2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
63 #include <asm/futex.h>
65 #include "rtmutex_common.h"
67 int __read_mostly futex_cmpxchg_enabled;
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
72 * Priority Inheritance state:
74 struct futex_pi_state {
76 * list of 'owned' pi_state instances - these have to be
77 * cleaned up in do_exit() if the task exits prematurely:
79 struct list_head list;
84 struct rt_mutex pi_mutex;
86 struct task_struct *owner;
93 * struct futex_q - The hashed futex queue entry, one per waiting task
94 * @task: the task waiting on the futex
95 * @lock_ptr: the hash bucket lock
96 * @key: the key the futex is hashed on
97 * @pi_state: optional priority inheritance state
98 * @rt_waiter: rt_waiter storage for use with requeue_pi
99 * @requeue_pi_key: the requeue_pi target futex key
100 * @bitset: bitset for the optional bitmasked wakeup
102 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103 * we can wake only the relevant ones (hashed queues may be shared).
105 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107 * The order of wakup is always to make the first condition true, then
110 * PI futexes are typically woken before they are removed from the hash list via
111 * the rt_mutex code. See unqueue_me_pi().
114 struct plist_node list;
116 struct task_struct *task;
117 spinlock_t *lock_ptr;
119 struct futex_pi_state *pi_state;
120 struct rt_mutex_waiter *rt_waiter;
121 union futex_key *requeue_pi_key;
126 * Hash buckets are shared by all the futex_keys that hash to the same
127 * location. Each key may have multiple futex_q structures, one for each task
128 * waiting on a futex.
130 struct futex_hash_bucket {
132 struct plist_head chain;
135 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
138 * We hash on the keys returned from get_futex_key (see below).
140 static struct futex_hash_bucket *hash_futex(union futex_key *key)
142 u32 hash = jhash2((u32*)&key->both.word,
143 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
145 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
149 * Return 1 if two futex_keys are equal, 0 otherwise.
151 static inline int match_futex(union futex_key *key1, union futex_key *key2)
154 && key1->both.word == key2->both.word
155 && key1->both.ptr == key2->both.ptr
156 && key1->both.offset == key2->both.offset);
160 * Take a reference to the resource addressed by a key.
161 * Can be called while holding spinlocks.
164 static void get_futex_key_refs(union futex_key *key)
169 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
171 atomic_inc(&key->shared.inode->i_count);
173 case FUT_OFF_MMSHARED:
174 atomic_inc(&key->private.mm->mm_count);
180 * Drop a reference to the resource addressed by a key.
181 * The hash bucket spinlock must not be held.
183 static void drop_futex_key_refs(union futex_key *key)
185 if (!key->both.ptr) {
186 /* If we're here then we tried to put a key we failed to get */
191 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
193 iput(key->shared.inode);
195 case FUT_OFF_MMSHARED:
196 mmdrop(key->private.mm);
202 * get_futex_key() - Get parameters which are the keys for a futex
203 * @uaddr: virtual address of the futex
204 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205 * @key: address where result is stored.
206 * @rw: mapping needs to be read/write (values: VERIFY_READ,
209 * Returns a negative error code or 0
210 * The key words are stored in *key on success.
212 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
213 * offset_within_page). For private mappings, it's (uaddr, current->mm).
214 * We can usually work out the index without swapping in the page.
216 * lock_page() might sleep, the caller should not hold a spinlock.
219 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
221 unsigned long address = (unsigned long)uaddr;
222 struct mm_struct *mm = current->mm;
227 * The futex address must be "naturally" aligned.
229 key->both.offset = address % PAGE_SIZE;
230 if (unlikely((address % sizeof(u32)) != 0))
232 address -= key->both.offset;
235 * PROCESS_PRIVATE futexes are fast.
236 * As the mm cannot disappear under us and the 'key' only needs
237 * virtual address, we dont even have to find the underlying vma.
238 * Note : We do have to check 'uaddr' is a valid user address,
239 * but access_ok() should be faster than find_vma()
242 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
244 key->private.mm = mm;
245 key->private.address = address;
246 get_futex_key_refs(key);
251 err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
255 page = compound_head(page);
257 if (!page->mapping) {
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.
270 if (PageAnon(page)) {
271 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
272 key->private.mm = mm;
273 key->private.address = address;
275 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
276 key->shared.inode = page->mapping->host;
277 key->shared.pgoff = page->index;
280 get_futex_key_refs(key);
288 void put_futex_key(int fshared, union futex_key *key)
290 drop_futex_key_refs(key);
294 * fault_in_user_writeable() - Fault in user address and verify RW access
295 * @uaddr: pointer to faulting user space address
297 * Slow path to fixup the fault we just took in the atomic write
300 * We have no generic implementation of a non destructive write to the
301 * user address. We know that we faulted in the atomic pagefault
302 * disabled section so we can as well avoid the #PF overhead by
303 * calling get_user_pages() right away.
305 static int fault_in_user_writeable(u32 __user *uaddr)
307 int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
308 1, 1, 0, NULL, NULL);
309 return ret < 0 ? ret : 0;
313 * futex_top_waiter() - Return the highest priority waiter on a futex
314 * @hb: the hash bucket the futex_q's reside in
315 * @key: the futex key (to distinguish it from other futex futex_q's)
317 * Must be called with the hb lock held.
319 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
320 union futex_key *key)
322 struct futex_q *this;
324 plist_for_each_entry(this, &hb->chain, list) {
325 if (match_futex(&this->key, key))
331 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
336 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
342 static int get_futex_value_locked(u32 *dest, u32 __user *from)
347 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
350 return ret ? -EFAULT : 0;
357 static int refill_pi_state_cache(void)
359 struct futex_pi_state *pi_state;
361 if (likely(current->pi_state_cache))
364 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
369 INIT_LIST_HEAD(&pi_state->list);
370 /* pi_mutex gets initialized later */
371 pi_state->owner = NULL;
372 atomic_set(&pi_state->refcount, 1);
373 pi_state->key = FUTEX_KEY_INIT;
375 current->pi_state_cache = pi_state;
380 static struct futex_pi_state * alloc_pi_state(void)
382 struct futex_pi_state *pi_state = current->pi_state_cache;
385 current->pi_state_cache = NULL;
390 static void free_pi_state(struct futex_pi_state *pi_state)
392 if (!atomic_dec_and_test(&pi_state->refcount))
396 * If pi_state->owner is NULL, the owner is most probably dying
397 * and has cleaned up the pi_state already
399 if (pi_state->owner) {
400 spin_lock_irq(&pi_state->owner->pi_lock);
401 list_del_init(&pi_state->list);
402 spin_unlock_irq(&pi_state->owner->pi_lock);
404 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
407 if (current->pi_state_cache)
411 * pi_state->list is already empty.
412 * clear pi_state->owner.
413 * refcount is at 0 - put it back to 1.
415 pi_state->owner = NULL;
416 atomic_set(&pi_state->refcount, 1);
417 current->pi_state_cache = pi_state;
422 * Look up the task based on what TID userspace gave us.
425 static struct task_struct * futex_find_get_task(pid_t pid)
427 struct task_struct *p;
428 const struct cred *cred = current_cred(), *pcred;
431 p = find_task_by_vpid(pid);
435 pcred = __task_cred(p);
436 if (cred->euid != pcred->euid &&
437 cred->euid != pcred->uid)
449 * This task is holding PI mutexes at exit time => bad.
450 * Kernel cleans up PI-state, but userspace is likely hosed.
451 * (Robust-futex cleanup is separate and might save the day for userspace.)
453 void exit_pi_state_list(struct task_struct *curr)
455 struct list_head *next, *head = &curr->pi_state_list;
456 struct futex_pi_state *pi_state;
457 struct futex_hash_bucket *hb;
458 union futex_key key = FUTEX_KEY_INIT;
460 if (!futex_cmpxchg_enabled)
463 * We are a ZOMBIE and nobody can enqueue itself on
464 * pi_state_list anymore, but we have to be careful
465 * versus waiters unqueueing themselves:
467 spin_lock_irq(&curr->pi_lock);
468 while (!list_empty(head)) {
471 pi_state = list_entry(next, struct futex_pi_state, list);
473 hb = hash_futex(&key);
474 spin_unlock_irq(&curr->pi_lock);
476 spin_lock(&hb->lock);
478 spin_lock_irq(&curr->pi_lock);
480 * We dropped the pi-lock, so re-check whether this
481 * task still owns the PI-state:
483 if (head->next != next) {
484 spin_unlock(&hb->lock);
488 WARN_ON(pi_state->owner != curr);
489 WARN_ON(list_empty(&pi_state->list));
490 list_del_init(&pi_state->list);
491 pi_state->owner = NULL;
492 spin_unlock_irq(&curr->pi_lock);
494 rt_mutex_unlock(&pi_state->pi_mutex);
496 spin_unlock(&hb->lock);
498 spin_lock_irq(&curr->pi_lock);
500 spin_unlock_irq(&curr->pi_lock);
504 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
505 union futex_key *key, struct futex_pi_state **ps)
507 struct futex_pi_state *pi_state = NULL;
508 struct futex_q *this, *next;
509 struct plist_head *head;
510 struct task_struct *p;
511 pid_t pid = uval & FUTEX_TID_MASK;
515 plist_for_each_entry_safe(this, next, head, list) {
516 if (match_futex(&this->key, key)) {
518 * Another waiter already exists - bump up
519 * the refcount and return its pi_state:
521 pi_state = this->pi_state;
523 * Userspace might have messed up non PI and PI futexes
525 if (unlikely(!pi_state))
528 WARN_ON(!atomic_read(&pi_state->refcount));
529 WARN_ON(pid && pi_state->owner &&
530 pi_state->owner->pid != pid);
532 atomic_inc(&pi_state->refcount);
540 * We are the first waiter - try to look up the real owner and attach
541 * the new pi_state to it, but bail out when TID = 0
545 p = futex_find_get_task(pid);
550 * We need to look at the task state flags to figure out,
551 * whether the task is exiting. To protect against the do_exit
552 * change of the task flags, we do this protected by
555 spin_lock_irq(&p->pi_lock);
556 if (unlikely(p->flags & PF_EXITING)) {
558 * The task is on the way out. When PF_EXITPIDONE is
559 * set, we know that the task has finished the
562 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
564 spin_unlock_irq(&p->pi_lock);
569 pi_state = alloc_pi_state();
572 * Initialize the pi_mutex in locked state and make 'p'
575 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
577 /* Store the key for possible exit cleanups: */
578 pi_state->key = *key;
580 WARN_ON(!list_empty(&pi_state->list));
581 list_add(&pi_state->list, &p->pi_state_list);
583 spin_unlock_irq(&p->pi_lock);
593 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
594 * @uaddr: the pi futex user address
595 * @hb: the pi futex hash bucket
596 * @key: the futex key associated with uaddr and hb
597 * @ps: the pi_state pointer where we store the result of the
599 * @task: the task to perform the atomic lock work for. This will
600 * be "current" except in the case of requeue pi.
601 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
605 * 1 - acquired the lock
608 * The hb->lock and futex_key refs shall be held by the caller.
610 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
611 union futex_key *key,
612 struct futex_pi_state **ps,
613 struct task_struct *task, int set_waiters)
615 int lock_taken, ret, ownerdied = 0;
616 u32 uval, newval, curval;
619 ret = lock_taken = 0;
622 * To avoid races, we attempt to take the lock here again
623 * (by doing a 0 -> TID atomic cmpxchg), while holding all
624 * the locks. It will most likely not succeed.
626 newval = task_pid_vnr(task);
628 newval |= FUTEX_WAITERS;
630 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
632 if (unlikely(curval == -EFAULT))
638 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
642 * Surprise - we got the lock. Just return to userspace:
644 if (unlikely(!curval))
650 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
651 * to wake at the next unlock.
653 newval = curval | FUTEX_WAITERS;
656 * There are two cases, where a futex might have no owner (the
657 * owner TID is 0): OWNER_DIED. We take over the futex in this
658 * case. We also do an unconditional take over, when the owner
661 * This is safe as we are protected by the hash bucket lock !
663 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
664 /* Keep the OWNER_DIED bit */
665 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
670 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
672 if (unlikely(curval == -EFAULT))
674 if (unlikely(curval != uval))
678 * We took the lock due to owner died take over.
680 if (unlikely(lock_taken))
684 * We dont have the lock. Look up the PI state (or create it if
685 * we are the first waiter):
687 ret = lookup_pi_state(uval, hb, key, ps);
693 * No owner found for this futex. Check if the
694 * OWNER_DIED bit is set to figure out whether
695 * this is a robust futex or not.
697 if (get_futex_value_locked(&curval, uaddr))
701 * We simply start over in case of a robust
702 * futex. The code above will take the futex
705 if (curval & FUTEX_OWNER_DIED) {
718 * The hash bucket lock must be held when this is called.
719 * Afterwards, the futex_q must not be accessed.
721 static void wake_futex(struct futex_q *q)
723 struct task_struct *p = q->task;
726 * We set q->lock_ptr = NULL _before_ we wake up the task. If
727 * a non futex wake up happens on another CPU then the task
728 * might exit and p would dereference a non existing task
729 * struct. Prevent this by holding a reference on p across the
734 plist_del(&q->list, &q->list.plist);
736 * The waiting task can free the futex_q as soon as
737 * q->lock_ptr = NULL is written, without taking any locks. A
738 * memory barrier is required here to prevent the following
739 * store to lock_ptr from getting ahead of the plist_del.
744 wake_up_state(p, TASK_NORMAL);
748 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
750 struct task_struct *new_owner;
751 struct futex_pi_state *pi_state = this->pi_state;
757 spin_lock(&pi_state->pi_mutex.wait_lock);
758 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
761 * This happens when we have stolen the lock and the original
762 * pending owner did not enqueue itself back on the rt_mutex.
763 * Thats not a tragedy. We know that way, that a lock waiter
764 * is on the fly. We make the futex_q waiter the pending owner.
767 new_owner = this->task;
770 * We pass it to the next owner. (The WAITERS bit is always
771 * kept enabled while there is PI state around. We must also
772 * preserve the owner died bit.)
774 if (!(uval & FUTEX_OWNER_DIED)) {
777 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
779 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
781 if (curval == -EFAULT)
783 else if (curval != uval)
786 spin_unlock(&pi_state->pi_mutex.wait_lock);
791 spin_lock_irq(&pi_state->owner->pi_lock);
792 WARN_ON(list_empty(&pi_state->list));
793 list_del_init(&pi_state->list);
794 spin_unlock_irq(&pi_state->owner->pi_lock);
796 spin_lock_irq(&new_owner->pi_lock);
797 WARN_ON(!list_empty(&pi_state->list));
798 list_add(&pi_state->list, &new_owner->pi_state_list);
799 pi_state->owner = new_owner;
800 spin_unlock_irq(&new_owner->pi_lock);
802 spin_unlock(&pi_state->pi_mutex.wait_lock);
803 rt_mutex_unlock(&pi_state->pi_mutex);
808 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
813 * There is no waiter, so we unlock the futex. The owner died
814 * bit has not to be preserved here. We are the owner:
816 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
818 if (oldval == -EFAULT)
827 * Express the locking dependencies for lockdep:
830 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
833 spin_lock(&hb1->lock);
835 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
836 } else { /* hb1 > hb2 */
837 spin_lock(&hb2->lock);
838 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
843 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
845 spin_unlock(&hb1->lock);
847 spin_unlock(&hb2->lock);
851 * Wake up waiters matching bitset queued on this futex (uaddr).
853 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
855 struct futex_hash_bucket *hb;
856 struct futex_q *this, *next;
857 struct plist_head *head;
858 union futex_key key = FUTEX_KEY_INIT;
864 ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
865 if (unlikely(ret != 0))
868 hb = hash_futex(&key);
869 spin_lock(&hb->lock);
872 plist_for_each_entry_safe(this, next, head, list) {
873 if (match_futex (&this->key, &key)) {
874 if (this->pi_state || this->rt_waiter) {
879 /* Check if one of the bits is set in both bitsets */
880 if (!(this->bitset & bitset))
884 if (++ret >= nr_wake)
889 spin_unlock(&hb->lock);
890 put_futex_key(fshared, &key);
896 * Wake up all waiters hashed on the physical page that is mapped
897 * to this virtual address:
900 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
901 int nr_wake, int nr_wake2, int op)
903 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
904 struct futex_hash_bucket *hb1, *hb2;
905 struct plist_head *head;
906 struct futex_q *this, *next;
910 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
911 if (unlikely(ret != 0))
913 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
914 if (unlikely(ret != 0))
917 hb1 = hash_futex(&key1);
918 hb2 = hash_futex(&key2);
921 double_lock_hb(hb1, hb2);
922 op_ret = futex_atomic_op_inuser(op, uaddr2);
923 if (unlikely(op_ret < 0)) {
925 double_unlock_hb(hb1, hb2);
929 * we don't get EFAULT from MMU faults if we don't have an MMU,
930 * but we might get them from range checking
936 if (unlikely(op_ret != -EFAULT)) {
941 ret = fault_in_user_writeable(uaddr2);
948 put_futex_key(fshared, &key2);
949 put_futex_key(fshared, &key1);
955 plist_for_each_entry_safe(this, next, head, list) {
956 if (match_futex (&this->key, &key1)) {
958 if (++ret >= nr_wake)
967 plist_for_each_entry_safe(this, next, head, list) {
968 if (match_futex (&this->key, &key2)) {
970 if (++op_ret >= nr_wake2)
977 double_unlock_hb(hb1, hb2);
979 put_futex_key(fshared, &key2);
981 put_futex_key(fshared, &key1);
987 * requeue_futex() - Requeue a futex_q from one hb to another
988 * @q: the futex_q to requeue
989 * @hb1: the source hash_bucket
990 * @hb2: the target hash_bucket
991 * @key2: the new key for the requeued futex_q
994 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
995 struct futex_hash_bucket *hb2, union futex_key *key2)
999 * If key1 and key2 hash to the same bucket, no need to
1002 if (likely(&hb1->chain != &hb2->chain)) {
1003 plist_del(&q->list, &hb1->chain);
1004 plist_add(&q->list, &hb2->chain);
1005 q->lock_ptr = &hb2->lock;
1006 #ifdef CONFIG_DEBUG_PI_LIST
1007 q->list.plist.lock = &hb2->lock;
1010 get_futex_key_refs(key2);
1015 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1017 * @key: the key of the requeue target futex
1018 * @hb: the hash_bucket of the requeue target futex
1020 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1021 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1022 * to the requeue target futex so the waiter can detect the wakeup on the right
1023 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1024 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1025 * to protect access to the pi_state to fixup the owner later. Must be called
1026 * with both q->lock_ptr and hb->lock held.
1029 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1030 struct futex_hash_bucket *hb)
1032 drop_futex_key_refs(&q->key);
1033 get_futex_key_refs(key);
1036 WARN_ON(plist_node_empty(&q->list));
1037 plist_del(&q->list, &q->list.plist);
1039 WARN_ON(!q->rt_waiter);
1040 q->rt_waiter = NULL;
1042 q->lock_ptr = &hb->lock;
1043 #ifdef CONFIG_DEBUG_PI_LIST
1044 q->list.plist.lock = &hb->lock;
1047 wake_up_state(q->task, TASK_NORMAL);
1051 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1052 * @pifutex: the user address of the to futex
1053 * @hb1: the from futex hash bucket, must be locked by the caller
1054 * @hb2: the to futex hash bucket, must be locked by the caller
1055 * @key1: the from futex key
1056 * @key2: the to futex key
1057 * @ps: address to store the pi_state pointer
1058 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1060 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1061 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1062 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1063 * hb1 and hb2 must be held by the caller.
1066 * 0 - failed to acquire the lock atomicly
1067 * 1 - acquired the lock
1070 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1071 struct futex_hash_bucket *hb1,
1072 struct futex_hash_bucket *hb2,
1073 union futex_key *key1, union futex_key *key2,
1074 struct futex_pi_state **ps, int set_waiters)
1076 struct futex_q *top_waiter = NULL;
1080 if (get_futex_value_locked(&curval, pifutex))
1084 * Find the top_waiter and determine if there are additional waiters.
1085 * If the caller intends to requeue more than 1 waiter to pifutex,
1086 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1087 * as we have means to handle the possible fault. If not, don't set
1088 * the bit unecessarily as it will force the subsequent unlock to enter
1091 top_waiter = futex_top_waiter(hb1, key1);
1093 /* There are no waiters, nothing for us to do. */
1097 /* Ensure we requeue to the expected futex. */
1098 if (!match_futex(top_waiter->requeue_pi_key, key2))
1102 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1103 * the contended case or if set_waiters is 1. The pi_state is returned
1104 * in ps in contended cases.
1106 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1109 requeue_pi_wake_futex(top_waiter, key2, hb2);
1115 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1116 * uaddr1: source futex user address
1117 * uaddr2: target futex user address
1118 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1119 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1120 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1121 * pi futex (pi to pi requeue is not supported)
1123 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1124 * uaddr2 atomically on behalf of the top waiter.
1127 * >=0 - on success, the number of tasks requeued or woken
1130 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1131 int nr_wake, int nr_requeue, u32 *cmpval,
1134 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1135 int drop_count = 0, task_count = 0, ret;
1136 struct futex_pi_state *pi_state = NULL;
1137 struct futex_hash_bucket *hb1, *hb2;
1138 struct plist_head *head1;
1139 struct futex_q *this, *next;
1144 * requeue_pi requires a pi_state, try to allocate it now
1145 * without any locks in case it fails.
1147 if (refill_pi_state_cache())
1150 * requeue_pi must wake as many tasks as it can, up to nr_wake
1151 * + nr_requeue, since it acquires the rt_mutex prior to
1152 * returning to userspace, so as to not leave the rt_mutex with
1153 * waiters and no owner. However, second and third wake-ups
1154 * cannot be predicted as they involve race conditions with the
1155 * first wake and a fault while looking up the pi_state. Both
1156 * pthread_cond_signal() and pthread_cond_broadcast() should
1164 if (pi_state != NULL) {
1166 * We will have to lookup the pi_state again, so free this one
1167 * to keep the accounting correct.
1169 free_pi_state(pi_state);
1173 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1174 if (unlikely(ret != 0))
1176 ret = get_futex_key(uaddr2, fshared, &key2,
1177 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1178 if (unlikely(ret != 0))
1181 hb1 = hash_futex(&key1);
1182 hb2 = hash_futex(&key2);
1185 double_lock_hb(hb1, hb2);
1187 if (likely(cmpval != NULL)) {
1190 ret = get_futex_value_locked(&curval, uaddr1);
1192 if (unlikely(ret)) {
1193 double_unlock_hb(hb1, hb2);
1195 ret = get_user(curval, uaddr1);
1202 put_futex_key(fshared, &key2);
1203 put_futex_key(fshared, &key1);
1206 if (curval != *cmpval) {
1212 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1214 * Attempt to acquire uaddr2 and wake the top waiter. If we
1215 * intend to requeue waiters, force setting the FUTEX_WAITERS
1216 * bit. We force this here where we are able to easily handle
1217 * faults rather in the requeue loop below.
1219 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1220 &key2, &pi_state, nr_requeue);
1223 * At this point the top_waiter has either taken uaddr2 or is
1224 * waiting on it. If the former, then the pi_state will not
1225 * exist yet, look it up one more time to ensure we have a
1231 ret = get_futex_value_locked(&curval2, uaddr2);
1233 ret = lookup_pi_state(curval2, hb2, &key2,
1241 double_unlock_hb(hb1, hb2);
1242 put_futex_key(fshared, &key2);
1243 put_futex_key(fshared, &key1);
1244 ret = fault_in_user_writeable(uaddr2);
1249 /* The owner was exiting, try again. */
1250 double_unlock_hb(hb1, hb2);
1251 put_futex_key(fshared, &key2);
1252 put_futex_key(fshared, &key1);
1260 head1 = &hb1->chain;
1261 plist_for_each_entry_safe(this, next, head1, list) {
1262 if (task_count - nr_wake >= nr_requeue)
1265 if (!match_futex(&this->key, &key1))
1269 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1270 * be paired with each other and no other futex ops.
1272 if ((requeue_pi && !this->rt_waiter) ||
1273 (!requeue_pi && this->rt_waiter)) {
1279 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1280 * lock, we already woke the top_waiter. If not, it will be
1281 * woken by futex_unlock_pi().
1283 if (++task_count <= nr_wake && !requeue_pi) {
1288 /* Ensure we requeue to the expected futex for requeue_pi. */
1289 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1295 * Requeue nr_requeue waiters and possibly one more in the case
1296 * of requeue_pi if we couldn't acquire the lock atomically.
1299 /* Prepare the waiter to take the rt_mutex. */
1300 atomic_inc(&pi_state->refcount);
1301 this->pi_state = pi_state;
1302 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1306 /* We got the lock. */
1307 requeue_pi_wake_futex(this, &key2, hb2);
1311 this->pi_state = NULL;
1312 free_pi_state(pi_state);
1316 requeue_futex(this, hb1, hb2, &key2);
1321 double_unlock_hb(hb1, hb2);
1324 * drop_futex_key_refs() must be called outside the spinlocks. During
1325 * the requeue we moved futex_q's from the hash bucket at key1 to the
1326 * one at key2 and updated their key pointer. We no longer need to
1327 * hold the references to key1.
1329 while (--drop_count >= 0)
1330 drop_futex_key_refs(&key1);
1333 put_futex_key(fshared, &key2);
1335 put_futex_key(fshared, &key1);
1337 if (pi_state != NULL)
1338 free_pi_state(pi_state);
1339 return ret ? ret : task_count;
1342 /* The key must be already stored in q->key. */
1343 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1345 struct futex_hash_bucket *hb;
1347 get_futex_key_refs(&q->key);
1348 hb = hash_futex(&q->key);
1349 q->lock_ptr = &hb->lock;
1351 spin_lock(&hb->lock);
1356 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1358 spin_unlock(&hb->lock);
1359 drop_futex_key_refs(&q->key);
1363 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1364 * @q: The futex_q to enqueue
1365 * @hb: The destination hash bucket
1367 * The hb->lock must be held by the caller, and is released here. A call to
1368 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1369 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1370 * or nothing if the unqueue is done as part of the wake process and the unqueue
1371 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1374 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1379 * The priority used to register this element is
1380 * - either the real thread-priority for the real-time threads
1381 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1382 * - or MAX_RT_PRIO for non-RT threads.
1383 * Thus, all RT-threads are woken first in priority order, and
1384 * the others are woken last, in FIFO order.
1386 prio = min(current->normal_prio, MAX_RT_PRIO);
1388 plist_node_init(&q->list, prio);
1389 #ifdef CONFIG_DEBUG_PI_LIST
1390 q->list.plist.lock = &hb->lock;
1392 plist_add(&q->list, &hb->chain);
1394 spin_unlock(&hb->lock);
1398 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1399 * @q: The futex_q to unqueue
1401 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1402 * be paired with exactly one earlier call to queue_me().
1405 * 1 - if the futex_q was still queued (and we removed unqueued it)
1406 * 0 - if the futex_q was already removed by the waking thread
1408 static int unqueue_me(struct futex_q *q)
1410 spinlock_t *lock_ptr;
1413 /* In the common case we don't take the spinlock, which is nice. */
1415 lock_ptr = q->lock_ptr;
1417 if (lock_ptr != NULL) {
1418 spin_lock(lock_ptr);
1420 * q->lock_ptr can change between reading it and
1421 * spin_lock(), causing us to take the wrong lock. This
1422 * corrects the race condition.
1424 * Reasoning goes like this: if we have the wrong lock,
1425 * q->lock_ptr must have changed (maybe several times)
1426 * between reading it and the spin_lock(). It can
1427 * change again after the spin_lock() but only if it was
1428 * already changed before the spin_lock(). It cannot,
1429 * however, change back to the original value. Therefore
1430 * we can detect whether we acquired the correct lock.
1432 if (unlikely(lock_ptr != q->lock_ptr)) {
1433 spin_unlock(lock_ptr);
1436 WARN_ON(plist_node_empty(&q->list));
1437 plist_del(&q->list, &q->list.plist);
1439 BUG_ON(q->pi_state);
1441 spin_unlock(lock_ptr);
1445 drop_futex_key_refs(&q->key);
1450 * PI futexes can not be requeued and must remove themself from the
1451 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1454 static void unqueue_me_pi(struct futex_q *q)
1456 WARN_ON(plist_node_empty(&q->list));
1457 plist_del(&q->list, &q->list.plist);
1459 BUG_ON(!q->pi_state);
1460 free_pi_state(q->pi_state);
1463 spin_unlock(q->lock_ptr);
1465 drop_futex_key_refs(&q->key);
1469 * Fixup the pi_state owner with the new owner.
1471 * Must be called with hash bucket lock held and mm->sem held for non
1474 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1475 struct task_struct *newowner, int fshared)
1477 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1478 struct futex_pi_state *pi_state = q->pi_state;
1479 struct task_struct *oldowner = pi_state->owner;
1480 u32 uval, curval, newval;
1484 if (!pi_state->owner)
1485 newtid |= FUTEX_OWNER_DIED;
1488 * We are here either because we stole the rtmutex from the
1489 * pending owner or we are the pending owner which failed to
1490 * get the rtmutex. We have to replace the pending owner TID
1491 * in the user space variable. This must be atomic as we have
1492 * to preserve the owner died bit here.
1494 * Note: We write the user space value _before_ changing the pi_state
1495 * because we can fault here. Imagine swapped out pages or a fork
1496 * that marked all the anonymous memory readonly for cow.
1498 * Modifying pi_state _before_ the user space value would
1499 * leave the pi_state in an inconsistent state when we fault
1500 * here, because we need to drop the hash bucket lock to
1501 * handle the fault. This might be observed in the PID check
1502 * in lookup_pi_state.
1505 if (get_futex_value_locked(&uval, uaddr))
1509 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1511 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1513 if (curval == -EFAULT)
1521 * We fixed up user space. Now we need to fix the pi_state
1524 if (pi_state->owner != NULL) {
1525 spin_lock_irq(&pi_state->owner->pi_lock);
1526 WARN_ON(list_empty(&pi_state->list));
1527 list_del_init(&pi_state->list);
1528 spin_unlock_irq(&pi_state->owner->pi_lock);
1531 pi_state->owner = newowner;
1533 spin_lock_irq(&newowner->pi_lock);
1534 WARN_ON(!list_empty(&pi_state->list));
1535 list_add(&pi_state->list, &newowner->pi_state_list);
1536 spin_unlock_irq(&newowner->pi_lock);
1540 * To handle the page fault we need to drop the hash bucket
1541 * lock here. That gives the other task (either the pending
1542 * owner itself or the task which stole the rtmutex) the
1543 * chance to try the fixup of the pi_state. So once we are
1544 * back from handling the fault we need to check the pi_state
1545 * after reacquiring the hash bucket lock and before trying to
1546 * do another fixup. When the fixup has been done already we
1550 spin_unlock(q->lock_ptr);
1552 ret = fault_in_user_writeable(uaddr);
1554 spin_lock(q->lock_ptr);
1557 * Check if someone else fixed it for us:
1559 if (pi_state->owner != oldowner)
1569 * In case we must use restart_block to restart a futex_wait,
1570 * we encode in the 'flags' shared capability
1572 #define FLAGS_SHARED 0x01
1573 #define FLAGS_CLOCKRT 0x02
1574 #define FLAGS_HAS_TIMEOUT 0x04
1576 static long futex_wait_restart(struct restart_block *restart);
1579 * fixup_owner() - Post lock pi_state and corner case management
1580 * @uaddr: user address of the futex
1581 * @fshared: whether the futex is shared (1) or not (0)
1582 * @q: futex_q (contains pi_state and access to the rt_mutex)
1583 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1585 * After attempting to lock an rt_mutex, this function is called to cleanup
1586 * the pi_state owner as well as handle race conditions that may allow us to
1587 * acquire the lock. Must be called with the hb lock held.
1590 * 1 - success, lock taken
1591 * 0 - success, lock not taken
1592 * <0 - on error (-EFAULT)
1594 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1597 struct task_struct *owner;
1602 * Got the lock. We might not be the anticipated owner if we
1603 * did a lock-steal - fix up the PI-state in that case:
1605 if (q->pi_state->owner != current)
1606 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1611 * Catch the rare case, where the lock was released when we were on the
1612 * way back before we locked the hash bucket.
1614 if (q->pi_state->owner == current) {
1616 * Try to get the rt_mutex now. This might fail as some other
1617 * task acquired the rt_mutex after we removed ourself from the
1618 * rt_mutex waiters list.
1620 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1626 * pi_state is incorrect, some other task did a lock steal and
1627 * we returned due to timeout or signal without taking the
1628 * rt_mutex. Too late. We can access the rt_mutex_owner without
1629 * locking, as the other task is now blocked on the hash bucket
1630 * lock. Fix the state up.
1632 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1633 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1638 * Paranoia check. If we did not take the lock, then we should not be
1639 * the owner, nor the pending owner, of the rt_mutex.
1641 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1642 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1643 "pi-state %p\n", ret,
1644 q->pi_state->pi_mutex.owner,
1645 q->pi_state->owner);
1648 return ret ? ret : locked;
1652 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1653 * @hb: the futex hash bucket, must be locked by the caller
1654 * @q: the futex_q to queue up on
1655 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1657 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1658 struct hrtimer_sleeper *timeout)
1661 * The task state is guaranteed to be set before another task can
1662 * wake it. set_current_state() is implemented using set_mb() and
1663 * queue_me() calls spin_unlock() upon completion, both serializing
1664 * access to the hash list and forcing another memory barrier.
1666 set_current_state(TASK_INTERRUPTIBLE);
1671 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1672 if (!hrtimer_active(&timeout->timer))
1673 timeout->task = NULL;
1677 * If we have been removed from the hash list, then another task
1678 * has tried to wake us, and we can skip the call to schedule().
1680 if (likely(!plist_node_empty(&q->list))) {
1682 * If the timer has already expired, current will already be
1683 * flagged for rescheduling. Only call schedule if there
1684 * is no timeout, or if it has yet to expire.
1686 if (!timeout || timeout->task)
1689 __set_current_state(TASK_RUNNING);
1693 * futex_wait_setup() - Prepare to wait on a futex
1694 * @uaddr: the futex userspace address
1695 * @val: the expected value
1696 * @fshared: whether the futex is shared (1) or not (0)
1697 * @q: the associated futex_q
1698 * @hb: storage for hash_bucket pointer to be returned to caller
1700 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1701 * compare it with the expected value. Handle atomic faults internally.
1702 * Return with the hb lock held and a q.key reference on success, and unlocked
1703 * with no q.key reference on failure.
1706 * 0 - uaddr contains val and hb has been locked
1707 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1709 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1710 struct futex_q *q, struct futex_hash_bucket **hb)
1716 * Access the page AFTER the hash-bucket is locked.
1717 * Order is important:
1719 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1720 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1722 * The basic logical guarantee of a futex is that it blocks ONLY
1723 * if cond(var) is known to be true at the time of blocking, for
1724 * any cond. If we queued after testing *uaddr, that would open
1725 * a race condition where we could block indefinitely with
1726 * cond(var) false, which would violate the guarantee.
1728 * A consequence is that futex_wait() can return zero and absorb
1729 * a wakeup when *uaddr != val on entry to the syscall. This is
1733 q->key = FUTEX_KEY_INIT;
1734 ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1735 if (unlikely(ret != 0))
1739 *hb = queue_lock(q);
1741 ret = get_futex_value_locked(&uval, uaddr);
1744 queue_unlock(q, *hb);
1746 ret = get_user(uval, uaddr);
1753 put_futex_key(fshared, &q->key);
1758 queue_unlock(q, *hb);
1764 put_futex_key(fshared, &q->key);
1768 static int futex_wait(u32 __user *uaddr, int fshared,
1769 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1771 struct hrtimer_sleeper timeout, *to = NULL;
1772 struct restart_block *restart;
1773 struct futex_hash_bucket *hb;
1783 q.requeue_pi_key = NULL;
1788 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1789 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1790 hrtimer_init_sleeper(to, current);
1791 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1792 current->timer_slack_ns);
1796 /* Prepare to wait on uaddr. */
1797 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1801 /* queue_me and wait for wakeup, timeout, or a signal. */
1802 futex_wait_queue_me(hb, &q, to);
1804 /* If we were woken (and unqueued), we succeeded, whatever. */
1806 if (!unqueue_me(&q))
1809 if (to && !to->task)
1813 * We expect signal_pending(current), but we might be the
1814 * victim of a spurious wakeup as well.
1816 if (!signal_pending(current)) {
1817 put_futex_key(fshared, &q.key);
1825 restart = ¤t_thread_info()->restart_block;
1826 restart->fn = futex_wait_restart;
1827 restart->futex.uaddr = (u32 *)uaddr;
1828 restart->futex.val = val;
1829 restart->futex.time = abs_time->tv64;
1830 restart->futex.bitset = bitset;
1831 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1834 restart->futex.flags |= FLAGS_SHARED;
1836 restart->futex.flags |= FLAGS_CLOCKRT;
1838 ret = -ERESTART_RESTARTBLOCK;
1841 put_futex_key(fshared, &q.key);
1844 hrtimer_cancel(&to->timer);
1845 destroy_hrtimer_on_stack(&to->timer);
1851 static long futex_wait_restart(struct restart_block *restart)
1853 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1855 ktime_t t, *tp = NULL;
1857 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1858 t.tv64 = restart->futex.time;
1861 restart->fn = do_no_restart_syscall;
1862 if (restart->futex.flags & FLAGS_SHARED)
1864 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1865 restart->futex.bitset,
1866 restart->futex.flags & FLAGS_CLOCKRT);
1871 * Userspace tried a 0 -> TID atomic transition of the futex value
1872 * and failed. The kernel side here does the whole locking operation:
1873 * if there are waiters then it will block, it does PI, etc. (Due to
1874 * races the kernel might see a 0 value of the futex too.)
1876 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1877 int detect, ktime_t *time, int trylock)
1879 struct hrtimer_sleeper timeout, *to = NULL;
1880 struct futex_hash_bucket *hb;
1884 if (refill_pi_state_cache())
1889 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1891 hrtimer_init_sleeper(to, current);
1892 hrtimer_set_expires(&to->timer, *time);
1897 q.requeue_pi_key = NULL;
1899 q.key = FUTEX_KEY_INIT;
1900 ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1901 if (unlikely(ret != 0))
1905 hb = queue_lock(&q);
1907 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1908 if (unlikely(ret)) {
1911 /* We got the lock. */
1913 goto out_unlock_put_key;
1918 * Task is exiting and we just wait for the
1921 queue_unlock(&q, hb);
1922 put_futex_key(fshared, &q.key);
1926 goto out_unlock_put_key;
1931 * Only actually queue now that the atomic ops are done:
1935 WARN_ON(!q.pi_state);
1937 * Block on the PI mutex:
1940 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1942 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1943 /* Fixup the trylock return value: */
1944 ret = ret ? 0 : -EWOULDBLOCK;
1947 spin_lock(q.lock_ptr);
1949 * Fixup the pi_state owner and possibly acquire the lock if we
1952 res = fixup_owner(uaddr, fshared, &q, !ret);
1954 * If fixup_owner() returned an error, proprogate that. If it acquired
1955 * the lock, clear our -ETIMEDOUT or -EINTR.
1958 ret = (res < 0) ? res : 0;
1961 * If fixup_owner() faulted and was unable to handle the fault, unlock
1962 * it and return the fault to userspace.
1964 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1965 rt_mutex_unlock(&q.pi_state->pi_mutex);
1967 /* Unqueue and drop the lock */
1973 queue_unlock(&q, hb);
1976 put_futex_key(fshared, &q.key);
1979 destroy_hrtimer_on_stack(&to->timer);
1980 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1983 queue_unlock(&q, hb);
1985 ret = fault_in_user_writeable(uaddr);
1992 put_futex_key(fshared, &q.key);
1997 * Userspace attempted a TID -> 0 atomic transition, and failed.
1998 * This is the in-kernel slowpath: we look up the PI state (if any),
1999 * and do the rt-mutex unlock.
2001 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2003 struct futex_hash_bucket *hb;
2004 struct futex_q *this, *next;
2006 struct plist_head *head;
2007 union futex_key key = FUTEX_KEY_INIT;
2011 if (get_user(uval, uaddr))
2014 * We release only a lock we actually own:
2016 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2019 ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
2020 if (unlikely(ret != 0))
2023 hb = hash_futex(&key);
2024 spin_lock(&hb->lock);
2027 * To avoid races, try to do the TID -> 0 atomic transition
2028 * again. If it succeeds then we can return without waking
2031 if (!(uval & FUTEX_OWNER_DIED))
2032 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2035 if (unlikely(uval == -EFAULT))
2038 * Rare case: we managed to release the lock atomically,
2039 * no need to wake anyone else up:
2041 if (unlikely(uval == task_pid_vnr(current)))
2045 * Ok, other tasks may need to be woken up - check waiters
2046 * and do the wakeup if necessary:
2050 plist_for_each_entry_safe(this, next, head, list) {
2051 if (!match_futex (&this->key, &key))
2053 ret = wake_futex_pi(uaddr, uval, this);
2055 * The atomic access to the futex value
2056 * generated a pagefault, so retry the
2057 * user-access and the wakeup:
2064 * No waiters - kernel unlocks the futex:
2066 if (!(uval & FUTEX_OWNER_DIED)) {
2067 ret = unlock_futex_pi(uaddr, uval);
2073 spin_unlock(&hb->lock);
2074 put_futex_key(fshared, &key);
2080 spin_unlock(&hb->lock);
2081 put_futex_key(fshared, &key);
2083 ret = fault_in_user_writeable(uaddr);
2091 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2092 * @hb: the hash_bucket futex_q was original enqueued on
2093 * @q: the futex_q woken while waiting to be requeued
2094 * @key2: the futex_key of the requeue target futex
2095 * @timeout: the timeout associated with the wait (NULL if none)
2097 * Detect if the task was woken on the initial futex as opposed to the requeue
2098 * target futex. If so, determine if it was a timeout or a signal that caused
2099 * the wakeup and return the appropriate error code to the caller. Must be
2100 * called with the hb lock held.
2103 * 0 - no early wakeup detected
2104 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2107 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2108 struct futex_q *q, union futex_key *key2,
2109 struct hrtimer_sleeper *timeout)
2114 * With the hb lock held, we avoid races while we process the wakeup.
2115 * We only need to hold hb (and not hb2) to ensure atomicity as the
2116 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2117 * It can't be requeued from uaddr2 to something else since we don't
2118 * support a PI aware source futex for requeue.
2120 if (!match_futex(&q->key, key2)) {
2121 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2123 * We were woken prior to requeue by a timeout or a signal.
2124 * Unqueue the futex_q and determine which it was.
2126 plist_del(&q->list, &q->list.plist);
2128 /* Handle spurious wakeups gracefully */
2130 if (timeout && !timeout->task)
2132 else if (signal_pending(current))
2133 ret = -ERESTARTNOINTR;
2139 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2140 * @uaddr: the futex we initially wait on (non-pi)
2141 * @fshared: whether the futexes are shared (1) or not (0). They must be
2142 * the same type, no requeueing from private to shared, etc.
2143 * @val: the expected value of uaddr
2144 * @abs_time: absolute timeout
2145 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2146 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2147 * @uaddr2: the pi futex we will take prior to returning to user-space
2149 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2150 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2151 * complete the acquisition of the rt_mutex prior to returning to userspace.
2152 * This ensures the rt_mutex maintains an owner when it has waiters; without
2153 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2156 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2157 * via the following:
2158 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2159 * 2) wakeup on uaddr2 after a requeue
2163 * If 3, cleanup and return -ERESTARTNOINTR.
2165 * If 2, we may then block on trying to take the rt_mutex and return via:
2166 * 5) successful lock
2169 * 8) other lock acquisition failure
2171 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2173 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2179 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2180 u32 val, ktime_t *abs_time, u32 bitset,
2181 int clockrt, u32 __user *uaddr2)
2183 struct hrtimer_sleeper timeout, *to = NULL;
2184 struct rt_mutex_waiter rt_waiter;
2185 struct rt_mutex *pi_mutex = NULL;
2186 struct futex_hash_bucket *hb;
2187 union futex_key key2;
2196 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2197 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2198 hrtimer_init_sleeper(to, current);
2199 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2200 current->timer_slack_ns);
2204 * The waiter is allocated on our stack, manipulated by the requeue
2205 * code while we sleep on uaddr.
2207 debug_rt_mutex_init_waiter(&rt_waiter);
2208 rt_waiter.task = NULL;
2211 key2 = FUTEX_KEY_INIT;
2212 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2213 if (unlikely(ret != 0))
2218 q.rt_waiter = &rt_waiter;
2219 q.requeue_pi_key = &key2;
2221 /* Prepare to wait on uaddr. */
2222 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2226 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2227 futex_wait_queue_me(hb, &q, to);
2229 spin_lock(&hb->lock);
2230 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2231 spin_unlock(&hb->lock);
2236 * In order for us to be here, we know our q.key == key2, and since
2237 * we took the hb->lock above, we also know that futex_requeue() has
2238 * completed and we no longer have to concern ourselves with a wakeup
2239 * race with the atomic proxy lock acquition by the requeue code.
2242 /* Check if the requeue code acquired the second futex for us. */
2245 * Got the lock. We might not be the anticipated owner if we
2246 * did a lock-steal - fix up the PI-state in that case.
2248 if (q.pi_state && (q.pi_state->owner != current)) {
2249 spin_lock(q.lock_ptr);
2250 ret = fixup_pi_state_owner(uaddr2, &q, current,
2252 spin_unlock(q.lock_ptr);
2256 * We have been woken up by futex_unlock_pi(), a timeout, or a
2257 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2260 WARN_ON(!&q.pi_state);
2261 pi_mutex = &q.pi_state->pi_mutex;
2262 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2263 debug_rt_mutex_free_waiter(&rt_waiter);
2265 spin_lock(q.lock_ptr);
2267 * Fixup the pi_state owner and possibly acquire the lock if we
2270 res = fixup_owner(uaddr2, fshared, &q, !ret);
2272 * If fixup_owner() returned an error, proprogate that. If it
2273 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2276 ret = (res < 0) ? res : 0;
2278 /* Unqueue and drop the lock. */
2283 * If fixup_pi_state_owner() faulted and was unable to handle the
2284 * fault, unlock the rt_mutex and return the fault to userspace.
2286 if (ret == -EFAULT) {
2287 if (rt_mutex_owner(pi_mutex) == current)
2288 rt_mutex_unlock(pi_mutex);
2289 } else if (ret == -EINTR) {
2291 * We've already been requeued, but cannot restart by calling
2292 * futex_lock_pi() directly. We could restart this syscall, but
2293 * it would detect that the user space "val" changed and return
2294 * -EWOULDBLOCK. Save the overhead of the restart and return
2295 * -EWOULDBLOCK directly.
2301 put_futex_key(fshared, &q.key);
2303 put_futex_key(fshared, &key2);
2305 /* Spurious wakeup ? */
2310 hrtimer_cancel(&to->timer);
2311 destroy_hrtimer_on_stack(&to->timer);
2317 * Support for robust futexes: the kernel cleans up held futexes at
2320 * Implementation: user-space maintains a per-thread list of locks it
2321 * is holding. Upon do_exit(), the kernel carefully walks this list,
2322 * and marks all locks that are owned by this thread with the
2323 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2324 * always manipulated with the lock held, so the list is private and
2325 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2326 * field, to allow the kernel to clean up if the thread dies after
2327 * acquiring the lock, but just before it could have added itself to
2328 * the list. There can only be one such pending lock.
2332 * sys_set_robust_list() - Set the robust-futex list head of a task
2333 * @head: pointer to the list-head
2334 * @len: length of the list-head, as userspace expects
2336 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2339 if (!futex_cmpxchg_enabled)
2342 * The kernel knows only one size for now:
2344 if (unlikely(len != sizeof(*head)))
2347 current->robust_list = head;
2353 * sys_get_robust_list() - Get the robust-futex list head of a task
2354 * @pid: pid of the process [zero for current task]
2355 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2356 * @len_ptr: pointer to a length field, the kernel fills in the header size
2358 SYSCALL_DEFINE3(get_robust_list, int, pid,
2359 struct robust_list_head __user * __user *, head_ptr,
2360 size_t __user *, len_ptr)
2362 struct robust_list_head __user *head;
2364 const struct cred *cred = current_cred(), *pcred;
2366 if (!futex_cmpxchg_enabled)
2370 head = current->robust_list;
2372 struct task_struct *p;
2376 p = find_task_by_vpid(pid);
2380 pcred = __task_cred(p);
2381 if (cred->euid != pcred->euid &&
2382 cred->euid != pcred->uid &&
2383 !capable(CAP_SYS_PTRACE))
2385 head = p->robust_list;
2389 if (put_user(sizeof(*head), len_ptr))
2391 return put_user(head, head_ptr);
2400 * Process a futex-list entry, check whether it's owned by the
2401 * dying task, and do notification if so:
2403 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2405 u32 uval, nval, mval;
2408 if (get_user(uval, uaddr))
2411 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2413 * Ok, this dying thread is truly holding a futex
2414 * of interest. Set the OWNER_DIED bit atomically
2415 * via cmpxchg, and if the value had FUTEX_WAITERS
2416 * set, wake up a waiter (if any). (We have to do a
2417 * futex_wake() even if OWNER_DIED is already set -
2418 * to handle the rare but possible case of recursive
2419 * thread-death.) The rest of the cleanup is done in
2422 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2423 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2425 if (nval == -EFAULT)
2432 * Wake robust non-PI futexes here. The wakeup of
2433 * PI futexes happens in exit_pi_state():
2435 if (!pi && (uval & FUTEX_WAITERS))
2436 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2442 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2444 static inline int fetch_robust_entry(struct robust_list __user **entry,
2445 struct robust_list __user * __user *head,
2448 unsigned long uentry;
2450 if (get_user(uentry, (unsigned long __user *)head))
2453 *entry = (void __user *)(uentry & ~1UL);
2460 * Walk curr->robust_list (very carefully, it's a userspace list!)
2461 * and mark any locks found there dead, and notify any waiters.
2463 * We silently return on any sign of list-walking problem.
2465 void exit_robust_list(struct task_struct *curr)
2467 struct robust_list_head __user *head = curr->robust_list;
2468 struct robust_list __user *entry, *next_entry, *pending;
2469 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2470 unsigned long futex_offset;
2473 if (!futex_cmpxchg_enabled)
2477 * Fetch the list head (which was registered earlier, via
2478 * sys_set_robust_list()):
2480 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2483 * Fetch the relative futex offset:
2485 if (get_user(futex_offset, &head->futex_offset))
2488 * Fetch any possibly pending lock-add first, and handle it
2491 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2494 next_entry = NULL; /* avoid warning with gcc */
2495 while (entry != &head->list) {
2497 * Fetch the next entry in the list before calling
2498 * handle_futex_death:
2500 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2502 * A pending lock might already be on the list, so
2503 * don't process it twice:
2505 if (entry != pending)
2506 if (handle_futex_death((void __user *)entry + futex_offset,
2514 * Avoid excessively long or circular lists:
2523 handle_futex_death((void __user *)pending + futex_offset,
2527 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2528 u32 __user *uaddr2, u32 val2, u32 val3)
2530 int clockrt, ret = -ENOSYS;
2531 int cmd = op & FUTEX_CMD_MASK;
2534 if (!(op & FUTEX_PRIVATE_FLAG))
2537 clockrt = op & FUTEX_CLOCK_REALTIME;
2538 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2543 val3 = FUTEX_BITSET_MATCH_ANY;
2544 case FUTEX_WAIT_BITSET:
2545 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2548 val3 = FUTEX_BITSET_MATCH_ANY;
2549 case FUTEX_WAKE_BITSET:
2550 ret = futex_wake(uaddr, fshared, val, val3);
2553 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2555 case FUTEX_CMP_REQUEUE:
2556 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2560 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2563 if (futex_cmpxchg_enabled)
2564 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2566 case FUTEX_UNLOCK_PI:
2567 if (futex_cmpxchg_enabled)
2568 ret = futex_unlock_pi(uaddr, fshared);
2570 case FUTEX_TRYLOCK_PI:
2571 if (futex_cmpxchg_enabled)
2572 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2574 case FUTEX_WAIT_REQUEUE_PI:
2575 val3 = FUTEX_BITSET_MATCH_ANY;
2576 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2579 case FUTEX_CMP_REQUEUE_PI:
2580 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2590 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2591 struct timespec __user *, utime, u32 __user *, uaddr2,
2595 ktime_t t, *tp = NULL;
2597 int cmd = op & FUTEX_CMD_MASK;
2599 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2600 cmd == FUTEX_WAIT_BITSET ||
2601 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2602 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2604 if (!timespec_valid(&ts))
2607 t = timespec_to_ktime(ts);
2608 if (cmd == FUTEX_WAIT)
2609 t = ktime_add_safe(ktime_get(), t);
2613 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2614 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2616 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2617 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2618 val2 = (u32) (unsigned long) utime;
2620 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2623 static int __init futex_init(void)
2629 * This will fail and we want it. Some arch implementations do
2630 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2631 * functionality. We want to know that before we call in any
2632 * of the complex code paths. Also we want to prevent
2633 * registration of robust lists in that case. NULL is
2634 * guaranteed to fault and we get -EFAULT on functional
2635 * implementation, the non functional ones will return
2638 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2639 if (curval == -EFAULT)
2640 futex_cmpxchg_enabled = 1;
2642 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2643 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2644 spin_lock_init(&futex_queues[i].lock);
2649 __initcall(futex_init);