Make suspend abort reason logging depend on CONFIG_PM_SLEEP
[linux-3.10.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/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/freezer.h>
65
66 #include <asm/futex.h>
67
68 #include "rtmutex_common.h"
69
70 int __read_mostly futex_cmpxchg_enabled;
71
72 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
73
74 /*
75  * Futex flags used to encode options to functions and preserve them across
76  * restarts.
77  */
78 #define FLAGS_SHARED            0x01
79 #define FLAGS_CLOCKRT           0x02
80 #define FLAGS_HAS_TIMEOUT       0x04
81
82 /*
83  * Priority Inheritance state:
84  */
85 struct futex_pi_state {
86         /*
87          * list of 'owned' pi_state instances - these have to be
88          * cleaned up in do_exit() if the task exits prematurely:
89          */
90         struct list_head list;
91
92         /*
93          * The PI object:
94          */
95         struct rt_mutex pi_mutex;
96
97         struct task_struct *owner;
98         atomic_t refcount;
99
100         union futex_key key;
101 };
102
103 /**
104  * struct futex_q - The hashed futex queue entry, one per waiting task
105  * @list:               priority-sorted list of tasks waiting on this futex
106  * @task:               the task waiting on the futex
107  * @lock_ptr:           the hash bucket lock
108  * @key:                the key the futex is hashed on
109  * @pi_state:           optional priority inheritance state
110  * @rt_waiter:          rt_waiter storage for use with requeue_pi
111  * @requeue_pi_key:     the requeue_pi target futex key
112  * @bitset:             bitset for the optional bitmasked wakeup
113  *
114  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
115  * we can wake only the relevant ones (hashed queues may be shared).
116  *
117  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
118  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
119  * The order of wakeup is always to make the first condition true, then
120  * the second.
121  *
122  * PI futexes are typically woken before they are removed from the hash list via
123  * the rt_mutex code. See unqueue_me_pi().
124  */
125 struct futex_q {
126         struct plist_node list;
127
128         struct task_struct *task;
129         spinlock_t *lock_ptr;
130         union futex_key key;
131         struct futex_pi_state *pi_state;
132         struct rt_mutex_waiter *rt_waiter;
133         union futex_key *requeue_pi_key;
134         u32 bitset;
135 };
136
137 static const struct futex_q futex_q_init = {
138         /* list gets initialized in queue_me()*/
139         .key = FUTEX_KEY_INIT,
140         .bitset = FUTEX_BITSET_MATCH_ANY
141 };
142
143 /*
144  * Hash buckets are shared by all the futex_keys that hash to the same
145  * location.  Each key may have multiple futex_q structures, one for each task
146  * waiting on a futex.
147  */
148 struct futex_hash_bucket {
149         spinlock_t lock;
150         struct plist_head chain;
151 };
152
153 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
154
155 /*
156  * We hash on the keys returned from get_futex_key (see below).
157  */
158 static struct futex_hash_bucket *hash_futex(union futex_key *key)
159 {
160         u32 hash = jhash2((u32*)&key->both.word,
161                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
162                           key->both.offset);
163         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
164 }
165
166 /*
167  * Return 1 if two futex_keys are equal, 0 otherwise.
168  */
169 static inline int match_futex(union futex_key *key1, union futex_key *key2)
170 {
171         return (key1 && key2
172                 && key1->both.word == key2->both.word
173                 && key1->both.ptr == key2->both.ptr
174                 && key1->both.offset == key2->both.offset);
175 }
176
177 /*
178  * Take a reference to the resource addressed by a key.
179  * Can be called while holding spinlocks.
180  *
181  */
182 static void get_futex_key_refs(union futex_key *key)
183 {
184         if (!key->both.ptr)
185                 return;
186
187         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
188         case FUT_OFF_INODE:
189                 ihold(key->shared.inode);
190                 break;
191         case FUT_OFF_MMSHARED:
192                 atomic_inc(&key->private.mm->mm_count);
193                 break;
194         }
195 }
196
197 /*
198  * Drop a reference to the resource addressed by a key.
199  * The hash bucket spinlock must not be held.
200  */
201 static void drop_futex_key_refs(union futex_key *key)
202 {
203         if (!key->both.ptr) {
204                 /* If we're here then we tried to put a key we failed to get */
205                 WARN_ON_ONCE(1);
206                 return;
207         }
208
209         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
210         case FUT_OFF_INODE:
211                 iput(key->shared.inode);
212                 break;
213         case FUT_OFF_MMSHARED:
214                 mmdrop(key->private.mm);
215                 break;
216         }
217 }
218
219 /**
220  * get_futex_key() - Get parameters which are the keys for a futex
221  * @uaddr:      virtual address of the futex
222  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
223  * @key:        address where result is stored.
224  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
225  *              VERIFY_WRITE)
226  *
227  * Return: a negative error code or 0
228  *
229  * The key words are stored in *key on success.
230  *
231  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
232  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
233  * We can usually work out the index without swapping in the page.
234  *
235  * lock_page() might sleep, the caller should not hold a spinlock.
236  */
237 static int
238 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
239 {
240         unsigned long address = (unsigned long)uaddr;
241         struct mm_struct *mm = current->mm;
242         struct page *page, *page_head;
243         int err, ro = 0;
244
245         /*
246          * The futex address must be "naturally" aligned.
247          */
248         key->both.offset = address % PAGE_SIZE;
249         if (unlikely((address % sizeof(u32)) != 0))
250                 return -EINVAL;
251         address -= key->both.offset;
252
253         /*
254          * PROCESS_PRIVATE futexes are fast.
255          * As the mm cannot disappear under us and the 'key' only needs
256          * virtual address, we dont even have to find the underlying vma.
257          * Note : We do have to check 'uaddr' is a valid user address,
258          *        but access_ok() should be faster than find_vma()
259          */
260         if (!fshared) {
261                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
262                         return -EFAULT;
263                 key->private.mm = mm;
264                 key->private.address = address;
265                 get_futex_key_refs(key);
266                 return 0;
267         }
268
269 again:
270         err = get_user_pages_fast(address, 1, 1, &page);
271         /*
272          * If write access is not required (eg. FUTEX_WAIT), try
273          * and get read-only access.
274          */
275         if (err == -EFAULT && rw == VERIFY_READ) {
276                 err = get_user_pages_fast(address, 1, 0, &page);
277                 ro = 1;
278         }
279         if (err < 0)
280                 return err;
281         else
282                 err = 0;
283
284 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
285         page_head = page;
286         if (unlikely(PageTail(page))) {
287                 put_page(page);
288                 /* serialize against __split_huge_page_splitting() */
289                 local_irq_disable();
290                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
291                         page_head = compound_head(page);
292                         /*
293                          * page_head is valid pointer but we must pin
294                          * it before taking the PG_lock and/or
295                          * PG_compound_lock. The moment we re-enable
296                          * irqs __split_huge_page_splitting() can
297                          * return and the head page can be freed from
298                          * under us. We can't take the PG_lock and/or
299                          * PG_compound_lock on a page that could be
300                          * freed from under us.
301                          */
302                         if (page != page_head) {
303                                 get_page(page_head);
304                                 put_page(page);
305                         }
306                         local_irq_enable();
307                 } else {
308                         local_irq_enable();
309                         goto again;
310                 }
311         }
312 #else
313         page_head = compound_head(page);
314         if (page != page_head) {
315                 get_page(page_head);
316                 put_page(page);
317         }
318 #endif
319
320         lock_page(page_head);
321
322         /*
323          * If page_head->mapping is NULL, then it cannot be a PageAnon
324          * page; but it might be the ZERO_PAGE or in the gate area or
325          * in a special mapping (all cases which we are happy to fail);
326          * or it may have been a good file page when get_user_pages_fast
327          * found it, but truncated or holepunched or subjected to
328          * invalidate_complete_page2 before we got the page lock (also
329          * cases which we are happy to fail).  And we hold a reference,
330          * so refcount care in invalidate_complete_page's remove_mapping
331          * prevents drop_caches from setting mapping to NULL beneath us.
332          *
333          * The case we do have to guard against is when memory pressure made
334          * shmem_writepage move it from filecache to swapcache beneath us:
335          * an unlikely race, but we do need to retry for page_head->mapping.
336          */
337         if (!page_head->mapping) {
338                 int shmem_swizzled = PageSwapCache(page_head);
339                 unlock_page(page_head);
340                 put_page(page_head);
341                 if (shmem_swizzled)
342                         goto again;
343                 return -EFAULT;
344         }
345
346         /*
347          * Private mappings are handled in a simple way.
348          *
349          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
350          * it's a read-only handle, it's expected that futexes attach to
351          * the object not the particular process.
352          */
353         if (PageAnon(page_head)) {
354                 /*
355                  * A RO anonymous page will never change and thus doesn't make
356                  * sense for futex operations.
357                  */
358                 if (ro) {
359                         err = -EFAULT;
360                         goto out;
361                 }
362
363                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
364                 key->private.mm = mm;
365                 key->private.address = address;
366         } else {
367                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
368                 key->shared.inode = page_head->mapping->host;
369                 key->shared.pgoff = page_head->index;
370         }
371
372         get_futex_key_refs(key);
373
374 out:
375         unlock_page(page_head);
376         put_page(page_head);
377         return err;
378 }
379
380 static inline void put_futex_key(union futex_key *key)
381 {
382         drop_futex_key_refs(key);
383 }
384
385 /**
386  * fault_in_user_writeable() - Fault in user address and verify RW access
387  * @uaddr:      pointer to faulting user space address
388  *
389  * Slow path to fixup the fault we just took in the atomic write
390  * access to @uaddr.
391  *
392  * We have no generic implementation of a non-destructive write to the
393  * user address. We know that we faulted in the atomic pagefault
394  * disabled section so we can as well avoid the #PF overhead by
395  * calling get_user_pages() right away.
396  */
397 static int fault_in_user_writeable(u32 __user *uaddr)
398 {
399         struct mm_struct *mm = current->mm;
400         int ret;
401
402         down_read(&mm->mmap_sem);
403         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
404                                FAULT_FLAG_WRITE);
405         up_read(&mm->mmap_sem);
406
407         return ret < 0 ? ret : 0;
408 }
409
410 /**
411  * futex_top_waiter() - Return the highest priority waiter on a futex
412  * @hb:         the hash bucket the futex_q's reside in
413  * @key:        the futex key (to distinguish it from other futex futex_q's)
414  *
415  * Must be called with the hb lock held.
416  */
417 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
418                                         union futex_key *key)
419 {
420         struct futex_q *this;
421
422         plist_for_each_entry(this, &hb->chain, list) {
423                 if (match_futex(&this->key, key))
424                         return this;
425         }
426         return NULL;
427 }
428
429 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
430                                       u32 uval, u32 newval)
431 {
432         int ret;
433
434         pagefault_disable();
435         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
436         pagefault_enable();
437
438         return ret;
439 }
440
441 static int get_futex_value_locked(u32 *dest, u32 __user *from)
442 {
443         int ret;
444
445         pagefault_disable();
446         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
447         pagefault_enable();
448
449         return ret ? -EFAULT : 0;
450 }
451
452
453 /*
454  * PI code:
455  */
456 static int refill_pi_state_cache(void)
457 {
458         struct futex_pi_state *pi_state;
459
460         if (likely(current->pi_state_cache))
461                 return 0;
462
463         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
464
465         if (!pi_state)
466                 return -ENOMEM;
467
468         INIT_LIST_HEAD(&pi_state->list);
469         /* pi_mutex gets initialized later */
470         pi_state->owner = NULL;
471         atomic_set(&pi_state->refcount, 1);
472         pi_state->key = FUTEX_KEY_INIT;
473
474         current->pi_state_cache = pi_state;
475
476         return 0;
477 }
478
479 static struct futex_pi_state * alloc_pi_state(void)
480 {
481         struct futex_pi_state *pi_state = current->pi_state_cache;
482
483         WARN_ON(!pi_state);
484         current->pi_state_cache = NULL;
485
486         return pi_state;
487 }
488
489 static void free_pi_state(struct futex_pi_state *pi_state)
490 {
491         if (!atomic_dec_and_test(&pi_state->refcount))
492                 return;
493
494         /*
495          * If pi_state->owner is NULL, the owner is most probably dying
496          * and has cleaned up the pi_state already
497          */
498         if (pi_state->owner) {
499                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
500                 list_del_init(&pi_state->list);
501                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
502
503                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
504         }
505
506         if (current->pi_state_cache)
507                 kfree(pi_state);
508         else {
509                 /*
510                  * pi_state->list is already empty.
511                  * clear pi_state->owner.
512                  * refcount is at 0 - put it back to 1.
513                  */
514                 pi_state->owner = NULL;
515                 atomic_set(&pi_state->refcount, 1);
516                 current->pi_state_cache = pi_state;
517         }
518 }
519
520 /*
521  * Look up the task based on what TID userspace gave us.
522  * We dont trust it.
523  */
524 static struct task_struct * futex_find_get_task(pid_t pid)
525 {
526         struct task_struct *p;
527
528         rcu_read_lock();
529         p = find_task_by_vpid(pid);
530         if (p)
531                 get_task_struct(p);
532
533         rcu_read_unlock();
534
535         return p;
536 }
537
538 /*
539  * This task is holding PI mutexes at exit time => bad.
540  * Kernel cleans up PI-state, but userspace is likely hosed.
541  * (Robust-futex cleanup is separate and might save the day for userspace.)
542  */
543 void exit_pi_state_list(struct task_struct *curr)
544 {
545         struct list_head *next, *head = &curr->pi_state_list;
546         struct futex_pi_state *pi_state;
547         struct futex_hash_bucket *hb;
548         union futex_key key = FUTEX_KEY_INIT;
549
550         if (!futex_cmpxchg_enabled)
551                 return;
552         /*
553          * We are a ZOMBIE and nobody can enqueue itself on
554          * pi_state_list anymore, but we have to be careful
555          * versus waiters unqueueing themselves:
556          */
557         raw_spin_lock_irq(&curr->pi_lock);
558         while (!list_empty(head)) {
559
560                 next = head->next;
561                 pi_state = list_entry(next, struct futex_pi_state, list);
562                 key = pi_state->key;
563                 hb = hash_futex(&key);
564                 raw_spin_unlock_irq(&curr->pi_lock);
565
566                 spin_lock(&hb->lock);
567
568                 raw_spin_lock_irq(&curr->pi_lock);
569                 /*
570                  * We dropped the pi-lock, so re-check whether this
571                  * task still owns the PI-state:
572                  */
573                 if (head->next != next) {
574                         spin_unlock(&hb->lock);
575                         continue;
576                 }
577
578                 WARN_ON(pi_state->owner != curr);
579                 WARN_ON(list_empty(&pi_state->list));
580                 list_del_init(&pi_state->list);
581                 pi_state->owner = NULL;
582                 raw_spin_unlock_irq(&curr->pi_lock);
583
584                 rt_mutex_unlock(&pi_state->pi_mutex);
585
586                 spin_unlock(&hb->lock);
587
588                 raw_spin_lock_irq(&curr->pi_lock);
589         }
590         raw_spin_unlock_irq(&curr->pi_lock);
591 }
592
593 /*
594  * We need to check the following states:
595  *
596  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
597  *
598  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
599  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
600  *
601  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
602  *
603  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
604  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
605  *
606  * [6]  Found  | Found    | task      | 0         | 1      | Valid
607  *
608  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
609  *
610  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
611  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
612  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
613  *
614  * [1]  Indicates that the kernel can acquire the futex atomically. We
615  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
616  *
617  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
618  *      thread is found then it indicates that the owner TID has died.
619  *
620  * [3]  Invalid. The waiter is queued on a non PI futex
621  *
622  * [4]  Valid state after exit_robust_list(), which sets the user space
623  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
624  *
625  * [5]  The user space value got manipulated between exit_robust_list()
626  *      and exit_pi_state_list()
627  *
628  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
629  *      the pi_state but cannot access the user space value.
630  *
631  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
632  *
633  * [8]  Owner and user space value match
634  *
635  * [9]  There is no transient state which sets the user space TID to 0
636  *      except exit_robust_list(), but this is indicated by the
637  *      FUTEX_OWNER_DIED bit. See [4]
638  *
639  * [10] There is no transient state which leaves owner and user space
640  *      TID out of sync.
641  */
642 static int
643 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
644                 union futex_key *key, struct futex_pi_state **ps)
645 {
646         struct futex_pi_state *pi_state = NULL;
647         struct futex_q *this, *next;
648         struct plist_head *head;
649         struct task_struct *p;
650         pid_t pid = uval & FUTEX_TID_MASK;
651
652         head = &hb->chain;
653
654         plist_for_each_entry_safe(this, next, head, list) {
655                 if (match_futex(&this->key, key)) {
656                         /*
657                          * Sanity check the waiter before increasing
658                          * the refcount and attaching to it.
659                          */
660                         pi_state = this->pi_state;
661                         /*
662                          * Userspace might have messed up non-PI and
663                          * PI futexes [3]
664                          */
665                         if (unlikely(!pi_state))
666                                 return -EINVAL;
667
668                         WARN_ON(!atomic_read(&pi_state->refcount));
669
670                         /*
671                          * Handle the owner died case:
672                          */
673                         if (uval & FUTEX_OWNER_DIED) {
674                                 /*
675                                  * exit_pi_state_list sets owner to NULL and
676                                  * wakes the topmost waiter. The task which
677                                  * acquires the pi_state->rt_mutex will fixup
678                                  * owner.
679                                  */
680                                 if (!pi_state->owner) {
681                                         /*
682                                          * No pi state owner, but the user
683                                          * space TID is not 0. Inconsistent
684                                          * state. [5]
685                                          */
686                                         if (pid)
687                                                 return -EINVAL;
688                                         /*
689                                          * Take a ref on the state and
690                                          * return. [4]
691                                          */
692                                         goto out_state;
693                                 }
694
695                                 /*
696                                  * If TID is 0, then either the dying owner
697                                  * has not yet executed exit_pi_state_list()
698                                  * or some waiter acquired the rtmutex in the
699                                  * pi state, but did not yet fixup the TID in
700                                  * user space.
701                                  *
702                                  * Take a ref on the state and return. [6]
703                                  */
704                                 if (!pid)
705                                         goto out_state;
706                         } else {
707                                 /*
708                                  * If the owner died bit is not set,
709                                  * then the pi_state must have an
710                                  * owner. [7]
711                                  */
712                                 if (!pi_state->owner)
713                                         return -EINVAL;
714                         }
715
716                         /*
717                          * Bail out if user space manipulated the
718                          * futex value. If pi state exists then the
719                          * owner TID must be the same as the user
720                          * space TID. [9/10]
721                          */
722                         if (pid != task_pid_vnr(pi_state->owner))
723                                 return -EINVAL;
724
725                 out_state:
726                         atomic_inc(&pi_state->refcount);
727                         *ps = pi_state;
728                         return 0;
729                 }
730         }
731
732         /*
733          * We are the first waiter - try to look up the real owner and attach
734          * the new pi_state to it, but bail out when TID = 0 [1]
735          */
736         if (!pid)
737                 return -ESRCH;
738         p = futex_find_get_task(pid);
739         if (!p)
740                 return -ESRCH;
741
742         /*
743          * We need to look at the task state flags to figure out,
744          * whether the task is exiting. To protect against the do_exit
745          * change of the task flags, we do this protected by
746          * p->pi_lock:
747          */
748         raw_spin_lock_irq(&p->pi_lock);
749         if (unlikely(p->flags & PF_EXITING)) {
750                 /*
751                  * The task is on the way out. When PF_EXITPIDONE is
752                  * set, we know that the task has finished the
753                  * cleanup:
754                  */
755                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
756
757                 raw_spin_unlock_irq(&p->pi_lock);
758                 put_task_struct(p);
759                 return ret;
760         }
761
762         /*
763          * No existing pi state. First waiter. [2]
764          */
765         pi_state = alloc_pi_state();
766
767         /*
768          * Initialize the pi_mutex in locked state and make 'p'
769          * the owner of it:
770          */
771         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
772
773         /* Store the key for possible exit cleanups: */
774         pi_state->key = *key;
775
776         WARN_ON(!list_empty(&pi_state->list));
777         list_add(&pi_state->list, &p->pi_state_list);
778         pi_state->owner = p;
779         raw_spin_unlock_irq(&p->pi_lock);
780
781         put_task_struct(p);
782
783         *ps = pi_state;
784
785         return 0;
786 }
787
788 /**
789  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
790  * @uaddr:              the pi futex user address
791  * @hb:                 the pi futex hash bucket
792  * @key:                the futex key associated with uaddr and hb
793  * @ps:                 the pi_state pointer where we store the result of the
794  *                      lookup
795  * @task:               the task to perform the atomic lock work for.  This will
796  *                      be "current" except in the case of requeue pi.
797  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
798  *
799  * Return:
800  *  0 - ready to wait;
801  *  1 - acquired the lock;
802  * <0 - error
803  *
804  * The hb->lock and futex_key refs shall be held by the caller.
805  */
806 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
807                                 union futex_key *key,
808                                 struct futex_pi_state **ps,
809                                 struct task_struct *task, int set_waiters)
810 {
811         int lock_taken, ret, force_take = 0;
812         u32 uval, newval, curval, vpid = task_pid_vnr(task);
813
814 retry:
815         ret = lock_taken = 0;
816
817         /*
818          * To avoid races, we attempt to take the lock here again
819          * (by doing a 0 -> TID atomic cmpxchg), while holding all
820          * the locks. It will most likely not succeed.
821          */
822         newval = vpid;
823         if (set_waiters)
824                 newval |= FUTEX_WAITERS;
825
826         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
827                 return -EFAULT;
828
829         /*
830          * Detect deadlocks.
831          */
832         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
833                 return -EDEADLK;
834
835         /*
836          * Surprise - we got the lock, but we do not trust user space at all.
837          */
838         if (unlikely(!curval)) {
839                 /*
840                  * We verify whether there is kernel state for this
841                  * futex. If not, we can safely assume, that the 0 ->
842                  * TID transition is correct. If state exists, we do
843                  * not bother to fixup the user space state as it was
844                  * corrupted already.
845                  */
846                 return futex_top_waiter(hb, key) ? -EINVAL : 1;
847         }
848
849         uval = curval;
850
851         /*
852          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
853          * to wake at the next unlock.
854          */
855         newval = curval | FUTEX_WAITERS;
856
857         /*
858          * Should we force take the futex? See below.
859          */
860         if (unlikely(force_take)) {
861                 /*
862                  * Keep the OWNER_DIED and the WAITERS bit and set the
863                  * new TID value.
864                  */
865                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
866                 force_take = 0;
867                 lock_taken = 1;
868         }
869
870         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
871                 return -EFAULT;
872         if (unlikely(curval != uval))
873                 goto retry;
874
875         /*
876          * We took the lock due to forced take over.
877          */
878         if (unlikely(lock_taken))
879                 return 1;
880
881         /*
882          * We dont have the lock. Look up the PI state (or create it if
883          * we are the first waiter):
884          */
885         ret = lookup_pi_state(uval, hb, key, ps);
886
887         if (unlikely(ret)) {
888                 switch (ret) {
889                 case -ESRCH:
890                         /*
891                          * We failed to find an owner for this
892                          * futex. So we have no pi_state to block
893                          * on. This can happen in two cases:
894                          *
895                          * 1) The owner died
896                          * 2) A stale FUTEX_WAITERS bit
897                          *
898                          * Re-read the futex value.
899                          */
900                         if (get_futex_value_locked(&curval, uaddr))
901                                 return -EFAULT;
902
903                         /*
904                          * If the owner died or we have a stale
905                          * WAITERS bit the owner TID in the user space
906                          * futex is 0.
907                          */
908                         if (!(curval & FUTEX_TID_MASK)) {
909                                 force_take = 1;
910                                 goto retry;
911                         }
912                 default:
913                         break;
914                 }
915         }
916
917         return ret;
918 }
919
920 /**
921  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
922  * @q:  The futex_q to unqueue
923  *
924  * The q->lock_ptr must not be NULL and must be held by the caller.
925  */
926 static void __unqueue_futex(struct futex_q *q)
927 {
928         struct futex_hash_bucket *hb;
929
930         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
931             || WARN_ON(plist_node_empty(&q->list)))
932                 return;
933
934         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
935         plist_del(&q->list, &hb->chain);
936 }
937
938 /*
939  * The hash bucket lock must be held when this is called.
940  * Afterwards, the futex_q must not be accessed.
941  */
942 static void wake_futex(struct futex_q *q)
943 {
944         struct task_struct *p = q->task;
945
946         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
947                 return;
948
949         /*
950          * We set q->lock_ptr = NULL _before_ we wake up the task. If
951          * a non-futex wake up happens on another CPU then the task
952          * might exit and p would dereference a non-existing task
953          * struct. Prevent this by holding a reference on p across the
954          * wake up.
955          */
956         get_task_struct(p);
957
958         __unqueue_futex(q);
959         /*
960          * The waiting task can free the futex_q as soon as
961          * q->lock_ptr = NULL is written, without taking any locks. A
962          * memory barrier is required here to prevent the following
963          * store to lock_ptr from getting ahead of the plist_del.
964          */
965         smp_wmb();
966         q->lock_ptr = NULL;
967
968         wake_up_state(p, TASK_NORMAL);
969         put_task_struct(p);
970 }
971
972 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
973 {
974         struct task_struct *new_owner;
975         struct futex_pi_state *pi_state = this->pi_state;
976         u32 uninitialized_var(curval), newval;
977         int ret = 0;
978
979         if (!pi_state)
980                 return -EINVAL;
981
982         /*
983          * If current does not own the pi_state then the futex is
984          * inconsistent and user space fiddled with the futex value.
985          */
986         if (pi_state->owner != current)
987                 return -EINVAL;
988
989         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
990         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
991
992         /*
993          * It is possible that the next waiter (the one that brought
994          * this owner to the kernel) timed out and is no longer
995          * waiting on the lock.
996          */
997         if (!new_owner)
998                 new_owner = this->task;
999
1000         /*
1001          * We pass it to the next owner. The WAITERS bit is always
1002          * kept enabled while there is PI state around. We cleanup the
1003          * owner died bit, because we are the owner.
1004          */
1005         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1006
1007         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1008                 ret = -EFAULT;
1009         else if (curval != uval)
1010                 ret = -EINVAL;
1011         if (ret) {
1012                 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1013                 return ret;
1014         }
1015
1016         raw_spin_lock_irq(&pi_state->owner->pi_lock);
1017         WARN_ON(list_empty(&pi_state->list));
1018         list_del_init(&pi_state->list);
1019         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1020
1021         raw_spin_lock_irq(&new_owner->pi_lock);
1022         WARN_ON(!list_empty(&pi_state->list));
1023         list_add(&pi_state->list, &new_owner->pi_state_list);
1024         pi_state->owner = new_owner;
1025         raw_spin_unlock_irq(&new_owner->pi_lock);
1026
1027         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1028         rt_mutex_unlock(&pi_state->pi_mutex);
1029
1030         return 0;
1031 }
1032
1033 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1034 {
1035         u32 uninitialized_var(oldval);
1036
1037         /*
1038          * There is no waiter, so we unlock the futex. The owner died
1039          * bit has not to be preserved here. We are the owner:
1040          */
1041         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1042                 return -EFAULT;
1043         if (oldval != uval)
1044                 return -EAGAIN;
1045
1046         return 0;
1047 }
1048
1049 /*
1050  * Express the locking dependencies for lockdep:
1051  */
1052 static inline void
1053 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1054 {
1055         if (hb1 <= hb2) {
1056                 spin_lock(&hb1->lock);
1057                 if (hb1 < hb2)
1058                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1059         } else { /* hb1 > hb2 */
1060                 spin_lock(&hb2->lock);
1061                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1062         }
1063 }
1064
1065 static inline void
1066 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1067 {
1068         spin_unlock(&hb1->lock);
1069         if (hb1 != hb2)
1070                 spin_unlock(&hb2->lock);
1071 }
1072
1073 /*
1074  * Wake up waiters matching bitset queued on this futex (uaddr).
1075  */
1076 static int
1077 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1078 {
1079         struct futex_hash_bucket *hb;
1080         struct futex_q *this, *next;
1081         struct plist_head *head;
1082         union futex_key key = FUTEX_KEY_INIT;
1083         int ret;
1084
1085         if (!bitset)
1086                 return -EINVAL;
1087
1088         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1089         if (unlikely(ret != 0))
1090                 goto out;
1091
1092         hb = hash_futex(&key);
1093         spin_lock(&hb->lock);
1094         head = &hb->chain;
1095
1096         plist_for_each_entry_safe(this, next, head, list) {
1097                 if (match_futex (&this->key, &key)) {
1098                         if (this->pi_state || this->rt_waiter) {
1099                                 ret = -EINVAL;
1100                                 break;
1101                         }
1102
1103                         /* Check if one of the bits is set in both bitsets */
1104                         if (!(this->bitset & bitset))
1105                                 continue;
1106
1107                         wake_futex(this);
1108                         if (++ret >= nr_wake)
1109                                 break;
1110                 }
1111         }
1112
1113         spin_unlock(&hb->lock);
1114         put_futex_key(&key);
1115 out:
1116         return ret;
1117 }
1118
1119 /*
1120  * Wake up all waiters hashed on the physical page that is mapped
1121  * to this virtual address:
1122  */
1123 static int
1124 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1125               int nr_wake, int nr_wake2, int op)
1126 {
1127         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1128         struct futex_hash_bucket *hb1, *hb2;
1129         struct plist_head *head;
1130         struct futex_q *this, *next;
1131         int ret, op_ret;
1132
1133 retry:
1134         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1135         if (unlikely(ret != 0))
1136                 goto out;
1137         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1138         if (unlikely(ret != 0))
1139                 goto out_put_key1;
1140
1141         hb1 = hash_futex(&key1);
1142         hb2 = hash_futex(&key2);
1143
1144 retry_private:
1145         double_lock_hb(hb1, hb2);
1146         op_ret = futex_atomic_op_inuser(op, uaddr2);
1147         if (unlikely(op_ret < 0)) {
1148
1149                 double_unlock_hb(hb1, hb2);
1150
1151 #ifndef CONFIG_MMU
1152                 /*
1153                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1154                  * but we might get them from range checking
1155                  */
1156                 ret = op_ret;
1157                 goto out_put_keys;
1158 #endif
1159
1160                 if (unlikely(op_ret != -EFAULT)) {
1161                         ret = op_ret;
1162                         goto out_put_keys;
1163                 }
1164
1165                 ret = fault_in_user_writeable(uaddr2);
1166                 if (ret)
1167                         goto out_put_keys;
1168
1169                 if (!(flags & FLAGS_SHARED))
1170                         goto retry_private;
1171
1172                 put_futex_key(&key2);
1173                 put_futex_key(&key1);
1174                 goto retry;
1175         }
1176
1177         head = &hb1->chain;
1178
1179         plist_for_each_entry_safe(this, next, head, list) {
1180                 if (match_futex (&this->key, &key1)) {
1181                         if (this->pi_state || this->rt_waiter) {
1182                                 ret = -EINVAL;
1183                                 goto out_unlock;
1184                         }
1185                         wake_futex(this);
1186                         if (++ret >= nr_wake)
1187                                 break;
1188                 }
1189         }
1190
1191         if (op_ret > 0) {
1192                 head = &hb2->chain;
1193
1194                 op_ret = 0;
1195                 plist_for_each_entry_safe(this, next, head, list) {
1196                         if (match_futex (&this->key, &key2)) {
1197                                 if (this->pi_state || this->rt_waiter) {
1198                                         ret = -EINVAL;
1199                                         goto out_unlock;
1200                                 }
1201                                 wake_futex(this);
1202                                 if (++op_ret >= nr_wake2)
1203                                         break;
1204                         }
1205                 }
1206                 ret += op_ret;
1207         }
1208
1209 out_unlock:
1210         double_unlock_hb(hb1, hb2);
1211 out_put_keys:
1212         put_futex_key(&key2);
1213 out_put_key1:
1214         put_futex_key(&key1);
1215 out:
1216         return ret;
1217 }
1218
1219 /**
1220  * requeue_futex() - Requeue a futex_q from one hb to another
1221  * @q:          the futex_q to requeue
1222  * @hb1:        the source hash_bucket
1223  * @hb2:        the target hash_bucket
1224  * @key2:       the new key for the requeued futex_q
1225  */
1226 static inline
1227 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1228                    struct futex_hash_bucket *hb2, union futex_key *key2)
1229 {
1230
1231         /*
1232          * If key1 and key2 hash to the same bucket, no need to
1233          * requeue.
1234          */
1235         if (likely(&hb1->chain != &hb2->chain)) {
1236                 plist_del(&q->list, &hb1->chain);
1237                 plist_add(&q->list, &hb2->chain);
1238                 q->lock_ptr = &hb2->lock;
1239         }
1240         get_futex_key_refs(key2);
1241         q->key = *key2;
1242 }
1243
1244 /**
1245  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1246  * @q:          the futex_q
1247  * @key:        the key of the requeue target futex
1248  * @hb:         the hash_bucket of the requeue target futex
1249  *
1250  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1251  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1252  * to the requeue target futex so the waiter can detect the wakeup on the right
1253  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1254  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1255  * to protect access to the pi_state to fixup the owner later.  Must be called
1256  * with both q->lock_ptr and hb->lock held.
1257  */
1258 static inline
1259 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1260                            struct futex_hash_bucket *hb)
1261 {
1262         get_futex_key_refs(key);
1263         q->key = *key;
1264
1265         __unqueue_futex(q);
1266
1267         WARN_ON(!q->rt_waiter);
1268         q->rt_waiter = NULL;
1269
1270         q->lock_ptr = &hb->lock;
1271
1272         wake_up_state(q->task, TASK_NORMAL);
1273 }
1274
1275 /**
1276  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1277  * @pifutex:            the user address of the to futex
1278  * @hb1:                the from futex hash bucket, must be locked by the caller
1279  * @hb2:                the to futex hash bucket, must be locked by the caller
1280  * @key1:               the from futex key
1281  * @key2:               the to futex key
1282  * @ps:                 address to store the pi_state pointer
1283  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1284  *
1285  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1286  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1287  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1288  * hb1 and hb2 must be held by the caller.
1289  *
1290  * Return:
1291  *  0 - failed to acquire the lock atomically;
1292  *  1 - acquired the lock;
1293  * <0 - error
1294  */
1295 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1296                                  struct futex_hash_bucket *hb1,
1297                                  struct futex_hash_bucket *hb2,
1298                                  union futex_key *key1, union futex_key *key2,
1299                                  struct futex_pi_state **ps, int set_waiters)
1300 {
1301         struct futex_q *top_waiter = NULL;
1302         u32 curval;
1303         int ret;
1304
1305         if (get_futex_value_locked(&curval, pifutex))
1306                 return -EFAULT;
1307
1308         /*
1309          * Find the top_waiter and determine if there are additional waiters.
1310          * If the caller intends to requeue more than 1 waiter to pifutex,
1311          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1312          * as we have means to handle the possible fault.  If not, don't set
1313          * the bit unecessarily as it will force the subsequent unlock to enter
1314          * the kernel.
1315          */
1316         top_waiter = futex_top_waiter(hb1, key1);
1317
1318         /* There are no waiters, nothing for us to do. */
1319         if (!top_waiter)
1320                 return 0;
1321
1322         /* Ensure we requeue to the expected futex. */
1323         if (!match_futex(top_waiter->requeue_pi_key, key2))
1324                 return -EINVAL;
1325
1326         /*
1327          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1328          * the contended case or if set_waiters is 1.  The pi_state is returned
1329          * in ps in contended cases.
1330          */
1331         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1332                                    set_waiters);
1333         if (ret == 1)
1334                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1335
1336         return ret;
1337 }
1338
1339 /**
1340  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1341  * @uaddr1:     source futex user address
1342  * @flags:      futex flags (FLAGS_SHARED, etc.)
1343  * @uaddr2:     target futex user address
1344  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1345  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1346  * @cmpval:     @uaddr1 expected value (or %NULL)
1347  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1348  *              pi futex (pi to pi requeue is not supported)
1349  *
1350  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1351  * uaddr2 atomically on behalf of the top waiter.
1352  *
1353  * Return:
1354  * >=0 - on success, the number of tasks requeued or woken;
1355  *  <0 - on error
1356  */
1357 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1358                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1359                          u32 *cmpval, int requeue_pi)
1360 {
1361         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1362         int drop_count = 0, task_count = 0, ret;
1363         struct futex_pi_state *pi_state = NULL;
1364         struct futex_hash_bucket *hb1, *hb2;
1365         struct plist_head *head1;
1366         struct futex_q *this, *next;
1367         u32 curval2;
1368
1369         if (requeue_pi) {
1370                 /*
1371                  * Requeue PI only works on two distinct uaddrs. This
1372                  * check is only valid for private futexes. See below.
1373                  */
1374                 if (uaddr1 == uaddr2)
1375                         return -EINVAL;
1376
1377                 /*
1378                  * requeue_pi requires a pi_state, try to allocate it now
1379                  * without any locks in case it fails.
1380                  */
1381                 if (refill_pi_state_cache())
1382                         return -ENOMEM;
1383                 /*
1384                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1385                  * + nr_requeue, since it acquires the rt_mutex prior to
1386                  * returning to userspace, so as to not leave the rt_mutex with
1387                  * waiters and no owner.  However, second and third wake-ups
1388                  * cannot be predicted as they involve race conditions with the
1389                  * first wake and a fault while looking up the pi_state.  Both
1390                  * pthread_cond_signal() and pthread_cond_broadcast() should
1391                  * use nr_wake=1.
1392                  */
1393                 if (nr_wake != 1)
1394                         return -EINVAL;
1395         }
1396
1397 retry:
1398         if (pi_state != NULL) {
1399                 /*
1400                  * We will have to lookup the pi_state again, so free this one
1401                  * to keep the accounting correct.
1402                  */
1403                 free_pi_state(pi_state);
1404                 pi_state = NULL;
1405         }
1406
1407         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1408         if (unlikely(ret != 0))
1409                 goto out;
1410         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1411                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1412         if (unlikely(ret != 0))
1413                 goto out_put_key1;
1414
1415         /*
1416          * The check above which compares uaddrs is not sufficient for
1417          * shared futexes. We need to compare the keys:
1418          */
1419         if (requeue_pi && match_futex(&key1, &key2)) {
1420                 ret = -EINVAL;
1421                 goto out_put_keys;
1422         }
1423
1424         hb1 = hash_futex(&key1);
1425         hb2 = hash_futex(&key2);
1426
1427 retry_private:
1428         double_lock_hb(hb1, hb2);
1429
1430         if (likely(cmpval != NULL)) {
1431                 u32 curval;
1432
1433                 ret = get_futex_value_locked(&curval, uaddr1);
1434
1435                 if (unlikely(ret)) {
1436                         double_unlock_hb(hb1, hb2);
1437
1438                         ret = get_user(curval, uaddr1);
1439                         if (ret)
1440                                 goto out_put_keys;
1441
1442                         if (!(flags & FLAGS_SHARED))
1443                                 goto retry_private;
1444
1445                         put_futex_key(&key2);
1446                         put_futex_key(&key1);
1447                         goto retry;
1448                 }
1449                 if (curval != *cmpval) {
1450                         ret = -EAGAIN;
1451                         goto out_unlock;
1452                 }
1453         }
1454
1455         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1456                 /*
1457                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1458                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1459                  * bit.  We force this here where we are able to easily handle
1460                  * faults rather in the requeue loop below.
1461                  */
1462                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1463                                                  &key2, &pi_state, nr_requeue);
1464
1465                 /*
1466                  * At this point the top_waiter has either taken uaddr2 or is
1467                  * waiting on it.  If the former, then the pi_state will not
1468                  * exist yet, look it up one more time to ensure we have a
1469                  * reference to it.
1470                  */
1471                 if (ret == 1) {
1472                         WARN_ON(pi_state);
1473                         drop_count++;
1474                         task_count++;
1475                         ret = get_futex_value_locked(&curval2, uaddr2);
1476                         if (!ret)
1477                                 ret = lookup_pi_state(curval2, hb2, &key2,
1478                                                       &pi_state);
1479                 }
1480
1481                 switch (ret) {
1482                 case 0:
1483                         break;
1484                 case -EFAULT:
1485                         double_unlock_hb(hb1, hb2);
1486                         put_futex_key(&key2);
1487                         put_futex_key(&key1);
1488                         ret = fault_in_user_writeable(uaddr2);
1489                         if (!ret)
1490                                 goto retry;
1491                         goto out;
1492                 case -EAGAIN:
1493                         /* The owner was exiting, try again. */
1494                         double_unlock_hb(hb1, hb2);
1495                         put_futex_key(&key2);
1496                         put_futex_key(&key1);
1497                         cond_resched();
1498                         goto retry;
1499                 default:
1500                         goto out_unlock;
1501                 }
1502         }
1503
1504         head1 = &hb1->chain;
1505         plist_for_each_entry_safe(this, next, head1, list) {
1506                 if (task_count - nr_wake >= nr_requeue)
1507                         break;
1508
1509                 if (!match_futex(&this->key, &key1))
1510                         continue;
1511
1512                 /*
1513                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1514                  * be paired with each other and no other futex ops.
1515                  *
1516                  * We should never be requeueing a futex_q with a pi_state,
1517                  * which is awaiting a futex_unlock_pi().
1518                  */
1519                 if ((requeue_pi && !this->rt_waiter) ||
1520                     (!requeue_pi && this->rt_waiter) ||
1521                     this->pi_state) {
1522                         ret = -EINVAL;
1523                         break;
1524                 }
1525
1526                 /*
1527                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1528                  * lock, we already woke the top_waiter.  If not, it will be
1529                  * woken by futex_unlock_pi().
1530                  */
1531                 if (++task_count <= nr_wake && !requeue_pi) {
1532                         wake_futex(this);
1533                         continue;
1534                 }
1535
1536                 /* Ensure we requeue to the expected futex for requeue_pi. */
1537                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1538                         ret = -EINVAL;
1539                         break;
1540                 }
1541
1542                 /*
1543                  * Requeue nr_requeue waiters and possibly one more in the case
1544                  * of requeue_pi if we couldn't acquire the lock atomically.
1545                  */
1546                 if (requeue_pi) {
1547                         /* Prepare the waiter to take the rt_mutex. */
1548                         atomic_inc(&pi_state->refcount);
1549                         this->pi_state = pi_state;
1550                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1551                                                         this->rt_waiter,
1552                                                         this->task, 1);
1553                         if (ret == 1) {
1554                                 /* We got the lock. */
1555                                 requeue_pi_wake_futex(this, &key2, hb2);
1556                                 drop_count++;
1557                                 continue;
1558                         } else if (ret) {
1559                                 /* -EDEADLK */
1560                                 this->pi_state = NULL;
1561                                 free_pi_state(pi_state);
1562                                 goto out_unlock;
1563                         }
1564                 }
1565                 requeue_futex(this, hb1, hb2, &key2);
1566                 drop_count++;
1567         }
1568
1569 out_unlock:
1570         double_unlock_hb(hb1, hb2);
1571
1572         /*
1573          * drop_futex_key_refs() must be called outside the spinlocks. During
1574          * the requeue we moved futex_q's from the hash bucket at key1 to the
1575          * one at key2 and updated their key pointer.  We no longer need to
1576          * hold the references to key1.
1577          */
1578         while (--drop_count >= 0)
1579                 drop_futex_key_refs(&key1);
1580
1581 out_put_keys:
1582         put_futex_key(&key2);
1583 out_put_key1:
1584         put_futex_key(&key1);
1585 out:
1586         if (pi_state != NULL)
1587                 free_pi_state(pi_state);
1588         return ret ? ret : task_count;
1589 }
1590
1591 /* The key must be already stored in q->key. */
1592 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1593         __acquires(&hb->lock)
1594 {
1595         struct futex_hash_bucket *hb;
1596
1597         hb = hash_futex(&q->key);
1598         q->lock_ptr = &hb->lock;
1599
1600         spin_lock(&hb->lock);
1601         return hb;
1602 }
1603
1604 static inline void
1605 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1606         __releases(&hb->lock)
1607 {
1608         spin_unlock(&hb->lock);
1609 }
1610
1611 /**
1612  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1613  * @q:  The futex_q to enqueue
1614  * @hb: The destination hash bucket
1615  *
1616  * The hb->lock must be held by the caller, and is released here. A call to
1617  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1618  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1619  * or nothing if the unqueue is done as part of the wake process and the unqueue
1620  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1621  * an example).
1622  */
1623 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1624         __releases(&hb->lock)
1625 {
1626         int prio;
1627
1628         /*
1629          * The priority used to register this element is
1630          * - either the real thread-priority for the real-time threads
1631          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1632          * - or MAX_RT_PRIO for non-RT threads.
1633          * Thus, all RT-threads are woken first in priority order, and
1634          * the others are woken last, in FIFO order.
1635          */
1636         prio = min(current->normal_prio, MAX_RT_PRIO);
1637
1638         plist_node_init(&q->list, prio);
1639         plist_add(&q->list, &hb->chain);
1640         q->task = current;
1641         spin_unlock(&hb->lock);
1642 }
1643
1644 /**
1645  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1646  * @q:  The futex_q to unqueue
1647  *
1648  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1649  * be paired with exactly one earlier call to queue_me().
1650  *
1651  * Return:
1652  *   1 - if the futex_q was still queued (and we removed unqueued it);
1653  *   0 - if the futex_q was already removed by the waking thread
1654  */
1655 static int unqueue_me(struct futex_q *q)
1656 {
1657         spinlock_t *lock_ptr;
1658         int ret = 0;
1659
1660         /* In the common case we don't take the spinlock, which is nice. */
1661 retry:
1662         lock_ptr = q->lock_ptr;
1663         barrier();
1664         if (lock_ptr != NULL) {
1665                 spin_lock(lock_ptr);
1666                 /*
1667                  * q->lock_ptr can change between reading it and
1668                  * spin_lock(), causing us to take the wrong lock.  This
1669                  * corrects the race condition.
1670                  *
1671                  * Reasoning goes like this: if we have the wrong lock,
1672                  * q->lock_ptr must have changed (maybe several times)
1673                  * between reading it and the spin_lock().  It can
1674                  * change again after the spin_lock() but only if it was
1675                  * already changed before the spin_lock().  It cannot,
1676                  * however, change back to the original value.  Therefore
1677                  * we can detect whether we acquired the correct lock.
1678                  */
1679                 if (unlikely(lock_ptr != q->lock_ptr)) {
1680                         spin_unlock(lock_ptr);
1681                         goto retry;
1682                 }
1683                 __unqueue_futex(q);
1684
1685                 BUG_ON(q->pi_state);
1686
1687                 spin_unlock(lock_ptr);
1688                 ret = 1;
1689         }
1690
1691         drop_futex_key_refs(&q->key);
1692         return ret;
1693 }
1694
1695 /*
1696  * PI futexes can not be requeued and must remove themself from the
1697  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1698  * and dropped here.
1699  */
1700 static void unqueue_me_pi(struct futex_q *q)
1701         __releases(q->lock_ptr)
1702 {
1703         __unqueue_futex(q);
1704
1705         BUG_ON(!q->pi_state);
1706         free_pi_state(q->pi_state);
1707         q->pi_state = NULL;
1708
1709         spin_unlock(q->lock_ptr);
1710 }
1711
1712 /*
1713  * Fixup the pi_state owner with the new owner.
1714  *
1715  * Must be called with hash bucket lock held and mm->sem held for non
1716  * private futexes.
1717  */
1718 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1719                                 struct task_struct *newowner)
1720 {
1721         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1722         struct futex_pi_state *pi_state = q->pi_state;
1723         struct task_struct *oldowner = pi_state->owner;
1724         u32 uval, uninitialized_var(curval), newval;
1725         int ret;
1726
1727         /* Owner died? */
1728         if (!pi_state->owner)
1729                 newtid |= FUTEX_OWNER_DIED;
1730
1731         /*
1732          * We are here either because we stole the rtmutex from the
1733          * previous highest priority waiter or we are the highest priority
1734          * waiter but failed to get the rtmutex the first time.
1735          * We have to replace the newowner TID in the user space variable.
1736          * This must be atomic as we have to preserve the owner died bit here.
1737          *
1738          * Note: We write the user space value _before_ changing the pi_state
1739          * because we can fault here. Imagine swapped out pages or a fork
1740          * that marked all the anonymous memory readonly for cow.
1741          *
1742          * Modifying pi_state _before_ the user space value would
1743          * leave the pi_state in an inconsistent state when we fault
1744          * here, because we need to drop the hash bucket lock to
1745          * handle the fault. This might be observed in the PID check
1746          * in lookup_pi_state.
1747          */
1748 retry:
1749         if (get_futex_value_locked(&uval, uaddr))
1750                 goto handle_fault;
1751
1752         while (1) {
1753                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1754
1755                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1756                         goto handle_fault;
1757                 if (curval == uval)
1758                         break;
1759                 uval = curval;
1760         }
1761
1762         /*
1763          * We fixed up user space. Now we need to fix the pi_state
1764          * itself.
1765          */
1766         if (pi_state->owner != NULL) {
1767                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1768                 WARN_ON(list_empty(&pi_state->list));
1769                 list_del_init(&pi_state->list);
1770                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1771         }
1772
1773         pi_state->owner = newowner;
1774
1775         raw_spin_lock_irq(&newowner->pi_lock);
1776         WARN_ON(!list_empty(&pi_state->list));
1777         list_add(&pi_state->list, &newowner->pi_state_list);
1778         raw_spin_unlock_irq(&newowner->pi_lock);
1779         return 0;
1780
1781         /*
1782          * To handle the page fault we need to drop the hash bucket
1783          * lock here. That gives the other task (either the highest priority
1784          * waiter itself or the task which stole the rtmutex) the
1785          * chance to try the fixup of the pi_state. So once we are
1786          * back from handling the fault we need to check the pi_state
1787          * after reacquiring the hash bucket lock and before trying to
1788          * do another fixup. When the fixup has been done already we
1789          * simply return.
1790          */
1791 handle_fault:
1792         spin_unlock(q->lock_ptr);
1793
1794         ret = fault_in_user_writeable(uaddr);
1795
1796         spin_lock(q->lock_ptr);
1797
1798         /*
1799          * Check if someone else fixed it for us:
1800          */
1801         if (pi_state->owner != oldowner)
1802                 return 0;
1803
1804         if (ret)
1805                 return ret;
1806
1807         goto retry;
1808 }
1809
1810 static long futex_wait_restart(struct restart_block *restart);
1811
1812 /**
1813  * fixup_owner() - Post lock pi_state and corner case management
1814  * @uaddr:      user address of the futex
1815  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1816  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1817  *
1818  * After attempting to lock an rt_mutex, this function is called to cleanup
1819  * the pi_state owner as well as handle race conditions that may allow us to
1820  * acquire the lock. Must be called with the hb lock held.
1821  *
1822  * Return:
1823  *  1 - success, lock taken;
1824  *  0 - success, lock not taken;
1825  * <0 - on error (-EFAULT)
1826  */
1827 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1828 {
1829         struct task_struct *owner;
1830         int ret = 0;
1831
1832         if (locked) {
1833                 /*
1834                  * Got the lock. We might not be the anticipated owner if we
1835                  * did a lock-steal - fix up the PI-state in that case:
1836                  */
1837                 if (q->pi_state->owner != current)
1838                         ret = fixup_pi_state_owner(uaddr, q, current);
1839                 goto out;
1840         }
1841
1842         /*
1843          * Catch the rare case, where the lock was released when we were on the
1844          * way back before we locked the hash bucket.
1845          */
1846         if (q->pi_state->owner == current) {
1847                 /*
1848                  * Try to get the rt_mutex now. This might fail as some other
1849                  * task acquired the rt_mutex after we removed ourself from the
1850                  * rt_mutex waiters list.
1851                  */
1852                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1853                         locked = 1;
1854                         goto out;
1855                 }
1856
1857                 /*
1858                  * pi_state is incorrect, some other task did a lock steal and
1859                  * we returned due to timeout or signal without taking the
1860                  * rt_mutex. Too late.
1861                  */
1862                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1863                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1864                 if (!owner)
1865                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1866                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1867                 ret = fixup_pi_state_owner(uaddr, q, owner);
1868                 goto out;
1869         }
1870
1871         /*
1872          * Paranoia check. If we did not take the lock, then we should not be
1873          * the owner of the rt_mutex.
1874          */
1875         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1876                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1877                                 "pi-state %p\n", ret,
1878                                 q->pi_state->pi_mutex.owner,
1879                                 q->pi_state->owner);
1880
1881 out:
1882         return ret ? ret : locked;
1883 }
1884
1885 /**
1886  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1887  * @hb:         the futex hash bucket, must be locked by the caller
1888  * @q:          the futex_q to queue up on
1889  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1890  */
1891 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1892                                 struct hrtimer_sleeper *timeout)
1893 {
1894         /*
1895          * The task state is guaranteed to be set before another task can
1896          * wake it. set_current_state() is implemented using set_mb() and
1897          * queue_me() calls spin_unlock() upon completion, both serializing
1898          * access to the hash list and forcing another memory barrier.
1899          */
1900         set_current_state(TASK_INTERRUPTIBLE);
1901         queue_me(q, hb);
1902
1903         /* Arm the timer */
1904         if (timeout) {
1905                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1906                 if (!hrtimer_active(&timeout->timer))
1907                         timeout->task = NULL;
1908         }
1909
1910         /*
1911          * If we have been removed from the hash list, then another task
1912          * has tried to wake us, and we can skip the call to schedule().
1913          */
1914         if (likely(!plist_node_empty(&q->list))) {
1915                 /*
1916                  * If the timer has already expired, current will already be
1917                  * flagged for rescheduling. Only call schedule if there
1918                  * is no timeout, or if it has yet to expire.
1919                  */
1920                 if (!timeout || timeout->task)
1921                         freezable_schedule();
1922         }
1923         __set_current_state(TASK_RUNNING);
1924 }
1925
1926 /**
1927  * futex_wait_setup() - Prepare to wait on a futex
1928  * @uaddr:      the futex userspace address
1929  * @val:        the expected value
1930  * @flags:      futex flags (FLAGS_SHARED, etc.)
1931  * @q:          the associated futex_q
1932  * @hb:         storage for hash_bucket pointer to be returned to caller
1933  *
1934  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1935  * compare it with the expected value.  Handle atomic faults internally.
1936  * Return with the hb lock held and a q.key reference on success, and unlocked
1937  * with no q.key reference on failure.
1938  *
1939  * Return:
1940  *  0 - uaddr contains val and hb has been locked;
1941  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1942  */
1943 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1944                            struct futex_q *q, struct futex_hash_bucket **hb)
1945 {
1946         u32 uval;
1947         int ret;
1948
1949         /*
1950          * Access the page AFTER the hash-bucket is locked.
1951          * Order is important:
1952          *
1953          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1954          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1955          *
1956          * The basic logical guarantee of a futex is that it blocks ONLY
1957          * if cond(var) is known to be true at the time of blocking, for
1958          * any cond.  If we locked the hash-bucket after testing *uaddr, that
1959          * would open a race condition where we could block indefinitely with
1960          * cond(var) false, which would violate the guarantee.
1961          *
1962          * On the other hand, we insert q and release the hash-bucket only
1963          * after testing *uaddr.  This guarantees that futex_wait() will NOT
1964          * absorb a wakeup if *uaddr does not match the desired values
1965          * while the syscall executes.
1966          */
1967 retry:
1968         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1969         if (unlikely(ret != 0))
1970                 return ret;
1971
1972 retry_private:
1973         *hb = queue_lock(q);
1974
1975         ret = get_futex_value_locked(&uval, uaddr);
1976
1977         if (ret) {
1978                 queue_unlock(q, *hb);
1979
1980                 ret = get_user(uval, uaddr);
1981                 if (ret)
1982                         goto out;
1983
1984                 if (!(flags & FLAGS_SHARED))
1985                         goto retry_private;
1986
1987                 put_futex_key(&q->key);
1988                 goto retry;
1989         }
1990
1991         if (uval != val) {
1992                 queue_unlock(q, *hb);
1993                 ret = -EWOULDBLOCK;
1994         }
1995
1996 out:
1997         if (ret)
1998                 put_futex_key(&q->key);
1999         return ret;
2000 }
2001
2002 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2003                       ktime_t *abs_time, u32 bitset)
2004 {
2005         struct hrtimer_sleeper timeout, *to = NULL;
2006         struct restart_block *restart;
2007         struct futex_hash_bucket *hb;
2008         struct futex_q q = futex_q_init;
2009         int ret;
2010
2011         if (!bitset)
2012                 return -EINVAL;
2013         q.bitset = bitset;
2014
2015         if (abs_time) {
2016                 to = &timeout;
2017
2018                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2019                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2020                                       HRTIMER_MODE_ABS);
2021                 hrtimer_init_sleeper(to, current);
2022                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2023                                              current->timer_slack_ns);
2024         }
2025
2026 retry:
2027         /*
2028          * Prepare to wait on uaddr. On success, holds hb lock and increments
2029          * q.key refs.
2030          */
2031         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2032         if (ret)
2033                 goto out;
2034
2035         /* queue_me and wait for wakeup, timeout, or a signal. */
2036         futex_wait_queue_me(hb, &q, to);
2037
2038         /* If we were woken (and unqueued), we succeeded, whatever. */
2039         ret = 0;
2040         /* unqueue_me() drops q.key ref */
2041         if (!unqueue_me(&q))
2042                 goto out;
2043         ret = -ETIMEDOUT;
2044         if (to && !to->task)
2045                 goto out;
2046
2047         /*
2048          * We expect signal_pending(current), but we might be the
2049          * victim of a spurious wakeup as well.
2050          */
2051         if (!signal_pending(current))
2052                 goto retry;
2053
2054         ret = -ERESTARTSYS;
2055         if (!abs_time)
2056                 goto out;
2057
2058         restart = &current_thread_info()->restart_block;
2059         restart->fn = futex_wait_restart;
2060         restart->futex.uaddr = uaddr;
2061         restart->futex.val = val;
2062         restart->futex.time = abs_time->tv64;
2063         restart->futex.bitset = bitset;
2064         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2065
2066         ret = -ERESTART_RESTARTBLOCK;
2067
2068 out:
2069         if (to) {
2070                 hrtimer_cancel(&to->timer);
2071                 destroy_hrtimer_on_stack(&to->timer);
2072         }
2073         return ret;
2074 }
2075
2076
2077 static long futex_wait_restart(struct restart_block *restart)
2078 {
2079         u32 __user *uaddr = restart->futex.uaddr;
2080         ktime_t t, *tp = NULL;
2081
2082         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2083                 t.tv64 = restart->futex.time;
2084                 tp = &t;
2085         }
2086         restart->fn = do_no_restart_syscall;
2087
2088         return (long)futex_wait(uaddr, restart->futex.flags,
2089                                 restart->futex.val, tp, restart->futex.bitset);
2090 }
2091
2092
2093 /*
2094  * Userspace tried a 0 -> TID atomic transition of the futex value
2095  * and failed. The kernel side here does the whole locking operation:
2096  * if there are waiters then it will block, it does PI, etc. (Due to
2097  * races the kernel might see a 0 value of the futex too.)
2098  */
2099 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2100                          ktime_t *time, int trylock)
2101 {
2102         struct hrtimer_sleeper timeout, *to = NULL;
2103         struct futex_hash_bucket *hb;
2104         struct futex_q q = futex_q_init;
2105         int res, ret;
2106
2107         if (refill_pi_state_cache())
2108                 return -ENOMEM;
2109
2110         if (time) {
2111                 to = &timeout;
2112                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2113                                       HRTIMER_MODE_ABS);
2114                 hrtimer_init_sleeper(to, current);
2115                 hrtimer_set_expires(&to->timer, *time);
2116         }
2117
2118 retry:
2119         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2120         if (unlikely(ret != 0))
2121                 goto out;
2122
2123 retry_private:
2124         hb = queue_lock(&q);
2125
2126         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2127         if (unlikely(ret)) {
2128                 switch (ret) {
2129                 case 1:
2130                         /* We got the lock. */
2131                         ret = 0;
2132                         goto out_unlock_put_key;
2133                 case -EFAULT:
2134                         goto uaddr_faulted;
2135                 case -EAGAIN:
2136                         /*
2137                          * Task is exiting and we just wait for the
2138                          * exit to complete.
2139                          */
2140                         queue_unlock(&q, hb);
2141                         put_futex_key(&q.key);
2142                         cond_resched();
2143                         goto retry;
2144                 default:
2145                         goto out_unlock_put_key;
2146                 }
2147         }
2148
2149         /*
2150          * Only actually queue now that the atomic ops are done:
2151          */
2152         queue_me(&q, hb);
2153
2154         WARN_ON(!q.pi_state);
2155         /*
2156          * Block on the PI mutex:
2157          */
2158         if (!trylock)
2159                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2160         else {
2161                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2162                 /* Fixup the trylock return value: */
2163                 ret = ret ? 0 : -EWOULDBLOCK;
2164         }
2165
2166         spin_lock(q.lock_ptr);
2167         /*
2168          * Fixup the pi_state owner and possibly acquire the lock if we
2169          * haven't already.
2170          */
2171         res = fixup_owner(uaddr, &q, !ret);
2172         /*
2173          * If fixup_owner() returned an error, proprogate that.  If it acquired
2174          * the lock, clear our -ETIMEDOUT or -EINTR.
2175          */
2176         if (res)
2177                 ret = (res < 0) ? res : 0;
2178
2179         /*
2180          * If fixup_owner() faulted and was unable to handle the fault, unlock
2181          * it and return the fault to userspace.
2182          */
2183         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2184                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2185
2186         /* Unqueue and drop the lock */
2187         unqueue_me_pi(&q);
2188
2189         goto out_put_key;
2190
2191 out_unlock_put_key:
2192         queue_unlock(&q, hb);
2193
2194 out_put_key:
2195         put_futex_key(&q.key);
2196 out:
2197         if (to)
2198                 destroy_hrtimer_on_stack(&to->timer);
2199         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2200
2201 uaddr_faulted:
2202         queue_unlock(&q, hb);
2203
2204         ret = fault_in_user_writeable(uaddr);
2205         if (ret)
2206                 goto out_put_key;
2207
2208         if (!(flags & FLAGS_SHARED))
2209                 goto retry_private;
2210
2211         put_futex_key(&q.key);
2212         goto retry;
2213 }
2214
2215 /*
2216  * Userspace attempted a TID -> 0 atomic transition, and failed.
2217  * This is the in-kernel slowpath: we look up the PI state (if any),
2218  * and do the rt-mutex unlock.
2219  */
2220 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2221 {
2222         struct futex_hash_bucket *hb;
2223         struct futex_q *this, *next;
2224         struct plist_head *head;
2225         union futex_key key = FUTEX_KEY_INIT;
2226         u32 uval, vpid = task_pid_vnr(current);
2227         int ret;
2228
2229 retry:
2230         if (get_user(uval, uaddr))
2231                 return -EFAULT;
2232         /*
2233          * We release only a lock we actually own:
2234          */
2235         if ((uval & FUTEX_TID_MASK) != vpid)
2236                 return -EPERM;
2237
2238         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2239         if (unlikely(ret != 0))
2240                 goto out;
2241
2242         hb = hash_futex(&key);
2243         spin_lock(&hb->lock);
2244
2245         /*
2246          * To avoid races, try to do the TID -> 0 atomic transition
2247          * again. If it succeeds then we can return without waking
2248          * anyone else up. We only try this if neither the waiters nor
2249          * the owner died bit are set.
2250          */
2251         if (!(uval & ~FUTEX_TID_MASK) &&
2252             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2253                 goto pi_faulted;
2254         /*
2255          * Rare case: we managed to release the lock atomically,
2256          * no need to wake anyone else up:
2257          */
2258         if (unlikely(uval == vpid))
2259                 goto out_unlock;
2260
2261         /*
2262          * Ok, other tasks may need to be woken up - check waiters
2263          * and do the wakeup if necessary:
2264          */
2265         head = &hb->chain;
2266
2267         plist_for_each_entry_safe(this, next, head, list) {
2268                 if (!match_futex (&this->key, &key))
2269                         continue;
2270                 ret = wake_futex_pi(uaddr, uval, this);
2271                 /*
2272                  * The atomic access to the futex value
2273                  * generated a pagefault, so retry the
2274                  * user-access and the wakeup:
2275                  */
2276                 if (ret == -EFAULT)
2277                         goto pi_faulted;
2278                 goto out_unlock;
2279         }
2280         /*
2281          * No waiters - kernel unlocks the futex:
2282          */
2283         ret = unlock_futex_pi(uaddr, uval);
2284         if (ret == -EFAULT)
2285                 goto pi_faulted;
2286
2287 out_unlock:
2288         spin_unlock(&hb->lock);
2289         put_futex_key(&key);
2290
2291 out:
2292         return ret;
2293
2294 pi_faulted:
2295         spin_unlock(&hb->lock);
2296         put_futex_key(&key);
2297
2298         ret = fault_in_user_writeable(uaddr);
2299         if (!ret)
2300                 goto retry;
2301
2302         return ret;
2303 }
2304
2305 /**
2306  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2307  * @hb:         the hash_bucket futex_q was original enqueued on
2308  * @q:          the futex_q woken while waiting to be requeued
2309  * @key2:       the futex_key of the requeue target futex
2310  * @timeout:    the timeout associated with the wait (NULL if none)
2311  *
2312  * Detect if the task was woken on the initial futex as opposed to the requeue
2313  * target futex.  If so, determine if it was a timeout or a signal that caused
2314  * the wakeup and return the appropriate error code to the caller.  Must be
2315  * called with the hb lock held.
2316  *
2317  * Return:
2318  *  0 = no early wakeup detected;
2319  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2320  */
2321 static inline
2322 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2323                                    struct futex_q *q, union futex_key *key2,
2324                                    struct hrtimer_sleeper *timeout)
2325 {
2326         int ret = 0;
2327
2328         /*
2329          * With the hb lock held, we avoid races while we process the wakeup.
2330          * We only need to hold hb (and not hb2) to ensure atomicity as the
2331          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2332          * It can't be requeued from uaddr2 to something else since we don't
2333          * support a PI aware source futex for requeue.
2334          */
2335         if (!match_futex(&q->key, key2)) {
2336                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2337                 /*
2338                  * We were woken prior to requeue by a timeout or a signal.
2339                  * Unqueue the futex_q and determine which it was.
2340                  */
2341                 plist_del(&q->list, &hb->chain);
2342
2343                 /* Handle spurious wakeups gracefully */
2344                 ret = -EWOULDBLOCK;
2345                 if (timeout && !timeout->task)
2346                         ret = -ETIMEDOUT;
2347                 else if (signal_pending(current))
2348                         ret = -ERESTARTNOINTR;
2349         }
2350         return ret;
2351 }
2352
2353 /**
2354  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2355  * @uaddr:      the futex we initially wait on (non-pi)
2356  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2357  *              the same type, no requeueing from private to shared, etc.
2358  * @val:        the expected value of uaddr
2359  * @abs_time:   absolute timeout
2360  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2361  * @uaddr2:     the pi futex we will take prior to returning to user-space
2362  *
2363  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2364  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2365  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2366  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2367  * without one, the pi logic would not know which task to boost/deboost, if
2368  * there was a need to.
2369  *
2370  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2371  * via the following--
2372  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2373  * 2) wakeup on uaddr2 after a requeue
2374  * 3) signal
2375  * 4) timeout
2376  *
2377  * If 3, cleanup and return -ERESTARTNOINTR.
2378  *
2379  * If 2, we may then block on trying to take the rt_mutex and return via:
2380  * 5) successful lock
2381  * 6) signal
2382  * 7) timeout
2383  * 8) other lock acquisition failure
2384  *
2385  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2386  *
2387  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2388  *
2389  * Return:
2390  *  0 - On success;
2391  * <0 - On error
2392  */
2393 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2394                                  u32 val, ktime_t *abs_time, u32 bitset,
2395                                  u32 __user *uaddr2)
2396 {
2397         struct hrtimer_sleeper timeout, *to = NULL;
2398         struct rt_mutex_waiter rt_waiter;
2399         struct rt_mutex *pi_mutex = NULL;
2400         struct futex_hash_bucket *hb;
2401         union futex_key key2 = FUTEX_KEY_INIT;
2402         struct futex_q q = futex_q_init;
2403         int res, ret;
2404
2405         if (uaddr == uaddr2)
2406                 return -EINVAL;
2407
2408         if (!bitset)
2409                 return -EINVAL;
2410
2411         if (abs_time) {
2412                 to = &timeout;
2413                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2414                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2415                                       HRTIMER_MODE_ABS);
2416                 hrtimer_init_sleeper(to, current);
2417                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2418                                              current->timer_slack_ns);
2419         }
2420
2421         /*
2422          * The waiter is allocated on our stack, manipulated by the requeue
2423          * code while we sleep on uaddr.
2424          */
2425         debug_rt_mutex_init_waiter(&rt_waiter);
2426         rt_waiter.task = NULL;
2427
2428         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2429         if (unlikely(ret != 0))
2430                 goto out;
2431
2432         q.bitset = bitset;
2433         q.rt_waiter = &rt_waiter;
2434         q.requeue_pi_key = &key2;
2435
2436         /*
2437          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2438          * count.
2439          */
2440         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2441         if (ret)
2442                 goto out_key2;
2443
2444         /*
2445          * The check above which compares uaddrs is not sufficient for
2446          * shared futexes. We need to compare the keys:
2447          */
2448         if (match_futex(&q.key, &key2)) {
2449                 ret = -EINVAL;
2450                 goto out_put_keys;
2451         }
2452
2453         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2454         futex_wait_queue_me(hb, &q, to);
2455
2456         spin_lock(&hb->lock);
2457         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2458         spin_unlock(&hb->lock);
2459         if (ret)
2460                 goto out_put_keys;
2461
2462         /*
2463          * In order for us to be here, we know our q.key == key2, and since
2464          * we took the hb->lock above, we also know that futex_requeue() has
2465          * completed and we no longer have to concern ourselves with a wakeup
2466          * race with the atomic proxy lock acquisition by the requeue code. The
2467          * futex_requeue dropped our key1 reference and incremented our key2
2468          * reference count.
2469          */
2470
2471         /* Check if the requeue code acquired the second futex for us. */
2472         if (!q.rt_waiter) {
2473                 /*
2474                  * Got the lock. We might not be the anticipated owner if we
2475                  * did a lock-steal - fix up the PI-state in that case.
2476                  */
2477                 if (q.pi_state && (q.pi_state->owner != current)) {
2478                         spin_lock(q.lock_ptr);
2479                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2480                         spin_unlock(q.lock_ptr);
2481                 }
2482         } else {
2483                 /*
2484                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2485                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2486                  * the pi_state.
2487                  */
2488                 WARN_ON(!q.pi_state);
2489                 pi_mutex = &q.pi_state->pi_mutex;
2490                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2491                 debug_rt_mutex_free_waiter(&rt_waiter);
2492
2493                 spin_lock(q.lock_ptr);
2494                 /*
2495                  * Fixup the pi_state owner and possibly acquire the lock if we
2496                  * haven't already.
2497                  */
2498                 res = fixup_owner(uaddr2, &q, !ret);
2499                 /*
2500                  * If fixup_owner() returned an error, proprogate that.  If it
2501                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2502                  */
2503                 if (res)
2504                         ret = (res < 0) ? res : 0;
2505
2506                 /* Unqueue and drop the lock. */
2507                 unqueue_me_pi(&q);
2508         }
2509
2510         /*
2511          * If fixup_pi_state_owner() faulted and was unable to handle the
2512          * fault, unlock the rt_mutex and return the fault to userspace.
2513          */
2514         if (ret == -EFAULT) {
2515                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2516                         rt_mutex_unlock(pi_mutex);
2517         } else if (ret == -EINTR) {
2518                 /*
2519                  * We've already been requeued, but cannot restart by calling
2520                  * futex_lock_pi() directly. We could restart this syscall, but
2521                  * it would detect that the user space "val" changed and return
2522                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2523                  * -EWOULDBLOCK directly.
2524                  */
2525                 ret = -EWOULDBLOCK;
2526         }
2527
2528 out_put_keys:
2529         put_futex_key(&q.key);
2530 out_key2:
2531         put_futex_key(&key2);
2532
2533 out:
2534         if (to) {
2535                 hrtimer_cancel(&to->timer);
2536                 destroy_hrtimer_on_stack(&to->timer);
2537         }
2538         return ret;
2539 }
2540
2541 /*
2542  * Support for robust futexes: the kernel cleans up held futexes at
2543  * thread exit time.
2544  *
2545  * Implementation: user-space maintains a per-thread list of locks it
2546  * is holding. Upon do_exit(), the kernel carefully walks this list,
2547  * and marks all locks that are owned by this thread with the
2548  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2549  * always manipulated with the lock held, so the list is private and
2550  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2551  * field, to allow the kernel to clean up if the thread dies after
2552  * acquiring the lock, but just before it could have added itself to
2553  * the list. There can only be one such pending lock.
2554  */
2555
2556 /**
2557  * sys_set_robust_list() - Set the robust-futex list head of a task
2558  * @head:       pointer to the list-head
2559  * @len:        length of the list-head, as userspace expects
2560  */
2561 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2562                 size_t, len)
2563 {
2564         if (!futex_cmpxchg_enabled)
2565                 return -ENOSYS;
2566         /*
2567          * The kernel knows only one size for now:
2568          */
2569         if (unlikely(len != sizeof(*head)))
2570                 return -EINVAL;
2571
2572         current->robust_list = head;
2573
2574         return 0;
2575 }
2576
2577 /**
2578  * sys_get_robust_list() - Get the robust-futex list head of a task
2579  * @pid:        pid of the process [zero for current task]
2580  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2581  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2582  */
2583 SYSCALL_DEFINE3(get_robust_list, int, pid,
2584                 struct robust_list_head __user * __user *, head_ptr,
2585                 size_t __user *, len_ptr)
2586 {
2587         struct robust_list_head __user *head;
2588         unsigned long ret;
2589         struct task_struct *p;
2590
2591         if (!futex_cmpxchg_enabled)
2592                 return -ENOSYS;
2593
2594         rcu_read_lock();
2595
2596         ret = -ESRCH;
2597         if (!pid)
2598                 p = current;
2599         else {
2600                 p = find_task_by_vpid(pid);
2601                 if (!p)
2602                         goto err_unlock;
2603         }
2604
2605         ret = -EPERM;
2606         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2607                 goto err_unlock;
2608
2609         head = p->robust_list;
2610         rcu_read_unlock();
2611
2612         if (put_user(sizeof(*head), len_ptr))
2613                 return -EFAULT;
2614         return put_user(head, head_ptr);
2615
2616 err_unlock:
2617         rcu_read_unlock();
2618
2619         return ret;
2620 }
2621
2622 /*
2623  * Process a futex-list entry, check whether it's owned by the
2624  * dying task, and do notification if so:
2625  */
2626 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2627 {
2628         u32 uval, uninitialized_var(nval), mval;
2629
2630 retry:
2631         if (get_user(uval, uaddr))
2632                 return -1;
2633
2634         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2635                 /*
2636                  * Ok, this dying thread is truly holding a futex
2637                  * of interest. Set the OWNER_DIED bit atomically
2638                  * via cmpxchg, and if the value had FUTEX_WAITERS
2639                  * set, wake up a waiter (if any). (We have to do a
2640                  * futex_wake() even if OWNER_DIED is already set -
2641                  * to handle the rare but possible case of recursive
2642                  * thread-death.) The rest of the cleanup is done in
2643                  * userspace.
2644                  */
2645                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2646                 /*
2647                  * We are not holding a lock here, but we want to have
2648                  * the pagefault_disable/enable() protection because
2649                  * we want to handle the fault gracefully. If the
2650                  * access fails we try to fault in the futex with R/W
2651                  * verification via get_user_pages. get_user() above
2652                  * does not guarantee R/W access. If that fails we
2653                  * give up and leave the futex locked.
2654                  */
2655                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2656                         if (fault_in_user_writeable(uaddr))
2657                                 return -1;
2658                         goto retry;
2659                 }
2660                 if (nval != uval)
2661                         goto retry;
2662
2663                 /*
2664                  * Wake robust non-PI futexes here. The wakeup of
2665                  * PI futexes happens in exit_pi_state():
2666                  */
2667                 if (!pi && (uval & FUTEX_WAITERS))
2668                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2669         }
2670         return 0;
2671 }
2672
2673 /*
2674  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2675  */
2676 static inline int fetch_robust_entry(struct robust_list __user **entry,
2677                                      struct robust_list __user * __user *head,
2678                                      unsigned int *pi)
2679 {
2680         unsigned long uentry;
2681
2682         if (get_user(uentry, (unsigned long __user *)head))
2683                 return -EFAULT;
2684
2685         *entry = (void __user *)(uentry & ~1UL);
2686         *pi = uentry & 1;
2687
2688         return 0;
2689 }
2690
2691 /*
2692  * Walk curr->robust_list (very carefully, it's a userspace list!)
2693  * and mark any locks found there dead, and notify any waiters.
2694  *
2695  * We silently return on any sign of list-walking problem.
2696  */
2697 void exit_robust_list(struct task_struct *curr)
2698 {
2699         struct robust_list_head __user *head = curr->robust_list;
2700         struct robust_list __user *entry, *next_entry, *pending;
2701         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2702         unsigned int uninitialized_var(next_pi);
2703         unsigned long futex_offset;
2704         int rc;
2705
2706         if (!futex_cmpxchg_enabled)
2707                 return;
2708
2709         /*
2710          * Fetch the list head (which was registered earlier, via
2711          * sys_set_robust_list()):
2712          */
2713         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2714                 return;
2715         /*
2716          * Fetch the relative futex offset:
2717          */
2718         if (get_user(futex_offset, &head->futex_offset))
2719                 return;
2720         /*
2721          * Fetch any possibly pending lock-add first, and handle it
2722          * if it exists:
2723          */
2724         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2725                 return;
2726
2727         next_entry = NULL;      /* avoid warning with gcc */
2728         while (entry != &head->list) {
2729                 /*
2730                  * Fetch the next entry in the list before calling
2731                  * handle_futex_death:
2732                  */
2733                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2734                 /*
2735                  * A pending lock might already be on the list, so
2736                  * don't process it twice:
2737                  */
2738                 if (entry != pending)
2739                         if (handle_futex_death((void __user *)entry + futex_offset,
2740                                                 curr, pi))
2741                                 return;
2742                 if (rc)
2743                         return;
2744                 entry = next_entry;
2745                 pi = next_pi;
2746                 /*
2747                  * Avoid excessively long or circular lists:
2748                  */
2749                 if (!--limit)
2750                         break;
2751
2752                 cond_resched();
2753         }
2754
2755         if (pending)
2756                 handle_futex_death((void __user *)pending + futex_offset,
2757                                    curr, pip);
2758 }
2759
2760 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2761                 u32 __user *uaddr2, u32 val2, u32 val3)
2762 {
2763         int cmd = op & FUTEX_CMD_MASK;
2764         unsigned int flags = 0;
2765
2766         if (!(op & FUTEX_PRIVATE_FLAG))
2767                 flags |= FLAGS_SHARED;
2768
2769         if (op & FUTEX_CLOCK_REALTIME) {
2770                 flags |= FLAGS_CLOCKRT;
2771                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2772                         return -ENOSYS;
2773         }
2774
2775         switch (cmd) {
2776         case FUTEX_LOCK_PI:
2777         case FUTEX_UNLOCK_PI:
2778         case FUTEX_TRYLOCK_PI:
2779         case FUTEX_WAIT_REQUEUE_PI:
2780         case FUTEX_CMP_REQUEUE_PI:
2781                 if (!futex_cmpxchg_enabled)
2782                         return -ENOSYS;
2783         }
2784
2785         switch (cmd) {
2786         case FUTEX_WAIT:
2787                 val3 = FUTEX_BITSET_MATCH_ANY;
2788         case FUTEX_WAIT_BITSET:
2789                 return futex_wait(uaddr, flags, val, timeout, val3);
2790         case FUTEX_WAKE:
2791                 val3 = FUTEX_BITSET_MATCH_ANY;
2792         case FUTEX_WAKE_BITSET:
2793                 return futex_wake(uaddr, flags, val, val3);
2794         case FUTEX_REQUEUE:
2795                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2796         case FUTEX_CMP_REQUEUE:
2797                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2798         case FUTEX_WAKE_OP:
2799                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2800         case FUTEX_LOCK_PI:
2801                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2802         case FUTEX_UNLOCK_PI:
2803                 return futex_unlock_pi(uaddr, flags);
2804         case FUTEX_TRYLOCK_PI:
2805                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2806         case FUTEX_WAIT_REQUEUE_PI:
2807                 val3 = FUTEX_BITSET_MATCH_ANY;
2808                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2809                                              uaddr2);
2810         case FUTEX_CMP_REQUEUE_PI:
2811                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2812         }
2813         return -ENOSYS;
2814 }
2815
2816
2817 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2818                 struct timespec __user *, utime, u32 __user *, uaddr2,
2819                 u32, val3)
2820 {
2821         struct timespec ts;
2822         ktime_t t, *tp = NULL;
2823         u32 val2 = 0;
2824         int cmd = op & FUTEX_CMD_MASK;
2825
2826         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2827                       cmd == FUTEX_WAIT_BITSET ||
2828                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2829                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2830                         return -EFAULT;
2831                 if (!timespec_valid(&ts))
2832                         return -EINVAL;
2833
2834                 t = timespec_to_ktime(ts);
2835                 if (cmd == FUTEX_WAIT)
2836                         t = ktime_add_safe(ktime_get(), t);
2837                 tp = &t;
2838         }
2839         /*
2840          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2841          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2842          */
2843         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2844             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2845                 val2 = (u32) (unsigned long) utime;
2846
2847         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2848 }
2849
2850 static int __init futex_init(void)
2851 {
2852         u32 curval;
2853         int i;
2854
2855         /*
2856          * This will fail and we want it. Some arch implementations do
2857          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2858          * functionality. We want to know that before we call in any
2859          * of the complex code paths. Also we want to prevent
2860          * registration of robust lists in that case. NULL is
2861          * guaranteed to fault and we get -EFAULT on functional
2862          * implementation, the non-functional ones will return
2863          * -ENOSYS.
2864          */
2865         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2866                 futex_cmpxchg_enabled = 1;
2867
2868         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2869                 plist_head_init(&futex_queues[i].chain);
2870                 spin_lock_init(&futex_queues[i].lock);
2871         }
2872
2873         return 0;
2874 }
2875 __initcall(futex_init);