futex: Move drop_futex_key_refs out of spinlock'ed region
[linux-2.6/verdex.git] / kernel / futex.c
blobf435af120a67409f5170d374d2c792d524fd935c
1 /*
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>
49 #include <linux/fs.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;
82 * The PI object:
84 struct rt_mutex pi_mutex;
86 struct task_struct *owner;
87 atomic_t refcount;
89 union futex_key key;
93 * We use this hashed waitqueue instead of a normal wait_queue_t, so
94 * we can wake only the relevant ones (hashed queues may be shared).
96 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
97 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
98 * The order of wakup is always to make the first condition true, then
99 * wake up q->waiter, then make the second condition true.
101 struct futex_q {
102 struct plist_node list;
103 /* Waiter reference */
104 struct task_struct *task;
106 /* Which hash list lock to use: */
107 spinlock_t *lock_ptr;
109 /* Key which the futex is hashed on: */
110 union futex_key key;
112 /* Optional priority inheritance state: */
113 struct futex_pi_state *pi_state;
115 /* rt_waiter storage for requeue_pi: */
116 struct rt_mutex_waiter *rt_waiter;
118 /* The expected requeue pi target futex key: */
119 union futex_key *requeue_pi_key;
121 /* Bitset for the optional bitmasked wakeup */
122 u32 bitset;
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 {
131 spinlock_t lock;
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,
144 key->both.offset);
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)
153 return (key1 && 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)
166 if (!key->both.ptr)
167 return;
169 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170 case FUT_OFF_INODE:
171 atomic_inc(&key->shared.inode->i_count);
172 break;
173 case FUT_OFF_MMSHARED:
174 atomic_inc(&key->private.mm->mm_count);
175 break;
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 */
187 WARN_ON_ONCE(1);
188 return;
191 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192 case FUT_OFF_INODE:
193 iput(key->shared.inode);
194 break;
195 case FUT_OFF_MMSHARED:
196 mmdrop(key->private.mm);
197 break;
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, VERIFY_WRITE)
208 * Returns a negative error code or 0
209 * The key words are stored in *key on success.
211 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
212 * offset_within_page). For private mappings, it's (uaddr, current->mm).
213 * We can usually work out the index without swapping in the page.
215 * lock_page() might sleep, the caller should not hold a spinlock.
217 static int
218 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
220 unsigned long address = (unsigned long)uaddr;
221 struct mm_struct *mm = current->mm;
222 struct page *page;
223 int err;
226 * The futex address must be "naturally" aligned.
228 key->both.offset = address % PAGE_SIZE;
229 if (unlikely((address % sizeof(u32)) != 0))
230 return -EINVAL;
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()
240 if (!fshared) {
241 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
242 return -EFAULT;
243 key->private.mm = mm;
244 key->private.address = address;
245 get_futex_key_refs(key);
246 return 0;
249 again:
250 err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
251 if (err < 0)
252 return err;
254 page = compound_head(page);
255 lock_page(page);
256 if (!page->mapping) {
257 unlock_page(page);
258 put_page(page);
259 goto again;
263 * Private mappings are handled in a simple way.
265 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
266 * it's a read-only handle, it's expected that futexes attach to
267 * the object not the particular process.
269 if (PageAnon(page)) {
270 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
271 key->private.mm = mm;
272 key->private.address = address;
273 } else {
274 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
275 key->shared.inode = page->mapping->host;
276 key->shared.pgoff = page->index;
279 get_futex_key_refs(key);
281 unlock_page(page);
282 put_page(page);
283 return 0;
286 static inline
287 void put_futex_key(int fshared, union futex_key *key)
289 drop_futex_key_refs(key);
293 * fault_in_user_writeable - fault in user address and verify RW access
294 * @uaddr: pointer to faulting user space address
296 * Slow path to fixup the fault we just took in the atomic write
297 * access to @uaddr.
299 * We have no generic implementation of a non destructive write to the
300 * user address. We know that we faulted in the atomic pagefault
301 * disabled section so we can as well avoid the #PF overhead by
302 * calling get_user_pages() right away.
304 static int fault_in_user_writeable(u32 __user *uaddr)
306 int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
307 1, 1, 0, NULL, NULL);
308 return ret < 0 ? ret : 0;
312 * futex_top_waiter() - Return the highest priority waiter on a futex
313 * @hb: the hash bucket the futex_q's reside in
314 * @key: the futex key (to distinguish it from other futex futex_q's)
316 * Must be called with the hb lock held.
318 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
319 union futex_key *key)
321 struct futex_q *this;
323 plist_for_each_entry(this, &hb->chain, list) {
324 if (match_futex(&this->key, key))
325 return this;
327 return NULL;
330 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
332 u32 curval;
334 pagefault_disable();
335 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
336 pagefault_enable();
338 return curval;
341 static int get_futex_value_locked(u32 *dest, u32 __user *from)
343 int ret;
345 pagefault_disable();
346 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
347 pagefault_enable();
349 return ret ? -EFAULT : 0;
354 * PI code:
356 static int refill_pi_state_cache(void)
358 struct futex_pi_state *pi_state;
360 if (likely(current->pi_state_cache))
361 return 0;
363 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
365 if (!pi_state)
366 return -ENOMEM;
368 INIT_LIST_HEAD(&pi_state->list);
369 /* pi_mutex gets initialized later */
370 pi_state->owner = NULL;
371 atomic_set(&pi_state->refcount, 1);
372 pi_state->key = FUTEX_KEY_INIT;
374 current->pi_state_cache = pi_state;
376 return 0;
379 static struct futex_pi_state * alloc_pi_state(void)
381 struct futex_pi_state *pi_state = current->pi_state_cache;
383 WARN_ON(!pi_state);
384 current->pi_state_cache = NULL;
386 return pi_state;
389 static void free_pi_state(struct futex_pi_state *pi_state)
391 if (!atomic_dec_and_test(&pi_state->refcount))
392 return;
395 * If pi_state->owner is NULL, the owner is most probably dying
396 * and has cleaned up the pi_state already
398 if (pi_state->owner) {
399 spin_lock_irq(&pi_state->owner->pi_lock);
400 list_del_init(&pi_state->list);
401 spin_unlock_irq(&pi_state->owner->pi_lock);
403 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
406 if (current->pi_state_cache)
407 kfree(pi_state);
408 else {
410 * pi_state->list is already empty.
411 * clear pi_state->owner.
412 * refcount is at 0 - put it back to 1.
414 pi_state->owner = NULL;
415 atomic_set(&pi_state->refcount, 1);
416 current->pi_state_cache = pi_state;
421 * Look up the task based on what TID userspace gave us.
422 * We dont trust it.
424 static struct task_struct * futex_find_get_task(pid_t pid)
426 struct task_struct *p;
427 const struct cred *cred = current_cred(), *pcred;
429 rcu_read_lock();
430 p = find_task_by_vpid(pid);
431 if (!p) {
432 p = ERR_PTR(-ESRCH);
433 } else {
434 pcred = __task_cred(p);
435 if (cred->euid != pcred->euid &&
436 cred->euid != pcred->uid)
437 p = ERR_PTR(-ESRCH);
438 else
439 get_task_struct(p);
442 rcu_read_unlock();
444 return p;
448 * This task is holding PI mutexes at exit time => bad.
449 * Kernel cleans up PI-state, but userspace is likely hosed.
450 * (Robust-futex cleanup is separate and might save the day for userspace.)
452 void exit_pi_state_list(struct task_struct *curr)
454 struct list_head *next, *head = &curr->pi_state_list;
455 struct futex_pi_state *pi_state;
456 struct futex_hash_bucket *hb;
457 union futex_key key = FUTEX_KEY_INIT;
459 if (!futex_cmpxchg_enabled)
460 return;
462 * We are a ZOMBIE and nobody can enqueue itself on
463 * pi_state_list anymore, but we have to be careful
464 * versus waiters unqueueing themselves:
466 spin_lock_irq(&curr->pi_lock);
467 while (!list_empty(head)) {
469 next = head->next;
470 pi_state = list_entry(next, struct futex_pi_state, list);
471 key = pi_state->key;
472 hb = hash_futex(&key);
473 spin_unlock_irq(&curr->pi_lock);
475 spin_lock(&hb->lock);
477 spin_lock_irq(&curr->pi_lock);
479 * We dropped the pi-lock, so re-check whether this
480 * task still owns the PI-state:
482 if (head->next != next) {
483 spin_unlock(&hb->lock);
484 continue;
487 WARN_ON(pi_state->owner != curr);
488 WARN_ON(list_empty(&pi_state->list));
489 list_del_init(&pi_state->list);
490 pi_state->owner = NULL;
491 spin_unlock_irq(&curr->pi_lock);
493 rt_mutex_unlock(&pi_state->pi_mutex);
495 spin_unlock(&hb->lock);
497 spin_lock_irq(&curr->pi_lock);
499 spin_unlock_irq(&curr->pi_lock);
502 static int
503 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
504 union futex_key *key, struct futex_pi_state **ps)
506 struct futex_pi_state *pi_state = NULL;
507 struct futex_q *this, *next;
508 struct plist_head *head;
509 struct task_struct *p;
510 pid_t pid = uval & FUTEX_TID_MASK;
512 head = &hb->chain;
514 plist_for_each_entry_safe(this, next, head, list) {
515 if (match_futex(&this->key, key)) {
517 * Another waiter already exists - bump up
518 * the refcount and return its pi_state:
520 pi_state = this->pi_state;
522 * Userspace might have messed up non PI and PI futexes
524 if (unlikely(!pi_state))
525 return -EINVAL;
527 WARN_ON(!atomic_read(&pi_state->refcount));
528 WARN_ON(pid && pi_state->owner &&
529 pi_state->owner->pid != pid);
531 atomic_inc(&pi_state->refcount);
532 *ps = pi_state;
534 return 0;
539 * We are the first waiter - try to look up the real owner and attach
540 * the new pi_state to it, but bail out when TID = 0
542 if (!pid)
543 return -ESRCH;
544 p = futex_find_get_task(pid);
545 if (IS_ERR(p))
546 return PTR_ERR(p);
549 * We need to look at the task state flags to figure out,
550 * whether the task is exiting. To protect against the do_exit
551 * change of the task flags, we do this protected by
552 * p->pi_lock:
554 spin_lock_irq(&p->pi_lock);
555 if (unlikely(p->flags & PF_EXITING)) {
557 * The task is on the way out. When PF_EXITPIDONE is
558 * set, we know that the task has finished the
559 * cleanup:
561 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
563 spin_unlock_irq(&p->pi_lock);
564 put_task_struct(p);
565 return ret;
568 pi_state = alloc_pi_state();
571 * Initialize the pi_mutex in locked state and make 'p'
572 * the owner of it:
574 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
576 /* Store the key for possible exit cleanups: */
577 pi_state->key = *key;
579 WARN_ON(!list_empty(&pi_state->list));
580 list_add(&pi_state->list, &p->pi_state_list);
581 pi_state->owner = p;
582 spin_unlock_irq(&p->pi_lock);
584 put_task_struct(p);
586 *ps = pi_state;
588 return 0;
592 * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
593 * @uaddr: the pi futex user address
594 * @hb: the pi futex hash bucket
595 * @key: the futex key associated with uaddr and hb
596 * @ps: the pi_state pointer where we store the result of the
597 * lookup
598 * @task: the task to perform the atomic lock work for. This will
599 * be "current" except in the case of requeue pi.
600 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
602 * Returns:
603 * 0 - ready to wait
604 * 1 - acquired the lock
605 * <0 - error
607 * The hb->lock and futex_key refs shall be held by the caller.
609 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
610 union futex_key *key,
611 struct futex_pi_state **ps,
612 struct task_struct *task, int set_waiters)
614 int lock_taken, ret, ownerdied = 0;
615 u32 uval, newval, curval;
617 retry:
618 ret = lock_taken = 0;
621 * To avoid races, we attempt to take the lock here again
622 * (by doing a 0 -> TID atomic cmpxchg), while holding all
623 * the locks. It will most likely not succeed.
625 newval = task_pid_vnr(task);
626 if (set_waiters)
627 newval |= FUTEX_WAITERS;
629 curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
631 if (unlikely(curval == -EFAULT))
632 return -EFAULT;
635 * Detect deadlocks.
637 if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
638 return -EDEADLK;
641 * Surprise - we got the lock. Just return to userspace:
643 if (unlikely(!curval))
644 return 1;
646 uval = curval;
649 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
650 * to wake at the next unlock.
652 newval = curval | FUTEX_WAITERS;
655 * There are two cases, where a futex might have no owner (the
656 * owner TID is 0): OWNER_DIED. We take over the futex in this
657 * case. We also do an unconditional take over, when the owner
658 * of the futex died.
660 * This is safe as we are protected by the hash bucket lock !
662 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
663 /* Keep the OWNER_DIED bit */
664 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
665 ownerdied = 0;
666 lock_taken = 1;
669 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
671 if (unlikely(curval == -EFAULT))
672 return -EFAULT;
673 if (unlikely(curval != uval))
674 goto retry;
677 * We took the lock due to owner died take over.
679 if (unlikely(lock_taken))
680 return 1;
683 * We dont have the lock. Look up the PI state (or create it if
684 * we are the first waiter):
686 ret = lookup_pi_state(uval, hb, key, ps);
688 if (unlikely(ret)) {
689 switch (ret) {
690 case -ESRCH:
692 * No owner found for this futex. Check if the
693 * OWNER_DIED bit is set to figure out whether
694 * this is a robust futex or not.
696 if (get_futex_value_locked(&curval, uaddr))
697 return -EFAULT;
700 * We simply start over in case of a robust
701 * futex. The code above will take the futex
702 * and return happy.
704 if (curval & FUTEX_OWNER_DIED) {
705 ownerdied = 1;
706 goto retry;
708 default:
709 break;
713 return ret;
717 * The hash bucket lock must be held when this is called.
718 * Afterwards, the futex_q must not be accessed.
720 static void wake_futex(struct futex_q *q)
722 struct task_struct *p = q->task;
725 * We set q->lock_ptr = NULL _before_ we wake up the task. If
726 * a non futex wake up happens on another CPU then the task
727 * might exit and p would dereference a non existing task
728 * struct. Prevent this by holding a reference on p across the
729 * wake up.
731 get_task_struct(p);
733 plist_del(&q->list, &q->list.plist);
735 * The waiting task can free the futex_q as soon as
736 * q->lock_ptr = NULL is written, without taking any locks. A
737 * memory barrier is required here to prevent the following
738 * store to lock_ptr from getting ahead of the plist_del.
740 smp_wmb();
741 q->lock_ptr = NULL;
743 wake_up_state(p, TASK_NORMAL);
744 put_task_struct(p);
747 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
749 struct task_struct *new_owner;
750 struct futex_pi_state *pi_state = this->pi_state;
751 u32 curval, newval;
753 if (!pi_state)
754 return -EINVAL;
756 spin_lock(&pi_state->pi_mutex.wait_lock);
757 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
760 * This happens when we have stolen the lock and the original
761 * pending owner did not enqueue itself back on the rt_mutex.
762 * Thats not a tragedy. We know that way, that a lock waiter
763 * is on the fly. We make the futex_q waiter the pending owner.
765 if (!new_owner)
766 new_owner = this->task;
769 * We pass it to the next owner. (The WAITERS bit is always
770 * kept enabled while there is PI state around. We must also
771 * preserve the owner died bit.)
773 if (!(uval & FUTEX_OWNER_DIED)) {
774 int ret = 0;
776 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
778 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
780 if (curval == -EFAULT)
781 ret = -EFAULT;
782 else if (curval != uval)
783 ret = -EINVAL;
784 if (ret) {
785 spin_unlock(&pi_state->pi_mutex.wait_lock);
786 return ret;
790 spin_lock_irq(&pi_state->owner->pi_lock);
791 WARN_ON(list_empty(&pi_state->list));
792 list_del_init(&pi_state->list);
793 spin_unlock_irq(&pi_state->owner->pi_lock);
795 spin_lock_irq(&new_owner->pi_lock);
796 WARN_ON(!list_empty(&pi_state->list));
797 list_add(&pi_state->list, &new_owner->pi_state_list);
798 pi_state->owner = new_owner;
799 spin_unlock_irq(&new_owner->pi_lock);
801 spin_unlock(&pi_state->pi_mutex.wait_lock);
802 rt_mutex_unlock(&pi_state->pi_mutex);
804 return 0;
807 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
809 u32 oldval;
812 * There is no waiter, so we unlock the futex. The owner died
813 * bit has not to be preserved here. We are the owner:
815 oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
817 if (oldval == -EFAULT)
818 return oldval;
819 if (oldval != uval)
820 return -EAGAIN;
822 return 0;
826 * Express the locking dependencies for lockdep:
828 static inline void
829 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
831 if (hb1 <= hb2) {
832 spin_lock(&hb1->lock);
833 if (hb1 < hb2)
834 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
835 } else { /* hb1 > hb2 */
836 spin_lock(&hb2->lock);
837 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
841 static inline void
842 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
844 spin_unlock(&hb1->lock);
845 if (hb1 != hb2)
846 spin_unlock(&hb2->lock);
850 * Wake up waiters matching bitset queued on this futex (uaddr).
852 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
854 struct futex_hash_bucket *hb;
855 struct futex_q *this, *next;
856 struct plist_head *head;
857 union futex_key key = FUTEX_KEY_INIT;
858 int ret;
860 if (!bitset)
861 return -EINVAL;
863 ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
864 if (unlikely(ret != 0))
865 goto out;
867 hb = hash_futex(&key);
868 spin_lock(&hb->lock);
869 head = &hb->chain;
871 plist_for_each_entry_safe(this, next, head, list) {
872 if (match_futex (&this->key, &key)) {
873 if (this->pi_state || this->rt_waiter) {
874 ret = -EINVAL;
875 break;
878 /* Check if one of the bits is set in both bitsets */
879 if (!(this->bitset & bitset))
880 continue;
882 wake_futex(this);
883 if (++ret >= nr_wake)
884 break;
888 spin_unlock(&hb->lock);
889 put_futex_key(fshared, &key);
890 out:
891 return ret;
895 * Wake up all waiters hashed on the physical page that is mapped
896 * to this virtual address:
898 static int
899 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
900 int nr_wake, int nr_wake2, int op)
902 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
903 struct futex_hash_bucket *hb1, *hb2;
904 struct plist_head *head;
905 struct futex_q *this, *next;
906 int ret, op_ret;
908 retry:
909 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
910 if (unlikely(ret != 0))
911 goto out;
912 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
913 if (unlikely(ret != 0))
914 goto out_put_key1;
916 hb1 = hash_futex(&key1);
917 hb2 = hash_futex(&key2);
919 retry_private:
920 double_lock_hb(hb1, hb2);
921 op_ret = futex_atomic_op_inuser(op, uaddr2);
922 if (unlikely(op_ret < 0)) {
924 double_unlock_hb(hb1, hb2);
926 #ifndef CONFIG_MMU
928 * we don't get EFAULT from MMU faults if we don't have an MMU,
929 * but we might get them from range checking
931 ret = op_ret;
932 goto out_put_keys;
933 #endif
935 if (unlikely(op_ret != -EFAULT)) {
936 ret = op_ret;
937 goto out_put_keys;
940 ret = fault_in_user_writeable(uaddr2);
941 if (ret)
942 goto out_put_keys;
944 if (!fshared)
945 goto retry_private;
947 put_futex_key(fshared, &key2);
948 put_futex_key(fshared, &key1);
949 goto retry;
952 head = &hb1->chain;
954 plist_for_each_entry_safe(this, next, head, list) {
955 if (match_futex (&this->key, &key1)) {
956 wake_futex(this);
957 if (++ret >= nr_wake)
958 break;
962 if (op_ret > 0) {
963 head = &hb2->chain;
965 op_ret = 0;
966 plist_for_each_entry_safe(this, next, head, list) {
967 if (match_futex (&this->key, &key2)) {
968 wake_futex(this);
969 if (++op_ret >= nr_wake2)
970 break;
973 ret += op_ret;
976 double_unlock_hb(hb1, hb2);
977 out_put_keys:
978 put_futex_key(fshared, &key2);
979 out_put_key1:
980 put_futex_key(fshared, &key1);
981 out:
982 return ret;
986 * requeue_futex() - Requeue a futex_q from one hb to another
987 * @q: the futex_q to requeue
988 * @hb1: the source hash_bucket
989 * @hb2: the target hash_bucket
990 * @key2: the new key for the requeued futex_q
992 static inline
993 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
994 struct futex_hash_bucket *hb2, union futex_key *key2)
998 * If key1 and key2 hash to the same bucket, no need to
999 * requeue.
1001 if (likely(&hb1->chain != &hb2->chain)) {
1002 plist_del(&q->list, &hb1->chain);
1003 plist_add(&q->list, &hb2->chain);
1004 q->lock_ptr = &hb2->lock;
1005 #ifdef CONFIG_DEBUG_PI_LIST
1006 q->list.plist.lock = &hb2->lock;
1007 #endif
1009 get_futex_key_refs(key2);
1010 q->key = *key2;
1014 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1015 * q: the futex_q
1016 * key: the key of the requeue target futex
1017 * hb: the hash_bucket of the requeue target futex
1019 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1020 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1021 * to the requeue target futex so the waiter can detect the wakeup on the right
1022 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1023 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1024 * to protect access to the pi_state to fixup the owner later. Must be called
1025 * with both q->lock_ptr and hb->lock held.
1027 static inline
1028 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1029 struct futex_hash_bucket *hb)
1031 get_futex_key_refs(key);
1032 q->key = *key;
1034 WARN_ON(plist_node_empty(&q->list));
1035 plist_del(&q->list, &q->list.plist);
1037 WARN_ON(!q->rt_waiter);
1038 q->rt_waiter = NULL;
1040 q->lock_ptr = &hb->lock;
1041 #ifdef CONFIG_DEBUG_PI_LIST
1042 q->list.plist.lock = &hb->lock;
1043 #endif
1045 wake_up_state(q->task, TASK_NORMAL);
1049 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1050 * @pifutex: the user address of the to futex
1051 * @hb1: the from futex hash bucket, must be locked by the caller
1052 * @hb2: the to futex hash bucket, must be locked by the caller
1053 * @key1: the from futex key
1054 * @key2: the to futex key
1055 * @ps: address to store the pi_state pointer
1056 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1058 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1059 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1060 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1061 * hb1 and hb2 must be held by the caller.
1063 * Returns:
1064 * 0 - failed to acquire the lock atomicly
1065 * 1 - acquired the lock
1066 * <0 - error
1068 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1069 struct futex_hash_bucket *hb1,
1070 struct futex_hash_bucket *hb2,
1071 union futex_key *key1, union futex_key *key2,
1072 struct futex_pi_state **ps, int set_waiters)
1074 struct futex_q *top_waiter = NULL;
1075 u32 curval;
1076 int ret;
1078 if (get_futex_value_locked(&curval, pifutex))
1079 return -EFAULT;
1082 * Find the top_waiter and determine if there are additional waiters.
1083 * If the caller intends to requeue more than 1 waiter to pifutex,
1084 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1085 * as we have means to handle the possible fault. If not, don't set
1086 * the bit unecessarily as it will force the subsequent unlock to enter
1087 * the kernel.
1089 top_waiter = futex_top_waiter(hb1, key1);
1091 /* There are no waiters, nothing for us to do. */
1092 if (!top_waiter)
1093 return 0;
1095 /* Ensure we requeue to the expected futex. */
1096 if (!match_futex(top_waiter->requeue_pi_key, key2))
1097 return -EINVAL;
1100 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1101 * the contended case or if set_waiters is 1. The pi_state is returned
1102 * in ps in contended cases.
1104 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1105 set_waiters);
1106 if (ret == 1)
1107 requeue_pi_wake_futex(top_waiter, key2, hb2);
1109 return ret;
1113 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1114 * uaddr1: source futex user address
1115 * uaddr2: target futex user address
1116 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1117 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1118 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1119 * pi futex (pi to pi requeue is not supported)
1121 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1122 * uaddr2 atomically on behalf of the top waiter.
1124 * Returns:
1125 * >=0 - on success, the number of tasks requeued or woken
1126 * <0 - on error
1128 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1129 int nr_wake, int nr_requeue, u32 *cmpval,
1130 int requeue_pi)
1132 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1133 int drop_count = 0, task_count = 0, ret;
1134 struct futex_pi_state *pi_state = NULL;
1135 struct futex_hash_bucket *hb1, *hb2;
1136 struct plist_head *head1;
1137 struct futex_q *this, *next;
1138 u32 curval2;
1140 if (requeue_pi) {
1142 * requeue_pi requires a pi_state, try to allocate it now
1143 * without any locks in case it fails.
1145 if (refill_pi_state_cache())
1146 return -ENOMEM;
1148 * requeue_pi must wake as many tasks as it can, up to nr_wake
1149 * + nr_requeue, since it acquires the rt_mutex prior to
1150 * returning to userspace, so as to not leave the rt_mutex with
1151 * waiters and no owner. However, second and third wake-ups
1152 * cannot be predicted as they involve race conditions with the
1153 * first wake and a fault while looking up the pi_state. Both
1154 * pthread_cond_signal() and pthread_cond_broadcast() should
1155 * use nr_wake=1.
1157 if (nr_wake != 1)
1158 return -EINVAL;
1161 retry:
1162 if (pi_state != NULL) {
1164 * We will have to lookup the pi_state again, so free this one
1165 * to keep the accounting correct.
1167 free_pi_state(pi_state);
1168 pi_state = NULL;
1171 ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1172 if (unlikely(ret != 0))
1173 goto out;
1174 ret = get_futex_key(uaddr2, fshared, &key2,
1175 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1176 if (unlikely(ret != 0))
1177 goto out_put_key1;
1179 hb1 = hash_futex(&key1);
1180 hb2 = hash_futex(&key2);
1182 retry_private:
1183 double_lock_hb(hb1, hb2);
1185 if (likely(cmpval != NULL)) {
1186 u32 curval;
1188 ret = get_futex_value_locked(&curval, uaddr1);
1190 if (unlikely(ret)) {
1191 double_unlock_hb(hb1, hb2);
1193 ret = get_user(curval, uaddr1);
1194 if (ret)
1195 goto out_put_keys;
1197 if (!fshared)
1198 goto retry_private;
1200 put_futex_key(fshared, &key2);
1201 put_futex_key(fshared, &key1);
1202 goto retry;
1204 if (curval != *cmpval) {
1205 ret = -EAGAIN;
1206 goto out_unlock;
1210 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1212 * Attempt to acquire uaddr2 and wake the top waiter. If we
1213 * intend to requeue waiters, force setting the FUTEX_WAITERS
1214 * bit. We force this here where we are able to easily handle
1215 * faults rather in the requeue loop below.
1217 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1218 &key2, &pi_state, nr_requeue);
1221 * At this point the top_waiter has either taken uaddr2 or is
1222 * waiting on it. If the former, then the pi_state will not
1223 * exist yet, look it up one more time to ensure we have a
1224 * reference to it.
1226 if (ret == 1) {
1227 WARN_ON(pi_state);
1228 drop_count++;
1229 task_count++;
1230 ret = get_futex_value_locked(&curval2, uaddr2);
1231 if (!ret)
1232 ret = lookup_pi_state(curval2, hb2, &key2,
1233 &pi_state);
1236 switch (ret) {
1237 case 0:
1238 break;
1239 case -EFAULT:
1240 double_unlock_hb(hb1, hb2);
1241 put_futex_key(fshared, &key2);
1242 put_futex_key(fshared, &key1);
1243 ret = fault_in_user_writeable(uaddr2);
1244 if (!ret)
1245 goto retry;
1246 goto out;
1247 case -EAGAIN:
1248 /* The owner was exiting, try again. */
1249 double_unlock_hb(hb1, hb2);
1250 put_futex_key(fshared, &key2);
1251 put_futex_key(fshared, &key1);
1252 cond_resched();
1253 goto retry;
1254 default:
1255 goto out_unlock;
1259 head1 = &hb1->chain;
1260 plist_for_each_entry_safe(this, next, head1, list) {
1261 if (task_count - nr_wake >= nr_requeue)
1262 break;
1264 if (!match_futex(&this->key, &key1))
1265 continue;
1268 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1269 * be paired with each other and no other futex ops.
1271 if ((requeue_pi && !this->rt_waiter) ||
1272 (!requeue_pi && this->rt_waiter)) {
1273 ret = -EINVAL;
1274 break;
1278 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1279 * lock, we already woke the top_waiter. If not, it will be
1280 * woken by futex_unlock_pi().
1282 if (++task_count <= nr_wake && !requeue_pi) {
1283 wake_futex(this);
1284 continue;
1287 /* Ensure we requeue to the expected futex for requeue_pi. */
1288 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1289 ret = -EINVAL;
1290 break;
1294 * Requeue nr_requeue waiters and possibly one more in the case
1295 * of requeue_pi if we couldn't acquire the lock atomically.
1297 if (requeue_pi) {
1298 /* Prepare the waiter to take the rt_mutex. */
1299 atomic_inc(&pi_state->refcount);
1300 this->pi_state = pi_state;
1301 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1302 this->rt_waiter,
1303 this->task, 1);
1304 if (ret == 1) {
1305 /* We got the lock. */
1306 requeue_pi_wake_futex(this, &key2, hb2);
1307 drop_count++;
1308 continue;
1309 } else if (ret) {
1310 /* -EDEADLK */
1311 this->pi_state = NULL;
1312 free_pi_state(pi_state);
1313 goto out_unlock;
1316 requeue_futex(this, hb1, hb2, &key2);
1317 drop_count++;
1320 out_unlock:
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);
1332 out_put_keys:
1333 put_futex_key(fshared, &key2);
1334 out_put_key1:
1335 put_futex_key(fshared, &key1);
1336 out:
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);
1352 return hb;
1355 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1357 int prio;
1360 * The priority used to register this element is
1361 * - either the real thread-priority for the real-time threads
1362 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1363 * - or MAX_RT_PRIO for non-RT threads.
1364 * Thus, all RT-threads are woken first in priority order, and
1365 * the others are woken last, in FIFO order.
1367 prio = min(current->normal_prio, MAX_RT_PRIO);
1369 plist_node_init(&q->list, prio);
1370 #ifdef CONFIG_DEBUG_PI_LIST
1371 q->list.plist.lock = &hb->lock;
1372 #endif
1373 plist_add(&q->list, &hb->chain);
1374 q->task = current;
1375 spin_unlock(&hb->lock);
1378 static inline void
1379 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1381 spin_unlock(&hb->lock);
1382 drop_futex_key_refs(&q->key);
1386 * queue_me and unqueue_me must be called as a pair, each
1387 * exactly once. They are called with the hashed spinlock held.
1390 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1391 static int unqueue_me(struct futex_q *q)
1393 spinlock_t *lock_ptr;
1394 int ret = 0;
1396 /* In the common case we don't take the spinlock, which is nice. */
1397 retry:
1398 lock_ptr = q->lock_ptr;
1399 barrier();
1400 if (lock_ptr != NULL) {
1401 spin_lock(lock_ptr);
1403 * q->lock_ptr can change between reading it and
1404 * spin_lock(), causing us to take the wrong lock. This
1405 * corrects the race condition.
1407 * Reasoning goes like this: if we have the wrong lock,
1408 * q->lock_ptr must have changed (maybe several times)
1409 * between reading it and the spin_lock(). It can
1410 * change again after the spin_lock() but only if it was
1411 * already changed before the spin_lock(). It cannot,
1412 * however, change back to the original value. Therefore
1413 * we can detect whether we acquired the correct lock.
1415 if (unlikely(lock_ptr != q->lock_ptr)) {
1416 spin_unlock(lock_ptr);
1417 goto retry;
1419 WARN_ON(plist_node_empty(&q->list));
1420 plist_del(&q->list, &q->list.plist);
1422 BUG_ON(q->pi_state);
1424 spin_unlock(lock_ptr);
1425 ret = 1;
1428 drop_futex_key_refs(&q->key);
1429 return ret;
1433 * PI futexes can not be requeued and must remove themself from the
1434 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1435 * and dropped here.
1437 static void unqueue_me_pi(struct futex_q *q)
1439 WARN_ON(plist_node_empty(&q->list));
1440 plist_del(&q->list, &q->list.plist);
1442 BUG_ON(!q->pi_state);
1443 free_pi_state(q->pi_state);
1444 q->pi_state = NULL;
1446 spin_unlock(q->lock_ptr);
1448 drop_futex_key_refs(&q->key);
1452 * Fixup the pi_state owner with the new owner.
1454 * Must be called with hash bucket lock held and mm->sem held for non
1455 * private futexes.
1457 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1458 struct task_struct *newowner, int fshared)
1460 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1461 struct futex_pi_state *pi_state = q->pi_state;
1462 struct task_struct *oldowner = pi_state->owner;
1463 u32 uval, curval, newval;
1464 int ret;
1466 /* Owner died? */
1467 if (!pi_state->owner)
1468 newtid |= FUTEX_OWNER_DIED;
1471 * We are here either because we stole the rtmutex from the
1472 * pending owner or we are the pending owner which failed to
1473 * get the rtmutex. We have to replace the pending owner TID
1474 * in the user space variable. This must be atomic as we have
1475 * to preserve the owner died bit here.
1477 * Note: We write the user space value _before_ changing the pi_state
1478 * because we can fault here. Imagine swapped out pages or a fork
1479 * that marked all the anonymous memory readonly for cow.
1481 * Modifying pi_state _before_ the user space value would
1482 * leave the pi_state in an inconsistent state when we fault
1483 * here, because we need to drop the hash bucket lock to
1484 * handle the fault. This might be observed in the PID check
1485 * in lookup_pi_state.
1487 retry:
1488 if (get_futex_value_locked(&uval, uaddr))
1489 goto handle_fault;
1491 while (1) {
1492 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1494 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1496 if (curval == -EFAULT)
1497 goto handle_fault;
1498 if (curval == uval)
1499 break;
1500 uval = curval;
1504 * We fixed up user space. Now we need to fix the pi_state
1505 * itself.
1507 if (pi_state->owner != NULL) {
1508 spin_lock_irq(&pi_state->owner->pi_lock);
1509 WARN_ON(list_empty(&pi_state->list));
1510 list_del_init(&pi_state->list);
1511 spin_unlock_irq(&pi_state->owner->pi_lock);
1514 pi_state->owner = newowner;
1516 spin_lock_irq(&newowner->pi_lock);
1517 WARN_ON(!list_empty(&pi_state->list));
1518 list_add(&pi_state->list, &newowner->pi_state_list);
1519 spin_unlock_irq(&newowner->pi_lock);
1520 return 0;
1523 * To handle the page fault we need to drop the hash bucket
1524 * lock here. That gives the other task (either the pending
1525 * owner itself or the task which stole the rtmutex) the
1526 * chance to try the fixup of the pi_state. So once we are
1527 * back from handling the fault we need to check the pi_state
1528 * after reacquiring the hash bucket lock and before trying to
1529 * do another fixup. When the fixup has been done already we
1530 * simply return.
1532 handle_fault:
1533 spin_unlock(q->lock_ptr);
1535 ret = fault_in_user_writeable(uaddr);
1537 spin_lock(q->lock_ptr);
1540 * Check if someone else fixed it for us:
1542 if (pi_state->owner != oldowner)
1543 return 0;
1545 if (ret)
1546 return ret;
1548 goto retry;
1552 * In case we must use restart_block to restart a futex_wait,
1553 * we encode in the 'flags' shared capability
1555 #define FLAGS_SHARED 0x01
1556 #define FLAGS_CLOCKRT 0x02
1557 #define FLAGS_HAS_TIMEOUT 0x04
1559 static long futex_wait_restart(struct restart_block *restart);
1562 * fixup_owner() - Post lock pi_state and corner case management
1563 * @uaddr: user address of the futex
1564 * @fshared: whether the futex is shared (1) or not (0)
1565 * @q: futex_q (contains pi_state and access to the rt_mutex)
1566 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1568 * After attempting to lock an rt_mutex, this function is called to cleanup
1569 * the pi_state owner as well as handle race conditions that may allow us to
1570 * acquire the lock. Must be called with the hb lock held.
1572 * Returns:
1573 * 1 - success, lock taken
1574 * 0 - success, lock not taken
1575 * <0 - on error (-EFAULT)
1577 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1578 int locked)
1580 struct task_struct *owner;
1581 int ret = 0;
1583 if (locked) {
1585 * Got the lock. We might not be the anticipated owner if we
1586 * did a lock-steal - fix up the PI-state in that case:
1588 if (q->pi_state->owner != current)
1589 ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1590 goto out;
1594 * Catch the rare case, where the lock was released when we were on the
1595 * way back before we locked the hash bucket.
1597 if (q->pi_state->owner == current) {
1599 * Try to get the rt_mutex now. This might fail as some other
1600 * task acquired the rt_mutex after we removed ourself from the
1601 * rt_mutex waiters list.
1603 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1604 locked = 1;
1605 goto out;
1609 * pi_state is incorrect, some other task did a lock steal and
1610 * we returned due to timeout or signal without taking the
1611 * rt_mutex. Too late. We can access the rt_mutex_owner without
1612 * locking, as the other task is now blocked on the hash bucket
1613 * lock. Fix the state up.
1615 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1616 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1617 goto out;
1621 * Paranoia check. If we did not take the lock, then we should not be
1622 * the owner, nor the pending owner, of the rt_mutex.
1624 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1625 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1626 "pi-state %p\n", ret,
1627 q->pi_state->pi_mutex.owner,
1628 q->pi_state->owner);
1630 out:
1631 return ret ? ret : locked;
1635 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1636 * @hb: the futex hash bucket, must be locked by the caller
1637 * @q: the futex_q to queue up on
1638 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1640 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1641 struct hrtimer_sleeper *timeout)
1643 set_current_state(TASK_INTERRUPTIBLE);
1644 queue_me(q, hb);
1646 /* Arm the timer */
1647 if (timeout) {
1648 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1649 if (!hrtimer_active(&timeout->timer))
1650 timeout->task = NULL;
1654 * If we have been removed from the hash list, then another task
1655 * has tried to wake us, and we can skip the call to schedule().
1657 if (likely(!plist_node_empty(&q->list))) {
1659 * If the timer has already expired, current will already be
1660 * flagged for rescheduling. Only call schedule if there
1661 * is no timeout, or if it has yet to expire.
1663 if (!timeout || timeout->task)
1664 schedule();
1666 __set_current_state(TASK_RUNNING);
1670 * futex_wait_setup() - Prepare to wait on a futex
1671 * @uaddr: the futex userspace address
1672 * @val: the expected value
1673 * @fshared: whether the futex is shared (1) or not (0)
1674 * @q: the associated futex_q
1675 * @hb: storage for hash_bucket pointer to be returned to caller
1677 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1678 * compare it with the expected value. Handle atomic faults internally.
1679 * Return with the hb lock held and a q.key reference on success, and unlocked
1680 * with no q.key reference on failure.
1682 * Returns:
1683 * 0 - uaddr contains val and hb has been locked
1684 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1686 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1687 struct futex_q *q, struct futex_hash_bucket **hb)
1689 u32 uval;
1690 int ret;
1693 * Access the page AFTER the hash-bucket is locked.
1694 * Order is important:
1696 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1697 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1699 * The basic logical guarantee of a futex is that it blocks ONLY
1700 * if cond(var) is known to be true at the time of blocking, for
1701 * any cond. If we queued after testing *uaddr, that would open
1702 * a race condition where we could block indefinitely with
1703 * cond(var) false, which would violate the guarantee.
1705 * A consequence is that futex_wait() can return zero and absorb
1706 * a wakeup when *uaddr != val on entry to the syscall. This is
1707 * rare, but normal.
1709 retry:
1710 q->key = FUTEX_KEY_INIT;
1711 ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1712 if (unlikely(ret != 0))
1713 return ret;
1715 retry_private:
1716 *hb = queue_lock(q);
1718 ret = get_futex_value_locked(&uval, uaddr);
1720 if (ret) {
1721 queue_unlock(q, *hb);
1723 ret = get_user(uval, uaddr);
1724 if (ret)
1725 goto out;
1727 if (!fshared)
1728 goto retry_private;
1730 put_futex_key(fshared, &q->key);
1731 goto retry;
1734 if (uval != val) {
1735 queue_unlock(q, *hb);
1736 ret = -EWOULDBLOCK;
1739 out:
1740 if (ret)
1741 put_futex_key(fshared, &q->key);
1742 return ret;
1745 static int futex_wait(u32 __user *uaddr, int fshared,
1746 u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1748 struct hrtimer_sleeper timeout, *to = NULL;
1749 struct restart_block *restart;
1750 struct futex_hash_bucket *hb;
1751 struct futex_q q;
1752 int ret;
1754 if (!bitset)
1755 return -EINVAL;
1757 q.pi_state = NULL;
1758 q.bitset = bitset;
1759 q.rt_waiter = NULL;
1760 q.requeue_pi_key = NULL;
1762 if (abs_time) {
1763 to = &timeout;
1765 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1766 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1767 hrtimer_init_sleeper(to, current);
1768 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1769 current->timer_slack_ns);
1772 retry:
1773 /* Prepare to wait on uaddr. */
1774 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1775 if (ret)
1776 goto out;
1778 /* queue_me and wait for wakeup, timeout, or a signal. */
1779 futex_wait_queue_me(hb, &q, to);
1781 /* If we were woken (and unqueued), we succeeded, whatever. */
1782 ret = 0;
1783 if (!unqueue_me(&q))
1784 goto out_put_key;
1785 ret = -ETIMEDOUT;
1786 if (to && !to->task)
1787 goto out_put_key;
1790 * We expect signal_pending(current), but we might be the
1791 * victim of a spurious wakeup as well.
1793 if (!signal_pending(current)) {
1794 put_futex_key(fshared, &q.key);
1795 goto retry;
1798 ret = -ERESTARTSYS;
1799 if (!abs_time)
1800 goto out_put_key;
1802 restart = &current_thread_info()->restart_block;
1803 restart->fn = futex_wait_restart;
1804 restart->futex.uaddr = (u32 *)uaddr;
1805 restart->futex.val = val;
1806 restart->futex.time = abs_time->tv64;
1807 restart->futex.bitset = bitset;
1808 restart->futex.flags = FLAGS_HAS_TIMEOUT;
1810 if (fshared)
1811 restart->futex.flags |= FLAGS_SHARED;
1812 if (clockrt)
1813 restart->futex.flags |= FLAGS_CLOCKRT;
1815 ret = -ERESTART_RESTARTBLOCK;
1817 out_put_key:
1818 put_futex_key(fshared, &q.key);
1819 out:
1820 if (to) {
1821 hrtimer_cancel(&to->timer);
1822 destroy_hrtimer_on_stack(&to->timer);
1824 return ret;
1828 static long futex_wait_restart(struct restart_block *restart)
1830 u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1831 int fshared = 0;
1832 ktime_t t, *tp = NULL;
1834 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1835 t.tv64 = restart->futex.time;
1836 tp = &t;
1838 restart->fn = do_no_restart_syscall;
1839 if (restart->futex.flags & FLAGS_SHARED)
1840 fshared = 1;
1841 return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1842 restart->futex.bitset,
1843 restart->futex.flags & FLAGS_CLOCKRT);
1848 * Userspace tried a 0 -> TID atomic transition of the futex value
1849 * and failed. The kernel side here does the whole locking operation:
1850 * if there are waiters then it will block, it does PI, etc. (Due to
1851 * races the kernel might see a 0 value of the futex too.)
1853 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1854 int detect, ktime_t *time, int trylock)
1856 struct hrtimer_sleeper timeout, *to = NULL;
1857 struct futex_hash_bucket *hb;
1858 struct futex_q q;
1859 int res, ret;
1861 if (refill_pi_state_cache())
1862 return -ENOMEM;
1864 if (time) {
1865 to = &timeout;
1866 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1867 HRTIMER_MODE_ABS);
1868 hrtimer_init_sleeper(to, current);
1869 hrtimer_set_expires(&to->timer, *time);
1872 q.pi_state = NULL;
1873 q.rt_waiter = NULL;
1874 q.requeue_pi_key = NULL;
1875 retry:
1876 q.key = FUTEX_KEY_INIT;
1877 ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1878 if (unlikely(ret != 0))
1879 goto out;
1881 retry_private:
1882 hb = queue_lock(&q);
1884 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1885 if (unlikely(ret)) {
1886 switch (ret) {
1887 case 1:
1888 /* We got the lock. */
1889 ret = 0;
1890 goto out_unlock_put_key;
1891 case -EFAULT:
1892 goto uaddr_faulted;
1893 case -EAGAIN:
1895 * Task is exiting and we just wait for the
1896 * exit to complete.
1898 queue_unlock(&q, hb);
1899 put_futex_key(fshared, &q.key);
1900 cond_resched();
1901 goto retry;
1902 default:
1903 goto out_unlock_put_key;
1908 * Only actually queue now that the atomic ops are done:
1910 queue_me(&q, hb);
1912 WARN_ON(!q.pi_state);
1914 * Block on the PI mutex:
1916 if (!trylock)
1917 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1918 else {
1919 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1920 /* Fixup the trylock return value: */
1921 ret = ret ? 0 : -EWOULDBLOCK;
1924 spin_lock(q.lock_ptr);
1926 * Fixup the pi_state owner and possibly acquire the lock if we
1927 * haven't already.
1929 res = fixup_owner(uaddr, fshared, &q, !ret);
1931 * If fixup_owner() returned an error, proprogate that. If it acquired
1932 * the lock, clear our -ETIMEDOUT or -EINTR.
1934 if (res)
1935 ret = (res < 0) ? res : 0;
1938 * If fixup_owner() faulted and was unable to handle the fault, unlock
1939 * it and return the fault to userspace.
1941 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1942 rt_mutex_unlock(&q.pi_state->pi_mutex);
1944 /* Unqueue and drop the lock */
1945 unqueue_me_pi(&q);
1947 goto out;
1949 out_unlock_put_key:
1950 queue_unlock(&q, hb);
1952 out_put_key:
1953 put_futex_key(fshared, &q.key);
1954 out:
1955 if (to)
1956 destroy_hrtimer_on_stack(&to->timer);
1957 return ret != -EINTR ? ret : -ERESTARTNOINTR;
1959 uaddr_faulted:
1960 queue_unlock(&q, hb);
1962 ret = fault_in_user_writeable(uaddr);
1963 if (ret)
1964 goto out_put_key;
1966 if (!fshared)
1967 goto retry_private;
1969 put_futex_key(fshared, &q.key);
1970 goto retry;
1974 * Userspace attempted a TID -> 0 atomic transition, and failed.
1975 * This is the in-kernel slowpath: we look up the PI state (if any),
1976 * and do the rt-mutex unlock.
1978 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1980 struct futex_hash_bucket *hb;
1981 struct futex_q *this, *next;
1982 u32 uval;
1983 struct plist_head *head;
1984 union futex_key key = FUTEX_KEY_INIT;
1985 int ret;
1987 retry:
1988 if (get_user(uval, uaddr))
1989 return -EFAULT;
1991 * We release only a lock we actually own:
1993 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1994 return -EPERM;
1996 ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
1997 if (unlikely(ret != 0))
1998 goto out;
2000 hb = hash_futex(&key);
2001 spin_lock(&hb->lock);
2004 * To avoid races, try to do the TID -> 0 atomic transition
2005 * again. If it succeeds then we can return without waking
2006 * anyone else up:
2008 if (!(uval & FUTEX_OWNER_DIED))
2009 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2012 if (unlikely(uval == -EFAULT))
2013 goto pi_faulted;
2015 * Rare case: we managed to release the lock atomically,
2016 * no need to wake anyone else up:
2018 if (unlikely(uval == task_pid_vnr(current)))
2019 goto out_unlock;
2022 * Ok, other tasks may need to be woken up - check waiters
2023 * and do the wakeup if necessary:
2025 head = &hb->chain;
2027 plist_for_each_entry_safe(this, next, head, list) {
2028 if (!match_futex (&this->key, &key))
2029 continue;
2030 ret = wake_futex_pi(uaddr, uval, this);
2032 * The atomic access to the futex value
2033 * generated a pagefault, so retry the
2034 * user-access and the wakeup:
2036 if (ret == -EFAULT)
2037 goto pi_faulted;
2038 goto out_unlock;
2041 * No waiters - kernel unlocks the futex:
2043 if (!(uval & FUTEX_OWNER_DIED)) {
2044 ret = unlock_futex_pi(uaddr, uval);
2045 if (ret == -EFAULT)
2046 goto pi_faulted;
2049 out_unlock:
2050 spin_unlock(&hb->lock);
2051 put_futex_key(fshared, &key);
2053 out:
2054 return ret;
2056 pi_faulted:
2057 spin_unlock(&hb->lock);
2058 put_futex_key(fshared, &key);
2060 ret = fault_in_user_writeable(uaddr);
2061 if (!ret)
2062 goto retry;
2064 return ret;
2068 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2069 * @hb: the hash_bucket futex_q was original enqueued on
2070 * @q: the futex_q woken while waiting to be requeued
2071 * @key2: the futex_key of the requeue target futex
2072 * @timeout: the timeout associated with the wait (NULL if none)
2074 * Detect if the task was woken on the initial futex as opposed to the requeue
2075 * target futex. If so, determine if it was a timeout or a signal that caused
2076 * the wakeup and return the appropriate error code to the caller. Must be
2077 * called with the hb lock held.
2079 * Returns
2080 * 0 - no early wakeup detected
2081 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2083 static inline
2084 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2085 struct futex_q *q, union futex_key *key2,
2086 struct hrtimer_sleeper *timeout)
2088 int ret = 0;
2091 * With the hb lock held, we avoid races while we process the wakeup.
2092 * We only need to hold hb (and not hb2) to ensure atomicity as the
2093 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2094 * It can't be requeued from uaddr2 to something else since we don't
2095 * support a PI aware source futex for requeue.
2097 if (!match_futex(&q->key, key2)) {
2098 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2100 * We were woken prior to requeue by a timeout or a signal.
2101 * Unqueue the futex_q and determine which it was.
2103 plist_del(&q->list, &q->list.plist);
2105 /* Handle spurious wakeups gracefully */
2106 ret = -EAGAIN;
2107 if (timeout && !timeout->task)
2108 ret = -ETIMEDOUT;
2109 else if (signal_pending(current))
2110 ret = -ERESTARTNOINTR;
2112 return ret;
2116 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2117 * @uaddr: the futex we initialyl wait on (non-pi)
2118 * @fshared: whether the futexes are shared (1) or not (0). They must be
2119 * the same type, no requeueing from private to shared, etc.
2120 * @val: the expected value of uaddr
2121 * @abs_time: absolute timeout
2122 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all.
2123 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2124 * @uaddr2: the pi futex we will take prior to returning to user-space
2126 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2127 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2128 * complete the acquisition of the rt_mutex prior to returning to userspace.
2129 * This ensures the rt_mutex maintains an owner when it has waiters; without
2130 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2131 * need to.
2133 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2134 * via the following:
2135 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2136 * 2) wakeup on uaddr2 after a requeue and subsequent unlock
2137 * 3) signal (before or after requeue)
2138 * 4) timeout (before or after requeue)
2140 * If 3, we setup a restart_block with futex_wait_requeue_pi() as the function.
2142 * If 2, we may then block on trying to take the rt_mutex and return via:
2143 * 5) successful lock
2144 * 6) signal
2145 * 7) timeout
2146 * 8) other lock acquisition failure
2148 * If 6, we setup a restart_block with futex_lock_pi() as the function.
2150 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2152 * Returns:
2153 * 0 - On success
2154 * <0 - On error
2156 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2157 u32 val, ktime_t *abs_time, u32 bitset,
2158 int clockrt, u32 __user *uaddr2)
2160 struct hrtimer_sleeper timeout, *to = NULL;
2161 struct rt_mutex_waiter rt_waiter;
2162 struct rt_mutex *pi_mutex = NULL;
2163 struct futex_hash_bucket *hb;
2164 union futex_key key2;
2165 struct futex_q q;
2166 int res, ret;
2168 if (!bitset)
2169 return -EINVAL;
2171 if (abs_time) {
2172 to = &timeout;
2173 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2174 CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2175 hrtimer_init_sleeper(to, current);
2176 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2177 current->timer_slack_ns);
2181 * The waiter is allocated on our stack, manipulated by the requeue
2182 * code while we sleep on uaddr.
2184 debug_rt_mutex_init_waiter(&rt_waiter);
2185 rt_waiter.task = NULL;
2187 retry:
2188 key2 = FUTEX_KEY_INIT;
2189 ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2190 if (unlikely(ret != 0))
2191 goto out;
2193 q.pi_state = NULL;
2194 q.bitset = bitset;
2195 q.rt_waiter = &rt_waiter;
2196 q.requeue_pi_key = &key2;
2198 /* Prepare to wait on uaddr. */
2199 ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2200 if (ret)
2201 goto out_key2;
2203 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2204 futex_wait_queue_me(hb, &q, to);
2206 spin_lock(&hb->lock);
2207 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2208 spin_unlock(&hb->lock);
2209 if (ret)
2210 goto out_put_keys;
2213 * In order for us to be here, we know our q.key == key2, and since
2214 * we took the hb->lock above, we also know that futex_requeue() has
2215 * completed and we no longer have to concern ourselves with a wakeup
2216 * race with the atomic proxy lock acquition by the requeue code.
2219 /* Check if the requeue code acquired the second futex for us. */
2220 if (!q.rt_waiter) {
2222 * Got the lock. We might not be the anticipated owner if we
2223 * did a lock-steal - fix up the PI-state in that case.
2225 if (q.pi_state && (q.pi_state->owner != current)) {
2226 spin_lock(q.lock_ptr);
2227 ret = fixup_pi_state_owner(uaddr2, &q, current,
2228 fshared);
2229 spin_unlock(q.lock_ptr);
2231 } else {
2233 * We have been woken up by futex_unlock_pi(), a timeout, or a
2234 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2235 * the pi_state.
2237 WARN_ON(!&q.pi_state);
2238 pi_mutex = &q.pi_state->pi_mutex;
2239 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2240 debug_rt_mutex_free_waiter(&rt_waiter);
2242 spin_lock(q.lock_ptr);
2244 * Fixup the pi_state owner and possibly acquire the lock if we
2245 * haven't already.
2247 res = fixup_owner(uaddr2, fshared, &q, !ret);
2249 * If fixup_owner() returned an error, proprogate that. If it
2250 * acquired the lock, clear our -ETIMEDOUT or -EINTR.
2252 if (res)
2253 ret = (res < 0) ? res : 0;
2255 /* Unqueue and drop the lock. */
2256 unqueue_me_pi(&q);
2260 * If fixup_pi_state_owner() faulted and was unable to handle the
2261 * fault, unlock the rt_mutex and return the fault to userspace.
2263 if (ret == -EFAULT) {
2264 if (rt_mutex_owner(pi_mutex) == current)
2265 rt_mutex_unlock(pi_mutex);
2266 } else if (ret == -EINTR) {
2268 * We've already been requeued, but we have no way to
2269 * restart by calling futex_lock_pi() directly. We
2270 * could restart the syscall, but that will look at
2271 * the user space value and return right away. So we
2272 * drop back with EWOULDBLOCK to tell user space that
2273 * "val" has been changed. That's the same what the
2274 * restart of the syscall would do in
2275 * futex_wait_setup().
2277 ret = -EWOULDBLOCK;
2280 out_put_keys:
2281 put_futex_key(fshared, &q.key);
2282 out_key2:
2283 put_futex_key(fshared, &key2);
2285 /* Spurious wakeup ? */
2286 if (ret == -EAGAIN)
2287 goto retry;
2288 out:
2289 if (to) {
2290 hrtimer_cancel(&to->timer);
2291 destroy_hrtimer_on_stack(&to->timer);
2293 return ret;
2297 * Support for robust futexes: the kernel cleans up held futexes at
2298 * thread exit time.
2300 * Implementation: user-space maintains a per-thread list of locks it
2301 * is holding. Upon do_exit(), the kernel carefully walks this list,
2302 * and marks all locks that are owned by this thread with the
2303 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2304 * always manipulated with the lock held, so the list is private and
2305 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2306 * field, to allow the kernel to clean up if the thread dies after
2307 * acquiring the lock, but just before it could have added itself to
2308 * the list. There can only be one such pending lock.
2312 * sys_set_robust_list - set the robust-futex list head of a task
2313 * @head: pointer to the list-head
2314 * @len: length of the list-head, as userspace expects
2316 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2317 size_t, len)
2319 if (!futex_cmpxchg_enabled)
2320 return -ENOSYS;
2322 * The kernel knows only one size for now:
2324 if (unlikely(len != sizeof(*head)))
2325 return -EINVAL;
2327 current->robust_list = head;
2329 return 0;
2333 * sys_get_robust_list - get the robust-futex list head of a task
2334 * @pid: pid of the process [zero for current task]
2335 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2336 * @len_ptr: pointer to a length field, the kernel fills in the header size
2338 SYSCALL_DEFINE3(get_robust_list, int, pid,
2339 struct robust_list_head __user * __user *, head_ptr,
2340 size_t __user *, len_ptr)
2342 struct robust_list_head __user *head;
2343 unsigned long ret;
2344 const struct cred *cred = current_cred(), *pcred;
2346 if (!futex_cmpxchg_enabled)
2347 return -ENOSYS;
2349 if (!pid)
2350 head = current->robust_list;
2351 else {
2352 struct task_struct *p;
2354 ret = -ESRCH;
2355 rcu_read_lock();
2356 p = find_task_by_vpid(pid);
2357 if (!p)
2358 goto err_unlock;
2359 ret = -EPERM;
2360 pcred = __task_cred(p);
2361 if (cred->euid != pcred->euid &&
2362 cred->euid != pcred->uid &&
2363 !capable(CAP_SYS_PTRACE))
2364 goto err_unlock;
2365 head = p->robust_list;
2366 rcu_read_unlock();
2369 if (put_user(sizeof(*head), len_ptr))
2370 return -EFAULT;
2371 return put_user(head, head_ptr);
2373 err_unlock:
2374 rcu_read_unlock();
2376 return ret;
2380 * Process a futex-list entry, check whether it's owned by the
2381 * dying task, and do notification if so:
2383 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2385 u32 uval, nval, mval;
2387 retry:
2388 if (get_user(uval, uaddr))
2389 return -1;
2391 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2393 * Ok, this dying thread is truly holding a futex
2394 * of interest. Set the OWNER_DIED bit atomically
2395 * via cmpxchg, and if the value had FUTEX_WAITERS
2396 * set, wake up a waiter (if any). (We have to do a
2397 * futex_wake() even if OWNER_DIED is already set -
2398 * to handle the rare but possible case of recursive
2399 * thread-death.) The rest of the cleanup is done in
2400 * userspace.
2402 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2403 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2405 if (nval == -EFAULT)
2406 return -1;
2408 if (nval != uval)
2409 goto retry;
2412 * Wake robust non-PI futexes here. The wakeup of
2413 * PI futexes happens in exit_pi_state():
2415 if (!pi && (uval & FUTEX_WAITERS))
2416 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2418 return 0;
2422 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2424 static inline int fetch_robust_entry(struct robust_list __user **entry,
2425 struct robust_list __user * __user *head,
2426 int *pi)
2428 unsigned long uentry;
2430 if (get_user(uentry, (unsigned long __user *)head))
2431 return -EFAULT;
2433 *entry = (void __user *)(uentry & ~1UL);
2434 *pi = uentry & 1;
2436 return 0;
2440 * Walk curr->robust_list (very carefully, it's a userspace list!)
2441 * and mark any locks found there dead, and notify any waiters.
2443 * We silently return on any sign of list-walking problem.
2445 void exit_robust_list(struct task_struct *curr)
2447 struct robust_list_head __user *head = curr->robust_list;
2448 struct robust_list __user *entry, *next_entry, *pending;
2449 unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2450 unsigned long futex_offset;
2451 int rc;
2453 if (!futex_cmpxchg_enabled)
2454 return;
2457 * Fetch the list head (which was registered earlier, via
2458 * sys_set_robust_list()):
2460 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2461 return;
2463 * Fetch the relative futex offset:
2465 if (get_user(futex_offset, &head->futex_offset))
2466 return;
2468 * Fetch any possibly pending lock-add first, and handle it
2469 * if it exists:
2471 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2472 return;
2474 next_entry = NULL; /* avoid warning with gcc */
2475 while (entry != &head->list) {
2477 * Fetch the next entry in the list before calling
2478 * handle_futex_death:
2480 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2482 * A pending lock might already be on the list, so
2483 * don't process it twice:
2485 if (entry != pending)
2486 if (handle_futex_death((void __user *)entry + futex_offset,
2487 curr, pi))
2488 return;
2489 if (rc)
2490 return;
2491 entry = next_entry;
2492 pi = next_pi;
2494 * Avoid excessively long or circular lists:
2496 if (!--limit)
2497 break;
2499 cond_resched();
2502 if (pending)
2503 handle_futex_death((void __user *)pending + futex_offset,
2504 curr, pip);
2507 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2508 u32 __user *uaddr2, u32 val2, u32 val3)
2510 int clockrt, ret = -ENOSYS;
2511 int cmd = op & FUTEX_CMD_MASK;
2512 int fshared = 0;
2514 if (!(op & FUTEX_PRIVATE_FLAG))
2515 fshared = 1;
2517 clockrt = op & FUTEX_CLOCK_REALTIME;
2518 if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2519 return -ENOSYS;
2521 switch (cmd) {
2522 case FUTEX_WAIT:
2523 val3 = FUTEX_BITSET_MATCH_ANY;
2524 case FUTEX_WAIT_BITSET:
2525 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2526 break;
2527 case FUTEX_WAKE:
2528 val3 = FUTEX_BITSET_MATCH_ANY;
2529 case FUTEX_WAKE_BITSET:
2530 ret = futex_wake(uaddr, fshared, val, val3);
2531 break;
2532 case FUTEX_REQUEUE:
2533 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2534 break;
2535 case FUTEX_CMP_REQUEUE:
2536 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2538 break;
2539 case FUTEX_WAKE_OP:
2540 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2541 break;
2542 case FUTEX_LOCK_PI:
2543 if (futex_cmpxchg_enabled)
2544 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2545 break;
2546 case FUTEX_UNLOCK_PI:
2547 if (futex_cmpxchg_enabled)
2548 ret = futex_unlock_pi(uaddr, fshared);
2549 break;
2550 case FUTEX_TRYLOCK_PI:
2551 if (futex_cmpxchg_enabled)
2552 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2553 break;
2554 case FUTEX_WAIT_REQUEUE_PI:
2555 val3 = FUTEX_BITSET_MATCH_ANY;
2556 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2557 clockrt, uaddr2);
2558 break;
2559 case FUTEX_CMP_REQUEUE_PI:
2560 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2562 break;
2563 default:
2564 ret = -ENOSYS;
2566 return ret;
2570 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2571 struct timespec __user *, utime, u32 __user *, uaddr2,
2572 u32, val3)
2574 struct timespec ts;
2575 ktime_t t, *tp = NULL;
2576 u32 val2 = 0;
2577 int cmd = op & FUTEX_CMD_MASK;
2579 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2580 cmd == FUTEX_WAIT_BITSET ||
2581 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2582 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2583 return -EFAULT;
2584 if (!timespec_valid(&ts))
2585 return -EINVAL;
2587 t = timespec_to_ktime(ts);
2588 if (cmd == FUTEX_WAIT)
2589 t = ktime_add_safe(ktime_get(), t);
2590 tp = &t;
2593 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2594 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2596 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2597 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2598 val2 = (u32) (unsigned long) utime;
2600 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2603 static int __init futex_init(void)
2605 u32 curval;
2606 int i;
2609 * This will fail and we want it. Some arch implementations do
2610 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2611 * functionality. We want to know that before we call in any
2612 * of the complex code paths. Also we want to prevent
2613 * registration of robust lists in that case. NULL is
2614 * guaranteed to fault and we get -EFAULT on functional
2615 * implementation, the non functional ones will return
2616 * -ENOSYS.
2618 curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2619 if (curval == -EFAULT)
2620 futex_cmpxchg_enabled = 1;
2622 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2623 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2624 spin_lock_init(&futex_queues[i].lock);
2627 return 0;
2629 __initcall(futex_init);