futex: fixup get_futex_key() for private futexes
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
58
59 #include <asm/futex.h>
60
61 #include "rtmutex_common.h"
62
63 int __read_mostly futex_cmpxchg_enabled;
64
65 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
66
67 /*
68  * Priority Inheritance state:
69  */
70 struct futex_pi_state {
71         /*
72          * list of 'owned' pi_state instances - these have to be
73          * cleaned up in do_exit() if the task exits prematurely:
74          */
75         struct list_head list;
76
77         /*
78          * The PI object:
79          */
80         struct rt_mutex pi_mutex;
81
82         struct task_struct *owner;
83         atomic_t refcount;
84
85         union futex_key key;
86 };
87
88 /*
89  * We use this hashed waitqueue instead of a normal wait_queue_t, so
90  * we can wake only the relevant ones (hashed queues may be shared).
91  *
92  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
93  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
94  * The order of wakup is always to make the first condition true, then
95  * wake up q->waiters, then make the second condition true.
96  */
97 struct futex_q {
98         struct plist_node list;
99         wait_queue_head_t waiters;
100
101         /* Which hash list lock to use: */
102         spinlock_t *lock_ptr;
103
104         /* Key which the futex is hashed on: */
105         union futex_key key;
106
107         /* Optional priority inheritance state: */
108         struct futex_pi_state *pi_state;
109         struct task_struct *task;
110
111         /* Bitset for the optional bitmasked wakeup */
112         u32 bitset;
113 };
114
115 /*
116  * Split the global futex_lock into every hash list lock.
117  */
118 struct futex_hash_bucket {
119         spinlock_t lock;
120         struct plist_head chain;
121 };
122
123 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
124
125 /*
126  * We hash on the keys returned from get_futex_key (see below).
127  */
128 static struct futex_hash_bucket *hash_futex(union futex_key *key)
129 {
130         u32 hash = jhash2((u32*)&key->both.word,
131                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
132                           key->both.offset);
133         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
134 }
135
136 /*
137  * Return 1 if two futex_keys are equal, 0 otherwise.
138  */
139 static inline int match_futex(union futex_key *key1, union futex_key *key2)
140 {
141         return (key1->both.word == key2->both.word
142                 && key1->both.ptr == key2->both.ptr
143                 && key1->both.offset == key2->both.offset);
144 }
145
146 /*
147  * Take a reference to the resource addressed by a key.
148  * Can be called while holding spinlocks.
149  *
150  */
151 static void get_futex_key_refs(union futex_key *key)
152 {
153         if (!key->both.ptr)
154                 return;
155
156         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
157         case FUT_OFF_INODE:
158                 atomic_inc(&key->shared.inode->i_count);
159                 break;
160         case FUT_OFF_MMSHARED:
161                 atomic_inc(&key->private.mm->mm_count);
162                 break;
163         }
164 }
165
166 /*
167  * Drop a reference to the resource addressed by a key.
168  * The hash bucket spinlock must not be held.
169  */
170 static void drop_futex_key_refs(union futex_key *key)
171 {
172         if (!key->both.ptr)
173                 return;
174
175         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
176         case FUT_OFF_INODE:
177                 iput(key->shared.inode);
178                 break;
179         case FUT_OFF_MMSHARED:
180                 mmdrop(key->private.mm);
181                 break;
182         }
183 }
184
185 /**
186  * get_futex_key - Get parameters which are the keys for a futex.
187  * @uaddr: virtual address of the futex
188  * @shared: NULL for a PROCESS_PRIVATE futex,
189  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
190  * @key: address where result is stored.
191  *
192  * Returns a negative error code or 0
193  * The key words are stored in *key on success.
194  *
195  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
196  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
197  * We can usually work out the index without swapping in the page.
198  *
199  * fshared is NULL for PROCESS_PRIVATE futexes
200  * For other futexes, it points to &current->mm->mmap_sem and
201  * caller must have taken the reader lock. but NOT any spinlocks.
202  */
203 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
204 {
205         unsigned long address = (unsigned long)uaddr;
206         struct mm_struct *mm = current->mm;
207         struct page *page;
208         int err;
209
210         /*
211          * The futex address must be "naturally" aligned.
212          */
213         key->both.offset = address % PAGE_SIZE;
214         if (unlikely((address % sizeof(u32)) != 0))
215                 return -EINVAL;
216         address -= key->both.offset;
217
218         /*
219          * PROCESS_PRIVATE futexes are fast.
220          * As the mm cannot disappear under us and the 'key' only needs
221          * virtual address, we dont even have to find the underlying vma.
222          * Note : We do have to check 'uaddr' is a valid user address,
223          *        but access_ok() should be faster than find_vma()
224          */
225         if (!fshared) {
226                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
227                         return -EFAULT;
228                 key->private.mm = mm;
229                 key->private.address = address;
230                 get_futex_key_refs(key);
231                 return 0;
232         }
233
234 again:
235         err = get_user_pages_fast(address, 1, 0, &page);
236         if (err < 0)
237                 return err;
238
239         lock_page(page);
240         if (!page->mapping) {
241                 unlock_page(page);
242                 put_page(page);
243                 goto again;
244         }
245
246         /*
247          * Private mappings are handled in a simple way.
248          *
249          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
250          * it's a read-only handle, it's expected that futexes attach to
251          * the object not the particular process.
252          */
253         if (PageAnon(page)) {
254                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
255                 key->private.mm = mm;
256                 key->private.address = address;
257         } else {
258                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
259                 key->shared.inode = page->mapping->host;
260                 key->shared.pgoff = page->index;
261         }
262
263         get_futex_key_refs(key);
264
265         unlock_page(page);
266         put_page(page);
267         return 0;
268 }
269
270 static inline
271 void put_futex_key(int fshared, union futex_key *key)
272 {
273         drop_futex_key_refs(key);
274 }
275
276 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
277 {
278         u32 curval;
279
280         pagefault_disable();
281         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
282         pagefault_enable();
283
284         return curval;
285 }
286
287 static int get_futex_value_locked(u32 *dest, u32 __user *from)
288 {
289         int ret;
290
291         pagefault_disable();
292         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
293         pagefault_enable();
294
295         return ret ? -EFAULT : 0;
296 }
297
298 /*
299  * Fault handling.
300  */
301 static int futex_handle_fault(unsigned long address, int attempt)
302 {
303         struct vm_area_struct * vma;
304         struct mm_struct *mm = current->mm;
305         int ret = -EFAULT;
306
307         if (attempt > 2)
308                 return ret;
309
310         down_read(&mm->mmap_sem);
311         vma = find_vma(mm, address);
312         if (vma && address >= vma->vm_start &&
313             (vma->vm_flags & VM_WRITE)) {
314                 int fault;
315                 fault = handle_mm_fault(mm, vma, address, 1);
316                 if (unlikely((fault & VM_FAULT_ERROR))) {
317 #if 0
318                         /* XXX: let's do this when we verify it is OK */
319                         if (ret & VM_FAULT_OOM)
320                                 ret = -ENOMEM;
321 #endif
322                 } else {
323                         ret = 0;
324                         if (fault & VM_FAULT_MAJOR)
325                                 current->maj_flt++;
326                         else
327                                 current->min_flt++;
328                 }
329         }
330         up_read(&mm->mmap_sem);
331         return ret;
332 }
333
334 /*
335  * PI code:
336  */
337 static int refill_pi_state_cache(void)
338 {
339         struct futex_pi_state *pi_state;
340
341         if (likely(current->pi_state_cache))
342                 return 0;
343
344         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
345
346         if (!pi_state)
347                 return -ENOMEM;
348
349         INIT_LIST_HEAD(&pi_state->list);
350         /* pi_mutex gets initialized later */
351         pi_state->owner = NULL;
352         atomic_set(&pi_state->refcount, 1);
353         pi_state->key = FUTEX_KEY_INIT;
354
355         current->pi_state_cache = pi_state;
356
357         return 0;
358 }
359
360 static struct futex_pi_state * alloc_pi_state(void)
361 {
362         struct futex_pi_state *pi_state = current->pi_state_cache;
363
364         WARN_ON(!pi_state);
365         current->pi_state_cache = NULL;
366
367         return pi_state;
368 }
369
370 static void free_pi_state(struct futex_pi_state *pi_state)
371 {
372         if (!atomic_dec_and_test(&pi_state->refcount))
373                 return;
374
375         /*
376          * If pi_state->owner is NULL, the owner is most probably dying
377          * and has cleaned up the pi_state already
378          */
379         if (pi_state->owner) {
380                 spin_lock_irq(&pi_state->owner->pi_lock);
381                 list_del_init(&pi_state->list);
382                 spin_unlock_irq(&pi_state->owner->pi_lock);
383
384                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
385         }
386
387         if (current->pi_state_cache)
388                 kfree(pi_state);
389         else {
390                 /*
391                  * pi_state->list is already empty.
392                  * clear pi_state->owner.
393                  * refcount is at 0 - put it back to 1.
394                  */
395                 pi_state->owner = NULL;
396                 atomic_set(&pi_state->refcount, 1);
397                 current->pi_state_cache = pi_state;
398         }
399 }
400
401 /*
402  * Look up the task based on what TID userspace gave us.
403  * We dont trust it.
404  */
405 static struct task_struct * futex_find_get_task(pid_t pid)
406 {
407         struct task_struct *p;
408
409         rcu_read_lock();
410         p = find_task_by_vpid(pid);
411         if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
412                 p = ERR_PTR(-ESRCH);
413         else
414                 get_task_struct(p);
415
416         rcu_read_unlock();
417
418         return p;
419 }
420
421 /*
422  * This task is holding PI mutexes at exit time => bad.
423  * Kernel cleans up PI-state, but userspace is likely hosed.
424  * (Robust-futex cleanup is separate and might save the day for userspace.)
425  */
426 void exit_pi_state_list(struct task_struct *curr)
427 {
428         struct list_head *next, *head = &curr->pi_state_list;
429         struct futex_pi_state *pi_state;
430         struct futex_hash_bucket *hb;
431         union futex_key key = FUTEX_KEY_INIT;
432
433         if (!futex_cmpxchg_enabled)
434                 return;
435         /*
436          * We are a ZOMBIE and nobody can enqueue itself on
437          * pi_state_list anymore, but we have to be careful
438          * versus waiters unqueueing themselves:
439          */
440         spin_lock_irq(&curr->pi_lock);
441         while (!list_empty(head)) {
442
443                 next = head->next;
444                 pi_state = list_entry(next, struct futex_pi_state, list);
445                 key = pi_state->key;
446                 hb = hash_futex(&key);
447                 spin_unlock_irq(&curr->pi_lock);
448
449                 spin_lock(&hb->lock);
450
451                 spin_lock_irq(&curr->pi_lock);
452                 /*
453                  * We dropped the pi-lock, so re-check whether this
454                  * task still owns the PI-state:
455                  */
456                 if (head->next != next) {
457                         spin_unlock(&hb->lock);
458                         continue;
459                 }
460
461                 WARN_ON(pi_state->owner != curr);
462                 WARN_ON(list_empty(&pi_state->list));
463                 list_del_init(&pi_state->list);
464                 pi_state->owner = NULL;
465                 spin_unlock_irq(&curr->pi_lock);
466
467                 rt_mutex_unlock(&pi_state->pi_mutex);
468
469                 spin_unlock(&hb->lock);
470
471                 spin_lock_irq(&curr->pi_lock);
472         }
473         spin_unlock_irq(&curr->pi_lock);
474 }
475
476 static int
477 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
478                 union futex_key *key, struct futex_pi_state **ps)
479 {
480         struct futex_pi_state *pi_state = NULL;
481         struct futex_q *this, *next;
482         struct plist_head *head;
483         struct task_struct *p;
484         pid_t pid = uval & FUTEX_TID_MASK;
485
486         head = &hb->chain;
487
488         plist_for_each_entry_safe(this, next, head, list) {
489                 if (match_futex(&this->key, key)) {
490                         /*
491                          * Another waiter already exists - bump up
492                          * the refcount and return its pi_state:
493                          */
494                         pi_state = this->pi_state;
495                         /*
496                          * Userspace might have messed up non PI and PI futexes
497                          */
498                         if (unlikely(!pi_state))
499                                 return -EINVAL;
500
501                         WARN_ON(!atomic_read(&pi_state->refcount));
502                         WARN_ON(pid && pi_state->owner &&
503                                 pi_state->owner->pid != pid);
504
505                         atomic_inc(&pi_state->refcount);
506                         *ps = pi_state;
507
508                         return 0;
509                 }
510         }
511
512         /*
513          * We are the first waiter - try to look up the real owner and attach
514          * the new pi_state to it, but bail out when TID = 0
515          */
516         if (!pid)
517                 return -ESRCH;
518         p = futex_find_get_task(pid);
519         if (IS_ERR(p))
520                 return PTR_ERR(p);
521
522         /*
523          * We need to look at the task state flags to figure out,
524          * whether the task is exiting. To protect against the do_exit
525          * change of the task flags, we do this protected by
526          * p->pi_lock:
527          */
528         spin_lock_irq(&p->pi_lock);
529         if (unlikely(p->flags & PF_EXITING)) {
530                 /*
531                  * The task is on the way out. When PF_EXITPIDONE is
532                  * set, we know that the task has finished the
533                  * cleanup:
534                  */
535                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
536
537                 spin_unlock_irq(&p->pi_lock);
538                 put_task_struct(p);
539                 return ret;
540         }
541
542         pi_state = alloc_pi_state();
543
544         /*
545          * Initialize the pi_mutex in locked state and make 'p'
546          * the owner of it:
547          */
548         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
549
550         /* Store the key for possible exit cleanups: */
551         pi_state->key = *key;
552
553         WARN_ON(!list_empty(&pi_state->list));
554         list_add(&pi_state->list, &p->pi_state_list);
555         pi_state->owner = p;
556         spin_unlock_irq(&p->pi_lock);
557
558         put_task_struct(p);
559
560         *ps = pi_state;
561
562         return 0;
563 }
564
565 /*
566  * The hash bucket lock must be held when this is called.
567  * Afterwards, the futex_q must not be accessed.
568  */
569 static void wake_futex(struct futex_q *q)
570 {
571         plist_del(&q->list, &q->list.plist);
572         /*
573          * The lock in wake_up_all() is a crucial memory barrier after the
574          * plist_del() and also before assigning to q->lock_ptr.
575          */
576         wake_up_all(&q->waiters);
577         /*
578          * The waiting task can free the futex_q as soon as this is written,
579          * without taking any locks.  This must come last.
580          *
581          * A memory barrier is required here to prevent the following store
582          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
583          * at the end of wake_up_all() does not prevent this store from
584          * moving.
585          */
586         smp_wmb();
587         q->lock_ptr = NULL;
588 }
589
590 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
591 {
592         struct task_struct *new_owner;
593         struct futex_pi_state *pi_state = this->pi_state;
594         u32 curval, newval;
595
596         if (!pi_state)
597                 return -EINVAL;
598
599         spin_lock(&pi_state->pi_mutex.wait_lock);
600         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
601
602         /*
603          * This happens when we have stolen the lock and the original
604          * pending owner did not enqueue itself back on the rt_mutex.
605          * Thats not a tragedy. We know that way, that a lock waiter
606          * is on the fly. We make the futex_q waiter the pending owner.
607          */
608         if (!new_owner)
609                 new_owner = this->task;
610
611         /*
612          * We pass it to the next owner. (The WAITERS bit is always
613          * kept enabled while there is PI state around. We must also
614          * preserve the owner died bit.)
615          */
616         if (!(uval & FUTEX_OWNER_DIED)) {
617                 int ret = 0;
618
619                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
620
621                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
622
623                 if (curval == -EFAULT)
624                         ret = -EFAULT;
625                 else if (curval != uval)
626                         ret = -EINVAL;
627                 if (ret) {
628                         spin_unlock(&pi_state->pi_mutex.wait_lock);
629                         return ret;
630                 }
631         }
632
633         spin_lock_irq(&pi_state->owner->pi_lock);
634         WARN_ON(list_empty(&pi_state->list));
635         list_del_init(&pi_state->list);
636         spin_unlock_irq(&pi_state->owner->pi_lock);
637
638         spin_lock_irq(&new_owner->pi_lock);
639         WARN_ON(!list_empty(&pi_state->list));
640         list_add(&pi_state->list, &new_owner->pi_state_list);
641         pi_state->owner = new_owner;
642         spin_unlock_irq(&new_owner->pi_lock);
643
644         spin_unlock(&pi_state->pi_mutex.wait_lock);
645         rt_mutex_unlock(&pi_state->pi_mutex);
646
647         return 0;
648 }
649
650 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
651 {
652         u32 oldval;
653
654         /*
655          * There is no waiter, so we unlock the futex. The owner died
656          * bit has not to be preserved here. We are the owner:
657          */
658         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
659
660         if (oldval == -EFAULT)
661                 return oldval;
662         if (oldval != uval)
663                 return -EAGAIN;
664
665         return 0;
666 }
667
668 /*
669  * Express the locking dependencies for lockdep:
670  */
671 static inline void
672 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
673 {
674         if (hb1 <= hb2) {
675                 spin_lock(&hb1->lock);
676                 if (hb1 < hb2)
677                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
678         } else { /* hb1 > hb2 */
679                 spin_lock(&hb2->lock);
680                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
681         }
682 }
683
684 /*
685  * Wake up all waiters hashed on the physical page that is mapped
686  * to this virtual address:
687  */
688 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
689 {
690         struct futex_hash_bucket *hb;
691         struct futex_q *this, *next;
692         struct plist_head *head;
693         union futex_key key = FUTEX_KEY_INIT;
694         int ret;
695
696         if (!bitset)
697                 return -EINVAL;
698
699         ret = get_futex_key(uaddr, fshared, &key);
700         if (unlikely(ret != 0))
701                 goto out;
702
703         hb = hash_futex(&key);
704         spin_lock(&hb->lock);
705         head = &hb->chain;
706
707         plist_for_each_entry_safe(this, next, head, list) {
708                 if (match_futex (&this->key, &key)) {
709                         if (this->pi_state) {
710                                 ret = -EINVAL;
711                                 break;
712                         }
713
714                         /* Check if one of the bits is set in both bitsets */
715                         if (!(this->bitset & bitset))
716                                 continue;
717
718                         wake_futex(this);
719                         if (++ret >= nr_wake)
720                                 break;
721                 }
722         }
723
724         spin_unlock(&hb->lock);
725 out:
726         put_futex_key(fshared, &key);
727         return ret;
728 }
729
730 /*
731  * Wake up all waiters hashed on the physical page that is mapped
732  * to this virtual address:
733  */
734 static int
735 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
736               int nr_wake, int nr_wake2, int op)
737 {
738         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
739         struct futex_hash_bucket *hb1, *hb2;
740         struct plist_head *head;
741         struct futex_q *this, *next;
742         int ret, op_ret, attempt = 0;
743
744 retryfull:
745         ret = get_futex_key(uaddr1, fshared, &key1);
746         if (unlikely(ret != 0))
747                 goto out;
748         ret = get_futex_key(uaddr2, fshared, &key2);
749         if (unlikely(ret != 0))
750                 goto out;
751
752         hb1 = hash_futex(&key1);
753         hb2 = hash_futex(&key2);
754
755 retry:
756         double_lock_hb(hb1, hb2);
757
758         op_ret = futex_atomic_op_inuser(op, uaddr2);
759         if (unlikely(op_ret < 0)) {
760                 u32 dummy;
761
762                 spin_unlock(&hb1->lock);
763                 if (hb1 != hb2)
764                         spin_unlock(&hb2->lock);
765
766 #ifndef CONFIG_MMU
767                 /*
768                  * we don't get EFAULT from MMU faults if we don't have an MMU,
769                  * but we might get them from range checking
770                  */
771                 ret = op_ret;
772                 goto out;
773 #endif
774
775                 if (unlikely(op_ret != -EFAULT)) {
776                         ret = op_ret;
777                         goto out;
778                 }
779
780                 /*
781                  * futex_atomic_op_inuser needs to both read and write
782                  * *(int __user *)uaddr2, but we can't modify it
783                  * non-atomically.  Therefore, if get_user below is not
784                  * enough, we need to handle the fault ourselves, while
785                  * still holding the mmap_sem.
786                  */
787                 if (attempt++) {
788                         ret = futex_handle_fault((unsigned long)uaddr2,
789                                                  attempt);
790                         if (ret)
791                                 goto out;
792                         goto retry;
793                 }
794
795                 ret = get_user(dummy, uaddr2);
796                 if (ret)
797                         return ret;
798
799                 goto retryfull;
800         }
801
802         head = &hb1->chain;
803
804         plist_for_each_entry_safe(this, next, head, list) {
805                 if (match_futex (&this->key, &key1)) {
806                         wake_futex(this);
807                         if (++ret >= nr_wake)
808                                 break;
809                 }
810         }
811
812         if (op_ret > 0) {
813                 head = &hb2->chain;
814
815                 op_ret = 0;
816                 plist_for_each_entry_safe(this, next, head, list) {
817                         if (match_futex (&this->key, &key2)) {
818                                 wake_futex(this);
819                                 if (++op_ret >= nr_wake2)
820                                         break;
821                         }
822                 }
823                 ret += op_ret;
824         }
825
826         spin_unlock(&hb1->lock);
827         if (hb1 != hb2)
828                 spin_unlock(&hb2->lock);
829 out:
830         put_futex_key(fshared, &key2);
831         put_futex_key(fshared, &key1);
832
833         return ret;
834 }
835
836 /*
837  * Requeue all waiters hashed on one physical page to another
838  * physical page.
839  */
840 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
841                          int nr_wake, int nr_requeue, u32 *cmpval)
842 {
843         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
844         struct futex_hash_bucket *hb1, *hb2;
845         struct plist_head *head1;
846         struct futex_q *this, *next;
847         int ret, drop_count = 0;
848
849  retry:
850         ret = get_futex_key(uaddr1, fshared, &key1);
851         if (unlikely(ret != 0))
852                 goto out;
853         ret = get_futex_key(uaddr2, fshared, &key2);
854         if (unlikely(ret != 0))
855                 goto out;
856
857         hb1 = hash_futex(&key1);
858         hb2 = hash_futex(&key2);
859
860         double_lock_hb(hb1, hb2);
861
862         if (likely(cmpval != NULL)) {
863                 u32 curval;
864
865                 ret = get_futex_value_locked(&curval, uaddr1);
866
867                 if (unlikely(ret)) {
868                         spin_unlock(&hb1->lock);
869                         if (hb1 != hb2)
870                                 spin_unlock(&hb2->lock);
871
872                         ret = get_user(curval, uaddr1);
873
874                         if (!ret)
875                                 goto retry;
876
877                         return ret;
878                 }
879                 if (curval != *cmpval) {
880                         ret = -EAGAIN;
881                         goto out_unlock;
882                 }
883         }
884
885         head1 = &hb1->chain;
886         plist_for_each_entry_safe(this, next, head1, list) {
887                 if (!match_futex (&this->key, &key1))
888                         continue;
889                 if (++ret <= nr_wake) {
890                         wake_futex(this);
891                 } else {
892                         /*
893                          * If key1 and key2 hash to the same bucket, no need to
894                          * requeue.
895                          */
896                         if (likely(head1 != &hb2->chain)) {
897                                 plist_del(&this->list, &hb1->chain);
898                                 plist_add(&this->list, &hb2->chain);
899                                 this->lock_ptr = &hb2->lock;
900 #ifdef CONFIG_DEBUG_PI_LIST
901                                 this->list.plist.lock = &hb2->lock;
902 #endif
903                         }
904                         this->key = key2;
905                         get_futex_key_refs(&key2);
906                         drop_count++;
907
908                         if (ret - nr_wake >= nr_requeue)
909                                 break;
910                 }
911         }
912
913 out_unlock:
914         spin_unlock(&hb1->lock);
915         if (hb1 != hb2)
916                 spin_unlock(&hb2->lock);
917
918         /* drop_futex_key_refs() must be called outside the spinlocks. */
919         while (--drop_count >= 0)
920                 drop_futex_key_refs(&key1);
921
922 out:
923         put_futex_key(fshared, &key2);
924         put_futex_key(fshared, &key1);
925         return ret;
926 }
927
928 /* The key must be already stored in q->key. */
929 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
930 {
931         struct futex_hash_bucket *hb;
932
933         init_waitqueue_head(&q->waiters);
934
935         get_futex_key_refs(&q->key);
936         hb = hash_futex(&q->key);
937         q->lock_ptr = &hb->lock;
938
939         spin_lock(&hb->lock);
940         return hb;
941 }
942
943 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
944 {
945         int prio;
946
947         /*
948          * The priority used to register this element is
949          * - either the real thread-priority for the real-time threads
950          * (i.e. threads with a priority lower than MAX_RT_PRIO)
951          * - or MAX_RT_PRIO for non-RT threads.
952          * Thus, all RT-threads are woken first in priority order, and
953          * the others are woken last, in FIFO order.
954          */
955         prio = min(current->normal_prio, MAX_RT_PRIO);
956
957         plist_node_init(&q->list, prio);
958 #ifdef CONFIG_DEBUG_PI_LIST
959         q->list.plist.lock = &hb->lock;
960 #endif
961         plist_add(&q->list, &hb->chain);
962         q->task = current;
963         spin_unlock(&hb->lock);
964 }
965
966 static inline void
967 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
968 {
969         spin_unlock(&hb->lock);
970         drop_futex_key_refs(&q->key);
971 }
972
973 /*
974  * queue_me and unqueue_me must be called as a pair, each
975  * exactly once.  They are called with the hashed spinlock held.
976  */
977
978 /* Return 1 if we were still queued (ie. 0 means we were woken) */
979 static int unqueue_me(struct futex_q *q)
980 {
981         spinlock_t *lock_ptr;
982         int ret = 0;
983
984         /* In the common case we don't take the spinlock, which is nice. */
985  retry:
986         lock_ptr = q->lock_ptr;
987         barrier();
988         if (lock_ptr != NULL) {
989                 spin_lock(lock_ptr);
990                 /*
991                  * q->lock_ptr can change between reading it and
992                  * spin_lock(), causing us to take the wrong lock.  This
993                  * corrects the race condition.
994                  *
995                  * Reasoning goes like this: if we have the wrong lock,
996                  * q->lock_ptr must have changed (maybe several times)
997                  * between reading it and the spin_lock().  It can
998                  * change again after the spin_lock() but only if it was
999                  * already changed before the spin_lock().  It cannot,
1000                  * however, change back to the original value.  Therefore
1001                  * we can detect whether we acquired the correct lock.
1002                  */
1003                 if (unlikely(lock_ptr != q->lock_ptr)) {
1004                         spin_unlock(lock_ptr);
1005                         goto retry;
1006                 }
1007                 WARN_ON(plist_node_empty(&q->list));
1008                 plist_del(&q->list, &q->list.plist);
1009
1010                 BUG_ON(q->pi_state);
1011
1012                 spin_unlock(lock_ptr);
1013                 ret = 1;
1014         }
1015
1016         drop_futex_key_refs(&q->key);
1017         return ret;
1018 }
1019
1020 /*
1021  * PI futexes can not be requeued and must remove themself from the
1022  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1023  * and dropped here.
1024  */
1025 static void unqueue_me_pi(struct futex_q *q)
1026 {
1027         WARN_ON(plist_node_empty(&q->list));
1028         plist_del(&q->list, &q->list.plist);
1029
1030         BUG_ON(!q->pi_state);
1031         free_pi_state(q->pi_state);
1032         q->pi_state = NULL;
1033
1034         spin_unlock(q->lock_ptr);
1035
1036         drop_futex_key_refs(&q->key);
1037 }
1038
1039 /*
1040  * Fixup the pi_state owner with the new owner.
1041  *
1042  * Must be called with hash bucket lock held and mm->sem held for non
1043  * private futexes.
1044  */
1045 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1046                                 struct task_struct *newowner, int fshared)
1047 {
1048         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1049         struct futex_pi_state *pi_state = q->pi_state;
1050         struct task_struct *oldowner = pi_state->owner;
1051         u32 uval, curval, newval;
1052         int ret, attempt = 0;
1053
1054         /* Owner died? */
1055         if (!pi_state->owner)
1056                 newtid |= FUTEX_OWNER_DIED;
1057
1058         /*
1059          * We are here either because we stole the rtmutex from the
1060          * pending owner or we are the pending owner which failed to
1061          * get the rtmutex. We have to replace the pending owner TID
1062          * in the user space variable. This must be atomic as we have
1063          * to preserve the owner died bit here.
1064          *
1065          * Note: We write the user space value _before_ changing the
1066          * pi_state because we can fault here. Imagine swapped out
1067          * pages or a fork, which was running right before we acquired
1068          * mmap_sem, that marked all the anonymous memory readonly for
1069          * cow.
1070          *
1071          * Modifying pi_state _before_ the user space value would
1072          * leave the pi_state in an inconsistent state when we fault
1073          * here, because we need to drop the hash bucket lock to
1074          * handle the fault. This might be observed in the PID check
1075          * in lookup_pi_state.
1076          */
1077 retry:
1078         if (get_futex_value_locked(&uval, uaddr))
1079                 goto handle_fault;
1080
1081         while (1) {
1082                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1083
1084                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1085
1086                 if (curval == -EFAULT)
1087                         goto handle_fault;
1088                 if (curval == uval)
1089                         break;
1090                 uval = curval;
1091         }
1092
1093         /*
1094          * We fixed up user space. Now we need to fix the pi_state
1095          * itself.
1096          */
1097         if (pi_state->owner != NULL) {
1098                 spin_lock_irq(&pi_state->owner->pi_lock);
1099                 WARN_ON(list_empty(&pi_state->list));
1100                 list_del_init(&pi_state->list);
1101                 spin_unlock_irq(&pi_state->owner->pi_lock);
1102         }
1103
1104         pi_state->owner = newowner;
1105
1106         spin_lock_irq(&newowner->pi_lock);
1107         WARN_ON(!list_empty(&pi_state->list));
1108         list_add(&pi_state->list, &newowner->pi_state_list);
1109         spin_unlock_irq(&newowner->pi_lock);
1110         return 0;
1111
1112         /*
1113          * To handle the page fault we need to drop the hash bucket
1114          * lock here. That gives the other task (either the pending
1115          * owner itself or the task which stole the rtmutex) the
1116          * chance to try the fixup of the pi_state. So once we are
1117          * back from handling the fault we need to check the pi_state
1118          * after reacquiring the hash bucket lock and before trying to
1119          * do another fixup. When the fixup has been done already we
1120          * simply return.
1121          */
1122 handle_fault:
1123         spin_unlock(q->lock_ptr);
1124
1125         ret = futex_handle_fault((unsigned long)uaddr, attempt++);
1126
1127         spin_lock(q->lock_ptr);
1128
1129         /*
1130          * Check if someone else fixed it for us:
1131          */
1132         if (pi_state->owner != oldowner)
1133                 return 0;
1134
1135         if (ret)
1136                 return ret;
1137
1138         goto retry;
1139 }
1140
1141 /*
1142  * In case we must use restart_block to restart a futex_wait,
1143  * we encode in the 'flags' shared capability
1144  */
1145 #define FLAGS_SHARED  1
1146
1147 static long futex_wait_restart(struct restart_block *restart);
1148
1149 static int futex_wait(u32 __user *uaddr, int fshared,
1150                       u32 val, ktime_t *abs_time, u32 bitset)
1151 {
1152         struct task_struct *curr = current;
1153         DECLARE_WAITQUEUE(wait, curr);
1154         struct futex_hash_bucket *hb;
1155         struct futex_q q;
1156         u32 uval;
1157         int ret;
1158         struct hrtimer_sleeper t;
1159         int rem = 0;
1160
1161         if (!bitset)
1162                 return -EINVAL;
1163
1164         q.pi_state = NULL;
1165         q.bitset = bitset;
1166  retry:
1167         q.key = FUTEX_KEY_INIT;
1168         ret = get_futex_key(uaddr, fshared, &q.key);
1169         if (unlikely(ret != 0))
1170                 goto out_release_sem;
1171
1172         hb = queue_lock(&q);
1173
1174         /*
1175          * Access the page AFTER the futex is queued.
1176          * Order is important:
1177          *
1178          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1179          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1180          *
1181          * The basic logical guarantee of a futex is that it blocks ONLY
1182          * if cond(var) is known to be true at the time of blocking, for
1183          * any cond.  If we queued after testing *uaddr, that would open
1184          * a race condition where we could block indefinitely with
1185          * cond(var) false, which would violate the guarantee.
1186          *
1187          * A consequence is that futex_wait() can return zero and absorb
1188          * a wakeup when *uaddr != val on entry to the syscall.  This is
1189          * rare, but normal.
1190          *
1191          * for shared futexes, we hold the mmap semaphore, so the mapping
1192          * cannot have changed since we looked it up in get_futex_key.
1193          */
1194         ret = get_futex_value_locked(&uval, uaddr);
1195
1196         if (unlikely(ret)) {
1197                 queue_unlock(&q, hb);
1198
1199                 ret = get_user(uval, uaddr);
1200
1201                 if (!ret)
1202                         goto retry;
1203                 return ret;
1204         }
1205         ret = -EWOULDBLOCK;
1206         if (uval != val)
1207                 goto out_unlock_release_sem;
1208
1209         /* Only actually queue if *uaddr contained val.  */
1210         queue_me(&q, hb);
1211
1212         /*
1213          * There might have been scheduling since the queue_me(), as we
1214          * cannot hold a spinlock across the get_user() in case it
1215          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1216          * queueing ourselves into the futex hash.  This code thus has to
1217          * rely on the futex_wake() code removing us from hash when it
1218          * wakes us up.
1219          */
1220
1221         /* add_wait_queue is the barrier after __set_current_state. */
1222         __set_current_state(TASK_INTERRUPTIBLE);
1223         add_wait_queue(&q.waiters, &wait);
1224         /*
1225          * !plist_node_empty() is safe here without any lock.
1226          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1227          */
1228         if (likely(!plist_node_empty(&q.list))) {
1229                 if (!abs_time)
1230                         schedule();
1231                 else {
1232                         hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC,
1233                                                 HRTIMER_MODE_ABS);
1234                         hrtimer_init_sleeper(&t, current);
1235                         t.timer.expires = *abs_time;
1236
1237                         hrtimer_start(&t.timer, t.timer.expires,
1238                                                 HRTIMER_MODE_ABS);
1239                         if (!hrtimer_active(&t.timer))
1240                                 t.task = NULL;
1241
1242                         /*
1243                          * the timer could have already expired, in which
1244                          * case current would be flagged for rescheduling.
1245                          * Don't bother calling schedule.
1246                          */
1247                         if (likely(t.task))
1248                                 schedule();
1249
1250                         hrtimer_cancel(&t.timer);
1251
1252                         /* Flag if a timeout occured */
1253                         rem = (t.task == NULL);
1254
1255                         destroy_hrtimer_on_stack(&t.timer);
1256                 }
1257         }
1258         __set_current_state(TASK_RUNNING);
1259
1260         /*
1261          * NOTE: we don't remove ourselves from the waitqueue because
1262          * we are the only user of it.
1263          */
1264
1265         /* If we were woken (and unqueued), we succeeded, whatever. */
1266         if (!unqueue_me(&q))
1267                 return 0;
1268         if (rem)
1269                 return -ETIMEDOUT;
1270
1271         /*
1272          * We expect signal_pending(current), but another thread may
1273          * have handled it for us already.
1274          */
1275         if (!abs_time)
1276                 return -ERESTARTSYS;
1277         else {
1278                 struct restart_block *restart;
1279                 restart = &current_thread_info()->restart_block;
1280                 restart->fn = futex_wait_restart;
1281                 restart->futex.uaddr = (u32 *)uaddr;
1282                 restart->futex.val = val;
1283                 restart->futex.time = abs_time->tv64;
1284                 restart->futex.bitset = bitset;
1285                 restart->futex.flags = 0;
1286
1287                 if (fshared)
1288                         restart->futex.flags |= FLAGS_SHARED;
1289                 return -ERESTART_RESTARTBLOCK;
1290         }
1291
1292  out_unlock_release_sem:
1293         queue_unlock(&q, hb);
1294
1295  out_release_sem:
1296         put_futex_key(fshared, &q.key);
1297         return ret;
1298 }
1299
1300
1301 static long futex_wait_restart(struct restart_block *restart)
1302 {
1303         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1304         int fshared = 0;
1305         ktime_t t;
1306
1307         t.tv64 = restart->futex.time;
1308         restart->fn = do_no_restart_syscall;
1309         if (restart->futex.flags & FLAGS_SHARED)
1310                 fshared = 1;
1311         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
1312                                 restart->futex.bitset);
1313 }
1314
1315
1316 /*
1317  * Userspace tried a 0 -> TID atomic transition of the futex value
1318  * and failed. The kernel side here does the whole locking operation:
1319  * if there are waiters then it will block, it does PI, etc. (Due to
1320  * races the kernel might see a 0 value of the futex too.)
1321  */
1322 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1323                          int detect, ktime_t *time, int trylock)
1324 {
1325         struct hrtimer_sleeper timeout, *to = NULL;
1326         struct task_struct *curr = current;
1327         struct futex_hash_bucket *hb;
1328         u32 uval, newval, curval;
1329         struct futex_q q;
1330         int ret, lock_taken, ownerdied = 0, attempt = 0;
1331
1332         if (refill_pi_state_cache())
1333                 return -ENOMEM;
1334
1335         if (time) {
1336                 to = &timeout;
1337                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1338                                       HRTIMER_MODE_ABS);
1339                 hrtimer_init_sleeper(to, current);
1340                 to->timer.expires = *time;
1341         }
1342
1343         q.pi_state = NULL;
1344  retry:
1345         q.key = FUTEX_KEY_INIT;
1346         ret = get_futex_key(uaddr, fshared, &q.key);
1347         if (unlikely(ret != 0))
1348                 goto out_release_sem;
1349
1350  retry_unlocked:
1351         hb = queue_lock(&q);
1352
1353  retry_locked:
1354         ret = lock_taken = 0;
1355
1356         /*
1357          * To avoid races, we attempt to take the lock here again
1358          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1359          * the locks. It will most likely not succeed.
1360          */
1361         newval = task_pid_vnr(current);
1362
1363         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1364
1365         if (unlikely(curval == -EFAULT))
1366                 goto uaddr_faulted;
1367
1368         /*
1369          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1370          * situation and we return success to user space.
1371          */
1372         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1373                 ret = -EDEADLK;
1374                 goto out_unlock_release_sem;
1375         }
1376
1377         /*
1378          * Surprise - we got the lock. Just return to userspace:
1379          */
1380         if (unlikely(!curval))
1381                 goto out_unlock_release_sem;
1382
1383         uval = curval;
1384
1385         /*
1386          * Set the WAITERS flag, so the owner will know it has someone
1387          * to wake at next unlock
1388          */
1389         newval = curval | FUTEX_WAITERS;
1390
1391         /*
1392          * There are two cases, where a futex might have no owner (the
1393          * owner TID is 0): OWNER_DIED. We take over the futex in this
1394          * case. We also do an unconditional take over, when the owner
1395          * of the futex died.
1396          *
1397          * This is safe as we are protected by the hash bucket lock !
1398          */
1399         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1400                 /* Keep the OWNER_DIED bit */
1401                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1402                 ownerdied = 0;
1403                 lock_taken = 1;
1404         }
1405
1406         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1407
1408         if (unlikely(curval == -EFAULT))
1409                 goto uaddr_faulted;
1410         if (unlikely(curval != uval))
1411                 goto retry_locked;
1412
1413         /*
1414          * We took the lock due to owner died take over.
1415          */
1416         if (unlikely(lock_taken))
1417                 goto out_unlock_release_sem;
1418
1419         /*
1420          * We dont have the lock. Look up the PI state (or create it if
1421          * we are the first waiter):
1422          */
1423         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1424
1425         if (unlikely(ret)) {
1426                 switch (ret) {
1427
1428                 case -EAGAIN:
1429                         /*
1430                          * Task is exiting and we just wait for the
1431                          * exit to complete.
1432                          */
1433                         queue_unlock(&q, hb);
1434                         cond_resched();
1435                         goto retry;
1436
1437                 case -ESRCH:
1438                         /*
1439                          * No owner found for this futex. Check if the
1440                          * OWNER_DIED bit is set to figure out whether
1441                          * this is a robust futex or not.
1442                          */
1443                         if (get_futex_value_locked(&curval, uaddr))
1444                                 goto uaddr_faulted;
1445
1446                         /*
1447                          * We simply start over in case of a robust
1448                          * futex. The code above will take the futex
1449                          * and return happy.
1450                          */
1451                         if (curval & FUTEX_OWNER_DIED) {
1452                                 ownerdied = 1;
1453                                 goto retry_locked;
1454                         }
1455                 default:
1456                         goto out_unlock_release_sem;
1457                 }
1458         }
1459
1460         /*
1461          * Only actually queue now that the atomic ops are done:
1462          */
1463         queue_me(&q, hb);
1464
1465         WARN_ON(!q.pi_state);
1466         /*
1467          * Block on the PI mutex:
1468          */
1469         if (!trylock)
1470                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1471         else {
1472                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1473                 /* Fixup the trylock return value: */
1474                 ret = ret ? 0 : -EWOULDBLOCK;
1475         }
1476
1477         spin_lock(q.lock_ptr);
1478
1479         if (!ret) {
1480                 /*
1481                  * Got the lock. We might not be the anticipated owner
1482                  * if we did a lock-steal - fix up the PI-state in
1483                  * that case:
1484                  */
1485                 if (q.pi_state->owner != curr)
1486                         ret = fixup_pi_state_owner(uaddr, &q, curr, fshared);
1487         } else {
1488                 /*
1489                  * Catch the rare case, where the lock was released
1490                  * when we were on the way back before we locked the
1491                  * hash bucket.
1492                  */
1493                 if (q.pi_state->owner == curr) {
1494                         /*
1495                          * Try to get the rt_mutex now. This might
1496                          * fail as some other task acquired the
1497                          * rt_mutex after we removed ourself from the
1498                          * rt_mutex waiters list.
1499                          */
1500                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1501                                 ret = 0;
1502                         else {
1503                                 /*
1504                                  * pi_state is incorrect, some other
1505                                  * task did a lock steal and we
1506                                  * returned due to timeout or signal
1507                                  * without taking the rt_mutex. Too
1508                                  * late. We can access the
1509                                  * rt_mutex_owner without locking, as
1510                                  * the other task is now blocked on
1511                                  * the hash bucket lock. Fix the state
1512                                  * up.
1513                                  */
1514                                 struct task_struct *owner;
1515                                 int res;
1516
1517                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1518                                 res = fixup_pi_state_owner(uaddr, &q, owner,
1519                                                            fshared);
1520
1521                                 /* propagate -EFAULT, if the fixup failed */
1522                                 if (res)
1523                                         ret = res;
1524                         }
1525                 } else {
1526                         /*
1527                          * Paranoia check. If we did not take the lock
1528                          * in the trylock above, then we should not be
1529                          * the owner of the rtmutex, neither the real
1530                          * nor the pending one:
1531                          */
1532                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1533                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1534                                        "pi-mutex: %p pi-state %p\n", ret,
1535                                        q.pi_state->pi_mutex.owner,
1536                                        q.pi_state->owner);
1537                 }
1538         }
1539
1540         /* Unqueue and drop the lock */
1541         unqueue_me_pi(&q);
1542
1543         if (to)
1544                 destroy_hrtimer_on_stack(&to->timer);
1545         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1546
1547  out_unlock_release_sem:
1548         queue_unlock(&q, hb);
1549
1550  out_release_sem:
1551         put_futex_key(fshared, &q.key);
1552         if (to)
1553                 destroy_hrtimer_on_stack(&to->timer);
1554         return ret;
1555
1556  uaddr_faulted:
1557         /*
1558          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1559          * non-atomically.  Therefore, if get_user below is not
1560          * enough, we need to handle the fault ourselves, while
1561          * still holding the mmap_sem.
1562          *
1563          * ... and hb->lock. :-) --ANK
1564          */
1565         queue_unlock(&q, hb);
1566
1567         if (attempt++) {
1568                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1569                 if (ret)
1570                         goto out_release_sem;
1571                 goto retry_unlocked;
1572         }
1573
1574         ret = get_user(uval, uaddr);
1575         if (!ret && (uval != -EFAULT))
1576                 goto retry;
1577
1578         if (to)
1579                 destroy_hrtimer_on_stack(&to->timer);
1580         return ret;
1581 }
1582
1583 /*
1584  * Userspace attempted a TID -> 0 atomic transition, and failed.
1585  * This is the in-kernel slowpath: we look up the PI state (if any),
1586  * and do the rt-mutex unlock.
1587  */
1588 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1589 {
1590         struct futex_hash_bucket *hb;
1591         struct futex_q *this, *next;
1592         u32 uval;
1593         struct plist_head *head;
1594         union futex_key key = FUTEX_KEY_INIT;
1595         int ret, attempt = 0;
1596
1597 retry:
1598         if (get_user(uval, uaddr))
1599                 return -EFAULT;
1600         /*
1601          * We release only a lock we actually own:
1602          */
1603         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1604                 return -EPERM;
1605
1606         ret = get_futex_key(uaddr, fshared, &key);
1607         if (unlikely(ret != 0))
1608                 goto out;
1609
1610         hb = hash_futex(&key);
1611 retry_unlocked:
1612         spin_lock(&hb->lock);
1613
1614         /*
1615          * To avoid races, try to do the TID -> 0 atomic transition
1616          * again. If it succeeds then we can return without waking
1617          * anyone else up:
1618          */
1619         if (!(uval & FUTEX_OWNER_DIED))
1620                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1621
1622
1623         if (unlikely(uval == -EFAULT))
1624                 goto pi_faulted;
1625         /*
1626          * Rare case: we managed to release the lock atomically,
1627          * no need to wake anyone else up:
1628          */
1629         if (unlikely(uval == task_pid_vnr(current)))
1630                 goto out_unlock;
1631
1632         /*
1633          * Ok, other tasks may need to be woken up - check waiters
1634          * and do the wakeup if necessary:
1635          */
1636         head = &hb->chain;
1637
1638         plist_for_each_entry_safe(this, next, head, list) {
1639                 if (!match_futex (&this->key, &key))
1640                         continue;
1641                 ret = wake_futex_pi(uaddr, uval, this);
1642                 /*
1643                  * The atomic access to the futex value
1644                  * generated a pagefault, so retry the
1645                  * user-access and the wakeup:
1646                  */
1647                 if (ret == -EFAULT)
1648                         goto pi_faulted;
1649                 goto out_unlock;
1650         }
1651         /*
1652          * No waiters - kernel unlocks the futex:
1653          */
1654         if (!(uval & FUTEX_OWNER_DIED)) {
1655                 ret = unlock_futex_pi(uaddr, uval);
1656                 if (ret == -EFAULT)
1657                         goto pi_faulted;
1658         }
1659
1660 out_unlock:
1661         spin_unlock(&hb->lock);
1662 out:
1663         put_futex_key(fshared, &key);
1664
1665         return ret;
1666
1667 pi_faulted:
1668         /*
1669          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1670          * non-atomically.  Therefore, if get_user below is not
1671          * enough, we need to handle the fault ourselves, while
1672          * still holding the mmap_sem.
1673          *
1674          * ... and hb->lock. --ANK
1675          */
1676         spin_unlock(&hb->lock);
1677
1678         if (attempt++) {
1679                 ret = futex_handle_fault((unsigned long)uaddr, attempt);
1680                 if (ret)
1681                         goto out;
1682                 uval = 0;
1683                 goto retry_unlocked;
1684         }
1685
1686         ret = get_user(uval, uaddr);
1687         if (!ret && (uval != -EFAULT))
1688                 goto retry;
1689
1690         return ret;
1691 }
1692
1693 /*
1694  * Support for robust futexes: the kernel cleans up held futexes at
1695  * thread exit time.
1696  *
1697  * Implementation: user-space maintains a per-thread list of locks it
1698  * is holding. Upon do_exit(), the kernel carefully walks this list,
1699  * and marks all locks that are owned by this thread with the
1700  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1701  * always manipulated with the lock held, so the list is private and
1702  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1703  * field, to allow the kernel to clean up if the thread dies after
1704  * acquiring the lock, but just before it could have added itself to
1705  * the list. There can only be one such pending lock.
1706  */
1707
1708 /**
1709  * sys_set_robust_list - set the robust-futex list head of a task
1710  * @head: pointer to the list-head
1711  * @len: length of the list-head, as userspace expects
1712  */
1713 asmlinkage long
1714 sys_set_robust_list(struct robust_list_head __user *head,
1715                     size_t len)
1716 {
1717         if (!futex_cmpxchg_enabled)
1718                 return -ENOSYS;
1719         /*
1720          * The kernel knows only one size for now:
1721          */
1722         if (unlikely(len != sizeof(*head)))
1723                 return -EINVAL;
1724
1725         current->robust_list = head;
1726
1727         return 0;
1728 }
1729
1730 /**
1731  * sys_get_robust_list - get the robust-futex list head of a task
1732  * @pid: pid of the process [zero for current task]
1733  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1734  * @len_ptr: pointer to a length field, the kernel fills in the header size
1735  */
1736 asmlinkage long
1737 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1738                     size_t __user *len_ptr)
1739 {
1740         struct robust_list_head __user *head;
1741         unsigned long ret;
1742
1743         if (!futex_cmpxchg_enabled)
1744                 return -ENOSYS;
1745
1746         if (!pid)
1747                 head = current->robust_list;
1748         else {
1749                 struct task_struct *p;
1750
1751                 ret = -ESRCH;
1752                 rcu_read_lock();
1753                 p = find_task_by_vpid(pid);
1754                 if (!p)
1755                         goto err_unlock;
1756                 ret = -EPERM;
1757                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1758                                 !capable(CAP_SYS_PTRACE))
1759                         goto err_unlock;
1760                 head = p->robust_list;
1761                 rcu_read_unlock();
1762         }
1763
1764         if (put_user(sizeof(*head), len_ptr))
1765                 return -EFAULT;
1766         return put_user(head, head_ptr);
1767
1768 err_unlock:
1769         rcu_read_unlock();
1770
1771         return ret;
1772 }
1773
1774 /*
1775  * Process a futex-list entry, check whether it's owned by the
1776  * dying task, and do notification if so:
1777  */
1778 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1779 {
1780         u32 uval, nval, mval;
1781
1782 retry:
1783         if (get_user(uval, uaddr))
1784                 return -1;
1785
1786         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1787                 /*
1788                  * Ok, this dying thread is truly holding a futex
1789                  * of interest. Set the OWNER_DIED bit atomically
1790                  * via cmpxchg, and if the value had FUTEX_WAITERS
1791                  * set, wake up a waiter (if any). (We have to do a
1792                  * futex_wake() even if OWNER_DIED is already set -
1793                  * to handle the rare but possible case of recursive
1794                  * thread-death.) The rest of the cleanup is done in
1795                  * userspace.
1796                  */
1797                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1798                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1799
1800                 if (nval == -EFAULT)
1801                         return -1;
1802
1803                 if (nval != uval)
1804                         goto retry;
1805
1806                 /*
1807                  * Wake robust non-PI futexes here. The wakeup of
1808                  * PI futexes happens in exit_pi_state():
1809                  */
1810                 if (!pi && (uval & FUTEX_WAITERS))
1811                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
1812         }
1813         return 0;
1814 }
1815
1816 /*
1817  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1818  */
1819 static inline int fetch_robust_entry(struct robust_list __user **entry,
1820                                      struct robust_list __user * __user *head,
1821                                      int *pi)
1822 {
1823         unsigned long uentry;
1824
1825         if (get_user(uentry, (unsigned long __user *)head))
1826                 return -EFAULT;
1827
1828         *entry = (void __user *)(uentry & ~1UL);
1829         *pi = uentry & 1;
1830
1831         return 0;
1832 }
1833
1834 /*
1835  * Walk curr->robust_list (very carefully, it's a userspace list!)
1836  * and mark any locks found there dead, and notify any waiters.
1837  *
1838  * We silently return on any sign of list-walking problem.
1839  */
1840 void exit_robust_list(struct task_struct *curr)
1841 {
1842         struct robust_list_head __user *head = curr->robust_list;
1843         struct robust_list __user *entry, *next_entry, *pending;
1844         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1845         unsigned long futex_offset;
1846         int rc;
1847
1848         if (!futex_cmpxchg_enabled)
1849                 return;
1850
1851         /*
1852          * Fetch the list head (which was registered earlier, via
1853          * sys_set_robust_list()):
1854          */
1855         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1856                 return;
1857         /*
1858          * Fetch the relative futex offset:
1859          */
1860         if (get_user(futex_offset, &head->futex_offset))
1861                 return;
1862         /*
1863          * Fetch any possibly pending lock-add first, and handle it
1864          * if it exists:
1865          */
1866         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1867                 return;
1868
1869         next_entry = NULL;      /* avoid warning with gcc */
1870         while (entry != &head->list) {
1871                 /*
1872                  * Fetch the next entry in the list before calling
1873                  * handle_futex_death:
1874                  */
1875                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
1876                 /*
1877                  * A pending lock might already be on the list, so
1878                  * don't process it twice:
1879                  */
1880                 if (entry != pending)
1881                         if (handle_futex_death((void __user *)entry + futex_offset,
1882                                                 curr, pi))
1883                                 return;
1884                 if (rc)
1885                         return;
1886                 entry = next_entry;
1887                 pi = next_pi;
1888                 /*
1889                  * Avoid excessively long or circular lists:
1890                  */
1891                 if (!--limit)
1892                         break;
1893
1894                 cond_resched();
1895         }
1896
1897         if (pending)
1898                 handle_futex_death((void __user *)pending + futex_offset,
1899                                    curr, pip);
1900 }
1901
1902 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1903                 u32 __user *uaddr2, u32 val2, u32 val3)
1904 {
1905         int ret = -ENOSYS;
1906         int cmd = op & FUTEX_CMD_MASK;
1907         int fshared = 0;
1908
1909         if (!(op & FUTEX_PRIVATE_FLAG))
1910                 fshared = 1;
1911
1912         switch (cmd) {
1913         case FUTEX_WAIT:
1914                 val3 = FUTEX_BITSET_MATCH_ANY;
1915         case FUTEX_WAIT_BITSET:
1916                 ret = futex_wait(uaddr, fshared, val, timeout, val3);
1917                 break;
1918         case FUTEX_WAKE:
1919                 val3 = FUTEX_BITSET_MATCH_ANY;
1920         case FUTEX_WAKE_BITSET:
1921                 ret = futex_wake(uaddr, fshared, val, val3);
1922                 break;
1923         case FUTEX_REQUEUE:
1924                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
1925                 break;
1926         case FUTEX_CMP_REQUEUE:
1927                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
1928                 break;
1929         case FUTEX_WAKE_OP:
1930                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
1931                 break;
1932         case FUTEX_LOCK_PI:
1933                 if (futex_cmpxchg_enabled)
1934                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
1935                 break;
1936         case FUTEX_UNLOCK_PI:
1937                 if (futex_cmpxchg_enabled)
1938                         ret = futex_unlock_pi(uaddr, fshared);
1939                 break;
1940         case FUTEX_TRYLOCK_PI:
1941                 if (futex_cmpxchg_enabled)
1942                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
1943                 break;
1944         default:
1945                 ret = -ENOSYS;
1946         }
1947         return ret;
1948 }
1949
1950
1951 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1952                           struct timespec __user *utime, u32 __user *uaddr2,
1953                           u32 val3)
1954 {
1955         struct timespec ts;
1956         ktime_t t, *tp = NULL;
1957         u32 val2 = 0;
1958         int cmd = op & FUTEX_CMD_MASK;
1959
1960         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
1961                       cmd == FUTEX_WAIT_BITSET)) {
1962                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
1963                         return -EFAULT;
1964                 if (!timespec_valid(&ts))
1965                         return -EINVAL;
1966
1967                 t = timespec_to_ktime(ts);
1968                 if (cmd == FUTEX_WAIT)
1969                         t = ktime_add_safe(ktime_get(), t);
1970                 tp = &t;
1971         }
1972         /*
1973          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
1974          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
1975          */
1976         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
1977             cmd == FUTEX_WAKE_OP)
1978                 val2 = (u32) (unsigned long) utime;
1979
1980         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
1981 }
1982
1983 static int __init futex_init(void)
1984 {
1985         u32 curval;
1986         int i;
1987
1988         /*
1989          * This will fail and we want it. Some arch implementations do
1990          * runtime detection of the futex_atomic_cmpxchg_inatomic()
1991          * functionality. We want to know that before we call in any
1992          * of the complex code paths. Also we want to prevent
1993          * registration of robust lists in that case. NULL is
1994          * guaranteed to fault and we get -EFAULT on functional
1995          * implementation, the non functional ones will return
1996          * -ENOSYS.
1997          */
1998         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
1999         if (curval == -EFAULT)
2000                 futex_cmpxchg_enabled = 1;
2001
2002         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2003                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2004                 spin_lock_init(&futex_queues[i].lock);
2005         }
2006
2007         return 0;
2008 }
2009 __initcall(futex_init);