Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael...
[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  * @list:               priority-sorted list of tasks waiting on this futex
95  * @task:               the task waiting on the futex
96  * @lock_ptr:           the hash bucket lock
97  * @key:                the key the futex is hashed on
98  * @pi_state:           optional priority inheritance state
99  * @rt_waiter:          rt_waiter storage for use with requeue_pi
100  * @requeue_pi_key:     the requeue_pi target futex key
101  * @bitset:             bitset for the optional bitmasked wakeup
102  *
103  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
104  * we can wake only the relevant ones (hashed queues may be shared).
105  *
106  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
107  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
108  * The order of wakeup is always to make the first condition true, then
109  * the second.
110  *
111  * PI futexes are typically woken before they are removed from the hash list via
112  * the rt_mutex code. See unqueue_me_pi().
113  */
114 struct futex_q {
115         struct plist_node list;
116
117         struct task_struct *task;
118         spinlock_t *lock_ptr;
119         union futex_key key;
120         struct futex_pi_state *pi_state;
121         struct rt_mutex_waiter *rt_waiter;
122         union futex_key *requeue_pi_key;
123         u32 bitset;
124 };
125
126 /*
127  * Hash buckets are shared by all the futex_keys that hash to the same
128  * location.  Each key may have multiple futex_q structures, one for each task
129  * waiting on a futex.
130  */
131 struct futex_hash_bucket {
132         spinlock_t lock;
133         struct plist_head chain;
134 };
135
136 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
137
138 /*
139  * We hash on the keys returned from get_futex_key (see below).
140  */
141 static struct futex_hash_bucket *hash_futex(union futex_key *key)
142 {
143         u32 hash = jhash2((u32*)&key->both.word,
144                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
145                           key->both.offset);
146         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
147 }
148
149 /*
150  * Return 1 if two futex_keys are equal, 0 otherwise.
151  */
152 static inline int match_futex(union futex_key *key1, union futex_key *key2)
153 {
154         return (key1 && key2
155                 && key1->both.word == key2->both.word
156                 && key1->both.ptr == key2->both.ptr
157                 && key1->both.offset == key2->both.offset);
158 }
159
160 /*
161  * Take a reference to the resource addressed by a key.
162  * Can be called while holding spinlocks.
163  *
164  */
165 static void get_futex_key_refs(union futex_key *key)
166 {
167         if (!key->both.ptr)
168                 return;
169
170         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
171         case FUT_OFF_INODE:
172                 atomic_inc(&key->shared.inode->i_count);
173                 break;
174         case FUT_OFF_MMSHARED:
175                 atomic_inc(&key->private.mm->mm_count);
176                 break;
177         }
178 }
179
180 /*
181  * Drop a reference to the resource addressed by a key.
182  * The hash bucket spinlock must not be held.
183  */
184 static void drop_futex_key_refs(union futex_key *key)
185 {
186         if (!key->both.ptr) {
187                 /* If we're here then we tried to put a key we failed to get */
188                 WARN_ON_ONCE(1);
189                 return;
190         }
191
192         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
193         case FUT_OFF_INODE:
194                 iput(key->shared.inode);
195                 break;
196         case FUT_OFF_MMSHARED:
197                 mmdrop(key->private.mm);
198                 break;
199         }
200 }
201
202 /**
203  * get_futex_key() - Get parameters which are the keys for a futex
204  * @uaddr:      virtual address of the futex
205  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
206  * @key:        address where result is stored.
207  *
208  * Returns a negative error code or 0
209  * The key words are stored in *key on success.
210  *
211  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
212  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
213  * We can usually work out the index without swapping in the page.
214  *
215  * lock_page() might sleep, the caller should not hold a spinlock.
216  */
217 static int
218 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
219 {
220         unsigned long address = (unsigned long)uaddr;
221         struct mm_struct *mm = current->mm;
222         struct page *page;
223         int err;
224
225         /*
226          * The futex address must be "naturally" aligned.
227          */
228         key->both.offset = address % PAGE_SIZE;
229         if (unlikely((address % sizeof(u32)) != 0))
230                 return -EINVAL;
231         address -= key->both.offset;
232
233         /*
234          * PROCESS_PRIVATE futexes are fast.
235          * As the mm cannot disappear under us and the 'key' only needs
236          * virtual address, we dont even have to find the underlying vma.
237          * Note : We do have to check 'uaddr' is a valid user address,
238          *        but access_ok() should be faster than find_vma()
239          */
240         if (!fshared) {
241                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
242                         return -EFAULT;
243                 key->private.mm = mm;
244                 key->private.address = address;
245                 get_futex_key_refs(key);
246                 return 0;
247         }
248
249 again:
250         err = get_user_pages_fast(address, 1, 1, &page);
251         if (err < 0)
252                 return err;
253
254         page = compound_head(page);
255         lock_page(page);
256         if (!page->mapping) {
257                 unlock_page(page);
258                 put_page(page);
259                 goto again;
260         }
261
262         /*
263          * Private mappings are handled in a simple way.
264          *
265          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
266          * it's a read-only handle, it's expected that futexes attach to
267          * the object not the particular process.
268          */
269         if (PageAnon(page)) {
270                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
271                 key->private.mm = mm;
272                 key->private.address = address;
273         } else {
274                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
275                 key->shared.inode = page->mapping->host;
276                 key->shared.pgoff = page->index;
277         }
278
279         get_futex_key_refs(key);
280
281         unlock_page(page);
282         put_page(page);
283         return 0;
284 }
285
286 static inline
287 void put_futex_key(int fshared, union futex_key *key)
288 {
289         drop_futex_key_refs(key);
290 }
291
292 /**
293  * fault_in_user_writeable() - Fault in user address and verify RW access
294  * @uaddr:      pointer to faulting user space address
295  *
296  * Slow path to fixup the fault we just took in the atomic write
297  * access to @uaddr.
298  *
299  * We have no generic implementation of a non-destructive write to the
300  * user address. We know that we faulted in the atomic pagefault
301  * disabled section so we can as well avoid the #PF overhead by
302  * calling get_user_pages() right away.
303  */
304 static int fault_in_user_writeable(u32 __user *uaddr)
305 {
306         struct mm_struct *mm = current->mm;
307         int ret;
308
309         down_read(&mm->mmap_sem);
310         ret = get_user_pages(current, mm, (unsigned long)uaddr,
311                              1, 1, 0, NULL, NULL);
312         up_read(&mm->mmap_sem);
313
314         return ret < 0 ? ret : 0;
315 }
316
317 /**
318  * futex_top_waiter() - Return the highest priority waiter on a futex
319  * @hb:         the hash bucket the futex_q's reside in
320  * @key:        the futex key (to distinguish it from other futex futex_q's)
321  *
322  * Must be called with the hb lock held.
323  */
324 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
325                                         union futex_key *key)
326 {
327         struct futex_q *this;
328
329         plist_for_each_entry(this, &hb->chain, list) {
330                 if (match_futex(&this->key, key))
331                         return this;
332         }
333         return NULL;
334 }
335
336 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
337 {
338         u32 curval;
339
340         pagefault_disable();
341         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
342         pagefault_enable();
343
344         return curval;
345 }
346
347 static int get_futex_value_locked(u32 *dest, u32 __user *from)
348 {
349         int ret;
350
351         pagefault_disable();
352         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
353         pagefault_enable();
354
355         return ret ? -EFAULT : 0;
356 }
357
358
359 /*
360  * PI code:
361  */
362 static int refill_pi_state_cache(void)
363 {
364         struct futex_pi_state *pi_state;
365
366         if (likely(current->pi_state_cache))
367                 return 0;
368
369         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
370
371         if (!pi_state)
372                 return -ENOMEM;
373
374         INIT_LIST_HEAD(&pi_state->list);
375         /* pi_mutex gets initialized later */
376         pi_state->owner = NULL;
377         atomic_set(&pi_state->refcount, 1);
378         pi_state->key = FUTEX_KEY_INIT;
379
380         current->pi_state_cache = pi_state;
381
382         return 0;
383 }
384
385 static struct futex_pi_state * alloc_pi_state(void)
386 {
387         struct futex_pi_state *pi_state = current->pi_state_cache;
388
389         WARN_ON(!pi_state);
390         current->pi_state_cache = NULL;
391
392         return pi_state;
393 }
394
395 static void free_pi_state(struct futex_pi_state *pi_state)
396 {
397         if (!atomic_dec_and_test(&pi_state->refcount))
398                 return;
399
400         /*
401          * If pi_state->owner is NULL, the owner is most probably dying
402          * and has cleaned up the pi_state already
403          */
404         if (pi_state->owner) {
405                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
406                 list_del_init(&pi_state->list);
407                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
408
409                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
410         }
411
412         if (current->pi_state_cache)
413                 kfree(pi_state);
414         else {
415                 /*
416                  * pi_state->list is already empty.
417                  * clear pi_state->owner.
418                  * refcount is at 0 - put it back to 1.
419                  */
420                 pi_state->owner = NULL;
421                 atomic_set(&pi_state->refcount, 1);
422                 current->pi_state_cache = pi_state;
423         }
424 }
425
426 /*
427  * Look up the task based on what TID userspace gave us.
428  * We dont trust it.
429  */
430 static struct task_struct * futex_find_get_task(pid_t pid)
431 {
432         struct task_struct *p;
433
434         rcu_read_lock();
435         p = find_task_by_vpid(pid);
436         if (p)
437                 get_task_struct(p);
438
439         rcu_read_unlock();
440
441         return p;
442 }
443
444 /*
445  * This task is holding PI mutexes at exit time => bad.
446  * Kernel cleans up PI-state, but userspace is likely hosed.
447  * (Robust-futex cleanup is separate and might save the day for userspace.)
448  */
449 void exit_pi_state_list(struct task_struct *curr)
450 {
451         struct list_head *next, *head = &curr->pi_state_list;
452         struct futex_pi_state *pi_state;
453         struct futex_hash_bucket *hb;
454         union futex_key key = FUTEX_KEY_INIT;
455
456         if (!futex_cmpxchg_enabled)
457                 return;
458         /*
459          * We are a ZOMBIE and nobody can enqueue itself on
460          * pi_state_list anymore, but we have to be careful
461          * versus waiters unqueueing themselves:
462          */
463         raw_spin_lock_irq(&curr->pi_lock);
464         while (!list_empty(head)) {
465
466                 next = head->next;
467                 pi_state = list_entry(next, struct futex_pi_state, list);
468                 key = pi_state->key;
469                 hb = hash_futex(&key);
470                 raw_spin_unlock_irq(&curr->pi_lock);
471
472                 spin_lock(&hb->lock);
473
474                 raw_spin_lock_irq(&curr->pi_lock);
475                 /*
476                  * We dropped the pi-lock, so re-check whether this
477                  * task still owns the PI-state:
478                  */
479                 if (head->next != next) {
480                         spin_unlock(&hb->lock);
481                         continue;
482                 }
483
484                 WARN_ON(pi_state->owner != curr);
485                 WARN_ON(list_empty(&pi_state->list));
486                 list_del_init(&pi_state->list);
487                 pi_state->owner = NULL;
488                 raw_spin_unlock_irq(&curr->pi_lock);
489
490                 rt_mutex_unlock(&pi_state->pi_mutex);
491
492                 spin_unlock(&hb->lock);
493
494                 raw_spin_lock_irq(&curr->pi_lock);
495         }
496         raw_spin_unlock_irq(&curr->pi_lock);
497 }
498
499 static int
500 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
501                 union futex_key *key, struct futex_pi_state **ps)
502 {
503         struct futex_pi_state *pi_state = NULL;
504         struct futex_q *this, *next;
505         struct plist_head *head;
506         struct task_struct *p;
507         pid_t pid = uval & FUTEX_TID_MASK;
508
509         head = &hb->chain;
510
511         plist_for_each_entry_safe(this, next, head, list) {
512                 if (match_futex(&this->key, key)) {
513                         /*
514                          * Another waiter already exists - bump up
515                          * the refcount and return its pi_state:
516                          */
517                         pi_state = this->pi_state;
518                         /*
519                          * Userspace might have messed up non-PI and PI futexes
520                          */
521                         if (unlikely(!pi_state))
522                                 return -EINVAL;
523
524                         WARN_ON(!atomic_read(&pi_state->refcount));
525
526                         /*
527                          * When pi_state->owner is NULL then the owner died
528                          * and another waiter is on the fly. pi_state->owner
529                          * is fixed up by the task which acquires
530                          * pi_state->rt_mutex.
531                          *
532                          * We do not check for pid == 0 which can happen when
533                          * the owner died and robust_list_exit() cleared the
534                          * TID.
535                          */
536                         if (pid && pi_state->owner) {
537                                 /*
538                                  * Bail out if user space manipulated the
539                                  * futex value.
540                                  */
541                                 if (pid != task_pid_vnr(pi_state->owner))
542                                         return -EINVAL;
543                         }
544
545                         atomic_inc(&pi_state->refcount);
546                         *ps = pi_state;
547
548                         return 0;
549                 }
550         }
551
552         /*
553          * We are the first waiter - try to look up the real owner and attach
554          * the new pi_state to it, but bail out when TID = 0
555          */
556         if (!pid)
557                 return -ESRCH;
558         p = futex_find_get_task(pid);
559         if (!p)
560                 return -ESRCH;
561
562         /*
563          * We need to look at the task state flags to figure out,
564          * whether the task is exiting. To protect against the do_exit
565          * change of the task flags, we do this protected by
566          * p->pi_lock:
567          */
568         raw_spin_lock_irq(&p->pi_lock);
569         if (unlikely(p->flags & PF_EXITING)) {
570                 /*
571                  * The task is on the way out. When PF_EXITPIDONE is
572                  * set, we know that the task has finished the
573                  * cleanup:
574                  */
575                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
576
577                 raw_spin_unlock_irq(&p->pi_lock);
578                 put_task_struct(p);
579                 return ret;
580         }
581
582         pi_state = alloc_pi_state();
583
584         /*
585          * Initialize the pi_mutex in locked state and make 'p'
586          * the owner of it:
587          */
588         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
589
590         /* Store the key for possible exit cleanups: */
591         pi_state->key = *key;
592
593         WARN_ON(!list_empty(&pi_state->list));
594         list_add(&pi_state->list, &p->pi_state_list);
595         pi_state->owner = p;
596         raw_spin_unlock_irq(&p->pi_lock);
597
598         put_task_struct(p);
599
600         *ps = pi_state;
601
602         return 0;
603 }
604
605 /**
606  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
607  * @uaddr:              the pi futex user address
608  * @hb:                 the pi futex hash bucket
609  * @key:                the futex key associated with uaddr and hb
610  * @ps:                 the pi_state pointer where we store the result of the
611  *                      lookup
612  * @task:               the task to perform the atomic lock work for.  This will
613  *                      be "current" except in the case of requeue pi.
614  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
615  *
616  * Returns:
617  *  0 - ready to wait
618  *  1 - acquired the lock
619  * <0 - error
620  *
621  * The hb->lock and futex_key refs shall be held by the caller.
622  */
623 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
624                                 union futex_key *key,
625                                 struct futex_pi_state **ps,
626                                 struct task_struct *task, int set_waiters)
627 {
628         int lock_taken, ret, ownerdied = 0;
629         u32 uval, newval, curval;
630
631 retry:
632         ret = lock_taken = 0;
633
634         /*
635          * To avoid races, we attempt to take the lock here again
636          * (by doing a 0 -> TID atomic cmpxchg), while holding all
637          * the locks. It will most likely not succeed.
638          */
639         newval = task_pid_vnr(task);
640         if (set_waiters)
641                 newval |= FUTEX_WAITERS;
642
643         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
644
645         if (unlikely(curval == -EFAULT))
646                 return -EFAULT;
647
648         /*
649          * Detect deadlocks.
650          */
651         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
652                 return -EDEADLK;
653
654         /*
655          * Surprise - we got the lock. Just return to userspace:
656          */
657         if (unlikely(!curval))
658                 return 1;
659
660         uval = curval;
661
662         /*
663          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
664          * to wake at the next unlock.
665          */
666         newval = curval | FUTEX_WAITERS;
667
668         /*
669          * There are two cases, where a futex might have no owner (the
670          * owner TID is 0): OWNER_DIED. We take over the futex in this
671          * case. We also do an unconditional take over, when the owner
672          * of the futex died.
673          *
674          * This is safe as we are protected by the hash bucket lock !
675          */
676         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
677                 /* Keep the OWNER_DIED bit */
678                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
679                 ownerdied = 0;
680                 lock_taken = 1;
681         }
682
683         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
684
685         if (unlikely(curval == -EFAULT))
686                 return -EFAULT;
687         if (unlikely(curval != uval))
688                 goto retry;
689
690         /*
691          * We took the lock due to owner died take over.
692          */
693         if (unlikely(lock_taken))
694                 return 1;
695
696         /*
697          * We dont have the lock. Look up the PI state (or create it if
698          * we are the first waiter):
699          */
700         ret = lookup_pi_state(uval, hb, key, ps);
701
702         if (unlikely(ret)) {
703                 switch (ret) {
704                 case -ESRCH:
705                         /*
706                          * No owner found for this futex. Check if the
707                          * OWNER_DIED bit is set to figure out whether
708                          * this is a robust futex or not.
709                          */
710                         if (get_futex_value_locked(&curval, uaddr))
711                                 return -EFAULT;
712
713                         /*
714                          * We simply start over in case of a robust
715                          * futex. The code above will take the futex
716                          * and return happy.
717                          */
718                         if (curval & FUTEX_OWNER_DIED) {
719                                 ownerdied = 1;
720                                 goto retry;
721                         }
722                 default:
723                         break;
724                 }
725         }
726
727         return ret;
728 }
729
730 /*
731  * The hash bucket lock must be held when this is called.
732  * Afterwards, the futex_q must not be accessed.
733  */
734 static void wake_futex(struct futex_q *q)
735 {
736         struct task_struct *p = q->task;
737
738         /*
739          * We set q->lock_ptr = NULL _before_ we wake up the task. If
740          * a non-futex wake up happens on another CPU then the task
741          * might exit and p would dereference a non-existing task
742          * struct. Prevent this by holding a reference on p across the
743          * wake up.
744          */
745         get_task_struct(p);
746
747         plist_del(&q->list, &q->list.plist);
748         /*
749          * The waiting task can free the futex_q as soon as
750          * q->lock_ptr = NULL is written, without taking any locks. A
751          * memory barrier is required here to prevent the following
752          * store to lock_ptr from getting ahead of the plist_del.
753          */
754         smp_wmb();
755         q->lock_ptr = NULL;
756
757         wake_up_state(p, TASK_NORMAL);
758         put_task_struct(p);
759 }
760
761 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
762 {
763         struct task_struct *new_owner;
764         struct futex_pi_state *pi_state = this->pi_state;
765         u32 curval, newval;
766
767         if (!pi_state)
768                 return -EINVAL;
769
770         /*
771          * If current does not own the pi_state then the futex is
772          * inconsistent and user space fiddled with the futex value.
773          */
774         if (pi_state->owner != current)
775                 return -EINVAL;
776
777         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
778         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
779
780         /*
781          * This happens when we have stolen the lock and the original
782          * pending owner did not enqueue itself back on the rt_mutex.
783          * Thats not a tragedy. We know that way, that a lock waiter
784          * is on the fly. We make the futex_q waiter the pending owner.
785          */
786         if (!new_owner)
787                 new_owner = this->task;
788
789         /*
790          * We pass it to the next owner. (The WAITERS bit is always
791          * kept enabled while there is PI state around. We must also
792          * preserve the owner died bit.)
793          */
794         if (!(uval & FUTEX_OWNER_DIED)) {
795                 int ret = 0;
796
797                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
798
799                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
800
801                 if (curval == -EFAULT)
802                         ret = -EFAULT;
803                 else if (curval != uval)
804                         ret = -EINVAL;
805                 if (ret) {
806                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
807                         return ret;
808                 }
809         }
810
811         raw_spin_lock_irq(&pi_state->owner->pi_lock);
812         WARN_ON(list_empty(&pi_state->list));
813         list_del_init(&pi_state->list);
814         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
815
816         raw_spin_lock_irq(&new_owner->pi_lock);
817         WARN_ON(!list_empty(&pi_state->list));
818         list_add(&pi_state->list, &new_owner->pi_state_list);
819         pi_state->owner = new_owner;
820         raw_spin_unlock_irq(&new_owner->pi_lock);
821
822         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
823         rt_mutex_unlock(&pi_state->pi_mutex);
824
825         return 0;
826 }
827
828 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
829 {
830         u32 oldval;
831
832         /*
833          * There is no waiter, so we unlock the futex. The owner died
834          * bit has not to be preserved here. We are the owner:
835          */
836         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
837
838         if (oldval == -EFAULT)
839                 return oldval;
840         if (oldval != uval)
841                 return -EAGAIN;
842
843         return 0;
844 }
845
846 /*
847  * Express the locking dependencies for lockdep:
848  */
849 static inline void
850 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
851 {
852         if (hb1 <= hb2) {
853                 spin_lock(&hb1->lock);
854                 if (hb1 < hb2)
855                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
856         } else { /* hb1 > hb2 */
857                 spin_lock(&hb2->lock);
858                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
859         }
860 }
861
862 static inline void
863 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
864 {
865         spin_unlock(&hb1->lock);
866         if (hb1 != hb2)
867                 spin_unlock(&hb2->lock);
868 }
869
870 /*
871  * Wake up waiters matching bitset queued on this futex (uaddr).
872  */
873 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
874 {
875         struct futex_hash_bucket *hb;
876         struct futex_q *this, *next;
877         struct plist_head *head;
878         union futex_key key = FUTEX_KEY_INIT;
879         int ret;
880
881         if (!bitset)
882                 return -EINVAL;
883
884         ret = get_futex_key(uaddr, fshared, &key);
885         if (unlikely(ret != 0))
886                 goto out;
887
888         hb = hash_futex(&key);
889         spin_lock(&hb->lock);
890         head = &hb->chain;
891
892         plist_for_each_entry_safe(this, next, head, list) {
893                 if (match_futex (&this->key, &key)) {
894                         if (this->pi_state || this->rt_waiter) {
895                                 ret = -EINVAL;
896                                 break;
897                         }
898
899                         /* Check if one of the bits is set in both bitsets */
900                         if (!(this->bitset & bitset))
901                                 continue;
902
903                         wake_futex(this);
904                         if (++ret >= nr_wake)
905                                 break;
906                 }
907         }
908
909         spin_unlock(&hb->lock);
910         put_futex_key(fshared, &key);
911 out:
912         return ret;
913 }
914
915 /*
916  * Wake up all waiters hashed on the physical page that is mapped
917  * to this virtual address:
918  */
919 static int
920 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
921               int nr_wake, int nr_wake2, int op)
922 {
923         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
924         struct futex_hash_bucket *hb1, *hb2;
925         struct plist_head *head;
926         struct futex_q *this, *next;
927         int ret, op_ret;
928
929 retry:
930         ret = get_futex_key(uaddr1, fshared, &key1);
931         if (unlikely(ret != 0))
932                 goto out;
933         ret = get_futex_key(uaddr2, fshared, &key2);
934         if (unlikely(ret != 0))
935                 goto out_put_key1;
936
937         hb1 = hash_futex(&key1);
938         hb2 = hash_futex(&key2);
939
940 retry_private:
941         double_lock_hb(hb1, hb2);
942         op_ret = futex_atomic_op_inuser(op, uaddr2);
943         if (unlikely(op_ret < 0)) {
944
945                 double_unlock_hb(hb1, hb2);
946
947 #ifndef CONFIG_MMU
948                 /*
949                  * we don't get EFAULT from MMU faults if we don't have an MMU,
950                  * but we might get them from range checking
951                  */
952                 ret = op_ret;
953                 goto out_put_keys;
954 #endif
955
956                 if (unlikely(op_ret != -EFAULT)) {
957                         ret = op_ret;
958                         goto out_put_keys;
959                 }
960
961                 ret = fault_in_user_writeable(uaddr2);
962                 if (ret)
963                         goto out_put_keys;
964
965                 if (!fshared)
966                         goto retry_private;
967
968                 put_futex_key(fshared, &key2);
969                 put_futex_key(fshared, &key1);
970                 goto retry;
971         }
972
973         head = &hb1->chain;
974
975         plist_for_each_entry_safe(this, next, head, list) {
976                 if (match_futex (&this->key, &key1)) {
977                         wake_futex(this);
978                         if (++ret >= nr_wake)
979                                 break;
980                 }
981         }
982
983         if (op_ret > 0) {
984                 head = &hb2->chain;
985
986                 op_ret = 0;
987                 plist_for_each_entry_safe(this, next, head, list) {
988                         if (match_futex (&this->key, &key2)) {
989                                 wake_futex(this);
990                                 if (++op_ret >= nr_wake2)
991                                         break;
992                         }
993                 }
994                 ret += op_ret;
995         }
996
997         double_unlock_hb(hb1, hb2);
998 out_put_keys:
999         put_futex_key(fshared, &key2);
1000 out_put_key1:
1001         put_futex_key(fshared, &key1);
1002 out:
1003         return ret;
1004 }
1005
1006 /**
1007  * requeue_futex() - Requeue a futex_q from one hb to another
1008  * @q:          the futex_q to requeue
1009  * @hb1:        the source hash_bucket
1010  * @hb2:        the target hash_bucket
1011  * @key2:       the new key for the requeued futex_q
1012  */
1013 static inline
1014 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1015                    struct futex_hash_bucket *hb2, union futex_key *key2)
1016 {
1017
1018         /*
1019          * If key1 and key2 hash to the same bucket, no need to
1020          * requeue.
1021          */
1022         if (likely(&hb1->chain != &hb2->chain)) {
1023                 plist_del(&q->list, &hb1->chain);
1024                 plist_add(&q->list, &hb2->chain);
1025                 q->lock_ptr = &hb2->lock;
1026 #ifdef CONFIG_DEBUG_PI_LIST
1027                 q->list.plist.spinlock = &hb2->lock;
1028 #endif
1029         }
1030         get_futex_key_refs(key2);
1031         q->key = *key2;
1032 }
1033
1034 /**
1035  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1036  * @q:          the futex_q
1037  * @key:        the key of the requeue target futex
1038  * @hb:         the hash_bucket of the requeue target futex
1039  *
1040  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1041  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1042  * to the requeue target futex so the waiter can detect the wakeup on the right
1043  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1044  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1045  * to protect access to the pi_state to fixup the owner later.  Must be called
1046  * with both q->lock_ptr and hb->lock held.
1047  */
1048 static inline
1049 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1050                            struct futex_hash_bucket *hb)
1051 {
1052         get_futex_key_refs(key);
1053         q->key = *key;
1054
1055         WARN_ON(plist_node_empty(&q->list));
1056         plist_del(&q->list, &q->list.plist);
1057
1058         WARN_ON(!q->rt_waiter);
1059         q->rt_waiter = NULL;
1060
1061         q->lock_ptr = &hb->lock;
1062 #ifdef CONFIG_DEBUG_PI_LIST
1063         q->list.plist.spinlock = &hb->lock;
1064 #endif
1065
1066         wake_up_state(q->task, TASK_NORMAL);
1067 }
1068
1069 /**
1070  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1071  * @pifutex:            the user address of the to futex
1072  * @hb1:                the from futex hash bucket, must be locked by the caller
1073  * @hb2:                the to futex hash bucket, must be locked by the caller
1074  * @key1:               the from futex key
1075  * @key2:               the to futex key
1076  * @ps:                 address to store the pi_state pointer
1077  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1078  *
1079  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1080  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1081  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1082  * hb1 and hb2 must be held by the caller.
1083  *
1084  * Returns:
1085  *  0 - failed to acquire the lock atomicly
1086  *  1 - acquired the lock
1087  * <0 - error
1088  */
1089 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1090                                  struct futex_hash_bucket *hb1,
1091                                  struct futex_hash_bucket *hb2,
1092                                  union futex_key *key1, union futex_key *key2,
1093                                  struct futex_pi_state **ps, int set_waiters)
1094 {
1095         struct futex_q *top_waiter = NULL;
1096         u32 curval;
1097         int ret;
1098
1099         if (get_futex_value_locked(&curval, pifutex))
1100                 return -EFAULT;
1101
1102         /*
1103          * Find the top_waiter and determine if there are additional waiters.
1104          * If the caller intends to requeue more than 1 waiter to pifutex,
1105          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1106          * as we have means to handle the possible fault.  If not, don't set
1107          * the bit unecessarily as it will force the subsequent unlock to enter
1108          * the kernel.
1109          */
1110         top_waiter = futex_top_waiter(hb1, key1);
1111
1112         /* There are no waiters, nothing for us to do. */
1113         if (!top_waiter)
1114                 return 0;
1115
1116         /* Ensure we requeue to the expected futex. */
1117         if (!match_futex(top_waiter->requeue_pi_key, key2))
1118                 return -EINVAL;
1119
1120         /*
1121          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1122          * the contended case or if set_waiters is 1.  The pi_state is returned
1123          * in ps in contended cases.
1124          */
1125         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1126                                    set_waiters);
1127         if (ret == 1)
1128                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1129
1130         return ret;
1131 }
1132
1133 /**
1134  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1135  * @uaddr1:     source futex user address
1136  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
1137  * @uaddr2:     target futex user address
1138  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1139  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1140  * @cmpval:     @uaddr1 expected value (or %NULL)
1141  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1142  *              pi futex (pi to pi requeue is not supported)
1143  *
1144  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1145  * uaddr2 atomically on behalf of the top waiter.
1146  *
1147  * Returns:
1148  * >=0 - on success, the number of tasks requeued or woken
1149  *  <0 - on error
1150  */
1151 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1152                          int nr_wake, int nr_requeue, u32 *cmpval,
1153                          int requeue_pi)
1154 {
1155         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1156         int drop_count = 0, task_count = 0, ret;
1157         struct futex_pi_state *pi_state = NULL;
1158         struct futex_hash_bucket *hb1, *hb2;
1159         struct plist_head *head1;
1160         struct futex_q *this, *next;
1161         u32 curval2;
1162
1163         if (requeue_pi) {
1164                 /*
1165                  * requeue_pi requires a pi_state, try to allocate it now
1166                  * without any locks in case it fails.
1167                  */
1168                 if (refill_pi_state_cache())
1169                         return -ENOMEM;
1170                 /*
1171                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1172                  * + nr_requeue, since it acquires the rt_mutex prior to
1173                  * returning to userspace, so as to not leave the rt_mutex with
1174                  * waiters and no owner.  However, second and third wake-ups
1175                  * cannot be predicted as they involve race conditions with the
1176                  * first wake and a fault while looking up the pi_state.  Both
1177                  * pthread_cond_signal() and pthread_cond_broadcast() should
1178                  * use nr_wake=1.
1179                  */
1180                 if (nr_wake != 1)
1181                         return -EINVAL;
1182         }
1183
1184 retry:
1185         if (pi_state != NULL) {
1186                 /*
1187                  * We will have to lookup the pi_state again, so free this one
1188                  * to keep the accounting correct.
1189                  */
1190                 free_pi_state(pi_state);
1191                 pi_state = NULL;
1192         }
1193
1194         ret = get_futex_key(uaddr1, fshared, &key1);
1195         if (unlikely(ret != 0))
1196                 goto out;
1197         ret = get_futex_key(uaddr2, fshared, &key2);
1198         if (unlikely(ret != 0))
1199                 goto out_put_key1;
1200
1201         hb1 = hash_futex(&key1);
1202         hb2 = hash_futex(&key2);
1203
1204 retry_private:
1205         double_lock_hb(hb1, hb2);
1206
1207         if (likely(cmpval != NULL)) {
1208                 u32 curval;
1209
1210                 ret = get_futex_value_locked(&curval, uaddr1);
1211
1212                 if (unlikely(ret)) {
1213                         double_unlock_hb(hb1, hb2);
1214
1215                         ret = get_user(curval, uaddr1);
1216                         if (ret)
1217                                 goto out_put_keys;
1218
1219                         if (!fshared)
1220                                 goto retry_private;
1221
1222                         put_futex_key(fshared, &key2);
1223                         put_futex_key(fshared, &key1);
1224                         goto retry;
1225                 }
1226                 if (curval != *cmpval) {
1227                         ret = -EAGAIN;
1228                         goto out_unlock;
1229                 }
1230         }
1231
1232         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1233                 /*
1234                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1235                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1236                  * bit.  We force this here where we are able to easily handle
1237                  * faults rather in the requeue loop below.
1238                  */
1239                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1240                                                  &key2, &pi_state, nr_requeue);
1241
1242                 /*
1243                  * At this point the top_waiter has either taken uaddr2 or is
1244                  * waiting on it.  If the former, then the pi_state will not
1245                  * exist yet, look it up one more time to ensure we have a
1246                  * reference to it.
1247                  */
1248                 if (ret == 1) {
1249                         WARN_ON(pi_state);
1250                         drop_count++;
1251                         task_count++;
1252                         ret = get_futex_value_locked(&curval2, uaddr2);
1253                         if (!ret)
1254                                 ret = lookup_pi_state(curval2, hb2, &key2,
1255                                                       &pi_state);
1256                 }
1257
1258                 switch (ret) {
1259                 case 0:
1260                         break;
1261                 case -EFAULT:
1262                         double_unlock_hb(hb1, hb2);
1263                         put_futex_key(fshared, &key2);
1264                         put_futex_key(fshared, &key1);
1265                         ret = fault_in_user_writeable(uaddr2);
1266                         if (!ret)
1267                                 goto retry;
1268                         goto out;
1269                 case -EAGAIN:
1270                         /* The owner was exiting, try again. */
1271                         double_unlock_hb(hb1, hb2);
1272                         put_futex_key(fshared, &key2);
1273                         put_futex_key(fshared, &key1);
1274                         cond_resched();
1275                         goto retry;
1276                 default:
1277                         goto out_unlock;
1278                 }
1279         }
1280
1281         head1 = &hb1->chain;
1282         plist_for_each_entry_safe(this, next, head1, list) {
1283                 if (task_count - nr_wake >= nr_requeue)
1284                         break;
1285
1286                 if (!match_futex(&this->key, &key1))
1287                         continue;
1288
1289                 /*
1290                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1291                  * be paired with each other and no other futex ops.
1292                  */
1293                 if ((requeue_pi && !this->rt_waiter) ||
1294                     (!requeue_pi && this->rt_waiter)) {
1295                         ret = -EINVAL;
1296                         break;
1297                 }
1298
1299                 /*
1300                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1301                  * lock, we already woke the top_waiter.  If not, it will be
1302                  * woken by futex_unlock_pi().
1303                  */
1304                 if (++task_count <= nr_wake && !requeue_pi) {
1305                         wake_futex(this);
1306                         continue;
1307                 }
1308
1309                 /* Ensure we requeue to the expected futex for requeue_pi. */
1310                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1311                         ret = -EINVAL;
1312                         break;
1313                 }
1314
1315                 /*
1316                  * Requeue nr_requeue waiters and possibly one more in the case
1317                  * of requeue_pi if we couldn't acquire the lock atomically.
1318                  */
1319                 if (requeue_pi) {
1320                         /* Prepare the waiter to take the rt_mutex. */
1321                         atomic_inc(&pi_state->refcount);
1322                         this->pi_state = pi_state;
1323                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1324                                                         this->rt_waiter,
1325                                                         this->task, 1);
1326                         if (ret == 1) {
1327                                 /* We got the lock. */
1328                                 requeue_pi_wake_futex(this, &key2, hb2);
1329                                 drop_count++;
1330                                 continue;
1331                         } else if (ret) {
1332                                 /* -EDEADLK */
1333                                 this->pi_state = NULL;
1334                                 free_pi_state(pi_state);
1335                                 goto out_unlock;
1336                         }
1337                 }
1338                 requeue_futex(this, hb1, hb2, &key2);
1339                 drop_count++;
1340         }
1341
1342 out_unlock:
1343         double_unlock_hb(hb1, hb2);
1344
1345         /*
1346          * drop_futex_key_refs() must be called outside the spinlocks. During
1347          * the requeue we moved futex_q's from the hash bucket at key1 to the
1348          * one at key2 and updated their key pointer.  We no longer need to
1349          * hold the references to key1.
1350          */
1351         while (--drop_count >= 0)
1352                 drop_futex_key_refs(&key1);
1353
1354 out_put_keys:
1355         put_futex_key(fshared, &key2);
1356 out_put_key1:
1357         put_futex_key(fshared, &key1);
1358 out:
1359         if (pi_state != NULL)
1360                 free_pi_state(pi_state);
1361         return ret ? ret : task_count;
1362 }
1363
1364 /* The key must be already stored in q->key. */
1365 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1366         __acquires(&hb->lock)
1367 {
1368         struct futex_hash_bucket *hb;
1369
1370         hb = hash_futex(&q->key);
1371         q->lock_ptr = &hb->lock;
1372
1373         spin_lock(&hb->lock);
1374         return hb;
1375 }
1376
1377 static inline void
1378 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1379         __releases(&hb->lock)
1380 {
1381         spin_unlock(&hb->lock);
1382 }
1383
1384 /**
1385  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1386  * @q:  The futex_q to enqueue
1387  * @hb: The destination hash bucket
1388  *
1389  * The hb->lock must be held by the caller, and is released here. A call to
1390  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1391  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1392  * or nothing if the unqueue is done as part of the wake process and the unqueue
1393  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1394  * an example).
1395  */
1396 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1397         __releases(&hb->lock)
1398 {
1399         int prio;
1400
1401         /*
1402          * The priority used to register this element is
1403          * - either the real thread-priority for the real-time threads
1404          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1405          * - or MAX_RT_PRIO for non-RT threads.
1406          * Thus, all RT-threads are woken first in priority order, and
1407          * the others are woken last, in FIFO order.
1408          */
1409         prio = min(current->normal_prio, MAX_RT_PRIO);
1410
1411         plist_node_init(&q->list, prio);
1412 #ifdef CONFIG_DEBUG_PI_LIST
1413         q->list.plist.spinlock = &hb->lock;
1414 #endif
1415         plist_add(&q->list, &hb->chain);
1416         q->task = current;
1417         spin_unlock(&hb->lock);
1418 }
1419
1420 /**
1421  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1422  * @q:  The futex_q to unqueue
1423  *
1424  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1425  * be paired with exactly one earlier call to queue_me().
1426  *
1427  * Returns:
1428  *   1 - if the futex_q was still queued (and we removed unqueued it)
1429  *   0 - if the futex_q was already removed by the waking thread
1430  */
1431 static int unqueue_me(struct futex_q *q)
1432 {
1433         spinlock_t *lock_ptr;
1434         int ret = 0;
1435
1436         /* In the common case we don't take the spinlock, which is nice. */
1437 retry:
1438         lock_ptr = q->lock_ptr;
1439         barrier();
1440         if (lock_ptr != NULL) {
1441                 spin_lock(lock_ptr);
1442                 /*
1443                  * q->lock_ptr can change between reading it and
1444                  * spin_lock(), causing us to take the wrong lock.  This
1445                  * corrects the race condition.
1446                  *
1447                  * Reasoning goes like this: if we have the wrong lock,
1448                  * q->lock_ptr must have changed (maybe several times)
1449                  * between reading it and the spin_lock().  It can
1450                  * change again after the spin_lock() but only if it was
1451                  * already changed before the spin_lock().  It cannot,
1452                  * however, change back to the original value.  Therefore
1453                  * we can detect whether we acquired the correct lock.
1454                  */
1455                 if (unlikely(lock_ptr != q->lock_ptr)) {
1456                         spin_unlock(lock_ptr);
1457                         goto retry;
1458                 }
1459                 WARN_ON(plist_node_empty(&q->list));
1460                 plist_del(&q->list, &q->list.plist);
1461
1462                 BUG_ON(q->pi_state);
1463
1464                 spin_unlock(lock_ptr);
1465                 ret = 1;
1466         }
1467
1468         drop_futex_key_refs(&q->key);
1469         return ret;
1470 }
1471
1472 /*
1473  * PI futexes can not be requeued and must remove themself from the
1474  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1475  * and dropped here.
1476  */
1477 static void unqueue_me_pi(struct futex_q *q)
1478         __releases(q->lock_ptr)
1479 {
1480         WARN_ON(plist_node_empty(&q->list));
1481         plist_del(&q->list, &q->list.plist);
1482
1483         BUG_ON(!q->pi_state);
1484         free_pi_state(q->pi_state);
1485         q->pi_state = NULL;
1486
1487         spin_unlock(q->lock_ptr);
1488 }
1489
1490 /*
1491  * Fixup the pi_state owner with the new owner.
1492  *
1493  * Must be called with hash bucket lock held and mm->sem held for non
1494  * private futexes.
1495  */
1496 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1497                                 struct task_struct *newowner, int fshared)
1498 {
1499         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1500         struct futex_pi_state *pi_state = q->pi_state;
1501         struct task_struct *oldowner = pi_state->owner;
1502         u32 uval, curval, newval;
1503         int ret;
1504
1505         /* Owner died? */
1506         if (!pi_state->owner)
1507                 newtid |= FUTEX_OWNER_DIED;
1508
1509         /*
1510          * We are here either because we stole the rtmutex from the
1511          * pending owner or we are the pending owner which failed to
1512          * get the rtmutex. We have to replace the pending owner TID
1513          * in the user space variable. This must be atomic as we have
1514          * to preserve the owner died bit here.
1515          *
1516          * Note: We write the user space value _before_ changing the pi_state
1517          * because we can fault here. Imagine swapped out pages or a fork
1518          * that marked all the anonymous memory readonly for cow.
1519          *
1520          * Modifying pi_state _before_ the user space value would
1521          * leave the pi_state in an inconsistent state when we fault
1522          * here, because we need to drop the hash bucket lock to
1523          * handle the fault. This might be observed in the PID check
1524          * in lookup_pi_state.
1525          */
1526 retry:
1527         if (get_futex_value_locked(&uval, uaddr))
1528                 goto handle_fault;
1529
1530         while (1) {
1531                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1532
1533                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1534
1535                 if (curval == -EFAULT)
1536                         goto handle_fault;
1537                 if (curval == uval)
1538                         break;
1539                 uval = curval;
1540         }
1541
1542         /*
1543          * We fixed up user space. Now we need to fix the pi_state
1544          * itself.
1545          */
1546         if (pi_state->owner != NULL) {
1547                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1548                 WARN_ON(list_empty(&pi_state->list));
1549                 list_del_init(&pi_state->list);
1550                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1551         }
1552
1553         pi_state->owner = newowner;
1554
1555         raw_spin_lock_irq(&newowner->pi_lock);
1556         WARN_ON(!list_empty(&pi_state->list));
1557         list_add(&pi_state->list, &newowner->pi_state_list);
1558         raw_spin_unlock_irq(&newowner->pi_lock);
1559         return 0;
1560
1561         /*
1562          * To handle the page fault we need to drop the hash bucket
1563          * lock here. That gives the other task (either the pending
1564          * owner itself or the task which stole the rtmutex) the
1565          * chance to try the fixup of the pi_state. So once we are
1566          * back from handling the fault we need to check the pi_state
1567          * after reacquiring the hash bucket lock and before trying to
1568          * do another fixup. When the fixup has been done already we
1569          * simply return.
1570          */
1571 handle_fault:
1572         spin_unlock(q->lock_ptr);
1573
1574         ret = fault_in_user_writeable(uaddr);
1575
1576         spin_lock(q->lock_ptr);
1577
1578         /*
1579          * Check if someone else fixed it for us:
1580          */
1581         if (pi_state->owner != oldowner)
1582                 return 0;
1583
1584         if (ret)
1585                 return ret;
1586
1587         goto retry;
1588 }
1589
1590 /*
1591  * In case we must use restart_block to restart a futex_wait,
1592  * we encode in the 'flags' shared capability
1593  */
1594 #define FLAGS_SHARED            0x01
1595 #define FLAGS_CLOCKRT           0x02
1596 #define FLAGS_HAS_TIMEOUT       0x04
1597
1598 static long futex_wait_restart(struct restart_block *restart);
1599
1600 /**
1601  * fixup_owner() - Post lock pi_state and corner case management
1602  * @uaddr:      user address of the futex
1603  * @fshared:    whether the futex is shared (1) or not (0)
1604  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1605  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1606  *
1607  * After attempting to lock an rt_mutex, this function is called to cleanup
1608  * the pi_state owner as well as handle race conditions that may allow us to
1609  * acquire the lock. Must be called with the hb lock held.
1610  *
1611  * Returns:
1612  *  1 - success, lock taken
1613  *  0 - success, lock not taken
1614  * <0 - on error (-EFAULT)
1615  */
1616 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1617                        int locked)
1618 {
1619         struct task_struct *owner;
1620         int ret = 0;
1621
1622         if (locked) {
1623                 /*
1624                  * Got the lock. We might not be the anticipated owner if we
1625                  * did a lock-steal - fix up the PI-state in that case:
1626                  */
1627                 if (q->pi_state->owner != current)
1628                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1629                 goto out;
1630         }
1631
1632         /*
1633          * Catch the rare case, where the lock was released when we were on the
1634          * way back before we locked the hash bucket.
1635          */
1636         if (q->pi_state->owner == current) {
1637                 /*
1638                  * Try to get the rt_mutex now. This might fail as some other
1639                  * task acquired the rt_mutex after we removed ourself from the
1640                  * rt_mutex waiters list.
1641                  */
1642                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1643                         locked = 1;
1644                         goto out;
1645                 }
1646
1647                 /*
1648                  * pi_state is incorrect, some other task did a lock steal and
1649                  * we returned due to timeout or signal without taking the
1650                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1651                  * locking, as the other task is now blocked on the hash bucket
1652                  * lock. Fix the state up.
1653                  */
1654                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1655                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1656                 goto out;
1657         }
1658
1659         /*
1660          * Paranoia check. If we did not take the lock, then we should not be
1661          * the owner, nor the pending owner, of the rt_mutex.
1662          */
1663         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1664                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1665                                 "pi-state %p\n", ret,
1666                                 q->pi_state->pi_mutex.owner,
1667                                 q->pi_state->owner);
1668
1669 out:
1670         return ret ? ret : locked;
1671 }
1672
1673 /**
1674  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1675  * @hb:         the futex hash bucket, must be locked by the caller
1676  * @q:          the futex_q to queue up on
1677  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1678  */
1679 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1680                                 struct hrtimer_sleeper *timeout)
1681 {
1682         /*
1683          * The task state is guaranteed to be set before another task can
1684          * wake it. set_current_state() is implemented using set_mb() and
1685          * queue_me() calls spin_unlock() upon completion, both serializing
1686          * access to the hash list and forcing another memory barrier.
1687          */
1688         set_current_state(TASK_INTERRUPTIBLE);
1689         queue_me(q, hb);
1690
1691         /* Arm the timer */
1692         if (timeout) {
1693                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1694                 if (!hrtimer_active(&timeout->timer))
1695                         timeout->task = NULL;
1696         }
1697
1698         /*
1699          * If we have been removed from the hash list, then another task
1700          * has tried to wake us, and we can skip the call to schedule().
1701          */
1702         if (likely(!plist_node_empty(&q->list))) {
1703                 /*
1704                  * If the timer has already expired, current will already be
1705                  * flagged for rescheduling. Only call schedule if there
1706                  * is no timeout, or if it has yet to expire.
1707                  */
1708                 if (!timeout || timeout->task)
1709                         schedule();
1710         }
1711         __set_current_state(TASK_RUNNING);
1712 }
1713
1714 /**
1715  * futex_wait_setup() - Prepare to wait on a futex
1716  * @uaddr:      the futex userspace address
1717  * @val:        the expected value
1718  * @fshared:    whether the futex is shared (1) or not (0)
1719  * @q:          the associated futex_q
1720  * @hb:         storage for hash_bucket pointer to be returned to caller
1721  *
1722  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1723  * compare it with the expected value.  Handle atomic faults internally.
1724  * Return with the hb lock held and a q.key reference on success, and unlocked
1725  * with no q.key reference on failure.
1726  *
1727  * Returns:
1728  *  0 - uaddr contains val and hb has been locked
1729  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1730  */
1731 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1732                            struct futex_q *q, struct futex_hash_bucket **hb)
1733 {
1734         u32 uval;
1735         int ret;
1736
1737         /*
1738          * Access the page AFTER the hash-bucket is locked.
1739          * Order is important:
1740          *
1741          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1742          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1743          *
1744          * The basic logical guarantee of a futex is that it blocks ONLY
1745          * if cond(var) is known to be true at the time of blocking, for
1746          * any cond.  If we queued after testing *uaddr, that would open
1747          * a race condition where we could block indefinitely with
1748          * cond(var) false, which would violate the guarantee.
1749          *
1750          * A consequence is that futex_wait() can return zero and absorb
1751          * a wakeup when *uaddr != val on entry to the syscall.  This is
1752          * rare, but normal.
1753          */
1754 retry:
1755         q->key = FUTEX_KEY_INIT;
1756         ret = get_futex_key(uaddr, fshared, &q->key);
1757         if (unlikely(ret != 0))
1758                 return ret;
1759
1760 retry_private:
1761         *hb = queue_lock(q);
1762
1763         ret = get_futex_value_locked(&uval, uaddr);
1764
1765         if (ret) {
1766                 queue_unlock(q, *hb);
1767
1768                 ret = get_user(uval, uaddr);
1769                 if (ret)
1770                         goto out;
1771
1772                 if (!fshared)
1773                         goto retry_private;
1774
1775                 put_futex_key(fshared, &q->key);
1776                 goto retry;
1777         }
1778
1779         if (uval != val) {
1780                 queue_unlock(q, *hb);
1781                 ret = -EWOULDBLOCK;
1782         }
1783
1784 out:
1785         if (ret)
1786                 put_futex_key(fshared, &q->key);
1787         return ret;
1788 }
1789
1790 static int futex_wait(u32 __user *uaddr, int fshared,
1791                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1792 {
1793         struct hrtimer_sleeper timeout, *to = NULL;
1794         struct restart_block *restart;
1795         struct futex_hash_bucket *hb;
1796         struct futex_q q;
1797         int ret;
1798
1799         if (!bitset)
1800                 return -EINVAL;
1801
1802         q.pi_state = NULL;
1803         q.bitset = bitset;
1804         q.rt_waiter = NULL;
1805         q.requeue_pi_key = NULL;
1806
1807         if (abs_time) {
1808                 to = &timeout;
1809
1810                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1811                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1812                 hrtimer_init_sleeper(to, current);
1813                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1814                                              current->timer_slack_ns);
1815         }
1816
1817 retry:
1818         /*
1819          * Prepare to wait on uaddr. On success, holds hb lock and increments
1820          * q.key refs.
1821          */
1822         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1823         if (ret)
1824                 goto out;
1825
1826         /* queue_me and wait for wakeup, timeout, or a signal. */
1827         futex_wait_queue_me(hb, &q, to);
1828
1829         /* If we were woken (and unqueued), we succeeded, whatever. */
1830         ret = 0;
1831         /* unqueue_me() drops q.key ref */
1832         if (!unqueue_me(&q))
1833                 goto out;
1834         ret = -ETIMEDOUT;
1835         if (to && !to->task)
1836                 goto out;
1837
1838         /*
1839          * We expect signal_pending(current), but we might be the
1840          * victim of a spurious wakeup as well.
1841          */
1842         if (!signal_pending(current))
1843                 goto retry;
1844
1845         ret = -ERESTARTSYS;
1846         if (!abs_time)
1847                 goto out;
1848
1849         restart = &current_thread_info()->restart_block;
1850         restart->fn = futex_wait_restart;
1851         restart->futex.uaddr = uaddr;
1852         restart->futex.val = val;
1853         restart->futex.time = abs_time->tv64;
1854         restart->futex.bitset = bitset;
1855         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1856
1857         if (fshared)
1858                 restart->futex.flags |= FLAGS_SHARED;
1859         if (clockrt)
1860                 restart->futex.flags |= FLAGS_CLOCKRT;
1861
1862         ret = -ERESTART_RESTARTBLOCK;
1863
1864 out:
1865         if (to) {
1866                 hrtimer_cancel(&to->timer);
1867                 destroy_hrtimer_on_stack(&to->timer);
1868         }
1869         return ret;
1870 }
1871
1872
1873 static long futex_wait_restart(struct restart_block *restart)
1874 {
1875         u32 __user *uaddr = restart->futex.uaddr;
1876         int fshared = 0;
1877         ktime_t t, *tp = NULL;
1878
1879         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1880                 t.tv64 = restart->futex.time;
1881                 tp = &t;
1882         }
1883         restart->fn = do_no_restart_syscall;
1884         if (restart->futex.flags & FLAGS_SHARED)
1885                 fshared = 1;
1886         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1887                                 restart->futex.bitset,
1888                                 restart->futex.flags & FLAGS_CLOCKRT);
1889 }
1890
1891
1892 /*
1893  * Userspace tried a 0 -> TID atomic transition of the futex value
1894  * and failed. The kernel side here does the whole locking operation:
1895  * if there are waiters then it will block, it does PI, etc. (Due to
1896  * races the kernel might see a 0 value of the futex too.)
1897  */
1898 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1899                          int detect, ktime_t *time, int trylock)
1900 {
1901         struct hrtimer_sleeper timeout, *to = NULL;
1902         struct futex_hash_bucket *hb;
1903         struct futex_q q;
1904         int res, ret;
1905
1906         if (refill_pi_state_cache())
1907                 return -ENOMEM;
1908
1909         if (time) {
1910                 to = &timeout;
1911                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1912                                       HRTIMER_MODE_ABS);
1913                 hrtimer_init_sleeper(to, current);
1914                 hrtimer_set_expires(&to->timer, *time);
1915         }
1916
1917         q.pi_state = NULL;
1918         q.rt_waiter = NULL;
1919         q.requeue_pi_key = NULL;
1920 retry:
1921         q.key = FUTEX_KEY_INIT;
1922         ret = get_futex_key(uaddr, fshared, &q.key);
1923         if (unlikely(ret != 0))
1924                 goto out;
1925
1926 retry_private:
1927         hb = queue_lock(&q);
1928
1929         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1930         if (unlikely(ret)) {
1931                 switch (ret) {
1932                 case 1:
1933                         /* We got the lock. */
1934                         ret = 0;
1935                         goto out_unlock_put_key;
1936                 case -EFAULT:
1937                         goto uaddr_faulted;
1938                 case -EAGAIN:
1939                         /*
1940                          * Task is exiting and we just wait for the
1941                          * exit to complete.
1942                          */
1943                         queue_unlock(&q, hb);
1944                         put_futex_key(fshared, &q.key);
1945                         cond_resched();
1946                         goto retry;
1947                 default:
1948                         goto out_unlock_put_key;
1949                 }
1950         }
1951
1952         /*
1953          * Only actually queue now that the atomic ops are done:
1954          */
1955         queue_me(&q, hb);
1956
1957         WARN_ON(!q.pi_state);
1958         /*
1959          * Block on the PI mutex:
1960          */
1961         if (!trylock)
1962                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1963         else {
1964                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1965                 /* Fixup the trylock return value: */
1966                 ret = ret ? 0 : -EWOULDBLOCK;
1967         }
1968
1969         spin_lock(q.lock_ptr);
1970         /*
1971          * Fixup the pi_state owner and possibly acquire the lock if we
1972          * haven't already.
1973          */
1974         res = fixup_owner(uaddr, fshared, &q, !ret);
1975         /*
1976          * If fixup_owner() returned an error, proprogate that.  If it acquired
1977          * the lock, clear our -ETIMEDOUT or -EINTR.
1978          */
1979         if (res)
1980                 ret = (res < 0) ? res : 0;
1981
1982         /*
1983          * If fixup_owner() faulted and was unable to handle the fault, unlock
1984          * it and return the fault to userspace.
1985          */
1986         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1987                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1988
1989         /* Unqueue and drop the lock */
1990         unqueue_me_pi(&q);
1991
1992         goto out_put_key;
1993
1994 out_unlock_put_key:
1995         queue_unlock(&q, hb);
1996
1997 out_put_key:
1998         put_futex_key(fshared, &q.key);
1999 out:
2000         if (to)
2001                 destroy_hrtimer_on_stack(&to->timer);
2002         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2003
2004 uaddr_faulted:
2005         queue_unlock(&q, hb);
2006
2007         ret = fault_in_user_writeable(uaddr);
2008         if (ret)
2009                 goto out_put_key;
2010
2011         if (!fshared)
2012                 goto retry_private;
2013
2014         put_futex_key(fshared, &q.key);
2015         goto retry;
2016 }
2017
2018 /*
2019  * Userspace attempted a TID -> 0 atomic transition, and failed.
2020  * This is the in-kernel slowpath: we look up the PI state (if any),
2021  * and do the rt-mutex unlock.
2022  */
2023 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2024 {
2025         struct futex_hash_bucket *hb;
2026         struct futex_q *this, *next;
2027         u32 uval;
2028         struct plist_head *head;
2029         union futex_key key = FUTEX_KEY_INIT;
2030         int ret;
2031
2032 retry:
2033         if (get_user(uval, uaddr))
2034                 return -EFAULT;
2035         /*
2036          * We release only a lock we actually own:
2037          */
2038         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2039                 return -EPERM;
2040
2041         ret = get_futex_key(uaddr, fshared, &key);
2042         if (unlikely(ret != 0))
2043                 goto out;
2044
2045         hb = hash_futex(&key);
2046         spin_lock(&hb->lock);
2047
2048         /*
2049          * To avoid races, try to do the TID -> 0 atomic transition
2050          * again. If it succeeds then we can return without waking
2051          * anyone else up:
2052          */
2053         if (!(uval & FUTEX_OWNER_DIED))
2054                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2055
2056
2057         if (unlikely(uval == -EFAULT))
2058                 goto pi_faulted;
2059         /*
2060          * Rare case: we managed to release the lock atomically,
2061          * no need to wake anyone else up:
2062          */
2063         if (unlikely(uval == task_pid_vnr(current)))
2064                 goto out_unlock;
2065
2066         /*
2067          * Ok, other tasks may need to be woken up - check waiters
2068          * and do the wakeup if necessary:
2069          */
2070         head = &hb->chain;
2071
2072         plist_for_each_entry_safe(this, next, head, list) {
2073                 if (!match_futex (&this->key, &key))
2074                         continue;
2075                 ret = wake_futex_pi(uaddr, uval, this);
2076                 /*
2077                  * The atomic access to the futex value
2078                  * generated a pagefault, so retry the
2079                  * user-access and the wakeup:
2080                  */
2081                 if (ret == -EFAULT)
2082                         goto pi_faulted;
2083                 goto out_unlock;
2084         }
2085         /*
2086          * No waiters - kernel unlocks the futex:
2087          */
2088         if (!(uval & FUTEX_OWNER_DIED)) {
2089                 ret = unlock_futex_pi(uaddr, uval);
2090                 if (ret == -EFAULT)
2091                         goto pi_faulted;
2092         }
2093
2094 out_unlock:
2095         spin_unlock(&hb->lock);
2096         put_futex_key(fshared, &key);
2097
2098 out:
2099         return ret;
2100
2101 pi_faulted:
2102         spin_unlock(&hb->lock);
2103         put_futex_key(fshared, &key);
2104
2105         ret = fault_in_user_writeable(uaddr);
2106         if (!ret)
2107                 goto retry;
2108
2109         return ret;
2110 }
2111
2112 /**
2113  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2114  * @hb:         the hash_bucket futex_q was original enqueued on
2115  * @q:          the futex_q woken while waiting to be requeued
2116  * @key2:       the futex_key of the requeue target futex
2117  * @timeout:    the timeout associated with the wait (NULL if none)
2118  *
2119  * Detect if the task was woken on the initial futex as opposed to the requeue
2120  * target futex.  If so, determine if it was a timeout or a signal that caused
2121  * the wakeup and return the appropriate error code to the caller.  Must be
2122  * called with the hb lock held.
2123  *
2124  * Returns
2125  *  0 - no early wakeup detected
2126  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2127  */
2128 static inline
2129 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2130                                    struct futex_q *q, union futex_key *key2,
2131                                    struct hrtimer_sleeper *timeout)
2132 {
2133         int ret = 0;
2134
2135         /*
2136          * With the hb lock held, we avoid races while we process the wakeup.
2137          * We only need to hold hb (and not hb2) to ensure atomicity as the
2138          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2139          * It can't be requeued from uaddr2 to something else since we don't
2140          * support a PI aware source futex for requeue.
2141          */
2142         if (!match_futex(&q->key, key2)) {
2143                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2144                 /*
2145                  * We were woken prior to requeue by a timeout or a signal.
2146                  * Unqueue the futex_q and determine which it was.
2147                  */
2148                 plist_del(&q->list, &q->list.plist);
2149
2150                 /* Handle spurious wakeups gracefully */
2151                 ret = -EWOULDBLOCK;
2152                 if (timeout && !timeout->task)
2153                         ret = -ETIMEDOUT;
2154                 else if (signal_pending(current))
2155                         ret = -ERESTARTNOINTR;
2156         }
2157         return ret;
2158 }
2159
2160 /**
2161  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2162  * @uaddr:      the futex we initially wait on (non-pi)
2163  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2164  *              the same type, no requeueing from private to shared, etc.
2165  * @val:        the expected value of uaddr
2166  * @abs_time:   absolute timeout
2167  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2168  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2169  * @uaddr2:     the pi futex we will take prior to returning to user-space
2170  *
2171  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2172  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2173  * complete the acquisition of the rt_mutex prior to returning to userspace.
2174  * This ensures the rt_mutex maintains an owner when it has waiters; without
2175  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2176  * need to.
2177  *
2178  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2179  * via the following:
2180  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2181  * 2) wakeup on uaddr2 after a requeue
2182  * 3) signal
2183  * 4) timeout
2184  *
2185  * If 3, cleanup and return -ERESTARTNOINTR.
2186  *
2187  * If 2, we may then block on trying to take the rt_mutex and return via:
2188  * 5) successful lock
2189  * 6) signal
2190  * 7) timeout
2191  * 8) other lock acquisition failure
2192  *
2193  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2194  *
2195  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2196  *
2197  * Returns:
2198  *  0 - On success
2199  * <0 - On error
2200  */
2201 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2202                                  u32 val, ktime_t *abs_time, u32 bitset,
2203                                  int clockrt, u32 __user *uaddr2)
2204 {
2205         struct hrtimer_sleeper timeout, *to = NULL;
2206         struct rt_mutex_waiter rt_waiter;
2207         struct rt_mutex *pi_mutex = NULL;
2208         struct futex_hash_bucket *hb;
2209         union futex_key key2;
2210         struct futex_q q;
2211         int res, ret;
2212
2213         if (!bitset)
2214                 return -EINVAL;
2215
2216         if (abs_time) {
2217                 to = &timeout;
2218                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2219                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2220                 hrtimer_init_sleeper(to, current);
2221                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2222                                              current->timer_slack_ns);
2223         }
2224
2225         /*
2226          * The waiter is allocated on our stack, manipulated by the requeue
2227          * code while we sleep on uaddr.
2228          */
2229         debug_rt_mutex_init_waiter(&rt_waiter);
2230         rt_waiter.task = NULL;
2231
2232         key2 = FUTEX_KEY_INIT;
2233         ret = get_futex_key(uaddr2, fshared, &key2);
2234         if (unlikely(ret != 0))
2235                 goto out;
2236
2237         q.pi_state = NULL;
2238         q.bitset = bitset;
2239         q.rt_waiter = &rt_waiter;
2240         q.requeue_pi_key = &key2;
2241
2242         /*
2243          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2244          * count.
2245          */
2246         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2247         if (ret)
2248                 goto out_key2;
2249
2250         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2251         futex_wait_queue_me(hb, &q, to);
2252
2253         spin_lock(&hb->lock);
2254         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2255         spin_unlock(&hb->lock);
2256         if (ret)
2257                 goto out_put_keys;
2258
2259         /*
2260          * In order for us to be here, we know our q.key == key2, and since
2261          * we took the hb->lock above, we also know that futex_requeue() has
2262          * completed and we no longer have to concern ourselves with a wakeup
2263          * race with the atomic proxy lock acquisition by the requeue code. The
2264          * futex_requeue dropped our key1 reference and incremented our key2
2265          * reference count.
2266          */
2267
2268         /* Check if the requeue code acquired the second futex for us. */
2269         if (!q.rt_waiter) {
2270                 /*
2271                  * Got the lock. We might not be the anticipated owner if we
2272                  * did a lock-steal - fix up the PI-state in that case.
2273                  */
2274                 if (q.pi_state && (q.pi_state->owner != current)) {
2275                         spin_lock(q.lock_ptr);
2276                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2277                                                    fshared);
2278                         spin_unlock(q.lock_ptr);
2279                 }
2280         } else {
2281                 /*
2282                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2283                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2284                  * the pi_state.
2285                  */
2286                 WARN_ON(!&q.pi_state);
2287                 pi_mutex = &q.pi_state->pi_mutex;
2288                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2289                 debug_rt_mutex_free_waiter(&rt_waiter);
2290
2291                 spin_lock(q.lock_ptr);
2292                 /*
2293                  * Fixup the pi_state owner and possibly acquire the lock if we
2294                  * haven't already.
2295                  */
2296                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2297                 /*
2298                  * If fixup_owner() returned an error, proprogate that.  If it
2299                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2300                  */
2301                 if (res)
2302                         ret = (res < 0) ? res : 0;
2303
2304                 /* Unqueue and drop the lock. */
2305                 unqueue_me_pi(&q);
2306         }
2307
2308         /*
2309          * If fixup_pi_state_owner() faulted and was unable to handle the
2310          * fault, unlock the rt_mutex and return the fault to userspace.
2311          */
2312         if (ret == -EFAULT) {
2313                 if (rt_mutex_owner(pi_mutex) == current)
2314                         rt_mutex_unlock(pi_mutex);
2315         } else if (ret == -EINTR) {
2316                 /*
2317                  * We've already been requeued, but cannot restart by calling
2318                  * futex_lock_pi() directly. We could restart this syscall, but
2319                  * it would detect that the user space "val" changed and return
2320                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2321                  * -EWOULDBLOCK directly.
2322                  */
2323                 ret = -EWOULDBLOCK;
2324         }
2325
2326 out_put_keys:
2327         put_futex_key(fshared, &q.key);
2328 out_key2:
2329         put_futex_key(fshared, &key2);
2330
2331 out:
2332         if (to) {
2333                 hrtimer_cancel(&to->timer);
2334                 destroy_hrtimer_on_stack(&to->timer);
2335         }
2336         return ret;
2337 }
2338
2339 /*
2340  * Support for robust futexes: the kernel cleans up held futexes at
2341  * thread exit time.
2342  *
2343  * Implementation: user-space maintains a per-thread list of locks it
2344  * is holding. Upon do_exit(), the kernel carefully walks this list,
2345  * and marks all locks that are owned by this thread with the
2346  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2347  * always manipulated with the lock held, so the list is private and
2348  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2349  * field, to allow the kernel to clean up if the thread dies after
2350  * acquiring the lock, but just before it could have added itself to
2351  * the list. There can only be one such pending lock.
2352  */
2353
2354 /**
2355  * sys_set_robust_list() - Set the robust-futex list head of a task
2356  * @head:       pointer to the list-head
2357  * @len:        length of the list-head, as userspace expects
2358  */
2359 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2360                 size_t, len)
2361 {
2362         if (!futex_cmpxchg_enabled)
2363                 return -ENOSYS;
2364         /*
2365          * The kernel knows only one size for now:
2366          */
2367         if (unlikely(len != sizeof(*head)))
2368                 return -EINVAL;
2369
2370         current->robust_list = head;
2371
2372         return 0;
2373 }
2374
2375 /**
2376  * sys_get_robust_list() - Get the robust-futex list head of a task
2377  * @pid:        pid of the process [zero for current task]
2378  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2379  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2380  */
2381 SYSCALL_DEFINE3(get_robust_list, int, pid,
2382                 struct robust_list_head __user * __user *, head_ptr,
2383                 size_t __user *, len_ptr)
2384 {
2385         struct robust_list_head __user *head;
2386         unsigned long ret;
2387         const struct cred *cred = current_cred(), *pcred;
2388
2389         if (!futex_cmpxchg_enabled)
2390                 return -ENOSYS;
2391
2392         if (!pid)
2393                 head = current->robust_list;
2394         else {
2395                 struct task_struct *p;
2396
2397                 ret = -ESRCH;
2398                 rcu_read_lock();
2399                 p = find_task_by_vpid(pid);
2400                 if (!p)
2401                         goto err_unlock;
2402                 ret = -EPERM;
2403                 pcred = __task_cred(p);
2404                 if (cred->euid != pcred->euid &&
2405                     cred->euid != pcred->uid &&
2406                     !capable(CAP_SYS_PTRACE))
2407                         goto err_unlock;
2408                 head = p->robust_list;
2409                 rcu_read_unlock();
2410         }
2411
2412         if (put_user(sizeof(*head), len_ptr))
2413                 return -EFAULT;
2414         return put_user(head, head_ptr);
2415
2416 err_unlock:
2417         rcu_read_unlock();
2418
2419         return ret;
2420 }
2421
2422 /*
2423  * Process a futex-list entry, check whether it's owned by the
2424  * dying task, and do notification if so:
2425  */
2426 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2427 {
2428         u32 uval, nval, mval;
2429
2430 retry:
2431         if (get_user(uval, uaddr))
2432                 return -1;
2433
2434         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2435                 /*
2436                  * Ok, this dying thread is truly holding a futex
2437                  * of interest. Set the OWNER_DIED bit atomically
2438                  * via cmpxchg, and if the value had FUTEX_WAITERS
2439                  * set, wake up a waiter (if any). (We have to do a
2440                  * futex_wake() even if OWNER_DIED is already set -
2441                  * to handle the rare but possible case of recursive
2442                  * thread-death.) The rest of the cleanup is done in
2443                  * userspace.
2444                  */
2445                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2446                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2447
2448                 if (nval == -EFAULT)
2449                         return -1;
2450
2451                 if (nval != uval)
2452                         goto retry;
2453
2454                 /*
2455                  * Wake robust non-PI futexes here. The wakeup of
2456                  * PI futexes happens in exit_pi_state():
2457                  */
2458                 if (!pi && (uval & FUTEX_WAITERS))
2459                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2460         }
2461         return 0;
2462 }
2463
2464 /*
2465  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2466  */
2467 static inline int fetch_robust_entry(struct robust_list __user **entry,
2468                                      struct robust_list __user * __user *head,
2469                                      unsigned int *pi)
2470 {
2471         unsigned long uentry;
2472
2473         if (get_user(uentry, (unsigned long __user *)head))
2474                 return -EFAULT;
2475
2476         *entry = (void __user *)(uentry & ~1UL);
2477         *pi = uentry & 1;
2478
2479         return 0;
2480 }
2481
2482 /*
2483  * Walk curr->robust_list (very carefully, it's a userspace list!)
2484  * and mark any locks found there dead, and notify any waiters.
2485  *
2486  * We silently return on any sign of list-walking problem.
2487  */
2488 void exit_robust_list(struct task_struct *curr)
2489 {
2490         struct robust_list_head __user *head = curr->robust_list;
2491         struct robust_list __user *entry, *next_entry, *pending;
2492         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2493         unsigned long futex_offset;
2494         int rc;
2495
2496         if (!futex_cmpxchg_enabled)
2497                 return;
2498
2499         /*
2500          * Fetch the list head (which was registered earlier, via
2501          * sys_set_robust_list()):
2502          */
2503         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2504                 return;
2505         /*
2506          * Fetch the relative futex offset:
2507          */
2508         if (get_user(futex_offset, &head->futex_offset))
2509                 return;
2510         /*
2511          * Fetch any possibly pending lock-add first, and handle it
2512          * if it exists:
2513          */
2514         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2515                 return;
2516
2517         next_entry = NULL;      /* avoid warning with gcc */
2518         while (entry != &head->list) {
2519                 /*
2520                  * Fetch the next entry in the list before calling
2521                  * handle_futex_death:
2522                  */
2523                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2524                 /*
2525                  * A pending lock might already be on the list, so
2526                  * don't process it twice:
2527                  */
2528                 if (entry != pending)
2529                         if (handle_futex_death((void __user *)entry + futex_offset,
2530                                                 curr, pi))
2531                                 return;
2532                 if (rc)
2533                         return;
2534                 entry = next_entry;
2535                 pi = next_pi;
2536                 /*
2537                  * Avoid excessively long or circular lists:
2538                  */
2539                 if (!--limit)
2540                         break;
2541
2542                 cond_resched();
2543         }
2544
2545         if (pending)
2546                 handle_futex_death((void __user *)pending + futex_offset,
2547                                    curr, pip);
2548 }
2549
2550 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2551                 u32 __user *uaddr2, u32 val2, u32 val3)
2552 {
2553         int clockrt, ret = -ENOSYS;
2554         int cmd = op & FUTEX_CMD_MASK;
2555         int fshared = 0;
2556
2557         if (!(op & FUTEX_PRIVATE_FLAG))
2558                 fshared = 1;
2559
2560         clockrt = op & FUTEX_CLOCK_REALTIME;
2561         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2562                 return -ENOSYS;
2563
2564         switch (cmd) {
2565         case FUTEX_WAIT:
2566                 val3 = FUTEX_BITSET_MATCH_ANY;
2567         case FUTEX_WAIT_BITSET:
2568                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2569                 break;
2570         case FUTEX_WAKE:
2571                 val3 = FUTEX_BITSET_MATCH_ANY;
2572         case FUTEX_WAKE_BITSET:
2573                 ret = futex_wake(uaddr, fshared, val, val3);
2574                 break;
2575         case FUTEX_REQUEUE:
2576                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2577                 break;
2578         case FUTEX_CMP_REQUEUE:
2579                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2580                                     0);
2581                 break;
2582         case FUTEX_WAKE_OP:
2583                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2584                 break;
2585         case FUTEX_LOCK_PI:
2586                 if (futex_cmpxchg_enabled)
2587                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2588                 break;
2589         case FUTEX_UNLOCK_PI:
2590                 if (futex_cmpxchg_enabled)
2591                         ret = futex_unlock_pi(uaddr, fshared);
2592                 break;
2593         case FUTEX_TRYLOCK_PI:
2594                 if (futex_cmpxchg_enabled)
2595                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2596                 break;
2597         case FUTEX_WAIT_REQUEUE_PI:
2598                 val3 = FUTEX_BITSET_MATCH_ANY;
2599                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2600                                             clockrt, uaddr2);
2601                 break;
2602         case FUTEX_CMP_REQUEUE_PI:
2603                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2604                                     1);
2605                 break;
2606         default:
2607                 ret = -ENOSYS;
2608         }
2609         return ret;
2610 }
2611
2612
2613 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2614                 struct timespec __user *, utime, u32 __user *, uaddr2,
2615                 u32, val3)
2616 {
2617         struct timespec ts;
2618         ktime_t t, *tp = NULL;
2619         u32 val2 = 0;
2620         int cmd = op & FUTEX_CMD_MASK;
2621
2622         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2623                       cmd == FUTEX_WAIT_BITSET ||
2624                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2625                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2626                         return -EFAULT;
2627                 if (!timespec_valid(&ts))
2628                         return -EINVAL;
2629
2630                 t = timespec_to_ktime(ts);
2631                 if (cmd == FUTEX_WAIT)
2632                         t = ktime_add_safe(ktime_get(), t);
2633                 tp = &t;
2634         }
2635         /*
2636          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2637          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2638          */
2639         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2640             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2641                 val2 = (u32) (unsigned long) utime;
2642
2643         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2644 }
2645
2646 static int __init futex_init(void)
2647 {
2648         u32 curval;
2649         int i;
2650
2651         /*
2652          * This will fail and we want it. Some arch implementations do
2653          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2654          * functionality. We want to know that before we call in any
2655          * of the complex code paths. Also we want to prevent
2656          * registration of robust lists in that case. NULL is
2657          * guaranteed to fault and we get -EFAULT on functional
2658          * implementation, the non-functional ones will return
2659          * -ENOSYS.
2660          */
2661         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2662         if (curval == -EFAULT)
2663                 futex_cmpxchg_enabled = 1;
2664
2665         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2666                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2667                 spin_lock_init(&futex_queues[i].lock);
2668         }
2669
2670         return 0;
2671 }
2672 __initcall(futex_init);