futex: unlock before returning -EFAULT
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
58
59 #include <asm/futex.h>
60
61 #include "rtmutex_common.h"
62
63 int __read_mostly futex_cmpxchg_enabled;
64
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
66
67 /*
68  * Priority Inheritance state:
69  */
70 struct futex_pi_state {
71         /*
72          * list of 'owned' pi_state instances - these have to be
73          * cleaned up in do_exit() if the task exits prematurely:
74          */
75         struct list_head list;
76
77         /*
78          * The PI object:
79          */
80         struct rt_mutex pi_mutex;
81
82         struct task_struct *owner;
83         atomic_t refcount;
84
85         union futex_key key;
86 };
87
88 /*
89  * We use this hashed waitqueue instead of a normal wait_queue_t, so
90  * we can wake only the relevant ones (hashed queues may be shared).
91  *
92  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94  * The order of wakup is always to make the first condition true, then
95  * wake up q->waiter, then make the second condition true.
96  */
97 struct futex_q {
98         struct plist_node list;
99         /* There can only be a single waiter */
100         wait_queue_head_t waiter;
101
102         /* Which hash list lock to use: */
103         spinlock_t *lock_ptr;
104
105         /* Key which the futex is hashed on: */
106         union futex_key key;
107
108         /* Optional priority inheritance state: */
109         struct futex_pi_state *pi_state;
110         struct task_struct *task;
111
112         /* Bitset for the optional bitmasked wakeup */
113         u32 bitset;
114 };
115
116 /*
117  * Hash buckets are shared by all the futex_keys that hash to the same
118  * location.  Each key may have multiple futex_q structures, one for each task
119  * waiting on a futex.
120  */
121 struct futex_hash_bucket {
122         spinlock_t lock;
123         struct plist_head chain;
124 };
125
126 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
127
128 /*
129  * We hash on the keys returned from get_futex_key (see below).
130  */
131 static struct futex_hash_bucket *hash_futex(union futex_key *key)
132 {
133         u32 hash = jhash2((u32*)&key->both.word,
134                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
135                           key->both.offset);
136         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
137 }
138
139 /*
140  * Return 1 if two futex_keys are equal, 0 otherwise.
141  */
142 static inline int match_futex(union futex_key *key1, union futex_key *key2)
143 {
144         return (key1->both.word == key2->both.word
145                 && key1->both.ptr == key2->both.ptr
146                 && key1->both.offset == key2->both.offset);
147 }
148
149 /*
150  * Take a reference to the resource addressed by a key.
151  * Can be called while holding spinlocks.
152  *
153  */
154 static void get_futex_key_refs(union futex_key *key)
155 {
156         if (!key->both.ptr)
157                 return;
158
159         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
160         case FUT_OFF_INODE:
161                 atomic_inc(&key->shared.inode->i_count);
162                 break;
163         case FUT_OFF_MMSHARED:
164                 atomic_inc(&key->private.mm->mm_count);
165                 break;
166         }
167 }
168
169 /*
170  * Drop a reference to the resource addressed by a key.
171  * The hash bucket spinlock must not be held.
172  */
173 static void drop_futex_key_refs(union futex_key *key)
174 {
175         if (!key->both.ptr) {
176                 /* If we're here then we tried to put a key we failed to get */
177                 WARN_ON_ONCE(1);
178                 return;
179         }
180
181         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
182         case FUT_OFF_INODE:
183                 iput(key->shared.inode);
184                 break;
185         case FUT_OFF_MMSHARED:
186                 mmdrop(key->private.mm);
187                 break;
188         }
189 }
190
191 /**
192  * get_futex_key - Get parameters which are the keys for a futex.
193  * @uaddr: virtual address of the futex
194  * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
195  * @key: address where result is stored.
196  *
197  * Returns a negative error code or 0
198  * The key words are stored in *key on success.
199  *
200  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
201  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
202  * We can usually work out the index without swapping in the page.
203  *
204  * lock_page() might sleep, the caller should not hold a spinlock.
205  */
206 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
207 {
208         unsigned long address = (unsigned long)uaddr;
209         struct mm_struct *mm = current->mm;
210         struct page *page;
211         int err;
212
213         /*
214          * The futex address must be "naturally" aligned.
215          */
216         key->both.offset = address % PAGE_SIZE;
217         if (unlikely((address % sizeof(u32)) != 0))
218                 return -EINVAL;
219         address -= key->both.offset;
220
221         /*
222          * PROCESS_PRIVATE futexes are fast.
223          * As the mm cannot disappear under us and the 'key' only needs
224          * virtual address, we dont even have to find the underlying vma.
225          * Note : We do have to check 'uaddr' is a valid user address,
226          *        but access_ok() should be faster than find_vma()
227          */
228         if (!fshared) {
229                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
230                         return -EFAULT;
231                 key->private.mm = mm;
232                 key->private.address = address;
233                 get_futex_key_refs(key);
234                 return 0;
235         }
236
237 again:
238         err = get_user_pages_fast(address, 1, 0, &page);
239         if (err < 0)
240                 return err;
241
242         lock_page(page);
243         if (!page->mapping) {
244                 unlock_page(page);
245                 put_page(page);
246                 goto again;
247         }
248
249         /*
250          * Private mappings are handled in a simple way.
251          *
252          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
253          * it's a read-only handle, it's expected that futexes attach to
254          * the object not the particular process.
255          */
256         if (PageAnon(page)) {
257                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
258                 key->private.mm = mm;
259                 key->private.address = address;
260         } else {
261                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
262                 key->shared.inode = page->mapping->host;
263                 key->shared.pgoff = page->index;
264         }
265
266         get_futex_key_refs(key);
267
268         unlock_page(page);
269         put_page(page);
270         return 0;
271 }
272
273 static inline
274 void put_futex_key(int fshared, union futex_key *key)
275 {
276         drop_futex_key_refs(key);
277 }
278
279 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
280 {
281         u32 curval;
282
283         pagefault_disable();
284         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
285         pagefault_enable();
286
287         return curval;
288 }
289
290 static int get_futex_value_locked(u32 *dest, u32 __user *from)
291 {
292         int ret;
293
294         pagefault_disable();
295         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
296         pagefault_enable();
297
298         return ret ? -EFAULT : 0;
299 }
300
301 /*
302  * Fault handling.
303  */
304 static int futex_handle_fault(unsigned long address, int attempt)
305 {
306         struct vm_area_struct * vma;
307         struct mm_struct *mm = current->mm;
308         int ret = -EFAULT;
309
310         if (attempt > 2)
311                 return ret;
312
313         down_read(&mm->mmap_sem);
314         vma = find_vma(mm, address);
315         if (vma && address >= vma->vm_start &&
316             (vma->vm_flags & VM_WRITE)) {
317                 int fault;
318                 fault = handle_mm_fault(mm, vma, address, 1);
319                 if (unlikely((fault & VM_FAULT_ERROR))) {
320 #if 0
321                         /* XXX: let's do this when we verify it is OK */
322                         if (ret & VM_FAULT_OOM)
323                                 ret = -ENOMEM;
324 #endif
325                 } else {
326                         ret = 0;
327                         if (fault & VM_FAULT_MAJOR)
328                                 current->maj_flt++;
329                         else
330                                 current->min_flt++;
331                 }
332         }
333         up_read(&mm->mmap_sem);
334         return ret;
335 }
336
337 /*
338  * PI code:
339  */
340 static int refill_pi_state_cache(void)
341 {
342         struct futex_pi_state *pi_state;
343
344         if (likely(current->pi_state_cache))
345                 return 0;
346
347         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
348
349         if (!pi_state)
350                 return -ENOMEM;
351
352         INIT_LIST_HEAD(&pi_state->list);
353         /* pi_mutex gets initialized later */
354         pi_state->owner = NULL;
355         atomic_set(&pi_state->refcount, 1);
356         pi_state->key = FUTEX_KEY_INIT;
357
358         current->pi_state_cache = pi_state;
359
360         return 0;
361 }
362
363 static struct futex_pi_state * alloc_pi_state(void)
364 {
365         struct futex_pi_state *pi_state = current->pi_state_cache;
366
367         WARN_ON(!pi_state);
368         current->pi_state_cache = NULL;
369
370         return pi_state;
371 }
372
373 static void free_pi_state(struct futex_pi_state *pi_state)
374 {
375         if (!atomic_dec_and_test(&pi_state->refcount))
376                 return;
377
378         /*
379          * If pi_state->owner is NULL, the owner is most probably dying
380          * and has cleaned up the pi_state already
381          */
382         if (pi_state->owner) {
383                 spin_lock_irq(&pi_state->owner->pi_lock);
384                 list_del_init(&pi_state->list);
385                 spin_unlock_irq(&pi_state->owner->pi_lock);
386
387                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
388         }
389
390         if (current->pi_state_cache)
391                 kfree(pi_state);
392         else {
393                 /*
394                  * pi_state->list is already empty.
395                  * clear pi_state->owner.
396                  * refcount is at 0 - put it back to 1.
397                  */
398                 pi_state->owner = NULL;
399                 atomic_set(&pi_state->refcount, 1);
400                 current->pi_state_cache = pi_state;
401         }
402 }
403
404 /*
405  * Look up the task based on what TID userspace gave us.
406  * We dont trust it.
407  */
408 static struct task_struct * futex_find_get_task(pid_t pid)
409 {
410         struct task_struct *p;
411         const struct cred *cred = current_cred(), *pcred;
412
413         rcu_read_lock();
414         p = find_task_by_vpid(pid);
415         if (!p) {
416                 p = ERR_PTR(-ESRCH);
417         } else {
418                 pcred = __task_cred(p);
419                 if (cred->euid != pcred->euid &&
420                     cred->euid != pcred->uid)
421                         p = ERR_PTR(-ESRCH);
422                 else
423                         get_task_struct(p);
424         }
425
426         rcu_read_unlock();
427
428         return p;
429 }
430
431 /*
432  * This task is holding PI mutexes at exit time => bad.
433  * Kernel cleans up PI-state, but userspace is likely hosed.
434  * (Robust-futex cleanup is separate and might save the day for userspace.)
435  */
436 void exit_pi_state_list(struct task_struct *curr)
437 {
438         struct list_head *next, *head = &curr->pi_state_list;
439         struct futex_pi_state *pi_state;
440         struct futex_hash_bucket *hb;
441         union futex_key key = FUTEX_KEY_INIT;
442
443         if (!futex_cmpxchg_enabled)
444                 return;
445         /*
446          * We are a ZOMBIE and nobody can enqueue itself on
447          * pi_state_list anymore, but we have to be careful
448          * versus waiters unqueueing themselves:
449          */
450         spin_lock_irq(&curr->pi_lock);
451         while (!list_empty(head)) {
452
453                 next = head->next;
454                 pi_state = list_entry(next, struct futex_pi_state, list);
455                 key = pi_state->key;
456                 hb = hash_futex(&key);
457                 spin_unlock_irq(&curr->pi_lock);
458
459                 spin_lock(&hb->lock);
460
461                 spin_lock_irq(&curr->pi_lock);
462                 /*
463                  * We dropped the pi-lock, so re-check whether this
464                  * task still owns the PI-state:
465                  */
466                 if (head->next != next) {
467                         spin_unlock(&hb->lock);
468                         continue;
469                 }
470
471                 WARN_ON(pi_state->owner != curr);
472                 WARN_ON(list_empty(&pi_state->list));
473                 list_del_init(&pi_state->list);
474                 pi_state->owner = NULL;
475                 spin_unlock_irq(&curr->pi_lock);
476
477                 rt_mutex_unlock(&pi_state->pi_mutex);
478
479                 spin_unlock(&hb->lock);
480
481                 spin_lock_irq(&curr->pi_lock);
482         }
483         spin_unlock_irq(&curr->pi_lock);
484 }
485
486 static int
487 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
488                 union futex_key *key, struct futex_pi_state **ps)
489 {
490         struct futex_pi_state *pi_state = NULL;
491         struct futex_q *this, *next;
492         struct plist_head *head;
493         struct task_struct *p;
494         pid_t pid = uval & FUTEX_TID_MASK;
495
496         head = &hb->chain;
497
498         plist_for_each_entry_safe(this, next, head, list) {
499                 if (match_futex(&this->key, key)) {
500                         /*
501                          * Another waiter already exists - bump up
502                          * the refcount and return its pi_state:
503                          */
504                         pi_state = this->pi_state;
505                         /*
506                          * Userspace might have messed up non PI and PI futexes
507                          */
508                         if (unlikely(!pi_state))
509                                 return -EINVAL;
510
511                         WARN_ON(!atomic_read(&pi_state->refcount));
512                         WARN_ON(pid && pi_state->owner &&
513                                 pi_state->owner->pid != pid);
514
515                         atomic_inc(&pi_state->refcount);
516                         *ps = pi_state;
517
518                         return 0;
519                 }
520         }
521
522         /*
523          * We are the first waiter - try to look up the real owner and attach
524          * the new pi_state to it, but bail out when TID = 0
525          */
526         if (!pid)
527                 return -ESRCH;
528         p = futex_find_get_task(pid);
529         if (IS_ERR(p))
530                 return PTR_ERR(p);
531
532         /*
533          * We need to look at the task state flags to figure out,
534          * whether the task is exiting. To protect against the do_exit
535          * change of the task flags, we do this protected by
536          * p->pi_lock:
537          */
538         spin_lock_irq(&p->pi_lock);
539         if (unlikely(p->flags & PF_EXITING)) {
540                 /*
541                  * The task is on the way out. When PF_EXITPIDONE is
542                  * set, we know that the task has finished the
543                  * cleanup:
544                  */
545                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
546
547                 spin_unlock_irq(&p->pi_lock);
548                 put_task_struct(p);
549                 return ret;
550         }
551
552         pi_state = alloc_pi_state();
553
554         /*
555          * Initialize the pi_mutex in locked state and make 'p'
556          * the owner of it:
557          */
558         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
559
560         /* Store the key for possible exit cleanups: */
561         pi_state->key = *key;
562
563         WARN_ON(!list_empty(&pi_state->list));
564         list_add(&pi_state->list, &p->pi_state_list);
565         pi_state->owner = p;
566         spin_unlock_irq(&p->pi_lock);
567
568         put_task_struct(p);
569
570         *ps = pi_state;
571
572         return 0;
573 }
574
575 /*
576  * The hash bucket lock must be held when this is called.
577  * Afterwards, the futex_q must not be accessed.
578  */
579 static void wake_futex(struct futex_q *q)
580 {
581         plist_del(&q->list, &q->list.plist);
582         /*
583          * The lock in wake_up_all() is a crucial memory barrier after the
584          * plist_del() and also before assigning to q->lock_ptr.
585          */
586         wake_up(&q->waiter);
587         /*
588          * The waiting task can free the futex_q as soon as this is written,
589          * without taking any locks.  This must come last.
590          *
591          * A memory barrier is required here to prevent the following store to
592          * lock_ptr from getting ahead of the wakeup. Clearing the lock at the
593          * end of wake_up() does not prevent this store from moving.
594          */
595         smp_wmb();
596         q->lock_ptr = NULL;
597 }
598
599 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
600 {
601         struct task_struct *new_owner;
602         struct futex_pi_state *pi_state = this->pi_state;
603         u32 curval, newval;
604
605         if (!pi_state)
606                 return -EINVAL;
607
608         spin_lock(&pi_state->pi_mutex.wait_lock);
609         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
610
611         /*
612          * This happens when we have stolen the lock and the original
613          * pending owner did not enqueue itself back on the rt_mutex.
614          * Thats not a tragedy. We know that way, that a lock waiter
615          * is on the fly. We make the futex_q waiter the pending owner.
616          */
617         if (!new_owner)
618                 new_owner = this->task;
619
620         /*
621          * We pass it to the next owner. (The WAITERS bit is always
622          * kept enabled while there is PI state around. We must also
623          * preserve the owner died bit.)
624          */
625         if (!(uval & FUTEX_OWNER_DIED)) {
626                 int ret = 0;
627
628                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
629
630                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
631
632                 if (curval == -EFAULT)
633                         ret = -EFAULT;
634                 else if (curval != uval)
635                         ret = -EINVAL;
636                 if (ret) {
637                         spin_unlock(&pi_state->pi_mutex.wait_lock);
638                         return ret;
639                 }
640         }
641
642         spin_lock_irq(&pi_state->owner->pi_lock);
643         WARN_ON(list_empty(&pi_state->list));
644         list_del_init(&pi_state->list);
645         spin_unlock_irq(&pi_state->owner->pi_lock);
646
647         spin_lock_irq(&new_owner->pi_lock);
648         WARN_ON(!list_empty(&pi_state->list));
649         list_add(&pi_state->list, &new_owner->pi_state_list);
650         pi_state->owner = new_owner;
651         spin_unlock_irq(&new_owner->pi_lock);
652
653         spin_unlock(&pi_state->pi_mutex.wait_lock);
654         rt_mutex_unlock(&pi_state->pi_mutex);
655
656         return 0;
657 }
658
659 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
660 {
661         u32 oldval;
662
663         /*
664          * There is no waiter, so we unlock the futex. The owner died
665          * bit has not to be preserved here. We are the owner:
666          */
667         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
668
669         if (oldval == -EFAULT)
670                 return oldval;
671         if (oldval != uval)
672                 return -EAGAIN;
673
674         return 0;
675 }
676
677 /*
678  * Express the locking dependencies for lockdep:
679  */
680 static inline void
681 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
682 {
683         if (hb1 <= hb2) {
684                 spin_lock(&hb1->lock);
685                 if (hb1 < hb2)
686                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
687         } else { /* hb1 > hb2 */
688                 spin_lock(&hb2->lock);
689                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
690         }
691 }
692
693 static inline void
694 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
695 {
696         if (hb1 <= hb2) {
697                 spin_unlock(&hb2->lock);
698                 if (hb1 < hb2)
699                         spin_unlock(&hb1->lock);
700         } else { /* hb1 > hb2 */
701                 spin_unlock(&hb1->lock);
702                 spin_unlock(&hb2->lock);
703         }
704 }
705
706 /*
707  * Wake up waiters matching bitset queued on this futex (uaddr).
708  */
709 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
710 {
711         struct futex_hash_bucket *hb;
712         struct futex_q *this, *next;
713         struct plist_head *head;
714         union futex_key key = FUTEX_KEY_INIT;
715         int ret;
716
717         if (!bitset)
718                 return -EINVAL;
719
720         ret = get_futex_key(uaddr, fshared, &key);
721         if (unlikely(ret != 0))
722                 goto out;
723
724         hb = hash_futex(&key);
725         spin_lock(&hb->lock);
726         head = &hb->chain;
727
728         plist_for_each_entry_safe(this, next, head, list) {
729                 if (match_futex (&this->key, &key)) {
730                         if (this->pi_state) {
731                                 ret = -EINVAL;
732                                 break;
733                         }
734
735                         /* Check if one of the bits is set in both bitsets */
736                         if (!(this->bitset & bitset))
737                                 continue;
738
739                         wake_futex(this);
740                         if (++ret >= nr_wake)
741                                 break;
742                 }
743         }
744
745         spin_unlock(&hb->lock);
746         put_futex_key(fshared, &key);
747 out:
748         return ret;
749 }
750
751 /*
752  * Wake up all waiters hashed on the physical page that is mapped
753  * to this virtual address:
754  */
755 static int
756 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
757               int nr_wake, int nr_wake2, int op)
758 {
759         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
760         struct futex_hash_bucket *hb1, *hb2;
761         struct plist_head *head;
762         struct futex_q *this, *next;
763         int ret, op_ret, attempt = 0;
764
765 retryfull:
766         ret = get_futex_key(uaddr1, fshared, &key1);
767         if (unlikely(ret != 0))
768                 goto out;
769         ret = get_futex_key(uaddr2, fshared, &key2);
770         if (unlikely(ret != 0))
771                 goto out_put_key1;
772
773         hb1 = hash_futex(&key1);
774         hb2 = hash_futex(&key2);
775
776 retry:
777         double_lock_hb(hb1, hb2);
778
779         op_ret = futex_atomic_op_inuser(op, uaddr2);
780         if (unlikely(op_ret < 0)) {
781                 u32 dummy;
782
783                 double_unlock_hb(hb1, hb2);
784
785 #ifndef CONFIG_MMU
786                 /*
787                  * we don't get EFAULT from MMU faults if we don't have an MMU,
788                  * but we might get them from range checking
789                  */
790                 ret = op_ret;
791                 goto out_put_keys;
792 #endif
793
794                 if (unlikely(op_ret != -EFAULT)) {
795                         ret = op_ret;
796                         goto out_put_keys;
797                 }
798
799                 /*
800                  * futex_atomic_op_inuser needs to both read and write
801                  * *(int __user *)uaddr2, but we can't modify it
802                  * non-atomically.  Therefore, if get_user below is not
803                  * enough, we need to handle the fault ourselves, while
804                  * still holding the mmap_sem.
805                  */
806                 if (attempt++) {
807                         ret = futex_handle_fault((unsigned long)uaddr2,
808                                                  attempt);
809                         if (ret)
810                                 goto out_put_keys;
811                         goto retry;
812                 }
813
814                 ret = get_user(dummy, uaddr2);
815                 if (ret)
816                         goto out_put_keys;
817
818                 put_futex_key(fshared, &key2);
819                 put_futex_key(fshared, &key1);
820                 goto retryfull;
821         }
822
823         head = &hb1->chain;
824
825         plist_for_each_entry_safe(this, next, head, list) {
826                 if (match_futex (&this->key, &key1)) {
827                         wake_futex(this);
828                         if (++ret >= nr_wake)
829                                 break;
830                 }
831         }
832
833         if (op_ret > 0) {
834                 head = &hb2->chain;
835
836                 op_ret = 0;
837                 plist_for_each_entry_safe(this, next, head, list) {
838                         if (match_futex (&this->key, &key2)) {
839                                 wake_futex(this);
840                                 if (++op_ret >= nr_wake2)
841                                         break;
842                         }
843                 }
844                 ret += op_ret;
845         }
846
847         double_unlock_hb(hb1, hb2);
848 out_put_keys:
849         put_futex_key(fshared, &key2);
850 out_put_key1:
851         put_futex_key(fshared, &key1);
852 out:
853         return ret;
854 }
855
856 /*
857  * Requeue all waiters hashed on one physical page to another
858  * physical page.
859  */
860 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
861                          int nr_wake, int nr_requeue, u32 *cmpval)
862 {
863         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
864         struct futex_hash_bucket *hb1, *hb2;
865         struct plist_head *head1;
866         struct futex_q *this, *next;
867         int ret, drop_count = 0;
868
869 retry:
870         ret = get_futex_key(uaddr1, fshared, &key1);
871         if (unlikely(ret != 0))
872                 goto out;
873         ret = get_futex_key(uaddr2, fshared, &key2);
874         if (unlikely(ret != 0))
875                 goto out_put_key1;
876
877         hb1 = hash_futex(&key1);
878         hb2 = hash_futex(&key2);
879
880         double_lock_hb(hb1, hb2);
881
882         if (likely(cmpval != NULL)) {
883                 u32 curval;
884
885                 ret = get_futex_value_locked(&curval, uaddr1);
886
887                 if (unlikely(ret)) {
888                         double_unlock_hb(hb1, hb2);
889
890                         put_futex_key(fshared, &key2);
891                         put_futex_key(fshared, &key1);
892
893                         ret = get_user(curval, uaddr1);
894
895                         if (!ret)
896                                 goto retry;
897
898                         goto out_put_keys;
899                 }
900                 if (curval != *cmpval) {
901                         ret = -EAGAIN;
902                         goto out_unlock;
903                 }
904         }
905
906         head1 = &hb1->chain;
907         plist_for_each_entry_safe(this, next, head1, list) {
908                 if (!match_futex (&this->key, &key1))
909                         continue;
910                 if (++ret <= nr_wake) {
911                         wake_futex(this);
912                 } else {
913                         /*
914                          * If key1 and key2 hash to the same bucket, no need to
915                          * requeue.
916                          */
917                         if (likely(head1 != &hb2->chain)) {
918                                 plist_del(&this->list, &hb1->chain);
919                                 plist_add(&this->list, &hb2->chain);
920                                 this->lock_ptr = &hb2->lock;
921 #ifdef CONFIG_DEBUG_PI_LIST
922                                 this->list.plist.lock = &hb2->lock;
923 #endif
924                         }
925                         this->key = key2;
926                         get_futex_key_refs(&key2);
927                         drop_count++;
928
929                         if (ret - nr_wake >= nr_requeue)
930                                 break;
931                 }
932         }
933
934 out_unlock:
935         double_unlock_hb(hb1, hb2);
936
937         /* drop_futex_key_refs() must be called outside the spinlocks. */
938         while (--drop_count >= 0)
939                 drop_futex_key_refs(&key1);
940
941 out_put_keys:
942         put_futex_key(fshared, &key2);
943 out_put_key1:
944         put_futex_key(fshared, &key1);
945 out:
946         return ret;
947 }
948
949 /* The key must be already stored in q->key. */
950 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
951 {
952         struct futex_hash_bucket *hb;
953
954         init_waitqueue_head(&q->waiter);
955
956         get_futex_key_refs(&q->key);
957         hb = hash_futex(&q->key);
958         q->lock_ptr = &hb->lock;
959
960         spin_lock(&hb->lock);
961         return hb;
962 }
963
964 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
965 {
966         int prio;
967
968         /*
969          * The priority used to register this element is
970          * - either the real thread-priority for the real-time threads
971          * (i.e. threads with a priority lower than MAX_RT_PRIO)
972          * - or MAX_RT_PRIO for non-RT threads.
973          * Thus, all RT-threads are woken first in priority order, and
974          * the others are woken last, in FIFO order.
975          */
976         prio = min(current->normal_prio, MAX_RT_PRIO);
977
978         plist_node_init(&q->list, prio);
979 #ifdef CONFIG_DEBUG_PI_LIST
980         q->list.plist.lock = &hb->lock;
981 #endif
982         plist_add(&q->list, &hb->chain);
983         q->task = current;
984         spin_unlock(&hb->lock);
985 }
986
987 static inline void
988 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
989 {
990         spin_unlock(&hb->lock);
991         drop_futex_key_refs(&q->key);
992 }
993
994 /*
995  * queue_me and unqueue_me must be called as a pair, each
996  * exactly once.  They are called with the hashed spinlock held.
997  */
998
999 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1000 static int unqueue_me(struct futex_q *q)
1001 {
1002         spinlock_t *lock_ptr;
1003         int ret = 0;
1004
1005         /* In the common case we don't take the spinlock, which is nice. */
1006 retry:
1007         lock_ptr = q->lock_ptr;
1008         barrier();
1009         if (lock_ptr != NULL) {
1010                 spin_lock(lock_ptr);
1011                 /*
1012                  * q->lock_ptr can change between reading it and
1013                  * spin_lock(), causing us to take the wrong lock.  This
1014                  * corrects the race condition.
1015                  *
1016                  * Reasoning goes like this: if we have the wrong lock,
1017                  * q->lock_ptr must have changed (maybe several times)
1018                  * between reading it and the spin_lock().  It can
1019                  * change again after the spin_lock() but only if it was
1020                  * already changed before the spin_lock().  It cannot,
1021                  * however, change back to the original value.  Therefore
1022                  * we can detect whether we acquired the correct lock.
1023                  */
1024                 if (unlikely(lock_ptr != q->lock_ptr)) {
1025                         spin_unlock(lock_ptr);
1026                         goto retry;
1027                 }
1028                 WARN_ON(plist_node_empty(&q->list));
1029                 plist_del(&q->list, &q->list.plist);
1030
1031                 BUG_ON(q->pi_state);
1032
1033                 spin_unlock(lock_ptr);
1034                 ret = 1;
1035         }
1036
1037         drop_futex_key_refs(&q->key);
1038         return ret;
1039 }
1040
1041 /*
1042  * PI futexes can not be requeued and must remove themself from the
1043  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1044  * and dropped here.
1045  */
1046 static void unqueue_me_pi(struct futex_q *q)
1047 {
1048         WARN_ON(plist_node_empty(&q->list));
1049         plist_del(&q->list, &q->list.plist);
1050
1051         BUG_ON(!q->pi_state);
1052         free_pi_state(q->pi_state);
1053         q->pi_state = NULL;
1054
1055         spin_unlock(q->lock_ptr);
1056
1057         drop_futex_key_refs(&q->key);
1058 }
1059
1060 /*
1061  * Fixup the pi_state owner with the new owner.
1062  *
1063  * Must be called with hash bucket lock held and mm->sem held for non
1064  * private futexes.
1065  */
1066 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1067                                 struct task_struct *newowner, int fshared)
1068 {
1069         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1070         struct futex_pi_state *pi_state = q->pi_state;
1071         struct task_struct *oldowner = pi_state->owner;
1072         u32 uval, curval, newval;
1073         int ret, attempt = 0;
1074
1075         /* Owner died? */
1076         if (!pi_state->owner)
1077                 newtid |= FUTEX_OWNER_DIED;
1078
1079         /*
1080          * We are here either because we stole the rtmutex from the
1081          * pending owner or we are the pending owner which failed to
1082          * get the rtmutex. We have to replace the pending owner TID
1083          * in the user space variable. This must be atomic as we have
1084          * to preserve the owner died bit here.
1085          *
1086          * Note: We write the user space value _before_ changing the pi_state
1087          * because we can fault here. Imagine swapped out pages or a fork
1088          * that marked all the anonymous memory readonly for cow.
1089          *
1090          * Modifying pi_state _before_ the user space value would
1091          * leave the pi_state in an inconsistent state when we fault
1092          * here, because we need to drop the hash bucket lock to
1093          * handle the fault. This might be observed in the PID check
1094          * in lookup_pi_state.
1095          */
1096 retry:
1097         if (get_futex_value_locked(&uval, uaddr))
1098                 goto handle_fault;
1099
1100         while (1) {
1101                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1102
1103                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1104
1105                 if (curval == -EFAULT)
1106                         goto handle_fault;
1107                 if (curval == uval)
1108                         break;
1109                 uval = curval;
1110         }
1111
1112         /*
1113          * We fixed up user space. Now we need to fix the pi_state
1114          * itself.
1115          */
1116         if (pi_state->owner != NULL) {
1117                 spin_lock_irq(&pi_state->owner->pi_lock);
1118                 WARN_ON(list_empty(&pi_state->list));
1119                 list_del_init(&pi_state->list);
1120                 spin_unlock_irq(&pi_state->owner->pi_lock);
1121         }
1122
1123         pi_state->owner = newowner;
1124
1125         spin_lock_irq(&newowner->pi_lock);
1126         WARN_ON(!list_empty(&pi_state->list));
1127         list_add(&pi_state->list, &newowner->pi_state_list);
1128         spin_unlock_irq(&newowner->pi_lock);
1129         return 0;
1130
1131         /*
1132          * To handle the page fault we need to drop the hash bucket
1133          * lock here. That gives the other task (either the pending
1134          * owner itself or the task which stole the rtmutex) the
1135          * chance to try the fixup of the pi_state. So once we are
1136          * back from handling the fault we need to check the pi_state
1137          * after reacquiring the hash bucket lock and before trying to
1138          * do another fixup. When the fixup has been done already we
1139          * simply return.
1140          */
1141 handle_fault:
1142         spin_unlock(q->lock_ptr);
1143
1144         ret = futex_handle_fault((unsigned long)uaddr, attempt++);
1145
1146         spin_lock(q->lock_ptr);
1147
1148         /*
1149          * Check if someone else fixed it for us:
1150          */
1151         if (pi_state->owner != oldowner)
1152                 return 0;
1153
1154         if (ret)
1155                 return ret;
1156
1157         goto retry;
1158 }
1159
1160 /*
1161  * In case we must use restart_block to restart a futex_wait,
1162  * we encode in the 'flags' shared capability
1163  */
1164 #define FLAGS_SHARED            0x01
1165 #define FLAGS_CLOCKRT           0x02
1166
1167 static long futex_wait_restart(struct restart_block *restart);
1168
1169 static int futex_wait(u32 __user *uaddr, int fshared,
1170                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1171 {
1172         struct task_struct *curr = current;
1173         struct restart_block *restart;
1174         DECLARE_WAITQUEUE(wait, curr);
1175         struct futex_hash_bucket *hb;
1176         struct futex_q q;
1177         u32 uval;
1178         int ret;
1179         struct hrtimer_sleeper t;
1180         int rem = 0;
1181
1182         if (!bitset)
1183                 return -EINVAL;
1184
1185         q.pi_state = NULL;
1186         q.bitset = bitset;
1187 retry:
1188         q.key = FUTEX_KEY_INIT;
1189         ret = get_futex_key(uaddr, fshared, &q.key);
1190         if (unlikely(ret != 0))
1191                 goto out;
1192
1193         hb = queue_lock(&q);
1194
1195         /*
1196          * Access the page AFTER the hash-bucket is locked.
1197          * Order is important:
1198          *
1199          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1200          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1201          *
1202          * The basic logical guarantee of a futex is that it blocks ONLY
1203          * if cond(var) is known to be true at the time of blocking, for
1204          * any cond.  If we queued after testing *uaddr, that would open
1205          * a race condition where we could block indefinitely with
1206          * cond(var) false, which would violate the guarantee.
1207          *
1208          * A consequence is that futex_wait() can return zero and absorb
1209          * a wakeup when *uaddr != val on entry to the syscall.  This is
1210          * rare, but normal.
1211          *
1212          * For shared futexes, we hold the mmap semaphore, so the mapping
1213          * cannot have changed since we looked it up in get_futex_key.
1214          */
1215         ret = get_futex_value_locked(&uval, uaddr);
1216
1217         if (unlikely(ret)) {
1218                 queue_unlock(&q, hb);
1219                 put_futex_key(fshared, &q.key);
1220
1221                 ret = get_user(uval, uaddr);
1222
1223                 if (!ret)
1224                         goto retry;
1225                 goto out;
1226         }
1227         ret = -EWOULDBLOCK;
1228         if (unlikely(uval != val)) {
1229                 queue_unlock(&q, hb);
1230                 goto out_put_key;
1231         }
1232
1233         /* Only actually queue if *uaddr contained val.  */
1234         queue_me(&q, hb);
1235
1236         /*
1237          * There might have been scheduling since the queue_me(), as we
1238          * cannot hold a spinlock across the get_user() in case it
1239          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1240          * queueing ourselves into the futex hash.  This code thus has to
1241          * rely on the futex_wake() code removing us from hash when it
1242          * wakes us up.
1243          */
1244
1245         /* add_wait_queue is the barrier after __set_current_state. */
1246         __set_current_state(TASK_INTERRUPTIBLE);
1247         add_wait_queue(&q.waiter, &wait);
1248         /*
1249          * !plist_node_empty() is safe here without any lock.
1250          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1251          */
1252         if (likely(!plist_node_empty(&q.list))) {
1253                 if (!abs_time)
1254                         schedule();
1255                 else {
1256                         hrtimer_init_on_stack(&t.timer,
1257                                               clockrt ? CLOCK_REALTIME :
1258                                               CLOCK_MONOTONIC,
1259                                               HRTIMER_MODE_ABS);
1260                         hrtimer_init_sleeper(&t, current);
1261                         hrtimer_set_expires_range_ns(&t.timer, *abs_time,
1262                                                      current->timer_slack_ns);
1263
1264                         hrtimer_start_expires(&t.timer, HRTIMER_MODE_ABS);
1265                         if (!hrtimer_active(&t.timer))
1266                                 t.task = NULL;
1267
1268                         /*
1269                          * the timer could have already expired, in which
1270                          * case current would be flagged for rescheduling.
1271                          * Don't bother calling schedule.
1272                          */
1273                         if (likely(t.task))
1274                                 schedule();
1275
1276                         hrtimer_cancel(&t.timer);
1277
1278                         /* Flag if a timeout occured */
1279                         rem = (t.task == NULL);
1280
1281                         destroy_hrtimer_on_stack(&t.timer);
1282                 }
1283         }
1284         __set_current_state(TASK_RUNNING);
1285
1286         /*
1287          * NOTE: we don't remove ourselves from the waitqueue because
1288          * we are the only user of it.
1289          */
1290
1291         /* If we were woken (and unqueued), we succeeded, whatever. */
1292         ret = 0;
1293         if (!unqueue_me(&q))
1294                 goto out_put_key;
1295         ret = -ETIMEDOUT;
1296         if (rem)
1297                 goto out_put_key;
1298
1299         /*
1300          * We expect signal_pending(current), but another thread may
1301          * have handled it for us already.
1302          */
1303         ret = -ERESTARTSYS;
1304         if (!abs_time)
1305                 goto out_put_key;
1306
1307         restart = &current_thread_info()->restart_block;
1308         restart->fn = futex_wait_restart;
1309         restart->futex.uaddr = (u32 *)uaddr;
1310         restart->futex.val = val;
1311         restart->futex.time = abs_time->tv64;
1312         restart->futex.bitset = bitset;
1313         restart->futex.flags = 0;
1314
1315         if (fshared)
1316                 restart->futex.flags |= FLAGS_SHARED;
1317         if (clockrt)
1318                 restart->futex.flags |= FLAGS_CLOCKRT;
1319
1320         ret = -ERESTART_RESTARTBLOCK;
1321
1322 out_put_key:
1323         put_futex_key(fshared, &q.key);
1324 out:
1325         return ret;
1326 }
1327
1328
1329 static long futex_wait_restart(struct restart_block *restart)
1330 {
1331         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1332         int fshared = 0;
1333         ktime_t t;
1334
1335         t.tv64 = restart->futex.time;
1336         restart->fn = do_no_restart_syscall;
1337         if (restart->futex.flags & FLAGS_SHARED)
1338                 fshared = 1;
1339         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1340                                 restart->futex.bitset,
1341                                 restart->futex.flags & FLAGS_CLOCKRT);
1342 }
1343
1344
1345 /*
1346  * Userspace tried a 0 -> TID atomic transition of the futex value
1347  * and failed. The kernel side here does the whole locking operation:
1348  * if there are waiters then it will block, it does PI, etc. (Due to
1349  * races the kernel might see a 0 value of the futex too.)
1350  */
1351 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1352                          int detect, ktime_t *time, int trylock)
1353 {
1354         struct hrtimer_sleeper timeout, *to = NULL;
1355         struct task_struct *curr = current;
1356         struct futex_hash_bucket *hb;
1357         u32 uval, newval, curval;
1358         struct futex_q q;
1359         int ret, lock_taken, ownerdied = 0, attempt = 0;
1360
1361         if (refill_pi_state_cache())
1362                 return -ENOMEM;
1363
1364         if (time) {
1365                 to = &timeout;
1366                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1367                                       HRTIMER_MODE_ABS);
1368                 hrtimer_init_sleeper(to, current);
1369                 hrtimer_set_expires(&to->timer, *time);
1370         }
1371
1372         q.pi_state = NULL;
1373 retry:
1374         q.key = FUTEX_KEY_INIT;
1375         ret = get_futex_key(uaddr, fshared, &q.key);
1376         if (unlikely(ret != 0))
1377                 goto out;
1378
1379 retry_unlocked:
1380         hb = queue_lock(&q);
1381
1382 retry_locked:
1383         ret = lock_taken = 0;
1384
1385         /*
1386          * To avoid races, we attempt to take the lock here again
1387          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1388          * the locks. It will most likely not succeed.
1389          */
1390         newval = task_pid_vnr(current);
1391
1392         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1393
1394         if (unlikely(curval == -EFAULT))
1395                 goto uaddr_faulted;
1396
1397         /*
1398          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1399          * situation and we return success to user space.
1400          */
1401         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1402                 ret = -EDEADLK;
1403                 goto out_unlock_put_key;
1404         }
1405
1406         /*
1407          * Surprise - we got the lock. Just return to userspace:
1408          */
1409         if (unlikely(!curval))
1410                 goto out_unlock_put_key;
1411
1412         uval = curval;
1413
1414         /*
1415          * Set the WAITERS flag, so the owner will know it has someone
1416          * to wake at next unlock
1417          */
1418         newval = curval | FUTEX_WAITERS;
1419
1420         /*
1421          * There are two cases, where a futex might have no owner (the
1422          * owner TID is 0): OWNER_DIED. We take over the futex in this
1423          * case. We also do an unconditional take over, when the owner
1424          * of the futex died.
1425          *
1426          * This is safe as we are protected by the hash bucket lock !
1427          */
1428         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1429                 /* Keep the OWNER_DIED bit */
1430                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1431                 ownerdied = 0;
1432                 lock_taken = 1;
1433         }
1434
1435         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1436
1437         if (unlikely(curval == -EFAULT))
1438                 goto uaddr_faulted;
1439         if (unlikely(curval != uval))
1440                 goto retry_locked;
1441
1442         /*
1443          * We took the lock due to owner died take over.
1444          */
1445         if (unlikely(lock_taken))
1446                 goto out_unlock_put_key;
1447
1448         /*
1449          * We dont have the lock. Look up the PI state (or create it if
1450          * we are the first waiter):
1451          */
1452         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1453
1454         if (unlikely(ret)) {
1455                 switch (ret) {
1456
1457                 case -EAGAIN:
1458                         /*
1459                          * Task is exiting and we just wait for the
1460                          * exit to complete.
1461                          */
1462                         queue_unlock(&q, hb);
1463                         put_futex_key(fshared, &q.key);
1464                         cond_resched();
1465                         goto retry;
1466
1467                 case -ESRCH:
1468                         /*
1469                          * No owner found for this futex. Check if the
1470                          * OWNER_DIED bit is set to figure out whether
1471                          * this is a robust futex or not.
1472                          */
1473                         if (get_futex_value_locked(&curval, uaddr))
1474                                 goto uaddr_faulted;
1475
1476                         /*
1477                          * We simply start over in case of a robust
1478                          * futex. The code above will take the futex
1479                          * and return happy.
1480                          */
1481                         if (curval & FUTEX_OWNER_DIED) {
1482                                 ownerdied = 1;
1483                                 goto retry_locked;
1484                         }
1485                 default:
1486                         goto out_unlock_put_key;
1487                 }
1488         }
1489
1490         /*
1491          * Only actually queue now that the atomic ops are done:
1492          */
1493         queue_me(&q, hb);
1494
1495         WARN_ON(!q.pi_state);
1496         /*
1497          * Block on the PI mutex:
1498          */
1499         if (!trylock)
1500                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1501         else {
1502                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1503                 /* Fixup the trylock return value: */
1504                 ret = ret ? 0 : -EWOULDBLOCK;
1505         }
1506
1507         spin_lock(q.lock_ptr);
1508
1509         if (!ret) {
1510                 /*
1511                  * Got the lock. We might not be the anticipated owner
1512                  * if we did a lock-steal - fix up the PI-state in
1513                  * that case:
1514                  */
1515                 if (q.pi_state->owner != curr)
1516                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1517         } else {
1518                 /*
1519                  * Catch the rare case, where the lock was released
1520                  * when we were on the way back before we locked the
1521                  * hash bucket.
1522                  */
1523                 if (q.pi_state->owner == curr) {
1524                         /*
1525                          * Try to get the rt_mutex now. This might
1526                          * fail as some other task acquired the
1527                          * rt_mutex after we removed ourself from the
1528                          * rt_mutex waiters list.
1529                          */
1530                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1531                                 ret = 0;
1532                         else {
1533                                 /*
1534                                  * pi_state is incorrect, some other
1535                                  * task did a lock steal and we
1536                                  * returned due to timeout or signal
1537                                  * without taking the rt_mutex. Too
1538                                  * late. We can access the
1539                                  * rt_mutex_owner without locking, as
1540                                  * the other task is now blocked on
1541                                  * the hash bucket lock. Fix the state
1542                                  * up.
1543                                  */
1544                                 struct task_struct *owner;
1545                                 int res;
1546
1547                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1548                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1549                                                            fshared);
1550
1551                                 /* propagate -EFAULT, if the fixup failed */
1552                                 if (res)
1553                                         ret = res;
1554                         }
1555                 } else {
1556                         /*
1557                          * Paranoia check. If we did not take the lock
1558                          * in the trylock above, then we should not be
1559                          * the owner of the rtmutex, neither the real
1560                          * nor the pending one:
1561                          */
1562                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1563                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1564                                        "pi-mutex: %p pi-state %p\n", ret,
1565                                        q.pi_state->pi_mutex.owner,
1566                                        q.pi_state->owner);
1567                 }
1568         }
1569
1570         /*
1571          * If fixup_pi_state_owner() faulted and was unable to handle the
1572          * fault, unlock it and return the fault to userspace.
1573          */
1574         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1575                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1576
1577         /* Unqueue and drop the lock */
1578         unqueue_me_pi(&q);
1579
1580         if (to)
1581                 destroy_hrtimer_on_stack(&to->timer);
1582         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1583
1584 out_unlock_put_key:
1585         queue_unlock(&q, hb);
1586
1587 out_put_key:
1588         put_futex_key(fshared, &q.key);
1589 out:
1590         if (to)
1591                 destroy_hrtimer_on_stack(&to->timer);
1592         return ret;
1593
1594 uaddr_faulted:
1595         /*
1596          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1597          * atomically.  Therefore, if we continue to fault after get_user()
1598          * below, we need to handle the fault ourselves, while still holding
1599          * the mmap_sem.  This can occur if the uaddr is under contention as
1600          * we have to drop the mmap_sem in order to call get_user().
1601          */
1602         queue_unlock(&q, hb);
1603
1604         if (attempt++) {
1605                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1606                 if (ret)
1607                         goto out_put_key;
1608                 goto retry_unlocked;
1609         }
1610
1611         ret = get_user(uval, uaddr);
1612         if (!ret)
1613                 goto retry_unlocked;
1614
1615         goto out_put_key;
1616 }
1617
1618
1619 /*
1620  * Userspace attempted a TID -> 0 atomic transition, and failed.
1621  * This is the in-kernel slowpath: we look up the PI state (if any),
1622  * and do the rt-mutex unlock.
1623  */
1624 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1625 {
1626         struct futex_hash_bucket *hb;
1627         struct futex_q *this, *next;
1628         u32 uval;
1629         struct plist_head *head;
1630         union futex_key key = FUTEX_KEY_INIT;
1631         int ret, attempt = 0;
1632
1633 retry:
1634         if (get_user(uval, uaddr))
1635                 return -EFAULT;
1636         /*
1637          * We release only a lock we actually own:
1638          */
1639         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1640                 return -EPERM;
1641
1642         ret = get_futex_key(uaddr, fshared, &key);
1643         if (unlikely(ret != 0))
1644                 goto out;
1645
1646         hb = hash_futex(&key);
1647 retry_unlocked:
1648         spin_lock(&hb->lock);
1649
1650         /*
1651          * To avoid races, try to do the TID -> 0 atomic transition
1652          * again. If it succeeds then we can return without waking
1653          * anyone else up:
1654          */
1655         if (!(uval & FUTEX_OWNER_DIED))
1656                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1657
1658
1659         if (unlikely(uval == -EFAULT))
1660                 goto pi_faulted;
1661         /*
1662          * Rare case: we managed to release the lock atomically,
1663          * no need to wake anyone else up:
1664          */
1665         if (unlikely(uval == task_pid_vnr(current)))
1666                 goto out_unlock;
1667
1668         /*
1669          * Ok, other tasks may need to be woken up - check waiters
1670          * and do the wakeup if necessary:
1671          */
1672         head = &hb->chain;
1673
1674         plist_for_each_entry_safe(this, next, head, list) {
1675                 if (!match_futex (&this->key, &key))
1676                         continue;
1677                 ret = wake_futex_pi(uaddr, uval, this);
1678                 /*
1679                  * The atomic access to the futex value
1680                  * generated a pagefault, so retry the
1681                  * user-access and the wakeup:
1682                  */
1683                 if (ret == -EFAULT)
1684                         goto pi_faulted;
1685                 goto out_unlock;
1686         }
1687         /*
1688          * No waiters - kernel unlocks the futex:
1689          */
1690         if (!(uval & FUTEX_OWNER_DIED)) {
1691                 ret = unlock_futex_pi(uaddr, uval);
1692                 if (ret == -EFAULT)
1693                         goto pi_faulted;
1694         }
1695
1696 out_unlock:
1697         spin_unlock(&hb->lock);
1698         put_futex_key(fshared, &key);
1699
1700 out:
1701         return ret;
1702
1703 pi_faulted:
1704         /*
1705          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1706          * atomically.  Therefore, if we continue to fault after get_user()
1707          * below, we need to handle the fault ourselves, while still holding
1708          * the mmap_sem.  This can occur if the uaddr is under contention as
1709          * we have to drop the mmap_sem in order to call get_user().
1710          */
1711         spin_unlock(&hb->lock);
1712
1713         if (attempt++) {
1714                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1715                 if (ret)
1716                         goto out;
1717                 uval = 0;
1718                 goto retry_unlocked;
1719         }
1720
1721         ret = get_user(uval, uaddr);
1722         put_futex_key(fshared, &key);
1723         if (!ret)
1724                 goto retry;
1725
1726         return ret;
1727 }
1728
1729 /*
1730  * Support for robust futexes: the kernel cleans up held futexes at
1731  * thread exit time.
1732  *
1733  * Implementation: user-space maintains a per-thread list of locks it
1734  * is holding. Upon do_exit(), the kernel carefully walks this list,
1735  * and marks all locks that are owned by this thread with the
1736  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1737  * always manipulated with the lock held, so the list is private and
1738  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1739  * field, to allow the kernel to clean up if the thread dies after
1740  * acquiring the lock, but just before it could have added itself to
1741  * the list. There can only be one such pending lock.
1742  */
1743
1744 /**
1745  * sys_set_robust_list - set the robust-futex list head of a task
1746  * @head: pointer to the list-head
1747  * @len: length of the list-head, as userspace expects
1748  */
1749 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
1750                 size_t, len)
1751 {
1752         if (!futex_cmpxchg_enabled)
1753                 return -ENOSYS;
1754         /*
1755          * The kernel knows only one size for now:
1756          */
1757         if (unlikely(len != sizeof(*head)))
1758                 return -EINVAL;
1759
1760         current->robust_list = head;
1761
1762         return 0;
1763 }
1764
1765 /**
1766  * sys_get_robust_list - get the robust-futex list head of a task
1767  * @pid: pid of the process [zero for current task]
1768  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1769  * @len_ptr: pointer to a length field, the kernel fills in the header size
1770  */
1771 SYSCALL_DEFINE3(get_robust_list, int, pid,
1772                 struct robust_list_head __user * __user *, head_ptr,
1773                 size_t __user *, len_ptr)
1774 {
1775         struct robust_list_head __user *head;
1776         unsigned long ret;
1777         const struct cred *cred = current_cred(), *pcred;
1778
1779         if (!futex_cmpxchg_enabled)
1780                 return -ENOSYS;
1781
1782         if (!pid)
1783                 head = current->robust_list;
1784         else {
1785                 struct task_struct *p;
1786
1787                 ret = -ESRCH;
1788                 rcu_read_lock();
1789                 p = find_task_by_vpid(pid);
1790                 if (!p)
1791                         goto err_unlock;
1792                 ret = -EPERM;
1793                 pcred = __task_cred(p);
1794                 if (cred->euid != pcred->euid &&
1795                     cred->euid != pcred->uid &&
1796                     !capable(CAP_SYS_PTRACE))
1797                         goto err_unlock;
1798                 head = p->robust_list;
1799                 rcu_read_unlock();
1800         }
1801
1802         if (put_user(sizeof(*head), len_ptr))
1803                 return -EFAULT;
1804         return put_user(head, head_ptr);
1805
1806 err_unlock:
1807         rcu_read_unlock();
1808
1809         return ret;
1810 }
1811
1812 /*
1813  * Process a futex-list entry, check whether it's owned by the
1814  * dying task, and do notification if so:
1815  */
1816 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1817 {
1818         u32 uval, nval, mval;
1819
1820 retry:
1821         if (get_user(uval, uaddr))
1822                 return -1;
1823
1824         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1825                 /*
1826                  * Ok, this dying thread is truly holding a futex
1827                  * of interest. Set the OWNER_DIED bit atomically
1828                  * via cmpxchg, and if the value had FUTEX_WAITERS
1829                  * set, wake up a waiter (if any). (We have to do a
1830                  * futex_wake() even if OWNER_DIED is already set -
1831                  * to handle the rare but possible case of recursive
1832                  * thread-death.) The rest of the cleanup is done in
1833                  * userspace.
1834                  */
1835                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1836                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1837
1838                 if (nval == -EFAULT)
1839                         return -1;
1840
1841                 if (nval != uval)
1842                         goto retry;
1843
1844                 /*
1845                  * Wake robust non-PI futexes here. The wakeup of
1846                  * PI futexes happens in exit_pi_state():
1847                  */
1848                 if (!pi && (uval & FUTEX_WAITERS))
1849                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1850         }
1851         return 0;
1852 }
1853
1854 /*
1855  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1856  */
1857 static inline int fetch_robust_entry(struct robust_list __user **entry,
1858                                      struct robust_list __user * __user *head,
1859                                      int *pi)
1860 {
1861         unsigned long uentry;
1862
1863         if (get_user(uentry, (unsigned long __user *)head))
1864                 return -EFAULT;
1865
1866         *entry = (void __user *)(uentry & ~1UL);
1867         *pi = uentry & 1;
1868
1869         return 0;
1870 }
1871
1872 /*
1873  * Walk curr->robust_list (very carefully, it's a userspace list!)
1874  * and mark any locks found there dead, and notify any waiters.
1875  *
1876  * We silently return on any sign of list-walking problem.
1877  */
1878 void exit_robust_list(struct task_struct *curr)
1879 {
1880         struct robust_list_head __user *head = curr->robust_list;
1881         struct robust_list __user *entry, *next_entry, *pending;
1882         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1883         unsigned long futex_offset;
1884         int rc;
1885
1886         if (!futex_cmpxchg_enabled)
1887                 return;
1888
1889         /*
1890          * Fetch the list head (which was registered earlier, via
1891          * sys_set_robust_list()):
1892          */
1893         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1894                 return;
1895         /*
1896          * Fetch the relative futex offset:
1897          */
1898         if (get_user(futex_offset, &head->futex_offset))
1899                 return;
1900         /*
1901          * Fetch any possibly pending lock-add first, and handle it
1902          * if it exists:
1903          */
1904         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1905                 return;
1906
1907         next_entry = NULL;      /* avoid warning with gcc */
1908         while (entry != &head->list) {
1909                 /*
1910                  * Fetch the next entry in the list before calling
1911                  * handle_futex_death:
1912                  */
1913                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1914                 /*
1915                  * A pending lock might already be on the list, so
1916                  * don't process it twice:
1917                  */
1918                 if (entry != pending)
1919                         if (handle_futex_death((void __user *)entry + futex_offset,
1920                                                 curr, pi))
1921                                 return;
1922                 if (rc)
1923                         return;
1924                 entry = next_entry;
1925                 pi = next_pi;
1926                 /*
1927                  * Avoid excessively long or circular lists:
1928                  */
1929                 if (!--limit)
1930                         break;
1931
1932                 cond_resched();
1933         }
1934
1935         if (pending)
1936                 handle_futex_death((void __user *)pending + futex_offset,
1937                                    curr, pip);
1938 }
1939
1940 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1941                 u32 __user *uaddr2, u32 val2, u32 val3)
1942 {
1943         int clockrt, ret = -ENOSYS;
1944         int cmd = op & FUTEX_CMD_MASK;
1945         int fshared = 0;
1946
1947         if (!(op & FUTEX_PRIVATE_FLAG))
1948                 fshared = 1;
1949
1950         clockrt = op & FUTEX_CLOCK_REALTIME;
1951         if (clockrt && cmd != FUTEX_WAIT_BITSET)
1952                 return -ENOSYS;
1953
1954         switch (cmd) {
1955         case FUTEX_WAIT:
1956                 val3 = FUTEX_BITSET_MATCH_ANY;
1957         case FUTEX_WAIT_BITSET:
1958                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
1959                 break;
1960         case FUTEX_WAKE:
1961                 val3 = FUTEX_BITSET_MATCH_ANY;
1962         case FUTEX_WAKE_BITSET:
1963                 ret = futex_wake(uaddr, fshared, val, val3);
1964                 break;
1965         case FUTEX_REQUEUE:
1966                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1967                 break;
1968         case FUTEX_CMP_REQUEUE:
1969                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1970                 break;
1971         case FUTEX_WAKE_OP:
1972                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1973                 break;
1974         case FUTEX_LOCK_PI:
1975                 if (futex_cmpxchg_enabled)
1976                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1977                 break;
1978         case FUTEX_UNLOCK_PI:
1979                 if (futex_cmpxchg_enabled)
1980                         ret = futex_unlock_pi(uaddr, fshared);
1981                 break;
1982         case FUTEX_TRYLOCK_PI:
1983                 if (futex_cmpxchg_enabled)
1984                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1985                 break;
1986         default:
1987                 ret = -ENOSYS;
1988         }
1989         return ret;
1990 }
1991
1992
1993 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
1994                 struct timespec __user *, utime, u32 __user *, uaddr2,
1995                 u32, val3)
1996 {
1997         struct timespec ts;
1998         ktime_t t, *tp = NULL;
1999         u32 val2 = 0;
2000         int cmd = op & FUTEX_CMD_MASK;
2001
2002         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2003                       cmd == FUTEX_WAIT_BITSET)) {
2004                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2005                         return -EFAULT;
2006                 if (!timespec_valid(&ts))
2007                         return -EINVAL;
2008
2009                 t = timespec_to_ktime(ts);
2010                 if (cmd == FUTEX_WAIT)
2011                         t = ktime_add_safe(ktime_get(), t);
2012                 tp = &t;
2013         }
2014         /*
2015          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2016          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2017          */
2018         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2019             cmd == FUTEX_WAKE_OP)
2020                 val2 = (u32) (unsigned long) utime;
2021
2022         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2023 }
2024
2025 static int __init futex_init(void)
2026 {
2027         u32 curval;
2028         int i;
2029
2030         /*
2031          * This will fail and we want it. Some arch implementations do
2032          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2033          * functionality. We want to know that before we call in any
2034          * of the complex code paths. Also we want to prevent
2035          * registration of robust lists in that case. NULL is
2036          * guaranteed to fault and we get -EFAULT on functional
2037          * implementation, the non functional ones will return
2038          * -ENOSYS.
2039          */
2040         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2041         if (curval == -EFAULT)
2042                 futex_cmpxchg_enabled = 1;
2043
2044         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2045                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2046                 spin_lock_init(&futex_queues[i].lock);
2047         }
2048
2049         return 0;
2050 }
2051 __initcall(futex_init);