tracelevel module: Prioritize trace events
[linux-3.10.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 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66
67 #include <asm/futex.h>
68
69 #include "rtmutex_common.h"
70
71 int __read_mostly futex_cmpxchg_enabled;
72
73 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
74
75 /*
76  * Futex flags used to encode options to functions and preserve them across
77  * restarts.
78  */
79 #define FLAGS_SHARED            0x01
80 #define FLAGS_CLOCKRT           0x02
81 #define FLAGS_HAS_TIMEOUT       0x04
82
83 /*
84  * Priority Inheritance state:
85  */
86 struct futex_pi_state {
87         /*
88          * list of 'owned' pi_state instances - these have to be
89          * cleaned up in do_exit() if the task exits prematurely:
90          */
91         struct list_head list;
92
93         /*
94          * The PI object:
95          */
96         struct rt_mutex pi_mutex;
97
98         struct task_struct *owner;
99         atomic_t refcount;
100
101         union futex_key key;
102 };
103
104 /**
105  * struct futex_q - The hashed futex queue entry, one per waiting task
106  * @list:               priority-sorted list of tasks waiting on this futex
107  * @task:               the task waiting on the futex
108  * @lock_ptr:           the hash bucket lock
109  * @key:                the key the futex is hashed on
110  * @pi_state:           optional priority inheritance state
111  * @rt_waiter:          rt_waiter storage for use with requeue_pi
112  * @requeue_pi_key:     the requeue_pi target futex key
113  * @bitset:             bitset for the optional bitmasked wakeup
114  *
115  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
116  * we can wake only the relevant ones (hashed queues may be shared).
117  *
118  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
119  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
120  * The order of wakeup is always to make the first condition true, then
121  * the second.
122  *
123  * PI futexes are typically woken before they are removed from the hash list via
124  * the rt_mutex code. See unqueue_me_pi().
125  */
126 struct futex_q {
127         struct plist_node list;
128
129         struct task_struct *task;
130         spinlock_t *lock_ptr;
131         union futex_key key;
132         struct futex_pi_state *pi_state;
133         struct rt_mutex_waiter *rt_waiter;
134         union futex_key *requeue_pi_key;
135         u32 bitset;
136 };
137
138 static const struct futex_q futex_q_init = {
139         /* list gets initialized in queue_me()*/
140         .key = FUTEX_KEY_INIT,
141         .bitset = FUTEX_BITSET_MATCH_ANY
142 };
143
144 /*
145  * Hash buckets are shared by all the futex_keys that hash to the same
146  * location.  Each key may have multiple futex_q structures, one for each task
147  * waiting on a futex.
148  */
149 struct futex_hash_bucket {
150         spinlock_t lock;
151         struct plist_head chain;
152 };
153
154 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
155
156 /*
157  * We hash on the keys returned from get_futex_key (see below).
158  */
159 static struct futex_hash_bucket *hash_futex(union futex_key *key)
160 {
161         u32 hash = jhash2((u32*)&key->both.word,
162                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
163                           key->both.offset);
164         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
165 }
166
167 /*
168  * Return 1 if two futex_keys are equal, 0 otherwise.
169  */
170 static inline int match_futex(union futex_key *key1, union futex_key *key2)
171 {
172         return (key1 && key2
173                 && key1->both.word == key2->both.word
174                 && key1->both.ptr == key2->both.ptr
175                 && key1->both.offset == key2->both.offset);
176 }
177
178 /*
179  * Take a reference to the resource addressed by a key.
180  * Can be called while holding spinlocks.
181  *
182  */
183 static void get_futex_key_refs(union futex_key *key)
184 {
185         if (!key->both.ptr)
186                 return;
187
188         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
189         case FUT_OFF_INODE:
190                 ihold(key->shared.inode);
191                 break;
192         case FUT_OFF_MMSHARED:
193                 atomic_inc(&key->private.mm->mm_count);
194                 break;
195         }
196 }
197
198 /*
199  * Drop a reference to the resource addressed by a key.
200  * The hash bucket spinlock must not be held.
201  */
202 static void drop_futex_key_refs(union futex_key *key)
203 {
204         if (!key->both.ptr) {
205                 /* If we're here then we tried to put a key we failed to get */
206                 WARN_ON_ONCE(1);
207                 return;
208         }
209
210         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
211         case FUT_OFF_INODE:
212                 iput(key->shared.inode);
213                 break;
214         case FUT_OFF_MMSHARED:
215                 mmdrop(key->private.mm);
216                 break;
217         }
218 }
219
220 /**
221  * get_futex_key() - Get parameters which are the keys for a futex
222  * @uaddr:      virtual address of the futex
223  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
224  * @key:        address where result is stored.
225  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
226  *              VERIFY_WRITE)
227  *
228  * Return: a negative error code or 0
229  *
230  * The key words are stored in *key on success.
231  *
232  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
233  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
234  * We can usually work out the index without swapping in the page.
235  *
236  * lock_page() might sleep, the caller should not hold a spinlock.
237  */
238 static int
239 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
240 {
241         unsigned long address = (unsigned long)uaddr;
242         struct mm_struct *mm = current->mm;
243         struct page *page, *page_head;
244         int err, ro = 0;
245
246         /*
247          * The futex address must be "naturally" aligned.
248          */
249         key->both.offset = address % PAGE_SIZE;
250         if (unlikely((address % sizeof(u32)) != 0))
251                 return -EINVAL;
252         address -= key->both.offset;
253
254         /*
255          * PROCESS_PRIVATE futexes are fast.
256          * As the mm cannot disappear under us and the 'key' only needs
257          * virtual address, we dont even have to find the underlying vma.
258          * Note : We do have to check 'uaddr' is a valid user address,
259          *        but access_ok() should be faster than find_vma()
260          */
261         if (!fshared) {
262                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
263                         return -EFAULT;
264                 key->private.mm = mm;
265                 key->private.address = address;
266                 get_futex_key_refs(key);
267                 return 0;
268         }
269
270 again:
271         err = get_user_pages_fast(address, 1, 1, &page);
272         /*
273          * If write access is not required (eg. FUTEX_WAIT), try
274          * and get read-only access.
275          */
276         if (err == -EFAULT && rw == VERIFY_READ) {
277                 err = get_user_pages_fast(address, 1, 0, &page);
278                 ro = 1;
279         }
280         if (err < 0)
281                 return err;
282         else
283                 err = 0;
284
285 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
286         page_head = page;
287         if (unlikely(PageTail(page))) {
288                 put_page(page);
289                 /* serialize against __split_huge_page_splitting() */
290                 local_irq_disable();
291                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
292                         page_head = compound_head(page);
293                         /*
294                          * page_head is valid pointer but we must pin
295                          * it before taking the PG_lock and/or
296                          * PG_compound_lock. The moment we re-enable
297                          * irqs __split_huge_page_splitting() can
298                          * return and the head page can be freed from
299                          * under us. We can't take the PG_lock and/or
300                          * PG_compound_lock on a page that could be
301                          * freed from under us.
302                          */
303                         if (page != page_head) {
304                                 get_page(page_head);
305                                 put_page(page);
306                         }
307                         local_irq_enable();
308                 } else {
309                         local_irq_enable();
310                         goto again;
311                 }
312         }
313 #else
314         page_head = compound_head(page);
315         if (page != page_head) {
316                 get_page(page_head);
317                 put_page(page);
318         }
319 #endif
320
321         lock_page(page_head);
322
323         /*
324          * If page_head->mapping is NULL, then it cannot be a PageAnon
325          * page; but it might be the ZERO_PAGE or in the gate area or
326          * in a special mapping (all cases which we are happy to fail);
327          * or it may have been a good file page when get_user_pages_fast
328          * found it, but truncated or holepunched or subjected to
329          * invalidate_complete_page2 before we got the page lock (also
330          * cases which we are happy to fail).  And we hold a reference,
331          * so refcount care in invalidate_complete_page's remove_mapping
332          * prevents drop_caches from setting mapping to NULL beneath us.
333          *
334          * The case we do have to guard against is when memory pressure made
335          * shmem_writepage move it from filecache to swapcache beneath us:
336          * an unlikely race, but we do need to retry for page_head->mapping.
337          */
338         if (!page_head->mapping) {
339                 int shmem_swizzled = PageSwapCache(page_head);
340                 unlock_page(page_head);
341                 put_page(page_head);
342                 if (shmem_swizzled)
343                         goto again;
344                 return -EFAULT;
345         }
346
347         /*
348          * Private mappings are handled in a simple way.
349          *
350          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
351          * it's a read-only handle, it's expected that futexes attach to
352          * the object not the particular process.
353          */
354         if (PageAnon(page_head)) {
355                 /*
356                  * A RO anonymous page will never change and thus doesn't make
357                  * sense for futex operations.
358                  */
359                 if (ro) {
360                         err = -EFAULT;
361                         goto out;
362                 }
363
364                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
365                 key->private.mm = mm;
366                 key->private.address = address;
367         } else {
368                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
369                 key->shared.inode = page_head->mapping->host;
370                 key->shared.pgoff = basepage_index(page);
371         }
372
373         get_futex_key_refs(key);
374
375 out:
376         unlock_page(page_head);
377         put_page(page_head);
378         return err;
379 }
380
381 static inline void put_futex_key(union futex_key *key)
382 {
383         drop_futex_key_refs(key);
384 }
385
386 /**
387  * fault_in_user_writeable() - Fault in user address and verify RW access
388  * @uaddr:      pointer to faulting user space address
389  *
390  * Slow path to fixup the fault we just took in the atomic write
391  * access to @uaddr.
392  *
393  * We have no generic implementation of a non-destructive write to the
394  * user address. We know that we faulted in the atomic pagefault
395  * disabled section so we can as well avoid the #PF overhead by
396  * calling get_user_pages() right away.
397  */
398 static int fault_in_user_writeable(u32 __user *uaddr)
399 {
400         struct mm_struct *mm = current->mm;
401         int ret;
402
403         down_read(&mm->mmap_sem);
404         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
405                                FAULT_FLAG_WRITE);
406         up_read(&mm->mmap_sem);
407
408         return ret < 0 ? ret : 0;
409 }
410
411 /**
412  * futex_top_waiter() - Return the highest priority waiter on a futex
413  * @hb:         the hash bucket the futex_q's reside in
414  * @key:        the futex key (to distinguish it from other futex futex_q's)
415  *
416  * Must be called with the hb lock held.
417  */
418 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
419                                         union futex_key *key)
420 {
421         struct futex_q *this;
422
423         plist_for_each_entry(this, &hb->chain, list) {
424                 if (match_futex(&this->key, key))
425                         return this;
426         }
427         return NULL;
428 }
429
430 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
431                                       u32 uval, u32 newval)
432 {
433         int ret;
434
435         pagefault_disable();
436         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
437         pagefault_enable();
438
439         return ret;
440 }
441
442 static int get_futex_value_locked(u32 *dest, u32 __user *from)
443 {
444         int ret;
445
446         pagefault_disable();
447         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
448         pagefault_enable();
449
450         return ret ? -EFAULT : 0;
451 }
452
453
454 /*
455  * PI code:
456  */
457 static int refill_pi_state_cache(void)
458 {
459         struct futex_pi_state *pi_state;
460
461         if (likely(current->pi_state_cache))
462                 return 0;
463
464         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
465
466         if (!pi_state)
467                 return -ENOMEM;
468
469         INIT_LIST_HEAD(&pi_state->list);
470         /* pi_mutex gets initialized later */
471         pi_state->owner = NULL;
472         atomic_set(&pi_state->refcount, 1);
473         pi_state->key = FUTEX_KEY_INIT;
474
475         current->pi_state_cache = pi_state;
476
477         return 0;
478 }
479
480 static struct futex_pi_state * alloc_pi_state(void)
481 {
482         struct futex_pi_state *pi_state = current->pi_state_cache;
483
484         WARN_ON(!pi_state);
485         current->pi_state_cache = NULL;
486
487         return pi_state;
488 }
489
490 static void free_pi_state(struct futex_pi_state *pi_state)
491 {
492         if (!atomic_dec_and_test(&pi_state->refcount))
493                 return;
494
495         /*
496          * If pi_state->owner is NULL, the owner is most probably dying
497          * and has cleaned up the pi_state already
498          */
499         if (pi_state->owner) {
500                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
501                 list_del_init(&pi_state->list);
502                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
503
504                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
505         }
506
507         if (current->pi_state_cache)
508                 kfree(pi_state);
509         else {
510                 /*
511                  * pi_state->list is already empty.
512                  * clear pi_state->owner.
513                  * refcount is at 0 - put it back to 1.
514                  */
515                 pi_state->owner = NULL;
516                 atomic_set(&pi_state->refcount, 1);
517                 current->pi_state_cache = pi_state;
518         }
519 }
520
521 /*
522  * Look up the task based on what TID userspace gave us.
523  * We dont trust it.
524  */
525 static struct task_struct * futex_find_get_task(pid_t pid)
526 {
527         struct task_struct *p;
528
529         rcu_read_lock();
530         p = find_task_by_vpid(pid);
531         if (p)
532                 get_task_struct(p);
533
534         rcu_read_unlock();
535
536         return p;
537 }
538
539 /*
540  * This task is holding PI mutexes at exit time => bad.
541  * Kernel cleans up PI-state, but userspace is likely hosed.
542  * (Robust-futex cleanup is separate and might save the day for userspace.)
543  */
544 void exit_pi_state_list(struct task_struct *curr)
545 {
546         struct list_head *next, *head = &curr->pi_state_list;
547         struct futex_pi_state *pi_state;
548         struct futex_hash_bucket *hb;
549         union futex_key key = FUTEX_KEY_INIT;
550
551         if (!futex_cmpxchg_enabled)
552                 return;
553         /*
554          * We are a ZOMBIE and nobody can enqueue itself on
555          * pi_state_list anymore, but we have to be careful
556          * versus waiters unqueueing themselves:
557          */
558         raw_spin_lock_irq(&curr->pi_lock);
559         while (!list_empty(head)) {
560
561                 next = head->next;
562                 pi_state = list_entry(next, struct futex_pi_state, list);
563                 key = pi_state->key;
564                 hb = hash_futex(&key);
565                 raw_spin_unlock_irq(&curr->pi_lock);
566
567                 spin_lock(&hb->lock);
568
569                 raw_spin_lock_irq(&curr->pi_lock);
570                 /*
571                  * We dropped the pi-lock, so re-check whether this
572                  * task still owns the PI-state:
573                  */
574                 if (head->next != next) {
575                         spin_unlock(&hb->lock);
576                         continue;
577                 }
578
579                 WARN_ON(pi_state->owner != curr);
580                 WARN_ON(list_empty(&pi_state->list));
581                 list_del_init(&pi_state->list);
582                 pi_state->owner = NULL;
583                 raw_spin_unlock_irq(&curr->pi_lock);
584
585                 rt_mutex_unlock(&pi_state->pi_mutex);
586
587                 spin_unlock(&hb->lock);
588
589                 raw_spin_lock_irq(&curr->pi_lock);
590         }
591         raw_spin_unlock_irq(&curr->pi_lock);
592 }
593
594 static int
595 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
596                 union futex_key *key, struct futex_pi_state **ps)
597 {
598         struct futex_pi_state *pi_state = NULL;
599         struct futex_q *this, *next;
600         struct plist_head *head;
601         struct task_struct *p;
602         pid_t pid = uval & FUTEX_TID_MASK;
603
604         head = &hb->chain;
605
606         plist_for_each_entry_safe(this, next, head, list) {
607                 if (match_futex(&this->key, key)) {
608                         /*
609                          * Another waiter already exists - bump up
610                          * the refcount and return its pi_state:
611                          */
612                         pi_state = this->pi_state;
613                         /*
614                          * Userspace might have messed up non-PI and PI futexes
615                          */
616                         if (unlikely(!pi_state))
617                                 return -EINVAL;
618
619                         WARN_ON(!atomic_read(&pi_state->refcount));
620
621                         /*
622                          * When pi_state->owner is NULL then the owner died
623                          * and another waiter is on the fly. pi_state->owner
624                          * is fixed up by the task which acquires
625                          * pi_state->rt_mutex.
626                          *
627                          * We do not check for pid == 0 which can happen when
628                          * the owner died and robust_list_exit() cleared the
629                          * TID.
630                          */
631                         if (pid && pi_state->owner) {
632                                 /*
633                                  * Bail out if user space manipulated the
634                                  * futex value.
635                                  */
636                                 if (pid != task_pid_vnr(pi_state->owner))
637                                         return -EINVAL;
638                         }
639
640                         atomic_inc(&pi_state->refcount);
641                         *ps = pi_state;
642
643                         return 0;
644                 }
645         }
646
647         /*
648          * We are the first waiter - try to look up the real owner and attach
649          * the new pi_state to it, but bail out when TID = 0
650          */
651         if (!pid)
652                 return -ESRCH;
653         p = futex_find_get_task(pid);
654         if (!p)
655                 return -ESRCH;
656
657         /*
658          * We need to look at the task state flags to figure out,
659          * whether the task is exiting. To protect against the do_exit
660          * change of the task flags, we do this protected by
661          * p->pi_lock:
662          */
663         raw_spin_lock_irq(&p->pi_lock);
664         if (unlikely(p->flags & PF_EXITING)) {
665                 /*
666                  * The task is on the way out. When PF_EXITPIDONE is
667                  * set, we know that the task has finished the
668                  * cleanup:
669                  */
670                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
671
672                 raw_spin_unlock_irq(&p->pi_lock);
673                 put_task_struct(p);
674                 return ret;
675         }
676
677         pi_state = alloc_pi_state();
678
679         /*
680          * Initialize the pi_mutex in locked state and make 'p'
681          * the owner of it:
682          */
683         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
684
685         /* Store the key for possible exit cleanups: */
686         pi_state->key = *key;
687
688         WARN_ON(!list_empty(&pi_state->list));
689         list_add(&pi_state->list, &p->pi_state_list);
690         pi_state->owner = p;
691         raw_spin_unlock_irq(&p->pi_lock);
692
693         put_task_struct(p);
694
695         *ps = pi_state;
696
697         return 0;
698 }
699
700 /**
701  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
702  * @uaddr:              the pi futex user address
703  * @hb:                 the pi futex hash bucket
704  * @key:                the futex key associated with uaddr and hb
705  * @ps:                 the pi_state pointer where we store the result of the
706  *                      lookup
707  * @task:               the task to perform the atomic lock work for.  This will
708  *                      be "current" except in the case of requeue pi.
709  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
710  *
711  * Return:
712  *  0 - ready to wait;
713  *  1 - acquired the lock;
714  * <0 - error
715  *
716  * The hb->lock and futex_key refs shall be held by the caller.
717  */
718 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
719                                 union futex_key *key,
720                                 struct futex_pi_state **ps,
721                                 struct task_struct *task, int set_waiters)
722 {
723         int lock_taken, ret, force_take = 0;
724         u32 uval, newval, curval, vpid = task_pid_vnr(task);
725
726 retry:
727         ret = lock_taken = 0;
728
729         /*
730          * To avoid races, we attempt to take the lock here again
731          * (by doing a 0 -> TID atomic cmpxchg), while holding all
732          * the locks. It will most likely not succeed.
733          */
734         newval = vpid;
735         if (set_waiters)
736                 newval |= FUTEX_WAITERS;
737
738         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
739                 return -EFAULT;
740
741         /*
742          * Detect deadlocks.
743          */
744         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
745                 return -EDEADLK;
746
747         /*
748          * Surprise - we got the lock. Just return to userspace:
749          */
750         if (unlikely(!curval))
751                 return 1;
752
753         uval = curval;
754
755         /*
756          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
757          * to wake at the next unlock.
758          */
759         newval = curval | FUTEX_WAITERS;
760
761         /*
762          * Should we force take the futex? See below.
763          */
764         if (unlikely(force_take)) {
765                 /*
766                  * Keep the OWNER_DIED and the WAITERS bit and set the
767                  * new TID value.
768                  */
769                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
770                 force_take = 0;
771                 lock_taken = 1;
772         }
773
774         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
775                 return -EFAULT;
776         if (unlikely(curval != uval))
777                 goto retry;
778
779         /*
780          * We took the lock due to forced take over.
781          */
782         if (unlikely(lock_taken))
783                 return 1;
784
785         /*
786          * We dont have the lock. Look up the PI state (or create it if
787          * we are the first waiter):
788          */
789         ret = lookup_pi_state(uval, hb, key, ps);
790
791         if (unlikely(ret)) {
792                 switch (ret) {
793                 case -ESRCH:
794                         /*
795                          * We failed to find an owner for this
796                          * futex. So we have no pi_state to block
797                          * on. This can happen in two cases:
798                          *
799                          * 1) The owner died
800                          * 2) A stale FUTEX_WAITERS bit
801                          *
802                          * Re-read the futex value.
803                          */
804                         if (get_futex_value_locked(&curval, uaddr))
805                                 return -EFAULT;
806
807                         /*
808                          * If the owner died or we have a stale
809                          * WAITERS bit the owner TID in the user space
810                          * futex is 0.
811                          */
812                         if (!(curval & FUTEX_TID_MASK)) {
813                                 force_take = 1;
814                                 goto retry;
815                         }
816                 default:
817                         break;
818                 }
819         }
820
821         return ret;
822 }
823
824 /**
825  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
826  * @q:  The futex_q to unqueue
827  *
828  * The q->lock_ptr must not be NULL and must be held by the caller.
829  */
830 static void __unqueue_futex(struct futex_q *q)
831 {
832         struct futex_hash_bucket *hb;
833
834         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
835             || WARN_ON(plist_node_empty(&q->list)))
836                 return;
837
838         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
839         plist_del(&q->list, &hb->chain);
840 }
841
842 /*
843  * The hash bucket lock must be held when this is called.
844  * Afterwards, the futex_q must not be accessed.
845  */
846 static void wake_futex(struct futex_q *q)
847 {
848         struct task_struct *p = q->task;
849
850         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
851                 return;
852
853         /*
854          * We set q->lock_ptr = NULL _before_ we wake up the task. If
855          * a non-futex wake up happens on another CPU then the task
856          * might exit and p would dereference a non-existing task
857          * struct. Prevent this by holding a reference on p across the
858          * wake up.
859          */
860         get_task_struct(p);
861
862         __unqueue_futex(q);
863         /*
864          * The waiting task can free the futex_q as soon as
865          * q->lock_ptr = NULL is written, without taking any locks. A
866          * memory barrier is required here to prevent the following
867          * store to lock_ptr from getting ahead of the plist_del.
868          */
869         smp_wmb();
870         q->lock_ptr = NULL;
871
872         wake_up_state(p, TASK_NORMAL);
873         put_task_struct(p);
874 }
875
876 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
877 {
878         struct task_struct *new_owner;
879         struct futex_pi_state *pi_state = this->pi_state;
880         u32 uninitialized_var(curval), newval;
881
882         if (!pi_state)
883                 return -EINVAL;
884
885         /*
886          * If current does not own the pi_state then the futex is
887          * inconsistent and user space fiddled with the futex value.
888          */
889         if (pi_state->owner != current)
890                 return -EINVAL;
891
892         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
893         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
894
895         /*
896          * It is possible that the next waiter (the one that brought
897          * this owner to the kernel) timed out and is no longer
898          * waiting on the lock.
899          */
900         if (!new_owner)
901                 new_owner = this->task;
902
903         /*
904          * We pass it to the next owner. (The WAITERS bit is always
905          * kept enabled while there is PI state around. We must also
906          * preserve the owner died bit.)
907          */
908         if (!(uval & FUTEX_OWNER_DIED)) {
909                 int ret = 0;
910
911                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
912
913                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
914                         ret = -EFAULT;
915                 else if (curval != uval)
916                         ret = -EINVAL;
917                 if (ret) {
918                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
919                         return ret;
920                 }
921         }
922
923         raw_spin_lock_irq(&pi_state->owner->pi_lock);
924         WARN_ON(list_empty(&pi_state->list));
925         list_del_init(&pi_state->list);
926         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
927
928         raw_spin_lock_irq(&new_owner->pi_lock);
929         WARN_ON(!list_empty(&pi_state->list));
930         list_add(&pi_state->list, &new_owner->pi_state_list);
931         pi_state->owner = new_owner;
932         raw_spin_unlock_irq(&new_owner->pi_lock);
933
934         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
935         rt_mutex_unlock(&pi_state->pi_mutex);
936
937         return 0;
938 }
939
940 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
941 {
942         u32 uninitialized_var(oldval);
943
944         /*
945          * There is no waiter, so we unlock the futex. The owner died
946          * bit has not to be preserved here. We are the owner:
947          */
948         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
949                 return -EFAULT;
950         if (oldval != uval)
951                 return -EAGAIN;
952
953         return 0;
954 }
955
956 /*
957  * Express the locking dependencies for lockdep:
958  */
959 static inline void
960 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
961 {
962         if (hb1 <= hb2) {
963                 spin_lock(&hb1->lock);
964                 if (hb1 < hb2)
965                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
966         } else { /* hb1 > hb2 */
967                 spin_lock(&hb2->lock);
968                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
969         }
970 }
971
972 static inline void
973 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
974 {
975         spin_unlock(&hb1->lock);
976         if (hb1 != hb2)
977                 spin_unlock(&hb2->lock);
978 }
979
980 /*
981  * Wake up waiters matching bitset queued on this futex (uaddr).
982  */
983 static int
984 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
985 {
986         struct futex_hash_bucket *hb;
987         struct futex_q *this, *next;
988         struct plist_head *head;
989         union futex_key key = FUTEX_KEY_INIT;
990         int ret;
991
992         if (!bitset)
993                 return -EINVAL;
994
995         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
996         if (unlikely(ret != 0))
997                 goto out;
998
999         hb = hash_futex(&key);
1000         spin_lock(&hb->lock);
1001         head = &hb->chain;
1002
1003         plist_for_each_entry_safe(this, next, head, list) {
1004                 if (match_futex (&this->key, &key)) {
1005                         if (this->pi_state || this->rt_waiter) {
1006                                 ret = -EINVAL;
1007                                 break;
1008                         }
1009
1010                         /* Check if one of the bits is set in both bitsets */
1011                         if (!(this->bitset & bitset))
1012                                 continue;
1013
1014                         wake_futex(this);
1015                         if (++ret >= nr_wake)
1016                                 break;
1017                 }
1018         }
1019
1020         spin_unlock(&hb->lock);
1021         put_futex_key(&key);
1022 out:
1023         return ret;
1024 }
1025
1026 /*
1027  * Wake up all waiters hashed on the physical page that is mapped
1028  * to this virtual address:
1029  */
1030 static int
1031 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1032               int nr_wake, int nr_wake2, int op)
1033 {
1034         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1035         struct futex_hash_bucket *hb1, *hb2;
1036         struct plist_head *head;
1037         struct futex_q *this, *next;
1038         int ret, op_ret;
1039
1040 retry:
1041         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1042         if (unlikely(ret != 0))
1043                 goto out;
1044         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1045         if (unlikely(ret != 0))
1046                 goto out_put_key1;
1047
1048         hb1 = hash_futex(&key1);
1049         hb2 = hash_futex(&key2);
1050
1051 retry_private:
1052         double_lock_hb(hb1, hb2);
1053         op_ret = futex_atomic_op_inuser(op, uaddr2);
1054         if (unlikely(op_ret < 0)) {
1055
1056                 double_unlock_hb(hb1, hb2);
1057
1058 #ifndef CONFIG_MMU
1059                 /*
1060                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1061                  * but we might get them from range checking
1062                  */
1063                 ret = op_ret;
1064                 goto out_put_keys;
1065 #endif
1066
1067                 if (unlikely(op_ret != -EFAULT)) {
1068                         ret = op_ret;
1069                         goto out_put_keys;
1070                 }
1071
1072                 ret = fault_in_user_writeable(uaddr2);
1073                 if (ret)
1074                         goto out_put_keys;
1075
1076                 if (!(flags & FLAGS_SHARED))
1077                         goto retry_private;
1078
1079                 put_futex_key(&key2);
1080                 put_futex_key(&key1);
1081                 goto retry;
1082         }
1083
1084         head = &hb1->chain;
1085
1086         plist_for_each_entry_safe(this, next, head, list) {
1087                 if (match_futex (&this->key, &key1)) {
1088                         if (this->pi_state || this->rt_waiter) {
1089                                 ret = -EINVAL;
1090                                 goto out_unlock;
1091                         }
1092                         wake_futex(this);
1093                         if (++ret >= nr_wake)
1094                                 break;
1095                 }
1096         }
1097
1098         if (op_ret > 0) {
1099                 head = &hb2->chain;
1100
1101                 op_ret = 0;
1102                 plist_for_each_entry_safe(this, next, head, list) {
1103                         if (match_futex (&this->key, &key2)) {
1104                                 if (this->pi_state || this->rt_waiter) {
1105                                         ret = -EINVAL;
1106                                         goto out_unlock;
1107                                 }
1108                                 wake_futex(this);
1109                                 if (++op_ret >= nr_wake2)
1110                                         break;
1111                         }
1112                 }
1113                 ret += op_ret;
1114         }
1115
1116 out_unlock:
1117         double_unlock_hb(hb1, hb2);
1118 out_put_keys:
1119         put_futex_key(&key2);
1120 out_put_key1:
1121         put_futex_key(&key1);
1122 out:
1123         return ret;
1124 }
1125
1126 /**
1127  * requeue_futex() - Requeue a futex_q from one hb to another
1128  * @q:          the futex_q to requeue
1129  * @hb1:        the source hash_bucket
1130  * @hb2:        the target hash_bucket
1131  * @key2:       the new key for the requeued futex_q
1132  */
1133 static inline
1134 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1135                    struct futex_hash_bucket *hb2, union futex_key *key2)
1136 {
1137
1138         /*
1139          * If key1 and key2 hash to the same bucket, no need to
1140          * requeue.
1141          */
1142         if (likely(&hb1->chain != &hb2->chain)) {
1143                 plist_del(&q->list, &hb1->chain);
1144                 plist_add(&q->list, &hb2->chain);
1145                 q->lock_ptr = &hb2->lock;
1146         }
1147         get_futex_key_refs(key2);
1148         q->key = *key2;
1149 }
1150
1151 /**
1152  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1153  * @q:          the futex_q
1154  * @key:        the key of the requeue target futex
1155  * @hb:         the hash_bucket of the requeue target futex
1156  *
1157  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1158  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1159  * to the requeue target futex so the waiter can detect the wakeup on the right
1160  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1161  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1162  * to protect access to the pi_state to fixup the owner later.  Must be called
1163  * with both q->lock_ptr and hb->lock held.
1164  */
1165 static inline
1166 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1167                            struct futex_hash_bucket *hb)
1168 {
1169         get_futex_key_refs(key);
1170         q->key = *key;
1171
1172         __unqueue_futex(q);
1173
1174         WARN_ON(!q->rt_waiter);
1175         q->rt_waiter = NULL;
1176
1177         q->lock_ptr = &hb->lock;
1178
1179         wake_up_state(q->task, TASK_NORMAL);
1180 }
1181
1182 /**
1183  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1184  * @pifutex:            the user address of the to futex
1185  * @hb1:                the from futex hash bucket, must be locked by the caller
1186  * @hb2:                the to futex hash bucket, must be locked by the caller
1187  * @key1:               the from futex key
1188  * @key2:               the to futex key
1189  * @ps:                 address to store the pi_state pointer
1190  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1191  *
1192  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1193  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1194  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1195  * hb1 and hb2 must be held by the caller.
1196  *
1197  * Return:
1198  *  0 - failed to acquire the lock atomically;
1199  *  1 - acquired the lock;
1200  * <0 - error
1201  */
1202 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1203                                  struct futex_hash_bucket *hb1,
1204                                  struct futex_hash_bucket *hb2,
1205                                  union futex_key *key1, union futex_key *key2,
1206                                  struct futex_pi_state **ps, int set_waiters)
1207 {
1208         struct futex_q *top_waiter = NULL;
1209         u32 curval;
1210         int ret;
1211
1212         if (get_futex_value_locked(&curval, pifutex))
1213                 return -EFAULT;
1214
1215         /*
1216          * Find the top_waiter and determine if there are additional waiters.
1217          * If the caller intends to requeue more than 1 waiter to pifutex,
1218          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1219          * as we have means to handle the possible fault.  If not, don't set
1220          * the bit unecessarily as it will force the subsequent unlock to enter
1221          * the kernel.
1222          */
1223         top_waiter = futex_top_waiter(hb1, key1);
1224
1225         /* There are no waiters, nothing for us to do. */
1226         if (!top_waiter)
1227                 return 0;
1228
1229         /* Ensure we requeue to the expected futex. */
1230         if (!match_futex(top_waiter->requeue_pi_key, key2))
1231                 return -EINVAL;
1232
1233         /*
1234          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1235          * the contended case or if set_waiters is 1.  The pi_state is returned
1236          * in ps in contended cases.
1237          */
1238         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1239                                    set_waiters);
1240         if (ret == 1)
1241                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1242
1243         return ret;
1244 }
1245
1246 /**
1247  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1248  * @uaddr1:     source futex user address
1249  * @flags:      futex flags (FLAGS_SHARED, etc.)
1250  * @uaddr2:     target futex user address
1251  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1252  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1253  * @cmpval:     @uaddr1 expected value (or %NULL)
1254  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1255  *              pi futex (pi to pi requeue is not supported)
1256  *
1257  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1258  * uaddr2 atomically on behalf of the top waiter.
1259  *
1260  * Return:
1261  * >=0 - on success, the number of tasks requeued or woken;
1262  *  <0 - on error
1263  */
1264 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1265                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1266                          u32 *cmpval, int requeue_pi)
1267 {
1268         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1269         int drop_count = 0, task_count = 0, ret;
1270         struct futex_pi_state *pi_state = NULL;
1271         struct futex_hash_bucket *hb1, *hb2;
1272         struct plist_head *head1;
1273         struct futex_q *this, *next;
1274         u32 curval2;
1275
1276         if (requeue_pi) {
1277                 /*
1278                  * requeue_pi requires a pi_state, try to allocate it now
1279                  * without any locks in case it fails.
1280                  */
1281                 if (refill_pi_state_cache())
1282                         return -ENOMEM;
1283                 /*
1284                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1285                  * + nr_requeue, since it acquires the rt_mutex prior to
1286                  * returning to userspace, so as to not leave the rt_mutex with
1287                  * waiters and no owner.  However, second and third wake-ups
1288                  * cannot be predicted as they involve race conditions with the
1289                  * first wake and a fault while looking up the pi_state.  Both
1290                  * pthread_cond_signal() and pthread_cond_broadcast() should
1291                  * use nr_wake=1.
1292                  */
1293                 if (nr_wake != 1)
1294                         return -EINVAL;
1295         }
1296
1297 retry:
1298         if (pi_state != NULL) {
1299                 /*
1300                  * We will have to lookup the pi_state again, so free this one
1301                  * to keep the accounting correct.
1302                  */
1303                 free_pi_state(pi_state);
1304                 pi_state = NULL;
1305         }
1306
1307         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1308         if (unlikely(ret != 0))
1309                 goto out;
1310         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1311                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1312         if (unlikely(ret != 0))
1313                 goto out_put_key1;
1314
1315         hb1 = hash_futex(&key1);
1316         hb2 = hash_futex(&key2);
1317
1318 retry_private:
1319         double_lock_hb(hb1, hb2);
1320
1321         if (likely(cmpval != NULL)) {
1322                 u32 curval;
1323
1324                 ret = get_futex_value_locked(&curval, uaddr1);
1325
1326                 if (unlikely(ret)) {
1327                         double_unlock_hb(hb1, hb2);
1328
1329                         ret = get_user(curval, uaddr1);
1330                         if (ret)
1331                                 goto out_put_keys;
1332
1333                         if (!(flags & FLAGS_SHARED))
1334                                 goto retry_private;
1335
1336                         put_futex_key(&key2);
1337                         put_futex_key(&key1);
1338                         goto retry;
1339                 }
1340                 if (curval != *cmpval) {
1341                         ret = -EAGAIN;
1342                         goto out_unlock;
1343                 }
1344         }
1345
1346         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1347                 /*
1348                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1349                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1350                  * bit.  We force this here where we are able to easily handle
1351                  * faults rather in the requeue loop below.
1352                  */
1353                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1354                                                  &key2, &pi_state, nr_requeue);
1355
1356                 /*
1357                  * At this point the top_waiter has either taken uaddr2 or is
1358                  * waiting on it.  If the former, then the pi_state will not
1359                  * exist yet, look it up one more time to ensure we have a
1360                  * reference to it.
1361                  */
1362                 if (ret == 1) {
1363                         WARN_ON(pi_state);
1364                         drop_count++;
1365                         task_count++;
1366                         ret = get_futex_value_locked(&curval2, uaddr2);
1367                         if (!ret)
1368                                 ret = lookup_pi_state(curval2, hb2, &key2,
1369                                                       &pi_state);
1370                 }
1371
1372                 switch (ret) {
1373                 case 0:
1374                         break;
1375                 case -EFAULT:
1376                         double_unlock_hb(hb1, hb2);
1377                         put_futex_key(&key2);
1378                         put_futex_key(&key1);
1379                         ret = fault_in_user_writeable(uaddr2);
1380                         if (!ret)
1381                                 goto retry;
1382                         goto out;
1383                 case -EAGAIN:
1384                         /* The owner was exiting, try again. */
1385                         double_unlock_hb(hb1, hb2);
1386                         put_futex_key(&key2);
1387                         put_futex_key(&key1);
1388                         cond_resched();
1389                         goto retry;
1390                 default:
1391                         goto out_unlock;
1392                 }
1393         }
1394
1395         head1 = &hb1->chain;
1396         plist_for_each_entry_safe(this, next, head1, list) {
1397                 if (task_count - nr_wake >= nr_requeue)
1398                         break;
1399
1400                 if (!match_futex(&this->key, &key1))
1401                         continue;
1402
1403                 /*
1404                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1405                  * be paired with each other and no other futex ops.
1406                  *
1407                  * We should never be requeueing a futex_q with a pi_state,
1408                  * which is awaiting a futex_unlock_pi().
1409                  */
1410                 if ((requeue_pi && !this->rt_waiter) ||
1411                     (!requeue_pi && this->rt_waiter) ||
1412                     this->pi_state) {
1413                         ret = -EINVAL;
1414                         break;
1415                 }
1416
1417                 /*
1418                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1419                  * lock, we already woke the top_waiter.  If not, it will be
1420                  * woken by futex_unlock_pi().
1421                  */
1422                 if (++task_count <= nr_wake && !requeue_pi) {
1423                         wake_futex(this);
1424                         continue;
1425                 }
1426
1427                 /* Ensure we requeue to the expected futex for requeue_pi. */
1428                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1429                         ret = -EINVAL;
1430                         break;
1431                 }
1432
1433                 /*
1434                  * Requeue nr_requeue waiters and possibly one more in the case
1435                  * of requeue_pi if we couldn't acquire the lock atomically.
1436                  */
1437                 if (requeue_pi) {
1438                         /* Prepare the waiter to take the rt_mutex. */
1439                         atomic_inc(&pi_state->refcount);
1440                         this->pi_state = pi_state;
1441                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1442                                                         this->rt_waiter,
1443                                                         this->task, 1);
1444                         if (ret == 1) {
1445                                 /* We got the lock. */
1446                                 requeue_pi_wake_futex(this, &key2, hb2);
1447                                 drop_count++;
1448                                 continue;
1449                         } else if (ret) {
1450                                 /* -EDEADLK */
1451                                 this->pi_state = NULL;
1452                                 free_pi_state(pi_state);
1453                                 goto out_unlock;
1454                         }
1455                 }
1456                 requeue_futex(this, hb1, hb2, &key2);
1457                 drop_count++;
1458         }
1459
1460 out_unlock:
1461         double_unlock_hb(hb1, hb2);
1462
1463         /*
1464          * drop_futex_key_refs() must be called outside the spinlocks. During
1465          * the requeue we moved futex_q's from the hash bucket at key1 to the
1466          * one at key2 and updated their key pointer.  We no longer need to
1467          * hold the references to key1.
1468          */
1469         while (--drop_count >= 0)
1470                 drop_futex_key_refs(&key1);
1471
1472 out_put_keys:
1473         put_futex_key(&key2);
1474 out_put_key1:
1475         put_futex_key(&key1);
1476 out:
1477         if (pi_state != NULL)
1478                 free_pi_state(pi_state);
1479         return ret ? ret : task_count;
1480 }
1481
1482 /* The key must be already stored in q->key. */
1483 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1484         __acquires(&hb->lock)
1485 {
1486         struct futex_hash_bucket *hb;
1487
1488         hb = hash_futex(&q->key);
1489         q->lock_ptr = &hb->lock;
1490
1491         spin_lock(&hb->lock);
1492         return hb;
1493 }
1494
1495 static inline void
1496 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1497         __releases(&hb->lock)
1498 {
1499         spin_unlock(&hb->lock);
1500 }
1501
1502 /**
1503  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1504  * @q:  The futex_q to enqueue
1505  * @hb: The destination hash bucket
1506  *
1507  * The hb->lock must be held by the caller, and is released here. A call to
1508  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1509  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1510  * or nothing if the unqueue is done as part of the wake process and the unqueue
1511  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1512  * an example).
1513  */
1514 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1515         __releases(&hb->lock)
1516 {
1517         int prio;
1518
1519         /*
1520          * The priority used to register this element is
1521          * - either the real thread-priority for the real-time threads
1522          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1523          * - or MAX_RT_PRIO for non-RT threads.
1524          * Thus, all RT-threads are woken first in priority order, and
1525          * the others are woken last, in FIFO order.
1526          */
1527         prio = min(current->normal_prio, MAX_RT_PRIO);
1528
1529         plist_node_init(&q->list, prio);
1530         plist_add(&q->list, &hb->chain);
1531         q->task = current;
1532         spin_unlock(&hb->lock);
1533 }
1534
1535 /**
1536  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1537  * @q:  The futex_q to unqueue
1538  *
1539  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1540  * be paired with exactly one earlier call to queue_me().
1541  *
1542  * Return:
1543  *   1 - if the futex_q was still queued (and we removed unqueued it);
1544  *   0 - if the futex_q was already removed by the waking thread
1545  */
1546 static int unqueue_me(struct futex_q *q)
1547 {
1548         spinlock_t *lock_ptr;
1549         int ret = 0;
1550
1551         /* In the common case we don't take the spinlock, which is nice. */
1552 retry:
1553         lock_ptr = q->lock_ptr;
1554         barrier();
1555         if (lock_ptr != NULL) {
1556                 spin_lock(lock_ptr);
1557                 /*
1558                  * q->lock_ptr can change between reading it and
1559                  * spin_lock(), causing us to take the wrong lock.  This
1560                  * corrects the race condition.
1561                  *
1562                  * Reasoning goes like this: if we have the wrong lock,
1563                  * q->lock_ptr must have changed (maybe several times)
1564                  * between reading it and the spin_lock().  It can
1565                  * change again after the spin_lock() but only if it was
1566                  * already changed before the spin_lock().  It cannot,
1567                  * however, change back to the original value.  Therefore
1568                  * we can detect whether we acquired the correct lock.
1569                  */
1570                 if (unlikely(lock_ptr != q->lock_ptr)) {
1571                         spin_unlock(lock_ptr);
1572                         goto retry;
1573                 }
1574                 __unqueue_futex(q);
1575
1576                 BUG_ON(q->pi_state);
1577
1578                 spin_unlock(lock_ptr);
1579                 ret = 1;
1580         }
1581
1582         drop_futex_key_refs(&q->key);
1583         return ret;
1584 }
1585
1586 /*
1587  * PI futexes can not be requeued and must remove themself from the
1588  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1589  * and dropped here.
1590  */
1591 static void unqueue_me_pi(struct futex_q *q)
1592         __releases(q->lock_ptr)
1593 {
1594         __unqueue_futex(q);
1595
1596         BUG_ON(!q->pi_state);
1597         free_pi_state(q->pi_state);
1598         q->pi_state = NULL;
1599
1600         spin_unlock(q->lock_ptr);
1601 }
1602
1603 /*
1604  * Fixup the pi_state owner with the new owner.
1605  *
1606  * Must be called with hash bucket lock held and mm->sem held for non
1607  * private futexes.
1608  */
1609 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1610                                 struct task_struct *newowner)
1611 {
1612         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1613         struct futex_pi_state *pi_state = q->pi_state;
1614         struct task_struct *oldowner = pi_state->owner;
1615         u32 uval, uninitialized_var(curval), newval;
1616         int ret;
1617
1618         /* Owner died? */
1619         if (!pi_state->owner)
1620                 newtid |= FUTEX_OWNER_DIED;
1621
1622         /*
1623          * We are here either because we stole the rtmutex from the
1624          * previous highest priority waiter or we are the highest priority
1625          * waiter but failed to get the rtmutex the first time.
1626          * We have to replace the newowner TID in the user space variable.
1627          * This must be atomic as we have to preserve the owner died bit here.
1628          *
1629          * Note: We write the user space value _before_ changing the pi_state
1630          * because we can fault here. Imagine swapped out pages or a fork
1631          * that marked all the anonymous memory readonly for cow.
1632          *
1633          * Modifying pi_state _before_ the user space value would
1634          * leave the pi_state in an inconsistent state when we fault
1635          * here, because we need to drop the hash bucket lock to
1636          * handle the fault. This might be observed in the PID check
1637          * in lookup_pi_state.
1638          */
1639 retry:
1640         if (get_futex_value_locked(&uval, uaddr))
1641                 goto handle_fault;
1642
1643         while (1) {
1644                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1645
1646                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1647                         goto handle_fault;
1648                 if (curval == uval)
1649                         break;
1650                 uval = curval;
1651         }
1652
1653         /*
1654          * We fixed up user space. Now we need to fix the pi_state
1655          * itself.
1656          */
1657         if (pi_state->owner != NULL) {
1658                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1659                 WARN_ON(list_empty(&pi_state->list));
1660                 list_del_init(&pi_state->list);
1661                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1662         }
1663
1664         pi_state->owner = newowner;
1665
1666         raw_spin_lock_irq(&newowner->pi_lock);
1667         WARN_ON(!list_empty(&pi_state->list));
1668         list_add(&pi_state->list, &newowner->pi_state_list);
1669         raw_spin_unlock_irq(&newowner->pi_lock);
1670         return 0;
1671
1672         /*
1673          * To handle the page fault we need to drop the hash bucket
1674          * lock here. That gives the other task (either the highest priority
1675          * waiter itself or the task which stole the rtmutex) the
1676          * chance to try the fixup of the pi_state. So once we are
1677          * back from handling the fault we need to check the pi_state
1678          * after reacquiring the hash bucket lock and before trying to
1679          * do another fixup. When the fixup has been done already we
1680          * simply return.
1681          */
1682 handle_fault:
1683         spin_unlock(q->lock_ptr);
1684
1685         ret = fault_in_user_writeable(uaddr);
1686
1687         spin_lock(q->lock_ptr);
1688
1689         /*
1690          * Check if someone else fixed it for us:
1691          */
1692         if (pi_state->owner != oldowner)
1693                 return 0;
1694
1695         if (ret)
1696                 return ret;
1697
1698         goto retry;
1699 }
1700
1701 static long futex_wait_restart(struct restart_block *restart);
1702
1703 /**
1704  * fixup_owner() - Post lock pi_state and corner case management
1705  * @uaddr:      user address of the futex
1706  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1707  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1708  *
1709  * After attempting to lock an rt_mutex, this function is called to cleanup
1710  * the pi_state owner as well as handle race conditions that may allow us to
1711  * acquire the lock. Must be called with the hb lock held.
1712  *
1713  * Return:
1714  *  1 - success, lock taken;
1715  *  0 - success, lock not taken;
1716  * <0 - on error (-EFAULT)
1717  */
1718 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1719 {
1720         struct task_struct *owner;
1721         int ret = 0;
1722
1723         if (locked) {
1724                 /*
1725                  * Got the lock. We might not be the anticipated owner if we
1726                  * did a lock-steal - fix up the PI-state in that case:
1727                  */
1728                 if (q->pi_state->owner != current)
1729                         ret = fixup_pi_state_owner(uaddr, q, current);
1730                 goto out;
1731         }
1732
1733         /*
1734          * Catch the rare case, where the lock was released when we were on the
1735          * way back before we locked the hash bucket.
1736          */
1737         if (q->pi_state->owner == current) {
1738                 /*
1739                  * Try to get the rt_mutex now. This might fail as some other
1740                  * task acquired the rt_mutex after we removed ourself from the
1741                  * rt_mutex waiters list.
1742                  */
1743                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1744                         locked = 1;
1745                         goto out;
1746                 }
1747
1748                 /*
1749                  * pi_state is incorrect, some other task did a lock steal and
1750                  * we returned due to timeout or signal without taking the
1751                  * rt_mutex. Too late.
1752                  */
1753                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1754                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1755                 if (!owner)
1756                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1757                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1758                 ret = fixup_pi_state_owner(uaddr, q, owner);
1759                 goto out;
1760         }
1761
1762         /*
1763          * Paranoia check. If we did not take the lock, then we should not be
1764          * the owner of the rt_mutex.
1765          */
1766         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1767                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1768                                 "pi-state %p\n", ret,
1769                                 q->pi_state->pi_mutex.owner,
1770                                 q->pi_state->owner);
1771
1772 out:
1773         return ret ? ret : locked;
1774 }
1775
1776 /**
1777  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1778  * @hb:         the futex hash bucket, must be locked by the caller
1779  * @q:          the futex_q to queue up on
1780  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1781  */
1782 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1783                                 struct hrtimer_sleeper *timeout)
1784 {
1785         /*
1786          * The task state is guaranteed to be set before another task can
1787          * wake it. set_current_state() is implemented using set_mb() and
1788          * queue_me() calls spin_unlock() upon completion, both serializing
1789          * access to the hash list and forcing another memory barrier.
1790          */
1791         set_current_state(TASK_INTERRUPTIBLE);
1792         queue_me(q, hb);
1793
1794         /* Arm the timer */
1795         if (timeout) {
1796                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1797                 if (!hrtimer_active(&timeout->timer))
1798                         timeout->task = NULL;
1799         }
1800
1801         /*
1802          * If we have been removed from the hash list, then another task
1803          * has tried to wake us, and we can skip the call to schedule().
1804          */
1805         if (likely(!plist_node_empty(&q->list))) {
1806                 /*
1807                  * If the timer has already expired, current will already be
1808                  * flagged for rescheduling. Only call schedule if there
1809                  * is no timeout, or if it has yet to expire.
1810                  */
1811                 if (!timeout || timeout->task)
1812                         freezable_schedule();
1813         }
1814         __set_current_state(TASK_RUNNING);
1815 }
1816
1817 /**
1818  * futex_wait_setup() - Prepare to wait on a futex
1819  * @uaddr:      the futex userspace address
1820  * @val:        the expected value
1821  * @flags:      futex flags (FLAGS_SHARED, etc.)
1822  * @q:          the associated futex_q
1823  * @hb:         storage for hash_bucket pointer to be returned to caller
1824  *
1825  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1826  * compare it with the expected value.  Handle atomic faults internally.
1827  * Return with the hb lock held and a q.key reference on success, and unlocked
1828  * with no q.key reference on failure.
1829  *
1830  * Return:
1831  *  0 - uaddr contains val and hb has been locked;
1832  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1833  */
1834 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1835                            struct futex_q *q, struct futex_hash_bucket **hb)
1836 {
1837         u32 uval;
1838         int ret;
1839
1840         /*
1841          * Access the page AFTER the hash-bucket is locked.
1842          * Order is important:
1843          *
1844          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1845          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1846          *
1847          * The basic logical guarantee of a futex is that it blocks ONLY
1848          * if cond(var) is known to be true at the time of blocking, for
1849          * any cond.  If we locked the hash-bucket after testing *uaddr, that
1850          * would open a race condition where we could block indefinitely with
1851          * cond(var) false, which would violate the guarantee.
1852          *
1853          * On the other hand, we insert q and release the hash-bucket only
1854          * after testing *uaddr.  This guarantees that futex_wait() will NOT
1855          * absorb a wakeup if *uaddr does not match the desired values
1856          * while the syscall executes.
1857          */
1858 retry:
1859         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1860         if (unlikely(ret != 0))
1861                 return ret;
1862
1863 retry_private:
1864         *hb = queue_lock(q);
1865
1866         ret = get_futex_value_locked(&uval, uaddr);
1867
1868         if (ret) {
1869                 queue_unlock(q, *hb);
1870
1871                 ret = get_user(uval, uaddr);
1872                 if (ret)
1873                         goto out;
1874
1875                 if (!(flags & FLAGS_SHARED))
1876                         goto retry_private;
1877
1878                 put_futex_key(&q->key);
1879                 goto retry;
1880         }
1881
1882         if (uval != val) {
1883                 queue_unlock(q, *hb);
1884                 ret = -EWOULDBLOCK;
1885         }
1886
1887 out:
1888         if (ret)
1889                 put_futex_key(&q->key);
1890         return ret;
1891 }
1892
1893 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1894                       ktime_t *abs_time, u32 bitset)
1895 {
1896         struct hrtimer_sleeper timeout, *to = NULL;
1897         struct restart_block *restart;
1898         struct futex_hash_bucket *hb;
1899         struct futex_q q = futex_q_init;
1900         int ret;
1901
1902         if (!bitset)
1903                 return -EINVAL;
1904         q.bitset = bitset;
1905
1906         if (abs_time) {
1907                 to = &timeout;
1908
1909                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1910                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
1911                                       HRTIMER_MODE_ABS);
1912                 hrtimer_init_sleeper(to, current);
1913                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1914                                              current->timer_slack_ns);
1915         }
1916
1917 retry:
1918         /*
1919          * Prepare to wait on uaddr. On success, holds hb lock and increments
1920          * q.key refs.
1921          */
1922         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1923         if (ret)
1924                 goto out;
1925
1926         /* queue_me and wait for wakeup, timeout, or a signal. */
1927         futex_wait_queue_me(hb, &q, to);
1928
1929         /* If we were woken (and unqueued), we succeeded, whatever. */
1930         ret = 0;
1931         /* unqueue_me() drops q.key ref */
1932         if (!unqueue_me(&q))
1933                 goto out;
1934         ret = -ETIMEDOUT;
1935         if (to && !to->task)
1936                 goto out;
1937
1938         /*
1939          * We expect signal_pending(current), but we might be the
1940          * victim of a spurious wakeup as well.
1941          */
1942         if (!signal_pending(current))
1943                 goto retry;
1944
1945         ret = -ERESTARTSYS;
1946         if (!abs_time)
1947                 goto out;
1948
1949         restart = &current_thread_info()->restart_block;
1950         restart->fn = futex_wait_restart;
1951         restart->futex.uaddr = uaddr;
1952         restart->futex.val = val;
1953         restart->futex.time = abs_time->tv64;
1954         restart->futex.bitset = bitset;
1955         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1956
1957         ret = -ERESTART_RESTARTBLOCK;
1958
1959 out:
1960         if (to) {
1961                 hrtimer_cancel(&to->timer);
1962                 destroy_hrtimer_on_stack(&to->timer);
1963         }
1964         return ret;
1965 }
1966
1967
1968 static long futex_wait_restart(struct restart_block *restart)
1969 {
1970         u32 __user *uaddr = restart->futex.uaddr;
1971         ktime_t t, *tp = NULL;
1972
1973         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1974                 t.tv64 = restart->futex.time;
1975                 tp = &t;
1976         }
1977         restart->fn = do_no_restart_syscall;
1978
1979         return (long)futex_wait(uaddr, restart->futex.flags,
1980                                 restart->futex.val, tp, restart->futex.bitset);
1981 }
1982
1983
1984 /*
1985  * Userspace tried a 0 -> TID atomic transition of the futex value
1986  * and failed. The kernel side here does the whole locking operation:
1987  * if there are waiters then it will block, it does PI, etc. (Due to
1988  * races the kernel might see a 0 value of the futex too.)
1989  */
1990 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1991                          ktime_t *time, int trylock)
1992 {
1993         struct hrtimer_sleeper timeout, *to = NULL;
1994         struct futex_hash_bucket *hb;
1995         struct futex_q q = futex_q_init;
1996         int res, ret;
1997
1998         if (refill_pi_state_cache())
1999                 return -ENOMEM;
2000
2001         if (time) {
2002                 to = &timeout;
2003                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2004                                       HRTIMER_MODE_ABS);
2005                 hrtimer_init_sleeper(to, current);
2006                 hrtimer_set_expires(&to->timer, *time);
2007         }
2008
2009 retry:
2010         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2011         if (unlikely(ret != 0))
2012                 goto out;
2013
2014 retry_private:
2015         hb = queue_lock(&q);
2016
2017         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2018         if (unlikely(ret)) {
2019                 switch (ret) {
2020                 case 1:
2021                         /* We got the lock. */
2022                         ret = 0;
2023                         goto out_unlock_put_key;
2024                 case -EFAULT:
2025                         goto uaddr_faulted;
2026                 case -EAGAIN:
2027                         /*
2028                          * Task is exiting and we just wait for the
2029                          * exit to complete.
2030                          */
2031                         queue_unlock(&q, hb);
2032                         put_futex_key(&q.key);
2033                         cond_resched();
2034                         goto retry;
2035                 default:
2036                         goto out_unlock_put_key;
2037                 }
2038         }
2039
2040         /*
2041          * Only actually queue now that the atomic ops are done:
2042          */
2043         queue_me(&q, hb);
2044
2045         WARN_ON(!q.pi_state);
2046         /*
2047          * Block on the PI mutex:
2048          */
2049         if (!trylock)
2050                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2051         else {
2052                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2053                 /* Fixup the trylock return value: */
2054                 ret = ret ? 0 : -EWOULDBLOCK;
2055         }
2056
2057         spin_lock(q.lock_ptr);
2058         /*
2059          * Fixup the pi_state owner and possibly acquire the lock if we
2060          * haven't already.
2061          */
2062         res = fixup_owner(uaddr, &q, !ret);
2063         /*
2064          * If fixup_owner() returned an error, proprogate that.  If it acquired
2065          * the lock, clear our -ETIMEDOUT or -EINTR.
2066          */
2067         if (res)
2068                 ret = (res < 0) ? res : 0;
2069
2070         /*
2071          * If fixup_owner() faulted and was unable to handle the fault, unlock
2072          * it and return the fault to userspace.
2073          */
2074         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2075                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2076
2077         /* Unqueue and drop the lock */
2078         unqueue_me_pi(&q);
2079
2080         goto out_put_key;
2081
2082 out_unlock_put_key:
2083         queue_unlock(&q, hb);
2084
2085 out_put_key:
2086         put_futex_key(&q.key);
2087 out:
2088         if (to)
2089                 destroy_hrtimer_on_stack(&to->timer);
2090         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2091
2092 uaddr_faulted:
2093         queue_unlock(&q, hb);
2094
2095         ret = fault_in_user_writeable(uaddr);
2096         if (ret)
2097                 goto out_put_key;
2098
2099         if (!(flags & FLAGS_SHARED))
2100                 goto retry_private;
2101
2102         put_futex_key(&q.key);
2103         goto retry;
2104 }
2105
2106 /*
2107  * Userspace attempted a TID -> 0 atomic transition, and failed.
2108  * This is the in-kernel slowpath: we look up the PI state (if any),
2109  * and do the rt-mutex unlock.
2110  */
2111 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2112 {
2113         struct futex_hash_bucket *hb;
2114         struct futex_q *this, *next;
2115         struct plist_head *head;
2116         union futex_key key = FUTEX_KEY_INIT;
2117         u32 uval, vpid = task_pid_vnr(current);
2118         int ret;
2119
2120 retry:
2121         if (get_user(uval, uaddr))
2122                 return -EFAULT;
2123         /*
2124          * We release only a lock we actually own:
2125          */
2126         if ((uval & FUTEX_TID_MASK) != vpid)
2127                 return -EPERM;
2128
2129         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2130         if (unlikely(ret != 0))
2131                 goto out;
2132
2133         hb = hash_futex(&key);
2134         spin_lock(&hb->lock);
2135
2136         /*
2137          * To avoid races, try to do the TID -> 0 atomic transition
2138          * again. If it succeeds then we can return without waking
2139          * anyone else up:
2140          */
2141         if (!(uval & FUTEX_OWNER_DIED) &&
2142             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2143                 goto pi_faulted;
2144         /*
2145          * Rare case: we managed to release the lock atomically,
2146          * no need to wake anyone else up:
2147          */
2148         if (unlikely(uval == vpid))
2149                 goto out_unlock;
2150
2151         /*
2152          * Ok, other tasks may need to be woken up - check waiters
2153          * and do the wakeup if necessary:
2154          */
2155         head = &hb->chain;
2156
2157         plist_for_each_entry_safe(this, next, head, list) {
2158                 if (!match_futex (&this->key, &key))
2159                         continue;
2160                 ret = wake_futex_pi(uaddr, uval, this);
2161                 /*
2162                  * The atomic access to the futex value
2163                  * generated a pagefault, so retry the
2164                  * user-access and the wakeup:
2165                  */
2166                 if (ret == -EFAULT)
2167                         goto pi_faulted;
2168                 goto out_unlock;
2169         }
2170         /*
2171          * No waiters - kernel unlocks the futex:
2172          */
2173         if (!(uval & FUTEX_OWNER_DIED)) {
2174                 ret = unlock_futex_pi(uaddr, uval);
2175                 if (ret == -EFAULT)
2176                         goto pi_faulted;
2177         }
2178
2179 out_unlock:
2180         spin_unlock(&hb->lock);
2181         put_futex_key(&key);
2182
2183 out:
2184         return ret;
2185
2186 pi_faulted:
2187         spin_unlock(&hb->lock);
2188         put_futex_key(&key);
2189
2190         ret = fault_in_user_writeable(uaddr);
2191         if (!ret)
2192                 goto retry;
2193
2194         return ret;
2195 }
2196
2197 /**
2198  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2199  * @hb:         the hash_bucket futex_q was original enqueued on
2200  * @q:          the futex_q woken while waiting to be requeued
2201  * @key2:       the futex_key of the requeue target futex
2202  * @timeout:    the timeout associated with the wait (NULL if none)
2203  *
2204  * Detect if the task was woken on the initial futex as opposed to the requeue
2205  * target futex.  If so, determine if it was a timeout or a signal that caused
2206  * the wakeup and return the appropriate error code to the caller.  Must be
2207  * called with the hb lock held.
2208  *
2209  * Return:
2210  *  0 = no early wakeup detected;
2211  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2212  */
2213 static inline
2214 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2215                                    struct futex_q *q, union futex_key *key2,
2216                                    struct hrtimer_sleeper *timeout)
2217 {
2218         int ret = 0;
2219
2220         /*
2221          * With the hb lock held, we avoid races while we process the wakeup.
2222          * We only need to hold hb (and not hb2) to ensure atomicity as the
2223          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2224          * It can't be requeued from uaddr2 to something else since we don't
2225          * support a PI aware source futex for requeue.
2226          */
2227         if (!match_futex(&q->key, key2)) {
2228                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2229                 /*
2230                  * We were woken prior to requeue by a timeout or a signal.
2231                  * Unqueue the futex_q and determine which it was.
2232                  */
2233                 plist_del(&q->list, &hb->chain);
2234
2235                 /* Handle spurious wakeups gracefully */
2236                 ret = -EWOULDBLOCK;
2237                 if (timeout && !timeout->task)
2238                         ret = -ETIMEDOUT;
2239                 else if (signal_pending(current))
2240                         ret = -ERESTARTNOINTR;
2241         }
2242         return ret;
2243 }
2244
2245 /**
2246  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2247  * @uaddr:      the futex we initially wait on (non-pi)
2248  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2249  *              the same type, no requeueing from private to shared, etc.
2250  * @val:        the expected value of uaddr
2251  * @abs_time:   absolute timeout
2252  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2253  * @uaddr2:     the pi futex we will take prior to returning to user-space
2254  *
2255  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2256  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2257  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2258  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2259  * without one, the pi logic would not know which task to boost/deboost, if
2260  * there was a need to.
2261  *
2262  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2263  * via the following--
2264  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2265  * 2) wakeup on uaddr2 after a requeue
2266  * 3) signal
2267  * 4) timeout
2268  *
2269  * If 3, cleanup and return -ERESTARTNOINTR.
2270  *
2271  * If 2, we may then block on trying to take the rt_mutex and return via:
2272  * 5) successful lock
2273  * 6) signal
2274  * 7) timeout
2275  * 8) other lock acquisition failure
2276  *
2277  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2278  *
2279  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2280  *
2281  * Return:
2282  *  0 - On success;
2283  * <0 - On error
2284  */
2285 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2286                                  u32 val, ktime_t *abs_time, u32 bitset,
2287                                  u32 __user *uaddr2)
2288 {
2289         struct hrtimer_sleeper timeout, *to = NULL;
2290         struct rt_mutex_waiter rt_waiter;
2291         struct rt_mutex *pi_mutex = NULL;
2292         struct futex_hash_bucket *hb;
2293         union futex_key key2 = FUTEX_KEY_INIT;
2294         struct futex_q q = futex_q_init;
2295         int res, ret;
2296
2297         if (uaddr == uaddr2)
2298                 return -EINVAL;
2299
2300         if (!bitset)
2301                 return -EINVAL;
2302
2303         if (abs_time) {
2304                 to = &timeout;
2305                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2306                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2307                                       HRTIMER_MODE_ABS);
2308                 hrtimer_init_sleeper(to, current);
2309                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2310                                              current->timer_slack_ns);
2311         }
2312
2313         /*
2314          * The waiter is allocated on our stack, manipulated by the requeue
2315          * code while we sleep on uaddr.
2316          */
2317         debug_rt_mutex_init_waiter(&rt_waiter);
2318         rt_waiter.task = NULL;
2319
2320         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2321         if (unlikely(ret != 0))
2322                 goto out;
2323
2324         q.bitset = bitset;
2325         q.rt_waiter = &rt_waiter;
2326         q.requeue_pi_key = &key2;
2327
2328         /*
2329          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2330          * count.
2331          */
2332         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2333         if (ret)
2334                 goto out_key2;
2335
2336         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2337         futex_wait_queue_me(hb, &q, to);
2338
2339         spin_lock(&hb->lock);
2340         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2341         spin_unlock(&hb->lock);
2342         if (ret)
2343                 goto out_put_keys;
2344
2345         /*
2346          * In order for us to be here, we know our q.key == key2, and since
2347          * we took the hb->lock above, we also know that futex_requeue() has
2348          * completed and we no longer have to concern ourselves with a wakeup
2349          * race with the atomic proxy lock acquisition by the requeue code. The
2350          * futex_requeue dropped our key1 reference and incremented our key2
2351          * reference count.
2352          */
2353
2354         /* Check if the requeue code acquired the second futex for us. */
2355         if (!q.rt_waiter) {
2356                 /*
2357                  * Got the lock. We might not be the anticipated owner if we
2358                  * did a lock-steal - fix up the PI-state in that case.
2359                  */
2360                 if (q.pi_state && (q.pi_state->owner != current)) {
2361                         spin_lock(q.lock_ptr);
2362                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2363                         spin_unlock(q.lock_ptr);
2364                 }
2365         } else {
2366                 /*
2367                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2368                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2369                  * the pi_state.
2370                  */
2371                 WARN_ON(!q.pi_state);
2372                 pi_mutex = &q.pi_state->pi_mutex;
2373                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2374                 debug_rt_mutex_free_waiter(&rt_waiter);
2375
2376                 spin_lock(q.lock_ptr);
2377                 /*
2378                  * Fixup the pi_state owner and possibly acquire the lock if we
2379                  * haven't already.
2380                  */
2381                 res = fixup_owner(uaddr2, &q, !ret);
2382                 /*
2383                  * If fixup_owner() returned an error, proprogate that.  If it
2384                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2385                  */
2386                 if (res)
2387                         ret = (res < 0) ? res : 0;
2388
2389                 /* Unqueue and drop the lock. */
2390                 unqueue_me_pi(&q);
2391         }
2392
2393         /*
2394          * If fixup_pi_state_owner() faulted and was unable to handle the
2395          * fault, unlock the rt_mutex and return the fault to userspace.
2396          */
2397         if (ret == -EFAULT) {
2398                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2399                         rt_mutex_unlock(pi_mutex);
2400         } else if (ret == -EINTR) {
2401                 /*
2402                  * We've already been requeued, but cannot restart by calling
2403                  * futex_lock_pi() directly. We could restart this syscall, but
2404                  * it would detect that the user space "val" changed and return
2405                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2406                  * -EWOULDBLOCK directly.
2407                  */
2408                 ret = -EWOULDBLOCK;
2409         }
2410
2411 out_put_keys:
2412         put_futex_key(&q.key);
2413 out_key2:
2414         put_futex_key(&key2);
2415
2416 out:
2417         if (to) {
2418                 hrtimer_cancel(&to->timer);
2419                 destroy_hrtimer_on_stack(&to->timer);
2420         }
2421         return ret;
2422 }
2423
2424 /*
2425  * Support for robust futexes: the kernel cleans up held futexes at
2426  * thread exit time.
2427  *
2428  * Implementation: user-space maintains a per-thread list of locks it
2429  * is holding. Upon do_exit(), the kernel carefully walks this list,
2430  * and marks all locks that are owned by this thread with the
2431  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2432  * always manipulated with the lock held, so the list is private and
2433  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2434  * field, to allow the kernel to clean up if the thread dies after
2435  * acquiring the lock, but just before it could have added itself to
2436  * the list. There can only be one such pending lock.
2437  */
2438
2439 /**
2440  * sys_set_robust_list() - Set the robust-futex list head of a task
2441  * @head:       pointer to the list-head
2442  * @len:        length of the list-head, as userspace expects
2443  */
2444 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2445                 size_t, len)
2446 {
2447         if (!futex_cmpxchg_enabled)
2448                 return -ENOSYS;
2449         /*
2450          * The kernel knows only one size for now:
2451          */
2452         if (unlikely(len != sizeof(*head)))
2453                 return -EINVAL;
2454
2455         current->robust_list = head;
2456
2457         return 0;
2458 }
2459
2460 /**
2461  * sys_get_robust_list() - Get the robust-futex list head of a task
2462  * @pid:        pid of the process [zero for current task]
2463  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2464  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2465  */
2466 SYSCALL_DEFINE3(get_robust_list, int, pid,
2467                 struct robust_list_head __user * __user *, head_ptr,
2468                 size_t __user *, len_ptr)
2469 {
2470         struct robust_list_head __user *head;
2471         unsigned long ret;
2472         struct task_struct *p;
2473
2474         if (!futex_cmpxchg_enabled)
2475                 return -ENOSYS;
2476
2477         rcu_read_lock();
2478
2479         ret = -ESRCH;
2480         if (!pid)
2481                 p = current;
2482         else {
2483                 p = find_task_by_vpid(pid);
2484                 if (!p)
2485                         goto err_unlock;
2486         }
2487
2488         ret = -EPERM;
2489         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2490                 goto err_unlock;
2491
2492         head = p->robust_list;
2493         rcu_read_unlock();
2494
2495         if (put_user(sizeof(*head), len_ptr))
2496                 return -EFAULT;
2497         return put_user(head, head_ptr);
2498
2499 err_unlock:
2500         rcu_read_unlock();
2501
2502         return ret;
2503 }
2504
2505 /*
2506  * Process a futex-list entry, check whether it's owned by the
2507  * dying task, and do notification if so:
2508  */
2509 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2510 {
2511         u32 uval, uninitialized_var(nval), mval;
2512
2513 retry:
2514         if (get_user(uval, uaddr))
2515                 return -1;
2516
2517         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2518                 /*
2519                  * Ok, this dying thread is truly holding a futex
2520                  * of interest. Set the OWNER_DIED bit atomically
2521                  * via cmpxchg, and if the value had FUTEX_WAITERS
2522                  * set, wake up a waiter (if any). (We have to do a
2523                  * futex_wake() even if OWNER_DIED is already set -
2524                  * to handle the rare but possible case of recursive
2525                  * thread-death.) The rest of the cleanup is done in
2526                  * userspace.
2527                  */
2528                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2529                 /*
2530                  * We are not holding a lock here, but we want to have
2531                  * the pagefault_disable/enable() protection because
2532                  * we want to handle the fault gracefully. If the
2533                  * access fails we try to fault in the futex with R/W
2534                  * verification via get_user_pages. get_user() above
2535                  * does not guarantee R/W access. If that fails we
2536                  * give up and leave the futex locked.
2537                  */
2538                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2539                         if (fault_in_user_writeable(uaddr))
2540                                 return -1;
2541                         goto retry;
2542                 }
2543                 if (nval != uval)
2544                         goto retry;
2545
2546                 /*
2547                  * Wake robust non-PI futexes here. The wakeup of
2548                  * PI futexes happens in exit_pi_state():
2549                  */
2550                 if (!pi && (uval & FUTEX_WAITERS))
2551                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2552         }
2553         return 0;
2554 }
2555
2556 /*
2557  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2558  */
2559 static inline int fetch_robust_entry(struct robust_list __user **entry,
2560                                      struct robust_list __user * __user *head,
2561                                      unsigned int *pi)
2562 {
2563         unsigned long uentry;
2564
2565         if (get_user(uentry, (unsigned long __user *)head))
2566                 return -EFAULT;
2567
2568         *entry = (void __user *)(uentry & ~1UL);
2569         *pi = uentry & 1;
2570
2571         return 0;
2572 }
2573
2574 /*
2575  * Walk curr->robust_list (very carefully, it's a userspace list!)
2576  * and mark any locks found there dead, and notify any waiters.
2577  *
2578  * We silently return on any sign of list-walking problem.
2579  */
2580 void exit_robust_list(struct task_struct *curr)
2581 {
2582         struct robust_list_head __user *head = curr->robust_list;
2583         struct robust_list __user *entry, *next_entry, *pending;
2584         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2585         unsigned int uninitialized_var(next_pi);
2586         unsigned long futex_offset;
2587         int rc;
2588
2589         if (!futex_cmpxchg_enabled)
2590                 return;
2591
2592         /*
2593          * Fetch the list head (which was registered earlier, via
2594          * sys_set_robust_list()):
2595          */
2596         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2597                 return;
2598         /*
2599          * Fetch the relative futex offset:
2600          */
2601         if (get_user(futex_offset, &head->futex_offset))
2602                 return;
2603         /*
2604          * Fetch any possibly pending lock-add first, and handle it
2605          * if it exists:
2606          */
2607         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2608                 return;
2609
2610         next_entry = NULL;      /* avoid warning with gcc */
2611         while (entry != &head->list) {
2612                 /*
2613                  * Fetch the next entry in the list before calling
2614                  * handle_futex_death:
2615                  */
2616                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2617                 /*
2618                  * A pending lock might already be on the list, so
2619                  * don't process it twice:
2620                  */
2621                 if (entry != pending)
2622                         if (handle_futex_death((void __user *)entry + futex_offset,
2623                                                 curr, pi))
2624                                 return;
2625                 if (rc)
2626                         return;
2627                 entry = next_entry;
2628                 pi = next_pi;
2629                 /*
2630                  * Avoid excessively long or circular lists:
2631                  */
2632                 if (!--limit)
2633                         break;
2634
2635                 cond_resched();
2636         }
2637
2638         if (pending)
2639                 handle_futex_death((void __user *)pending + futex_offset,
2640                                    curr, pip);
2641 }
2642
2643 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2644                 u32 __user *uaddr2, u32 val2, u32 val3)
2645 {
2646         int cmd = op & FUTEX_CMD_MASK;
2647         unsigned int flags = 0;
2648
2649         if (!(op & FUTEX_PRIVATE_FLAG))
2650                 flags |= FLAGS_SHARED;
2651
2652         if (op & FUTEX_CLOCK_REALTIME) {
2653                 flags |= FLAGS_CLOCKRT;
2654                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2655                         return -ENOSYS;
2656         }
2657
2658         switch (cmd) {
2659         case FUTEX_LOCK_PI:
2660         case FUTEX_UNLOCK_PI:
2661         case FUTEX_TRYLOCK_PI:
2662         case FUTEX_WAIT_REQUEUE_PI:
2663         case FUTEX_CMP_REQUEUE_PI:
2664                 if (!futex_cmpxchg_enabled)
2665                         return -ENOSYS;
2666         }
2667
2668         switch (cmd) {
2669         case FUTEX_WAIT:
2670                 val3 = FUTEX_BITSET_MATCH_ANY;
2671         case FUTEX_WAIT_BITSET:
2672                 return futex_wait(uaddr, flags, val, timeout, val3);
2673         case FUTEX_WAKE:
2674                 val3 = FUTEX_BITSET_MATCH_ANY;
2675         case FUTEX_WAKE_BITSET:
2676                 return futex_wake(uaddr, flags, val, val3);
2677         case FUTEX_REQUEUE:
2678                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2679         case FUTEX_CMP_REQUEUE:
2680                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2681         case FUTEX_WAKE_OP:
2682                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2683         case FUTEX_LOCK_PI:
2684                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2685         case FUTEX_UNLOCK_PI:
2686                 return futex_unlock_pi(uaddr, flags);
2687         case FUTEX_TRYLOCK_PI:
2688                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2689         case FUTEX_WAIT_REQUEUE_PI:
2690                 val3 = FUTEX_BITSET_MATCH_ANY;
2691                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2692                                              uaddr2);
2693         case FUTEX_CMP_REQUEUE_PI:
2694                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2695         }
2696         return -ENOSYS;
2697 }
2698
2699
2700 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2701                 struct timespec __user *, utime, u32 __user *, uaddr2,
2702                 u32, val3)
2703 {
2704         struct timespec ts;
2705         ktime_t t, *tp = NULL;
2706         u32 val2 = 0;
2707         int cmd = op & FUTEX_CMD_MASK;
2708
2709         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2710                       cmd == FUTEX_WAIT_BITSET ||
2711                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2712                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2713                         return -EFAULT;
2714                 if (!timespec_valid(&ts))
2715                         return -EINVAL;
2716
2717                 t = timespec_to_ktime(ts);
2718                 if (cmd == FUTEX_WAIT)
2719                         t = ktime_add_safe(ktime_get(), t);
2720                 tp = &t;
2721         }
2722         /*
2723          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2724          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2725          */
2726         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2727             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2728                 val2 = (u32) (unsigned long) utime;
2729
2730         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2731 }
2732
2733 static int __init futex_init(void)
2734 {
2735         u32 curval;
2736         int i;
2737
2738         /*
2739          * This will fail and we want it. Some arch implementations do
2740          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2741          * functionality. We want to know that before we call in any
2742          * of the complex code paths. Also we want to prevent
2743          * registration of robust lists in that case. NULL is
2744          * guaranteed to fault and we get -EFAULT on functional
2745          * implementation, the non-functional ones will return
2746          * -ENOSYS.
2747          */
2748         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2749                 futex_cmpxchg_enabled = 1;
2750
2751         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2752                 plist_head_init(&futex_queues[i].chain);
2753                 spin_lock_init(&futex_queues[i].lock);
2754         }
2755
2756         return 0;
2757 }
2758 __initcall(futex_init);