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 * struct futex_q - The hashed futex queue entry, one per waiting task
94 * @task: the task waiting on the futex
95 * @lock_ptr: the hash bucket lock
96 * @key: the key the futex is hashed on
97 * @pi_state: optional priority inheritance state
98 * @rt_waiter: rt_waiter storage for use with requeue_pi
99 * @requeue_pi_key: the requeue_pi target futex key
100 * @bitset: bitset for the optional bitmasked wakeup
102 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103 * we can wake only the relevant ones (hashed queues may be shared).
105 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107 * The order of wakup is always to make the first condition true, then
110 * PI futexes are typically woken before they are removed from the hash list via
111 * the rt_mutex code. See unqueue_me_pi().
114 struct plist_node list;
116 struct task_struct *task;
117 spinlock_t *lock_ptr;
119 struct futex_pi_state *pi_state;
120 struct rt_mutex_waiter *rt_waiter;
121 union futex_key *requeue_pi_key;
126 * Hash buckets are shared by all the futex_keys that hash to the same
127 * location. Each key may have multiple futex_q structures, one for each task
128 * waiting on a futex.
130 struct futex_hash_bucket {
132 struct plist_head chain;
135 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
138 * We hash on the keys returned from get_futex_key (see below).
140 static struct futex_hash_bucket *hash_futex(union futex_key *key)
142 u32 hash = jhash2((u32*)&key->both.word,
143 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
145 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
149 * Return 1 if two futex_keys are equal, 0 otherwise.
151 static inline int match_futex(union futex_key *key1, union futex_key *key2)
153 return (key1->both.word == key2->both.word
154 && key1->both.ptr == key2->both.ptr
155 && key1->both.offset == key2->both.offset);
159 * Take a reference to the resource addressed by a key.
160 * Can be called while holding spinlocks.
163 static void get_futex_key_refs(union futex_key *key)
168 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170 atomic_inc(&key->shared.inode->i_count);
172 case FUT_OFF_MMSHARED:
173 atomic_inc(&key->private.mm->mm_count);
179 * Drop a reference to the resource addressed by a key.
180 * The hash bucket spinlock must not be held.
182 static void drop_futex_key_refs(union futex_key *key)
184 if (!key->both.ptr) {
185 /* If we're here then we tried to put a key we failed to get */
190 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192 iput(key->shared.inode);
194 case FUT_OFF_MMSHARED:
195 mmdrop(key->private.mm);
201 * get_futex_key() - Get parameters which are the keys for a futex
202 * @uaddr: virtual address of the futex
203 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
204 * @key: address where result is stored.
205 * @rw: mapping needs to be read/write (values: VERIFY_READ,
208 * Returns a negative error code or 0
209 * The key words are stored in *key on success.
211 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
212 * offset_within_page). For private mappings, it's (uaddr, current->mm).
213 * We can usually work out the index without swapping in the page.
215 * lock_page() might sleep, the caller should not hold a spinlock.
218 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
220 unsigned long address = (unsigned long)uaddr;
221 struct mm_struct *mm = current->mm;
226 * The futex address must be "naturally" aligned.
228 key->both.offset = address % PAGE_SIZE;
229 if (unlikely((address % sizeof(u32)) != 0))
231 address -= key->both.offset;
234 * PROCESS_PRIVATE futexes are fast.
235 * As the mm cannot disappear under us and the 'key' only needs
236 * virtual address, we dont even have to find the underlying vma.
237 * Note : We do have to check 'uaddr' is a valid user address,
238 * but access_ok() should be faster than find_vma()
241 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
243 key->private.mm = mm;
244 key->private.address = address;
245 get_futex_key_refs(key);
250 err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
254 page = compound_head(page);
256 if (!page->mapping) {
263 * Private mappings are handled in a simple way.
265 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
266 * it's a read-only handle, it's expected that futexes attach to
267 * the object not the particular process.
269 if (PageAnon(page)) {
270 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
271 key->private.mm = mm;
272 key->private.address = address;
274 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
275 key->shared.inode = page->mapping->host;
276 key->shared.pgoff = page->index;
279 get_futex_key_refs(key);
287 void put_futex_key(int fshared, union futex_key *key)
289 drop_futex_key_refs(key);
293 * fault_in_user_writeable() - Fault in user address and verify RW access
294 * @uaddr: pointer to faulting user space address
296 * Slow path to fixup the fault we just took in the atomic write
299 * We have no generic implementation of a non destructive write to the
300 * user address. We know that we faulted in the atomic pagefault
301 * disabled section so we can as well avoid the #PF overhead by
302 * calling get_user_pages() right away.
304 static int fault_in_user_writeable(u32 __user *uaddr)
306 int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
307 1, 1, 0, NULL, NULL);
308 return ret < 0 ? ret : 0;
312 * futex_top_waiter() - Return the highest priority waiter on a futex
313 * @hb: the hash bucket the futex_q's reside in
314 * @key: the futex key (to distinguish it from other futex futex_q's)
316 * Must be called with the hb lock held.
318 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
319 union futex_key *key)
321 struct futex_q *this;
323 plist_for_each_entry(this, &hb->chain, list) {
324 if (match_futex(&this->key, key))
330 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
335 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
341 static int get_futex_value_locked(u32 *dest, u32 __user *from)
346 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
349 return ret ? -EFAULT : 0;
356 static int refill_pi_state_cache(void)
358 struct futex_pi_state *pi_state;
360 if (likely(current->pi_state_cache))
363 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
368 INIT_LIST_HEAD(&pi_state->list);
369 /* pi_mutex gets initialized later */
370 pi_state->owner = NULL;
371 atomic_set(&pi_state->refcount, 1);
372 pi_state->key = FUTEX_KEY_INIT;
374 current->pi_state_cache = pi_state;
379 static struct futex_pi_state * alloc_pi_state(void)
381 struct futex_pi_state *pi_state = current->pi_state_cache;
384 current->pi_state_cache = NULL;
389 static void free_pi_state(struct futex_pi_state *pi_state)
391 if (!atomic_dec_and_test(&pi_state->refcount))
395 * If pi_state->owner is NULL, the owner is most probably dying
396 * and has cleaned up the pi_state already
398 if (pi_state->owner) {
399 spin_lock_irq(&pi_state->owner->pi_lock);
400 list_del_init(&pi_state->list);
401 spin_unlock_irq(&pi_state->owner->pi_lock);
403 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
406 if (current->pi_state_cache)
410 * pi_state->list is already empty.
411 * clear pi_state->owner.
412 * refcount is at 0 - put it back to 1.
414 pi_state->owner = NULL;
415 atomic_set(&pi_state->refcount, 1);
416 current->pi_state_cache = pi_state;
421 * Look up the task based on what TID userspace gave us.
424 static struct task_struct * futex_find_get_task(pid_t pid)
426 struct task_struct *p;
427 const struct cred *cred = current_cred(), *pcred;
430 p = find_task_by_vpid(pid);
434 pcred = __task_cred(p);
435 if (cred->euid != pcred->euid &&
436 cred->euid != pcred->uid)
448 * This task is holding PI mutexes at exit time => bad.
449 * Kernel cleans up PI-state, but userspace is likely hosed.
450 * (Robust-futex cleanup is separate and might save the day for userspace.)
452 void exit_pi_state_list(struct task_struct *curr)
454 struct list_head *next, *head = &curr->pi_state_list;
455 struct futex_pi_state *pi_state;
456 struct futex_hash_bucket *hb;
457 union futex_key key = FUTEX_KEY_INIT;
459 if (!futex_cmpxchg_enabled)
462 * We are a ZOMBIE and nobody can enqueue itself on
463 * pi_state_list anymore, but we have to be careful
464 * versus waiters unqueueing themselves:
466 spin_lock_irq(&curr->pi_lock);
467 while (!list_empty(head)) {
470 pi_state = list_entry(next, struct futex_pi_state, list);
472 hb = hash_futex(&key);
473 spin_unlock_irq(&curr->pi_lock);
475 spin_lock(&hb->lock);
477 spin_lock_irq(&curr->pi_lock);
479 * We dropped the pi-lock, so re-check whether this
480 * task still owns the PI-state:
482 if (head->next != next) {
483 spin_unlock(&hb->lock);
487 WARN_ON(pi_state->owner != curr);
488 WARN_ON(list_empty(&pi_state->list));
489 list_del_init(&pi_state->list);
490 pi_state->owner = NULL;
491 spin_unlock_irq(&curr->pi_lock);
493 rt_mutex_unlock(&pi_state->pi_mutex);
495 spin_unlock(&hb->lock);
497 spin_lock_irq(&curr->pi_lock);
499 spin_unlock_irq(&curr->pi_lock);
503 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
504 union futex_key *key, struct futex_pi_state **ps)
506 struct futex_pi_state *pi_state = NULL;
507 struct futex_q *this, *next;
508 struct plist_head *head;
509 struct task_struct *p;
510 pid_t pid = uval & FUTEX_TID_MASK;
514 plist_for_each_entry_safe(this, next, head, list) {
515 if (match_futex(&this->key, key)) {
517 * Another waiter already exists - bump up
518 * the refcount and return its pi_state:
520 pi_state = this->pi_state;
522 * Userspace might have messed up non PI and PI futexes
524 if (unlikely(!pi_state))
527 WARN_ON(!atomic_read(&pi_state->refcount));
528 WARN_ON(pid && pi_state->owner &&
529 pi_state->owner->pid != pid);
531 atomic_inc(&pi_state->refcount);
539 * We are the first waiter - try to look up the real owner and attach
540 * the new pi_state to it, but bail out when TID = 0
544 p = futex_find_get_task(pid);
549 * We need to look at the task state flags to figure out,
550 * whether the task is exiting. To protect against the do_exit
551 * change of the task flags, we do this protected by
554 spin_lock_irq(&p->pi_lock);
555 if (unlikely(p->flags & PF_EXITING)) {
557 * The task is on the way out. When PF_EXITPIDONE is
558 * set, we know that the task has finished the
561 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
563 spin_unlock_irq(&p->pi_lock);
568 pi_state = alloc_pi_state();
571 * Initialize the pi_mutex in locked state and make 'p'
574 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
576 /* Store the key for possible exit cleanups: */
577 pi_state->key = *key;
579 WARN_ON(!list_empty(&pi_state->list));
580 list_add(&pi_state->list, &p->pi_state_list);
582 spin_unlock_irq(&p->pi_lock);
592 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
593 * @uaddr: the pi futex user address
594 * @hb: the pi futex hash bucket
595 * @key: the futex key associated with uaddr and hb
596 * @ps: the pi_state pointer where we store the result of the
598 * @task: the task to perform the atomic lock work for. This will
599 * be "current" except in the case of requeue pi.
600 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
604 * 1 - acquired the lock
607 * The hb->lock and futex_key refs shall be held by the caller.
609 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
610 union futex_key *key,
611 struct futex_pi_state **ps,
612 struct task_struct *task, int set_waiters)
614 int lock_taken, ret, ownerdied = 0;
615 u32 uval, newval, curval;
618 ret = lock_taken = 0;
621 * To avoid races, we attempt to take the lock here again
622 * (by doing a 0 -> TID atomic cmpxchg), while holding all
623 * the locks. It will most likely not succeed.
625 newval = task_pid_vnr(task);
627 newval |= FUTEX_WAITERS;
629 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
631 if (unlikely(curval == -EFAULT))
637 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
641 * Surprise - we got the lock. Just return to userspace:
643 if (unlikely(!curval))
649 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
650 * to wake at the next unlock.
652 newval = curval | FUTEX_WAITERS;
655 * There are two cases, where a futex might have no owner (the
656 * owner TID is 0): OWNER_DIED. We take over the futex in this
657 * case. We also do an unconditional take over, when the owner
660 * This is safe as we are protected by the hash bucket lock !
662 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
663 /* Keep the OWNER_DIED bit */
664 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
669 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
671 if (unlikely(curval == -EFAULT))
673 if (unlikely(curval != uval))
677 * We took the lock due to owner died take over.
679 if (unlikely(lock_taken))
683 * We dont have the lock. Look up the PI state (or create it if
684 * we are the first waiter):
686 ret = lookup_pi_state(uval, hb, key, ps);
692 * No owner found for this futex. Check if the
693 * OWNER_DIED bit is set to figure out whether
694 * this is a robust futex or not.
696 if (get_futex_value_locked(&curval, uaddr))
700 * We simply start over in case of a robust
701 * futex. The code above will take the futex
704 if (curval & FUTEX_OWNER_DIED) {
717 * The hash bucket lock must be held when this is called.
718 * Afterwards, the futex_q must not be accessed.
720 static void wake_futex(struct futex_q *q)
722 struct task_struct *p = q->task;
725 * We set q->lock_ptr = NULL _before_ we wake up the task. If
726 * a non futex wake up happens on another CPU then the task
727 * might exit and p would dereference a non existing task
728 * struct. Prevent this by holding a reference on p across the
733 plist_del(&q->list, &q->list.plist);
735 * The waiting task can free the futex_q as soon as
736 * q->lock_ptr = NULL is written, without taking any locks. A
737 * memory barrier is required here to prevent the following
738 * store to lock_ptr from getting ahead of the plist_del.
743 wake_up_state(p, TASK_NORMAL);
747 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
749 struct task_struct *new_owner;
750 struct futex_pi_state *pi_state = this->pi_state;
756 spin_lock(&pi_state->pi_mutex.wait_lock);
757 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
760 * This happens when we have stolen the lock and the original
761 * pending owner did not enqueue itself back on the rt_mutex.
762 * Thats not a tragedy. We know that way, that a lock waiter
763 * is on the fly. We make the futex_q waiter the pending owner.
766 new_owner = this->task;
769 * We pass it to the next owner. (The WAITERS bit is always
770 * kept enabled while there is PI state around. We must also
771 * preserve the owner died bit.)
773 if (!(uval & FUTEX_OWNER_DIED)) {
776 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
778 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
780 if (curval == -EFAULT)
782 else if (curval != uval)
785 spin_unlock(&pi_state->pi_mutex.wait_lock);
790 spin_lock_irq(&pi_state->owner->pi_lock);
791 WARN_ON(list_empty(&pi_state->list));
792 list_del_init(&pi_state->list);
793 spin_unlock_irq(&pi_state->owner->pi_lock);
795 spin_lock_irq(&new_owner->pi_lock);
796 WARN_ON(!list_empty(&pi_state->list));
797 list_add(&pi_state->list, &new_owner->pi_state_list);
798 pi_state->owner = new_owner;
799 spin_unlock_irq(&new_owner->pi_lock);
801 spin_unlock(&pi_state->pi_mutex.wait_lock);
802 rt_mutex_unlock(&pi_state->pi_mutex);
807 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
812 * There is no waiter, so we unlock the futex. The owner died
813 * bit has not to be preserved here. We are the owner:
815 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
817 if (oldval == -EFAULT)
826 * Express the locking dependencies for lockdep:
829 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
832 spin_lock(&hb1->lock);
834 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
835 } else { /* hb1 > hb2 */
836 spin_lock(&hb2->lock);
837 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
842 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
844 spin_unlock(&hb1->lock);
846 spin_unlock(&hb2->lock);
850 * Wake up waiters matching bitset queued on this futex (uaddr).
852 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
854 struct futex_hash_bucket *hb;
855 struct futex_q *this, *next;
856 struct plist_head *head;
857 union futex_key key = FUTEX_KEY_INIT;
863 ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
864 if (unlikely(ret != 0))
867 hb = hash_futex(&key);
868 spin_lock(&hb->lock);
871 plist_for_each_entry_safe(this, next, head, list) {
872 if (match_futex (&this->key, &key)) {
873 if (this->pi_state || this->rt_waiter) {
878 /* Check if one of the bits is set in both bitsets */
879 if (!(this->bitset & bitset))
883 if (++ret >= nr_wake)
888 spin_unlock(&hb->lock);
889 put_futex_key(fshared, &key);
895 * Wake up all waiters hashed on the physical page that is mapped
896 * to this virtual address:
899 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
900 int nr_wake, int nr_wake2, int op)
902 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
903 struct futex_hash_bucket *hb1, *hb2;
904 struct plist_head *head;
905 struct futex_q *this, *next;
909 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
910 if (unlikely(ret != 0))
912 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
913 if (unlikely(ret != 0))
916 hb1 = hash_futex(&key1);
917 hb2 = hash_futex(&key2);
920 double_lock_hb(hb1, hb2);
921 op_ret = futex_atomic_op_inuser(op, uaddr2);
922 if (unlikely(op_ret < 0)) {
924 double_unlock_hb(hb1, hb2);
928 * we don't get EFAULT from MMU faults if we don't have an MMU,
929 * but we might get them from range checking
935 if (unlikely(op_ret != -EFAULT)) {
940 ret = fault_in_user_writeable(uaddr2);
947 put_futex_key(fshared, &key2);
948 put_futex_key(fshared, &key1);
954 plist_for_each_entry_safe(this, next, head, list) {
955 if (match_futex (&this->key, &key1)) {
957 if (++ret >= nr_wake)
966 plist_for_each_entry_safe(this, next, head, list) {
967 if (match_futex (&this->key, &key2)) {
969 if (++op_ret >= nr_wake2)
976 double_unlock_hb(hb1, hb2);
978 put_futex_key(fshared, &key2);
980 put_futex_key(fshared, &key1);
986 * requeue_futex() - Requeue a futex_q from one hb to another
987 * @q: the futex_q to requeue
988 * @hb1: the source hash_bucket
989 * @hb2: the target hash_bucket
990 * @key2: the new key for the requeued futex_q
993 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
994 struct futex_hash_bucket *hb2, union futex_key *key2)
998 * If key1 and key2 hash to the same bucket, no need to
1001 if (likely(&hb1->chain != &hb2->chain)) {
1002 plist_del(&q->list, &hb1->chain);
1003 plist_add(&q->list, &hb2->chain);
1004 q->lock_ptr = &hb2->lock;
1005 #ifdef CONFIG_DEBUG_PI_LIST
1006 q->list.plist.lock = &hb2->lock;
1009 get_futex_key_refs(key2);
1014 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1016 * @key: the key of the requeue target futex
1017 * @hb: the hash_bucket of the requeue target futex
1019 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1020 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1021 * to the requeue target futex so the waiter can detect the wakeup on the right
1022 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1023 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1024 * to protect access to the pi_state to fixup the owner later. Must be called
1025 * with both q->lock_ptr and hb->lock held.
1028 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1029 struct futex_hash_bucket *hb)
1031 drop_futex_key_refs(&q->key);
1032 get_futex_key_refs(key);
1035 WARN_ON(plist_node_empty(&q->list));
1036 plist_del(&q->list, &q->list.plist);
1038 WARN_ON(!q->rt_waiter);
1039 q->rt_waiter = NULL;
1041 q->lock_ptr = &hb->lock;
1042 #ifdef CONFIG_DEBUG_PI_LIST
1043 q->list.plist.lock = &hb->lock;
1046 wake_up_state(q->task, TASK_NORMAL);
1050 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1051 * @pifutex: the user address of the to futex
1052 * @hb1: the from futex hash bucket, must be locked by the caller
1053 * @hb2: the to futex hash bucket, must be locked by the caller
1054 * @key1: the from futex key
1055 * @key2: the to futex key
1056 * @ps: address to store the pi_state pointer
1057 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1059 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1060 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1061 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1062 * hb1 and hb2 must be held by the caller.
1065 * 0 - failed to acquire the lock atomicly
1066 * 1 - acquired the lock
1069 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1070 struct futex_hash_bucket *hb1,
1071 struct futex_hash_bucket *hb2,
1072 union futex_key *key1, union futex_key *key2,
1073 struct futex_pi_state **ps, int set_waiters)
1075 struct futex_q *top_waiter = NULL;
1079 if (get_futex_value_locked(&curval, pifutex))
1083 * Find the top_waiter and determine if there are additional waiters.
1084 * If the caller intends to requeue more than 1 waiter to pifutex,
1085 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1086 * as we have means to handle the possible fault. If not, don't set
1087 * the bit unecessarily as it will force the subsequent unlock to enter
1090 top_waiter = futex_top_waiter(hb1, key1);
1092 /* There are no waiters, nothing for us to do. */
1096 /* Ensure we requeue to the expected futex. */
1097 if (!match_futex(top_waiter->requeue_pi_key, key2))
1101 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1102 * the contended case or if set_waiters is 1. The pi_state is returned
1103 * in ps in contended cases.
1105 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1108 requeue_pi_wake_futex(top_waiter, key2, hb2);
1114 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1115 * uaddr1: source futex user address
1116 * uaddr2: target futex user address
1117 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1118 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1119 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1120 * pi futex (pi to pi requeue is not supported)
1122 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1123 * uaddr2 atomically on behalf of the top waiter.
1126 * >=0 - on success, the number of tasks requeued or woken
1129 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1130 int nr_wake, int nr_requeue, u32 *cmpval,
1133 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1134 int drop_count = 0, task_count = 0, ret;
1135 struct futex_pi_state *pi_state = NULL;
1136 struct futex_hash_bucket *hb1, *hb2;
1137 struct plist_head *head1;
1138 struct futex_q *this, *next;
1143 * requeue_pi requires a pi_state, try to allocate it now
1144 * without any locks in case it fails.
1146 if (refill_pi_state_cache())
1149 * requeue_pi must wake as many tasks as it can, up to nr_wake
1150 * + nr_requeue, since it acquires the rt_mutex prior to
1151 * returning to userspace, so as to not leave the rt_mutex with
1152 * waiters and no owner. However, second and third wake-ups
1153 * cannot be predicted as they involve race conditions with the
1154 * first wake and a fault while looking up the pi_state. Both
1155 * pthread_cond_signal() and pthread_cond_broadcast() should
1163 if (pi_state != NULL) {
1165 * We will have to lookup the pi_state again, so free this one
1166 * to keep the accounting correct.
1168 free_pi_state(pi_state);
1172 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1173 if (unlikely(ret != 0))
1175 ret = get_futex_key(uaddr2, fshared, &key2,
1176 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1177 if (unlikely(ret != 0))
1180 hb1 = hash_futex(&key1);
1181 hb2 = hash_futex(&key2);
1184 double_lock_hb(hb1, hb2);
1186 if (likely(cmpval != NULL)) {
1189 ret = get_futex_value_locked(&curval, uaddr1);
1191 if (unlikely(ret)) {
1192 double_unlock_hb(hb1, hb2);
1194 ret = get_user(curval, uaddr1);
1201 put_futex_key(fshared, &key2);
1202 put_futex_key(fshared, &key1);
1205 if (curval != *cmpval) {
1211 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1213 * Attempt to acquire uaddr2 and wake the top waiter. If we
1214 * intend to requeue waiters, force setting the FUTEX_WAITERS
1215 * bit. We force this here where we are able to easily handle
1216 * faults rather in the requeue loop below.
1218 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1219 &key2, &pi_state, nr_requeue);
1222 * At this point the top_waiter has either taken uaddr2 or is
1223 * waiting on it. If the former, then the pi_state will not
1224 * exist yet, look it up one more time to ensure we have a
1230 ret = get_futex_value_locked(&curval2, uaddr2);
1232 ret = lookup_pi_state(curval2, hb2, &key2,
1240 double_unlock_hb(hb1, hb2);
1241 put_futex_key(fshared, &key2);
1242 put_futex_key(fshared, &key1);
1243 ret = fault_in_user_writeable(uaddr2);
1248 /* The owner was exiting, try again. */
1249 double_unlock_hb(hb1, hb2);
1250 put_futex_key(fshared, &key2);
1251 put_futex_key(fshared, &key1);
1259 head1 = &hb1->chain;
1260 plist_for_each_entry_safe(this, next, head1, list) {
1261 if (task_count - nr_wake >= nr_requeue)
1264 if (!match_futex(&this->key, &key1))
1268 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1269 * be paired with each other and no other futex ops.
1271 if ((requeue_pi && !this->rt_waiter) ||
1272 (!requeue_pi && this->rt_waiter)) {
1278 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1279 * lock, we already woke the top_waiter. If not, it will be
1280 * woken by futex_unlock_pi().
1282 if (++task_count <= nr_wake && !requeue_pi) {
1287 /* Ensure we requeue to the expected futex for requeue_pi. */
1288 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1294 * Requeue nr_requeue waiters and possibly one more in the case
1295 * of requeue_pi if we couldn't acquire the lock atomically.
1298 /* Prepare the waiter to take the rt_mutex. */
1299 atomic_inc(&pi_state->refcount);
1300 this->pi_state = pi_state;
1301 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1305 /* We got the lock. */
1306 requeue_pi_wake_futex(this, &key2, hb2);
1310 this->pi_state = NULL;
1311 free_pi_state(pi_state);
1315 requeue_futex(this, hb1, hb2, &key2);
1320 double_unlock_hb(hb1, hb2);
1323 * drop_futex_key_refs() must be called outside the spinlocks. During
1324 * the requeue we moved futex_q's from the hash bucket at key1 to the
1325 * one at key2 and updated their key pointer. We no longer need to
1326 * hold the references to key1.
1328 while (--drop_count >= 0)
1329 drop_futex_key_refs(&key1);
1332 put_futex_key(fshared, &key2);
1334 put_futex_key(fshared, &key1);
1336 if (pi_state != NULL)
1337 free_pi_state(pi_state);
1338 return ret ? ret : task_count;
1341 /* The key must be already stored in q->key. */
1342 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1344 struct futex_hash_bucket *hb;
1346 get_futex_key_refs(&q->key);
1347 hb = hash_futex(&q->key);
1348 q->lock_ptr = &hb->lock;
1350 spin_lock(&hb->lock);
1355 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1357 spin_unlock(&hb->lock);
1358 drop_futex_key_refs(&q->key);
1362 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1363 * @q: The futex_q to enqueue
1364 * @hb: The destination hash bucket
1366 * The hb->lock must be held by the caller, and is released here. A call to
1367 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1368 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1369 * or nothing if the unqueue is done as part of the wake process and the unqueue
1370 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1373 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1378 * The priority used to register this element is
1379 * - either the real thread-priority for the real-time threads
1380 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1381 * - or MAX_RT_PRIO for non-RT threads.
1382 * Thus, all RT-threads are woken first in priority order, and
1383 * the others are woken last, in FIFO order.
1385 prio = min(current->normal_prio, MAX_RT_PRIO);
1387 plist_node_init(&q->list, prio);
1388 #ifdef CONFIG_DEBUG_PI_LIST
1389 q->list.plist.lock = &hb->lock;
1391 plist_add(&q->list, &hb->chain);
1393 spin_unlock(&hb->lock);
1397 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1398 * @q: The futex_q to unqueue
1400 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1401 * be paired with exactly one earlier call to queue_me().
1404 * 1 - if the futex_q was still queued (and we removed unqueued it)
1405 * 0 - if the futex_q was already removed by the waking thread
1407 static int unqueue_me(struct futex_q *q)
1409 spinlock_t *lock_ptr;
1412 /* In the common case we don't take the spinlock, which is nice. */
1414 lock_ptr = q->lock_ptr;
1416 if (lock_ptr != NULL) {
1417 spin_lock(lock_ptr);
1419 * q->lock_ptr can change between reading it and
1420 * spin_lock(), causing us to take the wrong lock. This
1421 * corrects the race condition.
1423 * Reasoning goes like this: if we have the wrong lock,
1424 * q->lock_ptr must have changed (maybe several times)
1425 * between reading it and the spin_lock(). It can
1426 * change again after the spin_lock() but only if it was
1427 * already changed before the spin_lock(). It cannot,
1428 * however, change back to the original value. Therefore
1429 * we can detect whether we acquired the correct lock.
1431 if (unlikely(lock_ptr != q->lock_ptr)) {
1432 spin_unlock(lock_ptr);
1435 WARN_ON(plist_node_empty(&q->list));
1436 plist_del(&q->list, &q->list.plist);
1438 BUG_ON(q->pi_state);
1440 spin_unlock(lock_ptr);
1444 drop_futex_key_refs(&q->key);
1449 * PI futexes can not be requeued and must remove themself from the
1450 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1453 static void unqueue_me_pi(struct futex_q *q)
1455 WARN_ON(plist_node_empty(&q->list));
1456 plist_del(&q->list, &q->list.plist);
1458 BUG_ON(!q->pi_state);
1459 free_pi_state(q->pi_state);
1462 spin_unlock(q->lock_ptr);
1464 drop_futex_key_refs(&q->key);
1468 * Fixup the pi_state owner with the new owner.
1470 * Must be called with hash bucket lock held and mm->sem held for non
1473 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1474 struct task_struct *newowner, int fshared)
1476 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1477 struct futex_pi_state *pi_state = q->pi_state;
1478 struct task_struct *oldowner = pi_state->owner;
1479 u32 uval, curval, newval;
1483 if (!pi_state->owner)
1484 newtid |= FUTEX_OWNER_DIED;
1487 * We are here either because we stole the rtmutex from the
1488 * pending owner or we are the pending owner which failed to
1489 * get the rtmutex. We have to replace the pending owner TID
1490 * in the user space variable. This must be atomic as we have
1491 * to preserve the owner died bit here.
1493 * Note: We write the user space value _before_ changing the pi_state
1494 * because we can fault here. Imagine swapped out pages or a fork
1495 * that marked all the anonymous memory readonly for cow.
1497 * Modifying pi_state _before_ the user space value would
1498 * leave the pi_state in an inconsistent state when we fault
1499 * here, because we need to drop the hash bucket lock to
1500 * handle the fault. This might be observed in the PID check
1501 * in lookup_pi_state.
1504 if (get_futex_value_locked(&uval, uaddr))
1508 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1510 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1512 if (curval == -EFAULT)
1520 * We fixed up user space. Now we need to fix the pi_state
1523 if (pi_state->owner != NULL) {
1524 spin_lock_irq(&pi_state->owner->pi_lock);
1525 WARN_ON(list_empty(&pi_state->list));
1526 list_del_init(&pi_state->list);
1527 spin_unlock_irq(&pi_state->owner->pi_lock);
1530 pi_state->owner = newowner;
1532 spin_lock_irq(&newowner->pi_lock);
1533 WARN_ON(!list_empty(&pi_state->list));
1534 list_add(&pi_state->list, &newowner->pi_state_list);
1535 spin_unlock_irq(&newowner->pi_lock);
1539 * To handle the page fault we need to drop the hash bucket
1540 * lock here. That gives the other task (either the pending
1541 * owner itself or the task which stole the rtmutex) the
1542 * chance to try the fixup of the pi_state. So once we are
1543 * back from handling the fault we need to check the pi_state
1544 * after reacquiring the hash bucket lock and before trying to
1545 * do another fixup. When the fixup has been done already we
1549 spin_unlock(q->lock_ptr);
1551 ret = fault_in_user_writeable(uaddr);
1553 spin_lock(q->lock_ptr);
1556 * Check if someone else fixed it for us:
1558 if (pi_state->owner != oldowner)
1568 * In case we must use restart_block to restart a futex_wait,
1569 * we encode in the 'flags' shared capability
1571 #define FLAGS_SHARED 0x01
1572 #define FLAGS_CLOCKRT 0x02
1573 #define FLAGS_HAS_TIMEOUT 0x04
1575 static long futex_wait_restart(struct restart_block *restart);
1578 * fixup_owner() - Post lock pi_state and corner case management
1579 * @uaddr: user address of the futex
1580 * @fshared: whether the futex is shared (1) or not (0)
1581 * @q: futex_q (contains pi_state and access to the rt_mutex)
1582 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1584 * After attempting to lock an rt_mutex, this function is called to cleanup
1585 * the pi_state owner as well as handle race conditions that may allow us to
1586 * acquire the lock. Must be called with the hb lock held.
1589 * 1 - success, lock taken
1590 * 0 - success, lock not taken
1591 * <0 - on error (-EFAULT)
1593 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1596 struct task_struct *owner;
1601 * Got the lock. We might not be the anticipated owner if we
1602 * did a lock-steal - fix up the PI-state in that case:
1604 if (q->pi_state->owner != current)
1605 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1610 * Catch the rare case, where the lock was released when we were on the
1611 * way back before we locked the hash bucket.
1613 if (q->pi_state->owner == current) {
1615 * Try to get the rt_mutex now. This might fail as some other
1616 * task acquired the rt_mutex after we removed ourself from the
1617 * rt_mutex waiters list.
1619 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1625 * pi_state is incorrect, some other task did a lock steal and
1626 * we returned due to timeout or signal without taking the
1627 * rt_mutex. Too late. We can access the rt_mutex_owner without
1628 * locking, as the other task is now blocked on the hash bucket
1629 * lock. Fix the state up.
1631 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1632 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1637 * Paranoia check. If we did not take the lock, then we should not be
1638 * the owner, nor the pending owner, of the rt_mutex.
1640 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1641 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1642 "pi-state %p\n", ret,
1643 q->pi_state->pi_mutex.owner,
1644 q->pi_state->owner);
1647 return ret ? ret : locked;
1651 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1652 * @hb: the futex hash bucket, must be locked by the caller
1653 * @q: the futex_q to queue up on
1654 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1656 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1657 struct hrtimer_sleeper *timeout)
1660 * The task state is guaranteed to be set before another task can
1661 * wake it. set_current_state() is implemented using set_mb() and
1662 * queue_me() calls spin_unlock() upon completion, both serializing
1663 * access to the hash list and forcing another memory barrier.
1665 set_current_state(TASK_INTERRUPTIBLE);
1670 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1671 if (!hrtimer_active(&timeout->timer))
1672 timeout->task = NULL;
1676 * If we have been removed from the hash list, then another task
1677 * has tried to wake us, and we can skip the call to schedule().
1679 if (likely(!plist_node_empty(&q->list))) {
1681 * If the timer has already expired, current will already be
1682 * flagged for rescheduling. Only call schedule if there
1683 * is no timeout, or if it has yet to expire.
1685 if (!timeout || timeout->task)
1688 __set_current_state(TASK_RUNNING);
1692 * futex_wait_setup() - Prepare to wait on a futex
1693 * @uaddr: the futex userspace address
1694 * @val: the expected value
1695 * @fshared: whether the futex is shared (1) or not (0)
1696 * @q: the associated futex_q
1697 * @hb: storage for hash_bucket pointer to be returned to caller
1699 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1700 * compare it with the expected value. Handle atomic faults internally.
1701 * Return with the hb lock held and a q.key reference on success, and unlocked
1702 * with no q.key reference on failure.
1705 * 0 - uaddr contains val and hb has been locked
1706 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1708 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1709 struct futex_q *q, struct futex_hash_bucket **hb)
1715 * Access the page AFTER the hash-bucket is locked.
1716 * Order is important:
1718 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1719 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1721 * The basic logical guarantee of a futex is that it blocks ONLY
1722 * if cond(var) is known to be true at the time of blocking, for
1723 * any cond. If we queued after testing *uaddr, that would open
1724 * a race condition where we could block indefinitely with
1725 * cond(var) false, which would violate the guarantee.
1727 * A consequence is that futex_wait() can return zero and absorb
1728 * a wakeup when *uaddr != val on entry to the syscall. This is
1732 q->key = FUTEX_KEY_INIT;
1733 ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1734 if (unlikely(ret != 0))
1738 *hb = queue_lock(q);
1740 ret = get_futex_value_locked(&uval, uaddr);
1743 queue_unlock(q, *hb);
1745 ret = get_user(uval, uaddr);
1752 put_futex_key(fshared, &q->key);
1757 queue_unlock(q, *hb);
1763 put_futex_key(fshared, &q->key);
1767 static int futex_wait(u32 __user *uaddr, int fshared,
1768 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1770 struct hrtimer_sleeper timeout, *to = NULL;
1771 struct restart_block *restart;
1772 struct futex_hash_bucket *hb;
1782 q.requeue_pi_key = NULL;
1787 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1788 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1789 hrtimer_init_sleeper(to, current);
1790 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1791 current->timer_slack_ns);
1794 /* Prepare to wait on uaddr. */
1795 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1799 /* queue_me and wait for wakeup, timeout, or a signal. */
1800 futex_wait_queue_me(hb, &q, to);
1802 /* If we were woken (and unqueued), we succeeded, whatever. */
1804 if (!unqueue_me(&q))
1807 if (to && !to->task)
1811 * We expect signal_pending(current), but another thread may
1812 * have handled it for us already.
1818 restart = ¤t_thread_info()->restart_block;
1819 restart->fn = futex_wait_restart;
1820 restart->futex.uaddr = (u32 *)uaddr;
1821 restart->futex.val = val;
1822 restart->futex.time = abs_time->tv64;
1823 restart->futex.bitset = bitset;
1824 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1827 restart->futex.flags |= FLAGS_SHARED;
1829 restart->futex.flags |= FLAGS_CLOCKRT;
1831 ret = -ERESTART_RESTARTBLOCK;
1834 put_futex_key(fshared, &q.key);
1837 hrtimer_cancel(&to->timer);
1838 destroy_hrtimer_on_stack(&to->timer);
1844 static long futex_wait_restart(struct restart_block *restart)
1846 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1848 ktime_t t, *tp = NULL;
1850 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1851 t.tv64 = restart->futex.time;
1854 restart->fn = do_no_restart_syscall;
1855 if (restart->futex.flags & FLAGS_SHARED)
1857 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1858 restart->futex.bitset,
1859 restart->futex.flags & FLAGS_CLOCKRT);
1864 * Userspace tried a 0 -> TID atomic transition of the futex value
1865 * and failed. The kernel side here does the whole locking operation:
1866 * if there are waiters then it will block, it does PI, etc. (Due to
1867 * races the kernel might see a 0 value of the futex too.)
1869 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1870 int detect, ktime_t *time, int trylock)
1872 struct hrtimer_sleeper timeout, *to = NULL;
1873 struct futex_hash_bucket *hb;
1877 if (refill_pi_state_cache())
1882 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1884 hrtimer_init_sleeper(to, current);
1885 hrtimer_set_expires(&to->timer, *time);
1890 q.requeue_pi_key = NULL;
1892 q.key = FUTEX_KEY_INIT;
1893 ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1894 if (unlikely(ret != 0))
1898 hb = queue_lock(&q);
1900 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1901 if (unlikely(ret)) {
1904 /* We got the lock. */
1906 goto out_unlock_put_key;
1911 * Task is exiting and we just wait for the
1914 queue_unlock(&q, hb);
1915 put_futex_key(fshared, &q.key);
1919 goto out_unlock_put_key;
1924 * Only actually queue now that the atomic ops are done:
1928 WARN_ON(!q.pi_state);
1930 * Block on the PI mutex:
1933 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1935 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1936 /* Fixup the trylock return value: */
1937 ret = ret ? 0 : -EWOULDBLOCK;
1940 spin_lock(q.lock_ptr);
1942 * Fixup the pi_state owner and possibly acquire the lock if we
1945 res = fixup_owner(uaddr, fshared, &q, !ret);
1947 * If fixup_owner() returned an error, proprogate that. If it acquired
1948 * the lock, clear our -ETIMEDOUT or -EINTR.
1951 ret = (res < 0) ? res : 0;
1954 * If fixup_owner() faulted and was unable to handle the fault, unlock
1955 * it and return the fault to userspace.
1957 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1958 rt_mutex_unlock(&q.pi_state->pi_mutex);
1960 /* Unqueue and drop the lock */
1966 queue_unlock(&q, hb);
1969 put_futex_key(fshared, &q.key);
1972 destroy_hrtimer_on_stack(&to->timer);
1973 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1976 queue_unlock(&q, hb);
1978 ret = fault_in_user_writeable(uaddr);
1985 put_futex_key(fshared, &q.key);
1990 * Userspace attempted a TID -> 0 atomic transition, and failed.
1991 * This is the in-kernel slowpath: we look up the PI state (if any),
1992 * and do the rt-mutex unlock.
1994 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1996 struct futex_hash_bucket *hb;
1997 struct futex_q *this, *next;
1999 struct plist_head *head;
2000 union futex_key key = FUTEX_KEY_INIT;
2004 if (get_user(uval, uaddr))
2007 * We release only a lock we actually own:
2009 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2012 ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
2013 if (unlikely(ret != 0))
2016 hb = hash_futex(&key);
2017 spin_lock(&hb->lock);
2020 * To avoid races, try to do the TID -> 0 atomic transition
2021 * again. If it succeeds then we can return without waking
2024 if (!(uval & FUTEX_OWNER_DIED))
2025 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2028 if (unlikely(uval == -EFAULT))
2031 * Rare case: we managed to release the lock atomically,
2032 * no need to wake anyone else up:
2034 if (unlikely(uval == task_pid_vnr(current)))
2038 * Ok, other tasks may need to be woken up - check waiters
2039 * and do the wakeup if necessary:
2043 plist_for_each_entry_safe(this, next, head, list) {
2044 if (!match_futex (&this->key, &key))
2046 ret = wake_futex_pi(uaddr, uval, this);
2048 * The atomic access to the futex value
2049 * generated a pagefault, so retry the
2050 * user-access and the wakeup:
2057 * No waiters - kernel unlocks the futex:
2059 if (!(uval & FUTEX_OWNER_DIED)) {
2060 ret = unlock_futex_pi(uaddr, uval);
2066 spin_unlock(&hb->lock);
2067 put_futex_key(fshared, &key);
2073 spin_unlock(&hb->lock);
2074 put_futex_key(fshared, &key);
2076 ret = fault_in_user_writeable(uaddr);
2084 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2085 * @hb: the hash_bucket futex_q was original enqueued on
2086 * @q: the futex_q woken while waiting to be requeued
2087 * @key2: the futex_key of the requeue target futex
2088 * @timeout: the timeout associated with the wait (NULL if none)
2090 * Detect if the task was woken on the initial futex as opposed to the requeue
2091 * target futex. If so, determine if it was a timeout or a signal that caused
2092 * the wakeup and return the appropriate error code to the caller. Must be
2093 * called with the hb lock held.
2096 * 0 - no early wakeup detected
2097 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2100 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2101 struct futex_q *q, union futex_key *key2,
2102 struct hrtimer_sleeper *timeout)
2107 * With the hb lock held, we avoid races while we process the wakeup.
2108 * We only need to hold hb (and not hb2) to ensure atomicity as the
2109 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2110 * It can't be requeued from uaddr2 to something else since we don't
2111 * support a PI aware source futex for requeue.
2113 if (!match_futex(&q->key, key2)) {
2114 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2116 * We were woken prior to requeue by a timeout or a signal.
2117 * Unqueue the futex_q and determine which it was.
2119 plist_del(&q->list, &q->list.plist);
2121 if (timeout && !timeout->task)
2124 ret = -ERESTARTNOINTR;
2130 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2131 * @uaddr: the futex we initially wait on (non-pi)
2132 * @fshared: whether the futexes are shared (1) or not (0). They must be
2133 * the same type, no requeueing from private to shared, etc.
2134 * @val: the expected value of uaddr
2135 * @abs_time: absolute timeout
2136 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2137 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2138 * @uaddr2: the pi futex we will take prior to returning to user-space
2140 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2141 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2142 * complete the acquisition of the rt_mutex prior to returning to userspace.
2143 * This ensures the rt_mutex maintains an owner when it has waiters; without
2144 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2147 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2148 * via the following:
2149 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2150 * 2) wakeup on uaddr2 after a requeue
2154 * If 3, cleanup and return -ERESTARTNOINTR.
2156 * If 2, we may then block on trying to take the rt_mutex and return via:
2157 * 5) successful lock
2160 * 8) other lock acquisition failure
2162 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2164 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2170 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2171 u32 val, ktime_t *abs_time, u32 bitset,
2172 int clockrt, u32 __user *uaddr2)
2174 struct hrtimer_sleeper timeout, *to = NULL;
2175 struct rt_mutex_waiter rt_waiter;
2176 struct rt_mutex *pi_mutex = NULL;
2177 struct futex_hash_bucket *hb;
2178 union futex_key key2;
2187 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2188 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2189 hrtimer_init_sleeper(to, current);
2190 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2191 current->timer_slack_ns);
2195 * The waiter is allocated on our stack, manipulated by the requeue
2196 * code while we sleep on uaddr.
2198 debug_rt_mutex_init_waiter(&rt_waiter);
2199 rt_waiter.task = NULL;
2201 key2 = FUTEX_KEY_INIT;
2202 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2203 if (unlikely(ret != 0))
2208 q.rt_waiter = &rt_waiter;
2209 q.requeue_pi_key = &key2;
2211 /* Prepare to wait on uaddr. */
2212 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2216 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2217 futex_wait_queue_me(hb, &q, to);
2219 spin_lock(&hb->lock);
2220 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2221 spin_unlock(&hb->lock);
2226 * In order for us to be here, we know our q.key == key2, and since
2227 * we took the hb->lock above, we also know that futex_requeue() has
2228 * completed and we no longer have to concern ourselves with a wakeup
2229 * race with the atomic proxy lock acquition by the requeue code.
2232 /* Check if the requeue code acquired the second futex for us. */
2235 * Got the lock. We might not be the anticipated owner if we
2236 * did a lock-steal - fix up the PI-state in that case.
2238 if (q.pi_state && (q.pi_state->owner != current)) {
2239 spin_lock(q.lock_ptr);
2240 ret = fixup_pi_state_owner(uaddr2, &q, current,
2242 spin_unlock(q.lock_ptr);
2246 * We have been woken up by futex_unlock_pi(), a timeout, or a
2247 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2250 WARN_ON(!&q.pi_state);
2251 pi_mutex = &q.pi_state->pi_mutex;
2252 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2253 debug_rt_mutex_free_waiter(&rt_waiter);
2255 spin_lock(q.lock_ptr);
2257 * Fixup the pi_state owner and possibly acquire the lock if we
2260 res = fixup_owner(uaddr2, fshared, &q, !ret);
2262 * If fixup_owner() returned an error, proprogate that. If it
2263 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2266 ret = (res < 0) ? res : 0;
2268 /* Unqueue and drop the lock. */
2273 * If fixup_pi_state_owner() faulted and was unable to handle the
2274 * fault, unlock the rt_mutex and return the fault to userspace.
2276 if (ret == -EFAULT) {
2277 if (rt_mutex_owner(pi_mutex) == current)
2278 rt_mutex_unlock(pi_mutex);
2279 } else if (ret == -EINTR) {
2281 * We've already been requeued, but cannot restart by calling
2282 * futex_lock_pi() directly. We could restart this syscall, but
2283 * it would detect that the user space "val" changed and return
2284 * -EWOULDBLOCK. Save the overhead of the restart and return
2285 * -EWOULDBLOCK directly.
2291 put_futex_key(fshared, &q.key);
2293 put_futex_key(fshared, &key2);
2297 hrtimer_cancel(&to->timer);
2298 destroy_hrtimer_on_stack(&to->timer);
2304 * Support for robust futexes: the kernel cleans up held futexes at
2307 * Implementation: user-space maintains a per-thread list of locks it
2308 * is holding. Upon do_exit(), the kernel carefully walks this list,
2309 * and marks all locks that are owned by this thread with the
2310 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2311 * always manipulated with the lock held, so the list is private and
2312 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2313 * field, to allow the kernel to clean up if the thread dies after
2314 * acquiring the lock, but just before it could have added itself to
2315 * the list. There can only be one such pending lock.
2319 * sys_set_robust_list() - Set the robust-futex list head of a task
2320 * @head: pointer to the list-head
2321 * @len: length of the list-head, as userspace expects
2323 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2326 if (!futex_cmpxchg_enabled)
2329 * The kernel knows only one size for now:
2331 if (unlikely(len != sizeof(*head)))
2334 current->robust_list = head;
2340 * sys_get_robust_list() - Get the robust-futex list head of a task
2341 * @pid: pid of the process [zero for current task]
2342 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2343 * @len_ptr: pointer to a length field, the kernel fills in the header size
2345 SYSCALL_DEFINE3(get_robust_list, int, pid,
2346 struct robust_list_head __user * __user *, head_ptr,
2347 size_t __user *, len_ptr)
2349 struct robust_list_head __user *head;
2351 const struct cred *cred = current_cred(), *pcred;
2353 if (!futex_cmpxchg_enabled)
2357 head = current->robust_list;
2359 struct task_struct *p;
2363 p = find_task_by_vpid(pid);
2367 pcred = __task_cred(p);
2368 if (cred->euid != pcred->euid &&
2369 cred->euid != pcred->uid &&
2370 !capable(CAP_SYS_PTRACE))
2372 head = p->robust_list;
2376 if (put_user(sizeof(*head), len_ptr))
2378 return put_user(head, head_ptr);
2387 * Process a futex-list entry, check whether it's owned by the
2388 * dying task, and do notification if so:
2390 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2392 u32 uval, nval, mval;
2395 if (get_user(uval, uaddr))
2398 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2400 * Ok, this dying thread is truly holding a futex
2401 * of interest. Set the OWNER_DIED bit atomically
2402 * via cmpxchg, and if the value had FUTEX_WAITERS
2403 * set, wake up a waiter (if any). (We have to do a
2404 * futex_wake() even if OWNER_DIED is already set -
2405 * to handle the rare but possible case of recursive
2406 * thread-death.) The rest of the cleanup is done in
2409 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2410 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2412 if (nval == -EFAULT)
2419 * Wake robust non-PI futexes here. The wakeup of
2420 * PI futexes happens in exit_pi_state():
2422 if (!pi && (uval & FUTEX_WAITERS))
2423 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2429 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2431 static inline int fetch_robust_entry(struct robust_list __user **entry,
2432 struct robust_list __user * __user *head,
2435 unsigned long uentry;
2437 if (get_user(uentry, (unsigned long __user *)head))
2440 *entry = (void __user *)(uentry & ~1UL);
2447 * Walk curr->robust_list (very carefully, it's a userspace list!)
2448 * and mark any locks found there dead, and notify any waiters.
2450 * We silently return on any sign of list-walking problem.
2452 void exit_robust_list(struct task_struct *curr)
2454 struct robust_list_head __user *head = curr->robust_list;
2455 struct robust_list __user *entry, *next_entry, *pending;
2456 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2457 unsigned long futex_offset;
2460 if (!futex_cmpxchg_enabled)
2464 * Fetch the list head (which was registered earlier, via
2465 * sys_set_robust_list()):
2467 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2470 * Fetch the relative futex offset:
2472 if (get_user(futex_offset, &head->futex_offset))
2475 * Fetch any possibly pending lock-add first, and handle it
2478 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2481 next_entry = NULL; /* avoid warning with gcc */
2482 while (entry != &head->list) {
2484 * Fetch the next entry in the list before calling
2485 * handle_futex_death:
2487 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2489 * A pending lock might already be on the list, so
2490 * don't process it twice:
2492 if (entry != pending)
2493 if (handle_futex_death((void __user *)entry + futex_offset,
2501 * Avoid excessively long or circular lists:
2510 handle_futex_death((void __user *)pending + futex_offset,
2514 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2515 u32 __user *uaddr2, u32 val2, u32 val3)
2517 int clockrt, ret = -ENOSYS;
2518 int cmd = op & FUTEX_CMD_MASK;
2521 if (!(op & FUTEX_PRIVATE_FLAG))
2524 clockrt = op & FUTEX_CLOCK_REALTIME;
2525 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2530 val3 = FUTEX_BITSET_MATCH_ANY;
2531 case FUTEX_WAIT_BITSET:
2532 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2535 val3 = FUTEX_BITSET_MATCH_ANY;
2536 case FUTEX_WAKE_BITSET:
2537 ret = futex_wake(uaddr, fshared, val, val3);
2540 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2542 case FUTEX_CMP_REQUEUE:
2543 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2547 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2550 if (futex_cmpxchg_enabled)
2551 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2553 case FUTEX_UNLOCK_PI:
2554 if (futex_cmpxchg_enabled)
2555 ret = futex_unlock_pi(uaddr, fshared);
2557 case FUTEX_TRYLOCK_PI:
2558 if (futex_cmpxchg_enabled)
2559 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2561 case FUTEX_WAIT_REQUEUE_PI:
2562 val3 = FUTEX_BITSET_MATCH_ANY;
2563 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2566 case FUTEX_CMP_REQUEUE_PI:
2567 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2577 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2578 struct timespec __user *, utime, u32 __user *, uaddr2,
2582 ktime_t t, *tp = NULL;
2584 int cmd = op & FUTEX_CMD_MASK;
2586 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2587 cmd == FUTEX_WAIT_BITSET ||
2588 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2589 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2591 if (!timespec_valid(&ts))
2594 t = timespec_to_ktime(ts);
2595 if (cmd == FUTEX_WAIT)
2596 t = ktime_add_safe(ktime_get(), t);
2600 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2601 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2603 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2604 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2605 val2 = (u32) (unsigned long) utime;
2607 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2610 static int __init futex_init(void)
2616 * This will fail and we want it. Some arch implementations do
2617 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2618 * functionality. We want to know that before we call in any
2619 * of the complex code paths. Also we want to prevent
2620 * registration of robust lists in that case. NULL is
2621 * guaranteed to fault and we get -EFAULT on functional
2622 * implementation, the non functional ones will return
2625 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2626 if (curval == -EFAULT)
2627 futex_cmpxchg_enabled = 1;
2629 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2630 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2631 spin_lock_init(&futex_queues[i].lock);
2636 __initcall(futex_init);