futex: Fix errors in nested key ref-counting
[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  * struct futex_q - The hashed futex queue entry, one per waiting task
94  * @task:               the task waiting on the futex
95  * @lock_ptr:           the hash bucket lock
96  * @key:                the key the futex is hashed on
97  * @pi_state:           optional priority inheritance state
98  * @rt_waiter:          rt_waiter storage for use with requeue_pi
99  * @requeue_pi_key:     the requeue_pi target futex key
100  * @bitset:             bitset for the optional bitmasked wakeup
101  *
102  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103  * we can wake only the relevant ones (hashed queues may be shared).
104  *
105  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107  * The order of wakup is always to make the first condition true, then
108  * the second.
109  *
110  * PI futexes are typically woken before they are removed from the hash list via
111  * the rt_mutex code. See unqueue_me_pi().
112  */
113 struct futex_q {
114         struct plist_node list;
115
116         struct task_struct *task;
117         spinlock_t *lock_ptr;
118         union futex_key key;
119         struct futex_pi_state *pi_state;
120         struct rt_mutex_waiter *rt_waiter;
121         union futex_key *requeue_pi_key;
122         u32 bitset;
123 };
124
125 /*
126  * Hash buckets are shared by all the futex_keys that hash to the same
127  * location.  Each key may have multiple futex_q structures, one for each task
128  * waiting on a futex.
129  */
130 struct futex_hash_bucket {
131         spinlock_t lock;
132         struct plist_head chain;
133 };
134
135 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
136
137 /*
138  * We hash on the keys returned from get_futex_key (see below).
139  */
140 static struct futex_hash_bucket *hash_futex(union futex_key *key)
141 {
142         u32 hash = jhash2((u32*)&key->both.word,
143                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
144                           key->both.offset);
145         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146 }
147
148 /*
149  * Return 1 if two futex_keys are equal, 0 otherwise.
150  */
151 static inline int match_futex(union futex_key *key1, union futex_key *key2)
152 {
153         return (key1 && key2
154                 && key1->both.word == key2->both.word
155                 && key1->both.ptr == key2->both.ptr
156                 && key1->both.offset == key2->both.offset);
157 }
158
159 /*
160  * Take a reference to the resource addressed by a key.
161  * Can be called while holding spinlocks.
162  *
163  */
164 static void get_futex_key_refs(union futex_key *key)
165 {
166         if (!key->both.ptr)
167                 return;
168
169         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170         case FUT_OFF_INODE:
171                 atomic_inc(&key->shared.inode->i_count);
172                 break;
173         case FUT_OFF_MMSHARED:
174                 atomic_inc(&key->private.mm->mm_count);
175                 break;
176         }
177 }
178
179 /*
180  * Drop a reference to the resource addressed by a key.
181  * The hash bucket spinlock must not be held.
182  */
183 static void drop_futex_key_refs(union futex_key *key)
184 {
185         if (!key->both.ptr) {
186                 /* If we're here then we tried to put a key we failed to get */
187                 WARN_ON_ONCE(1);
188                 return;
189         }
190
191         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192         case FUT_OFF_INODE:
193                 iput(key->shared.inode);
194                 break;
195         case FUT_OFF_MMSHARED:
196                 mmdrop(key->private.mm);
197                 break;
198         }
199 }
200
201 /**
202  * get_futex_key() - Get parameters which are the keys for a futex
203  * @uaddr:      virtual address of the futex
204  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205  * @key:        address where result is stored.
206  *
207  * Returns a negative error code or 0
208  * The key words are stored in *key on success.
209  *
210  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
211  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
212  * We can usually work out the index without swapping in the page.
213  *
214  * lock_page() might sleep, the caller should not hold a spinlock.
215  */
216 static int
217 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
218 {
219         unsigned long address = (unsigned long)uaddr;
220         struct mm_struct *mm = current->mm;
221         struct page *page;
222         int err;
223
224         /*
225          * The futex address must be "naturally" aligned.
226          */
227         key->both.offset = address % PAGE_SIZE;
228         if (unlikely((address % sizeof(u32)) != 0))
229                 return -EINVAL;
230         address -= key->both.offset;
231
232         /*
233          * PROCESS_PRIVATE futexes are fast.
234          * As the mm cannot disappear under us and the 'key' only needs
235          * virtual address, we dont even have to find the underlying vma.
236          * Note : We do have to check 'uaddr' is a valid user address,
237          *        but access_ok() should be faster than find_vma()
238          */
239         if (!fshared) {
240                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
241                         return -EFAULT;
242                 key->private.mm = mm;
243                 key->private.address = address;
244                 get_futex_key_refs(key);
245                 return 0;
246         }
247
248 again:
249         err = get_user_pages_fast(address, 1, 1, &page);
250         if (err < 0)
251                 return err;
252
253         page = compound_head(page);
254         lock_page(page);
255         if (!page->mapping) {
256                 unlock_page(page);
257                 put_page(page);
258                 goto again;
259         }
260
261         /*
262          * Private mappings are handled in a simple way.
263          *
264          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
265          * it's a read-only handle, it's expected that futexes attach to
266          * the object not the particular process.
267          */
268         if (PageAnon(page)) {
269                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
270                 key->private.mm = mm;
271                 key->private.address = address;
272         } else {
273                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
274                 key->shared.inode = page->mapping->host;
275                 key->shared.pgoff = page->index;
276         }
277
278         get_futex_key_refs(key);
279
280         unlock_page(page);
281         put_page(page);
282         return 0;
283 }
284
285 static inline
286 void put_futex_key(int fshared, union futex_key *key)
287 {
288         drop_futex_key_refs(key);
289 }
290
291 /**
292  * fault_in_user_writeable() - Fault in user address and verify RW access
293  * @uaddr:      pointer to faulting user space address
294  *
295  * Slow path to fixup the fault we just took in the atomic write
296  * access to @uaddr.
297  *
298  * We have no generic implementation of a non destructive write to the
299  * user address. We know that we faulted in the atomic pagefault
300  * disabled section so we can as well avoid the #PF overhead by
301  * calling get_user_pages() right away.
302  */
303 static int fault_in_user_writeable(u32 __user *uaddr)
304 {
305         struct mm_struct *mm = current->mm;
306         int ret;
307
308         down_read(&mm->mmap_sem);
309         ret = get_user_pages(current, mm, (unsigned long)uaddr,
310                              1, 1, 0, NULL, NULL);
311         up_read(&mm->mmap_sem);
312
313         return ret < 0 ? ret : 0;
314 }
315
316 /**
317  * futex_top_waiter() - Return the highest priority waiter on a futex
318  * @hb:         the hash bucket the futex_q's reside in
319  * @key:        the futex key (to distinguish it from other futex futex_q's)
320  *
321  * Must be called with the hb lock held.
322  */
323 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
324                                         union futex_key *key)
325 {
326         struct futex_q *this;
327
328         plist_for_each_entry(this, &hb->chain, list) {
329                 if (match_futex(&this->key, key))
330                         return this;
331         }
332         return NULL;
333 }
334
335 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
336 {
337         u32 curval;
338
339         pagefault_disable();
340         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
341         pagefault_enable();
342
343         return curval;
344 }
345
346 static int get_futex_value_locked(u32 *dest, u32 __user *from)
347 {
348         int ret;
349
350         pagefault_disable();
351         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
352         pagefault_enable();
353
354         return ret ? -EFAULT : 0;
355 }
356
357
358 /*
359  * PI code:
360  */
361 static int refill_pi_state_cache(void)
362 {
363         struct futex_pi_state *pi_state;
364
365         if (likely(current->pi_state_cache))
366                 return 0;
367
368         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
369
370         if (!pi_state)
371                 return -ENOMEM;
372
373         INIT_LIST_HEAD(&pi_state->list);
374         /* pi_mutex gets initialized later */
375         pi_state->owner = NULL;
376         atomic_set(&pi_state->refcount, 1);
377         pi_state->key = FUTEX_KEY_INIT;
378
379         current->pi_state_cache = pi_state;
380
381         return 0;
382 }
383
384 static struct futex_pi_state * alloc_pi_state(void)
385 {
386         struct futex_pi_state *pi_state = current->pi_state_cache;
387
388         WARN_ON(!pi_state);
389         current->pi_state_cache = NULL;
390
391         return pi_state;
392 }
393
394 static void free_pi_state(struct futex_pi_state *pi_state)
395 {
396         if (!atomic_dec_and_test(&pi_state->refcount))
397                 return;
398
399         /*
400          * If pi_state->owner is NULL, the owner is most probably dying
401          * and has cleaned up the pi_state already
402          */
403         if (pi_state->owner) {
404                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
405                 list_del_init(&pi_state->list);
406                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
407
408                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
409         }
410
411         if (current->pi_state_cache)
412                 kfree(pi_state);
413         else {
414                 /*
415                  * pi_state->list is already empty.
416                  * clear pi_state->owner.
417                  * refcount is at 0 - put it back to 1.
418                  */
419                 pi_state->owner = NULL;
420                 atomic_set(&pi_state->refcount, 1);
421                 current->pi_state_cache = pi_state;
422         }
423 }
424
425 /*
426  * Look up the task based on what TID userspace gave us.
427  * We dont trust it.
428  */
429 static struct task_struct * futex_find_get_task(pid_t pid)
430 {
431         struct task_struct *p;
432
433         rcu_read_lock();
434         p = find_task_by_vpid(pid);
435         if (p)
436                 get_task_struct(p);
437
438         rcu_read_unlock();
439
440         return p;
441 }
442
443 /*
444  * This task is holding PI mutexes at exit time => bad.
445  * Kernel cleans up PI-state, but userspace is likely hosed.
446  * (Robust-futex cleanup is separate and might save the day for userspace.)
447  */
448 void exit_pi_state_list(struct task_struct *curr)
449 {
450         struct list_head *next, *head = &curr->pi_state_list;
451         struct futex_pi_state *pi_state;
452         struct futex_hash_bucket *hb;
453         union futex_key key = FUTEX_KEY_INIT;
454
455         if (!futex_cmpxchg_enabled)
456                 return;
457         /*
458          * We are a ZOMBIE and nobody can enqueue itself on
459          * pi_state_list anymore, but we have to be careful
460          * versus waiters unqueueing themselves:
461          */
462         raw_spin_lock_irq(&curr->pi_lock);
463         while (!list_empty(head)) {
464
465                 next = head->next;
466                 pi_state = list_entry(next, struct futex_pi_state, list);
467                 key = pi_state->key;
468                 hb = hash_futex(&key);
469                 raw_spin_unlock_irq(&curr->pi_lock);
470
471                 spin_lock(&hb->lock);
472
473                 raw_spin_lock_irq(&curr->pi_lock);
474                 /*
475                  * We dropped the pi-lock, so re-check whether this
476                  * task still owns the PI-state:
477                  */
478                 if (head->next != next) {
479                         spin_unlock(&hb->lock);
480                         continue;
481                 }
482
483                 WARN_ON(pi_state->owner != curr);
484                 WARN_ON(list_empty(&pi_state->list));
485                 list_del_init(&pi_state->list);
486                 pi_state->owner = NULL;
487                 raw_spin_unlock_irq(&curr->pi_lock);
488
489                 rt_mutex_unlock(&pi_state->pi_mutex);
490
491                 spin_unlock(&hb->lock);
492
493                 raw_spin_lock_irq(&curr->pi_lock);
494         }
495         raw_spin_unlock_irq(&curr->pi_lock);
496 }
497
498 static int
499 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
500                 union futex_key *key, struct futex_pi_state **ps)
501 {
502         struct futex_pi_state *pi_state = NULL;
503         struct futex_q *this, *next;
504         struct plist_head *head;
505         struct task_struct *p;
506         pid_t pid = uval & FUTEX_TID_MASK;
507
508         head = &hb->chain;
509
510         plist_for_each_entry_safe(this, next, head, list) {
511                 if (match_futex(&this->key, key)) {
512                         /*
513                          * Another waiter already exists - bump up
514                          * the refcount and return its pi_state:
515                          */
516                         pi_state = this->pi_state;
517                         /*
518                          * Userspace might have messed up non PI and PI futexes
519                          */
520                         if (unlikely(!pi_state))
521                                 return -EINVAL;
522
523                         WARN_ON(!atomic_read(&pi_state->refcount));
524
525                         /*
526                          * When pi_state->owner is NULL then the owner died
527                          * and another waiter is on the fly. pi_state->owner
528                          * is fixed up by the task which acquires
529                          * pi_state->rt_mutex.
530                          *
531                          * We do not check for pid == 0 which can happen when
532                          * the owner died and robust_list_exit() cleared the
533                          * TID.
534                          */
535                         if (pid && pi_state->owner) {
536                                 /*
537                                  * Bail out if user space manipulated the
538                                  * futex value.
539                                  */
540                                 if (pid != task_pid_vnr(pi_state->owner))
541                                         return -EINVAL;
542                         }
543
544                         atomic_inc(&pi_state->refcount);
545                         *ps = pi_state;
546
547                         return 0;
548                 }
549         }
550
551         /*
552          * We are the first waiter - try to look up the real owner and attach
553          * the new pi_state to it, but bail out when TID = 0
554          */
555         if (!pid)
556                 return -ESRCH;
557         p = futex_find_get_task(pid);
558         if (!p)
559                 return -ESRCH;
560
561         /*
562          * We need to look at the task state flags to figure out,
563          * whether the task is exiting. To protect against the do_exit
564          * change of the task flags, we do this protected by
565          * p->pi_lock:
566          */
567         raw_spin_lock_irq(&p->pi_lock);
568         if (unlikely(p->flags & PF_EXITING)) {
569                 /*
570                  * The task is on the way out. When PF_EXITPIDONE is
571                  * set, we know that the task has finished the
572                  * cleanup:
573                  */
574                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
575
576                 raw_spin_unlock_irq(&p->pi_lock);
577                 put_task_struct(p);
578                 return ret;
579         }
580
581         pi_state = alloc_pi_state();
582
583         /*
584          * Initialize the pi_mutex in locked state and make 'p'
585          * the owner of it:
586          */
587         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
588
589         /* Store the key for possible exit cleanups: */
590         pi_state->key = *key;
591
592         WARN_ON(!list_empty(&pi_state->list));
593         list_add(&pi_state->list, &p->pi_state_list);
594         pi_state->owner = p;
595         raw_spin_unlock_irq(&p->pi_lock);
596
597         put_task_struct(p);
598
599         *ps = pi_state;
600
601         return 0;
602 }
603
604 /**
605  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
606  * @uaddr:              the pi futex user address
607  * @hb:                 the pi futex hash bucket
608  * @key:                the futex key associated with uaddr and hb
609  * @ps:                 the pi_state pointer where we store the result of the
610  *                      lookup
611  * @task:               the task to perform the atomic lock work for.  This will
612  *                      be "current" except in the case of requeue pi.
613  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
614  *
615  * Returns:
616  *  0 - ready to wait
617  *  1 - acquired the lock
618  * <0 - error
619  *
620  * The hb->lock and futex_key refs shall be held by the caller.
621  */
622 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
623                                 union futex_key *key,
624                                 struct futex_pi_state **ps,
625                                 struct task_struct *task, int set_waiters)
626 {
627         int lock_taken, ret, ownerdied = 0;
628         u32 uval, newval, curval;
629
630 retry:
631         ret = lock_taken = 0;
632
633         /*
634          * To avoid races, we attempt to take the lock here again
635          * (by doing a 0 -> TID atomic cmpxchg), while holding all
636          * the locks. It will most likely not succeed.
637          */
638         newval = task_pid_vnr(task);
639         if (set_waiters)
640                 newval |= FUTEX_WAITERS;
641
642         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
643
644         if (unlikely(curval == -EFAULT))
645                 return -EFAULT;
646
647         /*
648          * Detect deadlocks.
649          */
650         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
651                 return -EDEADLK;
652
653         /*
654          * Surprise - we got the lock. Just return to userspace:
655          */
656         if (unlikely(!curval))
657                 return 1;
658
659         uval = curval;
660
661         /*
662          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
663          * to wake at the next unlock.
664          */
665         newval = curval | FUTEX_WAITERS;
666
667         /*
668          * There are two cases, where a futex might have no owner (the
669          * owner TID is 0): OWNER_DIED. We take over the futex in this
670          * case. We also do an unconditional take over, when the owner
671          * of the futex died.
672          *
673          * This is safe as we are protected by the hash bucket lock !
674          */
675         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
676                 /* Keep the OWNER_DIED bit */
677                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
678                 ownerdied = 0;
679                 lock_taken = 1;
680         }
681
682         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
683
684         if (unlikely(curval == -EFAULT))
685                 return -EFAULT;
686         if (unlikely(curval != uval))
687                 goto retry;
688
689         /*
690          * We took the lock due to owner died take over.
691          */
692         if (unlikely(lock_taken))
693                 return 1;
694
695         /*
696          * We dont have the lock. Look up the PI state (or create it if
697          * we are the first waiter):
698          */
699         ret = lookup_pi_state(uval, hb, key, ps);
700
701         if (unlikely(ret)) {
702                 switch (ret) {
703                 case -ESRCH:
704                         /*
705                          * No owner found for this futex. Check if the
706                          * OWNER_DIED bit is set to figure out whether
707                          * this is a robust futex or not.
708                          */
709                         if (get_futex_value_locked(&curval, uaddr))
710                                 return -EFAULT;
711
712                         /*
713                          * We simply start over in case of a robust
714                          * futex. The code above will take the futex
715                          * and return happy.
716                          */
717                         if (curval & FUTEX_OWNER_DIED) {
718                                 ownerdied = 1;
719                                 goto retry;
720                         }
721                 default:
722                         break;
723                 }
724         }
725
726         return ret;
727 }
728
729 /*
730  * The hash bucket lock must be held when this is called.
731  * Afterwards, the futex_q must not be accessed.
732  */
733 static void wake_futex(struct futex_q *q)
734 {
735         struct task_struct *p = q->task;
736
737         /*
738          * We set q->lock_ptr = NULL _before_ we wake up the task. If
739          * a non futex wake up happens on another CPU then the task
740          * might exit and p would dereference a non existing task
741          * struct. Prevent this by holding a reference on p across the
742          * wake up.
743          */
744         get_task_struct(p);
745
746         plist_del(&q->list, &q->list.plist);
747         /*
748          * The waiting task can free the futex_q as soon as
749          * q->lock_ptr = NULL is written, without taking any locks. A
750          * memory barrier is required here to prevent the following
751          * store to lock_ptr from getting ahead of the plist_del.
752          */
753         smp_wmb();
754         q->lock_ptr = NULL;
755
756         wake_up_state(p, TASK_NORMAL);
757         put_task_struct(p);
758 }
759
760 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
761 {
762         struct task_struct *new_owner;
763         struct futex_pi_state *pi_state = this->pi_state;
764         u32 curval, newval;
765
766         if (!pi_state)
767                 return -EINVAL;
768
769         /*
770          * If current does not own the pi_state then the futex is
771          * inconsistent and user space fiddled with the futex value.
772          */
773         if (pi_state->owner != current)
774                 return -EINVAL;
775
776         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
777         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
778
779         /*
780          * This happens when we have stolen the lock and the original
781          * pending owner did not enqueue itself back on the rt_mutex.
782          * Thats not a tragedy. We know that way, that a lock waiter
783          * is on the fly. We make the futex_q waiter the pending owner.
784          */
785         if (!new_owner)
786                 new_owner = this->task;
787
788         /*
789          * We pass it to the next owner. (The WAITERS bit is always
790          * kept enabled while there is PI state around. We must also
791          * preserve the owner died bit.)
792          */
793         if (!(uval & FUTEX_OWNER_DIED)) {
794                 int ret = 0;
795
796                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
797
798                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
799
800                 if (curval == -EFAULT)
801                         ret = -EFAULT;
802                 else if (curval != uval)
803                         ret = -EINVAL;
804                 if (ret) {
805                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
806                         return ret;
807                 }
808         }
809
810         raw_spin_lock_irq(&pi_state->owner->pi_lock);
811         WARN_ON(list_empty(&pi_state->list));
812         list_del_init(&pi_state->list);
813         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
814
815         raw_spin_lock_irq(&new_owner->pi_lock);
816         WARN_ON(!list_empty(&pi_state->list));
817         list_add(&pi_state->list, &new_owner->pi_state_list);
818         pi_state->owner = new_owner;
819         raw_spin_unlock_irq(&new_owner->pi_lock);
820
821         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
822         rt_mutex_unlock(&pi_state->pi_mutex);
823
824         return 0;
825 }
826
827 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
828 {
829         u32 oldval;
830
831         /*
832          * There is no waiter, so we unlock the futex. The owner died
833          * bit has not to be preserved here. We are the owner:
834          */
835         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
836
837         if (oldval == -EFAULT)
838                 return oldval;
839         if (oldval != uval)
840                 return -EAGAIN;
841
842         return 0;
843 }
844
845 /*
846  * Express the locking dependencies for lockdep:
847  */
848 static inline void
849 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
850 {
851         if (hb1 <= hb2) {
852                 spin_lock(&hb1->lock);
853                 if (hb1 < hb2)
854                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
855         } else { /* hb1 > hb2 */
856                 spin_lock(&hb2->lock);
857                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
858         }
859 }
860
861 static inline void
862 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
863 {
864         spin_unlock(&hb1->lock);
865         if (hb1 != hb2)
866                 spin_unlock(&hb2->lock);
867 }
868
869 /*
870  * Wake up waiters matching bitset queued on this futex (uaddr).
871  */
872 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
873 {
874         struct futex_hash_bucket *hb;
875         struct futex_q *this, *next;
876         struct plist_head *head;
877         union futex_key key = FUTEX_KEY_INIT;
878         int ret;
879
880         if (!bitset)
881                 return -EINVAL;
882
883         ret = get_futex_key(uaddr, fshared, &key);
884         if (unlikely(ret != 0))
885                 goto out;
886
887         hb = hash_futex(&key);
888         spin_lock(&hb->lock);
889         head = &hb->chain;
890
891         plist_for_each_entry_safe(this, next, head, list) {
892                 if (match_futex (&this->key, &key)) {
893                         if (this->pi_state || this->rt_waiter) {
894                                 ret = -EINVAL;
895                                 break;
896                         }
897
898                         /* Check if one of the bits is set in both bitsets */
899                         if (!(this->bitset & bitset))
900                                 continue;
901
902                         wake_futex(this);
903                         if (++ret >= nr_wake)
904                                 break;
905                 }
906         }
907
908         spin_unlock(&hb->lock);
909         put_futex_key(fshared, &key);
910 out:
911         return ret;
912 }
913
914 /*
915  * Wake up all waiters hashed on the physical page that is mapped
916  * to this virtual address:
917  */
918 static int
919 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
920               int nr_wake, int nr_wake2, int op)
921 {
922         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
923         struct futex_hash_bucket *hb1, *hb2;
924         struct plist_head *head;
925         struct futex_q *this, *next;
926         int ret, op_ret;
927
928 retry:
929         ret = get_futex_key(uaddr1, fshared, &key1);
930         if (unlikely(ret != 0))
931                 goto out;
932         ret = get_futex_key(uaddr2, fshared, &key2);
933         if (unlikely(ret != 0))
934                 goto out_put_key1;
935
936         hb1 = hash_futex(&key1);
937         hb2 = hash_futex(&key2);
938
939 retry_private:
940         double_lock_hb(hb1, hb2);
941         op_ret = futex_atomic_op_inuser(op, uaddr2);
942         if (unlikely(op_ret < 0)) {
943
944                 double_unlock_hb(hb1, hb2);
945
946 #ifndef CONFIG_MMU
947                 /*
948                  * we don't get EFAULT from MMU faults if we don't have an MMU,
949                  * but we might get them from range checking
950                  */
951                 ret = op_ret;
952                 goto out_put_keys;
953 #endif
954
955                 if (unlikely(op_ret != -EFAULT)) {
956                         ret = op_ret;
957                         goto out_put_keys;
958                 }
959
960                 ret = fault_in_user_writeable(uaddr2);
961                 if (ret)
962                         goto out_put_keys;
963
964                 if (!fshared)
965                         goto retry_private;
966
967                 put_futex_key(fshared, &key2);
968                 put_futex_key(fshared, &key1);
969                 goto retry;
970         }
971
972         head = &hb1->chain;
973
974         plist_for_each_entry_safe(this, next, head, list) {
975                 if (match_futex (&this->key, &key1)) {
976                         wake_futex(this);
977                         if (++ret >= nr_wake)
978                                 break;
979                 }
980         }
981
982         if (op_ret > 0) {
983                 head = &hb2->chain;
984
985                 op_ret = 0;
986                 plist_for_each_entry_safe(this, next, head, list) {
987                         if (match_futex (&this->key, &key2)) {
988                                 wake_futex(this);
989                                 if (++op_ret >= nr_wake2)
990                                         break;
991                         }
992                 }
993                 ret += op_ret;
994         }
995
996         double_unlock_hb(hb1, hb2);
997 out_put_keys:
998         put_futex_key(fshared, &key2);
999 out_put_key1:
1000         put_futex_key(fshared, &key1);
1001 out:
1002         return ret;
1003 }
1004
1005 /**
1006  * requeue_futex() - Requeue a futex_q from one hb to another
1007  * @q:          the futex_q to requeue
1008  * @hb1:        the source hash_bucket
1009  * @hb2:        the target hash_bucket
1010  * @key2:       the new key for the requeued futex_q
1011  */
1012 static inline
1013 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1014                    struct futex_hash_bucket *hb2, union futex_key *key2)
1015 {
1016
1017         /*
1018          * If key1 and key2 hash to the same bucket, no need to
1019          * requeue.
1020          */
1021         if (likely(&hb1->chain != &hb2->chain)) {
1022                 plist_del(&q->list, &hb1->chain);
1023                 plist_add(&q->list, &hb2->chain);
1024                 q->lock_ptr = &hb2->lock;
1025 #ifdef CONFIG_DEBUG_PI_LIST
1026                 q->list.plist.spinlock = &hb2->lock;
1027 #endif
1028         }
1029         get_futex_key_refs(key2);
1030         q->key = *key2;
1031 }
1032
1033 /**
1034  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1035  * @q:          the futex_q
1036  * @key:        the key of the requeue target futex
1037  * @hb:         the hash_bucket of the requeue target futex
1038  *
1039  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1040  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1041  * to the requeue target futex so the waiter can detect the wakeup on the right
1042  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1043  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1044  * to protect access to the pi_state to fixup the owner later.  Must be called
1045  * with both q->lock_ptr and hb->lock held.
1046  */
1047 static inline
1048 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1049                            struct futex_hash_bucket *hb)
1050 {
1051         get_futex_key_refs(key);
1052         q->key = *key;
1053
1054         WARN_ON(plist_node_empty(&q->list));
1055         plist_del(&q->list, &q->list.plist);
1056
1057         WARN_ON(!q->rt_waiter);
1058         q->rt_waiter = NULL;
1059
1060         q->lock_ptr = &hb->lock;
1061 #ifdef CONFIG_DEBUG_PI_LIST
1062         q->list.plist.spinlock = &hb->lock;
1063 #endif
1064
1065         wake_up_state(q->task, TASK_NORMAL);
1066 }
1067
1068 /**
1069  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1070  * @pifutex:            the user address of the to futex
1071  * @hb1:                the from futex hash bucket, must be locked by the caller
1072  * @hb2:                the to futex hash bucket, must be locked by the caller
1073  * @key1:               the from futex key
1074  * @key2:               the to futex key
1075  * @ps:                 address to store the pi_state pointer
1076  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1077  *
1078  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1079  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1080  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1081  * hb1 and hb2 must be held by the caller.
1082  *
1083  * Returns:
1084  *  0 - failed to acquire the lock atomicly
1085  *  1 - acquired the lock
1086  * <0 - error
1087  */
1088 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1089                                  struct futex_hash_bucket *hb1,
1090                                  struct futex_hash_bucket *hb2,
1091                                  union futex_key *key1, union futex_key *key2,
1092                                  struct futex_pi_state **ps, int set_waiters)
1093 {
1094         struct futex_q *top_waiter = NULL;
1095         u32 curval;
1096         int ret;
1097
1098         if (get_futex_value_locked(&curval, pifutex))
1099                 return -EFAULT;
1100
1101         /*
1102          * Find the top_waiter and determine if there are additional waiters.
1103          * If the caller intends to requeue more than 1 waiter to pifutex,
1104          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1105          * as we have means to handle the possible fault.  If not, don't set
1106          * the bit unecessarily as it will force the subsequent unlock to enter
1107          * the kernel.
1108          */
1109         top_waiter = futex_top_waiter(hb1, key1);
1110
1111         /* There are no waiters, nothing for us to do. */
1112         if (!top_waiter)
1113                 return 0;
1114
1115         /* Ensure we requeue to the expected futex. */
1116         if (!match_futex(top_waiter->requeue_pi_key, key2))
1117                 return -EINVAL;
1118
1119         /*
1120          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1121          * the contended case or if set_waiters is 1.  The pi_state is returned
1122          * in ps in contended cases.
1123          */
1124         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1125                                    set_waiters);
1126         if (ret == 1)
1127                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1128
1129         return ret;
1130 }
1131
1132 /**
1133  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1134  * uaddr1:      source futex user address
1135  * uaddr2:      target futex user address
1136  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1137  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1138  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1139  *              pi futex (pi to pi requeue is not supported)
1140  *
1141  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1142  * uaddr2 atomically on behalf of the top waiter.
1143  *
1144  * Returns:
1145  * >=0 - on success, the number of tasks requeued or woken
1146  *  <0 - on error
1147  */
1148 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1149                          int nr_wake, int nr_requeue, u32 *cmpval,
1150                          int requeue_pi)
1151 {
1152         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1153         int drop_count = 0, task_count = 0, ret;
1154         struct futex_pi_state *pi_state = NULL;
1155         struct futex_hash_bucket *hb1, *hb2;
1156         struct plist_head *head1;
1157         struct futex_q *this, *next;
1158         u32 curval2;
1159
1160         if (requeue_pi) {
1161                 /*
1162                  * requeue_pi requires a pi_state, try to allocate it now
1163                  * without any locks in case it fails.
1164                  */
1165                 if (refill_pi_state_cache())
1166                         return -ENOMEM;
1167                 /*
1168                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1169                  * + nr_requeue, since it acquires the rt_mutex prior to
1170                  * returning to userspace, so as to not leave the rt_mutex with
1171                  * waiters and no owner.  However, second and third wake-ups
1172                  * cannot be predicted as they involve race conditions with the
1173                  * first wake and a fault while looking up the pi_state.  Both
1174                  * pthread_cond_signal() and pthread_cond_broadcast() should
1175                  * use nr_wake=1.
1176                  */
1177                 if (nr_wake != 1)
1178                         return -EINVAL;
1179         }
1180
1181 retry:
1182         if (pi_state != NULL) {
1183                 /*
1184                  * We will have to lookup the pi_state again, so free this one
1185                  * to keep the accounting correct.
1186                  */
1187                 free_pi_state(pi_state);
1188                 pi_state = NULL;
1189         }
1190
1191         ret = get_futex_key(uaddr1, fshared, &key1);
1192         if (unlikely(ret != 0))
1193                 goto out;
1194         ret = get_futex_key(uaddr2, fshared, &key2);
1195         if (unlikely(ret != 0))
1196                 goto out_put_key1;
1197
1198         hb1 = hash_futex(&key1);
1199         hb2 = hash_futex(&key2);
1200
1201 retry_private:
1202         double_lock_hb(hb1, hb2);
1203
1204         if (likely(cmpval != NULL)) {
1205                 u32 curval;
1206
1207                 ret = get_futex_value_locked(&curval, uaddr1);
1208
1209                 if (unlikely(ret)) {
1210                         double_unlock_hb(hb1, hb2);
1211
1212                         ret = get_user(curval, uaddr1);
1213                         if (ret)
1214                                 goto out_put_keys;
1215
1216                         if (!fshared)
1217                                 goto retry_private;
1218
1219                         put_futex_key(fshared, &key2);
1220                         put_futex_key(fshared, &key1);
1221                         goto retry;
1222                 }
1223                 if (curval != *cmpval) {
1224                         ret = -EAGAIN;
1225                         goto out_unlock;
1226                 }
1227         }
1228
1229         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1230                 /*
1231                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1232                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1233                  * bit.  We force this here where we are able to easily handle
1234                  * faults rather in the requeue loop below.
1235                  */
1236                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1237                                                  &key2, &pi_state, nr_requeue);
1238
1239                 /*
1240                  * At this point the top_waiter has either taken uaddr2 or is
1241                  * waiting on it.  If the former, then the pi_state will not
1242                  * exist yet, look it up one more time to ensure we have a
1243                  * reference to it.
1244                  */
1245                 if (ret == 1) {
1246                         WARN_ON(pi_state);
1247                         drop_count++;
1248                         task_count++;
1249                         ret = get_futex_value_locked(&curval2, uaddr2);
1250                         if (!ret)
1251                                 ret = lookup_pi_state(curval2, hb2, &key2,
1252                                                       &pi_state);
1253                 }
1254
1255                 switch (ret) {
1256                 case 0:
1257                         break;
1258                 case -EFAULT:
1259                         double_unlock_hb(hb1, hb2);
1260                         put_futex_key(fshared, &key2);
1261                         put_futex_key(fshared, &key1);
1262                         ret = fault_in_user_writeable(uaddr2);
1263                         if (!ret)
1264                                 goto retry;
1265                         goto out;
1266                 case -EAGAIN:
1267                         /* The owner was exiting, try again. */
1268                         double_unlock_hb(hb1, hb2);
1269                         put_futex_key(fshared, &key2);
1270                         put_futex_key(fshared, &key1);
1271                         cond_resched();
1272                         goto retry;
1273                 default:
1274                         goto out_unlock;
1275                 }
1276         }
1277
1278         head1 = &hb1->chain;
1279         plist_for_each_entry_safe(this, next, head1, list) {
1280                 if (task_count - nr_wake >= nr_requeue)
1281                         break;
1282
1283                 if (!match_futex(&this->key, &key1))
1284                         continue;
1285
1286                 /*
1287                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1288                  * be paired with each other and no other futex ops.
1289                  */
1290                 if ((requeue_pi && !this->rt_waiter) ||
1291                     (!requeue_pi && this->rt_waiter)) {
1292                         ret = -EINVAL;
1293                         break;
1294                 }
1295
1296                 /*
1297                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1298                  * lock, we already woke the top_waiter.  If not, it will be
1299                  * woken by futex_unlock_pi().
1300                  */
1301                 if (++task_count <= nr_wake && !requeue_pi) {
1302                         wake_futex(this);
1303                         continue;
1304                 }
1305
1306                 /* Ensure we requeue to the expected futex for requeue_pi. */
1307                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1308                         ret = -EINVAL;
1309                         break;
1310                 }
1311
1312                 /*
1313                  * Requeue nr_requeue waiters and possibly one more in the case
1314                  * of requeue_pi if we couldn't acquire the lock atomically.
1315                  */
1316                 if (requeue_pi) {
1317                         /* Prepare the waiter to take the rt_mutex. */
1318                         atomic_inc(&pi_state->refcount);
1319                         this->pi_state = pi_state;
1320                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1321                                                         this->rt_waiter,
1322                                                         this->task, 1);
1323                         if (ret == 1) {
1324                                 /* We got the lock. */
1325                                 requeue_pi_wake_futex(this, &key2, hb2);
1326                                 drop_count++;
1327                                 continue;
1328                         } else if (ret) {
1329                                 /* -EDEADLK */
1330                                 this->pi_state = NULL;
1331                                 free_pi_state(pi_state);
1332                                 goto out_unlock;
1333                         }
1334                 }
1335                 requeue_futex(this, hb1, hb2, &key2);
1336                 drop_count++;
1337         }
1338
1339 out_unlock:
1340         double_unlock_hb(hb1, hb2);
1341
1342         /*
1343          * drop_futex_key_refs() must be called outside the spinlocks. During
1344          * the requeue we moved futex_q's from the hash bucket at key1 to the
1345          * one at key2 and updated their key pointer.  We no longer need to
1346          * hold the references to key1.
1347          */
1348         while (--drop_count >= 0)
1349                 drop_futex_key_refs(&key1);
1350
1351 out_put_keys:
1352         put_futex_key(fshared, &key2);
1353 out_put_key1:
1354         put_futex_key(fshared, &key1);
1355 out:
1356         if (pi_state != NULL)
1357                 free_pi_state(pi_state);
1358         return ret ? ret : task_count;
1359 }
1360
1361 /* The key must be already stored in q->key. */
1362 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1363 {
1364         struct futex_hash_bucket *hb;
1365
1366         hb = hash_futex(&q->key);
1367         q->lock_ptr = &hb->lock;
1368
1369         spin_lock(&hb->lock);
1370         return hb;
1371 }
1372
1373 static inline void
1374 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1375 {
1376         spin_unlock(&hb->lock);
1377 }
1378
1379 /**
1380  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1381  * @q:  The futex_q to enqueue
1382  * @hb: The destination hash bucket
1383  *
1384  * The hb->lock must be held by the caller, and is released here. A call to
1385  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1386  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1387  * or nothing if the unqueue is done as part of the wake process and the unqueue
1388  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1389  * an example).
1390  */
1391 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1392 {
1393         int prio;
1394
1395         /*
1396          * The priority used to register this element is
1397          * - either the real thread-priority for the real-time threads
1398          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1399          * - or MAX_RT_PRIO for non-RT threads.
1400          * Thus, all RT-threads are woken first in priority order, and
1401          * the others are woken last, in FIFO order.
1402          */
1403         prio = min(current->normal_prio, MAX_RT_PRIO);
1404
1405         plist_node_init(&q->list, prio);
1406 #ifdef CONFIG_DEBUG_PI_LIST
1407         q->list.plist.spinlock = &hb->lock;
1408 #endif
1409         plist_add(&q->list, &hb->chain);
1410         q->task = current;
1411         spin_unlock(&hb->lock);
1412 }
1413
1414 /**
1415  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1416  * @q:  The futex_q to unqueue
1417  *
1418  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1419  * be paired with exactly one earlier call to queue_me().
1420  *
1421  * Returns:
1422  *   1 - if the futex_q was still queued (and we removed unqueued it)
1423  *   0 - if the futex_q was already removed by the waking thread
1424  */
1425 static int unqueue_me(struct futex_q *q)
1426 {
1427         spinlock_t *lock_ptr;
1428         int ret = 0;
1429
1430         /* In the common case we don't take the spinlock, which is nice. */
1431 retry:
1432         lock_ptr = q->lock_ptr;
1433         barrier();
1434         if (lock_ptr != NULL) {
1435                 spin_lock(lock_ptr);
1436                 /*
1437                  * q->lock_ptr can change between reading it and
1438                  * spin_lock(), causing us to take the wrong lock.  This
1439                  * corrects the race condition.
1440                  *
1441                  * Reasoning goes like this: if we have the wrong lock,
1442                  * q->lock_ptr must have changed (maybe several times)
1443                  * between reading it and the spin_lock().  It can
1444                  * change again after the spin_lock() but only if it was
1445                  * already changed before the spin_lock().  It cannot,
1446                  * however, change back to the original value.  Therefore
1447                  * we can detect whether we acquired the correct lock.
1448                  */
1449                 if (unlikely(lock_ptr != q->lock_ptr)) {
1450                         spin_unlock(lock_ptr);
1451                         goto retry;
1452                 }
1453                 WARN_ON(plist_node_empty(&q->list));
1454                 plist_del(&q->list, &q->list.plist);
1455
1456                 BUG_ON(q->pi_state);
1457
1458                 spin_unlock(lock_ptr);
1459                 ret = 1;
1460         }
1461
1462         drop_futex_key_refs(&q->key);
1463         return ret;
1464 }
1465
1466 /*
1467  * PI futexes can not be requeued and must remove themself from the
1468  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1469  * and dropped here.
1470  */
1471 static void unqueue_me_pi(struct futex_q *q)
1472 {
1473         WARN_ON(plist_node_empty(&q->list));
1474         plist_del(&q->list, &q->list.plist);
1475
1476         BUG_ON(!q->pi_state);
1477         free_pi_state(q->pi_state);
1478         q->pi_state = NULL;
1479
1480         spin_unlock(q->lock_ptr);
1481 }
1482
1483 /*
1484  * Fixup the pi_state owner with the new owner.
1485  *
1486  * Must be called with hash bucket lock held and mm->sem held for non
1487  * private futexes.
1488  */
1489 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1490                                 struct task_struct *newowner, int fshared)
1491 {
1492         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1493         struct futex_pi_state *pi_state = q->pi_state;
1494         struct task_struct *oldowner = pi_state->owner;
1495         u32 uval, curval, newval;
1496         int ret;
1497
1498         /* Owner died? */
1499         if (!pi_state->owner)
1500                 newtid |= FUTEX_OWNER_DIED;
1501
1502         /*
1503          * We are here either because we stole the rtmutex from the
1504          * pending owner or we are the pending owner which failed to
1505          * get the rtmutex. We have to replace the pending owner TID
1506          * in the user space variable. This must be atomic as we have
1507          * to preserve the owner died bit here.
1508          *
1509          * Note: We write the user space value _before_ changing the pi_state
1510          * because we can fault here. Imagine swapped out pages or a fork
1511          * that marked all the anonymous memory readonly for cow.
1512          *
1513          * Modifying pi_state _before_ the user space value would
1514          * leave the pi_state in an inconsistent state when we fault
1515          * here, because we need to drop the hash bucket lock to
1516          * handle the fault. This might be observed in the PID check
1517          * in lookup_pi_state.
1518          */
1519 retry:
1520         if (get_futex_value_locked(&uval, uaddr))
1521                 goto handle_fault;
1522
1523         while (1) {
1524                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1525
1526                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1527
1528                 if (curval == -EFAULT)
1529                         goto handle_fault;
1530                 if (curval == uval)
1531                         break;
1532                 uval = curval;
1533         }
1534
1535         /*
1536          * We fixed up user space. Now we need to fix the pi_state
1537          * itself.
1538          */
1539         if (pi_state->owner != NULL) {
1540                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1541                 WARN_ON(list_empty(&pi_state->list));
1542                 list_del_init(&pi_state->list);
1543                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1544         }
1545
1546         pi_state->owner = newowner;
1547
1548         raw_spin_lock_irq(&newowner->pi_lock);
1549         WARN_ON(!list_empty(&pi_state->list));
1550         list_add(&pi_state->list, &newowner->pi_state_list);
1551         raw_spin_unlock_irq(&newowner->pi_lock);
1552         return 0;
1553
1554         /*
1555          * To handle the page fault we need to drop the hash bucket
1556          * lock here. That gives the other task (either the pending
1557          * owner itself or the task which stole the rtmutex) the
1558          * chance to try the fixup of the pi_state. So once we are
1559          * back from handling the fault we need to check the pi_state
1560          * after reacquiring the hash bucket lock and before trying to
1561          * do another fixup. When the fixup has been done already we
1562          * simply return.
1563          */
1564 handle_fault:
1565         spin_unlock(q->lock_ptr);
1566
1567         ret = fault_in_user_writeable(uaddr);
1568
1569         spin_lock(q->lock_ptr);
1570
1571         /*
1572          * Check if someone else fixed it for us:
1573          */
1574         if (pi_state->owner != oldowner)
1575                 return 0;
1576
1577         if (ret)
1578                 return ret;
1579
1580         goto retry;
1581 }
1582
1583 /*
1584  * In case we must use restart_block to restart a futex_wait,
1585  * we encode in the 'flags' shared capability
1586  */
1587 #define FLAGS_SHARED            0x01
1588 #define FLAGS_CLOCKRT           0x02
1589 #define FLAGS_HAS_TIMEOUT       0x04
1590
1591 static long futex_wait_restart(struct restart_block *restart);
1592
1593 /**
1594  * fixup_owner() - Post lock pi_state and corner case management
1595  * @uaddr:      user address of the futex
1596  * @fshared:    whether the futex is shared (1) or not (0)
1597  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1598  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1599  *
1600  * After attempting to lock an rt_mutex, this function is called to cleanup
1601  * the pi_state owner as well as handle race conditions that may allow us to
1602  * acquire the lock. Must be called with the hb lock held.
1603  *
1604  * Returns:
1605  *  1 - success, lock taken
1606  *  0 - success, lock not taken
1607  * <0 - on error (-EFAULT)
1608  */
1609 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1610                        int locked)
1611 {
1612         struct task_struct *owner;
1613         int ret = 0;
1614
1615         if (locked) {
1616                 /*
1617                  * Got the lock. We might not be the anticipated owner if we
1618                  * did a lock-steal - fix up the PI-state in that case:
1619                  */
1620                 if (q->pi_state->owner != current)
1621                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1622                 goto out;
1623         }
1624
1625         /*
1626          * Catch the rare case, where the lock was released when we were on the
1627          * way back before we locked the hash bucket.
1628          */
1629         if (q->pi_state->owner == current) {
1630                 /*
1631                  * Try to get the rt_mutex now. This might fail as some other
1632                  * task acquired the rt_mutex after we removed ourself from the
1633                  * rt_mutex waiters list.
1634                  */
1635                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1636                         locked = 1;
1637                         goto out;
1638                 }
1639
1640                 /*
1641                  * pi_state is incorrect, some other task did a lock steal and
1642                  * we returned due to timeout or signal without taking the
1643                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1644                  * locking, as the other task is now blocked on the hash bucket
1645                  * lock. Fix the state up.
1646                  */
1647                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1648                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1649                 goto out;
1650         }
1651
1652         /*
1653          * Paranoia check. If we did not take the lock, then we should not be
1654          * the owner, nor the pending owner, of the rt_mutex.
1655          */
1656         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1657                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1658                                 "pi-state %p\n", ret,
1659                                 q->pi_state->pi_mutex.owner,
1660                                 q->pi_state->owner);
1661
1662 out:
1663         return ret ? ret : locked;
1664 }
1665
1666 /**
1667  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1668  * @hb:         the futex hash bucket, must be locked by the caller
1669  * @q:          the futex_q to queue up on
1670  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1671  */
1672 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1673                                 struct hrtimer_sleeper *timeout)
1674 {
1675         /*
1676          * The task state is guaranteed to be set before another task can
1677          * wake it. set_current_state() is implemented using set_mb() and
1678          * queue_me() calls spin_unlock() upon completion, both serializing
1679          * access to the hash list and forcing another memory barrier.
1680          */
1681         set_current_state(TASK_INTERRUPTIBLE);
1682         queue_me(q, hb);
1683
1684         /* Arm the timer */
1685         if (timeout) {
1686                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1687                 if (!hrtimer_active(&timeout->timer))
1688                         timeout->task = NULL;
1689         }
1690
1691         /*
1692          * If we have been removed from the hash list, then another task
1693          * has tried to wake us, and we can skip the call to schedule().
1694          */
1695         if (likely(!plist_node_empty(&q->list))) {
1696                 /*
1697                  * If the timer has already expired, current will already be
1698                  * flagged for rescheduling. Only call schedule if there
1699                  * is no timeout, or if it has yet to expire.
1700                  */
1701                 if (!timeout || timeout->task)
1702                         schedule();
1703         }
1704         __set_current_state(TASK_RUNNING);
1705 }
1706
1707 /**
1708  * futex_wait_setup() - Prepare to wait on a futex
1709  * @uaddr:      the futex userspace address
1710  * @val:        the expected value
1711  * @fshared:    whether the futex is shared (1) or not (0)
1712  * @q:          the associated futex_q
1713  * @hb:         storage for hash_bucket pointer to be returned to caller
1714  *
1715  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1716  * compare it with the expected value.  Handle atomic faults internally.
1717  * Return with the hb lock held and a q.key reference on success, and unlocked
1718  * with no q.key reference on failure.
1719  *
1720  * Returns:
1721  *  0 - uaddr contains val and hb has been locked
1722  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1723  */
1724 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1725                            struct futex_q *q, struct futex_hash_bucket **hb)
1726 {
1727         u32 uval;
1728         int ret;
1729
1730         /*
1731          * Access the page AFTER the hash-bucket is locked.
1732          * Order is important:
1733          *
1734          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1735          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1736          *
1737          * The basic logical guarantee of a futex is that it blocks ONLY
1738          * if cond(var) is known to be true at the time of blocking, for
1739          * any cond.  If we queued after testing *uaddr, that would open
1740          * a race condition where we could block indefinitely with
1741          * cond(var) false, which would violate the guarantee.
1742          *
1743          * A consequence is that futex_wait() can return zero and absorb
1744          * a wakeup when *uaddr != val on entry to the syscall.  This is
1745          * rare, but normal.
1746          */
1747 retry:
1748         q->key = FUTEX_KEY_INIT;
1749         ret = get_futex_key(uaddr, fshared, &q->key);
1750         if (unlikely(ret != 0))
1751                 return ret;
1752
1753 retry_private:
1754         *hb = queue_lock(q);
1755
1756         ret = get_futex_value_locked(&uval, uaddr);
1757
1758         if (ret) {
1759                 queue_unlock(q, *hb);
1760
1761                 ret = get_user(uval, uaddr);
1762                 if (ret)
1763                         goto out;
1764
1765                 if (!fshared)
1766                         goto retry_private;
1767
1768                 put_futex_key(fshared, &q->key);
1769                 goto retry;
1770         }
1771
1772         if (uval != val) {
1773                 queue_unlock(q, *hb);
1774                 ret = -EWOULDBLOCK;
1775         }
1776
1777 out:
1778         if (ret)
1779                 put_futex_key(fshared, &q->key);
1780         return ret;
1781 }
1782
1783 static int futex_wait(u32 __user *uaddr, int fshared,
1784                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1785 {
1786         struct hrtimer_sleeper timeout, *to = NULL;
1787         struct restart_block *restart;
1788         struct futex_hash_bucket *hb;
1789         struct futex_q q;
1790         int ret;
1791
1792         if (!bitset)
1793                 return -EINVAL;
1794
1795         q.pi_state = NULL;
1796         q.bitset = bitset;
1797         q.rt_waiter = NULL;
1798         q.requeue_pi_key = NULL;
1799
1800         if (abs_time) {
1801                 to = &timeout;
1802
1803                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1804                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1805                 hrtimer_init_sleeper(to, current);
1806                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1807                                              current->timer_slack_ns);
1808         }
1809
1810 retry:
1811         /*
1812          * Prepare to wait on uaddr. On success, holds hb lock and increments
1813          * q.key refs.
1814          */
1815         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1816         if (ret)
1817                 goto out;
1818
1819         /* queue_me and wait for wakeup, timeout, or a signal. */
1820         futex_wait_queue_me(hb, &q, to);
1821
1822         /* If we were woken (and unqueued), we succeeded, whatever. */
1823         ret = 0;
1824         /* unqueue_me() drops q.key ref */
1825         if (!unqueue_me(&q))
1826                 goto out;
1827         ret = -ETIMEDOUT;
1828         if (to && !to->task)
1829                 goto out;
1830
1831         /*
1832          * We expect signal_pending(current), but we might be the
1833          * victim of a spurious wakeup as well.
1834          */
1835         if (!signal_pending(current))
1836                 goto retry;
1837
1838         ret = -ERESTARTSYS;
1839         if (!abs_time)
1840                 goto out;
1841
1842         restart = &current_thread_info()->restart_block;
1843         restart->fn = futex_wait_restart;
1844         restart->futex.uaddr = (u32 *)uaddr;
1845         restart->futex.val = val;
1846         restart->futex.time = abs_time->tv64;
1847         restart->futex.bitset = bitset;
1848         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1849
1850         if (fshared)
1851                 restart->futex.flags |= FLAGS_SHARED;
1852         if (clockrt)
1853                 restart->futex.flags |= FLAGS_CLOCKRT;
1854
1855         ret = -ERESTART_RESTARTBLOCK;
1856
1857 out:
1858         if (to) {
1859                 hrtimer_cancel(&to->timer);
1860                 destroy_hrtimer_on_stack(&to->timer);
1861         }
1862         return ret;
1863 }
1864
1865
1866 static long futex_wait_restart(struct restart_block *restart)
1867 {
1868         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1869         int fshared = 0;
1870         ktime_t t, *tp = NULL;
1871
1872         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1873                 t.tv64 = restart->futex.time;
1874                 tp = &t;
1875         }
1876         restart->fn = do_no_restart_syscall;
1877         if (restart->futex.flags & FLAGS_SHARED)
1878                 fshared = 1;
1879         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1880                                 restart->futex.bitset,
1881                                 restart->futex.flags & FLAGS_CLOCKRT);
1882 }
1883
1884
1885 /*
1886  * Userspace tried a 0 -> TID atomic transition of the futex value
1887  * and failed. The kernel side here does the whole locking operation:
1888  * if there are waiters then it will block, it does PI, etc. (Due to
1889  * races the kernel might see a 0 value of the futex too.)
1890  */
1891 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1892                          int detect, ktime_t *time, int trylock)
1893 {
1894         struct hrtimer_sleeper timeout, *to = NULL;
1895         struct futex_hash_bucket *hb;
1896         struct futex_q q;
1897         int res, ret;
1898
1899         if (refill_pi_state_cache())
1900                 return -ENOMEM;
1901
1902         if (time) {
1903                 to = &timeout;
1904                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1905                                       HRTIMER_MODE_ABS);
1906                 hrtimer_init_sleeper(to, current);
1907                 hrtimer_set_expires(&to->timer, *time);
1908         }
1909
1910         q.pi_state = NULL;
1911         q.rt_waiter = NULL;
1912         q.requeue_pi_key = NULL;
1913 retry:
1914         q.key = FUTEX_KEY_INIT;
1915         ret = get_futex_key(uaddr, fshared, &q.key);
1916         if (unlikely(ret != 0))
1917                 goto out;
1918
1919 retry_private:
1920         hb = queue_lock(&q);
1921
1922         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1923         if (unlikely(ret)) {
1924                 switch (ret) {
1925                 case 1:
1926                         /* We got the lock. */
1927                         ret = 0;
1928                         goto out_unlock_put_key;
1929                 case -EFAULT:
1930                         goto uaddr_faulted;
1931                 case -EAGAIN:
1932                         /*
1933                          * Task is exiting and we just wait for the
1934                          * exit to complete.
1935                          */
1936                         queue_unlock(&q, hb);
1937                         put_futex_key(fshared, &q.key);
1938                         cond_resched();
1939                         goto retry;
1940                 default:
1941                         goto out_unlock_put_key;
1942                 }
1943         }
1944
1945         /*
1946          * Only actually queue now that the atomic ops are done:
1947          */
1948         queue_me(&q, hb);
1949
1950         WARN_ON(!q.pi_state);
1951         /*
1952          * Block on the PI mutex:
1953          */
1954         if (!trylock)
1955                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1956         else {
1957                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1958                 /* Fixup the trylock return value: */
1959                 ret = ret ? 0 : -EWOULDBLOCK;
1960         }
1961
1962         spin_lock(q.lock_ptr);
1963         /*
1964          * Fixup the pi_state owner and possibly acquire the lock if we
1965          * haven't already.
1966          */
1967         res = fixup_owner(uaddr, fshared, &q, !ret);
1968         /*
1969          * If fixup_owner() returned an error, proprogate that.  If it acquired
1970          * the lock, clear our -ETIMEDOUT or -EINTR.
1971          */
1972         if (res)
1973                 ret = (res < 0) ? res : 0;
1974
1975         /*
1976          * If fixup_owner() faulted and was unable to handle the fault, unlock
1977          * it and return the fault to userspace.
1978          */
1979         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1980                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1981
1982         /* Unqueue and drop the lock */
1983         unqueue_me_pi(&q);
1984
1985         goto out_put_key;
1986
1987 out_unlock_put_key:
1988         queue_unlock(&q, hb);
1989
1990 out_put_key:
1991         put_futex_key(fshared, &q.key);
1992 out:
1993         if (to)
1994                 destroy_hrtimer_on_stack(&to->timer);
1995         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1996
1997 uaddr_faulted:
1998         queue_unlock(&q, hb);
1999
2000         ret = fault_in_user_writeable(uaddr);
2001         if (ret)
2002                 goto out_put_key;
2003
2004         if (!fshared)
2005                 goto retry_private;
2006
2007         put_futex_key(fshared, &q.key);
2008         goto retry;
2009 }
2010
2011 /*
2012  * Userspace attempted a TID -> 0 atomic transition, and failed.
2013  * This is the in-kernel slowpath: we look up the PI state (if any),
2014  * and do the rt-mutex unlock.
2015  */
2016 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2017 {
2018         struct futex_hash_bucket *hb;
2019         struct futex_q *this, *next;
2020         u32 uval;
2021         struct plist_head *head;
2022         union futex_key key = FUTEX_KEY_INIT;
2023         int ret;
2024
2025 retry:
2026         if (get_user(uval, uaddr))
2027                 return -EFAULT;
2028         /*
2029          * We release only a lock we actually own:
2030          */
2031         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2032                 return -EPERM;
2033
2034         ret = get_futex_key(uaddr, fshared, &key);
2035         if (unlikely(ret != 0))
2036                 goto out;
2037
2038         hb = hash_futex(&key);
2039         spin_lock(&hb->lock);
2040
2041         /*
2042          * To avoid races, try to do the TID -> 0 atomic transition
2043          * again. If it succeeds then we can return without waking
2044          * anyone else up:
2045          */
2046         if (!(uval & FUTEX_OWNER_DIED))
2047                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2048
2049
2050         if (unlikely(uval == -EFAULT))
2051                 goto pi_faulted;
2052         /*
2053          * Rare case: we managed to release the lock atomically,
2054          * no need to wake anyone else up:
2055          */
2056         if (unlikely(uval == task_pid_vnr(current)))
2057                 goto out_unlock;
2058
2059         /*
2060          * Ok, other tasks may need to be woken up - check waiters
2061          * and do the wakeup if necessary:
2062          */
2063         head = &hb->chain;
2064
2065         plist_for_each_entry_safe(this, next, head, list) {
2066                 if (!match_futex (&this->key, &key))
2067                         continue;
2068                 ret = wake_futex_pi(uaddr, uval, this);
2069                 /*
2070                  * The atomic access to the futex value
2071                  * generated a pagefault, so retry the
2072                  * user-access and the wakeup:
2073                  */
2074                 if (ret == -EFAULT)
2075                         goto pi_faulted;
2076                 goto out_unlock;
2077         }
2078         /*
2079          * No waiters - kernel unlocks the futex:
2080          */
2081         if (!(uval & FUTEX_OWNER_DIED)) {
2082                 ret = unlock_futex_pi(uaddr, uval);
2083                 if (ret == -EFAULT)
2084                         goto pi_faulted;
2085         }
2086
2087 out_unlock:
2088         spin_unlock(&hb->lock);
2089         put_futex_key(fshared, &key);
2090
2091 out:
2092         return ret;
2093
2094 pi_faulted:
2095         spin_unlock(&hb->lock);
2096         put_futex_key(fshared, &key);
2097
2098         ret = fault_in_user_writeable(uaddr);
2099         if (!ret)
2100                 goto retry;
2101
2102         return ret;
2103 }
2104
2105 /**
2106  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2107  * @hb:         the hash_bucket futex_q was original enqueued on
2108  * @q:          the futex_q woken while waiting to be requeued
2109  * @key2:       the futex_key of the requeue target futex
2110  * @timeout:    the timeout associated with the wait (NULL if none)
2111  *
2112  * Detect if the task was woken on the initial futex as opposed to the requeue
2113  * target futex.  If so, determine if it was a timeout or a signal that caused
2114  * the wakeup and return the appropriate error code to the caller.  Must be
2115  * called with the hb lock held.
2116  *
2117  * Returns
2118  *  0 - no early wakeup detected
2119  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2120  */
2121 static inline
2122 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2123                                    struct futex_q *q, union futex_key *key2,
2124                                    struct hrtimer_sleeper *timeout)
2125 {
2126         int ret = 0;
2127
2128         /*
2129          * With the hb lock held, we avoid races while we process the wakeup.
2130          * We only need to hold hb (and not hb2) to ensure atomicity as the
2131          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2132          * It can't be requeued from uaddr2 to something else since we don't
2133          * support a PI aware source futex for requeue.
2134          */
2135         if (!match_futex(&q->key, key2)) {
2136                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2137                 /*
2138                  * We were woken prior to requeue by a timeout or a signal.
2139                  * Unqueue the futex_q and determine which it was.
2140                  */
2141                 plist_del(&q->list, &q->list.plist);
2142
2143                 /* Handle spurious wakeups gracefully */
2144                 ret = -EWOULDBLOCK;
2145                 if (timeout && !timeout->task)
2146                         ret = -ETIMEDOUT;
2147                 else if (signal_pending(current))
2148                         ret = -ERESTARTNOINTR;
2149         }
2150         return ret;
2151 }
2152
2153 /**
2154  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2155  * @uaddr:      the futex we initially wait on (non-pi)
2156  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2157  *              the same type, no requeueing from private to shared, etc.
2158  * @val:        the expected value of uaddr
2159  * @abs_time:   absolute timeout
2160  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2161  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2162  * @uaddr2:     the pi futex we will take prior to returning to user-space
2163  *
2164  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2165  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2166  * complete the acquisition of the rt_mutex prior to returning to userspace.
2167  * This ensures the rt_mutex maintains an owner when it has waiters; without
2168  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2169  * need to.
2170  *
2171  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2172  * via the following:
2173  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2174  * 2) wakeup on uaddr2 after a requeue
2175  * 3) signal
2176  * 4) timeout
2177  *
2178  * If 3, cleanup and return -ERESTARTNOINTR.
2179  *
2180  * If 2, we may then block on trying to take the rt_mutex and return via:
2181  * 5) successful lock
2182  * 6) signal
2183  * 7) timeout
2184  * 8) other lock acquisition failure
2185  *
2186  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2187  *
2188  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2189  *
2190  * Returns:
2191  *  0 - On success
2192  * <0 - On error
2193  */
2194 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2195                                  u32 val, ktime_t *abs_time, u32 bitset,
2196                                  int clockrt, u32 __user *uaddr2)
2197 {
2198         struct hrtimer_sleeper timeout, *to = NULL;
2199         struct rt_mutex_waiter rt_waiter;
2200         struct rt_mutex *pi_mutex = NULL;
2201         struct futex_hash_bucket *hb;
2202         union futex_key key2;
2203         struct futex_q q;
2204         int res, ret;
2205
2206         if (!bitset)
2207                 return -EINVAL;
2208
2209         if (abs_time) {
2210                 to = &timeout;
2211                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2212                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2213                 hrtimer_init_sleeper(to, current);
2214                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2215                                              current->timer_slack_ns);
2216         }
2217
2218         /*
2219          * The waiter is allocated on our stack, manipulated by the requeue
2220          * code while we sleep on uaddr.
2221          */
2222         debug_rt_mutex_init_waiter(&rt_waiter);
2223         rt_waiter.task = NULL;
2224
2225         key2 = FUTEX_KEY_INIT;
2226         ret = get_futex_key(uaddr2, fshared, &key2);
2227         if (unlikely(ret != 0))
2228                 goto out;
2229
2230         q.pi_state = NULL;
2231         q.bitset = bitset;
2232         q.rt_waiter = &rt_waiter;
2233         q.requeue_pi_key = &key2;
2234
2235         /*
2236          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2237          * count.
2238          */
2239         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2240         if (ret)
2241                 goto out_key2;
2242
2243         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2244         futex_wait_queue_me(hb, &q, to);
2245
2246         spin_lock(&hb->lock);
2247         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2248         spin_unlock(&hb->lock);
2249         if (ret)
2250                 goto out_put_keys;
2251
2252         /*
2253          * In order for us to be here, we know our q.key == key2, and since
2254          * we took the hb->lock above, we also know that futex_requeue() has
2255          * completed and we no longer have to concern ourselves with a wakeup
2256          * race with the atomic proxy lock acquisition by the requeue code. The
2257          * futex_requeue dropped our key1 reference and incremented our key2
2258          * reference count.
2259          */
2260
2261         /* Check if the requeue code acquired the second futex for us. */
2262         if (!q.rt_waiter) {
2263                 /*
2264                  * Got the lock. We might not be the anticipated owner if we
2265                  * did a lock-steal - fix up the PI-state in that case.
2266                  */
2267                 if (q.pi_state && (q.pi_state->owner != current)) {
2268                         spin_lock(q.lock_ptr);
2269                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2270                                                    fshared);
2271                         spin_unlock(q.lock_ptr);
2272                 }
2273         } else {
2274                 /*
2275                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2276                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2277                  * the pi_state.
2278                  */
2279                 WARN_ON(!&q.pi_state);
2280                 pi_mutex = &q.pi_state->pi_mutex;
2281                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2282                 debug_rt_mutex_free_waiter(&rt_waiter);
2283
2284                 spin_lock(q.lock_ptr);
2285                 /*
2286                  * Fixup the pi_state owner and possibly acquire the lock if we
2287                  * haven't already.
2288                  */
2289                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2290                 /*
2291                  * If fixup_owner() returned an error, proprogate that.  If it
2292                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2293                  */
2294                 if (res)
2295                         ret = (res < 0) ? res : 0;
2296
2297                 /* Unqueue and drop the lock. */
2298                 unqueue_me_pi(&q);
2299         }
2300
2301         /*
2302          * If fixup_pi_state_owner() faulted and was unable to handle the
2303          * fault, unlock the rt_mutex and return the fault to userspace.
2304          */
2305         if (ret == -EFAULT) {
2306                 if (rt_mutex_owner(pi_mutex) == current)
2307                         rt_mutex_unlock(pi_mutex);
2308         } else if (ret == -EINTR) {
2309                 /*
2310                  * We've already been requeued, but cannot restart by calling
2311                  * futex_lock_pi() directly. We could restart this syscall, but
2312                  * it would detect that the user space "val" changed and return
2313                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2314                  * -EWOULDBLOCK directly.
2315                  */
2316                 ret = -EWOULDBLOCK;
2317         }
2318
2319 out_put_keys:
2320         put_futex_key(fshared, &q.key);
2321 out_key2:
2322         put_futex_key(fshared, &key2);
2323
2324 out:
2325         if (to) {
2326                 hrtimer_cancel(&to->timer);
2327                 destroy_hrtimer_on_stack(&to->timer);
2328         }
2329         return ret;
2330 }
2331
2332 /*
2333  * Support for robust futexes: the kernel cleans up held futexes at
2334  * thread exit time.
2335  *
2336  * Implementation: user-space maintains a per-thread list of locks it
2337  * is holding. Upon do_exit(), the kernel carefully walks this list,
2338  * and marks all locks that are owned by this thread with the
2339  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2340  * always manipulated with the lock held, so the list is private and
2341  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2342  * field, to allow the kernel to clean up if the thread dies after
2343  * acquiring the lock, but just before it could have added itself to
2344  * the list. There can only be one such pending lock.
2345  */
2346
2347 /**
2348  * sys_set_robust_list() - Set the robust-futex list head of a task
2349  * @head:       pointer to the list-head
2350  * @len:        length of the list-head, as userspace expects
2351  */
2352 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2353                 size_t, len)
2354 {
2355         if (!futex_cmpxchg_enabled)
2356                 return -ENOSYS;
2357         /*
2358          * The kernel knows only one size for now:
2359          */
2360         if (unlikely(len != sizeof(*head)))
2361                 return -EINVAL;
2362
2363         current->robust_list = head;
2364
2365         return 0;
2366 }
2367
2368 /**
2369  * sys_get_robust_list() - Get the robust-futex list head of a task
2370  * @pid:        pid of the process [zero for current task]
2371  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2372  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2373  */
2374 SYSCALL_DEFINE3(get_robust_list, int, pid,
2375                 struct robust_list_head __user * __user *, head_ptr,
2376                 size_t __user *, len_ptr)
2377 {
2378         struct robust_list_head __user *head;
2379         unsigned long ret;
2380         const struct cred *cred = current_cred(), *pcred;
2381
2382         if (!futex_cmpxchg_enabled)
2383                 return -ENOSYS;
2384
2385         if (!pid)
2386                 head = current->robust_list;
2387         else {
2388                 struct task_struct *p;
2389
2390                 ret = -ESRCH;
2391                 rcu_read_lock();
2392                 p = find_task_by_vpid(pid);
2393                 if (!p)
2394                         goto err_unlock;
2395                 ret = -EPERM;
2396                 pcred = __task_cred(p);
2397                 if (cred->euid != pcred->euid &&
2398                     cred->euid != pcred->uid &&
2399                     !capable(CAP_SYS_PTRACE))
2400                         goto err_unlock;
2401                 head = p->robust_list;
2402                 rcu_read_unlock();
2403         }
2404
2405         if (put_user(sizeof(*head), len_ptr))
2406                 return -EFAULT;
2407         return put_user(head, head_ptr);
2408
2409 err_unlock:
2410         rcu_read_unlock();
2411
2412         return ret;
2413 }
2414
2415 /*
2416  * Process a futex-list entry, check whether it's owned by the
2417  * dying task, and do notification if so:
2418  */
2419 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2420 {
2421         u32 uval, nval, mval;
2422
2423 retry:
2424         if (get_user(uval, uaddr))
2425                 return -1;
2426
2427         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2428                 /*
2429                  * Ok, this dying thread is truly holding a futex
2430                  * of interest. Set the OWNER_DIED bit atomically
2431                  * via cmpxchg, and if the value had FUTEX_WAITERS
2432                  * set, wake up a waiter (if any). (We have to do a
2433                  * futex_wake() even if OWNER_DIED is already set -
2434                  * to handle the rare but possible case of recursive
2435                  * thread-death.) The rest of the cleanup is done in
2436                  * userspace.
2437                  */
2438                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2439                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2440
2441                 if (nval == -EFAULT)
2442                         return -1;
2443
2444                 if (nval != uval)
2445                         goto retry;
2446
2447                 /*
2448                  * Wake robust non-PI futexes here. The wakeup of
2449                  * PI futexes happens in exit_pi_state():
2450                  */
2451                 if (!pi && (uval & FUTEX_WAITERS))
2452                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2453         }
2454         return 0;
2455 }
2456
2457 /*
2458  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2459  */
2460 static inline int fetch_robust_entry(struct robust_list __user **entry,
2461                                      struct robust_list __user * __user *head,
2462                                      int *pi)
2463 {
2464         unsigned long uentry;
2465
2466         if (get_user(uentry, (unsigned long __user *)head))
2467                 return -EFAULT;
2468
2469         *entry = (void __user *)(uentry & ~1UL);
2470         *pi = uentry & 1;
2471
2472         return 0;
2473 }
2474
2475 /*
2476  * Walk curr->robust_list (very carefully, it's a userspace list!)
2477  * and mark any locks found there dead, and notify any waiters.
2478  *
2479  * We silently return on any sign of list-walking problem.
2480  */
2481 void exit_robust_list(struct task_struct *curr)
2482 {
2483         struct robust_list_head __user *head = curr->robust_list;
2484         struct robust_list __user *entry, *next_entry, *pending;
2485         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2486         unsigned long futex_offset;
2487         int rc;
2488
2489         if (!futex_cmpxchg_enabled)
2490                 return;
2491
2492         /*
2493          * Fetch the list head (which was registered earlier, via
2494          * sys_set_robust_list()):
2495          */
2496         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2497                 return;
2498         /*
2499          * Fetch the relative futex offset:
2500          */
2501         if (get_user(futex_offset, &head->futex_offset))
2502                 return;
2503         /*
2504          * Fetch any possibly pending lock-add first, and handle it
2505          * if it exists:
2506          */
2507         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2508                 return;
2509
2510         next_entry = NULL;      /* avoid warning with gcc */
2511         while (entry != &head->list) {
2512                 /*
2513                  * Fetch the next entry in the list before calling
2514                  * handle_futex_death:
2515                  */
2516                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2517                 /*
2518                  * A pending lock might already be on the list, so
2519                  * don't process it twice:
2520                  */
2521                 if (entry != pending)
2522                         if (handle_futex_death((void __user *)entry + futex_offset,
2523                                                 curr, pi))
2524                                 return;
2525                 if (rc)
2526                         return;
2527                 entry = next_entry;
2528                 pi = next_pi;
2529                 /*
2530                  * Avoid excessively long or circular lists:
2531                  */
2532                 if (!--limit)
2533                         break;
2534
2535                 cond_resched();
2536         }
2537
2538         if (pending)
2539                 handle_futex_death((void __user *)pending + futex_offset,
2540                                    curr, pip);
2541 }
2542
2543 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2544                 u32 __user *uaddr2, u32 val2, u32 val3)
2545 {
2546         int clockrt, ret = -ENOSYS;
2547         int cmd = op & FUTEX_CMD_MASK;
2548         int fshared = 0;
2549
2550         if (!(op & FUTEX_PRIVATE_FLAG))
2551                 fshared = 1;
2552
2553         clockrt = op & FUTEX_CLOCK_REALTIME;
2554         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2555                 return -ENOSYS;
2556
2557         switch (cmd) {
2558         case FUTEX_WAIT:
2559                 val3 = FUTEX_BITSET_MATCH_ANY;
2560         case FUTEX_WAIT_BITSET:
2561                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2562                 break;
2563         case FUTEX_WAKE:
2564                 val3 = FUTEX_BITSET_MATCH_ANY;
2565         case FUTEX_WAKE_BITSET:
2566                 ret = futex_wake(uaddr, fshared, val, val3);
2567                 break;
2568         case FUTEX_REQUEUE:
2569                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2570                 break;
2571         case FUTEX_CMP_REQUEUE:
2572                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2573                                     0);
2574                 break;
2575         case FUTEX_WAKE_OP:
2576                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2577                 break;
2578         case FUTEX_LOCK_PI:
2579                 if (futex_cmpxchg_enabled)
2580                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2581                 break;
2582         case FUTEX_UNLOCK_PI:
2583                 if (futex_cmpxchg_enabled)
2584                         ret = futex_unlock_pi(uaddr, fshared);
2585                 break;
2586         case FUTEX_TRYLOCK_PI:
2587                 if (futex_cmpxchg_enabled)
2588                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2589                 break;
2590         case FUTEX_WAIT_REQUEUE_PI:
2591                 val3 = FUTEX_BITSET_MATCH_ANY;
2592                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2593                                             clockrt, uaddr2);
2594                 break;
2595         case FUTEX_CMP_REQUEUE_PI:
2596                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2597                                     1);
2598                 break;
2599         default:
2600                 ret = -ENOSYS;
2601         }
2602         return ret;
2603 }
2604
2605
2606 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2607                 struct timespec __user *, utime, u32 __user *, uaddr2,
2608                 u32, val3)
2609 {
2610         struct timespec ts;
2611         ktime_t t, *tp = NULL;
2612         u32 val2 = 0;
2613         int cmd = op & FUTEX_CMD_MASK;
2614
2615         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2616                       cmd == FUTEX_WAIT_BITSET ||
2617                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2618                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2619                         return -EFAULT;
2620                 if (!timespec_valid(&ts))
2621                         return -EINVAL;
2622
2623                 t = timespec_to_ktime(ts);
2624                 if (cmd == FUTEX_WAIT)
2625                         t = ktime_add_safe(ktime_get(), t);
2626                 tp = &t;
2627         }
2628         /*
2629          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2630          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2631          */
2632         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2633             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2634                 val2 = (u32) (unsigned long) utime;
2635
2636         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2637 }
2638
2639 static int __init futex_init(void)
2640 {
2641         u32 curval;
2642         int i;
2643
2644         /*
2645          * This will fail and we want it. Some arch implementations do
2646          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2647          * functionality. We want to know that before we call in any
2648          * of the complex code paths. Also we want to prevent
2649          * registration of robust lists in that case. NULL is
2650          * guaranteed to fault and we get -EFAULT on functional
2651          * implementation, the non functional ones will return
2652          * -ENOSYS.
2653          */
2654         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2655         if (curval == -EFAULT)
2656                 futex_cmpxchg_enabled = 1;
2657
2658         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2659                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2660                 spin_lock_init(&futex_queues[i].lock);
2661         }
2662
2663         return 0;
2664 }
2665 __initcall(futex_init);