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;
464 const struct cred *cred = current_cred(), *pcred;
467 p = find_task_by_vpid(pid);
471 pcred = __task_cred(p);
472 if (cred->euid != pcred->euid &&
473 cred->euid != pcred->uid)
485 * This task is holding PI mutexes at exit time => bad.
486 * Kernel cleans up PI-state, but userspace is likely hosed.
487 * (Robust-futex cleanup is separate and might save the day for userspace.)
489 void exit_pi_state_list(struct task_struct *curr)
491 struct list_head *next, *head = &curr->pi_state_list;
492 struct futex_pi_state *pi_state;
493 struct futex_hash_bucket *hb;
494 union futex_key key = FUTEX_KEY_INIT;
496 if (!futex_cmpxchg_enabled)
499 * We are a ZOMBIE and nobody can enqueue itself on
500 * pi_state_list anymore, but we have to be careful
501 * versus waiters unqueueing themselves:
503 spin_lock_irq(&curr->pi_lock);
504 while (!list_empty(head)) {
507 pi_state = list_entry(next, struct futex_pi_state, list);
509 hb = hash_futex(&key);
510 spin_unlock_irq(&curr->pi_lock);
512 spin_lock(&hb->lock);
514 spin_lock_irq(&curr->pi_lock);
516 * We dropped the pi-lock, so re-check whether this
517 * task still owns the PI-state:
519 if (head->next != next) {
520 spin_unlock(&hb->lock);
524 WARN_ON(pi_state->owner != curr);
525 WARN_ON(list_empty(&pi_state->list));
526 list_del_init(&pi_state->list);
527 pi_state->owner = NULL;
528 spin_unlock_irq(&curr->pi_lock);
530 rt_mutex_unlock(&pi_state->pi_mutex);
532 spin_unlock(&hb->lock);
534 spin_lock_irq(&curr->pi_lock);
536 spin_unlock_irq(&curr->pi_lock);
540 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
541 union futex_key *key, struct futex_pi_state **ps)
543 struct futex_pi_state *pi_state = NULL;
544 struct futex_q *this, *next;
545 struct plist_head *head;
546 struct task_struct *p;
547 pid_t pid = uval & FUTEX_TID_MASK;
551 plist_for_each_entry_safe(this, next, head, list) {
552 if (match_futex(&this->key, key)) {
554 * Another waiter already exists - bump up
555 * the refcount and return its pi_state:
557 pi_state = this->pi_state;
559 * Userspace might have messed up non PI and PI futexes
561 if (unlikely(!pi_state))
564 WARN_ON(!atomic_read(&pi_state->refcount));
567 * When pi_state->owner is NULL then the owner died
568 * and another waiter is on the fly. pi_state->owner
569 * is fixed up by the task which acquires
570 * pi_state->rt_mutex.
572 * We do not check for pid == 0 which can happen when
573 * the owner died and robust_list_exit() cleared the
576 if (pid && pi_state->owner) {
578 * Bail out if user space manipulated the
581 if (pid != task_pid_vnr(pi_state->owner))
585 atomic_inc(&pi_state->refcount);
593 * We are the first waiter - try to look up the real owner and attach
594 * the new pi_state to it, but bail out when TID = 0
598 p = futex_find_get_task(pid);
603 * We need to look at the task state flags to figure out,
604 * whether the task is exiting. To protect against the do_exit
605 * change of the task flags, we do this protected by
608 spin_lock_irq(&p->pi_lock);
609 if (unlikely(p->flags & PF_EXITING)) {
611 * The task is on the way out. When PF_EXITPIDONE is
612 * set, we know that the task has finished the
615 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
617 spin_unlock_irq(&p->pi_lock);
622 pi_state = alloc_pi_state();
625 * Initialize the pi_mutex in locked state and make 'p'
628 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
630 /* Store the key for possible exit cleanups: */
631 pi_state->key = *key;
633 WARN_ON(!list_empty(&pi_state->list));
634 list_add(&pi_state->list, &p->pi_state_list);
636 spin_unlock_irq(&p->pi_lock);
646 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
647 * @uaddr: the pi futex user address
648 * @hb: the pi futex hash bucket
649 * @key: the futex key associated with uaddr and hb
650 * @ps: the pi_state pointer where we store the result of the
652 * @task: the task to perform the atomic lock work for. This will
653 * be "current" except in the case of requeue pi.
654 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
658 * 1 - acquired the lock
661 * The hb->lock and futex_key refs shall be held by the caller.
663 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
664 union futex_key *key,
665 struct futex_pi_state **ps,
666 struct task_struct *task, int set_waiters)
668 int lock_taken, ret, ownerdied = 0;
669 u32 uval, newval, curval;
672 ret = lock_taken = 0;
675 * To avoid races, we attempt to take the lock here again
676 * (by doing a 0 -> TID atomic cmpxchg), while holding all
677 * the locks. It will most likely not succeed.
679 newval = task_pid_vnr(task);
681 newval |= FUTEX_WAITERS;
683 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
685 if (unlikely(curval == -EFAULT))
691 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
695 * Surprise - we got the lock. Just return to userspace:
697 if (unlikely(!curval))
703 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
704 * to wake at the next unlock.
706 newval = curval | FUTEX_WAITERS;
709 * There are two cases, where a futex might have no owner (the
710 * owner TID is 0): OWNER_DIED. We take over the futex in this
711 * case. We also do an unconditional take over, when the owner
714 * This is safe as we are protected by the hash bucket lock !
716 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
717 /* Keep the OWNER_DIED bit */
718 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
723 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
725 if (unlikely(curval == -EFAULT))
727 if (unlikely(curval != uval))
731 * We took the lock due to owner died take over.
733 if (unlikely(lock_taken))
737 * We dont have the lock. Look up the PI state (or create it if
738 * we are the first waiter):
740 ret = lookup_pi_state(uval, hb, key, ps);
746 * No owner found for this futex. Check if the
747 * OWNER_DIED bit is set to figure out whether
748 * this is a robust futex or not.
750 if (get_futex_value_locked(&curval, uaddr))
754 * We simply start over in case of a robust
755 * futex. The code above will take the futex
758 if (curval & FUTEX_OWNER_DIED) {
771 * The hash bucket lock must be held when this is called.
772 * Afterwards, the futex_q must not be accessed.
774 static void wake_futex(struct futex_q *q)
776 struct task_struct *p = q->task;
779 * We set q->lock_ptr = NULL _before_ we wake up the task. If
780 * a non futex wake up happens on another CPU then the task
781 * might exit and p would dereference a non existing task
782 * struct. Prevent this by holding a reference on p across the
787 plist_del(&q->list, &q->list.plist);
789 * The waiting task can free the futex_q as soon as
790 * q->lock_ptr = NULL is written, without taking any locks. A
791 * memory barrier is required here to prevent the following
792 * store to lock_ptr from getting ahead of the plist_del.
797 wake_up_state(p, TASK_NORMAL);
801 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
803 struct task_struct *new_owner;
804 struct futex_pi_state *pi_state = this->pi_state;
811 * If current does not own the pi_state then the futex is
812 * inconsistent and user space fiddled with the futex value.
814 if (pi_state->owner != current)
817 spin_lock(&pi_state->pi_mutex.wait_lock);
818 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
821 * This happens when we have stolen the lock and the original
822 * pending owner did not enqueue itself back on the rt_mutex.
823 * Thats not a tragedy. We know that way, that a lock waiter
824 * is on the fly. We make the futex_q waiter the pending owner.
827 new_owner = this->task;
830 * We pass it to the next owner. (The WAITERS bit is always
831 * kept enabled while there is PI state around. We must also
832 * preserve the owner died bit.)
834 if (!(uval & FUTEX_OWNER_DIED)) {
837 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
839 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
841 if (curval == -EFAULT)
843 else if (curval != uval)
846 spin_unlock(&pi_state->pi_mutex.wait_lock);
851 spin_lock_irq(&pi_state->owner->pi_lock);
852 WARN_ON(list_empty(&pi_state->list));
853 list_del_init(&pi_state->list);
854 spin_unlock_irq(&pi_state->owner->pi_lock);
856 spin_lock_irq(&new_owner->pi_lock);
857 WARN_ON(!list_empty(&pi_state->list));
858 list_add(&pi_state->list, &new_owner->pi_state_list);
859 pi_state->owner = new_owner;
860 spin_unlock_irq(&new_owner->pi_lock);
862 spin_unlock(&pi_state->pi_mutex.wait_lock);
863 rt_mutex_unlock(&pi_state->pi_mutex);
868 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
873 * There is no waiter, so we unlock the futex. The owner died
874 * bit has not to be preserved here. We are the owner:
876 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
878 if (oldval == -EFAULT)
887 * Express the locking dependencies for lockdep:
890 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
893 spin_lock(&hb1->lock);
895 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
896 } else { /* hb1 > hb2 */
897 spin_lock(&hb2->lock);
898 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
903 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
905 spin_unlock(&hb1->lock);
907 spin_unlock(&hb2->lock);
911 * Wake up waiters matching bitset queued on this futex (uaddr).
913 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
915 struct futex_hash_bucket *hb;
916 struct futex_q *this, *next;
917 struct plist_head *head;
918 union futex_key key = FUTEX_KEY_INIT;
924 ret = get_futex_key(uaddr, fshared, &key);
925 if (unlikely(ret != 0))
928 hb = hash_futex(&key);
929 spin_lock(&hb->lock);
932 plist_for_each_entry_safe(this, next, head, list) {
933 if (match_futex (&this->key, &key)) {
934 if (this->pi_state || this->rt_waiter) {
939 /* Check if one of the bits is set in both bitsets */
940 if (!(this->bitset & bitset))
944 if (++ret >= nr_wake)
949 spin_unlock(&hb->lock);
950 put_futex_key(fshared, &key);
956 * Wake up all waiters hashed on the physical page that is mapped
957 * to this virtual address:
960 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
961 int nr_wake, int nr_wake2, int op)
963 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
964 struct futex_hash_bucket *hb1, *hb2;
965 struct plist_head *head;
966 struct futex_q *this, *next;
970 ret = get_futex_key(uaddr1, fshared, &key1);
971 if (unlikely(ret != 0))
973 ret = get_futex_key(uaddr2, fshared, &key2);
974 if (unlikely(ret != 0))
977 hb1 = hash_futex(&key1);
978 hb2 = hash_futex(&key2);
981 double_lock_hb(hb1, hb2);
982 op_ret = futex_atomic_op_inuser(op, uaddr2);
983 if (unlikely(op_ret < 0)) {
985 double_unlock_hb(hb1, hb2);
989 * we don't get EFAULT from MMU faults if we don't have an MMU,
990 * but we might get them from range checking
996 if (unlikely(op_ret != -EFAULT)) {
1001 ret = fault_in_user_writeable(uaddr2);
1008 put_futex_key(fshared, &key2);
1009 put_futex_key(fshared, &key1);
1015 plist_for_each_entry_safe(this, next, head, list) {
1016 if (match_futex (&this->key, &key1)) {
1018 if (++ret >= nr_wake)
1027 plist_for_each_entry_safe(this, next, head, list) {
1028 if (match_futex (&this->key, &key2)) {
1030 if (++op_ret >= nr_wake2)
1037 double_unlock_hb(hb1, hb2);
1039 put_futex_key(fshared, &key2);
1041 put_futex_key(fshared, &key1);
1047 * requeue_futex() - Requeue a futex_q from one hb to another
1048 * @q: the futex_q to requeue
1049 * @hb1: the source hash_bucket
1050 * @hb2: the target hash_bucket
1051 * @key2: the new key for the requeued futex_q
1054 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1055 struct futex_hash_bucket *hb2, union futex_key *key2)
1059 * If key1 and key2 hash to the same bucket, no need to
1062 if (likely(&hb1->chain != &hb2->chain)) {
1063 plist_del(&q->list, &hb1->chain);
1064 plist_add(&q->list, &hb2->chain);
1065 q->lock_ptr = &hb2->lock;
1066 #ifdef CONFIG_DEBUG_PI_LIST
1067 q->list.plist.lock = &hb2->lock;
1070 get_futex_key_refs(key2);
1075 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1077 * @key: the key of the requeue target futex
1078 * @hb: the hash_bucket of the requeue target futex
1080 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1081 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1082 * to the requeue target futex so the waiter can detect the wakeup on the right
1083 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1084 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1085 * to protect access to the pi_state to fixup the owner later. Must be called
1086 * with both q->lock_ptr and hb->lock held.
1089 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1090 struct futex_hash_bucket *hb)
1092 get_futex_key_refs(key);
1095 WARN_ON(plist_node_empty(&q->list));
1096 plist_del(&q->list, &q->list.plist);
1098 WARN_ON(!q->rt_waiter);
1099 q->rt_waiter = NULL;
1101 q->lock_ptr = &hb->lock;
1102 #ifdef CONFIG_DEBUG_PI_LIST
1103 q->list.plist.lock = &hb->lock;
1106 wake_up_state(q->task, TASK_NORMAL);
1110 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1111 * @pifutex: the user address of the to futex
1112 * @hb1: the from futex hash bucket, must be locked by the caller
1113 * @hb2: the to futex hash bucket, must be locked by the caller
1114 * @key1: the from futex key
1115 * @key2: the to futex key
1116 * @ps: address to store the pi_state pointer
1117 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1119 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1120 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1121 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1122 * hb1 and hb2 must be held by the caller.
1125 * 0 - failed to acquire the lock atomicly
1126 * 1 - acquired the lock
1129 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1130 struct futex_hash_bucket *hb1,
1131 struct futex_hash_bucket *hb2,
1132 union futex_key *key1, union futex_key *key2,
1133 struct futex_pi_state **ps, int set_waiters)
1135 struct futex_q *top_waiter = NULL;
1139 if (get_futex_value_locked(&curval, pifutex))
1143 * Find the top_waiter and determine if there are additional waiters.
1144 * If the caller intends to requeue more than 1 waiter to pifutex,
1145 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1146 * as we have means to handle the possible fault. If not, don't set
1147 * the bit unecessarily as it will force the subsequent unlock to enter
1150 top_waiter = futex_top_waiter(hb1, key1);
1152 /* There are no waiters, nothing for us to do. */
1156 /* Ensure we requeue to the expected futex. */
1157 if (!match_futex(top_waiter->requeue_pi_key, key2))
1161 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1162 * the contended case or if set_waiters is 1. The pi_state is returned
1163 * in ps in contended cases.
1165 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1168 requeue_pi_wake_futex(top_waiter, key2, hb2);
1174 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1175 * uaddr1: source futex user address
1176 * uaddr2: target futex user address
1177 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1178 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1179 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1180 * pi futex (pi to pi requeue is not supported)
1182 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1183 * uaddr2 atomically on behalf of the top waiter.
1186 * >=0 - on success, the number of tasks requeued or woken
1189 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1190 int nr_wake, int nr_requeue, u32 *cmpval,
1193 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1194 int drop_count = 0, task_count = 0, ret;
1195 struct futex_pi_state *pi_state = NULL;
1196 struct futex_hash_bucket *hb1, *hb2;
1197 struct plist_head *head1;
1198 struct futex_q *this, *next;
1203 * requeue_pi requires a pi_state, try to allocate it now
1204 * without any locks in case it fails.
1206 if (refill_pi_state_cache())
1209 * requeue_pi must wake as many tasks as it can, up to nr_wake
1210 * + nr_requeue, since it acquires the rt_mutex prior to
1211 * returning to userspace, so as to not leave the rt_mutex with
1212 * waiters and no owner. However, second and third wake-ups
1213 * cannot be predicted as they involve race conditions with the
1214 * first wake and a fault while looking up the pi_state. Both
1215 * pthread_cond_signal() and pthread_cond_broadcast() should
1223 if (pi_state != NULL) {
1225 * We will have to lookup the pi_state again, so free this one
1226 * to keep the accounting correct.
1228 free_pi_state(pi_state);
1232 ret = get_futex_key(uaddr1, fshared, &key1);
1233 if (unlikely(ret != 0))
1235 ret = get_futex_key(uaddr2, fshared, &key2);
1236 if (unlikely(ret != 0))
1239 hb1 = hash_futex(&key1);
1240 hb2 = hash_futex(&key2);
1243 double_lock_hb(hb1, hb2);
1245 if (likely(cmpval != NULL)) {
1248 ret = get_futex_value_locked(&curval, uaddr1);
1250 if (unlikely(ret)) {
1251 double_unlock_hb(hb1, hb2);
1253 ret = get_user(curval, uaddr1);
1260 put_futex_key(fshared, &key2);
1261 put_futex_key(fshared, &key1);
1264 if (curval != *cmpval) {
1270 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1272 * Attempt to acquire uaddr2 and wake the top waiter. If we
1273 * intend to requeue waiters, force setting the FUTEX_WAITERS
1274 * bit. We force this here where we are able to easily handle
1275 * faults rather in the requeue loop below.
1277 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1278 &key2, &pi_state, nr_requeue);
1281 * At this point the top_waiter has either taken uaddr2 or is
1282 * waiting on it. If the former, then the pi_state will not
1283 * exist yet, look it up one more time to ensure we have a
1290 ret = get_futex_value_locked(&curval2, uaddr2);
1292 ret = lookup_pi_state(curval2, hb2, &key2,
1300 double_unlock_hb(hb1, hb2);
1301 put_futex_key(fshared, &key2);
1302 put_futex_key(fshared, &key1);
1303 ret = fault_in_user_writeable(uaddr2);
1308 /* The owner was exiting, try again. */
1309 double_unlock_hb(hb1, hb2);
1310 put_futex_key(fshared, &key2);
1311 put_futex_key(fshared, &key1);
1319 head1 = &hb1->chain;
1320 plist_for_each_entry_safe(this, next, head1, list) {
1321 if (task_count - nr_wake >= nr_requeue)
1324 if (!match_futex(&this->key, &key1))
1328 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1329 * be paired with each other and no other futex ops.
1331 if ((requeue_pi && !this->rt_waiter) ||
1332 (!requeue_pi && this->rt_waiter)) {
1338 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1339 * lock, we already woke the top_waiter. If not, it will be
1340 * woken by futex_unlock_pi().
1342 if (++task_count <= nr_wake && !requeue_pi) {
1347 /* Ensure we requeue to the expected futex for requeue_pi. */
1348 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1354 * Requeue nr_requeue waiters and possibly one more in the case
1355 * of requeue_pi if we couldn't acquire the lock atomically.
1358 /* Prepare the waiter to take the rt_mutex. */
1359 atomic_inc(&pi_state->refcount);
1360 this->pi_state = pi_state;
1361 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1365 /* We got the lock. */
1366 requeue_pi_wake_futex(this, &key2, hb2);
1371 this->pi_state = NULL;
1372 free_pi_state(pi_state);
1376 requeue_futex(this, hb1, hb2, &key2);
1381 double_unlock_hb(hb1, hb2);
1384 * drop_futex_key_refs() must be called outside the spinlocks. During
1385 * the requeue we moved futex_q's from the hash bucket at key1 to the
1386 * one at key2 and updated their key pointer. We no longer need to
1387 * hold the references to key1.
1389 while (--drop_count >= 0)
1390 drop_futex_key_refs(&key1);
1393 put_futex_key(fshared, &key2);
1395 put_futex_key(fshared, &key1);
1397 if (pi_state != NULL)
1398 free_pi_state(pi_state);
1399 return ret ? ret : task_count;
1402 /* The key must be already stored in q->key. */
1403 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1405 struct futex_hash_bucket *hb;
1407 get_futex_key_refs(&q->key);
1408 hb = hash_futex(&q->key);
1409 q->lock_ptr = &hb->lock;
1411 spin_lock(&hb->lock);
1416 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1418 spin_unlock(&hb->lock);
1419 drop_futex_key_refs(&q->key);
1423 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1424 * @q: The futex_q to enqueue
1425 * @hb: The destination hash bucket
1427 * The hb->lock must be held by the caller, and is released here. A call to
1428 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1429 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1430 * or nothing if the unqueue is done as part of the wake process and the unqueue
1431 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1434 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1439 * The priority used to register this element is
1440 * - either the real thread-priority for the real-time threads
1441 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1442 * - or MAX_RT_PRIO for non-RT threads.
1443 * Thus, all RT-threads are woken first in priority order, and
1444 * the others are woken last, in FIFO order.
1446 prio = min(current->normal_prio, MAX_RT_PRIO);
1448 plist_node_init(&q->list, prio);
1449 #ifdef CONFIG_DEBUG_PI_LIST
1450 q->list.plist.lock = &hb->lock;
1452 plist_add(&q->list, &hb->chain);
1454 spin_unlock(&hb->lock);
1458 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1459 * @q: The futex_q to unqueue
1461 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1462 * be paired with exactly one earlier call to queue_me().
1465 * 1 - if the futex_q was still queued (and we removed unqueued it)
1466 * 0 - if the futex_q was already removed by the waking thread
1468 static int unqueue_me(struct futex_q *q)
1470 spinlock_t *lock_ptr;
1473 /* In the common case we don't take the spinlock, which is nice. */
1475 lock_ptr = q->lock_ptr;
1477 if (lock_ptr != NULL) {
1478 spin_lock(lock_ptr);
1480 * q->lock_ptr can change between reading it and
1481 * spin_lock(), causing us to take the wrong lock. This
1482 * corrects the race condition.
1484 * Reasoning goes like this: if we have the wrong lock,
1485 * q->lock_ptr must have changed (maybe several times)
1486 * between reading it and the spin_lock(). It can
1487 * change again after the spin_lock() but only if it was
1488 * already changed before the spin_lock(). It cannot,
1489 * however, change back to the original value. Therefore
1490 * we can detect whether we acquired the correct lock.
1492 if (unlikely(lock_ptr != q->lock_ptr)) {
1493 spin_unlock(lock_ptr);
1496 WARN_ON(plist_node_empty(&q->list));
1497 plist_del(&q->list, &q->list.plist);
1499 BUG_ON(q->pi_state);
1501 spin_unlock(lock_ptr);
1505 drop_futex_key_refs(&q->key);
1510 * PI futexes can not be requeued and must remove themself from the
1511 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1514 static void unqueue_me_pi(struct futex_q *q)
1516 WARN_ON(plist_node_empty(&q->list));
1517 plist_del(&q->list, &q->list.plist);
1519 BUG_ON(!q->pi_state);
1520 free_pi_state(q->pi_state);
1523 spin_unlock(q->lock_ptr);
1525 drop_futex_key_refs(&q->key);
1529 * Fixup the pi_state owner with the new owner.
1531 * Must be called with hash bucket lock held and mm->sem held for non
1534 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1535 struct task_struct *newowner, int fshared)
1537 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1538 struct futex_pi_state *pi_state = q->pi_state;
1539 struct task_struct *oldowner = pi_state->owner;
1540 u32 uval, curval, newval;
1544 if (!pi_state->owner)
1545 newtid |= FUTEX_OWNER_DIED;
1548 * We are here either because we stole the rtmutex from the
1549 * pending owner or we are the pending owner which failed to
1550 * get the rtmutex. We have to replace the pending owner TID
1551 * in the user space variable. This must be atomic as we have
1552 * to preserve the owner died bit here.
1554 * Note: We write the user space value _before_ changing the pi_state
1555 * because we can fault here. Imagine swapped out pages or a fork
1556 * that marked all the anonymous memory readonly for cow.
1558 * Modifying pi_state _before_ the user space value would
1559 * leave the pi_state in an inconsistent state when we fault
1560 * here, because we need to drop the hash bucket lock to
1561 * handle the fault. This might be observed in the PID check
1562 * in lookup_pi_state.
1565 if (get_futex_value_locked(&uval, uaddr))
1569 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1571 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1573 if (curval == -EFAULT)
1581 * We fixed up user space. Now we need to fix the pi_state
1584 if (pi_state->owner != NULL) {
1585 spin_lock_irq(&pi_state->owner->pi_lock);
1586 WARN_ON(list_empty(&pi_state->list));
1587 list_del_init(&pi_state->list);
1588 spin_unlock_irq(&pi_state->owner->pi_lock);
1591 pi_state->owner = newowner;
1593 spin_lock_irq(&newowner->pi_lock);
1594 WARN_ON(!list_empty(&pi_state->list));
1595 list_add(&pi_state->list, &newowner->pi_state_list);
1596 spin_unlock_irq(&newowner->pi_lock);
1600 * To handle the page fault we need to drop the hash bucket
1601 * lock here. That gives the other task (either the pending
1602 * owner itself or the task which stole the rtmutex) the
1603 * chance to try the fixup of the pi_state. So once we are
1604 * back from handling the fault we need to check the pi_state
1605 * after reacquiring the hash bucket lock and before trying to
1606 * do another fixup. When the fixup has been done already we
1610 spin_unlock(q->lock_ptr);
1612 ret = fault_in_user_writeable(uaddr);
1614 spin_lock(q->lock_ptr);
1617 * Check if someone else fixed it for us:
1619 if (pi_state->owner != oldowner)
1629 * In case we must use restart_block to restart a futex_wait,
1630 * we encode in the 'flags' shared capability
1632 #define FLAGS_SHARED 0x01
1633 #define FLAGS_CLOCKRT 0x02
1634 #define FLAGS_HAS_TIMEOUT 0x04
1636 static long futex_wait_restart(struct restart_block *restart);
1639 * fixup_owner() - Post lock pi_state and corner case management
1640 * @uaddr: user address of the futex
1641 * @fshared: whether the futex is shared (1) or not (0)
1642 * @q: futex_q (contains pi_state and access to the rt_mutex)
1643 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1645 * After attempting to lock an rt_mutex, this function is called to cleanup
1646 * the pi_state owner as well as handle race conditions that may allow us to
1647 * acquire the lock. Must be called with the hb lock held.
1650 * 1 - success, lock taken
1651 * 0 - success, lock not taken
1652 * <0 - on error (-EFAULT)
1654 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1657 struct task_struct *owner;
1662 * Got the lock. We might not be the anticipated owner if we
1663 * did a lock-steal - fix up the PI-state in that case:
1665 if (q->pi_state->owner != current)
1666 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1671 * Catch the rare case, where the lock was released when we were on the
1672 * way back before we locked the hash bucket.
1674 if (q->pi_state->owner == current) {
1676 * Try to get the rt_mutex now. This might fail as some other
1677 * task acquired the rt_mutex after we removed ourself from the
1678 * rt_mutex waiters list.
1680 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1686 * pi_state is incorrect, some other task did a lock steal and
1687 * we returned due to timeout or signal without taking the
1688 * rt_mutex. Too late. We can access the rt_mutex_owner without
1689 * locking, as the other task is now blocked on the hash bucket
1690 * lock. Fix the state up.
1692 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1693 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1698 * Paranoia check. If we did not take the lock, then we should not be
1699 * the owner, nor the pending owner, of the rt_mutex.
1701 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1702 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1703 "pi-state %p\n", ret,
1704 q->pi_state->pi_mutex.owner,
1705 q->pi_state->owner);
1708 return ret ? ret : locked;
1712 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1713 * @hb: the futex hash bucket, must be locked by the caller
1714 * @q: the futex_q to queue up on
1715 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1717 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1718 struct hrtimer_sleeper *timeout)
1721 * The task state is guaranteed to be set before another task can
1722 * wake it. set_current_state() is implemented using set_mb() and
1723 * queue_me() calls spin_unlock() upon completion, both serializing
1724 * access to the hash list and forcing another memory barrier.
1726 set_current_state(TASK_INTERRUPTIBLE);
1731 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1732 if (!hrtimer_active(&timeout->timer))
1733 timeout->task = NULL;
1737 * If we have been removed from the hash list, then another task
1738 * has tried to wake us, and we can skip the call to schedule().
1740 if (likely(!plist_node_empty(&q->list))) {
1742 * If the timer has already expired, current will already be
1743 * flagged for rescheduling. Only call schedule if there
1744 * is no timeout, or if it has yet to expire.
1746 if (!timeout || timeout->task)
1749 __set_current_state(TASK_RUNNING);
1753 * futex_wait_setup() - Prepare to wait on a futex
1754 * @uaddr: the futex userspace address
1755 * @val: the expected value
1756 * @fshared: whether the futex is shared (1) or not (0)
1757 * @q: the associated futex_q
1758 * @hb: storage for hash_bucket pointer to be returned to caller
1760 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1761 * compare it with the expected value. Handle atomic faults internally.
1762 * Return with the hb lock held and a q.key reference on success, and unlocked
1763 * with no q.key reference on failure.
1766 * 0 - uaddr contains val and hb has been locked
1767 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1769 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1770 struct futex_q *q, struct futex_hash_bucket **hb)
1776 * Access the page AFTER the hash-bucket is locked.
1777 * Order is important:
1779 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1780 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1782 * The basic logical guarantee of a futex is that it blocks ONLY
1783 * if cond(var) is known to be true at the time of blocking, for
1784 * any cond. If we queued after testing *uaddr, that would open
1785 * a race condition where we could block indefinitely with
1786 * cond(var) false, which would violate the guarantee.
1788 * A consequence is that futex_wait() can return zero and absorb
1789 * a wakeup when *uaddr != val on entry to the syscall. This is
1793 q->key = FUTEX_KEY_INIT;
1794 ret = get_futex_key(uaddr, fshared, &q->key);
1795 if (unlikely(ret != 0))
1799 *hb = queue_lock(q);
1801 ret = get_futex_value_locked(&uval, uaddr);
1804 queue_unlock(q, *hb);
1806 ret = get_user(uval, uaddr);
1813 put_futex_key(fshared, &q->key);
1818 queue_unlock(q, *hb);
1824 put_futex_key(fshared, &q->key);
1828 static int futex_wait(u32 __user *uaddr, int fshared,
1829 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1831 struct hrtimer_sleeper timeout, *to = NULL;
1832 struct restart_block *restart;
1833 struct futex_hash_bucket *hb;
1843 q.requeue_pi_key = NULL;
1848 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1849 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1850 hrtimer_init_sleeper(to, current);
1851 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1852 current->timer_slack_ns);
1856 /* Prepare to wait on uaddr. */
1857 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1861 /* queue_me and wait for wakeup, timeout, or a signal. */
1862 futex_wait_queue_me(hb, &q, to);
1864 /* If we were woken (and unqueued), we succeeded, whatever. */
1866 if (!unqueue_me(&q))
1869 if (to && !to->task)
1873 * We expect signal_pending(current), but we might be the
1874 * victim of a spurious wakeup as well.
1876 if (!signal_pending(current)) {
1877 put_futex_key(fshared, &q.key);
1885 restart = ¤t_thread_info()->restart_block;
1886 restart->fn = futex_wait_restart;
1887 restart->futex.uaddr = (u32 *)uaddr;
1888 restart->futex.val = val;
1889 restart->futex.time = abs_time->tv64;
1890 restart->futex.bitset = bitset;
1891 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1894 restart->futex.flags |= FLAGS_SHARED;
1896 restart->futex.flags |= FLAGS_CLOCKRT;
1898 ret = -ERESTART_RESTARTBLOCK;
1901 put_futex_key(fshared, &q.key);
1904 hrtimer_cancel(&to->timer);
1905 destroy_hrtimer_on_stack(&to->timer);
1911 static long futex_wait_restart(struct restart_block *restart)
1913 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1915 ktime_t t, *tp = NULL;
1917 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1918 t.tv64 = restart->futex.time;
1921 restart->fn = do_no_restart_syscall;
1922 if (restart->futex.flags & FLAGS_SHARED)
1924 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1925 restart->futex.bitset,
1926 restart->futex.flags & FLAGS_CLOCKRT);
1931 * Userspace tried a 0 -> TID atomic transition of the futex value
1932 * and failed. The kernel side here does the whole locking operation:
1933 * if there are waiters then it will block, it does PI, etc. (Due to
1934 * races the kernel might see a 0 value of the futex too.)
1936 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1937 int detect, ktime_t *time, int trylock)
1939 struct hrtimer_sleeper timeout, *to = NULL;
1940 struct futex_hash_bucket *hb;
1944 if (refill_pi_state_cache())
1949 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1951 hrtimer_init_sleeper(to, current);
1952 hrtimer_set_expires(&to->timer, *time);
1957 q.requeue_pi_key = NULL;
1959 q.key = FUTEX_KEY_INIT;
1960 ret = get_futex_key(uaddr, fshared, &q.key);
1961 if (unlikely(ret != 0))
1965 hb = queue_lock(&q);
1967 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1968 if (unlikely(ret)) {
1971 /* We got the lock. */
1973 goto out_unlock_put_key;
1978 * Task is exiting and we just wait for the
1981 queue_unlock(&q, hb);
1982 put_futex_key(fshared, &q.key);
1986 goto out_unlock_put_key;
1991 * Only actually queue now that the atomic ops are done:
1995 WARN_ON(!q.pi_state);
1997 * Block on the PI mutex:
2000 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2002 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2003 /* Fixup the trylock return value: */
2004 ret = ret ? 0 : -EWOULDBLOCK;
2007 spin_lock(q.lock_ptr);
2009 * Fixup the pi_state owner and possibly acquire the lock if we
2012 res = fixup_owner(uaddr, fshared, &q, !ret);
2014 * If fixup_owner() returned an error, proprogate that. If it acquired
2015 * the lock, clear our -ETIMEDOUT or -EINTR.
2018 ret = (res < 0) ? res : 0;
2021 * If fixup_owner() faulted and was unable to handle the fault, unlock
2022 * it and return the fault to userspace.
2024 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2025 rt_mutex_unlock(&q.pi_state->pi_mutex);
2027 /* Unqueue and drop the lock */
2033 queue_unlock(&q, hb);
2036 put_futex_key(fshared, &q.key);
2039 destroy_hrtimer_on_stack(&to->timer);
2040 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2043 queue_unlock(&q, hb);
2045 ret = fault_in_user_writeable(uaddr);
2052 put_futex_key(fshared, &q.key);
2057 * Userspace attempted a TID -> 0 atomic transition, and failed.
2058 * This is the in-kernel slowpath: we look up the PI state (if any),
2059 * and do the rt-mutex unlock.
2061 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2063 struct futex_hash_bucket *hb;
2064 struct futex_q *this, *next;
2066 struct plist_head *head;
2067 union futex_key key = FUTEX_KEY_INIT;
2071 if (get_user(uval, uaddr))
2074 * We release only a lock we actually own:
2076 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2079 ret = get_futex_key(uaddr, fshared, &key);
2080 if (unlikely(ret != 0))
2083 hb = hash_futex(&key);
2084 spin_lock(&hb->lock);
2087 * To avoid races, try to do the TID -> 0 atomic transition
2088 * again. If it succeeds then we can return without waking
2091 if (!(uval & FUTEX_OWNER_DIED))
2092 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2095 if (unlikely(uval == -EFAULT))
2098 * Rare case: we managed to release the lock atomically,
2099 * no need to wake anyone else up:
2101 if (unlikely(uval == task_pid_vnr(current)))
2105 * Ok, other tasks may need to be woken up - check waiters
2106 * and do the wakeup if necessary:
2110 plist_for_each_entry_safe(this, next, head, list) {
2111 if (!match_futex (&this->key, &key))
2113 ret = wake_futex_pi(uaddr, uval, this);
2115 * The atomic access to the futex value
2116 * generated a pagefault, so retry the
2117 * user-access and the wakeup:
2124 * No waiters - kernel unlocks the futex:
2126 if (!(uval & FUTEX_OWNER_DIED)) {
2127 ret = unlock_futex_pi(uaddr, uval);
2133 spin_unlock(&hb->lock);
2134 put_futex_key(fshared, &key);
2140 spin_unlock(&hb->lock);
2141 put_futex_key(fshared, &key);
2143 ret = fault_in_user_writeable(uaddr);
2151 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2152 * @hb: the hash_bucket futex_q was original enqueued on
2153 * @q: the futex_q woken while waiting to be requeued
2154 * @key2: the futex_key of the requeue target futex
2155 * @timeout: the timeout associated with the wait (NULL if none)
2157 * Detect if the task was woken on the initial futex as opposed to the requeue
2158 * target futex. If so, determine if it was a timeout or a signal that caused
2159 * the wakeup and return the appropriate error code to the caller. Must be
2160 * called with the hb lock held.
2163 * 0 - no early wakeup detected
2164 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2167 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2168 struct futex_q *q, union futex_key *key2,
2169 struct hrtimer_sleeper *timeout)
2174 * With the hb lock held, we avoid races while we process the wakeup.
2175 * We only need to hold hb (and not hb2) to ensure atomicity as the
2176 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2177 * It can't be requeued from uaddr2 to something else since we don't
2178 * support a PI aware source futex for requeue.
2180 if (!match_futex(&q->key, key2)) {
2181 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2183 * We were woken prior to requeue by a timeout or a signal.
2184 * Unqueue the futex_q and determine which it was.
2186 plist_del(&q->list, &q->list.plist);
2188 /* Handle spurious wakeups gracefully */
2190 if (timeout && !timeout->task)
2192 else if (signal_pending(current))
2193 ret = -ERESTARTNOINTR;
2199 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2200 * @uaddr: the futex we initially wait on (non-pi)
2201 * @fshared: whether the futexes are shared (1) or not (0). They must be
2202 * the same type, no requeueing from private to shared, etc.
2203 * @val: the expected value of uaddr
2204 * @abs_time: absolute timeout
2205 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2206 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2207 * @uaddr2: the pi futex we will take prior to returning to user-space
2209 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2210 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2211 * complete the acquisition of the rt_mutex prior to returning to userspace.
2212 * This ensures the rt_mutex maintains an owner when it has waiters; without
2213 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2216 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2217 * via the following:
2218 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2219 * 2) wakeup on uaddr2 after a requeue
2223 * If 3, cleanup and return -ERESTARTNOINTR.
2225 * If 2, we may then block on trying to take the rt_mutex and return via:
2226 * 5) successful lock
2229 * 8) other lock acquisition failure
2231 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2233 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2239 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2240 u32 val, ktime_t *abs_time, u32 bitset,
2241 int clockrt, u32 __user *uaddr2)
2243 struct hrtimer_sleeper timeout, *to = NULL;
2244 struct rt_mutex_waiter rt_waiter;
2245 struct rt_mutex *pi_mutex = NULL;
2246 struct futex_hash_bucket *hb;
2247 union futex_key key2;
2256 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2257 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2258 hrtimer_init_sleeper(to, current);
2259 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2260 current->timer_slack_ns);
2264 * The waiter is allocated on our stack, manipulated by the requeue
2265 * code while we sleep on uaddr.
2267 debug_rt_mutex_init_waiter(&rt_waiter);
2268 rt_waiter.task = NULL;
2270 key2 = FUTEX_KEY_INIT;
2271 ret = get_futex_key(uaddr2, fshared, &key2);
2272 if (unlikely(ret != 0))
2277 q.rt_waiter = &rt_waiter;
2278 q.requeue_pi_key = &key2;
2280 /* Prepare to wait on uaddr. */
2281 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2285 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2286 futex_wait_queue_me(hb, &q, to);
2288 spin_lock(&hb->lock);
2289 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2290 spin_unlock(&hb->lock);
2295 * In order for us to be here, we know our q.key == key2, and since
2296 * we took the hb->lock above, we also know that futex_requeue() has
2297 * completed and we no longer have to concern ourselves with a wakeup
2298 * race with the atomic proxy lock acquition by the requeue code.
2301 /* Check if the requeue code acquired the second futex for us. */
2304 * Got the lock. We might not be the anticipated owner if we
2305 * did a lock-steal - fix up the PI-state in that case.
2307 if (q.pi_state && (q.pi_state->owner != current)) {
2308 spin_lock(q.lock_ptr);
2309 ret = fixup_pi_state_owner(uaddr2, &q, current,
2311 spin_unlock(q.lock_ptr);
2315 * We have been woken up by futex_unlock_pi(), a timeout, or a
2316 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2319 WARN_ON(!&q.pi_state);
2320 pi_mutex = &q.pi_state->pi_mutex;
2321 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2322 debug_rt_mutex_free_waiter(&rt_waiter);
2324 spin_lock(q.lock_ptr);
2326 * Fixup the pi_state owner and possibly acquire the lock if we
2329 res = fixup_owner(uaddr2, fshared, &q, !ret);
2331 * If fixup_owner() returned an error, proprogate that. If it
2332 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2335 ret = (res < 0) ? res : 0;
2337 /* Unqueue and drop the lock. */
2342 * If fixup_pi_state_owner() faulted and was unable to handle the
2343 * fault, unlock the rt_mutex and return the fault to userspace.
2345 if (ret == -EFAULT) {
2346 if (rt_mutex_owner(pi_mutex) == current)
2347 rt_mutex_unlock(pi_mutex);
2348 } else if (ret == -EINTR) {
2350 * We've already been requeued, but cannot restart by calling
2351 * futex_lock_pi() directly. We could restart this syscall, but
2352 * it would detect that the user space "val" changed and return
2353 * -EWOULDBLOCK. Save the overhead of the restart and return
2354 * -EWOULDBLOCK directly.
2360 put_futex_key(fshared, &q.key);
2362 put_futex_key(fshared, &key2);
2366 hrtimer_cancel(&to->timer);
2367 destroy_hrtimer_on_stack(&to->timer);
2373 * Support for robust futexes: the kernel cleans up held futexes at
2376 * Implementation: user-space maintains a per-thread list of locks it
2377 * is holding. Upon do_exit(), the kernel carefully walks this list,
2378 * and marks all locks that are owned by this thread with the
2379 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2380 * always manipulated with the lock held, so the list is private and
2381 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2382 * field, to allow the kernel to clean up if the thread dies after
2383 * acquiring the lock, but just before it could have added itself to
2384 * the list. There can only be one such pending lock.
2388 * sys_set_robust_list() - Set the robust-futex list head of a task
2389 * @head: pointer to the list-head
2390 * @len: length of the list-head, as userspace expects
2392 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2395 if (!futex_cmpxchg_enabled)
2398 * The kernel knows only one size for now:
2400 if (unlikely(len != sizeof(*head)))
2403 current->robust_list = head;
2409 * sys_get_robust_list() - Get the robust-futex list head of a task
2410 * @pid: pid of the process [zero for current task]
2411 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2412 * @len_ptr: pointer to a length field, the kernel fills in the header size
2414 SYSCALL_DEFINE3(get_robust_list, int, pid,
2415 struct robust_list_head __user * __user *, head_ptr,
2416 size_t __user *, len_ptr)
2418 struct robust_list_head __user *head;
2420 const struct cred *cred = current_cred(), *pcred;
2422 if (!futex_cmpxchg_enabled)
2426 head = current->robust_list;
2428 struct task_struct *p;
2432 p = find_task_by_vpid(pid);
2436 pcred = __task_cred(p);
2437 if (cred->euid != pcred->euid &&
2438 cred->euid != pcred->uid &&
2439 !capable(CAP_SYS_PTRACE))
2441 head = p->robust_list;
2445 if (put_user(sizeof(*head), len_ptr))
2447 return put_user(head, head_ptr);
2456 * Process a futex-list entry, check whether it's owned by the
2457 * dying task, and do notification if so:
2459 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2461 u32 uval, nval, mval;
2464 if (get_user(uval, uaddr))
2467 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2469 * Ok, this dying thread is truly holding a futex
2470 * of interest. Set the OWNER_DIED bit atomically
2471 * via cmpxchg, and if the value had FUTEX_WAITERS
2472 * set, wake up a waiter (if any). (We have to do a
2473 * futex_wake() even if OWNER_DIED is already set -
2474 * to handle the rare but possible case of recursive
2475 * thread-death.) The rest of the cleanup is done in
2478 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2479 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2481 if (nval == -EFAULT)
2488 * Wake robust non-PI futexes here. The wakeup of
2489 * PI futexes happens in exit_pi_state():
2491 if (!pi && (uval & FUTEX_WAITERS))
2492 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2498 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2500 static inline int fetch_robust_entry(struct robust_list __user **entry,
2501 struct robust_list __user * __user *head,
2504 unsigned long uentry;
2506 if (get_user(uentry, (unsigned long __user *)head))
2509 *entry = (void __user *)(uentry & ~1UL);
2516 * Walk curr->robust_list (very carefully, it's a userspace list!)
2517 * and mark any locks found there dead, and notify any waiters.
2519 * We silently return on any sign of list-walking problem.
2521 void exit_robust_list(struct task_struct *curr)
2523 struct robust_list_head __user *head = curr->robust_list;
2524 struct robust_list __user *entry, *next_entry, *pending;
2525 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2526 unsigned long futex_offset;
2529 if (!futex_cmpxchg_enabled)
2533 * Fetch the list head (which was registered earlier, via
2534 * sys_set_robust_list()):
2536 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2539 * Fetch the relative futex offset:
2541 if (get_user(futex_offset, &head->futex_offset))
2544 * Fetch any possibly pending lock-add first, and handle it
2547 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2550 next_entry = NULL; /* avoid warning with gcc */
2551 while (entry != &head->list) {
2553 * Fetch the next entry in the list before calling
2554 * handle_futex_death:
2556 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2558 * A pending lock might already be on the list, so
2559 * don't process it twice:
2561 if (entry != pending)
2562 if (handle_futex_death((void __user *)entry + futex_offset,
2570 * Avoid excessively long or circular lists:
2579 handle_futex_death((void __user *)pending + futex_offset,
2583 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2584 u32 __user *uaddr2, u32 val2, u32 val3)
2586 int clockrt, ret = -ENOSYS;
2587 int cmd = op & FUTEX_CMD_MASK;
2590 if (!(op & FUTEX_PRIVATE_FLAG))
2593 clockrt = op & FUTEX_CLOCK_REALTIME;
2594 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2599 val3 = FUTEX_BITSET_MATCH_ANY;
2600 case FUTEX_WAIT_BITSET:
2601 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2604 val3 = FUTEX_BITSET_MATCH_ANY;
2605 case FUTEX_WAKE_BITSET:
2606 ret = futex_wake(uaddr, fshared, val, val3);
2609 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2611 case FUTEX_CMP_REQUEUE:
2612 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2616 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2619 if (futex_cmpxchg_enabled)
2620 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2622 case FUTEX_UNLOCK_PI:
2623 if (futex_cmpxchg_enabled)
2624 ret = futex_unlock_pi(uaddr, fshared);
2626 case FUTEX_TRYLOCK_PI:
2627 if (futex_cmpxchg_enabled)
2628 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2630 case FUTEX_WAIT_REQUEUE_PI:
2631 val3 = FUTEX_BITSET_MATCH_ANY;
2632 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2635 case FUTEX_CMP_REQUEUE_PI:
2636 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2646 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2647 struct timespec __user *, utime, u32 __user *, uaddr2,
2651 ktime_t t, *tp = NULL;
2653 int cmd = op & FUTEX_CMD_MASK;
2655 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2656 cmd == FUTEX_WAIT_BITSET ||
2657 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2658 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2660 if (!timespec_valid(&ts))
2663 t = timespec_to_ktime(ts);
2664 if (cmd == FUTEX_WAIT)
2665 t = ktime_add_safe(ktime_get(), t);
2669 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2670 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2672 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2673 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2674 val2 = (u32) (unsigned long) utime;
2676 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2679 static int __init futex_init(void)
2685 * This will fail and we want it. Some arch implementations do
2686 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2687 * functionality. We want to know that before we call in any
2688 * of the complex code paths. Also we want to prevent
2689 * registration of robust lists in that case. NULL is
2690 * guaranteed to fault and we get -EFAULT on functional
2691 * implementation, the non functional ones will return
2694 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2695 if (curval == -EFAULT)
2696 futex_cmpxchg_enabled = 1;
2698 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2699 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2700 spin_lock_init(&futex_queues[i].lock);
2705 __initcall(futex_init);