Merge branch 'core-locking-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[~shefty/rdma-dev.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/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64
65 #include <asm/futex.h>
66
67 #include "rtmutex_common.h"
68
69 int __read_mostly futex_cmpxchg_enabled;
70
71 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
72
73 /*
74  * Futex flags used to encode options to functions and preserve them across
75  * restarts.
76  */
77 #define FLAGS_SHARED            0x01
78 #define FLAGS_CLOCKRT           0x02
79 #define FLAGS_HAS_TIMEOUT       0x04
80
81 /*
82  * Priority Inheritance state:
83  */
84 struct futex_pi_state {
85         /*
86          * list of 'owned' pi_state instances - these have to be
87          * cleaned up in do_exit() if the task exits prematurely:
88          */
89         struct list_head list;
90
91         /*
92          * The PI object:
93          */
94         struct rt_mutex pi_mutex;
95
96         struct task_struct *owner;
97         atomic_t refcount;
98
99         union futex_key key;
100 };
101
102 /**
103  * struct futex_q - The hashed futex queue entry, one per waiting task
104  * @list:               priority-sorted list of tasks waiting on this futex
105  * @task:               the task waiting on the futex
106  * @lock_ptr:           the hash bucket lock
107  * @key:                the key the futex is hashed on
108  * @pi_state:           optional priority inheritance state
109  * @rt_waiter:          rt_waiter storage for use with requeue_pi
110  * @requeue_pi_key:     the requeue_pi target futex key
111  * @bitset:             bitset for the optional bitmasked wakeup
112  *
113  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
114  * we can wake only the relevant ones (hashed queues may be shared).
115  *
116  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
117  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
118  * The order of wakeup is always to make the first condition true, then
119  * the second.
120  *
121  * PI futexes are typically woken before they are removed from the hash list via
122  * the rt_mutex code. See unqueue_me_pi().
123  */
124 struct futex_q {
125         struct plist_node list;
126
127         struct task_struct *task;
128         spinlock_t *lock_ptr;
129         union futex_key key;
130         struct futex_pi_state *pi_state;
131         struct rt_mutex_waiter *rt_waiter;
132         union futex_key *requeue_pi_key;
133         u32 bitset;
134 };
135
136 static const struct futex_q futex_q_init = {
137         /* list gets initialized in queue_me()*/
138         .key = FUTEX_KEY_INIT,
139         .bitset = FUTEX_BITSET_MATCH_ANY
140 };
141
142 /*
143  * Hash buckets are shared by all the futex_keys that hash to the same
144  * location.  Each key may have multiple futex_q structures, one for each task
145  * waiting on a futex.
146  */
147 struct futex_hash_bucket {
148         spinlock_t lock;
149         struct plist_head chain;
150 };
151
152 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
153
154 /*
155  * We hash on the keys returned from get_futex_key (see below).
156  */
157 static struct futex_hash_bucket *hash_futex(union futex_key *key)
158 {
159         u32 hash = jhash2((u32*)&key->both.word,
160                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
161                           key->both.offset);
162         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
163 }
164
165 /*
166  * Return 1 if two futex_keys are equal, 0 otherwise.
167  */
168 static inline int match_futex(union futex_key *key1, union futex_key *key2)
169 {
170         return (key1 && key2
171                 && key1->both.word == key2->both.word
172                 && key1->both.ptr == key2->both.ptr
173                 && key1->both.offset == key2->both.offset);
174 }
175
176 /*
177  * Take a reference to the resource addressed by a key.
178  * Can be called while holding spinlocks.
179  *
180  */
181 static void get_futex_key_refs(union futex_key *key)
182 {
183         if (!key->both.ptr)
184                 return;
185
186         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
187         case FUT_OFF_INODE:
188                 ihold(key->shared.inode);
189                 break;
190         case FUT_OFF_MMSHARED:
191                 atomic_inc(&key->private.mm->mm_count);
192                 break;
193         }
194 }
195
196 /*
197  * Drop a reference to the resource addressed by a key.
198  * The hash bucket spinlock must not be held.
199  */
200 static void drop_futex_key_refs(union futex_key *key)
201 {
202         if (!key->both.ptr) {
203                 /* If we're here then we tried to put a key we failed to get */
204                 WARN_ON_ONCE(1);
205                 return;
206         }
207
208         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
209         case FUT_OFF_INODE:
210                 iput(key->shared.inode);
211                 break;
212         case FUT_OFF_MMSHARED:
213                 mmdrop(key->private.mm);
214                 break;
215         }
216 }
217
218 /**
219  * get_futex_key() - Get parameters which are the keys for a futex
220  * @uaddr:      virtual address of the futex
221  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
222  * @key:        address where result is stored.
223  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
224  *              VERIFY_WRITE)
225  *
226  * Returns a negative error code or 0
227  * The key words are stored in *key on success.
228  *
229  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
230  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
231  * We can usually work out the index without swapping in the page.
232  *
233  * lock_page() might sleep, the caller should not hold a spinlock.
234  */
235 static int
236 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
237 {
238         unsigned long address = (unsigned long)uaddr;
239         struct mm_struct *mm = current->mm;
240         struct page *page, *page_head;
241         int err, ro = 0;
242
243         /*
244          * The futex address must be "naturally" aligned.
245          */
246         key->both.offset = address % PAGE_SIZE;
247         if (unlikely((address % sizeof(u32)) != 0))
248                 return -EINVAL;
249         address -= key->both.offset;
250
251         /*
252          * PROCESS_PRIVATE futexes are fast.
253          * As the mm cannot disappear under us and the 'key' only needs
254          * virtual address, we dont even have to find the underlying vma.
255          * Note : We do have to check 'uaddr' is a valid user address,
256          *        but access_ok() should be faster than find_vma()
257          */
258         if (!fshared) {
259                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
260                         return -EFAULT;
261                 key->private.mm = mm;
262                 key->private.address = address;
263                 get_futex_key_refs(key);
264                 return 0;
265         }
266
267 again:
268         err = get_user_pages_fast(address, 1, 1, &page);
269         /*
270          * If write access is not required (eg. FUTEX_WAIT), try
271          * and get read-only access.
272          */
273         if (err == -EFAULT && rw == VERIFY_READ) {
274                 err = get_user_pages_fast(address, 1, 0, &page);
275                 ro = 1;
276         }
277         if (err < 0)
278                 return err;
279         else
280                 err = 0;
281
282 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
283         page_head = page;
284         if (unlikely(PageTail(page))) {
285                 put_page(page);
286                 /* serialize against __split_huge_page_splitting() */
287                 local_irq_disable();
288                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
289                         page_head = compound_head(page);
290                         /*
291                          * page_head is valid pointer but we must pin
292                          * it before taking the PG_lock and/or
293                          * PG_compound_lock. The moment we re-enable
294                          * irqs __split_huge_page_splitting() can
295                          * return and the head page can be freed from
296                          * under us. We can't take the PG_lock and/or
297                          * PG_compound_lock on a page that could be
298                          * freed from under us.
299                          */
300                         if (page != page_head) {
301                                 get_page(page_head);
302                                 put_page(page);
303                         }
304                         local_irq_enable();
305                 } else {
306                         local_irq_enable();
307                         goto again;
308                 }
309         }
310 #else
311         page_head = compound_head(page);
312         if (page != page_head) {
313                 get_page(page_head);
314                 put_page(page);
315         }
316 #endif
317
318         lock_page(page_head);
319
320         /*
321          * If page_head->mapping is NULL, then it cannot be a PageAnon
322          * page; but it might be the ZERO_PAGE or in the gate area or
323          * in a special mapping (all cases which we are happy to fail);
324          * or it may have been a good file page when get_user_pages_fast
325          * found it, but truncated or holepunched or subjected to
326          * invalidate_complete_page2 before we got the page lock (also
327          * cases which we are happy to fail).  And we hold a reference,
328          * so refcount care in invalidate_complete_page's remove_mapping
329          * prevents drop_caches from setting mapping to NULL beneath us.
330          *
331          * The case we do have to guard against is when memory pressure made
332          * shmem_writepage move it from filecache to swapcache beneath us:
333          * an unlikely race, but we do need to retry for page_head->mapping.
334          */
335         if (!page_head->mapping) {
336                 int shmem_swizzled = PageSwapCache(page_head);
337                 unlock_page(page_head);
338                 put_page(page_head);
339                 if (shmem_swizzled)
340                         goto again;
341                 return -EFAULT;
342         }
343
344         /*
345          * Private mappings are handled in a simple way.
346          *
347          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
348          * it's a read-only handle, it's expected that futexes attach to
349          * the object not the particular process.
350          */
351         if (PageAnon(page_head)) {
352                 /*
353                  * A RO anonymous page will never change and thus doesn't make
354                  * sense for futex operations.
355                  */
356                 if (ro) {
357                         err = -EFAULT;
358                         goto out;
359                 }
360
361                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
362                 key->private.mm = mm;
363                 key->private.address = address;
364         } else {
365                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
366                 key->shared.inode = page_head->mapping->host;
367                 key->shared.pgoff = page_head->index;
368         }
369
370         get_futex_key_refs(key);
371
372 out:
373         unlock_page(page_head);
374         put_page(page_head);
375         return err;
376 }
377
378 static inline void put_futex_key(union futex_key *key)
379 {
380         drop_futex_key_refs(key);
381 }
382
383 /**
384  * fault_in_user_writeable() - Fault in user address and verify RW access
385  * @uaddr:      pointer to faulting user space address
386  *
387  * Slow path to fixup the fault we just took in the atomic write
388  * access to @uaddr.
389  *
390  * We have no generic implementation of a non-destructive write to the
391  * user address. We know that we faulted in the atomic pagefault
392  * disabled section so we can as well avoid the #PF overhead by
393  * calling get_user_pages() right away.
394  */
395 static int fault_in_user_writeable(u32 __user *uaddr)
396 {
397         struct mm_struct *mm = current->mm;
398         int ret;
399
400         down_read(&mm->mmap_sem);
401         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
402                                FAULT_FLAG_WRITE);
403         up_read(&mm->mmap_sem);
404
405         return ret < 0 ? ret : 0;
406 }
407
408 /**
409  * futex_top_waiter() - Return the highest priority waiter on a futex
410  * @hb:         the hash bucket the futex_q's reside in
411  * @key:        the futex key (to distinguish it from other futex futex_q's)
412  *
413  * Must be called with the hb lock held.
414  */
415 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
416                                         union futex_key *key)
417 {
418         struct futex_q *this;
419
420         plist_for_each_entry(this, &hb->chain, list) {
421                 if (match_futex(&this->key, key))
422                         return this;
423         }
424         return NULL;
425 }
426
427 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
428                                       u32 uval, u32 newval)
429 {
430         int ret;
431
432         pagefault_disable();
433         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
434         pagefault_enable();
435
436         return ret;
437 }
438
439 static int get_futex_value_locked(u32 *dest, u32 __user *from)
440 {
441         int ret;
442
443         pagefault_disable();
444         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
445         pagefault_enable();
446
447         return ret ? -EFAULT : 0;
448 }
449
450
451 /*
452  * PI code:
453  */
454 static int refill_pi_state_cache(void)
455 {
456         struct futex_pi_state *pi_state;
457
458         if (likely(current->pi_state_cache))
459                 return 0;
460
461         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
462
463         if (!pi_state)
464                 return -ENOMEM;
465
466         INIT_LIST_HEAD(&pi_state->list);
467         /* pi_mutex gets initialized later */
468         pi_state->owner = NULL;
469         atomic_set(&pi_state->refcount, 1);
470         pi_state->key = FUTEX_KEY_INIT;
471
472         current->pi_state_cache = pi_state;
473
474         return 0;
475 }
476
477 static struct futex_pi_state * alloc_pi_state(void)
478 {
479         struct futex_pi_state *pi_state = current->pi_state_cache;
480
481         WARN_ON(!pi_state);
482         current->pi_state_cache = NULL;
483
484         return pi_state;
485 }
486
487 static void free_pi_state(struct futex_pi_state *pi_state)
488 {
489         if (!atomic_dec_and_test(&pi_state->refcount))
490                 return;
491
492         /*
493          * If pi_state->owner is NULL, the owner is most probably dying
494          * and has cleaned up the pi_state already
495          */
496         if (pi_state->owner) {
497                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
498                 list_del_init(&pi_state->list);
499                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
500
501                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
502         }
503
504         if (current->pi_state_cache)
505                 kfree(pi_state);
506         else {
507                 /*
508                  * pi_state->list is already empty.
509                  * clear pi_state->owner.
510                  * refcount is at 0 - put it back to 1.
511                  */
512                 pi_state->owner = NULL;
513                 atomic_set(&pi_state->refcount, 1);
514                 current->pi_state_cache = pi_state;
515         }
516 }
517
518 /*
519  * Look up the task based on what TID userspace gave us.
520  * We dont trust it.
521  */
522 static struct task_struct * futex_find_get_task(pid_t pid)
523 {
524         struct task_struct *p;
525
526         rcu_read_lock();
527         p = find_task_by_vpid(pid);
528         if (p)
529                 get_task_struct(p);
530
531         rcu_read_unlock();
532
533         return p;
534 }
535
536 /*
537  * This task is holding PI mutexes at exit time => bad.
538  * Kernel cleans up PI-state, but userspace is likely hosed.
539  * (Robust-futex cleanup is separate and might save the day for userspace.)
540  */
541 void exit_pi_state_list(struct task_struct *curr)
542 {
543         struct list_head *next, *head = &curr->pi_state_list;
544         struct futex_pi_state *pi_state;
545         struct futex_hash_bucket *hb;
546         union futex_key key = FUTEX_KEY_INIT;
547
548         if (!futex_cmpxchg_enabled)
549                 return;
550         /*
551          * We are a ZOMBIE and nobody can enqueue itself on
552          * pi_state_list anymore, but we have to be careful
553          * versus waiters unqueueing themselves:
554          */
555         raw_spin_lock_irq(&curr->pi_lock);
556         while (!list_empty(head)) {
557
558                 next = head->next;
559                 pi_state = list_entry(next, struct futex_pi_state, list);
560                 key = pi_state->key;
561                 hb = hash_futex(&key);
562                 raw_spin_unlock_irq(&curr->pi_lock);
563
564                 spin_lock(&hb->lock);
565
566                 raw_spin_lock_irq(&curr->pi_lock);
567                 /*
568                  * We dropped the pi-lock, so re-check whether this
569                  * task still owns the PI-state:
570                  */
571                 if (head->next != next) {
572                         spin_unlock(&hb->lock);
573                         continue;
574                 }
575
576                 WARN_ON(pi_state->owner != curr);
577                 WARN_ON(list_empty(&pi_state->list));
578                 list_del_init(&pi_state->list);
579                 pi_state->owner = NULL;
580                 raw_spin_unlock_irq(&curr->pi_lock);
581
582                 rt_mutex_unlock(&pi_state->pi_mutex);
583
584                 spin_unlock(&hb->lock);
585
586                 raw_spin_lock_irq(&curr->pi_lock);
587         }
588         raw_spin_unlock_irq(&curr->pi_lock);
589 }
590
591 static int
592 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
593                 union futex_key *key, struct futex_pi_state **ps)
594 {
595         struct futex_pi_state *pi_state = NULL;
596         struct futex_q *this, *next;
597         struct plist_head *head;
598         struct task_struct *p;
599         pid_t pid = uval & FUTEX_TID_MASK;
600
601         head = &hb->chain;
602
603         plist_for_each_entry_safe(this, next, head, list) {
604                 if (match_futex(&this->key, key)) {
605                         /*
606                          * Another waiter already exists - bump up
607                          * the refcount and return its pi_state:
608                          */
609                         pi_state = this->pi_state;
610                         /*
611                          * Userspace might have messed up non-PI and PI futexes
612                          */
613                         if (unlikely(!pi_state))
614                                 return -EINVAL;
615
616                         WARN_ON(!atomic_read(&pi_state->refcount));
617
618                         /*
619                          * When pi_state->owner is NULL then the owner died
620                          * and another waiter is on the fly. pi_state->owner
621                          * is fixed up by the task which acquires
622                          * pi_state->rt_mutex.
623                          *
624                          * We do not check for pid == 0 which can happen when
625                          * the owner died and robust_list_exit() cleared the
626                          * TID.
627                          */
628                         if (pid && pi_state->owner) {
629                                 /*
630                                  * Bail out if user space manipulated the
631                                  * futex value.
632                                  */
633                                 if (pid != task_pid_vnr(pi_state->owner))
634                                         return -EINVAL;
635                         }
636
637                         atomic_inc(&pi_state->refcount);
638                         *ps = pi_state;
639
640                         return 0;
641                 }
642         }
643
644         /*
645          * We are the first waiter - try to look up the real owner and attach
646          * the new pi_state to it, but bail out when TID = 0
647          */
648         if (!pid)
649                 return -ESRCH;
650         p = futex_find_get_task(pid);
651         if (!p)
652                 return -ESRCH;
653
654         /*
655          * We need to look at the task state flags to figure out,
656          * whether the task is exiting. To protect against the do_exit
657          * change of the task flags, we do this protected by
658          * p->pi_lock:
659          */
660         raw_spin_lock_irq(&p->pi_lock);
661         if (unlikely(p->flags & PF_EXITING)) {
662                 /*
663                  * The task is on the way out. When PF_EXITPIDONE is
664                  * set, we know that the task has finished the
665                  * cleanup:
666                  */
667                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
668
669                 raw_spin_unlock_irq(&p->pi_lock);
670                 put_task_struct(p);
671                 return ret;
672         }
673
674         pi_state = alloc_pi_state();
675
676         /*
677          * Initialize the pi_mutex in locked state and make 'p'
678          * the owner of it:
679          */
680         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
681
682         /* Store the key for possible exit cleanups: */
683         pi_state->key = *key;
684
685         WARN_ON(!list_empty(&pi_state->list));
686         list_add(&pi_state->list, &p->pi_state_list);
687         pi_state->owner = p;
688         raw_spin_unlock_irq(&p->pi_lock);
689
690         put_task_struct(p);
691
692         *ps = pi_state;
693
694         return 0;
695 }
696
697 /**
698  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
699  * @uaddr:              the pi futex user address
700  * @hb:                 the pi futex hash bucket
701  * @key:                the futex key associated with uaddr and hb
702  * @ps:                 the pi_state pointer where we store the result of the
703  *                      lookup
704  * @task:               the task to perform the atomic lock work for.  This will
705  *                      be "current" except in the case of requeue pi.
706  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
707  *
708  * Returns:
709  *  0 - ready to wait
710  *  1 - acquired the lock
711  * <0 - error
712  *
713  * The hb->lock and futex_key refs shall be held by the caller.
714  */
715 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
716                                 union futex_key *key,
717                                 struct futex_pi_state **ps,
718                                 struct task_struct *task, int set_waiters)
719 {
720         int lock_taken, ret, force_take = 0;
721         u32 uval, newval, curval, vpid = task_pid_vnr(task);
722
723 retry:
724         ret = lock_taken = 0;
725
726         /*
727          * To avoid races, we attempt to take the lock here again
728          * (by doing a 0 -> TID atomic cmpxchg), while holding all
729          * the locks. It will most likely not succeed.
730          */
731         newval = vpid;
732         if (set_waiters)
733                 newval |= FUTEX_WAITERS;
734
735         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
736                 return -EFAULT;
737
738         /*
739          * Detect deadlocks.
740          */
741         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
742                 return -EDEADLK;
743
744         /*
745          * Surprise - we got the lock. Just return to userspace:
746          */
747         if (unlikely(!curval))
748                 return 1;
749
750         uval = curval;
751
752         /*
753          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
754          * to wake at the next unlock.
755          */
756         newval = curval | FUTEX_WAITERS;
757
758         /*
759          * Should we force take the futex? See below.
760          */
761         if (unlikely(force_take)) {
762                 /*
763                  * Keep the OWNER_DIED and the WAITERS bit and set the
764                  * new TID value.
765                  */
766                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
767                 force_take = 0;
768                 lock_taken = 1;
769         }
770
771         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
772                 return -EFAULT;
773         if (unlikely(curval != uval))
774                 goto retry;
775
776         /*
777          * We took the lock due to forced take over.
778          */
779         if (unlikely(lock_taken))
780                 return 1;
781
782         /*
783          * We dont have the lock. Look up the PI state (or create it if
784          * we are the first waiter):
785          */
786         ret = lookup_pi_state(uval, hb, key, ps);
787
788         if (unlikely(ret)) {
789                 switch (ret) {
790                 case -ESRCH:
791                         /*
792                          * We failed to find an owner for this
793                          * futex. So we have no pi_state to block
794                          * on. This can happen in two cases:
795                          *
796                          * 1) The owner died
797                          * 2) A stale FUTEX_WAITERS bit
798                          *
799                          * Re-read the futex value.
800                          */
801                         if (get_futex_value_locked(&curval, uaddr))
802                                 return -EFAULT;
803
804                         /*
805                          * If the owner died or we have a stale
806                          * WAITERS bit the owner TID in the user space
807                          * futex is 0.
808                          */
809                         if (!(curval & FUTEX_TID_MASK)) {
810                                 force_take = 1;
811                                 goto retry;
812                         }
813                 default:
814                         break;
815                 }
816         }
817
818         return ret;
819 }
820
821 /**
822  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
823  * @q:  The futex_q to unqueue
824  *
825  * The q->lock_ptr must not be NULL and must be held by the caller.
826  */
827 static void __unqueue_futex(struct futex_q *q)
828 {
829         struct futex_hash_bucket *hb;
830
831         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
832             || WARN_ON(plist_node_empty(&q->list)))
833                 return;
834
835         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
836         plist_del(&q->list, &hb->chain);
837 }
838
839 /*
840  * The hash bucket lock must be held when this is called.
841  * Afterwards, the futex_q must not be accessed.
842  */
843 static void wake_futex(struct futex_q *q)
844 {
845         struct task_struct *p = q->task;
846
847         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
848                 return;
849
850         /*
851          * We set q->lock_ptr = NULL _before_ we wake up the task. If
852          * a non-futex wake up happens on another CPU then the task
853          * might exit and p would dereference a non-existing task
854          * struct. Prevent this by holding a reference on p across the
855          * wake up.
856          */
857         get_task_struct(p);
858
859         __unqueue_futex(q);
860         /*
861          * The waiting task can free the futex_q as soon as
862          * q->lock_ptr = NULL is written, without taking any locks. A
863          * memory barrier is required here to prevent the following
864          * store to lock_ptr from getting ahead of the plist_del.
865          */
866         smp_wmb();
867         q->lock_ptr = NULL;
868
869         wake_up_state(p, TASK_NORMAL);
870         put_task_struct(p);
871 }
872
873 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
874 {
875         struct task_struct *new_owner;
876         struct futex_pi_state *pi_state = this->pi_state;
877         u32 uninitialized_var(curval), newval;
878
879         if (!pi_state)
880                 return -EINVAL;
881
882         /*
883          * If current does not own the pi_state then the futex is
884          * inconsistent and user space fiddled with the futex value.
885          */
886         if (pi_state->owner != current)
887                 return -EINVAL;
888
889         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
890         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
891
892         /*
893          * It is possible that the next waiter (the one that brought
894          * this owner to the kernel) timed out and is no longer
895          * waiting on the lock.
896          */
897         if (!new_owner)
898                 new_owner = this->task;
899
900         /*
901          * We pass it to the next owner. (The WAITERS bit is always
902          * kept enabled while there is PI state around. We must also
903          * preserve the owner died bit.)
904          */
905         if (!(uval & FUTEX_OWNER_DIED)) {
906                 int ret = 0;
907
908                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
909
910                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
911                         ret = -EFAULT;
912                 else if (curval != uval)
913                         ret = -EINVAL;
914                 if (ret) {
915                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
916                         return ret;
917                 }
918         }
919
920         raw_spin_lock_irq(&pi_state->owner->pi_lock);
921         WARN_ON(list_empty(&pi_state->list));
922         list_del_init(&pi_state->list);
923         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
924
925         raw_spin_lock_irq(&new_owner->pi_lock);
926         WARN_ON(!list_empty(&pi_state->list));
927         list_add(&pi_state->list, &new_owner->pi_state_list);
928         pi_state->owner = new_owner;
929         raw_spin_unlock_irq(&new_owner->pi_lock);
930
931         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
932         rt_mutex_unlock(&pi_state->pi_mutex);
933
934         return 0;
935 }
936
937 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
938 {
939         u32 uninitialized_var(oldval);
940
941         /*
942          * There is no waiter, so we unlock the futex. The owner died
943          * bit has not to be preserved here. We are the owner:
944          */
945         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
946                 return -EFAULT;
947         if (oldval != uval)
948                 return -EAGAIN;
949
950         return 0;
951 }
952
953 /*
954  * Express the locking dependencies for lockdep:
955  */
956 static inline void
957 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
958 {
959         if (hb1 <= hb2) {
960                 spin_lock(&hb1->lock);
961                 if (hb1 < hb2)
962                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
963         } else { /* hb1 > hb2 */
964                 spin_lock(&hb2->lock);
965                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
966         }
967 }
968
969 static inline void
970 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
971 {
972         spin_unlock(&hb1->lock);
973         if (hb1 != hb2)
974                 spin_unlock(&hb2->lock);
975 }
976
977 /*
978  * Wake up waiters matching bitset queued on this futex (uaddr).
979  */
980 static int
981 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
982 {
983         struct futex_hash_bucket *hb;
984         struct futex_q *this, *next;
985         struct plist_head *head;
986         union futex_key key = FUTEX_KEY_INIT;
987         int ret;
988
989         if (!bitset)
990                 return -EINVAL;
991
992         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
993         if (unlikely(ret != 0))
994                 goto out;
995
996         hb = hash_futex(&key);
997         spin_lock(&hb->lock);
998         head = &hb->chain;
999
1000         plist_for_each_entry_safe(this, next, head, list) {
1001                 if (match_futex (&this->key, &key)) {
1002                         if (this->pi_state || this->rt_waiter) {
1003                                 ret = -EINVAL;
1004                                 break;
1005                         }
1006
1007                         /* Check if one of the bits is set in both bitsets */
1008                         if (!(this->bitset & bitset))
1009                                 continue;
1010
1011                         wake_futex(this);
1012                         if (++ret >= nr_wake)
1013                                 break;
1014                 }
1015         }
1016
1017         spin_unlock(&hb->lock);
1018         put_futex_key(&key);
1019 out:
1020         return ret;
1021 }
1022
1023 /*
1024  * Wake up all waiters hashed on the physical page that is mapped
1025  * to this virtual address:
1026  */
1027 static int
1028 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1029               int nr_wake, int nr_wake2, int op)
1030 {
1031         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1032         struct futex_hash_bucket *hb1, *hb2;
1033         struct plist_head *head;
1034         struct futex_q *this, *next;
1035         int ret, op_ret;
1036
1037 retry:
1038         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1039         if (unlikely(ret != 0))
1040                 goto out;
1041         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1042         if (unlikely(ret != 0))
1043                 goto out_put_key1;
1044
1045         hb1 = hash_futex(&key1);
1046         hb2 = hash_futex(&key2);
1047
1048 retry_private:
1049         double_lock_hb(hb1, hb2);
1050         op_ret = futex_atomic_op_inuser(op, uaddr2);
1051         if (unlikely(op_ret < 0)) {
1052
1053                 double_unlock_hb(hb1, hb2);
1054
1055 #ifndef CONFIG_MMU
1056                 /*
1057                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1058                  * but we might get them from range checking
1059                  */
1060                 ret = op_ret;
1061                 goto out_put_keys;
1062 #endif
1063
1064                 if (unlikely(op_ret != -EFAULT)) {
1065                         ret = op_ret;
1066                         goto out_put_keys;
1067                 }
1068
1069                 ret = fault_in_user_writeable(uaddr2);
1070                 if (ret)
1071                         goto out_put_keys;
1072
1073                 if (!(flags & FLAGS_SHARED))
1074                         goto retry_private;
1075
1076                 put_futex_key(&key2);
1077                 put_futex_key(&key1);
1078                 goto retry;
1079         }
1080
1081         head = &hb1->chain;
1082
1083         plist_for_each_entry_safe(this, next, head, list) {
1084                 if (match_futex (&this->key, &key1)) {
1085                         if (this->pi_state || this->rt_waiter) {
1086                                 ret = -EINVAL;
1087                                 goto out_unlock;
1088                         }
1089                         wake_futex(this);
1090                         if (++ret >= nr_wake)
1091                                 break;
1092                 }
1093         }
1094
1095         if (op_ret > 0) {
1096                 head = &hb2->chain;
1097
1098                 op_ret = 0;
1099                 plist_for_each_entry_safe(this, next, head, list) {
1100                         if (match_futex (&this->key, &key2)) {
1101                                 if (this->pi_state || this->rt_waiter) {
1102                                         ret = -EINVAL;
1103                                         goto out_unlock;
1104                                 }
1105                                 wake_futex(this);
1106                                 if (++op_ret >= nr_wake2)
1107                                         break;
1108                         }
1109                 }
1110                 ret += op_ret;
1111         }
1112
1113 out_unlock:
1114         double_unlock_hb(hb1, hb2);
1115 out_put_keys:
1116         put_futex_key(&key2);
1117 out_put_key1:
1118         put_futex_key(&key1);
1119 out:
1120         return ret;
1121 }
1122
1123 /**
1124  * requeue_futex() - Requeue a futex_q from one hb to another
1125  * @q:          the futex_q to requeue
1126  * @hb1:        the source hash_bucket
1127  * @hb2:        the target hash_bucket
1128  * @key2:       the new key for the requeued futex_q
1129  */
1130 static inline
1131 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1132                    struct futex_hash_bucket *hb2, union futex_key *key2)
1133 {
1134
1135         /*
1136          * If key1 and key2 hash to the same bucket, no need to
1137          * requeue.
1138          */
1139         if (likely(&hb1->chain != &hb2->chain)) {
1140                 plist_del(&q->list, &hb1->chain);
1141                 plist_add(&q->list, &hb2->chain);
1142                 q->lock_ptr = &hb2->lock;
1143         }
1144         get_futex_key_refs(key2);
1145         q->key = *key2;
1146 }
1147
1148 /**
1149  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1150  * @q:          the futex_q
1151  * @key:        the key of the requeue target futex
1152  * @hb:         the hash_bucket of the requeue target futex
1153  *
1154  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1155  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1156  * to the requeue target futex so the waiter can detect the wakeup on the right
1157  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1158  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1159  * to protect access to the pi_state to fixup the owner later.  Must be called
1160  * with both q->lock_ptr and hb->lock held.
1161  */
1162 static inline
1163 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1164                            struct futex_hash_bucket *hb)
1165 {
1166         get_futex_key_refs(key);
1167         q->key = *key;
1168
1169         __unqueue_futex(q);
1170
1171         WARN_ON(!q->rt_waiter);
1172         q->rt_waiter = NULL;
1173
1174         q->lock_ptr = &hb->lock;
1175
1176         wake_up_state(q->task, TASK_NORMAL);
1177 }
1178
1179 /**
1180  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1181  * @pifutex:            the user address of the to futex
1182  * @hb1:                the from futex hash bucket, must be locked by the caller
1183  * @hb2:                the to futex hash bucket, must be locked by the caller
1184  * @key1:               the from futex key
1185  * @key2:               the to futex key
1186  * @ps:                 address to store the pi_state pointer
1187  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1188  *
1189  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1190  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1191  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1192  * hb1 and hb2 must be held by the caller.
1193  *
1194  * Returns:
1195  *  0 - failed to acquire the lock atomicly
1196  *  1 - acquired the lock
1197  * <0 - error
1198  */
1199 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1200                                  struct futex_hash_bucket *hb1,
1201                                  struct futex_hash_bucket *hb2,
1202                                  union futex_key *key1, union futex_key *key2,
1203                                  struct futex_pi_state **ps, int set_waiters)
1204 {
1205         struct futex_q *top_waiter = NULL;
1206         u32 curval;
1207         int ret;
1208
1209         if (get_futex_value_locked(&curval, pifutex))
1210                 return -EFAULT;
1211
1212         /*
1213          * Find the top_waiter and determine if there are additional waiters.
1214          * If the caller intends to requeue more than 1 waiter to pifutex,
1215          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1216          * as we have means to handle the possible fault.  If not, don't set
1217          * the bit unecessarily as it will force the subsequent unlock to enter
1218          * the kernel.
1219          */
1220         top_waiter = futex_top_waiter(hb1, key1);
1221
1222         /* There are no waiters, nothing for us to do. */
1223         if (!top_waiter)
1224                 return 0;
1225
1226         /* Ensure we requeue to the expected futex. */
1227         if (!match_futex(top_waiter->requeue_pi_key, key2))
1228                 return -EINVAL;
1229
1230         /*
1231          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1232          * the contended case or if set_waiters is 1.  The pi_state is returned
1233          * in ps in contended cases.
1234          */
1235         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1236                                    set_waiters);
1237         if (ret == 1)
1238                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1239
1240         return ret;
1241 }
1242
1243 /**
1244  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1245  * @uaddr1:     source futex user address
1246  * @flags:      futex flags (FLAGS_SHARED, etc.)
1247  * @uaddr2:     target futex user address
1248  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1249  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1250  * @cmpval:     @uaddr1 expected value (or %NULL)
1251  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1252  *              pi futex (pi to pi requeue is not supported)
1253  *
1254  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1255  * uaddr2 atomically on behalf of the top waiter.
1256  *
1257  * Returns:
1258  * >=0 - on success, the number of tasks requeued or woken
1259  *  <0 - on error
1260  */
1261 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1262                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1263                          u32 *cmpval, int requeue_pi)
1264 {
1265         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1266         int drop_count = 0, task_count = 0, ret;
1267         struct futex_pi_state *pi_state = NULL;
1268         struct futex_hash_bucket *hb1, *hb2;
1269         struct plist_head *head1;
1270         struct futex_q *this, *next;
1271         u32 curval2;
1272
1273         if (requeue_pi) {
1274                 /*
1275                  * requeue_pi requires a pi_state, try to allocate it now
1276                  * without any locks in case it fails.
1277                  */
1278                 if (refill_pi_state_cache())
1279                         return -ENOMEM;
1280                 /*
1281                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1282                  * + nr_requeue, since it acquires the rt_mutex prior to
1283                  * returning to userspace, so as to not leave the rt_mutex with
1284                  * waiters and no owner.  However, second and third wake-ups
1285                  * cannot be predicted as they involve race conditions with the
1286                  * first wake and a fault while looking up the pi_state.  Both
1287                  * pthread_cond_signal() and pthread_cond_broadcast() should
1288                  * use nr_wake=1.
1289                  */
1290                 if (nr_wake != 1)
1291                         return -EINVAL;
1292         }
1293
1294 retry:
1295         if (pi_state != NULL) {
1296                 /*
1297                  * We will have to lookup the pi_state again, so free this one
1298                  * to keep the accounting correct.
1299                  */
1300                 free_pi_state(pi_state);
1301                 pi_state = NULL;
1302         }
1303
1304         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1305         if (unlikely(ret != 0))
1306                 goto out;
1307         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1308                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1309         if (unlikely(ret != 0))
1310                 goto out_put_key1;
1311
1312         hb1 = hash_futex(&key1);
1313         hb2 = hash_futex(&key2);
1314
1315 retry_private:
1316         double_lock_hb(hb1, hb2);
1317
1318         if (likely(cmpval != NULL)) {
1319                 u32 curval;
1320
1321                 ret = get_futex_value_locked(&curval, uaddr1);
1322
1323                 if (unlikely(ret)) {
1324                         double_unlock_hb(hb1, hb2);
1325
1326                         ret = get_user(curval, uaddr1);
1327                         if (ret)
1328                                 goto out_put_keys;
1329
1330                         if (!(flags & FLAGS_SHARED))
1331                                 goto retry_private;
1332
1333                         put_futex_key(&key2);
1334                         put_futex_key(&key1);
1335                         goto retry;
1336                 }
1337                 if (curval != *cmpval) {
1338                         ret = -EAGAIN;
1339                         goto out_unlock;
1340                 }
1341         }
1342
1343         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1344                 /*
1345                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1346                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1347                  * bit.  We force this here where we are able to easily handle
1348                  * faults rather in the requeue loop below.
1349                  */
1350                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1351                                                  &key2, &pi_state, nr_requeue);
1352
1353                 /*
1354                  * At this point the top_waiter has either taken uaddr2 or is
1355                  * waiting on it.  If the former, then the pi_state will not
1356                  * exist yet, look it up one more time to ensure we have a
1357                  * reference to it.
1358                  */
1359                 if (ret == 1) {
1360                         WARN_ON(pi_state);
1361                         drop_count++;
1362                         task_count++;
1363                         ret = get_futex_value_locked(&curval2, uaddr2);
1364                         if (!ret)
1365                                 ret = lookup_pi_state(curval2, hb2, &key2,
1366                                                       &pi_state);
1367                 }
1368
1369                 switch (ret) {
1370                 case 0:
1371                         break;
1372                 case -EFAULT:
1373                         double_unlock_hb(hb1, hb2);
1374                         put_futex_key(&key2);
1375                         put_futex_key(&key1);
1376                         ret = fault_in_user_writeable(uaddr2);
1377                         if (!ret)
1378                                 goto retry;
1379                         goto out;
1380                 case -EAGAIN:
1381                         /* The owner was exiting, try again. */
1382                         double_unlock_hb(hb1, hb2);
1383                         put_futex_key(&key2);
1384                         put_futex_key(&key1);
1385                         cond_resched();
1386                         goto retry;
1387                 default:
1388                         goto out_unlock;
1389                 }
1390         }
1391
1392         head1 = &hb1->chain;
1393         plist_for_each_entry_safe(this, next, head1, list) {
1394                 if (task_count - nr_wake >= nr_requeue)
1395                         break;
1396
1397                 if (!match_futex(&this->key, &key1))
1398                         continue;
1399
1400                 /*
1401                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1402                  * be paired with each other and no other futex ops.
1403                  *
1404                  * We should never be requeueing a futex_q with a pi_state,
1405                  * which is awaiting a futex_unlock_pi().
1406                  */
1407                 if ((requeue_pi && !this->rt_waiter) ||
1408                     (!requeue_pi && this->rt_waiter) ||
1409                     this->pi_state) {
1410                         ret = -EINVAL;
1411                         break;
1412                 }
1413
1414                 /*
1415                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1416                  * lock, we already woke the top_waiter.  If not, it will be
1417                  * woken by futex_unlock_pi().
1418                  */
1419                 if (++task_count <= nr_wake && !requeue_pi) {
1420                         wake_futex(this);
1421                         continue;
1422                 }
1423
1424                 /* Ensure we requeue to the expected futex for requeue_pi. */
1425                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1426                         ret = -EINVAL;
1427                         break;
1428                 }
1429
1430                 /*
1431                  * Requeue nr_requeue waiters and possibly one more in the case
1432                  * of requeue_pi if we couldn't acquire the lock atomically.
1433                  */
1434                 if (requeue_pi) {
1435                         /* Prepare the waiter to take the rt_mutex. */
1436                         atomic_inc(&pi_state->refcount);
1437                         this->pi_state = pi_state;
1438                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1439                                                         this->rt_waiter,
1440                                                         this->task, 1);
1441                         if (ret == 1) {
1442                                 /* We got the lock. */
1443                                 requeue_pi_wake_futex(this, &key2, hb2);
1444                                 drop_count++;
1445                                 continue;
1446                         } else if (ret) {
1447                                 /* -EDEADLK */
1448                                 this->pi_state = NULL;
1449                                 free_pi_state(pi_state);
1450                                 goto out_unlock;
1451                         }
1452                 }
1453                 requeue_futex(this, hb1, hb2, &key2);
1454                 drop_count++;
1455         }
1456
1457 out_unlock:
1458         double_unlock_hb(hb1, hb2);
1459
1460         /*
1461          * drop_futex_key_refs() must be called outside the spinlocks. During
1462          * the requeue we moved futex_q's from the hash bucket at key1 to the
1463          * one at key2 and updated their key pointer.  We no longer need to
1464          * hold the references to key1.
1465          */
1466         while (--drop_count >= 0)
1467                 drop_futex_key_refs(&key1);
1468
1469 out_put_keys:
1470         put_futex_key(&key2);
1471 out_put_key1:
1472         put_futex_key(&key1);
1473 out:
1474         if (pi_state != NULL)
1475                 free_pi_state(pi_state);
1476         return ret ? ret : task_count;
1477 }
1478
1479 /* The key must be already stored in q->key. */
1480 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1481         __acquires(&hb->lock)
1482 {
1483         struct futex_hash_bucket *hb;
1484
1485         hb = hash_futex(&q->key);
1486         q->lock_ptr = &hb->lock;
1487
1488         spin_lock(&hb->lock);
1489         return hb;
1490 }
1491
1492 static inline void
1493 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1494         __releases(&hb->lock)
1495 {
1496         spin_unlock(&hb->lock);
1497 }
1498
1499 /**
1500  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1501  * @q:  The futex_q to enqueue
1502  * @hb: The destination hash bucket
1503  *
1504  * The hb->lock must be held by the caller, and is released here. A call to
1505  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1506  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1507  * or nothing if the unqueue is done as part of the wake process and the unqueue
1508  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1509  * an example).
1510  */
1511 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1512         __releases(&hb->lock)
1513 {
1514         int prio;
1515
1516         /*
1517          * The priority used to register this element is
1518          * - either the real thread-priority for the real-time threads
1519          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1520          * - or MAX_RT_PRIO for non-RT threads.
1521          * Thus, all RT-threads are woken first in priority order, and
1522          * the others are woken last, in FIFO order.
1523          */
1524         prio = min(current->normal_prio, MAX_RT_PRIO);
1525
1526         plist_node_init(&q->list, prio);
1527         plist_add(&q->list, &hb->chain);
1528         q->task = current;
1529         spin_unlock(&hb->lock);
1530 }
1531
1532 /**
1533  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1534  * @q:  The futex_q to unqueue
1535  *
1536  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1537  * be paired with exactly one earlier call to queue_me().
1538  *
1539  * Returns:
1540  *   1 - if the futex_q was still queued (and we removed unqueued it)
1541  *   0 - if the futex_q was already removed by the waking thread
1542  */
1543 static int unqueue_me(struct futex_q *q)
1544 {
1545         spinlock_t *lock_ptr;
1546         int ret = 0;
1547
1548         /* In the common case we don't take the spinlock, which is nice. */
1549 retry:
1550         lock_ptr = q->lock_ptr;
1551         barrier();
1552         if (lock_ptr != NULL) {
1553                 spin_lock(lock_ptr);
1554                 /*
1555                  * q->lock_ptr can change between reading it and
1556                  * spin_lock(), causing us to take the wrong lock.  This
1557                  * corrects the race condition.
1558                  *
1559                  * Reasoning goes like this: if we have the wrong lock,
1560                  * q->lock_ptr must have changed (maybe several times)
1561                  * between reading it and the spin_lock().  It can
1562                  * change again after the spin_lock() but only if it was
1563                  * already changed before the spin_lock().  It cannot,
1564                  * however, change back to the original value.  Therefore
1565                  * we can detect whether we acquired the correct lock.
1566                  */
1567                 if (unlikely(lock_ptr != q->lock_ptr)) {
1568                         spin_unlock(lock_ptr);
1569                         goto retry;
1570                 }
1571                 __unqueue_futex(q);
1572
1573                 BUG_ON(q->pi_state);
1574
1575                 spin_unlock(lock_ptr);
1576                 ret = 1;
1577         }
1578
1579         drop_futex_key_refs(&q->key);
1580         return ret;
1581 }
1582
1583 /*
1584  * PI futexes can not be requeued and must remove themself from the
1585  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1586  * and dropped here.
1587  */
1588 static void unqueue_me_pi(struct futex_q *q)
1589         __releases(q->lock_ptr)
1590 {
1591         __unqueue_futex(q);
1592
1593         BUG_ON(!q->pi_state);
1594         free_pi_state(q->pi_state);
1595         q->pi_state = NULL;
1596
1597         spin_unlock(q->lock_ptr);
1598 }
1599
1600 /*
1601  * Fixup the pi_state owner with the new owner.
1602  *
1603  * Must be called with hash bucket lock held and mm->sem held for non
1604  * private futexes.
1605  */
1606 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1607                                 struct task_struct *newowner)
1608 {
1609         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1610         struct futex_pi_state *pi_state = q->pi_state;
1611         struct task_struct *oldowner = pi_state->owner;
1612         u32 uval, uninitialized_var(curval), newval;
1613         int ret;
1614
1615         /* Owner died? */
1616         if (!pi_state->owner)
1617                 newtid |= FUTEX_OWNER_DIED;
1618
1619         /*
1620          * We are here either because we stole the rtmutex from the
1621          * previous highest priority waiter or we are the highest priority
1622          * waiter but failed to get the rtmutex the first time.
1623          * We have to replace the newowner TID in the user space variable.
1624          * This must be atomic as we have to preserve the owner died bit here.
1625          *
1626          * Note: We write the user space value _before_ changing the pi_state
1627          * because we can fault here. Imagine swapped out pages or a fork
1628          * that marked all the anonymous memory readonly for cow.
1629          *
1630          * Modifying pi_state _before_ the user space value would
1631          * leave the pi_state in an inconsistent state when we fault
1632          * here, because we need to drop the hash bucket lock to
1633          * handle the fault. This might be observed in the PID check
1634          * in lookup_pi_state.
1635          */
1636 retry:
1637         if (get_futex_value_locked(&uval, uaddr))
1638                 goto handle_fault;
1639
1640         while (1) {
1641                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1642
1643                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1644                         goto handle_fault;
1645                 if (curval == uval)
1646                         break;
1647                 uval = curval;
1648         }
1649
1650         /*
1651          * We fixed up user space. Now we need to fix the pi_state
1652          * itself.
1653          */
1654         if (pi_state->owner != NULL) {
1655                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1656                 WARN_ON(list_empty(&pi_state->list));
1657                 list_del_init(&pi_state->list);
1658                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1659         }
1660
1661         pi_state->owner = newowner;
1662
1663         raw_spin_lock_irq(&newowner->pi_lock);
1664         WARN_ON(!list_empty(&pi_state->list));
1665         list_add(&pi_state->list, &newowner->pi_state_list);
1666         raw_spin_unlock_irq(&newowner->pi_lock);
1667         return 0;
1668
1669         /*
1670          * To handle the page fault we need to drop the hash bucket
1671          * lock here. That gives the other task (either the highest priority
1672          * waiter itself or the task which stole the rtmutex) the
1673          * chance to try the fixup of the pi_state. So once we are
1674          * back from handling the fault we need to check the pi_state
1675          * after reacquiring the hash bucket lock and before trying to
1676          * do another fixup. When the fixup has been done already we
1677          * simply return.
1678          */
1679 handle_fault:
1680         spin_unlock(q->lock_ptr);
1681
1682         ret = fault_in_user_writeable(uaddr);
1683
1684         spin_lock(q->lock_ptr);
1685
1686         /*
1687          * Check if someone else fixed it for us:
1688          */
1689         if (pi_state->owner != oldowner)
1690                 return 0;
1691
1692         if (ret)
1693                 return ret;
1694
1695         goto retry;
1696 }
1697
1698 static long futex_wait_restart(struct restart_block *restart);
1699
1700 /**
1701  * fixup_owner() - Post lock pi_state and corner case management
1702  * @uaddr:      user address of the futex
1703  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1704  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1705  *
1706  * After attempting to lock an rt_mutex, this function is called to cleanup
1707  * the pi_state owner as well as handle race conditions that may allow us to
1708  * acquire the lock. Must be called with the hb lock held.
1709  *
1710  * Returns:
1711  *  1 - success, lock taken
1712  *  0 - success, lock not taken
1713  * <0 - on error (-EFAULT)
1714  */
1715 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1716 {
1717         struct task_struct *owner;
1718         int ret = 0;
1719
1720         if (locked) {
1721                 /*
1722                  * Got the lock. We might not be the anticipated owner if we
1723                  * did a lock-steal - fix up the PI-state in that case:
1724                  */
1725                 if (q->pi_state->owner != current)
1726                         ret = fixup_pi_state_owner(uaddr, q, current);
1727                 goto out;
1728         }
1729
1730         /*
1731          * Catch the rare case, where the lock was released when we were on the
1732          * way back before we locked the hash bucket.
1733          */
1734         if (q->pi_state->owner == current) {
1735                 /*
1736                  * Try to get the rt_mutex now. This might fail as some other
1737                  * task acquired the rt_mutex after we removed ourself from the
1738                  * rt_mutex waiters list.
1739                  */
1740                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1741                         locked = 1;
1742                         goto out;
1743                 }
1744
1745                 /*
1746                  * pi_state is incorrect, some other task did a lock steal and
1747                  * we returned due to timeout or signal without taking the
1748                  * rt_mutex. Too late.
1749                  */
1750                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1751                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1752                 if (!owner)
1753                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1754                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1755                 ret = fixup_pi_state_owner(uaddr, q, owner);
1756                 goto out;
1757         }
1758
1759         /*
1760          * Paranoia check. If we did not take the lock, then we should not be
1761          * the owner of the rt_mutex.
1762          */
1763         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1764                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1765                                 "pi-state %p\n", ret,
1766                                 q->pi_state->pi_mutex.owner,
1767                                 q->pi_state->owner);
1768
1769 out:
1770         return ret ? ret : locked;
1771 }
1772
1773 /**
1774  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1775  * @hb:         the futex hash bucket, must be locked by the caller
1776  * @q:          the futex_q to queue up on
1777  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1778  */
1779 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1780                                 struct hrtimer_sleeper *timeout)
1781 {
1782         /*
1783          * The task state is guaranteed to be set before another task can
1784          * wake it. set_current_state() is implemented using set_mb() and
1785          * queue_me() calls spin_unlock() upon completion, both serializing
1786          * access to the hash list and forcing another memory barrier.
1787          */
1788         set_current_state(TASK_INTERRUPTIBLE);
1789         queue_me(q, hb);
1790
1791         /* Arm the timer */
1792         if (timeout) {
1793                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1794                 if (!hrtimer_active(&timeout->timer))
1795                         timeout->task = NULL;
1796         }
1797
1798         /*
1799          * If we have been removed from the hash list, then another task
1800          * has tried to wake us, and we can skip the call to schedule().
1801          */
1802         if (likely(!plist_node_empty(&q->list))) {
1803                 /*
1804                  * If the timer has already expired, current will already be
1805                  * flagged for rescheduling. Only call schedule if there
1806                  * is no timeout, or if it has yet to expire.
1807                  */
1808                 if (!timeout || timeout->task)
1809                         schedule();
1810         }
1811         __set_current_state(TASK_RUNNING);
1812 }
1813
1814 /**
1815  * futex_wait_setup() - Prepare to wait on a futex
1816  * @uaddr:      the futex userspace address
1817  * @val:        the expected value
1818  * @flags:      futex flags (FLAGS_SHARED, etc.)
1819  * @q:          the associated futex_q
1820  * @hb:         storage for hash_bucket pointer to be returned to caller
1821  *
1822  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1823  * compare it with the expected value.  Handle atomic faults internally.
1824  * Return with the hb lock held and a q.key reference on success, and unlocked
1825  * with no q.key reference on failure.
1826  *
1827  * Returns:
1828  *  0 - uaddr contains val and hb has been locked
1829  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1830  */
1831 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1832                            struct futex_q *q, struct futex_hash_bucket **hb)
1833 {
1834         u32 uval;
1835         int ret;
1836
1837         /*
1838          * Access the page AFTER the hash-bucket is locked.
1839          * Order is important:
1840          *
1841          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1842          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1843          *
1844          * The basic logical guarantee of a futex is that it blocks ONLY
1845          * if cond(var) is known to be true at the time of blocking, for
1846          * any cond.  If we locked the hash-bucket after testing *uaddr, that
1847          * would open a race condition where we could block indefinitely with
1848          * cond(var) false, which would violate the guarantee.
1849          *
1850          * On the other hand, we insert q and release the hash-bucket only
1851          * after testing *uaddr.  This guarantees that futex_wait() will NOT
1852          * absorb a wakeup if *uaddr does not match the desired values
1853          * while the syscall executes.
1854          */
1855 retry:
1856         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1857         if (unlikely(ret != 0))
1858                 return ret;
1859
1860 retry_private:
1861         *hb = queue_lock(q);
1862
1863         ret = get_futex_value_locked(&uval, uaddr);
1864
1865         if (ret) {
1866                 queue_unlock(q, *hb);
1867
1868                 ret = get_user(uval, uaddr);
1869                 if (ret)
1870                         goto out;
1871
1872                 if (!(flags & FLAGS_SHARED))
1873                         goto retry_private;
1874
1875                 put_futex_key(&q->key);
1876                 goto retry;
1877         }
1878
1879         if (uval != val) {
1880                 queue_unlock(q, *hb);
1881                 ret = -EWOULDBLOCK;
1882         }
1883
1884 out:
1885         if (ret)
1886                 put_futex_key(&q->key);
1887         return ret;
1888 }
1889
1890 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1891                       ktime_t *abs_time, u32 bitset)
1892 {
1893         struct hrtimer_sleeper timeout, *to = NULL;
1894         struct restart_block *restart;
1895         struct futex_hash_bucket *hb;
1896         struct futex_q q = futex_q_init;
1897         int ret;
1898
1899         if (!bitset)
1900                 return -EINVAL;
1901         q.bitset = bitset;
1902
1903         if (abs_time) {
1904                 to = &timeout;
1905
1906                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1907                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
1908                                       HRTIMER_MODE_ABS);
1909                 hrtimer_init_sleeper(to, current);
1910                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1911                                              current->timer_slack_ns);
1912         }
1913
1914 retry:
1915         /*
1916          * Prepare to wait on uaddr. On success, holds hb lock and increments
1917          * q.key refs.
1918          */
1919         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1920         if (ret)
1921                 goto out;
1922
1923         /* queue_me and wait for wakeup, timeout, or a signal. */
1924         futex_wait_queue_me(hb, &q, to);
1925
1926         /* If we were woken (and unqueued), we succeeded, whatever. */
1927         ret = 0;
1928         /* unqueue_me() drops q.key ref */
1929         if (!unqueue_me(&q))
1930                 goto out;
1931         ret = -ETIMEDOUT;
1932         if (to && !to->task)
1933                 goto out;
1934
1935         /*
1936          * We expect signal_pending(current), but we might be the
1937          * victim of a spurious wakeup as well.
1938          */
1939         if (!signal_pending(current))
1940                 goto retry;
1941
1942         ret = -ERESTARTSYS;
1943         if (!abs_time)
1944                 goto out;
1945
1946         restart = &current_thread_info()->restart_block;
1947         restart->fn = futex_wait_restart;
1948         restart->futex.uaddr = uaddr;
1949         restart->futex.val = val;
1950         restart->futex.time = abs_time->tv64;
1951         restart->futex.bitset = bitset;
1952         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1953
1954         ret = -ERESTART_RESTARTBLOCK;
1955
1956 out:
1957         if (to) {
1958                 hrtimer_cancel(&to->timer);
1959                 destroy_hrtimer_on_stack(&to->timer);
1960         }
1961         return ret;
1962 }
1963
1964
1965 static long futex_wait_restart(struct restart_block *restart)
1966 {
1967         u32 __user *uaddr = restart->futex.uaddr;
1968         ktime_t t, *tp = NULL;
1969
1970         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1971                 t.tv64 = restart->futex.time;
1972                 tp = &t;
1973         }
1974         restart->fn = do_no_restart_syscall;
1975
1976         return (long)futex_wait(uaddr, restart->futex.flags,
1977                                 restart->futex.val, tp, restart->futex.bitset);
1978 }
1979
1980
1981 /*
1982  * Userspace tried a 0 -> TID atomic transition of the futex value
1983  * and failed. The kernel side here does the whole locking operation:
1984  * if there are waiters then it will block, it does PI, etc. (Due to
1985  * races the kernel might see a 0 value of the futex too.)
1986  */
1987 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1988                          ktime_t *time, int trylock)
1989 {
1990         struct hrtimer_sleeper timeout, *to = NULL;
1991         struct futex_hash_bucket *hb;
1992         struct futex_q q = futex_q_init;
1993         int res, ret;
1994
1995         if (refill_pi_state_cache())
1996                 return -ENOMEM;
1997
1998         if (time) {
1999                 to = &timeout;
2000                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2001                                       HRTIMER_MODE_ABS);
2002                 hrtimer_init_sleeper(to, current);
2003                 hrtimer_set_expires(&to->timer, *time);
2004         }
2005
2006 retry:
2007         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2008         if (unlikely(ret != 0))
2009                 goto out;
2010
2011 retry_private:
2012         hb = queue_lock(&q);
2013
2014         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2015         if (unlikely(ret)) {
2016                 switch (ret) {
2017                 case 1:
2018                         /* We got the lock. */
2019                         ret = 0;
2020                         goto out_unlock_put_key;
2021                 case -EFAULT:
2022                         goto uaddr_faulted;
2023                 case -EAGAIN:
2024                         /*
2025                          * Task is exiting and we just wait for the
2026                          * exit to complete.
2027                          */
2028                         queue_unlock(&q, hb);
2029                         put_futex_key(&q.key);
2030                         cond_resched();
2031                         goto retry;
2032                 default:
2033                         goto out_unlock_put_key;
2034                 }
2035         }
2036
2037         /*
2038          * Only actually queue now that the atomic ops are done:
2039          */
2040         queue_me(&q, hb);
2041
2042         WARN_ON(!q.pi_state);
2043         /*
2044          * Block on the PI mutex:
2045          */
2046         if (!trylock)
2047                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2048         else {
2049                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2050                 /* Fixup the trylock return value: */
2051                 ret = ret ? 0 : -EWOULDBLOCK;
2052         }
2053
2054         spin_lock(q.lock_ptr);
2055         /*
2056          * Fixup the pi_state owner and possibly acquire the lock if we
2057          * haven't already.
2058          */
2059         res = fixup_owner(uaddr, &q, !ret);
2060         /*
2061          * If fixup_owner() returned an error, proprogate that.  If it acquired
2062          * the lock, clear our -ETIMEDOUT or -EINTR.
2063          */
2064         if (res)
2065                 ret = (res < 0) ? res : 0;
2066
2067         /*
2068          * If fixup_owner() faulted and was unable to handle the fault, unlock
2069          * it and return the fault to userspace.
2070          */
2071         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2072                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2073
2074         /* Unqueue and drop the lock */
2075         unqueue_me_pi(&q);
2076
2077         goto out_put_key;
2078
2079 out_unlock_put_key:
2080         queue_unlock(&q, hb);
2081
2082 out_put_key:
2083         put_futex_key(&q.key);
2084 out:
2085         if (to)
2086                 destroy_hrtimer_on_stack(&to->timer);
2087         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2088
2089 uaddr_faulted:
2090         queue_unlock(&q, hb);
2091
2092         ret = fault_in_user_writeable(uaddr);
2093         if (ret)
2094                 goto out_put_key;
2095
2096         if (!(flags & FLAGS_SHARED))
2097                 goto retry_private;
2098
2099         put_futex_key(&q.key);
2100         goto retry;
2101 }
2102
2103 /*
2104  * Userspace attempted a TID -> 0 atomic transition, and failed.
2105  * This is the in-kernel slowpath: we look up the PI state (if any),
2106  * and do the rt-mutex unlock.
2107  */
2108 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2109 {
2110         struct futex_hash_bucket *hb;
2111         struct futex_q *this, *next;
2112         struct plist_head *head;
2113         union futex_key key = FUTEX_KEY_INIT;
2114         u32 uval, vpid = task_pid_vnr(current);
2115         int ret;
2116
2117 retry:
2118         if (get_user(uval, uaddr))
2119                 return -EFAULT;
2120         /*
2121          * We release only a lock we actually own:
2122          */
2123         if ((uval & FUTEX_TID_MASK) != vpid)
2124                 return -EPERM;
2125
2126         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2127         if (unlikely(ret != 0))
2128                 goto out;
2129
2130         hb = hash_futex(&key);
2131         spin_lock(&hb->lock);
2132
2133         /*
2134          * To avoid races, try to do the TID -> 0 atomic transition
2135          * again. If it succeeds then we can return without waking
2136          * anyone else up:
2137          */
2138         if (!(uval & FUTEX_OWNER_DIED) &&
2139             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2140                 goto pi_faulted;
2141         /*
2142          * Rare case: we managed to release the lock atomically,
2143          * no need to wake anyone else up:
2144          */
2145         if (unlikely(uval == vpid))
2146                 goto out_unlock;
2147
2148         /*
2149          * Ok, other tasks may need to be woken up - check waiters
2150          * and do the wakeup if necessary:
2151          */
2152         head = &hb->chain;
2153
2154         plist_for_each_entry_safe(this, next, head, list) {
2155                 if (!match_futex (&this->key, &key))
2156                         continue;
2157                 ret = wake_futex_pi(uaddr, uval, this);
2158                 /*
2159                  * The atomic access to the futex value
2160                  * generated a pagefault, so retry the
2161                  * user-access and the wakeup:
2162                  */
2163                 if (ret == -EFAULT)
2164                         goto pi_faulted;
2165                 goto out_unlock;
2166         }
2167         /*
2168          * No waiters - kernel unlocks the futex:
2169          */
2170         if (!(uval & FUTEX_OWNER_DIED)) {
2171                 ret = unlock_futex_pi(uaddr, uval);
2172                 if (ret == -EFAULT)
2173                         goto pi_faulted;
2174         }
2175
2176 out_unlock:
2177         spin_unlock(&hb->lock);
2178         put_futex_key(&key);
2179
2180 out:
2181         return ret;
2182
2183 pi_faulted:
2184         spin_unlock(&hb->lock);
2185         put_futex_key(&key);
2186
2187         ret = fault_in_user_writeable(uaddr);
2188         if (!ret)
2189                 goto retry;
2190
2191         return ret;
2192 }
2193
2194 /**
2195  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2196  * @hb:         the hash_bucket futex_q was original enqueued on
2197  * @q:          the futex_q woken while waiting to be requeued
2198  * @key2:       the futex_key of the requeue target futex
2199  * @timeout:    the timeout associated with the wait (NULL if none)
2200  *
2201  * Detect if the task was woken on the initial futex as opposed to the requeue
2202  * target futex.  If so, determine if it was a timeout or a signal that caused
2203  * the wakeup and return the appropriate error code to the caller.  Must be
2204  * called with the hb lock held.
2205  *
2206  * Returns
2207  *  0 - no early wakeup detected
2208  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2209  */
2210 static inline
2211 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2212                                    struct futex_q *q, union futex_key *key2,
2213                                    struct hrtimer_sleeper *timeout)
2214 {
2215         int ret = 0;
2216
2217         /*
2218          * With the hb lock held, we avoid races while we process the wakeup.
2219          * We only need to hold hb (and not hb2) to ensure atomicity as the
2220          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2221          * It can't be requeued from uaddr2 to something else since we don't
2222          * support a PI aware source futex for requeue.
2223          */
2224         if (!match_futex(&q->key, key2)) {
2225                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2226                 /*
2227                  * We were woken prior to requeue by a timeout or a signal.
2228                  * Unqueue the futex_q and determine which it was.
2229                  */
2230                 plist_del(&q->list, &hb->chain);
2231
2232                 /* Handle spurious wakeups gracefully */
2233                 ret = -EWOULDBLOCK;
2234                 if (timeout && !timeout->task)
2235                         ret = -ETIMEDOUT;
2236                 else if (signal_pending(current))
2237                         ret = -ERESTARTNOINTR;
2238         }
2239         return ret;
2240 }
2241
2242 /**
2243  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2244  * @uaddr:      the futex we initially wait on (non-pi)
2245  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2246  *              the same type, no requeueing from private to shared, etc.
2247  * @val:        the expected value of uaddr
2248  * @abs_time:   absolute timeout
2249  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2250  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2251  * @uaddr2:     the pi futex we will take prior to returning to user-space
2252  *
2253  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2254  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2255  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2256  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2257  * without one, the pi logic would not know which task to boost/deboost, if
2258  * there was a need to.
2259  *
2260  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2261  * via the following:
2262  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2263  * 2) wakeup on uaddr2 after a requeue
2264  * 3) signal
2265  * 4) timeout
2266  *
2267  * If 3, cleanup and return -ERESTARTNOINTR.
2268  *
2269  * If 2, we may then block on trying to take the rt_mutex and return via:
2270  * 5) successful lock
2271  * 6) signal
2272  * 7) timeout
2273  * 8) other lock acquisition failure
2274  *
2275  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2276  *
2277  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2278  *
2279  * Returns:
2280  *  0 - On success
2281  * <0 - On error
2282  */
2283 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2284                                  u32 val, ktime_t *abs_time, u32 bitset,
2285                                  u32 __user *uaddr2)
2286 {
2287         struct hrtimer_sleeper timeout, *to = NULL;
2288         struct rt_mutex_waiter rt_waiter;
2289         struct rt_mutex *pi_mutex = NULL;
2290         struct futex_hash_bucket *hb;
2291         union futex_key key2 = FUTEX_KEY_INIT;
2292         struct futex_q q = futex_q_init;
2293         int res, ret;
2294
2295         if (uaddr == uaddr2)
2296                 return -EINVAL;
2297
2298         if (!bitset)
2299                 return -EINVAL;
2300
2301         if (abs_time) {
2302                 to = &timeout;
2303                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2304                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2305                                       HRTIMER_MODE_ABS);
2306                 hrtimer_init_sleeper(to, current);
2307                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2308                                              current->timer_slack_ns);
2309         }
2310
2311         /*
2312          * The waiter is allocated on our stack, manipulated by the requeue
2313          * code while we sleep on uaddr.
2314          */
2315         debug_rt_mutex_init_waiter(&rt_waiter);
2316         rt_waiter.task = NULL;
2317
2318         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2319         if (unlikely(ret != 0))
2320                 goto out;
2321
2322         q.bitset = bitset;
2323         q.rt_waiter = &rt_waiter;
2324         q.requeue_pi_key = &key2;
2325
2326         /*
2327          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2328          * count.
2329          */
2330         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2331         if (ret)
2332                 goto out_key2;
2333
2334         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2335         futex_wait_queue_me(hb, &q, to);
2336
2337         spin_lock(&hb->lock);
2338         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2339         spin_unlock(&hb->lock);
2340         if (ret)
2341                 goto out_put_keys;
2342
2343         /*
2344          * In order for us to be here, we know our q.key == key2, and since
2345          * we took the hb->lock above, we also know that futex_requeue() has
2346          * completed and we no longer have to concern ourselves with a wakeup
2347          * race with the atomic proxy lock acquisition by the requeue code. The
2348          * futex_requeue dropped our key1 reference and incremented our key2
2349          * reference count.
2350          */
2351
2352         /* Check if the requeue code acquired the second futex for us. */
2353         if (!q.rt_waiter) {
2354                 /*
2355                  * Got the lock. We might not be the anticipated owner if we
2356                  * did a lock-steal - fix up the PI-state in that case.
2357                  */
2358                 if (q.pi_state && (q.pi_state->owner != current)) {
2359                         spin_lock(q.lock_ptr);
2360                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2361                         spin_unlock(q.lock_ptr);
2362                 }
2363         } else {
2364                 /*
2365                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2366                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2367                  * the pi_state.
2368                  */
2369                 WARN_ON(!q.pi_state);
2370                 pi_mutex = &q.pi_state->pi_mutex;
2371                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2372                 debug_rt_mutex_free_waiter(&rt_waiter);
2373
2374                 spin_lock(q.lock_ptr);
2375                 /*
2376                  * Fixup the pi_state owner and possibly acquire the lock if we
2377                  * haven't already.
2378                  */
2379                 res = fixup_owner(uaddr2, &q, !ret);
2380                 /*
2381                  * If fixup_owner() returned an error, proprogate that.  If it
2382                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2383                  */
2384                 if (res)
2385                         ret = (res < 0) ? res : 0;
2386
2387                 /* Unqueue and drop the lock. */
2388                 unqueue_me_pi(&q);
2389         }
2390
2391         /*
2392          * If fixup_pi_state_owner() faulted and was unable to handle the
2393          * fault, unlock the rt_mutex and return the fault to userspace.
2394          */
2395         if (ret == -EFAULT) {
2396                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2397                         rt_mutex_unlock(pi_mutex);
2398         } else if (ret == -EINTR) {
2399                 /*
2400                  * We've already been requeued, but cannot restart by calling
2401                  * futex_lock_pi() directly. We could restart this syscall, but
2402                  * it would detect that the user space "val" changed and return
2403                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2404                  * -EWOULDBLOCK directly.
2405                  */
2406                 ret = -EWOULDBLOCK;
2407         }
2408
2409 out_put_keys:
2410         put_futex_key(&q.key);
2411 out_key2:
2412         put_futex_key(&key2);
2413
2414 out:
2415         if (to) {
2416                 hrtimer_cancel(&to->timer);
2417                 destroy_hrtimer_on_stack(&to->timer);
2418         }
2419         return ret;
2420 }
2421
2422 /*
2423  * Support for robust futexes: the kernel cleans up held futexes at
2424  * thread exit time.
2425  *
2426  * Implementation: user-space maintains a per-thread list of locks it
2427  * is holding. Upon do_exit(), the kernel carefully walks this list,
2428  * and marks all locks that are owned by this thread with the
2429  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2430  * always manipulated with the lock held, so the list is private and
2431  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2432  * field, to allow the kernel to clean up if the thread dies after
2433  * acquiring the lock, but just before it could have added itself to
2434  * the list. There can only be one such pending lock.
2435  */
2436
2437 /**
2438  * sys_set_robust_list() - Set the robust-futex list head of a task
2439  * @head:       pointer to the list-head
2440  * @len:        length of the list-head, as userspace expects
2441  */
2442 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2443                 size_t, len)
2444 {
2445         if (!futex_cmpxchg_enabled)
2446                 return -ENOSYS;
2447         /*
2448          * The kernel knows only one size for now:
2449          */
2450         if (unlikely(len != sizeof(*head)))
2451                 return -EINVAL;
2452
2453         current->robust_list = head;
2454
2455         return 0;
2456 }
2457
2458 /**
2459  * sys_get_robust_list() - Get the robust-futex list head of a task
2460  * @pid:        pid of the process [zero for current task]
2461  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2462  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2463  */
2464 SYSCALL_DEFINE3(get_robust_list, int, pid,
2465                 struct robust_list_head __user * __user *, head_ptr,
2466                 size_t __user *, len_ptr)
2467 {
2468         struct robust_list_head __user *head;
2469         unsigned long ret;
2470         struct task_struct *p;
2471
2472         if (!futex_cmpxchg_enabled)
2473                 return -ENOSYS;
2474
2475         rcu_read_lock();
2476
2477         ret = -ESRCH;
2478         if (!pid)
2479                 p = current;
2480         else {
2481                 p = find_task_by_vpid(pid);
2482                 if (!p)
2483                         goto err_unlock;
2484         }
2485
2486         ret = -EPERM;
2487         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2488                 goto err_unlock;
2489
2490         head = p->robust_list;
2491         rcu_read_unlock();
2492
2493         if (put_user(sizeof(*head), len_ptr))
2494                 return -EFAULT;
2495         return put_user(head, head_ptr);
2496
2497 err_unlock:
2498         rcu_read_unlock();
2499
2500         return ret;
2501 }
2502
2503 /*
2504  * Process a futex-list entry, check whether it's owned by the
2505  * dying task, and do notification if so:
2506  */
2507 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2508 {
2509         u32 uval, uninitialized_var(nval), mval;
2510
2511 retry:
2512         if (get_user(uval, uaddr))
2513                 return -1;
2514
2515         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2516                 /*
2517                  * Ok, this dying thread is truly holding a futex
2518                  * of interest. Set the OWNER_DIED bit atomically
2519                  * via cmpxchg, and if the value had FUTEX_WAITERS
2520                  * set, wake up a waiter (if any). (We have to do a
2521                  * futex_wake() even if OWNER_DIED is already set -
2522                  * to handle the rare but possible case of recursive
2523                  * thread-death.) The rest of the cleanup is done in
2524                  * userspace.
2525                  */
2526                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2527                 /*
2528                  * We are not holding a lock here, but we want to have
2529                  * the pagefault_disable/enable() protection because
2530                  * we want to handle the fault gracefully. If the
2531                  * access fails we try to fault in the futex with R/W
2532                  * verification via get_user_pages. get_user() above
2533                  * does not guarantee R/W access. If that fails we
2534                  * give up and leave the futex locked.
2535                  */
2536                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2537                         if (fault_in_user_writeable(uaddr))
2538                                 return -1;
2539                         goto retry;
2540                 }
2541                 if (nval != uval)
2542                         goto retry;
2543
2544                 /*
2545                  * Wake robust non-PI futexes here. The wakeup of
2546                  * PI futexes happens in exit_pi_state():
2547                  */
2548                 if (!pi && (uval & FUTEX_WAITERS))
2549                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2550         }
2551         return 0;
2552 }
2553
2554 /*
2555  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2556  */
2557 static inline int fetch_robust_entry(struct robust_list __user **entry,
2558                                      struct robust_list __user * __user *head,
2559                                      unsigned int *pi)
2560 {
2561         unsigned long uentry;
2562
2563         if (get_user(uentry, (unsigned long __user *)head))
2564                 return -EFAULT;
2565
2566         *entry = (void __user *)(uentry & ~1UL);
2567         *pi = uentry & 1;
2568
2569         return 0;
2570 }
2571
2572 /*
2573  * Walk curr->robust_list (very carefully, it's a userspace list!)
2574  * and mark any locks found there dead, and notify any waiters.
2575  *
2576  * We silently return on any sign of list-walking problem.
2577  */
2578 void exit_robust_list(struct task_struct *curr)
2579 {
2580         struct robust_list_head __user *head = curr->robust_list;
2581         struct robust_list __user *entry, *next_entry, *pending;
2582         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2583         unsigned int uninitialized_var(next_pi);
2584         unsigned long futex_offset;
2585         int rc;
2586
2587         if (!futex_cmpxchg_enabled)
2588                 return;
2589
2590         /*
2591          * Fetch the list head (which was registered earlier, via
2592          * sys_set_robust_list()):
2593          */
2594         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2595                 return;
2596         /*
2597          * Fetch the relative futex offset:
2598          */
2599         if (get_user(futex_offset, &head->futex_offset))
2600                 return;
2601         /*
2602          * Fetch any possibly pending lock-add first, and handle it
2603          * if it exists:
2604          */
2605         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2606                 return;
2607
2608         next_entry = NULL;      /* avoid warning with gcc */
2609         while (entry != &head->list) {
2610                 /*
2611                  * Fetch the next entry in the list before calling
2612                  * handle_futex_death:
2613                  */
2614                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2615                 /*
2616                  * A pending lock might already be on the list, so
2617                  * don't process it twice:
2618                  */
2619                 if (entry != pending)
2620                         if (handle_futex_death((void __user *)entry + futex_offset,
2621                                                 curr, pi))
2622                                 return;
2623                 if (rc)
2624                         return;
2625                 entry = next_entry;
2626                 pi = next_pi;
2627                 /*
2628                  * Avoid excessively long or circular lists:
2629                  */
2630                 if (!--limit)
2631                         break;
2632
2633                 cond_resched();
2634         }
2635
2636         if (pending)
2637                 handle_futex_death((void __user *)pending + futex_offset,
2638                                    curr, pip);
2639 }
2640
2641 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2642                 u32 __user *uaddr2, u32 val2, u32 val3)
2643 {
2644         int cmd = op & FUTEX_CMD_MASK;
2645         unsigned int flags = 0;
2646
2647         if (!(op & FUTEX_PRIVATE_FLAG))
2648                 flags |= FLAGS_SHARED;
2649
2650         if (op & FUTEX_CLOCK_REALTIME) {
2651                 flags |= FLAGS_CLOCKRT;
2652                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2653                         return -ENOSYS;
2654         }
2655
2656         switch (cmd) {
2657         case FUTEX_LOCK_PI:
2658         case FUTEX_UNLOCK_PI:
2659         case FUTEX_TRYLOCK_PI:
2660         case FUTEX_WAIT_REQUEUE_PI:
2661         case FUTEX_CMP_REQUEUE_PI:
2662                 if (!futex_cmpxchg_enabled)
2663                         return -ENOSYS;
2664         }
2665
2666         switch (cmd) {
2667         case FUTEX_WAIT:
2668                 val3 = FUTEX_BITSET_MATCH_ANY;
2669         case FUTEX_WAIT_BITSET:
2670                 return futex_wait(uaddr, flags, val, timeout, val3);
2671         case FUTEX_WAKE:
2672                 val3 = FUTEX_BITSET_MATCH_ANY;
2673         case FUTEX_WAKE_BITSET:
2674                 return futex_wake(uaddr, flags, val, val3);
2675         case FUTEX_REQUEUE:
2676                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2677         case FUTEX_CMP_REQUEUE:
2678                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2679         case FUTEX_WAKE_OP:
2680                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2681         case FUTEX_LOCK_PI:
2682                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2683         case FUTEX_UNLOCK_PI:
2684                 return futex_unlock_pi(uaddr, flags);
2685         case FUTEX_TRYLOCK_PI:
2686                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2687         case FUTEX_WAIT_REQUEUE_PI:
2688                 val3 = FUTEX_BITSET_MATCH_ANY;
2689                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2690                                              uaddr2);
2691         case FUTEX_CMP_REQUEUE_PI:
2692                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2693         }
2694         return -ENOSYS;
2695 }
2696
2697
2698 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2699                 struct timespec __user *, utime, u32 __user *, uaddr2,
2700                 u32, val3)
2701 {
2702         struct timespec ts;
2703         ktime_t t, *tp = NULL;
2704         u32 val2 = 0;
2705         int cmd = op & FUTEX_CMD_MASK;
2706
2707         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2708                       cmd == FUTEX_WAIT_BITSET ||
2709                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2710                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2711                         return -EFAULT;
2712                 if (!timespec_valid(&ts))
2713                         return -EINVAL;
2714
2715                 t = timespec_to_ktime(ts);
2716                 if (cmd == FUTEX_WAIT)
2717                         t = ktime_add_safe(ktime_get(), t);
2718                 tp = &t;
2719         }
2720         /*
2721          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2722          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2723          */
2724         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2725             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2726                 val2 = (u32) (unsigned long) utime;
2727
2728         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2729 }
2730
2731 static int __init futex_init(void)
2732 {
2733         u32 curval;
2734         int i;
2735
2736         /*
2737          * This will fail and we want it. Some arch implementations do
2738          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2739          * functionality. We want to know that before we call in any
2740          * of the complex code paths. Also we want to prevent
2741          * registration of robust lists in that case. NULL is
2742          * guaranteed to fault and we get -EFAULT on functional
2743          * implementation, the non-functional ones will return
2744          * -ENOSYS.
2745          */
2746         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2747                 futex_cmpxchg_enabled = 1;
2748
2749         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2750                 plist_head_init(&futex_queues[i].chain);
2751                 spin_lock_init(&futex_queues[i].lock);
2752         }
2753
2754         return 0;
2755 }
2756 __initcall(futex_init);