edd: add default mode CONFIG_EDD_OFF=n, override with edd={on,off}
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
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.
14  *
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>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
58
59 #include <asm/futex.h>
60
61 #include "rtmutex_common.h"
62
63 int __read_mostly futex_cmpxchg_enabled;
64
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
66
67 /*
68  * Priority Inheritance state:
69  */
70 struct futex_pi_state {
71         /*
72          * list of 'owned' pi_state instances - these have to be
73          * cleaned up in do_exit() if the task exits prematurely:
74          */
75         struct list_head list;
76
77         /*
78          * The PI object:
79          */
80         struct rt_mutex pi_mutex;
81
82         struct task_struct *owner;
83         atomic_t refcount;
84
85         union futex_key key;
86 };
87
88 /*
89  * We use this hashed waitqueue instead of a normal wait_queue_t, so
90  * we can wake only the relevant ones (hashed queues may be shared).
91  *
92  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94  * The order of wakup is always to make the first condition true, then
95  * wake up q->waiters, then make the second condition true.
96  */
97 struct futex_q {
98         struct plist_node list;
99         wait_queue_head_t waiters;
100
101         /* Which hash list lock to use: */
102         spinlock_t *lock_ptr;
103
104         /* Key which the futex is hashed on: */
105         union futex_key key;
106
107         /* For fd, sigio sent using these: */
108         int fd;
109         struct file *filp;
110
111         /* Optional priority inheritance state: */
112         struct futex_pi_state *pi_state;
113         struct task_struct *task;
114
115         /* Bitset for the optional bitmasked wakeup */
116         u32 bitset;
117 };
118
119 /*
120  * Split the global futex_lock into every hash list lock.
121  */
122 struct futex_hash_bucket {
123         spinlock_t lock;
124         struct plist_head chain;
125 };
126
127 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
128
129 /* Futex-fs vfsmount entry: */
130 static struct vfsmount *futex_mnt;
131
132 /*
133  * Take mm->mmap_sem, when futex is shared
134  */
135 static inline void futex_lock_mm(struct rw_semaphore *fshared)
136 {
137         if (fshared)
138                 down_read(fshared);
139 }
140
141 /*
142  * Release mm->mmap_sem, when the futex is shared
143  */
144 static inline void futex_unlock_mm(struct rw_semaphore *fshared)
145 {
146         if (fshared)
147                 up_read(fshared);
148 }
149
150 /*
151  * We hash on the keys returned from get_futex_key (see below).
152  */
153 static struct futex_hash_bucket *hash_futex(union futex_key *key)
154 {
155         u32 hash = jhash2((u32*)&key->both.word,
156                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
157                           key->both.offset);
158         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
159 }
160
161 /*
162  * Return 1 if two futex_keys are equal, 0 otherwise.
163  */
164 static inline int match_futex(union futex_key *key1, union futex_key *key2)
165 {
166         return (key1->both.word == key2->both.word
167                 && key1->both.ptr == key2->both.ptr
168                 && key1->both.offset == key2->both.offset);
169 }
170
171 /**
172  * get_futex_key - Get parameters which are the keys for a futex.
173  * @uaddr: virtual address of the futex
174  * @shared: NULL for a PROCESS_PRIVATE futex,
175  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
176  * @key: address where result is stored.
177  *
178  * Returns a negative error code or 0
179  * The key words are stored in *key on success.
180  *
181  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
182  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
183  * We can usually work out the index without swapping in the page.
184  *
185  * fshared is NULL for PROCESS_PRIVATE futexes
186  * For other futexes, it points to &current->mm->mmap_sem and
187  * caller must have taken the reader lock. but NOT any spinlocks.
188  */
189 static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
190                          union futex_key *key)
191 {
192         unsigned long address = (unsigned long)uaddr;
193         struct mm_struct *mm = current->mm;
194         struct vm_area_struct *vma;
195         struct page *page;
196         int err;
197
198         /*
199          * The futex address must be "naturally" aligned.
200          */
201         key->both.offset = address % PAGE_SIZE;
202         if (unlikely((address % sizeof(u32)) != 0))
203                 return -EINVAL;
204         address -= key->both.offset;
205
206         /*
207          * PROCESS_PRIVATE futexes are fast.
208          * As the mm cannot disappear under us and the 'key' only needs
209          * virtual address, we dont even have to find the underlying vma.
210          * Note : We do have to check 'uaddr' is a valid user address,
211          *        but access_ok() should be faster than find_vma()
212          */
213         if (!fshared) {
214                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
215                         return -EFAULT;
216                 key->private.mm = mm;
217                 key->private.address = address;
218                 return 0;
219         }
220         /*
221          * The futex is hashed differently depending on whether
222          * it's in a shared or private mapping.  So check vma first.
223          */
224         vma = find_extend_vma(mm, address);
225         if (unlikely(!vma))
226                 return -EFAULT;
227
228         /*
229          * Permissions.
230          */
231         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
232                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
233
234         /*
235          * Private mappings are handled in a simple way.
236          *
237          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
238          * it's a read-only handle, it's expected that futexes attach to
239          * the object not the particular process.  Therefore we use
240          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
241          * mappings of _writable_ handles.
242          */
243         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
244                 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
245                 key->private.mm = mm;
246                 key->private.address = address;
247                 return 0;
248         }
249
250         /*
251          * Linear file mappings are also simple.
252          */
253         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
254         key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
255         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
256                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
257                                      + vma->vm_pgoff);
258                 return 0;
259         }
260
261         /*
262          * We could walk the page table to read the non-linear
263          * pte, and get the page index without fetching the page
264          * from swap.  But that's a lot of code to duplicate here
265          * for a rare case, so we simply fetch the page.
266          */
267         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
268         if (err >= 0) {
269                 key->shared.pgoff =
270                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
271                 put_page(page);
272                 return 0;
273         }
274         return err;
275 }
276
277 /*
278  * Take a reference to the resource addressed by a key.
279  * Can be called while holding spinlocks.
280  *
281  */
282 static void get_futex_key_refs(union futex_key *key)
283 {
284         if (key->both.ptr == NULL)
285                 return;
286         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
287                 case FUT_OFF_INODE:
288                         atomic_inc(&key->shared.inode->i_count);
289                         break;
290                 case FUT_OFF_MMSHARED:
291                         atomic_inc(&key->private.mm->mm_count);
292                         break;
293         }
294 }
295
296 /*
297  * Drop a reference to the resource addressed by a key.
298  * The hash bucket spinlock must not be held.
299  */
300 static void drop_futex_key_refs(union futex_key *key)
301 {
302         if (!key->both.ptr)
303                 return;
304         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
305                 case FUT_OFF_INODE:
306                         iput(key->shared.inode);
307                         break;
308                 case FUT_OFF_MMSHARED:
309                         mmdrop(key->private.mm);
310                         break;
311         }
312 }
313
314 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
315 {
316         u32 curval;
317
318         pagefault_disable();
319         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
320         pagefault_enable();
321
322         return curval;
323 }
324
325 static int get_futex_value_locked(u32 *dest, u32 __user *from)
326 {
327         int ret;
328
329         pagefault_disable();
330         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
331         pagefault_enable();
332
333         return ret ? -EFAULT : 0;
334 }
335
336 /*
337  * Fault handling.
338  * if fshared is non NULL, current->mm->mmap_sem is already held
339  */
340 static int futex_handle_fault(unsigned long address,
341                               struct rw_semaphore *fshared, int attempt)
342 {
343         struct vm_area_struct * vma;
344         struct mm_struct *mm = current->mm;
345         int ret = -EFAULT;
346
347         if (attempt > 2)
348                 return ret;
349
350         if (!fshared)
351                 down_read(&mm->mmap_sem);
352         vma = find_vma(mm, address);
353         if (vma && address >= vma->vm_start &&
354             (vma->vm_flags & VM_WRITE)) {
355                 int fault;
356                 fault = handle_mm_fault(mm, vma, address, 1);
357                 if (unlikely((fault & VM_FAULT_ERROR))) {
358 #if 0
359                         /* XXX: let's do this when we verify it is OK */
360                         if (ret & VM_FAULT_OOM)
361                                 ret = -ENOMEM;
362 #endif
363                 } else {
364                         ret = 0;
365                         if (fault & VM_FAULT_MAJOR)
366                                 current->maj_flt++;
367                         else
368                                 current->min_flt++;
369                 }
370         }
371         if (!fshared)
372                 up_read(&mm->mmap_sem);
373         return ret;
374 }
375
376 /*
377  * PI code:
378  */
379 static int refill_pi_state_cache(void)
380 {
381         struct futex_pi_state *pi_state;
382
383         if (likely(current->pi_state_cache))
384                 return 0;
385
386         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
387
388         if (!pi_state)
389                 return -ENOMEM;
390
391         INIT_LIST_HEAD(&pi_state->list);
392         /* pi_mutex gets initialized later */
393         pi_state->owner = NULL;
394         atomic_set(&pi_state->refcount, 1);
395
396         current->pi_state_cache = pi_state;
397
398         return 0;
399 }
400
401 static struct futex_pi_state * alloc_pi_state(void)
402 {
403         struct futex_pi_state *pi_state = current->pi_state_cache;
404
405         WARN_ON(!pi_state);
406         current->pi_state_cache = NULL;
407
408         return pi_state;
409 }
410
411 static void free_pi_state(struct futex_pi_state *pi_state)
412 {
413         if (!atomic_dec_and_test(&pi_state->refcount))
414                 return;
415
416         /*
417          * If pi_state->owner is NULL, the owner is most probably dying
418          * and has cleaned up the pi_state already
419          */
420         if (pi_state->owner) {
421                 spin_lock_irq(&pi_state->owner->pi_lock);
422                 list_del_init(&pi_state->list);
423                 spin_unlock_irq(&pi_state->owner->pi_lock);
424
425                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
426         }
427
428         if (current->pi_state_cache)
429                 kfree(pi_state);
430         else {
431                 /*
432                  * pi_state->list is already empty.
433                  * clear pi_state->owner.
434                  * refcount is at 0 - put it back to 1.
435                  */
436                 pi_state->owner = NULL;
437                 atomic_set(&pi_state->refcount, 1);
438                 current->pi_state_cache = pi_state;
439         }
440 }
441
442 /*
443  * Look up the task based on what TID userspace gave us.
444  * We dont trust it.
445  */
446 static struct task_struct * futex_find_get_task(pid_t pid)
447 {
448         struct task_struct *p;
449
450         rcu_read_lock();
451         p = find_task_by_vpid(pid);
452         if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
453                 p = ERR_PTR(-ESRCH);
454         else
455                 get_task_struct(p);
456
457         rcu_read_unlock();
458
459         return p;
460 }
461
462 /*
463  * This task is holding PI mutexes at exit time => bad.
464  * Kernel cleans up PI-state, but userspace is likely hosed.
465  * (Robust-futex cleanup is separate and might save the day for userspace.)
466  */
467 void exit_pi_state_list(struct task_struct *curr)
468 {
469         struct list_head *next, *head = &curr->pi_state_list;
470         struct futex_pi_state *pi_state;
471         struct futex_hash_bucket *hb;
472         union futex_key key;
473
474         if (!futex_cmpxchg_enabled)
475                 return;
476         /*
477          * We are a ZOMBIE and nobody can enqueue itself on
478          * pi_state_list anymore, but we have to be careful
479          * versus waiters unqueueing themselves:
480          */
481         spin_lock_irq(&curr->pi_lock);
482         while (!list_empty(head)) {
483
484                 next = head->next;
485                 pi_state = list_entry(next, struct futex_pi_state, list);
486                 key = pi_state->key;
487                 hb = hash_futex(&key);
488                 spin_unlock_irq(&curr->pi_lock);
489
490                 spin_lock(&hb->lock);
491
492                 spin_lock_irq(&curr->pi_lock);
493                 /*
494                  * We dropped the pi-lock, so re-check whether this
495                  * task still owns the PI-state:
496                  */
497                 if (head->next != next) {
498                         spin_unlock(&hb->lock);
499                         continue;
500                 }
501
502                 WARN_ON(pi_state->owner != curr);
503                 WARN_ON(list_empty(&pi_state->list));
504                 list_del_init(&pi_state->list);
505                 pi_state->owner = NULL;
506                 spin_unlock_irq(&curr->pi_lock);
507
508                 rt_mutex_unlock(&pi_state->pi_mutex);
509
510                 spin_unlock(&hb->lock);
511
512                 spin_lock_irq(&curr->pi_lock);
513         }
514         spin_unlock_irq(&curr->pi_lock);
515 }
516
517 static int
518 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
519                 union futex_key *key, struct futex_pi_state **ps)
520 {
521         struct futex_pi_state *pi_state = NULL;
522         struct futex_q *this, *next;
523         struct plist_head *head;
524         struct task_struct *p;
525         pid_t pid = uval & FUTEX_TID_MASK;
526
527         head = &hb->chain;
528
529         plist_for_each_entry_safe(this, next, head, list) {
530                 if (match_futex(&this->key, key)) {
531                         /*
532                          * Another waiter already exists - bump up
533                          * the refcount and return its pi_state:
534                          */
535                         pi_state = this->pi_state;
536                         /*
537                          * Userspace might have messed up non PI and PI futexes
538                          */
539                         if (unlikely(!pi_state))
540                                 return -EINVAL;
541
542                         WARN_ON(!atomic_read(&pi_state->refcount));
543                         WARN_ON(pid && pi_state->owner &&
544                                 pi_state->owner->pid != pid);
545
546                         atomic_inc(&pi_state->refcount);
547                         *ps = pi_state;
548
549                         return 0;
550                 }
551         }
552
553         /*
554          * We are the first waiter - try to look up the real owner and attach
555          * the new pi_state to it, but bail out when TID = 0
556          */
557         if (!pid)
558                 return -ESRCH;
559         p = futex_find_get_task(pid);
560         if (IS_ERR(p))
561                 return PTR_ERR(p);
562
563         /*
564          * We need to look at the task state flags to figure out,
565          * whether the task is exiting. To protect against the do_exit
566          * change of the task flags, we do this protected by
567          * p->pi_lock:
568          */
569         spin_lock_irq(&p->pi_lock);
570         if (unlikely(p->flags & PF_EXITING)) {
571                 /*
572                  * The task is on the way out. When PF_EXITPIDONE is
573                  * set, we know that the task has finished the
574                  * cleanup:
575                  */
576                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
577
578                 spin_unlock_irq(&p->pi_lock);
579                 put_task_struct(p);
580                 return ret;
581         }
582
583         pi_state = alloc_pi_state();
584
585         /*
586          * Initialize the pi_mutex in locked state and make 'p'
587          * the owner of it:
588          */
589         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
590
591         /* Store the key for possible exit cleanups: */
592         pi_state->key = *key;
593
594         WARN_ON(!list_empty(&pi_state->list));
595         list_add(&pi_state->list, &p->pi_state_list);
596         pi_state->owner = p;
597         spin_unlock_irq(&p->pi_lock);
598
599         put_task_struct(p);
600
601         *ps = pi_state;
602
603         return 0;
604 }
605
606 /*
607  * The hash bucket lock must be held when this is called.
608  * Afterwards, the futex_q must not be accessed.
609  */
610 static void wake_futex(struct futex_q *q)
611 {
612         plist_del(&q->list, &q->list.plist);
613         if (q->filp)
614                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
615         /*
616          * The lock in wake_up_all() is a crucial memory barrier after the
617          * plist_del() and also before assigning to q->lock_ptr.
618          */
619         wake_up_all(&q->waiters);
620         /*
621          * The waiting task can free the futex_q as soon as this is written,
622          * without taking any locks.  This must come last.
623          *
624          * A memory barrier is required here to prevent the following store
625          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
626          * at the end of wake_up_all() does not prevent this store from
627          * moving.
628          */
629         smp_wmb();
630         q->lock_ptr = NULL;
631 }
632
633 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
634 {
635         struct task_struct *new_owner;
636         struct futex_pi_state *pi_state = this->pi_state;
637         u32 curval, newval;
638
639         if (!pi_state)
640                 return -EINVAL;
641
642         spin_lock(&pi_state->pi_mutex.wait_lock);
643         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
644
645         /*
646          * This happens when we have stolen the lock and the original
647          * pending owner did not enqueue itself back on the rt_mutex.
648          * Thats not a tragedy. We know that way, that a lock waiter
649          * is on the fly. We make the futex_q waiter the pending owner.
650          */
651         if (!new_owner)
652                 new_owner = this->task;
653
654         /*
655          * We pass it to the next owner. (The WAITERS bit is always
656          * kept enabled while there is PI state around. We must also
657          * preserve the owner died bit.)
658          */
659         if (!(uval & FUTEX_OWNER_DIED)) {
660                 int ret = 0;
661
662                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
663
664                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
665
666                 if (curval == -EFAULT)
667                         ret = -EFAULT;
668                 else if (curval != uval)
669                         ret = -EINVAL;
670                 if (ret) {
671                         spin_unlock(&pi_state->pi_mutex.wait_lock);
672                         return ret;
673                 }
674         }
675
676         spin_lock_irq(&pi_state->owner->pi_lock);
677         WARN_ON(list_empty(&pi_state->list));
678         list_del_init(&pi_state->list);
679         spin_unlock_irq(&pi_state->owner->pi_lock);
680
681         spin_lock_irq(&new_owner->pi_lock);
682         WARN_ON(!list_empty(&pi_state->list));
683         list_add(&pi_state->list, &new_owner->pi_state_list);
684         pi_state->owner = new_owner;
685         spin_unlock_irq(&new_owner->pi_lock);
686
687         spin_unlock(&pi_state->pi_mutex.wait_lock);
688         rt_mutex_unlock(&pi_state->pi_mutex);
689
690         return 0;
691 }
692
693 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
694 {
695         u32 oldval;
696
697         /*
698          * There is no waiter, so we unlock the futex. The owner died
699          * bit has not to be preserved here. We are the owner:
700          */
701         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
702
703         if (oldval == -EFAULT)
704                 return oldval;
705         if (oldval != uval)
706                 return -EAGAIN;
707
708         return 0;
709 }
710
711 /*
712  * Express the locking dependencies for lockdep:
713  */
714 static inline void
715 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
716 {
717         if (hb1 <= hb2) {
718                 spin_lock(&hb1->lock);
719                 if (hb1 < hb2)
720                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
721         } else { /* hb1 > hb2 */
722                 spin_lock(&hb2->lock);
723                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
724         }
725 }
726
727 /*
728  * Wake up all waiters hashed on the physical page that is mapped
729  * to this virtual address:
730  */
731 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
732                       int nr_wake, u32 bitset)
733 {
734         struct futex_hash_bucket *hb;
735         struct futex_q *this, *next;
736         struct plist_head *head;
737         union futex_key key;
738         int ret;
739
740         if (!bitset)
741                 return -EINVAL;
742
743         futex_lock_mm(fshared);
744
745         ret = get_futex_key(uaddr, fshared, &key);
746         if (unlikely(ret != 0))
747                 goto out;
748
749         hb = hash_futex(&key);
750         spin_lock(&hb->lock);
751         head = &hb->chain;
752
753         plist_for_each_entry_safe(this, next, head, list) {
754                 if (match_futex (&this->key, &key)) {
755                         if (this->pi_state) {
756                                 ret = -EINVAL;
757                                 break;
758                         }
759
760                         /* Check if one of the bits is set in both bitsets */
761                         if (!(this->bitset & bitset))
762                                 continue;
763
764                         wake_futex(this);
765                         if (++ret >= nr_wake)
766                                 break;
767                 }
768         }
769
770         spin_unlock(&hb->lock);
771 out:
772         futex_unlock_mm(fshared);
773         return ret;
774 }
775
776 /*
777  * Wake up all waiters hashed on the physical page that is mapped
778  * to this virtual address:
779  */
780 static int
781 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
782               u32 __user *uaddr2,
783               int nr_wake, int nr_wake2, int op)
784 {
785         union futex_key key1, key2;
786         struct futex_hash_bucket *hb1, *hb2;
787         struct plist_head *head;
788         struct futex_q *this, *next;
789         int ret, op_ret, attempt = 0;
790
791 retryfull:
792         futex_lock_mm(fshared);
793
794         ret = get_futex_key(uaddr1, fshared, &key1);
795         if (unlikely(ret != 0))
796                 goto out;
797         ret = get_futex_key(uaddr2, fshared, &key2);
798         if (unlikely(ret != 0))
799                 goto out;
800
801         hb1 = hash_futex(&key1);
802         hb2 = hash_futex(&key2);
803
804 retry:
805         double_lock_hb(hb1, hb2);
806
807         op_ret = futex_atomic_op_inuser(op, uaddr2);
808         if (unlikely(op_ret < 0)) {
809                 u32 dummy;
810
811                 spin_unlock(&hb1->lock);
812                 if (hb1 != hb2)
813                         spin_unlock(&hb2->lock);
814
815 #ifndef CONFIG_MMU
816                 /*
817                  * we don't get EFAULT from MMU faults if we don't have an MMU,
818                  * but we might get them from range checking
819                  */
820                 ret = op_ret;
821                 goto out;
822 #endif
823
824                 if (unlikely(op_ret != -EFAULT)) {
825                         ret = op_ret;
826                         goto out;
827                 }
828
829                 /*
830                  * futex_atomic_op_inuser needs to both read and write
831                  * *(int __user *)uaddr2, but we can't modify it
832                  * non-atomically.  Therefore, if get_user below is not
833                  * enough, we need to handle the fault ourselves, while
834                  * still holding the mmap_sem.
835                  */
836                 if (attempt++) {
837                         ret = futex_handle_fault((unsigned long)uaddr2,
838                                                  fshared, attempt);
839                         if (ret)
840                                 goto out;
841                         goto retry;
842                 }
843
844                 /*
845                  * If we would have faulted, release mmap_sem,
846                  * fault it in and start all over again.
847                  */
848                 futex_unlock_mm(fshared);
849
850                 ret = get_user(dummy, uaddr2);
851                 if (ret)
852                         return ret;
853
854                 goto retryfull;
855         }
856
857         head = &hb1->chain;
858
859         plist_for_each_entry_safe(this, next, head, list) {
860                 if (match_futex (&this->key, &key1)) {
861                         wake_futex(this);
862                         if (++ret >= nr_wake)
863                                 break;
864                 }
865         }
866
867         if (op_ret > 0) {
868                 head = &hb2->chain;
869
870                 op_ret = 0;
871                 plist_for_each_entry_safe(this, next, head, list) {
872                         if (match_futex (&this->key, &key2)) {
873                                 wake_futex(this);
874                                 if (++op_ret >= nr_wake2)
875                                         break;
876                         }
877                 }
878                 ret += op_ret;
879         }
880
881         spin_unlock(&hb1->lock);
882         if (hb1 != hb2)
883                 spin_unlock(&hb2->lock);
884 out:
885         futex_unlock_mm(fshared);
886
887         return ret;
888 }
889
890 /*
891  * Requeue all waiters hashed on one physical page to another
892  * physical page.
893  */
894 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
895                          u32 __user *uaddr2,
896                          int nr_wake, int nr_requeue, u32 *cmpval)
897 {
898         union futex_key key1, key2;
899         struct futex_hash_bucket *hb1, *hb2;
900         struct plist_head *head1;
901         struct futex_q *this, *next;
902         int ret, drop_count = 0;
903
904  retry:
905         futex_lock_mm(fshared);
906
907         ret = get_futex_key(uaddr1, fshared, &key1);
908         if (unlikely(ret != 0))
909                 goto out;
910         ret = get_futex_key(uaddr2, fshared, &key2);
911         if (unlikely(ret != 0))
912                 goto out;
913
914         hb1 = hash_futex(&key1);
915         hb2 = hash_futex(&key2);
916
917         double_lock_hb(hb1, hb2);
918
919         if (likely(cmpval != NULL)) {
920                 u32 curval;
921
922                 ret = get_futex_value_locked(&curval, uaddr1);
923
924                 if (unlikely(ret)) {
925                         spin_unlock(&hb1->lock);
926                         if (hb1 != hb2)
927                                 spin_unlock(&hb2->lock);
928
929                         /*
930                          * If we would have faulted, release mmap_sem, fault
931                          * it in and start all over again.
932                          */
933                         futex_unlock_mm(fshared);
934
935                         ret = get_user(curval, uaddr1);
936
937                         if (!ret)
938                                 goto retry;
939
940                         return ret;
941                 }
942                 if (curval != *cmpval) {
943                         ret = -EAGAIN;
944                         goto out_unlock;
945                 }
946         }
947
948         head1 = &hb1->chain;
949         plist_for_each_entry_safe(this, next, head1, list) {
950                 if (!match_futex (&this->key, &key1))
951                         continue;
952                 if (++ret <= nr_wake) {
953                         wake_futex(this);
954                 } else {
955                         /*
956                          * If key1 and key2 hash to the same bucket, no need to
957                          * requeue.
958                          */
959                         if (likely(head1 != &hb2->chain)) {
960                                 plist_del(&this->list, &hb1->chain);
961                                 plist_add(&this->list, &hb2->chain);
962                                 this->lock_ptr = &hb2->lock;
963 #ifdef CONFIG_DEBUG_PI_LIST
964                                 this->list.plist.lock = &hb2->lock;
965 #endif
966                         }
967                         this->key = key2;
968                         get_futex_key_refs(&key2);
969                         drop_count++;
970
971                         if (ret - nr_wake >= nr_requeue)
972                                 break;
973                 }
974         }
975
976 out_unlock:
977         spin_unlock(&hb1->lock);
978         if (hb1 != hb2)
979                 spin_unlock(&hb2->lock);
980
981         /* drop_futex_key_refs() must be called outside the spinlocks. */
982         while (--drop_count >= 0)
983                 drop_futex_key_refs(&key1);
984
985 out:
986         futex_unlock_mm(fshared);
987         return ret;
988 }
989
990 /* The key must be already stored in q->key. */
991 static inline struct futex_hash_bucket *
992 queue_lock(struct futex_q *q, int fd, struct file *filp)
993 {
994         struct futex_hash_bucket *hb;
995
996         q->fd = fd;
997         q->filp = filp;
998
999         init_waitqueue_head(&q->waiters);
1000
1001         get_futex_key_refs(&q->key);
1002         hb = hash_futex(&q->key);
1003         q->lock_ptr = &hb->lock;
1004
1005         spin_lock(&hb->lock);
1006         return hb;
1007 }
1008
1009 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1010 {
1011         int prio;
1012
1013         /*
1014          * The priority used to register this element is
1015          * - either the real thread-priority for the real-time threads
1016          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1017          * - or MAX_RT_PRIO for non-RT threads.
1018          * Thus, all RT-threads are woken first in priority order, and
1019          * the others are woken last, in FIFO order.
1020          */
1021         prio = min(current->normal_prio, MAX_RT_PRIO);
1022
1023         plist_node_init(&q->list, prio);
1024 #ifdef CONFIG_DEBUG_PI_LIST
1025         q->list.plist.lock = &hb->lock;
1026 #endif
1027         plist_add(&q->list, &hb->chain);
1028         q->task = current;
1029         spin_unlock(&hb->lock);
1030 }
1031
1032 static inline void
1033 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1034 {
1035         spin_unlock(&hb->lock);
1036         drop_futex_key_refs(&q->key);
1037 }
1038
1039 /*
1040  * queue_me and unqueue_me must be called as a pair, each
1041  * exactly once.  They are called with the hashed spinlock held.
1042  */
1043
1044 /* The key must be already stored in q->key. */
1045 static void queue_me(struct futex_q *q, int fd, struct file *filp)
1046 {
1047         struct futex_hash_bucket *hb;
1048
1049         hb = queue_lock(q, fd, filp);
1050         __queue_me(q, hb);
1051 }
1052
1053 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1054 static int unqueue_me(struct futex_q *q)
1055 {
1056         spinlock_t *lock_ptr;
1057         int ret = 0;
1058
1059         /* In the common case we don't take the spinlock, which is nice. */
1060  retry:
1061         lock_ptr = q->lock_ptr;
1062         barrier();
1063         if (lock_ptr != NULL) {
1064                 spin_lock(lock_ptr);
1065                 /*
1066                  * q->lock_ptr can change between reading it and
1067                  * spin_lock(), causing us to take the wrong lock.  This
1068                  * corrects the race condition.
1069                  *
1070                  * Reasoning goes like this: if we have the wrong lock,
1071                  * q->lock_ptr must have changed (maybe several times)
1072                  * between reading it and the spin_lock().  It can
1073                  * change again after the spin_lock() but only if it was
1074                  * already changed before the spin_lock().  It cannot,
1075                  * however, change back to the original value.  Therefore
1076                  * we can detect whether we acquired the correct lock.
1077                  */
1078                 if (unlikely(lock_ptr != q->lock_ptr)) {
1079                         spin_unlock(lock_ptr);
1080                         goto retry;
1081                 }
1082                 WARN_ON(plist_node_empty(&q->list));
1083                 plist_del(&q->list, &q->list.plist);
1084
1085                 BUG_ON(q->pi_state);
1086
1087                 spin_unlock(lock_ptr);
1088                 ret = 1;
1089         }
1090
1091         drop_futex_key_refs(&q->key);
1092         return ret;
1093 }
1094
1095 /*
1096  * PI futexes can not be requeued and must remove themself from the
1097  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1098  * and dropped here.
1099  */
1100 static void unqueue_me_pi(struct futex_q *q)
1101 {
1102         WARN_ON(plist_node_empty(&q->list));
1103         plist_del(&q->list, &q->list.plist);
1104
1105         BUG_ON(!q->pi_state);
1106         free_pi_state(q->pi_state);
1107         q->pi_state = NULL;
1108
1109         spin_unlock(q->lock_ptr);
1110
1111         drop_futex_key_refs(&q->key);
1112 }
1113
1114 /*
1115  * Fixup the pi_state owner with the new owner.
1116  *
1117  * Must be called with hash bucket lock held and mm->sem held for non
1118  * private futexes.
1119  */
1120 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1121                                 struct task_struct *newowner)
1122 {
1123         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1124         struct futex_pi_state *pi_state = q->pi_state;
1125         u32 uval, curval, newval;
1126         int ret;
1127
1128         /* Owner died? */
1129         if (pi_state->owner != NULL) {
1130                 spin_lock_irq(&pi_state->owner->pi_lock);
1131                 WARN_ON(list_empty(&pi_state->list));
1132                 list_del_init(&pi_state->list);
1133                 spin_unlock_irq(&pi_state->owner->pi_lock);
1134         } else
1135                 newtid |= FUTEX_OWNER_DIED;
1136
1137         pi_state->owner = newowner;
1138
1139         spin_lock_irq(&newowner->pi_lock);
1140         WARN_ON(!list_empty(&pi_state->list));
1141         list_add(&pi_state->list, &newowner->pi_state_list);
1142         spin_unlock_irq(&newowner->pi_lock);
1143
1144         /*
1145          * We own it, so we have to replace the pending owner
1146          * TID. This must be atomic as we have preserve the
1147          * owner died bit here.
1148          */
1149         ret = get_futex_value_locked(&uval, uaddr);
1150
1151         while (!ret) {
1152                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1153
1154                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1155
1156                 if (curval == -EFAULT)
1157                         ret = -EFAULT;
1158                 if (curval == uval)
1159                         break;
1160                 uval = curval;
1161         }
1162         return ret;
1163 }
1164
1165 /*
1166  * In case we must use restart_block to restart a futex_wait,
1167  * we encode in the 'flags' shared capability
1168  */
1169 #define FLAGS_SHARED  1
1170
1171 static long futex_wait_restart(struct restart_block *restart);
1172
1173 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1174                       u32 val, ktime_t *abs_time, u32 bitset)
1175 {
1176         struct task_struct *curr = current;
1177         DECLARE_WAITQUEUE(wait, curr);
1178         struct futex_hash_bucket *hb;
1179         struct futex_q q;
1180         u32 uval;
1181         int ret;
1182         struct hrtimer_sleeper t;
1183         int rem = 0;
1184
1185         if (!bitset)
1186                 return -EINVAL;
1187
1188         q.pi_state = NULL;
1189         q.bitset = bitset;
1190  retry:
1191         futex_lock_mm(fshared);
1192
1193         ret = get_futex_key(uaddr, fshared, &q.key);
1194         if (unlikely(ret != 0))
1195                 goto out_release_sem;
1196
1197         hb = queue_lock(&q, -1, NULL);
1198
1199         /*
1200          * Access the page AFTER the futex is queued.
1201          * Order is important:
1202          *
1203          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1204          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1205          *
1206          * The basic logical guarantee of a futex is that it blocks ONLY
1207          * if cond(var) is known to be true at the time of blocking, for
1208          * any cond.  If we queued after testing *uaddr, that would open
1209          * a race condition where we could block indefinitely with
1210          * cond(var) false, which would violate the guarantee.
1211          *
1212          * A consequence is that futex_wait() can return zero and absorb
1213          * a wakeup when *uaddr != val on entry to the syscall.  This is
1214          * rare, but normal.
1215          *
1216          * for shared futexes, we hold the mmap semaphore, so the mapping
1217          * cannot have changed since we looked it up in get_futex_key.
1218          */
1219         ret = get_futex_value_locked(&uval, uaddr);
1220
1221         if (unlikely(ret)) {
1222                 queue_unlock(&q, hb);
1223
1224                 /*
1225                  * If we would have faulted, release mmap_sem, fault it in and
1226                  * start all over again.
1227                  */
1228                 futex_unlock_mm(fshared);
1229
1230                 ret = get_user(uval, uaddr);
1231
1232                 if (!ret)
1233                         goto retry;
1234                 return ret;
1235         }
1236         ret = -EWOULDBLOCK;
1237         if (uval != val)
1238                 goto out_unlock_release_sem;
1239
1240         /* Only actually queue if *uaddr contained val.  */
1241         __queue_me(&q, hb);
1242
1243         /*
1244          * Now the futex is queued and we have checked the data, we
1245          * don't want to hold mmap_sem while we sleep.
1246          */
1247         futex_unlock_mm(fshared);
1248
1249         /*
1250          * There might have been scheduling since the queue_me(), as we
1251          * cannot hold a spinlock across the get_user() in case it
1252          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1253          * queueing ourselves into the futex hash.  This code thus has to
1254          * rely on the futex_wake() code removing us from hash when it
1255          * wakes us up.
1256          */
1257
1258         /* add_wait_queue is the barrier after __set_current_state. */
1259         __set_current_state(TASK_INTERRUPTIBLE);
1260         add_wait_queue(&q.waiters, &wait);
1261         /*
1262          * !plist_node_empty() is safe here without any lock.
1263          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1264          */
1265         if (likely(!plist_node_empty(&q.list))) {
1266                 if (!abs_time)
1267                         schedule();
1268                 else {
1269                         hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1270                         hrtimer_init_sleeper(&t, current);
1271                         t.timer.expires = *abs_time;
1272
1273                         hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
1274                         if (!hrtimer_active(&t.timer))
1275                                 t.task = NULL;
1276
1277                         /*
1278                          * the timer could have already expired, in which
1279                          * case current would be flagged for rescheduling.
1280                          * Don't bother calling schedule.
1281                          */
1282                         if (likely(t.task))
1283                                 schedule();
1284
1285                         hrtimer_cancel(&t.timer);
1286
1287                         /* Flag if a timeout occured */
1288                         rem = (t.task == NULL);
1289                 }
1290         }
1291         __set_current_state(TASK_RUNNING);
1292
1293         /*
1294          * NOTE: we don't remove ourselves from the waitqueue because
1295          * we are the only user of it.
1296          */
1297
1298         /* If we were woken (and unqueued), we succeeded, whatever. */
1299         if (!unqueue_me(&q))
1300                 return 0;
1301         if (rem)
1302                 return -ETIMEDOUT;
1303
1304         /*
1305          * We expect signal_pending(current), but another thread may
1306          * have handled it for us already.
1307          */
1308         if (!abs_time)
1309                 return -ERESTARTSYS;
1310         else {
1311                 struct restart_block *restart;
1312                 restart = &current_thread_info()->restart_block;
1313                 restart->fn = futex_wait_restart;
1314                 restart->futex.uaddr = (u32 *)uaddr;
1315                 restart->futex.val = val;
1316                 restart->futex.time = abs_time->tv64;
1317                 restart->futex.bitset = bitset;
1318                 restart->futex.flags = 0;
1319
1320                 if (fshared)
1321                         restart->futex.flags |= FLAGS_SHARED;
1322                 return -ERESTART_RESTARTBLOCK;
1323         }
1324
1325  out_unlock_release_sem:
1326         queue_unlock(&q, hb);
1327
1328  out_release_sem:
1329         futex_unlock_mm(fshared);
1330         return ret;
1331 }
1332
1333
1334 static long futex_wait_restart(struct restart_block *restart)
1335 {
1336         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1337         struct rw_semaphore *fshared = NULL;
1338         ktime_t t;
1339
1340         t.tv64 = restart->futex.time;
1341         restart->fn = do_no_restart_syscall;
1342         if (restart->futex.flags & FLAGS_SHARED)
1343                 fshared = &current->mm->mmap_sem;
1344         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1345                                 restart->futex.bitset);
1346 }
1347
1348
1349 /*
1350  * Userspace tried a 0 -> TID atomic transition of the futex value
1351  * and failed. The kernel side here does the whole locking operation:
1352  * if there are waiters then it will block, it does PI, etc. (Due to
1353  * races the kernel might see a 0 value of the futex too.)
1354  */
1355 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1356                          int detect, ktime_t *time, int trylock)
1357 {
1358         struct hrtimer_sleeper timeout, *to = NULL;
1359         struct task_struct *curr = current;
1360         struct futex_hash_bucket *hb;
1361         u32 uval, newval, curval;
1362         struct futex_q q;
1363         int ret, lock_taken, ownerdied = 0, attempt = 0;
1364
1365         if (refill_pi_state_cache())
1366                 return -ENOMEM;
1367
1368         if (time) {
1369                 to = &timeout;
1370                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1371                 hrtimer_init_sleeper(to, current);
1372                 to->timer.expires = *time;
1373         }
1374
1375         q.pi_state = NULL;
1376  retry:
1377         futex_lock_mm(fshared);
1378
1379         ret = get_futex_key(uaddr, fshared, &q.key);
1380         if (unlikely(ret != 0))
1381                 goto out_release_sem;
1382
1383  retry_unlocked:
1384         hb = queue_lock(&q, -1, NULL);
1385
1386  retry_locked:
1387         ret = lock_taken = 0;
1388
1389         /*
1390          * To avoid races, we attempt to take the lock here again
1391          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1392          * the locks. It will most likely not succeed.
1393          */
1394         newval = task_pid_vnr(current);
1395
1396         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1397
1398         if (unlikely(curval == -EFAULT))
1399                 goto uaddr_faulted;
1400
1401         /*
1402          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1403          * situation and we return success to user space.
1404          */
1405         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1406                 ret = -EDEADLK;
1407                 goto out_unlock_release_sem;
1408         }
1409
1410         /*
1411          * Surprise - we got the lock. Just return to userspace:
1412          */
1413         if (unlikely(!curval))
1414                 goto out_unlock_release_sem;
1415
1416         uval = curval;
1417
1418         /*
1419          * Set the WAITERS flag, so the owner will know it has someone
1420          * to wake at next unlock
1421          */
1422         newval = curval | FUTEX_WAITERS;
1423
1424         /*
1425          * There are two cases, where a futex might have no owner (the
1426          * owner TID is 0): OWNER_DIED. We take over the futex in this
1427          * case. We also do an unconditional take over, when the owner
1428          * of the futex died.
1429          *
1430          * This is safe as we are protected by the hash bucket lock !
1431          */
1432         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1433                 /* Keep the OWNER_DIED bit */
1434                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1435                 ownerdied = 0;
1436                 lock_taken = 1;
1437         }
1438
1439         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1440
1441         if (unlikely(curval == -EFAULT))
1442                 goto uaddr_faulted;
1443         if (unlikely(curval != uval))
1444                 goto retry_locked;
1445
1446         /*
1447          * We took the lock due to owner died take over.
1448          */
1449         if (unlikely(lock_taken))
1450                 goto out_unlock_release_sem;
1451
1452         /*
1453          * We dont have the lock. Look up the PI state (or create it if
1454          * we are the first waiter):
1455          */
1456         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1457
1458         if (unlikely(ret)) {
1459                 switch (ret) {
1460
1461                 case -EAGAIN:
1462                         /*
1463                          * Task is exiting and we just wait for the
1464                          * exit to complete.
1465                          */
1466                         queue_unlock(&q, hb);
1467                         futex_unlock_mm(fshared);
1468                         cond_resched();
1469                         goto retry;
1470
1471                 case -ESRCH:
1472                         /*
1473                          * No owner found for this futex. Check if the
1474                          * OWNER_DIED bit is set to figure out whether
1475                          * this is a robust futex or not.
1476                          */
1477                         if (get_futex_value_locked(&curval, uaddr))
1478                                 goto uaddr_faulted;
1479
1480                         /*
1481                          * We simply start over in case of a robust
1482                          * futex. The code above will take the futex
1483                          * and return happy.
1484                          */
1485                         if (curval & FUTEX_OWNER_DIED) {
1486                                 ownerdied = 1;
1487                                 goto retry_locked;
1488                         }
1489                 default:
1490                         goto out_unlock_release_sem;
1491                 }
1492         }
1493
1494         /*
1495          * Only actually queue now that the atomic ops are done:
1496          */
1497         __queue_me(&q, hb);
1498
1499         /*
1500          * Now the futex is queued and we have checked the data, we
1501          * don't want to hold mmap_sem while we sleep.
1502          */
1503         futex_unlock_mm(fshared);
1504
1505         WARN_ON(!q.pi_state);
1506         /*
1507          * Block on the PI mutex:
1508          */
1509         if (!trylock)
1510                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1511         else {
1512                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1513                 /* Fixup the trylock return value: */
1514                 ret = ret ? 0 : -EWOULDBLOCK;
1515         }
1516
1517         futex_lock_mm(fshared);
1518         spin_lock(q.lock_ptr);
1519
1520         if (!ret) {
1521                 /*
1522                  * Got the lock. We might not be the anticipated owner
1523                  * if we did a lock-steal - fix up the PI-state in
1524                  * that case:
1525                  */
1526                 if (q.pi_state->owner != curr)
1527                         ret = fixup_pi_state_owner(uaddr, &q, curr);
1528         } else {
1529                 /*
1530                  * Catch the rare case, where the lock was released
1531                  * when we were on the way back before we locked the
1532                  * hash bucket.
1533                  */
1534                 if (q.pi_state->owner == curr) {
1535                         /*
1536                          * Try to get the rt_mutex now. This might
1537                          * fail as some other task acquired the
1538                          * rt_mutex after we removed ourself from the
1539                          * rt_mutex waiters list.
1540                          */
1541                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1542                                 ret = 0;
1543                         else {
1544                                 /*
1545                                  * pi_state is incorrect, some other
1546                                  * task did a lock steal and we
1547                                  * returned due to timeout or signal
1548                                  * without taking the rt_mutex. Too
1549                                  * late. We can access the
1550                                  * rt_mutex_owner without locking, as
1551                                  * the other task is now blocked on
1552                                  * the hash bucket lock. Fix the state
1553                                  * up.
1554                                  */
1555                                 struct task_struct *owner;
1556                                 int res;
1557
1558                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1559                                 res = fixup_pi_state_owner(uaddr, &q, owner);
1560
1561                                 /* propagate -EFAULT, if the fixup failed */
1562                                 if (res)
1563                                         ret = res;
1564                         }
1565                 } else {
1566                         /*
1567                          * Paranoia check. If we did not take the lock
1568                          * in the trylock above, then we should not be
1569                          * the owner of the rtmutex, neither the real
1570                          * nor the pending one:
1571                          */
1572                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1573                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1574                                        "pi-mutex: %p pi-state %p\n", ret,
1575                                        q.pi_state->pi_mutex.owner,
1576                                        q.pi_state->owner);
1577                 }
1578         }
1579
1580         /* Unqueue and drop the lock */
1581         unqueue_me_pi(&q);
1582         futex_unlock_mm(fshared);
1583
1584         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1585
1586  out_unlock_release_sem:
1587         queue_unlock(&q, hb);
1588
1589  out_release_sem:
1590         futex_unlock_mm(fshared);
1591         return ret;
1592
1593  uaddr_faulted:
1594         /*
1595          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1596          * non-atomically.  Therefore, if get_user below is not
1597          * enough, we need to handle the fault ourselves, while
1598          * still holding the mmap_sem.
1599          *
1600          * ... and hb->lock. :-) --ANK
1601          */
1602         queue_unlock(&q, hb);
1603
1604         if (attempt++) {
1605                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1606                                          attempt);
1607                 if (ret)
1608                         goto out_release_sem;
1609                 goto retry_unlocked;
1610         }
1611
1612         futex_unlock_mm(fshared);
1613
1614         ret = get_user(uval, uaddr);
1615         if (!ret && (uval != -EFAULT))
1616                 goto retry;
1617
1618         return ret;
1619 }
1620
1621 /*
1622  * Userspace attempted a TID -> 0 atomic transition, and failed.
1623  * This is the in-kernel slowpath: we look up the PI state (if any),
1624  * and do the rt-mutex unlock.
1625  */
1626 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1627 {
1628         struct futex_hash_bucket *hb;
1629         struct futex_q *this, *next;
1630         u32 uval;
1631         struct plist_head *head;
1632         union futex_key key;
1633         int ret, attempt = 0;
1634
1635 retry:
1636         if (get_user(uval, uaddr))
1637                 return -EFAULT;
1638         /*
1639          * We release only a lock we actually own:
1640          */
1641         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1642                 return -EPERM;
1643         /*
1644          * First take all the futex related locks:
1645          */
1646         futex_lock_mm(fshared);
1647
1648         ret = get_futex_key(uaddr, fshared, &key);
1649         if (unlikely(ret != 0))
1650                 goto out;
1651
1652         hb = hash_futex(&key);
1653 retry_unlocked:
1654         spin_lock(&hb->lock);
1655
1656         /*
1657          * To avoid races, try to do the TID -> 0 atomic transition
1658          * again. If it succeeds then we can return without waking
1659          * anyone else up:
1660          */
1661         if (!(uval & FUTEX_OWNER_DIED))
1662                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1663
1664
1665         if (unlikely(uval == -EFAULT))
1666                 goto pi_faulted;
1667         /*
1668          * Rare case: we managed to release the lock atomically,
1669          * no need to wake anyone else up:
1670          */
1671         if (unlikely(uval == task_pid_vnr(current)))
1672                 goto out_unlock;
1673
1674         /*
1675          * Ok, other tasks may need to be woken up - check waiters
1676          * and do the wakeup if necessary:
1677          */
1678         head = &hb->chain;
1679
1680         plist_for_each_entry_safe(this, next, head, list) {
1681                 if (!match_futex (&this->key, &key))
1682                         continue;
1683                 ret = wake_futex_pi(uaddr, uval, this);
1684                 /*
1685                  * The atomic access to the futex value
1686                  * generated a pagefault, so retry the
1687                  * user-access and the wakeup:
1688                  */
1689                 if (ret == -EFAULT)
1690                         goto pi_faulted;
1691                 goto out_unlock;
1692         }
1693         /*
1694          * No waiters - kernel unlocks the futex:
1695          */
1696         if (!(uval & FUTEX_OWNER_DIED)) {
1697                 ret = unlock_futex_pi(uaddr, uval);
1698                 if (ret == -EFAULT)
1699                         goto pi_faulted;
1700         }
1701
1702 out_unlock:
1703         spin_unlock(&hb->lock);
1704 out:
1705         futex_unlock_mm(fshared);
1706
1707         return ret;
1708
1709 pi_faulted:
1710         /*
1711          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1712          * non-atomically.  Therefore, if get_user below is not
1713          * enough, we need to handle the fault ourselves, while
1714          * still holding the mmap_sem.
1715          *
1716          * ... and hb->lock. --ANK
1717          */
1718         spin_unlock(&hb->lock);
1719
1720         if (attempt++) {
1721                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1722                                          attempt);
1723                 if (ret)
1724                         goto out;
1725                 uval = 0;
1726                 goto retry_unlocked;
1727         }
1728
1729         futex_unlock_mm(fshared);
1730
1731         ret = get_user(uval, uaddr);
1732         if (!ret && (uval != -EFAULT))
1733                 goto retry;
1734
1735         return ret;
1736 }
1737
1738 static int futex_close(struct inode *inode, struct file *filp)
1739 {
1740         struct futex_q *q = filp->private_data;
1741
1742         unqueue_me(q);
1743         kfree(q);
1744
1745         return 0;
1746 }
1747
1748 /* This is one-shot: once it's gone off you need a new fd */
1749 static unsigned int futex_poll(struct file *filp,
1750                                struct poll_table_struct *wait)
1751 {
1752         struct futex_q *q = filp->private_data;
1753         int ret = 0;
1754
1755         poll_wait(filp, &q->waiters, wait);
1756
1757         /*
1758          * plist_node_empty() is safe here without any lock.
1759          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1760          */
1761         if (plist_node_empty(&q->list))
1762                 ret = POLLIN | POLLRDNORM;
1763
1764         return ret;
1765 }
1766
1767 static const struct file_operations futex_fops = {
1768         .release        = futex_close,
1769         .poll           = futex_poll,
1770 };
1771
1772 /*
1773  * Signal allows caller to avoid the race which would occur if they
1774  * set the sigio stuff up afterwards.
1775  */
1776 static int futex_fd(u32 __user *uaddr, int signal)
1777 {
1778         struct futex_q *q;
1779         struct file *filp;
1780         int ret, err;
1781         struct rw_semaphore *fshared;
1782         static unsigned long printk_interval;
1783
1784         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1785                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1786                        "will be removed from the kernel in June 2007\n",
1787                        current->comm);
1788         }
1789
1790         ret = -EINVAL;
1791         if (!valid_signal(signal))
1792                 goto out;
1793
1794         ret = get_unused_fd();
1795         if (ret < 0)
1796                 goto out;
1797         filp = get_empty_filp();
1798         if (!filp) {
1799                 put_unused_fd(ret);
1800                 ret = -ENFILE;
1801                 goto out;
1802         }
1803         filp->f_op = &futex_fops;
1804         filp->f_path.mnt = mntget(futex_mnt);
1805         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1806         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1807
1808         if (signal) {
1809                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1810                 if (err < 0) {
1811                         goto error;
1812                 }
1813                 filp->f_owner.signum = signal;
1814         }
1815
1816         q = kmalloc(sizeof(*q), GFP_KERNEL);
1817         if (!q) {
1818                 err = -ENOMEM;
1819                 goto error;
1820         }
1821         q->pi_state = NULL;
1822
1823         fshared = &current->mm->mmap_sem;
1824         down_read(fshared);
1825         err = get_futex_key(uaddr, fshared, &q->key);
1826
1827         if (unlikely(err != 0)) {
1828                 up_read(fshared);
1829                 kfree(q);
1830                 goto error;
1831         }
1832
1833         /*
1834          * queue_me() must be called before releasing mmap_sem, because
1835          * key->shared.inode needs to be referenced while holding it.
1836          */
1837         filp->private_data = q;
1838
1839         queue_me(q, ret, filp);
1840         up_read(fshared);
1841
1842         /* Now we map fd to filp, so userspace can access it */
1843         fd_install(ret, filp);
1844 out:
1845         return ret;
1846 error:
1847         put_unused_fd(ret);
1848         put_filp(filp);
1849         ret = err;
1850         goto out;
1851 }
1852
1853 /*
1854  * Support for robust futexes: the kernel cleans up held futexes at
1855  * thread exit time.
1856  *
1857  * Implementation: user-space maintains a per-thread list of locks it
1858  * is holding. Upon do_exit(), the kernel carefully walks this list,
1859  * and marks all locks that are owned by this thread with the
1860  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1861  * always manipulated with the lock held, so the list is private and
1862  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1863  * field, to allow the kernel to clean up if the thread dies after
1864  * acquiring the lock, but just before it could have added itself to
1865  * the list. There can only be one such pending lock.
1866  */
1867
1868 /**
1869  * sys_set_robust_list - set the robust-futex list head of a task
1870  * @head: pointer to the list-head
1871  * @len: length of the list-head, as userspace expects
1872  */
1873 asmlinkage long
1874 sys_set_robust_list(struct robust_list_head __user *head,
1875                     size_t len)
1876 {
1877         if (!futex_cmpxchg_enabled)
1878                 return -ENOSYS;
1879         /*
1880          * The kernel knows only one size for now:
1881          */
1882         if (unlikely(len != sizeof(*head)))
1883                 return -EINVAL;
1884
1885         current->robust_list = head;
1886
1887         return 0;
1888 }
1889
1890 /**
1891  * sys_get_robust_list - get the robust-futex list head of a task
1892  * @pid: pid of the process [zero for current task]
1893  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1894  * @len_ptr: pointer to a length field, the kernel fills in the header size
1895  */
1896 asmlinkage long
1897 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1898                     size_t __user *len_ptr)
1899 {
1900         struct robust_list_head __user *head;
1901         unsigned long ret;
1902
1903         if (!futex_cmpxchg_enabled)
1904                 return -ENOSYS;
1905
1906         if (!pid)
1907                 head = current->robust_list;
1908         else {
1909                 struct task_struct *p;
1910
1911                 ret = -ESRCH;
1912                 rcu_read_lock();
1913                 p = find_task_by_vpid(pid);
1914                 if (!p)
1915                         goto err_unlock;
1916                 ret = -EPERM;
1917                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1918                                 !capable(CAP_SYS_PTRACE))
1919                         goto err_unlock;
1920                 head = p->robust_list;
1921                 rcu_read_unlock();
1922         }
1923
1924         if (put_user(sizeof(*head), len_ptr))
1925                 return -EFAULT;
1926         return put_user(head, head_ptr);
1927
1928 err_unlock:
1929         rcu_read_unlock();
1930
1931         return ret;
1932 }
1933
1934 /*
1935  * Process a futex-list entry, check whether it's owned by the
1936  * dying task, and do notification if so:
1937  */
1938 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1939 {
1940         u32 uval, nval, mval;
1941
1942 retry:
1943         if (get_user(uval, uaddr))
1944                 return -1;
1945
1946         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1947                 /*
1948                  * Ok, this dying thread is truly holding a futex
1949                  * of interest. Set the OWNER_DIED bit atomically
1950                  * via cmpxchg, and if the value had FUTEX_WAITERS
1951                  * set, wake up a waiter (if any). (We have to do a
1952                  * futex_wake() even if OWNER_DIED is already set -
1953                  * to handle the rare but possible case of recursive
1954                  * thread-death.) The rest of the cleanup is done in
1955                  * userspace.
1956                  */
1957                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1958                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1959
1960                 if (nval == -EFAULT)
1961                         return -1;
1962
1963                 if (nval != uval)
1964                         goto retry;
1965
1966                 /*
1967                  * Wake robust non-PI futexes here. The wakeup of
1968                  * PI futexes happens in exit_pi_state():
1969                  */
1970                 if (!pi && (uval & FUTEX_WAITERS))
1971                         futex_wake(uaddr, &curr->mm->mmap_sem, 1,
1972                                    FUTEX_BITSET_MATCH_ANY);
1973         }
1974         return 0;
1975 }
1976
1977 /*
1978  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1979  */
1980 static inline int fetch_robust_entry(struct robust_list __user **entry,
1981                                      struct robust_list __user * __user *head,
1982                                      int *pi)
1983 {
1984         unsigned long uentry;
1985
1986         if (get_user(uentry, (unsigned long __user *)head))
1987                 return -EFAULT;
1988
1989         *entry = (void __user *)(uentry & ~1UL);
1990         *pi = uentry & 1;
1991
1992         return 0;
1993 }
1994
1995 /*
1996  * Walk curr->robust_list (very carefully, it's a userspace list!)
1997  * and mark any locks found there dead, and notify any waiters.
1998  *
1999  * We silently return on any sign of list-walking problem.
2000  */
2001 void exit_robust_list(struct task_struct *curr)
2002 {
2003         struct robust_list_head __user *head = curr->robust_list;
2004         struct robust_list __user *entry, *next_entry, *pending;
2005         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2006         unsigned long futex_offset;
2007         int rc;
2008
2009         if (!futex_cmpxchg_enabled)
2010                 return;
2011
2012         /*
2013          * Fetch the list head (which was registered earlier, via
2014          * sys_set_robust_list()):
2015          */
2016         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2017                 return;
2018         /*
2019          * Fetch the relative futex offset:
2020          */
2021         if (get_user(futex_offset, &head->futex_offset))
2022                 return;
2023         /*
2024          * Fetch any possibly pending lock-add first, and handle it
2025          * if it exists:
2026          */
2027         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2028                 return;
2029
2030         next_entry = NULL;      /* avoid warning with gcc */
2031         while (entry != &head->list) {
2032                 /*
2033                  * Fetch the next entry in the list before calling
2034                  * handle_futex_death:
2035                  */
2036                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2037                 /*
2038                  * A pending lock might already be on the list, so
2039                  * don't process it twice:
2040                  */
2041                 if (entry != pending)
2042                         if (handle_futex_death((void __user *)entry + futex_offset,
2043                                                 curr, pi))
2044                                 return;
2045                 if (rc)
2046                         return;
2047                 entry = next_entry;
2048                 pi = next_pi;
2049                 /*
2050                  * Avoid excessively long or circular lists:
2051                  */
2052                 if (!--limit)
2053                         break;
2054
2055                 cond_resched();
2056         }
2057
2058         if (pending)
2059                 handle_futex_death((void __user *)pending + futex_offset,
2060                                    curr, pip);
2061 }
2062
2063 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2064                 u32 __user *uaddr2, u32 val2, u32 val3)
2065 {
2066         int ret = -ENOSYS;
2067         int cmd = op & FUTEX_CMD_MASK;
2068         struct rw_semaphore *fshared = NULL;
2069
2070         if (!(op & FUTEX_PRIVATE_FLAG))
2071                 fshared = &current->mm->mmap_sem;
2072
2073         switch (cmd) {
2074         case FUTEX_WAIT:
2075                 val3 = FUTEX_BITSET_MATCH_ANY;
2076         case FUTEX_WAIT_BITSET:
2077                 ret = futex_wait(uaddr, fshared, val, timeout, val3);
2078                 break;
2079         case FUTEX_WAKE:
2080                 val3 = FUTEX_BITSET_MATCH_ANY;
2081         case FUTEX_WAKE_BITSET:
2082                 ret = futex_wake(uaddr, fshared, val, val3);
2083                 break;
2084         case FUTEX_FD:
2085                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
2086                 ret = futex_fd(uaddr, val);
2087                 break;
2088         case FUTEX_REQUEUE:
2089                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2090                 break;
2091         case FUTEX_CMP_REQUEUE:
2092                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2093                 break;
2094         case FUTEX_WAKE_OP:
2095                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2096                 break;
2097         case FUTEX_LOCK_PI:
2098                 if (futex_cmpxchg_enabled)
2099                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2100                 break;
2101         case FUTEX_UNLOCK_PI:
2102                 if (futex_cmpxchg_enabled)
2103                         ret = futex_unlock_pi(uaddr, fshared);
2104                 break;
2105         case FUTEX_TRYLOCK_PI:
2106                 if (futex_cmpxchg_enabled)
2107                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2108                 break;
2109         default:
2110                 ret = -ENOSYS;
2111         }
2112         return ret;
2113 }
2114
2115
2116 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2117                           struct timespec __user *utime, u32 __user *uaddr2,
2118                           u32 val3)
2119 {
2120         struct timespec ts;
2121         ktime_t t, *tp = NULL;
2122         u32 val2 = 0;
2123         int cmd = op & FUTEX_CMD_MASK;
2124
2125         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2126                       cmd == FUTEX_WAIT_BITSET)) {
2127                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2128                         return -EFAULT;
2129                 if (!timespec_valid(&ts))
2130                         return -EINVAL;
2131
2132                 t = timespec_to_ktime(ts);
2133                 if (cmd == FUTEX_WAIT)
2134                         t = ktime_add_safe(ktime_get(), t);
2135                 tp = &t;
2136         }
2137         /*
2138          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2139          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2140          */
2141         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2142             cmd == FUTEX_WAKE_OP)
2143                 val2 = (u32) (unsigned long) utime;
2144
2145         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2146 }
2147
2148 static int futexfs_get_sb(struct file_system_type *fs_type,
2149                           int flags, const char *dev_name, void *data,
2150                           struct vfsmount *mnt)
2151 {
2152         return get_sb_pseudo(fs_type, "futex", NULL, FUTEXFS_SUPER_MAGIC, mnt);
2153 }
2154
2155 static struct file_system_type futex_fs_type = {
2156         .name           = "futexfs",
2157         .get_sb         = futexfs_get_sb,
2158         .kill_sb        = kill_anon_super,
2159 };
2160
2161 static int __init futex_init(void)
2162 {
2163         u32 curval;
2164         int i;
2165
2166         /*
2167          * This will fail and we want it. Some arch implementations do
2168          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2169          * functionality. We want to know that before we call in any
2170          * of the complex code paths. Also we want to prevent
2171          * registration of robust lists in that case. NULL is
2172          * guaranteed to fault and we get -EFAULT on functional
2173          * implementation, the non functional ones will return
2174          * -ENOSYS.
2175          */
2176         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2177         if (curval == -EFAULT)
2178                 futex_cmpxchg_enabled = 1;
2179
2180         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2181                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2182                 spin_lock_init(&futex_queues[i].lock);
2183         }
2184
2185         i = register_filesystem(&futex_fs_type);
2186         if (i)
2187                 return i;
2188
2189         futex_mnt = kern_mount(&futex_fs_type);
2190         if (IS_ERR(futex_mnt)) {
2191                 unregister_filesystem(&futex_fs_type);
2192                 return PTR_ERR(futex_mnt);
2193         }
2194
2195         return 0;
2196 }
2197 __initcall(futex_init);