CRED: Wrap task credential accesses in the core kernel
[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  * get_futex_key - Get parameters which are the keys for a futex.
166  * @uaddr: virtual address of the futex
167  * @shared: NULL for a PROCESS_PRIVATE futex,
168  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
169  * @key: address where result is stored.
170  *
171  * Returns a negative error code or 0
172  * The key words are stored in *key on success.
173  *
174  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
175  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
176  * We can usually work out the index without swapping in the page.
177  *
178  * fshared is NULL for PROCESS_PRIVATE futexes
179  * For other futexes, it points to &current->mm->mmap_sem and
180  * caller must have taken the reader lock. but NOT any spinlocks.
181  */
182 static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
183                          union futex_key *key)
184 {
185         unsigned long address = (unsigned long)uaddr;
186         struct mm_struct *mm = current->mm;
187         struct vm_area_struct *vma;
188         struct page *page;
189         int err;
190
191         /*
192          * The futex address must be "naturally" aligned.
193          */
194         key->both.offset = address % PAGE_SIZE;
195         if (unlikely((address % sizeof(u32)) != 0))
196                 return -EINVAL;
197         address -= key->both.offset;
198
199         /*
200          * PROCESS_PRIVATE futexes are fast.
201          * As the mm cannot disappear under us and the 'key' only needs
202          * virtual address, we dont even have to find the underlying vma.
203          * Note : We do have to check 'uaddr' is a valid user address,
204          *        but access_ok() should be faster than find_vma()
205          */
206         if (!fshared) {
207                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
208                         return -EFAULT;
209                 key->private.mm = mm;
210                 key->private.address = address;
211                 return 0;
212         }
213         /*
214          * The futex is hashed differently depending on whether
215          * it's in a shared or private mapping.  So check vma first.
216          */
217         vma = find_extend_vma(mm, address);
218         if (unlikely(!vma))
219                 return -EFAULT;
220
221         /*
222          * Permissions.
223          */
224         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
225                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
226
227         /*
228          * Private mappings are handled in a simple way.
229          *
230          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
231          * it's a read-only handle, it's expected that futexes attach to
232          * the object not the particular process.  Therefore we use
233          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
234          * mappings of _writable_ handles.
235          */
236         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
237                 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
238                 key->private.mm = mm;
239                 key->private.address = address;
240                 return 0;
241         }
242
243         /*
244          * Linear file mappings are also simple.
245          */
246         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
247         key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
248         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
249                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
250                                      + vma->vm_pgoff);
251                 return 0;
252         }
253
254         /*
255          * We could walk the page table to read the non-linear
256          * pte, and get the page index without fetching the page
257          * from swap.  But that's a lot of code to duplicate here
258          * for a rare case, so we simply fetch the page.
259          */
260         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
261         if (err >= 0) {
262                 key->shared.pgoff =
263                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
264                 put_page(page);
265                 return 0;
266         }
267         return err;
268 }
269
270 /*
271  * Take a reference to the resource addressed by a key.
272  * Can be called while holding spinlocks.
273  *
274  */
275 static void get_futex_key_refs(union futex_key *key)
276 {
277         if (key->both.ptr == NULL)
278                 return;
279         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
280                 case FUT_OFF_INODE:
281                         atomic_inc(&key->shared.inode->i_count);
282                         break;
283                 case FUT_OFF_MMSHARED:
284                         atomic_inc(&key->private.mm->mm_count);
285                         break;
286         }
287 }
288
289 /*
290  * Drop a reference to the resource addressed by a key.
291  * The hash bucket spinlock must not be held.
292  */
293 static void drop_futex_key_refs(union futex_key *key)
294 {
295         if (!key->both.ptr)
296                 return;
297         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
298                 case FUT_OFF_INODE:
299                         iput(key->shared.inode);
300                         break;
301                 case FUT_OFF_MMSHARED:
302                         mmdrop(key->private.mm);
303                         break;
304         }
305 }
306
307 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
308 {
309         u32 curval;
310
311         pagefault_disable();
312         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
313         pagefault_enable();
314
315         return curval;
316 }
317
318 static int get_futex_value_locked(u32 *dest, u32 __user *from)
319 {
320         int ret;
321
322         pagefault_disable();
323         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
324         pagefault_enable();
325
326         return ret ? -EFAULT : 0;
327 }
328
329 /*
330  * Fault handling.
331  * if fshared is non NULL, current->mm->mmap_sem is already held
332  */
333 static int futex_handle_fault(unsigned long address,
334                               struct rw_semaphore *fshared, int attempt)
335 {
336         struct vm_area_struct * vma;
337         struct mm_struct *mm = current->mm;
338         int ret = -EFAULT;
339
340         if (attempt > 2)
341                 return ret;
342
343         if (!fshared)
344                 down_read(&mm->mmap_sem);
345         vma = find_vma(mm, address);
346         if (vma && address >= vma->vm_start &&
347             (vma->vm_flags & VM_WRITE)) {
348                 int fault;
349                 fault = handle_mm_fault(mm, vma, address, 1);
350                 if (unlikely((fault & VM_FAULT_ERROR))) {
351 #if 0
352                         /* XXX: let's do this when we verify it is OK */
353                         if (ret & VM_FAULT_OOM)
354                                 ret = -ENOMEM;
355 #endif
356                 } else {
357                         ret = 0;
358                         if (fault & VM_FAULT_MAJOR)
359                                 current->maj_flt++;
360                         else
361                                 current->min_flt++;
362                 }
363         }
364         if (!fshared)
365                 up_read(&mm->mmap_sem);
366         return ret;
367 }
368
369 /*
370  * PI code:
371  */
372 static int refill_pi_state_cache(void)
373 {
374         struct futex_pi_state *pi_state;
375
376         if (likely(current->pi_state_cache))
377                 return 0;
378
379         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
380
381         if (!pi_state)
382                 return -ENOMEM;
383
384         INIT_LIST_HEAD(&pi_state->list);
385         /* pi_mutex gets initialized later */
386         pi_state->owner = NULL;
387         atomic_set(&pi_state->refcount, 1);
388
389         current->pi_state_cache = pi_state;
390
391         return 0;
392 }
393
394 static struct futex_pi_state * alloc_pi_state(void)
395 {
396         struct futex_pi_state *pi_state = current->pi_state_cache;
397
398         WARN_ON(!pi_state);
399         current->pi_state_cache = NULL;
400
401         return pi_state;
402 }
403
404 static void free_pi_state(struct futex_pi_state *pi_state)
405 {
406         if (!atomic_dec_and_test(&pi_state->refcount))
407                 return;
408
409         /*
410          * If pi_state->owner is NULL, the owner is most probably dying
411          * and has cleaned up the pi_state already
412          */
413         if (pi_state->owner) {
414                 spin_lock_irq(&pi_state->owner->pi_lock);
415                 list_del_init(&pi_state->list);
416                 spin_unlock_irq(&pi_state->owner->pi_lock);
417
418                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
419         }
420
421         if (current->pi_state_cache)
422                 kfree(pi_state);
423         else {
424                 /*
425                  * pi_state->list is already empty.
426                  * clear pi_state->owner.
427                  * refcount is at 0 - put it back to 1.
428                  */
429                 pi_state->owner = NULL;
430                 atomic_set(&pi_state->refcount, 1);
431                 current->pi_state_cache = pi_state;
432         }
433 }
434
435 /*
436  * Look up the task based on what TID userspace gave us.
437  * We dont trust it.
438  */
439 static struct task_struct * futex_find_get_task(pid_t pid)
440 {
441         struct task_struct *p;
442         uid_t euid = current_euid();
443
444         rcu_read_lock();
445         p = find_task_by_vpid(pid);
446         if (!p || (euid != p->euid && euid != p->uid))
447                 p = ERR_PTR(-ESRCH);
448         else
449                 get_task_struct(p);
450
451         rcu_read_unlock();
452
453         return p;
454 }
455
456 /*
457  * This task is holding PI mutexes at exit time => bad.
458  * Kernel cleans up PI-state, but userspace is likely hosed.
459  * (Robust-futex cleanup is separate and might save the day for userspace.)
460  */
461 void exit_pi_state_list(struct task_struct *curr)
462 {
463         struct list_head *next, *head = &curr->pi_state_list;
464         struct futex_pi_state *pi_state;
465         struct futex_hash_bucket *hb;
466         union futex_key key;
467
468         if (!futex_cmpxchg_enabled)
469                 return;
470         /*
471          * We are a ZOMBIE and nobody can enqueue itself on
472          * pi_state_list anymore, but we have to be careful
473          * versus waiters unqueueing themselves:
474          */
475         spin_lock_irq(&curr->pi_lock);
476         while (!list_empty(head)) {
477
478                 next = head->next;
479                 pi_state = list_entry(next, struct futex_pi_state, list);
480                 key = pi_state->key;
481                 hb = hash_futex(&key);
482                 spin_unlock_irq(&curr->pi_lock);
483
484                 spin_lock(&hb->lock);
485
486                 spin_lock_irq(&curr->pi_lock);
487                 /*
488                  * We dropped the pi-lock, so re-check whether this
489                  * task still owns the PI-state:
490                  */
491                 if (head->next != next) {
492                         spin_unlock(&hb->lock);
493                         continue;
494                 }
495
496                 WARN_ON(pi_state->owner != curr);
497                 WARN_ON(list_empty(&pi_state->list));
498                 list_del_init(&pi_state->list);
499                 pi_state->owner = NULL;
500                 spin_unlock_irq(&curr->pi_lock);
501
502                 rt_mutex_unlock(&pi_state->pi_mutex);
503
504                 spin_unlock(&hb->lock);
505
506                 spin_lock_irq(&curr->pi_lock);
507         }
508         spin_unlock_irq(&curr->pi_lock);
509 }
510
511 static int
512 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
513                 union futex_key *key, struct futex_pi_state **ps)
514 {
515         struct futex_pi_state *pi_state = NULL;
516         struct futex_q *this, *next;
517         struct plist_head *head;
518         struct task_struct *p;
519         pid_t pid = uval & FUTEX_TID_MASK;
520
521         head = &hb->chain;
522
523         plist_for_each_entry_safe(this, next, head, list) {
524                 if (match_futex(&this->key, key)) {
525                         /*
526                          * Another waiter already exists - bump up
527                          * the refcount and return its pi_state:
528                          */
529                         pi_state = this->pi_state;
530                         /*
531                          * Userspace might have messed up non PI and PI futexes
532                          */
533                         if (unlikely(!pi_state))
534                                 return -EINVAL;
535
536                         WARN_ON(!atomic_read(&pi_state->refcount));
537                         WARN_ON(pid && pi_state->owner &&
538                                 pi_state->owner->pid != pid);
539
540                         atomic_inc(&pi_state->refcount);
541                         *ps = pi_state;
542
543                         return 0;
544                 }
545         }
546
547         /*
548          * We are the first waiter - try to look up the real owner and attach
549          * the new pi_state to it, but bail out when TID = 0
550          */
551         if (!pid)
552                 return -ESRCH;
553         p = futex_find_get_task(pid);
554         if (IS_ERR(p))
555                 return PTR_ERR(p);
556
557         /*
558          * We need to look at the task state flags to figure out,
559          * whether the task is exiting. To protect against the do_exit
560          * change of the task flags, we do this protected by
561          * p->pi_lock:
562          */
563         spin_lock_irq(&p->pi_lock);
564         if (unlikely(p->flags & PF_EXITING)) {
565                 /*
566                  * The task is on the way out. When PF_EXITPIDONE is
567                  * set, we know that the task has finished the
568                  * cleanup:
569                  */
570                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
571
572                 spin_unlock_irq(&p->pi_lock);
573                 put_task_struct(p);
574                 return ret;
575         }
576
577         pi_state = alloc_pi_state();
578
579         /*
580          * Initialize the pi_mutex in locked state and make 'p'
581          * the owner of it:
582          */
583         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
584
585         /* Store the key for possible exit cleanups: */
586         pi_state->key = *key;
587
588         WARN_ON(!list_empty(&pi_state->list));
589         list_add(&pi_state->list, &p->pi_state_list);
590         pi_state->owner = p;
591         spin_unlock_irq(&p->pi_lock);
592
593         put_task_struct(p);
594
595         *ps = pi_state;
596
597         return 0;
598 }
599
600 /*
601  * The hash bucket lock must be held when this is called.
602  * Afterwards, the futex_q must not be accessed.
603  */
604 static void wake_futex(struct futex_q *q)
605 {
606         plist_del(&q->list, &q->list.plist);
607         /*
608          * The lock in wake_up_all() is a crucial memory barrier after the
609          * plist_del() and also before assigning to q->lock_ptr.
610          */
611         wake_up_all(&q->waiters);
612         /*
613          * The waiting task can free the futex_q as soon as this is written,
614          * without taking any locks.  This must come last.
615          *
616          * A memory barrier is required here to prevent the following store
617          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
618          * at the end of wake_up_all() does not prevent this store from
619          * moving.
620          */
621         smp_wmb();
622         q->lock_ptr = NULL;
623 }
624
625 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
626 {
627         struct task_struct *new_owner;
628         struct futex_pi_state *pi_state = this->pi_state;
629         u32 curval, newval;
630
631         if (!pi_state)
632                 return -EINVAL;
633
634         spin_lock(&pi_state->pi_mutex.wait_lock);
635         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
636
637         /*
638          * This happens when we have stolen the lock and the original
639          * pending owner did not enqueue itself back on the rt_mutex.
640          * Thats not a tragedy. We know that way, that a lock waiter
641          * is on the fly. We make the futex_q waiter the pending owner.
642          */
643         if (!new_owner)
644                 new_owner = this->task;
645
646         /*
647          * We pass it to the next owner. (The WAITERS bit is always
648          * kept enabled while there is PI state around. We must also
649          * preserve the owner died bit.)
650          */
651         if (!(uval & FUTEX_OWNER_DIED)) {
652                 int ret = 0;
653
654                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
655
656                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
657
658                 if (curval == -EFAULT)
659                         ret = -EFAULT;
660                 else if (curval != uval)
661                         ret = -EINVAL;
662                 if (ret) {
663                         spin_unlock(&pi_state->pi_mutex.wait_lock);
664                         return ret;
665                 }
666         }
667
668         spin_lock_irq(&pi_state->owner->pi_lock);
669         WARN_ON(list_empty(&pi_state->list));
670         list_del_init(&pi_state->list);
671         spin_unlock_irq(&pi_state->owner->pi_lock);
672
673         spin_lock_irq(&new_owner->pi_lock);
674         WARN_ON(!list_empty(&pi_state->list));
675         list_add(&pi_state->list, &new_owner->pi_state_list);
676         pi_state->owner = new_owner;
677         spin_unlock_irq(&new_owner->pi_lock);
678
679         spin_unlock(&pi_state->pi_mutex.wait_lock);
680         rt_mutex_unlock(&pi_state->pi_mutex);
681
682         return 0;
683 }
684
685 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
686 {
687         u32 oldval;
688
689         /*
690          * There is no waiter, so we unlock the futex. The owner died
691          * bit has not to be preserved here. We are the owner:
692          */
693         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
694
695         if (oldval == -EFAULT)
696                 return oldval;
697         if (oldval != uval)
698                 return -EAGAIN;
699
700         return 0;
701 }
702
703 /*
704  * Express the locking dependencies for lockdep:
705  */
706 static inline void
707 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
708 {
709         if (hb1 <= hb2) {
710                 spin_lock(&hb1->lock);
711                 if (hb1 < hb2)
712                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
713         } else { /* hb1 > hb2 */
714                 spin_lock(&hb2->lock);
715                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
716         }
717 }
718
719 /*
720  * Wake up all waiters hashed on the physical page that is mapped
721  * to this virtual address:
722  */
723 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
724                       int nr_wake, u32 bitset)
725 {
726         struct futex_hash_bucket *hb;
727         struct futex_q *this, *next;
728         struct plist_head *head;
729         union futex_key key;
730         int ret;
731
732         if (!bitset)
733                 return -EINVAL;
734
735         futex_lock_mm(fshared);
736
737         ret = get_futex_key(uaddr, fshared, &key);
738         if (unlikely(ret != 0))
739                 goto out;
740
741         hb = hash_futex(&key);
742         spin_lock(&hb->lock);
743         head = &hb->chain;
744
745         plist_for_each_entry_safe(this, next, head, list) {
746                 if (match_futex (&this->key, &key)) {
747                         if (this->pi_state) {
748                                 ret = -EINVAL;
749                                 break;
750                         }
751
752                         /* Check if one of the bits is set in both bitsets */
753                         if (!(this->bitset & bitset))
754                                 continue;
755
756                         wake_futex(this);
757                         if (++ret >= nr_wake)
758                                 break;
759                 }
760         }
761
762         spin_unlock(&hb->lock);
763 out:
764         futex_unlock_mm(fshared);
765         return ret;
766 }
767
768 /*
769  * Wake up all waiters hashed on the physical page that is mapped
770  * to this virtual address:
771  */
772 static int
773 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
774               u32 __user *uaddr2,
775               int nr_wake, int nr_wake2, int op)
776 {
777         union futex_key key1, key2;
778         struct futex_hash_bucket *hb1, *hb2;
779         struct plist_head *head;
780         struct futex_q *this, *next;
781         int ret, op_ret, attempt = 0;
782
783 retryfull:
784         futex_lock_mm(fshared);
785
786         ret = get_futex_key(uaddr1, fshared, &key1);
787         if (unlikely(ret != 0))
788                 goto out;
789         ret = get_futex_key(uaddr2, fshared, &key2);
790         if (unlikely(ret != 0))
791                 goto out;
792
793         hb1 = hash_futex(&key1);
794         hb2 = hash_futex(&key2);
795
796 retry:
797         double_lock_hb(hb1, hb2);
798
799         op_ret = futex_atomic_op_inuser(op, uaddr2);
800         if (unlikely(op_ret < 0)) {
801                 u32 dummy;
802
803                 spin_unlock(&hb1->lock);
804                 if (hb1 != hb2)
805                         spin_unlock(&hb2->lock);
806
807 #ifndef CONFIG_MMU
808                 /*
809                  * we don't get EFAULT from MMU faults if we don't have an MMU,
810                  * but we might get them from range checking
811                  */
812                 ret = op_ret;
813                 goto out;
814 #endif
815
816                 if (unlikely(op_ret != -EFAULT)) {
817                         ret = op_ret;
818                         goto out;
819                 }
820
821                 /*
822                  * futex_atomic_op_inuser needs to both read and write
823                  * *(int __user *)uaddr2, but we can't modify it
824                  * non-atomically.  Therefore, if get_user below is not
825                  * enough, we need to handle the fault ourselves, while
826                  * still holding the mmap_sem.
827                  */
828                 if (attempt++) {
829                         ret = futex_handle_fault((unsigned long)uaddr2,
830                                                  fshared, attempt);
831                         if (ret)
832                                 goto out;
833                         goto retry;
834                 }
835
836                 /*
837                  * If we would have faulted, release mmap_sem,
838                  * fault it in and start all over again.
839                  */
840                 futex_unlock_mm(fshared);
841
842                 ret = get_user(dummy, uaddr2);
843                 if (ret)
844                         return ret;
845
846                 goto retryfull;
847         }
848
849         head = &hb1->chain;
850
851         plist_for_each_entry_safe(this, next, head, list) {
852                 if (match_futex (&this->key, &key1)) {
853                         wake_futex(this);
854                         if (++ret >= nr_wake)
855                                 break;
856                 }
857         }
858
859         if (op_ret > 0) {
860                 head = &hb2->chain;
861
862                 op_ret = 0;
863                 plist_for_each_entry_safe(this, next, head, list) {
864                         if (match_futex (&this->key, &key2)) {
865                                 wake_futex(this);
866                                 if (++op_ret >= nr_wake2)
867                                         break;
868                         }
869                 }
870                 ret += op_ret;
871         }
872
873         spin_unlock(&hb1->lock);
874         if (hb1 != hb2)
875                 spin_unlock(&hb2->lock);
876 out:
877         futex_unlock_mm(fshared);
878
879         return ret;
880 }
881
882 /*
883  * Requeue all waiters hashed on one physical page to another
884  * physical page.
885  */
886 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
887                          u32 __user *uaddr2,
888                          int nr_wake, int nr_requeue, u32 *cmpval)
889 {
890         union futex_key key1, key2;
891         struct futex_hash_bucket *hb1, *hb2;
892         struct plist_head *head1;
893         struct futex_q *this, *next;
894         int ret, drop_count = 0;
895
896  retry:
897         futex_lock_mm(fshared);
898
899         ret = get_futex_key(uaddr1, fshared, &key1);
900         if (unlikely(ret != 0))
901                 goto out;
902         ret = get_futex_key(uaddr2, fshared, &key2);
903         if (unlikely(ret != 0))
904                 goto out;
905
906         hb1 = hash_futex(&key1);
907         hb2 = hash_futex(&key2);
908
909         double_lock_hb(hb1, hb2);
910
911         if (likely(cmpval != NULL)) {
912                 u32 curval;
913
914                 ret = get_futex_value_locked(&curval, uaddr1);
915
916                 if (unlikely(ret)) {
917                         spin_unlock(&hb1->lock);
918                         if (hb1 != hb2)
919                                 spin_unlock(&hb2->lock);
920
921                         /*
922                          * If we would have faulted, release mmap_sem, fault
923                          * it in and start all over again.
924                          */
925                         futex_unlock_mm(fshared);
926
927                         ret = get_user(curval, uaddr1);
928
929                         if (!ret)
930                                 goto retry;
931
932                         return ret;
933                 }
934                 if (curval != *cmpval) {
935                         ret = -EAGAIN;
936                         goto out_unlock;
937                 }
938         }
939
940         head1 = &hb1->chain;
941         plist_for_each_entry_safe(this, next, head1, list) {
942                 if (!match_futex (&this->key, &key1))
943                         continue;
944                 if (++ret <= nr_wake) {
945                         wake_futex(this);
946                 } else {
947                         /*
948                          * If key1 and key2 hash to the same bucket, no need to
949                          * requeue.
950                          */
951                         if (likely(head1 != &hb2->chain)) {
952                                 plist_del(&this->list, &hb1->chain);
953                                 plist_add(&this->list, &hb2->chain);
954                                 this->lock_ptr = &hb2->lock;
955 #ifdef CONFIG_DEBUG_PI_LIST
956                                 this->list.plist.lock = &hb2->lock;
957 #endif
958                         }
959                         this->key = key2;
960                         get_futex_key_refs(&key2);
961                         drop_count++;
962
963                         if (ret - nr_wake >= nr_requeue)
964                                 break;
965                 }
966         }
967
968 out_unlock:
969         spin_unlock(&hb1->lock);
970         if (hb1 != hb2)
971                 spin_unlock(&hb2->lock);
972
973         /* drop_futex_key_refs() must be called outside the spinlocks. */
974         while (--drop_count >= 0)
975                 drop_futex_key_refs(&key1);
976
977 out:
978         futex_unlock_mm(fshared);
979         return ret;
980 }
981
982 /* The key must be already stored in q->key. */
983 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
984 {
985         struct futex_hash_bucket *hb;
986
987         init_waitqueue_head(&q->waiters);
988
989         get_futex_key_refs(&q->key);
990         hb = hash_futex(&q->key);
991         q->lock_ptr = &hb->lock;
992
993         spin_lock(&hb->lock);
994         return hb;
995 }
996
997 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
998 {
999         int prio;
1000
1001         /*
1002          * The priority used to register this element is
1003          * - either the real thread-priority for the real-time threads
1004          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1005          * - or MAX_RT_PRIO for non-RT threads.
1006          * Thus, all RT-threads are woken first in priority order, and
1007          * the others are woken last, in FIFO order.
1008          */
1009         prio = min(current->normal_prio, MAX_RT_PRIO);
1010
1011         plist_node_init(&q->list, prio);
1012 #ifdef CONFIG_DEBUG_PI_LIST
1013         q->list.plist.lock = &hb->lock;
1014 #endif
1015         plist_add(&q->list, &hb->chain);
1016         q->task = current;
1017         spin_unlock(&hb->lock);
1018 }
1019
1020 static inline void
1021 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1022 {
1023         spin_unlock(&hb->lock);
1024         drop_futex_key_refs(&q->key);
1025 }
1026
1027 /*
1028  * queue_me and unqueue_me must be called as a pair, each
1029  * exactly once.  They are called with the hashed spinlock held.
1030  */
1031
1032 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1033 static int unqueue_me(struct futex_q *q)
1034 {
1035         spinlock_t *lock_ptr;
1036         int ret = 0;
1037
1038         /* In the common case we don't take the spinlock, which is nice. */
1039  retry:
1040         lock_ptr = q->lock_ptr;
1041         barrier();
1042         if (lock_ptr != NULL) {
1043                 spin_lock(lock_ptr);
1044                 /*
1045                  * q->lock_ptr can change between reading it and
1046                  * spin_lock(), causing us to take the wrong lock.  This
1047                  * corrects the race condition.
1048                  *
1049                  * Reasoning goes like this: if we have the wrong lock,
1050                  * q->lock_ptr must have changed (maybe several times)
1051                  * between reading it and the spin_lock().  It can
1052                  * change again after the spin_lock() but only if it was
1053                  * already changed before the spin_lock().  It cannot,
1054                  * however, change back to the original value.  Therefore
1055                  * we can detect whether we acquired the correct lock.
1056                  */
1057                 if (unlikely(lock_ptr != q->lock_ptr)) {
1058                         spin_unlock(lock_ptr);
1059                         goto retry;
1060                 }
1061                 WARN_ON(plist_node_empty(&q->list));
1062                 plist_del(&q->list, &q->list.plist);
1063
1064                 BUG_ON(q->pi_state);
1065
1066                 spin_unlock(lock_ptr);
1067                 ret = 1;
1068         }
1069
1070         drop_futex_key_refs(&q->key);
1071         return ret;
1072 }
1073
1074 /*
1075  * PI futexes can not be requeued and must remove themself from the
1076  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1077  * and dropped here.
1078  */
1079 static void unqueue_me_pi(struct futex_q *q)
1080 {
1081         WARN_ON(plist_node_empty(&q->list));
1082         plist_del(&q->list, &q->list.plist);
1083
1084         BUG_ON(!q->pi_state);
1085         free_pi_state(q->pi_state);
1086         q->pi_state = NULL;
1087
1088         spin_unlock(q->lock_ptr);
1089
1090         drop_futex_key_refs(&q->key);
1091 }
1092
1093 /*
1094  * Fixup the pi_state owner with the new owner.
1095  *
1096  * Must be called with hash bucket lock held and mm->sem held for non
1097  * private futexes.
1098  */
1099 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1100                                 struct task_struct *newowner,
1101                                 struct rw_semaphore *fshared)
1102 {
1103         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1104         struct futex_pi_state *pi_state = q->pi_state;
1105         struct task_struct *oldowner = pi_state->owner;
1106         u32 uval, curval, newval;
1107         int ret, attempt = 0;
1108
1109         /* Owner died? */
1110         if (!pi_state->owner)
1111                 newtid |= FUTEX_OWNER_DIED;
1112
1113         /*
1114          * We are here either because we stole the rtmutex from the
1115          * pending owner or we are the pending owner which failed to
1116          * get the rtmutex. We have to replace the pending owner TID
1117          * in the user space variable. This must be atomic as we have
1118          * to preserve the owner died bit here.
1119          *
1120          * Note: We write the user space value _before_ changing the
1121          * pi_state because we can fault here. Imagine swapped out
1122          * pages or a fork, which was running right before we acquired
1123          * mmap_sem, that marked all the anonymous memory readonly for
1124          * cow.
1125          *
1126          * Modifying pi_state _before_ the user space value would
1127          * leave the pi_state in an inconsistent state when we fault
1128          * here, because we need to drop the hash bucket lock to
1129          * handle the fault. This might be observed in the PID check
1130          * in lookup_pi_state.
1131          */
1132 retry:
1133         if (get_futex_value_locked(&uval, uaddr))
1134                 goto handle_fault;
1135
1136         while (1) {
1137                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1138
1139                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1140
1141                 if (curval == -EFAULT)
1142                         goto handle_fault;
1143                 if (curval == uval)
1144                         break;
1145                 uval = curval;
1146         }
1147
1148         /*
1149          * We fixed up user space. Now we need to fix the pi_state
1150          * itself.
1151          */
1152         if (pi_state->owner != NULL) {
1153                 spin_lock_irq(&pi_state->owner->pi_lock);
1154                 WARN_ON(list_empty(&pi_state->list));
1155                 list_del_init(&pi_state->list);
1156                 spin_unlock_irq(&pi_state->owner->pi_lock);
1157         }
1158
1159         pi_state->owner = newowner;
1160
1161         spin_lock_irq(&newowner->pi_lock);
1162         WARN_ON(!list_empty(&pi_state->list));
1163         list_add(&pi_state->list, &newowner->pi_state_list);
1164         spin_unlock_irq(&newowner->pi_lock);
1165         return 0;
1166
1167         /*
1168          * To handle the page fault we need to drop the hash bucket
1169          * lock here. That gives the other task (either the pending
1170          * owner itself or the task which stole the rtmutex) the
1171          * chance to try the fixup of the pi_state. So once we are
1172          * back from handling the fault we need to check the pi_state
1173          * after reacquiring the hash bucket lock and before trying to
1174          * do another fixup. When the fixup has been done already we
1175          * simply return.
1176          */
1177 handle_fault:
1178         spin_unlock(q->lock_ptr);
1179
1180         ret = futex_handle_fault((unsigned long)uaddr, fshared, attempt++);
1181
1182         spin_lock(q->lock_ptr);
1183
1184         /*
1185          * Check if someone else fixed it for us:
1186          */
1187         if (pi_state->owner != oldowner)
1188                 return 0;
1189
1190         if (ret)
1191                 return ret;
1192
1193         goto retry;
1194 }
1195
1196 /*
1197  * In case we must use restart_block to restart a futex_wait,
1198  * we encode in the 'flags' shared capability
1199  */
1200 #define FLAGS_SHARED  1
1201
1202 static long futex_wait_restart(struct restart_block *restart);
1203
1204 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1205                       u32 val, ktime_t *abs_time, u32 bitset)
1206 {
1207         struct task_struct *curr = current;
1208         DECLARE_WAITQUEUE(wait, curr);
1209         struct futex_hash_bucket *hb;
1210         struct futex_q q;
1211         u32 uval;
1212         int ret;
1213         struct hrtimer_sleeper t;
1214         int rem = 0;
1215
1216         if (!bitset)
1217                 return -EINVAL;
1218
1219         q.pi_state = NULL;
1220         q.bitset = bitset;
1221  retry:
1222         futex_lock_mm(fshared);
1223
1224         ret = get_futex_key(uaddr, fshared, &q.key);
1225         if (unlikely(ret != 0))
1226                 goto out_release_sem;
1227
1228         hb = queue_lock(&q);
1229
1230         /*
1231          * Access the page AFTER the futex is queued.
1232          * Order is important:
1233          *
1234          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1235          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1236          *
1237          * The basic logical guarantee of a futex is that it blocks ONLY
1238          * if cond(var) is known to be true at the time of blocking, for
1239          * any cond.  If we queued after testing *uaddr, that would open
1240          * a race condition where we could block indefinitely with
1241          * cond(var) false, which would violate the guarantee.
1242          *
1243          * A consequence is that futex_wait() can return zero and absorb
1244          * a wakeup when *uaddr != val on entry to the syscall.  This is
1245          * rare, but normal.
1246          *
1247          * for shared futexes, we hold the mmap semaphore, so the mapping
1248          * cannot have changed since we looked it up in get_futex_key.
1249          */
1250         ret = get_futex_value_locked(&uval, uaddr);
1251
1252         if (unlikely(ret)) {
1253                 queue_unlock(&q, hb);
1254
1255                 /*
1256                  * If we would have faulted, release mmap_sem, fault it in and
1257                  * start all over again.
1258                  */
1259                 futex_unlock_mm(fshared);
1260
1261                 ret = get_user(uval, uaddr);
1262
1263                 if (!ret)
1264                         goto retry;
1265                 return ret;
1266         }
1267         ret = -EWOULDBLOCK;
1268         if (uval != val)
1269                 goto out_unlock_release_sem;
1270
1271         /* Only actually queue if *uaddr contained val.  */
1272         queue_me(&q, hb);
1273
1274         /*
1275          * Now the futex is queued and we have checked the data, we
1276          * don't want to hold mmap_sem while we sleep.
1277          */
1278         futex_unlock_mm(fshared);
1279
1280         /*
1281          * There might have been scheduling since the queue_me(), as we
1282          * cannot hold a spinlock across the get_user() in case it
1283          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1284          * queueing ourselves into the futex hash.  This code thus has to
1285          * rely on the futex_wake() code removing us from hash when it
1286          * wakes us up.
1287          */
1288
1289         /* add_wait_queue is the barrier after __set_current_state. */
1290         __set_current_state(TASK_INTERRUPTIBLE);
1291         add_wait_queue(&q.waiters, &wait);
1292         /*
1293          * !plist_node_empty() is safe here without any lock.
1294          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1295          */
1296         if (likely(!plist_node_empty(&q.list))) {
1297                 if (!abs_time)
1298                         schedule();
1299                 else {
1300                         unsigned long slack;
1301                         slack = current->timer_slack_ns;
1302                         if (rt_task(current))
1303                                 slack = 0;
1304                         hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1305                                                 HRTIMER_MODE_ABS);
1306                         hrtimer_init_sleeper(&t, current);
1307                         hrtimer_set_expires_range_ns(&t.timer, *abs_time, slack);
1308
1309                         hrtimer_start_expires(&t.timer, HRTIMER_MODE_ABS);
1310                         if (!hrtimer_active(&t.timer))
1311                                 t.task = NULL;
1312
1313                         /*
1314                          * the timer could have already expired, in which
1315                          * case current would be flagged for rescheduling.
1316                          * Don't bother calling schedule.
1317                          */
1318                         if (likely(t.task))
1319                                 schedule();
1320
1321                         hrtimer_cancel(&t.timer);
1322
1323                         /* Flag if a timeout occured */
1324                         rem = (t.task == NULL);
1325
1326                         destroy_hrtimer_on_stack(&t.timer);
1327                 }
1328         }
1329         __set_current_state(TASK_RUNNING);
1330
1331         /*
1332          * NOTE: we don't remove ourselves from the waitqueue because
1333          * we are the only user of it.
1334          */
1335
1336         /* If we were woken (and unqueued), we succeeded, whatever. */
1337         if (!unqueue_me(&q))
1338                 return 0;
1339         if (rem)
1340                 return -ETIMEDOUT;
1341
1342         /*
1343          * We expect signal_pending(current), but another thread may
1344          * have handled it for us already.
1345          */
1346         if (!abs_time)
1347                 return -ERESTARTSYS;
1348         else {
1349                 struct restart_block *restart;
1350                 restart = &current_thread_info()->restart_block;
1351                 restart->fn = futex_wait_restart;
1352                 restart->futex.uaddr = (u32 *)uaddr;
1353                 restart->futex.val = val;
1354                 restart->futex.time = abs_time->tv64;
1355                 restart->futex.bitset = bitset;
1356                 restart->futex.flags = 0;
1357
1358                 if (fshared)
1359                         restart->futex.flags |= FLAGS_SHARED;
1360                 return -ERESTART_RESTARTBLOCK;
1361         }
1362
1363  out_unlock_release_sem:
1364         queue_unlock(&q, hb);
1365
1366  out_release_sem:
1367         futex_unlock_mm(fshared);
1368         return ret;
1369 }
1370
1371
1372 static long futex_wait_restart(struct restart_block *restart)
1373 {
1374         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1375         struct rw_semaphore *fshared = NULL;
1376         ktime_t t;
1377
1378         t.tv64 = restart->futex.time;
1379         restart->fn = do_no_restart_syscall;
1380         if (restart->futex.flags & FLAGS_SHARED)
1381                 fshared = &current->mm->mmap_sem;
1382         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1383                                 restart->futex.bitset);
1384 }
1385
1386
1387 /*
1388  * Userspace tried a 0 -> TID atomic transition of the futex value
1389  * and failed. The kernel side here does the whole locking operation:
1390  * if there are waiters then it will block, it does PI, etc. (Due to
1391  * races the kernel might see a 0 value of the futex too.)
1392  */
1393 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1394                          int detect, ktime_t *time, int trylock)
1395 {
1396         struct hrtimer_sleeper timeout, *to = NULL;
1397         struct task_struct *curr = current;
1398         struct futex_hash_bucket *hb;
1399         u32 uval, newval, curval;
1400         struct futex_q q;
1401         int ret, lock_taken, ownerdied = 0, attempt = 0;
1402
1403         if (refill_pi_state_cache())
1404                 return -ENOMEM;
1405
1406         if (time) {
1407                 to = &timeout;
1408                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1409                                       HRTIMER_MODE_ABS);
1410                 hrtimer_init_sleeper(to, current);
1411                 hrtimer_set_expires(&to->timer, *time);
1412         }
1413
1414         q.pi_state = NULL;
1415  retry:
1416         futex_lock_mm(fshared);
1417
1418         ret = get_futex_key(uaddr, fshared, &q.key);
1419         if (unlikely(ret != 0))
1420                 goto out_release_sem;
1421
1422  retry_unlocked:
1423         hb = queue_lock(&q);
1424
1425  retry_locked:
1426         ret = lock_taken = 0;
1427
1428         /*
1429          * To avoid races, we attempt to take the lock here again
1430          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1431          * the locks. It will most likely not succeed.
1432          */
1433         newval = task_pid_vnr(current);
1434
1435         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1436
1437         if (unlikely(curval == -EFAULT))
1438                 goto uaddr_faulted;
1439
1440         /*
1441          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1442          * situation and we return success to user space.
1443          */
1444         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1445                 ret = -EDEADLK;
1446                 goto out_unlock_release_sem;
1447         }
1448
1449         /*
1450          * Surprise - we got the lock. Just return to userspace:
1451          */
1452         if (unlikely(!curval))
1453                 goto out_unlock_release_sem;
1454
1455         uval = curval;
1456
1457         /*
1458          * Set the WAITERS flag, so the owner will know it has someone
1459          * to wake at next unlock
1460          */
1461         newval = curval | FUTEX_WAITERS;
1462
1463         /*
1464          * There are two cases, where a futex might have no owner (the
1465          * owner TID is 0): OWNER_DIED. We take over the futex in this
1466          * case. We also do an unconditional take over, when the owner
1467          * of the futex died.
1468          *
1469          * This is safe as we are protected by the hash bucket lock !
1470          */
1471         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1472                 /* Keep the OWNER_DIED bit */
1473                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1474                 ownerdied = 0;
1475                 lock_taken = 1;
1476         }
1477
1478         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1479
1480         if (unlikely(curval == -EFAULT))
1481                 goto uaddr_faulted;
1482         if (unlikely(curval != uval))
1483                 goto retry_locked;
1484
1485         /*
1486          * We took the lock due to owner died take over.
1487          */
1488         if (unlikely(lock_taken))
1489                 goto out_unlock_release_sem;
1490
1491         /*
1492          * We dont have the lock. Look up the PI state (or create it if
1493          * we are the first waiter):
1494          */
1495         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1496
1497         if (unlikely(ret)) {
1498                 switch (ret) {
1499
1500                 case -EAGAIN:
1501                         /*
1502                          * Task is exiting and we just wait for the
1503                          * exit to complete.
1504                          */
1505                         queue_unlock(&q, hb);
1506                         futex_unlock_mm(fshared);
1507                         cond_resched();
1508                         goto retry;
1509
1510                 case -ESRCH:
1511                         /*
1512                          * No owner found for this futex. Check if the
1513                          * OWNER_DIED bit is set to figure out whether
1514                          * this is a robust futex or not.
1515                          */
1516                         if (get_futex_value_locked(&curval, uaddr))
1517                                 goto uaddr_faulted;
1518
1519                         /*
1520                          * We simply start over in case of a robust
1521                          * futex. The code above will take the futex
1522                          * and return happy.
1523                          */
1524                         if (curval & FUTEX_OWNER_DIED) {
1525                                 ownerdied = 1;
1526                                 goto retry_locked;
1527                         }
1528                 default:
1529                         goto out_unlock_release_sem;
1530                 }
1531         }
1532
1533         /*
1534          * Only actually queue now that the atomic ops are done:
1535          */
1536         queue_me(&q, hb);
1537
1538         /*
1539          * Now the futex is queued and we have checked the data, we
1540          * don't want to hold mmap_sem while we sleep.
1541          */
1542         futex_unlock_mm(fshared);
1543
1544         WARN_ON(!q.pi_state);
1545         /*
1546          * Block on the PI mutex:
1547          */
1548         if (!trylock)
1549                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1550         else {
1551                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1552                 /* Fixup the trylock return value: */
1553                 ret = ret ? 0 : -EWOULDBLOCK;
1554         }
1555
1556         futex_lock_mm(fshared);
1557         spin_lock(q.lock_ptr);
1558
1559         if (!ret) {
1560                 /*
1561                  * Got the lock. We might not be the anticipated owner
1562                  * if we did a lock-steal - fix up the PI-state in
1563                  * that case:
1564                  */
1565                 if (q.pi_state->owner != curr)
1566                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1567         } else {
1568                 /*
1569                  * Catch the rare case, where the lock was released
1570                  * when we were on the way back before we locked the
1571                  * hash bucket.
1572                  */
1573                 if (q.pi_state->owner == curr) {
1574                         /*
1575                          * Try to get the rt_mutex now. This might
1576                          * fail as some other task acquired the
1577                          * rt_mutex after we removed ourself from the
1578                          * rt_mutex waiters list.
1579                          */
1580                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1581                                 ret = 0;
1582                         else {
1583                                 /*
1584                                  * pi_state is incorrect, some other
1585                                  * task did a lock steal and we
1586                                  * returned due to timeout or signal
1587                                  * without taking the rt_mutex. Too
1588                                  * late. We can access the
1589                                  * rt_mutex_owner without locking, as
1590                                  * the other task is now blocked on
1591                                  * the hash bucket lock. Fix the state
1592                                  * up.
1593                                  */
1594                                 struct task_struct *owner;
1595                                 int res;
1596
1597                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1598                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1599                                                            fshared);
1600
1601                                 /* propagate -EFAULT, if the fixup failed */
1602                                 if (res)
1603                                         ret = res;
1604                         }
1605                 } else {
1606                         /*
1607                          * Paranoia check. If we did not take the lock
1608                          * in the trylock above, then we should not be
1609                          * the owner of the rtmutex, neither the real
1610                          * nor the pending one:
1611                          */
1612                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1613                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1614                                        "pi-mutex: %p pi-state %p\n", ret,
1615                                        q.pi_state->pi_mutex.owner,
1616                                        q.pi_state->owner);
1617                 }
1618         }
1619
1620         /* Unqueue and drop the lock */
1621         unqueue_me_pi(&q);
1622         futex_unlock_mm(fshared);
1623
1624         if (to)
1625                 destroy_hrtimer_on_stack(&to->timer);
1626         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1627
1628  out_unlock_release_sem:
1629         queue_unlock(&q, hb);
1630
1631  out_release_sem:
1632         futex_unlock_mm(fshared);
1633         if (to)
1634                 destroy_hrtimer_on_stack(&to->timer);
1635         return ret;
1636
1637  uaddr_faulted:
1638         /*
1639          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1640          * non-atomically.  Therefore, if get_user below is not
1641          * enough, we need to handle the fault ourselves, while
1642          * still holding the mmap_sem.
1643          *
1644          * ... and hb->lock. :-) --ANK
1645          */
1646         queue_unlock(&q, hb);
1647
1648         if (attempt++) {
1649                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1650                                          attempt);
1651                 if (ret)
1652                         goto out_release_sem;
1653                 goto retry_unlocked;
1654         }
1655
1656         futex_unlock_mm(fshared);
1657
1658         ret = get_user(uval, uaddr);
1659         if (!ret && (uval != -EFAULT))
1660                 goto retry;
1661
1662         if (to)
1663                 destroy_hrtimer_on_stack(&to->timer);
1664         return ret;
1665 }
1666
1667 /*
1668  * Userspace attempted a TID -> 0 atomic transition, and failed.
1669  * This is the in-kernel slowpath: we look up the PI state (if any),
1670  * and do the rt-mutex unlock.
1671  */
1672 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1673 {
1674         struct futex_hash_bucket *hb;
1675         struct futex_q *this, *next;
1676         u32 uval;
1677         struct plist_head *head;
1678         union futex_key key;
1679         int ret, attempt = 0;
1680
1681 retry:
1682         if (get_user(uval, uaddr))
1683                 return -EFAULT;
1684         /*
1685          * We release only a lock we actually own:
1686          */
1687         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1688                 return -EPERM;
1689         /*
1690          * First take all the futex related locks:
1691          */
1692         futex_lock_mm(fshared);
1693
1694         ret = get_futex_key(uaddr, fshared, &key);
1695         if (unlikely(ret != 0))
1696                 goto out;
1697
1698         hb = hash_futex(&key);
1699 retry_unlocked:
1700         spin_lock(&hb->lock);
1701
1702         /*
1703          * To avoid races, try to do the TID -> 0 atomic transition
1704          * again. If it succeeds then we can return without waking
1705          * anyone else up:
1706          */
1707         if (!(uval & FUTEX_OWNER_DIED))
1708                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1709
1710
1711         if (unlikely(uval == -EFAULT))
1712                 goto pi_faulted;
1713         /*
1714          * Rare case: we managed to release the lock atomically,
1715          * no need to wake anyone else up:
1716          */
1717         if (unlikely(uval == task_pid_vnr(current)))
1718                 goto out_unlock;
1719
1720         /*
1721          * Ok, other tasks may need to be woken up - check waiters
1722          * and do the wakeup if necessary:
1723          */
1724         head = &hb->chain;
1725
1726         plist_for_each_entry_safe(this, next, head, list) {
1727                 if (!match_futex (&this->key, &key))
1728                         continue;
1729                 ret = wake_futex_pi(uaddr, uval, this);
1730                 /*
1731                  * The atomic access to the futex value
1732                  * generated a pagefault, so retry the
1733                  * user-access and the wakeup:
1734                  */
1735                 if (ret == -EFAULT)
1736                         goto pi_faulted;
1737                 goto out_unlock;
1738         }
1739         /*
1740          * No waiters - kernel unlocks the futex:
1741          */
1742         if (!(uval & FUTEX_OWNER_DIED)) {
1743                 ret = unlock_futex_pi(uaddr, uval);
1744                 if (ret == -EFAULT)
1745                         goto pi_faulted;
1746         }
1747
1748 out_unlock:
1749         spin_unlock(&hb->lock);
1750 out:
1751         futex_unlock_mm(fshared);
1752
1753         return ret;
1754
1755 pi_faulted:
1756         /*
1757          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1758          * non-atomically.  Therefore, if get_user below is not
1759          * enough, we need to handle the fault ourselves, while
1760          * still holding the mmap_sem.
1761          *
1762          * ... and hb->lock. --ANK
1763          */
1764         spin_unlock(&hb->lock);
1765
1766         if (attempt++) {
1767                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1768                                          attempt);
1769                 if (ret)
1770                         goto out;
1771                 uval = 0;
1772                 goto retry_unlocked;
1773         }
1774
1775         futex_unlock_mm(fshared);
1776
1777         ret = get_user(uval, uaddr);
1778         if (!ret && (uval != -EFAULT))
1779                 goto retry;
1780
1781         return ret;
1782 }
1783
1784 /*
1785  * Support for robust futexes: the kernel cleans up held futexes at
1786  * thread exit time.
1787  *
1788  * Implementation: user-space maintains a per-thread list of locks it
1789  * is holding. Upon do_exit(), the kernel carefully walks this list,
1790  * and marks all locks that are owned by this thread with the
1791  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1792  * always manipulated with the lock held, so the list is private and
1793  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1794  * field, to allow the kernel to clean up if the thread dies after
1795  * acquiring the lock, but just before it could have added itself to
1796  * the list. There can only be one such pending lock.
1797  */
1798
1799 /**
1800  * sys_set_robust_list - set the robust-futex list head of a task
1801  * @head: pointer to the list-head
1802  * @len: length of the list-head, as userspace expects
1803  */
1804 asmlinkage long
1805 sys_set_robust_list(struct robust_list_head __user *head,
1806                     size_t len)
1807 {
1808         if (!futex_cmpxchg_enabled)
1809                 return -ENOSYS;
1810         /*
1811          * The kernel knows only one size for now:
1812          */
1813         if (unlikely(len != sizeof(*head)))
1814                 return -EINVAL;
1815
1816         current->robust_list = head;
1817
1818         return 0;
1819 }
1820
1821 /**
1822  * sys_get_robust_list - get the robust-futex list head of a task
1823  * @pid: pid of the process [zero for current task]
1824  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1825  * @len_ptr: pointer to a length field, the kernel fills in the header size
1826  */
1827 asmlinkage long
1828 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1829                     size_t __user *len_ptr)
1830 {
1831         struct robust_list_head __user *head;
1832         unsigned long ret;
1833         uid_t euid = current_euid();
1834
1835         if (!futex_cmpxchg_enabled)
1836                 return -ENOSYS;
1837
1838         if (!pid)
1839                 head = current->robust_list;
1840         else {
1841                 struct task_struct *p;
1842
1843                 ret = -ESRCH;
1844                 rcu_read_lock();
1845                 p = find_task_by_vpid(pid);
1846                 if (!p)
1847                         goto err_unlock;
1848                 ret = -EPERM;
1849                 if (euid != p->euid && euid != p->uid &&
1850                     !capable(CAP_SYS_PTRACE))
1851                         goto err_unlock;
1852                 head = p->robust_list;
1853                 rcu_read_unlock();
1854         }
1855
1856         if (put_user(sizeof(*head), len_ptr))
1857                 return -EFAULT;
1858         return put_user(head, head_ptr);
1859
1860 err_unlock:
1861         rcu_read_unlock();
1862
1863         return ret;
1864 }
1865
1866 /*
1867  * Process a futex-list entry, check whether it's owned by the
1868  * dying task, and do notification if so:
1869  */
1870 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1871 {
1872         u32 uval, nval, mval;
1873
1874 retry:
1875         if (get_user(uval, uaddr))
1876                 return -1;
1877
1878         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1879                 /*
1880                  * Ok, this dying thread is truly holding a futex
1881                  * of interest. Set the OWNER_DIED bit atomically
1882                  * via cmpxchg, and if the value had FUTEX_WAITERS
1883                  * set, wake up a waiter (if any). (We have to do a
1884                  * futex_wake() even if OWNER_DIED is already set -
1885                  * to handle the rare but possible case of recursive
1886                  * thread-death.) The rest of the cleanup is done in
1887                  * userspace.
1888                  */
1889                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1890                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1891
1892                 if (nval == -EFAULT)
1893                         return -1;
1894
1895                 if (nval != uval)
1896                         goto retry;
1897
1898                 /*
1899                  * Wake robust non-PI futexes here. The wakeup of
1900                  * PI futexes happens in exit_pi_state():
1901                  */
1902                 if (!pi && (uval & FUTEX_WAITERS))
1903                         futex_wake(uaddr, &curr->mm->mmap_sem, 1,
1904                                    FUTEX_BITSET_MATCH_ANY);
1905         }
1906         return 0;
1907 }
1908
1909 /*
1910  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1911  */
1912 static inline int fetch_robust_entry(struct robust_list __user **entry,
1913                                      struct robust_list __user * __user *head,
1914                                      int *pi)
1915 {
1916         unsigned long uentry;
1917
1918         if (get_user(uentry, (unsigned long __user *)head))
1919                 return -EFAULT;
1920
1921         *entry = (void __user *)(uentry & ~1UL);
1922         *pi = uentry & 1;
1923
1924         return 0;
1925 }
1926
1927 /*
1928  * Walk curr->robust_list (very carefully, it's a userspace list!)
1929  * and mark any locks found there dead, and notify any waiters.
1930  *
1931  * We silently return on any sign of list-walking problem.
1932  */
1933 void exit_robust_list(struct task_struct *curr)
1934 {
1935         struct robust_list_head __user *head = curr->robust_list;
1936         struct robust_list __user *entry, *next_entry, *pending;
1937         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1938         unsigned long futex_offset;
1939         int rc;
1940
1941         if (!futex_cmpxchg_enabled)
1942                 return;
1943
1944         /*
1945          * Fetch the list head (which was registered earlier, via
1946          * sys_set_robust_list()):
1947          */
1948         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1949                 return;
1950         /*
1951          * Fetch the relative futex offset:
1952          */
1953         if (get_user(futex_offset, &head->futex_offset))
1954                 return;
1955         /*
1956          * Fetch any possibly pending lock-add first, and handle it
1957          * if it exists:
1958          */
1959         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1960                 return;
1961
1962         next_entry = NULL;      /* avoid warning with gcc */
1963         while (entry != &head->list) {
1964                 /*
1965                  * Fetch the next entry in the list before calling
1966                  * handle_futex_death:
1967                  */
1968                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1969                 /*
1970                  * A pending lock might already be on the list, so
1971                  * don't process it twice:
1972                  */
1973                 if (entry != pending)
1974                         if (handle_futex_death((void __user *)entry + futex_offset,
1975                                                 curr, pi))
1976                                 return;
1977                 if (rc)
1978                         return;
1979                 entry = next_entry;
1980                 pi = next_pi;
1981                 /*
1982                  * Avoid excessively long or circular lists:
1983                  */
1984                 if (!--limit)
1985                         break;
1986
1987                 cond_resched();
1988         }
1989
1990         if (pending)
1991                 handle_futex_death((void __user *)pending + futex_offset,
1992                                    curr, pip);
1993 }
1994
1995 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1996                 u32 __user *uaddr2, u32 val2, u32 val3)
1997 {
1998         int ret = -ENOSYS;
1999         int cmd = op & FUTEX_CMD_MASK;
2000         struct rw_semaphore *fshared = NULL;
2001
2002         if (!(op & FUTEX_PRIVATE_FLAG))
2003                 fshared = &current->mm->mmap_sem;
2004
2005         switch (cmd) {
2006         case FUTEX_WAIT:
2007                 val3 = FUTEX_BITSET_MATCH_ANY;
2008         case FUTEX_WAIT_BITSET:
2009                 ret = futex_wait(uaddr, fshared, val, timeout, val3);
2010                 break;
2011         case FUTEX_WAKE:
2012                 val3 = FUTEX_BITSET_MATCH_ANY;
2013         case FUTEX_WAKE_BITSET:
2014                 ret = futex_wake(uaddr, fshared, val, val3);
2015                 break;
2016         case FUTEX_REQUEUE:
2017                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2018                 break;
2019         case FUTEX_CMP_REQUEUE:
2020                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2021                 break;
2022         case FUTEX_WAKE_OP:
2023                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2024                 break;
2025         case FUTEX_LOCK_PI:
2026                 if (futex_cmpxchg_enabled)
2027                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2028                 break;
2029         case FUTEX_UNLOCK_PI:
2030                 if (futex_cmpxchg_enabled)
2031                         ret = futex_unlock_pi(uaddr, fshared);
2032                 break;
2033         case FUTEX_TRYLOCK_PI:
2034                 if (futex_cmpxchg_enabled)
2035                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2036                 break;
2037         default:
2038                 ret = -ENOSYS;
2039         }
2040         return ret;
2041 }
2042
2043
2044 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2045                           struct timespec __user *utime, u32 __user *uaddr2,
2046                           u32 val3)
2047 {
2048         struct timespec ts;
2049         ktime_t t, *tp = NULL;
2050         u32 val2 = 0;
2051         int cmd = op & FUTEX_CMD_MASK;
2052
2053         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2054                       cmd == FUTEX_WAIT_BITSET)) {
2055                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2056                         return -EFAULT;
2057                 if (!timespec_valid(&ts))
2058                         return -EINVAL;
2059
2060                 t = timespec_to_ktime(ts);
2061                 if (cmd == FUTEX_WAIT)
2062                         t = ktime_add_safe(ktime_get(), t);
2063                 tp = &t;
2064         }
2065         /*
2066          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2067          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2068          */
2069         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2070             cmd == FUTEX_WAKE_OP)
2071                 val2 = (u32) (unsigned long) utime;
2072
2073         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2074 }
2075
2076 static int __init futex_init(void)
2077 {
2078         u32 curval;
2079         int i;
2080
2081         /*
2082          * This will fail and we want it. Some arch implementations do
2083          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2084          * functionality. We want to know that before we call in any
2085          * of the complex code paths. Also we want to prevent
2086          * registration of robust lists in that case. NULL is
2087          * guaranteed to fault and we get -EFAULT on functional
2088          * implementation, the non functional ones will return
2089          * -ENOSYS.
2090          */
2091         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2092         if (curval == -EFAULT)
2093                 futex_cmpxchg_enabled = 1;
2094
2095         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2096                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2097                 spin_lock_init(&futex_queues[i].lock);
2098         }
2099
2100         return 0;
2101 }
2102 __initcall(futex_init);