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.
207 * Returns a negative error code or 0
208 * The key words are stored in *key on success.
210 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
211 * offset_within_page). For private mappings, it's (uaddr, current->mm).
212 * We can usually work out the index without swapping in the page.
214 * lock_page() might sleep, the caller should not hold a spinlock.
217 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
219 unsigned long address = (unsigned long)uaddr;
220 struct mm_struct *mm = current->mm;
225 * The futex address must be "naturally" aligned.
227 key->both.offset = address % PAGE_SIZE;
228 if (unlikely((address % sizeof(u32)) != 0))
230 address -= key->both.offset;
233 * PROCESS_PRIVATE futexes are fast.
234 * As the mm cannot disappear under us and the 'key' only needs
235 * virtual address, we dont even have to find the underlying vma.
236 * Note : We do have to check 'uaddr' is a valid user address,
237 * but access_ok() should be faster than find_vma()
240 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
242 key->private.mm = mm;
243 key->private.address = address;
244 get_futex_key_refs(key);
249 err = get_user_pages_fast(address, 1, 1, &page);
253 page = compound_head(page);
255 if (!page->mapping) {
262 * Private mappings are handled in a simple way.
264 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
265 * it's a read-only handle, it's expected that futexes attach to
266 * the object not the particular process.
268 if (PageAnon(page)) {
269 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
270 key->private.mm = mm;
271 key->private.address = address;
273 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
274 key->shared.inode = page->mapping->host;
275 key->shared.pgoff = page->index;
278 get_futex_key_refs(key);
286 void put_futex_key(int fshared, union futex_key *key)
288 drop_futex_key_refs(key);
292 * fault_in_user_writeable() - Fault in user address and verify RW access
293 * @uaddr: pointer to faulting user space address
295 * Slow path to fixup the fault we just took in the atomic write
298 * We have no generic implementation of a non destructive write to the
299 * user address. We know that we faulted in the atomic pagefault
300 * disabled section so we can as well avoid the #PF overhead by
301 * calling get_user_pages() right away.
303 static int fault_in_user_writeable(u32 __user *uaddr)
305 struct mm_struct *mm = current->mm;
308 down_read(&mm->mmap_sem);
309 ret = get_user_pages(current, mm, (unsigned long)uaddr,
310 1, 1, 0, NULL, NULL);
311 up_read(&mm->mmap_sem);
313 return ret < 0 ? ret : 0;
317 * futex_top_waiter() - Return the highest priority waiter on a futex
318 * @hb: the hash bucket the futex_q's reside in
319 * @key: the futex key (to distinguish it from other futex futex_q's)
321 * Must be called with the hb lock held.
323 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
324 union futex_key *key)
326 struct futex_q *this;
328 plist_for_each_entry(this, &hb->chain, list) {
329 if (match_futex(&this->key, key))
335 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
336 u32 uval, u32 newval)
341 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
347 static int get_futex_value_locked(u32 *dest, u32 __user *from)
352 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
355 return ret ? -EFAULT : 0;
362 static int refill_pi_state_cache(void)
364 struct futex_pi_state *pi_state;
366 if (likely(current->pi_state_cache))
369 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
374 INIT_LIST_HEAD(&pi_state->list);
375 /* pi_mutex gets initialized later */
376 pi_state->owner = NULL;
377 atomic_set(&pi_state->refcount, 1);
378 pi_state->key = FUTEX_KEY_INIT;
380 current->pi_state_cache = pi_state;
385 static struct futex_pi_state * alloc_pi_state(void)
387 struct futex_pi_state *pi_state = current->pi_state_cache;
390 current->pi_state_cache = NULL;
395 static void free_pi_state(struct futex_pi_state *pi_state)
397 if (!atomic_dec_and_test(&pi_state->refcount))
401 * If pi_state->owner is NULL, the owner is most probably dying
402 * and has cleaned up the pi_state already
404 if (pi_state->owner) {
405 raw_spin_lock_irq(&pi_state->owner->pi_lock);
406 list_del_init(&pi_state->list);
407 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
409 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
412 if (current->pi_state_cache)
416 * pi_state->list is already empty.
417 * clear pi_state->owner.
418 * refcount is at 0 - put it back to 1.
420 pi_state->owner = NULL;
421 atomic_set(&pi_state->refcount, 1);
422 current->pi_state_cache = pi_state;
427 * Look up the task based on what TID userspace gave us.
430 static struct task_struct * futex_find_get_task(pid_t pid)
432 struct task_struct *p;
435 p = find_task_by_vpid(pid);
445 * This task is holding PI mutexes at exit time => bad.
446 * Kernel cleans up PI-state, but userspace is likely hosed.
447 * (Robust-futex cleanup is separate and might save the day for userspace.)
449 void exit_pi_state_list(struct task_struct *curr)
451 struct list_head *next, *head = &curr->pi_state_list;
452 struct futex_pi_state *pi_state;
453 struct futex_hash_bucket *hb;
454 union futex_key key = FUTEX_KEY_INIT;
456 if (!futex_cmpxchg_enabled)
459 * We are a ZOMBIE and nobody can enqueue itself on
460 * pi_state_list anymore, but we have to be careful
461 * versus waiters unqueueing themselves:
463 raw_spin_lock_irq(&curr->pi_lock);
464 while (!list_empty(head)) {
467 pi_state = list_entry(next, struct futex_pi_state, list);
469 hb = hash_futex(&key);
470 raw_spin_unlock_irq(&curr->pi_lock);
472 spin_lock(&hb->lock);
474 raw_spin_lock_irq(&curr->pi_lock);
476 * We dropped the pi-lock, so re-check whether this
477 * task still owns the PI-state:
479 if (head->next != next) {
480 spin_unlock(&hb->lock);
484 WARN_ON(pi_state->owner != curr);
485 WARN_ON(list_empty(&pi_state->list));
486 list_del_init(&pi_state->list);
487 pi_state->owner = NULL;
488 raw_spin_unlock_irq(&curr->pi_lock);
490 rt_mutex_unlock(&pi_state->pi_mutex);
492 spin_unlock(&hb->lock);
494 raw_spin_lock_irq(&curr->pi_lock);
496 raw_spin_unlock_irq(&curr->pi_lock);
500 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
501 union futex_key *key, struct futex_pi_state **ps)
503 struct futex_pi_state *pi_state = NULL;
504 struct futex_q *this, *next;
505 struct plist_head *head;
506 struct task_struct *p;
507 pid_t pid = uval & FUTEX_TID_MASK;
511 plist_for_each_entry_safe(this, next, head, list) {
512 if (match_futex(&this->key, key)) {
514 * Another waiter already exists - bump up
515 * the refcount and return its pi_state:
517 pi_state = this->pi_state;
519 * Userspace might have messed up non PI and PI futexes
521 if (unlikely(!pi_state))
524 WARN_ON(!atomic_read(&pi_state->refcount));
527 * When pi_state->owner is NULL then the owner died
528 * and another waiter is on the fly. pi_state->owner
529 * is fixed up by the task which acquires
530 * pi_state->rt_mutex.
532 * We do not check for pid == 0 which can happen when
533 * the owner died and robust_list_exit() cleared the
536 if (pid && pi_state->owner) {
538 * Bail out if user space manipulated the
541 if (pid != task_pid_vnr(pi_state->owner))
545 atomic_inc(&pi_state->refcount);
553 * We are the first waiter - try to look up the real owner and attach
554 * the new pi_state to it, but bail out when TID = 0
558 p = futex_find_get_task(pid);
563 * We need to look at the task state flags to figure out,
564 * whether the task is exiting. To protect against the do_exit
565 * change of the task flags, we do this protected by
568 raw_spin_lock_irq(&p->pi_lock);
569 if (unlikely(p->flags & PF_EXITING)) {
571 * The task is on the way out. When PF_EXITPIDONE is
572 * set, we know that the task has finished the
575 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
577 raw_spin_unlock_irq(&p->pi_lock);
582 pi_state = alloc_pi_state();
585 * Initialize the pi_mutex in locked state and make 'p'
588 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
590 /* Store the key for possible exit cleanups: */
591 pi_state->key = *key;
593 WARN_ON(!list_empty(&pi_state->list));
594 list_add(&pi_state->list, &p->pi_state_list);
596 raw_spin_unlock_irq(&p->pi_lock);
606 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
607 * @uaddr: the pi futex user address
608 * @hb: the pi futex hash bucket
609 * @key: the futex key associated with uaddr and hb
610 * @ps: the pi_state pointer where we store the result of the
612 * @task: the task to perform the atomic lock work for. This will
613 * be "current" except in the case of requeue pi.
614 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
618 * 1 - acquired the lock
621 * The hb->lock and futex_key refs shall be held by the caller.
623 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
624 union futex_key *key,
625 struct futex_pi_state **ps,
626 struct task_struct *task, int set_waiters)
628 int lock_taken, ret, ownerdied = 0;
629 u32 uval, newval, curval;
632 ret = lock_taken = 0;
635 * To avoid races, we attempt to take the lock here again
636 * (by doing a 0 -> TID atomic cmpxchg), while holding all
637 * the locks. It will most likely not succeed.
639 newval = task_pid_vnr(task);
641 newval |= FUTEX_WAITERS;
643 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
649 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
653 * Surprise - we got the lock. Just return to userspace:
655 if (unlikely(!curval))
661 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
662 * to wake at the next unlock.
664 newval = curval | FUTEX_WAITERS;
667 * There are two cases, where a futex might have no owner (the
668 * owner TID is 0): OWNER_DIED. We take over the futex in this
669 * case. We also do an unconditional take over, when the owner
672 * This is safe as we are protected by the hash bucket lock !
674 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
675 /* Keep the OWNER_DIED bit */
676 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
681 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
683 if (unlikely(curval != uval))
687 * We took the lock due to owner died take over.
689 if (unlikely(lock_taken))
693 * We dont have the lock. Look up the PI state (or create it if
694 * we are the first waiter):
696 ret = lookup_pi_state(uval, hb, key, ps);
702 * No owner found for this futex. Check if the
703 * OWNER_DIED bit is set to figure out whether
704 * this is a robust futex or not.
706 if (get_futex_value_locked(&curval, uaddr))
710 * We simply start over in case of a robust
711 * futex. The code above will take the futex
714 if (curval & FUTEX_OWNER_DIED) {
727 * The hash bucket lock must be held when this is called.
728 * Afterwards, the futex_q must not be accessed.
730 static void wake_futex(struct futex_q *q)
732 struct task_struct *p = q->task;
735 * We set q->lock_ptr = NULL _before_ we wake up the task. If
736 * a non futex wake up happens on another CPU then the task
737 * might exit and p would dereference a non existing task
738 * struct. Prevent this by holding a reference on p across the
743 plist_del(&q->list, &q->list.plist);
745 * The waiting task can free the futex_q as soon as
746 * q->lock_ptr = NULL is written, without taking any locks. A
747 * memory barrier is required here to prevent the following
748 * store to lock_ptr from getting ahead of the plist_del.
753 wake_up_state(p, TASK_NORMAL);
757 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
759 struct task_struct *new_owner;
760 struct futex_pi_state *pi_state = this->pi_state;
767 * If current does not own the pi_state then the futex is
768 * inconsistent and user space fiddled with the futex value.
770 if (pi_state->owner != current)
773 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
774 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
777 * This happens when we have stolen the lock and the original
778 * pending owner did not enqueue itself back on the rt_mutex.
779 * Thats not a tragedy. We know that way, that a lock waiter
780 * is on the fly. We make the futex_q waiter the pending owner.
783 new_owner = this->task;
786 * We pass it to the next owner. (The WAITERS bit is always
787 * kept enabled while there is PI state around. We must also
788 * preserve the owner died bit.)
790 if (!(uval & FUTEX_OWNER_DIED)) {
793 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
795 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
797 else if (curval != uval)
800 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
805 raw_spin_lock_irq(&pi_state->owner->pi_lock);
806 WARN_ON(list_empty(&pi_state->list));
807 list_del_init(&pi_state->list);
808 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
810 raw_spin_lock_irq(&new_owner->pi_lock);
811 WARN_ON(!list_empty(&pi_state->list));
812 list_add(&pi_state->list, &new_owner->pi_state_list);
813 pi_state->owner = new_owner;
814 raw_spin_unlock_irq(&new_owner->pi_lock);
816 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
817 rt_mutex_unlock(&pi_state->pi_mutex);
822 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
827 * There is no waiter, so we unlock the futex. The owner died
828 * bit has not to be preserved here. We are the owner:
830 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
839 * Express the locking dependencies for lockdep:
842 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
845 spin_lock(&hb1->lock);
847 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
848 } else { /* hb1 > hb2 */
849 spin_lock(&hb2->lock);
850 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
855 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
857 spin_unlock(&hb1->lock);
859 spin_unlock(&hb2->lock);
863 * Wake up waiters matching bitset queued on this futex (uaddr).
865 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
867 struct futex_hash_bucket *hb;
868 struct futex_q *this, *next;
869 struct plist_head *head;
870 union futex_key key = FUTEX_KEY_INIT;
876 ret = get_futex_key(uaddr, fshared, &key);
877 if (unlikely(ret != 0))
880 hb = hash_futex(&key);
881 spin_lock(&hb->lock);
884 plist_for_each_entry_safe(this, next, head, list) {
885 if (match_futex (&this->key, &key)) {
886 if (this->pi_state || this->rt_waiter) {
891 /* Check if one of the bits is set in both bitsets */
892 if (!(this->bitset & bitset))
896 if (++ret >= nr_wake)
901 spin_unlock(&hb->lock);
902 put_futex_key(fshared, &key);
908 * Wake up all waiters hashed on the physical page that is mapped
909 * to this virtual address:
912 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
913 int nr_wake, int nr_wake2, int op)
915 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
916 struct futex_hash_bucket *hb1, *hb2;
917 struct plist_head *head;
918 struct futex_q *this, *next;
922 ret = get_futex_key(uaddr1, fshared, &key1);
923 if (unlikely(ret != 0))
925 ret = get_futex_key(uaddr2, fshared, &key2);
926 if (unlikely(ret != 0))
929 hb1 = hash_futex(&key1);
930 hb2 = hash_futex(&key2);
933 double_lock_hb(hb1, hb2);
934 op_ret = futex_atomic_op_inuser(op, uaddr2);
935 if (unlikely(op_ret < 0)) {
937 double_unlock_hb(hb1, hb2);
941 * we don't get EFAULT from MMU faults if we don't have an MMU,
942 * but we might get them from range checking
948 if (unlikely(op_ret != -EFAULT)) {
953 ret = fault_in_user_writeable(uaddr2);
960 put_futex_key(fshared, &key2);
961 put_futex_key(fshared, &key1);
967 plist_for_each_entry_safe(this, next, head, list) {
968 if (match_futex (&this->key, &key1)) {
970 if (++ret >= nr_wake)
979 plist_for_each_entry_safe(this, next, head, list) {
980 if (match_futex (&this->key, &key2)) {
982 if (++op_ret >= nr_wake2)
989 double_unlock_hb(hb1, hb2);
991 put_futex_key(fshared, &key2);
993 put_futex_key(fshared, &key1);
999 * requeue_futex() - Requeue a futex_q from one hb to another
1000 * @q: the futex_q to requeue
1001 * @hb1: the source hash_bucket
1002 * @hb2: the target hash_bucket
1003 * @key2: the new key for the requeued futex_q
1006 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1007 struct futex_hash_bucket *hb2, union futex_key *key2)
1011 * If key1 and key2 hash to the same bucket, no need to
1014 if (likely(&hb1->chain != &hb2->chain)) {
1015 plist_del(&q->list, &hb1->chain);
1016 plist_add(&q->list, &hb2->chain);
1017 q->lock_ptr = &hb2->lock;
1018 #ifdef CONFIG_DEBUG_PI_LIST
1019 q->list.plist.spinlock = &hb2->lock;
1022 get_futex_key_refs(key2);
1027 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1029 * @key: the key of the requeue target futex
1030 * @hb: the hash_bucket of the requeue target futex
1032 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1033 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1034 * to the requeue target futex so the waiter can detect the wakeup on the right
1035 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1036 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1037 * to protect access to the pi_state to fixup the owner later. Must be called
1038 * with both q->lock_ptr and hb->lock held.
1041 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1042 struct futex_hash_bucket *hb)
1044 get_futex_key_refs(key);
1047 WARN_ON(plist_node_empty(&q->list));
1048 plist_del(&q->list, &q->list.plist);
1050 WARN_ON(!q->rt_waiter);
1051 q->rt_waiter = NULL;
1053 q->lock_ptr = &hb->lock;
1054 #ifdef CONFIG_DEBUG_PI_LIST
1055 q->list.plist.spinlock = &hb->lock;
1058 wake_up_state(q->task, TASK_NORMAL);
1062 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1063 * @pifutex: the user address of the to futex
1064 * @hb1: the from futex hash bucket, must be locked by the caller
1065 * @hb2: the to futex hash bucket, must be locked by the caller
1066 * @key1: the from futex key
1067 * @key2: the to futex key
1068 * @ps: address to store the pi_state pointer
1069 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1071 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1072 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1073 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1074 * hb1 and hb2 must be held by the caller.
1077 * 0 - failed to acquire the lock atomicly
1078 * 1 - acquired the lock
1081 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1082 struct futex_hash_bucket *hb1,
1083 struct futex_hash_bucket *hb2,
1084 union futex_key *key1, union futex_key *key2,
1085 struct futex_pi_state **ps, int set_waiters)
1087 struct futex_q *top_waiter = NULL;
1091 if (get_futex_value_locked(&curval, pifutex))
1095 * Find the top_waiter and determine if there are additional waiters.
1096 * If the caller intends to requeue more than 1 waiter to pifutex,
1097 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1098 * as we have means to handle the possible fault. If not, don't set
1099 * the bit unecessarily as it will force the subsequent unlock to enter
1102 top_waiter = futex_top_waiter(hb1, key1);
1104 /* There are no waiters, nothing for us to do. */
1108 /* Ensure we requeue to the expected futex. */
1109 if (!match_futex(top_waiter->requeue_pi_key, key2))
1113 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1114 * the contended case or if set_waiters is 1. The pi_state is returned
1115 * in ps in contended cases.
1117 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1120 requeue_pi_wake_futex(top_waiter, key2, hb2);
1126 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1127 * uaddr1: source futex user address
1128 * uaddr2: target futex user address
1129 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1130 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1131 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1132 * pi futex (pi to pi requeue is not supported)
1134 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1135 * uaddr2 atomically on behalf of the top waiter.
1138 * >=0 - on success, the number of tasks requeued or woken
1141 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1142 int nr_wake, int nr_requeue, u32 *cmpval,
1145 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1146 int drop_count = 0, task_count = 0, ret;
1147 struct futex_pi_state *pi_state = NULL;
1148 struct futex_hash_bucket *hb1, *hb2;
1149 struct plist_head *head1;
1150 struct futex_q *this, *next;
1155 * requeue_pi requires a pi_state, try to allocate it now
1156 * without any locks in case it fails.
1158 if (refill_pi_state_cache())
1161 * requeue_pi must wake as many tasks as it can, up to nr_wake
1162 * + nr_requeue, since it acquires the rt_mutex prior to
1163 * returning to userspace, so as to not leave the rt_mutex with
1164 * waiters and no owner. However, second and third wake-ups
1165 * cannot be predicted as they involve race conditions with the
1166 * first wake and a fault while looking up the pi_state. Both
1167 * pthread_cond_signal() and pthread_cond_broadcast() should
1175 if (pi_state != NULL) {
1177 * We will have to lookup the pi_state again, so free this one
1178 * to keep the accounting correct.
1180 free_pi_state(pi_state);
1184 ret = get_futex_key(uaddr1, fshared, &key1);
1185 if (unlikely(ret != 0))
1187 ret = get_futex_key(uaddr2, fshared, &key2);
1188 if (unlikely(ret != 0))
1191 hb1 = hash_futex(&key1);
1192 hb2 = hash_futex(&key2);
1195 double_lock_hb(hb1, hb2);
1197 if (likely(cmpval != NULL)) {
1200 ret = get_futex_value_locked(&curval, uaddr1);
1202 if (unlikely(ret)) {
1203 double_unlock_hb(hb1, hb2);
1205 ret = get_user(curval, uaddr1);
1212 put_futex_key(fshared, &key2);
1213 put_futex_key(fshared, &key1);
1216 if (curval != *cmpval) {
1222 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1224 * Attempt to acquire uaddr2 and wake the top waiter. If we
1225 * intend to requeue waiters, force setting the FUTEX_WAITERS
1226 * bit. We force this here where we are able to easily handle
1227 * faults rather in the requeue loop below.
1229 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1230 &key2, &pi_state, nr_requeue);
1233 * At this point the top_waiter has either taken uaddr2 or is
1234 * waiting on it. If the former, then the pi_state will not
1235 * exist yet, look it up one more time to ensure we have a
1242 ret = get_futex_value_locked(&curval2, uaddr2);
1244 ret = lookup_pi_state(curval2, hb2, &key2,
1252 double_unlock_hb(hb1, hb2);
1253 put_futex_key(fshared, &key2);
1254 put_futex_key(fshared, &key1);
1255 ret = fault_in_user_writeable(uaddr2);
1260 /* The owner was exiting, try again. */
1261 double_unlock_hb(hb1, hb2);
1262 put_futex_key(fshared, &key2);
1263 put_futex_key(fshared, &key1);
1271 head1 = &hb1->chain;
1272 plist_for_each_entry_safe(this, next, head1, list) {
1273 if (task_count - nr_wake >= nr_requeue)
1276 if (!match_futex(&this->key, &key1))
1280 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1281 * be paired with each other and no other futex ops.
1283 if ((requeue_pi && !this->rt_waiter) ||
1284 (!requeue_pi && this->rt_waiter)) {
1290 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1291 * lock, we already woke the top_waiter. If not, it will be
1292 * woken by futex_unlock_pi().
1294 if (++task_count <= nr_wake && !requeue_pi) {
1299 /* Ensure we requeue to the expected futex for requeue_pi. */
1300 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1306 * Requeue nr_requeue waiters and possibly one more in the case
1307 * of requeue_pi if we couldn't acquire the lock atomically.
1310 /* Prepare the waiter to take the rt_mutex. */
1311 atomic_inc(&pi_state->refcount);
1312 this->pi_state = pi_state;
1313 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1317 /* We got the lock. */
1318 requeue_pi_wake_futex(this, &key2, hb2);
1323 this->pi_state = NULL;
1324 free_pi_state(pi_state);
1328 requeue_futex(this, hb1, hb2, &key2);
1333 double_unlock_hb(hb1, hb2);
1336 * drop_futex_key_refs() must be called outside the spinlocks. During
1337 * the requeue we moved futex_q's from the hash bucket at key1 to the
1338 * one at key2 and updated their key pointer. We no longer need to
1339 * hold the references to key1.
1341 while (--drop_count >= 0)
1342 drop_futex_key_refs(&key1);
1345 put_futex_key(fshared, &key2);
1347 put_futex_key(fshared, &key1);
1349 if (pi_state != NULL)
1350 free_pi_state(pi_state);
1351 return ret ? ret : task_count;
1354 /* The key must be already stored in q->key. */
1355 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1357 struct futex_hash_bucket *hb;
1359 hb = hash_futex(&q->key);
1360 q->lock_ptr = &hb->lock;
1362 spin_lock(&hb->lock);
1367 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1369 spin_unlock(&hb->lock);
1373 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1374 * @q: The futex_q to enqueue
1375 * @hb: The destination hash bucket
1377 * The hb->lock must be held by the caller, and is released here. A call to
1378 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1379 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1380 * or nothing if the unqueue is done as part of the wake process and the unqueue
1381 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1384 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1389 * The priority used to register this element is
1390 * - either the real thread-priority for the real-time threads
1391 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1392 * - or MAX_RT_PRIO for non-RT threads.
1393 * Thus, all RT-threads are woken first in priority order, and
1394 * the others are woken last, in FIFO order.
1396 prio = min(current->normal_prio, MAX_RT_PRIO);
1398 plist_node_init(&q->list, prio);
1399 #ifdef CONFIG_DEBUG_PI_LIST
1400 q->list.plist.spinlock = &hb->lock;
1402 plist_add(&q->list, &hb->chain);
1404 spin_unlock(&hb->lock);
1408 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1409 * @q: The futex_q to unqueue
1411 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1412 * be paired with exactly one earlier call to queue_me().
1415 * 1 - if the futex_q was still queued (and we removed unqueued it)
1416 * 0 - if the futex_q was already removed by the waking thread
1418 static int unqueue_me(struct futex_q *q)
1420 spinlock_t *lock_ptr;
1423 /* In the common case we don't take the spinlock, which is nice. */
1425 lock_ptr = q->lock_ptr;
1427 if (lock_ptr != NULL) {
1428 spin_lock(lock_ptr);
1430 * q->lock_ptr can change between reading it and
1431 * spin_lock(), causing us to take the wrong lock. This
1432 * corrects the race condition.
1434 * Reasoning goes like this: if we have the wrong lock,
1435 * q->lock_ptr must have changed (maybe several times)
1436 * between reading it and the spin_lock(). It can
1437 * change again after the spin_lock() but only if it was
1438 * already changed before the spin_lock(). It cannot,
1439 * however, change back to the original value. Therefore
1440 * we can detect whether we acquired the correct lock.
1442 if (unlikely(lock_ptr != q->lock_ptr)) {
1443 spin_unlock(lock_ptr);
1446 WARN_ON(plist_node_empty(&q->list));
1447 plist_del(&q->list, &q->list.plist);
1449 BUG_ON(q->pi_state);
1451 spin_unlock(lock_ptr);
1455 drop_futex_key_refs(&q->key);
1460 * PI futexes can not be requeued and must remove themself from the
1461 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1464 static void unqueue_me_pi(struct futex_q *q)
1466 WARN_ON(plist_node_empty(&q->list));
1467 plist_del(&q->list, &q->list.plist);
1469 BUG_ON(!q->pi_state);
1470 free_pi_state(q->pi_state);
1473 spin_unlock(q->lock_ptr);
1477 * Fixup the pi_state owner with the new owner.
1479 * Must be called with hash bucket lock held and mm->sem held for non
1482 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1483 struct task_struct *newowner, int fshared)
1485 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1486 struct futex_pi_state *pi_state = q->pi_state;
1487 struct task_struct *oldowner = pi_state->owner;
1488 u32 uval, curval, newval;
1492 if (!pi_state->owner)
1493 newtid |= FUTEX_OWNER_DIED;
1496 * We are here either because we stole the rtmutex from the
1497 * pending owner or we are the pending owner which failed to
1498 * get the rtmutex. We have to replace the pending owner TID
1499 * in the user space variable. This must be atomic as we have
1500 * to preserve the owner died bit here.
1502 * Note: We write the user space value _before_ changing the pi_state
1503 * because we can fault here. Imagine swapped out pages or a fork
1504 * that marked all the anonymous memory readonly for cow.
1506 * Modifying pi_state _before_ the user space value would
1507 * leave the pi_state in an inconsistent state when we fault
1508 * here, because we need to drop the hash bucket lock to
1509 * handle the fault. This might be observed in the PID check
1510 * in lookup_pi_state.
1513 if (get_futex_value_locked(&uval, uaddr))
1517 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1519 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1527 * We fixed up user space. Now we need to fix the pi_state
1530 if (pi_state->owner != NULL) {
1531 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1532 WARN_ON(list_empty(&pi_state->list));
1533 list_del_init(&pi_state->list);
1534 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1537 pi_state->owner = newowner;
1539 raw_spin_lock_irq(&newowner->pi_lock);
1540 WARN_ON(!list_empty(&pi_state->list));
1541 list_add(&pi_state->list, &newowner->pi_state_list);
1542 raw_spin_unlock_irq(&newowner->pi_lock);
1546 * To handle the page fault we need to drop the hash bucket
1547 * lock here. That gives the other task (either the pending
1548 * owner itself or the task which stole the rtmutex) the
1549 * chance to try the fixup of the pi_state. So once we are
1550 * back from handling the fault we need to check the pi_state
1551 * after reacquiring the hash bucket lock and before trying to
1552 * do another fixup. When the fixup has been done already we
1556 spin_unlock(q->lock_ptr);
1558 ret = fault_in_user_writeable(uaddr);
1560 spin_lock(q->lock_ptr);
1563 * Check if someone else fixed it for us:
1565 if (pi_state->owner != oldowner)
1575 * In case we must use restart_block to restart a futex_wait,
1576 * we encode in the 'flags' shared capability
1578 #define FLAGS_SHARED 0x01
1579 #define FLAGS_CLOCKRT 0x02
1580 #define FLAGS_HAS_TIMEOUT 0x04
1582 static long futex_wait_restart(struct restart_block *restart);
1585 * fixup_owner() - Post lock pi_state and corner case management
1586 * @uaddr: user address of the futex
1587 * @fshared: whether the futex is shared (1) or not (0)
1588 * @q: futex_q (contains pi_state and access to the rt_mutex)
1589 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1591 * After attempting to lock an rt_mutex, this function is called to cleanup
1592 * the pi_state owner as well as handle race conditions that may allow us to
1593 * acquire the lock. Must be called with the hb lock held.
1596 * 1 - success, lock taken
1597 * 0 - success, lock not taken
1598 * <0 - on error (-EFAULT)
1600 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1603 struct task_struct *owner;
1608 * Got the lock. We might not be the anticipated owner if we
1609 * did a lock-steal - fix up the PI-state in that case:
1611 if (q->pi_state->owner != current)
1612 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1617 * Catch the rare case, where the lock was released when we were on the
1618 * way back before we locked the hash bucket.
1620 if (q->pi_state->owner == current) {
1622 * Try to get the rt_mutex now. This might fail as some other
1623 * task acquired the rt_mutex after we removed ourself from the
1624 * rt_mutex waiters list.
1626 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1632 * pi_state is incorrect, some other task did a lock steal and
1633 * we returned due to timeout or signal without taking the
1634 * rt_mutex. Too late. We can access the rt_mutex_owner without
1635 * locking, as the other task is now blocked on the hash bucket
1636 * lock. Fix the state up.
1638 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1639 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1644 * Paranoia check. If we did not take the lock, then we should not be
1645 * the owner, nor the pending owner, of the rt_mutex.
1647 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1648 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1649 "pi-state %p\n", ret,
1650 q->pi_state->pi_mutex.owner,
1651 q->pi_state->owner);
1654 return ret ? ret : locked;
1658 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1659 * @hb: the futex hash bucket, must be locked by the caller
1660 * @q: the futex_q to queue up on
1661 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1663 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1664 struct hrtimer_sleeper *timeout)
1667 * The task state is guaranteed to be set before another task can
1668 * wake it. set_current_state() is implemented using set_mb() and
1669 * queue_me() calls spin_unlock() upon completion, both serializing
1670 * access to the hash list and forcing another memory barrier.
1672 set_current_state(TASK_INTERRUPTIBLE);
1677 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1678 if (!hrtimer_active(&timeout->timer))
1679 timeout->task = NULL;
1683 * If we have been removed from the hash list, then another task
1684 * has tried to wake us, and we can skip the call to schedule().
1686 if (likely(!plist_node_empty(&q->list))) {
1688 * If the timer has already expired, current will already be
1689 * flagged for rescheduling. Only call schedule if there
1690 * is no timeout, or if it has yet to expire.
1692 if (!timeout || timeout->task)
1695 __set_current_state(TASK_RUNNING);
1699 * futex_wait_setup() - Prepare to wait on a futex
1700 * @uaddr: the futex userspace address
1701 * @val: the expected value
1702 * @fshared: whether the futex is shared (1) or not (0)
1703 * @q: the associated futex_q
1704 * @hb: storage for hash_bucket pointer to be returned to caller
1706 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1707 * compare it with the expected value. Handle atomic faults internally.
1708 * Return with the hb lock held and a q.key reference on success, and unlocked
1709 * with no q.key reference on failure.
1712 * 0 - uaddr contains val and hb has been locked
1713 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1715 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1716 struct futex_q *q, struct futex_hash_bucket **hb)
1722 * Access the page AFTER the hash-bucket is locked.
1723 * Order is important:
1725 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1726 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1728 * The basic logical guarantee of a futex is that it blocks ONLY
1729 * if cond(var) is known to be true at the time of blocking, for
1730 * any cond. If we queued after testing *uaddr, that would open
1731 * a race condition where we could block indefinitely with
1732 * cond(var) false, which would violate the guarantee.
1734 * A consequence is that futex_wait() can return zero and absorb
1735 * a wakeup when *uaddr != val on entry to the syscall. This is
1739 q->key = FUTEX_KEY_INIT;
1740 ret = get_futex_key(uaddr, fshared, &q->key);
1741 if (unlikely(ret != 0))
1745 *hb = queue_lock(q);
1747 ret = get_futex_value_locked(&uval, uaddr);
1750 queue_unlock(q, *hb);
1752 ret = get_user(uval, uaddr);
1759 put_futex_key(fshared, &q->key);
1764 queue_unlock(q, *hb);
1770 put_futex_key(fshared, &q->key);
1774 static int futex_wait(u32 __user *uaddr, int fshared,
1775 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1777 struct hrtimer_sleeper timeout, *to = NULL;
1778 struct restart_block *restart;
1779 struct futex_hash_bucket *hb;
1789 q.requeue_pi_key = NULL;
1794 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1795 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1796 hrtimer_init_sleeper(to, current);
1797 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1798 current->timer_slack_ns);
1803 * Prepare to wait on uaddr. On success, holds hb lock and increments
1806 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1810 /* queue_me and wait for wakeup, timeout, or a signal. */
1811 futex_wait_queue_me(hb, &q, to);
1813 /* If we were woken (and unqueued), we succeeded, whatever. */
1815 /* unqueue_me() drops q.key ref */
1816 if (!unqueue_me(&q))
1819 if (to && !to->task)
1823 * We expect signal_pending(current), but we might be the
1824 * victim of a spurious wakeup as well.
1826 if (!signal_pending(current))
1833 restart = ¤t_thread_info()->restart_block;
1834 restart->fn = futex_wait_restart;
1835 restart->futex.uaddr = (u32 *)uaddr;
1836 restart->futex.val = val;
1837 restart->futex.time = abs_time->tv64;
1838 restart->futex.bitset = bitset;
1839 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1842 restart->futex.flags |= FLAGS_SHARED;
1844 restart->futex.flags |= FLAGS_CLOCKRT;
1846 ret = -ERESTART_RESTARTBLOCK;
1850 hrtimer_cancel(&to->timer);
1851 destroy_hrtimer_on_stack(&to->timer);
1857 static long futex_wait_restart(struct restart_block *restart)
1859 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1861 ktime_t t, *tp = NULL;
1863 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1864 t.tv64 = restart->futex.time;
1867 restart->fn = do_no_restart_syscall;
1868 if (restart->futex.flags & FLAGS_SHARED)
1870 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1871 restart->futex.bitset,
1872 restart->futex.flags & FLAGS_CLOCKRT);
1877 * Userspace tried a 0 -> TID atomic transition of the futex value
1878 * and failed. The kernel side here does the whole locking operation:
1879 * if there are waiters then it will block, it does PI, etc. (Due to
1880 * races the kernel might see a 0 value of the futex too.)
1882 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1883 int detect, ktime_t *time, int trylock)
1885 struct hrtimer_sleeper timeout, *to = NULL;
1886 struct futex_hash_bucket *hb;
1890 if (refill_pi_state_cache())
1895 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1897 hrtimer_init_sleeper(to, current);
1898 hrtimer_set_expires(&to->timer, *time);
1903 q.requeue_pi_key = NULL;
1905 q.key = FUTEX_KEY_INIT;
1906 ret = get_futex_key(uaddr, fshared, &q.key);
1907 if (unlikely(ret != 0))
1911 hb = queue_lock(&q);
1913 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1914 if (unlikely(ret)) {
1917 /* We got the lock. */
1919 goto out_unlock_put_key;
1924 * Task is exiting and we just wait for the
1927 queue_unlock(&q, hb);
1928 put_futex_key(fshared, &q.key);
1932 goto out_unlock_put_key;
1937 * Only actually queue now that the atomic ops are done:
1941 WARN_ON(!q.pi_state);
1943 * Block on the PI mutex:
1946 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1948 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1949 /* Fixup the trylock return value: */
1950 ret = ret ? 0 : -EWOULDBLOCK;
1953 spin_lock(q.lock_ptr);
1955 * Fixup the pi_state owner and possibly acquire the lock if we
1958 res = fixup_owner(uaddr, fshared, &q, !ret);
1960 * If fixup_owner() returned an error, proprogate that. If it acquired
1961 * the lock, clear our -ETIMEDOUT or -EINTR.
1964 ret = (res < 0) ? res : 0;
1967 * If fixup_owner() faulted and was unable to handle the fault, unlock
1968 * it and return the fault to userspace.
1970 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1971 rt_mutex_unlock(&q.pi_state->pi_mutex);
1973 /* Unqueue and drop the lock */
1979 queue_unlock(&q, hb);
1982 put_futex_key(fshared, &q.key);
1985 destroy_hrtimer_on_stack(&to->timer);
1986 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1989 queue_unlock(&q, hb);
1991 ret = fault_in_user_writeable(uaddr);
1998 put_futex_key(fshared, &q.key);
2003 * Userspace attempted a TID -> 0 atomic transition, and failed.
2004 * This is the in-kernel slowpath: we look up the PI state (if any),
2005 * and do the rt-mutex unlock.
2007 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2009 struct futex_hash_bucket *hb;
2010 struct futex_q *this, *next;
2011 u32 uval, vpid = task_pid_vnr(current);
2012 struct plist_head *head;
2013 union futex_key key = FUTEX_KEY_INIT;
2017 if (get_user(uval, uaddr))
2020 * We release only a lock we actually own:
2022 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2025 ret = get_futex_key(uaddr, fshared, &key);
2026 if (unlikely(ret != 0))
2029 hb = hash_futex(&key);
2030 spin_lock(&hb->lock);
2033 * To avoid races, try to do the TID -> 0 atomic transition
2034 * again. If it succeeds then we can return without waking
2037 if (!(uval & FUTEX_OWNER_DIED) &&
2038 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2041 * Rare case: we managed to release the lock atomically,
2042 * no need to wake anyone else up:
2044 if (unlikely(uval == task_pid_vnr(current)))
2048 * Ok, other tasks may need to be woken up - check waiters
2049 * and do the wakeup if necessary:
2053 plist_for_each_entry_safe(this, next, head, list) {
2054 if (!match_futex (&this->key, &key))
2056 ret = wake_futex_pi(uaddr, uval, this);
2058 * The atomic access to the futex value
2059 * generated a pagefault, so retry the
2060 * user-access and the wakeup:
2067 * No waiters - kernel unlocks the futex:
2069 if (!(uval & FUTEX_OWNER_DIED)) {
2070 ret = unlock_futex_pi(uaddr, uval);
2076 spin_unlock(&hb->lock);
2077 put_futex_key(fshared, &key);
2083 spin_unlock(&hb->lock);
2084 put_futex_key(fshared, &key);
2086 ret = fault_in_user_writeable(uaddr);
2094 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2095 * @hb: the hash_bucket futex_q was original enqueued on
2096 * @q: the futex_q woken while waiting to be requeued
2097 * @key2: the futex_key of the requeue target futex
2098 * @timeout: the timeout associated with the wait (NULL if none)
2100 * Detect if the task was woken on the initial futex as opposed to the requeue
2101 * target futex. If so, determine if it was a timeout or a signal that caused
2102 * the wakeup and return the appropriate error code to the caller. Must be
2103 * called with the hb lock held.
2106 * 0 - no early wakeup detected
2107 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2110 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2111 struct futex_q *q, union futex_key *key2,
2112 struct hrtimer_sleeper *timeout)
2117 * With the hb lock held, we avoid races while we process the wakeup.
2118 * We only need to hold hb (and not hb2) to ensure atomicity as the
2119 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2120 * It can't be requeued from uaddr2 to something else since we don't
2121 * support a PI aware source futex for requeue.
2123 if (!match_futex(&q->key, key2)) {
2124 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2126 * We were woken prior to requeue by a timeout or a signal.
2127 * Unqueue the futex_q and determine which it was.
2129 plist_del(&q->list, &q->list.plist);
2131 /* Handle spurious wakeups gracefully */
2133 if (timeout && !timeout->task)
2135 else if (signal_pending(current))
2136 ret = -ERESTARTNOINTR;
2142 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2143 * @uaddr: the futex we initially wait on (non-pi)
2144 * @fshared: whether the futexes are shared (1) or not (0). They must be
2145 * the same type, no requeueing from private to shared, etc.
2146 * @val: the expected value of uaddr
2147 * @abs_time: absolute timeout
2148 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2149 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2150 * @uaddr2: the pi futex we will take prior to returning to user-space
2152 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2153 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2154 * complete the acquisition of the rt_mutex prior to returning to userspace.
2155 * This ensures the rt_mutex maintains an owner when it has waiters; without
2156 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2159 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2160 * via the following:
2161 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2162 * 2) wakeup on uaddr2 after a requeue
2166 * If 3, cleanup and return -ERESTARTNOINTR.
2168 * If 2, we may then block on trying to take the rt_mutex and return via:
2169 * 5) successful lock
2172 * 8) other lock acquisition failure
2174 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2176 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2182 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2183 u32 val, ktime_t *abs_time, u32 bitset,
2184 int clockrt, u32 __user *uaddr2)
2186 struct hrtimer_sleeper timeout, *to = NULL;
2187 struct rt_mutex_waiter rt_waiter;
2188 struct rt_mutex *pi_mutex = NULL;
2189 struct futex_hash_bucket *hb;
2190 union futex_key key2;
2199 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2200 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2201 hrtimer_init_sleeper(to, current);
2202 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2203 current->timer_slack_ns);
2207 * The waiter is allocated on our stack, manipulated by the requeue
2208 * code while we sleep on uaddr.
2210 debug_rt_mutex_init_waiter(&rt_waiter);
2211 rt_waiter.task = NULL;
2213 key2 = FUTEX_KEY_INIT;
2214 ret = get_futex_key(uaddr2, fshared, &key2);
2215 if (unlikely(ret != 0))
2220 q.rt_waiter = &rt_waiter;
2221 q.requeue_pi_key = &key2;
2224 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2227 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2231 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2232 futex_wait_queue_me(hb, &q, to);
2234 spin_lock(&hb->lock);
2235 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2236 spin_unlock(&hb->lock);
2241 * In order for us to be here, we know our q.key == key2, and since
2242 * we took the hb->lock above, we also know that futex_requeue() has
2243 * completed and we no longer have to concern ourselves with a wakeup
2244 * race with the atomic proxy lock acquisition by the requeue code. The
2245 * futex_requeue dropped our key1 reference and incremented our key2
2249 /* Check if the requeue code acquired the second futex for us. */
2252 * Got the lock. We might not be the anticipated owner if we
2253 * did a lock-steal - fix up the PI-state in that case.
2255 if (q.pi_state && (q.pi_state->owner != current)) {
2256 spin_lock(q.lock_ptr);
2257 ret = fixup_pi_state_owner(uaddr2, &q, current,
2259 spin_unlock(q.lock_ptr);
2263 * We have been woken up by futex_unlock_pi(), a timeout, or a
2264 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2267 WARN_ON(!&q.pi_state);
2268 pi_mutex = &q.pi_state->pi_mutex;
2269 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2270 debug_rt_mutex_free_waiter(&rt_waiter);
2272 spin_lock(q.lock_ptr);
2274 * Fixup the pi_state owner and possibly acquire the lock if we
2277 res = fixup_owner(uaddr2, fshared, &q, !ret);
2279 * If fixup_owner() returned an error, proprogate that. If it
2280 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2283 ret = (res < 0) ? res : 0;
2285 /* Unqueue and drop the lock. */
2290 * If fixup_pi_state_owner() faulted and was unable to handle the
2291 * fault, unlock the rt_mutex and return the fault to userspace.
2293 if (ret == -EFAULT) {
2294 if (rt_mutex_owner(pi_mutex) == current)
2295 rt_mutex_unlock(pi_mutex);
2296 } else if (ret == -EINTR) {
2298 * We've already been requeued, but cannot restart by calling
2299 * futex_lock_pi() directly. We could restart this syscall, but
2300 * it would detect that the user space "val" changed and return
2301 * -EWOULDBLOCK. Save the overhead of the restart and return
2302 * -EWOULDBLOCK directly.
2308 put_futex_key(fshared, &q.key);
2310 put_futex_key(fshared, &key2);
2314 hrtimer_cancel(&to->timer);
2315 destroy_hrtimer_on_stack(&to->timer);
2321 * Support for robust futexes: the kernel cleans up held futexes at
2324 * Implementation: user-space maintains a per-thread list of locks it
2325 * is holding. Upon do_exit(), the kernel carefully walks this list,
2326 * and marks all locks that are owned by this thread with the
2327 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2328 * always manipulated with the lock held, so the list is private and
2329 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2330 * field, to allow the kernel to clean up if the thread dies after
2331 * acquiring the lock, but just before it could have added itself to
2332 * the list. There can only be one such pending lock.
2336 * sys_set_robust_list() - Set the robust-futex list head of a task
2337 * @head: pointer to the list-head
2338 * @len: length of the list-head, as userspace expects
2340 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2343 if (!futex_cmpxchg_enabled)
2346 * The kernel knows only one size for now:
2348 if (unlikely(len != sizeof(*head)))
2351 current->robust_list = head;
2357 * sys_get_robust_list() - Get the robust-futex list head of a task
2358 * @pid: pid of the process [zero for current task]
2359 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2360 * @len_ptr: pointer to a length field, the kernel fills in the header size
2362 SYSCALL_DEFINE3(get_robust_list, int, pid,
2363 struct robust_list_head __user * __user *, head_ptr,
2364 size_t __user *, len_ptr)
2366 struct robust_list_head __user *head;
2368 const struct cred *cred = current_cred(), *pcred;
2370 if (!futex_cmpxchg_enabled)
2374 head = current->robust_list;
2376 struct task_struct *p;
2380 p = find_task_by_vpid(pid);
2384 pcred = __task_cred(p);
2385 if (cred->euid != pcred->euid &&
2386 cred->euid != pcred->uid &&
2387 !capable(CAP_SYS_PTRACE))
2389 head = p->robust_list;
2393 if (put_user(sizeof(*head), len_ptr))
2395 return put_user(head, head_ptr);
2404 * Process a futex-list entry, check whether it's owned by the
2405 * dying task, and do notification if so:
2407 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2409 u32 uval, nval, mval;
2412 if (get_user(uval, uaddr))
2415 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2417 * Ok, this dying thread is truly holding a futex
2418 * of interest. Set the OWNER_DIED bit atomically
2419 * via cmpxchg, and if the value had FUTEX_WAITERS
2420 * set, wake up a waiter (if any). (We have to do a
2421 * futex_wake() even if OWNER_DIED is already set -
2422 * to handle the rare but possible case of recursive
2423 * thread-death.) The rest of the cleanup is done in
2426 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2427 if (futex_atomic_cmpxchg_inatomic(&nval, uaddr, uval, mval))
2434 * Wake robust non-PI futexes here. The wakeup of
2435 * PI futexes happens in exit_pi_state():
2437 if (!pi && (uval & FUTEX_WAITERS))
2438 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2444 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2446 static inline int fetch_robust_entry(struct robust_list __user **entry,
2447 struct robust_list __user * __user *head,
2450 unsigned long uentry;
2452 if (get_user(uentry, (unsigned long __user *)head))
2455 *entry = (void __user *)(uentry & ~1UL);
2462 * Walk curr->robust_list (very carefully, it's a userspace list!)
2463 * and mark any locks found there dead, and notify any waiters.
2465 * We silently return on any sign of list-walking problem.
2467 void exit_robust_list(struct task_struct *curr)
2469 struct robust_list_head __user *head = curr->robust_list;
2470 struct robust_list __user *entry, *next_entry, *pending;
2471 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2472 unsigned long futex_offset;
2475 if (!futex_cmpxchg_enabled)
2479 * Fetch the list head (which was registered earlier, via
2480 * sys_set_robust_list()):
2482 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2485 * Fetch the relative futex offset:
2487 if (get_user(futex_offset, &head->futex_offset))
2490 * Fetch any possibly pending lock-add first, and handle it
2493 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2496 next_entry = NULL; /* avoid warning with gcc */
2497 while (entry != &head->list) {
2499 * Fetch the next entry in the list before calling
2500 * handle_futex_death:
2502 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2504 * A pending lock might already be on the list, so
2505 * don't process it twice:
2507 if (entry != pending)
2508 if (handle_futex_death((void __user *)entry + futex_offset,
2516 * Avoid excessively long or circular lists:
2525 handle_futex_death((void __user *)pending + futex_offset,
2529 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2530 u32 __user *uaddr2, u32 val2, u32 val3)
2532 int clockrt, ret = -ENOSYS;
2533 int cmd = op & FUTEX_CMD_MASK;
2536 if (!(op & FUTEX_PRIVATE_FLAG))
2539 clockrt = op & FUTEX_CLOCK_REALTIME;
2540 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2545 val3 = FUTEX_BITSET_MATCH_ANY;
2546 case FUTEX_WAIT_BITSET:
2547 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2550 val3 = FUTEX_BITSET_MATCH_ANY;
2551 case FUTEX_WAKE_BITSET:
2552 ret = futex_wake(uaddr, fshared, val, val3);
2555 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2557 case FUTEX_CMP_REQUEUE:
2558 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2562 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2565 if (futex_cmpxchg_enabled)
2566 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2568 case FUTEX_UNLOCK_PI:
2569 if (futex_cmpxchg_enabled)
2570 ret = futex_unlock_pi(uaddr, fshared);
2572 case FUTEX_TRYLOCK_PI:
2573 if (futex_cmpxchg_enabled)
2574 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2576 case FUTEX_WAIT_REQUEUE_PI:
2577 val3 = FUTEX_BITSET_MATCH_ANY;
2578 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2581 case FUTEX_CMP_REQUEUE_PI:
2582 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2592 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2593 struct timespec __user *, utime, u32 __user *, uaddr2,
2597 ktime_t t, *tp = NULL;
2599 int cmd = op & FUTEX_CMD_MASK;
2601 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2602 cmd == FUTEX_WAIT_BITSET ||
2603 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2604 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2606 if (!timespec_valid(&ts))
2609 t = timespec_to_ktime(ts);
2610 if (cmd == FUTEX_WAIT)
2611 t = ktime_add_safe(ktime_get(), t);
2615 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2616 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2618 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2619 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2620 val2 = (u32) (unsigned long) utime;
2622 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2625 static int __init futex_init(void)
2631 * This will fail and we want it. Some arch implementations do
2632 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2633 * functionality. We want to know that before we call in any
2634 * of the complex code paths. Also we want to prevent
2635 * registration of robust lists in that case. NULL is
2636 * guaranteed to fault and we get -EFAULT on functional
2637 * implementation, the non functional ones will return
2640 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2641 futex_cmpxchg_enabled = 1;
2643 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2644 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2645 spin_lock_init(&futex_queues[i].lock);
2650 __initcall(futex_init);