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 * @list: priority-sorted list of tasks waiting on this futex
95 * @task: the task waiting on the futex
96 * @lock_ptr: the hash bucket lock
97 * @key: the key the futex is hashed on
98 * @pi_state: optional priority inheritance state
99 * @rt_waiter: rt_waiter storage for use with requeue_pi
100 * @requeue_pi_key: the requeue_pi target futex key
101 * @bitset: bitset for the optional bitmasked wakeup
103 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
104 * we can wake only the relevant ones (hashed queues may be shared).
106 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
107 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
108 * The order of wakeup is always to make the first condition true, then
111 * PI futexes are typically woken before they are removed from the hash list via
112 * the rt_mutex code. See unqueue_me_pi().
115 struct plist_node list;
117 struct task_struct *task;
118 spinlock_t *lock_ptr;
120 struct futex_pi_state *pi_state;
121 struct rt_mutex_waiter *rt_waiter;
122 union futex_key *requeue_pi_key;
127 * Hash buckets are shared by all the futex_keys that hash to the same
128 * location. Each key may have multiple futex_q structures, one for each task
129 * waiting on a futex.
131 struct futex_hash_bucket {
133 struct plist_head chain;
136 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
139 * We hash on the keys returned from get_futex_key (see below).
141 static struct futex_hash_bucket *hash_futex(union futex_key *key)
143 u32 hash = jhash2((u32*)&key->both.word,
144 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
146 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
150 * Return 1 if two futex_keys are equal, 0 otherwise.
152 static inline int match_futex(union futex_key *key1, union futex_key *key2)
155 && key1->both.word == key2->both.word
156 && key1->both.ptr == key2->both.ptr
157 && key1->both.offset == key2->both.offset);
161 * Take a reference to the resource addressed by a key.
162 * Can be called while holding spinlocks.
165 static void get_futex_key_refs(union futex_key *key)
170 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
172 atomic_inc(&key->shared.inode->i_count);
174 case FUT_OFF_MMSHARED:
175 atomic_inc(&key->private.mm->mm_count);
181 * Drop a reference to the resource addressed by a key.
182 * The hash bucket spinlock must not be held.
184 static void drop_futex_key_refs(union futex_key *key)
186 if (!key->both.ptr) {
187 /* If we're here then we tried to put a key we failed to get */
192 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
194 iput(key->shared.inode);
196 case FUT_OFF_MMSHARED:
197 mmdrop(key->private.mm);
203 * get_futex_key() - Get parameters which are the keys for a futex
204 * @uaddr: virtual address of the futex
205 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
206 * @key: address where result is stored.
208 * Returns a negative error code or 0
209 * The key words are stored in *key on success.
211 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
212 * offset_within_page). For private mappings, it's (uaddr, current->mm).
213 * We can usually work out the index without swapping in the page.
215 * lock_page() might sleep, the caller should not hold a spinlock.
218 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
220 unsigned long address = (unsigned long)uaddr;
221 struct mm_struct *mm = current->mm;
226 * The futex address must be "naturally" aligned.
228 key->both.offset = address % PAGE_SIZE;
229 if (unlikely((address % sizeof(u32)) != 0))
231 address -= key->both.offset;
234 * PROCESS_PRIVATE futexes are fast.
235 * As the mm cannot disappear under us and the 'key' only needs
236 * virtual address, we dont even have to find the underlying vma.
237 * Note : We do have to check 'uaddr' is a valid user address,
238 * but access_ok() should be faster than find_vma()
241 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
243 key->private.mm = mm;
244 key->private.address = address;
245 get_futex_key_refs(key);
250 err = get_user_pages_fast(address, 1, 1, &page);
254 page = compound_head(page);
256 if (!page->mapping) {
263 * Private mappings are handled in a simple way.
265 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
266 * it's a read-only handle, it's expected that futexes attach to
267 * the object not the particular process.
269 if (PageAnon(page)) {
270 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
271 key->private.mm = mm;
272 key->private.address = address;
274 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
275 key->shared.inode = page->mapping->host;
276 key->shared.pgoff = page->index;
279 get_futex_key_refs(key);
287 void put_futex_key(int fshared, union futex_key *key)
289 drop_futex_key_refs(key);
293 * fault_in_user_writeable() - Fault in user address and verify RW access
294 * @uaddr: pointer to faulting user space address
296 * Slow path to fixup the fault we just took in the atomic write
299 * We have no generic implementation of a non-destructive write to the
300 * user address. We know that we faulted in the atomic pagefault
301 * disabled section so we can as well avoid the #PF overhead by
302 * calling get_user_pages() right away.
304 static int fault_in_user_writeable(u32 __user *uaddr)
306 struct mm_struct *mm = current->mm;
309 down_read(&mm->mmap_sem);
310 ret = get_user_pages(current, mm, (unsigned long)uaddr,
311 1, 1, 0, NULL, NULL);
312 up_read(&mm->mmap_sem);
314 return ret < 0 ? ret : 0;
318 * futex_top_waiter() - Return the highest priority waiter on a futex
319 * @hb: the hash bucket the futex_q's reside in
320 * @key: the futex key (to distinguish it from other futex futex_q's)
322 * Must be called with the hb lock held.
324 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
325 union futex_key *key)
327 struct futex_q *this;
329 plist_for_each_entry(this, &hb->chain, list) {
330 if (match_futex(&this->key, key))
336 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
341 curval = futex_atomic_cmpxchg_inatomic(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 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
645 if (unlikely(curval == -EFAULT))
651 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
655 * Surprise - we got the lock. Just return to userspace:
657 if (unlikely(!curval))
663 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
664 * to wake at the next unlock.
666 newval = curval | FUTEX_WAITERS;
669 * There are two cases, where a futex might have no owner (the
670 * owner TID is 0): OWNER_DIED. We take over the futex in this
671 * case. We also do an unconditional take over, when the owner
674 * This is safe as we are protected by the hash bucket lock !
676 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
677 /* Keep the OWNER_DIED bit */
678 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
683 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
685 if (unlikely(curval == -EFAULT))
687 if (unlikely(curval != uval))
691 * We took the lock due to owner died take over.
693 if (unlikely(lock_taken))
697 * We dont have the lock. Look up the PI state (or create it if
698 * we are the first waiter):
700 ret = lookup_pi_state(uval, hb, key, ps);
706 * No owner found for this futex. Check if the
707 * OWNER_DIED bit is set to figure out whether
708 * this is a robust futex or not.
710 if (get_futex_value_locked(&curval, uaddr))
714 * We simply start over in case of a robust
715 * futex. The code above will take the futex
718 if (curval & FUTEX_OWNER_DIED) {
731 * The hash bucket lock must be held when this is called.
732 * Afterwards, the futex_q must not be accessed.
734 static void wake_futex(struct futex_q *q)
736 struct task_struct *p = q->task;
739 * We set q->lock_ptr = NULL _before_ we wake up the task. If
740 * a non-futex wake up happens on another CPU then the task
741 * might exit and p would dereference a non-existing task
742 * struct. Prevent this by holding a reference on p across the
747 plist_del(&q->list, &q->list.plist);
749 * The waiting task can free the futex_q as soon as
750 * q->lock_ptr = NULL is written, without taking any locks. A
751 * memory barrier is required here to prevent the following
752 * store to lock_ptr from getting ahead of the plist_del.
757 wake_up_state(p, TASK_NORMAL);
761 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
763 struct task_struct *new_owner;
764 struct futex_pi_state *pi_state = this->pi_state;
771 * If current does not own the pi_state then the futex is
772 * inconsistent and user space fiddled with the futex value.
774 if (pi_state->owner != current)
777 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
778 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
781 * This happens when we have stolen the lock and the original
782 * pending owner did not enqueue itself back on the rt_mutex.
783 * Thats not a tragedy. We know that way, that a lock waiter
784 * is on the fly. We make the futex_q waiter the pending owner.
787 new_owner = this->task;
790 * We pass it to the next owner. (The WAITERS bit is always
791 * kept enabled while there is PI state around. We must also
792 * preserve the owner died bit.)
794 if (!(uval & FUTEX_OWNER_DIED)) {
797 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
799 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
801 if (curval == -EFAULT)
803 else if (curval != uval)
806 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
811 raw_spin_lock_irq(&pi_state->owner->pi_lock);
812 WARN_ON(list_empty(&pi_state->list));
813 list_del_init(&pi_state->list);
814 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
816 raw_spin_lock_irq(&new_owner->pi_lock);
817 WARN_ON(!list_empty(&pi_state->list));
818 list_add(&pi_state->list, &new_owner->pi_state_list);
819 pi_state->owner = new_owner;
820 raw_spin_unlock_irq(&new_owner->pi_lock);
822 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
823 rt_mutex_unlock(&pi_state->pi_mutex);
828 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
833 * There is no waiter, so we unlock the futex. The owner died
834 * bit has not to be preserved here. We are the owner:
836 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
838 if (oldval == -EFAULT)
847 * Express the locking dependencies for lockdep:
850 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
853 spin_lock(&hb1->lock);
855 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
856 } else { /* hb1 > hb2 */
857 spin_lock(&hb2->lock);
858 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
863 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
865 spin_unlock(&hb1->lock);
867 spin_unlock(&hb2->lock);
871 * Wake up waiters matching bitset queued on this futex (uaddr).
873 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
875 struct futex_hash_bucket *hb;
876 struct futex_q *this, *next;
877 struct plist_head *head;
878 union futex_key key = FUTEX_KEY_INIT;
884 ret = get_futex_key(uaddr, fshared, &key);
885 if (unlikely(ret != 0))
888 hb = hash_futex(&key);
889 spin_lock(&hb->lock);
892 plist_for_each_entry_safe(this, next, head, list) {
893 if (match_futex (&this->key, &key)) {
894 if (this->pi_state || this->rt_waiter) {
899 /* Check if one of the bits is set in both bitsets */
900 if (!(this->bitset & bitset))
904 if (++ret >= nr_wake)
909 spin_unlock(&hb->lock);
910 put_futex_key(fshared, &key);
916 * Wake up all waiters hashed on the physical page that is mapped
917 * to this virtual address:
920 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
921 int nr_wake, int nr_wake2, int op)
923 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
924 struct futex_hash_bucket *hb1, *hb2;
925 struct plist_head *head;
926 struct futex_q *this, *next;
930 ret = get_futex_key(uaddr1, fshared, &key1);
931 if (unlikely(ret != 0))
933 ret = get_futex_key(uaddr2, fshared, &key2);
934 if (unlikely(ret != 0))
937 hb1 = hash_futex(&key1);
938 hb2 = hash_futex(&key2);
941 double_lock_hb(hb1, hb2);
942 op_ret = futex_atomic_op_inuser(op, uaddr2);
943 if (unlikely(op_ret < 0)) {
945 double_unlock_hb(hb1, hb2);
949 * we don't get EFAULT from MMU faults if we don't have an MMU,
950 * but we might get them from range checking
956 if (unlikely(op_ret != -EFAULT)) {
961 ret = fault_in_user_writeable(uaddr2);
968 put_futex_key(fshared, &key2);
969 put_futex_key(fshared, &key1);
975 plist_for_each_entry_safe(this, next, head, list) {
976 if (match_futex (&this->key, &key1)) {
978 if (++ret >= nr_wake)
987 plist_for_each_entry_safe(this, next, head, list) {
988 if (match_futex (&this->key, &key2)) {
990 if (++op_ret >= nr_wake2)
997 double_unlock_hb(hb1, hb2);
999 put_futex_key(fshared, &key2);
1001 put_futex_key(fshared, &key1);
1007 * requeue_futex() - Requeue a futex_q from one hb to another
1008 * @q: the futex_q to requeue
1009 * @hb1: the source hash_bucket
1010 * @hb2: the target hash_bucket
1011 * @key2: the new key for the requeued futex_q
1014 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1015 struct futex_hash_bucket *hb2, union futex_key *key2)
1019 * If key1 and key2 hash to the same bucket, no need to
1022 if (likely(&hb1->chain != &hb2->chain)) {
1023 plist_del(&q->list, &hb1->chain);
1024 plist_add(&q->list, &hb2->chain);
1025 q->lock_ptr = &hb2->lock;
1026 #ifdef CONFIG_DEBUG_PI_LIST
1027 q->list.plist.spinlock = &hb2->lock;
1030 get_futex_key_refs(key2);
1035 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1037 * @key: the key of the requeue target futex
1038 * @hb: the hash_bucket of the requeue target futex
1040 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1041 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1042 * to the requeue target futex so the waiter can detect the wakeup on the right
1043 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1044 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1045 * to protect access to the pi_state to fixup the owner later. Must be called
1046 * with both q->lock_ptr and hb->lock held.
1049 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1050 struct futex_hash_bucket *hb)
1052 get_futex_key_refs(key);
1055 WARN_ON(plist_node_empty(&q->list));
1056 plist_del(&q->list, &q->list.plist);
1058 WARN_ON(!q->rt_waiter);
1059 q->rt_waiter = NULL;
1061 q->lock_ptr = &hb->lock;
1062 #ifdef CONFIG_DEBUG_PI_LIST
1063 q->list.plist.spinlock = &hb->lock;
1066 wake_up_state(q->task, TASK_NORMAL);
1070 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1071 * @pifutex: the user address of the to futex
1072 * @hb1: the from futex hash bucket, must be locked by the caller
1073 * @hb2: the to futex hash bucket, must be locked by the caller
1074 * @key1: the from futex key
1075 * @key2: the to futex key
1076 * @ps: address to store the pi_state pointer
1077 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1079 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1080 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1081 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1082 * hb1 and hb2 must be held by the caller.
1085 * 0 - failed to acquire the lock atomicly
1086 * 1 - acquired the lock
1089 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1090 struct futex_hash_bucket *hb1,
1091 struct futex_hash_bucket *hb2,
1092 union futex_key *key1, union futex_key *key2,
1093 struct futex_pi_state **ps, int set_waiters)
1095 struct futex_q *top_waiter = NULL;
1099 if (get_futex_value_locked(&curval, pifutex))
1103 * Find the top_waiter and determine if there are additional waiters.
1104 * If the caller intends to requeue more than 1 waiter to pifutex,
1105 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1106 * as we have means to handle the possible fault. If not, don't set
1107 * the bit unecessarily as it will force the subsequent unlock to enter
1110 top_waiter = futex_top_waiter(hb1, key1);
1112 /* There are no waiters, nothing for us to do. */
1116 /* Ensure we requeue to the expected futex. */
1117 if (!match_futex(top_waiter->requeue_pi_key, key2))
1121 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1122 * the contended case or if set_waiters is 1. The pi_state is returned
1123 * in ps in contended cases.
1125 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1128 requeue_pi_wake_futex(top_waiter, key2, hb2);
1134 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1135 * @uaddr1: source futex user address
1136 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
1137 * @uaddr2: target futex user address
1138 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1139 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1140 * @cmpval: @uaddr1 expected value (or %NULL)
1141 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1142 * pi futex (pi to pi requeue is not supported)
1144 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1145 * uaddr2 atomically on behalf of the top waiter.
1148 * >=0 - on success, the number of tasks requeued or woken
1151 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1152 int nr_wake, int nr_requeue, u32 *cmpval,
1155 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1156 int drop_count = 0, task_count = 0, ret;
1157 struct futex_pi_state *pi_state = NULL;
1158 struct futex_hash_bucket *hb1, *hb2;
1159 struct plist_head *head1;
1160 struct futex_q *this, *next;
1165 * requeue_pi requires a pi_state, try to allocate it now
1166 * without any locks in case it fails.
1168 if (refill_pi_state_cache())
1171 * requeue_pi must wake as many tasks as it can, up to nr_wake
1172 * + nr_requeue, since it acquires the rt_mutex prior to
1173 * returning to userspace, so as to not leave the rt_mutex with
1174 * waiters and no owner. However, second and third wake-ups
1175 * cannot be predicted as they involve race conditions with the
1176 * first wake and a fault while looking up the pi_state. Both
1177 * pthread_cond_signal() and pthread_cond_broadcast() should
1185 if (pi_state != NULL) {
1187 * We will have to lookup the pi_state again, so free this one
1188 * to keep the accounting correct.
1190 free_pi_state(pi_state);
1194 ret = get_futex_key(uaddr1, fshared, &key1);
1195 if (unlikely(ret != 0))
1197 ret = get_futex_key(uaddr2, fshared, &key2);
1198 if (unlikely(ret != 0))
1201 hb1 = hash_futex(&key1);
1202 hb2 = hash_futex(&key2);
1205 double_lock_hb(hb1, hb2);
1207 if (likely(cmpval != NULL)) {
1210 ret = get_futex_value_locked(&curval, uaddr1);
1212 if (unlikely(ret)) {
1213 double_unlock_hb(hb1, hb2);
1215 ret = get_user(curval, uaddr1);
1222 put_futex_key(fshared, &key2);
1223 put_futex_key(fshared, &key1);
1226 if (curval != *cmpval) {
1232 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1234 * Attempt to acquire uaddr2 and wake the top waiter. If we
1235 * intend to requeue waiters, force setting the FUTEX_WAITERS
1236 * bit. We force this here where we are able to easily handle
1237 * faults rather in the requeue loop below.
1239 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1240 &key2, &pi_state, nr_requeue);
1243 * At this point the top_waiter has either taken uaddr2 or is
1244 * waiting on it. If the former, then the pi_state will not
1245 * exist yet, look it up one more time to ensure we have a
1252 ret = get_futex_value_locked(&curval2, uaddr2);
1254 ret = lookup_pi_state(curval2, hb2, &key2,
1262 double_unlock_hb(hb1, hb2);
1263 put_futex_key(fshared, &key2);
1264 put_futex_key(fshared, &key1);
1265 ret = fault_in_user_writeable(uaddr2);
1270 /* The owner was exiting, try again. */
1271 double_unlock_hb(hb1, hb2);
1272 put_futex_key(fshared, &key2);
1273 put_futex_key(fshared, &key1);
1281 head1 = &hb1->chain;
1282 plist_for_each_entry_safe(this, next, head1, list) {
1283 if (task_count - nr_wake >= nr_requeue)
1286 if (!match_futex(&this->key, &key1))
1290 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1291 * be paired with each other and no other futex ops.
1293 if ((requeue_pi && !this->rt_waiter) ||
1294 (!requeue_pi && this->rt_waiter)) {
1300 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1301 * lock, we already woke the top_waiter. If not, it will be
1302 * woken by futex_unlock_pi().
1304 if (++task_count <= nr_wake && !requeue_pi) {
1309 /* Ensure we requeue to the expected futex for requeue_pi. */
1310 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1316 * Requeue nr_requeue waiters and possibly one more in the case
1317 * of requeue_pi if we couldn't acquire the lock atomically.
1320 /* Prepare the waiter to take the rt_mutex. */
1321 atomic_inc(&pi_state->refcount);
1322 this->pi_state = pi_state;
1323 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1327 /* We got the lock. */
1328 requeue_pi_wake_futex(this, &key2, hb2);
1333 this->pi_state = NULL;
1334 free_pi_state(pi_state);
1338 requeue_futex(this, hb1, hb2, &key2);
1343 double_unlock_hb(hb1, hb2);
1346 * drop_futex_key_refs() must be called outside the spinlocks. During
1347 * the requeue we moved futex_q's from the hash bucket at key1 to the
1348 * one at key2 and updated their key pointer. We no longer need to
1349 * hold the references to key1.
1351 while (--drop_count >= 0)
1352 drop_futex_key_refs(&key1);
1355 put_futex_key(fshared, &key2);
1357 put_futex_key(fshared, &key1);
1359 if (pi_state != NULL)
1360 free_pi_state(pi_state);
1361 return ret ? ret : task_count;
1364 /* The key must be already stored in q->key. */
1365 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1366 __acquires(&hb->lock)
1368 struct futex_hash_bucket *hb;
1370 get_futex_key_refs(&q->key);
1371 hb = hash_futex(&q->key);
1372 q->lock_ptr = &hb->lock;
1374 spin_lock(&hb->lock);
1379 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1380 __releases(&hb->lock)
1382 spin_unlock(&hb->lock);
1383 drop_futex_key_refs(&q->key);
1387 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1388 * @q: The futex_q to enqueue
1389 * @hb: The destination hash bucket
1391 * The hb->lock must be held by the caller, and is released here. A call to
1392 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1393 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1394 * or nothing if the unqueue is done as part of the wake process and the unqueue
1395 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1398 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1399 __releases(&hb->lock)
1404 * The priority used to register this element is
1405 * - either the real thread-priority for the real-time threads
1406 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1407 * - or MAX_RT_PRIO for non-RT threads.
1408 * Thus, all RT-threads are woken first in priority order, and
1409 * the others are woken last, in FIFO order.
1411 prio = min(current->normal_prio, MAX_RT_PRIO);
1413 plist_node_init(&q->list, prio);
1414 #ifdef CONFIG_DEBUG_PI_LIST
1415 q->list.plist.spinlock = &hb->lock;
1417 plist_add(&q->list, &hb->chain);
1419 spin_unlock(&hb->lock);
1423 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1424 * @q: The futex_q to unqueue
1426 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1427 * be paired with exactly one earlier call to queue_me().
1430 * 1 - if the futex_q was still queued (and we removed unqueued it)
1431 * 0 - if the futex_q was already removed by the waking thread
1433 static int unqueue_me(struct futex_q *q)
1435 spinlock_t *lock_ptr;
1438 /* In the common case we don't take the spinlock, which is nice. */
1440 lock_ptr = q->lock_ptr;
1442 if (lock_ptr != NULL) {
1443 spin_lock(lock_ptr);
1445 * q->lock_ptr can change between reading it and
1446 * spin_lock(), causing us to take the wrong lock. This
1447 * corrects the race condition.
1449 * Reasoning goes like this: if we have the wrong lock,
1450 * q->lock_ptr must have changed (maybe several times)
1451 * between reading it and the spin_lock(). It can
1452 * change again after the spin_lock() but only if it was
1453 * already changed before the spin_lock(). It cannot,
1454 * however, change back to the original value. Therefore
1455 * we can detect whether we acquired the correct lock.
1457 if (unlikely(lock_ptr != q->lock_ptr)) {
1458 spin_unlock(lock_ptr);
1461 WARN_ON(plist_node_empty(&q->list));
1462 plist_del(&q->list, &q->list.plist);
1464 BUG_ON(q->pi_state);
1466 spin_unlock(lock_ptr);
1470 drop_futex_key_refs(&q->key);
1475 * PI futexes can not be requeued and must remove themself from the
1476 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1479 static void unqueue_me_pi(struct futex_q *q)
1480 __releases(q->lock_ptr)
1482 WARN_ON(plist_node_empty(&q->list));
1483 plist_del(&q->list, &q->list.plist);
1485 BUG_ON(!q->pi_state);
1486 free_pi_state(q->pi_state);
1489 spin_unlock(q->lock_ptr);
1491 drop_futex_key_refs(&q->key);
1495 * Fixup the pi_state owner with the new owner.
1497 * Must be called with hash bucket lock held and mm->sem held for non
1500 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1501 struct task_struct *newowner, int fshared)
1503 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1504 struct futex_pi_state *pi_state = q->pi_state;
1505 struct task_struct *oldowner = pi_state->owner;
1506 u32 uval, curval, newval;
1510 if (!pi_state->owner)
1511 newtid |= FUTEX_OWNER_DIED;
1514 * We are here either because we stole the rtmutex from the
1515 * pending owner or we are the pending owner which failed to
1516 * get the rtmutex. We have to replace the pending owner TID
1517 * in the user space variable. This must be atomic as we have
1518 * to preserve the owner died bit here.
1520 * Note: We write the user space value _before_ changing the pi_state
1521 * because we can fault here. Imagine swapped out pages or a fork
1522 * that marked all the anonymous memory readonly for cow.
1524 * Modifying pi_state _before_ the user space value would
1525 * leave the pi_state in an inconsistent state when we fault
1526 * here, because we need to drop the hash bucket lock to
1527 * handle the fault. This might be observed in the PID check
1528 * in lookup_pi_state.
1531 if (get_futex_value_locked(&uval, uaddr))
1535 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1537 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1539 if (curval == -EFAULT)
1547 * We fixed up user space. Now we need to fix the pi_state
1550 if (pi_state->owner != NULL) {
1551 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1552 WARN_ON(list_empty(&pi_state->list));
1553 list_del_init(&pi_state->list);
1554 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1557 pi_state->owner = newowner;
1559 raw_spin_lock_irq(&newowner->pi_lock);
1560 WARN_ON(!list_empty(&pi_state->list));
1561 list_add(&pi_state->list, &newowner->pi_state_list);
1562 raw_spin_unlock_irq(&newowner->pi_lock);
1566 * To handle the page fault we need to drop the hash bucket
1567 * lock here. That gives the other task (either the pending
1568 * owner itself or the task which stole the rtmutex) the
1569 * chance to try the fixup of the pi_state. So once we are
1570 * back from handling the fault we need to check the pi_state
1571 * after reacquiring the hash bucket lock and before trying to
1572 * do another fixup. When the fixup has been done already we
1576 spin_unlock(q->lock_ptr);
1578 ret = fault_in_user_writeable(uaddr);
1580 spin_lock(q->lock_ptr);
1583 * Check if someone else fixed it for us:
1585 if (pi_state->owner != oldowner)
1595 * In case we must use restart_block to restart a futex_wait,
1596 * we encode in the 'flags' shared capability
1598 #define FLAGS_SHARED 0x01
1599 #define FLAGS_CLOCKRT 0x02
1600 #define FLAGS_HAS_TIMEOUT 0x04
1602 static long futex_wait_restart(struct restart_block *restart);
1605 * fixup_owner() - Post lock pi_state and corner case management
1606 * @uaddr: user address of the futex
1607 * @fshared: whether the futex is shared (1) or not (0)
1608 * @q: futex_q (contains pi_state and access to the rt_mutex)
1609 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1611 * After attempting to lock an rt_mutex, this function is called to cleanup
1612 * the pi_state owner as well as handle race conditions that may allow us to
1613 * acquire the lock. Must be called with the hb lock held.
1616 * 1 - success, lock taken
1617 * 0 - success, lock not taken
1618 * <0 - on error (-EFAULT)
1620 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1623 struct task_struct *owner;
1628 * Got the lock. We might not be the anticipated owner if we
1629 * did a lock-steal - fix up the PI-state in that case:
1631 if (q->pi_state->owner != current)
1632 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1637 * Catch the rare case, where the lock was released when we were on the
1638 * way back before we locked the hash bucket.
1640 if (q->pi_state->owner == current) {
1642 * Try to get the rt_mutex now. This might fail as some other
1643 * task acquired the rt_mutex after we removed ourself from the
1644 * rt_mutex waiters list.
1646 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1652 * pi_state is incorrect, some other task did a lock steal and
1653 * we returned due to timeout or signal without taking the
1654 * rt_mutex. Too late. We can access the rt_mutex_owner without
1655 * locking, as the other task is now blocked on the hash bucket
1656 * lock. Fix the state up.
1658 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1659 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1664 * Paranoia check. If we did not take the lock, then we should not be
1665 * the owner, nor the pending owner, of the rt_mutex.
1667 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1668 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1669 "pi-state %p\n", ret,
1670 q->pi_state->pi_mutex.owner,
1671 q->pi_state->owner);
1674 return ret ? ret : locked;
1678 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1679 * @hb: the futex hash bucket, must be locked by the caller
1680 * @q: the futex_q to queue up on
1681 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1683 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1684 struct hrtimer_sleeper *timeout)
1687 * The task state is guaranteed to be set before another task can
1688 * wake it. set_current_state() is implemented using set_mb() and
1689 * queue_me() calls spin_unlock() upon completion, both serializing
1690 * access to the hash list and forcing another memory barrier.
1692 set_current_state(TASK_INTERRUPTIBLE);
1697 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1698 if (!hrtimer_active(&timeout->timer))
1699 timeout->task = NULL;
1703 * If we have been removed from the hash list, then another task
1704 * has tried to wake us, and we can skip the call to schedule().
1706 if (likely(!plist_node_empty(&q->list))) {
1708 * If the timer has already expired, current will already be
1709 * flagged for rescheduling. Only call schedule if there
1710 * is no timeout, or if it has yet to expire.
1712 if (!timeout || timeout->task)
1715 __set_current_state(TASK_RUNNING);
1719 * futex_wait_setup() - Prepare to wait on a futex
1720 * @uaddr: the futex userspace address
1721 * @val: the expected value
1722 * @fshared: whether the futex is shared (1) or not (0)
1723 * @q: the associated futex_q
1724 * @hb: storage for hash_bucket pointer to be returned to caller
1726 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1727 * compare it with the expected value. Handle atomic faults internally.
1728 * Return with the hb lock held and a q.key reference on success, and unlocked
1729 * with no q.key reference on failure.
1732 * 0 - uaddr contains val and hb has been locked
1733 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1735 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1736 struct futex_q *q, struct futex_hash_bucket **hb)
1742 * Access the page AFTER the hash-bucket is locked.
1743 * Order is important:
1745 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1746 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1748 * The basic logical guarantee of a futex is that it blocks ONLY
1749 * if cond(var) is known to be true at the time of blocking, for
1750 * any cond. If we queued after testing *uaddr, that would open
1751 * a race condition where we could block indefinitely with
1752 * cond(var) false, which would violate the guarantee.
1754 * A consequence is that futex_wait() can return zero and absorb
1755 * a wakeup when *uaddr != val on entry to the syscall. This is
1759 q->key = FUTEX_KEY_INIT;
1760 ret = get_futex_key(uaddr, fshared, &q->key);
1761 if (unlikely(ret != 0))
1765 *hb = queue_lock(q);
1767 ret = get_futex_value_locked(&uval, uaddr);
1770 queue_unlock(q, *hb);
1772 ret = get_user(uval, uaddr);
1779 put_futex_key(fshared, &q->key);
1784 queue_unlock(q, *hb);
1790 put_futex_key(fshared, &q->key);
1794 static int futex_wait(u32 __user *uaddr, int fshared,
1795 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1797 struct hrtimer_sleeper timeout, *to = NULL;
1798 struct restart_block *restart;
1799 struct futex_hash_bucket *hb;
1809 q.requeue_pi_key = NULL;
1814 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1815 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1816 hrtimer_init_sleeper(to, current);
1817 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1818 current->timer_slack_ns);
1822 /* Prepare to wait on uaddr. */
1823 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1827 /* queue_me and wait for wakeup, timeout, or a signal. */
1828 futex_wait_queue_me(hb, &q, to);
1830 /* If we were woken (and unqueued), we succeeded, whatever. */
1832 if (!unqueue_me(&q))
1835 if (to && !to->task)
1839 * We expect signal_pending(current), but we might be the
1840 * victim of a spurious wakeup as well.
1842 if (!signal_pending(current)) {
1843 put_futex_key(fshared, &q.key);
1851 restart = ¤t_thread_info()->restart_block;
1852 restart->fn = futex_wait_restart;
1853 restart->futex.uaddr = uaddr;
1854 restart->futex.val = val;
1855 restart->futex.time = abs_time->tv64;
1856 restart->futex.bitset = bitset;
1857 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1860 restart->futex.flags |= FLAGS_SHARED;
1862 restart->futex.flags |= FLAGS_CLOCKRT;
1864 ret = -ERESTART_RESTARTBLOCK;
1867 put_futex_key(fshared, &q.key);
1870 hrtimer_cancel(&to->timer);
1871 destroy_hrtimer_on_stack(&to->timer);
1877 static long futex_wait_restart(struct restart_block *restart)
1879 u32 __user *uaddr = restart->futex.uaddr;
1881 ktime_t t, *tp = NULL;
1883 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1884 t.tv64 = restart->futex.time;
1887 restart->fn = do_no_restart_syscall;
1888 if (restart->futex.flags & FLAGS_SHARED)
1890 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1891 restart->futex.bitset,
1892 restart->futex.flags & FLAGS_CLOCKRT);
1897 * Userspace tried a 0 -> TID atomic transition of the futex value
1898 * and failed. The kernel side here does the whole locking operation:
1899 * if there are waiters then it will block, it does PI, etc. (Due to
1900 * races the kernel might see a 0 value of the futex too.)
1902 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1903 int detect, ktime_t *time, int trylock)
1905 struct hrtimer_sleeper timeout, *to = NULL;
1906 struct futex_hash_bucket *hb;
1910 if (refill_pi_state_cache())
1915 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1917 hrtimer_init_sleeper(to, current);
1918 hrtimer_set_expires(&to->timer, *time);
1923 q.requeue_pi_key = NULL;
1925 q.key = FUTEX_KEY_INIT;
1926 ret = get_futex_key(uaddr, fshared, &q.key);
1927 if (unlikely(ret != 0))
1931 hb = queue_lock(&q);
1933 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1934 if (unlikely(ret)) {
1937 /* We got the lock. */
1939 goto out_unlock_put_key;
1944 * Task is exiting and we just wait for the
1947 queue_unlock(&q, hb);
1948 put_futex_key(fshared, &q.key);
1952 goto out_unlock_put_key;
1957 * Only actually queue now that the atomic ops are done:
1961 WARN_ON(!q.pi_state);
1963 * Block on the PI mutex:
1966 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1968 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1969 /* Fixup the trylock return value: */
1970 ret = ret ? 0 : -EWOULDBLOCK;
1973 spin_lock(q.lock_ptr);
1975 * Fixup the pi_state owner and possibly acquire the lock if we
1978 res = fixup_owner(uaddr, fshared, &q, !ret);
1980 * If fixup_owner() returned an error, proprogate that. If it acquired
1981 * the lock, clear our -ETIMEDOUT or -EINTR.
1984 ret = (res < 0) ? res : 0;
1987 * If fixup_owner() faulted and was unable to handle the fault, unlock
1988 * it and return the fault to userspace.
1990 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1991 rt_mutex_unlock(&q.pi_state->pi_mutex);
1993 /* Unqueue and drop the lock */
1999 queue_unlock(&q, hb);
2002 put_futex_key(fshared, &q.key);
2005 destroy_hrtimer_on_stack(&to->timer);
2006 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2009 queue_unlock(&q, hb);
2011 ret = fault_in_user_writeable(uaddr);
2018 put_futex_key(fshared, &q.key);
2023 * Userspace attempted a TID -> 0 atomic transition, and failed.
2024 * This is the in-kernel slowpath: we look up the PI state (if any),
2025 * and do the rt-mutex unlock.
2027 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2029 struct futex_hash_bucket *hb;
2030 struct futex_q *this, *next;
2032 struct plist_head *head;
2033 union futex_key key = FUTEX_KEY_INIT;
2037 if (get_user(uval, uaddr))
2040 * We release only a lock we actually own:
2042 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2045 ret = get_futex_key(uaddr, fshared, &key);
2046 if (unlikely(ret != 0))
2049 hb = hash_futex(&key);
2050 spin_lock(&hb->lock);
2053 * To avoid races, try to do the TID -> 0 atomic transition
2054 * again. If it succeeds then we can return without waking
2057 if (!(uval & FUTEX_OWNER_DIED))
2058 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2061 if (unlikely(uval == -EFAULT))
2064 * Rare case: we managed to release the lock atomically,
2065 * no need to wake anyone else up:
2067 if (unlikely(uval == task_pid_vnr(current)))
2071 * Ok, other tasks may need to be woken up - check waiters
2072 * and do the wakeup if necessary:
2076 plist_for_each_entry_safe(this, next, head, list) {
2077 if (!match_futex (&this->key, &key))
2079 ret = wake_futex_pi(uaddr, uval, this);
2081 * The atomic access to the futex value
2082 * generated a pagefault, so retry the
2083 * user-access and the wakeup:
2090 * No waiters - kernel unlocks the futex:
2092 if (!(uval & FUTEX_OWNER_DIED)) {
2093 ret = unlock_futex_pi(uaddr, uval);
2099 spin_unlock(&hb->lock);
2100 put_futex_key(fshared, &key);
2106 spin_unlock(&hb->lock);
2107 put_futex_key(fshared, &key);
2109 ret = fault_in_user_writeable(uaddr);
2117 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2118 * @hb: the hash_bucket futex_q was original enqueued on
2119 * @q: the futex_q woken while waiting to be requeued
2120 * @key2: the futex_key of the requeue target futex
2121 * @timeout: the timeout associated with the wait (NULL if none)
2123 * Detect if the task was woken on the initial futex as opposed to the requeue
2124 * target futex. If so, determine if it was a timeout or a signal that caused
2125 * the wakeup and return the appropriate error code to the caller. Must be
2126 * called with the hb lock held.
2129 * 0 - no early wakeup detected
2130 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2133 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2134 struct futex_q *q, union futex_key *key2,
2135 struct hrtimer_sleeper *timeout)
2140 * With the hb lock held, we avoid races while we process the wakeup.
2141 * We only need to hold hb (and not hb2) to ensure atomicity as the
2142 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2143 * It can't be requeued from uaddr2 to something else since we don't
2144 * support a PI aware source futex for requeue.
2146 if (!match_futex(&q->key, key2)) {
2147 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2149 * We were woken prior to requeue by a timeout or a signal.
2150 * Unqueue the futex_q and determine which it was.
2152 plist_del(&q->list, &q->list.plist);
2154 /* Handle spurious wakeups gracefully */
2156 if (timeout && !timeout->task)
2158 else if (signal_pending(current))
2159 ret = -ERESTARTNOINTR;
2165 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2166 * @uaddr: the futex we initially wait on (non-pi)
2167 * @fshared: whether the futexes are shared (1) or not (0). They must be
2168 * the same type, no requeueing from private to shared, etc.
2169 * @val: the expected value of uaddr
2170 * @abs_time: absolute timeout
2171 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2172 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2173 * @uaddr2: the pi futex we will take prior to returning to user-space
2175 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2176 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2177 * complete the acquisition of the rt_mutex prior to returning to userspace.
2178 * This ensures the rt_mutex maintains an owner when it has waiters; without
2179 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2182 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2183 * via the following:
2184 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2185 * 2) wakeup on uaddr2 after a requeue
2189 * If 3, cleanup and return -ERESTARTNOINTR.
2191 * If 2, we may then block on trying to take the rt_mutex and return via:
2192 * 5) successful lock
2195 * 8) other lock acquisition failure
2197 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2199 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2205 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2206 u32 val, ktime_t *abs_time, u32 bitset,
2207 int clockrt, u32 __user *uaddr2)
2209 struct hrtimer_sleeper timeout, *to = NULL;
2210 struct rt_mutex_waiter rt_waiter;
2211 struct rt_mutex *pi_mutex = NULL;
2212 struct futex_hash_bucket *hb;
2213 union futex_key key2;
2222 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2223 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2224 hrtimer_init_sleeper(to, current);
2225 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2226 current->timer_slack_ns);
2230 * The waiter is allocated on our stack, manipulated by the requeue
2231 * code while we sleep on uaddr.
2233 debug_rt_mutex_init_waiter(&rt_waiter);
2234 rt_waiter.task = NULL;
2236 key2 = FUTEX_KEY_INIT;
2237 ret = get_futex_key(uaddr2, fshared, &key2);
2238 if (unlikely(ret != 0))
2243 q.rt_waiter = &rt_waiter;
2244 q.requeue_pi_key = &key2;
2246 /* Prepare to wait on uaddr. */
2247 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2251 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2252 futex_wait_queue_me(hb, &q, to);
2254 spin_lock(&hb->lock);
2255 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2256 spin_unlock(&hb->lock);
2261 * In order for us to be here, we know our q.key == key2, and since
2262 * we took the hb->lock above, we also know that futex_requeue() has
2263 * completed and we no longer have to concern ourselves with a wakeup
2264 * race with the atomic proxy lock acquition by the requeue code.
2267 /* Check if the requeue code acquired the second futex for us. */
2270 * Got the lock. We might not be the anticipated owner if we
2271 * did a lock-steal - fix up the PI-state in that case.
2273 if (q.pi_state && (q.pi_state->owner != current)) {
2274 spin_lock(q.lock_ptr);
2275 ret = fixup_pi_state_owner(uaddr2, &q, current,
2277 spin_unlock(q.lock_ptr);
2281 * We have been woken up by futex_unlock_pi(), a timeout, or a
2282 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2285 WARN_ON(!&q.pi_state);
2286 pi_mutex = &q.pi_state->pi_mutex;
2287 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2288 debug_rt_mutex_free_waiter(&rt_waiter);
2290 spin_lock(q.lock_ptr);
2292 * Fixup the pi_state owner and possibly acquire the lock if we
2295 res = fixup_owner(uaddr2, fshared, &q, !ret);
2297 * If fixup_owner() returned an error, proprogate that. If it
2298 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2301 ret = (res < 0) ? res : 0;
2303 /* Unqueue and drop the lock. */
2308 * If fixup_pi_state_owner() faulted and was unable to handle the
2309 * fault, unlock the rt_mutex and return the fault to userspace.
2311 if (ret == -EFAULT) {
2312 if (rt_mutex_owner(pi_mutex) == current)
2313 rt_mutex_unlock(pi_mutex);
2314 } else if (ret == -EINTR) {
2316 * We've already been requeued, but cannot restart by calling
2317 * futex_lock_pi() directly. We could restart this syscall, but
2318 * it would detect that the user space "val" changed and return
2319 * -EWOULDBLOCK. Save the overhead of the restart and return
2320 * -EWOULDBLOCK directly.
2326 put_futex_key(fshared, &q.key);
2328 put_futex_key(fshared, &key2);
2332 hrtimer_cancel(&to->timer);
2333 destroy_hrtimer_on_stack(&to->timer);
2339 * Support for robust futexes: the kernel cleans up held futexes at
2342 * Implementation: user-space maintains a per-thread list of locks it
2343 * is holding. Upon do_exit(), the kernel carefully walks this list,
2344 * and marks all locks that are owned by this thread with the
2345 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2346 * always manipulated with the lock held, so the list is private and
2347 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2348 * field, to allow the kernel to clean up if the thread dies after
2349 * acquiring the lock, but just before it could have added itself to
2350 * the list. There can only be one such pending lock.
2354 * sys_set_robust_list() - Set the robust-futex list head of a task
2355 * @head: pointer to the list-head
2356 * @len: length of the list-head, as userspace expects
2358 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2361 if (!futex_cmpxchg_enabled)
2364 * The kernel knows only one size for now:
2366 if (unlikely(len != sizeof(*head)))
2369 current->robust_list = head;
2375 * sys_get_robust_list() - Get the robust-futex list head of a task
2376 * @pid: pid of the process [zero for current task]
2377 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2378 * @len_ptr: pointer to a length field, the kernel fills in the header size
2380 SYSCALL_DEFINE3(get_robust_list, int, pid,
2381 struct robust_list_head __user * __user *, head_ptr,
2382 size_t __user *, len_ptr)
2384 struct robust_list_head __user *head;
2386 const struct cred *cred = current_cred(), *pcred;
2388 if (!futex_cmpxchg_enabled)
2392 head = current->robust_list;
2394 struct task_struct *p;
2398 p = find_task_by_vpid(pid);
2402 pcred = __task_cred(p);
2403 if (cred->euid != pcred->euid &&
2404 cred->euid != pcred->uid &&
2405 !capable(CAP_SYS_PTRACE))
2407 head = p->robust_list;
2411 if (put_user(sizeof(*head), len_ptr))
2413 return put_user(head, head_ptr);
2422 * Process a futex-list entry, check whether it's owned by the
2423 * dying task, and do notification if so:
2425 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2427 u32 uval, nval, mval;
2430 if (get_user(uval, uaddr))
2433 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2435 * Ok, this dying thread is truly holding a futex
2436 * of interest. Set the OWNER_DIED bit atomically
2437 * via cmpxchg, and if the value had FUTEX_WAITERS
2438 * set, wake up a waiter (if any). (We have to do a
2439 * futex_wake() even if OWNER_DIED is already set -
2440 * to handle the rare but possible case of recursive
2441 * thread-death.) The rest of the cleanup is done in
2444 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2445 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2447 if (nval == -EFAULT)
2454 * Wake robust non-PI futexes here. The wakeup of
2455 * PI futexes happens in exit_pi_state():
2457 if (!pi && (uval & FUTEX_WAITERS))
2458 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2464 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2466 static inline int fetch_robust_entry(struct robust_list __user **entry,
2467 struct robust_list __user * __user *head,
2470 unsigned long uentry;
2472 if (get_user(uentry, (unsigned long __user *)head))
2475 *entry = (void __user *)(uentry & ~1UL);
2482 * Walk curr->robust_list (very carefully, it's a userspace list!)
2483 * and mark any locks found there dead, and notify any waiters.
2485 * We silently return on any sign of list-walking problem.
2487 void exit_robust_list(struct task_struct *curr)
2489 struct robust_list_head __user *head = curr->robust_list;
2490 struct robust_list __user *entry, *next_entry, *pending;
2491 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2492 unsigned long futex_offset;
2495 if (!futex_cmpxchg_enabled)
2499 * Fetch the list head (which was registered earlier, via
2500 * sys_set_robust_list()):
2502 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2505 * Fetch the relative futex offset:
2507 if (get_user(futex_offset, &head->futex_offset))
2510 * Fetch any possibly pending lock-add first, and handle it
2513 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2516 next_entry = NULL; /* avoid warning with gcc */
2517 while (entry != &head->list) {
2519 * Fetch the next entry in the list before calling
2520 * handle_futex_death:
2522 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2524 * A pending lock might already be on the list, so
2525 * don't process it twice:
2527 if (entry != pending)
2528 if (handle_futex_death((void __user *)entry + futex_offset,
2536 * Avoid excessively long or circular lists:
2545 handle_futex_death((void __user *)pending + futex_offset,
2549 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2550 u32 __user *uaddr2, u32 val2, u32 val3)
2552 int clockrt, ret = -ENOSYS;
2553 int cmd = op & FUTEX_CMD_MASK;
2556 if (!(op & FUTEX_PRIVATE_FLAG))
2559 clockrt = op & FUTEX_CLOCK_REALTIME;
2560 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2565 val3 = FUTEX_BITSET_MATCH_ANY;
2566 case FUTEX_WAIT_BITSET:
2567 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2570 val3 = FUTEX_BITSET_MATCH_ANY;
2571 case FUTEX_WAKE_BITSET:
2572 ret = futex_wake(uaddr, fshared, val, val3);
2575 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2577 case FUTEX_CMP_REQUEUE:
2578 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2582 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2585 if (futex_cmpxchg_enabled)
2586 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2588 case FUTEX_UNLOCK_PI:
2589 if (futex_cmpxchg_enabled)
2590 ret = futex_unlock_pi(uaddr, fshared);
2592 case FUTEX_TRYLOCK_PI:
2593 if (futex_cmpxchg_enabled)
2594 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2596 case FUTEX_WAIT_REQUEUE_PI:
2597 val3 = FUTEX_BITSET_MATCH_ANY;
2598 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2601 case FUTEX_CMP_REQUEUE_PI:
2602 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2612 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2613 struct timespec __user *, utime, u32 __user *, uaddr2,
2617 ktime_t t, *tp = NULL;
2619 int cmd = op & FUTEX_CMD_MASK;
2621 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2622 cmd == FUTEX_WAIT_BITSET ||
2623 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2624 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2626 if (!timespec_valid(&ts))
2629 t = timespec_to_ktime(ts);
2630 if (cmd == FUTEX_WAIT)
2631 t = ktime_add_safe(ktime_get(), t);
2635 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2636 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2638 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2639 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2640 val2 = (u32) (unsigned long) utime;
2642 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2645 static int __init futex_init(void)
2651 * This will fail and we want it. Some arch implementations do
2652 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2653 * functionality. We want to know that before we call in any
2654 * of the complex code paths. Also we want to prevent
2655 * registration of robust lists in that case. NULL is
2656 * guaranteed to fault and we get -EFAULT on functional
2657 * implementation, the non-functional ones will return
2660 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2661 if (curval == -EFAULT)
2662 futex_cmpxchg_enabled = 1;
2664 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2665 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2666 spin_lock_init(&futex_queues[i].lock);
2671 __initcall(futex_init);