futex: remove the pointer math from double_unlock_hb
[linux-2.6] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
58
59 #include <asm/futex.h>
60
61 #include "rtmutex_common.h"
62
63 int __read_mostly futex_cmpxchg_enabled;
64
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
66
67 /*
68  * Priority Inheritance state:
69  */
70 struct futex_pi_state {
71         /*
72          * list of 'owned' pi_state instances - these have to be
73          * cleaned up in do_exit() if the task exits prematurely:
74          */
75         struct list_head list;
76
77         /*
78          * The PI object:
79          */
80         struct rt_mutex pi_mutex;
81
82         struct task_struct *owner;
83         atomic_t refcount;
84
85         union futex_key key;
86 };
87
88 /*
89  * We use this hashed waitqueue instead of a normal wait_queue_t, so
90  * we can wake only the relevant ones (hashed queues may be shared).
91  *
92  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94  * The order of wakup is always to make the first condition true, then
95  * wake up q->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         spin_unlock(&hb2->lock);
663 }
664
665 /*
666  * Wake up waiters matching bitset queued on this futex (uaddr).
667  */
668 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
669 {
670         struct futex_hash_bucket *hb;
671         struct futex_q *this, *next;
672         struct plist_head *head;
673         union futex_key key = FUTEX_KEY_INIT;
674         int ret;
675
676         if (!bitset)
677                 return -EINVAL;
678
679         ret = get_futex_key(uaddr, fshared, &key);
680         if (unlikely(ret != 0))
681                 goto out;
682
683         hb = hash_futex(&key);
684         spin_lock(&hb->lock);
685         head = &hb->chain;
686
687         plist_for_each_entry_safe(this, next, head, list) {
688                 if (match_futex (&this->key, &key)) {
689                         if (this->pi_state) {
690                                 ret = -EINVAL;
691                                 break;
692                         }
693
694                         /* Check if one of the bits is set in both bitsets */
695                         if (!(this->bitset & bitset))
696                                 continue;
697
698                         wake_futex(this);
699                         if (++ret >= nr_wake)
700                                 break;
701                 }
702         }
703
704         spin_unlock(&hb->lock);
705         put_futex_key(fshared, &key);
706 out:
707         return ret;
708 }
709
710 /*
711  * Wake up all waiters hashed on the physical page that is mapped
712  * to this virtual address:
713  */
714 static int
715 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
716               int nr_wake, int nr_wake2, int op)
717 {
718         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
719         struct futex_hash_bucket *hb1, *hb2;
720         struct plist_head *head;
721         struct futex_q *this, *next;
722         int ret, op_ret;
723
724 retry:
725         ret = get_futex_key(uaddr1, fshared, &key1);
726         if (unlikely(ret != 0))
727                 goto out;
728         ret = get_futex_key(uaddr2, fshared, &key2);
729         if (unlikely(ret != 0))
730                 goto out_put_key1;
731
732         hb1 = hash_futex(&key1);
733         hb2 = hash_futex(&key2);
734
735         double_lock_hb(hb1, hb2);
736 retry_private:
737         op_ret = futex_atomic_op_inuser(op, uaddr2);
738         if (unlikely(op_ret < 0)) {
739                 u32 dummy;
740
741                 double_unlock_hb(hb1, hb2);
742
743 #ifndef CONFIG_MMU
744                 /*
745                  * we don't get EFAULT from MMU faults if we don't have an MMU,
746                  * but we might get them from range checking
747                  */
748                 ret = op_ret;
749                 goto out_put_keys;
750 #endif
751
752                 if (unlikely(op_ret != -EFAULT)) {
753                         ret = op_ret;
754                         goto out_put_keys;
755                 }
756
757                 ret = get_user(dummy, uaddr2);
758                 if (ret)
759                         goto out_put_keys;
760
761                 if (!fshared)
762                         goto retry_private;
763
764                 put_futex_key(fshared, &key2);
765                 put_futex_key(fshared, &key1);
766                 goto retry;
767         }
768
769         head = &hb1->chain;
770
771         plist_for_each_entry_safe(this, next, head, list) {
772                 if (match_futex (&this->key, &key1)) {
773                         wake_futex(this);
774                         if (++ret >= nr_wake)
775                                 break;
776                 }
777         }
778
779         if (op_ret > 0) {
780                 head = &hb2->chain;
781
782                 op_ret = 0;
783                 plist_for_each_entry_safe(this, next, head, list) {
784                         if (match_futex (&this->key, &key2)) {
785                                 wake_futex(this);
786                                 if (++op_ret >= nr_wake2)
787                                         break;
788                         }
789                 }
790                 ret += op_ret;
791         }
792
793         double_unlock_hb(hb1, hb2);
794 out_put_keys:
795         put_futex_key(fshared, &key2);
796 out_put_key1:
797         put_futex_key(fshared, &key1);
798 out:
799         return ret;
800 }
801
802 /*
803  * Requeue all waiters hashed on one physical page to another
804  * physical page.
805  */
806 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
807                          int nr_wake, int nr_requeue, u32 *cmpval)
808 {
809         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
810         struct futex_hash_bucket *hb1, *hb2;
811         struct plist_head *head1;
812         struct futex_q *this, *next;
813         int ret, drop_count = 0;
814
815 retry:
816         ret = get_futex_key(uaddr1, fshared, &key1);
817         if (unlikely(ret != 0))
818                 goto out;
819         ret = get_futex_key(uaddr2, fshared, &key2);
820         if (unlikely(ret != 0))
821                 goto out_put_key1;
822
823         hb1 = hash_futex(&key1);
824         hb2 = hash_futex(&key2);
825
826 retry_private:
827         double_lock_hb(hb1, hb2);
828
829         if (likely(cmpval != NULL)) {
830                 u32 curval;
831
832                 ret = get_futex_value_locked(&curval, uaddr1);
833
834                 if (unlikely(ret)) {
835                         double_unlock_hb(hb1, hb2);
836
837                         ret = get_user(curval, uaddr1);
838                         if (ret)
839                                 goto out_put_keys;
840
841                         if (!fshared)
842                                 goto retry_private;
843
844                         put_futex_key(fshared, &key2);
845                         put_futex_key(fshared, &key1);
846                         goto retry;
847                 }
848                 if (curval != *cmpval) {
849                         ret = -EAGAIN;
850                         goto out_unlock;
851                 }
852         }
853
854         head1 = &hb1->chain;
855         plist_for_each_entry_safe(this, next, head1, list) {
856                 if (!match_futex (&this->key, &key1))
857                         continue;
858                 if (++ret <= nr_wake) {
859                         wake_futex(this);
860                 } else {
861                         /*
862                          * If key1 and key2 hash to the same bucket, no need to
863                          * requeue.
864                          */
865                         if (likely(head1 != &hb2->chain)) {
866                                 plist_del(&this->list, &hb1->chain);
867                                 plist_add(&this->list, &hb2->chain);
868                                 this->lock_ptr = &hb2->lock;
869 #ifdef CONFIG_DEBUG_PI_LIST
870                                 this->list.plist.lock = &hb2->lock;
871 #endif
872                         }
873                         this->key = key2;
874                         get_futex_key_refs(&key2);
875                         drop_count++;
876
877                         if (ret - nr_wake >= nr_requeue)
878                                 break;
879                 }
880         }
881
882 out_unlock:
883         double_unlock_hb(hb1, hb2);
884
885         /* drop_futex_key_refs() must be called outside the spinlocks. */
886         while (--drop_count >= 0)
887                 drop_futex_key_refs(&key1);
888
889 out_put_keys:
890         put_futex_key(fshared, &key2);
891 out_put_key1:
892         put_futex_key(fshared, &key1);
893 out:
894         return ret;
895 }
896
897 /* The key must be already stored in q->key. */
898 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
899 {
900         struct futex_hash_bucket *hb;
901
902         init_waitqueue_head(&q->waiter);
903
904         get_futex_key_refs(&q->key);
905         hb = hash_futex(&q->key);
906         q->lock_ptr = &hb->lock;
907
908         spin_lock(&hb->lock);
909         return hb;
910 }
911
912 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
913 {
914         int prio;
915
916         /*
917          * The priority used to register this element is
918          * - either the real thread-priority for the real-time threads
919          * (i.e. threads with a priority lower than MAX_RT_PRIO)
920          * - or MAX_RT_PRIO for non-RT threads.
921          * Thus, all RT-threads are woken first in priority order, and
922          * the others are woken last, in FIFO order.
923          */
924         prio = min(current->normal_prio, MAX_RT_PRIO);
925
926         plist_node_init(&q->list, prio);
927 #ifdef CONFIG_DEBUG_PI_LIST
928         q->list.plist.lock = &hb->lock;
929 #endif
930         plist_add(&q->list, &hb->chain);
931         q->task = current;
932         spin_unlock(&hb->lock);
933 }
934
935 static inline void
936 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
937 {
938         spin_unlock(&hb->lock);
939         drop_futex_key_refs(&q->key);
940 }
941
942 /*
943  * queue_me and unqueue_me must be called as a pair, each
944  * exactly once.  They are called with the hashed spinlock held.
945  */
946
947 /* Return 1 if we were still queued (ie. 0 means we were woken) */
948 static int unqueue_me(struct futex_q *q)
949 {
950         spinlock_t *lock_ptr;
951         int ret = 0;
952
953         /* In the common case we don't take the spinlock, which is nice. */
954 retry:
955         lock_ptr = q->lock_ptr;
956         barrier();
957         if (lock_ptr != NULL) {
958                 spin_lock(lock_ptr);
959                 /*
960                  * q->lock_ptr can change between reading it and
961                  * spin_lock(), causing us to take the wrong lock.  This
962                  * corrects the race condition.
963                  *
964                  * Reasoning goes like this: if we have the wrong lock,
965                  * q->lock_ptr must have changed (maybe several times)
966                  * between reading it and the spin_lock().  It can
967                  * change again after the spin_lock() but only if it was
968                  * already changed before the spin_lock().  It cannot,
969                  * however, change back to the original value.  Therefore
970                  * we can detect whether we acquired the correct lock.
971                  */
972                 if (unlikely(lock_ptr != q->lock_ptr)) {
973                         spin_unlock(lock_ptr);
974                         goto retry;
975                 }
976                 WARN_ON(plist_node_empty(&q->list));
977                 plist_del(&q->list, &q->list.plist);
978
979                 BUG_ON(q->pi_state);
980
981                 spin_unlock(lock_ptr);
982                 ret = 1;
983         }
984
985         drop_futex_key_refs(&q->key);
986         return ret;
987 }
988
989 /*
990  * PI futexes can not be requeued and must remove themself from the
991  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
992  * and dropped here.
993  */
994 static void unqueue_me_pi(struct futex_q *q)
995 {
996         WARN_ON(plist_node_empty(&q->list));
997         plist_del(&q->list, &q->list.plist);
998
999         BUG_ON(!q->pi_state);
1000         free_pi_state(q->pi_state);
1001         q->pi_state = NULL;
1002
1003         spin_unlock(q->lock_ptr);
1004
1005         drop_futex_key_refs(&q->key);
1006 }
1007
1008 /*
1009  * Fixup the pi_state owner with the new owner.
1010  *
1011  * Must be called with hash bucket lock held and mm->sem held for non
1012  * private futexes.
1013  */
1014 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1015                                 struct task_struct *newowner, int fshared)
1016 {
1017         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1018         struct futex_pi_state *pi_state = q->pi_state;
1019         struct task_struct *oldowner = pi_state->owner;
1020         u32 uval, curval, newval;
1021         int ret;
1022
1023         /* Owner died? */
1024         if (!pi_state->owner)
1025                 newtid |= FUTEX_OWNER_DIED;
1026
1027         /*
1028          * We are here either because we stole the rtmutex from the
1029          * pending owner or we are the pending owner which failed to
1030          * get the rtmutex. We have to replace the pending owner TID
1031          * in the user space variable. This must be atomic as we have
1032          * to preserve the owner died bit here.
1033          *
1034          * Note: We write the user space value _before_ changing the pi_state
1035          * because we can fault here. Imagine swapped out pages or a fork
1036          * that marked all the anonymous memory readonly for cow.
1037          *
1038          * Modifying pi_state _before_ the user space value would
1039          * leave the pi_state in an inconsistent state when we fault
1040          * here, because we need to drop the hash bucket lock to
1041          * handle the fault. This might be observed in the PID check
1042          * in lookup_pi_state.
1043          */
1044 retry:
1045         if (get_futex_value_locked(&uval, uaddr))
1046                 goto handle_fault;
1047
1048         while (1) {
1049                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1050
1051                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1052
1053                 if (curval == -EFAULT)
1054                         goto handle_fault;
1055                 if (curval == uval)
1056                         break;
1057                 uval = curval;
1058         }
1059
1060         /*
1061          * We fixed up user space. Now we need to fix the pi_state
1062          * itself.
1063          */
1064         if (pi_state->owner != NULL) {
1065                 spin_lock_irq(&pi_state->owner->pi_lock);
1066                 WARN_ON(list_empty(&pi_state->list));
1067                 list_del_init(&pi_state->list);
1068                 spin_unlock_irq(&pi_state->owner->pi_lock);
1069         }
1070
1071         pi_state->owner = newowner;
1072
1073         spin_lock_irq(&newowner->pi_lock);
1074         WARN_ON(!list_empty(&pi_state->list));
1075         list_add(&pi_state->list, &newowner->pi_state_list);
1076         spin_unlock_irq(&newowner->pi_lock);
1077         return 0;
1078
1079         /*
1080          * To handle the page fault we need to drop the hash bucket
1081          * lock here. That gives the other task (either the pending
1082          * owner itself or the task which stole the rtmutex) the
1083          * chance to try the fixup of the pi_state. So once we are
1084          * back from handling the fault we need to check the pi_state
1085          * after reacquiring the hash bucket lock and before trying to
1086          * do another fixup. When the fixup has been done already we
1087          * simply return.
1088          */
1089 handle_fault:
1090         spin_unlock(q->lock_ptr);
1091
1092         ret = get_user(uval, uaddr);
1093
1094         spin_lock(q->lock_ptr);
1095
1096         /*
1097          * Check if someone else fixed it for us:
1098          */
1099         if (pi_state->owner != oldowner)
1100                 return 0;
1101
1102         if (ret)
1103                 return ret;
1104
1105         goto retry;
1106 }
1107
1108 /*
1109  * In case we must use restart_block to restart a futex_wait,
1110  * we encode in the 'flags' shared capability
1111  */
1112 #define FLAGS_SHARED            0x01
1113 #define FLAGS_CLOCKRT           0x02
1114
1115 static long futex_wait_restart(struct restart_block *restart);
1116
1117 static int futex_wait(u32 __user *uaddr, int fshared,
1118                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1119 {
1120         struct task_struct *curr = current;
1121         struct restart_block *restart;
1122         DECLARE_WAITQUEUE(wait, curr);
1123         struct futex_hash_bucket *hb;
1124         struct futex_q q;
1125         u32 uval;
1126         int ret;
1127         struct hrtimer_sleeper t;
1128         int rem = 0;
1129
1130         if (!bitset)
1131                 return -EINVAL;
1132
1133         q.pi_state = NULL;
1134         q.bitset = bitset;
1135 retry:
1136         q.key = FUTEX_KEY_INIT;
1137         ret = get_futex_key(uaddr, fshared, &q.key);
1138         if (unlikely(ret != 0))
1139                 goto out;
1140
1141 retry_private:
1142         hb = queue_lock(&q);
1143
1144         /*
1145          * Access the page AFTER the hash-bucket is locked.
1146          * Order is important:
1147          *
1148          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1149          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1150          *
1151          * The basic logical guarantee of a futex is that it blocks ONLY
1152          * if cond(var) is known to be true at the time of blocking, for
1153          * any cond.  If we queued after testing *uaddr, that would open
1154          * a race condition where we could block indefinitely with
1155          * cond(var) false, which would violate the guarantee.
1156          *
1157          * A consequence is that futex_wait() can return zero and absorb
1158          * a wakeup when *uaddr != val on entry to the syscall.  This is
1159          * rare, but normal.
1160          *
1161          * For shared futexes, we hold the mmap semaphore, so the mapping
1162          * cannot have changed since we looked it up in get_futex_key.
1163          */
1164         ret = get_futex_value_locked(&uval, uaddr);
1165
1166         if (unlikely(ret)) {
1167                 queue_unlock(&q, hb);
1168
1169                 ret = get_user(uval, uaddr);
1170                 if (ret)
1171                         goto out_put_key;
1172
1173                 if (!fshared)
1174                         goto retry_private;
1175
1176                 put_futex_key(fshared, &q.key);
1177                 goto retry;
1178         }
1179         ret = -EWOULDBLOCK;
1180         if (unlikely(uval != val)) {
1181                 queue_unlock(&q, hb);
1182                 goto out_put_key;
1183         }
1184
1185         /* Only actually queue if *uaddr contained val.  */
1186         queue_me(&q, hb);
1187
1188         /*
1189          * There might have been scheduling since the queue_me(), as we
1190          * cannot hold a spinlock across the get_user() in case it
1191          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1192          * queueing ourselves into the futex hash.  This code thus has to
1193          * rely on the futex_wake() code removing us from hash when it
1194          * wakes us up.
1195          */
1196
1197         /* add_wait_queue is the barrier after __set_current_state. */
1198         __set_current_state(TASK_INTERRUPTIBLE);
1199         add_wait_queue(&q.waiter, &wait);
1200         /*
1201          * !plist_node_empty() is safe here without any lock.
1202          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1203          */
1204         if (likely(!plist_node_empty(&q.list))) {
1205                 if (!abs_time)
1206                         schedule();
1207                 else {
1208                         hrtimer_init_on_stack(&t.timer,
1209                                               clockrt ? CLOCK_REALTIME :
1210                                               CLOCK_MONOTONIC,
1211                                               HRTIMER_MODE_ABS);
1212                         hrtimer_init_sleeper(&t, current);
1213                         hrtimer_set_expires_range_ns(&t.timer, *abs_time,
1214                                                      current->timer_slack_ns);
1215
1216                         hrtimer_start_expires(&t.timer, HRTIMER_MODE_ABS);
1217                         if (!hrtimer_active(&t.timer))
1218                                 t.task = NULL;
1219
1220                         /*
1221                          * the timer could have already expired, in which
1222                          * case current would be flagged for rescheduling.
1223                          * Don't bother calling schedule.
1224                          */
1225                         if (likely(t.task))
1226                                 schedule();
1227
1228                         hrtimer_cancel(&t.timer);
1229
1230                         /* Flag if a timeout occured */
1231                         rem = (t.task == NULL);
1232
1233                         destroy_hrtimer_on_stack(&t.timer);
1234                 }
1235         }
1236         __set_current_state(TASK_RUNNING);
1237
1238         /*
1239          * NOTE: we don't remove ourselves from the waitqueue because
1240          * we are the only user of it.
1241          */
1242
1243         /* If we were woken (and unqueued), we succeeded, whatever. */
1244         ret = 0;
1245         if (!unqueue_me(&q))
1246                 goto out_put_key;
1247         ret = -ETIMEDOUT;
1248         if (rem)
1249                 goto out_put_key;
1250
1251         /*
1252          * We expect signal_pending(current), but another thread may
1253          * have handled it for us already.
1254          */
1255         ret = -ERESTARTSYS;
1256         if (!abs_time)
1257                 goto out_put_key;
1258
1259         restart = &current_thread_info()->restart_block;
1260         restart->fn = futex_wait_restart;
1261         restart->futex.uaddr = (u32 *)uaddr;
1262         restart->futex.val = val;
1263         restart->futex.time = abs_time->tv64;
1264         restart->futex.bitset = bitset;
1265         restart->futex.flags = 0;
1266
1267         if (fshared)
1268                 restart->futex.flags |= FLAGS_SHARED;
1269         if (clockrt)
1270                 restart->futex.flags |= FLAGS_CLOCKRT;
1271
1272         ret = -ERESTART_RESTARTBLOCK;
1273
1274 out_put_key:
1275         put_futex_key(fshared, &q.key);
1276 out:
1277         return ret;
1278 }
1279
1280
1281 static long futex_wait_restart(struct restart_block *restart)
1282 {
1283         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1284         int fshared = 0;
1285         ktime_t t;
1286
1287         t.tv64 = restart->futex.time;
1288         restart->fn = do_no_restart_syscall;
1289         if (restart->futex.flags & FLAGS_SHARED)
1290                 fshared = 1;
1291         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1292                                 restart->futex.bitset,
1293                                 restart->futex.flags & FLAGS_CLOCKRT);
1294 }
1295
1296
1297 /*
1298  * Userspace tried a 0 -> TID atomic transition of the futex value
1299  * and failed. The kernel side here does the whole locking operation:
1300  * if there are waiters then it will block, it does PI, etc. (Due to
1301  * races the kernel might see a 0 value of the futex too.)
1302  */
1303 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1304                          int detect, ktime_t *time, int trylock)
1305 {
1306         struct hrtimer_sleeper timeout, *to = NULL;
1307         struct task_struct *curr = current;
1308         struct futex_hash_bucket *hb;
1309         u32 uval, newval, curval;
1310         struct futex_q q;
1311         int ret, lock_taken, ownerdied = 0;
1312
1313         if (refill_pi_state_cache())
1314                 return -ENOMEM;
1315
1316         if (time) {
1317                 to = &timeout;
1318                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1319                                       HRTIMER_MODE_ABS);
1320                 hrtimer_init_sleeper(to, current);
1321                 hrtimer_set_expires(&to->timer, *time);
1322         }
1323
1324         q.pi_state = NULL;
1325 retry:
1326         q.key = FUTEX_KEY_INIT;
1327         ret = get_futex_key(uaddr, fshared, &q.key);
1328         if (unlikely(ret != 0))
1329                 goto out;
1330
1331 retry_private:
1332         hb = queue_lock(&q);
1333
1334 retry_locked:
1335         ret = lock_taken = 0;
1336
1337         /*
1338          * To avoid races, we attempt to take the lock here again
1339          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1340          * the locks. It will most likely not succeed.
1341          */
1342         newval = task_pid_vnr(current);
1343
1344         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1345
1346         if (unlikely(curval == -EFAULT))
1347                 goto uaddr_faulted;
1348
1349         /*
1350          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1351          * situation and we return success to user space.
1352          */
1353         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1354                 ret = -EDEADLK;
1355                 goto out_unlock_put_key;
1356         }
1357
1358         /*
1359          * Surprise - we got the lock. Just return to userspace:
1360          */
1361         if (unlikely(!curval))
1362                 goto out_unlock_put_key;
1363
1364         uval = curval;
1365
1366         /*
1367          * Set the WAITERS flag, so the owner will know it has someone
1368          * to wake at next unlock
1369          */
1370         newval = curval | FUTEX_WAITERS;
1371
1372         /*
1373          * There are two cases, where a futex might have no owner (the
1374          * owner TID is 0): OWNER_DIED. We take over the futex in this
1375          * case. We also do an unconditional take over, when the owner
1376          * of the futex died.
1377          *
1378          * This is safe as we are protected by the hash bucket lock !
1379          */
1380         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1381                 /* Keep the OWNER_DIED bit */
1382                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1383                 ownerdied = 0;
1384                 lock_taken = 1;
1385         }
1386
1387         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1388
1389         if (unlikely(curval == -EFAULT))
1390                 goto uaddr_faulted;
1391         if (unlikely(curval != uval))
1392                 goto retry_locked;
1393
1394         /*
1395          * We took the lock due to owner died take over.
1396          */
1397         if (unlikely(lock_taken))
1398                 goto out_unlock_put_key;
1399
1400         /*
1401          * We dont have the lock. Look up the PI state (or create it if
1402          * we are the first waiter):
1403          */
1404         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1405
1406         if (unlikely(ret)) {
1407                 switch (ret) {
1408
1409                 case -EAGAIN:
1410                         /*
1411                          * Task is exiting and we just wait for the
1412                          * exit to complete.
1413                          */
1414                         queue_unlock(&q, hb);
1415                         put_futex_key(fshared, &q.key);
1416                         cond_resched();
1417                         goto retry;
1418
1419                 case -ESRCH:
1420                         /*
1421                          * No owner found for this futex. Check if the
1422                          * OWNER_DIED bit is set to figure out whether
1423                          * this is a robust futex or not.
1424                          */
1425                         if (get_futex_value_locked(&curval, uaddr))
1426                                 goto uaddr_faulted;
1427
1428                         /*
1429                          * We simply start over in case of a robust
1430                          * futex. The code above will take the futex
1431                          * and return happy.
1432                          */
1433                         if (curval & FUTEX_OWNER_DIED) {
1434                                 ownerdied = 1;
1435                                 goto retry_locked;
1436                         }
1437                 default:
1438                         goto out_unlock_put_key;
1439                 }
1440         }
1441
1442         /*
1443          * Only actually queue now that the atomic ops are done:
1444          */
1445         queue_me(&q, hb);
1446
1447         WARN_ON(!q.pi_state);
1448         /*
1449          * Block on the PI mutex:
1450          */
1451         if (!trylock)
1452                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1453         else {
1454                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1455                 /* Fixup the trylock return value: */
1456                 ret = ret ? 0 : -EWOULDBLOCK;
1457         }
1458
1459         spin_lock(q.lock_ptr);
1460
1461         if (!ret) {
1462                 /*
1463                  * Got the lock. We might not be the anticipated owner
1464                  * if we did a lock-steal - fix up the PI-state in
1465                  * that case:
1466                  */
1467                 if (q.pi_state->owner != curr)
1468                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1469         } else {
1470                 /*
1471                  * Catch the rare case, where the lock was released
1472                  * when we were on the way back before we locked the
1473                  * hash bucket.
1474                  */
1475                 if (q.pi_state->owner == curr) {
1476                         /*
1477                          * Try to get the rt_mutex now. This might
1478                          * fail as some other task acquired the
1479                          * rt_mutex after we removed ourself from the
1480                          * rt_mutex waiters list.
1481                          */
1482                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1483                                 ret = 0;
1484                         else {
1485                                 /*
1486                                  * pi_state is incorrect, some other
1487                                  * task did a lock steal and we
1488                                  * returned due to timeout or signal
1489                                  * without taking the rt_mutex. Too
1490                                  * late. We can access the
1491                                  * rt_mutex_owner without locking, as
1492                                  * the other task is now blocked on
1493                                  * the hash bucket lock. Fix the state
1494                                  * up.
1495                                  */
1496                                 struct task_struct *owner;
1497                                 int res;
1498
1499                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1500                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1501                                                            fshared);
1502
1503                                 /* propagate -EFAULT, if the fixup failed */
1504                                 if (res)
1505                                         ret = res;
1506                         }
1507                 } else {
1508                         /*
1509                          * Paranoia check. If we did not take the lock
1510                          * in the trylock above, then we should not be
1511                          * the owner of the rtmutex, neither the real
1512                          * nor the pending one:
1513                          */
1514                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1515                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1516                                        "pi-mutex: %p pi-state %p\n", ret,
1517                                        q.pi_state->pi_mutex.owner,
1518                                        q.pi_state->owner);
1519                 }
1520         }
1521
1522         /*
1523          * If fixup_pi_state_owner() faulted and was unable to handle the
1524          * fault, unlock it and return the fault to userspace.
1525          */
1526         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1527                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1528
1529         /* Unqueue and drop the lock */
1530         unqueue_me_pi(&q);
1531
1532         if (to)
1533                 destroy_hrtimer_on_stack(&to->timer);
1534         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1535
1536 out_unlock_put_key:
1537         queue_unlock(&q, hb);
1538
1539 out_put_key:
1540         put_futex_key(fshared, &q.key);
1541 out:
1542         if (to)
1543                 destroy_hrtimer_on_stack(&to->timer);
1544         return ret;
1545
1546 uaddr_faulted:
1547         /*
1548          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1549          * atomically.  Therefore, if we continue to fault after get_user()
1550          * below, we need to handle the fault ourselves, while still holding
1551          * the mmap_sem.  This can occur if the uaddr is under contention as
1552          * we have to drop the mmap_sem in order to call get_user().
1553          */
1554         queue_unlock(&q, hb);
1555
1556         ret = get_user(uval, uaddr);
1557         if (ret)
1558                 goto out_put_key;
1559
1560         if (!fshared)
1561                 goto retry_private;
1562
1563         put_futex_key(fshared, &q.key);
1564         goto retry;
1565 }
1566
1567
1568 /*
1569  * Userspace attempted a TID -> 0 atomic transition, and failed.
1570  * This is the in-kernel slowpath: we look up the PI state (if any),
1571  * and do the rt-mutex unlock.
1572  */
1573 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1574 {
1575         struct futex_hash_bucket *hb;
1576         struct futex_q *this, *next;
1577         u32 uval;
1578         struct plist_head *head;
1579         union futex_key key = FUTEX_KEY_INIT;
1580         int ret;
1581
1582 retry:
1583         if (get_user(uval, uaddr))
1584                 return -EFAULT;
1585         /*
1586          * We release only a lock we actually own:
1587          */
1588         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1589                 return -EPERM;
1590
1591         ret = get_futex_key(uaddr, fshared, &key);
1592         if (unlikely(ret != 0))
1593                 goto out;
1594
1595         hb = hash_futex(&key);
1596         spin_lock(&hb->lock);
1597
1598         /*
1599          * To avoid races, try to do the TID -> 0 atomic transition
1600          * again. If it succeeds then we can return without waking
1601          * anyone else up:
1602          */
1603         if (!(uval & FUTEX_OWNER_DIED))
1604                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1605
1606
1607         if (unlikely(uval == -EFAULT))
1608                 goto pi_faulted;
1609         /*
1610          * Rare case: we managed to release the lock atomically,
1611          * no need to wake anyone else up:
1612          */
1613         if (unlikely(uval == task_pid_vnr(current)))
1614                 goto out_unlock;
1615
1616         /*
1617          * Ok, other tasks may need to be woken up - check waiters
1618          * and do the wakeup if necessary:
1619          */
1620         head = &hb->chain;
1621
1622         plist_for_each_entry_safe(this, next, head, list) {
1623                 if (!match_futex (&this->key, &key))
1624                         continue;
1625                 ret = wake_futex_pi(uaddr, uval, this);
1626                 /*
1627                  * The atomic access to the futex value
1628                  * generated a pagefault, so retry the
1629                  * user-access and the wakeup:
1630                  */
1631                 if (ret == -EFAULT)
1632                         goto pi_faulted;
1633                 goto out_unlock;
1634         }
1635         /*
1636          * No waiters - kernel unlocks the futex:
1637          */
1638         if (!(uval & FUTEX_OWNER_DIED)) {
1639                 ret = unlock_futex_pi(uaddr, uval);
1640                 if (ret == -EFAULT)
1641                         goto pi_faulted;
1642         }
1643
1644 out_unlock:
1645         spin_unlock(&hb->lock);
1646         put_futex_key(fshared, &key);
1647
1648 out:
1649         return ret;
1650
1651 pi_faulted:
1652         /*
1653          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1654          * atomically.  Therefore, if we continue to fault after get_user()
1655          * below, we need to handle the fault ourselves, while still holding
1656          * the mmap_sem.  This can occur if the uaddr is under contention as
1657          * we have to drop the mmap_sem in order to call get_user().
1658          */
1659         spin_unlock(&hb->lock);
1660         put_futex_key(fshared, &key);
1661
1662         ret = get_user(uval, uaddr);
1663         if (!ret)
1664                 goto retry;
1665
1666         return ret;
1667 }
1668
1669 /*
1670  * Support for robust futexes: the kernel cleans up held futexes at
1671  * thread exit time.
1672  *
1673  * Implementation: user-space maintains a per-thread list of locks it
1674  * is holding. Upon do_exit(), the kernel carefully walks this list,
1675  * and marks all locks that are owned by this thread with the
1676  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1677  * always manipulated with the lock held, so the list is private and
1678  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1679  * field, to allow the kernel to clean up if the thread dies after
1680  * acquiring the lock, but just before it could have added itself to
1681  * the list. There can only be one such pending lock.
1682  */
1683
1684 /**
1685  * sys_set_robust_list - set the robust-futex list head of a task
1686  * @head: pointer to the list-head
1687  * @len: length of the list-head, as userspace expects
1688  */
1689 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
1690                 size_t, len)
1691 {
1692         if (!futex_cmpxchg_enabled)
1693                 return -ENOSYS;
1694         /*
1695          * The kernel knows only one size for now:
1696          */
1697         if (unlikely(len != sizeof(*head)))
1698                 return -EINVAL;
1699
1700         current->robust_list = head;
1701
1702         return 0;
1703 }
1704
1705 /**
1706  * sys_get_robust_list - get the robust-futex list head of a task
1707  * @pid: pid of the process [zero for current task]
1708  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1709  * @len_ptr: pointer to a length field, the kernel fills in the header size
1710  */
1711 SYSCALL_DEFINE3(get_robust_list, int, pid,
1712                 struct robust_list_head __user * __user *, head_ptr,
1713                 size_t __user *, len_ptr)
1714 {
1715         struct robust_list_head __user *head;
1716         unsigned long ret;
1717         const struct cred *cred = current_cred(), *pcred;
1718
1719         if (!futex_cmpxchg_enabled)
1720                 return -ENOSYS;
1721
1722         if (!pid)
1723                 head = current->robust_list;
1724         else {
1725                 struct task_struct *p;
1726
1727                 ret = -ESRCH;
1728                 rcu_read_lock();
1729                 p = find_task_by_vpid(pid);
1730                 if (!p)
1731                         goto err_unlock;
1732                 ret = -EPERM;
1733                 pcred = __task_cred(p);
1734                 if (cred->euid != pcred->euid &&
1735                     cred->euid != pcred->uid &&
1736                     !capable(CAP_SYS_PTRACE))
1737                         goto err_unlock;
1738                 head = p->robust_list;
1739                 rcu_read_unlock();
1740         }
1741
1742         if (put_user(sizeof(*head), len_ptr))
1743                 return -EFAULT;
1744         return put_user(head, head_ptr);
1745
1746 err_unlock:
1747         rcu_read_unlock();
1748
1749         return ret;
1750 }
1751
1752 /*
1753  * Process a futex-list entry, check whether it's owned by the
1754  * dying task, and do notification if so:
1755  */
1756 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1757 {
1758         u32 uval, nval, mval;
1759
1760 retry:
1761         if (get_user(uval, uaddr))
1762                 return -1;
1763
1764         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1765                 /*
1766                  * Ok, this dying thread is truly holding a futex
1767                  * of interest. Set the OWNER_DIED bit atomically
1768                  * via cmpxchg, and if the value had FUTEX_WAITERS
1769                  * set, wake up a waiter (if any). (We have to do a
1770                  * futex_wake() even if OWNER_DIED is already set -
1771                  * to handle the rare but possible case of recursive
1772                  * thread-death.) The rest of the cleanup is done in
1773                  * userspace.
1774                  */
1775                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1776                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1777
1778                 if (nval == -EFAULT)
1779                         return -1;
1780
1781                 if (nval != uval)
1782                         goto retry;
1783
1784                 /*
1785                  * Wake robust non-PI futexes here. The wakeup of
1786                  * PI futexes happens in exit_pi_state():
1787                  */
1788                 if (!pi && (uval & FUTEX_WAITERS))
1789                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1790         }
1791         return 0;
1792 }
1793
1794 /*
1795  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1796  */
1797 static inline int fetch_robust_entry(struct robust_list __user **entry,
1798                                      struct robust_list __user * __user *head,
1799                                      int *pi)
1800 {
1801         unsigned long uentry;
1802
1803         if (get_user(uentry, (unsigned long __user *)head))
1804                 return -EFAULT;
1805
1806         *entry = (void __user *)(uentry & ~1UL);
1807         *pi = uentry & 1;
1808
1809         return 0;
1810 }
1811
1812 /*
1813  * Walk curr->robust_list (very carefully, it's a userspace list!)
1814  * and mark any locks found there dead, and notify any waiters.
1815  *
1816  * We silently return on any sign of list-walking problem.
1817  */
1818 void exit_robust_list(struct task_struct *curr)
1819 {
1820         struct robust_list_head __user *head = curr->robust_list;
1821         struct robust_list __user *entry, *next_entry, *pending;
1822         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1823         unsigned long futex_offset;
1824         int rc;
1825
1826         if (!futex_cmpxchg_enabled)
1827                 return;
1828
1829         /*
1830          * Fetch the list head (which was registered earlier, via
1831          * sys_set_robust_list()):
1832          */
1833         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1834                 return;
1835         /*
1836          * Fetch the relative futex offset:
1837          */
1838         if (get_user(futex_offset, &head->futex_offset))
1839                 return;
1840         /*
1841          * Fetch any possibly pending lock-add first, and handle it
1842          * if it exists:
1843          */
1844         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1845                 return;
1846
1847         next_entry = NULL;      /* avoid warning with gcc */
1848         while (entry != &head->list) {
1849                 /*
1850                  * Fetch the next entry in the list before calling
1851                  * handle_futex_death:
1852                  */
1853                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1854                 /*
1855                  * A pending lock might already be on the list, so
1856                  * don't process it twice:
1857                  */
1858                 if (entry != pending)
1859                         if (handle_futex_death((void __user *)entry + futex_offset,
1860                                                 curr, pi))
1861                                 return;
1862                 if (rc)
1863                         return;
1864                 entry = next_entry;
1865                 pi = next_pi;
1866                 /*
1867                  * Avoid excessively long or circular lists:
1868                  */
1869                 if (!--limit)
1870                         break;
1871
1872                 cond_resched();
1873         }
1874
1875         if (pending)
1876                 handle_futex_death((void __user *)pending + futex_offset,
1877                                    curr, pip);
1878 }
1879
1880 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1881                 u32 __user *uaddr2, u32 val2, u32 val3)
1882 {
1883         int clockrt, ret = -ENOSYS;
1884         int cmd = op & FUTEX_CMD_MASK;
1885         int fshared = 0;
1886
1887         if (!(op & FUTEX_PRIVATE_FLAG))
1888                 fshared = 1;
1889
1890         clockrt = op & FUTEX_CLOCK_REALTIME;
1891         if (clockrt && cmd != FUTEX_WAIT_BITSET)
1892                 return -ENOSYS;
1893
1894         switch (cmd) {
1895         case FUTEX_WAIT:
1896                 val3 = FUTEX_BITSET_MATCH_ANY;
1897         case FUTEX_WAIT_BITSET:
1898                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
1899                 break;
1900         case FUTEX_WAKE:
1901                 val3 = FUTEX_BITSET_MATCH_ANY;
1902         case FUTEX_WAKE_BITSET:
1903                 ret = futex_wake(uaddr, fshared, val, val3);
1904                 break;
1905         case FUTEX_REQUEUE:
1906                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1907                 break;
1908         case FUTEX_CMP_REQUEUE:
1909                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1910                 break;
1911         case FUTEX_WAKE_OP:
1912                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1913                 break;
1914         case FUTEX_LOCK_PI:
1915                 if (futex_cmpxchg_enabled)
1916                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1917                 break;
1918         case FUTEX_UNLOCK_PI:
1919                 if (futex_cmpxchg_enabled)
1920                         ret = futex_unlock_pi(uaddr, fshared);
1921                 break;
1922         case FUTEX_TRYLOCK_PI:
1923                 if (futex_cmpxchg_enabled)
1924                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1925                 break;
1926         default:
1927                 ret = -ENOSYS;
1928         }
1929         return ret;
1930 }
1931
1932
1933 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
1934                 struct timespec __user *, utime, u32 __user *, uaddr2,
1935                 u32, val3)
1936 {
1937         struct timespec ts;
1938         ktime_t t, *tp = NULL;
1939         u32 val2 = 0;
1940         int cmd = op & FUTEX_CMD_MASK;
1941
1942         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
1943                       cmd == FUTEX_WAIT_BITSET)) {
1944                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
1945                         return -EFAULT;
1946                 if (!timespec_valid(&ts))
1947                         return -EINVAL;
1948
1949                 t = timespec_to_ktime(ts);
1950                 if (cmd == FUTEX_WAIT)
1951                         t = ktime_add_safe(ktime_get(), t);
1952                 tp = &t;
1953         }
1954         /*
1955          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
1956          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
1957          */
1958         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
1959             cmd == FUTEX_WAKE_OP)
1960                 val2 = (u32) (unsigned long) utime;
1961
1962         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
1963 }
1964
1965 static int __init futex_init(void)
1966 {
1967         u32 curval;
1968         int i;
1969
1970         /*
1971          * This will fail and we want it. Some arch implementations do
1972          * runtime detection of the futex_atomic_cmpxchg_inatomic()
1973          * functionality. We want to know that before we call in any
1974          * of the complex code paths. Also we want to prevent
1975          * registration of robust lists in that case. NULL is
1976          * guaranteed to fault and we get -EFAULT on functional
1977          * implementation, the non functional ones will return
1978          * -ENOSYS.
1979          */
1980         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
1981         if (curval == -EFAULT)
1982                 futex_cmpxchg_enabled = 1;
1983
1984         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1985                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
1986                 spin_lock_init(&futex_queues[i].lock);
1987         }
1988
1989         return 0;
1990 }
1991 __initcall(futex_init);