hfs: add sanity check for file name length
[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  *  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.
25  *
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.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
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.
37  *
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.
42  *
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
46  */
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>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Futex flags used to encode options to functions and preserve them across
73  * restarts.
74  */
75 #define FLAGS_SHARED            0x01
76 #define FLAGS_CLOCKRT           0x02
77 #define FLAGS_HAS_TIMEOUT       0x04
78
79 /*
80  * Priority Inheritance state:
81  */
82 struct futex_pi_state {
83         /*
84          * list of 'owned' pi_state instances - these have to be
85          * cleaned up in do_exit() if the task exits prematurely:
86          */
87         struct list_head list;
88
89         /*
90          * The PI object:
91          */
92         struct rt_mutex pi_mutex;
93
94         struct task_struct *owner;
95         atomic_t refcount;
96
97         union futex_key key;
98 };
99
100 /**
101  * struct futex_q - The hashed futex queue entry, one per waiting task
102  * @list:               priority-sorted list of tasks waiting on this futex
103  * @task:               the task waiting on the futex
104  * @lock_ptr:           the hash bucket lock
105  * @key:                the key the futex is hashed on
106  * @pi_state:           optional priority inheritance state
107  * @rt_waiter:          rt_waiter storage for use with requeue_pi
108  * @requeue_pi_key:     the requeue_pi target futex key
109  * @bitset:             bitset for the optional bitmasked wakeup
110  *
111  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
112  * we can wake only the relevant ones (hashed queues may be shared).
113  *
114  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
115  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
116  * The order of wakeup is always to make the first condition true, then
117  * the second.
118  *
119  * PI futexes are typically woken before they are removed from the hash list via
120  * the rt_mutex code. See unqueue_me_pi().
121  */
122 struct futex_q {
123         struct plist_node list;
124
125         struct task_struct *task;
126         spinlock_t *lock_ptr;
127         union futex_key key;
128         struct futex_pi_state *pi_state;
129         struct rt_mutex_waiter *rt_waiter;
130         union futex_key *requeue_pi_key;
131         u32 bitset;
132 };
133
134 static const struct futex_q futex_q_init = {
135         /* list gets initialized in queue_me()*/
136         .key = FUTEX_KEY_INIT,
137         .bitset = FUTEX_BITSET_MATCH_ANY
138 };
139
140 /*
141  * Hash buckets are shared by all the futex_keys that hash to the same
142  * location.  Each key may have multiple futex_q structures, one for each task
143  * waiting on a futex.
144  */
145 struct futex_hash_bucket {
146         spinlock_t lock;
147         struct plist_head chain;
148 };
149
150 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
151
152 /*
153  * We hash on the keys returned from get_futex_key (see below).
154  */
155 static struct futex_hash_bucket *hash_futex(union futex_key *key)
156 {
157         u32 hash = jhash2((u32*)&key->both.word,
158                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
159                           key->both.offset);
160         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
161 }
162
163 /*
164  * Return 1 if two futex_keys are equal, 0 otherwise.
165  */
166 static inline int match_futex(union futex_key *key1, union futex_key *key2)
167 {
168         return (key1 && key2
169                 && key1->both.word == key2->both.word
170                 && key1->both.ptr == key2->both.ptr
171                 && key1->both.offset == key2->both.offset);
172 }
173
174 /*
175  * Take a reference to the resource addressed by a key.
176  * Can be called while holding spinlocks.
177  *
178  */
179 static void get_futex_key_refs(union futex_key *key)
180 {
181         if (!key->both.ptr)
182                 return;
183
184         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
185         case FUT_OFF_INODE:
186                 ihold(key->shared.inode);
187                 break;
188         case FUT_OFF_MMSHARED:
189                 atomic_inc(&key->private.mm->mm_count);
190                 break;
191         }
192 }
193
194 /*
195  * Drop a reference to the resource addressed by a key.
196  * The hash bucket spinlock must not be held.
197  */
198 static void drop_futex_key_refs(union futex_key *key)
199 {
200         if (!key->both.ptr) {
201                 /* If we're here then we tried to put a key we failed to get */
202                 WARN_ON_ONCE(1);
203                 return;
204         }
205
206         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
207         case FUT_OFF_INODE:
208                 iput(key->shared.inode);
209                 break;
210         case FUT_OFF_MMSHARED:
211                 mmdrop(key->private.mm);
212                 break;
213         }
214 }
215
216 /**
217  * get_futex_key() - Get parameters which are the keys for a futex
218  * @uaddr:      virtual address of the futex
219  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
220  * @key:        address where result is stored.
221  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
222  *              VERIFY_WRITE)
223  *
224  * Returns a negative error code or 0
225  * The key words are stored in *key on success.
226  *
227  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
228  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
229  * We can usually work out the index without swapping in the page.
230  *
231  * lock_page() might sleep, the caller should not hold a spinlock.
232  */
233 static int
234 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
235 {
236         unsigned long address = (unsigned long)uaddr;
237         struct mm_struct *mm = current->mm;
238         struct page *page, *page_head;
239         int err, ro = 0;
240
241         /*
242          * The futex address must be "naturally" aligned.
243          */
244         key->both.offset = address % PAGE_SIZE;
245         if (unlikely((address % sizeof(u32)) != 0))
246                 return -EINVAL;
247         address -= key->both.offset;
248
249         /*
250          * PROCESS_PRIVATE futexes are fast.
251          * As the mm cannot disappear under us and the 'key' only needs
252          * virtual address, we dont even have to find the underlying vma.
253          * Note : We do have to check 'uaddr' is a valid user address,
254          *        but access_ok() should be faster than find_vma()
255          */
256         if (!fshared) {
257                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
258                         return -EFAULT;
259                 key->private.mm = mm;
260                 key->private.address = address;
261                 get_futex_key_refs(key);
262                 return 0;
263         }
264
265 again:
266         err = get_user_pages_fast(address, 1, 1, &page);
267         /*
268          * If write access is not required (eg. FUTEX_WAIT), try
269          * and get read-only access.
270          */
271         if (err == -EFAULT && rw == VERIFY_READ) {
272                 err = get_user_pages_fast(address, 1, 0, &page);
273                 ro = 1;
274         }
275         if (err < 0)
276                 return err;
277         else
278                 err = 0;
279
280 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
281         page_head = page;
282         if (unlikely(PageTail(page))) {
283                 put_page(page);
284                 /* serialize against __split_huge_page_splitting() */
285                 local_irq_disable();
286                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
287                         page_head = compound_head(page);
288                         /*
289                          * page_head is valid pointer but we must pin
290                          * it before taking the PG_lock and/or
291                          * PG_compound_lock. The moment we re-enable
292                          * irqs __split_huge_page_splitting() can
293                          * return and the head page can be freed from
294                          * under us. We can't take the PG_lock and/or
295                          * PG_compound_lock on a page that could be
296                          * freed from under us.
297                          */
298                         if (page != page_head) {
299                                 get_page(page_head);
300                                 put_page(page);
301                         }
302                         local_irq_enable();
303                 } else {
304                         local_irq_enable();
305                         goto again;
306                 }
307         }
308 #else
309         page_head = compound_head(page);
310         if (page != page_head) {
311                 get_page(page_head);
312                 put_page(page);
313         }
314 #endif
315
316         lock_page(page_head);
317         if (!page_head->mapping) {
318                 unlock_page(page_head);
319                 put_page(page_head);
320                 /*
321                 * ZERO_PAGE pages don't have a mapping. Avoid a busy loop
322                 * trying to find one. RW mapping would have COW'd (and thus
323                 * have a mapping) so this page is RO and won't ever change.
324                 */
325                 if ((page_head == ZERO_PAGE(address)))
326                         return -EFAULT;
327                 goto again;
328         }
329
330         /*
331          * Private mappings are handled in a simple way.
332          *
333          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
334          * it's a read-only handle, it's expected that futexes attach to
335          * the object not the particular process.
336          */
337         if (PageAnon(page_head)) {
338                 /*
339                  * A RO anonymous page will never change and thus doesn't make
340                  * sense for futex operations.
341                  */
342                 if (ro) {
343                         err = -EFAULT;
344                         goto out;
345                 }
346
347                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
348                 key->private.mm = mm;
349                 key->private.address = address;
350         } else {
351                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
352                 key->shared.inode = page_head->mapping->host;
353                 key->shared.pgoff = page_head->index;
354         }
355
356         get_futex_key_refs(key);
357
358 out:
359         unlock_page(page_head);
360         put_page(page_head);
361         return err;
362 }
363
364 static inline void put_futex_key(union futex_key *key)
365 {
366         drop_futex_key_refs(key);
367 }
368
369 /**
370  * fault_in_user_writeable() - Fault in user address and verify RW access
371  * @uaddr:      pointer to faulting user space address
372  *
373  * Slow path to fixup the fault we just took in the atomic write
374  * access to @uaddr.
375  *
376  * We have no generic implementation of a non-destructive write to the
377  * user address. We know that we faulted in the atomic pagefault
378  * disabled section so we can as well avoid the #PF overhead by
379  * calling get_user_pages() right away.
380  */
381 static int fault_in_user_writeable(u32 __user *uaddr)
382 {
383         struct mm_struct *mm = current->mm;
384         int ret;
385
386         down_read(&mm->mmap_sem);
387         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
388                                FAULT_FLAG_WRITE);
389         up_read(&mm->mmap_sem);
390
391         return ret < 0 ? ret : 0;
392 }
393
394 /**
395  * futex_top_waiter() - Return the highest priority waiter on a futex
396  * @hb:         the hash bucket the futex_q's reside in
397  * @key:        the futex key (to distinguish it from other futex futex_q's)
398  *
399  * Must be called with the hb lock held.
400  */
401 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
402                                         union futex_key *key)
403 {
404         struct futex_q *this;
405
406         plist_for_each_entry(this, &hb->chain, list) {
407                 if (match_futex(&this->key, key))
408                         return this;
409         }
410         return NULL;
411 }
412
413 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
414                                       u32 uval, u32 newval)
415 {
416         int ret;
417
418         pagefault_disable();
419         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
420         pagefault_enable();
421
422         return ret;
423 }
424
425 static int get_futex_value_locked(u32 *dest, u32 __user *from)
426 {
427         int ret;
428
429         pagefault_disable();
430         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
431         pagefault_enable();
432
433         return ret ? -EFAULT : 0;
434 }
435
436
437 /*
438  * PI code:
439  */
440 static int refill_pi_state_cache(void)
441 {
442         struct futex_pi_state *pi_state;
443
444         if (likely(current->pi_state_cache))
445                 return 0;
446
447         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
448
449         if (!pi_state)
450                 return -ENOMEM;
451
452         INIT_LIST_HEAD(&pi_state->list);
453         /* pi_mutex gets initialized later */
454         pi_state->owner = NULL;
455         atomic_set(&pi_state->refcount, 1);
456         pi_state->key = FUTEX_KEY_INIT;
457
458         current->pi_state_cache = pi_state;
459
460         return 0;
461 }
462
463 static struct futex_pi_state * alloc_pi_state(void)
464 {
465         struct futex_pi_state *pi_state = current->pi_state_cache;
466
467         WARN_ON(!pi_state);
468         current->pi_state_cache = NULL;
469
470         return pi_state;
471 }
472
473 static void free_pi_state(struct futex_pi_state *pi_state)
474 {
475         if (!atomic_dec_and_test(&pi_state->refcount))
476                 return;
477
478         /*
479          * If pi_state->owner is NULL, the owner is most probably dying
480          * and has cleaned up the pi_state already
481          */
482         if (pi_state->owner) {
483                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
484                 list_del_init(&pi_state->list);
485                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
486
487                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
488         }
489
490         if (current->pi_state_cache)
491                 kfree(pi_state);
492         else {
493                 /*
494                  * pi_state->list is already empty.
495                  * clear pi_state->owner.
496                  * refcount is at 0 - put it back to 1.
497                  */
498                 pi_state->owner = NULL;
499                 atomic_set(&pi_state->refcount, 1);
500                 current->pi_state_cache = pi_state;
501         }
502 }
503
504 /*
505  * Look up the task based on what TID userspace gave us.
506  * We dont trust it.
507  */
508 static struct task_struct * futex_find_get_task(pid_t pid)
509 {
510         struct task_struct *p;
511
512         rcu_read_lock();
513         p = find_task_by_vpid(pid);
514         if (p)
515                 get_task_struct(p);
516
517         rcu_read_unlock();
518
519         return p;
520 }
521
522 /*
523  * This task is holding PI mutexes at exit time => bad.
524  * Kernel cleans up PI-state, but userspace is likely hosed.
525  * (Robust-futex cleanup is separate and might save the day for userspace.)
526  */
527 void exit_pi_state_list(struct task_struct *curr)
528 {
529         struct list_head *next, *head = &curr->pi_state_list;
530         struct futex_pi_state *pi_state;
531         struct futex_hash_bucket *hb;
532         union futex_key key = FUTEX_KEY_INIT;
533
534         if (!futex_cmpxchg_enabled)
535                 return;
536         /*
537          * We are a ZOMBIE and nobody can enqueue itself on
538          * pi_state_list anymore, but we have to be careful
539          * versus waiters unqueueing themselves:
540          */
541         raw_spin_lock_irq(&curr->pi_lock);
542         while (!list_empty(head)) {
543
544                 next = head->next;
545                 pi_state = list_entry(next, struct futex_pi_state, list);
546                 key = pi_state->key;
547                 hb = hash_futex(&key);
548                 raw_spin_unlock_irq(&curr->pi_lock);
549
550                 spin_lock(&hb->lock);
551
552                 raw_spin_lock_irq(&curr->pi_lock);
553                 /*
554                  * We dropped the pi-lock, so re-check whether this
555                  * task still owns the PI-state:
556                  */
557                 if (head->next != next) {
558                         spin_unlock(&hb->lock);
559                         continue;
560                 }
561
562                 WARN_ON(pi_state->owner != curr);
563                 WARN_ON(list_empty(&pi_state->list));
564                 list_del_init(&pi_state->list);
565                 pi_state->owner = NULL;
566                 raw_spin_unlock_irq(&curr->pi_lock);
567
568                 rt_mutex_unlock(&pi_state->pi_mutex);
569
570                 spin_unlock(&hb->lock);
571
572                 raw_spin_lock_irq(&curr->pi_lock);
573         }
574         raw_spin_unlock_irq(&curr->pi_lock);
575 }
576
577 static int
578 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
579                 union futex_key *key, struct futex_pi_state **ps)
580 {
581         struct futex_pi_state *pi_state = NULL;
582         struct futex_q *this, *next;
583         struct plist_head *head;
584         struct task_struct *p;
585         pid_t pid = uval & FUTEX_TID_MASK;
586
587         head = &hb->chain;
588
589         plist_for_each_entry_safe(this, next, head, list) {
590                 if (match_futex(&this->key, key)) {
591                         /*
592                          * Another waiter already exists - bump up
593                          * the refcount and return its pi_state:
594                          */
595                         pi_state = this->pi_state;
596                         /*
597                          * Userspace might have messed up non-PI and PI futexes
598                          */
599                         if (unlikely(!pi_state))
600                                 return -EINVAL;
601
602                         WARN_ON(!atomic_read(&pi_state->refcount));
603
604                         /*
605                          * When pi_state->owner is NULL then the owner died
606                          * and another waiter is on the fly. pi_state->owner
607                          * is fixed up by the task which acquires
608                          * pi_state->rt_mutex.
609                          *
610                          * We do not check for pid == 0 which can happen when
611                          * the owner died and robust_list_exit() cleared the
612                          * TID.
613                          */
614                         if (pid && pi_state->owner) {
615                                 /*
616                                  * Bail out if user space manipulated the
617                                  * futex value.
618                                  */
619                                 if (pid != task_pid_vnr(pi_state->owner))
620                                         return -EINVAL;
621                         }
622
623                         atomic_inc(&pi_state->refcount);
624                         *ps = pi_state;
625
626                         return 0;
627                 }
628         }
629
630         /*
631          * We are the first waiter - try to look up the real owner and attach
632          * the new pi_state to it, but bail out when TID = 0
633          */
634         if (!pid)
635                 return -ESRCH;
636         p = futex_find_get_task(pid);
637         if (!p)
638                 return -ESRCH;
639
640         /*
641          * We need to look at the task state flags to figure out,
642          * whether the task is exiting. To protect against the do_exit
643          * change of the task flags, we do this protected by
644          * p->pi_lock:
645          */
646         raw_spin_lock_irq(&p->pi_lock);
647         if (unlikely(p->flags & PF_EXITING)) {
648                 /*
649                  * The task is on the way out. When PF_EXITPIDONE is
650                  * set, we know that the task has finished the
651                  * cleanup:
652                  */
653                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
654
655                 raw_spin_unlock_irq(&p->pi_lock);
656                 put_task_struct(p);
657                 return ret;
658         }
659
660         pi_state = alloc_pi_state();
661
662         /*
663          * Initialize the pi_mutex in locked state and make 'p'
664          * the owner of it:
665          */
666         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
667
668         /* Store the key for possible exit cleanups: */
669         pi_state->key = *key;
670
671         WARN_ON(!list_empty(&pi_state->list));
672         list_add(&pi_state->list, &p->pi_state_list);
673         pi_state->owner = p;
674         raw_spin_unlock_irq(&p->pi_lock);
675
676         put_task_struct(p);
677
678         *ps = pi_state;
679
680         return 0;
681 }
682
683 /**
684  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
685  * @uaddr:              the pi futex user address
686  * @hb:                 the pi futex hash bucket
687  * @key:                the futex key associated with uaddr and hb
688  * @ps:                 the pi_state pointer where we store the result of the
689  *                      lookup
690  * @task:               the task to perform the atomic lock work for.  This will
691  *                      be "current" except in the case of requeue pi.
692  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
693  *
694  * Returns:
695  *  0 - ready to wait
696  *  1 - acquired the lock
697  * <0 - error
698  *
699  * The hb->lock and futex_key refs shall be held by the caller.
700  */
701 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
702                                 union futex_key *key,
703                                 struct futex_pi_state **ps,
704                                 struct task_struct *task, int set_waiters)
705 {
706         int lock_taken, ret, ownerdied = 0;
707         u32 uval, newval, curval, vpid = task_pid_vnr(task);
708
709 retry:
710         ret = lock_taken = 0;
711
712         /*
713          * To avoid races, we attempt to take the lock here again
714          * (by doing a 0 -> TID atomic cmpxchg), while holding all
715          * the locks. It will most likely not succeed.
716          */
717         newval = vpid;
718         if (set_waiters)
719                 newval |= FUTEX_WAITERS;
720
721         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
722                 return -EFAULT;
723
724         /*
725          * Detect deadlocks.
726          */
727         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
728                 return -EDEADLK;
729
730         /*
731          * Surprise - we got the lock. Just return to userspace:
732          */
733         if (unlikely(!curval))
734                 return 1;
735
736         uval = curval;
737
738         /*
739          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
740          * to wake at the next unlock.
741          */
742         newval = curval | FUTEX_WAITERS;
743
744         /*
745          * There are two cases, where a futex might have no owner (the
746          * owner TID is 0): OWNER_DIED. We take over the futex in this
747          * case. We also do an unconditional take over, when the owner
748          * of the futex died.
749          *
750          * This is safe as we are protected by the hash bucket lock !
751          */
752         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
753                 /* Keep the OWNER_DIED bit */
754                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
755                 ownerdied = 0;
756                 lock_taken = 1;
757         }
758
759         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
760                 return -EFAULT;
761         if (unlikely(curval != uval))
762                 goto retry;
763
764         /*
765          * We took the lock due to owner died take over.
766          */
767         if (unlikely(lock_taken))
768                 return 1;
769
770         /*
771          * We dont have the lock. Look up the PI state (or create it if
772          * we are the first waiter):
773          */
774         ret = lookup_pi_state(uval, hb, key, ps);
775
776         if (unlikely(ret)) {
777                 switch (ret) {
778                 case -ESRCH:
779                         /*
780                          * No owner found for this futex. Check if the
781                          * OWNER_DIED bit is set to figure out whether
782                          * this is a robust futex or not.
783                          */
784                         if (get_futex_value_locked(&curval, uaddr))
785                                 return -EFAULT;
786
787                         /*
788                          * We simply start over in case of a robust
789                          * futex. The code above will take the futex
790                          * and return happy.
791                          */
792                         if (curval & FUTEX_OWNER_DIED) {
793                                 ownerdied = 1;
794                                 goto retry;
795                         }
796                 default:
797                         break;
798                 }
799         }
800
801         return ret;
802 }
803
804 /**
805  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
806  * @q:  The futex_q to unqueue
807  *
808  * The q->lock_ptr must not be NULL and must be held by the caller.
809  */
810 static void __unqueue_futex(struct futex_q *q)
811 {
812         struct futex_hash_bucket *hb;
813
814         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
815             || WARN_ON(plist_node_empty(&q->list)))
816                 return;
817
818         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
819         plist_del(&q->list, &hb->chain);
820 }
821
822 /*
823  * The hash bucket lock must be held when this is called.
824  * Afterwards, the futex_q must not be accessed.
825  */
826 static void wake_futex(struct futex_q *q)
827 {
828         struct task_struct *p = q->task;
829
830         /*
831          * We set q->lock_ptr = NULL _before_ we wake up the task. If
832          * a non-futex wake up happens on another CPU then the task
833          * might exit and p would dereference a non-existing task
834          * struct. Prevent this by holding a reference on p across the
835          * wake up.
836          */
837         get_task_struct(p);
838
839         __unqueue_futex(q);
840         /*
841          * The waiting task can free the futex_q as soon as
842          * q->lock_ptr = NULL is written, without taking any locks. A
843          * memory barrier is required here to prevent the following
844          * store to lock_ptr from getting ahead of the plist_del.
845          */
846         smp_wmb();
847         q->lock_ptr = NULL;
848
849         wake_up_state(p, TASK_NORMAL);
850         put_task_struct(p);
851 }
852
853 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
854 {
855         struct task_struct *new_owner;
856         struct futex_pi_state *pi_state = this->pi_state;
857         u32 curval, newval;
858
859         if (!pi_state)
860                 return -EINVAL;
861
862         /*
863          * If current does not own the pi_state then the futex is
864          * inconsistent and user space fiddled with the futex value.
865          */
866         if (pi_state->owner != current)
867                 return -EINVAL;
868
869         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
870         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
871
872         /*
873          * It is possible that the next waiter (the one that brought
874          * this owner to the kernel) timed out and is no longer
875          * waiting on the lock.
876          */
877         if (!new_owner)
878                 new_owner = this->task;
879
880         /*
881          * We pass it to the next owner. (The WAITERS bit is always
882          * kept enabled while there is PI state around. We must also
883          * preserve the owner died bit.)
884          */
885         if (!(uval & FUTEX_OWNER_DIED)) {
886                 int ret = 0;
887
888                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
889
890                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
891                         ret = -EFAULT;
892                 else if (curval != uval)
893                         ret = -EINVAL;
894                 if (ret) {
895                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
896                         return ret;
897                 }
898         }
899
900         raw_spin_lock_irq(&pi_state->owner->pi_lock);
901         WARN_ON(list_empty(&pi_state->list));
902         list_del_init(&pi_state->list);
903         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
904
905         raw_spin_lock_irq(&new_owner->pi_lock);
906         WARN_ON(!list_empty(&pi_state->list));
907         list_add(&pi_state->list, &new_owner->pi_state_list);
908         pi_state->owner = new_owner;
909         raw_spin_unlock_irq(&new_owner->pi_lock);
910
911         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
912         rt_mutex_unlock(&pi_state->pi_mutex);
913
914         return 0;
915 }
916
917 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
918 {
919         u32 oldval;
920
921         /*
922          * There is no waiter, so we unlock the futex. The owner died
923          * bit has not to be preserved here. We are the owner:
924          */
925         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
926                 return -EFAULT;
927         if (oldval != uval)
928                 return -EAGAIN;
929
930         return 0;
931 }
932
933 /*
934  * Express the locking dependencies for lockdep:
935  */
936 static inline void
937 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
938 {
939         if (hb1 <= hb2) {
940                 spin_lock(&hb1->lock);
941                 if (hb1 < hb2)
942                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
943         } else { /* hb1 > hb2 */
944                 spin_lock(&hb2->lock);
945                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
946         }
947 }
948
949 static inline void
950 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
951 {
952         spin_unlock(&hb1->lock);
953         if (hb1 != hb2)
954                 spin_unlock(&hb2->lock);
955 }
956
957 /*
958  * Wake up waiters matching bitset queued on this futex (uaddr).
959  */
960 static int
961 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
962 {
963         struct futex_hash_bucket *hb;
964         struct futex_q *this, *next;
965         struct plist_head *head;
966         union futex_key key = FUTEX_KEY_INIT;
967         int ret;
968
969         if (!bitset)
970                 return -EINVAL;
971
972         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
973         if (unlikely(ret != 0))
974                 goto out;
975
976         hb = hash_futex(&key);
977         spin_lock(&hb->lock);
978         head = &hb->chain;
979
980         plist_for_each_entry_safe(this, next, head, list) {
981                 if (match_futex (&this->key, &key)) {
982                         if (this->pi_state || this->rt_waiter) {
983                                 ret = -EINVAL;
984                                 break;
985                         }
986
987                         /* Check if one of the bits is set in both bitsets */
988                         if (!(this->bitset & bitset))
989                                 continue;
990
991                         wake_futex(this);
992                         if (++ret >= nr_wake)
993                                 break;
994                 }
995         }
996
997         spin_unlock(&hb->lock);
998         put_futex_key(&key);
999 out:
1000         return ret;
1001 }
1002
1003 /*
1004  * Wake up all waiters hashed on the physical page that is mapped
1005  * to this virtual address:
1006  */
1007 static int
1008 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1009               int nr_wake, int nr_wake2, int op)
1010 {
1011         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1012         struct futex_hash_bucket *hb1, *hb2;
1013         struct plist_head *head;
1014         struct futex_q *this, *next;
1015         int ret, op_ret;
1016
1017 retry:
1018         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1019         if (unlikely(ret != 0))
1020                 goto out;
1021         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1022         if (unlikely(ret != 0))
1023                 goto out_put_key1;
1024
1025         hb1 = hash_futex(&key1);
1026         hb2 = hash_futex(&key2);
1027
1028 retry_private:
1029         double_lock_hb(hb1, hb2);
1030         op_ret = futex_atomic_op_inuser(op, uaddr2);
1031         if (unlikely(op_ret < 0)) {
1032
1033                 double_unlock_hb(hb1, hb2);
1034
1035 #ifndef CONFIG_MMU
1036                 /*
1037                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1038                  * but we might get them from range checking
1039                  */
1040                 ret = op_ret;
1041                 goto out_put_keys;
1042 #endif
1043
1044                 if (unlikely(op_ret != -EFAULT)) {
1045                         ret = op_ret;
1046                         goto out_put_keys;
1047                 }
1048
1049                 ret = fault_in_user_writeable(uaddr2);
1050                 if (ret)
1051                         goto out_put_keys;
1052
1053                 if (!(flags & FLAGS_SHARED))
1054                         goto retry_private;
1055
1056                 put_futex_key(&key2);
1057                 put_futex_key(&key1);
1058                 goto retry;
1059         }
1060
1061         head = &hb1->chain;
1062
1063         plist_for_each_entry_safe(this, next, head, list) {
1064                 if (match_futex (&this->key, &key1)) {
1065                         wake_futex(this);
1066                         if (++ret >= nr_wake)
1067                                 break;
1068                 }
1069         }
1070
1071         if (op_ret > 0) {
1072                 head = &hb2->chain;
1073
1074                 op_ret = 0;
1075                 plist_for_each_entry_safe(this, next, head, list) {
1076                         if (match_futex (&this->key, &key2)) {
1077                                 wake_futex(this);
1078                                 if (++op_ret >= nr_wake2)
1079                                         break;
1080                         }
1081                 }
1082                 ret += op_ret;
1083         }
1084
1085         double_unlock_hb(hb1, hb2);
1086 out_put_keys:
1087         put_futex_key(&key2);
1088 out_put_key1:
1089         put_futex_key(&key1);
1090 out:
1091         return ret;
1092 }
1093
1094 /**
1095  * requeue_futex() - Requeue a futex_q from one hb to another
1096  * @q:          the futex_q to requeue
1097  * @hb1:        the source hash_bucket
1098  * @hb2:        the target hash_bucket
1099  * @key2:       the new key for the requeued futex_q
1100  */
1101 static inline
1102 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1103                    struct futex_hash_bucket *hb2, union futex_key *key2)
1104 {
1105
1106         /*
1107          * If key1 and key2 hash to the same bucket, no need to
1108          * requeue.
1109          */
1110         if (likely(&hb1->chain != &hb2->chain)) {
1111                 plist_del(&q->list, &hb1->chain);
1112                 plist_add(&q->list, &hb2->chain);
1113                 q->lock_ptr = &hb2->lock;
1114         }
1115         get_futex_key_refs(key2);
1116         q->key = *key2;
1117 }
1118
1119 /**
1120  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1121  * @q:          the futex_q
1122  * @key:        the key of the requeue target futex
1123  * @hb:         the hash_bucket of the requeue target futex
1124  *
1125  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1126  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1127  * to the requeue target futex so the waiter can detect the wakeup on the right
1128  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1129  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1130  * to protect access to the pi_state to fixup the owner later.  Must be called
1131  * with both q->lock_ptr and hb->lock held.
1132  */
1133 static inline
1134 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1135                            struct futex_hash_bucket *hb)
1136 {
1137         get_futex_key_refs(key);
1138         q->key = *key;
1139
1140         __unqueue_futex(q);
1141
1142         WARN_ON(!q->rt_waiter);
1143         q->rt_waiter = NULL;
1144
1145         q->lock_ptr = &hb->lock;
1146
1147         wake_up_state(q->task, TASK_NORMAL);
1148 }
1149
1150 /**
1151  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1152  * @pifutex:            the user address of the to futex
1153  * @hb1:                the from futex hash bucket, must be locked by the caller
1154  * @hb2:                the to futex hash bucket, must be locked by the caller
1155  * @key1:               the from futex key
1156  * @key2:               the to futex key
1157  * @ps:                 address to store the pi_state pointer
1158  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1159  *
1160  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1161  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1162  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1163  * hb1 and hb2 must be held by the caller.
1164  *
1165  * Returns:
1166  *  0 - failed to acquire the lock atomicly
1167  *  1 - acquired the lock
1168  * <0 - error
1169  */
1170 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1171                                  struct futex_hash_bucket *hb1,
1172                                  struct futex_hash_bucket *hb2,
1173                                  union futex_key *key1, union futex_key *key2,
1174                                  struct futex_pi_state **ps, int set_waiters)
1175 {
1176         struct futex_q *top_waiter = NULL;
1177         u32 curval;
1178         int ret;
1179
1180         if (get_futex_value_locked(&curval, pifutex))
1181                 return -EFAULT;
1182
1183         /*
1184          * Find the top_waiter and determine if there are additional waiters.
1185          * If the caller intends to requeue more than 1 waiter to pifutex,
1186          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1187          * as we have means to handle the possible fault.  If not, don't set
1188          * the bit unecessarily as it will force the subsequent unlock to enter
1189          * the kernel.
1190          */
1191         top_waiter = futex_top_waiter(hb1, key1);
1192
1193         /* There are no waiters, nothing for us to do. */
1194         if (!top_waiter)
1195                 return 0;
1196
1197         /* Ensure we requeue to the expected futex. */
1198         if (!match_futex(top_waiter->requeue_pi_key, key2))
1199                 return -EINVAL;
1200
1201         /*
1202          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1203          * the contended case or if set_waiters is 1.  The pi_state is returned
1204          * in ps in contended cases.
1205          */
1206         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1207                                    set_waiters);
1208         if (ret == 1)
1209                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1210
1211         return ret;
1212 }
1213
1214 /**
1215  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1216  * @uaddr1:     source futex user address
1217  * @flags:      futex flags (FLAGS_SHARED, etc.)
1218  * @uaddr2:     target futex user address
1219  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1220  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1221  * @cmpval:     @uaddr1 expected value (or %NULL)
1222  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1223  *              pi futex (pi to pi requeue is not supported)
1224  *
1225  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1226  * uaddr2 atomically on behalf of the top waiter.
1227  *
1228  * Returns:
1229  * >=0 - on success, the number of tasks requeued or woken
1230  *  <0 - on error
1231  */
1232 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1233                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1234                          u32 *cmpval, int requeue_pi)
1235 {
1236         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1237         int drop_count = 0, task_count = 0, ret;
1238         struct futex_pi_state *pi_state = NULL;
1239         struct futex_hash_bucket *hb1, *hb2;
1240         struct plist_head *head1;
1241         struct futex_q *this, *next;
1242         u32 curval2;
1243
1244         if (requeue_pi) {
1245                 /*
1246                  * requeue_pi requires a pi_state, try to allocate it now
1247                  * without any locks in case it fails.
1248                  */
1249                 if (refill_pi_state_cache())
1250                         return -ENOMEM;
1251                 /*
1252                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1253                  * + nr_requeue, since it acquires the rt_mutex prior to
1254                  * returning to userspace, so as to not leave the rt_mutex with
1255                  * waiters and no owner.  However, second and third wake-ups
1256                  * cannot be predicted as they involve race conditions with the
1257                  * first wake and a fault while looking up the pi_state.  Both
1258                  * pthread_cond_signal() and pthread_cond_broadcast() should
1259                  * use nr_wake=1.
1260                  */
1261                 if (nr_wake != 1)
1262                         return -EINVAL;
1263         }
1264
1265 retry:
1266         if (pi_state != NULL) {
1267                 /*
1268                  * We will have to lookup the pi_state again, so free this one
1269                  * to keep the accounting correct.
1270                  */
1271                 free_pi_state(pi_state);
1272                 pi_state = NULL;
1273         }
1274
1275         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1276         if (unlikely(ret != 0))
1277                 goto out;
1278         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1279                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1280         if (unlikely(ret != 0))
1281                 goto out_put_key1;
1282
1283         hb1 = hash_futex(&key1);
1284         hb2 = hash_futex(&key2);
1285
1286 retry_private:
1287         double_lock_hb(hb1, hb2);
1288
1289         if (likely(cmpval != NULL)) {
1290                 u32 curval;
1291
1292                 ret = get_futex_value_locked(&curval, uaddr1);
1293
1294                 if (unlikely(ret)) {
1295                         double_unlock_hb(hb1, hb2);
1296
1297                         ret = get_user(curval, uaddr1);
1298                         if (ret)
1299                                 goto out_put_keys;
1300
1301                         if (!(flags & FLAGS_SHARED))
1302                                 goto retry_private;
1303
1304                         put_futex_key(&key2);
1305                         put_futex_key(&key1);
1306                         goto retry;
1307                 }
1308                 if (curval != *cmpval) {
1309                         ret = -EAGAIN;
1310                         goto out_unlock;
1311                 }
1312         }
1313
1314         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1315                 /*
1316                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1317                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1318                  * bit.  We force this here where we are able to easily handle
1319                  * faults rather in the requeue loop below.
1320                  */
1321                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1322                                                  &key2, &pi_state, nr_requeue);
1323
1324                 /*
1325                  * At this point the top_waiter has either taken uaddr2 or is
1326                  * waiting on it.  If the former, then the pi_state will not
1327                  * exist yet, look it up one more time to ensure we have a
1328                  * reference to it.
1329                  */
1330                 if (ret == 1) {
1331                         WARN_ON(pi_state);
1332                         drop_count++;
1333                         task_count++;
1334                         ret = get_futex_value_locked(&curval2, uaddr2);
1335                         if (!ret)
1336                                 ret = lookup_pi_state(curval2, hb2, &key2,
1337                                                       &pi_state);
1338                 }
1339
1340                 switch (ret) {
1341                 case 0:
1342                         break;
1343                 case -EFAULT:
1344                         double_unlock_hb(hb1, hb2);
1345                         put_futex_key(&key2);
1346                         put_futex_key(&key1);
1347                         ret = fault_in_user_writeable(uaddr2);
1348                         if (!ret)
1349                                 goto retry;
1350                         goto out;
1351                 case -EAGAIN:
1352                         /* The owner was exiting, try again. */
1353                         double_unlock_hb(hb1, hb2);
1354                         put_futex_key(&key2);
1355                         put_futex_key(&key1);
1356                         cond_resched();
1357                         goto retry;
1358                 default:
1359                         goto out_unlock;
1360                 }
1361         }
1362
1363         head1 = &hb1->chain;
1364         plist_for_each_entry_safe(this, next, head1, list) {
1365                 if (task_count - nr_wake >= nr_requeue)
1366                         break;
1367
1368                 if (!match_futex(&this->key, &key1))
1369                         continue;
1370
1371                 /*
1372                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1373                  * be paired with each other and no other futex ops.
1374                  */
1375                 if ((requeue_pi && !this->rt_waiter) ||
1376                     (!requeue_pi && this->rt_waiter)) {
1377                         ret = -EINVAL;
1378                         break;
1379                 }
1380
1381                 /*
1382                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1383                  * lock, we already woke the top_waiter.  If not, it will be
1384                  * woken by futex_unlock_pi().
1385                  */
1386                 if (++task_count <= nr_wake && !requeue_pi) {
1387                         wake_futex(this);
1388                         continue;
1389                 }
1390
1391                 /* Ensure we requeue to the expected futex for requeue_pi. */
1392                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1393                         ret = -EINVAL;
1394                         break;
1395                 }
1396
1397                 /*
1398                  * Requeue nr_requeue waiters and possibly one more in the case
1399                  * of requeue_pi if we couldn't acquire the lock atomically.
1400                  */
1401                 if (requeue_pi) {
1402                         /* Prepare the waiter to take the rt_mutex. */
1403                         atomic_inc(&pi_state->refcount);
1404                         this->pi_state = pi_state;
1405                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1406                                                         this->rt_waiter,
1407                                                         this->task, 1);
1408                         if (ret == 1) {
1409                                 /* We got the lock. */
1410                                 requeue_pi_wake_futex(this, &key2, hb2);
1411                                 drop_count++;
1412                                 continue;
1413                         } else if (ret) {
1414                                 /* -EDEADLK */
1415                                 this->pi_state = NULL;
1416                                 free_pi_state(pi_state);
1417                                 goto out_unlock;
1418                         }
1419                 }
1420                 requeue_futex(this, hb1, hb2, &key2);
1421                 drop_count++;
1422         }
1423
1424 out_unlock:
1425         double_unlock_hb(hb1, hb2);
1426
1427         /*
1428          * drop_futex_key_refs() must be called outside the spinlocks. During
1429          * the requeue we moved futex_q's from the hash bucket at key1 to the
1430          * one at key2 and updated their key pointer.  We no longer need to
1431          * hold the references to key1.
1432          */
1433         while (--drop_count >= 0)
1434                 drop_futex_key_refs(&key1);
1435
1436 out_put_keys:
1437         put_futex_key(&key2);
1438 out_put_key1:
1439         put_futex_key(&key1);
1440 out:
1441         if (pi_state != NULL)
1442                 free_pi_state(pi_state);
1443         return ret ? ret : task_count;
1444 }
1445
1446 /* The key must be already stored in q->key. */
1447 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1448         __acquires(&hb->lock)
1449 {
1450         struct futex_hash_bucket *hb;
1451
1452         hb = hash_futex(&q->key);
1453         q->lock_ptr = &hb->lock;
1454
1455         spin_lock(&hb->lock);
1456         return hb;
1457 }
1458
1459 static inline void
1460 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1461         __releases(&hb->lock)
1462 {
1463         spin_unlock(&hb->lock);
1464 }
1465
1466 /**
1467  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1468  * @q:  The futex_q to enqueue
1469  * @hb: The destination hash bucket
1470  *
1471  * The hb->lock must be held by the caller, and is released here. A call to
1472  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1473  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1474  * or nothing if the unqueue is done as part of the wake process and the unqueue
1475  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1476  * an example).
1477  */
1478 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1479         __releases(&hb->lock)
1480 {
1481         int prio;
1482
1483         /*
1484          * The priority used to register this element is
1485          * - either the real thread-priority for the real-time threads
1486          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1487          * - or MAX_RT_PRIO for non-RT threads.
1488          * Thus, all RT-threads are woken first in priority order, and
1489          * the others are woken last, in FIFO order.
1490          */
1491         prio = min(current->normal_prio, MAX_RT_PRIO);
1492
1493         plist_node_init(&q->list, prio);
1494         plist_add(&q->list, &hb->chain);
1495         q->task = current;
1496         spin_unlock(&hb->lock);
1497 }
1498
1499 /**
1500  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1501  * @q:  The futex_q to unqueue
1502  *
1503  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1504  * be paired with exactly one earlier call to queue_me().
1505  *
1506  * Returns:
1507  *   1 - if the futex_q was still queued (and we removed unqueued it)
1508  *   0 - if the futex_q was already removed by the waking thread
1509  */
1510 static int unqueue_me(struct futex_q *q)
1511 {
1512         spinlock_t *lock_ptr;
1513         int ret = 0;
1514
1515         /* In the common case we don't take the spinlock, which is nice. */
1516 retry:
1517         lock_ptr = q->lock_ptr;
1518         barrier();
1519         if (lock_ptr != NULL) {
1520                 spin_lock(lock_ptr);
1521                 /*
1522                  * q->lock_ptr can change between reading it and
1523                  * spin_lock(), causing us to take the wrong lock.  This
1524                  * corrects the race condition.
1525                  *
1526                  * Reasoning goes like this: if we have the wrong lock,
1527                  * q->lock_ptr must have changed (maybe several times)
1528                  * between reading it and the spin_lock().  It can
1529                  * change again after the spin_lock() but only if it was
1530                  * already changed before the spin_lock().  It cannot,
1531                  * however, change back to the original value.  Therefore
1532                  * we can detect whether we acquired the correct lock.
1533                  */
1534                 if (unlikely(lock_ptr != q->lock_ptr)) {
1535                         spin_unlock(lock_ptr);
1536                         goto retry;
1537                 }
1538                 __unqueue_futex(q);
1539
1540                 BUG_ON(q->pi_state);
1541
1542                 spin_unlock(lock_ptr);
1543                 ret = 1;
1544         }
1545
1546         drop_futex_key_refs(&q->key);
1547         return ret;
1548 }
1549
1550 /*
1551  * PI futexes can not be requeued and must remove themself from the
1552  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1553  * and dropped here.
1554  */
1555 static void unqueue_me_pi(struct futex_q *q)
1556         __releases(q->lock_ptr)
1557 {
1558         __unqueue_futex(q);
1559
1560         BUG_ON(!q->pi_state);
1561         free_pi_state(q->pi_state);
1562         q->pi_state = NULL;
1563
1564         spin_unlock(q->lock_ptr);
1565 }
1566
1567 /*
1568  * Fixup the pi_state owner with the new owner.
1569  *
1570  * Must be called with hash bucket lock held and mm->sem held for non
1571  * private futexes.
1572  */
1573 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1574                                 struct task_struct *newowner)
1575 {
1576         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1577         struct futex_pi_state *pi_state = q->pi_state;
1578         struct task_struct *oldowner = pi_state->owner;
1579         u32 uval, curval, newval;
1580         int ret;
1581
1582         /* Owner died? */
1583         if (!pi_state->owner)
1584                 newtid |= FUTEX_OWNER_DIED;
1585
1586         /*
1587          * We are here either because we stole the rtmutex from the
1588          * previous highest priority waiter or we are the highest priority
1589          * waiter but failed to get the rtmutex the first time.
1590          * We have to replace the newowner TID in the user space variable.
1591          * This must be atomic as we have to preserve the owner died bit here.
1592          *
1593          * Note: We write the user space value _before_ changing the pi_state
1594          * because we can fault here. Imagine swapped out pages or a fork
1595          * that marked all the anonymous memory readonly for cow.
1596          *
1597          * Modifying pi_state _before_ the user space value would
1598          * leave the pi_state in an inconsistent state when we fault
1599          * here, because we need to drop the hash bucket lock to
1600          * handle the fault. This might be observed in the PID check
1601          * in lookup_pi_state.
1602          */
1603 retry:
1604         if (get_futex_value_locked(&uval, uaddr))
1605                 goto handle_fault;
1606
1607         while (1) {
1608                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1609
1610                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1611                         goto handle_fault;
1612                 if (curval == uval)
1613                         break;
1614                 uval = curval;
1615         }
1616
1617         /*
1618          * We fixed up user space. Now we need to fix the pi_state
1619          * itself.
1620          */
1621         if (pi_state->owner != NULL) {
1622                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1623                 WARN_ON(list_empty(&pi_state->list));
1624                 list_del_init(&pi_state->list);
1625                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1626         }
1627
1628         pi_state->owner = newowner;
1629
1630         raw_spin_lock_irq(&newowner->pi_lock);
1631         WARN_ON(!list_empty(&pi_state->list));
1632         list_add(&pi_state->list, &newowner->pi_state_list);
1633         raw_spin_unlock_irq(&newowner->pi_lock);
1634         return 0;
1635
1636         /*
1637          * To handle the page fault we need to drop the hash bucket
1638          * lock here. That gives the other task (either the highest priority
1639          * waiter itself or the task which stole the rtmutex) the
1640          * chance to try the fixup of the pi_state. So once we are
1641          * back from handling the fault we need to check the pi_state
1642          * after reacquiring the hash bucket lock and before trying to
1643          * do another fixup. When the fixup has been done already we
1644          * simply return.
1645          */
1646 handle_fault:
1647         spin_unlock(q->lock_ptr);
1648
1649         ret = fault_in_user_writeable(uaddr);
1650
1651         spin_lock(q->lock_ptr);
1652
1653         /*
1654          * Check if someone else fixed it for us:
1655          */
1656         if (pi_state->owner != oldowner)
1657                 return 0;
1658
1659         if (ret)
1660                 return ret;
1661
1662         goto retry;
1663 }
1664
1665 static long futex_wait_restart(struct restart_block *restart);
1666
1667 /**
1668  * fixup_owner() - Post lock pi_state and corner case management
1669  * @uaddr:      user address of the futex
1670  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1671  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1672  *
1673  * After attempting to lock an rt_mutex, this function is called to cleanup
1674  * the pi_state owner as well as handle race conditions that may allow us to
1675  * acquire the lock. Must be called with the hb lock held.
1676  *
1677  * Returns:
1678  *  1 - success, lock taken
1679  *  0 - success, lock not taken
1680  * <0 - on error (-EFAULT)
1681  */
1682 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1683 {
1684         struct task_struct *owner;
1685         int ret = 0;
1686
1687         if (locked) {
1688                 /*
1689                  * Got the lock. We might not be the anticipated owner if we
1690                  * did a lock-steal - fix up the PI-state in that case:
1691                  */
1692                 if (q->pi_state->owner != current)
1693                         ret = fixup_pi_state_owner(uaddr, q, current);
1694                 goto out;
1695         }
1696
1697         /*
1698          * Catch the rare case, where the lock was released when we were on the
1699          * way back before we locked the hash bucket.
1700          */
1701         if (q->pi_state->owner == current) {
1702                 /*
1703                  * Try to get the rt_mutex now. This might fail as some other
1704                  * task acquired the rt_mutex after we removed ourself from the
1705                  * rt_mutex waiters list.
1706                  */
1707                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1708                         locked = 1;
1709                         goto out;
1710                 }
1711
1712                 /*
1713                  * pi_state is incorrect, some other task did a lock steal and
1714                  * we returned due to timeout or signal without taking the
1715                  * rt_mutex. Too late.
1716                  */
1717                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1718                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1719                 if (!owner)
1720                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1721                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1722                 ret = fixup_pi_state_owner(uaddr, q, owner);
1723                 goto out;
1724         }
1725
1726         /*
1727          * Paranoia check. If we did not take the lock, then we should not be
1728          * the owner of the rt_mutex.
1729          */
1730         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1731                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1732                                 "pi-state %p\n", ret,
1733                                 q->pi_state->pi_mutex.owner,
1734                                 q->pi_state->owner);
1735
1736 out:
1737         return ret ? ret : locked;
1738 }
1739
1740 /**
1741  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1742  * @hb:         the futex hash bucket, must be locked by the caller
1743  * @q:          the futex_q to queue up on
1744  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1745  */
1746 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1747                                 struct hrtimer_sleeper *timeout)
1748 {
1749         /*
1750          * The task state is guaranteed to be set before another task can
1751          * wake it. set_current_state() is implemented using set_mb() and
1752          * queue_me() calls spin_unlock() upon completion, both serializing
1753          * access to the hash list and forcing another memory barrier.
1754          */
1755         set_current_state(TASK_INTERRUPTIBLE);
1756         queue_me(q, hb);
1757
1758         /* Arm the timer */
1759         if (timeout) {
1760                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1761                 if (!hrtimer_active(&timeout->timer))
1762                         timeout->task = NULL;
1763         }
1764
1765         /*
1766          * If we have been removed from the hash list, then another task
1767          * has tried to wake us, and we can skip the call to schedule().
1768          */
1769         if (likely(!plist_node_empty(&q->list))) {
1770                 /*
1771                  * If the timer has already expired, current will already be
1772                  * flagged for rescheduling. Only call schedule if there
1773                  * is no timeout, or if it has yet to expire.
1774                  */
1775                 if (!timeout || timeout->task)
1776                         schedule();
1777         }
1778         __set_current_state(TASK_RUNNING);
1779 }
1780
1781 /**
1782  * futex_wait_setup() - Prepare to wait on a futex
1783  * @uaddr:      the futex userspace address
1784  * @val:        the expected value
1785  * @flags:      futex flags (FLAGS_SHARED, etc.)
1786  * @q:          the associated futex_q
1787  * @hb:         storage for hash_bucket pointer to be returned to caller
1788  *
1789  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1790  * compare it with the expected value.  Handle atomic faults internally.
1791  * Return with the hb lock held and a q.key reference on success, and unlocked
1792  * with no q.key reference on failure.
1793  *
1794  * Returns:
1795  *  0 - uaddr contains val and hb has been locked
1796  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1797  */
1798 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1799                            struct futex_q *q, struct futex_hash_bucket **hb)
1800 {
1801         u32 uval;
1802         int ret;
1803
1804         /*
1805          * Access the page AFTER the hash-bucket is locked.
1806          * Order is important:
1807          *
1808          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1809          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1810          *
1811          * The basic logical guarantee of a futex is that it blocks ONLY
1812          * if cond(var) is known to be true at the time of blocking, for
1813          * any cond.  If we locked the hash-bucket after testing *uaddr, that
1814          * would open a race condition where we could block indefinitely with
1815          * cond(var) false, which would violate the guarantee.
1816          *
1817          * On the other hand, we insert q and release the hash-bucket only
1818          * after testing *uaddr.  This guarantees that futex_wait() will NOT
1819          * absorb a wakeup if *uaddr does not match the desired values
1820          * while the syscall executes.
1821          */
1822 retry:
1823         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1824         if (unlikely(ret != 0))
1825                 return ret;
1826
1827 retry_private:
1828         *hb = queue_lock(q);
1829
1830         ret = get_futex_value_locked(&uval, uaddr);
1831
1832         if (ret) {
1833                 queue_unlock(q, *hb);
1834
1835                 ret = get_user(uval, uaddr);
1836                 if (ret)
1837                         goto out;
1838
1839                 if (!(flags & FLAGS_SHARED))
1840                         goto retry_private;
1841
1842                 put_futex_key(&q->key);
1843                 goto retry;
1844         }
1845
1846         if (uval != val) {
1847                 queue_unlock(q, *hb);
1848                 ret = -EWOULDBLOCK;
1849         }
1850
1851 out:
1852         if (ret)
1853                 put_futex_key(&q->key);
1854         return ret;
1855 }
1856
1857 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1858                       ktime_t *abs_time, u32 bitset)
1859 {
1860         struct hrtimer_sleeper timeout, *to = NULL;
1861         struct restart_block *restart;
1862         struct futex_hash_bucket *hb;
1863         struct futex_q q = futex_q_init;
1864         int ret;
1865
1866         if (!bitset)
1867                 return -EINVAL;
1868         q.bitset = bitset;
1869
1870         if (abs_time) {
1871                 to = &timeout;
1872
1873                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1874                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
1875                                       HRTIMER_MODE_ABS);
1876                 hrtimer_init_sleeper(to, current);
1877                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1878                                              current->timer_slack_ns);
1879         }
1880
1881 retry:
1882         /*
1883          * Prepare to wait on uaddr. On success, holds hb lock and increments
1884          * q.key refs.
1885          */
1886         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1887         if (ret)
1888                 goto out;
1889
1890         /* queue_me and wait for wakeup, timeout, or a signal. */
1891         futex_wait_queue_me(hb, &q, to);
1892
1893         /* If we were woken (and unqueued), we succeeded, whatever. */
1894         ret = 0;
1895         /* unqueue_me() drops q.key ref */
1896         if (!unqueue_me(&q))
1897                 goto out;
1898         ret = -ETIMEDOUT;
1899         if (to && !to->task)
1900                 goto out;
1901
1902         /*
1903          * We expect signal_pending(current), but we might be the
1904          * victim of a spurious wakeup as well.
1905          */
1906         if (!signal_pending(current))
1907                 goto retry;
1908
1909         ret = -ERESTARTSYS;
1910         if (!abs_time)
1911                 goto out;
1912
1913         restart = &current_thread_info()->restart_block;
1914         restart->fn = futex_wait_restart;
1915         restart->futex.uaddr = uaddr;
1916         restart->futex.val = val;
1917         restart->futex.time = abs_time->tv64;
1918         restart->futex.bitset = bitset;
1919         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1920
1921         ret = -ERESTART_RESTARTBLOCK;
1922
1923 out:
1924         if (to) {
1925                 hrtimer_cancel(&to->timer);
1926                 destroy_hrtimer_on_stack(&to->timer);
1927         }
1928         return ret;
1929 }
1930
1931
1932 static long futex_wait_restart(struct restart_block *restart)
1933 {
1934         u32 __user *uaddr = restart->futex.uaddr;
1935         ktime_t t, *tp = NULL;
1936
1937         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1938                 t.tv64 = restart->futex.time;
1939                 tp = &t;
1940         }
1941         restart->fn = do_no_restart_syscall;
1942
1943         return (long)futex_wait(uaddr, restart->futex.flags,
1944                                 restart->futex.val, tp, restart->futex.bitset);
1945 }
1946
1947
1948 /*
1949  * Userspace tried a 0 -> TID atomic transition of the futex value
1950  * and failed. The kernel side here does the whole locking operation:
1951  * if there are waiters then it will block, it does PI, etc. (Due to
1952  * races the kernel might see a 0 value of the futex too.)
1953  */
1954 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1955                          ktime_t *time, int trylock)
1956 {
1957         struct hrtimer_sleeper timeout, *to = NULL;
1958         struct futex_hash_bucket *hb;
1959         struct futex_q q = futex_q_init;
1960         int res, ret;
1961
1962         if (refill_pi_state_cache())
1963                 return -ENOMEM;
1964
1965         if (time) {
1966                 to = &timeout;
1967                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1968                                       HRTIMER_MODE_ABS);
1969                 hrtimer_init_sleeper(to, current);
1970                 hrtimer_set_expires(&to->timer, *time);
1971         }
1972
1973 retry:
1974         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
1975         if (unlikely(ret != 0))
1976                 goto out;
1977
1978 retry_private:
1979         hb = queue_lock(&q);
1980
1981         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1982         if (unlikely(ret)) {
1983                 switch (ret) {
1984                 case 1:
1985                         /* We got the lock. */
1986                         ret = 0;
1987                         goto out_unlock_put_key;
1988                 case -EFAULT:
1989                         goto uaddr_faulted;
1990                 case -EAGAIN:
1991                         /*
1992                          * Task is exiting and we just wait for the
1993                          * exit to complete.
1994                          */
1995                         queue_unlock(&q, hb);
1996                         put_futex_key(&q.key);
1997                         cond_resched();
1998                         goto retry;
1999                 default:
2000                         goto out_unlock_put_key;
2001                 }
2002         }
2003
2004         /*
2005          * Only actually queue now that the atomic ops are done:
2006          */
2007         queue_me(&q, hb);
2008
2009         WARN_ON(!q.pi_state);
2010         /*
2011          * Block on the PI mutex:
2012          */
2013         if (!trylock)
2014                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2015         else {
2016                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2017                 /* Fixup the trylock return value: */
2018                 ret = ret ? 0 : -EWOULDBLOCK;
2019         }
2020
2021         spin_lock(q.lock_ptr);
2022         /*
2023          * Fixup the pi_state owner and possibly acquire the lock if we
2024          * haven't already.
2025          */
2026         res = fixup_owner(uaddr, &q, !ret);
2027         /*
2028          * If fixup_owner() returned an error, proprogate that.  If it acquired
2029          * the lock, clear our -ETIMEDOUT or -EINTR.
2030          */
2031         if (res)
2032                 ret = (res < 0) ? res : 0;
2033
2034         /*
2035          * If fixup_owner() faulted and was unable to handle the fault, unlock
2036          * it and return the fault to userspace.
2037          */
2038         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2039                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2040
2041         /* Unqueue and drop the lock */
2042         unqueue_me_pi(&q);
2043
2044         goto out_put_key;
2045
2046 out_unlock_put_key:
2047         queue_unlock(&q, hb);
2048
2049 out_put_key:
2050         put_futex_key(&q.key);
2051 out:
2052         if (to)
2053                 destroy_hrtimer_on_stack(&to->timer);
2054         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2055
2056 uaddr_faulted:
2057         queue_unlock(&q, hb);
2058
2059         ret = fault_in_user_writeable(uaddr);
2060         if (ret)
2061                 goto out_put_key;
2062
2063         if (!(flags & FLAGS_SHARED))
2064                 goto retry_private;
2065
2066         put_futex_key(&q.key);
2067         goto retry;
2068 }
2069
2070 /*
2071  * Userspace attempted a TID -> 0 atomic transition, and failed.
2072  * This is the in-kernel slowpath: we look up the PI state (if any),
2073  * and do the rt-mutex unlock.
2074  */
2075 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2076 {
2077         struct futex_hash_bucket *hb;
2078         struct futex_q *this, *next;
2079         struct plist_head *head;
2080         union futex_key key = FUTEX_KEY_INIT;
2081         u32 uval, vpid = task_pid_vnr(current);
2082         int ret;
2083
2084 retry:
2085         if (get_user(uval, uaddr))
2086                 return -EFAULT;
2087         /*
2088          * We release only a lock we actually own:
2089          */
2090         if ((uval & FUTEX_TID_MASK) != vpid)
2091                 return -EPERM;
2092
2093         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2094         if (unlikely(ret != 0))
2095                 goto out;
2096
2097         hb = hash_futex(&key);
2098         spin_lock(&hb->lock);
2099
2100         /*
2101          * To avoid races, try to do the TID -> 0 atomic transition
2102          * again. If it succeeds then we can return without waking
2103          * anyone else up:
2104          */
2105         if (!(uval & FUTEX_OWNER_DIED) &&
2106             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2107                 goto pi_faulted;
2108         /*
2109          * Rare case: we managed to release the lock atomically,
2110          * no need to wake anyone else up:
2111          */
2112         if (unlikely(uval == vpid))
2113                 goto out_unlock;
2114
2115         /*
2116          * Ok, other tasks may need to be woken up - check waiters
2117          * and do the wakeup if necessary:
2118          */
2119         head = &hb->chain;
2120
2121         plist_for_each_entry_safe(this, next, head, list) {
2122                 if (!match_futex (&this->key, &key))
2123                         continue;
2124                 ret = wake_futex_pi(uaddr, uval, this);
2125                 /*
2126                  * The atomic access to the futex value
2127                  * generated a pagefault, so retry the
2128                  * user-access and the wakeup:
2129                  */
2130                 if (ret == -EFAULT)
2131                         goto pi_faulted;
2132                 goto out_unlock;
2133         }
2134         /*
2135          * No waiters - kernel unlocks the futex:
2136          */
2137         if (!(uval & FUTEX_OWNER_DIED)) {
2138                 ret = unlock_futex_pi(uaddr, uval);
2139                 if (ret == -EFAULT)
2140                         goto pi_faulted;
2141         }
2142
2143 out_unlock:
2144         spin_unlock(&hb->lock);
2145         put_futex_key(&key);
2146
2147 out:
2148         return ret;
2149
2150 pi_faulted:
2151         spin_unlock(&hb->lock);
2152         put_futex_key(&key);
2153
2154         ret = fault_in_user_writeable(uaddr);
2155         if (!ret)
2156                 goto retry;
2157
2158         return ret;
2159 }
2160
2161 /**
2162  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2163  * @hb:         the hash_bucket futex_q was original enqueued on
2164  * @q:          the futex_q woken while waiting to be requeued
2165  * @key2:       the futex_key of the requeue target futex
2166  * @timeout:    the timeout associated with the wait (NULL if none)
2167  *
2168  * Detect if the task was woken on the initial futex as opposed to the requeue
2169  * target futex.  If so, determine if it was a timeout or a signal that caused
2170  * the wakeup and return the appropriate error code to the caller.  Must be
2171  * called with the hb lock held.
2172  *
2173  * Returns
2174  *  0 - no early wakeup detected
2175  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2176  */
2177 static inline
2178 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2179                                    struct futex_q *q, union futex_key *key2,
2180                                    struct hrtimer_sleeper *timeout)
2181 {
2182         int ret = 0;
2183
2184         /*
2185          * With the hb lock held, we avoid races while we process the wakeup.
2186          * We only need to hold hb (and not hb2) to ensure atomicity as the
2187          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2188          * It can't be requeued from uaddr2 to something else since we don't
2189          * support a PI aware source futex for requeue.
2190          */
2191         if (!match_futex(&q->key, key2)) {
2192                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2193                 /*
2194                  * We were woken prior to requeue by a timeout or a signal.
2195                  * Unqueue the futex_q and determine which it was.
2196                  */
2197                 plist_del(&q->list, &hb->chain);
2198
2199                 /* Handle spurious wakeups gracefully */
2200                 ret = -EWOULDBLOCK;
2201                 if (timeout && !timeout->task)
2202                         ret = -ETIMEDOUT;
2203                 else if (signal_pending(current))
2204                         ret = -ERESTARTNOINTR;
2205         }
2206         return ret;
2207 }
2208
2209 /**
2210  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2211  * @uaddr:      the futex we initially wait on (non-pi)
2212  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2213  *              the same type, no requeueing from private to shared, etc.
2214  * @val:        the expected value of uaddr
2215  * @abs_time:   absolute timeout
2216  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2217  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2218  * @uaddr2:     the pi futex we will take prior to returning to user-space
2219  *
2220  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2221  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2222  * complete the acquisition of the rt_mutex prior to returning to userspace.
2223  * This ensures the rt_mutex maintains an owner when it has waiters; without
2224  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2225  * need to.
2226  *
2227  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2228  * via the following:
2229  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2230  * 2) wakeup on uaddr2 after a requeue
2231  * 3) signal
2232  * 4) timeout
2233  *
2234  * If 3, cleanup and return -ERESTARTNOINTR.
2235  *
2236  * If 2, we may then block on trying to take the rt_mutex and return via:
2237  * 5) successful lock
2238  * 6) signal
2239  * 7) timeout
2240  * 8) other lock acquisition failure
2241  *
2242  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2243  *
2244  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2245  *
2246  * Returns:
2247  *  0 - On success
2248  * <0 - On error
2249  */
2250 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2251                                  u32 val, ktime_t *abs_time, u32 bitset,
2252                                  u32 __user *uaddr2)
2253 {
2254         struct hrtimer_sleeper timeout, *to = NULL;
2255         struct rt_mutex_waiter rt_waiter;
2256         struct rt_mutex *pi_mutex = NULL;
2257         struct futex_hash_bucket *hb;
2258         union futex_key key2 = FUTEX_KEY_INIT;
2259         struct futex_q q = futex_q_init;
2260         int res, ret;
2261
2262         if (!bitset)
2263                 return -EINVAL;
2264
2265         if (abs_time) {
2266                 to = &timeout;
2267                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2268                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2269                                       HRTIMER_MODE_ABS);
2270                 hrtimer_init_sleeper(to, current);
2271                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2272                                              current->timer_slack_ns);
2273         }
2274
2275         /*
2276          * The waiter is allocated on our stack, manipulated by the requeue
2277          * code while we sleep on uaddr.
2278          */
2279         debug_rt_mutex_init_waiter(&rt_waiter);
2280         rt_waiter.task = NULL;
2281
2282         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2283         if (unlikely(ret != 0))
2284                 goto out;
2285
2286         q.bitset = bitset;
2287         q.rt_waiter = &rt_waiter;
2288         q.requeue_pi_key = &key2;
2289
2290         /*
2291          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2292          * count.
2293          */
2294         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2295         if (ret)
2296                 goto out_key2;
2297
2298         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2299         futex_wait_queue_me(hb, &q, to);
2300
2301         spin_lock(&hb->lock);
2302         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2303         spin_unlock(&hb->lock);
2304         if (ret)
2305                 goto out_put_keys;
2306
2307         /*
2308          * In order for us to be here, we know our q.key == key2, and since
2309          * we took the hb->lock above, we also know that futex_requeue() has
2310          * completed and we no longer have to concern ourselves with a wakeup
2311          * race with the atomic proxy lock acquisition by the requeue code. The
2312          * futex_requeue dropped our key1 reference and incremented our key2
2313          * reference count.
2314          */
2315
2316         /* Check if the requeue code acquired the second futex for us. */
2317         if (!q.rt_waiter) {
2318                 /*
2319                  * Got the lock. We might not be the anticipated owner if we
2320                  * did a lock-steal - fix up the PI-state in that case.
2321                  */
2322                 if (q.pi_state && (q.pi_state->owner != current)) {
2323                         spin_lock(q.lock_ptr);
2324                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2325                         spin_unlock(q.lock_ptr);
2326                 }
2327         } else {
2328                 /*
2329                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2330                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2331                  * the pi_state.
2332                  */
2333                 WARN_ON(!&q.pi_state);
2334                 pi_mutex = &q.pi_state->pi_mutex;
2335                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2336                 debug_rt_mutex_free_waiter(&rt_waiter);
2337
2338                 spin_lock(q.lock_ptr);
2339                 /*
2340                  * Fixup the pi_state owner and possibly acquire the lock if we
2341                  * haven't already.
2342                  */
2343                 res = fixup_owner(uaddr2, &q, !ret);
2344                 /*
2345                  * If fixup_owner() returned an error, proprogate that.  If it
2346                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2347                  */
2348                 if (res)
2349                         ret = (res < 0) ? res : 0;
2350
2351                 /* Unqueue and drop the lock. */
2352                 unqueue_me_pi(&q);
2353         }
2354
2355         /*
2356          * If fixup_pi_state_owner() faulted and was unable to handle the
2357          * fault, unlock the rt_mutex and return the fault to userspace.
2358          */
2359         if (ret == -EFAULT) {
2360                 if (rt_mutex_owner(pi_mutex) == current)
2361                         rt_mutex_unlock(pi_mutex);
2362         } else if (ret == -EINTR) {
2363                 /*
2364                  * We've already been requeued, but cannot restart by calling
2365                  * futex_lock_pi() directly. We could restart this syscall, but
2366                  * it would detect that the user space "val" changed and return
2367                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2368                  * -EWOULDBLOCK directly.
2369                  */
2370                 ret = -EWOULDBLOCK;
2371         }
2372
2373 out_put_keys:
2374         put_futex_key(&q.key);
2375 out_key2:
2376         put_futex_key(&key2);
2377
2378 out:
2379         if (to) {
2380                 hrtimer_cancel(&to->timer);
2381                 destroy_hrtimer_on_stack(&to->timer);
2382         }
2383         return ret;
2384 }
2385
2386 /*
2387  * Support for robust futexes: the kernel cleans up held futexes at
2388  * thread exit time.
2389  *
2390  * Implementation: user-space maintains a per-thread list of locks it
2391  * is holding. Upon do_exit(), the kernel carefully walks this list,
2392  * and marks all locks that are owned by this thread with the
2393  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2394  * always manipulated with the lock held, so the list is private and
2395  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2396  * field, to allow the kernel to clean up if the thread dies after
2397  * acquiring the lock, but just before it could have added itself to
2398  * the list. There can only be one such pending lock.
2399  */
2400
2401 /**
2402  * sys_set_robust_list() - Set the robust-futex list head of a task
2403  * @head:       pointer to the list-head
2404  * @len:        length of the list-head, as userspace expects
2405  */
2406 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2407                 size_t, len)
2408 {
2409         if (!futex_cmpxchg_enabled)
2410                 return -ENOSYS;
2411         /*
2412          * The kernel knows only one size for now:
2413          */
2414         if (unlikely(len != sizeof(*head)))
2415                 return -EINVAL;
2416
2417         current->robust_list = head;
2418
2419         return 0;
2420 }
2421
2422 /**
2423  * sys_get_robust_list() - Get the robust-futex list head of a task
2424  * @pid:        pid of the process [zero for current task]
2425  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2426  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2427  */
2428 SYSCALL_DEFINE3(get_robust_list, int, pid,
2429                 struct robust_list_head __user * __user *, head_ptr,
2430                 size_t __user *, len_ptr)
2431 {
2432         struct robust_list_head __user *head;
2433         unsigned long ret;
2434         const struct cred *cred = current_cred(), *pcred;
2435
2436         if (!futex_cmpxchg_enabled)
2437                 return -ENOSYS;
2438
2439         if (!pid)
2440                 head = current->robust_list;
2441         else {
2442                 struct task_struct *p;
2443
2444                 ret = -ESRCH;
2445                 rcu_read_lock();
2446                 p = find_task_by_vpid(pid);
2447                 if (!p)
2448                         goto err_unlock;
2449                 ret = -EPERM;
2450                 pcred = __task_cred(p);
2451                 /* If victim is in different user_ns, then uids are not
2452                    comparable, so we must have CAP_SYS_PTRACE */
2453                 if (cred->user->user_ns != pcred->user->user_ns) {
2454                         if (!ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2455                                 goto err_unlock;
2456                         goto ok;
2457                 }
2458                 /* If victim is in same user_ns, then uids are comparable */
2459                 if (cred->euid != pcred->euid &&
2460                     cred->euid != pcred->uid &&
2461                     !ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2462                         goto err_unlock;
2463 ok:
2464                 head = p->robust_list;
2465                 rcu_read_unlock();
2466         }
2467
2468         if (put_user(sizeof(*head), len_ptr))
2469                 return -EFAULT;
2470         return put_user(head, head_ptr);
2471
2472 err_unlock:
2473         rcu_read_unlock();
2474
2475         return ret;
2476 }
2477
2478 /*
2479  * Process a futex-list entry, check whether it's owned by the
2480  * dying task, and do notification if so:
2481  */
2482 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2483 {
2484         u32 uval, nval, mval;
2485
2486 retry:
2487         if (get_user(uval, uaddr))
2488                 return -1;
2489
2490         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2491                 /*
2492                  * Ok, this dying thread is truly holding a futex
2493                  * of interest. Set the OWNER_DIED bit atomically
2494                  * via cmpxchg, and if the value had FUTEX_WAITERS
2495                  * set, wake up a waiter (if any). (We have to do a
2496                  * futex_wake() even if OWNER_DIED is already set -
2497                  * to handle the rare but possible case of recursive
2498                  * thread-death.) The rest of the cleanup is done in
2499                  * userspace.
2500                  */
2501                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2502                 /*
2503                  * We are not holding a lock here, but we want to have
2504                  * the pagefault_disable/enable() protection because
2505                  * we want to handle the fault gracefully. If the
2506                  * access fails we try to fault in the futex with R/W
2507                  * verification via get_user_pages. get_user() above
2508                  * does not guarantee R/W access. If that fails we
2509                  * give up and leave the futex locked.
2510                  */
2511                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2512                         if (fault_in_user_writeable(uaddr))
2513                                 return -1;
2514                         goto retry;
2515                 }
2516                 if (nval != uval)
2517                         goto retry;
2518
2519                 /*
2520                  * Wake robust non-PI futexes here. The wakeup of
2521                  * PI futexes happens in exit_pi_state():
2522                  */
2523                 if (!pi && (uval & FUTEX_WAITERS))
2524                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2525         }
2526         return 0;
2527 }
2528
2529 /*
2530  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2531  */
2532 static inline int fetch_robust_entry(struct robust_list __user **entry,
2533                                      struct robust_list __user * __user *head,
2534                                      unsigned int *pi)
2535 {
2536         unsigned long uentry;
2537
2538         if (get_user(uentry, (unsigned long __user *)head))
2539                 return -EFAULT;
2540
2541         *entry = (void __user *)(uentry & ~1UL);
2542         *pi = uentry & 1;
2543
2544         return 0;
2545 }
2546
2547 /*
2548  * Walk curr->robust_list (very carefully, it's a userspace list!)
2549  * and mark any locks found there dead, and notify any waiters.
2550  *
2551  * We silently return on any sign of list-walking problem.
2552  */
2553 void exit_robust_list(struct task_struct *curr)
2554 {
2555         struct robust_list_head __user *head = curr->robust_list;
2556         struct robust_list __user *entry, *next_entry, *pending;
2557         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2558         unsigned int uninitialized_var(next_pi);
2559         unsigned long futex_offset;
2560         int rc;
2561
2562         if (!futex_cmpxchg_enabled)
2563                 return;
2564
2565         /*
2566          * Fetch the list head (which was registered earlier, via
2567          * sys_set_robust_list()):
2568          */
2569         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2570                 return;
2571         /*
2572          * Fetch the relative futex offset:
2573          */
2574         if (get_user(futex_offset, &head->futex_offset))
2575                 return;
2576         /*
2577          * Fetch any possibly pending lock-add first, and handle it
2578          * if it exists:
2579          */
2580         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2581                 return;
2582
2583         next_entry = NULL;      /* avoid warning with gcc */
2584         while (entry != &head->list) {
2585                 /*
2586                  * Fetch the next entry in the list before calling
2587                  * handle_futex_death:
2588                  */
2589                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2590                 /*
2591                  * A pending lock might already be on the list, so
2592                  * don't process it twice:
2593                  */
2594                 if (entry != pending)
2595                         if (handle_futex_death((void __user *)entry + futex_offset,
2596                                                 curr, pi))
2597                                 return;
2598                 if (rc)
2599                         return;
2600                 entry = next_entry;
2601                 pi = next_pi;
2602                 /*
2603                  * Avoid excessively long or circular lists:
2604                  */
2605                 if (!--limit)
2606                         break;
2607
2608                 cond_resched();
2609         }
2610
2611         if (pending)
2612                 handle_futex_death((void __user *)pending + futex_offset,
2613                                    curr, pip);
2614 }
2615
2616 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2617                 u32 __user *uaddr2, u32 val2, u32 val3)
2618 {
2619         int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2620         unsigned int flags = 0;
2621
2622         if (!(op & FUTEX_PRIVATE_FLAG))
2623                 flags |= FLAGS_SHARED;
2624
2625         if (op & FUTEX_CLOCK_REALTIME) {
2626                 flags |= FLAGS_CLOCKRT;
2627                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2628                         return -ENOSYS;
2629         }
2630
2631         switch (cmd) {
2632         case FUTEX_WAIT:
2633                 val3 = FUTEX_BITSET_MATCH_ANY;
2634         case FUTEX_WAIT_BITSET:
2635                 ret = futex_wait(uaddr, flags, val, timeout, val3);
2636                 break;
2637         case FUTEX_WAKE:
2638                 val3 = FUTEX_BITSET_MATCH_ANY;
2639         case FUTEX_WAKE_BITSET:
2640                 ret = futex_wake(uaddr, flags, val, val3);
2641                 break;
2642         case FUTEX_REQUEUE:
2643                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2644                 break;
2645         case FUTEX_CMP_REQUEUE:
2646                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2647                 break;
2648         case FUTEX_WAKE_OP:
2649                 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2650                 break;
2651         case FUTEX_LOCK_PI:
2652                 if (futex_cmpxchg_enabled)
2653                         ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2654                 break;
2655         case FUTEX_UNLOCK_PI:
2656                 if (futex_cmpxchg_enabled)
2657                         ret = futex_unlock_pi(uaddr, flags);
2658                 break;
2659         case FUTEX_TRYLOCK_PI:
2660                 if (futex_cmpxchg_enabled)
2661                         ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2662                 break;
2663         case FUTEX_WAIT_REQUEUE_PI:
2664                 val3 = FUTEX_BITSET_MATCH_ANY;
2665                 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2666                                             uaddr2);
2667                 break;
2668         case FUTEX_CMP_REQUEUE_PI:
2669                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2670                 break;
2671         default:
2672                 ret = -ENOSYS;
2673         }
2674         return ret;
2675 }
2676
2677
2678 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2679                 struct timespec __user *, utime, u32 __user *, uaddr2,
2680                 u32, val3)
2681 {
2682         struct timespec ts;
2683         ktime_t t, *tp = NULL;
2684         u32 val2 = 0;
2685         int cmd = op & FUTEX_CMD_MASK;
2686
2687         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2688                       cmd == FUTEX_WAIT_BITSET ||
2689                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2690                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2691                         return -EFAULT;
2692                 if (!timespec_valid(&ts))
2693                         return -EINVAL;
2694
2695                 t = timespec_to_ktime(ts);
2696                 if (cmd == FUTEX_WAIT)
2697                         t = ktime_add_safe(ktime_get(), t);
2698                 tp = &t;
2699         }
2700         /*
2701          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2702          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2703          */
2704         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2705             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2706                 val2 = (u32) (unsigned long) utime;
2707
2708         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2709 }
2710
2711 static int __init futex_init(void)
2712 {
2713         u32 curval;
2714         int i;
2715
2716         /*
2717          * This will fail and we want it. Some arch implementations do
2718          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2719          * functionality. We want to know that before we call in any
2720          * of the complex code paths. Also we want to prevent
2721          * registration of robust lists in that case. NULL is
2722          * guaranteed to fault and we get -EFAULT on functional
2723          * implementation, the non-functional ones will return
2724          * -ENOSYS.
2725          */
2726         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2727                 futex_cmpxchg_enabled = 1;
2728
2729         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2730                 plist_head_init(&futex_queues[i].chain);
2731                 spin_lock_init(&futex_queues[i].lock);
2732         }
2733
2734         return 0;
2735 }
2736 __initcall(futex_init);