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