Merge branches 'core/futexes', 'core/locking', 'core/rcu' and 'linus' into core/urgent
[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  * Split the global futex_lock into every hash list lock.
118  */
119 struct futex_hash_bucket {
120         spinlock_t lock;
121         struct plist_head chain;
122 };
123
124 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
125
126 /*
127  * We hash on the keys returned from get_futex_key (see below).
128  */
129 static struct futex_hash_bucket *hash_futex(union futex_key *key)
130 {
131         u32 hash = jhash2((u32*)&key->both.word,
132                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
133                           key->both.offset);
134         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
135 }
136
137 /*
138  * Return 1 if two futex_keys are equal, 0 otherwise.
139  */
140 static inline int match_futex(union futex_key *key1, union futex_key *key2)
141 {
142         return (key1->both.word == key2->both.word
143                 && key1->both.ptr == key2->both.ptr
144                 && key1->both.offset == key2->both.offset);
145 }
146
147 /*
148  * Take a reference to the resource addressed by a key.
149  * Can be called while holding spinlocks.
150  *
151  */
152 static void get_futex_key_refs(union futex_key *key)
153 {
154         if (!key->both.ptr)
155                 return;
156
157         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
158         case FUT_OFF_INODE:
159                 atomic_inc(&key->shared.inode->i_count);
160                 break;
161         case FUT_OFF_MMSHARED:
162                 atomic_inc(&key->private.mm->mm_count);
163                 break;
164         }
165 }
166
167 /*
168  * Drop a reference to the resource addressed by a key.
169  * The hash bucket spinlock must not be held.
170  */
171 static void drop_futex_key_refs(union futex_key *key)
172 {
173         if (!key->both.ptr) {
174                 /* If we're here then we tried to put a key we failed to get */
175                 WARN_ON_ONCE(1);
176                 return;
177         }
178
179         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
180         case FUT_OFF_INODE:
181                 iput(key->shared.inode);
182                 break;
183         case FUT_OFF_MMSHARED:
184                 mmdrop(key->private.mm);
185                 break;
186         }
187 }
188
189 /**
190  * get_futex_key - Get parameters which are the keys for a futex.
191  * @uaddr: virtual address of the futex
192  * @shared: NULL for a PROCESS_PRIVATE futex,
193  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
194  * @key: address where result is stored.
195  *
196  * Returns a negative error code or 0
197  * The key words are stored in *key on success.
198  *
199  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
200  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
201  * We can usually work out the index without swapping in the page.
202  *
203  * fshared is NULL for PROCESS_PRIVATE futexes
204  * For other futexes, it points to &current->mm->mmap_sem and
205  * caller must have taken the reader lock. but NOT any spinlocks.
206  */
207 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
208 {
209         unsigned long address = (unsigned long)uaddr;
210         struct mm_struct *mm = current->mm;
211         struct page *page;
212         int err;
213
214         /*
215          * The futex address must be "naturally" aligned.
216          */
217         key->both.offset = address % PAGE_SIZE;
218         if (unlikely((address % sizeof(u32)) != 0))
219                 return -EINVAL;
220         address -= key->both.offset;
221
222         /*
223          * PROCESS_PRIVATE futexes are fast.
224          * As the mm cannot disappear under us and the 'key' only needs
225          * virtual address, we dont even have to find the underlying vma.
226          * Note : We do have to check 'uaddr' is a valid user address,
227          *        but access_ok() should be faster than find_vma()
228          */
229         if (!fshared) {
230                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
231                         return -EFAULT;
232                 key->private.mm = mm;
233                 key->private.address = address;
234                 get_futex_key_refs(key);
235                 return 0;
236         }
237
238 again:
239         err = get_user_pages_fast(address, 1, 0, &page);
240         if (err < 0)
241                 return err;
242
243         lock_page(page);
244         if (!page->mapping) {
245                 unlock_page(page);
246                 put_page(page);
247                 goto again;
248         }
249
250         /*
251          * Private mappings are handled in a simple way.
252          *
253          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
254          * it's a read-only handle, it's expected that futexes attach to
255          * the object not the particular process.
256          */
257         if (PageAnon(page)) {
258                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
259                 key->private.mm = mm;
260                 key->private.address = address;
261         } else {
262                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
263                 key->shared.inode = page->mapping->host;
264                 key->shared.pgoff = page->index;
265         }
266
267         get_futex_key_refs(key);
268
269         unlock_page(page);
270         put_page(page);
271         return 0;
272 }
273
274 static inline
275 void put_futex_key(int fshared, union futex_key *key)
276 {
277         drop_futex_key_refs(key);
278 }
279
280 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
281 {
282         u32 curval;
283
284         pagefault_disable();
285         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
286         pagefault_enable();
287
288         return curval;
289 }
290
291 static int get_futex_value_locked(u32 *dest, u32 __user *from)
292 {
293         int ret;
294
295         pagefault_disable();
296         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
297         pagefault_enable();
298
299         return ret ? -EFAULT : 0;
300 }
301
302 /*
303  * Fault handling.
304  */
305 static int futex_handle_fault(unsigned long address, int attempt)
306 {
307         struct vm_area_struct * vma;
308         struct mm_struct *mm = current->mm;
309         int ret = -EFAULT;
310
311         if (attempt > 2)
312                 return ret;
313
314         down_read(&mm->mmap_sem);
315         vma = find_vma(mm, address);
316         if (vma && address >= vma->vm_start &&
317             (vma->vm_flags & VM_WRITE)) {
318                 int fault;
319                 fault = handle_mm_fault(mm, vma, address, 1);
320                 if (unlikely((fault & VM_FAULT_ERROR))) {
321 #if 0
322                         /* XXX: let's do this when we verify it is OK */
323                         if (ret & VM_FAULT_OOM)
324                                 ret = -ENOMEM;
325 #endif
326                 } else {
327                         ret = 0;
328                         if (fault & VM_FAULT_MAJOR)
329                                 current->maj_flt++;
330                         else
331                                 current->min_flt++;
332                 }
333         }
334         up_read(&mm->mmap_sem);
335         return ret;
336 }
337
338 /*
339  * PI code:
340  */
341 static int refill_pi_state_cache(void)
342 {
343         struct futex_pi_state *pi_state;
344
345         if (likely(current->pi_state_cache))
346                 return 0;
347
348         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
349
350         if (!pi_state)
351                 return -ENOMEM;
352
353         INIT_LIST_HEAD(&pi_state->list);
354         /* pi_mutex gets initialized later */
355         pi_state->owner = NULL;
356         atomic_set(&pi_state->refcount, 1);
357         pi_state->key = FUTEX_KEY_INIT;
358
359         current->pi_state_cache = pi_state;
360
361         return 0;
362 }
363
364 static struct futex_pi_state * alloc_pi_state(void)
365 {
366         struct futex_pi_state *pi_state = current->pi_state_cache;
367
368         WARN_ON(!pi_state);
369         current->pi_state_cache = NULL;
370
371         return pi_state;
372 }
373
374 static void free_pi_state(struct futex_pi_state *pi_state)
375 {
376         if (!atomic_dec_and_test(&pi_state->refcount))
377                 return;
378
379         /*
380          * If pi_state->owner is NULL, the owner is most probably dying
381          * and has cleaned up the pi_state already
382          */
383         if (pi_state->owner) {
384                 spin_lock_irq(&pi_state->owner->pi_lock);
385                 list_del_init(&pi_state->list);
386                 spin_unlock_irq(&pi_state->owner->pi_lock);
387
388                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
389         }
390
391         if (current->pi_state_cache)
392                 kfree(pi_state);
393         else {
394                 /*
395                  * pi_state->list is already empty.
396                  * clear pi_state->owner.
397                  * refcount is at 0 - put it back to 1.
398                  */
399                 pi_state->owner = NULL;
400                 atomic_set(&pi_state->refcount, 1);
401                 current->pi_state_cache = pi_state;
402         }
403 }
404
405 /*
406  * Look up the task based on what TID userspace gave us.
407  * We dont trust it.
408  */
409 static struct task_struct * futex_find_get_task(pid_t pid)
410 {
411         struct task_struct *p;
412         const struct cred *cred = current_cred(), *pcred;
413
414         rcu_read_lock();
415         p = find_task_by_vpid(pid);
416         if (!p) {
417                 p = ERR_PTR(-ESRCH);
418         } else {
419                 pcred = __task_cred(p);
420                 if (cred->euid != pcred->euid &&
421                     cred->euid != pcred->uid)
422                         p = ERR_PTR(-ESRCH);
423                 else
424                         get_task_struct(p);
425         }
426
427         rcu_read_unlock();
428
429         return p;
430 }
431
432 /*
433  * This task is holding PI mutexes at exit time => bad.
434  * Kernel cleans up PI-state, but userspace is likely hosed.
435  * (Robust-futex cleanup is separate and might save the day for userspace.)
436  */
437 void exit_pi_state_list(struct task_struct *curr)
438 {
439         struct list_head *next, *head = &curr->pi_state_list;
440         struct futex_pi_state *pi_state;
441         struct futex_hash_bucket *hb;
442         union futex_key key = FUTEX_KEY_INIT;
443
444         if (!futex_cmpxchg_enabled)
445                 return;
446         /*
447          * We are a ZOMBIE and nobody can enqueue itself on
448          * pi_state_list anymore, but we have to be careful
449          * versus waiters unqueueing themselves:
450          */
451         spin_lock_irq(&curr->pi_lock);
452         while (!list_empty(head)) {
453
454                 next = head->next;
455                 pi_state = list_entry(next, struct futex_pi_state, list);
456                 key = pi_state->key;
457                 hb = hash_futex(&key);
458                 spin_unlock_irq(&curr->pi_lock);
459
460                 spin_lock(&hb->lock);
461
462                 spin_lock_irq(&curr->pi_lock);
463                 /*
464                  * We dropped the pi-lock, so re-check whether this
465                  * task still owns the PI-state:
466                  */
467                 if (head->next != next) {
468                         spin_unlock(&hb->lock);
469                         continue;
470                 }
471
472                 WARN_ON(pi_state->owner != curr);
473                 WARN_ON(list_empty(&pi_state->list));
474                 list_del_init(&pi_state->list);
475                 pi_state->owner = NULL;
476                 spin_unlock_irq(&curr->pi_lock);
477
478                 rt_mutex_unlock(&pi_state->pi_mutex);
479
480                 spin_unlock(&hb->lock);
481
482                 spin_lock_irq(&curr->pi_lock);
483         }
484         spin_unlock_irq(&curr->pi_lock);
485 }
486
487 static int
488 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
489                 union futex_key *key, struct futex_pi_state **ps)
490 {
491         struct futex_pi_state *pi_state = NULL;
492         struct futex_q *this, *next;
493         struct plist_head *head;
494         struct task_struct *p;
495         pid_t pid = uval & FUTEX_TID_MASK;
496
497         head = &hb->chain;
498
499         plist_for_each_entry_safe(this, next, head, list) {
500                 if (match_futex(&this->key, key)) {
501                         /*
502                          * Another waiter already exists - bump up
503                          * the refcount and return its pi_state:
504                          */
505                         pi_state = this->pi_state;
506                         /*
507                          * Userspace might have messed up non PI and PI futexes
508                          */
509                         if (unlikely(!pi_state))
510                                 return -EINVAL;
511
512                         WARN_ON(!atomic_read(&pi_state->refcount));
513                         WARN_ON(pid && pi_state->owner &&
514                                 pi_state->owner->pid != pid);
515
516                         atomic_inc(&pi_state->refcount);
517                         *ps = pi_state;
518
519                         return 0;
520                 }
521         }
522
523         /*
524          * We are the first waiter - try to look up the real owner and attach
525          * the new pi_state to it, but bail out when TID = 0
526          */
527         if (!pid)
528                 return -ESRCH;
529         p = futex_find_get_task(pid);
530         if (IS_ERR(p))
531                 return PTR_ERR(p);
532
533         /*
534          * We need to look at the task state flags to figure out,
535          * whether the task is exiting. To protect against the do_exit
536          * change of the task flags, we do this protected by
537          * p->pi_lock:
538          */
539         spin_lock_irq(&p->pi_lock);
540         if (unlikely(p->flags & PF_EXITING)) {
541                 /*
542                  * The task is on the way out. When PF_EXITPIDONE is
543                  * set, we know that the task has finished the
544                  * cleanup:
545                  */
546                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
547
548                 spin_unlock_irq(&p->pi_lock);
549                 put_task_struct(p);
550                 return ret;
551         }
552
553         pi_state = alloc_pi_state();
554
555         /*
556          * Initialize the pi_mutex in locked state and make 'p'
557          * the owner of it:
558          */
559         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
560
561         /* Store the key for possible exit cleanups: */
562         pi_state->key = *key;
563
564         WARN_ON(!list_empty(&pi_state->list));
565         list_add(&pi_state->list, &p->pi_state_list);
566         pi_state->owner = p;
567         spin_unlock_irq(&p->pi_lock);
568
569         put_task_struct(p);
570
571         *ps = pi_state;
572
573         return 0;
574 }
575
576 /*
577  * The hash bucket lock must be held when this is called.
578  * Afterwards, the futex_q must not be accessed.
579  */
580 static void wake_futex(struct futex_q *q)
581 {
582         plist_del(&q->list, &q->list.plist);
583         /*
584          * The lock in wake_up_all() is a crucial memory barrier after the
585          * plist_del() and also before assigning to q->lock_ptr.
586          */
587         wake_up(&q->waiter);
588         /*
589          * The waiting task can free the futex_q as soon as this is written,
590          * without taking any locks.  This must come last.
591          *
592          * A memory barrier is required here to prevent the following store
593          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
594          * at the end of wake_up_all() does not prevent this store from
595          * moving.
596          */
597         smp_wmb();
598         q->lock_ptr = NULL;
599 }
600
601 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
602 {
603         struct task_struct *new_owner;
604         struct futex_pi_state *pi_state = this->pi_state;
605         u32 curval, newval;
606
607         if (!pi_state)
608                 return -EINVAL;
609
610         spin_lock(&pi_state->pi_mutex.wait_lock);
611         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
612
613         /*
614          * This happens when we have stolen the lock and the original
615          * pending owner did not enqueue itself back on the rt_mutex.
616          * Thats not a tragedy. We know that way, that a lock waiter
617          * is on the fly. We make the futex_q waiter the pending owner.
618          */
619         if (!new_owner)
620                 new_owner = this->task;
621
622         /*
623          * We pass it to the next owner. (The WAITERS bit is always
624          * kept enabled while there is PI state around. We must also
625          * preserve the owner died bit.)
626          */
627         if (!(uval & FUTEX_OWNER_DIED)) {
628                 int ret = 0;
629
630                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
631
632                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
633
634                 if (curval == -EFAULT)
635                         ret = -EFAULT;
636                 else if (curval != uval)
637                         ret = -EINVAL;
638                 if (ret) {
639                         spin_unlock(&pi_state->pi_mutex.wait_lock);
640                         return ret;
641                 }
642         }
643
644         spin_lock_irq(&pi_state->owner->pi_lock);
645         WARN_ON(list_empty(&pi_state->list));
646         list_del_init(&pi_state->list);
647         spin_unlock_irq(&pi_state->owner->pi_lock);
648
649         spin_lock_irq(&new_owner->pi_lock);
650         WARN_ON(!list_empty(&pi_state->list));
651         list_add(&pi_state->list, &new_owner->pi_state_list);
652         pi_state->owner = new_owner;
653         spin_unlock_irq(&new_owner->pi_lock);
654
655         spin_unlock(&pi_state->pi_mutex.wait_lock);
656         rt_mutex_unlock(&pi_state->pi_mutex);
657
658         return 0;
659 }
660
661 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
662 {
663         u32 oldval;
664
665         /*
666          * There is no waiter, so we unlock the futex. The owner died
667          * bit has not to be preserved here. We are the owner:
668          */
669         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
670
671         if (oldval == -EFAULT)
672                 return oldval;
673         if (oldval != uval)
674                 return -EAGAIN;
675
676         return 0;
677 }
678
679 /*
680  * Express the locking dependencies for lockdep:
681  */
682 static inline void
683 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
684 {
685         if (hb1 <= hb2) {
686                 spin_lock(&hb1->lock);
687                 if (hb1 < hb2)
688                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
689         } else { /* hb1 > hb2 */
690                 spin_lock(&hb2->lock);
691                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
692         }
693 }
694
695 /*
696  * Wake up all waiters hashed on the physical page that is mapped
697  * to this virtual address:
698  */
699 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
700 {
701         struct futex_hash_bucket *hb;
702         struct futex_q *this, *next;
703         struct plist_head *head;
704         union futex_key key = FUTEX_KEY_INIT;
705         int ret;
706
707         if (!bitset)
708                 return -EINVAL;
709
710         ret = get_futex_key(uaddr, fshared, &key);
711         if (unlikely(ret != 0))
712                 goto out;
713
714         hb = hash_futex(&key);
715         spin_lock(&hb->lock);
716         head = &hb->chain;
717
718         plist_for_each_entry_safe(this, next, head, list) {
719                 if (match_futex (&this->key, &key)) {
720                         if (this->pi_state) {
721                                 ret = -EINVAL;
722                                 break;
723                         }
724
725                         /* Check if one of the bits is set in both bitsets */
726                         if (!(this->bitset & bitset))
727                                 continue;
728
729                         wake_futex(this);
730                         if (++ret >= nr_wake)
731                                 break;
732                 }
733         }
734
735         spin_unlock(&hb->lock);
736         put_futex_key(fshared, &key);
737 out:
738         return ret;
739 }
740
741 /*
742  * Wake up all waiters hashed on the physical page that is mapped
743  * to this virtual address:
744  */
745 static int
746 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
747               int nr_wake, int nr_wake2, int op)
748 {
749         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
750         struct futex_hash_bucket *hb1, *hb2;
751         struct plist_head *head;
752         struct futex_q *this, *next;
753         int ret, op_ret, attempt = 0;
754
755 retryfull:
756         ret = get_futex_key(uaddr1, fshared, &key1);
757         if (unlikely(ret != 0))
758                 goto out;
759         ret = get_futex_key(uaddr2, fshared, &key2);
760         if (unlikely(ret != 0))
761                 goto out_put_key1;
762
763         hb1 = hash_futex(&key1);
764         hb2 = hash_futex(&key2);
765
766 retry:
767         double_lock_hb(hb1, hb2);
768
769         op_ret = futex_atomic_op_inuser(op, uaddr2);
770         if (unlikely(op_ret < 0)) {
771                 u32 dummy;
772
773                 spin_unlock(&hb1->lock);
774                 if (hb1 != hb2)
775                         spin_unlock(&hb2->lock);
776
777 #ifndef CONFIG_MMU
778                 /*
779                  * we don't get EFAULT from MMU faults if we don't have an MMU,
780                  * but we might get them from range checking
781                  */
782                 ret = op_ret;
783                 goto out_put_keys;
784 #endif
785
786                 if (unlikely(op_ret != -EFAULT)) {
787                         ret = op_ret;
788                         goto out_put_keys;
789                 }
790
791                 /*
792                  * futex_atomic_op_inuser needs to both read and write
793                  * *(int __user *)uaddr2, but we can't modify it
794                  * non-atomically.  Therefore, if get_user below is not
795                  * enough, we need to handle the fault ourselves, while
796                  * still holding the mmap_sem.
797                  */
798                 if (attempt++) {
799                         ret = futex_handle_fault((unsigned long)uaddr2,
800                                                  attempt);
801                         if (ret)
802                                 goto out_put_keys;
803                         goto retry;
804                 }
805
806                 ret = get_user(dummy, uaddr2);
807                 if (ret)
808                         return ret;
809
810                 goto retryfull;
811         }
812
813         head = &hb1->chain;
814
815         plist_for_each_entry_safe(this, next, head, list) {
816                 if (match_futex (&this->key, &key1)) {
817                         wake_futex(this);
818                         if (++ret >= nr_wake)
819                                 break;
820                 }
821         }
822
823         if (op_ret > 0) {
824                 head = &hb2->chain;
825
826                 op_ret = 0;
827                 plist_for_each_entry_safe(this, next, head, list) {
828                         if (match_futex (&this->key, &key2)) {
829                                 wake_futex(this);
830                                 if (++op_ret >= nr_wake2)
831                                         break;
832                         }
833                 }
834                 ret += op_ret;
835         }
836
837         spin_unlock(&hb1->lock);
838         if (hb1 != hb2)
839                 spin_unlock(&hb2->lock);
840 out_put_keys:
841         put_futex_key(fshared, &key2);
842 out_put_key1:
843         put_futex_key(fshared, &key1);
844 out:
845         return ret;
846 }
847
848 /*
849  * Requeue all waiters hashed on one physical page to another
850  * physical page.
851  */
852 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
853                          int nr_wake, int nr_requeue, u32 *cmpval)
854 {
855         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
856         struct futex_hash_bucket *hb1, *hb2;
857         struct plist_head *head1;
858         struct futex_q *this, *next;
859         int ret, drop_count = 0;
860
861 retry:
862         ret = get_futex_key(uaddr1, fshared, &key1);
863         if (unlikely(ret != 0))
864                 goto out;
865         ret = get_futex_key(uaddr2, fshared, &key2);
866         if (unlikely(ret != 0))
867                 goto out_put_key1;
868
869         hb1 = hash_futex(&key1);
870         hb2 = hash_futex(&key2);
871
872         double_lock_hb(hb1, hb2);
873
874         if (likely(cmpval != NULL)) {
875                 u32 curval;
876
877                 ret = get_futex_value_locked(&curval, uaddr1);
878
879                 if (unlikely(ret)) {
880                         spin_unlock(&hb1->lock);
881                         if (hb1 != hb2)
882                                 spin_unlock(&hb2->lock);
883
884                         ret = get_user(curval, uaddr1);
885
886                         if (!ret)
887                                 goto retry;
888
889                         goto out_put_keys;
890                 }
891                 if (curval != *cmpval) {
892                         ret = -EAGAIN;
893                         goto out_unlock;
894                 }
895         }
896
897         head1 = &hb1->chain;
898         plist_for_each_entry_safe(this, next, head1, list) {
899                 if (!match_futex (&this->key, &key1))
900                         continue;
901                 if (++ret <= nr_wake) {
902                         wake_futex(this);
903                 } else {
904                         /*
905                          * If key1 and key2 hash to the same bucket, no need to
906                          * requeue.
907                          */
908                         if (likely(head1 != &hb2->chain)) {
909                                 plist_del(&this->list, &hb1->chain);
910                                 plist_add(&this->list, &hb2->chain);
911                                 this->lock_ptr = &hb2->lock;
912 #ifdef CONFIG_DEBUG_PI_LIST
913                                 this->list.plist.lock = &hb2->lock;
914 #endif
915                         }
916                         this->key = key2;
917                         get_futex_key_refs(&key2);
918                         drop_count++;
919
920                         if (ret - nr_wake >= nr_requeue)
921                                 break;
922                 }
923         }
924
925 out_unlock:
926         spin_unlock(&hb1->lock);
927         if (hb1 != hb2)
928                 spin_unlock(&hb2->lock);
929
930         /* drop_futex_key_refs() must be called outside the spinlocks. */
931         while (--drop_count >= 0)
932                 drop_futex_key_refs(&key1);
933
934 out_put_keys:
935         put_futex_key(fshared, &key2);
936 out_put_key1:
937         put_futex_key(fshared, &key1);
938 out:
939         return ret;
940 }
941
942 /* The key must be already stored in q->key. */
943 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
944 {
945         struct futex_hash_bucket *hb;
946
947         init_waitqueue_head(&q->waiter);
948
949         get_futex_key_refs(&q->key);
950         hb = hash_futex(&q->key);
951         q->lock_ptr = &hb->lock;
952
953         spin_lock(&hb->lock);
954         return hb;
955 }
956
957 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
958 {
959         int prio;
960
961         /*
962          * The priority used to register this element is
963          * - either the real thread-priority for the real-time threads
964          * (i.e. threads with a priority lower than MAX_RT_PRIO)
965          * - or MAX_RT_PRIO for non-RT threads.
966          * Thus, all RT-threads are woken first in priority order, and
967          * the others are woken last, in FIFO order.
968          */
969         prio = min(current->normal_prio, MAX_RT_PRIO);
970
971         plist_node_init(&q->list, prio);
972 #ifdef CONFIG_DEBUG_PI_LIST
973         q->list.plist.lock = &hb->lock;
974 #endif
975         plist_add(&q->list, &hb->chain);
976         q->task = current;
977         spin_unlock(&hb->lock);
978 }
979
980 static inline void
981 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
982 {
983         spin_unlock(&hb->lock);
984         drop_futex_key_refs(&q->key);
985 }
986
987 /*
988  * queue_me and unqueue_me must be called as a pair, each
989  * exactly once.  They are called with the hashed spinlock held.
990  */
991
992 /* Return 1 if we were still queued (ie. 0 means we were woken) */
993 static int unqueue_me(struct futex_q *q)
994 {
995         spinlock_t *lock_ptr;
996         int ret = 0;
997
998         /* In the common case we don't take the spinlock, which is nice. */
999 retry:
1000         lock_ptr = q->lock_ptr;
1001         barrier();
1002         if (lock_ptr != NULL) {
1003                 spin_lock(lock_ptr);
1004                 /*
1005                  * q->lock_ptr can change between reading it and
1006                  * spin_lock(), causing us to take the wrong lock.  This
1007                  * corrects the race condition.
1008                  *
1009                  * Reasoning goes like this: if we have the wrong lock,
1010                  * q->lock_ptr must have changed (maybe several times)
1011                  * between reading it and the spin_lock().  It can
1012                  * change again after the spin_lock() but only if it was
1013                  * already changed before the spin_lock().  It cannot,
1014                  * however, change back to the original value.  Therefore
1015                  * we can detect whether we acquired the correct lock.
1016                  */
1017                 if (unlikely(lock_ptr != q->lock_ptr)) {
1018                         spin_unlock(lock_ptr);
1019                         goto retry;
1020                 }
1021                 WARN_ON(plist_node_empty(&q->list));
1022                 plist_del(&q->list, &q->list.plist);
1023
1024                 BUG_ON(q->pi_state);
1025
1026                 spin_unlock(lock_ptr);
1027                 ret = 1;
1028         }
1029
1030         drop_futex_key_refs(&q->key);
1031         return ret;
1032 }
1033
1034 /*
1035  * PI futexes can not be requeued and must remove themself from the
1036  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1037  * and dropped here.
1038  */
1039 static void unqueue_me_pi(struct futex_q *q)
1040 {
1041         WARN_ON(plist_node_empty(&q->list));
1042         plist_del(&q->list, &q->list.plist);
1043
1044         BUG_ON(!q->pi_state);
1045         free_pi_state(q->pi_state);
1046         q->pi_state = NULL;
1047
1048         spin_unlock(q->lock_ptr);
1049
1050         drop_futex_key_refs(&q->key);
1051 }
1052
1053 /*
1054  * Fixup the pi_state owner with the new owner.
1055  *
1056  * Must be called with hash bucket lock held and mm->sem held for non
1057  * private futexes.
1058  */
1059 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1060                                 struct task_struct *newowner, int fshared)
1061 {
1062         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1063         struct futex_pi_state *pi_state = q->pi_state;
1064         struct task_struct *oldowner = pi_state->owner;
1065         u32 uval, curval, newval;
1066         int ret, attempt = 0;
1067
1068         /* Owner died? */
1069         if (!pi_state->owner)
1070                 newtid |= FUTEX_OWNER_DIED;
1071
1072         /*
1073          * We are here either because we stole the rtmutex from the
1074          * pending owner or we are the pending owner which failed to
1075          * get the rtmutex. We have to replace the pending owner TID
1076          * in the user space variable. This must be atomic as we have
1077          * to preserve the owner died bit here.
1078          *
1079          * Note: We write the user space value _before_ changing the
1080          * pi_state because we can fault here. Imagine swapped out
1081          * pages or a fork, which was running right before we acquired
1082          * mmap_sem, that marked all the anonymous memory readonly for
1083          * cow.
1084          *
1085          * Modifying pi_state _before_ the user space value would
1086          * leave the pi_state in an inconsistent state when we fault
1087          * here, because we need to drop the hash bucket lock to
1088          * handle the fault. This might be observed in the PID check
1089          * in lookup_pi_state.
1090          */
1091 retry:
1092         if (get_futex_value_locked(&uval, uaddr))
1093                 goto handle_fault;
1094
1095         while (1) {
1096                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1097
1098                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1099
1100                 if (curval == -EFAULT)
1101                         goto handle_fault;
1102                 if (curval == uval)
1103                         break;
1104                 uval = curval;
1105         }
1106
1107         /*
1108          * We fixed up user space. Now we need to fix the pi_state
1109          * itself.
1110          */
1111         if (pi_state->owner != NULL) {
1112                 spin_lock_irq(&pi_state->owner->pi_lock);
1113                 WARN_ON(list_empty(&pi_state->list));
1114                 list_del_init(&pi_state->list);
1115                 spin_unlock_irq(&pi_state->owner->pi_lock);
1116         }
1117
1118         pi_state->owner = newowner;
1119
1120         spin_lock_irq(&newowner->pi_lock);
1121         WARN_ON(!list_empty(&pi_state->list));
1122         list_add(&pi_state->list, &newowner->pi_state_list);
1123         spin_unlock_irq(&newowner->pi_lock);
1124         return 0;
1125
1126         /*
1127          * To handle the page fault we need to drop the hash bucket
1128          * lock here. That gives the other task (either the pending
1129          * owner itself or the task which stole the rtmutex) the
1130          * chance to try the fixup of the pi_state. So once we are
1131          * back from handling the fault we need to check the pi_state
1132          * after reacquiring the hash bucket lock and before trying to
1133          * do another fixup. When the fixup has been done already we
1134          * simply return.
1135          */
1136 handle_fault:
1137         spin_unlock(q->lock_ptr);
1138
1139         ret = futex_handle_fault((unsigned long)uaddr, attempt++);
1140
1141         spin_lock(q->lock_ptr);
1142
1143         /*
1144          * Check if someone else fixed it for us:
1145          */
1146         if (pi_state->owner != oldowner)
1147                 return 0;
1148
1149         if (ret)
1150                 return ret;
1151
1152         goto retry;
1153 }
1154
1155 /*
1156  * In case we must use restart_block to restart a futex_wait,
1157  * we encode in the 'flags' shared capability
1158  */
1159 #define FLAGS_SHARED            0x01
1160 #define FLAGS_CLOCKRT           0x02
1161
1162 static long futex_wait_restart(struct restart_block *restart);
1163
1164 static int futex_wait(u32 __user *uaddr, int fshared,
1165                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1166 {
1167         struct task_struct *curr = current;
1168         DECLARE_WAITQUEUE(wait, curr);
1169         struct futex_hash_bucket *hb;
1170         struct futex_q q;
1171         u32 uval;
1172         int ret;
1173         struct hrtimer_sleeper t;
1174         int rem = 0;
1175
1176         if (!bitset)
1177                 return -EINVAL;
1178
1179         q.pi_state = NULL;
1180         q.bitset = bitset;
1181 retry:
1182         q.key = FUTEX_KEY_INIT;
1183         ret = get_futex_key(uaddr, fshared, &q.key);
1184         if (unlikely(ret != 0))
1185                 goto out;
1186
1187         hb = queue_lock(&q);
1188
1189         /*
1190          * Access the page AFTER the futex is queued.
1191          * Order is important:
1192          *
1193          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1194          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1195          *
1196          * The basic logical guarantee of a futex is that it blocks ONLY
1197          * if cond(var) is known to be true at the time of blocking, for
1198          * any cond.  If we queued after testing *uaddr, that would open
1199          * a race condition where we could block indefinitely with
1200          * cond(var) false, which would violate the guarantee.
1201          *
1202          * A consequence is that futex_wait() can return zero and absorb
1203          * a wakeup when *uaddr != val on entry to the syscall.  This is
1204          * rare, but normal.
1205          *
1206          * for shared futexes, we hold the mmap semaphore, so the mapping
1207          * cannot have changed since we looked it up in get_futex_key.
1208          */
1209         ret = get_futex_value_locked(&uval, uaddr);
1210
1211         if (unlikely(ret)) {
1212                 queue_unlock(&q, hb);
1213                 put_futex_key(fshared, &q.key);
1214
1215                 ret = get_user(uval, uaddr);
1216
1217                 if (!ret)
1218                         goto retry;
1219                 return ret;
1220         }
1221         ret = -EWOULDBLOCK;
1222         if (uval != val)
1223                 goto out_unlock_put_key;
1224
1225         /* Only actually queue if *uaddr contained val.  */
1226         queue_me(&q, hb);
1227
1228         /*
1229          * There might have been scheduling since the queue_me(), as we
1230          * cannot hold a spinlock across the get_user() in case it
1231          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1232          * queueing ourselves into the futex hash.  This code thus has to
1233          * rely on the futex_wake() code removing us from hash when it
1234          * wakes us up.
1235          */
1236
1237         /* add_wait_queue is the barrier after __set_current_state. */
1238         __set_current_state(TASK_INTERRUPTIBLE);
1239         add_wait_queue(&q.waiter, &wait);
1240         /*
1241          * !plist_node_empty() is safe here without any lock.
1242          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1243          */
1244         if (likely(!plist_node_empty(&q.list))) {
1245                 if (!abs_time)
1246                         schedule();
1247                 else {
1248                         unsigned long slack;
1249                         slack = current->timer_slack_ns;
1250                         if (rt_task(current))
1251                                 slack = 0;
1252                         hrtimer_init_on_stack(&t.timer,
1253                                               clockrt ? CLOCK_REALTIME :
1254                                               CLOCK_MONOTONIC,
1255                                               HRTIMER_MODE_ABS);
1256                         hrtimer_init_sleeper(&t, current);
1257                         hrtimer_set_expires_range_ns(&t.timer, *abs_time, slack);
1258
1259                         hrtimer_start_expires(&t.timer, HRTIMER_MODE_ABS);
1260                         if (!hrtimer_active(&t.timer))
1261                                 t.task = NULL;
1262
1263                         /*
1264                          * the timer could have already expired, in which
1265                          * case current would be flagged for rescheduling.
1266                          * Don't bother calling schedule.
1267                          */
1268                         if (likely(t.task))
1269                                 schedule();
1270
1271                         hrtimer_cancel(&t.timer);
1272
1273                         /* Flag if a timeout occured */
1274                         rem = (t.task == NULL);
1275
1276                         destroy_hrtimer_on_stack(&t.timer);
1277                 }
1278         }
1279         __set_current_state(TASK_RUNNING);
1280
1281         /*
1282          * NOTE: we don't remove ourselves from the waitqueue because
1283          * we are the only user of it.
1284          */
1285
1286         /* If we were woken (and unqueued), we succeeded, whatever. */
1287         if (!unqueue_me(&q))
1288                 return 0;
1289         if (rem)
1290                 return -ETIMEDOUT;
1291
1292         /*
1293          * We expect signal_pending(current), but another thread may
1294          * have handled it for us already.
1295          */
1296         if (!abs_time)
1297                 return -ERESTARTSYS;
1298         else {
1299                 struct restart_block *restart;
1300                 restart = &current_thread_info()->restart_block;
1301                 restart->fn = futex_wait_restart;
1302                 restart->futex.uaddr = (u32 *)uaddr;
1303                 restart->futex.val = val;
1304                 restart->futex.time = abs_time->tv64;
1305                 restart->futex.bitset = bitset;
1306                 restart->futex.flags = 0;
1307
1308                 if (fshared)
1309                         restart->futex.flags |= FLAGS_SHARED;
1310                 if (clockrt)
1311                         restart->futex.flags |= FLAGS_CLOCKRT;
1312                 return -ERESTART_RESTARTBLOCK;
1313         }
1314
1315 out_unlock_put_key:
1316         queue_unlock(&q, hb);
1317         put_futex_key(fshared, &q.key);
1318
1319 out:
1320         return ret;
1321 }
1322
1323
1324 static long futex_wait_restart(struct restart_block *restart)
1325 {
1326         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1327         int fshared = 0;
1328         ktime_t t;
1329
1330         t.tv64 = restart->futex.time;
1331         restart->fn = do_no_restart_syscall;
1332         if (restart->futex.flags & FLAGS_SHARED)
1333                 fshared = 1;
1334         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1335                                 restart->futex.bitset,
1336                                 restart->futex.flags & FLAGS_CLOCKRT);
1337 }
1338
1339
1340 /*
1341  * Userspace tried a 0 -> TID atomic transition of the futex value
1342  * and failed. The kernel side here does the whole locking operation:
1343  * if there are waiters then it will block, it does PI, etc. (Due to
1344  * races the kernel might see a 0 value of the futex too.)
1345  */
1346 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1347                          int detect, ktime_t *time, int trylock)
1348 {
1349         struct hrtimer_sleeper timeout, *to = NULL;
1350         struct task_struct *curr = current;
1351         struct futex_hash_bucket *hb;
1352         u32 uval, newval, curval;
1353         struct futex_q q;
1354         int ret, lock_taken, ownerdied = 0, attempt = 0;
1355
1356         if (refill_pi_state_cache())
1357                 return -ENOMEM;
1358
1359         if (time) {
1360                 to = &timeout;
1361                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1362                                       HRTIMER_MODE_ABS);
1363                 hrtimer_init_sleeper(to, current);
1364                 hrtimer_set_expires(&to->timer, *time);
1365         }
1366
1367         q.pi_state = NULL;
1368 retry:
1369         q.key = FUTEX_KEY_INIT;
1370         ret = get_futex_key(uaddr, fshared, &q.key);
1371         if (unlikely(ret != 0))
1372                 goto out;
1373
1374 retry_unlocked:
1375         hb = queue_lock(&q);
1376
1377 retry_locked:
1378         ret = lock_taken = 0;
1379
1380         /*
1381          * To avoid races, we attempt to take the lock here again
1382          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1383          * the locks. It will most likely not succeed.
1384          */
1385         newval = task_pid_vnr(current);
1386
1387         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1388
1389         if (unlikely(curval == -EFAULT))
1390                 goto uaddr_faulted;
1391
1392         /*
1393          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1394          * situation and we return success to user space.
1395          */
1396         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1397                 ret = -EDEADLK;
1398                 goto out_unlock_put_key;
1399         }
1400
1401         /*
1402          * Surprise - we got the lock. Just return to userspace:
1403          */
1404         if (unlikely(!curval))
1405                 goto out_unlock_put_key;
1406
1407         uval = curval;
1408
1409         /*
1410          * Set the WAITERS flag, so the owner will know it has someone
1411          * to wake at next unlock
1412          */
1413         newval = curval | FUTEX_WAITERS;
1414
1415         /*
1416          * There are two cases, where a futex might have no owner (the
1417          * owner TID is 0): OWNER_DIED. We take over the futex in this
1418          * case. We also do an unconditional take over, when the owner
1419          * of the futex died.
1420          *
1421          * This is safe as we are protected by the hash bucket lock !
1422          */
1423         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1424                 /* Keep the OWNER_DIED bit */
1425                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1426                 ownerdied = 0;
1427                 lock_taken = 1;
1428         }
1429
1430         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1431
1432         if (unlikely(curval == -EFAULT))
1433                 goto uaddr_faulted;
1434         if (unlikely(curval != uval))
1435                 goto retry_locked;
1436
1437         /*
1438          * We took the lock due to owner died take over.
1439          */
1440         if (unlikely(lock_taken))
1441                 goto out_unlock_put_key;
1442
1443         /*
1444          * We dont have the lock. Look up the PI state (or create it if
1445          * we are the first waiter):
1446          */
1447         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1448
1449         if (unlikely(ret)) {
1450                 switch (ret) {
1451
1452                 case -EAGAIN:
1453                         /*
1454                          * Task is exiting and we just wait for the
1455                          * exit to complete.
1456                          */
1457                         queue_unlock(&q, hb);
1458                         cond_resched();
1459                         goto retry;
1460
1461                 case -ESRCH:
1462                         /*
1463                          * No owner found for this futex. Check if the
1464                          * OWNER_DIED bit is set to figure out whether
1465                          * this is a robust futex or not.
1466                          */
1467                         if (get_futex_value_locked(&curval, uaddr))
1468                                 goto uaddr_faulted;
1469
1470                         /*
1471                          * We simply start over in case of a robust
1472                          * futex. The code above will take the futex
1473                          * and return happy.
1474                          */
1475                         if (curval & FUTEX_OWNER_DIED) {
1476                                 ownerdied = 1;
1477                                 goto retry_locked;
1478                         }
1479                 default:
1480                         goto out_unlock_put_key;
1481                 }
1482         }
1483
1484         /*
1485          * Only actually queue now that the atomic ops are done:
1486          */
1487         queue_me(&q, hb);
1488
1489         WARN_ON(!q.pi_state);
1490         /*
1491          * Block on the PI mutex:
1492          */
1493         if (!trylock)
1494                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1495         else {
1496                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1497                 /* Fixup the trylock return value: */
1498                 ret = ret ? 0 : -EWOULDBLOCK;
1499         }
1500
1501         spin_lock(q.lock_ptr);
1502
1503         if (!ret) {
1504                 /*
1505                  * Got the lock. We might not be the anticipated owner
1506                  * if we did a lock-steal - fix up the PI-state in
1507                  * that case:
1508                  */
1509                 if (q.pi_state->owner != curr)
1510                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1511         } else {
1512                 /*
1513                  * Catch the rare case, where the lock was released
1514                  * when we were on the way back before we locked the
1515                  * hash bucket.
1516                  */
1517                 if (q.pi_state->owner == curr) {
1518                         /*
1519                          * Try to get the rt_mutex now. This might
1520                          * fail as some other task acquired the
1521                          * rt_mutex after we removed ourself from the
1522                          * rt_mutex waiters list.
1523                          */
1524                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1525                                 ret = 0;
1526                         else {
1527                                 /*
1528                                  * pi_state is incorrect, some other
1529                                  * task did a lock steal and we
1530                                  * returned due to timeout or signal
1531                                  * without taking the rt_mutex. Too
1532                                  * late. We can access the
1533                                  * rt_mutex_owner without locking, as
1534                                  * the other task is now blocked on
1535                                  * the hash bucket lock. Fix the state
1536                                  * up.
1537                                  */
1538                                 struct task_struct *owner;
1539                                 int res;
1540
1541                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1542                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1543                                                            fshared);
1544
1545                                 /* propagate -EFAULT, if the fixup failed */
1546                                 if (res)
1547                                         ret = res;
1548                         }
1549                 } else {
1550                         /*
1551                          * Paranoia check. If we did not take the lock
1552                          * in the trylock above, then we should not be
1553                          * the owner of the rtmutex, neither the real
1554                          * nor the pending one:
1555                          */
1556                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1557                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1558                                        "pi-mutex: %p pi-state %p\n", ret,
1559                                        q.pi_state->pi_mutex.owner,
1560                                        q.pi_state->owner);
1561                 }
1562         }
1563
1564         /* Unqueue and drop the lock */
1565         unqueue_me_pi(&q);
1566
1567         if (to)
1568                 destroy_hrtimer_on_stack(&to->timer);
1569         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1570
1571 out_unlock_put_key:
1572         queue_unlock(&q, hb);
1573
1574 out_put_key:
1575         put_futex_key(fshared, &q.key);
1576 out:
1577         if (to)
1578                 destroy_hrtimer_on_stack(&to->timer);
1579         return ret;
1580
1581 uaddr_faulted:
1582         /*
1583          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1584          * atomically.  Therefore, if we continue to fault after get_user()
1585          * below, we need to handle the fault ourselves, while still holding
1586          * the mmap_sem.  This can occur if the uaddr is under contention as
1587          * we have to drop the mmap_sem in order to call get_user().
1588          */
1589         queue_unlock(&q, hb);
1590
1591         if (attempt++) {
1592                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1593                 if (ret)
1594                         goto out_put_key;
1595                 goto retry_unlocked;
1596         }
1597
1598         ret = get_user(uval, uaddr);
1599         if (!ret)
1600                 goto retry;
1601
1602         if (to)
1603                 destroy_hrtimer_on_stack(&to->timer);
1604         return ret;
1605 }
1606
1607 /*
1608  * Userspace attempted a TID -> 0 atomic transition, and failed.
1609  * This is the in-kernel slowpath: we look up the PI state (if any),
1610  * and do the rt-mutex unlock.
1611  */
1612 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1613 {
1614         struct futex_hash_bucket *hb;
1615         struct futex_q *this, *next;
1616         u32 uval;
1617         struct plist_head *head;
1618         union futex_key key = FUTEX_KEY_INIT;
1619         int ret, attempt = 0;
1620
1621 retry:
1622         if (get_user(uval, uaddr))
1623                 return -EFAULT;
1624         /*
1625          * We release only a lock we actually own:
1626          */
1627         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1628                 return -EPERM;
1629
1630         ret = get_futex_key(uaddr, fshared, &key);
1631         if (unlikely(ret != 0))
1632                 goto out;
1633
1634         hb = hash_futex(&key);
1635 retry_unlocked:
1636         spin_lock(&hb->lock);
1637
1638         /*
1639          * To avoid races, try to do the TID -> 0 atomic transition
1640          * again. If it succeeds then we can return without waking
1641          * anyone else up:
1642          */
1643         if (!(uval & FUTEX_OWNER_DIED))
1644                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1645
1646
1647         if (unlikely(uval == -EFAULT))
1648                 goto pi_faulted;
1649         /*
1650          * Rare case: we managed to release the lock atomically,
1651          * no need to wake anyone else up:
1652          */
1653         if (unlikely(uval == task_pid_vnr(current)))
1654                 goto out_unlock;
1655
1656         /*
1657          * Ok, other tasks may need to be woken up - check waiters
1658          * and do the wakeup if necessary:
1659          */
1660         head = &hb->chain;
1661
1662         plist_for_each_entry_safe(this, next, head, list) {
1663                 if (!match_futex (&this->key, &key))
1664                         continue;
1665                 ret = wake_futex_pi(uaddr, uval, this);
1666                 /*
1667                  * The atomic access to the futex value
1668                  * generated a pagefault, so retry the
1669                  * user-access and the wakeup:
1670                  */
1671                 if (ret == -EFAULT)
1672                         goto pi_faulted;
1673                 goto out_unlock;
1674         }
1675         /*
1676          * No waiters - kernel unlocks the futex:
1677          */
1678         if (!(uval & FUTEX_OWNER_DIED)) {
1679                 ret = unlock_futex_pi(uaddr, uval);
1680                 if (ret == -EFAULT)
1681                         goto pi_faulted;
1682         }
1683
1684 out_unlock:
1685         spin_unlock(&hb->lock);
1686         put_futex_key(fshared, &key);
1687
1688 out:
1689         return ret;
1690
1691 pi_faulted:
1692         /*
1693          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1694          * atomically.  Therefore, if we continue to fault after get_user()
1695          * below, we need to handle the fault ourselves, while still holding
1696          * the mmap_sem.  This can occur if the uaddr is under contention as
1697          * we have to drop the mmap_sem in order to call get_user().
1698          */
1699         spin_unlock(&hb->lock);
1700
1701         if (attempt++) {
1702                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1703                 if (ret)
1704                         goto out;
1705                 uval = 0;
1706                 goto retry_unlocked;
1707         }
1708
1709         ret = get_user(uval, uaddr);
1710         if (!ret)
1711                 goto retry;
1712
1713         return ret;
1714 }
1715
1716 /*
1717  * Support for robust futexes: the kernel cleans up held futexes at
1718  * thread exit time.
1719  *
1720  * Implementation: user-space maintains a per-thread list of locks it
1721  * is holding. Upon do_exit(), the kernel carefully walks this list,
1722  * and marks all locks that are owned by this thread with the
1723  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1724  * always manipulated with the lock held, so the list is private and
1725  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1726  * field, to allow the kernel to clean up if the thread dies after
1727  * acquiring the lock, but just before it could have added itself to
1728  * the list. There can only be one such pending lock.
1729  */
1730
1731 /**
1732  * sys_set_robust_list - set the robust-futex list head of a task
1733  * @head: pointer to the list-head
1734  * @len: length of the list-head, as userspace expects
1735  */
1736 asmlinkage long
1737 sys_set_robust_list(struct robust_list_head __user *head,
1738                     size_t len)
1739 {
1740         if (!futex_cmpxchg_enabled)
1741                 return -ENOSYS;
1742         /*
1743          * The kernel knows only one size for now:
1744          */
1745         if (unlikely(len != sizeof(*head)))
1746                 return -EINVAL;
1747
1748         current->robust_list = head;
1749
1750         return 0;
1751 }
1752
1753 /**
1754  * sys_get_robust_list - get the robust-futex list head of a task
1755  * @pid: pid of the process [zero for current task]
1756  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1757  * @len_ptr: pointer to a length field, the kernel fills in the header size
1758  */
1759 asmlinkage long
1760 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1761                     size_t __user *len_ptr)
1762 {
1763         struct robust_list_head __user *head;
1764         unsigned long ret;
1765         const struct cred *cred = current_cred(), *pcred;
1766
1767         if (!futex_cmpxchg_enabled)
1768                 return -ENOSYS;
1769
1770         if (!pid)
1771                 head = current->robust_list;
1772         else {
1773                 struct task_struct *p;
1774
1775                 ret = -ESRCH;
1776                 rcu_read_lock();
1777                 p = find_task_by_vpid(pid);
1778                 if (!p)
1779                         goto err_unlock;
1780                 ret = -EPERM;
1781                 pcred = __task_cred(p);
1782                 if (cred->euid != pcred->euid &&
1783                     cred->euid != pcred->uid &&
1784                     !capable(CAP_SYS_PTRACE))
1785                         goto err_unlock;
1786                 head = p->robust_list;
1787                 rcu_read_unlock();
1788         }
1789
1790         if (put_user(sizeof(*head), len_ptr))
1791                 return -EFAULT;
1792         return put_user(head, head_ptr);
1793
1794 err_unlock:
1795         rcu_read_unlock();
1796
1797         return ret;
1798 }
1799
1800 /*
1801  * Process a futex-list entry, check whether it's owned by the
1802  * dying task, and do notification if so:
1803  */
1804 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1805 {
1806         u32 uval, nval, mval;
1807
1808 retry:
1809         if (get_user(uval, uaddr))
1810                 return -1;
1811
1812         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1813                 /*
1814                  * Ok, this dying thread is truly holding a futex
1815                  * of interest. Set the OWNER_DIED bit atomically
1816                  * via cmpxchg, and if the value had FUTEX_WAITERS
1817                  * set, wake up a waiter (if any). (We have to do a
1818                  * futex_wake() even if OWNER_DIED is already set -
1819                  * to handle the rare but possible case of recursive
1820                  * thread-death.) The rest of the cleanup is done in
1821                  * userspace.
1822                  */
1823                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1824                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1825
1826                 if (nval == -EFAULT)
1827                         return -1;
1828
1829                 if (nval != uval)
1830                         goto retry;
1831
1832                 /*
1833                  * Wake robust non-PI futexes here. The wakeup of
1834                  * PI futexes happens in exit_pi_state():
1835                  */
1836                 if (!pi && (uval & FUTEX_WAITERS))
1837                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1838         }
1839         return 0;
1840 }
1841
1842 /*
1843  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1844  */
1845 static inline int fetch_robust_entry(struct robust_list __user **entry,
1846                                      struct robust_list __user * __user *head,
1847                                      int *pi)
1848 {
1849         unsigned long uentry;
1850
1851         if (get_user(uentry, (unsigned long __user *)head))
1852                 return -EFAULT;
1853
1854         *entry = (void __user *)(uentry & ~1UL);
1855         *pi = uentry & 1;
1856
1857         return 0;
1858 }
1859
1860 /*
1861  * Walk curr->robust_list (very carefully, it's a userspace list!)
1862  * and mark any locks found there dead, and notify any waiters.
1863  *
1864  * We silently return on any sign of list-walking problem.
1865  */
1866 void exit_robust_list(struct task_struct *curr)
1867 {
1868         struct robust_list_head __user *head = curr->robust_list;
1869         struct robust_list __user *entry, *next_entry, *pending;
1870         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1871         unsigned long futex_offset;
1872         int rc;
1873
1874         if (!futex_cmpxchg_enabled)
1875                 return;
1876
1877         /*
1878          * Fetch the list head (which was registered earlier, via
1879          * sys_set_robust_list()):
1880          */
1881         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1882                 return;
1883         /*
1884          * Fetch the relative futex offset:
1885          */
1886         if (get_user(futex_offset, &head->futex_offset))
1887                 return;
1888         /*
1889          * Fetch any possibly pending lock-add first, and handle it
1890          * if it exists:
1891          */
1892         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1893                 return;
1894
1895         next_entry = NULL;      /* avoid warning with gcc */
1896         while (entry != &head->list) {
1897                 /*
1898                  * Fetch the next entry in the list before calling
1899                  * handle_futex_death:
1900                  */
1901                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1902                 /*
1903                  * A pending lock might already be on the list, so
1904                  * don't process it twice:
1905                  */
1906                 if (entry != pending)
1907                         if (handle_futex_death((void __user *)entry + futex_offset,
1908                                                 curr, pi))
1909                                 return;
1910                 if (rc)
1911                         return;
1912                 entry = next_entry;
1913                 pi = next_pi;
1914                 /*
1915                  * Avoid excessively long or circular lists:
1916                  */
1917                 if (!--limit)
1918                         break;
1919
1920                 cond_resched();
1921         }
1922
1923         if (pending)
1924                 handle_futex_death((void __user *)pending + futex_offset,
1925                                    curr, pip);
1926 }
1927
1928 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1929                 u32 __user *uaddr2, u32 val2, u32 val3)
1930 {
1931         int clockrt, ret = -ENOSYS;
1932         int cmd = op & FUTEX_CMD_MASK;
1933         int fshared = 0;
1934
1935         if (!(op & FUTEX_PRIVATE_FLAG))
1936                 fshared = 1;
1937
1938         clockrt = op & FUTEX_CLOCK_REALTIME;
1939         if (clockrt && cmd != FUTEX_WAIT_BITSET)
1940                 return -ENOSYS;
1941
1942         switch (cmd) {
1943         case FUTEX_WAIT:
1944                 val3 = FUTEX_BITSET_MATCH_ANY;
1945         case FUTEX_WAIT_BITSET:
1946                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
1947                 break;
1948         case FUTEX_WAKE:
1949                 val3 = FUTEX_BITSET_MATCH_ANY;
1950         case FUTEX_WAKE_BITSET:
1951                 ret = futex_wake(uaddr, fshared, val, val3);
1952                 break;
1953         case FUTEX_REQUEUE:
1954                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1955                 break;
1956         case FUTEX_CMP_REQUEUE:
1957                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1958                 break;
1959         case FUTEX_WAKE_OP:
1960                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1961                 break;
1962         case FUTEX_LOCK_PI:
1963                 if (futex_cmpxchg_enabled)
1964                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1965                 break;
1966         case FUTEX_UNLOCK_PI:
1967                 if (futex_cmpxchg_enabled)
1968                         ret = futex_unlock_pi(uaddr, fshared);
1969                 break;
1970         case FUTEX_TRYLOCK_PI:
1971                 if (futex_cmpxchg_enabled)
1972                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1973                 break;
1974         default:
1975                 ret = -ENOSYS;
1976         }
1977         return ret;
1978 }
1979
1980
1981 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1982                           struct timespec __user *utime, u32 __user *uaddr2,
1983                           u32 val3)
1984 {
1985         struct timespec ts;
1986         ktime_t t, *tp = NULL;
1987         u32 val2 = 0;
1988         int cmd = op & FUTEX_CMD_MASK;
1989
1990         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
1991                       cmd == FUTEX_WAIT_BITSET)) {
1992                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
1993                         return -EFAULT;
1994                 if (!timespec_valid(&ts))
1995                         return -EINVAL;
1996
1997                 t = timespec_to_ktime(ts);
1998                 if (cmd == FUTEX_WAIT)
1999                         t = ktime_add_safe(ktime_get(), t);
2000                 tp = &t;
2001         }
2002         /*
2003          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2004          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2005          */
2006         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2007             cmd == FUTEX_WAKE_OP)
2008                 val2 = (u32) (unsigned long) utime;
2009
2010         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2011 }
2012
2013 static int __init futex_init(void)
2014 {
2015         u32 curval;
2016         int i;
2017
2018         /*
2019          * This will fail and we want it. Some arch implementations do
2020          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2021          * functionality. We want to know that before we call in any
2022          * of the complex code paths. Also we want to prevent
2023          * registration of robust lists in that case. NULL is
2024          * guaranteed to fault and we get -EFAULT on functional
2025          * implementation, the non functional ones will return
2026          * -ENOSYS.
2027          */
2028         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2029         if (curval == -EFAULT)
2030                 futex_cmpxchg_enabled = 1;
2031
2032         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2033                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2034                 spin_lock_init(&futex_queues[i].lock);
2035         }
2036
2037         return 0;
2038 }
2039 __initcall(futex_init);