futex: fix futex_wait_setup key handling
[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  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Priority Inheritance state:
73  */
74 struct futex_pi_state {
75         /*
76          * list of 'owned' pi_state instances - these have to be
77          * cleaned up in do_exit() if the task exits prematurely:
78          */
79         struct list_head list;
80
81         /*
82          * The PI object:
83          */
84         struct rt_mutex pi_mutex;
85
86         struct task_struct *owner;
87         atomic_t refcount;
88
89         union futex_key key;
90 };
91
92 /*
93  * We use this hashed waitqueue instead of a normal wait_queue_t, so
94  * we can wake only the relevant ones (hashed queues may be shared).
95  *
96  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
97  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
98  * The order of wakup is always to make the first condition true, then
99  * wake up q->waiter, then make the second condition true.
100  */
101 struct futex_q {
102         struct plist_node list;
103         /* There can only be a single waiter */
104         wait_queue_head_t waiter;
105
106         /* Which hash list lock to use: */
107         spinlock_t *lock_ptr;
108
109         /* Key which the futex is hashed on: */
110         union futex_key key;
111
112         /* Optional priority inheritance state: */
113         struct futex_pi_state *pi_state;
114         struct task_struct *task;
115
116         /* rt_waiter storage for requeue_pi: */
117         struct rt_mutex_waiter *rt_waiter;
118
119         /* Bitset for the optional bitmasked wakeup */
120         u32 bitset;
121 };
122
123 /*
124  * Hash buckets are shared by all the futex_keys that hash to the same
125  * location.  Each key may have multiple futex_q structures, one for each task
126  * waiting on a futex.
127  */
128 struct futex_hash_bucket {
129         spinlock_t lock;
130         struct plist_head chain;
131 };
132
133 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
134
135 /*
136  * We hash on the keys returned from get_futex_key (see below).
137  */
138 static struct futex_hash_bucket *hash_futex(union futex_key *key)
139 {
140         u32 hash = jhash2((u32*)&key->both.word,
141                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
142                           key->both.offset);
143         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
144 }
145
146 /*
147  * Return 1 if two futex_keys are equal, 0 otherwise.
148  */
149 static inline int match_futex(union futex_key *key1, union futex_key *key2)
150 {
151         return (key1->both.word == key2->both.word
152                 && key1->both.ptr == key2->both.ptr
153                 && key1->both.offset == key2->both.offset);
154 }
155
156 /*
157  * Take a reference to the resource addressed by a key.
158  * Can be called while holding spinlocks.
159  *
160  */
161 static void get_futex_key_refs(union futex_key *key)
162 {
163         if (!key->both.ptr)
164                 return;
165
166         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
167         case FUT_OFF_INODE:
168                 atomic_inc(&key->shared.inode->i_count);
169                 break;
170         case FUT_OFF_MMSHARED:
171                 atomic_inc(&key->private.mm->mm_count);
172                 break;
173         }
174 }
175
176 /*
177  * Drop a reference to the resource addressed by a key.
178  * The hash bucket spinlock must not be held.
179  */
180 static void drop_futex_key_refs(union futex_key *key)
181 {
182         if (!key->both.ptr) {
183                 /* If we're here then we tried to put a key we failed to get */
184                 WARN_ON_ONCE(1);
185                 return;
186         }
187
188         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
189         case FUT_OFF_INODE:
190                 iput(key->shared.inode);
191                 break;
192         case FUT_OFF_MMSHARED:
193                 mmdrop(key->private.mm);
194                 break;
195         }
196 }
197
198 /**
199  * get_futex_key - Get parameters which are the keys for a futex.
200  * @uaddr: virtual address of the futex
201  * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
202  * @key: address where result is stored.
203  *
204  * Returns a negative error code or 0
205  * The key words are stored in *key on success.
206  *
207  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
208  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
209  * We can usually work out the index without swapping in the page.
210  *
211  * lock_page() might sleep, the caller should not hold a spinlock.
212  */
213 static int get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
214 {
215         unsigned long address = (unsigned long)uaddr;
216         struct mm_struct *mm = current->mm;
217         struct page *page;
218         int err;
219
220         /*
221          * The futex address must be "naturally" aligned.
222          */
223         key->both.offset = address % PAGE_SIZE;
224         if (unlikely((address % sizeof(u32)) != 0))
225                 return -EINVAL;
226         address -= key->both.offset;
227
228         /*
229          * PROCESS_PRIVATE futexes are fast.
230          * As the mm cannot disappear under us and the 'key' only needs
231          * virtual address, we dont even have to find the underlying vma.
232          * Note : We do have to check 'uaddr' is a valid user address,
233          *        but access_ok() should be faster than find_vma()
234          */
235         if (!fshared) {
236                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
237                         return -EFAULT;
238                 key->private.mm = mm;
239                 key->private.address = address;
240                 get_futex_key_refs(key);
241                 return 0;
242         }
243
244 again:
245         err = get_user_pages_fast(address, 1, 0, &page);
246         if (err < 0)
247                 return err;
248
249         lock_page(page);
250         if (!page->mapping) {
251                 unlock_page(page);
252                 put_page(page);
253                 goto again;
254         }
255
256         /*
257          * Private mappings are handled in a simple way.
258          *
259          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
260          * it's a read-only handle, it's expected that futexes attach to
261          * the object not the particular process.
262          */
263         if (PageAnon(page)) {
264                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
265                 key->private.mm = mm;
266                 key->private.address = address;
267         } else {
268                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
269                 key->shared.inode = page->mapping->host;
270                 key->shared.pgoff = page->index;
271         }
272
273         get_futex_key_refs(key);
274
275         unlock_page(page);
276         put_page(page);
277         return 0;
278 }
279
280 static inline
281 void put_futex_key(int fshared, union futex_key *key)
282 {
283         drop_futex_key_refs(key);
284 }
285
286 /**
287  * futex_top_waiter() - Return the highest priority waiter on a futex
288  * @hb:     the hash bucket the futex_q's reside in
289  * @key:    the futex key (to distinguish it from other futex futex_q's)
290  *
291  * Must be called with the hb lock held.
292  */
293 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
294                                         union futex_key *key)
295 {
296         struct futex_q *this;
297
298         plist_for_each_entry(this, &hb->chain, list) {
299                 if (match_futex(&this->key, key))
300                         return this;
301         }
302         return NULL;
303 }
304
305 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
306 {
307         u32 curval;
308
309         pagefault_disable();
310         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
311         pagefault_enable();
312
313         return curval;
314 }
315
316 static int get_futex_value_locked(u32 *dest, u32 __user *from)
317 {
318         int ret;
319
320         pagefault_disable();
321         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
322         pagefault_enable();
323
324         return ret ? -EFAULT : 0;
325 }
326
327
328 /*
329  * PI code:
330  */
331 static int refill_pi_state_cache(void)
332 {
333         struct futex_pi_state *pi_state;
334
335         if (likely(current->pi_state_cache))
336                 return 0;
337
338         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
339
340         if (!pi_state)
341                 return -ENOMEM;
342
343         INIT_LIST_HEAD(&pi_state->list);
344         /* pi_mutex gets initialized later */
345         pi_state->owner = NULL;
346         atomic_set(&pi_state->refcount, 1);
347         pi_state->key = FUTEX_KEY_INIT;
348
349         current->pi_state_cache = pi_state;
350
351         return 0;
352 }
353
354 static struct futex_pi_state * alloc_pi_state(void)
355 {
356         struct futex_pi_state *pi_state = current->pi_state_cache;
357
358         WARN_ON(!pi_state);
359         current->pi_state_cache = NULL;
360
361         return pi_state;
362 }
363
364 static void free_pi_state(struct futex_pi_state *pi_state)
365 {
366         if (!atomic_dec_and_test(&pi_state->refcount))
367                 return;
368
369         /*
370          * If pi_state->owner is NULL, the owner is most probably dying
371          * and has cleaned up the pi_state already
372          */
373         if (pi_state->owner) {
374                 spin_lock_irq(&pi_state->owner->pi_lock);
375                 list_del_init(&pi_state->list);
376                 spin_unlock_irq(&pi_state->owner->pi_lock);
377
378                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
379         }
380
381         if (current->pi_state_cache)
382                 kfree(pi_state);
383         else {
384                 /*
385                  * pi_state->list is already empty.
386                  * clear pi_state->owner.
387                  * refcount is at 0 - put it back to 1.
388                  */
389                 pi_state->owner = NULL;
390                 atomic_set(&pi_state->refcount, 1);
391                 current->pi_state_cache = pi_state;
392         }
393 }
394
395 /*
396  * Look up the task based on what TID userspace gave us.
397  * We dont trust it.
398  */
399 static struct task_struct * futex_find_get_task(pid_t pid)
400 {
401         struct task_struct *p;
402         const struct cred *cred = current_cred(), *pcred;
403
404         rcu_read_lock();
405         p = find_task_by_vpid(pid);
406         if (!p) {
407                 p = ERR_PTR(-ESRCH);
408         } else {
409                 pcred = __task_cred(p);
410                 if (cred->euid != pcred->euid &&
411                     cred->euid != pcred->uid)
412                         p = ERR_PTR(-ESRCH);
413                 else
414                         get_task_struct(p);
415         }
416
417         rcu_read_unlock();
418
419         return p;
420 }
421
422 /*
423  * This task is holding PI mutexes at exit time => bad.
424  * Kernel cleans up PI-state, but userspace is likely hosed.
425  * (Robust-futex cleanup is separate and might save the day for userspace.)
426  */
427 void exit_pi_state_list(struct task_struct *curr)
428 {
429         struct list_head *next, *head = &curr->pi_state_list;
430         struct futex_pi_state *pi_state;
431         struct futex_hash_bucket *hb;
432         union futex_key key = FUTEX_KEY_INIT;
433
434         if (!futex_cmpxchg_enabled)
435                 return;
436         /*
437          * We are a ZOMBIE and nobody can enqueue itself on
438          * pi_state_list anymore, but we have to be careful
439          * versus waiters unqueueing themselves:
440          */
441         spin_lock_irq(&curr->pi_lock);
442         while (!list_empty(head)) {
443
444                 next = head->next;
445                 pi_state = list_entry(next, struct futex_pi_state, list);
446                 key = pi_state->key;
447                 hb = hash_futex(&key);
448                 spin_unlock_irq(&curr->pi_lock);
449
450                 spin_lock(&hb->lock);
451
452                 spin_lock_irq(&curr->pi_lock);
453                 /*
454                  * We dropped the pi-lock, so re-check whether this
455                  * task still owns the PI-state:
456                  */
457                 if (head->next != next) {
458                         spin_unlock(&hb->lock);
459                         continue;
460                 }
461
462                 WARN_ON(pi_state->owner != curr);
463                 WARN_ON(list_empty(&pi_state->list));
464                 list_del_init(&pi_state->list);
465                 pi_state->owner = NULL;
466                 spin_unlock_irq(&curr->pi_lock);
467
468                 rt_mutex_unlock(&pi_state->pi_mutex);
469
470                 spin_unlock(&hb->lock);
471
472                 spin_lock_irq(&curr->pi_lock);
473         }
474         spin_unlock_irq(&curr->pi_lock);
475 }
476
477 static int
478 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
479                 union futex_key *key, struct futex_pi_state **ps)
480 {
481         struct futex_pi_state *pi_state = NULL;
482         struct futex_q *this, *next;
483         struct plist_head *head;
484         struct task_struct *p;
485         pid_t pid = uval & FUTEX_TID_MASK;
486
487         head = &hb->chain;
488
489         plist_for_each_entry_safe(this, next, head, list) {
490                 if (match_futex(&this->key, key)) {
491                         /*
492                          * Another waiter already exists - bump up
493                          * the refcount and return its pi_state:
494                          */
495                         pi_state = this->pi_state;
496                         /*
497                          * Userspace might have messed up non PI and PI futexes
498                          */
499                         if (unlikely(!pi_state))
500                                 return -EINVAL;
501
502                         WARN_ON(!atomic_read(&pi_state->refcount));
503                         WARN_ON(pid && pi_state->owner &&
504                                 pi_state->owner->pid != pid);
505
506                         atomic_inc(&pi_state->refcount);
507                         *ps = pi_state;
508
509                         return 0;
510                 }
511         }
512
513         /*
514          * We are the first waiter - try to look up the real owner and attach
515          * the new pi_state to it, but bail out when TID = 0
516          */
517         if (!pid)
518                 return -ESRCH;
519         p = futex_find_get_task(pid);
520         if (IS_ERR(p))
521                 return PTR_ERR(p);
522
523         /*
524          * We need to look at the task state flags to figure out,
525          * whether the task is exiting. To protect against the do_exit
526          * change of the task flags, we do this protected by
527          * p->pi_lock:
528          */
529         spin_lock_irq(&p->pi_lock);
530         if (unlikely(p->flags & PF_EXITING)) {
531                 /*
532                  * The task is on the way out. When PF_EXITPIDONE is
533                  * set, we know that the task has finished the
534                  * cleanup:
535                  */
536                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
537
538                 spin_unlock_irq(&p->pi_lock);
539                 put_task_struct(p);
540                 return ret;
541         }
542
543         pi_state = alloc_pi_state();
544
545         /*
546          * Initialize the pi_mutex in locked state and make 'p'
547          * the owner of it:
548          */
549         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
550
551         /* Store the key for possible exit cleanups: */
552         pi_state->key = *key;
553
554         WARN_ON(!list_empty(&pi_state->list));
555         list_add(&pi_state->list, &p->pi_state_list);
556         pi_state->owner = p;
557         spin_unlock_irq(&p->pi_lock);
558
559         put_task_struct(p);
560
561         *ps = pi_state;
562
563         return 0;
564 }
565
566 /**
567  * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
568  * @uaddr:              the pi futex user address
569  * @hb:                 the pi futex hash bucket
570  * @key:                the futex key associated with uaddr and hb
571  * @ps:                 the pi_state pointer where we store the result of the
572  *                      lookup
573  * @task:               the task to perform the atomic lock work for.  This will
574  *                      be "current" except in the case of requeue pi.
575  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
576  *
577  * Returns:
578  *  0 - ready to wait
579  *  1 - acquired the lock
580  * <0 - error
581  *
582  * The hb->lock and futex_key refs shall be held by the caller.
583  */
584 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
585                                 union futex_key *key,
586                                 struct futex_pi_state **ps,
587                                 struct task_struct *task, int set_waiters)
588 {
589         int lock_taken, ret, ownerdied = 0;
590         u32 uval, newval, curval;
591
592 retry:
593         ret = lock_taken = 0;
594
595         /*
596          * To avoid races, we attempt to take the lock here again
597          * (by doing a 0 -> TID atomic cmpxchg), while holding all
598          * the locks. It will most likely not succeed.
599          */
600         newval = task_pid_vnr(task);
601         if (set_waiters)
602                 newval |= FUTEX_WAITERS;
603
604         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
605
606         if (unlikely(curval == -EFAULT))
607                 return -EFAULT;
608
609         /*
610          * Detect deadlocks.
611          */
612         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
613                 return -EDEADLK;
614
615         /*
616          * Surprise - we got the lock. Just return to userspace:
617          */
618         if (unlikely(!curval))
619                 return 1;
620
621         uval = curval;
622
623         /*
624          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
625          * to wake at the next unlock.
626          */
627         newval = curval | FUTEX_WAITERS;
628
629         /*
630          * There are two cases, where a futex might have no owner (the
631          * owner TID is 0): OWNER_DIED. We take over the futex in this
632          * case. We also do an unconditional take over, when the owner
633          * of the futex died.
634          *
635          * This is safe as we are protected by the hash bucket lock !
636          */
637         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
638                 /* Keep the OWNER_DIED bit */
639                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
640                 ownerdied = 0;
641                 lock_taken = 1;
642         }
643
644         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
645
646         if (unlikely(curval == -EFAULT))
647                 return -EFAULT;
648         if (unlikely(curval != uval))
649                 goto retry;
650
651         /*
652          * We took the lock due to owner died take over.
653          */
654         if (unlikely(lock_taken))
655                 return 1;
656
657         /*
658          * We dont have the lock. Look up the PI state (or create it if
659          * we are the first waiter):
660          */
661         ret = lookup_pi_state(uval, hb, key, ps);
662
663         if (unlikely(ret)) {
664                 switch (ret) {
665                 case -ESRCH:
666                         /*
667                          * No owner found for this futex. Check if the
668                          * OWNER_DIED bit is set to figure out whether
669                          * this is a robust futex or not.
670                          */
671                         if (get_futex_value_locked(&curval, uaddr))
672                                 return -EFAULT;
673
674                         /*
675                          * We simply start over in case of a robust
676                          * futex. The code above will take the futex
677                          * and return happy.
678                          */
679                         if (curval & FUTEX_OWNER_DIED) {
680                                 ownerdied = 1;
681                                 goto retry;
682                         }
683                 default:
684                         break;
685                 }
686         }
687
688         return ret;
689 }
690
691 /*
692  * The hash bucket lock must be held when this is called.
693  * Afterwards, the futex_q must not be accessed.
694  */
695 static void wake_futex(struct futex_q *q)
696 {
697         plist_del(&q->list, &q->list.plist);
698         /*
699          * The lock in wake_up_all() is a crucial memory barrier after the
700          * plist_del() and also before assigning to q->lock_ptr.
701          */
702         wake_up(&q->waiter);
703         /*
704          * The waiting task can free the futex_q as soon as this is written,
705          * without taking any locks.  This must come last.
706          *
707          * A memory barrier is required here to prevent the following store to
708          * lock_ptr from getting ahead of the wakeup. Clearing the lock at the
709          * end of wake_up() does not prevent this store from moving.
710          */
711         smp_wmb();
712         q->lock_ptr = NULL;
713 }
714
715 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
716 {
717         struct task_struct *new_owner;
718         struct futex_pi_state *pi_state = this->pi_state;
719         u32 curval, newval;
720
721         if (!pi_state)
722                 return -EINVAL;
723
724         spin_lock(&pi_state->pi_mutex.wait_lock);
725         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
726
727         /*
728          * This happens when we have stolen the lock and the original
729          * pending owner did not enqueue itself back on the rt_mutex.
730          * Thats not a tragedy. We know that way, that a lock waiter
731          * is on the fly. We make the futex_q waiter the pending owner.
732          */
733         if (!new_owner)
734                 new_owner = this->task;
735
736         /*
737          * We pass it to the next owner. (The WAITERS bit is always
738          * kept enabled while there is PI state around. We must also
739          * preserve the owner died bit.)
740          */
741         if (!(uval & FUTEX_OWNER_DIED)) {
742                 int ret = 0;
743
744                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
745
746                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
747
748                 if (curval == -EFAULT)
749                         ret = -EFAULT;
750                 else if (curval != uval)
751                         ret = -EINVAL;
752                 if (ret) {
753                         spin_unlock(&pi_state->pi_mutex.wait_lock);
754                         return ret;
755                 }
756         }
757
758         spin_lock_irq(&pi_state->owner->pi_lock);
759         WARN_ON(list_empty(&pi_state->list));
760         list_del_init(&pi_state->list);
761         spin_unlock_irq(&pi_state->owner->pi_lock);
762
763         spin_lock_irq(&new_owner->pi_lock);
764         WARN_ON(!list_empty(&pi_state->list));
765         list_add(&pi_state->list, &new_owner->pi_state_list);
766         pi_state->owner = new_owner;
767         spin_unlock_irq(&new_owner->pi_lock);
768
769         spin_unlock(&pi_state->pi_mutex.wait_lock);
770         rt_mutex_unlock(&pi_state->pi_mutex);
771
772         return 0;
773 }
774
775 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
776 {
777         u32 oldval;
778
779         /*
780          * There is no waiter, so we unlock the futex. The owner died
781          * bit has not to be preserved here. We are the owner:
782          */
783         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
784
785         if (oldval == -EFAULT)
786                 return oldval;
787         if (oldval != uval)
788                 return -EAGAIN;
789
790         return 0;
791 }
792
793 /*
794  * Express the locking dependencies for lockdep:
795  */
796 static inline void
797 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
798 {
799         if (hb1 <= hb2) {
800                 spin_lock(&hb1->lock);
801                 if (hb1 < hb2)
802                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
803         } else { /* hb1 > hb2 */
804                 spin_lock(&hb2->lock);
805                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
806         }
807 }
808
809 static inline void
810 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
811 {
812         spin_unlock(&hb1->lock);
813         if (hb1 != hb2)
814                 spin_unlock(&hb2->lock);
815 }
816
817 /*
818  * Wake up waiters matching bitset queued on this futex (uaddr).
819  */
820 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
821 {
822         struct futex_hash_bucket *hb;
823         struct futex_q *this, *next;
824         struct plist_head *head;
825         union futex_key key = FUTEX_KEY_INIT;
826         int ret;
827
828         if (!bitset)
829                 return -EINVAL;
830
831         ret = get_futex_key(uaddr, fshared, &key);
832         if (unlikely(ret != 0))
833                 goto out;
834
835         hb = hash_futex(&key);
836         spin_lock(&hb->lock);
837         head = &hb->chain;
838
839         plist_for_each_entry_safe(this, next, head, list) {
840                 if (match_futex (&this->key, &key)) {
841                         if (this->pi_state || this->rt_waiter) {
842                                 ret = -EINVAL;
843                                 break;
844                         }
845
846                         /* Check if one of the bits is set in both bitsets */
847                         if (!(this->bitset & bitset))
848                                 continue;
849
850                         wake_futex(this);
851                         if (++ret >= nr_wake)
852                                 break;
853                 }
854         }
855
856         spin_unlock(&hb->lock);
857         put_futex_key(fshared, &key);
858 out:
859         return ret;
860 }
861
862 /*
863  * Wake up all waiters hashed on the physical page that is mapped
864  * to this virtual address:
865  */
866 static int
867 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
868               int nr_wake, int nr_wake2, int op)
869 {
870         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
871         struct futex_hash_bucket *hb1, *hb2;
872         struct plist_head *head;
873         struct futex_q *this, *next;
874         int ret, op_ret;
875
876 retry:
877         ret = get_futex_key(uaddr1, fshared, &key1);
878         if (unlikely(ret != 0))
879                 goto out;
880         ret = get_futex_key(uaddr2, fshared, &key2);
881         if (unlikely(ret != 0))
882                 goto out_put_key1;
883
884         hb1 = hash_futex(&key1);
885         hb2 = hash_futex(&key2);
886
887         double_lock_hb(hb1, hb2);
888 retry_private:
889         op_ret = futex_atomic_op_inuser(op, uaddr2);
890         if (unlikely(op_ret < 0)) {
891                 u32 dummy;
892
893                 double_unlock_hb(hb1, hb2);
894
895 #ifndef CONFIG_MMU
896                 /*
897                  * we don't get EFAULT from MMU faults if we don't have an MMU,
898                  * but we might get them from range checking
899                  */
900                 ret = op_ret;
901                 goto out_put_keys;
902 #endif
903
904                 if (unlikely(op_ret != -EFAULT)) {
905                         ret = op_ret;
906                         goto out_put_keys;
907                 }
908
909                 ret = get_user(dummy, uaddr2);
910                 if (ret)
911                         goto out_put_keys;
912
913                 if (!fshared)
914                         goto retry_private;
915
916                 put_futex_key(fshared, &key2);
917                 put_futex_key(fshared, &key1);
918                 goto retry;
919         }
920
921         head = &hb1->chain;
922
923         plist_for_each_entry_safe(this, next, head, list) {
924                 if (match_futex (&this->key, &key1)) {
925                         wake_futex(this);
926                         if (++ret >= nr_wake)
927                                 break;
928                 }
929         }
930
931         if (op_ret > 0) {
932                 head = &hb2->chain;
933
934                 op_ret = 0;
935                 plist_for_each_entry_safe(this, next, head, list) {
936                         if (match_futex (&this->key, &key2)) {
937                                 wake_futex(this);
938                                 if (++op_ret >= nr_wake2)
939                                         break;
940                         }
941                 }
942                 ret += op_ret;
943         }
944
945         double_unlock_hb(hb1, hb2);
946 out_put_keys:
947         put_futex_key(fshared, &key2);
948 out_put_key1:
949         put_futex_key(fshared, &key1);
950 out:
951         return ret;
952 }
953
954 /**
955  * requeue_futex() - Requeue a futex_q from one hb to another
956  * @q:          the futex_q to requeue
957  * @hb1:        the source hash_bucket
958  * @hb2:        the target hash_bucket
959  * @key2:       the new key for the requeued futex_q
960  */
961 static inline
962 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
963                    struct futex_hash_bucket *hb2, union futex_key *key2)
964 {
965
966         /*
967          * If key1 and key2 hash to the same bucket, no need to
968          * requeue.
969          */
970         if (likely(&hb1->chain != &hb2->chain)) {
971                 plist_del(&q->list, &hb1->chain);
972                 plist_add(&q->list, &hb2->chain);
973                 q->lock_ptr = &hb2->lock;
974 #ifdef CONFIG_DEBUG_PI_LIST
975                 q->list.plist.lock = &hb2->lock;
976 #endif
977         }
978         get_futex_key_refs(key2);
979         q->key = *key2;
980 }
981
982 /**
983  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
984  * q:   the futex_q
985  * key: the key of the requeue target futex
986  *
987  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
988  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
989  * to the requeue target futex so the waiter can detect the wakeup on the right
990  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
991  * atomic lock acquisition.  Must be called with the q->lock_ptr held.
992  */
993 static inline
994 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key)
995 {
996         drop_futex_key_refs(&q->key);
997         get_futex_key_refs(key);
998         q->key = *key;
999
1000         WARN_ON(plist_node_empty(&q->list));
1001         plist_del(&q->list, &q->list.plist);
1002
1003         WARN_ON(!q->rt_waiter);
1004         q->rt_waiter = NULL;
1005
1006         wake_up(&q->waiter);
1007 }
1008
1009 /**
1010  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1011  * @pifutex:            the user address of the to futex
1012  * @hb1:                the from futex hash bucket, must be locked by the caller
1013  * @hb2:                the to futex hash bucket, must be locked by the caller
1014  * @key1:               the from futex key
1015  * @key2:               the to futex key
1016  * @ps:                 address to store the pi_state pointer
1017  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1018  *
1019  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1020  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1021  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1022  * hb1 and hb2 must be held by the caller.
1023  *
1024  * Returns:
1025  *  0 - failed to acquire the lock atomicly
1026  *  1 - acquired the lock
1027  * <0 - error
1028  */
1029 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1030                                  struct futex_hash_bucket *hb1,
1031                                  struct futex_hash_bucket *hb2,
1032                                  union futex_key *key1, union futex_key *key2,
1033                                  struct futex_pi_state **ps, int set_waiters)
1034 {
1035         struct futex_q *top_waiter = NULL;
1036         u32 curval;
1037         int ret;
1038
1039         if (get_futex_value_locked(&curval, pifutex))
1040                 return -EFAULT;
1041
1042         /*
1043          * Find the top_waiter and determine if there are additional waiters.
1044          * If the caller intends to requeue more than 1 waiter to pifutex,
1045          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1046          * as we have means to handle the possible fault.  If not, don't set
1047          * the bit unecessarily as it will force the subsequent unlock to enter
1048          * the kernel.
1049          */
1050         top_waiter = futex_top_waiter(hb1, key1);
1051
1052         /* There are no waiters, nothing for us to do. */
1053         if (!top_waiter)
1054                 return 0;
1055
1056         /*
1057          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1058          * the contended case or if set_waiters is 1.  The pi_state is returned
1059          * in ps in contended cases.
1060          */
1061         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1062                                    set_waiters);
1063         if (ret == 1)
1064                 requeue_pi_wake_futex(top_waiter, key2);
1065
1066         return ret;
1067 }
1068
1069 /**
1070  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1071  * uaddr1:      source futex user address
1072  * uaddr2:      target futex user address
1073  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1074  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1075  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1076  *              pi futex (pi to pi requeue is not supported)
1077  *
1078  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1079  * uaddr2 atomically on behalf of the top waiter.
1080  *
1081  * Returns:
1082  * >=0 - on success, the number of tasks requeued or woken
1083  *  <0 - on error
1084  */
1085 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1086                          int nr_wake, int nr_requeue, u32 *cmpval,
1087                          int requeue_pi)
1088 {
1089         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1090         int drop_count = 0, task_count = 0, ret;
1091         struct futex_pi_state *pi_state = NULL;
1092         struct futex_hash_bucket *hb1, *hb2;
1093         struct plist_head *head1;
1094         struct futex_q *this, *next;
1095         u32 curval2;
1096
1097         if (requeue_pi) {
1098                 /*
1099                  * requeue_pi requires a pi_state, try to allocate it now
1100                  * without any locks in case it fails.
1101                  */
1102                 if (refill_pi_state_cache())
1103                         return -ENOMEM;
1104                 /*
1105                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1106                  * + nr_requeue, since it acquires the rt_mutex prior to
1107                  * returning to userspace, so as to not leave the rt_mutex with
1108                  * waiters and no owner.  However, second and third wake-ups
1109                  * cannot be predicted as they involve race conditions with the
1110                  * first wake and a fault while looking up the pi_state.  Both
1111                  * pthread_cond_signal() and pthread_cond_broadcast() should
1112                  * use nr_wake=1.
1113                  */
1114                 if (nr_wake != 1)
1115                         return -EINVAL;
1116         }
1117
1118 retry:
1119         if (pi_state != NULL) {
1120                 /*
1121                  * We will have to lookup the pi_state again, so free this one
1122                  * to keep the accounting correct.
1123                  */
1124                 free_pi_state(pi_state);
1125                 pi_state = NULL;
1126         }
1127
1128         ret = get_futex_key(uaddr1, fshared, &key1);
1129         if (unlikely(ret != 0))
1130                 goto out;
1131         ret = get_futex_key(uaddr2, fshared, &key2);
1132         if (unlikely(ret != 0))
1133                 goto out_put_key1;
1134
1135         hb1 = hash_futex(&key1);
1136         hb2 = hash_futex(&key2);
1137
1138 retry_private:
1139         double_lock_hb(hb1, hb2);
1140
1141         if (likely(cmpval != NULL)) {
1142                 u32 curval;
1143
1144                 ret = get_futex_value_locked(&curval, uaddr1);
1145
1146                 if (unlikely(ret)) {
1147                         double_unlock_hb(hb1, hb2);
1148
1149                         ret = get_user(curval, uaddr1);
1150                         if (ret)
1151                                 goto out_put_keys;
1152
1153                         if (!fshared)
1154                                 goto retry_private;
1155
1156                         put_futex_key(fshared, &key2);
1157                         put_futex_key(fshared, &key1);
1158                         goto retry;
1159                 }
1160                 if (curval != *cmpval) {
1161                         ret = -EAGAIN;
1162                         goto out_unlock;
1163                 }
1164         }
1165
1166         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1167                 /*
1168                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1169                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1170                  * bit.  We force this here where we are able to easily handle
1171                  * faults rather in the requeue loop below.
1172                  */
1173                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1174                                                  &key2, &pi_state, nr_requeue);
1175
1176                 /*
1177                  * At this point the top_waiter has either taken uaddr2 or is
1178                  * waiting on it.  If the former, then the pi_state will not
1179                  * exist yet, look it up one more time to ensure we have a
1180                  * reference to it.
1181                  */
1182                 if (ret == 1) {
1183                         WARN_ON(pi_state);
1184                         task_count++;
1185                         ret = get_futex_value_locked(&curval2, uaddr2);
1186                         if (!ret)
1187                                 ret = lookup_pi_state(curval2, hb2, &key2,
1188                                                       &pi_state);
1189                 }
1190
1191                 switch (ret) {
1192                 case 0:
1193                         break;
1194                 case -EFAULT:
1195                         double_unlock_hb(hb1, hb2);
1196                         put_futex_key(fshared, &key2);
1197                         put_futex_key(fshared, &key1);
1198                         ret = get_user(curval2, uaddr2);
1199                         if (!ret)
1200                                 goto retry;
1201                         goto out;
1202                 case -EAGAIN:
1203                         /* The owner was exiting, try again. */
1204                         double_unlock_hb(hb1, hb2);
1205                         put_futex_key(fshared, &key2);
1206                         put_futex_key(fshared, &key1);
1207                         cond_resched();
1208                         goto retry;
1209                 default:
1210                         goto out_unlock;
1211                 }
1212         }
1213
1214         head1 = &hb1->chain;
1215         plist_for_each_entry_safe(this, next, head1, list) {
1216                 if (task_count - nr_wake >= nr_requeue)
1217                         break;
1218
1219                 if (!match_futex(&this->key, &key1))
1220                         continue;
1221
1222                 WARN_ON(!requeue_pi && this->rt_waiter);
1223                 WARN_ON(requeue_pi && !this->rt_waiter);
1224
1225                 /*
1226                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1227                  * lock, we already woke the top_waiter.  If not, it will be
1228                  * woken by futex_unlock_pi().
1229                  */
1230                 if (++task_count <= nr_wake && !requeue_pi) {
1231                         wake_futex(this);
1232                         continue;
1233                 }
1234
1235                 /*
1236                  * Requeue nr_requeue waiters and possibly one more in the case
1237                  * of requeue_pi if we couldn't acquire the lock atomically.
1238                  */
1239                 if (requeue_pi) {
1240                         /* Prepare the waiter to take the rt_mutex. */
1241                         atomic_inc(&pi_state->refcount);
1242                         this->pi_state = pi_state;
1243                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1244                                                         this->rt_waiter,
1245                                                         this->task, 1);
1246                         if (ret == 1) {
1247                                 /* We got the lock. */
1248                                 requeue_pi_wake_futex(this, &key2);
1249                                 continue;
1250                         } else if (ret) {
1251                                 /* -EDEADLK */
1252                                 this->pi_state = NULL;
1253                                 free_pi_state(pi_state);
1254                                 goto out_unlock;
1255                         }
1256                 }
1257                 requeue_futex(this, hb1, hb2, &key2);
1258                 drop_count++;
1259         }
1260
1261 out_unlock:
1262         double_unlock_hb(hb1, hb2);
1263
1264         /* drop_futex_key_refs() must be called outside the spinlocks. */
1265         while (--drop_count >= 0)
1266                 drop_futex_key_refs(&key1);
1267
1268 out_put_keys:
1269         put_futex_key(fshared, &key2);
1270 out_put_key1:
1271         put_futex_key(fshared, &key1);
1272 out:
1273         if (pi_state != NULL)
1274                 free_pi_state(pi_state);
1275         return ret ? ret : task_count;
1276 }
1277
1278 /* The key must be already stored in q->key. */
1279 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1280 {
1281         struct futex_hash_bucket *hb;
1282
1283         init_waitqueue_head(&q->waiter);
1284
1285         get_futex_key_refs(&q->key);
1286         hb = hash_futex(&q->key);
1287         q->lock_ptr = &hb->lock;
1288
1289         spin_lock(&hb->lock);
1290         return hb;
1291 }
1292
1293 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1294 {
1295         int prio;
1296
1297         /*
1298          * The priority used to register this element is
1299          * - either the real thread-priority for the real-time threads
1300          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1301          * - or MAX_RT_PRIO for non-RT threads.
1302          * Thus, all RT-threads are woken first in priority order, and
1303          * the others are woken last, in FIFO order.
1304          */
1305         prio = min(current->normal_prio, MAX_RT_PRIO);
1306
1307         plist_node_init(&q->list, prio);
1308 #ifdef CONFIG_DEBUG_PI_LIST
1309         q->list.plist.lock = &hb->lock;
1310 #endif
1311         plist_add(&q->list, &hb->chain);
1312         q->task = current;
1313         spin_unlock(&hb->lock);
1314 }
1315
1316 static inline void
1317 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1318 {
1319         spin_unlock(&hb->lock);
1320         drop_futex_key_refs(&q->key);
1321 }
1322
1323 /*
1324  * queue_me and unqueue_me must be called as a pair, each
1325  * exactly once.  They are called with the hashed spinlock held.
1326  */
1327
1328 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1329 static int unqueue_me(struct futex_q *q)
1330 {
1331         spinlock_t *lock_ptr;
1332         int ret = 0;
1333
1334         /* In the common case we don't take the spinlock, which is nice. */
1335 retry:
1336         lock_ptr = q->lock_ptr;
1337         barrier();
1338         if (lock_ptr != NULL) {
1339                 spin_lock(lock_ptr);
1340                 /*
1341                  * q->lock_ptr can change between reading it and
1342                  * spin_lock(), causing us to take the wrong lock.  This
1343                  * corrects the race condition.
1344                  *
1345                  * Reasoning goes like this: if we have the wrong lock,
1346                  * q->lock_ptr must have changed (maybe several times)
1347                  * between reading it and the spin_lock().  It can
1348                  * change again after the spin_lock() but only if it was
1349                  * already changed before the spin_lock().  It cannot,
1350                  * however, change back to the original value.  Therefore
1351                  * we can detect whether we acquired the correct lock.
1352                  */
1353                 if (unlikely(lock_ptr != q->lock_ptr)) {
1354                         spin_unlock(lock_ptr);
1355                         goto retry;
1356                 }
1357                 WARN_ON(plist_node_empty(&q->list));
1358                 plist_del(&q->list, &q->list.plist);
1359
1360                 BUG_ON(q->pi_state);
1361
1362                 spin_unlock(lock_ptr);
1363                 ret = 1;
1364         }
1365
1366         drop_futex_key_refs(&q->key);
1367         return ret;
1368 }
1369
1370 /*
1371  * PI futexes can not be requeued and must remove themself from the
1372  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1373  * and dropped here.
1374  */
1375 static void unqueue_me_pi(struct futex_q *q)
1376 {
1377         WARN_ON(plist_node_empty(&q->list));
1378         plist_del(&q->list, &q->list.plist);
1379
1380         BUG_ON(!q->pi_state);
1381         free_pi_state(q->pi_state);
1382         q->pi_state = NULL;
1383
1384         spin_unlock(q->lock_ptr);
1385
1386         drop_futex_key_refs(&q->key);
1387 }
1388
1389 /*
1390  * Fixup the pi_state owner with the new owner.
1391  *
1392  * Must be called with hash bucket lock held and mm->sem held for non
1393  * private futexes.
1394  */
1395 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1396                                 struct task_struct *newowner, int fshared)
1397 {
1398         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1399         struct futex_pi_state *pi_state = q->pi_state;
1400         struct task_struct *oldowner = pi_state->owner;
1401         u32 uval, curval, newval;
1402         int ret;
1403
1404         /* Owner died? */
1405         if (!pi_state->owner)
1406                 newtid |= FUTEX_OWNER_DIED;
1407
1408         /*
1409          * We are here either because we stole the rtmutex from the
1410          * pending owner or we are the pending owner which failed to
1411          * get the rtmutex. We have to replace the pending owner TID
1412          * in the user space variable. This must be atomic as we have
1413          * to preserve the owner died bit here.
1414          *
1415          * Note: We write the user space value _before_ changing the pi_state
1416          * because we can fault here. Imagine swapped out pages or a fork
1417          * that marked all the anonymous memory readonly for cow.
1418          *
1419          * Modifying pi_state _before_ the user space value would
1420          * leave the pi_state in an inconsistent state when we fault
1421          * here, because we need to drop the hash bucket lock to
1422          * handle the fault. This might be observed in the PID check
1423          * in lookup_pi_state.
1424          */
1425 retry:
1426         if (get_futex_value_locked(&uval, uaddr))
1427                 goto handle_fault;
1428
1429         while (1) {
1430                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1431
1432                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1433
1434                 if (curval == -EFAULT)
1435                         goto handle_fault;
1436                 if (curval == uval)
1437                         break;
1438                 uval = curval;
1439         }
1440
1441         /*
1442          * We fixed up user space. Now we need to fix the pi_state
1443          * itself.
1444          */
1445         if (pi_state->owner != NULL) {
1446                 spin_lock_irq(&pi_state->owner->pi_lock);
1447                 WARN_ON(list_empty(&pi_state->list));
1448                 list_del_init(&pi_state->list);
1449                 spin_unlock_irq(&pi_state->owner->pi_lock);
1450         }
1451
1452         pi_state->owner = newowner;
1453
1454         spin_lock_irq(&newowner->pi_lock);
1455         WARN_ON(!list_empty(&pi_state->list));
1456         list_add(&pi_state->list, &newowner->pi_state_list);
1457         spin_unlock_irq(&newowner->pi_lock);
1458         return 0;
1459
1460         /*
1461          * To handle the page fault we need to drop the hash bucket
1462          * lock here. That gives the other task (either the pending
1463          * owner itself or the task which stole the rtmutex) the
1464          * chance to try the fixup of the pi_state. So once we are
1465          * back from handling the fault we need to check the pi_state
1466          * after reacquiring the hash bucket lock and before trying to
1467          * do another fixup. When the fixup has been done already we
1468          * simply return.
1469          */
1470 handle_fault:
1471         spin_unlock(q->lock_ptr);
1472
1473         ret = get_user(uval, uaddr);
1474
1475         spin_lock(q->lock_ptr);
1476
1477         /*
1478          * Check if someone else fixed it for us:
1479          */
1480         if (pi_state->owner != oldowner)
1481                 return 0;
1482
1483         if (ret)
1484                 return ret;
1485
1486         goto retry;
1487 }
1488
1489 /*
1490  * In case we must use restart_block to restart a futex_wait,
1491  * we encode in the 'flags' shared capability
1492  */
1493 #define FLAGS_SHARED            0x01
1494 #define FLAGS_CLOCKRT           0x02
1495 #define FLAGS_HAS_TIMEOUT       0x04
1496
1497 static long futex_wait_restart(struct restart_block *restart);
1498 static long futex_lock_pi_restart(struct restart_block *restart);
1499
1500 /**
1501  * fixup_owner() - Post lock pi_state and corner case management
1502  * @uaddr:      user address of the futex
1503  * @fshared:    whether the futex is shared (1) or not (0)
1504  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1505  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1506  *
1507  * After attempting to lock an rt_mutex, this function is called to cleanup
1508  * the pi_state owner as well as handle race conditions that may allow us to
1509  * acquire the lock. Must be called with the hb lock held.
1510  *
1511  * Returns:
1512  *  1 - success, lock taken
1513  *  0 - success, lock not taken
1514  * <0 - on error (-EFAULT)
1515  */
1516 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1517                        int locked)
1518 {
1519         struct task_struct *owner;
1520         int ret = 0;
1521
1522         if (locked) {
1523                 /*
1524                  * Got the lock. We might not be the anticipated owner if we
1525                  * did a lock-steal - fix up the PI-state in that case:
1526                  */
1527                 if (q->pi_state->owner != current)
1528                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1529                 goto out;
1530         }
1531
1532         /*
1533          * Catch the rare case, where the lock was released when we were on the
1534          * way back before we locked the hash bucket.
1535          */
1536         if (q->pi_state->owner == current) {
1537                 /*
1538                  * Try to get the rt_mutex now. This might fail as some other
1539                  * task acquired the rt_mutex after we removed ourself from the
1540                  * rt_mutex waiters list.
1541                  */
1542                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1543                         locked = 1;
1544                         goto out;
1545                 }
1546
1547                 /*
1548                  * pi_state is incorrect, some other task did a lock steal and
1549                  * we returned due to timeout or signal without taking the
1550                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1551                  * locking, as the other task is now blocked on the hash bucket
1552                  * lock. Fix the state up.
1553                  */
1554                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1555                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1556                 goto out;
1557         }
1558
1559         /*
1560          * Paranoia check. If we did not take the lock, then we should not be
1561          * the owner, nor the pending owner, of the rt_mutex.
1562          */
1563         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1564                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1565                                 "pi-state %p\n", ret,
1566                                 q->pi_state->pi_mutex.owner,
1567                                 q->pi_state->owner);
1568
1569 out:
1570         return ret ? ret : locked;
1571 }
1572
1573 /**
1574  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1575  * @hb:         the futex hash bucket, must be locked by the caller
1576  * @q:          the futex_q to queue up on
1577  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1578  * @wait:       the wait_queue to add to the futex_q after queueing in the hb
1579  */
1580 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1581                                 struct hrtimer_sleeper *timeout,
1582                                 wait_queue_t *wait)
1583 {
1584         queue_me(q, hb);
1585
1586         /*
1587          * There might have been scheduling since the queue_me(), as we
1588          * cannot hold a spinlock across the get_user() in case it
1589          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1590          * queueing ourselves into the futex hash.  This code thus has to
1591          * rely on the futex_wake() code removing us from hash when it
1592          * wakes us up.
1593          */
1594
1595         /* add_wait_queue is the barrier after __set_current_state. */
1596         __set_current_state(TASK_INTERRUPTIBLE);
1597
1598         /*
1599          * Add current as the futex_q waiter.  We don't remove ourselves from
1600          * the wait_queue because we are the only user of it.
1601          */
1602         add_wait_queue(&q->waiter, wait);
1603
1604         /* Arm the timer */
1605         if (timeout) {
1606                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1607                 if (!hrtimer_active(&timeout->timer))
1608                         timeout->task = NULL;
1609         }
1610
1611         /*
1612          * !plist_node_empty() is safe here without any lock.
1613          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1614          */
1615         if (likely(!plist_node_empty(&q->list))) {
1616                 /*
1617                  * If the timer has already expired, current will already be
1618                  * flagged for rescheduling. Only call schedule if there
1619                  * is no timeout, or if it has yet to expire.
1620                  */
1621                 if (!timeout || timeout->task)
1622                         schedule();
1623         }
1624         __set_current_state(TASK_RUNNING);
1625 }
1626
1627 /**
1628  * futex_wait_setup() - Prepare to wait on a futex
1629  * @uaddr:      the futex userspace address
1630  * @val:        the expected value
1631  * @fshared:    whether the futex is shared (1) or not (0)
1632  * @q:          the associated futex_q
1633  * @hb:         storage for hash_bucket pointer to be returned to caller
1634  *
1635  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1636  * compare it with the expected value.  Handle atomic faults internally.
1637  * Return with the hb lock held and a q.key reference on success, and unlocked
1638  * with no q.key reference on failure.
1639  *
1640  * Returns:
1641  *  0 - uaddr contains val and hb has been locked
1642  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1643  */
1644 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1645                            struct futex_q *q, struct futex_hash_bucket **hb)
1646 {
1647         u32 uval;
1648         int ret;
1649
1650         /*
1651          * Access the page AFTER the hash-bucket is locked.
1652          * Order is important:
1653          *
1654          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1655          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1656          *
1657          * The basic logical guarantee of a futex is that it blocks ONLY
1658          * if cond(var) is known to be true at the time of blocking, for
1659          * any cond.  If we queued after testing *uaddr, that would open
1660          * a race condition where we could block indefinitely with
1661          * cond(var) false, which would violate the guarantee.
1662          *
1663          * A consequence is that futex_wait() can return zero and absorb
1664          * a wakeup when *uaddr != val on entry to the syscall.  This is
1665          * rare, but normal.
1666          */
1667 retry:
1668         q->key = FUTEX_KEY_INIT;
1669         ret = get_futex_key(uaddr, fshared, &q->key);
1670         if (unlikely(ret != 0))
1671                 return ret;
1672
1673 retry_private:
1674         *hb = queue_lock(q);
1675
1676         ret = get_futex_value_locked(&uval, uaddr);
1677
1678         if (ret) {
1679                 queue_unlock(q, *hb);
1680
1681                 ret = get_user(uval, uaddr);
1682                 if (ret)
1683                         goto out;
1684
1685                 if (!fshared)
1686                         goto retry_private;
1687
1688                 put_futex_key(fshared, &q->key);
1689                 goto retry;
1690         }
1691
1692         if (uval != val) {
1693                 queue_unlock(q, *hb);
1694                 ret = -EWOULDBLOCK;
1695         }
1696
1697 out:
1698         if (ret)
1699                 put_futex_key(fshared, &q->key);
1700         return ret;
1701 }
1702
1703 static int futex_wait(u32 __user *uaddr, int fshared,
1704                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1705 {
1706         struct hrtimer_sleeper timeout, *to = NULL;
1707         DECLARE_WAITQUEUE(wait, current);
1708         struct restart_block *restart;
1709         struct futex_hash_bucket *hb;
1710         struct futex_q q;
1711         int ret;
1712
1713         if (!bitset)
1714                 return -EINVAL;
1715
1716         q.pi_state = NULL;
1717         q.bitset = bitset;
1718         q.rt_waiter = NULL;
1719
1720         if (abs_time) {
1721                 to = &timeout;
1722
1723                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1724                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1725                 hrtimer_init_sleeper(to, current);
1726                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1727                                              current->timer_slack_ns);
1728         }
1729
1730         /* Prepare to wait on uaddr. */
1731         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1732         if (ret)
1733                 goto out;
1734
1735         /* queue_me and wait for wakeup, timeout, or a signal. */
1736         futex_wait_queue_me(hb, &q, to, &wait);
1737
1738         /* If we were woken (and unqueued), we succeeded, whatever. */
1739         ret = 0;
1740         if (!unqueue_me(&q))
1741                 goto out_put_key;
1742         ret = -ETIMEDOUT;
1743         if (to && !to->task)
1744                 goto out_put_key;
1745
1746         /*
1747          * We expect signal_pending(current), but another thread may
1748          * have handled it for us already.
1749          */
1750         ret = -ERESTARTSYS;
1751         if (!abs_time)
1752                 goto out_put_key;
1753
1754         restart = &current_thread_info()->restart_block;
1755         restart->fn = futex_wait_restart;
1756         restart->futex.uaddr = (u32 *)uaddr;
1757         restart->futex.val = val;
1758         restart->futex.time = abs_time->tv64;
1759         restart->futex.bitset = bitset;
1760         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1761
1762         if (fshared)
1763                 restart->futex.flags |= FLAGS_SHARED;
1764         if (clockrt)
1765                 restart->futex.flags |= FLAGS_CLOCKRT;
1766
1767         ret = -ERESTART_RESTARTBLOCK;
1768
1769 out_put_key:
1770         put_futex_key(fshared, &q.key);
1771 out:
1772         if (to) {
1773                 hrtimer_cancel(&to->timer);
1774                 destroy_hrtimer_on_stack(&to->timer);
1775         }
1776         return ret;
1777 }
1778
1779
1780 static long futex_wait_restart(struct restart_block *restart)
1781 {
1782         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1783         int fshared = 0;
1784         ktime_t t, *tp = NULL;
1785
1786         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1787                 t.tv64 = restart->futex.time;
1788                 tp = &t;
1789         }
1790         restart->fn = do_no_restart_syscall;
1791         if (restart->futex.flags & FLAGS_SHARED)
1792                 fshared = 1;
1793         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1794                                 restart->futex.bitset,
1795                                 restart->futex.flags & FLAGS_CLOCKRT);
1796 }
1797
1798
1799 /*
1800  * Userspace tried a 0 -> TID atomic transition of the futex value
1801  * and failed. The kernel side here does the whole locking operation:
1802  * if there are waiters then it will block, it does PI, etc. (Due to
1803  * races the kernel might see a 0 value of the futex too.)
1804  */
1805 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1806                          int detect, ktime_t *time, int trylock)
1807 {
1808         struct hrtimer_sleeper timeout, *to = NULL;
1809         struct futex_hash_bucket *hb;
1810         u32 uval;
1811         struct futex_q q;
1812         int res, ret;
1813
1814         if (refill_pi_state_cache())
1815                 return -ENOMEM;
1816
1817         if (time) {
1818                 to = &timeout;
1819                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1820                                       HRTIMER_MODE_ABS);
1821                 hrtimer_init_sleeper(to, current);
1822                 hrtimer_set_expires(&to->timer, *time);
1823         }
1824
1825         q.pi_state = NULL;
1826         q.rt_waiter = NULL;
1827 retry:
1828         q.key = FUTEX_KEY_INIT;
1829         ret = get_futex_key(uaddr, fshared, &q.key);
1830         if (unlikely(ret != 0))
1831                 goto out;
1832
1833 retry_private:
1834         hb = queue_lock(&q);
1835
1836         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1837         if (unlikely(ret)) {
1838                 switch (ret) {
1839                 case 1:
1840                         /* We got the lock. */
1841                         ret = 0;
1842                         goto out_unlock_put_key;
1843                 case -EFAULT:
1844                         goto uaddr_faulted;
1845                 case -EAGAIN:
1846                         /*
1847                          * Task is exiting and we just wait for the
1848                          * exit to complete.
1849                          */
1850                         queue_unlock(&q, hb);
1851                         put_futex_key(fshared, &q.key);
1852                         cond_resched();
1853                         goto retry;
1854                 default:
1855                         goto out_unlock_put_key;
1856                 }
1857         }
1858
1859         /*
1860          * Only actually queue now that the atomic ops are done:
1861          */
1862         queue_me(&q, hb);
1863
1864         WARN_ON(!q.pi_state);
1865         /*
1866          * Block on the PI mutex:
1867          */
1868         if (!trylock)
1869                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1870         else {
1871                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1872                 /* Fixup the trylock return value: */
1873                 ret = ret ? 0 : -EWOULDBLOCK;
1874         }
1875
1876         spin_lock(q.lock_ptr);
1877         /*
1878          * Fixup the pi_state owner and possibly acquire the lock if we
1879          * haven't already.
1880          */
1881         res = fixup_owner(uaddr, fshared, &q, !ret);
1882         /*
1883          * If fixup_owner() returned an error, proprogate that.  If it acquired
1884          * the lock, clear our -ETIMEDOUT or -EINTR.
1885          */
1886         if (res)
1887                 ret = (res < 0) ? res : 0;
1888
1889         /*
1890          * If fixup_owner() faulted and was unable to handle the fault, unlock
1891          * it and return the fault to userspace.
1892          */
1893         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1894                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1895
1896         /* Unqueue and drop the lock */
1897         unqueue_me_pi(&q);
1898
1899         goto out;
1900
1901 out_unlock_put_key:
1902         queue_unlock(&q, hb);
1903
1904 out_put_key:
1905         put_futex_key(fshared, &q.key);
1906 out:
1907         if (to)
1908                 destroy_hrtimer_on_stack(&to->timer);
1909         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1910
1911 uaddr_faulted:
1912         /*
1913          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1914          * atomically.  Therefore, if we continue to fault after get_user()
1915          * below, we need to handle the fault ourselves, while still holding
1916          * the mmap_sem.  This can occur if the uaddr is under contention as
1917          * we have to drop the mmap_sem in order to call get_user().
1918          */
1919         queue_unlock(&q, hb);
1920
1921         ret = get_user(uval, uaddr);
1922         if (ret)
1923                 goto out_put_key;
1924
1925         if (!fshared)
1926                 goto retry_private;
1927
1928         put_futex_key(fshared, &q.key);
1929         goto retry;
1930 }
1931
1932 static long futex_lock_pi_restart(struct restart_block *restart)
1933 {
1934         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1935         ktime_t t, *tp = NULL;
1936         int fshared = restart->futex.flags & FLAGS_SHARED;
1937
1938         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1939                 t.tv64 = restart->futex.time;
1940                 tp = &t;
1941         }
1942         restart->fn = do_no_restart_syscall;
1943
1944         return (long)futex_lock_pi(uaddr, fshared, restart->futex.val, tp, 0);
1945 }
1946
1947 /*
1948  * Userspace attempted a TID -> 0 atomic transition, and failed.
1949  * This is the in-kernel slowpath: we look up the PI state (if any),
1950  * and do the rt-mutex unlock.
1951  */
1952 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1953 {
1954         struct futex_hash_bucket *hb;
1955         struct futex_q *this, *next;
1956         u32 uval;
1957         struct plist_head *head;
1958         union futex_key key = FUTEX_KEY_INIT;
1959         int ret;
1960
1961 retry:
1962         if (get_user(uval, uaddr))
1963                 return -EFAULT;
1964         /*
1965          * We release only a lock we actually own:
1966          */
1967         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1968                 return -EPERM;
1969
1970         ret = get_futex_key(uaddr, fshared, &key);
1971         if (unlikely(ret != 0))
1972                 goto out;
1973
1974         hb = hash_futex(&key);
1975         spin_lock(&hb->lock);
1976
1977         /*
1978          * To avoid races, try to do the TID -> 0 atomic transition
1979          * again. If it succeeds then we can return without waking
1980          * anyone else up:
1981          */
1982         if (!(uval & FUTEX_OWNER_DIED))
1983                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1984
1985
1986         if (unlikely(uval == -EFAULT))
1987                 goto pi_faulted;
1988         /*
1989          * Rare case: we managed to release the lock atomically,
1990          * no need to wake anyone else up:
1991          */
1992         if (unlikely(uval == task_pid_vnr(current)))
1993                 goto out_unlock;
1994
1995         /*
1996          * Ok, other tasks may need to be woken up - check waiters
1997          * and do the wakeup if necessary:
1998          */
1999         head = &hb->chain;
2000
2001         plist_for_each_entry_safe(this, next, head, list) {
2002                 if (!match_futex (&this->key, &key))
2003                         continue;
2004                 ret = wake_futex_pi(uaddr, uval, this);
2005                 /*
2006                  * The atomic access to the futex value
2007                  * generated a pagefault, so retry the
2008                  * user-access and the wakeup:
2009                  */
2010                 if (ret == -EFAULT)
2011                         goto pi_faulted;
2012                 goto out_unlock;
2013         }
2014         /*
2015          * No waiters - kernel unlocks the futex:
2016          */
2017         if (!(uval & FUTEX_OWNER_DIED)) {
2018                 ret = unlock_futex_pi(uaddr, uval);
2019                 if (ret == -EFAULT)
2020                         goto pi_faulted;
2021         }
2022
2023 out_unlock:
2024         spin_unlock(&hb->lock);
2025         put_futex_key(fshared, &key);
2026
2027 out:
2028         return ret;
2029
2030 pi_faulted:
2031         /*
2032          * We have to r/w  *(int __user *)uaddr, and we have to modify it
2033          * atomically.  Therefore, if we continue to fault after get_user()
2034          * below, we need to handle the fault ourselves, while still holding
2035          * the mmap_sem.  This can occur if the uaddr is under contention as
2036          * we have to drop the mmap_sem in order to call get_user().
2037          */
2038         spin_unlock(&hb->lock);
2039         put_futex_key(fshared, &key);
2040
2041         ret = get_user(uval, uaddr);
2042         if (!ret)
2043                 goto retry;
2044
2045         return ret;
2046 }
2047
2048 /**
2049  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2050  * @hb:         the hash_bucket futex_q was original enqueued on
2051  * @q:          the futex_q woken while waiting to be requeued
2052  * @key2:       the futex_key of the requeue target futex
2053  * @timeout:    the timeout associated with the wait (NULL if none)
2054  *
2055  * Detect if the task was woken on the initial futex as opposed to the requeue
2056  * target futex.  If so, determine if it was a timeout or a signal that caused
2057  * the wakeup and return the appropriate error code to the caller.  Must be
2058  * called with the hb lock held.
2059  *
2060  * Returns
2061  *  0 - no early wakeup detected
2062  * <0 - -ETIMEDOUT or -ERESTARTSYS (FIXME: or ERESTARTNOINTR?)
2063  */
2064 static inline
2065 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2066                                    struct futex_q *q, union futex_key *key2,
2067                                    struct hrtimer_sleeper *timeout)
2068 {
2069         int ret = 0;
2070
2071         /*
2072          * With the hb lock held, we avoid races while we process the wakeup.
2073          * We only need to hold hb (and not hb2) to ensure atomicity as the
2074          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2075          * It can't be requeued from uaddr2 to something else since we don't
2076          * support a PI aware source futex for requeue.
2077          */
2078         if (!match_futex(&q->key, key2)) {
2079                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2080                 /*
2081                  * We were woken prior to requeue by a timeout or a signal.
2082                  * Unqueue the futex_q and determine which it was.
2083                  */
2084                 plist_del(&q->list, &q->list.plist);
2085                 drop_futex_key_refs(&q->key);
2086
2087                 if (timeout && !timeout->task)
2088                         ret = -ETIMEDOUT;
2089                 else {
2090                         /*
2091                          * We expect signal_pending(current), but another
2092                          * thread may have handled it for us already.
2093                          */
2094                         /* FIXME: ERESTARTSYS or ERESTARTNOINTR?  Do we care if
2095                          * the user specified SA_RESTART or not? */
2096                         ret = -ERESTARTSYS;
2097                 }
2098         }
2099         return ret;
2100 }
2101
2102 /**
2103  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2104  * @uaddr:      the futex we initialyl wait on (non-pi)
2105  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2106  *              the same type, no requeueing from private to shared, etc.
2107  * @val:        the expected value of uaddr
2108  * @abs_time:   absolute timeout
2109  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all.
2110  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2111  * @uaddr2:     the pi futex we will take prior to returning to user-space
2112  *
2113  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2114  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2115  * complete the acquisition of the rt_mutex prior to returning to userspace.
2116  * This ensures the rt_mutex maintains an owner when it has waiters; without
2117  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2118  * need to.
2119  *
2120  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2121  * via the following:
2122  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2123  * 2) wakeup on uaddr2 after a requeue and subsequent unlock
2124  * 3) signal (before or after requeue)
2125  * 4) timeout (before or after requeue)
2126  *
2127  * If 3, we setup a restart_block with futex_wait_requeue_pi() as the function.
2128  *
2129  * If 2, we may then block on trying to take the rt_mutex and return via:
2130  * 5) successful lock
2131  * 6) signal
2132  * 7) timeout
2133  * 8) other lock acquisition failure
2134  *
2135  * If 6, we setup a restart_block with futex_lock_pi() as the function.
2136  *
2137  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2138  *
2139  * Returns:
2140  *  0 - On success
2141  * <0 - On error
2142  */
2143 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2144                                  u32 val, ktime_t *abs_time, u32 bitset,
2145                                  int clockrt, u32 __user *uaddr2)
2146 {
2147         struct hrtimer_sleeper timeout, *to = NULL;
2148         struct rt_mutex_waiter rt_waiter;
2149         struct rt_mutex *pi_mutex = NULL;
2150         DECLARE_WAITQUEUE(wait, current);
2151         struct restart_block *restart;
2152         struct futex_hash_bucket *hb;
2153         union futex_key key2;
2154         struct futex_q q;
2155         int res, ret;
2156         u32 uval;
2157
2158         if (!bitset)
2159                 return -EINVAL;
2160
2161         if (abs_time) {
2162                 to = &timeout;
2163                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2164                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2165                 hrtimer_init_sleeper(to, current);
2166                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2167                                              current->timer_slack_ns);
2168         }
2169
2170         /*
2171          * The waiter is allocated on our stack, manipulated by the requeue
2172          * code while we sleep on uaddr.
2173          */
2174         debug_rt_mutex_init_waiter(&rt_waiter);
2175         rt_waiter.task = NULL;
2176
2177         q.pi_state = NULL;
2178         q.bitset = bitset;
2179         q.rt_waiter = &rt_waiter;
2180
2181         key2 = FUTEX_KEY_INIT;
2182         ret = get_futex_key(uaddr2, fshared, &key2);
2183         if (unlikely(ret != 0))
2184                 goto out;
2185
2186         /* Prepare to wait on uaddr. */
2187         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2188         if (ret) {
2189                 put_futex_key(fshared, &key2);
2190                 goto out;
2191         }
2192
2193         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2194         futex_wait_queue_me(hb, &q, to, &wait);
2195
2196         spin_lock(&hb->lock);
2197         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2198         spin_unlock(&hb->lock);
2199         if (ret)
2200                 goto out_put_keys;
2201
2202         /*
2203          * In order for us to be here, we know our q.key == key2, and since
2204          * we took the hb->lock above, we also know that futex_requeue() has
2205          * completed and we no longer have to concern ourselves with a wakeup
2206          * race with the atomic proxy lock acquition by the requeue code.
2207          */
2208
2209         /* Check if the requeue code acquired the second futex for us. */
2210         if (!q.rt_waiter) {
2211                 /*
2212                  * Got the lock. We might not be the anticipated owner if we
2213                  * did a lock-steal - fix up the PI-state in that case.
2214                  */
2215                 if (q.pi_state && (q.pi_state->owner != current)) {
2216                         spin_lock(q.lock_ptr);
2217                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2218                                                    fshared);
2219                         spin_unlock(q.lock_ptr);
2220                 }
2221         } else {
2222                 /*
2223                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2224                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2225                  * the pi_state.
2226                  */
2227                 WARN_ON(!&q.pi_state);
2228                 pi_mutex = &q.pi_state->pi_mutex;
2229                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2230                 debug_rt_mutex_free_waiter(&rt_waiter);
2231
2232                 spin_lock(q.lock_ptr);
2233                 /*
2234                  * Fixup the pi_state owner and possibly acquire the lock if we
2235                  * haven't already.
2236                  */
2237                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2238                 /*
2239                  * If fixup_owner() returned an error, proprogate that.  If it
2240                  * acquired the lock, clear our -ETIMEDOUT or -EINTR.
2241                  */
2242                 if (res)
2243                         ret = (res < 0) ? res : 0;
2244
2245                 /* Unqueue and drop the lock. */
2246                 unqueue_me_pi(&q);
2247         }
2248
2249         /*
2250          * If fixup_pi_state_owner() faulted and was unable to handle the
2251          * fault, unlock the rt_mutex and return the fault to userspace.
2252          */
2253         if (ret == -EFAULT) {
2254                 if (rt_mutex_owner(pi_mutex) == current)
2255                         rt_mutex_unlock(pi_mutex);
2256         } else if (ret == -EINTR) {
2257                 ret = -EFAULT;
2258                 if (get_user(uval, uaddr2))
2259                         goto out_put_keys;
2260
2261                 /*
2262                  * We've already been requeued, so restart by calling
2263                  * futex_lock_pi() directly, rather then returning to this
2264                  * function.
2265                  */
2266                 ret = -ERESTART_RESTARTBLOCK;
2267                 restart = &current_thread_info()->restart_block;
2268                 restart->fn = futex_lock_pi_restart;
2269                 restart->futex.uaddr = (u32 *)uaddr2;
2270                 restart->futex.val = uval;
2271                 restart->futex.flags = 0;
2272                 if (abs_time) {
2273                         restart->futex.flags |= FLAGS_HAS_TIMEOUT;
2274                         restart->futex.time = abs_time->tv64;
2275                 }
2276
2277                 if (fshared)
2278                         restart->futex.flags |= FLAGS_SHARED;
2279                 if (clockrt)
2280                         restart->futex.flags |= FLAGS_CLOCKRT;
2281         }
2282
2283 out_put_keys:
2284         put_futex_key(fshared, &q.key);
2285         put_futex_key(fshared, &key2);
2286
2287 out:
2288         if (to) {
2289                 hrtimer_cancel(&to->timer);
2290                 destroy_hrtimer_on_stack(&to->timer);
2291         }
2292         return ret;
2293 }
2294
2295 /*
2296  * Support for robust futexes: the kernel cleans up held futexes at
2297  * thread exit time.
2298  *
2299  * Implementation: user-space maintains a per-thread list of locks it
2300  * is holding. Upon do_exit(), the kernel carefully walks this list,
2301  * and marks all locks that are owned by this thread with the
2302  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2303  * always manipulated with the lock held, so the list is private and
2304  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2305  * field, to allow the kernel to clean up if the thread dies after
2306  * acquiring the lock, but just before it could have added itself to
2307  * the list. There can only be one such pending lock.
2308  */
2309
2310 /**
2311  * sys_set_robust_list - set the robust-futex list head of a task
2312  * @head: pointer to the list-head
2313  * @len: length of the list-head, as userspace expects
2314  */
2315 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2316                 size_t, len)
2317 {
2318         if (!futex_cmpxchg_enabled)
2319                 return -ENOSYS;
2320         /*
2321          * The kernel knows only one size for now:
2322          */
2323         if (unlikely(len != sizeof(*head)))
2324                 return -EINVAL;
2325
2326         current->robust_list = head;
2327
2328         return 0;
2329 }
2330
2331 /**
2332  * sys_get_robust_list - get the robust-futex list head of a task
2333  * @pid: pid of the process [zero for current task]
2334  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2335  * @len_ptr: pointer to a length field, the kernel fills in the header size
2336  */
2337 SYSCALL_DEFINE3(get_robust_list, int, pid,
2338                 struct robust_list_head __user * __user *, head_ptr,
2339                 size_t __user *, len_ptr)
2340 {
2341         struct robust_list_head __user *head;
2342         unsigned long ret;
2343         const struct cred *cred = current_cred(), *pcred;
2344
2345         if (!futex_cmpxchg_enabled)
2346                 return -ENOSYS;
2347
2348         if (!pid)
2349                 head = current->robust_list;
2350         else {
2351                 struct task_struct *p;
2352
2353                 ret = -ESRCH;
2354                 rcu_read_lock();
2355                 p = find_task_by_vpid(pid);
2356                 if (!p)
2357                         goto err_unlock;
2358                 ret = -EPERM;
2359                 pcred = __task_cred(p);
2360                 if (cred->euid != pcred->euid &&
2361                     cred->euid != pcred->uid &&
2362                     !capable(CAP_SYS_PTRACE))
2363                         goto err_unlock;
2364                 head = p->robust_list;
2365                 rcu_read_unlock();
2366         }
2367
2368         if (put_user(sizeof(*head), len_ptr))
2369                 return -EFAULT;
2370         return put_user(head, head_ptr);
2371
2372 err_unlock:
2373         rcu_read_unlock();
2374
2375         return ret;
2376 }
2377
2378 /*
2379  * Process a futex-list entry, check whether it's owned by the
2380  * dying task, and do notification if so:
2381  */
2382 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2383 {
2384         u32 uval, nval, mval;
2385
2386 retry:
2387         if (get_user(uval, uaddr))
2388                 return -1;
2389
2390         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2391                 /*
2392                  * Ok, this dying thread is truly holding a futex
2393                  * of interest. Set the OWNER_DIED bit atomically
2394                  * via cmpxchg, and if the value had FUTEX_WAITERS
2395                  * set, wake up a waiter (if any). (We have to do a
2396                  * futex_wake() even if OWNER_DIED is already set -
2397                  * to handle the rare but possible case of recursive
2398                  * thread-death.) The rest of the cleanup is done in
2399                  * userspace.
2400                  */
2401                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2402                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2403
2404                 if (nval == -EFAULT)
2405                         return -1;
2406
2407                 if (nval != uval)
2408                         goto retry;
2409
2410                 /*
2411                  * Wake robust non-PI futexes here. The wakeup of
2412                  * PI futexes happens in exit_pi_state():
2413                  */
2414                 if (!pi && (uval & FUTEX_WAITERS))
2415                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2416         }
2417         return 0;
2418 }
2419
2420 /*
2421  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2422  */
2423 static inline int fetch_robust_entry(struct robust_list __user **entry,
2424                                      struct robust_list __user * __user *head,
2425                                      int *pi)
2426 {
2427         unsigned long uentry;
2428
2429         if (get_user(uentry, (unsigned long __user *)head))
2430                 return -EFAULT;
2431
2432         *entry = (void __user *)(uentry & ~1UL);
2433         *pi = uentry & 1;
2434
2435         return 0;
2436 }
2437
2438 /*
2439  * Walk curr->robust_list (very carefully, it's a userspace list!)
2440  * and mark any locks found there dead, and notify any waiters.
2441  *
2442  * We silently return on any sign of list-walking problem.
2443  */
2444 void exit_robust_list(struct task_struct *curr)
2445 {
2446         struct robust_list_head __user *head = curr->robust_list;
2447         struct robust_list __user *entry, *next_entry, *pending;
2448         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2449         unsigned long futex_offset;
2450         int rc;
2451
2452         if (!futex_cmpxchg_enabled)
2453                 return;
2454
2455         /*
2456          * Fetch the list head (which was registered earlier, via
2457          * sys_set_robust_list()):
2458          */
2459         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2460                 return;
2461         /*
2462          * Fetch the relative futex offset:
2463          */
2464         if (get_user(futex_offset, &head->futex_offset))
2465                 return;
2466         /*
2467          * Fetch any possibly pending lock-add first, and handle it
2468          * if it exists:
2469          */
2470         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2471                 return;
2472
2473         next_entry = NULL;      /* avoid warning with gcc */
2474         while (entry != &head->list) {
2475                 /*
2476                  * Fetch the next entry in the list before calling
2477                  * handle_futex_death:
2478                  */
2479                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2480                 /*
2481                  * A pending lock might already be on the list, so
2482                  * don't process it twice:
2483                  */
2484                 if (entry != pending)
2485                         if (handle_futex_death((void __user *)entry + futex_offset,
2486                                                 curr, pi))
2487                                 return;
2488                 if (rc)
2489                         return;
2490                 entry = next_entry;
2491                 pi = next_pi;
2492                 /*
2493                  * Avoid excessively long or circular lists:
2494                  */
2495                 if (!--limit)
2496                         break;
2497
2498                 cond_resched();
2499         }
2500
2501         if (pending)
2502                 handle_futex_death((void __user *)pending + futex_offset,
2503                                    curr, pip);
2504 }
2505
2506 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2507                 u32 __user *uaddr2, u32 val2, u32 val3)
2508 {
2509         int clockrt, ret = -ENOSYS;
2510         int cmd = op & FUTEX_CMD_MASK;
2511         int fshared = 0;
2512
2513         if (!(op & FUTEX_PRIVATE_FLAG))
2514                 fshared = 1;
2515
2516         clockrt = op & FUTEX_CLOCK_REALTIME;
2517         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2518                 return -ENOSYS;
2519
2520         switch (cmd) {
2521         case FUTEX_WAIT:
2522                 val3 = FUTEX_BITSET_MATCH_ANY;
2523         case FUTEX_WAIT_BITSET:
2524                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2525                 break;
2526         case FUTEX_WAKE:
2527                 val3 = FUTEX_BITSET_MATCH_ANY;
2528         case FUTEX_WAKE_BITSET:
2529                 ret = futex_wake(uaddr, fshared, val, val3);
2530                 break;
2531         case FUTEX_REQUEUE:
2532                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2533                 break;
2534         case FUTEX_CMP_REQUEUE:
2535                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2536                                     0);
2537                 break;
2538         case FUTEX_WAKE_OP:
2539                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2540                 break;
2541         case FUTEX_LOCK_PI:
2542                 if (futex_cmpxchg_enabled)
2543                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2544                 break;
2545         case FUTEX_UNLOCK_PI:
2546                 if (futex_cmpxchg_enabled)
2547                         ret = futex_unlock_pi(uaddr, fshared);
2548                 break;
2549         case FUTEX_TRYLOCK_PI:
2550                 if (futex_cmpxchg_enabled)
2551                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2552                 break;
2553         case FUTEX_WAIT_REQUEUE_PI:
2554                 val3 = FUTEX_BITSET_MATCH_ANY;
2555                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2556                                             clockrt, uaddr2);
2557                 break;
2558         case FUTEX_REQUEUE_PI:
2559                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 1);
2560                 break;
2561         case FUTEX_CMP_REQUEUE_PI:
2562                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2563                                     1);
2564                 break;
2565         default:
2566                 ret = -ENOSYS;
2567         }
2568         return ret;
2569 }
2570
2571
2572 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2573                 struct timespec __user *, utime, u32 __user *, uaddr2,
2574                 u32, val3)
2575 {
2576         struct timespec ts;
2577         ktime_t t, *tp = NULL;
2578         u32 val2 = 0;
2579         int cmd = op & FUTEX_CMD_MASK;
2580
2581         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2582                       cmd == FUTEX_WAIT_BITSET ||
2583                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2584                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2585                         return -EFAULT;
2586                 if (!timespec_valid(&ts))
2587                         return -EINVAL;
2588
2589                 t = timespec_to_ktime(ts);
2590                 if (cmd == FUTEX_WAIT)
2591                         t = ktime_add_safe(ktime_get(), t);
2592                 tp = &t;
2593         }
2594         /*
2595          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2596          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2597          */
2598         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2599             cmd == FUTEX_REQUEUE_PI || cmd == FUTEX_CMP_REQUEUE_PI ||
2600             cmd == FUTEX_WAKE_OP)
2601                 val2 = (u32) (unsigned long) utime;
2602
2603         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2604 }
2605
2606 static int __init futex_init(void)
2607 {
2608         u32 curval;
2609         int i;
2610
2611         /*
2612          * This will fail and we want it. Some arch implementations do
2613          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2614          * functionality. We want to know that before we call in any
2615          * of the complex code paths. Also we want to prevent
2616          * registration of robust lists in that case. NULL is
2617          * guaranteed to fault and we get -EFAULT on functional
2618          * implementation, the non functional ones will return
2619          * -ENOSYS.
2620          */
2621         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2622         if (curval == -EFAULT)
2623                 futex_cmpxchg_enabled = 1;
2624
2625         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2626                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2627                 spin_lock_init(&futex_queues[i].lock);
2628         }
2629
2630         return 0;
2631 }
2632 __initcall(futex_init);