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)
154 && key1->both.word == key2->both.word
155 && key1->both.ptr == key2->both.ptr
156 && key1->both.offset == key2->both.offset);
160 * Take a reference to the resource addressed by a key.
161 * Can be called while holding spinlocks.
164 static void get_futex_key_refs(union futex_key *key)
169 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
171 atomic_inc(&key->shared.inode->i_count);
173 case FUT_OFF_MMSHARED:
174 atomic_inc(&key->private.mm->mm_count);
180 * Drop a reference to the resource addressed by a key.
181 * The hash bucket spinlock must not be held.
183 static void drop_futex_key_refs(union futex_key *key)
185 if (!key->both.ptr) {
186 /* If we're here then we tried to put a key we failed to get */
191 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
193 iput(key->shared.inode);
195 case FUT_OFF_MMSHARED:
196 mmdrop(key->private.mm);
202 * get_futex_key() - Get parameters which are the keys for a futex
203 * @uaddr: virtual address of the futex
204 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205 * @key: address where result is stored.
206 * @rw: mapping needs to be read/write (values: VERIFY_READ,
209 * Returns a negative error code or 0
210 * The key words are stored in *key on success.
212 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
213 * offset_within_page). For private mappings, it's (uaddr, current->mm).
214 * We can usually work out the index without swapping in the page.
216 * lock_page() might sleep, the caller should not hold a spinlock.
219 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
221 unsigned long address = (unsigned long)uaddr;
222 struct mm_struct *mm = current->mm;
227 * The futex address must be "naturally" aligned.
229 key->both.offset = address % PAGE_SIZE;
230 if (unlikely((address % sizeof(u32)) != 0))
232 address -= key->both.offset;
235 * PROCESS_PRIVATE futexes are fast.
236 * As the mm cannot disappear under us and the 'key' only needs
237 * virtual address, we dont even have to find the underlying vma.
238 * Note : We do have to check 'uaddr' is a valid user address,
239 * but access_ok() should be faster than find_vma()
242 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
244 key->private.mm = mm;
245 key->private.address = address;
246 get_futex_key_refs(key);
251 err = get_user_pages_fast(address, 1, 1, &page);
253 * If write access is not required (eg. FUTEX_WAIT), try
254 * and get read-only access.
256 if (err == -EFAULT && rw == VERIFY_READ) {
257 err = get_user_pages_fast(address, 1, 0, &page);
265 page = compound_head(page);
267 if (!page->mapping) {
271 * ZERO_PAGE pages don't have a mapping. Avoid a busy loop
272 * trying to find one. RW mapping would have COW'd (and thus
273 * have a mapping) so this page is RO and won't ever change.
275 if ((page == ZERO_PAGE(address)))
281 * Private mappings are handled in a simple way.
283 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
284 * it's a read-only handle, it's expected that futexes attach to
285 * the object not the particular process.
287 if (PageAnon(page)) {
289 * A RO anonymous page will never change and thus doesn't make
290 * sense for futex operations.
297 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
298 key->private.mm = mm;
299 key->private.address = address;
301 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
302 key->shared.inode = page->mapping->host;
303 key->shared.pgoff = page->index;
306 get_futex_key_refs(key);
315 void put_futex_key(int fshared, union futex_key *key)
317 drop_futex_key_refs(key);
321 * fault_in_user_writeable() - Fault in user address and verify RW access
322 * @uaddr: pointer to faulting user space address
324 * Slow path to fixup the fault we just took in the atomic write
327 * We have no generic implementation of a non destructive write to the
328 * user address. We know that we faulted in the atomic pagefault
329 * disabled section so we can as well avoid the #PF overhead by
330 * calling get_user_pages() right away.
332 static int fault_in_user_writeable(u32 __user *uaddr)
334 struct mm_struct *mm = current->mm;
337 down_read(&mm->mmap_sem);
338 ret = get_user_pages(current, mm, (unsigned long)uaddr,
339 1, 1, 0, NULL, NULL);
340 up_read(&mm->mmap_sem);
342 return ret < 0 ? ret : 0;
346 * futex_top_waiter() - Return the highest priority waiter on a futex
347 * @hb: the hash bucket the futex_q's reside in
348 * @key: the futex key (to distinguish it from other futex futex_q's)
350 * Must be called with the hb lock held.
352 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
353 union futex_key *key)
355 struct futex_q *this;
357 plist_for_each_entry(this, &hb->chain, list) {
358 if (match_futex(&this->key, key))
364 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
369 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
375 static int get_futex_value_locked(u32 *dest, u32 __user *from)
380 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
383 return ret ? -EFAULT : 0;
390 static int refill_pi_state_cache(void)
392 struct futex_pi_state *pi_state;
394 if (likely(current->pi_state_cache))
397 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
402 INIT_LIST_HEAD(&pi_state->list);
403 /* pi_mutex gets initialized later */
404 pi_state->owner = NULL;
405 atomic_set(&pi_state->refcount, 1);
406 pi_state->key = FUTEX_KEY_INIT;
408 current->pi_state_cache = pi_state;
413 static struct futex_pi_state * alloc_pi_state(void)
415 struct futex_pi_state *pi_state = current->pi_state_cache;
418 current->pi_state_cache = NULL;
423 static void free_pi_state(struct futex_pi_state *pi_state)
425 if (!atomic_dec_and_test(&pi_state->refcount))
429 * If pi_state->owner is NULL, the owner is most probably dying
430 * and has cleaned up the pi_state already
432 if (pi_state->owner) {
433 spin_lock_irq(&pi_state->owner->pi_lock);
434 list_del_init(&pi_state->list);
435 spin_unlock_irq(&pi_state->owner->pi_lock);
437 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
440 if (current->pi_state_cache)
444 * pi_state->list is already empty.
445 * clear pi_state->owner.
446 * refcount is at 0 - put it back to 1.
448 pi_state->owner = NULL;
449 atomic_set(&pi_state->refcount, 1);
450 current->pi_state_cache = pi_state;
455 * Look up the task based on what TID userspace gave us.
458 static struct task_struct * futex_find_get_task(pid_t pid)
460 struct task_struct *p;
463 p = find_task_by_vpid(pid);
473 * This task is holding PI mutexes at exit time => bad.
474 * Kernel cleans up PI-state, but userspace is likely hosed.
475 * (Robust-futex cleanup is separate and might save the day for userspace.)
477 void exit_pi_state_list(struct task_struct *curr)
479 struct list_head *next, *head = &curr->pi_state_list;
480 struct futex_pi_state *pi_state;
481 struct futex_hash_bucket *hb;
482 union futex_key key = FUTEX_KEY_INIT;
484 if (!futex_cmpxchg_enabled)
487 * We are a ZOMBIE and nobody can enqueue itself on
488 * pi_state_list anymore, but we have to be careful
489 * versus waiters unqueueing themselves:
491 spin_lock_irq(&curr->pi_lock);
492 while (!list_empty(head)) {
495 pi_state = list_entry(next, struct futex_pi_state, list);
497 hb = hash_futex(&key);
498 spin_unlock_irq(&curr->pi_lock);
500 spin_lock(&hb->lock);
502 spin_lock_irq(&curr->pi_lock);
504 * We dropped the pi-lock, so re-check whether this
505 * task still owns the PI-state:
507 if (head->next != next) {
508 spin_unlock(&hb->lock);
512 WARN_ON(pi_state->owner != curr);
513 WARN_ON(list_empty(&pi_state->list));
514 list_del_init(&pi_state->list);
515 pi_state->owner = NULL;
516 spin_unlock_irq(&curr->pi_lock);
518 rt_mutex_unlock(&pi_state->pi_mutex);
520 spin_unlock(&hb->lock);
522 spin_lock_irq(&curr->pi_lock);
524 spin_unlock_irq(&curr->pi_lock);
528 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
529 union futex_key *key, struct futex_pi_state **ps)
531 struct futex_pi_state *pi_state = NULL;
532 struct futex_q *this, *next;
533 struct plist_head *head;
534 struct task_struct *p;
535 pid_t pid = uval & FUTEX_TID_MASK;
539 plist_for_each_entry_safe(this, next, head, list) {
540 if (match_futex(&this->key, key)) {
542 * Another waiter already exists - bump up
543 * the refcount and return its pi_state:
545 pi_state = this->pi_state;
547 * Userspace might have messed up non PI and PI futexes
549 if (unlikely(!pi_state))
552 WARN_ON(!atomic_read(&pi_state->refcount));
555 * When pi_state->owner is NULL then the owner died
556 * and another waiter is on the fly. pi_state->owner
557 * is fixed up by the task which acquires
558 * pi_state->rt_mutex.
560 * We do not check for pid == 0 which can happen when
561 * the owner died and robust_list_exit() cleared the
564 if (pid && pi_state->owner) {
566 * Bail out if user space manipulated the
569 if (pid != task_pid_vnr(pi_state->owner))
573 atomic_inc(&pi_state->refcount);
581 * We are the first waiter - try to look up the real owner and attach
582 * the new pi_state to it, but bail out when TID = 0
586 p = futex_find_get_task(pid);
591 * We need to look at the task state flags to figure out,
592 * whether the task is exiting. To protect against the do_exit
593 * change of the task flags, we do this protected by
596 spin_lock_irq(&p->pi_lock);
597 if (unlikely(p->flags & PF_EXITING)) {
599 * The task is on the way out. When PF_EXITPIDONE is
600 * set, we know that the task has finished the
603 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
605 spin_unlock_irq(&p->pi_lock);
610 pi_state = alloc_pi_state();
613 * Initialize the pi_mutex in locked state and make 'p'
616 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
618 /* Store the key for possible exit cleanups: */
619 pi_state->key = *key;
621 WARN_ON(!list_empty(&pi_state->list));
622 list_add(&pi_state->list, &p->pi_state_list);
624 spin_unlock_irq(&p->pi_lock);
634 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
635 * @uaddr: the pi futex user address
636 * @hb: the pi futex hash bucket
637 * @key: the futex key associated with uaddr and hb
638 * @ps: the pi_state pointer where we store the result of the
640 * @task: the task to perform the atomic lock work for. This will
641 * be "current" except in the case of requeue pi.
642 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
646 * 1 - acquired the lock
649 * The hb->lock and futex_key refs shall be held by the caller.
651 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
652 union futex_key *key,
653 struct futex_pi_state **ps,
654 struct task_struct *task, int set_waiters)
656 int lock_taken, ret, ownerdied = 0;
657 u32 uval, newval, curval;
660 ret = lock_taken = 0;
663 * To avoid races, we attempt to take the lock here again
664 * (by doing a 0 -> TID atomic cmpxchg), while holding all
665 * the locks. It will most likely not succeed.
667 newval = task_pid_vnr(task);
669 newval |= FUTEX_WAITERS;
671 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
673 if (unlikely(curval == -EFAULT))
679 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
683 * Surprise - we got the lock. Just return to userspace:
685 if (unlikely(!curval))
691 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
692 * to wake at the next unlock.
694 newval = curval | FUTEX_WAITERS;
697 * There are two cases, where a futex might have no owner (the
698 * owner TID is 0): OWNER_DIED. We take over the futex in this
699 * case. We also do an unconditional take over, when the owner
702 * This is safe as we are protected by the hash bucket lock !
704 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
705 /* Keep the OWNER_DIED bit */
706 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
711 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
713 if (unlikely(curval == -EFAULT))
715 if (unlikely(curval != uval))
719 * We took the lock due to owner died take over.
721 if (unlikely(lock_taken))
725 * We dont have the lock. Look up the PI state (or create it if
726 * we are the first waiter):
728 ret = lookup_pi_state(uval, hb, key, ps);
734 * No owner found for this futex. Check if the
735 * OWNER_DIED bit is set to figure out whether
736 * this is a robust futex or not.
738 if (get_futex_value_locked(&curval, uaddr))
742 * We simply start over in case of a robust
743 * futex. The code above will take the futex
746 if (curval & FUTEX_OWNER_DIED) {
759 * The hash bucket lock must be held when this is called.
760 * Afterwards, the futex_q must not be accessed.
762 static void wake_futex(struct futex_q *q)
764 struct task_struct *p = q->task;
767 * We set q->lock_ptr = NULL _before_ we wake up the task. If
768 * a non futex wake up happens on another CPU then the task
769 * might exit and p would dereference a non existing task
770 * struct. Prevent this by holding a reference on p across the
775 plist_del(&q->list, &q->list.plist);
777 * The waiting task can free the futex_q as soon as
778 * q->lock_ptr = NULL is written, without taking any locks. A
779 * memory barrier is required here to prevent the following
780 * store to lock_ptr from getting ahead of the plist_del.
785 wake_up_state(p, TASK_NORMAL);
789 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
791 struct task_struct *new_owner;
792 struct futex_pi_state *pi_state = this->pi_state;
799 * If current does not own the pi_state then the futex is
800 * inconsistent and user space fiddled with the futex value.
802 if (pi_state->owner != current)
805 spin_lock(&pi_state->pi_mutex.wait_lock);
806 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
809 * This happens when we have stolen the lock and the original
810 * pending owner did not enqueue itself back on the rt_mutex.
811 * Thats not a tragedy. We know that way, that a lock waiter
812 * is on the fly. We make the futex_q waiter the pending owner.
815 new_owner = this->task;
818 * We pass it to the next owner. (The WAITERS bit is always
819 * kept enabled while there is PI state around. We must also
820 * preserve the owner died bit.)
822 if (!(uval & FUTEX_OWNER_DIED)) {
825 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
827 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
829 if (curval == -EFAULT)
831 else if (curval != uval)
834 spin_unlock(&pi_state->pi_mutex.wait_lock);
839 spin_lock_irq(&pi_state->owner->pi_lock);
840 WARN_ON(list_empty(&pi_state->list));
841 list_del_init(&pi_state->list);
842 spin_unlock_irq(&pi_state->owner->pi_lock);
844 spin_lock_irq(&new_owner->pi_lock);
845 WARN_ON(!list_empty(&pi_state->list));
846 list_add(&pi_state->list, &new_owner->pi_state_list);
847 pi_state->owner = new_owner;
848 spin_unlock_irq(&new_owner->pi_lock);
850 spin_unlock(&pi_state->pi_mutex.wait_lock);
851 rt_mutex_unlock(&pi_state->pi_mutex);
856 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
861 * There is no waiter, so we unlock the futex. The owner died
862 * bit has not to be preserved here. We are the owner:
864 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
866 if (oldval == -EFAULT)
875 * Express the locking dependencies for lockdep:
878 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
881 spin_lock(&hb1->lock);
883 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
884 } else { /* hb1 > hb2 */
885 spin_lock(&hb2->lock);
886 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
891 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
893 spin_unlock(&hb1->lock);
895 spin_unlock(&hb2->lock);
899 * Wake up waiters matching bitset queued on this futex (uaddr).
901 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
903 struct futex_hash_bucket *hb;
904 struct futex_q *this, *next;
905 struct plist_head *head;
906 union futex_key key = FUTEX_KEY_INIT;
912 ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
913 if (unlikely(ret != 0))
916 hb = hash_futex(&key);
917 spin_lock(&hb->lock);
920 plist_for_each_entry_safe(this, next, head, list) {
921 if (match_futex (&this->key, &key)) {
922 if (this->pi_state || this->rt_waiter) {
927 /* Check if one of the bits is set in both bitsets */
928 if (!(this->bitset & bitset))
932 if (++ret >= nr_wake)
937 spin_unlock(&hb->lock);
938 put_futex_key(fshared, &key);
944 * Wake up all waiters hashed on the physical page that is mapped
945 * to this virtual address:
948 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
949 int nr_wake, int nr_wake2, int op)
951 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
952 struct futex_hash_bucket *hb1, *hb2;
953 struct plist_head *head;
954 struct futex_q *this, *next;
958 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
959 if (unlikely(ret != 0))
961 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
962 if (unlikely(ret != 0))
965 hb1 = hash_futex(&key1);
966 hb2 = hash_futex(&key2);
969 double_lock_hb(hb1, hb2);
970 op_ret = futex_atomic_op_inuser(op, uaddr2);
971 if (unlikely(op_ret < 0)) {
973 double_unlock_hb(hb1, hb2);
977 * we don't get EFAULT from MMU faults if we don't have an MMU,
978 * but we might get them from range checking
984 if (unlikely(op_ret != -EFAULT)) {
989 ret = fault_in_user_writeable(uaddr2);
996 put_futex_key(fshared, &key2);
997 put_futex_key(fshared, &key1);
1003 plist_for_each_entry_safe(this, next, head, list) {
1004 if (match_futex (&this->key, &key1)) {
1006 if (++ret >= nr_wake)
1015 plist_for_each_entry_safe(this, next, head, list) {
1016 if (match_futex (&this->key, &key2)) {
1018 if (++op_ret >= nr_wake2)
1025 double_unlock_hb(hb1, hb2);
1027 put_futex_key(fshared, &key2);
1029 put_futex_key(fshared, &key1);
1035 * requeue_futex() - Requeue a futex_q from one hb to another
1036 * @q: the futex_q to requeue
1037 * @hb1: the source hash_bucket
1038 * @hb2: the target hash_bucket
1039 * @key2: the new key for the requeued futex_q
1042 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1043 struct futex_hash_bucket *hb2, union futex_key *key2)
1047 * If key1 and key2 hash to the same bucket, no need to
1050 if (likely(&hb1->chain != &hb2->chain)) {
1051 plist_del(&q->list, &hb1->chain);
1052 plist_add(&q->list, &hb2->chain);
1053 q->lock_ptr = &hb2->lock;
1054 #ifdef CONFIG_DEBUG_PI_LIST
1055 q->list.plist.lock = &hb2->lock;
1058 get_futex_key_refs(key2);
1063 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1065 * @key: the key of the requeue target futex
1066 * @hb: the hash_bucket of the requeue target futex
1068 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1069 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1070 * to the requeue target futex so the waiter can detect the wakeup on the right
1071 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1072 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1073 * to protect access to the pi_state to fixup the owner later. Must be called
1074 * with both q->lock_ptr and hb->lock held.
1077 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1078 struct futex_hash_bucket *hb)
1080 get_futex_key_refs(key);
1083 WARN_ON(plist_node_empty(&q->list));
1084 plist_del(&q->list, &q->list.plist);
1086 WARN_ON(!q->rt_waiter);
1087 q->rt_waiter = NULL;
1089 q->lock_ptr = &hb->lock;
1090 #ifdef CONFIG_DEBUG_PI_LIST
1091 q->list.plist.lock = &hb->lock;
1094 wake_up_state(q->task, TASK_NORMAL);
1098 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1099 * @pifutex: the user address of the to futex
1100 * @hb1: the from futex hash bucket, must be locked by the caller
1101 * @hb2: the to futex hash bucket, must be locked by the caller
1102 * @key1: the from futex key
1103 * @key2: the to futex key
1104 * @ps: address to store the pi_state pointer
1105 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1107 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1108 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1109 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1110 * hb1 and hb2 must be held by the caller.
1113 * 0 - failed to acquire the lock atomicly
1114 * 1 - acquired the lock
1117 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1118 struct futex_hash_bucket *hb1,
1119 struct futex_hash_bucket *hb2,
1120 union futex_key *key1, union futex_key *key2,
1121 struct futex_pi_state **ps, int set_waiters)
1123 struct futex_q *top_waiter = NULL;
1127 if (get_futex_value_locked(&curval, pifutex))
1131 * Find the top_waiter and determine if there are additional waiters.
1132 * If the caller intends to requeue more than 1 waiter to pifutex,
1133 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1134 * as we have means to handle the possible fault. If not, don't set
1135 * the bit unecessarily as it will force the subsequent unlock to enter
1138 top_waiter = futex_top_waiter(hb1, key1);
1140 /* There are no waiters, nothing for us to do. */
1144 /* Ensure we requeue to the expected futex. */
1145 if (!match_futex(top_waiter->requeue_pi_key, key2))
1149 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1150 * the contended case or if set_waiters is 1. The pi_state is returned
1151 * in ps in contended cases.
1153 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1156 requeue_pi_wake_futex(top_waiter, key2, hb2);
1162 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1163 * uaddr1: source futex user address
1164 * uaddr2: target futex user address
1165 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1166 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1167 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1168 * pi futex (pi to pi requeue is not supported)
1170 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1171 * uaddr2 atomically on behalf of the top waiter.
1174 * >=0 - on success, the number of tasks requeued or woken
1177 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1178 int nr_wake, int nr_requeue, u32 *cmpval,
1181 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1182 int drop_count = 0, task_count = 0, ret;
1183 struct futex_pi_state *pi_state = NULL;
1184 struct futex_hash_bucket *hb1, *hb2;
1185 struct plist_head *head1;
1186 struct futex_q *this, *next;
1191 * requeue_pi requires a pi_state, try to allocate it now
1192 * without any locks in case it fails.
1194 if (refill_pi_state_cache())
1197 * requeue_pi must wake as many tasks as it can, up to nr_wake
1198 * + nr_requeue, since it acquires the rt_mutex prior to
1199 * returning to userspace, so as to not leave the rt_mutex with
1200 * waiters and no owner. However, second and third wake-ups
1201 * cannot be predicted as they involve race conditions with the
1202 * first wake and a fault while looking up the pi_state. Both
1203 * pthread_cond_signal() and pthread_cond_broadcast() should
1211 if (pi_state != NULL) {
1213 * We will have to lookup the pi_state again, so free this one
1214 * to keep the accounting correct.
1216 free_pi_state(pi_state);
1220 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1221 if (unlikely(ret != 0))
1223 ret = get_futex_key(uaddr2, fshared, &key2,
1224 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1225 if (unlikely(ret != 0))
1228 hb1 = hash_futex(&key1);
1229 hb2 = hash_futex(&key2);
1232 double_lock_hb(hb1, hb2);
1234 if (likely(cmpval != NULL)) {
1237 ret = get_futex_value_locked(&curval, uaddr1);
1239 if (unlikely(ret)) {
1240 double_unlock_hb(hb1, hb2);
1242 ret = get_user(curval, uaddr1);
1249 put_futex_key(fshared, &key2);
1250 put_futex_key(fshared, &key1);
1253 if (curval != *cmpval) {
1259 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1261 * Attempt to acquire uaddr2 and wake the top waiter. If we
1262 * intend to requeue waiters, force setting the FUTEX_WAITERS
1263 * bit. We force this here where we are able to easily handle
1264 * faults rather in the requeue loop below.
1266 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1267 &key2, &pi_state, nr_requeue);
1270 * At this point the top_waiter has either taken uaddr2 or is
1271 * waiting on it. If the former, then the pi_state will not
1272 * exist yet, look it up one more time to ensure we have a
1279 ret = get_futex_value_locked(&curval2, uaddr2);
1281 ret = lookup_pi_state(curval2, hb2, &key2,
1289 double_unlock_hb(hb1, hb2);
1290 put_futex_key(fshared, &key2);
1291 put_futex_key(fshared, &key1);
1292 ret = fault_in_user_writeable(uaddr2);
1297 /* The owner was exiting, try again. */
1298 double_unlock_hb(hb1, hb2);
1299 put_futex_key(fshared, &key2);
1300 put_futex_key(fshared, &key1);
1308 head1 = &hb1->chain;
1309 plist_for_each_entry_safe(this, next, head1, list) {
1310 if (task_count - nr_wake >= nr_requeue)
1313 if (!match_futex(&this->key, &key1))
1317 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1318 * be paired with each other and no other futex ops.
1320 if ((requeue_pi && !this->rt_waiter) ||
1321 (!requeue_pi && this->rt_waiter)) {
1327 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1328 * lock, we already woke the top_waiter. If not, it will be
1329 * woken by futex_unlock_pi().
1331 if (++task_count <= nr_wake && !requeue_pi) {
1336 /* Ensure we requeue to the expected futex for requeue_pi. */
1337 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1343 * Requeue nr_requeue waiters and possibly one more in the case
1344 * of requeue_pi if we couldn't acquire the lock atomically.
1347 /* Prepare the waiter to take the rt_mutex. */
1348 atomic_inc(&pi_state->refcount);
1349 this->pi_state = pi_state;
1350 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1354 /* We got the lock. */
1355 requeue_pi_wake_futex(this, &key2, hb2);
1360 this->pi_state = NULL;
1361 free_pi_state(pi_state);
1365 requeue_futex(this, hb1, hb2, &key2);
1370 double_unlock_hb(hb1, hb2);
1373 * drop_futex_key_refs() must be called outside the spinlocks. During
1374 * the requeue we moved futex_q's from the hash bucket at key1 to the
1375 * one at key2 and updated their key pointer. We no longer need to
1376 * hold the references to key1.
1378 while (--drop_count >= 0)
1379 drop_futex_key_refs(&key1);
1382 put_futex_key(fshared, &key2);
1384 put_futex_key(fshared, &key1);
1386 if (pi_state != NULL)
1387 free_pi_state(pi_state);
1388 return ret ? ret : task_count;
1391 /* The key must be already stored in q->key. */
1392 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1394 struct futex_hash_bucket *hb;
1396 hb = hash_futex(&q->key);
1397 q->lock_ptr = &hb->lock;
1399 spin_lock(&hb->lock);
1404 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1406 spin_unlock(&hb->lock);
1410 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1411 * @q: The futex_q to enqueue
1412 * @hb: The destination hash bucket
1414 * The hb->lock must be held by the caller, and is released here. A call to
1415 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1416 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1417 * or nothing if the unqueue is done as part of the wake process and the unqueue
1418 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1421 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1426 * The priority used to register this element is
1427 * - either the real thread-priority for the real-time threads
1428 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1429 * - or MAX_RT_PRIO for non-RT threads.
1430 * Thus, all RT-threads are woken first in priority order, and
1431 * the others are woken last, in FIFO order.
1433 prio = min(current->normal_prio, MAX_RT_PRIO);
1435 plist_node_init(&q->list, prio);
1436 #ifdef CONFIG_DEBUG_PI_LIST
1437 q->list.plist.lock = &hb->lock;
1439 plist_add(&q->list, &hb->chain);
1441 spin_unlock(&hb->lock);
1445 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1446 * @q: The futex_q to unqueue
1448 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1449 * be paired with exactly one earlier call to queue_me().
1452 * 1 - if the futex_q was still queued (and we removed unqueued it)
1453 * 0 - if the futex_q was already removed by the waking thread
1455 static int unqueue_me(struct futex_q *q)
1457 spinlock_t *lock_ptr;
1460 /* In the common case we don't take the spinlock, which is nice. */
1462 lock_ptr = q->lock_ptr;
1464 if (lock_ptr != NULL) {
1465 spin_lock(lock_ptr);
1467 * q->lock_ptr can change between reading it and
1468 * spin_lock(), causing us to take the wrong lock. This
1469 * corrects the race condition.
1471 * Reasoning goes like this: if we have the wrong lock,
1472 * q->lock_ptr must have changed (maybe several times)
1473 * between reading it and the spin_lock(). It can
1474 * change again after the spin_lock() but only if it was
1475 * already changed before the spin_lock(). It cannot,
1476 * however, change back to the original value. Therefore
1477 * we can detect whether we acquired the correct lock.
1479 if (unlikely(lock_ptr != q->lock_ptr)) {
1480 spin_unlock(lock_ptr);
1483 WARN_ON(plist_node_empty(&q->list));
1484 plist_del(&q->list, &q->list.plist);
1486 BUG_ON(q->pi_state);
1488 spin_unlock(lock_ptr);
1492 drop_futex_key_refs(&q->key);
1497 * PI futexes can not be requeued and must remove themself from the
1498 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1501 static void unqueue_me_pi(struct futex_q *q)
1503 WARN_ON(plist_node_empty(&q->list));
1504 plist_del(&q->list, &q->list.plist);
1506 BUG_ON(!q->pi_state);
1507 free_pi_state(q->pi_state);
1510 spin_unlock(q->lock_ptr);
1514 * Fixup the pi_state owner with the new owner.
1516 * Must be called with hash bucket lock held and mm->sem held for non
1519 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1520 struct task_struct *newowner, int fshared)
1522 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1523 struct futex_pi_state *pi_state = q->pi_state;
1524 struct task_struct *oldowner = pi_state->owner;
1525 u32 uval, curval, newval;
1529 if (!pi_state->owner)
1530 newtid |= FUTEX_OWNER_DIED;
1533 * We are here either because we stole the rtmutex from the
1534 * pending owner or we are the pending owner which failed to
1535 * get the rtmutex. We have to replace the pending owner TID
1536 * in the user space variable. This must be atomic as we have
1537 * to preserve the owner died bit here.
1539 * Note: We write the user space value _before_ changing the pi_state
1540 * because we can fault here. Imagine swapped out pages or a fork
1541 * that marked all the anonymous memory readonly for cow.
1543 * Modifying pi_state _before_ the user space value would
1544 * leave the pi_state in an inconsistent state when we fault
1545 * here, because we need to drop the hash bucket lock to
1546 * handle the fault. This might be observed in the PID check
1547 * in lookup_pi_state.
1550 if (get_futex_value_locked(&uval, uaddr))
1554 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1556 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1558 if (curval == -EFAULT)
1566 * We fixed up user space. Now we need to fix the pi_state
1569 if (pi_state->owner != NULL) {
1570 spin_lock_irq(&pi_state->owner->pi_lock);
1571 WARN_ON(list_empty(&pi_state->list));
1572 list_del_init(&pi_state->list);
1573 spin_unlock_irq(&pi_state->owner->pi_lock);
1576 pi_state->owner = newowner;
1578 spin_lock_irq(&newowner->pi_lock);
1579 WARN_ON(!list_empty(&pi_state->list));
1580 list_add(&pi_state->list, &newowner->pi_state_list);
1581 spin_unlock_irq(&newowner->pi_lock);
1585 * To handle the page fault we need to drop the hash bucket
1586 * lock here. That gives the other task (either the pending
1587 * owner itself or the task which stole the rtmutex) the
1588 * chance to try the fixup of the pi_state. So once we are
1589 * back from handling the fault we need to check the pi_state
1590 * after reacquiring the hash bucket lock and before trying to
1591 * do another fixup. When the fixup has been done already we
1595 spin_unlock(q->lock_ptr);
1597 ret = fault_in_user_writeable(uaddr);
1599 spin_lock(q->lock_ptr);
1602 * Check if someone else fixed it for us:
1604 if (pi_state->owner != oldowner)
1614 * In case we must use restart_block to restart a futex_wait,
1615 * we encode in the 'flags' shared capability
1617 #define FLAGS_SHARED 0x01
1618 #define FLAGS_CLOCKRT 0x02
1619 #define FLAGS_HAS_TIMEOUT 0x04
1621 static long futex_wait_restart(struct restart_block *restart);
1624 * fixup_owner() - Post lock pi_state and corner case management
1625 * @uaddr: user address of the futex
1626 * @fshared: whether the futex is shared (1) or not (0)
1627 * @q: futex_q (contains pi_state and access to the rt_mutex)
1628 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1630 * After attempting to lock an rt_mutex, this function is called to cleanup
1631 * the pi_state owner as well as handle race conditions that may allow us to
1632 * acquire the lock. Must be called with the hb lock held.
1635 * 1 - success, lock taken
1636 * 0 - success, lock not taken
1637 * <0 - on error (-EFAULT)
1639 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1642 struct task_struct *owner;
1647 * Got the lock. We might not be the anticipated owner if we
1648 * did a lock-steal - fix up the PI-state in that case:
1650 if (q->pi_state->owner != current)
1651 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1656 * Catch the rare case, where the lock was released when we were on the
1657 * way back before we locked the hash bucket.
1659 if (q->pi_state->owner == current) {
1661 * Try to get the rt_mutex now. This might fail as some other
1662 * task acquired the rt_mutex after we removed ourself from the
1663 * rt_mutex waiters list.
1665 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1671 * pi_state is incorrect, some other task did a lock steal and
1672 * we returned due to timeout or signal without taking the
1673 * rt_mutex. Too late. We can access the rt_mutex_owner without
1674 * locking, as the other task is now blocked on the hash bucket
1675 * lock. Fix the state up.
1677 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1678 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1683 * Paranoia check. If we did not take the lock, then we should not be
1684 * the owner, nor the pending owner, of the rt_mutex.
1686 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1687 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1688 "pi-state %p\n", ret,
1689 q->pi_state->pi_mutex.owner,
1690 q->pi_state->owner);
1693 return ret ? ret : locked;
1697 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1698 * @hb: the futex hash bucket, must be locked by the caller
1699 * @q: the futex_q to queue up on
1700 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1702 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1703 struct hrtimer_sleeper *timeout)
1706 * The task state is guaranteed to be set before another task can
1707 * wake it. set_current_state() is implemented using set_mb() and
1708 * queue_me() calls spin_unlock() upon completion, both serializing
1709 * access to the hash list and forcing another memory barrier.
1711 set_current_state(TASK_INTERRUPTIBLE);
1716 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1717 if (!hrtimer_active(&timeout->timer))
1718 timeout->task = NULL;
1722 * If we have been removed from the hash list, then another task
1723 * has tried to wake us, and we can skip the call to schedule().
1725 if (likely(!plist_node_empty(&q->list))) {
1727 * If the timer has already expired, current will already be
1728 * flagged for rescheduling. Only call schedule if there
1729 * is no timeout, or if it has yet to expire.
1731 if (!timeout || timeout->task)
1734 __set_current_state(TASK_RUNNING);
1738 * futex_wait_setup() - Prepare to wait on a futex
1739 * @uaddr: the futex userspace address
1740 * @val: the expected value
1741 * @fshared: whether the futex is shared (1) or not (0)
1742 * @q: the associated futex_q
1743 * @hb: storage for hash_bucket pointer to be returned to caller
1745 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1746 * compare it with the expected value. Handle atomic faults internally.
1747 * Return with the hb lock held and a q.key reference on success, and unlocked
1748 * with no q.key reference on failure.
1751 * 0 - uaddr contains val and hb has been locked
1752 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1754 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1755 struct futex_q *q, struct futex_hash_bucket **hb)
1761 * Access the page AFTER the hash-bucket is locked.
1762 * Order is important:
1764 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1765 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1767 * The basic logical guarantee of a futex is that it blocks ONLY
1768 * if cond(var) is known to be true at the time of blocking, for
1769 * any cond. If we queued after testing *uaddr, that would open
1770 * a race condition where we could block indefinitely with
1771 * cond(var) false, which would violate the guarantee.
1773 * A consequence is that futex_wait() can return zero and absorb
1774 * a wakeup when *uaddr != val on entry to the syscall. This is
1778 q->key = FUTEX_KEY_INIT;
1779 ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1780 if (unlikely(ret != 0))
1784 *hb = queue_lock(q);
1786 ret = get_futex_value_locked(&uval, uaddr);
1789 queue_unlock(q, *hb);
1791 ret = get_user(uval, uaddr);
1798 put_futex_key(fshared, &q->key);
1803 queue_unlock(q, *hb);
1809 put_futex_key(fshared, &q->key);
1813 static int futex_wait(u32 __user *uaddr, int fshared,
1814 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1816 struct hrtimer_sleeper timeout, *to = NULL;
1817 struct restart_block *restart;
1818 struct futex_hash_bucket *hb;
1828 q.requeue_pi_key = NULL;
1833 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1834 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1835 hrtimer_init_sleeper(to, current);
1836 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1837 current->timer_slack_ns);
1842 * Prepare to wait on uaddr. On success, holds hb lock and increments
1845 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1849 /* queue_me and wait for wakeup, timeout, or a signal. */
1850 futex_wait_queue_me(hb, &q, to);
1852 /* If we were woken (and unqueued), we succeeded, whatever. */
1854 /* unqueue_me() drops q.key ref */
1855 if (!unqueue_me(&q))
1858 if (to && !to->task)
1862 * We expect signal_pending(current), but we might be the
1863 * victim of a spurious wakeup as well.
1865 if (!signal_pending(current))
1872 restart = ¤t_thread_info()->restart_block;
1873 restart->fn = futex_wait_restart;
1874 restart->futex.uaddr = (u32 *)uaddr;
1875 restart->futex.val = val;
1876 restart->futex.time = abs_time->tv64;
1877 restart->futex.bitset = bitset;
1878 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1881 restart->futex.flags |= FLAGS_SHARED;
1883 restart->futex.flags |= FLAGS_CLOCKRT;
1885 ret = -ERESTART_RESTARTBLOCK;
1889 hrtimer_cancel(&to->timer);
1890 destroy_hrtimer_on_stack(&to->timer);
1896 static long futex_wait_restart(struct restart_block *restart)
1898 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1900 ktime_t t, *tp = NULL;
1902 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1903 t.tv64 = restart->futex.time;
1906 restart->fn = do_no_restart_syscall;
1907 if (restart->futex.flags & FLAGS_SHARED)
1909 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1910 restart->futex.bitset,
1911 restart->futex.flags & FLAGS_CLOCKRT);
1916 * Userspace tried a 0 -> TID atomic transition of the futex value
1917 * and failed. The kernel side here does the whole locking operation:
1918 * if there are waiters then it will block, it does PI, etc. (Due to
1919 * races the kernel might see a 0 value of the futex too.)
1921 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1922 int detect, ktime_t *time, int trylock)
1924 struct hrtimer_sleeper timeout, *to = NULL;
1925 struct futex_hash_bucket *hb;
1929 if (refill_pi_state_cache())
1934 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1936 hrtimer_init_sleeper(to, current);
1937 hrtimer_set_expires(&to->timer, *time);
1942 q.requeue_pi_key = NULL;
1944 q.key = FUTEX_KEY_INIT;
1945 ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1946 if (unlikely(ret != 0))
1950 hb = queue_lock(&q);
1952 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1953 if (unlikely(ret)) {
1956 /* We got the lock. */
1958 goto out_unlock_put_key;
1963 * Task is exiting and we just wait for the
1966 queue_unlock(&q, hb);
1967 put_futex_key(fshared, &q.key);
1971 goto out_unlock_put_key;
1976 * Only actually queue now that the atomic ops are done:
1980 WARN_ON(!q.pi_state);
1982 * Block on the PI mutex:
1985 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1987 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1988 /* Fixup the trylock return value: */
1989 ret = ret ? 0 : -EWOULDBLOCK;
1992 spin_lock(q.lock_ptr);
1994 * Fixup the pi_state owner and possibly acquire the lock if we
1997 res = fixup_owner(uaddr, fshared, &q, !ret);
1999 * If fixup_owner() returned an error, proprogate that. If it acquired
2000 * the lock, clear our -ETIMEDOUT or -EINTR.
2003 ret = (res < 0) ? res : 0;
2006 * If fixup_owner() faulted and was unable to handle the fault, unlock
2007 * it and return the fault to userspace.
2009 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2010 rt_mutex_unlock(&q.pi_state->pi_mutex);
2012 /* Unqueue and drop the lock */
2018 queue_unlock(&q, hb);
2021 put_futex_key(fshared, &q.key);
2024 destroy_hrtimer_on_stack(&to->timer);
2025 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2028 queue_unlock(&q, hb);
2030 ret = fault_in_user_writeable(uaddr);
2037 put_futex_key(fshared, &q.key);
2042 * Userspace attempted a TID -> 0 atomic transition, and failed.
2043 * This is the in-kernel slowpath: we look up the PI state (if any),
2044 * and do the rt-mutex unlock.
2046 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2048 struct futex_hash_bucket *hb;
2049 struct futex_q *this, *next;
2051 struct plist_head *head;
2052 union futex_key key = FUTEX_KEY_INIT;
2056 if (get_user(uval, uaddr))
2059 * We release only a lock we actually own:
2061 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2064 ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
2065 if (unlikely(ret != 0))
2068 hb = hash_futex(&key);
2069 spin_lock(&hb->lock);
2072 * To avoid races, try to do the TID -> 0 atomic transition
2073 * again. If it succeeds then we can return without waking
2076 if (!(uval & FUTEX_OWNER_DIED))
2077 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2080 if (unlikely(uval == -EFAULT))
2083 * Rare case: we managed to release the lock atomically,
2084 * no need to wake anyone else up:
2086 if (unlikely(uval == task_pid_vnr(current)))
2090 * Ok, other tasks may need to be woken up - check waiters
2091 * and do the wakeup if necessary:
2095 plist_for_each_entry_safe(this, next, head, list) {
2096 if (!match_futex (&this->key, &key))
2098 ret = wake_futex_pi(uaddr, uval, this);
2100 * The atomic access to the futex value
2101 * generated a pagefault, so retry the
2102 * user-access and the wakeup:
2109 * No waiters - kernel unlocks the futex:
2111 if (!(uval & FUTEX_OWNER_DIED)) {
2112 ret = unlock_futex_pi(uaddr, uval);
2118 spin_unlock(&hb->lock);
2119 put_futex_key(fshared, &key);
2125 spin_unlock(&hb->lock);
2126 put_futex_key(fshared, &key);
2128 ret = fault_in_user_writeable(uaddr);
2136 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2137 * @hb: the hash_bucket futex_q was original enqueued on
2138 * @q: the futex_q woken while waiting to be requeued
2139 * @key2: the futex_key of the requeue target futex
2140 * @timeout: the timeout associated with the wait (NULL if none)
2142 * Detect if the task was woken on the initial futex as opposed to the requeue
2143 * target futex. If so, determine if it was a timeout or a signal that caused
2144 * the wakeup and return the appropriate error code to the caller. Must be
2145 * called with the hb lock held.
2148 * 0 - no early wakeup detected
2149 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2152 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2153 struct futex_q *q, union futex_key *key2,
2154 struct hrtimer_sleeper *timeout)
2159 * With the hb lock held, we avoid races while we process the wakeup.
2160 * We only need to hold hb (and not hb2) to ensure atomicity as the
2161 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2162 * It can't be requeued from uaddr2 to something else since we don't
2163 * support a PI aware source futex for requeue.
2165 if (!match_futex(&q->key, key2)) {
2166 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2168 * We were woken prior to requeue by a timeout or a signal.
2169 * Unqueue the futex_q and determine which it was.
2171 plist_del(&q->list, &q->list.plist);
2173 /* Handle spurious wakeups gracefully */
2175 if (timeout && !timeout->task)
2177 else if (signal_pending(current))
2178 ret = -ERESTARTNOINTR;
2184 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2185 * @uaddr: the futex we initially wait on (non-pi)
2186 * @fshared: whether the futexes are shared (1) or not (0). They must be
2187 * the same type, no requeueing from private to shared, etc.
2188 * @val: the expected value of uaddr
2189 * @abs_time: absolute timeout
2190 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2191 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2192 * @uaddr2: the pi futex we will take prior to returning to user-space
2194 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2195 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2196 * complete the acquisition of the rt_mutex prior to returning to userspace.
2197 * This ensures the rt_mutex maintains an owner when it has waiters; without
2198 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2201 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2202 * via the following:
2203 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2204 * 2) wakeup on uaddr2 after a requeue
2208 * If 3, cleanup and return -ERESTARTNOINTR.
2210 * If 2, we may then block on trying to take the rt_mutex and return via:
2211 * 5) successful lock
2214 * 8) other lock acquisition failure
2216 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2218 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2224 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2225 u32 val, ktime_t *abs_time, u32 bitset,
2226 int clockrt, u32 __user *uaddr2)
2228 struct hrtimer_sleeper timeout, *to = NULL;
2229 struct rt_mutex_waiter rt_waiter;
2230 struct rt_mutex *pi_mutex = NULL;
2231 struct futex_hash_bucket *hb;
2232 union futex_key key2;
2241 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2242 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2243 hrtimer_init_sleeper(to, current);
2244 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2245 current->timer_slack_ns);
2249 * The waiter is allocated on our stack, manipulated by the requeue
2250 * code while we sleep on uaddr.
2252 debug_rt_mutex_init_waiter(&rt_waiter);
2253 rt_waiter.task = NULL;
2255 key2 = FUTEX_KEY_INIT;
2256 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2257 if (unlikely(ret != 0))
2262 q.rt_waiter = &rt_waiter;
2263 q.requeue_pi_key = &key2;
2266 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2269 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2273 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2274 futex_wait_queue_me(hb, &q, to);
2276 spin_lock(&hb->lock);
2277 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2278 spin_unlock(&hb->lock);
2283 * In order for us to be here, we know our q.key == key2, and since
2284 * we took the hb->lock above, we also know that futex_requeue() has
2285 * completed and we no longer have to concern ourselves with a wakeup
2286 * race with the atomic proxy lock acquisition by the requeue code. The
2287 * futex_requeue dropped our key1 reference and incremented our key2
2291 /* Check if the requeue code acquired the second futex for us. */
2294 * Got the lock. We might not be the anticipated owner if we
2295 * did a lock-steal - fix up the PI-state in that case.
2297 if (q.pi_state && (q.pi_state->owner != current)) {
2298 spin_lock(q.lock_ptr);
2299 ret = fixup_pi_state_owner(uaddr2, &q, current,
2301 spin_unlock(q.lock_ptr);
2305 * We have been woken up by futex_unlock_pi(), a timeout, or a
2306 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2309 WARN_ON(!&q.pi_state);
2310 pi_mutex = &q.pi_state->pi_mutex;
2311 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2312 debug_rt_mutex_free_waiter(&rt_waiter);
2314 spin_lock(q.lock_ptr);
2316 * Fixup the pi_state owner and possibly acquire the lock if we
2319 res = fixup_owner(uaddr2, fshared, &q, !ret);
2321 * If fixup_owner() returned an error, proprogate that. If it
2322 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2325 ret = (res < 0) ? res : 0;
2327 /* Unqueue and drop the lock. */
2332 * If fixup_pi_state_owner() faulted and was unable to handle the
2333 * fault, unlock the rt_mutex and return the fault to userspace.
2335 if (ret == -EFAULT) {
2336 if (rt_mutex_owner(pi_mutex) == current)
2337 rt_mutex_unlock(pi_mutex);
2338 } else if (ret == -EINTR) {
2340 * We've already been requeued, but cannot restart by calling
2341 * futex_lock_pi() directly. We could restart this syscall, but
2342 * it would detect that the user space "val" changed and return
2343 * -EWOULDBLOCK. Save the overhead of the restart and return
2344 * -EWOULDBLOCK directly.
2350 put_futex_key(fshared, &q.key);
2352 put_futex_key(fshared, &key2);
2356 hrtimer_cancel(&to->timer);
2357 destroy_hrtimer_on_stack(&to->timer);
2363 * Support for robust futexes: the kernel cleans up held futexes at
2366 * Implementation: user-space maintains a per-thread list of locks it
2367 * is holding. Upon do_exit(), the kernel carefully walks this list,
2368 * and marks all locks that are owned by this thread with the
2369 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2370 * always manipulated with the lock held, so the list is private and
2371 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2372 * field, to allow the kernel to clean up if the thread dies after
2373 * acquiring the lock, but just before it could have added itself to
2374 * the list. There can only be one such pending lock.
2378 * sys_set_robust_list() - Set the robust-futex list head of a task
2379 * @head: pointer to the list-head
2380 * @len: length of the list-head, as userspace expects
2382 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2385 if (!futex_cmpxchg_enabled)
2388 * The kernel knows only one size for now:
2390 if (unlikely(len != sizeof(*head)))
2393 current->robust_list = head;
2399 * sys_get_robust_list() - Get the robust-futex list head of a task
2400 * @pid: pid of the process [zero for current task]
2401 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2402 * @len_ptr: pointer to a length field, the kernel fills in the header size
2404 SYSCALL_DEFINE3(get_robust_list, int, pid,
2405 struct robust_list_head __user * __user *, head_ptr,
2406 size_t __user *, len_ptr)
2408 struct robust_list_head __user *head;
2410 const struct cred *cred = current_cred(), *pcred;
2412 if (!futex_cmpxchg_enabled)
2416 head = current->robust_list;
2418 struct task_struct *p;
2422 p = find_task_by_vpid(pid);
2426 pcred = __task_cred(p);
2427 if (cred->euid != pcred->euid &&
2428 cred->euid != pcred->uid &&
2429 !capable(CAP_SYS_PTRACE))
2431 head = p->robust_list;
2435 if (put_user(sizeof(*head), len_ptr))
2437 return put_user(head, head_ptr);
2446 * Process a futex-list entry, check whether it's owned by the
2447 * dying task, and do notification if so:
2449 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2451 u32 uval, nval, mval;
2454 if (get_user(uval, uaddr))
2457 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2459 * Ok, this dying thread is truly holding a futex
2460 * of interest. Set the OWNER_DIED bit atomically
2461 * via cmpxchg, and if the value had FUTEX_WAITERS
2462 * set, wake up a waiter (if any). (We have to do a
2463 * futex_wake() even if OWNER_DIED is already set -
2464 * to handle the rare but possible case of recursive
2465 * thread-death.) The rest of the cleanup is done in
2468 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2469 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2471 if (nval == -EFAULT)
2478 * Wake robust non-PI futexes here. The wakeup of
2479 * PI futexes happens in exit_pi_state():
2481 if (!pi && (uval & FUTEX_WAITERS))
2482 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2488 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2490 static inline int fetch_robust_entry(struct robust_list __user **entry,
2491 struct robust_list __user * __user *head,
2494 unsigned long uentry;
2496 if (get_user(uentry, (unsigned long __user *)head))
2499 *entry = (void __user *)(uentry & ~1UL);
2506 * Walk curr->robust_list (very carefully, it's a userspace list!)
2507 * and mark any locks found there dead, and notify any waiters.
2509 * We silently return on any sign of list-walking problem.
2511 void exit_robust_list(struct task_struct *curr)
2513 struct robust_list_head __user *head = curr->robust_list;
2514 struct robust_list __user *entry, *next_entry, *pending;
2515 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2516 unsigned long futex_offset;
2519 if (!futex_cmpxchg_enabled)
2523 * Fetch the list head (which was registered earlier, via
2524 * sys_set_robust_list()):
2526 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2529 * Fetch the relative futex offset:
2531 if (get_user(futex_offset, &head->futex_offset))
2534 * Fetch any possibly pending lock-add first, and handle it
2537 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2540 next_entry = NULL; /* avoid warning with gcc */
2541 while (entry != &head->list) {
2543 * Fetch the next entry in the list before calling
2544 * handle_futex_death:
2546 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2548 * A pending lock might already be on the list, so
2549 * don't process it twice:
2551 if (entry != pending)
2552 if (handle_futex_death((void __user *)entry + futex_offset,
2560 * Avoid excessively long or circular lists:
2569 handle_futex_death((void __user *)pending + futex_offset,
2573 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2574 u32 __user *uaddr2, u32 val2, u32 val3)
2576 int clockrt, ret = -ENOSYS;
2577 int cmd = op & FUTEX_CMD_MASK;
2580 if (!(op & FUTEX_PRIVATE_FLAG))
2583 clockrt = op & FUTEX_CLOCK_REALTIME;
2584 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2589 val3 = FUTEX_BITSET_MATCH_ANY;
2590 case FUTEX_WAIT_BITSET:
2591 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2594 val3 = FUTEX_BITSET_MATCH_ANY;
2595 case FUTEX_WAKE_BITSET:
2596 ret = futex_wake(uaddr, fshared, val, val3);
2599 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2601 case FUTEX_CMP_REQUEUE:
2602 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2606 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2609 if (futex_cmpxchg_enabled)
2610 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2612 case FUTEX_UNLOCK_PI:
2613 if (futex_cmpxchg_enabled)
2614 ret = futex_unlock_pi(uaddr, fshared);
2616 case FUTEX_TRYLOCK_PI:
2617 if (futex_cmpxchg_enabled)
2618 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2620 case FUTEX_WAIT_REQUEUE_PI:
2621 val3 = FUTEX_BITSET_MATCH_ANY;
2622 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2625 case FUTEX_CMP_REQUEUE_PI:
2626 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2636 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2637 struct timespec __user *, utime, u32 __user *, uaddr2,
2641 ktime_t t, *tp = NULL;
2643 int cmd = op & FUTEX_CMD_MASK;
2645 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2646 cmd == FUTEX_WAIT_BITSET ||
2647 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2648 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2650 if (!timespec_valid(&ts))
2653 t = timespec_to_ktime(ts);
2654 if (cmd == FUTEX_WAIT)
2655 t = ktime_add_safe(ktime_get(), t);
2659 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2660 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2662 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2663 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2664 val2 = (u32) (unsigned long) utime;
2666 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2669 static int __init futex_init(void)
2675 * This will fail and we want it. Some arch implementations do
2676 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2677 * functionality. We want to know that before we call in any
2678 * of the complex code paths. Also we want to prevent
2679 * registration of robust lists in that case. NULL is
2680 * guaranteed to fault and we get -EFAULT on functional
2681 * implementation, the non functional ones will return
2684 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2685 if (curval == -EFAULT)
2686 futex_cmpxchg_enabled = 1;
2688 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2689 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2690 spin_lock_init(&futex_queues[i].lock);
2695 __initcall(futex_init);