futex: rely on get_user_pages() for shared futexes
[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->waiters, then make the second condition true.
96  */
97 struct futex_q {
98         struct plist_node list;
99         wait_queue_head_t waiters;
100
101         /* Which hash list lock to use: */
102         spinlock_t *lock_ptr;
103
104         /* Key which the futex is hashed on: */
105         union futex_key key;
106
107         /* Optional priority inheritance state: */
108         struct futex_pi_state *pi_state;
109         struct task_struct *task;
110
111         /* Bitset for the optional bitmasked wakeup */
112         u32 bitset;
113 };
114
115 /*
116  * Split the global futex_lock into every hash list lock.
117  */
118 struct futex_hash_bucket {
119         spinlock_t lock;
120         struct plist_head chain;
121 };
122
123 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
124
125 /*
126  * Take mm->mmap_sem, when futex is shared
127  */
128 static inline void futex_lock_mm(struct rw_semaphore *fshared)
129 {
130         if (fshared)
131                 down_read(fshared);
132 }
133
134 /*
135  * Release mm->mmap_sem, when the futex is shared
136  */
137 static inline void futex_unlock_mm(struct rw_semaphore *fshared)
138 {
139         if (fshared)
140                 up_read(fshared);
141 }
142
143 /*
144  * We hash on the keys returned from get_futex_key (see below).
145  */
146 static struct futex_hash_bucket *hash_futex(union futex_key *key)
147 {
148         u32 hash = jhash2((u32*)&key->both.word,
149                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
150                           key->both.offset);
151         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
152 }
153
154 /*
155  * Return 1 if two futex_keys are equal, 0 otherwise.
156  */
157 static inline int match_futex(union futex_key *key1, union futex_key *key2)
158 {
159         return (key1->both.word == key2->both.word
160                 && key1->both.ptr == key2->both.ptr
161                 && key1->both.offset == key2->both.offset);
162 }
163
164 /*
165  * Take a reference to the resource addressed by a key.
166  * Can be called while holding spinlocks.
167  *
168  */
169 static void get_futex_key_refs(union futex_key *key)
170 {
171         if (!key->both.ptr)
172                 return;
173
174         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
175         case FUT_OFF_INODE:
176                 atomic_inc(&key->shared.inode->i_count);
177                 break;
178         case FUT_OFF_MMSHARED:
179                 atomic_inc(&key->private.mm->mm_count);
180                 break;
181         }
182 }
183
184 /*
185  * Drop a reference to the resource addressed by a key.
186  * The hash bucket spinlock must not be held.
187  */
188 static void drop_futex_key_refs(union futex_key *key)
189 {
190         if (!key->both.ptr)
191                 return;
192
193         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
194         case FUT_OFF_INODE:
195                 iput(key->shared.inode);
196                 break;
197         case FUT_OFF_MMSHARED:
198                 mmdrop(key->private.mm);
199                 break;
200         }
201 }
202
203 /**
204  * get_futex_key - Get parameters which are the keys for a futex.
205  * @uaddr: virtual address of the futex
206  * @shared: NULL for a PROCESS_PRIVATE futex,
207  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
208  * @key: address where result is stored.
209  *
210  * Returns a negative error code or 0
211  * The key words are stored in *key on success.
212  *
213  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
214  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
215  * We can usually work out the index without swapping in the page.
216  *
217  * fshared is NULL for PROCESS_PRIVATE futexes
218  * For other futexes, it points to &current->mm->mmap_sem and
219  * caller must have taken the reader lock. but NOT any spinlocks.
220  */
221 static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
222                          union futex_key *key)
223 {
224         unsigned long address = (unsigned long)uaddr;
225         struct mm_struct *mm = current->mm;
226         struct page *page;
227         int err;
228
229         /*
230          * The futex address must be "naturally" aligned.
231          */
232         key->both.offset = address % PAGE_SIZE;
233         if (unlikely((address % sizeof(u32)) != 0))
234                 return -EINVAL;
235         address -= key->both.offset;
236
237         /*
238          * PROCESS_PRIVATE futexes are fast.
239          * As the mm cannot disappear under us and the 'key' only needs
240          * virtual address, we dont even have to find the underlying vma.
241          * Note : We do have to check 'uaddr' is a valid user address,
242          *        but access_ok() should be faster than find_vma()
243          */
244         if (!fshared) {
245                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
246                         return -EFAULT;
247                 key->private.mm = mm;
248                 key->private.address = address;
249                 return 0;
250         }
251
252 again:
253         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
254         if (err < 0)
255                 return err;
256
257         lock_page(page);
258         if (!page->mapping) {
259                 unlock_page(page);
260                 put_page(page);
261                 goto again;
262         }
263
264         /*
265          * Private mappings are handled in a simple way.
266          *
267          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
268          * it's a read-only handle, it's expected that futexes attach to
269          * the object not the particular process.
270          */
271         if (PageAnon(page)) {
272                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
273                 key->private.mm = mm;
274                 key->private.address = address;
275         } else {
276                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
277                 key->shared.inode = page->mapping->host;
278                 key->shared.pgoff = page->index;
279         }
280
281         get_futex_key_refs(key);
282
283         unlock_page(page);
284         put_page(page);
285         return 0;
286 }
287
288 static inline
289 void put_futex_key(struct rw_semaphore *fshared, union futex_key *key)
290 {
291         drop_futex_key_refs(key);
292 }
293
294 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
295 {
296         u32 curval;
297
298         pagefault_disable();
299         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
300         pagefault_enable();
301
302         return curval;
303 }
304
305 static int get_futex_value_locked(u32 *dest, u32 __user *from)
306 {
307         int ret;
308
309         pagefault_disable();
310         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
311         pagefault_enable();
312
313         return ret ? -EFAULT : 0;
314 }
315
316 /*
317  * Fault handling.
318  * if fshared is non NULL, current->mm->mmap_sem is already held
319  */
320 static int futex_handle_fault(unsigned long address,
321                               struct rw_semaphore *fshared, int attempt)
322 {
323         struct vm_area_struct * vma;
324         struct mm_struct *mm = current->mm;
325         int ret = -EFAULT;
326
327         if (attempt > 2)
328                 return ret;
329
330         if (!fshared)
331                 down_read(&mm->mmap_sem);
332         vma = find_vma(mm, address);
333         if (vma && address >= vma->vm_start &&
334             (vma->vm_flags & VM_WRITE)) {
335                 int fault;
336                 fault = handle_mm_fault(mm, vma, address, 1);
337                 if (unlikely((fault & VM_FAULT_ERROR))) {
338 #if 0
339                         /* XXX: let's do this when we verify it is OK */
340                         if (ret & VM_FAULT_OOM)
341                                 ret = -ENOMEM;
342 #endif
343                 } else {
344                         ret = 0;
345                         if (fault & VM_FAULT_MAJOR)
346                                 current->maj_flt++;
347                         else
348                                 current->min_flt++;
349                 }
350         }
351         if (!fshared)
352                 up_read(&mm->mmap_sem);
353         return ret;
354 }
355
356 /*
357  * PI code:
358  */
359 static int refill_pi_state_cache(void)
360 {
361         struct futex_pi_state *pi_state;
362
363         if (likely(current->pi_state_cache))
364                 return 0;
365
366         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
367
368         if (!pi_state)
369                 return -ENOMEM;
370
371         INIT_LIST_HEAD(&pi_state->list);
372         /* pi_mutex gets initialized later */
373         pi_state->owner = NULL;
374         atomic_set(&pi_state->refcount, 1);
375         pi_state->key = FUTEX_KEY_INIT;
376
377         current->pi_state_cache = pi_state;
378
379         return 0;
380 }
381
382 static struct futex_pi_state * alloc_pi_state(void)
383 {
384         struct futex_pi_state *pi_state = current->pi_state_cache;
385
386         WARN_ON(!pi_state);
387         current->pi_state_cache = NULL;
388
389         return pi_state;
390 }
391
392 static void free_pi_state(struct futex_pi_state *pi_state)
393 {
394         if (!atomic_dec_and_test(&pi_state->refcount))
395                 return;
396
397         /*
398          * If pi_state->owner is NULL, the owner is most probably dying
399          * and has cleaned up the pi_state already
400          */
401         if (pi_state->owner) {
402                 spin_lock_irq(&pi_state->owner->pi_lock);
403                 list_del_init(&pi_state->list);
404                 spin_unlock_irq(&pi_state->owner->pi_lock);
405
406                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
407         }
408
409         if (current->pi_state_cache)
410                 kfree(pi_state);
411         else {
412                 /*
413                  * pi_state->list is already empty.
414                  * clear pi_state->owner.
415                  * refcount is at 0 - put it back to 1.
416                  */
417                 pi_state->owner = NULL;
418                 atomic_set(&pi_state->refcount, 1);
419                 current->pi_state_cache = pi_state;
420         }
421 }
422
423 /*
424  * Look up the task based on what TID userspace gave us.
425  * We dont trust it.
426  */
427 static struct task_struct * futex_find_get_task(pid_t pid)
428 {
429         struct task_struct *p;
430
431         rcu_read_lock();
432         p = find_task_by_vpid(pid);
433         if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
434                 p = ERR_PTR(-ESRCH);
435         else
436                 get_task_struct(p);
437
438         rcu_read_unlock();
439
440         return p;
441 }
442
443 /*
444  * This task is holding PI mutexes at exit time => bad.
445  * Kernel cleans up PI-state, but userspace is likely hosed.
446  * (Robust-futex cleanup is separate and might save the day for userspace.)
447  */
448 void exit_pi_state_list(struct task_struct *curr)
449 {
450         struct list_head *next, *head = &curr->pi_state_list;
451         struct futex_pi_state *pi_state;
452         struct futex_hash_bucket *hb;
453         union futex_key key = FUTEX_KEY_INIT;
454
455         if (!futex_cmpxchg_enabled)
456                 return;
457         /*
458          * We are a ZOMBIE and nobody can enqueue itself on
459          * pi_state_list anymore, but we have to be careful
460          * versus waiters unqueueing themselves:
461          */
462         spin_lock_irq(&curr->pi_lock);
463         while (!list_empty(head)) {
464
465                 next = head->next;
466                 pi_state = list_entry(next, struct futex_pi_state, list);
467                 key = pi_state->key;
468                 hb = hash_futex(&key);
469                 spin_unlock_irq(&curr->pi_lock);
470
471                 spin_lock(&hb->lock);
472
473                 spin_lock_irq(&curr->pi_lock);
474                 /*
475                  * We dropped the pi-lock, so re-check whether this
476                  * task still owns the PI-state:
477                  */
478                 if (head->next != next) {
479                         spin_unlock(&hb->lock);
480                         continue;
481                 }
482
483                 WARN_ON(pi_state->owner != curr);
484                 WARN_ON(list_empty(&pi_state->list));
485                 list_del_init(&pi_state->list);
486                 pi_state->owner = NULL;
487                 spin_unlock_irq(&curr->pi_lock);
488
489                 rt_mutex_unlock(&pi_state->pi_mutex);
490
491                 spin_unlock(&hb->lock);
492
493                 spin_lock_irq(&curr->pi_lock);
494         }
495         spin_unlock_irq(&curr->pi_lock);
496 }
497
498 static int
499 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
500                 union futex_key *key, struct futex_pi_state **ps)
501 {
502         struct futex_pi_state *pi_state = NULL;
503         struct futex_q *this, *next;
504         struct plist_head *head;
505         struct task_struct *p;
506         pid_t pid = uval & FUTEX_TID_MASK;
507
508         head = &hb->chain;
509
510         plist_for_each_entry_safe(this, next, head, list) {
511                 if (match_futex(&this->key, key)) {
512                         /*
513                          * Another waiter already exists - bump up
514                          * the refcount and return its pi_state:
515                          */
516                         pi_state = this->pi_state;
517                         /*
518                          * Userspace might have messed up non PI and PI futexes
519                          */
520                         if (unlikely(!pi_state))
521                                 return -EINVAL;
522
523                         WARN_ON(!atomic_read(&pi_state->refcount));
524                         WARN_ON(pid && pi_state->owner &&
525                                 pi_state->owner->pid != pid);
526
527                         atomic_inc(&pi_state->refcount);
528                         *ps = pi_state;
529
530                         return 0;
531                 }
532         }
533
534         /*
535          * We are the first waiter - try to look up the real owner and attach
536          * the new pi_state to it, but bail out when TID = 0
537          */
538         if (!pid)
539                 return -ESRCH;
540         p = futex_find_get_task(pid);
541         if (IS_ERR(p))
542                 return PTR_ERR(p);
543
544         /*
545          * We need to look at the task state flags to figure out,
546          * whether the task is exiting. To protect against the do_exit
547          * change of the task flags, we do this protected by
548          * p->pi_lock:
549          */
550         spin_lock_irq(&p->pi_lock);
551         if (unlikely(p->flags & PF_EXITING)) {
552                 /*
553                  * The task is on the way out. When PF_EXITPIDONE is
554                  * set, we know that the task has finished the
555                  * cleanup:
556                  */
557                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
558
559                 spin_unlock_irq(&p->pi_lock);
560                 put_task_struct(p);
561                 return ret;
562         }
563
564         pi_state = alloc_pi_state();
565
566         /*
567          * Initialize the pi_mutex in locked state and make 'p'
568          * the owner of it:
569          */
570         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
571
572         /* Store the key for possible exit cleanups: */
573         pi_state->key = *key;
574
575         WARN_ON(!list_empty(&pi_state->list));
576         list_add(&pi_state->list, &p->pi_state_list);
577         pi_state->owner = p;
578         spin_unlock_irq(&p->pi_lock);
579
580         put_task_struct(p);
581
582         *ps = pi_state;
583
584         return 0;
585 }
586
587 /*
588  * The hash bucket lock must be held when this is called.
589  * Afterwards, the futex_q must not be accessed.
590  */
591 static void wake_futex(struct futex_q *q)
592 {
593         plist_del(&q->list, &q->list.plist);
594         /*
595          * The lock in wake_up_all() is a crucial memory barrier after the
596          * plist_del() and also before assigning to q->lock_ptr.
597          */
598         wake_up_all(&q->waiters);
599         /*
600          * The waiting task can free the futex_q as soon as this is written,
601          * without taking any locks.  This must come last.
602          *
603          * A memory barrier is required here to prevent the following store
604          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
605          * at the end of wake_up_all() does not prevent this store from
606          * moving.
607          */
608         smp_wmb();
609         q->lock_ptr = NULL;
610 }
611
612 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
613 {
614         struct task_struct *new_owner;
615         struct futex_pi_state *pi_state = this->pi_state;
616         u32 curval, newval;
617
618         if (!pi_state)
619                 return -EINVAL;
620
621         spin_lock(&pi_state->pi_mutex.wait_lock);
622         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
623
624         /*
625          * This happens when we have stolen the lock and the original
626          * pending owner did not enqueue itself back on the rt_mutex.
627          * Thats not a tragedy. We know that way, that a lock waiter
628          * is on the fly. We make the futex_q waiter the pending owner.
629          */
630         if (!new_owner)
631                 new_owner = this->task;
632
633         /*
634          * We pass it to the next owner. (The WAITERS bit is always
635          * kept enabled while there is PI state around. We must also
636          * preserve the owner died bit.)
637          */
638         if (!(uval & FUTEX_OWNER_DIED)) {
639                 int ret = 0;
640
641                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
642
643                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
644
645                 if (curval == -EFAULT)
646                         ret = -EFAULT;
647                 else if (curval != uval)
648                         ret = -EINVAL;
649                 if (ret) {
650                         spin_unlock(&pi_state->pi_mutex.wait_lock);
651                         return ret;
652                 }
653         }
654
655         spin_lock_irq(&pi_state->owner->pi_lock);
656         WARN_ON(list_empty(&pi_state->list));
657         list_del_init(&pi_state->list);
658         spin_unlock_irq(&pi_state->owner->pi_lock);
659
660         spin_lock_irq(&new_owner->pi_lock);
661         WARN_ON(!list_empty(&pi_state->list));
662         list_add(&pi_state->list, &new_owner->pi_state_list);
663         pi_state->owner = new_owner;
664         spin_unlock_irq(&new_owner->pi_lock);
665
666         spin_unlock(&pi_state->pi_mutex.wait_lock);
667         rt_mutex_unlock(&pi_state->pi_mutex);
668
669         return 0;
670 }
671
672 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
673 {
674         u32 oldval;
675
676         /*
677          * There is no waiter, so we unlock the futex. The owner died
678          * bit has not to be preserved here. We are the owner:
679          */
680         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
681
682         if (oldval == -EFAULT)
683                 return oldval;
684         if (oldval != uval)
685                 return -EAGAIN;
686
687         return 0;
688 }
689
690 /*
691  * Express the locking dependencies for lockdep:
692  */
693 static inline void
694 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
695 {
696         if (hb1 <= hb2) {
697                 spin_lock(&hb1->lock);
698                 if (hb1 < hb2)
699                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
700         } else { /* hb1 > hb2 */
701                 spin_lock(&hb2->lock);
702                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
703         }
704 }
705
706 /*
707  * Wake up all waiters hashed on the physical page that is mapped
708  * to this virtual address:
709  */
710 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
711                       int nr_wake, u32 bitset)
712 {
713         struct futex_hash_bucket *hb;
714         struct futex_q *this, *next;
715         struct plist_head *head;
716         union futex_key key = FUTEX_KEY_INIT;
717         int ret;
718
719         if (!bitset)
720                 return -EINVAL;
721
722         futex_lock_mm(fshared);
723
724         ret = get_futex_key(uaddr, fshared, &key);
725         if (unlikely(ret != 0))
726                 goto out;
727
728         hb = hash_futex(&key);
729         spin_lock(&hb->lock);
730         head = &hb->chain;
731
732         plist_for_each_entry_safe(this, next, head, list) {
733                 if (match_futex (&this->key, &key)) {
734                         if (this->pi_state) {
735                                 ret = -EINVAL;
736                                 break;
737                         }
738
739                         /* Check if one of the bits is set in both bitsets */
740                         if (!(this->bitset & bitset))
741                                 continue;
742
743                         wake_futex(this);
744                         if (++ret >= nr_wake)
745                                 break;
746                 }
747         }
748
749         spin_unlock(&hb->lock);
750 out:
751         put_futex_key(fshared, &key);
752         futex_unlock_mm(fshared);
753         return ret;
754 }
755
756 /*
757  * Wake up all waiters hashed on the physical page that is mapped
758  * to this virtual address:
759  */
760 static int
761 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
762               u32 __user *uaddr2,
763               int nr_wake, int nr_wake2, int op)
764 {
765         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
766         struct futex_hash_bucket *hb1, *hb2;
767         struct plist_head *head;
768         struct futex_q *this, *next;
769         int ret, op_ret, attempt = 0;
770
771 retryfull:
772         futex_lock_mm(fshared);
773
774         ret = get_futex_key(uaddr1, fshared, &key1);
775         if (unlikely(ret != 0))
776                 goto out;
777         ret = get_futex_key(uaddr2, fshared, &key2);
778         if (unlikely(ret != 0))
779                 goto out;
780
781         hb1 = hash_futex(&key1);
782         hb2 = hash_futex(&key2);
783
784 retry:
785         double_lock_hb(hb1, hb2);
786
787         op_ret = futex_atomic_op_inuser(op, uaddr2);
788         if (unlikely(op_ret < 0)) {
789                 u32 dummy;
790
791                 spin_unlock(&hb1->lock);
792                 if (hb1 != hb2)
793                         spin_unlock(&hb2->lock);
794
795 #ifndef CONFIG_MMU
796                 /*
797                  * we don't get EFAULT from MMU faults if we don't have an MMU,
798                  * but we might get them from range checking
799                  */
800                 ret = op_ret;
801                 goto out;
802 #endif
803
804                 if (unlikely(op_ret != -EFAULT)) {
805                         ret = op_ret;
806                         goto out;
807                 }
808
809                 /*
810                  * futex_atomic_op_inuser needs to both read and write
811                  * *(int __user *)uaddr2, but we can't modify it
812                  * non-atomically.  Therefore, if get_user below is not
813                  * enough, we need to handle the fault ourselves, while
814                  * still holding the mmap_sem.
815                  */
816                 if (attempt++) {
817                         ret = futex_handle_fault((unsigned long)uaddr2,
818                                                  fshared, attempt);
819                         if (ret)
820                                 goto out;
821                         goto retry;
822                 }
823
824                 /*
825                  * If we would have faulted, release mmap_sem,
826                  * fault it in and start all over again.
827                  */
828                 futex_unlock_mm(fshared);
829
830                 ret = get_user(dummy, uaddr2);
831                 if (ret)
832                         return ret;
833
834                 goto retryfull;
835         }
836
837         head = &hb1->chain;
838
839         plist_for_each_entry_safe(this, next, head, list) {
840                 if (match_futex (&this->key, &key1)) {
841                         wake_futex(this);
842                         if (++ret >= nr_wake)
843                                 break;
844                 }
845         }
846
847         if (op_ret > 0) {
848                 head = &hb2->chain;
849
850                 op_ret = 0;
851                 plist_for_each_entry_safe(this, next, head, list) {
852                         if (match_futex (&this->key, &key2)) {
853                                 wake_futex(this);
854                                 if (++op_ret >= nr_wake2)
855                                         break;
856                         }
857                 }
858                 ret += op_ret;
859         }
860
861         spin_unlock(&hb1->lock);
862         if (hb1 != hb2)
863                 spin_unlock(&hb2->lock);
864 out:
865         put_futex_key(fshared, &key2);
866         put_futex_key(fshared, &key1);
867         futex_unlock_mm(fshared);
868
869         return ret;
870 }
871
872 /*
873  * Requeue all waiters hashed on one physical page to another
874  * physical page.
875  */
876 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
877                          u32 __user *uaddr2,
878                          int nr_wake, int nr_requeue, u32 *cmpval)
879 {
880         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
881         struct futex_hash_bucket *hb1, *hb2;
882         struct plist_head *head1;
883         struct futex_q *this, *next;
884         int ret, drop_count = 0;
885
886  retry:
887         futex_lock_mm(fshared);
888
889         ret = get_futex_key(uaddr1, fshared, &key1);
890         if (unlikely(ret != 0))
891                 goto out;
892         ret = get_futex_key(uaddr2, fshared, &key2);
893         if (unlikely(ret != 0))
894                 goto out;
895
896         hb1 = hash_futex(&key1);
897         hb2 = hash_futex(&key2);
898
899         double_lock_hb(hb1, hb2);
900
901         if (likely(cmpval != NULL)) {
902                 u32 curval;
903
904                 ret = get_futex_value_locked(&curval, uaddr1);
905
906                 if (unlikely(ret)) {
907                         spin_unlock(&hb1->lock);
908                         if (hb1 != hb2)
909                                 spin_unlock(&hb2->lock);
910
911                         /*
912                          * If we would have faulted, release mmap_sem, fault
913                          * it in and start all over again.
914                          */
915                         futex_unlock_mm(fshared);
916
917                         ret = get_user(curval, uaddr1);
918
919                         if (!ret)
920                                 goto retry;
921
922                         return ret;
923                 }
924                 if (curval != *cmpval) {
925                         ret = -EAGAIN;
926                         goto out_unlock;
927                 }
928         }
929
930         head1 = &hb1->chain;
931         plist_for_each_entry_safe(this, next, head1, list) {
932                 if (!match_futex (&this->key, &key1))
933                         continue;
934                 if (++ret <= nr_wake) {
935                         wake_futex(this);
936                 } else {
937                         /*
938                          * If key1 and key2 hash to the same bucket, no need to
939                          * requeue.
940                          */
941                         if (likely(head1 != &hb2->chain)) {
942                                 plist_del(&this->list, &hb1->chain);
943                                 plist_add(&this->list, &hb2->chain);
944                                 this->lock_ptr = &hb2->lock;
945 #ifdef CONFIG_DEBUG_PI_LIST
946                                 this->list.plist.lock = &hb2->lock;
947 #endif
948                         }
949                         this->key = key2;
950                         get_futex_key_refs(&key2);
951                         drop_count++;
952
953                         if (ret - nr_wake >= nr_requeue)
954                                 break;
955                 }
956         }
957
958 out_unlock:
959         spin_unlock(&hb1->lock);
960         if (hb1 != hb2)
961                 spin_unlock(&hb2->lock);
962
963         /* drop_futex_key_refs() must be called outside the spinlocks. */
964         while (--drop_count >= 0)
965                 drop_futex_key_refs(&key1);
966
967 out:
968         put_futex_key(fshared, &key2);
969         put_futex_key(fshared, &key1);
970         futex_unlock_mm(fshared);
971         return ret;
972 }
973
974 /* The key must be already stored in q->key. */
975 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
976 {
977         struct futex_hash_bucket *hb;
978
979         init_waitqueue_head(&q->waiters);
980
981         get_futex_key_refs(&q->key);
982         hb = hash_futex(&q->key);
983         q->lock_ptr = &hb->lock;
984
985         spin_lock(&hb->lock);
986         return hb;
987 }
988
989 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
990 {
991         int prio;
992
993         /*
994          * The priority used to register this element is
995          * - either the real thread-priority for the real-time threads
996          * (i.e. threads with a priority lower than MAX_RT_PRIO)
997          * - or MAX_RT_PRIO for non-RT threads.
998          * Thus, all RT-threads are woken first in priority order, and
999          * the others are woken last, in FIFO order.
1000          */
1001         prio = min(current->normal_prio, MAX_RT_PRIO);
1002
1003         plist_node_init(&q->list, prio);
1004 #ifdef CONFIG_DEBUG_PI_LIST
1005         q->list.plist.lock = &hb->lock;
1006 #endif
1007         plist_add(&q->list, &hb->chain);
1008         q->task = current;
1009         spin_unlock(&hb->lock);
1010 }
1011
1012 static inline void
1013 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1014 {
1015         spin_unlock(&hb->lock);
1016         drop_futex_key_refs(&q->key);
1017 }
1018
1019 /*
1020  * queue_me and unqueue_me must be called as a pair, each
1021  * exactly once.  They are called with the hashed spinlock held.
1022  */
1023
1024 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1025 static int unqueue_me(struct futex_q *q)
1026 {
1027         spinlock_t *lock_ptr;
1028         int ret = 0;
1029
1030         /* In the common case we don't take the spinlock, which is nice. */
1031  retry:
1032         lock_ptr = q->lock_ptr;
1033         barrier();
1034         if (lock_ptr != NULL) {
1035                 spin_lock(lock_ptr);
1036                 /*
1037                  * q->lock_ptr can change between reading it and
1038                  * spin_lock(), causing us to take the wrong lock.  This
1039                  * corrects the race condition.
1040                  *
1041                  * Reasoning goes like this: if we have the wrong lock,
1042                  * q->lock_ptr must have changed (maybe several times)
1043                  * between reading it and the spin_lock().  It can
1044                  * change again after the spin_lock() but only if it was
1045                  * already changed before the spin_lock().  It cannot,
1046                  * however, change back to the original value.  Therefore
1047                  * we can detect whether we acquired the correct lock.
1048                  */
1049                 if (unlikely(lock_ptr != q->lock_ptr)) {
1050                         spin_unlock(lock_ptr);
1051                         goto retry;
1052                 }
1053                 WARN_ON(plist_node_empty(&q->list));
1054                 plist_del(&q->list, &q->list.plist);
1055
1056                 BUG_ON(q->pi_state);
1057
1058                 spin_unlock(lock_ptr);
1059                 ret = 1;
1060         }
1061
1062         drop_futex_key_refs(&q->key);
1063         return ret;
1064 }
1065
1066 /*
1067  * PI futexes can not be requeued and must remove themself from the
1068  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1069  * and dropped here.
1070  */
1071 static void unqueue_me_pi(struct futex_q *q)
1072 {
1073         WARN_ON(plist_node_empty(&q->list));
1074         plist_del(&q->list, &q->list.plist);
1075
1076         BUG_ON(!q->pi_state);
1077         free_pi_state(q->pi_state);
1078         q->pi_state = NULL;
1079
1080         spin_unlock(q->lock_ptr);
1081
1082         drop_futex_key_refs(&q->key);
1083 }
1084
1085 /*
1086  * Fixup the pi_state owner with the new owner.
1087  *
1088  * Must be called with hash bucket lock held and mm->sem held for non
1089  * private futexes.
1090  */
1091 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1092                                 struct task_struct *newowner,
1093                                 struct rw_semaphore *fshared)
1094 {
1095         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1096         struct futex_pi_state *pi_state = q->pi_state;
1097         struct task_struct *oldowner = pi_state->owner;
1098         u32 uval, curval, newval;
1099         int ret, attempt = 0;
1100
1101         /* Owner died? */
1102         if (!pi_state->owner)
1103                 newtid |= FUTEX_OWNER_DIED;
1104
1105         /*
1106          * We are here either because we stole the rtmutex from the
1107          * pending owner or we are the pending owner which failed to
1108          * get the rtmutex. We have to replace the pending owner TID
1109          * in the user space variable. This must be atomic as we have
1110          * to preserve the owner died bit here.
1111          *
1112          * Note: We write the user space value _before_ changing the
1113          * pi_state because we can fault here. Imagine swapped out
1114          * pages or a fork, which was running right before we acquired
1115          * mmap_sem, that marked all the anonymous memory readonly for
1116          * cow.
1117          *
1118          * Modifying pi_state _before_ the user space value would
1119          * leave the pi_state in an inconsistent state when we fault
1120          * here, because we need to drop the hash bucket lock to
1121          * handle the fault. This might be observed in the PID check
1122          * in lookup_pi_state.
1123          */
1124 retry:
1125         if (get_futex_value_locked(&uval, uaddr))
1126                 goto handle_fault;
1127
1128         while (1) {
1129                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1130
1131                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1132
1133                 if (curval == -EFAULT)
1134                         goto handle_fault;
1135                 if (curval == uval)
1136                         break;
1137                 uval = curval;
1138         }
1139
1140         /*
1141          * We fixed up user space. Now we need to fix the pi_state
1142          * itself.
1143          */
1144         if (pi_state->owner != NULL) {
1145                 spin_lock_irq(&pi_state->owner->pi_lock);
1146                 WARN_ON(list_empty(&pi_state->list));
1147                 list_del_init(&pi_state->list);
1148                 spin_unlock_irq(&pi_state->owner->pi_lock);
1149         }
1150
1151         pi_state->owner = newowner;
1152
1153         spin_lock_irq(&newowner->pi_lock);
1154         WARN_ON(!list_empty(&pi_state->list));
1155         list_add(&pi_state->list, &newowner->pi_state_list);
1156         spin_unlock_irq(&newowner->pi_lock);
1157         return 0;
1158
1159         /*
1160          * To handle the page fault we need to drop the hash bucket
1161          * lock here. That gives the other task (either the pending
1162          * owner itself or the task which stole the rtmutex) the
1163          * chance to try the fixup of the pi_state. So once we are
1164          * back from handling the fault we need to check the pi_state
1165          * after reacquiring the hash bucket lock and before trying to
1166          * do another fixup. When the fixup has been done already we
1167          * simply return.
1168          */
1169 handle_fault:
1170         spin_unlock(q->lock_ptr);
1171
1172         ret = futex_handle_fault((unsigned long)uaddr, fshared, attempt++);
1173
1174         spin_lock(q->lock_ptr);
1175
1176         /*
1177          * Check if someone else fixed it for us:
1178          */
1179         if (pi_state->owner != oldowner)
1180                 return 0;
1181
1182         if (ret)
1183                 return ret;
1184
1185         goto retry;
1186 }
1187
1188 /*
1189  * In case we must use restart_block to restart a futex_wait,
1190  * we encode in the 'flags' shared capability
1191  */
1192 #define FLAGS_SHARED  1
1193
1194 static long futex_wait_restart(struct restart_block *restart);
1195
1196 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1197                       u32 val, ktime_t *abs_time, u32 bitset)
1198 {
1199         struct task_struct *curr = current;
1200         DECLARE_WAITQUEUE(wait, curr);
1201         struct futex_hash_bucket *hb;
1202         struct futex_q q;
1203         u32 uval;
1204         int ret;
1205         struct hrtimer_sleeper t;
1206         int rem = 0;
1207
1208         if (!bitset)
1209                 return -EINVAL;
1210
1211         q.pi_state = NULL;
1212         q.bitset = bitset;
1213  retry:
1214         futex_lock_mm(fshared);
1215
1216         q.key = FUTEX_KEY_INIT;
1217         ret = get_futex_key(uaddr, fshared, &q.key);
1218         if (unlikely(ret != 0))
1219                 goto out_release_sem;
1220
1221         hb = queue_lock(&q);
1222
1223         /*
1224          * Access the page AFTER the futex is queued.
1225          * Order is important:
1226          *
1227          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1228          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1229          *
1230          * The basic logical guarantee of a futex is that it blocks ONLY
1231          * if cond(var) is known to be true at the time of blocking, for
1232          * any cond.  If we queued after testing *uaddr, that would open
1233          * a race condition where we could block indefinitely with
1234          * cond(var) false, which would violate the guarantee.
1235          *
1236          * A consequence is that futex_wait() can return zero and absorb
1237          * a wakeup when *uaddr != val on entry to the syscall.  This is
1238          * rare, but normal.
1239          *
1240          * for shared futexes, we hold the mmap semaphore, so the mapping
1241          * cannot have changed since we looked it up in get_futex_key.
1242          */
1243         ret = get_futex_value_locked(&uval, uaddr);
1244
1245         if (unlikely(ret)) {
1246                 queue_unlock(&q, hb);
1247
1248                 /*
1249                  * If we would have faulted, release mmap_sem, fault it in and
1250                  * start all over again.
1251                  */
1252                 futex_unlock_mm(fshared);
1253
1254                 ret = get_user(uval, uaddr);
1255
1256                 if (!ret)
1257                         goto retry;
1258                 return ret;
1259         }
1260         ret = -EWOULDBLOCK;
1261         if (uval != val)
1262                 goto out_unlock_release_sem;
1263
1264         /* Only actually queue if *uaddr contained val.  */
1265         queue_me(&q, hb);
1266
1267         /*
1268          * Now the futex is queued and we have checked the data, we
1269          * don't want to hold mmap_sem while we sleep.
1270          */
1271         futex_unlock_mm(fshared);
1272
1273         /*
1274          * There might have been scheduling since the queue_me(), as we
1275          * cannot hold a spinlock across the get_user() in case it
1276          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1277          * queueing ourselves into the futex hash.  This code thus has to
1278          * rely on the futex_wake() code removing us from hash when it
1279          * wakes us up.
1280          */
1281
1282         /* add_wait_queue is the barrier after __set_current_state. */
1283         __set_current_state(TASK_INTERRUPTIBLE);
1284         add_wait_queue(&q.waiters, &wait);
1285         /*
1286          * !plist_node_empty() is safe here without any lock.
1287          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1288          */
1289         if (likely(!plist_node_empty(&q.list))) {
1290                 if (!abs_time)
1291                         schedule();
1292                 else {
1293                         hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1294                                                 HRTIMER_MODE_ABS);
1295                         hrtimer_init_sleeper(&t, current);
1296                         t.timer.expires = *abs_time;
1297
1298                         hrtimer_start(&t.timer, t.timer.expires,
1299                                                 HRTIMER_MODE_ABS);
1300                         if (!hrtimer_active(&t.timer))
1301                                 t.task = NULL;
1302
1303                         /*
1304                          * the timer could have already expired, in which
1305                          * case current would be flagged for rescheduling.
1306                          * Don't bother calling schedule.
1307                          */
1308                         if (likely(t.task))
1309                                 schedule();
1310
1311                         hrtimer_cancel(&t.timer);
1312
1313                         /* Flag if a timeout occured */
1314                         rem = (t.task == NULL);
1315
1316                         destroy_hrtimer_on_stack(&t.timer);
1317                 }
1318         }
1319         __set_current_state(TASK_RUNNING);
1320
1321         /*
1322          * NOTE: we don't remove ourselves from the waitqueue because
1323          * we are the only user of it.
1324          */
1325
1326         /* If we were woken (and unqueued), we succeeded, whatever. */
1327         if (!unqueue_me(&q))
1328                 return 0;
1329         if (rem)
1330                 return -ETIMEDOUT;
1331
1332         /*
1333          * We expect signal_pending(current), but another thread may
1334          * have handled it for us already.
1335          */
1336         if (!abs_time)
1337                 return -ERESTARTSYS;
1338         else {
1339                 struct restart_block *restart;
1340                 restart = &current_thread_info()->restart_block;
1341                 restart->fn = futex_wait_restart;
1342                 restart->futex.uaddr = (u32 *)uaddr;
1343                 restart->futex.val = val;
1344                 restart->futex.time = abs_time->tv64;
1345                 restart->futex.bitset = bitset;
1346                 restart->futex.flags = 0;
1347
1348                 if (fshared)
1349                         restart->futex.flags |= FLAGS_SHARED;
1350                 return -ERESTART_RESTARTBLOCK;
1351         }
1352
1353  out_unlock_release_sem:
1354         queue_unlock(&q, hb);
1355
1356  out_release_sem:
1357         put_futex_key(fshared, &q.key);
1358         futex_unlock_mm(fshared);
1359         return ret;
1360 }
1361
1362
1363 static long futex_wait_restart(struct restart_block *restart)
1364 {
1365         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1366         struct rw_semaphore *fshared = NULL;
1367         ktime_t t;
1368
1369         t.tv64 = restart->futex.time;
1370         restart->fn = do_no_restart_syscall;
1371         if (restart->futex.flags & FLAGS_SHARED)
1372                 fshared = &current->mm->mmap_sem;
1373         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1374                                 restart->futex.bitset);
1375 }
1376
1377
1378 /*
1379  * Userspace tried a 0 -> TID atomic transition of the futex value
1380  * and failed. The kernel side here does the whole locking operation:
1381  * if there are waiters then it will block, it does PI, etc. (Due to
1382  * races the kernel might see a 0 value of the futex too.)
1383  */
1384 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1385                          int detect, ktime_t *time, int trylock)
1386 {
1387         struct hrtimer_sleeper timeout, *to = NULL;
1388         struct task_struct *curr = current;
1389         struct futex_hash_bucket *hb;
1390         u32 uval, newval, curval;
1391         struct futex_q q;
1392         int ret, lock_taken, ownerdied = 0, attempt = 0;
1393
1394         if (refill_pi_state_cache())
1395                 return -ENOMEM;
1396
1397         if (time) {
1398                 to = &timeout;
1399                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1400                                       HRTIMER_MODE_ABS);
1401                 hrtimer_init_sleeper(to, current);
1402                 to->timer.expires = *time;
1403         }
1404
1405         q.pi_state = NULL;
1406  retry:
1407         futex_lock_mm(fshared);
1408
1409         q.key = FUTEX_KEY_INIT;
1410         ret = get_futex_key(uaddr, fshared, &q.key);
1411         if (unlikely(ret != 0))
1412                 goto out_release_sem;
1413
1414  retry_unlocked:
1415         hb = queue_lock(&q);
1416
1417  retry_locked:
1418         ret = lock_taken = 0;
1419
1420         /*
1421          * To avoid races, we attempt to take the lock here again
1422          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1423          * the locks. It will most likely not succeed.
1424          */
1425         newval = task_pid_vnr(current);
1426
1427         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1428
1429         if (unlikely(curval == -EFAULT))
1430                 goto uaddr_faulted;
1431
1432         /*
1433          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1434          * situation and we return success to user space.
1435          */
1436         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1437                 ret = -EDEADLK;
1438                 goto out_unlock_release_sem;
1439         }
1440
1441         /*
1442          * Surprise - we got the lock. Just return to userspace:
1443          */
1444         if (unlikely(!curval))
1445                 goto out_unlock_release_sem;
1446
1447         uval = curval;
1448
1449         /*
1450          * Set the WAITERS flag, so the owner will know it has someone
1451          * to wake at next unlock
1452          */
1453         newval = curval | FUTEX_WAITERS;
1454
1455         /*
1456          * There are two cases, where a futex might have no owner (the
1457          * owner TID is 0): OWNER_DIED. We take over the futex in this
1458          * case. We also do an unconditional take over, when the owner
1459          * of the futex died.
1460          *
1461          * This is safe as we are protected by the hash bucket lock !
1462          */
1463         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1464                 /* Keep the OWNER_DIED bit */
1465                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1466                 ownerdied = 0;
1467                 lock_taken = 1;
1468         }
1469
1470         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1471
1472         if (unlikely(curval == -EFAULT))
1473                 goto uaddr_faulted;
1474         if (unlikely(curval != uval))
1475                 goto retry_locked;
1476
1477         /*
1478          * We took the lock due to owner died take over.
1479          */
1480         if (unlikely(lock_taken))
1481                 goto out_unlock_release_sem;
1482
1483         /*
1484          * We dont have the lock. Look up the PI state (or create it if
1485          * we are the first waiter):
1486          */
1487         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1488
1489         if (unlikely(ret)) {
1490                 switch (ret) {
1491
1492                 case -EAGAIN:
1493                         /*
1494                          * Task is exiting and we just wait for the
1495                          * exit to complete.
1496                          */
1497                         queue_unlock(&q, hb);
1498                         futex_unlock_mm(fshared);
1499                         cond_resched();
1500                         goto retry;
1501
1502                 case -ESRCH:
1503                         /*
1504                          * No owner found for this futex. Check if the
1505                          * OWNER_DIED bit is set to figure out whether
1506                          * this is a robust futex or not.
1507                          */
1508                         if (get_futex_value_locked(&curval, uaddr))
1509                                 goto uaddr_faulted;
1510
1511                         /*
1512                          * We simply start over in case of a robust
1513                          * futex. The code above will take the futex
1514                          * and return happy.
1515                          */
1516                         if (curval & FUTEX_OWNER_DIED) {
1517                                 ownerdied = 1;
1518                                 goto retry_locked;
1519                         }
1520                 default:
1521                         goto out_unlock_release_sem;
1522                 }
1523         }
1524
1525         /*
1526          * Only actually queue now that the atomic ops are done:
1527          */
1528         queue_me(&q, hb);
1529
1530         /*
1531          * Now the futex is queued and we have checked the data, we
1532          * don't want to hold mmap_sem while we sleep.
1533          */
1534         futex_unlock_mm(fshared);
1535
1536         WARN_ON(!q.pi_state);
1537         /*
1538          * Block on the PI mutex:
1539          */
1540         if (!trylock)
1541                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1542         else {
1543                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1544                 /* Fixup the trylock return value: */
1545                 ret = ret ? 0 : -EWOULDBLOCK;
1546         }
1547
1548         futex_lock_mm(fshared);
1549         spin_lock(q.lock_ptr);
1550
1551         if (!ret) {
1552                 /*
1553                  * Got the lock. We might not be the anticipated owner
1554                  * if we did a lock-steal - fix up the PI-state in
1555                  * that case:
1556                  */
1557                 if (q.pi_state->owner != curr)
1558                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1559         } else {
1560                 /*
1561                  * Catch the rare case, where the lock was released
1562                  * when we were on the way back before we locked the
1563                  * hash bucket.
1564                  */
1565                 if (q.pi_state->owner == curr) {
1566                         /*
1567                          * Try to get the rt_mutex now. This might
1568                          * fail as some other task acquired the
1569                          * rt_mutex after we removed ourself from the
1570                          * rt_mutex waiters list.
1571                          */
1572                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1573                                 ret = 0;
1574                         else {
1575                                 /*
1576                                  * pi_state is incorrect, some other
1577                                  * task did a lock steal and we
1578                                  * returned due to timeout or signal
1579                                  * without taking the rt_mutex. Too
1580                                  * late. We can access the
1581                                  * rt_mutex_owner without locking, as
1582                                  * the other task is now blocked on
1583                                  * the hash bucket lock. Fix the state
1584                                  * up.
1585                                  */
1586                                 struct task_struct *owner;
1587                                 int res;
1588
1589                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1590                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1591                                                            fshared);
1592
1593                                 /* propagate -EFAULT, if the fixup failed */
1594                                 if (res)
1595                                         ret = res;
1596                         }
1597                 } else {
1598                         /*
1599                          * Paranoia check. If we did not take the lock
1600                          * in the trylock above, then we should not be
1601                          * the owner of the rtmutex, neither the real
1602                          * nor the pending one:
1603                          */
1604                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1605                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1606                                        "pi-mutex: %p pi-state %p\n", ret,
1607                                        q.pi_state->pi_mutex.owner,
1608                                        q.pi_state->owner);
1609                 }
1610         }
1611
1612         /* Unqueue and drop the lock */
1613         unqueue_me_pi(&q);
1614         futex_unlock_mm(fshared);
1615
1616         if (to)
1617                 destroy_hrtimer_on_stack(&to->timer);
1618         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1619
1620  out_unlock_release_sem:
1621         queue_unlock(&q, hb);
1622
1623  out_release_sem:
1624         put_futex_key(fshared, &q.key);
1625         futex_unlock_mm(fshared);
1626         if (to)
1627                 destroy_hrtimer_on_stack(&to->timer);
1628         return ret;
1629
1630  uaddr_faulted:
1631         /*
1632          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1633          * non-atomically.  Therefore, if get_user below is not
1634          * enough, we need to handle the fault ourselves, while
1635          * still holding the mmap_sem.
1636          *
1637          * ... and hb->lock. :-) --ANK
1638          */
1639         queue_unlock(&q, hb);
1640
1641         if (attempt++) {
1642                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1643                                          attempt);
1644                 if (ret)
1645                         goto out_release_sem;
1646                 goto retry_unlocked;
1647         }
1648
1649         futex_unlock_mm(fshared);
1650
1651         ret = get_user(uval, uaddr);
1652         if (!ret && (uval != -EFAULT))
1653                 goto retry;
1654
1655         if (to)
1656                 destroy_hrtimer_on_stack(&to->timer);
1657         return ret;
1658 }
1659
1660 /*
1661  * Userspace attempted a TID -> 0 atomic transition, and failed.
1662  * This is the in-kernel slowpath: we look up the PI state (if any),
1663  * and do the rt-mutex unlock.
1664  */
1665 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1666 {
1667         struct futex_hash_bucket *hb;
1668         struct futex_q *this, *next;
1669         u32 uval;
1670         struct plist_head *head;
1671         union futex_key key = FUTEX_KEY_INIT;
1672         int ret, attempt = 0;
1673
1674 retry:
1675         if (get_user(uval, uaddr))
1676                 return -EFAULT;
1677         /*
1678          * We release only a lock we actually own:
1679          */
1680         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1681                 return -EPERM;
1682         /*
1683          * First take all the futex related locks:
1684          */
1685         futex_lock_mm(fshared);
1686
1687         ret = get_futex_key(uaddr, fshared, &key);
1688         if (unlikely(ret != 0))
1689                 goto out;
1690
1691         hb = hash_futex(&key);
1692 retry_unlocked:
1693         spin_lock(&hb->lock);
1694
1695         /*
1696          * To avoid races, try to do the TID -> 0 atomic transition
1697          * again. If it succeeds then we can return without waking
1698          * anyone else up:
1699          */
1700         if (!(uval & FUTEX_OWNER_DIED))
1701                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1702
1703
1704         if (unlikely(uval == -EFAULT))
1705                 goto pi_faulted;
1706         /*
1707          * Rare case: we managed to release the lock atomically,
1708          * no need to wake anyone else up:
1709          */
1710         if (unlikely(uval == task_pid_vnr(current)))
1711                 goto out_unlock;
1712
1713         /*
1714          * Ok, other tasks may need to be woken up - check waiters
1715          * and do the wakeup if necessary:
1716          */
1717         head = &hb->chain;
1718
1719         plist_for_each_entry_safe(this, next, head, list) {
1720                 if (!match_futex (&this->key, &key))
1721                         continue;
1722                 ret = wake_futex_pi(uaddr, uval, this);
1723                 /*
1724                  * The atomic access to the futex value
1725                  * generated a pagefault, so retry the
1726                  * user-access and the wakeup:
1727                  */
1728                 if (ret == -EFAULT)
1729                         goto pi_faulted;
1730                 goto out_unlock;
1731         }
1732         /*
1733          * No waiters - kernel unlocks the futex:
1734          */
1735         if (!(uval & FUTEX_OWNER_DIED)) {
1736                 ret = unlock_futex_pi(uaddr, uval);
1737                 if (ret == -EFAULT)
1738                         goto pi_faulted;
1739         }
1740
1741 out_unlock:
1742         spin_unlock(&hb->lock);
1743 out:
1744         put_futex_key(fshared, &key);
1745         futex_unlock_mm(fshared);
1746
1747         return ret;
1748
1749 pi_faulted:
1750         /*
1751          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1752          * non-atomically.  Therefore, if get_user below is not
1753          * enough, we need to handle the fault ourselves, while
1754          * still holding the mmap_sem.
1755          *
1756          * ... and hb->lock. --ANK
1757          */
1758         spin_unlock(&hb->lock);
1759
1760         if (attempt++) {
1761                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1762                                          attempt);
1763                 if (ret)
1764                         goto out;
1765                 uval = 0;
1766                 goto retry_unlocked;
1767         }
1768
1769         futex_unlock_mm(fshared);
1770
1771         ret = get_user(uval, uaddr);
1772         if (!ret && (uval != -EFAULT))
1773                 goto retry;
1774
1775         return ret;
1776 }
1777
1778 /*
1779  * Support for robust futexes: the kernel cleans up held futexes at
1780  * thread exit time.
1781  *
1782  * Implementation: user-space maintains a per-thread list of locks it
1783  * is holding. Upon do_exit(), the kernel carefully walks this list,
1784  * and marks all locks that are owned by this thread with the
1785  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1786  * always manipulated with the lock held, so the list is private and
1787  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1788  * field, to allow the kernel to clean up if the thread dies after
1789  * acquiring the lock, but just before it could have added itself to
1790  * the list. There can only be one such pending lock.
1791  */
1792
1793 /**
1794  * sys_set_robust_list - set the robust-futex list head of a task
1795  * @head: pointer to the list-head
1796  * @len: length of the list-head, as userspace expects
1797  */
1798 asmlinkage long
1799 sys_set_robust_list(struct robust_list_head __user *head,
1800                     size_t len)
1801 {
1802         if (!futex_cmpxchg_enabled)
1803                 return -ENOSYS;
1804         /*
1805          * The kernel knows only one size for now:
1806          */
1807         if (unlikely(len != sizeof(*head)))
1808                 return -EINVAL;
1809
1810         current->robust_list = head;
1811
1812         return 0;
1813 }
1814
1815 /**
1816  * sys_get_robust_list - get the robust-futex list head of a task
1817  * @pid: pid of the process [zero for current task]
1818  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1819  * @len_ptr: pointer to a length field, the kernel fills in the header size
1820  */
1821 asmlinkage long
1822 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1823                     size_t __user *len_ptr)
1824 {
1825         struct robust_list_head __user *head;
1826         unsigned long ret;
1827
1828         if (!futex_cmpxchg_enabled)
1829                 return -ENOSYS;
1830
1831         if (!pid)
1832                 head = current->robust_list;
1833         else {
1834                 struct task_struct *p;
1835
1836                 ret = -ESRCH;
1837                 rcu_read_lock();
1838                 p = find_task_by_vpid(pid);
1839                 if (!p)
1840                         goto err_unlock;
1841                 ret = -EPERM;
1842                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1843                                 !capable(CAP_SYS_PTRACE))
1844                         goto err_unlock;
1845                 head = p->robust_list;
1846                 rcu_read_unlock();
1847         }
1848
1849         if (put_user(sizeof(*head), len_ptr))
1850                 return -EFAULT;
1851         return put_user(head, head_ptr);
1852
1853 err_unlock:
1854         rcu_read_unlock();
1855
1856         return ret;
1857 }
1858
1859 /*
1860  * Process a futex-list entry, check whether it's owned by the
1861  * dying task, and do notification if so:
1862  */
1863 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1864 {
1865         u32 uval, nval, mval;
1866
1867 retry:
1868         if (get_user(uval, uaddr))
1869                 return -1;
1870
1871         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1872                 /*
1873                  * Ok, this dying thread is truly holding a futex
1874                  * of interest. Set the OWNER_DIED bit atomically
1875                  * via cmpxchg, and if the value had FUTEX_WAITERS
1876                  * set, wake up a waiter (if any). (We have to do a
1877                  * futex_wake() even if OWNER_DIED is already set -
1878                  * to handle the rare but possible case of recursive
1879                  * thread-death.) The rest of the cleanup is done in
1880                  * userspace.
1881                  */
1882                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1883                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1884
1885                 if (nval == -EFAULT)
1886                         return -1;
1887
1888                 if (nval != uval)
1889                         goto retry;
1890
1891                 /*
1892                  * Wake robust non-PI futexes here. The wakeup of
1893                  * PI futexes happens in exit_pi_state():
1894                  */
1895                 if (!pi && (uval & FUTEX_WAITERS))
1896                         futex_wake(uaddr, &curr->mm->mmap_sem, 1,
1897                                    FUTEX_BITSET_MATCH_ANY);
1898         }
1899         return 0;
1900 }
1901
1902 /*
1903  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1904  */
1905 static inline int fetch_robust_entry(struct robust_list __user **entry,
1906                                      struct robust_list __user * __user *head,
1907                                      int *pi)
1908 {
1909         unsigned long uentry;
1910
1911         if (get_user(uentry, (unsigned long __user *)head))
1912                 return -EFAULT;
1913
1914         *entry = (void __user *)(uentry & ~1UL);
1915         *pi = uentry & 1;
1916
1917         return 0;
1918 }
1919
1920 /*
1921  * Walk curr->robust_list (very carefully, it's a userspace list!)
1922  * and mark any locks found there dead, and notify any waiters.
1923  *
1924  * We silently return on any sign of list-walking problem.
1925  */
1926 void exit_robust_list(struct task_struct *curr)
1927 {
1928         struct robust_list_head __user *head = curr->robust_list;
1929         struct robust_list __user *entry, *next_entry, *pending;
1930         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1931         unsigned long futex_offset;
1932         int rc;
1933
1934         if (!futex_cmpxchg_enabled)
1935                 return;
1936
1937         /*
1938          * Fetch the list head (which was registered earlier, via
1939          * sys_set_robust_list()):
1940          */
1941         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1942                 return;
1943         /*
1944          * Fetch the relative futex offset:
1945          */
1946         if (get_user(futex_offset, &head->futex_offset))
1947                 return;
1948         /*
1949          * Fetch any possibly pending lock-add first, and handle it
1950          * if it exists:
1951          */
1952         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1953                 return;
1954
1955         next_entry = NULL;      /* avoid warning with gcc */
1956         while (entry != &head->list) {
1957                 /*
1958                  * Fetch the next entry in the list before calling
1959                  * handle_futex_death:
1960                  */
1961                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1962                 /*
1963                  * A pending lock might already be on the list, so
1964                  * don't process it twice:
1965                  */
1966                 if (entry != pending)
1967                         if (handle_futex_death((void __user *)entry + futex_offset,
1968                                                 curr, pi))
1969                                 return;
1970                 if (rc)
1971                         return;
1972                 entry = next_entry;
1973                 pi = next_pi;
1974                 /*
1975                  * Avoid excessively long or circular lists:
1976                  */
1977                 if (!--limit)
1978                         break;
1979
1980                 cond_resched();
1981         }
1982
1983         if (pending)
1984                 handle_futex_death((void __user *)pending + futex_offset,
1985                                    curr, pip);
1986 }
1987
1988 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1989                 u32 __user *uaddr2, u32 val2, u32 val3)
1990 {
1991         int ret = -ENOSYS;
1992         int cmd = op & FUTEX_CMD_MASK;
1993         struct rw_semaphore *fshared = NULL;
1994
1995         if (!(op & FUTEX_PRIVATE_FLAG))
1996                 fshared = &current->mm->mmap_sem;
1997
1998         switch (cmd) {
1999         case FUTEX_WAIT:
2000                 val3 = FUTEX_BITSET_MATCH_ANY;
2001         case FUTEX_WAIT_BITSET:
2002                 ret = futex_wait(uaddr, fshared, val, timeout, val3);
2003                 break;
2004         case FUTEX_WAKE:
2005                 val3 = FUTEX_BITSET_MATCH_ANY;
2006         case FUTEX_WAKE_BITSET:
2007                 ret = futex_wake(uaddr, fshared, val, val3);
2008                 break;
2009         case FUTEX_REQUEUE:
2010                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2011                 break;
2012         case FUTEX_CMP_REQUEUE:
2013                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2014                 break;
2015         case FUTEX_WAKE_OP:
2016                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2017                 break;
2018         case FUTEX_LOCK_PI:
2019                 if (futex_cmpxchg_enabled)
2020                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2021                 break;
2022         case FUTEX_UNLOCK_PI:
2023                 if (futex_cmpxchg_enabled)
2024                         ret = futex_unlock_pi(uaddr, fshared);
2025                 break;
2026         case FUTEX_TRYLOCK_PI:
2027                 if (futex_cmpxchg_enabled)
2028                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2029                 break;
2030         default:
2031                 ret = -ENOSYS;
2032         }
2033         return ret;
2034 }
2035
2036
2037 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2038                           struct timespec __user *utime, u32 __user *uaddr2,
2039                           u32 val3)
2040 {
2041         struct timespec ts;
2042         ktime_t t, *tp = NULL;
2043         u32 val2 = 0;
2044         int cmd = op & FUTEX_CMD_MASK;
2045
2046         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2047                       cmd == FUTEX_WAIT_BITSET)) {
2048                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2049                         return -EFAULT;
2050                 if (!timespec_valid(&ts))
2051                         return -EINVAL;
2052
2053                 t = timespec_to_ktime(ts);
2054                 if (cmd == FUTEX_WAIT)
2055                         t = ktime_add_safe(ktime_get(), t);
2056                 tp = &t;
2057         }
2058         /*
2059          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2060          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2061          */
2062         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2063             cmd == FUTEX_WAKE_OP)
2064                 val2 = (u32) (unsigned long) utime;
2065
2066         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2067 }
2068
2069 static int __init futex_init(void)
2070 {
2071         u32 curval;
2072         int i;
2073
2074         /*
2075          * This will fail and we want it. Some arch implementations do
2076          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2077          * functionality. We want to know that before we call in any
2078          * of the complex code paths. Also we want to prevent
2079          * registration of robust lists in that case. NULL is
2080          * guaranteed to fault and we get -EFAULT on functional
2081          * implementation, the non functional ones will return
2082          * -ENOSYS.
2083          */
2084         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2085         if (curval == -EFAULT)
2086                 futex_cmpxchg_enabled = 1;
2087
2088         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2089                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2090                 spin_lock_init(&futex_queues[i].lock);
2091         }
2092
2093         return 0;
2094 }
2095 __initcall(futex_init);