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