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;
223 struct vm_area_struct *vma;
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 * The futex is hashed differently depending on whether
251 * it's in a shared or private mapping. So check vma first.
253 vma = find_extend_vma(mm, address);
260 if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
261 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
264 * Private mappings are handled in a simple way.
266 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
267 * it's a read-only handle, it's expected that futexes attach to
268 * the object not the particular process. Therefore we use
269 * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
270 * mappings of _writable_ handles.
272 if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
273 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
274 key->private.mm = mm;
275 key->private.address = address;
276 get_futex_key_refs(key);
281 err = get_user_pages_fast(address, 1, 1, &page);
285 page = compound_head(page);
287 if (!page->mapping) {
294 * Private mappings are handled in a simple way.
296 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
297 * it's a read-only handle, it's expected that futexes attach to
298 * the object not the particular process.
300 if (PageAnon(page)) {
301 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
302 key->private.mm = mm;
303 key->private.address = address;
305 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
306 key->shared.inode = page->mapping->host;
307 key->shared.pgoff = page->index;
310 get_futex_key_refs(key);
318 void put_futex_key(int fshared, union futex_key *key)
320 drop_futex_key_refs(key);
324 * fault_in_user_writeable() - Fault in user address and verify RW access
325 * @uaddr: pointer to faulting user space address
327 * Slow path to fixup the fault we just took in the atomic write
330 * We have no generic implementation of a non destructive write to the
331 * user address. We know that we faulted in the atomic pagefault
332 * disabled section so we can as well avoid the #PF overhead by
333 * calling get_user_pages() right away.
335 static int fault_in_user_writeable(u32 __user *uaddr)
337 struct mm_struct *mm = current->mm;
340 down_read(&mm->mmap_sem);
341 ret = get_user_pages(current, mm, (unsigned long)uaddr,
342 1, 1, 0, NULL, NULL);
343 up_read(&mm->mmap_sem);
345 return ret < 0 ? ret : 0;
349 * futex_top_waiter() - Return the highest priority waiter on a futex
350 * @hb: the hash bucket the futex_q's reside in
351 * @key: the futex key (to distinguish it from other futex futex_q's)
353 * Must be called with the hb lock held.
355 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
356 union futex_key *key)
358 struct futex_q *this;
360 plist_for_each_entry(this, &hb->chain, list) {
361 if (match_futex(&this->key, key))
367 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
372 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
378 static int get_futex_value_locked(u32 *dest, u32 __user *from)
383 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
386 return ret ? -EFAULT : 0;
393 static int refill_pi_state_cache(void)
395 struct futex_pi_state *pi_state;
397 if (likely(current->pi_state_cache))
400 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
405 INIT_LIST_HEAD(&pi_state->list);
406 /* pi_mutex gets initialized later */
407 pi_state->owner = NULL;
408 atomic_set(&pi_state->refcount, 1);
409 pi_state->key = FUTEX_KEY_INIT;
411 current->pi_state_cache = pi_state;
416 static struct futex_pi_state * alloc_pi_state(void)
418 struct futex_pi_state *pi_state = current->pi_state_cache;
421 current->pi_state_cache = NULL;
426 static void free_pi_state(struct futex_pi_state *pi_state)
428 if (!atomic_dec_and_test(&pi_state->refcount))
432 * If pi_state->owner is NULL, the owner is most probably dying
433 * and has cleaned up the pi_state already
435 if (pi_state->owner) {
436 spin_lock_irq(&pi_state->owner->pi_lock);
437 list_del_init(&pi_state->list);
438 spin_unlock_irq(&pi_state->owner->pi_lock);
440 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
443 if (current->pi_state_cache)
447 * pi_state->list is already empty.
448 * clear pi_state->owner.
449 * refcount is at 0 - put it back to 1.
451 pi_state->owner = NULL;
452 atomic_set(&pi_state->refcount, 1);
453 current->pi_state_cache = pi_state;
458 * Look up the task based on what TID userspace gave us.
461 static struct task_struct * futex_find_get_task(pid_t pid)
463 struct task_struct *p;
466 p = find_task_by_vpid(pid);
476 * This task is holding PI mutexes at exit time => bad.
477 * Kernel cleans up PI-state, but userspace is likely hosed.
478 * (Robust-futex cleanup is separate and might save the day for userspace.)
480 void exit_pi_state_list(struct task_struct *curr)
482 struct list_head *next, *head = &curr->pi_state_list;
483 struct futex_pi_state *pi_state;
484 struct futex_hash_bucket *hb;
485 union futex_key key = FUTEX_KEY_INIT;
487 if (!futex_cmpxchg_enabled)
490 * We are a ZOMBIE and nobody can enqueue itself on
491 * pi_state_list anymore, but we have to be careful
492 * versus waiters unqueueing themselves:
494 spin_lock_irq(&curr->pi_lock);
495 while (!list_empty(head)) {
498 pi_state = list_entry(next, struct futex_pi_state, list);
500 hb = hash_futex(&key);
501 spin_unlock_irq(&curr->pi_lock);
503 spin_lock(&hb->lock);
505 spin_lock_irq(&curr->pi_lock);
507 * We dropped the pi-lock, so re-check whether this
508 * task still owns the PI-state:
510 if (head->next != next) {
511 spin_unlock(&hb->lock);
515 WARN_ON(pi_state->owner != curr);
516 WARN_ON(list_empty(&pi_state->list));
517 list_del_init(&pi_state->list);
518 pi_state->owner = NULL;
519 spin_unlock_irq(&curr->pi_lock);
521 rt_mutex_unlock(&pi_state->pi_mutex);
523 spin_unlock(&hb->lock);
525 spin_lock_irq(&curr->pi_lock);
527 spin_unlock_irq(&curr->pi_lock);
531 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
532 union futex_key *key, struct futex_pi_state **ps)
534 struct futex_pi_state *pi_state = NULL;
535 struct futex_q *this, *next;
536 struct plist_head *head;
537 struct task_struct *p;
538 pid_t pid = uval & FUTEX_TID_MASK;
542 plist_for_each_entry_safe(this, next, head, list) {
543 if (match_futex(&this->key, key)) {
545 * Another waiter already exists - bump up
546 * the refcount and return its pi_state:
548 pi_state = this->pi_state;
550 * Userspace might have messed up non PI and PI futexes
552 if (unlikely(!pi_state))
555 WARN_ON(!atomic_read(&pi_state->refcount));
558 * When pi_state->owner is NULL then the owner died
559 * and another waiter is on the fly. pi_state->owner
560 * is fixed up by the task which acquires
561 * pi_state->rt_mutex.
563 * We do not check for pid == 0 which can happen when
564 * the owner died and robust_list_exit() cleared the
567 if (pid && pi_state->owner) {
569 * Bail out if user space manipulated the
572 if (pid != task_pid_vnr(pi_state->owner))
576 atomic_inc(&pi_state->refcount);
584 * We are the first waiter - try to look up the real owner and attach
585 * the new pi_state to it, but bail out when TID = 0
589 p = futex_find_get_task(pid);
594 * We need to look at the task state flags to figure out,
595 * whether the task is exiting. To protect against the do_exit
596 * change of the task flags, we do this protected by
599 spin_lock_irq(&p->pi_lock);
600 if (unlikely(p->flags & PF_EXITING)) {
602 * The task is on the way out. When PF_EXITPIDONE is
603 * set, we know that the task has finished the
606 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
608 spin_unlock_irq(&p->pi_lock);
613 pi_state = alloc_pi_state();
616 * Initialize the pi_mutex in locked state and make 'p'
619 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
621 /* Store the key for possible exit cleanups: */
622 pi_state->key = *key;
624 WARN_ON(!list_empty(&pi_state->list));
625 list_add(&pi_state->list, &p->pi_state_list);
627 spin_unlock_irq(&p->pi_lock);
637 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
638 * @uaddr: the pi futex user address
639 * @hb: the pi futex hash bucket
640 * @key: the futex key associated with uaddr and hb
641 * @ps: the pi_state pointer where we store the result of the
643 * @task: the task to perform the atomic lock work for. This will
644 * be "current" except in the case of requeue pi.
645 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
649 * 1 - acquired the lock
652 * The hb->lock and futex_key refs shall be held by the caller.
654 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
655 union futex_key *key,
656 struct futex_pi_state **ps,
657 struct task_struct *task, int set_waiters)
659 int lock_taken, ret, ownerdied = 0;
660 u32 uval, newval, curval;
663 ret = lock_taken = 0;
666 * To avoid races, we attempt to take the lock here again
667 * (by doing a 0 -> TID atomic cmpxchg), while holding all
668 * the locks. It will most likely not succeed.
670 newval = task_pid_vnr(task);
672 newval |= FUTEX_WAITERS;
674 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
676 if (unlikely(curval == -EFAULT))
682 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
686 * Surprise - we got the lock. Just return to userspace:
688 if (unlikely(!curval))
694 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
695 * to wake at the next unlock.
697 newval = curval | FUTEX_WAITERS;
700 * There are two cases, where a futex might have no owner (the
701 * owner TID is 0): OWNER_DIED. We take over the futex in this
702 * case. We also do an unconditional take over, when the owner
705 * This is safe as we are protected by the hash bucket lock !
707 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
708 /* Keep the OWNER_DIED bit */
709 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
714 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
716 if (unlikely(curval == -EFAULT))
718 if (unlikely(curval != uval))
722 * We took the lock due to owner died take over.
724 if (unlikely(lock_taken))
728 * We dont have the lock. Look up the PI state (or create it if
729 * we are the first waiter):
731 ret = lookup_pi_state(uval, hb, key, ps);
737 * No owner found for this futex. Check if the
738 * OWNER_DIED bit is set to figure out whether
739 * this is a robust futex or not.
741 if (get_futex_value_locked(&curval, uaddr))
745 * We simply start over in case of a robust
746 * futex. The code above will take the futex
749 if (curval & FUTEX_OWNER_DIED) {
762 * The hash bucket lock must be held when this is called.
763 * Afterwards, the futex_q must not be accessed.
765 static void wake_futex(struct futex_q *q)
767 struct task_struct *p = q->task;
770 * We set q->lock_ptr = NULL _before_ we wake up the task. If
771 * a non futex wake up happens on another CPU then the task
772 * might exit and p would dereference a non existing task
773 * struct. Prevent this by holding a reference on p across the
778 plist_del(&q->list, &q->list.plist);
780 * The waiting task can free the futex_q as soon as
781 * q->lock_ptr = NULL is written, without taking any locks. A
782 * memory barrier is required here to prevent the following
783 * store to lock_ptr from getting ahead of the plist_del.
788 wake_up_state(p, TASK_NORMAL);
792 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
794 struct task_struct *new_owner;
795 struct futex_pi_state *pi_state = this->pi_state;
802 * If current does not own the pi_state then the futex is
803 * inconsistent and user space fiddled with the futex value.
805 if (pi_state->owner != current)
808 spin_lock(&pi_state->pi_mutex.wait_lock);
809 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
812 * This happens when we have stolen the lock and the original
813 * pending owner did not enqueue itself back on the rt_mutex.
814 * Thats not a tragedy. We know that way, that a lock waiter
815 * is on the fly. We make the futex_q waiter the pending owner.
818 new_owner = this->task;
821 * We pass it to the next owner. (The WAITERS bit is always
822 * kept enabled while there is PI state around. We must also
823 * preserve the owner died bit.)
825 if (!(uval & FUTEX_OWNER_DIED)) {
828 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
830 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
832 if (curval == -EFAULT)
834 else if (curval != uval)
837 spin_unlock(&pi_state->pi_mutex.wait_lock);
842 spin_lock_irq(&pi_state->owner->pi_lock);
843 WARN_ON(list_empty(&pi_state->list));
844 list_del_init(&pi_state->list);
845 spin_unlock_irq(&pi_state->owner->pi_lock);
847 spin_lock_irq(&new_owner->pi_lock);
848 WARN_ON(!list_empty(&pi_state->list));
849 list_add(&pi_state->list, &new_owner->pi_state_list);
850 pi_state->owner = new_owner;
851 spin_unlock_irq(&new_owner->pi_lock);
853 spin_unlock(&pi_state->pi_mutex.wait_lock);
854 rt_mutex_unlock(&pi_state->pi_mutex);
859 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
864 * There is no waiter, so we unlock the futex. The owner died
865 * bit has not to be preserved here. We are the owner:
867 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
869 if (oldval == -EFAULT)
878 * Express the locking dependencies for lockdep:
881 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
884 spin_lock(&hb1->lock);
886 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
887 } else { /* hb1 > hb2 */
888 spin_lock(&hb2->lock);
889 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
894 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
896 spin_unlock(&hb1->lock);
898 spin_unlock(&hb2->lock);
902 * Wake up waiters matching bitset queued on this futex (uaddr).
904 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
906 struct futex_hash_bucket *hb;
907 struct futex_q *this, *next;
908 struct plist_head *head;
909 union futex_key key = FUTEX_KEY_INIT;
915 ret = get_futex_key(uaddr, fshared, &key);
916 if (unlikely(ret != 0))
919 hb = hash_futex(&key);
920 spin_lock(&hb->lock);
923 plist_for_each_entry_safe(this, next, head, list) {
924 if (match_futex (&this->key, &key)) {
925 if (this->pi_state || this->rt_waiter) {
930 /* Check if one of the bits is set in both bitsets */
931 if (!(this->bitset & bitset))
935 if (++ret >= nr_wake)
940 spin_unlock(&hb->lock);
941 put_futex_key(fshared, &key);
947 * Wake up all waiters hashed on the physical page that is mapped
948 * to this virtual address:
951 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
952 int nr_wake, int nr_wake2, int op)
954 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
955 struct futex_hash_bucket *hb1, *hb2;
956 struct plist_head *head;
957 struct futex_q *this, *next;
961 ret = get_futex_key(uaddr1, fshared, &key1);
962 if (unlikely(ret != 0))
964 ret = get_futex_key(uaddr2, fshared, &key2);
965 if (unlikely(ret != 0))
968 hb1 = hash_futex(&key1);
969 hb2 = hash_futex(&key2);
972 double_lock_hb(hb1, hb2);
973 op_ret = futex_atomic_op_inuser(op, uaddr2);
974 if (unlikely(op_ret < 0)) {
976 double_unlock_hb(hb1, hb2);
980 * we don't get EFAULT from MMU faults if we don't have an MMU,
981 * but we might get them from range checking
987 if (unlikely(op_ret != -EFAULT)) {
992 ret = fault_in_user_writeable(uaddr2);
999 put_futex_key(fshared, &key2);
1000 put_futex_key(fshared, &key1);
1006 plist_for_each_entry_safe(this, next, head, list) {
1007 if (match_futex (&this->key, &key1)) {
1009 if (++ret >= nr_wake)
1018 plist_for_each_entry_safe(this, next, head, list) {
1019 if (match_futex (&this->key, &key2)) {
1021 if (++op_ret >= nr_wake2)
1028 double_unlock_hb(hb1, hb2);
1030 put_futex_key(fshared, &key2);
1032 put_futex_key(fshared, &key1);
1038 * requeue_futex() - Requeue a futex_q from one hb to another
1039 * @q: the futex_q to requeue
1040 * @hb1: the source hash_bucket
1041 * @hb2: the target hash_bucket
1042 * @key2: the new key for the requeued futex_q
1045 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1046 struct futex_hash_bucket *hb2, union futex_key *key2)
1050 * If key1 and key2 hash to the same bucket, no need to
1053 if (likely(&hb1->chain != &hb2->chain)) {
1054 plist_del(&q->list, &hb1->chain);
1055 plist_add(&q->list, &hb2->chain);
1056 q->lock_ptr = &hb2->lock;
1057 #ifdef CONFIG_DEBUG_PI_LIST
1058 q->list.plist.lock = &hb2->lock;
1061 get_futex_key_refs(key2);
1066 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1068 * @key: the key of the requeue target futex
1069 * @hb: the hash_bucket of the requeue target futex
1071 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1072 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1073 * to the requeue target futex so the waiter can detect the wakeup on the right
1074 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1075 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1076 * to protect access to the pi_state to fixup the owner later. Must be called
1077 * with both q->lock_ptr and hb->lock held.
1080 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1081 struct futex_hash_bucket *hb)
1083 get_futex_key_refs(key);
1086 WARN_ON(plist_node_empty(&q->list));
1087 plist_del(&q->list, &q->list.plist);
1089 WARN_ON(!q->rt_waiter);
1090 q->rt_waiter = NULL;
1092 q->lock_ptr = &hb->lock;
1093 #ifdef CONFIG_DEBUG_PI_LIST
1094 q->list.plist.lock = &hb->lock;
1097 wake_up_state(q->task, TASK_NORMAL);
1101 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1102 * @pifutex: the user address of the to futex
1103 * @hb1: the from futex hash bucket, must be locked by the caller
1104 * @hb2: the to futex hash bucket, must be locked by the caller
1105 * @key1: the from futex key
1106 * @key2: the to futex key
1107 * @ps: address to store the pi_state pointer
1108 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1110 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1111 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1112 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1113 * hb1 and hb2 must be held by the caller.
1116 * 0 - failed to acquire the lock atomicly
1117 * 1 - acquired the lock
1120 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1121 struct futex_hash_bucket *hb1,
1122 struct futex_hash_bucket *hb2,
1123 union futex_key *key1, union futex_key *key2,
1124 struct futex_pi_state **ps, int set_waiters)
1126 struct futex_q *top_waiter = NULL;
1130 if (get_futex_value_locked(&curval, pifutex))
1134 * Find the top_waiter and determine if there are additional waiters.
1135 * If the caller intends to requeue more than 1 waiter to pifutex,
1136 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1137 * as we have means to handle the possible fault. If not, don't set
1138 * the bit unecessarily as it will force the subsequent unlock to enter
1141 top_waiter = futex_top_waiter(hb1, key1);
1143 /* There are no waiters, nothing for us to do. */
1147 /* Ensure we requeue to the expected futex. */
1148 if (!match_futex(top_waiter->requeue_pi_key, key2))
1152 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1153 * the contended case or if set_waiters is 1. The pi_state is returned
1154 * in ps in contended cases.
1156 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1159 requeue_pi_wake_futex(top_waiter, key2, hb2);
1165 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1166 * uaddr1: source futex user address
1167 * uaddr2: target futex user address
1168 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1169 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1170 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1171 * pi futex (pi to pi requeue is not supported)
1173 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1174 * uaddr2 atomically on behalf of the top waiter.
1177 * >=0 - on success, the number of tasks requeued or woken
1180 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1181 int nr_wake, int nr_requeue, u32 *cmpval,
1184 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1185 int drop_count = 0, task_count = 0, ret;
1186 struct futex_pi_state *pi_state = NULL;
1187 struct futex_hash_bucket *hb1, *hb2;
1188 struct plist_head *head1;
1189 struct futex_q *this, *next;
1194 * requeue_pi requires a pi_state, try to allocate it now
1195 * without any locks in case it fails.
1197 if (refill_pi_state_cache())
1200 * requeue_pi must wake as many tasks as it can, up to nr_wake
1201 * + nr_requeue, since it acquires the rt_mutex prior to
1202 * returning to userspace, so as to not leave the rt_mutex with
1203 * waiters and no owner. However, second and third wake-ups
1204 * cannot be predicted as they involve race conditions with the
1205 * first wake and a fault while looking up the pi_state. Both
1206 * pthread_cond_signal() and pthread_cond_broadcast() should
1214 if (pi_state != NULL) {
1216 * We will have to lookup the pi_state again, so free this one
1217 * to keep the accounting correct.
1219 free_pi_state(pi_state);
1223 ret = get_futex_key(uaddr1, fshared, &key1);
1224 if (unlikely(ret != 0))
1226 ret = get_futex_key(uaddr2, fshared, &key2);
1227 if (unlikely(ret != 0))
1230 hb1 = hash_futex(&key1);
1231 hb2 = hash_futex(&key2);
1234 double_lock_hb(hb1, hb2);
1236 if (likely(cmpval != NULL)) {
1239 ret = get_futex_value_locked(&curval, uaddr1);
1241 if (unlikely(ret)) {
1242 double_unlock_hb(hb1, hb2);
1244 ret = get_user(curval, uaddr1);
1251 put_futex_key(fshared, &key2);
1252 put_futex_key(fshared, &key1);
1255 if (curval != *cmpval) {
1261 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1263 * Attempt to acquire uaddr2 and wake the top waiter. If we
1264 * intend to requeue waiters, force setting the FUTEX_WAITERS
1265 * bit. We force this here where we are able to easily handle
1266 * faults rather in the requeue loop below.
1268 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1269 &key2, &pi_state, nr_requeue);
1272 * At this point the top_waiter has either taken uaddr2 or is
1273 * waiting on it. If the former, then the pi_state will not
1274 * exist yet, look it up one more time to ensure we have a
1281 ret = get_futex_value_locked(&curval2, uaddr2);
1283 ret = lookup_pi_state(curval2, hb2, &key2,
1291 double_unlock_hb(hb1, hb2);
1292 put_futex_key(fshared, &key2);
1293 put_futex_key(fshared, &key1);
1294 ret = fault_in_user_writeable(uaddr2);
1299 /* The owner was exiting, try again. */
1300 double_unlock_hb(hb1, hb2);
1301 put_futex_key(fshared, &key2);
1302 put_futex_key(fshared, &key1);
1310 head1 = &hb1->chain;
1311 plist_for_each_entry_safe(this, next, head1, list) {
1312 if (task_count - nr_wake >= nr_requeue)
1315 if (!match_futex(&this->key, &key1))
1319 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1320 * be paired with each other and no other futex ops.
1322 if ((requeue_pi && !this->rt_waiter) ||
1323 (!requeue_pi && this->rt_waiter)) {
1329 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1330 * lock, we already woke the top_waiter. If not, it will be
1331 * woken by futex_unlock_pi().
1333 if (++task_count <= nr_wake && !requeue_pi) {
1338 /* Ensure we requeue to the expected futex for requeue_pi. */
1339 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1345 * Requeue nr_requeue waiters and possibly one more in the case
1346 * of requeue_pi if we couldn't acquire the lock atomically.
1349 /* Prepare the waiter to take the rt_mutex. */
1350 atomic_inc(&pi_state->refcount);
1351 this->pi_state = pi_state;
1352 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1356 /* We got the lock. */
1357 requeue_pi_wake_futex(this, &key2, hb2);
1362 this->pi_state = NULL;
1363 free_pi_state(pi_state);
1367 requeue_futex(this, hb1, hb2, &key2);
1372 double_unlock_hb(hb1, hb2);
1375 * drop_futex_key_refs() must be called outside the spinlocks. During
1376 * the requeue we moved futex_q's from the hash bucket at key1 to the
1377 * one at key2 and updated their key pointer. We no longer need to
1378 * hold the references to key1.
1380 while (--drop_count >= 0)
1381 drop_futex_key_refs(&key1);
1384 put_futex_key(fshared, &key2);
1386 put_futex_key(fshared, &key1);
1388 if (pi_state != NULL)
1389 free_pi_state(pi_state);
1390 return ret ? ret : task_count;
1393 /* The key must be already stored in q->key. */
1394 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1396 struct futex_hash_bucket *hb;
1398 hb = hash_futex(&q->key);
1399 q->lock_ptr = &hb->lock;
1401 spin_lock(&hb->lock);
1406 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1408 spin_unlock(&hb->lock);
1412 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1413 * @q: The futex_q to enqueue
1414 * @hb: The destination hash bucket
1416 * The hb->lock must be held by the caller, and is released here. A call to
1417 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1418 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1419 * or nothing if the unqueue is done as part of the wake process and the unqueue
1420 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1423 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1428 * The priority used to register this element is
1429 * - either the real thread-priority for the real-time threads
1430 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1431 * - or MAX_RT_PRIO for non-RT threads.
1432 * Thus, all RT-threads are woken first in priority order, and
1433 * the others are woken last, in FIFO order.
1435 prio = min(current->normal_prio, MAX_RT_PRIO);
1437 plist_node_init(&q->list, prio);
1438 #ifdef CONFIG_DEBUG_PI_LIST
1439 q->list.plist.lock = &hb->lock;
1441 plist_add(&q->list, &hb->chain);
1443 spin_unlock(&hb->lock);
1447 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1448 * @q: The futex_q to unqueue
1450 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1451 * be paired with exactly one earlier call to queue_me().
1454 * 1 - if the futex_q was still queued (and we removed unqueued it)
1455 * 0 - if the futex_q was already removed by the waking thread
1457 static int unqueue_me(struct futex_q *q)
1459 spinlock_t *lock_ptr;
1462 /* In the common case we don't take the spinlock, which is nice. */
1464 lock_ptr = q->lock_ptr;
1466 if (lock_ptr != NULL) {
1467 spin_lock(lock_ptr);
1469 * q->lock_ptr can change between reading it and
1470 * spin_lock(), causing us to take the wrong lock. This
1471 * corrects the race condition.
1473 * Reasoning goes like this: if we have the wrong lock,
1474 * q->lock_ptr must have changed (maybe several times)
1475 * between reading it and the spin_lock(). It can
1476 * change again after the spin_lock() but only if it was
1477 * already changed before the spin_lock(). It cannot,
1478 * however, change back to the original value. Therefore
1479 * we can detect whether we acquired the correct lock.
1481 if (unlikely(lock_ptr != q->lock_ptr)) {
1482 spin_unlock(lock_ptr);
1485 WARN_ON(plist_node_empty(&q->list));
1486 plist_del(&q->list, &q->list.plist);
1488 BUG_ON(q->pi_state);
1490 spin_unlock(lock_ptr);
1494 drop_futex_key_refs(&q->key);
1499 * PI futexes can not be requeued and must remove themself from the
1500 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1503 static void unqueue_me_pi(struct futex_q *q)
1505 WARN_ON(plist_node_empty(&q->list));
1506 plist_del(&q->list, &q->list.plist);
1508 BUG_ON(!q->pi_state);
1509 free_pi_state(q->pi_state);
1512 spin_unlock(q->lock_ptr);
1516 * Fixup the pi_state owner with the new owner.
1518 * Must be called with hash bucket lock held and mm->sem held for non
1521 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1522 struct task_struct *newowner, int fshared)
1524 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1525 struct futex_pi_state *pi_state = q->pi_state;
1526 struct task_struct *oldowner = pi_state->owner;
1527 u32 uval, curval, newval;
1531 if (!pi_state->owner)
1532 newtid |= FUTEX_OWNER_DIED;
1535 * We are here either because we stole the rtmutex from the
1536 * pending owner or we are the pending owner which failed to
1537 * get the rtmutex. We have to replace the pending owner TID
1538 * in the user space variable. This must be atomic as we have
1539 * to preserve the owner died bit here.
1541 * Note: We write the user space value _before_ changing the pi_state
1542 * because we can fault here. Imagine swapped out pages or a fork
1543 * that marked all the anonymous memory readonly for cow.
1545 * Modifying pi_state _before_ the user space value would
1546 * leave the pi_state in an inconsistent state when we fault
1547 * here, because we need to drop the hash bucket lock to
1548 * handle the fault. This might be observed in the PID check
1549 * in lookup_pi_state.
1552 if (get_futex_value_locked(&uval, uaddr))
1556 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1558 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1560 if (curval == -EFAULT)
1568 * We fixed up user space. Now we need to fix the pi_state
1571 if (pi_state->owner != NULL) {
1572 spin_lock_irq(&pi_state->owner->pi_lock);
1573 WARN_ON(list_empty(&pi_state->list));
1574 list_del_init(&pi_state->list);
1575 spin_unlock_irq(&pi_state->owner->pi_lock);
1578 pi_state->owner = newowner;
1580 spin_lock_irq(&newowner->pi_lock);
1581 WARN_ON(!list_empty(&pi_state->list));
1582 list_add(&pi_state->list, &newowner->pi_state_list);
1583 spin_unlock_irq(&newowner->pi_lock);
1587 * To handle the page fault we need to drop the hash bucket
1588 * lock here. That gives the other task (either the pending
1589 * owner itself or the task which stole the rtmutex) the
1590 * chance to try the fixup of the pi_state. So once we are
1591 * back from handling the fault we need to check the pi_state
1592 * after reacquiring the hash bucket lock and before trying to
1593 * do another fixup. When the fixup has been done already we
1597 spin_unlock(q->lock_ptr);
1599 ret = fault_in_user_writeable(uaddr);
1601 spin_lock(q->lock_ptr);
1604 * Check if someone else fixed it for us:
1606 if (pi_state->owner != oldowner)
1616 * In case we must use restart_block to restart a futex_wait,
1617 * we encode in the 'flags' shared capability
1619 #define FLAGS_SHARED 0x01
1620 #define FLAGS_CLOCKRT 0x02
1621 #define FLAGS_HAS_TIMEOUT 0x04
1623 static long futex_wait_restart(struct restart_block *restart);
1626 * fixup_owner() - Post lock pi_state and corner case management
1627 * @uaddr: user address of the futex
1628 * @fshared: whether the futex is shared (1) or not (0)
1629 * @q: futex_q (contains pi_state and access to the rt_mutex)
1630 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1632 * After attempting to lock an rt_mutex, this function is called to cleanup
1633 * the pi_state owner as well as handle race conditions that may allow us to
1634 * acquire the lock. Must be called with the hb lock held.
1637 * 1 - success, lock taken
1638 * 0 - success, lock not taken
1639 * <0 - on error (-EFAULT)
1641 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1644 struct task_struct *owner;
1649 * Got the lock. We might not be the anticipated owner if we
1650 * did a lock-steal - fix up the PI-state in that case:
1652 if (q->pi_state->owner != current)
1653 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1658 * Catch the rare case, where the lock was released when we were on the
1659 * way back before we locked the hash bucket.
1661 if (q->pi_state->owner == current) {
1663 * Try to get the rt_mutex now. This might fail as some other
1664 * task acquired the rt_mutex after we removed ourself from the
1665 * rt_mutex waiters list.
1667 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1673 * pi_state is incorrect, some other task did a lock steal and
1674 * we returned due to timeout or signal without taking the
1675 * rt_mutex. Too late. We can access the rt_mutex_owner without
1676 * locking, as the other task is now blocked on the hash bucket
1677 * lock. Fix the state up.
1679 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1680 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1685 * Paranoia check. If we did not take the lock, then we should not be
1686 * the owner, nor the pending owner, of the rt_mutex.
1688 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1689 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1690 "pi-state %p\n", ret,
1691 q->pi_state->pi_mutex.owner,
1692 q->pi_state->owner);
1695 return ret ? ret : locked;
1699 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1700 * @hb: the futex hash bucket, must be locked by the caller
1701 * @q: the futex_q to queue up on
1702 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1704 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1705 struct hrtimer_sleeper *timeout)
1708 * The task state is guaranteed to be set before another task can
1709 * wake it. set_current_state() is implemented using set_mb() and
1710 * queue_me() calls spin_unlock() upon completion, both serializing
1711 * access to the hash list and forcing another memory barrier.
1713 set_current_state(TASK_INTERRUPTIBLE);
1718 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1719 if (!hrtimer_active(&timeout->timer))
1720 timeout->task = NULL;
1724 * If we have been removed from the hash list, then another task
1725 * has tried to wake us, and we can skip the call to schedule().
1727 if (likely(!plist_node_empty(&q->list))) {
1729 * If the timer has already expired, current will already be
1730 * flagged for rescheduling. Only call schedule if there
1731 * is no timeout, or if it has yet to expire.
1733 if (!timeout || timeout->task)
1736 __set_current_state(TASK_RUNNING);
1740 * futex_wait_setup() - Prepare to wait on a futex
1741 * @uaddr: the futex userspace address
1742 * @val: the expected value
1743 * @fshared: whether the futex is shared (1) or not (0)
1744 * @q: the associated futex_q
1745 * @hb: storage for hash_bucket pointer to be returned to caller
1747 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1748 * compare it with the expected value. Handle atomic faults internally.
1749 * Return with the hb lock held and a q.key reference on success, and unlocked
1750 * with no q.key reference on failure.
1753 * 0 - uaddr contains val and hb has been locked
1754 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1756 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1757 struct futex_q *q, struct futex_hash_bucket **hb)
1763 * Access the page AFTER the hash-bucket is locked.
1764 * Order is important:
1766 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1767 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1769 * The basic logical guarantee of a futex is that it blocks ONLY
1770 * if cond(var) is known to be true at the time of blocking, for
1771 * any cond. If we queued after testing *uaddr, that would open
1772 * a race condition where we could block indefinitely with
1773 * cond(var) false, which would violate the guarantee.
1775 * A consequence is that futex_wait() can return zero and absorb
1776 * a wakeup when *uaddr != val on entry to the syscall. This is
1780 q->key = FUTEX_KEY_INIT;
1781 ret = get_futex_key(uaddr, fshared, &q->key);
1782 if (unlikely(ret != 0))
1786 *hb = queue_lock(q);
1788 ret = get_futex_value_locked(&uval, uaddr);
1791 queue_unlock(q, *hb);
1793 ret = get_user(uval, uaddr);
1800 put_futex_key(fshared, &q->key);
1805 queue_unlock(q, *hb);
1811 put_futex_key(fshared, &q->key);
1815 static int futex_wait(u32 __user *uaddr, int fshared,
1816 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1818 struct hrtimer_sleeper timeout, *to = NULL;
1819 struct restart_block *restart;
1820 struct futex_hash_bucket *hb;
1830 q.requeue_pi_key = NULL;
1835 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1836 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1837 hrtimer_init_sleeper(to, current);
1838 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1839 current->timer_slack_ns);
1844 * Prepare to wait on uaddr. On success, holds hb lock and increments
1847 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1851 /* queue_me and wait for wakeup, timeout, or a signal. */
1852 futex_wait_queue_me(hb, &q, to);
1854 /* If we were woken (and unqueued), we succeeded, whatever. */
1856 /* unqueue_me() drops q.key ref */
1857 if (!unqueue_me(&q))
1860 if (to && !to->task)
1864 * We expect signal_pending(current), but we might be the
1865 * victim of a spurious wakeup as well.
1867 if (!signal_pending(current))
1874 restart = ¤t_thread_info()->restart_block;
1875 restart->fn = futex_wait_restart;
1876 restart->futex.uaddr = (u32 *)uaddr;
1877 restart->futex.val = val;
1878 restart->futex.time = abs_time->tv64;
1879 restart->futex.bitset = bitset;
1880 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1883 restart->futex.flags |= FLAGS_SHARED;
1885 restart->futex.flags |= FLAGS_CLOCKRT;
1887 ret = -ERESTART_RESTARTBLOCK;
1891 hrtimer_cancel(&to->timer);
1892 destroy_hrtimer_on_stack(&to->timer);
1898 static long futex_wait_restart(struct restart_block *restart)
1900 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1902 ktime_t t, *tp = NULL;
1904 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1905 t.tv64 = restart->futex.time;
1908 restart->fn = do_no_restart_syscall;
1909 if (restart->futex.flags & FLAGS_SHARED)
1911 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1912 restart->futex.bitset,
1913 restart->futex.flags & FLAGS_CLOCKRT);
1918 * Userspace tried a 0 -> TID atomic transition of the futex value
1919 * and failed. The kernel side here does the whole locking operation:
1920 * if there are waiters then it will block, it does PI, etc. (Due to
1921 * races the kernel might see a 0 value of the futex too.)
1923 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1924 int detect, ktime_t *time, int trylock)
1926 struct hrtimer_sleeper timeout, *to = NULL;
1927 struct futex_hash_bucket *hb;
1931 if (refill_pi_state_cache())
1936 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1938 hrtimer_init_sleeper(to, current);
1939 hrtimer_set_expires(&to->timer, *time);
1944 q.requeue_pi_key = NULL;
1946 q.key = FUTEX_KEY_INIT;
1947 ret = get_futex_key(uaddr, fshared, &q.key);
1948 if (unlikely(ret != 0))
1952 hb = queue_lock(&q);
1954 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1955 if (unlikely(ret)) {
1958 /* We got the lock. */
1960 goto out_unlock_put_key;
1965 * Task is exiting and we just wait for the
1968 queue_unlock(&q, hb);
1969 put_futex_key(fshared, &q.key);
1973 goto out_unlock_put_key;
1978 * Only actually queue now that the atomic ops are done:
1982 WARN_ON(!q.pi_state);
1984 * Block on the PI mutex:
1987 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1989 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1990 /* Fixup the trylock return value: */
1991 ret = ret ? 0 : -EWOULDBLOCK;
1994 spin_lock(q.lock_ptr);
1996 * Fixup the pi_state owner and possibly acquire the lock if we
1999 res = fixup_owner(uaddr, fshared, &q, !ret);
2001 * If fixup_owner() returned an error, proprogate that. If it acquired
2002 * the lock, clear our -ETIMEDOUT or -EINTR.
2005 ret = (res < 0) ? res : 0;
2008 * If fixup_owner() faulted and was unable to handle the fault, unlock
2009 * it and return the fault to userspace.
2011 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2012 rt_mutex_unlock(&q.pi_state->pi_mutex);
2014 /* Unqueue and drop the lock */
2020 queue_unlock(&q, hb);
2023 put_futex_key(fshared, &q.key);
2026 destroy_hrtimer_on_stack(&to->timer);
2027 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2030 queue_unlock(&q, hb);
2032 ret = fault_in_user_writeable(uaddr);
2039 put_futex_key(fshared, &q.key);
2044 * Userspace attempted a TID -> 0 atomic transition, and failed.
2045 * This is the in-kernel slowpath: we look up the PI state (if any),
2046 * and do the rt-mutex unlock.
2048 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2050 struct futex_hash_bucket *hb;
2051 struct futex_q *this, *next;
2053 struct plist_head *head;
2054 union futex_key key = FUTEX_KEY_INIT;
2058 if (get_user(uval, uaddr))
2061 * We release only a lock we actually own:
2063 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2066 ret = get_futex_key(uaddr, fshared, &key);
2067 if (unlikely(ret != 0))
2070 hb = hash_futex(&key);
2071 spin_lock(&hb->lock);
2074 * To avoid races, try to do the TID -> 0 atomic transition
2075 * again. If it succeeds then we can return without waking
2078 if (!(uval & FUTEX_OWNER_DIED))
2079 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2082 if (unlikely(uval == -EFAULT))
2085 * Rare case: we managed to release the lock atomically,
2086 * no need to wake anyone else up:
2088 if (unlikely(uval == task_pid_vnr(current)))
2092 * Ok, other tasks may need to be woken up - check waiters
2093 * and do the wakeup if necessary:
2097 plist_for_each_entry_safe(this, next, head, list) {
2098 if (!match_futex (&this->key, &key))
2100 ret = wake_futex_pi(uaddr, uval, this);
2102 * The atomic access to the futex value
2103 * generated a pagefault, so retry the
2104 * user-access and the wakeup:
2111 * No waiters - kernel unlocks the futex:
2113 if (!(uval & FUTEX_OWNER_DIED)) {
2114 ret = unlock_futex_pi(uaddr, uval);
2120 spin_unlock(&hb->lock);
2121 put_futex_key(fshared, &key);
2127 spin_unlock(&hb->lock);
2128 put_futex_key(fshared, &key);
2130 ret = fault_in_user_writeable(uaddr);
2138 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2139 * @hb: the hash_bucket futex_q was original enqueued on
2140 * @q: the futex_q woken while waiting to be requeued
2141 * @key2: the futex_key of the requeue target futex
2142 * @timeout: the timeout associated with the wait (NULL if none)
2144 * Detect if the task was woken on the initial futex as opposed to the requeue
2145 * target futex. If so, determine if it was a timeout or a signal that caused
2146 * the wakeup and return the appropriate error code to the caller. Must be
2147 * called with the hb lock held.
2150 * 0 - no early wakeup detected
2151 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2154 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2155 struct futex_q *q, union futex_key *key2,
2156 struct hrtimer_sleeper *timeout)
2161 * With the hb lock held, we avoid races while we process the wakeup.
2162 * We only need to hold hb (and not hb2) to ensure atomicity as the
2163 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2164 * It can't be requeued from uaddr2 to something else since we don't
2165 * support a PI aware source futex for requeue.
2167 if (!match_futex(&q->key, key2)) {
2168 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2170 * We were woken prior to requeue by a timeout or a signal.
2171 * Unqueue the futex_q and determine which it was.
2173 plist_del(&q->list, &q->list.plist);
2175 /* Handle spurious wakeups gracefully */
2177 if (timeout && !timeout->task)
2179 else if (signal_pending(current))
2180 ret = -ERESTARTNOINTR;
2186 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2187 * @uaddr: the futex we initially wait on (non-pi)
2188 * @fshared: whether the futexes are shared (1) or not (0). They must be
2189 * the same type, no requeueing from private to shared, etc.
2190 * @val: the expected value of uaddr
2191 * @abs_time: absolute timeout
2192 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2193 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2194 * @uaddr2: the pi futex we will take prior to returning to user-space
2196 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2197 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2198 * complete the acquisition of the rt_mutex prior to returning to userspace.
2199 * This ensures the rt_mutex maintains an owner when it has waiters; without
2200 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2203 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2204 * via the following:
2205 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2206 * 2) wakeup on uaddr2 after a requeue
2210 * If 3, cleanup and return -ERESTARTNOINTR.
2212 * If 2, we may then block on trying to take the rt_mutex and return via:
2213 * 5) successful lock
2216 * 8) other lock acquisition failure
2218 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2220 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2226 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2227 u32 val, ktime_t *abs_time, u32 bitset,
2228 int clockrt, u32 __user *uaddr2)
2230 struct hrtimer_sleeper timeout, *to = NULL;
2231 struct rt_mutex_waiter rt_waiter;
2232 struct rt_mutex *pi_mutex = NULL;
2233 struct futex_hash_bucket *hb;
2234 union futex_key key2;
2243 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2244 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2245 hrtimer_init_sleeper(to, current);
2246 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2247 current->timer_slack_ns);
2251 * The waiter is allocated on our stack, manipulated by the requeue
2252 * code while we sleep on uaddr.
2254 debug_rt_mutex_init_waiter(&rt_waiter);
2255 rt_waiter.task = NULL;
2257 key2 = FUTEX_KEY_INIT;
2258 ret = get_futex_key(uaddr2, fshared, &key2);
2259 if (unlikely(ret != 0))
2264 q.rt_waiter = &rt_waiter;
2265 q.requeue_pi_key = &key2;
2268 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2271 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2275 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2276 futex_wait_queue_me(hb, &q, to);
2278 spin_lock(&hb->lock);
2279 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2280 spin_unlock(&hb->lock);
2285 * In order for us to be here, we know our q.key == key2, and since
2286 * we took the hb->lock above, we also know that futex_requeue() has
2287 * completed and we no longer have to concern ourselves with a wakeup
2288 * race with the atomic proxy lock acquisition by the requeue code. The
2289 * futex_requeue dropped our key1 reference and incremented our key2
2293 /* Check if the requeue code acquired the second futex for us. */
2296 * Got the lock. We might not be the anticipated owner if we
2297 * did a lock-steal - fix up the PI-state in that case.
2299 if (q.pi_state && (q.pi_state->owner != current)) {
2300 spin_lock(q.lock_ptr);
2301 ret = fixup_pi_state_owner(uaddr2, &q, current,
2303 spin_unlock(q.lock_ptr);
2307 * We have been woken up by futex_unlock_pi(), a timeout, or a
2308 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2311 WARN_ON(!&q.pi_state);
2312 pi_mutex = &q.pi_state->pi_mutex;
2313 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2314 debug_rt_mutex_free_waiter(&rt_waiter);
2316 spin_lock(q.lock_ptr);
2318 * Fixup the pi_state owner and possibly acquire the lock if we
2321 res = fixup_owner(uaddr2, fshared, &q, !ret);
2323 * If fixup_owner() returned an error, proprogate that. If it
2324 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2327 ret = (res < 0) ? res : 0;
2329 /* Unqueue and drop the lock. */
2334 * If fixup_pi_state_owner() faulted and was unable to handle the
2335 * fault, unlock the rt_mutex and return the fault to userspace.
2337 if (ret == -EFAULT) {
2338 if (rt_mutex_owner(pi_mutex) == current)
2339 rt_mutex_unlock(pi_mutex);
2340 } else if (ret == -EINTR) {
2342 * We've already been requeued, but cannot restart by calling
2343 * futex_lock_pi() directly. We could restart this syscall, but
2344 * it would detect that the user space "val" changed and return
2345 * -EWOULDBLOCK. Save the overhead of the restart and return
2346 * -EWOULDBLOCK directly.
2352 put_futex_key(fshared, &q.key);
2354 put_futex_key(fshared, &key2);
2358 hrtimer_cancel(&to->timer);
2359 destroy_hrtimer_on_stack(&to->timer);
2365 * Support for robust futexes: the kernel cleans up held futexes at
2368 * Implementation: user-space maintains a per-thread list of locks it
2369 * is holding. Upon do_exit(), the kernel carefully walks this list,
2370 * and marks all locks that are owned by this thread with the
2371 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2372 * always manipulated with the lock held, so the list is private and
2373 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2374 * field, to allow the kernel to clean up if the thread dies after
2375 * acquiring the lock, but just before it could have added itself to
2376 * the list. There can only be one such pending lock.
2380 * sys_set_robust_list() - Set the robust-futex list head of a task
2381 * @head: pointer to the list-head
2382 * @len: length of the list-head, as userspace expects
2384 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2387 if (!futex_cmpxchg_enabled)
2390 * The kernel knows only one size for now:
2392 if (unlikely(len != sizeof(*head)))
2395 current->robust_list = head;
2401 * sys_get_robust_list() - Get the robust-futex list head of a task
2402 * @pid: pid of the process [zero for current task]
2403 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2404 * @len_ptr: pointer to a length field, the kernel fills in the header size
2406 SYSCALL_DEFINE3(get_robust_list, int, pid,
2407 struct robust_list_head __user * __user *, head_ptr,
2408 size_t __user *, len_ptr)
2410 struct robust_list_head __user *head;
2412 const struct cred *cred = current_cred(), *pcred;
2414 if (!futex_cmpxchg_enabled)
2418 head = current->robust_list;
2420 struct task_struct *p;
2424 p = find_task_by_vpid(pid);
2428 pcred = __task_cred(p);
2429 if (cred->euid != pcred->euid &&
2430 cred->euid != pcred->uid &&
2431 !capable(CAP_SYS_PTRACE))
2433 head = p->robust_list;
2437 if (put_user(sizeof(*head), len_ptr))
2439 return put_user(head, head_ptr);
2448 * Process a futex-list entry, check whether it's owned by the
2449 * dying task, and do notification if so:
2451 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2453 u32 uval, nval, mval;
2456 if (get_user(uval, uaddr))
2459 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2461 * Ok, this dying thread is truly holding a futex
2462 * of interest. Set the OWNER_DIED bit atomically
2463 * via cmpxchg, and if the value had FUTEX_WAITERS
2464 * set, wake up a waiter (if any). (We have to do a
2465 * futex_wake() even if OWNER_DIED is already set -
2466 * to handle the rare but possible case of recursive
2467 * thread-death.) The rest of the cleanup is done in
2470 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2471 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2473 if (nval == -EFAULT)
2480 * Wake robust non-PI futexes here. The wakeup of
2481 * PI futexes happens in exit_pi_state():
2483 if (!pi && (uval & FUTEX_WAITERS))
2484 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2490 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2492 static inline int fetch_robust_entry(struct robust_list __user **entry,
2493 struct robust_list __user * __user *head,
2496 unsigned long uentry;
2498 if (get_user(uentry, (unsigned long __user *)head))
2501 *entry = (void __user *)(uentry & ~1UL);
2508 * Walk curr->robust_list (very carefully, it's a userspace list!)
2509 * and mark any locks found there dead, and notify any waiters.
2511 * We silently return on any sign of list-walking problem.
2513 void exit_robust_list(struct task_struct *curr)
2515 struct robust_list_head __user *head = curr->robust_list;
2516 struct robust_list __user *entry, *next_entry, *pending;
2517 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2518 unsigned long futex_offset;
2521 if (!futex_cmpxchg_enabled)
2525 * Fetch the list head (which was registered earlier, via
2526 * sys_set_robust_list()):
2528 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2531 * Fetch the relative futex offset:
2533 if (get_user(futex_offset, &head->futex_offset))
2536 * Fetch any possibly pending lock-add first, and handle it
2539 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2542 next_entry = NULL; /* avoid warning with gcc */
2543 while (entry != &head->list) {
2545 * Fetch the next entry in the list before calling
2546 * handle_futex_death:
2548 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2550 * A pending lock might already be on the list, so
2551 * don't process it twice:
2553 if (entry != pending)
2554 if (handle_futex_death((void __user *)entry + futex_offset,
2562 * Avoid excessively long or circular lists:
2571 handle_futex_death((void __user *)pending + futex_offset,
2575 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2576 u32 __user *uaddr2, u32 val2, u32 val3)
2578 int clockrt, ret = -ENOSYS;
2579 int cmd = op & FUTEX_CMD_MASK;
2582 if (!(op & FUTEX_PRIVATE_FLAG))
2585 clockrt = op & FUTEX_CLOCK_REALTIME;
2586 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2591 val3 = FUTEX_BITSET_MATCH_ANY;
2592 case FUTEX_WAIT_BITSET:
2593 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2596 val3 = FUTEX_BITSET_MATCH_ANY;
2597 case FUTEX_WAKE_BITSET:
2598 ret = futex_wake(uaddr, fshared, val, val3);
2601 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2603 case FUTEX_CMP_REQUEUE:
2604 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2608 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2611 if (futex_cmpxchg_enabled)
2612 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2614 case FUTEX_UNLOCK_PI:
2615 if (futex_cmpxchg_enabled)
2616 ret = futex_unlock_pi(uaddr, fshared);
2618 case FUTEX_TRYLOCK_PI:
2619 if (futex_cmpxchg_enabled)
2620 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2622 case FUTEX_WAIT_REQUEUE_PI:
2623 val3 = FUTEX_BITSET_MATCH_ANY;
2624 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2627 case FUTEX_CMP_REQUEUE_PI:
2628 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2638 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2639 struct timespec __user *, utime, u32 __user *, uaddr2,
2643 ktime_t t, *tp = NULL;
2645 int cmd = op & FUTEX_CMD_MASK;
2647 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2648 cmd == FUTEX_WAIT_BITSET ||
2649 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2650 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2652 if (!timespec_valid(&ts))
2655 t = timespec_to_ktime(ts);
2656 if (cmd == FUTEX_WAIT)
2657 t = ktime_add_safe(ktime_get(), t);
2661 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2662 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2664 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2665 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2666 val2 = (u32) (unsigned long) utime;
2668 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2671 static int __init futex_init(void)
2677 * This will fail and we want it. Some arch implementations do
2678 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2679 * functionality. We want to know that before we call in any
2680 * of the complex code paths. Also we want to prevent
2681 * registration of robust lists in that case. NULL is
2682 * guaranteed to fault and we get -EFAULT on functional
2683 * implementation, the non functional ones will return
2686 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2687 if (curval == -EFAULT)
2688 futex_cmpxchg_enabled = 1;
2690 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2691 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2692 spin_lock_init(&futex_queues[i].lock);
2697 __initcall(futex_init);