CRED: Use RCU to access another task's creds and to release a task's own creds
[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         const struct cred *cred = current_cred(), *pcred;
443
444         rcu_read_lock();
445         p = find_task_by_vpid(pid);
446         if (!p) {
447                 p = ERR_PTR(-ESRCH);
448         } else {
449                 pcred = __task_cred(p);
450                 if (cred->euid != pcred->euid &&
451                     cred->euid != pcred->uid)
452                         p = ERR_PTR(-ESRCH);
453                 else
454                         get_task_struct(p);
455         }
456
457         rcu_read_unlock();
458
459         return p;
460 }
461
462 /*
463  * This task is holding PI mutexes at exit time => bad.
464  * Kernel cleans up PI-state, but userspace is likely hosed.
465  * (Robust-futex cleanup is separate and might save the day for userspace.)
466  */
467 void exit_pi_state_list(struct task_struct *curr)
468 {
469         struct list_head *next, *head = &curr->pi_state_list;
470         struct futex_pi_state *pi_state;
471         struct futex_hash_bucket *hb;
472         union futex_key key;
473
474         if (!futex_cmpxchg_enabled)
475                 return;
476         /*
477          * We are a ZOMBIE and nobody can enqueue itself on
478          * pi_state_list anymore, but we have to be careful
479          * versus waiters unqueueing themselves:
480          */
481         spin_lock_irq(&curr->pi_lock);
482         while (!list_empty(head)) {
483
484                 next = head->next;
485                 pi_state = list_entry(next, struct futex_pi_state, list);
486                 key = pi_state->key;
487                 hb = hash_futex(&key);
488                 spin_unlock_irq(&curr->pi_lock);
489
490                 spin_lock(&hb->lock);
491
492                 spin_lock_irq(&curr->pi_lock);
493                 /*
494                  * We dropped the pi-lock, so re-check whether this
495                  * task still owns the PI-state:
496                  */
497                 if (head->next != next) {
498                         spin_unlock(&hb->lock);
499                         continue;
500                 }
501
502                 WARN_ON(pi_state->owner != curr);
503                 WARN_ON(list_empty(&pi_state->list));
504                 list_del_init(&pi_state->list);
505                 pi_state->owner = NULL;
506                 spin_unlock_irq(&curr->pi_lock);
507
508                 rt_mutex_unlock(&pi_state->pi_mutex);
509
510                 spin_unlock(&hb->lock);
511
512                 spin_lock_irq(&curr->pi_lock);
513         }
514         spin_unlock_irq(&curr->pi_lock);
515 }
516
517 static int
518 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
519                 union futex_key *key, struct futex_pi_state **ps)
520 {
521         struct futex_pi_state *pi_state = NULL;
522         struct futex_q *this, *next;
523         struct plist_head *head;
524         struct task_struct *p;
525         pid_t pid = uval & FUTEX_TID_MASK;
526
527         head = &hb->chain;
528
529         plist_for_each_entry_safe(this, next, head, list) {
530                 if (match_futex(&this->key, key)) {
531                         /*
532                          * Another waiter already exists - bump up
533                          * the refcount and return its pi_state:
534                          */
535                         pi_state = this->pi_state;
536                         /*
537                          * Userspace might have messed up non PI and PI futexes
538                          */
539                         if (unlikely(!pi_state))
540                                 return -EINVAL;
541
542                         WARN_ON(!atomic_read(&pi_state->refcount));
543                         WARN_ON(pid && pi_state->owner &&
544                                 pi_state->owner->pid != pid);
545
546                         atomic_inc(&pi_state->refcount);
547                         *ps = pi_state;
548
549                         return 0;
550                 }
551         }
552
553         /*
554          * We are the first waiter - try to look up the real owner and attach
555          * the new pi_state to it, but bail out when TID = 0
556          */
557         if (!pid)
558                 return -ESRCH;
559         p = futex_find_get_task(pid);
560         if (IS_ERR(p))
561                 return PTR_ERR(p);
562
563         /*
564          * We need to look at the task state flags to figure out,
565          * whether the task is exiting. To protect against the do_exit
566          * change of the task flags, we do this protected by
567          * p->pi_lock:
568          */
569         spin_lock_irq(&p->pi_lock);
570         if (unlikely(p->flags & PF_EXITING)) {
571                 /*
572                  * The task is on the way out. When PF_EXITPIDONE is
573                  * set, we know that the task has finished the
574                  * cleanup:
575                  */
576                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
577
578                 spin_unlock_irq(&p->pi_lock);
579                 put_task_struct(p);
580                 return ret;
581         }
582
583         pi_state = alloc_pi_state();
584
585         /*
586          * Initialize the pi_mutex in locked state and make 'p'
587          * the owner of it:
588          */
589         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
590
591         /* Store the key for possible exit cleanups: */
592         pi_state->key = *key;
593
594         WARN_ON(!list_empty(&pi_state->list));
595         list_add(&pi_state->list, &p->pi_state_list);
596         pi_state->owner = p;
597         spin_unlock_irq(&p->pi_lock);
598
599         put_task_struct(p);
600
601         *ps = pi_state;
602
603         return 0;
604 }
605
606 /*
607  * The hash bucket lock must be held when this is called.
608  * Afterwards, the futex_q must not be accessed.
609  */
610 static void wake_futex(struct futex_q *q)
611 {
612         plist_del(&q->list, &q->list.plist);
613         /*
614          * The lock in wake_up_all() is a crucial memory barrier after the
615          * plist_del() and also before assigning to q->lock_ptr.
616          */
617         wake_up_all(&q->waiters);
618         /*
619          * The waiting task can free the futex_q as soon as this is written,
620          * without taking any locks.  This must come last.
621          *
622          * A memory barrier is required here to prevent the following store
623          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
624          * at the end of wake_up_all() does not prevent this store from
625          * moving.
626          */
627         smp_wmb();
628         q->lock_ptr = NULL;
629 }
630
631 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
632 {
633         struct task_struct *new_owner;
634         struct futex_pi_state *pi_state = this->pi_state;
635         u32 curval, newval;
636
637         if (!pi_state)
638                 return -EINVAL;
639
640         spin_lock(&pi_state->pi_mutex.wait_lock);
641         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
642
643         /*
644          * This happens when we have stolen the lock and the original
645          * pending owner did not enqueue itself back on the rt_mutex.
646          * Thats not a tragedy. We know that way, that a lock waiter
647          * is on the fly. We make the futex_q waiter the pending owner.
648          */
649         if (!new_owner)
650                 new_owner = this->task;
651
652         /*
653          * We pass it to the next owner. (The WAITERS bit is always
654          * kept enabled while there is PI state around. We must also
655          * preserve the owner died bit.)
656          */
657         if (!(uval & FUTEX_OWNER_DIED)) {
658                 int ret = 0;
659
660                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
661
662                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
663
664                 if (curval == -EFAULT)
665                         ret = -EFAULT;
666                 else if (curval != uval)
667                         ret = -EINVAL;
668                 if (ret) {
669                         spin_unlock(&pi_state->pi_mutex.wait_lock);
670                         return ret;
671                 }
672         }
673
674         spin_lock_irq(&pi_state->owner->pi_lock);
675         WARN_ON(list_empty(&pi_state->list));
676         list_del_init(&pi_state->list);
677         spin_unlock_irq(&pi_state->owner->pi_lock);
678
679         spin_lock_irq(&new_owner->pi_lock);
680         WARN_ON(!list_empty(&pi_state->list));
681         list_add(&pi_state->list, &new_owner->pi_state_list);
682         pi_state->owner = new_owner;
683         spin_unlock_irq(&new_owner->pi_lock);
684
685         spin_unlock(&pi_state->pi_mutex.wait_lock);
686         rt_mutex_unlock(&pi_state->pi_mutex);
687
688         return 0;
689 }
690
691 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
692 {
693         u32 oldval;
694
695         /*
696          * There is no waiter, so we unlock the futex. The owner died
697          * bit has not to be preserved here. We are the owner:
698          */
699         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
700
701         if (oldval == -EFAULT)
702                 return oldval;
703         if (oldval != uval)
704                 return -EAGAIN;
705
706         return 0;
707 }
708
709 /*
710  * Express the locking dependencies for lockdep:
711  */
712 static inline void
713 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
714 {
715         if (hb1 <= hb2) {
716                 spin_lock(&hb1->lock);
717                 if (hb1 < hb2)
718                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
719         } else { /* hb1 > hb2 */
720                 spin_lock(&hb2->lock);
721                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
722         }
723 }
724
725 /*
726  * Wake up all waiters hashed on the physical page that is mapped
727  * to this virtual address:
728  */
729 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
730                       int nr_wake, u32 bitset)
731 {
732         struct futex_hash_bucket *hb;
733         struct futex_q *this, *next;
734         struct plist_head *head;
735         union futex_key key;
736         int ret;
737
738         if (!bitset)
739                 return -EINVAL;
740
741         futex_lock_mm(fshared);
742
743         ret = get_futex_key(uaddr, fshared, &key);
744         if (unlikely(ret != 0))
745                 goto out;
746
747         hb = hash_futex(&key);
748         spin_lock(&hb->lock);
749         head = &hb->chain;
750
751         plist_for_each_entry_safe(this, next, head, list) {
752                 if (match_futex (&this->key, &key)) {
753                         if (this->pi_state) {
754                                 ret = -EINVAL;
755                                 break;
756                         }
757
758                         /* Check if one of the bits is set in both bitsets */
759                         if (!(this->bitset & bitset))
760                                 continue;
761
762                         wake_futex(this);
763                         if (++ret >= nr_wake)
764                                 break;
765                 }
766         }
767
768         spin_unlock(&hb->lock);
769 out:
770         futex_unlock_mm(fshared);
771         return ret;
772 }
773
774 /*
775  * Wake up all waiters hashed on the physical page that is mapped
776  * to this virtual address:
777  */
778 static int
779 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
780               u32 __user *uaddr2,
781               int nr_wake, int nr_wake2, int op)
782 {
783         union futex_key key1, key2;
784         struct futex_hash_bucket *hb1, *hb2;
785         struct plist_head *head;
786         struct futex_q *this, *next;
787         int ret, op_ret, attempt = 0;
788
789 retryfull:
790         futex_lock_mm(fshared);
791
792         ret = get_futex_key(uaddr1, fshared, &key1);
793         if (unlikely(ret != 0))
794                 goto out;
795         ret = get_futex_key(uaddr2, fshared, &key2);
796         if (unlikely(ret != 0))
797                 goto out;
798
799         hb1 = hash_futex(&key1);
800         hb2 = hash_futex(&key2);
801
802 retry:
803         double_lock_hb(hb1, hb2);
804
805         op_ret = futex_atomic_op_inuser(op, uaddr2);
806         if (unlikely(op_ret < 0)) {
807                 u32 dummy;
808
809                 spin_unlock(&hb1->lock);
810                 if (hb1 != hb2)
811                         spin_unlock(&hb2->lock);
812
813 #ifndef CONFIG_MMU
814                 /*
815                  * we don't get EFAULT from MMU faults if we don't have an MMU,
816                  * but we might get them from range checking
817                  */
818                 ret = op_ret;
819                 goto out;
820 #endif
821
822                 if (unlikely(op_ret != -EFAULT)) {
823                         ret = op_ret;
824                         goto out;
825                 }
826
827                 /*
828                  * futex_atomic_op_inuser needs to both read and write
829                  * *(int __user *)uaddr2, but we can't modify it
830                  * non-atomically.  Therefore, if get_user below is not
831                  * enough, we need to handle the fault ourselves, while
832                  * still holding the mmap_sem.
833                  */
834                 if (attempt++) {
835                         ret = futex_handle_fault((unsigned long)uaddr2,
836                                                  fshared, attempt);
837                         if (ret)
838                                 goto out;
839                         goto retry;
840                 }
841
842                 /*
843                  * If we would have faulted, release mmap_sem,
844                  * fault it in and start all over again.
845                  */
846                 futex_unlock_mm(fshared);
847
848                 ret = get_user(dummy, uaddr2);
849                 if (ret)
850                         return ret;
851
852                 goto retryfull;
853         }
854
855         head = &hb1->chain;
856
857         plist_for_each_entry_safe(this, next, head, list) {
858                 if (match_futex (&this->key, &key1)) {
859                         wake_futex(this);
860                         if (++ret >= nr_wake)
861                                 break;
862                 }
863         }
864
865         if (op_ret > 0) {
866                 head = &hb2->chain;
867
868                 op_ret = 0;
869                 plist_for_each_entry_safe(this, next, head, list) {
870                         if (match_futex (&this->key, &key2)) {
871                                 wake_futex(this);
872                                 if (++op_ret >= nr_wake2)
873                                         break;
874                         }
875                 }
876                 ret += op_ret;
877         }
878
879         spin_unlock(&hb1->lock);
880         if (hb1 != hb2)
881                 spin_unlock(&hb2->lock);
882 out:
883         futex_unlock_mm(fshared);
884
885         return ret;
886 }
887
888 /*
889  * Requeue all waiters hashed on one physical page to another
890  * physical page.
891  */
892 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
893                          u32 __user *uaddr2,
894                          int nr_wake, int nr_requeue, u32 *cmpval)
895 {
896         union futex_key key1, key2;
897         struct futex_hash_bucket *hb1, *hb2;
898         struct plist_head *head1;
899         struct futex_q *this, *next;
900         int ret, drop_count = 0;
901
902  retry:
903         futex_lock_mm(fshared);
904
905         ret = get_futex_key(uaddr1, fshared, &key1);
906         if (unlikely(ret != 0))
907                 goto out;
908         ret = get_futex_key(uaddr2, fshared, &key2);
909         if (unlikely(ret != 0))
910                 goto out;
911
912         hb1 = hash_futex(&key1);
913         hb2 = hash_futex(&key2);
914
915         double_lock_hb(hb1, hb2);
916
917         if (likely(cmpval != NULL)) {
918                 u32 curval;
919
920                 ret = get_futex_value_locked(&curval, uaddr1);
921
922                 if (unlikely(ret)) {
923                         spin_unlock(&hb1->lock);
924                         if (hb1 != hb2)
925                                 spin_unlock(&hb2->lock);
926
927                         /*
928                          * If we would have faulted, release mmap_sem, fault
929                          * it in and start all over again.
930                          */
931                         futex_unlock_mm(fshared);
932
933                         ret = get_user(curval, uaddr1);
934
935                         if (!ret)
936                                 goto retry;
937
938                         return ret;
939                 }
940                 if (curval != *cmpval) {
941                         ret = -EAGAIN;
942                         goto out_unlock;
943                 }
944         }
945
946         head1 = &hb1->chain;
947         plist_for_each_entry_safe(this, next, head1, list) {
948                 if (!match_futex (&this->key, &key1))
949                         continue;
950                 if (++ret <= nr_wake) {
951                         wake_futex(this);
952                 } else {
953                         /*
954                          * If key1 and key2 hash to the same bucket, no need to
955                          * requeue.
956                          */
957                         if (likely(head1 != &hb2->chain)) {
958                                 plist_del(&this->list, &hb1->chain);
959                                 plist_add(&this->list, &hb2->chain);
960                                 this->lock_ptr = &hb2->lock;
961 #ifdef CONFIG_DEBUG_PI_LIST
962                                 this->list.plist.lock = &hb2->lock;
963 #endif
964                         }
965                         this->key = key2;
966                         get_futex_key_refs(&key2);
967                         drop_count++;
968
969                         if (ret - nr_wake >= nr_requeue)
970                                 break;
971                 }
972         }
973
974 out_unlock:
975         spin_unlock(&hb1->lock);
976         if (hb1 != hb2)
977                 spin_unlock(&hb2->lock);
978
979         /* drop_futex_key_refs() must be called outside the spinlocks. */
980         while (--drop_count >= 0)
981                 drop_futex_key_refs(&key1);
982
983 out:
984         futex_unlock_mm(fshared);
985         return ret;
986 }
987
988 /* The key must be already stored in q->key. */
989 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
990 {
991         struct futex_hash_bucket *hb;
992
993         init_waitqueue_head(&q->waiters);
994
995         get_futex_key_refs(&q->key);
996         hb = hash_futex(&q->key);
997         q->lock_ptr = &hb->lock;
998
999         spin_lock(&hb->lock);
1000         return hb;
1001 }
1002
1003 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1004 {
1005         int prio;
1006
1007         /*
1008          * The priority used to register this element is
1009          * - either the real thread-priority for the real-time threads
1010          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1011          * - or MAX_RT_PRIO for non-RT threads.
1012          * Thus, all RT-threads are woken first in priority order, and
1013          * the others are woken last, in FIFO order.
1014          */
1015         prio = min(current->normal_prio, MAX_RT_PRIO);
1016
1017         plist_node_init(&q->list, prio);
1018 #ifdef CONFIG_DEBUG_PI_LIST
1019         q->list.plist.lock = &hb->lock;
1020 #endif
1021         plist_add(&q->list, &hb->chain);
1022         q->task = current;
1023         spin_unlock(&hb->lock);
1024 }
1025
1026 static inline void
1027 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1028 {
1029         spin_unlock(&hb->lock);
1030         drop_futex_key_refs(&q->key);
1031 }
1032
1033 /*
1034  * queue_me and unqueue_me must be called as a pair, each
1035  * exactly once.  They are called with the hashed spinlock held.
1036  */
1037
1038 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1039 static int unqueue_me(struct futex_q *q)
1040 {
1041         spinlock_t *lock_ptr;
1042         int ret = 0;
1043
1044         /* In the common case we don't take the spinlock, which is nice. */
1045  retry:
1046         lock_ptr = q->lock_ptr;
1047         barrier();
1048         if (lock_ptr != NULL) {
1049                 spin_lock(lock_ptr);
1050                 /*
1051                  * q->lock_ptr can change between reading it and
1052                  * spin_lock(), causing us to take the wrong lock.  This
1053                  * corrects the race condition.
1054                  *
1055                  * Reasoning goes like this: if we have the wrong lock,
1056                  * q->lock_ptr must have changed (maybe several times)
1057                  * between reading it and the spin_lock().  It can
1058                  * change again after the spin_lock() but only if it was
1059                  * already changed before the spin_lock().  It cannot,
1060                  * however, change back to the original value.  Therefore
1061                  * we can detect whether we acquired the correct lock.
1062                  */
1063                 if (unlikely(lock_ptr != q->lock_ptr)) {
1064                         spin_unlock(lock_ptr);
1065                         goto retry;
1066                 }
1067                 WARN_ON(plist_node_empty(&q->list));
1068                 plist_del(&q->list, &q->list.plist);
1069
1070                 BUG_ON(q->pi_state);
1071
1072                 spin_unlock(lock_ptr);
1073                 ret = 1;
1074         }
1075
1076         drop_futex_key_refs(&q->key);
1077         return ret;
1078 }
1079
1080 /*
1081  * PI futexes can not be requeued and must remove themself from the
1082  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1083  * and dropped here.
1084  */
1085 static void unqueue_me_pi(struct futex_q *q)
1086 {
1087         WARN_ON(plist_node_empty(&q->list));
1088         plist_del(&q->list, &q->list.plist);
1089
1090         BUG_ON(!q->pi_state);
1091         free_pi_state(q->pi_state);
1092         q->pi_state = NULL;
1093
1094         spin_unlock(q->lock_ptr);
1095
1096         drop_futex_key_refs(&q->key);
1097 }
1098
1099 /*
1100  * Fixup the pi_state owner with the new owner.
1101  *
1102  * Must be called with hash bucket lock held and mm->sem held for non
1103  * private futexes.
1104  */
1105 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1106                                 struct task_struct *newowner,
1107                                 struct rw_semaphore *fshared)
1108 {
1109         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1110         struct futex_pi_state *pi_state = q->pi_state;
1111         struct task_struct *oldowner = pi_state->owner;
1112         u32 uval, curval, newval;
1113         int ret, attempt = 0;
1114
1115         /* Owner died? */
1116         if (!pi_state->owner)
1117                 newtid |= FUTEX_OWNER_DIED;
1118
1119         /*
1120          * We are here either because we stole the rtmutex from the
1121          * pending owner or we are the pending owner which failed to
1122          * get the rtmutex. We have to replace the pending owner TID
1123          * in the user space variable. This must be atomic as we have
1124          * to preserve the owner died bit here.
1125          *
1126          * Note: We write the user space value _before_ changing the
1127          * pi_state because we can fault here. Imagine swapped out
1128          * pages or a fork, which was running right before we acquired
1129          * mmap_sem, that marked all the anonymous memory readonly for
1130          * cow.
1131          *
1132          * Modifying pi_state _before_ the user space value would
1133          * leave the pi_state in an inconsistent state when we fault
1134          * here, because we need to drop the hash bucket lock to
1135          * handle the fault. This might be observed in the PID check
1136          * in lookup_pi_state.
1137          */
1138 retry:
1139         if (get_futex_value_locked(&uval, uaddr))
1140                 goto handle_fault;
1141
1142         while (1) {
1143                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1144
1145                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1146
1147                 if (curval == -EFAULT)
1148                         goto handle_fault;
1149                 if (curval == uval)
1150                         break;
1151                 uval = curval;
1152         }
1153
1154         /*
1155          * We fixed up user space. Now we need to fix the pi_state
1156          * itself.
1157          */
1158         if (pi_state->owner != NULL) {
1159                 spin_lock_irq(&pi_state->owner->pi_lock);
1160                 WARN_ON(list_empty(&pi_state->list));
1161                 list_del_init(&pi_state->list);
1162                 spin_unlock_irq(&pi_state->owner->pi_lock);
1163         }
1164
1165         pi_state->owner = newowner;
1166
1167         spin_lock_irq(&newowner->pi_lock);
1168         WARN_ON(!list_empty(&pi_state->list));
1169         list_add(&pi_state->list, &newowner->pi_state_list);
1170         spin_unlock_irq(&newowner->pi_lock);
1171         return 0;
1172
1173         /*
1174          * To handle the page fault we need to drop the hash bucket
1175          * lock here. That gives the other task (either the pending
1176          * owner itself or the task which stole the rtmutex) the
1177          * chance to try the fixup of the pi_state. So once we are
1178          * back from handling the fault we need to check the pi_state
1179          * after reacquiring the hash bucket lock and before trying to
1180          * do another fixup. When the fixup has been done already we
1181          * simply return.
1182          */
1183 handle_fault:
1184         spin_unlock(q->lock_ptr);
1185
1186         ret = futex_handle_fault((unsigned long)uaddr, fshared, attempt++);
1187
1188         spin_lock(q->lock_ptr);
1189
1190         /*
1191          * Check if someone else fixed it for us:
1192          */
1193         if (pi_state->owner != oldowner)
1194                 return 0;
1195
1196         if (ret)
1197                 return ret;
1198
1199         goto retry;
1200 }
1201
1202 /*
1203  * In case we must use restart_block to restart a futex_wait,
1204  * we encode in the 'flags' shared capability
1205  */
1206 #define FLAGS_SHARED  1
1207
1208 static long futex_wait_restart(struct restart_block *restart);
1209
1210 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1211                       u32 val, ktime_t *abs_time, u32 bitset)
1212 {
1213         struct task_struct *curr = current;
1214         DECLARE_WAITQUEUE(wait, curr);
1215         struct futex_hash_bucket *hb;
1216         struct futex_q q;
1217         u32 uval;
1218         int ret;
1219         struct hrtimer_sleeper t;
1220         int rem = 0;
1221
1222         if (!bitset)
1223                 return -EINVAL;
1224
1225         q.pi_state = NULL;
1226         q.bitset = bitset;
1227  retry:
1228         futex_lock_mm(fshared);
1229
1230         ret = get_futex_key(uaddr, fshared, &q.key);
1231         if (unlikely(ret != 0))
1232                 goto out_release_sem;
1233
1234         hb = queue_lock(&q);
1235
1236         /*
1237          * Access the page AFTER the futex is queued.
1238          * Order is important:
1239          *
1240          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1241          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1242          *
1243          * The basic logical guarantee of a futex is that it blocks ONLY
1244          * if cond(var) is known to be true at the time of blocking, for
1245          * any cond.  If we queued after testing *uaddr, that would open
1246          * a race condition where we could block indefinitely with
1247          * cond(var) false, which would violate the guarantee.
1248          *
1249          * A consequence is that futex_wait() can return zero and absorb
1250          * a wakeup when *uaddr != val on entry to the syscall.  This is
1251          * rare, but normal.
1252          *
1253          * for shared futexes, we hold the mmap semaphore, so the mapping
1254          * cannot have changed since we looked it up in get_futex_key.
1255          */
1256         ret = get_futex_value_locked(&uval, uaddr);
1257
1258         if (unlikely(ret)) {
1259                 queue_unlock(&q, hb);
1260
1261                 /*
1262                  * If we would have faulted, release mmap_sem, fault it in and
1263                  * start all over again.
1264                  */
1265                 futex_unlock_mm(fshared);
1266
1267                 ret = get_user(uval, uaddr);
1268
1269                 if (!ret)
1270                         goto retry;
1271                 return ret;
1272         }
1273         ret = -EWOULDBLOCK;
1274         if (uval != val)
1275                 goto out_unlock_release_sem;
1276
1277         /* Only actually queue if *uaddr contained val.  */
1278         queue_me(&q, hb);
1279
1280         /*
1281          * Now the futex is queued and we have checked the data, we
1282          * don't want to hold mmap_sem while we sleep.
1283          */
1284         futex_unlock_mm(fshared);
1285
1286         /*
1287          * There might have been scheduling since the queue_me(), as we
1288          * cannot hold a spinlock across the get_user() in case it
1289          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1290          * queueing ourselves into the futex hash.  This code thus has to
1291          * rely on the futex_wake() code removing us from hash when it
1292          * wakes us up.
1293          */
1294
1295         /* add_wait_queue is the barrier after __set_current_state. */
1296         __set_current_state(TASK_INTERRUPTIBLE);
1297         add_wait_queue(&q.waiters, &wait);
1298         /*
1299          * !plist_node_empty() is safe here without any lock.
1300          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1301          */
1302         if (likely(!plist_node_empty(&q.list))) {
1303                 if (!abs_time)
1304                         schedule();
1305                 else {
1306                         unsigned long slack;
1307                         slack = current->timer_slack_ns;
1308                         if (rt_task(current))
1309                                 slack = 0;
1310                         hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1311                                                 HRTIMER_MODE_ABS);
1312                         hrtimer_init_sleeper(&t, current);
1313                         hrtimer_set_expires_range_ns(&t.timer, *abs_time, slack);
1314
1315                         hrtimer_start_expires(&t.timer, HRTIMER_MODE_ABS);
1316                         if (!hrtimer_active(&t.timer))
1317                                 t.task = NULL;
1318
1319                         /*
1320                          * the timer could have already expired, in which
1321                          * case current would be flagged for rescheduling.
1322                          * Don't bother calling schedule.
1323                          */
1324                         if (likely(t.task))
1325                                 schedule();
1326
1327                         hrtimer_cancel(&t.timer);
1328
1329                         /* Flag if a timeout occured */
1330                         rem = (t.task == NULL);
1331
1332                         destroy_hrtimer_on_stack(&t.timer);
1333                 }
1334         }
1335         __set_current_state(TASK_RUNNING);
1336
1337         /*
1338          * NOTE: we don't remove ourselves from the waitqueue because
1339          * we are the only user of it.
1340          */
1341
1342         /* If we were woken (and unqueued), we succeeded, whatever. */
1343         if (!unqueue_me(&q))
1344                 return 0;
1345         if (rem)
1346                 return -ETIMEDOUT;
1347
1348         /*
1349          * We expect signal_pending(current), but another thread may
1350          * have handled it for us already.
1351          */
1352         if (!abs_time)
1353                 return -ERESTARTSYS;
1354         else {
1355                 struct restart_block *restart;
1356                 restart = &current_thread_info()->restart_block;
1357                 restart->fn = futex_wait_restart;
1358                 restart->futex.uaddr = (u32 *)uaddr;
1359                 restart->futex.val = val;
1360                 restart->futex.time = abs_time->tv64;
1361                 restart->futex.bitset = bitset;
1362                 restart->futex.flags = 0;
1363
1364                 if (fshared)
1365                         restart->futex.flags |= FLAGS_SHARED;
1366                 return -ERESTART_RESTARTBLOCK;
1367         }
1368
1369  out_unlock_release_sem:
1370         queue_unlock(&q, hb);
1371
1372  out_release_sem:
1373         futex_unlock_mm(fshared);
1374         return ret;
1375 }
1376
1377
1378 static long futex_wait_restart(struct restart_block *restart)
1379 {
1380         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1381         struct rw_semaphore *fshared = NULL;
1382         ktime_t t;
1383
1384         t.tv64 = restart->futex.time;
1385         restart->fn = do_no_restart_syscall;
1386         if (restart->futex.flags & FLAGS_SHARED)
1387                 fshared = &current->mm->mmap_sem;
1388         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1389                                 restart->futex.bitset);
1390 }
1391
1392
1393 /*
1394  * Userspace tried a 0 -> TID atomic transition of the futex value
1395  * and failed. The kernel side here does the whole locking operation:
1396  * if there are waiters then it will block, it does PI, etc. (Due to
1397  * races the kernel might see a 0 value of the futex too.)
1398  */
1399 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1400                          int detect, ktime_t *time, int trylock)
1401 {
1402         struct hrtimer_sleeper timeout, *to = NULL;
1403         struct task_struct *curr = current;
1404         struct futex_hash_bucket *hb;
1405         u32 uval, newval, curval;
1406         struct futex_q q;
1407         int ret, lock_taken, ownerdied = 0, attempt = 0;
1408
1409         if (refill_pi_state_cache())
1410                 return -ENOMEM;
1411
1412         if (time) {
1413                 to = &timeout;
1414                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1415                                       HRTIMER_MODE_ABS);
1416                 hrtimer_init_sleeper(to, current);
1417                 hrtimer_set_expires(&to->timer, *time);
1418         }
1419
1420         q.pi_state = NULL;
1421  retry:
1422         futex_lock_mm(fshared);
1423
1424         ret = get_futex_key(uaddr, fshared, &q.key);
1425         if (unlikely(ret != 0))
1426                 goto out_release_sem;
1427
1428  retry_unlocked:
1429         hb = queue_lock(&q);
1430
1431  retry_locked:
1432         ret = lock_taken = 0;
1433
1434         /*
1435          * To avoid races, we attempt to take the lock here again
1436          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1437          * the locks. It will most likely not succeed.
1438          */
1439         newval = task_pid_vnr(current);
1440
1441         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1442
1443         if (unlikely(curval == -EFAULT))
1444                 goto uaddr_faulted;
1445
1446         /*
1447          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1448          * situation and we return success to user space.
1449          */
1450         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1451                 ret = -EDEADLK;
1452                 goto out_unlock_release_sem;
1453         }
1454
1455         /*
1456          * Surprise - we got the lock. Just return to userspace:
1457          */
1458         if (unlikely(!curval))
1459                 goto out_unlock_release_sem;
1460
1461         uval = curval;
1462
1463         /*
1464          * Set the WAITERS flag, so the owner will know it has someone
1465          * to wake at next unlock
1466          */
1467         newval = curval | FUTEX_WAITERS;
1468
1469         /*
1470          * There are two cases, where a futex might have no owner (the
1471          * owner TID is 0): OWNER_DIED. We take over the futex in this
1472          * case. We also do an unconditional take over, when the owner
1473          * of the futex died.
1474          *
1475          * This is safe as we are protected by the hash bucket lock !
1476          */
1477         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1478                 /* Keep the OWNER_DIED bit */
1479                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1480                 ownerdied = 0;
1481                 lock_taken = 1;
1482         }
1483
1484         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1485
1486         if (unlikely(curval == -EFAULT))
1487                 goto uaddr_faulted;
1488         if (unlikely(curval != uval))
1489                 goto retry_locked;
1490
1491         /*
1492          * We took the lock due to owner died take over.
1493          */
1494         if (unlikely(lock_taken))
1495                 goto out_unlock_release_sem;
1496
1497         /*
1498          * We dont have the lock. Look up the PI state (or create it if
1499          * we are the first waiter):
1500          */
1501         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1502
1503         if (unlikely(ret)) {
1504                 switch (ret) {
1505
1506                 case -EAGAIN:
1507                         /*
1508                          * Task is exiting and we just wait for the
1509                          * exit to complete.
1510                          */
1511                         queue_unlock(&q, hb);
1512                         futex_unlock_mm(fshared);
1513                         cond_resched();
1514                         goto retry;
1515
1516                 case -ESRCH:
1517                         /*
1518                          * No owner found for this futex. Check if the
1519                          * OWNER_DIED bit is set to figure out whether
1520                          * this is a robust futex or not.
1521                          */
1522                         if (get_futex_value_locked(&curval, uaddr))
1523                                 goto uaddr_faulted;
1524
1525                         /*
1526                          * We simply start over in case of a robust
1527                          * futex. The code above will take the futex
1528                          * and return happy.
1529                          */
1530                         if (curval & FUTEX_OWNER_DIED) {
1531                                 ownerdied = 1;
1532                                 goto retry_locked;
1533                         }
1534                 default:
1535                         goto out_unlock_release_sem;
1536                 }
1537         }
1538
1539         /*
1540          * Only actually queue now that the atomic ops are done:
1541          */
1542         queue_me(&q, hb);
1543
1544         /*
1545          * Now the futex is queued and we have checked the data, we
1546          * don't want to hold mmap_sem while we sleep.
1547          */
1548         futex_unlock_mm(fshared);
1549
1550         WARN_ON(!q.pi_state);
1551         /*
1552          * Block on the PI mutex:
1553          */
1554         if (!trylock)
1555                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1556         else {
1557                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1558                 /* Fixup the trylock return value: */
1559                 ret = ret ? 0 : -EWOULDBLOCK;
1560         }
1561
1562         futex_lock_mm(fshared);
1563         spin_lock(q.lock_ptr);
1564
1565         if (!ret) {
1566                 /*
1567                  * Got the lock. We might not be the anticipated owner
1568                  * if we did a lock-steal - fix up the PI-state in
1569                  * that case:
1570                  */
1571                 if (q.pi_state->owner != curr)
1572                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1573         } else {
1574                 /*
1575                  * Catch the rare case, where the lock was released
1576                  * when we were on the way back before we locked the
1577                  * hash bucket.
1578                  */
1579                 if (q.pi_state->owner == curr) {
1580                         /*
1581                          * Try to get the rt_mutex now. This might
1582                          * fail as some other task acquired the
1583                          * rt_mutex after we removed ourself from the
1584                          * rt_mutex waiters list.
1585                          */
1586                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1587                                 ret = 0;
1588                         else {
1589                                 /*
1590                                  * pi_state is incorrect, some other
1591                                  * task did a lock steal and we
1592                                  * returned due to timeout or signal
1593                                  * without taking the rt_mutex. Too
1594                                  * late. We can access the
1595                                  * rt_mutex_owner without locking, as
1596                                  * the other task is now blocked on
1597                                  * the hash bucket lock. Fix the state
1598                                  * up.
1599                                  */
1600                                 struct task_struct *owner;
1601                                 int res;
1602
1603                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1604                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1605                                                            fshared);
1606
1607                                 /* propagate -EFAULT, if the fixup failed */
1608                                 if (res)
1609                                         ret = res;
1610                         }
1611                 } else {
1612                         /*
1613                          * Paranoia check. If we did not take the lock
1614                          * in the trylock above, then we should not be
1615                          * the owner of the rtmutex, neither the real
1616                          * nor the pending one:
1617                          */
1618                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1619                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1620                                        "pi-mutex: %p pi-state %p\n", ret,
1621                                        q.pi_state->pi_mutex.owner,
1622                                        q.pi_state->owner);
1623                 }
1624         }
1625
1626         /* Unqueue and drop the lock */
1627         unqueue_me_pi(&q);
1628         futex_unlock_mm(fshared);
1629
1630         if (to)
1631                 destroy_hrtimer_on_stack(&to->timer);
1632         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1633
1634  out_unlock_release_sem:
1635         queue_unlock(&q, hb);
1636
1637  out_release_sem:
1638         futex_unlock_mm(fshared);
1639         if (to)
1640                 destroy_hrtimer_on_stack(&to->timer);
1641         return ret;
1642
1643  uaddr_faulted:
1644         /*
1645          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1646          * non-atomically.  Therefore, if get_user below is not
1647          * enough, we need to handle the fault ourselves, while
1648          * still holding the mmap_sem.
1649          *
1650          * ... and hb->lock. :-) --ANK
1651          */
1652         queue_unlock(&q, hb);
1653
1654         if (attempt++) {
1655                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1656                                          attempt);
1657                 if (ret)
1658                         goto out_release_sem;
1659                 goto retry_unlocked;
1660         }
1661
1662         futex_unlock_mm(fshared);
1663
1664         ret = get_user(uval, uaddr);
1665         if (!ret && (uval != -EFAULT))
1666                 goto retry;
1667
1668         if (to)
1669                 destroy_hrtimer_on_stack(&to->timer);
1670         return ret;
1671 }
1672
1673 /*
1674  * Userspace attempted a TID -> 0 atomic transition, and failed.
1675  * This is the in-kernel slowpath: we look up the PI state (if any),
1676  * and do the rt-mutex unlock.
1677  */
1678 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1679 {
1680         struct futex_hash_bucket *hb;
1681         struct futex_q *this, *next;
1682         u32 uval;
1683         struct plist_head *head;
1684         union futex_key key;
1685         int ret, attempt = 0;
1686
1687 retry:
1688         if (get_user(uval, uaddr))
1689                 return -EFAULT;
1690         /*
1691          * We release only a lock we actually own:
1692          */
1693         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1694                 return -EPERM;
1695         /*
1696          * First take all the futex related locks:
1697          */
1698         futex_lock_mm(fshared);
1699
1700         ret = get_futex_key(uaddr, fshared, &key);
1701         if (unlikely(ret != 0))
1702                 goto out;
1703
1704         hb = hash_futex(&key);
1705 retry_unlocked:
1706         spin_lock(&hb->lock);
1707
1708         /*
1709          * To avoid races, try to do the TID -> 0 atomic transition
1710          * again. If it succeeds then we can return without waking
1711          * anyone else up:
1712          */
1713         if (!(uval & FUTEX_OWNER_DIED))
1714                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1715
1716
1717         if (unlikely(uval == -EFAULT))
1718                 goto pi_faulted;
1719         /*
1720          * Rare case: we managed to release the lock atomically,
1721          * no need to wake anyone else up:
1722          */
1723         if (unlikely(uval == task_pid_vnr(current)))
1724                 goto out_unlock;
1725
1726         /*
1727          * Ok, other tasks may need to be woken up - check waiters
1728          * and do the wakeup if necessary:
1729          */
1730         head = &hb->chain;
1731
1732         plist_for_each_entry_safe(this, next, head, list) {
1733                 if (!match_futex (&this->key, &key))
1734                         continue;
1735                 ret = wake_futex_pi(uaddr, uval, this);
1736                 /*
1737                  * The atomic access to the futex value
1738                  * generated a pagefault, so retry the
1739                  * user-access and the wakeup:
1740                  */
1741                 if (ret == -EFAULT)
1742                         goto pi_faulted;
1743                 goto out_unlock;
1744         }
1745         /*
1746          * No waiters - kernel unlocks the futex:
1747          */
1748         if (!(uval & FUTEX_OWNER_DIED)) {
1749                 ret = unlock_futex_pi(uaddr, uval);
1750                 if (ret == -EFAULT)
1751                         goto pi_faulted;
1752         }
1753
1754 out_unlock:
1755         spin_unlock(&hb->lock);
1756 out:
1757         futex_unlock_mm(fshared);
1758
1759         return ret;
1760
1761 pi_faulted:
1762         /*
1763          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1764          * non-atomically.  Therefore, if get_user below is not
1765          * enough, we need to handle the fault ourselves, while
1766          * still holding the mmap_sem.
1767          *
1768          * ... and hb->lock. --ANK
1769          */
1770         spin_unlock(&hb->lock);
1771
1772         if (attempt++) {
1773                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1774                                          attempt);
1775                 if (ret)
1776                         goto out;
1777                 uval = 0;
1778                 goto retry_unlocked;
1779         }
1780
1781         futex_unlock_mm(fshared);
1782
1783         ret = get_user(uval, uaddr);
1784         if (!ret && (uval != -EFAULT))
1785                 goto retry;
1786
1787         return ret;
1788 }
1789
1790 /*
1791  * Support for robust futexes: the kernel cleans up held futexes at
1792  * thread exit time.
1793  *
1794  * Implementation: user-space maintains a per-thread list of locks it
1795  * is holding. Upon do_exit(), the kernel carefully walks this list,
1796  * and marks all locks that are owned by this thread with the
1797  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1798  * always manipulated with the lock held, so the list is private and
1799  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1800  * field, to allow the kernel to clean up if the thread dies after
1801  * acquiring the lock, but just before it could have added itself to
1802  * the list. There can only be one such pending lock.
1803  */
1804
1805 /**
1806  * sys_set_robust_list - set the robust-futex list head of a task
1807  * @head: pointer to the list-head
1808  * @len: length of the list-head, as userspace expects
1809  */
1810 asmlinkage long
1811 sys_set_robust_list(struct robust_list_head __user *head,
1812                     size_t len)
1813 {
1814         if (!futex_cmpxchg_enabled)
1815                 return -ENOSYS;
1816         /*
1817          * The kernel knows only one size for now:
1818          */
1819         if (unlikely(len != sizeof(*head)))
1820                 return -EINVAL;
1821
1822         current->robust_list = head;
1823
1824         return 0;
1825 }
1826
1827 /**
1828  * sys_get_robust_list - get the robust-futex list head of a task
1829  * @pid: pid of the process [zero for current task]
1830  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1831  * @len_ptr: pointer to a length field, the kernel fills in the header size
1832  */
1833 asmlinkage long
1834 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1835                     size_t __user *len_ptr)
1836 {
1837         struct robust_list_head __user *head;
1838         unsigned long ret;
1839         const struct cred *cred = current_cred(), *pcred;
1840
1841         if (!futex_cmpxchg_enabled)
1842                 return -ENOSYS;
1843
1844         if (!pid)
1845                 head = current->robust_list;
1846         else {
1847                 struct task_struct *p;
1848
1849                 ret = -ESRCH;
1850                 rcu_read_lock();
1851                 p = find_task_by_vpid(pid);
1852                 if (!p)
1853                         goto err_unlock;
1854                 ret = -EPERM;
1855                 pcred = __task_cred(p);
1856                 if (cred->euid != pcred->euid &&
1857                     cred->euid != pcred->uid &&
1858                     !capable(CAP_SYS_PTRACE))
1859                         goto err_unlock;
1860                 head = p->robust_list;
1861                 rcu_read_unlock();
1862         }
1863
1864         if (put_user(sizeof(*head), len_ptr))
1865                 return -EFAULT;
1866         return put_user(head, head_ptr);
1867
1868 err_unlock:
1869         rcu_read_unlock();
1870
1871         return ret;
1872 }
1873
1874 /*
1875  * Process a futex-list entry, check whether it's owned by the
1876  * dying task, and do notification if so:
1877  */
1878 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1879 {
1880         u32 uval, nval, mval;
1881
1882 retry:
1883         if (get_user(uval, uaddr))
1884                 return -1;
1885
1886         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1887                 /*
1888                  * Ok, this dying thread is truly holding a futex
1889                  * of interest. Set the OWNER_DIED bit atomically
1890                  * via cmpxchg, and if the value had FUTEX_WAITERS
1891                  * set, wake up a waiter (if any). (We have to do a
1892                  * futex_wake() even if OWNER_DIED is already set -
1893                  * to handle the rare but possible case of recursive
1894                  * thread-death.) The rest of the cleanup is done in
1895                  * userspace.
1896                  */
1897                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1898                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1899
1900                 if (nval == -EFAULT)
1901                         return -1;
1902
1903                 if (nval != uval)
1904                         goto retry;
1905
1906                 /*
1907                  * Wake robust non-PI futexes here. The wakeup of
1908                  * PI futexes happens in exit_pi_state():
1909                  */
1910                 if (!pi && (uval & FUTEX_WAITERS))
1911                         futex_wake(uaddr, &curr->mm->mmap_sem, 1,
1912                                    FUTEX_BITSET_MATCH_ANY);
1913         }
1914         return 0;
1915 }
1916
1917 /*
1918  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1919  */
1920 static inline int fetch_robust_entry(struct robust_list __user **entry,
1921                                      struct robust_list __user * __user *head,
1922                                      int *pi)
1923 {
1924         unsigned long uentry;
1925
1926         if (get_user(uentry, (unsigned long __user *)head))
1927                 return -EFAULT;
1928
1929         *entry = (void __user *)(uentry & ~1UL);
1930         *pi = uentry & 1;
1931
1932         return 0;
1933 }
1934
1935 /*
1936  * Walk curr->robust_list (very carefully, it's a userspace list!)
1937  * and mark any locks found there dead, and notify any waiters.
1938  *
1939  * We silently return on any sign of list-walking problem.
1940  */
1941 void exit_robust_list(struct task_struct *curr)
1942 {
1943         struct robust_list_head __user *head = curr->robust_list;
1944         struct robust_list __user *entry, *next_entry, *pending;
1945         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1946         unsigned long futex_offset;
1947         int rc;
1948
1949         if (!futex_cmpxchg_enabled)
1950                 return;
1951
1952         /*
1953          * Fetch the list head (which was registered earlier, via
1954          * sys_set_robust_list()):
1955          */
1956         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1957                 return;
1958         /*
1959          * Fetch the relative futex offset:
1960          */
1961         if (get_user(futex_offset, &head->futex_offset))
1962                 return;
1963         /*
1964          * Fetch any possibly pending lock-add first, and handle it
1965          * if it exists:
1966          */
1967         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1968                 return;
1969
1970         next_entry = NULL;      /* avoid warning with gcc */
1971         while (entry != &head->list) {
1972                 /*
1973                  * Fetch the next entry in the list before calling
1974                  * handle_futex_death:
1975                  */
1976                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1977                 /*
1978                  * A pending lock might already be on the list, so
1979                  * don't process it twice:
1980                  */
1981                 if (entry != pending)
1982                         if (handle_futex_death((void __user *)entry + futex_offset,
1983                                                 curr, pi))
1984                                 return;
1985                 if (rc)
1986                         return;
1987                 entry = next_entry;
1988                 pi = next_pi;
1989                 /*
1990                  * Avoid excessively long or circular lists:
1991                  */
1992                 if (!--limit)
1993                         break;
1994
1995                 cond_resched();
1996         }
1997
1998         if (pending)
1999                 handle_futex_death((void __user *)pending + futex_offset,
2000                                    curr, pip);
2001 }
2002
2003 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2004                 u32 __user *uaddr2, u32 val2, u32 val3)
2005 {
2006         int ret = -ENOSYS;
2007         int cmd = op & FUTEX_CMD_MASK;
2008         struct rw_semaphore *fshared = NULL;
2009
2010         if (!(op & FUTEX_PRIVATE_FLAG))
2011                 fshared = &current->mm->mmap_sem;
2012
2013         switch (cmd) {
2014         case FUTEX_WAIT:
2015                 val3 = FUTEX_BITSET_MATCH_ANY;
2016         case FUTEX_WAIT_BITSET:
2017                 ret = futex_wait(uaddr, fshared, val, timeout, val3);
2018                 break;
2019         case FUTEX_WAKE:
2020                 val3 = FUTEX_BITSET_MATCH_ANY;
2021         case FUTEX_WAKE_BITSET:
2022                 ret = futex_wake(uaddr, fshared, val, val3);
2023                 break;
2024         case FUTEX_REQUEUE:
2025                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2026                 break;
2027         case FUTEX_CMP_REQUEUE:
2028                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2029                 break;
2030         case FUTEX_WAKE_OP:
2031                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2032                 break;
2033         case FUTEX_LOCK_PI:
2034                 if (futex_cmpxchg_enabled)
2035                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2036                 break;
2037         case FUTEX_UNLOCK_PI:
2038                 if (futex_cmpxchg_enabled)
2039                         ret = futex_unlock_pi(uaddr, fshared);
2040                 break;
2041         case FUTEX_TRYLOCK_PI:
2042                 if (futex_cmpxchg_enabled)
2043                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2044                 break;
2045         default:
2046                 ret = -ENOSYS;
2047         }
2048         return ret;
2049 }
2050
2051
2052 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2053                           struct timespec __user *utime, u32 __user *uaddr2,
2054                           u32 val3)
2055 {
2056         struct timespec ts;
2057         ktime_t t, *tp = NULL;
2058         u32 val2 = 0;
2059         int cmd = op & FUTEX_CMD_MASK;
2060
2061         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2062                       cmd == FUTEX_WAIT_BITSET)) {
2063                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2064                         return -EFAULT;
2065                 if (!timespec_valid(&ts))
2066                         return -EINVAL;
2067
2068                 t = timespec_to_ktime(ts);
2069                 if (cmd == FUTEX_WAIT)
2070                         t = ktime_add_safe(ktime_get(), t);
2071                 tp = &t;
2072         }
2073         /*
2074          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2075          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2076          */
2077         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2078             cmd == FUTEX_WAKE_OP)
2079                 val2 = (u32) (unsigned long) utime;
2080
2081         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2082 }
2083
2084 static int __init futex_init(void)
2085 {
2086         u32 curval;
2087         int i;
2088
2089         /*
2090          * This will fail and we want it. Some arch implementations do
2091          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2092          * functionality. We want to know that before we call in any
2093          * of the complex code paths. Also we want to prevent
2094          * registration of robust lists in that case. NULL is
2095          * guaranteed to fault and we get -EFAULT on functional
2096          * implementation, the non functional ones will return
2097          * -ENOSYS.
2098          */
2099         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2100         if (curval == -EFAULT)
2101                 futex_cmpxchg_enabled = 1;
2102
2103         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2104                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2105                 spin_lock_init(&futex_queues[i].lock);
2106         }
2107
2108         return 0;
2109 }
2110 __initcall(futex_init);