);
1898 put_futex_key(fshared, &q.key);
1901 destroy_hrtimer_on_stack(&to->timer);
1902 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1906 * We have to r/w *(int __user *)uaddr, and we have to modify it
1907 * atomically. Therefore, if we continue to fault after get_user()
1908 * below, we need to handle the fault ourselves, while still holding
1909 * the mmap_sem. This can occur if the uaddr is under contention as
1910 * we have to drop the mmap_sem in order to call get_user().
1912 queue_unlock(&q, hb);
1914 ret = get_user(uval, uaddr);
1921 put_futex_key(fshared, &q.key);
1925 static long futex_lock_pi_restart(struct restart_block *restart)
1927 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1928 ktime_t t, *tp = NULL;
1929 int fshared = restart->futex.flags & FLAGS_SHARED;
1931 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1932 t.tv64 = restart->futex.time;
1935 restart->fn = do_no_restart_syscall;
1937 return (long)futex_lock_pi(uaddr, fshared, restart->futex.val, tp, 0);
1941 * Userspace attempted a TID -> 0 atomic transition, and failed.
1942 * This is the in-kernel slowpath: we look up the PI state (if any),
1943 * and do the rt-mutex unlock.
1945 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1947 struct futex_hash_bucket *hb;
1948 struct futex_q *this, *next;
1950 struct plist_head *head;
1951 union futex_key key = FUTEX_KEY_INIT;
1955 if (get_user(uval, uaddr))
1958 * We release only a lock we actually own:
1960 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1963 ret = get_futex_key(uaddr, fshared, &key);
1964 if (unlikely(ret != 0))
1967 hb = hash_futex(&key);
1968 spin_lock(&hb->lock);
1971 * To avoid races, try to do the TID -> 0 atomic transition
1972 * again. If it succeeds then we can return without waking
1975 if (!(uval & FUTEX_OWNER_DIED))
1976 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1979 if (unlikely(uval == -EFAULT))
1982 * Rare case: we managed to release the lock atomically,
1983 * no need to wake anyone else up:
1985 if (unlikely(uval == task_pid_vnr(current)))
1989 * Ok, other tasks may need to be woken up - check waiters
1990 * and do the wakeup if necessary:
1994 plist_for_each_entry_safe(this, next, head, list) {
1995 if (!match_futex (&this->key, &key))
1997 ret = wake_futex_pi(uaddr, uval, this);
1999 * The atomic access to the futex value
2000 * generated a pagefault, so retry the
2001 * user-access and the wakeup:
2008 * No waiters - kernel unlocks the futex:
2010 if (!(uval & FUTEX_OWNER_DIED)) {
2011 ret = unlock_futex_pi(uaddr, uval);
2017 spin_unlock(&hb->lock);
2018 put_futex_key(fshared, &key);
2025 * We have to r/w *(int __user *)uaddr, and we have to modify it
2026 * atomically. Therefore, if we continue to fault after get_user()
2027 * below, we need to handle the fault ourselves, while still holding
2028 * the mmap_sem. This can occur if the uaddr is under contention as
2029 * we have to drop the mmap_sem in order to call get_user().
2031 spin_unlock(&hb->lock);
2032 put_futex_key(fshared, &key);
2034 ret = get_user(uval, uaddr);
2042 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2043 * @hb: the hash_bucket futex_q was original enqueued on
2044 * @q: the futex_q woken while waiting to be requeued
2045 * @key2: the futex_key of the requeue target futex
2046 * @timeout: the timeout associated with the wait (NULL if none)
2048 * Detect if the task was woken on the initial futex as opposed to the requeue
2049 * target futex. If so, determine if it was a timeout or a signal that caused
2050 * the wakeup and return the appropriate error code to the caller. Must be
2051 * called with the hb lock held.
2054 * 0 - no early wakeup detected
2055 * <0 - -ETIMEDOUT or -ERESTARTSYS (FIXME: or ERESTARTNOINTR?)
2058 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2059 struct futex_q *q, union futex_key *key2,
2060 struct hrtimer_sleeper *timeout)
2065 * With the hb lock held, we avoid races while we process the wakeup.
2066 * We only need to hold hb (and not hb2) to ensure atomicity as the
2067 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2068 * It can't be requeued from uaddr2 to something else since we don't
2069 * support a PI aware source futex for requeue.
2071 if (!match_futex(&q->key, key2)) {
2072 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2074 * We were woken prior to requeue by a timeout or a signal.
2075 * Unqueue the futex_q and determine which it was.
2077 plist_del(&q->list, &q->list.plist);
2078 drop_futex_key_refs(&q->key);
2080 if (timeout && !timeout->task)
2084 * We expect signal_pending(current), but another
2085 * thread may have handled it for us already.
2087 /* FIXME: ERESTARTSYS or ERESTARTNOINTR? Do we care if
2088 * the user specified SA_RESTART or not? */
2096 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2097 * @uaddr: the futex we initialyl wait on (non-pi)
2098 * @fshared: whether the futexes are shared (1) or not (0). They must be
2099 * the same type, no requeueing from private to shared, etc.
2100 * @val: the expected value of uaddr
2101 * @abs_time: absolute timeout
2102 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all.
2103 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2104 * @uaddr2: the pi futex we will take prior to returning to user-space
2106 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2107 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2108 * complete the acquisition of the rt_mutex prior to returning to userspace.
2109 * This ensures the rt_mutex maintains an owner when it has waiters; without
2110 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2113 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2114 * via the following:
2115 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2116 * 2) wakeup on uaddr2 after a requeue and subsequent unlock
2117 * 3) signal (before or after requeue)
2118 * 4) timeout (before or after requeue)
2120 * If 3, we setup a restart_block with futex_wait_requeue_pi() as the function.
2122 * If 2, we may then block on trying to take the rt_mutex and return via:
2123 * 5) successful lock
2126 * 8) other lock acquisition failure
2128 * If 6, we setup a restart_block with futex_lock_pi() as the function.
2130 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2136 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2137 u32 val, ktime_t *abs_time, u32 bitset,
2138 int clockrt, u32 __user *uaddr2)
2140 struct hrtimer_sleeper timeout, *to = NULL;
2141 struct rt_mutex_waiter rt_waiter;
2142 struct rt_mutex *pi_mutex = NULL;
2143 struct restart_block *restart;
2144 struct futex_hash_bucket *hb;
2145 union futex_key key2;
2155 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2156 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2157 hrtimer_init_sleeper(to, current);
2158 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2159 current->timer_slack_ns);
2163 * The waiter is allocated on our stack, manipulated by the requeue
2164 * code while we sleep on uaddr.
2166 debug_rt_mutex_init_waiter(&rt_waiter);
2167 rt_waiter.task = NULL;
2171 q.rt_waiter = &rt_waiter;
2173 key2 = FUTEX_KEY_INIT;
2174 ret = get_futex_key(uaddr2, fshared, &key2);
2175 if (unlikely(ret != 0))
2178 /* Prepare to wait on uaddr. */
2179 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2181 put_futex_key(fshared, &key2);
2185 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2186 futex_wait_queue_me(hb, &q, to);
2188 spin_lock(&hb->lock);
2189 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2190 spin_unlock(&hb->lock);
2195 * In order for us to be here, we know our q.key == key2, and since
2196 * we took the hb->lock above, we also know that futex_requeue() has
2197 * completed and we no longer have to concern ourselves with a wakeup
2198 * race with the atomic proxy lock acquition by the requeue code.
2201 /* Check if the requeue code acquired the second futex for us. */
2204 * Got the lock. We might not be the anticipated owner if we
2205 * did a lock-steal - fix up the PI-state in that case.
2207 if (q.pi_state && (q.pi_state->owner != current)) {
2208 spin_lock(q.lock_ptr);
2209 ret = fixup_pi_state_owner(uaddr2, &q, current,
2211 spin_unlock(q.lock_ptr);
2215 * We have been woken up by futex_unlock_pi(), a timeout, or a
2216 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2219 WARN_ON(!&q.pi_state);
2220 pi_mutex = &q.pi_state->pi_mutex;
2221 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2222 debug_rt_mutex_free_waiter(&rt_waiter);
2224 spin_lock(q.lock_ptr);
2226 * Fixup the pi_state owner and possibly acquire the lock if we
2229 res = fixup_owner(uaddr2, fshared, &q, !ret);
2231 * If fixup_owner() returned an error, proprogate that. If it
2232 * acquired the lock, clear our -ETIMEDOUT or -EINTR.
2235 ret = (res < 0) ? res : 0;
2237 /* Unqueue and drop the lock. */
2242 * If fixup_pi_state_owner() faulted and was unable to handle the
2243 * fault, unlock the rt_mutex and return the fault to userspace.
2245 if (ret == -EFAULT) {
2246 if (rt_mutex_owner(pi_mutex) == current)
2247 rt_mutex_unlock(pi_mutex);
2248 } else if (ret == -EINTR) {
2250 if (get_user(uval, uaddr2))
2254 * We've already been requeued, so restart by calling
2255 * futex_lock_pi() directly, rather then returning to this
2258 ret = -ERESTART_RESTARTBLOCK;
2259 restart = ¤t_thread_info()->restart_block;
2260 restart->fn = futex_lock_pi_restart;
2261 restart->futex.uaddr = (u32 *)uaddr2;
2262 restart->futex.val = uval;
2263 restart->futex.flags = 0;
2265 restart->futex.flags |= FLAGS_HAS_TIMEOUT;
2266 restart->futex.time = abs_time->tv64;
2270 restart->futex.flags |= FLAGS_SHARED;
2272 restart->futex.flags |= FLAGS_CLOCKRT;
2276 put_futex_key(fshared, &q.key);
2277 put_futex_key(fshared, &key2);
2281 hrtimer_cancel(&to->timer);
2282 destroy_hrtimer_on_stack(&to->timer);
2288 * Support for robust futexes: the kernel cleans up held futexes at
2291 * Implementation: user-space maintains a per-thread list of locks it
2292 * is holding. Upon do_exit(), the kernel carefully walks this list,
2293 * and marks all locks that are owned by this thread with the
2294 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2295 * always manipulated with the lock held, so the list is private and
2296 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2297 * field, to allow the kernel to clean up if the thread dies after
2298 * acquiring the lock, but just before it could have added itself to
2299 * the list. There can only be one such pending lock.
2303 * sys_set_robust_list - set the robust-futex list head of a task
2304 * @head: pointer to the list-head
2305 * @len: length of the list-head, as userspace expects
2307 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2310 if (!futex_cmpxchg_enabled)
2313 * The kernel knows only one size for now:
2315 if (unlikely(len != sizeof(*head)))
2318 current->robust_list = head;
2324 * sys_get_robust_list - get the robust-futex list head of a task
2325 * @pid: pid of the process [zero for current task]
2326 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2327 * @len_ptr: pointer to a length field, the kernel fills in the header size
2329 SYSCALL_DEFINE3(get_robust_list, int, pid,
2330 struct robust_list_head __user * __user *, head_ptr,
2331 size_t __user *, len_ptr)
2333 struct robust_list_head __user *head;
2335 const struct cred *cred = current_cred(), *pcred;
2337 if (!futex_cmpxchg_enabled)
2341 head = current->robust_list;
2343 struct task_struct *p;
2347 p = find_task_by_vpid(pid);
2351 pcred = __task_cred(p);
2352 if (cred->euid != pcred->euid &&
2353 cred->euid != pcred->uid &&
2354 !capable(CAP_SYS_PTRACE))
2356 head = p->robust_list;
2360 if (put_user(sizeof(*head), len_ptr))
2362 return put_user(head, head_ptr);
2371 * Process a futex-list entry, check whether it's owned by the
2372 * dying task, and do notification if so:
2374 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2376 u32 uval, nval, mval;
2379 if (get_user(uval, uaddr))
2382 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2384 * Ok, this dying thread is truly holding a futex
2385 * of interest. Set the OWNER_DIED bit atomically
2386 * via cmpxchg, and if the value had FUTEX_WAITERS
2387 * set, wake up a waiter (if any). (We have to do a
2388 * futex_wake() even if OWNER_DIED is already set -
2389 * to handle the rare but possible case of recursive
2390 * thread-death.) The rest of the cleanup is done in
2393 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2394 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2396 if (nval == -EFAULT)
2403 * Wake robust non-PI futexes here. The wakeup of
2404 * PI futexes happens in exit_pi_state():
2406 if (!pi && (uval & FUTEX_WAITERS))
2407 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2413 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2415 static inline int fetch_robust_entry(struct robust_list __user **entry,
2416 struct robust_list __user * __user *head,
2419 unsigned long uentry;
2421 if (get_user(uentry, (unsigned long __user *)head))
2424 *entry = (void __user *)(uentry & ~1UL);
2431 * Walk curr->robust_list (very carefully, it's a userspace list!)
2432 * and mark any locks found there dead, and notify any waiters.
2434 * We silently return on any sign of list-walking problem.
2436 void exit_robust_list(struct task_struct *curr)
2438 struct robust_list_head __user *head = curr->robust_list;
2439 struct robust_list __user *entry, *next_entry, *pending;
2440 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2441 unsigned long futex_offset;
2444 if (!futex_cmpxchg_enabled)
2448 * Fetch the list head (which was registered earlier, via
2449 * sys_set_robust_list()):
2451 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2454 * Fetch the relative futex offset:
2456 if (get_user(futex_offset, &head->futex_offset))
2459 * Fetch any possibly pending lock-add first, and handle it
2462 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2465 next_entry = NULL; /* avoid warning with gcc */
2466 while (entry != &head->list) {
2468 * Fetch the next entry in the list before calling
2469 * handle_futex_death:
2471 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2473 * A pending lock might already be on the list, so
2474 * don't process it twice:
2476 if (entry != pending)
2477 if (handle_futex_death((void __user *)entry + futex_offset,
2485 * Avoid excessively long or circular lists:
2494 handle_futex_death((void __user *)pending + futex_offset,
2498 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2499 u32 __user *uaddr2, u32 val2, u32 val3)
2501 int clockrt, ret = -ENOSYS;
2502 int cmd = op & FUTEX_CMD_MASK;
2505 if (!(op & FUTEX_PRIVATE_FLAG))
2508 clockrt = op & FUTEX_CLOCK_REALTIME;
2509 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2514 val3 = FUTEX_BITSET_MATCH_ANY;
2515 case FUTEX_WAIT_BITSET:
2516 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2519 val3 = FUTEX_BITSET_MATCH_ANY;
2520 case FUTEX_WAKE_BITSET:
2521 ret = futex_wake(uaddr, fshared, val, val3);
2524 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2526 case FUTEX_CMP_REQUEUE:
2527 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2531 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2534 if (futex_cmpxchg_enabled)
2535 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2537 case FUTEX_UNLOCK_PI:
2538 if (futex_cmpxchg_enabled)
2539 ret = futex_unlock_pi(uaddr, fshared);
2541 case FUTEX_TRYLOCK_PI:
2542 if (futex_cmpxchg_enabled)
2543 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2545 case FUTEX_WAIT_REQUEUE_PI:
2546 val3 = FUTEX_BITSET_MATCH_ANY;
2547 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2550 case FUTEX_CMP_REQUEUE_PI:
2551 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2561 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2562 struct timespec __user *, utime, u32 __user *, uaddr2,
2566 ktime_t t, *tp = NULL;
2568 int cmd = op & FUTEX_CMD_MASK;
2570 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2571 cmd == FUTEX_WAIT_BITSET ||
2572 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2573 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2575 if (!timespec_valid(&ts))
2578 t = timespec_to_ktime(ts);
2579 if (cmd == FUTEX_WAIT)
2580 t = ktime_add_safe(ktime_get(), t);
2584 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2585 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2587 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2588 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2589 val2 = (u32) (unsigned long) utime;
2591 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2594 static int __init futex_init(void)
2600 * This will fail and we want it. Some arch implementations do
2601 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2602 * functionality. We want to know that before we call in any
2603 * of the complex code paths. Also we want to prevent
2604 * registration of robust lists in that case. NULL is
2605 * guaranteed to fault and we get -EFAULT on functional
2606 * implementation, the non functional ones will return
2609 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2610 if (curval == -EFAULT)
2611 futex_cmpxchg_enabled = 1;
2613 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2614 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2615 spin_lock_init(&futex_queues[i].lock);
2620 __initcall(futex_init);