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