421fb5e42a10607fc110aca436a4e944bfe53792
[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 /**
280  * futex_top_waiter() - Return the highest priority waiter on a futex
281  * @hb:     the hash bucket the futex_q's reside in
282  * @key:    the futex key (to distinguish it from other futex futex_q's)
283  *
284  * Must be called with the hb lock held.
285  */
286 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
287                                         union futex_key *key)
288 {
289         struct futex_q *this;
290
291         plist_for_each_entry(this, &hb->chain, list) {
292                 if (match_futex(&this->key, key))
293                         return this;
294         }
295         return NULL;
296 }
297
298 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
299 {
300         u32 curval;
301
302         pagefault_disable();
303         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
304         pagefault_enable();
305
306         return curval;
307 }
308
309 static int get_futex_value_locked(u32 *dest, u32 __user *from)
310 {
311         int ret;
312
313         pagefault_disable();
314         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
315         pagefault_enable();
316
317         return ret ? -EFAULT : 0;
318 }
319
320
321 /*
322  * PI code:
323  */
324 static int refill_pi_state_cache(void)
325 {
326         struct futex_pi_state *pi_state;
327
328         if (likely(current->pi_state_cache))
329                 return 0;
330
331         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
332
333         if (!pi_state)
334                 return -ENOMEM;
335
336         INIT_LIST_HEAD(&pi_state->list);
337         /* pi_mutex gets initialized later */
338         pi_state->owner = NULL;
339         atomic_set(&pi_state->refcount, 1);
340         pi_state->key = FUTEX_KEY_INIT;
341
342         current->pi_state_cache = pi_state;
343
344         return 0;
345 }
346
347 static struct futex_pi_state * alloc_pi_state(void)
348 {
349         struct futex_pi_state *pi_state = current->pi_state_cache;
350
351         WARN_ON(!pi_state);
352         current->pi_state_cache = NULL;
353
354         return pi_state;
355 }
356
357 static void free_pi_state(struct futex_pi_state *pi_state)
358 {
359         if (!atomic_dec_and_test(&pi_state->refcount))
360                 return;
361
362         /*
363          * If pi_state->owner is NULL, the owner is most probably dying
364          * and has cleaned up the pi_state already
365          */
366         if (pi_state->owner) {
367                 spin_lock_irq(&pi_state->owner->pi_lock);
368                 list_del_init(&pi_state->list);
369                 spin_unlock_irq(&pi_state->owner->pi_lock);
370
371                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
372         }
373
374         if (current->pi_state_cache)
375                 kfree(pi_state);
376         else {
377                 /*
378                  * pi_state->list is already empty.
379                  * clear pi_state->owner.
380                  * refcount is at 0 - put it back to 1.
381                  */
382                 pi_state->owner = NULL;
383                 atomic_set(&pi_state->refcount, 1);
384                 current->pi_state_cache = pi_state;
385         }
386 }
387
388 /*
389  * Look up the task based on what TID userspace gave us.
390  * We dont trust it.
391  */
392 static struct task_struct * futex_find_get_task(pid_t pid)
393 {
394         struct task_struct *p;
395         const struct cred *cred = current_cred(), *pcred;
396
397         rcu_read_lock();
398         p = find_task_by_vpid(pid);
399         if (!p) {
400                 p = ERR_PTR(-ESRCH);
401         } else {
402                 pcred = __task_cred(p);
403                 if (cred->euid != pcred->euid &&
404                     cred->euid != pcred->uid)
405                         p = ERR_PTR(-ESRCH);
406                 else
407                         get_task_struct(p);
408         }
409
410         rcu_read_unlock();
411
412         return p;
413 }
414
415 /*
416  * This task is holding PI mutexes at exit time => bad.
417  * Kernel cleans up PI-state, but userspace is likely hosed.
418  * (Robust-futex cleanup is separate and might save the day for userspace.)
419  */
420 void exit_pi_state_list(struct task_struct *curr)
421 {
422         struct list_head *next, *head = &curr->pi_state_list;
423         struct futex_pi_state *pi_state;
424         struct futex_hash_bucket *hb;
425         union futex_key key = FUTEX_KEY_INIT;
426
427         if (!futex_cmpxchg_enabled)
428                 return;
429         /*
430          * We are a ZOMBIE and nobody can enqueue itself on
431          * pi_state_list anymore, but we have to be careful
432          * versus waiters unqueueing themselves:
433          */
434         spin_lock_irq(&curr->pi_lock);
435         while (!list_empty(head)) {
436
437                 next = head->next;
438                 pi_state = list_entry(next, struct futex_pi_state, list);
439                 key = pi_state->key;
440                 hb = hash_futex(&key);
441                 spin_unlock_irq(&curr->pi_lock);
442
443                 spin_lock(&hb->lock);
444
445                 spin_lock_irq(&curr->pi_lock);
446                 /*
447                  * We dropped the pi-lock, so re-check whether this
448                  * task still owns the PI-state:
449                  */
450                 if (head->next != next) {
451                         spin_unlock(&hb->lock);
452                         continue;
453                 }
454
455                 WARN_ON(pi_state->owner != curr);
456                 WARN_ON(list_empty(&pi_state->list));
457                 list_del_init(&pi_state->list);
458                 pi_state->owner = NULL;
459                 spin_unlock_irq(&curr->pi_lock);
460
461                 rt_mutex_unlock(&pi_state->pi_mutex);
462
463                 spin_unlock(&hb->lock);
464
465                 spin_lock_irq(&curr->pi_lock);
466         }
467         spin_unlock_irq(&curr->pi_lock);
468 }
469
470 static int
471 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
472                 union futex_key *key, struct futex_pi_state **ps)
473 {
474         struct futex_pi_state *pi_state = NULL;
475         struct futex_q *this, *next;
476         struct plist_head *head;
477         struct task_struct *p;
478         pid_t pid = uval & FUTEX_TID_MASK;
479
480         head = &hb->chain;
481
482         plist_for_each_entry_safe(this, next, head, list) {
483                 if (match_futex(&this->key, key)) {
484                         /*
485                          * Another waiter already exists - bump up
486                          * the refcount and return its pi_state:
487                          */
488                         pi_state = this->pi_state;
489                         /*
490                          * Userspace might have messed up non PI and PI futexes
491                          */
492                         if (unlikely(!pi_state))
493                                 return -EINVAL;
494
495                         WARN_ON(!atomic_read(&pi_state->refcount));
496                         WARN_ON(pid && pi_state->owner &&
497                                 pi_state->owner->pid != pid);
498
499                         atomic_inc(&pi_state->refcount);
500                         *ps = pi_state;
501
502                         return 0;
503                 }
504         }
505
506         /*
507          * We are the first waiter - try to look up the real owner and attach
508          * the new pi_state to it, but bail out when TID = 0
509          */
510         if (!pid)
511                 return -ESRCH;
512         p = futex_find_get_task(pid);
513         if (IS_ERR(p))
514                 return PTR_ERR(p);
515
516         /*
517          * We need to look at the task state flags to figure out,
518          * whether the task is exiting. To protect against the do_exit
519          * change of the task flags, we do this protected by
520          * p->pi_lock:
521          */
522         spin_lock_irq(&p->pi_lock);
523         if (unlikely(p->flags & PF_EXITING)) {
524                 /*
525                  * The task is on the way out. When PF_EXITPIDONE is
526                  * set, we know that the task has finished the
527                  * cleanup:
528                  */
529                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
530
531                 spin_unlock_irq(&p->pi_lock);
532                 put_task_struct(p);
533                 return ret;
534         }
535
536         pi_state = alloc_pi_state();
537
538         /*
539          * Initialize the pi_mutex in locked state and make 'p'
540          * the owner of it:
541          */
542         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
543
544         /* Store the key for possible exit cleanups: */
545         pi_state->key = *key;
546
547         WARN_ON(!list_empty(&pi_state->list));
548         list_add(&pi_state->list, &p->pi_state_list);
549         pi_state->owner = p;
550         spin_unlock_irq(&p->pi_lock);
551
552         put_task_struct(p);
553
554         *ps = pi_state;
555
556         return 0;
557 }
558
559 /*
560  * The hash bucket lock must be held when this is called.
561  * Afterwards, the futex_q must not be accessed.
562  */
563 static void wake_futex(struct futex_q *q)
564 {
565         plist_del(&q->list, &q->list.plist);
566         /*
567          * The lock in wake_up_all() is a crucial memory barrier after the
568          * plist_del() and also before assigning to q->lock_ptr.
569          */
570         wake_up(&q->waiter);
571         /*
572          * The waiting task can free the futex_q as soon as this is written,
573          * without taking any locks.  This must come last.
574          *
575          * A memory barrier is required here to prevent the following store to
576          * lock_ptr from getting ahead of the wakeup. Clearing the lock at the
577          * end of wake_up() does not prevent this store from moving.
578          */
579         smp_wmb();
580         q->lock_ptr = NULL;
581 }
582
583 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
584 {
585         struct task_struct *new_owner;
586         struct futex_pi_state *pi_state = this->pi_state;
587         u32 curval, newval;
588
589         if (!pi_state)
590                 return -EINVAL;
591
592         spin_lock(&pi_state->pi_mutex.wait_lock);
593         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
594
595         /*
596          * This happens when we have stolen the lock and the original
597          * pending owner did not enqueue itself back on the rt_mutex.
598          * Thats not a tragedy. We know that way, that a lock waiter
599          * is on the fly. We make the futex_q waiter the pending owner.
600          */
601         if (!new_owner)
602                 new_owner = this->task;
603
604         /*
605          * We pass it to the next owner. (The WAITERS bit is always
606          * kept enabled while there is PI state around. We must also
607          * preserve the owner died bit.)
608          */
609         if (!(uval & FUTEX_OWNER_DIED)) {
610                 int ret = 0;
611
612                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
613
614                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
615
616                 if (curval == -EFAULT)
617                         ret = -EFAULT;
618                 else if (curval != uval)
619                         ret = -EINVAL;
620                 if (ret) {
621                         spin_unlock(&pi_state->pi_mutex.wait_lock);
622                         return ret;
623                 }
624         }
625
626         spin_lock_irq(&pi_state->owner->pi_lock);
627         WARN_ON(list_empty(&pi_state->list));
628         list_del_init(&pi_state->list);
629         spin_unlock_irq(&pi_state->owner->pi_lock);
630
631         spin_lock_irq(&new_owner->pi_lock);
632         WARN_ON(!list_empty(&pi_state->list));
633         list_add(&pi_state->list, &new_owner->pi_state_list);
634         pi_state->owner = new_owner;
635         spin_unlock_irq(&new_owner->pi_lock);
636
637         spin_unlock(&pi_state->pi_mutex.wait_lock);
638         rt_mutex_unlock(&pi_state->pi_mutex);
639
640         return 0;
641 }
642
643 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
644 {
645         u32 oldval;
646
647         /*
648          * There is no waiter, so we unlock the futex. The owner died
649          * bit has not to be preserved here. We are the owner:
650          */
651         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
652
653         if (oldval == -EFAULT)
654                 return oldval;
655         if (oldval != uval)
656                 return -EAGAIN;
657
658         return 0;
659 }
660
661 /*
662  * Express the locking dependencies for lockdep:
663  */
664 static inline void
665 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
666 {
667         if (hb1 <= hb2) {
668                 spin_lock(&hb1->lock);
669                 if (hb1 < hb2)
670                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
671         } else { /* hb1 > hb2 */
672                 spin_lock(&hb2->lock);
673                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
674         }
675 }
676
677 static inline void
678 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
679 {
680         spin_unlock(&hb1->lock);
681         if (hb1 != hb2)
682                 spin_unlock(&hb2->lock);
683 }
684
685 /*
686  * Wake up waiters matching bitset queued on this futex (uaddr).
687  */
688 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
689 {
690         struct futex_hash_bucket *hb;
691         struct futex_q *this, *next;
692         struct plist_head *head;
693         union futex_key key = FUTEX_KEY_INIT;
694         int ret;
695
696         if (!bitset)
697                 return -EINVAL;
698
699         ret = get_futex_key(uaddr, fshared, &key);
700         if (unlikely(ret != 0))
701                 goto out;
702
703         hb = hash_futex(&key);
704         spin_lock(&hb->lock);
705         head = &hb->chain;
706
707         plist_for_each_entry_safe(this, next, head, list) {
708                 if (match_futex (&this->key, &key)) {
709                         if (this->pi_state) {
710                                 ret = -EINVAL;
711                                 break;
712                         }
713
714                         /* Check if one of the bits is set in both bitsets */
715                         if (!(this->bitset & bitset))
716                                 continue;
717
718                         wake_futex(this);
719                         if (++ret >= nr_wake)
720                                 break;
721                 }
722         }
723
724         spin_unlock(&hb->lock);
725         put_futex_key(fshared, &key);
726 out:
727         return ret;
728 }
729
730 /*
731  * Wake up all waiters hashed on the physical page that is mapped
732  * to this virtual address:
733  */
734 static int
735 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
736               int nr_wake, int nr_wake2, int op)
737 {
738         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
739         struct futex_hash_bucket *hb1, *hb2;
740         struct plist_head *head;
741         struct futex_q *this, *next;
742         int ret, op_ret;
743
744 retry:
745         ret = get_futex_key(uaddr1, fshared, &key1);
746         if (unlikely(ret != 0))
747                 goto out;
748         ret = get_futex_key(uaddr2, fshared, &key2);
749         if (unlikely(ret != 0))
750                 goto out_put_key1;
751
752         hb1 = hash_futex(&key1);
753         hb2 = hash_futex(&key2);
754
755         double_lock_hb(hb1, hb2);
756 retry_private:
757         op_ret = futex_atomic_op_inuser(op, uaddr2);
758         if (unlikely(op_ret < 0)) {
759                 u32 dummy;
760
761                 double_unlock_hb(hb1, hb2);
762
763 #ifndef CONFIG_MMU
764                 /*
765                  * we don't get EFAULT from MMU faults if we don't have an MMU,
766                  * but we might get them from range checking
767                  */
768                 ret = op_ret;
769                 goto out_put_keys;
770 #endif
771
772                 if (unlikely(op_ret != -EFAULT)) {
773                         ret = op_ret;
774                         goto out_put_keys;
775                 }
776
777                 ret = get_user(dummy, uaddr2);
778                 if (ret)
779                         goto out_put_keys;
780
781                 if (!fshared)
782                         goto retry_private;
783
784                 put_futex_key(fshared, &key2);
785                 put_futex_key(fshared, &key1);
786                 goto retry;
787         }
788
789         head = &hb1->chain;
790
791         plist_for_each_entry_safe(this, next, head, list) {
792                 if (match_futex (&this->key, &key1)) {
793                         wake_futex(this);
794                         if (++ret >= nr_wake)
795                                 break;
796                 }
797         }
798
799         if (op_ret > 0) {
800                 head = &hb2->chain;
801
802                 op_ret = 0;
803                 plist_for_each_entry_safe(this, next, head, list) {
804                         if (match_futex (&this->key, &key2)) {
805                                 wake_futex(this);
806                                 if (++op_ret >= nr_wake2)
807                                         break;
808                         }
809                 }
810                 ret += op_ret;
811         }
812
813         double_unlock_hb(hb1, hb2);
814 out_put_keys:
815         put_futex_key(fshared, &key2);
816 out_put_key1:
817         put_futex_key(fshared, &key1);
818 out:
819         return ret;
820 }
821
822 /*
823  * Requeue all waiters hashed on one physical page to another
824  * physical page.
825  */
826 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
827                          int nr_wake, int nr_requeue, u32 *cmpval)
828 {
829         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
830         struct futex_hash_bucket *hb1, *hb2;
831         struct plist_head *head1;
832         struct futex_q *this, *next;
833         int ret, drop_count = 0;
834
835 retry:
836         ret = get_futex_key(uaddr1, fshared, &key1);
837         if (unlikely(ret != 0))
838                 goto out;
839         ret = get_futex_key(uaddr2, fshared, &key2);
840         if (unlikely(ret != 0))
841                 goto out_put_key1;
842
843         hb1 = hash_futex(&key1);
844         hb2 = hash_futex(&key2);
845
846 retry_private:
847         double_lock_hb(hb1, hb2);
848
849         if (likely(cmpval != NULL)) {
850                 u32 curval;
851
852                 ret = get_futex_value_locked(&curval, uaddr1);
853
854                 if (unlikely(ret)) {
855                         double_unlock_hb(hb1, hb2);
856
857                         ret = get_user(curval, uaddr1);
858                         if (ret)
859                                 goto out_put_keys;
860
861                         if (!fshared)
862                                 goto retry_private;
863
864                         put_futex_key(fshared, &key2);
865                         put_futex_key(fshared, &key1);
866                         goto retry;
867                 }
868                 if (curval != *cmpval) {
869                         ret = -EAGAIN;
870                         goto out_unlock;
871                 }
872         }
873
874         head1 = &hb1->chain;
875         plist_for_each_entry_safe(this, next, head1, list) {
876                 if (!match_futex (&this->key, &key1))
877                         continue;
878                 if (++ret <= nr_wake) {
879                         wake_futex(this);
880                 } else {
881                         /*
882                          * If key1 and key2 hash to the same bucket, no need to
883                          * requeue.
884                          */
885                         if (likely(head1 != &hb2->chain)) {
886                                 plist_del(&this->list, &hb1->chain);
887                                 plist_add(&this->list, &hb2->chain);
888                                 this->lock_ptr = &hb2->lock;
889 #ifdef CONFIG_DEBUG_PI_LIST
890                                 this->list.plist.lock = &hb2->lock;
891 #endif
892                         }
893                         this->key = key2;
894                         get_futex_key_refs(&key2);
895                         drop_count++;
896
897                         if (ret - nr_wake >= nr_requeue)
898                                 break;
899                 }
900         }
901
902 out_unlock:
903         double_unlock_hb(hb1, hb2);
904
905         /* drop_futex_key_refs() must be called outside the spinlocks. */
906         while (--drop_count >= 0)
907                 drop_futex_key_refs(&key1);
908
909 out_put_keys:
910         put_futex_key(fshared, &key2);
911 out_put_key1:
912         put_futex_key(fshared, &key1);
913 out:
914         return ret;
915 }
916
917 /* The key must be already stored in q->key. */
918 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
919 {
920         struct futex_hash_bucket *hb;
921
922         init_waitqueue_head(&q->waiter);
923
924         get_futex_key_refs(&q->key);
925         hb = hash_futex(&q->key);
926         q->lock_ptr = &hb->lock;
927
928         spin_lock(&hb->lock);
929         return hb;
930 }
931
932 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
933 {
934         int prio;
935
936         /*
937          * The priority used to register this element is
938          * - either the real thread-priority for the real-time threads
939          * (i.e. threads with a priority lower than MAX_RT_PRIO)
940          * - or MAX_RT_PRIO for non-RT threads.
941          * Thus, all RT-threads are woken first in priority order, and
942          * the others are woken last, in FIFO order.
943          */
944         prio = min(current->normal_prio, MAX_RT_PRIO);
945
946         plist_node_init(&q->list, prio);
947 #ifdef CONFIG_DEBUG_PI_LIST
948         q->list.plist.lock = &hb->lock;
949 #endif
950         plist_add(&q->list, &hb->chain);
951         q->task = current;
952         spin_unlock(&hb->lock);
953 }
954
955 static inline void
956 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
957 {
958         spin_unlock(&hb->lock);
959         drop_futex_key_refs(&q->key);
960 }
961
962 /*
963  * queue_me and unqueue_me must be called as a pair, each
964  * exactly once.  They are called with the hashed spinlock held.
965  */
966
967 /* Return 1 if we were still queued (ie. 0 means we were woken) */
968 static int unqueue_me(struct futex_q *q)
969 {
970         spinlock_t *lock_ptr;
971         int ret = 0;
972
973         /* In the common case we don't take the spinlock, which is nice. */
974 retry:
975         lock_ptr = q->lock_ptr;
976         barrier();
977         if (lock_ptr != NULL) {
978                 spin_lock(lock_ptr);
979                 /*
980                  * q->lock_ptr can change between reading it and
981                  * spin_lock(), causing us to take the wrong lock.  This
982                  * corrects the race condition.
983                  *
984                  * Reasoning goes like this: if we have the wrong lock,
985                  * q->lock_ptr must have changed (maybe several times)
986                  * between reading it and the spin_lock().  It can
987                  * change again after the spin_lock() but only if it was
988                  * already changed before the spin_lock().  It cannot,
989                  * however, change back to the original value.  Therefore
990                  * we can detect whether we acquired the correct lock.
991                  */
992                 if (unlikely(lock_ptr != q->lock_ptr)) {
993                         spin_unlock(lock_ptr);
994                         goto retry;
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
1001                 spin_unlock(lock_ptr);
1002                 ret = 1;
1003         }
1004
1005         drop_futex_key_refs(&q->key);
1006         return ret;
1007 }
1008
1009 /*
1010  * PI futexes can not be requeued and must remove themself from the
1011  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1012  * and dropped here.
1013  */
1014 static void unqueue_me_pi(struct futex_q *q)
1015 {
1016         WARN_ON(plist_node_empty(&q->list));
1017         plist_del(&q->list, &q->list.plist);
1018
1019         BUG_ON(!q->pi_state);
1020         free_pi_state(q->pi_state);
1021         q->pi_state = NULL;
1022
1023         spin_unlock(q->lock_ptr);
1024
1025         drop_futex_key_refs(&q->key);
1026 }
1027
1028 /*
1029  * Fixup the pi_state owner with the new owner.
1030  *
1031  * Must be called with hash bucket lock held and mm->sem held for non
1032  * private futexes.
1033  */
1034 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1035                                 struct task_struct *newowner, int fshared)
1036 {
1037         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1038         struct futex_pi_state *pi_state = q->pi_state;
1039         struct task_struct *oldowner = pi_state->owner;
1040         u32 uval, curval, newval;
1041         int ret;
1042
1043         /* Owner died? */
1044         if (!pi_state->owner)
1045                 newtid |= FUTEX_OWNER_DIED;
1046
1047         /*
1048          * We are here either because we stole the rtmutex from the
1049          * pending owner or we are the pending owner which failed to
1050          * get the rtmutex. We have to replace the pending owner TID
1051          * in the user space variable. This must be atomic as we have
1052          * to preserve the owner died bit here.
1053          *
1054          * Note: We write the user space value _before_ changing the pi_state
1055          * because we can fault here. Imagine swapped out pages or a fork
1056          * that marked all the anonymous memory readonly for cow.
1057          *
1058          * Modifying pi_state _before_ the user space value would
1059          * leave the pi_state in an inconsistent state when we fault
1060          * here, because we need to drop the hash bucket lock to
1061          * handle the fault. This might be observed in the PID check
1062          * in lookup_pi_state.
1063          */
1064 retry:
1065         if (get_futex_value_locked(&uval, uaddr))
1066                 goto handle_fault;
1067
1068         while (1) {
1069                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1070
1071                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1072
1073                 if (curval == -EFAULT)
1074                         goto handle_fault;
1075                 if (curval == uval)
1076                         break;
1077                 uval = curval;
1078         }
1079
1080         /*
1081          * We fixed up user space. Now we need to fix the pi_state
1082          * itself.
1083          */
1084         if (pi_state->owner != NULL) {
1085                 spin_lock_irq(&pi_state->owner->pi_lock);
1086                 WARN_ON(list_empty(&pi_state->list));
1087                 list_del_init(&pi_state->list);
1088                 spin_unlock_irq(&pi_state->owner->pi_lock);
1089         }
1090
1091         pi_state->owner = newowner;
1092
1093         spin_lock_irq(&newowner->pi_lock);
1094         WARN_ON(!list_empty(&pi_state->list));
1095         list_add(&pi_state->list, &newowner->pi_state_list);
1096         spin_unlock_irq(&newowner->pi_lock);
1097         return 0;
1098
1099         /*
1100          * To handle the page fault we need to drop the hash bucket
1101          * lock here. That gives the other task (either the pending
1102          * owner itself or the task which stole the rtmutex) the
1103          * chance to try the fixup of the pi_state. So once we are
1104          * back from handling the fault we need to check the pi_state
1105          * after reacquiring the hash bucket lock and before trying to
1106          * do another fixup. When the fixup has been done already we
1107          * simply return.
1108          */
1109 handle_fault:
1110         spin_unlock(q->lock_ptr);
1111
1112         ret = get_user(uval, uaddr);
1113
1114         spin_lock(q->lock_ptr);
1115
1116         /*
1117          * Check if someone else fixed it for us:
1118          */
1119         if (pi_state->owner != oldowner)
1120                 return 0;
1121
1122         if (ret)
1123                 return ret;
1124
1125         goto retry;
1126 }
1127
1128 /*
1129  * In case we must use restart_block to restart a futex_wait,
1130  * we encode in the 'flags' shared capability
1131  */
1132 #define FLAGS_SHARED            0x01
1133 #define FLAGS_CLOCKRT           0x02
1134
1135 static long futex_wait_restart(struct restart_block *restart);
1136
1137 /**
1138  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1139  * @hb:         the futex hash bucket, must be locked by the caller
1140  * @q:          the futex_q to queue up on
1141  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1142  * @wait:       the wait_queue to add to the futex_q after queueing in the hb
1143  */
1144 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1145                                 struct hrtimer_sleeper *timeout,
1146                                 wait_queue_t *wait)
1147 {
1148         queue_me(q, hb);
1149
1150         /*
1151          * There might have been scheduling since the queue_me(), as we
1152          * cannot hold a spinlock across the get_user() in case it
1153          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1154          * queueing ourselves into the futex hash.  This code thus has to
1155          * rely on the futex_wake() code removing us from hash when it
1156          * wakes us up.
1157          */
1158
1159         /* add_wait_queue is the barrier after __set_current_state. */
1160         __set_current_state(TASK_INTERRUPTIBLE);
1161
1162         /*
1163          * Add current as the futex_q waiter.  We don't remove ourselves from
1164          * the wait_queue because we are the only user of it.
1165          */
1166         add_wait_queue(&q->waiter, wait);
1167
1168         /* Arm the timer */
1169         if (timeout) {
1170                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1171                 if (!hrtimer_active(&timeout->timer))
1172                         timeout->task = NULL;
1173         }
1174
1175         /*
1176          * !plist_node_empty() is safe here without any lock.
1177          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1178          */
1179         if (likely(!plist_node_empty(&q->list))) {
1180                 /*
1181                  * If the timer has already expired, current will already be
1182                  * flagged for rescheduling. Only call schedule if there
1183                  * is no timeout, or if it has yet to expire.
1184                  */
1185                 if (!timeout || timeout->task)
1186                         schedule();
1187         }
1188         __set_current_state(TASK_RUNNING);
1189 }
1190
1191 static int futex_wait(u32 __user *uaddr, int fshared,
1192                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1193 {
1194         struct hrtimer_sleeper timeout, *to = NULL;
1195         DECLARE_WAITQUEUE(wait, current);
1196         struct restart_block *restart;
1197         struct futex_hash_bucket *hb;
1198         struct futex_q q;
1199         u32 uval;
1200         int ret;
1201
1202         if (!bitset)
1203                 return -EINVAL;
1204
1205         q.pi_state = NULL;
1206         q.bitset = bitset;
1207
1208         if (abs_time) {
1209                 to = &timeout;
1210
1211                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1212                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1213                 hrtimer_init_sleeper(to, current);
1214                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1215                                              current->timer_slack_ns);
1216         }
1217
1218 retry:
1219         q.key = FUTEX_KEY_INIT;
1220         ret = get_futex_key(uaddr, fshared, &q.key);
1221         if (unlikely(ret != 0))
1222                 goto out;
1223
1224 retry_private:
1225         hb = queue_lock(&q);
1226
1227         /*
1228          * Access the page AFTER the hash-bucket is locked.
1229          * Order is important:
1230          *
1231          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1232          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1233          *
1234          * The basic logical guarantee of a futex is that it blocks ONLY
1235          * if cond(var) is known to be true at the time of blocking, for
1236          * any cond.  If we queued after testing *uaddr, that would open
1237          * a race condition where we could block indefinitely with
1238          * cond(var) false, which would violate the guarantee.
1239          *
1240          * A consequence is that futex_wait() can return zero and absorb
1241          * a wakeup when *uaddr != val on entry to the syscall.  This is
1242          * rare, but normal.
1243          *
1244          * For shared futexes, we hold the mmap semaphore, so the mapping
1245          * cannot have changed since we looked it up in get_futex_key.
1246          */
1247         ret = get_futex_value_locked(&uval, uaddr);
1248
1249         if (unlikely(ret)) {
1250                 queue_unlock(&q, hb);
1251
1252                 ret = get_user(uval, uaddr);
1253                 if (ret)
1254                         goto out_put_key;
1255
1256                 if (!fshared)
1257                         goto retry_private;
1258
1259                 put_futex_key(fshared, &q.key);
1260                 goto retry;
1261         }
1262         ret = -EWOULDBLOCK;
1263
1264         /* Only actually queue if *uaddr contained val.  */
1265         if (unlikely(uval != val)) {
1266                 queue_unlock(&q, hb);
1267                 goto out_put_key;
1268         }
1269
1270         /* queue_me and wait for wakeup, timeout, or a signal. */
1271         futex_wait_queue_me(hb, &q, to, &wait);
1272
1273         /* If we were woken (and unqueued), we succeeded, whatever. */
1274         ret = 0;
1275         if (!unqueue_me(&q))
1276                 goto out_put_key;
1277         ret = -ETIMEDOUT;
1278         if (to && !to->task)
1279                 goto out_put_key;
1280
1281         /*
1282          * We expect signal_pending(current), but another thread may
1283          * have handled it for us already.
1284          */
1285         ret = -ERESTARTSYS;
1286         if (!abs_time)
1287                 goto out_put_key;
1288
1289         restart = &current_thread_info()->restart_block;
1290         restart->fn = futex_wait_restart;
1291         restart->futex.uaddr = (u32 *)uaddr;
1292         restart->futex.val = val;
1293         restart->futex.time = abs_time->tv64;
1294         restart->futex.bitset = bitset;
1295         restart->futex.flags = 0;
1296
1297         if (fshared)
1298                 restart->futex.flags |= FLAGS_SHARED;
1299         if (clockrt)
1300                 restart->futex.flags |= FLAGS_CLOCKRT;
1301
1302         ret = -ERESTART_RESTARTBLOCK;
1303
1304 out_put_key:
1305         put_futex_key(fshared, &q.key);
1306 out:
1307         if (to) {
1308                 hrtimer_cancel(&to->timer);
1309                 destroy_hrtimer_on_stack(&to->timer);
1310         }
1311         return ret;
1312 }
1313
1314
1315 static long futex_wait_restart(struct restart_block *restart)
1316 {
1317         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1318         int fshared = 0;
1319         ktime_t t;
1320
1321         t.tv64 = restart->futex.time;
1322         restart->fn = do_no_restart_syscall;
1323         if (restart->futex.flags & FLAGS_SHARED)
1324                 fshared = 1;
1325         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1326                                 restart->futex.bitset,
1327                                 restart->futex.flags & FLAGS_CLOCKRT);
1328 }
1329
1330
1331 /*
1332  * Userspace tried a 0 -> TID atomic transition of the futex value
1333  * and failed. The kernel side here does the whole locking operation:
1334  * if there are waiters then it will block, it does PI, etc. (Due to
1335  * races the kernel might see a 0 value of the futex too.)
1336  */
1337 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1338                          int detect, ktime_t *time, int trylock)
1339 {
1340         struct hrtimer_sleeper timeout, *to = NULL;
1341         struct task_struct *curr = current;
1342         struct futex_hash_bucket *hb;
1343         u32 uval, newval, curval;
1344         struct futex_q q;
1345         int ret, lock_taken, ownerdied = 0;
1346
1347         if (refill_pi_state_cache())
1348                 return -ENOMEM;
1349
1350         if (time) {
1351                 to = &timeout;
1352                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1353                                       HRTIMER_MODE_ABS);
1354                 hrtimer_init_sleeper(to, current);
1355                 hrtimer_set_expires(&to->timer, *time);
1356         }
1357
1358         q.pi_state = NULL;
1359 retry:
1360         q.key = FUTEX_KEY_INIT;
1361         ret = get_futex_key(uaddr, fshared, &q.key);
1362         if (unlikely(ret != 0))
1363                 goto out;
1364
1365 retry_private:
1366         hb = queue_lock(&q);
1367
1368 retry_locked:
1369         ret = lock_taken = 0;
1370
1371         /*
1372          * To avoid races, we attempt to take the lock here again
1373          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1374          * the locks. It will most likely not succeed.
1375          */
1376         newval = task_pid_vnr(current);
1377
1378         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1379
1380         if (unlikely(curval == -EFAULT))
1381                 goto uaddr_faulted;
1382
1383         /*
1384          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1385          * situation and we return success to user space.
1386          */
1387         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1388                 ret = -EDEADLK;
1389                 goto out_unlock_put_key;
1390         }
1391
1392         /*
1393          * Surprise - we got the lock. Just return to userspace:
1394          */
1395         if (unlikely(!curval))
1396                 goto out_unlock_put_key;
1397
1398         uval = curval;
1399
1400         /*
1401          * Set the WAITERS flag, so the owner will know it has someone
1402          * to wake at next unlock
1403          */
1404         newval = curval | FUTEX_WAITERS;
1405
1406         /*
1407          * There are two cases, where a futex might have no owner (the
1408          * owner TID is 0): OWNER_DIED. We take over the futex in this
1409          * case. We also do an unconditional take over, when the owner
1410          * of the futex died.
1411          *
1412          * This is safe as we are protected by the hash bucket lock !
1413          */
1414         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1415                 /* Keep the OWNER_DIED bit */
1416                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1417                 ownerdied = 0;
1418                 lock_taken = 1;
1419         }
1420
1421         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1422
1423         if (unlikely(curval == -EFAULT))
1424                 goto uaddr_faulted;
1425         if (unlikely(curval != uval))
1426                 goto retry_locked;
1427
1428         /*
1429          * We took the lock due to owner died take over.
1430          */
1431         if (unlikely(lock_taken))
1432                 goto out_unlock_put_key;
1433
1434         /*
1435          * We dont have the lock. Look up the PI state (or create it if
1436          * we are the first waiter):
1437          */
1438         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1439
1440         if (unlikely(ret)) {
1441                 switch (ret) {
1442
1443                 case -EAGAIN:
1444                         /*
1445                          * Task is exiting and we just wait for the
1446                          * exit to complete.
1447                          */
1448                         queue_unlock(&q, hb);
1449                         put_futex_key(fshared, &q.key);
1450                         cond_resched();
1451                         goto retry;
1452
1453                 case -ESRCH:
1454                         /*
1455                          * No owner found for this futex. Check if the
1456                          * OWNER_DIED bit is set to figure out whether
1457                          * this is a robust futex or not.
1458                          */
1459                         if (get_futex_value_locked(&curval, uaddr))
1460                                 goto uaddr_faulted;
1461
1462                         /*
1463                          * We simply start over in case of a robust
1464                          * futex. The code above will take the futex
1465                          * and return happy.
1466                          */
1467                         if (curval & FUTEX_OWNER_DIED) {
1468                                 ownerdied = 1;
1469                                 goto retry_locked;
1470                         }
1471                 default:
1472                         goto out_unlock_put_key;
1473                 }
1474         }
1475
1476         /*
1477          * Only actually queue now that the atomic ops are done:
1478          */
1479         queue_me(&q, hb);
1480
1481         WARN_ON(!q.pi_state);
1482         /*
1483          * Block on the PI mutex:
1484          */
1485         if (!trylock)
1486                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1487         else {
1488                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1489                 /* Fixup the trylock return value: */
1490                 ret = ret ? 0 : -EWOULDBLOCK;
1491         }
1492
1493         spin_lock(q.lock_ptr);
1494
1495         if (!ret) {
1496                 /*
1497                  * Got the lock. We might not be the anticipated owner
1498                  * if we did a lock-steal - fix up the PI-state in
1499                  * that case:
1500                  */
1501                 if (q.pi_state->owner != curr)
1502                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1503         } else {
1504                 /*
1505                  * Catch the rare case, where the lock was released
1506                  * when we were on the way back before we locked the
1507                  * hash bucket.
1508                  */
1509                 if (q.pi_state->owner == curr) {
1510                         /*
1511                          * Try to get the rt_mutex now. This might
1512                          * fail as some other task acquired the
1513                          * rt_mutex after we removed ourself from the
1514                          * rt_mutex waiters list.
1515                          */
1516                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1517                                 ret = 0;
1518                         else {
1519                                 /*
1520                                  * pi_state is incorrect, some other
1521                                  * task did a lock steal and we
1522                                  * returned due to timeout or signal
1523                                  * without taking the rt_mutex. Too
1524                                  * late. We can access the
1525                                  * rt_mutex_owner without locking, as
1526                                  * the other task is now blocked on
1527                                  * the hash bucket lock. Fix the state
1528                                  * up.
1529                                  */
1530                                 struct task_struct *owner;
1531                                 int res;
1532
1533                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1534                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1535                                                            fshared);
1536
1537                                 /* propagate -EFAULT, if the fixup failed */
1538                                 if (res)
1539                                         ret = res;
1540                         }
1541                 } else {
1542                         /*
1543                          * Paranoia check. If we did not take the lock
1544                          * in the trylock above, then we should not be
1545                          * the owner of the rtmutex, neither the real
1546                          * nor the pending one:
1547                          */
1548                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1549                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1550                                        "pi-mutex: %p pi-state %p\n", ret,
1551                                        q.pi_state->pi_mutex.owner,
1552                                        q.pi_state->owner);
1553                 }
1554         }
1555
1556         /*
1557          * If fixup_pi_state_owner() faulted and was unable to handle the
1558          * fault, unlock it and return the fault to userspace.
1559          */
1560         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1561                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1562
1563         /* Unqueue and drop the lock */
1564         unqueue_me_pi(&q);
1565
1566         if (to)
1567                 destroy_hrtimer_on_stack(&to->timer);
1568         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1569
1570 out_unlock_put_key:
1571         queue_unlock(&q, hb);
1572
1573 out_put_key:
1574         put_futex_key(fshared, &q.key);
1575 out:
1576         if (to)
1577                 destroy_hrtimer_on_stack(&to->timer);
1578         return ret;
1579
1580 uaddr_faulted:
1581         /*
1582          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1583          * atomically.  Therefore, if we continue to fault after get_user()
1584          * below, we need to handle the fault ourselves, while still holding
1585          * the mmap_sem.  This can occur if the uaddr is under contention as
1586          * we have to drop the mmap_sem in order to call get_user().
1587          */
1588         queue_unlock(&q, hb);
1589
1590         ret = get_user(uval, uaddr);
1591         if (ret)
1592                 goto out_put_key;
1593
1594         if (!fshared)
1595                 goto retry_private;
1596
1597         put_futex_key(fshared, &q.key);
1598         goto retry;
1599 }
1600
1601
1602 /*
1603  * Userspace attempted a TID -> 0 atomic transition, and failed.
1604  * This is the in-kernel slowpath: we look up the PI state (if any),
1605  * and do the rt-mutex unlock.
1606  */
1607 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1608 {
1609         struct futex_hash_bucket *hb;
1610         struct futex_q *this, *next;
1611         u32 uval;
1612         struct plist_head *head;
1613         union futex_key key = FUTEX_KEY_INIT;
1614         int ret;
1615
1616 retry:
1617         if (get_user(uval, uaddr))
1618                 return -EFAULT;
1619         /*
1620          * We release only a lock we actually own:
1621          */
1622         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1623                 return -EPERM;
1624
1625         ret = get_futex_key(uaddr, fshared, &key);
1626         if (unlikely(ret != 0))
1627                 goto out;
1628
1629         hb = hash_futex(&key);
1630         spin_lock(&hb->lock);
1631
1632         /*
1633          * To avoid races, try to do the TID -> 0 atomic transition
1634          * again. If it succeeds then we can return without waking
1635          * anyone else up:
1636          */
1637         if (!(uval & FUTEX_OWNER_DIED))
1638                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1639
1640
1641         if (unlikely(uval == -EFAULT))
1642                 goto pi_faulted;
1643         /*
1644          * Rare case: we managed to release the lock atomically,
1645          * no need to wake anyone else up:
1646          */
1647         if (unlikely(uval == task_pid_vnr(current)))
1648                 goto out_unlock;
1649
1650         /*
1651          * Ok, other tasks may need to be woken up - check waiters
1652          * and do the wakeup if necessary:
1653          */
1654         head = &hb->chain;
1655
1656         plist_for_each_entry_safe(this, next, head, list) {
1657                 if (!match_futex (&this->key, &key))
1658                         continue;
1659                 ret = wake_futex_pi(uaddr, uval, this);
1660                 /*
1661                  * The atomic access to the futex value
1662                  * generated a pagefault, so retry the
1663                  * user-access and the wakeup:
1664                  */
1665                 if (ret == -EFAULT)
1666                         goto pi_faulted;
1667                 goto out_unlock;
1668         }
1669         /*
1670          * No waiters - kernel unlocks the futex:
1671          */
1672         if (!(uval & FUTEX_OWNER_DIED)) {
1673                 ret = unlock_futex_pi(uaddr, uval);
1674                 if (ret == -EFAULT)
1675                         goto pi_faulted;
1676         }
1677
1678 out_unlock:
1679         spin_unlock(&hb->lock);
1680         put_futex_key(fshared, &key);
1681
1682 out:
1683         return ret;
1684
1685 pi_faulted:
1686         /*
1687          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1688          * atomically.  Therefore, if we continue to fault after get_user()
1689          * below, we need to handle the fault ourselves, while still holding
1690          * the mmap_sem.  This can occur if the uaddr is under contention as
1691          * we have to drop the mmap_sem in order to call get_user().
1692          */
1693         spin_unlock(&hb->lock);
1694         put_futex_key(fshared, &key);
1695
1696         ret = get_user(uval, uaddr);
1697         if (!ret)
1698                 goto retry;
1699
1700         return ret;
1701 }
1702
1703 /*
1704  * Support for robust futexes: the kernel cleans up held futexes at
1705  * thread exit time.
1706  *
1707  * Implementation: user-space maintains a per-thread list of locks it
1708  * is holding. Upon do_exit(), the kernel carefully walks this list,
1709  * and marks all locks that are owned by this thread with the
1710  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1711  * always manipulated with the lock held, so the list is private and
1712  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1713  * field, to allow the kernel to clean up if the thread dies after
1714  * acquiring the lock, but just before it could have added itself to
1715  * the list. There can only be one such pending lock.
1716  */
1717
1718 /**
1719  * sys_set_robust_list - set the robust-futex list head of a task
1720  * @head: pointer to the list-head
1721  * @len: length of the list-head, as userspace expects
1722  */
1723 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
1724                 size_t, len)
1725 {
1726         if (!futex_cmpxchg_enabled)
1727                 return -ENOSYS;
1728         /*
1729          * The kernel knows only one size for now:
1730          */
1731         if (unlikely(len != sizeof(*head)))
1732                 return -EINVAL;
1733
1734         current->robust_list = head;
1735
1736         return 0;
1737 }
1738
1739 /**
1740  * sys_get_robust_list - get the robust-futex list head of a task
1741  * @pid: pid of the process [zero for current task]
1742  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1743  * @len_ptr: pointer to a length field, the kernel fills in the header size
1744  */
1745 SYSCALL_DEFINE3(get_robust_list, int, pid,
1746                 struct robust_list_head __user * __user *, head_ptr,
1747                 size_t __user *, len_ptr)
1748 {
1749         struct robust_list_head __user *head;
1750         unsigned long ret;
1751         const struct cred *cred = current_cred(), *pcred;
1752
1753         if (!futex_cmpxchg_enabled)
1754                 return -ENOSYS;
1755
1756         if (!pid)
1757                 head = current->robust_list;
1758         else {
1759                 struct task_struct *p;
1760
1761                 ret = -ESRCH;
1762                 rcu_read_lock();
1763                 p = find_task_by_vpid(pid);
1764                 if (!p)
1765                         goto err_unlock;
1766                 ret = -EPERM;
1767                 pcred = __task_cred(p);
1768                 if (cred->euid != pcred->euid &&
1769                     cred->euid != pcred->uid &&
1770                     !capable(CAP_SYS_PTRACE))
1771                         goto err_unlock;
1772                 head = p->robust_list;
1773                 rcu_read_unlock();
1774         }
1775
1776         if (put_user(sizeof(*head), len_ptr))
1777                 return -EFAULT;
1778         return put_user(head, head_ptr);
1779
1780 err_unlock:
1781         rcu_read_unlock();
1782
1783         return ret;
1784 }
1785
1786 /*
1787  * Process a futex-list entry, check whether it's owned by the
1788  * dying task, and do notification if so:
1789  */
1790 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1791 {
1792         u32 uval, nval, mval;
1793
1794 retry:
1795         if (get_user(uval, uaddr))
1796                 return -1;
1797
1798         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1799                 /*
1800                  * Ok, this dying thread is truly holding a futex
1801                  * of interest. Set the OWNER_DIED bit atomically
1802                  * via cmpxchg, and if the value had FUTEX_WAITERS
1803                  * set, wake up a waiter (if any). (We have to do a
1804                  * futex_wake() even if OWNER_DIED is already set -
1805                  * to handle the rare but possible case of recursive
1806                  * thread-death.) The rest of the cleanup is done in
1807                  * userspace.
1808                  */
1809                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1810                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1811
1812                 if (nval == -EFAULT)
1813                         return -1;
1814
1815                 if (nval != uval)
1816                         goto retry;
1817
1818                 /*
1819                  * Wake robust non-PI futexes here. The wakeup of
1820                  * PI futexes happens in exit_pi_state():
1821                  */
1822                 if (!pi && (uval & FUTEX_WAITERS))
1823                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1824         }
1825         return 0;
1826 }
1827
1828 /*
1829  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1830  */
1831 static inline int fetch_robust_entry(struct robust_list __user **entry,
1832                                      struct robust_list __user * __user *head,
1833                                      int *pi)
1834 {
1835         unsigned long uentry;
1836
1837         if (get_user(uentry, (unsigned long __user *)head))
1838                 return -EFAULT;
1839
1840         *entry = (void __user *)(uentry & ~1UL);
1841         *pi = uentry & 1;
1842
1843         return 0;
1844 }
1845
1846 /*
1847  * Walk curr->robust_list (very carefully, it's a userspace list!)
1848  * and mark any locks found there dead, and notify any waiters.
1849  *
1850  * We silently return on any sign of list-walking problem.
1851  */
1852 void exit_robust_list(struct task_struct *curr)
1853 {
1854         struct robust_list_head __user *head = curr->robust_list;
1855         struct robust_list __user *entry, *next_entry, *pending;
1856         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1857         unsigned long futex_offset;
1858         int rc;
1859
1860         if (!futex_cmpxchg_enabled)
1861                 return;
1862
1863         /*
1864          * Fetch the list head (which was registered earlier, via
1865          * sys_set_robust_list()):
1866          */
1867         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1868                 return;
1869         /*
1870          * Fetch the relative futex offset:
1871          */
1872         if (get_user(futex_offset, &head->futex_offset))
1873                 return;
1874         /*
1875          * Fetch any possibly pending lock-add first, and handle it
1876          * if it exists:
1877          */
1878         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1879                 return;
1880
1881         next_entry = NULL;      /* avoid warning with gcc */
1882         while (entry != &head->list) {
1883                 /*
1884                  * Fetch the next entry in the list before calling
1885                  * handle_futex_death:
1886                  */
1887                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1888                 /*
1889                  * A pending lock might already be on the list, so
1890                  * don't process it twice:
1891                  */
1892                 if (entry != pending)
1893                         if (handle_futex_death((void __user *)entry + futex_offset,
1894                                                 curr, pi))
1895                                 return;
1896                 if (rc)
1897                         return;
1898                 entry = next_entry;
1899                 pi = next_pi;
1900                 /*
1901                  * Avoid excessively long or circular lists:
1902                  */
1903                 if (!--limit)
1904                         break;
1905
1906                 cond_resched();
1907         }
1908
1909         if (pending)
1910                 handle_futex_death((void __user *)pending + futex_offset,
1911                                    curr, pip);
1912 }
1913
1914 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1915                 u32 __user *uaddr2, u32 val2, u32 val3)
1916 {
1917         int clockrt, ret = -ENOSYS;
1918         int cmd = op & FUTEX_CMD_MASK;
1919         int fshared = 0;
1920
1921         if (!(op & FUTEX_PRIVATE_FLAG))
1922                 fshared = 1;
1923
1924         clockrt = op & FUTEX_CLOCK_REALTIME;
1925         if (clockrt && cmd != FUTEX_WAIT_BITSET)
1926                 return -ENOSYS;
1927
1928         switch (cmd) {
1929         case FUTEX_WAIT:
1930                 val3 = FUTEX_BITSET_MATCH_ANY;
1931         case FUTEX_WAIT_BITSET:
1932                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
1933                 break;
1934         case FUTEX_WAKE:
1935                 val3 = FUTEX_BITSET_MATCH_ANY;
1936         case FUTEX_WAKE_BITSET:
1937                 ret = futex_wake(uaddr, fshared, val, val3);
1938                 break;
1939         case FUTEX_REQUEUE:
1940                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1941                 break;
1942         case FUTEX_CMP_REQUEUE:
1943                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1944                 break;
1945         case FUTEX_WAKE_OP:
1946                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1947                 break;
1948         case FUTEX_LOCK_PI:
1949                 if (futex_cmpxchg_enabled)
1950                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1951                 break;
1952         case FUTEX_UNLOCK_PI:
1953                 if (futex_cmpxchg_enabled)
1954                         ret = futex_unlock_pi(uaddr, fshared);
1955                 break;
1956         case FUTEX_TRYLOCK_PI:
1957                 if (futex_cmpxchg_enabled)
1958                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1959                 break;
1960         default:
1961                 ret = -ENOSYS;
1962         }
1963         return ret;
1964 }
1965
1966
1967 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
1968                 struct timespec __user *, utime, u32 __user *, uaddr2,
1969                 u32, val3)
1970 {
1971         struct timespec ts;
1972         ktime_t t, *tp = NULL;
1973         u32 val2 = 0;
1974         int cmd = op & FUTEX_CMD_MASK;
1975
1976         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
1977                       cmd == FUTEX_WAIT_BITSET)) {
1978                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
1979                         return -EFAULT;
1980                 if (!timespec_valid(&ts))
1981                         return -EINVAL;
1982
1983                 t = timespec_to_ktime(ts);
1984                 if (cmd == FUTEX_WAIT)
1985                         t = ktime_add_safe(ktime_get(), t);
1986                 tp = &t;
1987         }
1988         /*
1989          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
1990          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
1991          */
1992         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
1993             cmd == FUTEX_WAKE_OP)
1994                 val2 = (u32) (unsigned long) utime;
1995
1996         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
1997 }
1998
1999 static int __init futex_init(void)
2000 {
2001         u32 curval;
2002         int i;
2003
2004         /*
2005          * This will fail and we want it. Some arch implementations do
2006          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2007          * functionality. We want to know that before we call in any
2008          * of the complex code paths. Also we want to prevent
2009          * registration of robust lists in that case. NULL is
2010          * guaranteed to fault and we get -EFAULT on functional
2011          * implementation, the non functional ones will return
2012          * -ENOSYS.
2013          */
2014         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2015         if (curval == -EFAULT)
2016                 futex_cmpxchg_enabled = 1;
2017
2018         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2019                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2020                 spin_lock_init(&futex_queues[i].lock);
2021         }
2022
2023         return 0;
2024 }
2025 __initcall(futex_init);