Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / 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/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
67
68 #include <asm/futex.h>
69
70 #include "locking/rtmutex_common.h"
71
72 /*
73  * READ this before attempting to hack on futexes!
74  *
75  * Basic futex operation and ordering guarantees
76  * =============================================
77  *
78  * The waiter reads the futex value in user space and calls
79  * futex_wait(). This function computes the hash bucket and acquires
80  * the hash bucket lock. After that it reads the futex user space value
81  * again and verifies that the data has not changed. If it has not changed
82  * it enqueues itself into the hash bucket, releases the hash bucket lock
83  * and schedules.
84  *
85  * The waker side modifies the user space value of the futex and calls
86  * futex_wake(). This function computes the hash bucket and acquires the
87  * hash bucket lock. Then it looks for waiters on that futex in the hash
88  * bucket and wakes them.
89  *
90  * In futex wake up scenarios where no tasks are blocked on a futex, taking
91  * the hb spinlock can be avoided and simply return. In order for this
92  * optimization to work, ordering guarantees must exist so that the waiter
93  * being added to the list is acknowledged when the list is concurrently being
94  * checked by the waker, avoiding scenarios like the following:
95  *
96  * CPU 0                               CPU 1
97  * val = *futex;
98  * sys_futex(WAIT, futex, val);
99  *   futex_wait(futex, val);
100  *   uval = *futex;
101  *                                     *futex = newval;
102  *                                     sys_futex(WAKE, futex);
103  *                                       futex_wake(futex);
104  *                                       if (queue_empty())
105  *                                         return;
106  *   if (uval == val)
107  *      lock(hash_bucket(futex));
108  *      queue();
109  *     unlock(hash_bucket(futex));
110  *     schedule();
111  *
112  * This would cause the waiter on CPU 0 to wait forever because it
113  * missed the transition of the user space value from val to newval
114  * and the waker did not find the waiter in the hash bucket queue.
115  *
116  * The correct serialization ensures that a waiter either observes
117  * the changed user space value before blocking or is woken by a
118  * concurrent waker:
119  *
120  * CPU 0                                 CPU 1
121  * val = *futex;
122  * sys_futex(WAIT, futex, val);
123  *   futex_wait(futex, val);
124  *
125  *   waiters++; (a)
126  *   mb(); (A) <-- paired with -.
127  *                              |
128  *   lock(hash_bucket(futex));  |
129  *                              |
130  *   uval = *futex;             |
131  *                              |        *futex = newval;
132  *                              |        sys_futex(WAKE, futex);
133  *                              |          futex_wake(futex);
134  *                              |
135  *                              `------->  mb(); (B)
136  *   if (uval == val)
137  *     queue();
138  *     unlock(hash_bucket(futex));
139  *     schedule();                         if (waiters)
140  *                                           lock(hash_bucket(futex));
141  *   else                                    wake_waiters(futex);
142  *     waiters--; (b)                        unlock(hash_bucket(futex));
143  *
144  * Where (A) orders the waiters increment and the futex value read through
145  * atomic operations (see hb_waiters_inc) and where (B) orders the write
146  * to futex and the waiters read -- this is done by the barriers for both
147  * shared and private futexes in get_futex_key_refs().
148  *
149  * This yields the following case (where X:=waiters, Y:=futex):
150  *
151  *      X = Y = 0
152  *
153  *      w[X]=1          w[Y]=1
154  *      MB              MB
155  *      r[Y]=y          r[X]=x
156  *
157  * Which guarantees that x==0 && y==0 is impossible; which translates back into
158  * the guarantee that we cannot both miss the futex variable change and the
159  * enqueue.
160  *
161  * Note that a new waiter is accounted for in (a) even when it is possible that
162  * the wait call can return error, in which case we backtrack from it in (b).
163  * Refer to the comment in queue_lock().
164  *
165  * Similarly, in order to account for waiters being requeued on another
166  * address we always increment the waiters for the destination bucket before
167  * acquiring the lock. It then decrements them again  after releasing it -
168  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
169  * will do the additional required waiter count housekeeping. This is done for
170  * double_lock_hb() and double_unlock_hb(), respectively.
171  */
172
173 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
174 int __read_mostly futex_cmpxchg_enabled;
175 #endif
176
177 /*
178  * Futex flags used to encode options to functions and preserve them across
179  * restarts.
180  */
181 #define FLAGS_SHARED            0x01
182 #define FLAGS_CLOCKRT           0x02
183 #define FLAGS_HAS_TIMEOUT       0x04
184
185 /*
186  * Priority Inheritance state:
187  */
188 struct futex_pi_state {
189         /*
190          * list of 'owned' pi_state instances - these have to be
191          * cleaned up in do_exit() if the task exits prematurely:
192          */
193         struct list_head list;
194
195         /*
196          * The PI object:
197          */
198         struct rt_mutex pi_mutex;
199
200         struct task_struct *owner;
201         atomic_t refcount;
202
203         union futex_key key;
204 };
205
206 /**
207  * struct futex_q - The hashed futex queue entry, one per waiting task
208  * @list:               priority-sorted list of tasks waiting on this futex
209  * @task:               the task waiting on the futex
210  * @lock_ptr:           the hash bucket lock
211  * @key:                the key the futex is hashed on
212  * @pi_state:           optional priority inheritance state
213  * @rt_waiter:          rt_waiter storage for use with requeue_pi
214  * @requeue_pi_key:     the requeue_pi target futex key
215  * @bitset:             bitset for the optional bitmasked wakeup
216  *
217  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
218  * we can wake only the relevant ones (hashed queues may be shared).
219  *
220  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
221  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
222  * The order of wakeup is always to make the first condition true, then
223  * the second.
224  *
225  * PI futexes are typically woken before they are removed from the hash list via
226  * the rt_mutex code. See unqueue_me_pi().
227  */
228 struct futex_q {
229         struct plist_node list;
230
231         struct task_struct *task;
232         spinlock_t *lock_ptr;
233         union futex_key key;
234         struct futex_pi_state *pi_state;
235         struct rt_mutex_waiter *rt_waiter;
236         union futex_key *requeue_pi_key;
237         u32 bitset;
238 };
239
240 static const struct futex_q futex_q_init = {
241         /* list gets initialized in queue_me()*/
242         .key = FUTEX_KEY_INIT,
243         .bitset = FUTEX_BITSET_MATCH_ANY
244 };
245
246 /*
247  * Hash buckets are shared by all the futex_keys that hash to the same
248  * location.  Each key may have multiple futex_q structures, one for each task
249  * waiting on a futex.
250  */
251 struct futex_hash_bucket {
252         atomic_t waiters;
253         spinlock_t lock;
254         struct plist_head chain;
255 } ____cacheline_aligned_in_smp;
256
257 static unsigned long __read_mostly futex_hashsize;
258
259 static struct futex_hash_bucket *futex_queues;
260
261 static inline void futex_get_mm(union futex_key *key)
262 {
263         atomic_inc(&key->private.mm->mm_count);
264         /*
265          * Ensure futex_get_mm() implies a full barrier such that
266          * get_futex_key() implies a full barrier. This is relied upon
267          * as full barrier (B), see the ordering comment above.
268          */
269         smp_mb__after_atomic();
270 }
271
272 /*
273  * Reflects a new waiter being added to the waitqueue.
274  */
275 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
276 {
277 #ifdef CONFIG_SMP
278         atomic_inc(&hb->waiters);
279         /*
280          * Full barrier (A), see the ordering comment above.
281          */
282         smp_mb__after_atomic();
283 #endif
284 }
285
286 /*
287  * Reflects a waiter being removed from the waitqueue by wakeup
288  * paths.
289  */
290 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
291 {
292 #ifdef CONFIG_SMP
293         atomic_dec(&hb->waiters);
294 #endif
295 }
296
297 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
298 {
299 #ifdef CONFIG_SMP
300         return atomic_read(&hb->waiters);
301 #else
302         return 1;
303 #endif
304 }
305
306 /*
307  * We hash on the keys returned from get_futex_key (see below).
308  */
309 static struct futex_hash_bucket *hash_futex(union futex_key *key)
310 {
311         u32 hash = jhash2((u32*)&key->both.word,
312                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
313                           key->both.offset);
314         return &futex_queues[hash & (futex_hashsize - 1)];
315 }
316
317 /*
318  * Return 1 if two futex_keys are equal, 0 otherwise.
319  */
320 static inline int match_futex(union futex_key *key1, union futex_key *key2)
321 {
322         return (key1 && key2
323                 && key1->both.word == key2->both.word
324                 && key1->both.ptr == key2->both.ptr
325                 && key1->both.offset == key2->both.offset);
326 }
327
328 /*
329  * Take a reference to the resource addressed by a key.
330  * Can be called while holding spinlocks.
331  *
332  */
333 static void get_futex_key_refs(union futex_key *key)
334 {
335         if (!key->both.ptr)
336                 return;
337
338         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
339         case FUT_OFF_INODE:
340                 ihold(key->shared.inode); /* implies MB (B) */
341                 break;
342         case FUT_OFF_MMSHARED:
343                 futex_get_mm(key); /* implies MB (B) */
344                 break;
345         default:
346                 /*
347                  * Private futexes do not hold reference on an inode or
348                  * mm, therefore the only purpose of calling get_futex_key_refs
349                  * is because we need the barrier for the lockless waiter check.
350                  */
351                 smp_mb(); /* explicit MB (B) */
352         }
353 }
354
355 /*
356  * Drop a reference to the resource addressed by a key.
357  * The hash bucket spinlock must not be held. This is
358  * a no-op for private futexes, see comment in the get
359  * counterpart.
360  */
361 static void drop_futex_key_refs(union futex_key *key)
362 {
363         if (!key->both.ptr) {
364                 /* If we're here then we tried to put a key we failed to get */
365                 WARN_ON_ONCE(1);
366                 return;
367         }
368
369         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
370         case FUT_OFF_INODE:
371                 iput(key->shared.inode);
372                 break;
373         case FUT_OFF_MMSHARED:
374                 mmdrop(key->private.mm);
375                 break;
376         }
377 }
378
379 /**
380  * get_futex_key() - Get parameters which are the keys for a futex
381  * @uaddr:      virtual address of the futex
382  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
383  * @key:        address where result is stored.
384  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
385  *              VERIFY_WRITE)
386  *
387  * Return: a negative error code or 0
388  *
389  * The key words are stored in *key on success.
390  *
391  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
392  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
393  * We can usually work out the index without swapping in the page.
394  *
395  * lock_page() might sleep, the caller should not hold a spinlock.
396  */
397 static int
398 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
399 {
400         unsigned long address = (unsigned long)uaddr;
401         struct mm_struct *mm = current->mm;
402         struct page *page, *page_head;
403         int err, ro = 0;
404
405         /*
406          * The futex address must be "naturally" aligned.
407          */
408         key->both.offset = address % PAGE_SIZE;
409         if (unlikely((address % sizeof(u32)) != 0))
410                 return -EINVAL;
411         address -= key->both.offset;
412
413         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
414                 return -EFAULT;
415
416         /*
417          * PROCESS_PRIVATE futexes are fast.
418          * As the mm cannot disappear under us and the 'key' only needs
419          * virtual address, we dont even have to find the underlying vma.
420          * Note : We do have to check 'uaddr' is a valid user address,
421          *        but access_ok() should be faster than find_vma()
422          */
423         if (!fshared) {
424                 key->private.mm = mm;
425                 key->private.address = address;
426                 get_futex_key_refs(key);  /* implies MB (B) */
427                 return 0;
428         }
429
430 again:
431         err = get_user_pages_fast(address, 1, 1, &page);
432         /*
433          * If write access is not required (eg. FUTEX_WAIT), try
434          * and get read-only access.
435          */
436         if (err == -EFAULT && rw == VERIFY_READ) {
437                 err = get_user_pages_fast(address, 1, 0, &page);
438                 ro = 1;
439         }
440         if (err < 0)
441                 return err;
442         else
443                 err = 0;
444
445 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
446         page_head = page;
447         if (unlikely(PageTail(page))) {
448                 put_page(page);
449                 /* serialize against __split_huge_page_splitting() */
450                 local_irq_disable();
451                 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
452                         page_head = compound_head(page);
453                         /*
454                          * page_head is valid pointer but we must pin
455                          * it before taking the PG_lock and/or
456                          * PG_compound_lock. The moment we re-enable
457                          * irqs __split_huge_page_splitting() can
458                          * return and the head page can be freed from
459                          * under us. We can't take the PG_lock and/or
460                          * PG_compound_lock on a page that could be
461                          * freed from under us.
462                          */
463                         if (page != page_head) {
464                                 get_page(page_head);
465                                 put_page(page);
466                         }
467                         local_irq_enable();
468                 } else {
469                         local_irq_enable();
470                         goto again;
471                 }
472         }
473 #else
474         page_head = compound_head(page);
475         if (page != page_head) {
476                 get_page(page_head);
477                 put_page(page);
478         }
479 #endif
480
481         lock_page(page_head);
482
483         /*
484          * If page_head->mapping is NULL, then it cannot be a PageAnon
485          * page; but it might be the ZERO_PAGE or in the gate area or
486          * in a special mapping (all cases which we are happy to fail);
487          * or it may have been a good file page when get_user_pages_fast
488          * found it, but truncated or holepunched or subjected to
489          * invalidate_complete_page2 before we got the page lock (also
490          * cases which we are happy to fail).  And we hold a reference,
491          * so refcount care in invalidate_complete_page's remove_mapping
492          * prevents drop_caches from setting mapping to NULL beneath us.
493          *
494          * The case we do have to guard against is when memory pressure made
495          * shmem_writepage move it from filecache to swapcache beneath us:
496          * an unlikely race, but we do need to retry for page_head->mapping.
497          */
498         if (!page_head->mapping) {
499                 int shmem_swizzled = PageSwapCache(page_head);
500                 unlock_page(page_head);
501                 put_page(page_head);
502                 if (shmem_swizzled)
503                         goto again;
504                 return -EFAULT;
505         }
506
507         /*
508          * Private mappings are handled in a simple way.
509          *
510          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
511          * it's a read-only handle, it's expected that futexes attach to
512          * the object not the particular process.
513          */
514         if (PageAnon(page_head)) {
515                 /*
516                  * A RO anonymous page will never change and thus doesn't make
517                  * sense for futex operations.
518                  */
519                 if (ro) {
520                         err = -EFAULT;
521                         goto out;
522                 }
523
524                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
525                 key->private.mm = mm;
526                 key->private.address = address;
527         } else {
528                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
529                 key->shared.inode = page_head->mapping->host;
530                 key->shared.pgoff = basepage_index(page);
531         }
532
533         get_futex_key_refs(key); /* implies MB (B) */
534
535 out:
536         unlock_page(page_head);
537         put_page(page_head);
538         return err;
539 }
540
541 static inline void put_futex_key(union futex_key *key)
542 {
543         drop_futex_key_refs(key);
544 }
545
546 /**
547  * fault_in_user_writeable() - Fault in user address and verify RW access
548  * @uaddr:      pointer to faulting user space address
549  *
550  * Slow path to fixup the fault we just took in the atomic write
551  * access to @uaddr.
552  *
553  * We have no generic implementation of a non-destructive write to the
554  * user address. We know that we faulted in the atomic pagefault
555  * disabled section so we can as well avoid the #PF overhead by
556  * calling get_user_pages() right away.
557  */
558 static int fault_in_user_writeable(u32 __user *uaddr)
559 {
560         struct mm_struct *mm = current->mm;
561         int ret;
562
563         down_read(&mm->mmap_sem);
564         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
565                                FAULT_FLAG_WRITE);
566         up_read(&mm->mmap_sem);
567
568         return ret < 0 ? ret : 0;
569 }
570
571 /**
572  * futex_top_waiter() - Return the highest priority waiter on a futex
573  * @hb:         the hash bucket the futex_q's reside in
574  * @key:        the futex key (to distinguish it from other futex futex_q's)
575  *
576  * Must be called with the hb lock held.
577  */
578 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
579                                         union futex_key *key)
580 {
581         struct futex_q *this;
582
583         plist_for_each_entry(this, &hb->chain, list) {
584                 if (match_futex(&this->key, key))
585                         return this;
586         }
587         return NULL;
588 }
589
590 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
591                                       u32 uval, u32 newval)
592 {
593         int ret;
594
595         pagefault_disable();
596         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
597         pagefault_enable();
598
599         return ret;
600 }
601
602 static int get_futex_value_locked(u32 *dest, u32 __user *from)
603 {
604         int ret;
605
606         pagefault_disable();
607         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
608         pagefault_enable();
609
610         return ret ? -EFAULT : 0;
611 }
612
613
614 /*
615  * PI code:
616  */
617 static int refill_pi_state_cache(void)
618 {
619         struct futex_pi_state *pi_state;
620
621         if (likely(current->pi_state_cache))
622                 return 0;
623
624         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
625
626         if (!pi_state)
627                 return -ENOMEM;
628
629         INIT_LIST_HEAD(&pi_state->list);
630         /* pi_mutex gets initialized later */
631         pi_state->owner = NULL;
632         atomic_set(&pi_state->refcount, 1);
633         pi_state->key = FUTEX_KEY_INIT;
634
635         current->pi_state_cache = pi_state;
636
637         return 0;
638 }
639
640 static struct futex_pi_state * alloc_pi_state(void)
641 {
642         struct futex_pi_state *pi_state = current->pi_state_cache;
643
644         WARN_ON(!pi_state);
645         current->pi_state_cache = NULL;
646
647         return pi_state;
648 }
649
650 /*
651  * Must be called with the hb lock held.
652  */
653 static void free_pi_state(struct futex_pi_state *pi_state)
654 {
655         if (!pi_state)
656                 return;
657
658         if (!atomic_dec_and_test(&pi_state->refcount))
659                 return;
660
661         /*
662          * If pi_state->owner is NULL, the owner is most probably dying
663          * and has cleaned up the pi_state already
664          */
665         if (pi_state->owner) {
666                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
667                 list_del_init(&pi_state->list);
668                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
669
670                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
671         }
672
673         if (current->pi_state_cache)
674                 kfree(pi_state);
675         else {
676                 /*
677                  * pi_state->list is already empty.
678                  * clear pi_state->owner.
679                  * refcount is at 0 - put it back to 1.
680                  */
681                 pi_state->owner = NULL;
682                 atomic_set(&pi_state->refcount, 1);
683                 current->pi_state_cache = pi_state;
684         }
685 }
686
687 /*
688  * Look up the task based on what TID userspace gave us.
689  * We dont trust it.
690  */
691 static struct task_struct * futex_find_get_task(pid_t pid)
692 {
693         struct task_struct *p;
694
695         rcu_read_lock();
696         p = find_task_by_vpid(pid);
697         if (p)
698                 get_task_struct(p);
699
700         rcu_read_unlock();
701
702         return p;
703 }
704
705 /*
706  * This task is holding PI mutexes at exit time => bad.
707  * Kernel cleans up PI-state, but userspace is likely hosed.
708  * (Robust-futex cleanup is separate and might save the day for userspace.)
709  */
710 void exit_pi_state_list(struct task_struct *curr)
711 {
712         struct list_head *next, *head = &curr->pi_state_list;
713         struct futex_pi_state *pi_state;
714         struct futex_hash_bucket *hb;
715         union futex_key key = FUTEX_KEY_INIT;
716
717         if (!futex_cmpxchg_enabled)
718                 return;
719         /*
720          * We are a ZOMBIE and nobody can enqueue itself on
721          * pi_state_list anymore, but we have to be careful
722          * versus waiters unqueueing themselves:
723          */
724         raw_spin_lock_irq(&curr->pi_lock);
725         while (!list_empty(head)) {
726
727                 next = head->next;
728                 pi_state = list_entry(next, struct futex_pi_state, list);
729                 key = pi_state->key;
730                 hb = hash_futex(&key);
731                 raw_spin_unlock_irq(&curr->pi_lock);
732
733                 spin_lock(&hb->lock);
734
735                 raw_spin_lock_irq(&curr->pi_lock);
736                 /*
737                  * We dropped the pi-lock, so re-check whether this
738                  * task still owns the PI-state:
739                  */
740                 if (head->next != next) {
741                         raw_spin_unlock_irq(&curr->pi_lock);
742                         spin_unlock(&hb->lock);
743                         raw_spin_lock_irq(&curr->pi_lock);
744                         continue;
745                 }
746
747                 WARN_ON(pi_state->owner != curr);
748                 WARN_ON(list_empty(&pi_state->list));
749                 list_del_init(&pi_state->list);
750                 pi_state->owner = NULL;
751                 raw_spin_unlock_irq(&curr->pi_lock);
752
753                 rt_mutex_unlock(&pi_state->pi_mutex);
754
755                 spin_unlock(&hb->lock);
756
757                 raw_spin_lock_irq(&curr->pi_lock);
758         }
759         raw_spin_unlock_irq(&curr->pi_lock);
760 }
761
762 /*
763  * We need to check the following states:
764  *
765  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
766  *
767  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
768  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
769  *
770  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
771  *
772  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
773  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
774  *
775  * [6]  Found  | Found    | task      | 0         | 1      | Valid
776  *
777  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
778  *
779  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
780  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
781  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
782  *
783  * [1]  Indicates that the kernel can acquire the futex atomically. We
784  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
785  *
786  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
787  *      thread is found then it indicates that the owner TID has died.
788  *
789  * [3]  Invalid. The waiter is queued on a non PI futex
790  *
791  * [4]  Valid state after exit_robust_list(), which sets the user space
792  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
793  *
794  * [5]  The user space value got manipulated between exit_robust_list()
795  *      and exit_pi_state_list()
796  *
797  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
798  *      the pi_state but cannot access the user space value.
799  *
800  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
801  *
802  * [8]  Owner and user space value match
803  *
804  * [9]  There is no transient state which sets the user space TID to 0
805  *      except exit_robust_list(), but this is indicated by the
806  *      FUTEX_OWNER_DIED bit. See [4]
807  *
808  * [10] There is no transient state which leaves owner and user space
809  *      TID out of sync.
810  */
811
812 /*
813  * Validate that the existing waiter has a pi_state and sanity check
814  * the pi_state against the user space value. If correct, attach to
815  * it.
816  */
817 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
818                               struct futex_pi_state **ps)
819 {
820         pid_t pid = uval & FUTEX_TID_MASK;
821
822         /*
823          * Userspace might have messed up non-PI and PI futexes [3]
824          */
825         if (unlikely(!pi_state))
826                 return -EINVAL;
827
828         WARN_ON(!atomic_read(&pi_state->refcount));
829
830         /*
831          * Handle the owner died case:
832          */
833         if (uval & FUTEX_OWNER_DIED) {
834                 /*
835                  * exit_pi_state_list sets owner to NULL and wakes the
836                  * topmost waiter. The task which acquires the
837                  * pi_state->rt_mutex will fixup owner.
838                  */
839                 if (!pi_state->owner) {
840                         /*
841                          * No pi state owner, but the user space TID
842                          * is not 0. Inconsistent state. [5]
843                          */
844                         if (pid)
845                                 return -EINVAL;
846                         /*
847                          * Take a ref on the state and return success. [4]
848                          */
849                         goto out_state;
850                 }
851
852                 /*
853                  * If TID is 0, then either the dying owner has not
854                  * yet executed exit_pi_state_list() or some waiter
855                  * acquired the rtmutex in the pi state, but did not
856                  * yet fixup the TID in user space.
857                  *
858                  * Take a ref on the state and return success. [6]
859                  */
860                 if (!pid)
861                         goto out_state;
862         } else {
863                 /*
864                  * If the owner died bit is not set, then the pi_state
865                  * must have an owner. [7]
866                  */
867                 if (!pi_state->owner)
868                         return -EINVAL;
869         }
870
871         /*
872          * Bail out if user space manipulated the futex value. If pi
873          * state exists then the owner TID must be the same as the
874          * user space TID. [9/10]
875          */
876         if (pid != task_pid_vnr(pi_state->owner))
877                 return -EINVAL;
878 out_state:
879         atomic_inc(&pi_state->refcount);
880         *ps = pi_state;
881         return 0;
882 }
883
884 /*
885  * Lookup the task for the TID provided from user space and attach to
886  * it after doing proper sanity checks.
887  */
888 static int attach_to_pi_owner(u32 uval, union futex_key *key,
889                               struct futex_pi_state **ps)
890 {
891         pid_t pid = uval & FUTEX_TID_MASK;
892         struct futex_pi_state *pi_state;
893         struct task_struct *p;
894
895         /*
896          * We are the first waiter - try to look up the real owner and attach
897          * the new pi_state to it, but bail out when TID = 0 [1]
898          */
899         if (!pid)
900                 return -ESRCH;
901         p = futex_find_get_task(pid);
902         if (!p)
903                 return -ESRCH;
904
905         if (unlikely(p->flags & PF_KTHREAD)) {
906                 put_task_struct(p);
907                 return -EPERM;
908         }
909
910         /*
911          * We need to look at the task state flags to figure out,
912          * whether the task is exiting. To protect against the do_exit
913          * change of the task flags, we do this protected by
914          * p->pi_lock:
915          */
916         raw_spin_lock_irq(&p->pi_lock);
917         if (unlikely(p->flags & PF_EXITING)) {
918                 /*
919                  * The task is on the way out. When PF_EXITPIDONE is
920                  * set, we know that the task has finished the
921                  * cleanup:
922                  */
923                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
924
925                 raw_spin_unlock_irq(&p->pi_lock);
926                 put_task_struct(p);
927                 return ret;
928         }
929
930         /*
931          * No existing pi state. First waiter. [2]
932          */
933         pi_state = alloc_pi_state();
934
935         /*
936          * Initialize the pi_mutex in locked state and make @p
937          * the owner of it:
938          */
939         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
940
941         /* Store the key for possible exit cleanups: */
942         pi_state->key = *key;
943
944         WARN_ON(!list_empty(&pi_state->list));
945         list_add(&pi_state->list, &p->pi_state_list);
946         pi_state->owner = p;
947         raw_spin_unlock_irq(&p->pi_lock);
948
949         put_task_struct(p);
950
951         *ps = pi_state;
952
953         return 0;
954 }
955
956 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
957                            union futex_key *key, struct futex_pi_state **ps)
958 {
959         struct futex_q *match = futex_top_waiter(hb, key);
960
961         /*
962          * If there is a waiter on that futex, validate it and
963          * attach to the pi_state when the validation succeeds.
964          */
965         if (match)
966                 return attach_to_pi_state(uval, match->pi_state, ps);
967
968         /*
969          * We are the first waiter - try to look up the owner based on
970          * @uval and attach to it.
971          */
972         return attach_to_pi_owner(uval, key, ps);
973 }
974
975 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
976 {
977         u32 uninitialized_var(curval);
978
979         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
980                 return -EFAULT;
981
982         /*If user space value changed, let the caller retry */
983         return curval != uval ? -EAGAIN : 0;
984 }
985
986 /**
987  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
988  * @uaddr:              the pi futex user address
989  * @hb:                 the pi futex hash bucket
990  * @key:                the futex key associated with uaddr and hb
991  * @ps:                 the pi_state pointer where we store the result of the
992  *                      lookup
993  * @task:               the task to perform the atomic lock work for.  This will
994  *                      be "current" except in the case of requeue pi.
995  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
996  *
997  * Return:
998  *  0 - ready to wait;
999  *  1 - acquired the lock;
1000  * <0 - error
1001  *
1002  * The hb->lock and futex_key refs shall be held by the caller.
1003  */
1004 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1005                                 union futex_key *key,
1006                                 struct futex_pi_state **ps,
1007                                 struct task_struct *task, int set_waiters)
1008 {
1009         u32 uval, newval, vpid = task_pid_vnr(task);
1010         struct futex_q *match;
1011         int ret;
1012
1013         /*
1014          * Read the user space value first so we can validate a few
1015          * things before proceeding further.
1016          */
1017         if (get_futex_value_locked(&uval, uaddr))
1018                 return -EFAULT;
1019
1020         /*
1021          * Detect deadlocks.
1022          */
1023         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1024                 return -EDEADLK;
1025
1026         /*
1027          * Lookup existing state first. If it exists, try to attach to
1028          * its pi_state.
1029          */
1030         match = futex_top_waiter(hb, key);
1031         if (match)
1032                 return attach_to_pi_state(uval, match->pi_state, ps);
1033
1034         /*
1035          * No waiter and user TID is 0. We are here because the
1036          * waiters or the owner died bit is set or called from
1037          * requeue_cmp_pi or for whatever reason something took the
1038          * syscall.
1039          */
1040         if (!(uval & FUTEX_TID_MASK)) {
1041                 /*
1042                  * We take over the futex. No other waiters and the user space
1043                  * TID is 0. We preserve the owner died bit.
1044                  */
1045                 newval = uval & FUTEX_OWNER_DIED;
1046                 newval |= vpid;
1047
1048                 /* The futex requeue_pi code can enforce the waiters bit */
1049                 if (set_waiters)
1050                         newval |= FUTEX_WAITERS;
1051
1052                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1053                 /* If the take over worked, return 1 */
1054                 return ret < 0 ? ret : 1;
1055         }
1056
1057         /*
1058          * First waiter. Set the waiters bit before attaching ourself to
1059          * the owner. If owner tries to unlock, it will be forced into
1060          * the kernel and blocked on hb->lock.
1061          */
1062         newval = uval | FUTEX_WAITERS;
1063         ret = lock_pi_update_atomic(uaddr, uval, newval);
1064         if (ret)
1065                 return ret;
1066         /*
1067          * If the update of the user space value succeeded, we try to
1068          * attach to the owner. If that fails, no harm done, we only
1069          * set the FUTEX_WAITERS bit in the user space variable.
1070          */
1071         return attach_to_pi_owner(uval, key, ps);
1072 }
1073
1074 /**
1075  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1076  * @q:  The futex_q to unqueue
1077  *
1078  * The q->lock_ptr must not be NULL and must be held by the caller.
1079  */
1080 static void __unqueue_futex(struct futex_q *q)
1081 {
1082         struct futex_hash_bucket *hb;
1083
1084         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1085             || WARN_ON(plist_node_empty(&q->list)))
1086                 return;
1087
1088         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1089         plist_del(&q->list, &hb->chain);
1090         hb_waiters_dec(hb);
1091 }
1092
1093 /*
1094  * The hash bucket lock must be held when this is called.
1095  * Afterwards, the futex_q must not be accessed. Callers
1096  * must ensure to later call wake_up_q() for the actual
1097  * wakeups to occur.
1098  */
1099 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1100 {
1101         struct task_struct *p = q->task;
1102
1103         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1104                 return;
1105
1106         /*
1107          * Queue the task for later wakeup for after we've released
1108          * the hb->lock. wake_q_add() grabs reference to p.
1109          */
1110         wake_q_add(wake_q, p);
1111         __unqueue_futex(q);
1112         /*
1113          * The waiting task can free the futex_q as soon as
1114          * q->lock_ptr = NULL is written, without taking any locks. A
1115          * memory barrier is required here to prevent the following
1116          * store to lock_ptr from getting ahead of the plist_del.
1117          */
1118         smp_wmb();
1119         q->lock_ptr = NULL;
1120 }
1121
1122 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1123                          struct futex_hash_bucket *hb)
1124 {
1125         struct task_struct *new_owner;
1126         struct futex_pi_state *pi_state = this->pi_state;
1127         u32 uninitialized_var(curval), newval;
1128         bool deboost;
1129         int ret = 0;
1130
1131         if (!pi_state)
1132                 return -EINVAL;
1133
1134         /*
1135          * If current does not own the pi_state then the futex is
1136          * inconsistent and user space fiddled with the futex value.
1137          */
1138         if (pi_state->owner != current)
1139                 return -EINVAL;
1140
1141         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1142         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1143
1144         /*
1145          * It is possible that the next waiter (the one that brought
1146          * this owner to the kernel) timed out and is no longer
1147          * waiting on the lock.
1148          */
1149         if (!new_owner)
1150                 new_owner = this->task;
1151
1152         /*
1153          * We pass it to the next owner. The WAITERS bit is always
1154          * kept enabled while there is PI state around. We cleanup the
1155          * owner died bit, because we are the owner.
1156          */
1157         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1158
1159         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1160                 ret = -EFAULT;
1161         else if (curval != uval)
1162                 ret = -EINVAL;
1163         if (ret) {
1164                 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1165                 return ret;
1166         }
1167
1168         raw_spin_lock_irq(&pi_state->owner->pi_lock);
1169         WARN_ON(list_empty(&pi_state->list));
1170         list_del_init(&pi_state->list);
1171         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1172
1173         raw_spin_lock_irq(&new_owner->pi_lock);
1174         WARN_ON(!list_empty(&pi_state->list));
1175         list_add(&pi_state->list, &new_owner->pi_state_list);
1176         pi_state->owner = new_owner;
1177         raw_spin_unlock_irq(&new_owner->pi_lock);
1178
1179         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1180
1181         deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex);
1182
1183         /*
1184          * We deboost after dropping hb->lock. That prevents a double
1185          * wakeup on RT.
1186          */
1187         spin_unlock(&hb->lock);
1188
1189         if (deboost)
1190                 rt_mutex_adjust_prio(current);
1191
1192         return 0;
1193 }
1194
1195 /*
1196  * Express the locking dependencies for lockdep:
1197  */
1198 static inline void
1199 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1200 {
1201         if (hb1 <= hb2) {
1202                 spin_lock(&hb1->lock);
1203                 if (hb1 < hb2)
1204                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1205         } else { /* hb1 > hb2 */
1206                 spin_lock(&hb2->lock);
1207                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1208         }
1209 }
1210
1211 static inline void
1212 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1213 {
1214         spin_unlock(&hb1->lock);
1215         if (hb1 != hb2)
1216                 spin_unlock(&hb2->lock);
1217 }
1218
1219 /*
1220  * Wake up waiters matching bitset queued on this futex (uaddr).
1221  */
1222 static int
1223 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1224 {
1225         struct futex_hash_bucket *hb;
1226         struct futex_q *this, *next;
1227         union futex_key key = FUTEX_KEY_INIT;
1228         int ret;
1229         WAKE_Q(wake_q);
1230
1231         if (!bitset)
1232                 return -EINVAL;
1233
1234         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1235         if (unlikely(ret != 0))
1236                 goto out;
1237
1238         hb = hash_futex(&key);
1239
1240         /* Make sure we really have tasks to wakeup */
1241         if (!hb_waiters_pending(hb))
1242                 goto out_put_key;
1243
1244         spin_lock(&hb->lock);
1245
1246         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1247                 if (match_futex (&this->key, &key)) {
1248                         if (this->pi_state || this->rt_waiter) {
1249                                 ret = -EINVAL;
1250                                 break;
1251                         }
1252
1253                         /* Check if one of the bits is set in both bitsets */
1254                         if (!(this->bitset & bitset))
1255                                 continue;
1256
1257                         mark_wake_futex(&wake_q, this);
1258                         if (++ret >= nr_wake)
1259                                 break;
1260                 }
1261         }
1262
1263         spin_unlock(&hb->lock);
1264         wake_up_q(&wake_q);
1265 out_put_key:
1266         put_futex_key(&key);
1267 out:
1268         return ret;
1269 }
1270
1271 /*
1272  * Wake up all waiters hashed on the physical page that is mapped
1273  * to this virtual address:
1274  */
1275 static int
1276 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1277               int nr_wake, int nr_wake2, int op)
1278 {
1279         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1280         struct futex_hash_bucket *hb1, *hb2;
1281         struct futex_q *this, *next;
1282         int ret, op_ret;
1283         WAKE_Q(wake_q);
1284
1285 retry:
1286         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1287         if (unlikely(ret != 0))
1288                 goto out;
1289         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1290         if (unlikely(ret != 0))
1291                 goto out_put_key1;
1292
1293         hb1 = hash_futex(&key1);
1294         hb2 = hash_futex(&key2);
1295
1296 retry_private:
1297         double_lock_hb(hb1, hb2);
1298         op_ret = futex_atomic_op_inuser(op, uaddr2);
1299         if (unlikely(op_ret < 0)) {
1300
1301                 double_unlock_hb(hb1, hb2);
1302
1303 #ifndef CONFIG_MMU
1304                 /*
1305                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1306                  * but we might get them from range checking
1307                  */
1308                 ret = op_ret;
1309                 goto out_put_keys;
1310 #endif
1311
1312                 if (unlikely(op_ret != -EFAULT)) {
1313                         ret = op_ret;
1314                         goto out_put_keys;
1315                 }
1316
1317                 ret = fault_in_user_writeable(uaddr2);
1318                 if (ret)
1319                         goto out_put_keys;
1320
1321                 if (!(flags & FLAGS_SHARED))
1322                         goto retry_private;
1323
1324                 put_futex_key(&key2);
1325                 put_futex_key(&key1);
1326                 goto retry;
1327         }
1328
1329         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1330                 if (match_futex (&this->key, &key1)) {
1331                         if (this->pi_state || this->rt_waiter) {
1332                                 ret = -EINVAL;
1333                                 goto out_unlock;
1334                         }
1335                         mark_wake_futex(&wake_q, this);
1336                         if (++ret >= nr_wake)
1337                                 break;
1338                 }
1339         }
1340
1341         if (op_ret > 0) {
1342                 op_ret = 0;
1343                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1344                         if (match_futex (&this->key, &key2)) {
1345                                 if (this->pi_state || this->rt_waiter) {
1346                                         ret = -EINVAL;
1347                                         goto out_unlock;
1348                                 }
1349                                 mark_wake_futex(&wake_q, this);
1350                                 if (++op_ret >= nr_wake2)
1351                                         break;
1352                         }
1353                 }
1354                 ret += op_ret;
1355         }
1356
1357 out_unlock:
1358         double_unlock_hb(hb1, hb2);
1359         wake_up_q(&wake_q);
1360 out_put_keys:
1361         put_futex_key(&key2);
1362 out_put_key1:
1363         put_futex_key(&key1);
1364 out:
1365         return ret;
1366 }
1367
1368 /**
1369  * requeue_futex() - Requeue a futex_q from one hb to another
1370  * @q:          the futex_q to requeue
1371  * @hb1:        the source hash_bucket
1372  * @hb2:        the target hash_bucket
1373  * @key2:       the new key for the requeued futex_q
1374  */
1375 static inline
1376 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1377                    struct futex_hash_bucket *hb2, union futex_key *key2)
1378 {
1379
1380         /*
1381          * If key1 and key2 hash to the same bucket, no need to
1382          * requeue.
1383          */
1384         if (likely(&hb1->chain != &hb2->chain)) {
1385                 plist_del(&q->list, &hb1->chain);
1386                 hb_waiters_dec(hb1);
1387                 plist_add(&q->list, &hb2->chain);
1388                 hb_waiters_inc(hb2);
1389                 q->lock_ptr = &hb2->lock;
1390         }
1391         get_futex_key_refs(key2);
1392         q->key = *key2;
1393 }
1394
1395 /**
1396  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1397  * @q:          the futex_q
1398  * @key:        the key of the requeue target futex
1399  * @hb:         the hash_bucket of the requeue target futex
1400  *
1401  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1402  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1403  * to the requeue target futex so the waiter can detect the wakeup on the right
1404  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1405  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1406  * to protect access to the pi_state to fixup the owner later.  Must be called
1407  * with both q->lock_ptr and hb->lock held.
1408  */
1409 static inline
1410 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1411                            struct futex_hash_bucket *hb)
1412 {
1413         get_futex_key_refs(key);
1414         q->key = *key;
1415
1416         __unqueue_futex(q);
1417
1418         WARN_ON(!q->rt_waiter);
1419         q->rt_waiter = NULL;
1420
1421         q->lock_ptr = &hb->lock;
1422
1423         wake_up_state(q->task, TASK_NORMAL);
1424 }
1425
1426 /**
1427  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1428  * @pifutex:            the user address of the to futex
1429  * @hb1:                the from futex hash bucket, must be locked by the caller
1430  * @hb2:                the to futex hash bucket, must be locked by the caller
1431  * @key1:               the from futex key
1432  * @key2:               the to futex key
1433  * @ps:                 address to store the pi_state pointer
1434  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1435  *
1436  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1437  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1438  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1439  * hb1 and hb2 must be held by the caller.
1440  *
1441  * Return:
1442  *  0 - failed to acquire the lock atomically;
1443  * >0 - acquired the lock, return value is vpid of the top_waiter
1444  * <0 - error
1445  */
1446 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1447                                  struct futex_hash_bucket *hb1,
1448                                  struct futex_hash_bucket *hb2,
1449                                  union futex_key *key1, union futex_key *key2,
1450                                  struct futex_pi_state **ps, int set_waiters)
1451 {
1452         struct futex_q *top_waiter = NULL;
1453         u32 curval;
1454         int ret, vpid;
1455
1456         if (get_futex_value_locked(&curval, pifutex))
1457                 return -EFAULT;
1458
1459         /*
1460          * Find the top_waiter and determine if there are additional waiters.
1461          * If the caller intends to requeue more than 1 waiter to pifutex,
1462          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1463          * as we have means to handle the possible fault.  If not, don't set
1464          * the bit unecessarily as it will force the subsequent unlock to enter
1465          * the kernel.
1466          */
1467         top_waiter = futex_top_waiter(hb1, key1);
1468
1469         /* There are no waiters, nothing for us to do. */
1470         if (!top_waiter)
1471                 return 0;
1472
1473         /* Ensure we requeue to the expected futex. */
1474         if (!match_futex(top_waiter->requeue_pi_key, key2))
1475                 return -EINVAL;
1476
1477         /*
1478          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1479          * the contended case or if set_waiters is 1.  The pi_state is returned
1480          * in ps in contended cases.
1481          */
1482         vpid = task_pid_vnr(top_waiter->task);
1483         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1484                                    set_waiters);
1485         if (ret == 1) {
1486                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1487                 return vpid;
1488         }
1489         return ret;
1490 }
1491
1492 /**
1493  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1494  * @uaddr1:     source futex user address
1495  * @flags:      futex flags (FLAGS_SHARED, etc.)
1496  * @uaddr2:     target futex user address
1497  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1498  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1499  * @cmpval:     @uaddr1 expected value (or %NULL)
1500  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1501  *              pi futex (pi to pi requeue is not supported)
1502  *
1503  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1504  * uaddr2 atomically on behalf of the top waiter.
1505  *
1506  * Return:
1507  * >=0 - on success, the number of tasks requeued or woken;
1508  *  <0 - on error
1509  */
1510 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1511                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1512                          u32 *cmpval, int requeue_pi)
1513 {
1514         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1515         int drop_count = 0, task_count = 0, ret;
1516         struct futex_pi_state *pi_state = NULL;
1517         struct futex_hash_bucket *hb1, *hb2;
1518         struct futex_q *this, *next;
1519         WAKE_Q(wake_q);
1520
1521         if (requeue_pi) {
1522                 /*
1523                  * Requeue PI only works on two distinct uaddrs. This
1524                  * check is only valid for private futexes. See below.
1525                  */
1526                 if (uaddr1 == uaddr2)
1527                         return -EINVAL;
1528
1529                 /*
1530                  * requeue_pi requires a pi_state, try to allocate it now
1531                  * without any locks in case it fails.
1532                  */
1533                 if (refill_pi_state_cache())
1534                         return -ENOMEM;
1535                 /*
1536                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1537                  * + nr_requeue, since it acquires the rt_mutex prior to
1538                  * returning to userspace, so as to not leave the rt_mutex with
1539                  * waiters and no owner.  However, second and third wake-ups
1540                  * cannot be predicted as they involve race conditions with the
1541                  * first wake and a fault while looking up the pi_state.  Both
1542                  * pthread_cond_signal() and pthread_cond_broadcast() should
1543                  * use nr_wake=1.
1544                  */
1545                 if (nr_wake != 1)
1546                         return -EINVAL;
1547         }
1548
1549 retry:
1550         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1551         if (unlikely(ret != 0))
1552                 goto out;
1553         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1554                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1555         if (unlikely(ret != 0))
1556                 goto out_put_key1;
1557
1558         /*
1559          * The check above which compares uaddrs is not sufficient for
1560          * shared futexes. We need to compare the keys:
1561          */
1562         if (requeue_pi && match_futex(&key1, &key2)) {
1563                 ret = -EINVAL;
1564                 goto out_put_keys;
1565         }
1566
1567         hb1 = hash_futex(&key1);
1568         hb2 = hash_futex(&key2);
1569
1570 retry_private:
1571         hb_waiters_inc(hb2);
1572         double_lock_hb(hb1, hb2);
1573
1574         if (likely(cmpval != NULL)) {
1575                 u32 curval;
1576
1577                 ret = get_futex_value_locked(&curval, uaddr1);
1578
1579                 if (unlikely(ret)) {
1580                         double_unlock_hb(hb1, hb2);
1581                         hb_waiters_dec(hb2);
1582
1583                         ret = get_user(curval, uaddr1);
1584                         if (ret)
1585                                 goto out_put_keys;
1586
1587                         if (!(flags & FLAGS_SHARED))
1588                                 goto retry_private;
1589
1590                         put_futex_key(&key2);
1591                         put_futex_key(&key1);
1592                         goto retry;
1593                 }
1594                 if (curval != *cmpval) {
1595                         ret = -EAGAIN;
1596                         goto out_unlock;
1597                 }
1598         }
1599
1600         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1601                 /*
1602                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1603                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1604                  * bit.  We force this here where we are able to easily handle
1605                  * faults rather in the requeue loop below.
1606                  */
1607                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1608                                                  &key2, &pi_state, nr_requeue);
1609
1610                 /*
1611                  * At this point the top_waiter has either taken uaddr2 or is
1612                  * waiting on it.  If the former, then the pi_state will not
1613                  * exist yet, look it up one more time to ensure we have a
1614                  * reference to it. If the lock was taken, ret contains the
1615                  * vpid of the top waiter task.
1616                  */
1617                 if (ret > 0) {
1618                         WARN_ON(pi_state);
1619                         drop_count++;
1620                         task_count++;
1621                         /*
1622                          * If we acquired the lock, then the user
1623                          * space value of uaddr2 should be vpid. It
1624                          * cannot be changed by the top waiter as it
1625                          * is blocked on hb2 lock if it tries to do
1626                          * so. If something fiddled with it behind our
1627                          * back the pi state lookup might unearth
1628                          * it. So we rather use the known value than
1629                          * rereading and handing potential crap to
1630                          * lookup_pi_state.
1631                          */
1632                         ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1633                 }
1634
1635                 switch (ret) {
1636                 case 0:
1637                         break;
1638                 case -EFAULT:
1639                         free_pi_state(pi_state);
1640                         pi_state = NULL;
1641                         double_unlock_hb(hb1, hb2);
1642                         hb_waiters_dec(hb2);
1643                         put_futex_key(&key2);
1644                         put_futex_key(&key1);
1645                         ret = fault_in_user_writeable(uaddr2);
1646                         if (!ret)
1647                                 goto retry;
1648                         goto out;
1649                 case -EAGAIN:
1650                         /*
1651                          * Two reasons for this:
1652                          * - Owner is exiting and we just wait for the
1653                          *   exit to complete.
1654                          * - The user space value changed.
1655                          */
1656                         free_pi_state(pi_state);
1657                         pi_state = NULL;
1658                         double_unlock_hb(hb1, hb2);
1659                         hb_waiters_dec(hb2);
1660                         put_futex_key(&key2);
1661                         put_futex_key(&key1);
1662                         cond_resched();
1663                         goto retry;
1664                 default:
1665                         goto out_unlock;
1666                 }
1667         }
1668
1669         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1670                 if (task_count - nr_wake >= nr_requeue)
1671                         break;
1672
1673                 if (!match_futex(&this->key, &key1))
1674                         continue;
1675
1676                 /*
1677                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1678                  * be paired with each other and no other futex ops.
1679                  *
1680                  * We should never be requeueing a futex_q with a pi_state,
1681                  * which is awaiting a futex_unlock_pi().
1682                  */
1683                 if ((requeue_pi && !this->rt_waiter) ||
1684                     (!requeue_pi && this->rt_waiter) ||
1685                     this->pi_state) {
1686                         ret = -EINVAL;
1687                         break;
1688                 }
1689
1690                 /*
1691                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1692                  * lock, we already woke the top_waiter.  If not, it will be
1693                  * woken by futex_unlock_pi().
1694                  */
1695                 if (++task_count <= nr_wake && !requeue_pi) {
1696                         mark_wake_futex(&wake_q, this);
1697                         continue;
1698                 }
1699
1700                 /* Ensure we requeue to the expected futex for requeue_pi. */
1701                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1702                         ret = -EINVAL;
1703                         break;
1704                 }
1705
1706                 /*
1707                  * Requeue nr_requeue waiters and possibly one more in the case
1708                  * of requeue_pi if we couldn't acquire the lock atomically.
1709                  */
1710                 if (requeue_pi) {
1711                         /* Prepare the waiter to take the rt_mutex. */
1712                         atomic_inc(&pi_state->refcount);
1713                         this->pi_state = pi_state;
1714                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1715                                                         this->rt_waiter,
1716                                                         this->task);
1717                         if (ret == 1) {
1718                                 /* We got the lock. */
1719                                 requeue_pi_wake_futex(this, &key2, hb2);
1720                                 drop_count++;
1721                                 continue;
1722                         } else if (ret == -EAGAIN) {
1723                                 /*
1724                                  * Waiter was woken by timeout or
1725                                  * signal and has set pi_blocked_on to
1726                                  * PI_WAKEUP_INPROGRESS before we
1727                                  * tried to enqueue it on the rtmutex.
1728                                  */
1729                                 this->pi_state = NULL;
1730                                 free_pi_state(pi_state);
1731                                 continue;
1732                         } else if (ret) {
1733                                 /* -EDEADLK */
1734                                 this->pi_state = NULL;
1735                                 free_pi_state(pi_state);
1736                                 goto out_unlock;
1737                         }
1738                 }
1739                 requeue_futex(this, hb1, hb2, &key2);
1740                 drop_count++;
1741         }
1742
1743 out_unlock:
1744         free_pi_state(pi_state);
1745         double_unlock_hb(hb1, hb2);
1746         wake_up_q(&wake_q);
1747         hb_waiters_dec(hb2);
1748
1749         /*
1750          * drop_futex_key_refs() must be called outside the spinlocks. During
1751          * the requeue we moved futex_q's from the hash bucket at key1 to the
1752          * one at key2 and updated their key pointer.  We no longer need to
1753          * hold the references to key1.
1754          */
1755         while (--drop_count >= 0)
1756                 drop_futex_key_refs(&key1);
1757
1758 out_put_keys:
1759         put_futex_key(&key2);
1760 out_put_key1:
1761         put_futex_key(&key1);
1762 out:
1763         return ret ? ret : task_count;
1764 }
1765
1766 /* The key must be already stored in q->key. */
1767 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1768         __acquires(&hb->lock)
1769 {
1770         struct futex_hash_bucket *hb;
1771
1772         hb = hash_futex(&q->key);
1773
1774         /*
1775          * Increment the counter before taking the lock so that
1776          * a potential waker won't miss a to-be-slept task that is
1777          * waiting for the spinlock. This is safe as all queue_lock()
1778          * users end up calling queue_me(). Similarly, for housekeeping,
1779          * decrement the counter at queue_unlock() when some error has
1780          * occurred and we don't end up adding the task to the list.
1781          */
1782         hb_waiters_inc(hb);
1783
1784         q->lock_ptr = &hb->lock;
1785
1786         spin_lock(&hb->lock); /* implies MB (A) */
1787         return hb;
1788 }
1789
1790 static inline void
1791 queue_unlock(struct futex_hash_bucket *hb)
1792         __releases(&hb->lock)
1793 {
1794         spin_unlock(&hb->lock);
1795         hb_waiters_dec(hb);
1796 }
1797
1798 /**
1799  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1800  * @q:  The futex_q to enqueue
1801  * @hb: The destination hash bucket
1802  *
1803  * The hb->lock must be held by the caller, and is released here. A call to
1804  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1805  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1806  * or nothing if the unqueue is done as part of the wake process and the unqueue
1807  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1808  * an example).
1809  */
1810 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1811         __releases(&hb->lock)
1812 {
1813         int prio;
1814
1815         /*
1816          * The priority used to register this element is
1817          * - either the real thread-priority for the real-time threads
1818          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1819          * - or MAX_RT_PRIO for non-RT threads.
1820          * Thus, all RT-threads are woken first in priority order, and
1821          * the others are woken last, in FIFO order.
1822          */
1823         prio = min(current->normal_prio, MAX_RT_PRIO);
1824
1825         plist_node_init(&q->list, prio);
1826         plist_add(&q->list, &hb->chain);
1827         q->task = current;
1828         spin_unlock(&hb->lock);
1829 }
1830
1831 /**
1832  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1833  * @q:  The futex_q to unqueue
1834  *
1835  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1836  * be paired with exactly one earlier call to queue_me().
1837  *
1838  * Return:
1839  *   1 - if the futex_q was still queued (and we removed unqueued it);
1840  *   0 - if the futex_q was already removed by the waking thread
1841  */
1842 static int unqueue_me(struct futex_q *q)
1843 {
1844         spinlock_t *lock_ptr;
1845         int ret = 0;
1846
1847         /* In the common case we don't take the spinlock, which is nice. */
1848 retry:
1849         lock_ptr = q->lock_ptr;
1850         barrier();
1851         if (lock_ptr != NULL) {
1852                 spin_lock(lock_ptr);
1853                 /*
1854                  * q->lock_ptr can change between reading it and
1855                  * spin_lock(), causing us to take the wrong lock.  This
1856                  * corrects the race condition.
1857                  *
1858                  * Reasoning goes like this: if we have the wrong lock,
1859                  * q->lock_ptr must have changed (maybe several times)
1860                  * between reading it and the spin_lock().  It can
1861                  * change again after the spin_lock() but only if it was
1862                  * already changed before the spin_lock().  It cannot,
1863                  * however, change back to the original value.  Therefore
1864                  * we can detect whether we acquired the correct lock.
1865                  */
1866                 if (unlikely(lock_ptr != q->lock_ptr)) {
1867                         spin_unlock(lock_ptr);
1868                         goto retry;
1869                 }
1870                 __unqueue_futex(q);
1871
1872                 BUG_ON(q->pi_state);
1873
1874                 spin_unlock(lock_ptr);
1875                 ret = 1;
1876         }
1877
1878         drop_futex_key_refs(&q->key);
1879         return ret;
1880 }
1881
1882 /*
1883  * PI futexes can not be requeued and must remove themself from the
1884  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1885  * and dropped here.
1886  */
1887 static void unqueue_me_pi(struct futex_q *q)
1888         __releases(q->lock_ptr)
1889 {
1890         __unqueue_futex(q);
1891
1892         BUG_ON(!q->pi_state);
1893         free_pi_state(q->pi_state);
1894         q->pi_state = NULL;
1895
1896         spin_unlock(q->lock_ptr);
1897 }
1898
1899 /*
1900  * Fixup the pi_state owner with the new owner.
1901  *
1902  * Must be called with hash bucket lock held and mm->sem held for non
1903  * private futexes.
1904  */
1905 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1906                                 struct task_struct *newowner)
1907 {
1908         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1909         struct futex_pi_state *pi_state = q->pi_state;
1910         struct task_struct *oldowner = pi_state->owner;
1911         u32 uval, uninitialized_var(curval), newval;
1912         int ret;
1913
1914         /* Owner died? */
1915         if (!pi_state->owner)
1916                 newtid |= FUTEX_OWNER_DIED;
1917
1918         /*
1919          * We are here either because we stole the rtmutex from the
1920          * previous highest priority waiter or we are the highest priority
1921          * waiter but failed to get the rtmutex the first time.
1922          * We have to replace the newowner TID in the user space variable.
1923          * This must be atomic as we have to preserve the owner died bit here.
1924          *
1925          * Note: We write the user space value _before_ changing the pi_state
1926          * because we can fault here. Imagine swapped out pages or a fork
1927          * that marked all the anonymous memory readonly for cow.
1928          *
1929          * Modifying pi_state _before_ the user space value would
1930          * leave the pi_state in an inconsistent state when we fault
1931          * here, because we need to drop the hash bucket lock to
1932          * handle the fault. This might be observed in the PID check
1933          * in lookup_pi_state.
1934          */
1935 retry:
1936         if (get_futex_value_locked(&uval, uaddr))
1937                 goto handle_fault;
1938
1939         while (1) {
1940                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1941
1942                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1943                         goto handle_fault;
1944                 if (curval == uval)
1945                         break;
1946                 uval = curval;
1947         }
1948
1949         /*
1950          * We fixed up user space. Now we need to fix the pi_state
1951          * itself.
1952          */
1953         if (pi_state->owner != NULL) {
1954                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1955                 WARN_ON(list_empty(&pi_state->list));
1956                 list_del_init(&pi_state->list);
1957                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1958         }
1959
1960         pi_state->owner = newowner;
1961
1962         raw_spin_lock_irq(&newowner->pi_lock);
1963         WARN_ON(!list_empty(&pi_state->list));
1964         list_add(&pi_state->list, &newowner->pi_state_list);
1965         raw_spin_unlock_irq(&newowner->pi_lock);
1966         return 0;
1967
1968         /*
1969          * To handle the page fault we need to drop the hash bucket
1970          * lock here. That gives the other task (either the highest priority
1971          * waiter itself or the task which stole the rtmutex) the
1972          * chance to try the fixup of the pi_state. So once we are
1973          * back from handling the fault we need to check the pi_state
1974          * after reacquiring the hash bucket lock and before trying to
1975          * do another fixup. When the fixup has been done already we
1976          * simply return.
1977          */
1978 handle_fault:
1979         spin_unlock(q->lock_ptr);
1980
1981         ret = fault_in_user_writeable(uaddr);
1982
1983         spin_lock(q->lock_ptr);
1984
1985         /*
1986          * Check if someone else fixed it for us:
1987          */
1988         if (pi_state->owner != oldowner)
1989                 return 0;
1990
1991         if (ret)
1992                 return ret;
1993
1994         goto retry;
1995 }
1996
1997 static long futex_wait_restart(struct restart_block *restart);
1998
1999 /**
2000  * fixup_owner() - Post lock pi_state and corner case management
2001  * @uaddr:      user address of the futex
2002  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2003  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2004  *
2005  * After attempting to lock an rt_mutex, this function is called to cleanup
2006  * the pi_state owner as well as handle race conditions that may allow us to
2007  * acquire the lock. Must be called with the hb lock held.
2008  *
2009  * Return:
2010  *  1 - success, lock taken;
2011  *  0 - success, lock not taken;
2012  * <0 - on error (-EFAULT)
2013  */
2014 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2015 {
2016         struct task_struct *owner;
2017         int ret = 0;
2018
2019         if (locked) {
2020                 /*
2021                  * Got the lock. We might not be the anticipated owner if we
2022                  * did a lock-steal - fix up the PI-state in that case:
2023                  */
2024                 if (q->pi_state->owner != current)
2025                         ret = fixup_pi_state_owner(uaddr, q, current);
2026                 goto out;
2027         }
2028
2029         /*
2030          * Catch the rare case, where the lock was released when we were on the
2031          * way back before we locked the hash bucket.
2032          */
2033         if (q->pi_state->owner == current) {
2034                 /*
2035                  * Try to get the rt_mutex now. This might fail as some other
2036                  * task acquired the rt_mutex after we removed ourself from the
2037                  * rt_mutex waiters list.
2038                  */
2039                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2040                         locked = 1;
2041                         goto out;
2042                 }
2043
2044                 /*
2045                  * pi_state is incorrect, some other task did a lock steal and
2046                  * we returned due to timeout or signal without taking the
2047                  * rt_mutex. Too late.
2048                  */
2049                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2050                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2051                 if (!owner)
2052                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2053                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2054                 ret = fixup_pi_state_owner(uaddr, q, owner);
2055                 goto out;
2056         }
2057
2058         /*
2059          * Paranoia check. If we did not take the lock, then we should not be
2060          * the owner of the rt_mutex.
2061          */
2062         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2063                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2064                                 "pi-state %p\n", ret,
2065                                 q->pi_state->pi_mutex.owner,
2066                                 q->pi_state->owner);
2067
2068 out:
2069         return ret ? ret : locked;
2070 }
2071
2072 /**
2073  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2074  * @hb:         the futex hash bucket, must be locked by the caller
2075  * @q:          the futex_q to queue up on
2076  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2077  */
2078 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2079                                 struct hrtimer_sleeper *timeout)
2080 {
2081         /*
2082          * The task state is guaranteed to be set before another task can
2083          * wake it. set_current_state() is implemented using set_mb() and
2084          * queue_me() calls spin_unlock() upon completion, both serializing
2085          * access to the hash list and forcing another memory barrier.
2086          */
2087         set_current_state(TASK_INTERRUPTIBLE);
2088         queue_me(q, hb);
2089
2090         /* Arm the timer */
2091         if (timeout) {
2092                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2093                 if (!hrtimer_active(&timeout->timer))
2094                         timeout->task = NULL;
2095         }
2096
2097         /*
2098          * If we have been removed from the hash list, then another task
2099          * has tried to wake us, and we can skip the call to schedule().
2100          */
2101         if (likely(!plist_node_empty(&q->list))) {
2102                 /*
2103                  * If the timer has already expired, current will already be
2104                  * flagged for rescheduling. Only call schedule if there
2105                  * is no timeout, or if it has yet to expire.
2106                  */
2107                 if (!timeout || timeout->task)
2108                         freezable_schedule();
2109         }
2110         __set_current_state(TASK_RUNNING);
2111 }
2112
2113 /**
2114  * futex_wait_setup() - Prepare to wait on a futex
2115  * @uaddr:      the futex userspace address
2116  * @val:        the expected value
2117  * @flags:      futex flags (FLAGS_SHARED, etc.)
2118  * @q:          the associated futex_q
2119  * @hb:         storage for hash_bucket pointer to be returned to caller
2120  *
2121  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2122  * compare it with the expected value.  Handle atomic faults internally.
2123  * Return with the hb lock held and a q.key reference on success, and unlocked
2124  * with no q.key reference on failure.
2125  *
2126  * Return:
2127  *  0 - uaddr contains val and hb has been locked;
2128  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2129  */
2130 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2131                            struct futex_q *q, struct futex_hash_bucket **hb)
2132 {
2133         u32 uval;
2134         int ret;
2135
2136         /*
2137          * Access the page AFTER the hash-bucket is locked.
2138          * Order is important:
2139          *
2140          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2141          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2142          *
2143          * The basic logical guarantee of a futex is that it blocks ONLY
2144          * if cond(var) is known to be true at the time of blocking, for
2145          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2146          * would open a race condition where we could block indefinitely with
2147          * cond(var) false, which would violate the guarantee.
2148          *
2149          * On the other hand, we insert q and release the hash-bucket only
2150          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2151          * absorb a wakeup if *uaddr does not match the desired values
2152          * while the syscall executes.
2153          */
2154 retry:
2155         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2156         if (unlikely(ret != 0))
2157                 return ret;
2158
2159 retry_private:
2160         *hb = queue_lock(q);
2161
2162         ret = get_futex_value_locked(&uval, uaddr);
2163
2164         if (ret) {
2165                 queue_unlock(*hb);
2166
2167                 ret = get_user(uval, uaddr);
2168                 if (ret)
2169                         goto out;
2170
2171                 if (!(flags & FLAGS_SHARED))
2172                         goto retry_private;
2173
2174                 put_futex_key(&q->key);
2175                 goto retry;
2176         }
2177
2178         if (uval != val) {
2179                 queue_unlock(*hb);
2180                 ret = -EWOULDBLOCK;
2181         }
2182
2183 out:
2184         if (ret)
2185                 put_futex_key(&q->key);
2186         return ret;
2187 }
2188
2189 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2190                       ktime_t *abs_time, u32 bitset)
2191 {
2192         struct hrtimer_sleeper timeout, *to = NULL;
2193         struct restart_block *restart;
2194         struct futex_hash_bucket *hb;
2195         struct futex_q q = futex_q_init;
2196         int ret;
2197
2198         if (!bitset)
2199                 return -EINVAL;
2200         q.bitset = bitset;
2201
2202         if (abs_time) {
2203                 to = &timeout;
2204
2205                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2206                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2207                                       HRTIMER_MODE_ABS);
2208                 hrtimer_init_sleeper(to, current);
2209                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2210                                              current->timer_slack_ns);
2211         }
2212
2213 retry:
2214         /*
2215          * Prepare to wait on uaddr. On success, holds hb lock and increments
2216          * q.key refs.
2217          */
2218         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2219         if (ret)
2220                 goto out;
2221
2222         /* queue_me and wait for wakeup, timeout, or a signal. */
2223         futex_wait_queue_me(hb, &q, to);
2224
2225         /* If we were woken (and unqueued), we succeeded, whatever. */
2226         ret = 0;
2227         /* unqueue_me() drops q.key ref */
2228         if (!unqueue_me(&q))
2229                 goto out;
2230         ret = -ETIMEDOUT;
2231         if (to && !to->task)
2232                 goto out;
2233
2234         /*
2235          * We expect signal_pending(current), but we might be the
2236          * victim of a spurious wakeup as well.
2237          */
2238         if (!signal_pending(current))
2239                 goto retry;
2240
2241         ret = -ERESTARTSYS;
2242         if (!abs_time)
2243                 goto out;
2244
2245         restart = &current->restart_block;
2246         restart->fn = futex_wait_restart;
2247         restart->futex.uaddr = uaddr;
2248         restart->futex.val = val;
2249         restart->futex.time = abs_time->tv64;
2250         restart->futex.bitset = bitset;
2251         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2252
2253         ret = -ERESTART_RESTARTBLOCK;
2254
2255 out:
2256         if (to) {
2257                 hrtimer_cancel(&to->timer);
2258                 destroy_hrtimer_on_stack(&to->timer);
2259         }
2260         return ret;
2261 }
2262
2263
2264 static long futex_wait_restart(struct restart_block *restart)
2265 {
2266         u32 __user *uaddr = restart->futex.uaddr;
2267         ktime_t t, *tp = NULL;
2268
2269         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2270                 t.tv64 = restart->futex.time;
2271                 tp = &t;
2272         }
2273         restart->fn = do_no_restart_syscall;
2274
2275         return (long)futex_wait(uaddr, restart->futex.flags,
2276                                 restart->futex.val, tp, restart->futex.bitset);
2277 }
2278
2279
2280 /*
2281  * Userspace tried a 0 -> TID atomic transition of the futex value
2282  * and failed. The kernel side here does the whole locking operation:
2283  * if there are waiters then it will block, it does PI, etc. (Due to
2284  * races the kernel might see a 0 value of the futex too.)
2285  */
2286 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2287                          ktime_t *time, int trylock)
2288 {
2289         struct hrtimer_sleeper timeout, *to = NULL;
2290         struct futex_hash_bucket *hb;
2291         struct futex_q q = futex_q_init;
2292         int res, ret;
2293
2294         if (refill_pi_state_cache())
2295                 return -ENOMEM;
2296
2297         if (time) {
2298                 to = &timeout;
2299                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2300                                       HRTIMER_MODE_ABS);
2301                 hrtimer_init_sleeper(to, current);
2302                 hrtimer_set_expires(&to->timer, *time);
2303         }
2304
2305 retry:
2306         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2307         if (unlikely(ret != 0))
2308                 goto out;
2309
2310 retry_private:
2311         hb = queue_lock(&q);
2312
2313         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2314         if (unlikely(ret)) {
2315                 switch (ret) {
2316                 case 1:
2317                         /* We got the lock. */
2318                         ret = 0;
2319                         goto out_unlock_put_key;
2320                 case -EFAULT:
2321                         goto uaddr_faulted;
2322                 case -EAGAIN:
2323                         /*
2324                          * Two reasons for this:
2325                          * - Task is exiting and we just wait for the
2326                          *   exit to complete.
2327                          * - The user space value changed.
2328                          */
2329                         queue_unlock(hb);
2330                         put_futex_key(&q.key);
2331                         cond_resched();
2332                         goto retry;
2333                 default:
2334                         goto out_unlock_put_key;
2335                 }
2336         }
2337
2338         /*
2339          * Only actually queue now that the atomic ops are done:
2340          */
2341         queue_me(&q, hb);
2342
2343         WARN_ON(!q.pi_state);
2344         /*
2345          * Block on the PI mutex:
2346          */
2347         if (!trylock) {
2348                 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2349         } else {
2350                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2351                 /* Fixup the trylock return value: */
2352                 ret = ret ? 0 : -EWOULDBLOCK;
2353         }
2354
2355         spin_lock(q.lock_ptr);
2356         /*
2357          * Fixup the pi_state owner and possibly acquire the lock if we
2358          * haven't already.
2359          */
2360         res = fixup_owner(uaddr, &q, !ret);
2361         /*
2362          * If fixup_owner() returned an error, proprogate that.  If it acquired
2363          * the lock, clear our -ETIMEDOUT or -EINTR.
2364          */
2365         if (res)
2366                 ret = (res < 0) ? res : 0;
2367
2368         /*
2369          * If fixup_owner() faulted and was unable to handle the fault, unlock
2370          * it and return the fault to userspace.
2371          */
2372         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2373                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2374
2375         /* Unqueue and drop the lock */
2376         unqueue_me_pi(&q);
2377
2378         goto out_put_key;
2379
2380 out_unlock_put_key:
2381         queue_unlock(hb);
2382
2383 out_put_key:
2384         put_futex_key(&q.key);
2385 out:
2386         if (to)
2387                 destroy_hrtimer_on_stack(&to->timer);
2388         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2389
2390 uaddr_faulted:
2391         queue_unlock(hb);
2392
2393         ret = fault_in_user_writeable(uaddr);
2394         if (ret)
2395                 goto out_put_key;
2396
2397         if (!(flags & FLAGS_SHARED))
2398                 goto retry_private;
2399
2400         put_futex_key(&q.key);
2401         goto retry;
2402 }
2403
2404 /*
2405  * Userspace attempted a TID -> 0 atomic transition, and failed.
2406  * This is the in-kernel slowpath: we look up the PI state (if any),
2407  * and do the rt-mutex unlock.
2408  */
2409 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2410 {
2411         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2412         union futex_key key = FUTEX_KEY_INIT;
2413         struct futex_hash_bucket *hb;
2414         struct futex_q *match;
2415         int ret;
2416
2417 retry:
2418         if (get_user(uval, uaddr))
2419                 return -EFAULT;
2420         /*
2421          * We release only a lock we actually own:
2422          */
2423         if ((uval & FUTEX_TID_MASK) != vpid)
2424                 return -EPERM;
2425
2426         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2427         if (ret)
2428                 return ret;
2429
2430         hb = hash_futex(&key);
2431         spin_lock(&hb->lock);
2432
2433         /*
2434          * Check waiters first. We do not trust user space values at
2435          * all and we at least want to know if user space fiddled
2436          * with the futex value instead of blindly unlocking.
2437          */
2438         match = futex_top_waiter(hb, &key);
2439         if (match) {
2440                 ret = wake_futex_pi(uaddr, uval, match, hb);
2441
2442                 /*
2443                  * In case of success wake_futex_pi dropped the hash
2444                  * bucket lock.
2445                  */
2446                 if (!ret)
2447                         goto out_putkey;
2448
2449                 /*
2450                  * The atomic access to the futex value generated a
2451                  * pagefault, so retry the user-access and the wakeup:
2452                  */
2453                 if (ret == -EFAULT)
2454                         goto pi_faulted;
2455
2456                 /*
2457                  * wake_futex_pi has detected invalid state. Tell user
2458                  * space.
2459                  */
2460                 goto out_unlock;
2461         }
2462
2463         /*
2464          * We have no kernel internal state, i.e. no waiters in the
2465          * kernel. Waiters which are about to queue themselves are stuck
2466          * on hb->lock. So we can safely ignore them. We do neither
2467          * preserve the WAITERS bit not the OWNER_DIED one. We are the
2468          * owner.
2469          */
2470         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2471                 goto pi_faulted;
2472
2473         /*
2474          * If uval has changed, let user space handle it.
2475          */
2476         ret = (curval == uval) ? 0 : -EAGAIN;
2477
2478 out_unlock:
2479         spin_unlock(&hb->lock);
2480 out_putkey:
2481         put_futex_key(&key);
2482         return ret;
2483
2484 pi_faulted:
2485         spin_unlock(&hb->lock);
2486         put_futex_key(&key);
2487
2488         ret = fault_in_user_writeable(uaddr);
2489         if (!ret)
2490                 goto retry;
2491
2492         return ret;
2493 }
2494
2495 /**
2496  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2497  * @hb:         the hash_bucket futex_q was original enqueued on
2498  * @q:          the futex_q woken while waiting to be requeued
2499  * @key2:       the futex_key of the requeue target futex
2500  * @timeout:    the timeout associated with the wait (NULL if none)
2501  *
2502  * Detect if the task was woken on the initial futex as opposed to the requeue
2503  * target futex.  If so, determine if it was a timeout or a signal that caused
2504  * the wakeup and return the appropriate error code to the caller.  Must be
2505  * called with the hb lock held.
2506  *
2507  * Return:
2508  *  0 = no early wakeup detected;
2509  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2510  */
2511 static inline
2512 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2513                                    struct futex_q *q, union futex_key *key2,
2514                                    struct hrtimer_sleeper *timeout)
2515 {
2516         int ret = 0;
2517
2518         /*
2519          * With the hb lock held, we avoid races while we process the wakeup.
2520          * We only need to hold hb (and not hb2) to ensure atomicity as the
2521          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2522          * It can't be requeued from uaddr2 to something else since we don't
2523          * support a PI aware source futex for requeue.
2524          */
2525         if (!match_futex(&q->key, key2)) {
2526                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2527                 /*
2528                  * We were woken prior to requeue by a timeout or a signal.
2529                  * Unqueue the futex_q and determine which it was.
2530                  */
2531                 plist_del(&q->list, &hb->chain);
2532                 hb_waiters_dec(hb);
2533
2534                 /* Handle spurious wakeups gracefully */
2535                 ret = -EWOULDBLOCK;
2536                 if (timeout && !timeout->task)
2537                         ret = -ETIMEDOUT;
2538                 else if (signal_pending(current))
2539                         ret = -ERESTARTNOINTR;
2540         }
2541         return ret;
2542 }
2543
2544 /**
2545  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2546  * @uaddr:      the futex we initially wait on (non-pi)
2547  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2548  *              the same type, no requeueing from private to shared, etc.
2549  * @val:        the expected value of uaddr
2550  * @abs_time:   absolute timeout
2551  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2552  * @uaddr2:     the pi futex we will take prior to returning to user-space
2553  *
2554  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2555  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2556  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2557  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2558  * without one, the pi logic would not know which task to boost/deboost, if
2559  * there was a need to.
2560  *
2561  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2562  * via the following--
2563  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2564  * 2) wakeup on uaddr2 after a requeue
2565  * 3) signal
2566  * 4) timeout
2567  *
2568  * If 3, cleanup and return -ERESTARTNOINTR.
2569  *
2570  * If 2, we may then block on trying to take the rt_mutex and return via:
2571  * 5) successful lock
2572  * 6) signal
2573  * 7) timeout
2574  * 8) other lock acquisition failure
2575  *
2576  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2577  *
2578  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2579  *
2580  * Return:
2581  *  0 - On success;
2582  * <0 - On error
2583  */
2584 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2585                                  u32 val, ktime_t *abs_time, u32 bitset,
2586                                  u32 __user *uaddr2)
2587 {
2588         struct hrtimer_sleeper timeout, *to = NULL;
2589         struct rt_mutex_waiter rt_waiter;
2590         struct rt_mutex *pi_mutex = NULL;
2591         struct futex_hash_bucket *hb, *hb2;
2592         union futex_key key2 = FUTEX_KEY_INIT;
2593         struct futex_q q = futex_q_init;
2594         int res, ret;
2595
2596         if (uaddr == uaddr2)
2597                 return -EINVAL;
2598
2599         if (!bitset)
2600                 return -EINVAL;
2601
2602         if (abs_time) {
2603                 to = &timeout;
2604                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2605                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2606                                       HRTIMER_MODE_ABS);
2607                 hrtimer_init_sleeper(to, current);
2608                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2609                                              current->timer_slack_ns);
2610         }
2611
2612         /*
2613          * The waiter is allocated on our stack, manipulated by the requeue
2614          * code while we sleep on uaddr.
2615          */
2616         rt_mutex_init_waiter(&rt_waiter, false);
2617
2618         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2619         if (unlikely(ret != 0))
2620                 goto out;
2621
2622         q.bitset = bitset;
2623         q.rt_waiter = &rt_waiter;
2624         q.requeue_pi_key = &key2;
2625
2626         /*
2627          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2628          * count.
2629          */
2630         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2631         if (ret)
2632                 goto out_key2;
2633
2634         /*
2635          * The check above which compares uaddrs is not sufficient for
2636          * shared futexes. We need to compare the keys:
2637          */
2638         if (match_futex(&q.key, &key2)) {
2639                 queue_unlock(hb);
2640                 ret = -EINVAL;
2641                 goto out_put_keys;
2642         }
2643
2644         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2645         futex_wait_queue_me(hb, &q, to);
2646
2647         /*
2648          * On RT we must avoid races with requeue and trying to block
2649          * on two mutexes (hb->lock and uaddr2's rtmutex) by
2650          * serializing access to pi_blocked_on with pi_lock.
2651          */
2652         raw_spin_lock_irq(&current->pi_lock);
2653         if (current->pi_blocked_on) {
2654                 /*
2655                  * We have been requeued or are in the process of
2656                  * being requeued.
2657                  */
2658                 raw_spin_unlock_irq(&current->pi_lock);
2659         } else {
2660                 /*
2661                  * Setting pi_blocked_on to PI_WAKEUP_INPROGRESS
2662                  * prevents a concurrent requeue from moving us to the
2663                  * uaddr2 rtmutex. After that we can safely acquire
2664                  * (and possibly block on) hb->lock.
2665                  */
2666                 current->pi_blocked_on = PI_WAKEUP_INPROGRESS;
2667                 raw_spin_unlock_irq(&current->pi_lock);
2668
2669                 spin_lock(&hb->lock);
2670
2671                 /*
2672                  * Clean up pi_blocked_on. We might leak it otherwise
2673                  * when we succeeded with the hb->lock in the fast
2674                  * path.
2675                  */
2676                 raw_spin_lock_irq(&current->pi_lock);
2677                 current->pi_blocked_on = NULL;
2678                 raw_spin_unlock_irq(&current->pi_lock);
2679
2680                 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2681                 spin_unlock(&hb->lock);
2682                 if (ret)
2683                         goto out_put_keys;
2684         }
2685
2686         /*
2687          * In order to be here, we have either been requeued, are in
2688          * the process of being requeued, or requeue successfully
2689          * acquired uaddr2 on our behalf.  If pi_blocked_on was
2690          * non-null above, we may be racing with a requeue.  Do not
2691          * rely on q->lock_ptr to be hb2->lock until after blocking on
2692          * hb->lock or hb2->lock. The futex_requeue dropped our key1
2693          * reference and incremented our key2 reference count.
2694          */
2695         hb2 = hash_futex(&key2);
2696
2697         /* Check if the requeue code acquired the second futex for us. */
2698         if (!q.rt_waiter) {
2699                 /*
2700                  * Got the lock. We might not be the anticipated owner if we
2701                  * did a lock-steal - fix up the PI-state in that case.
2702                  */
2703                 if (q.pi_state && (q.pi_state->owner != current)) {
2704                         spin_lock(&hb2->lock);
2705                         BUG_ON(&hb2->lock != q.lock_ptr);
2706                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2707                         spin_unlock(&hb2->lock);
2708                 }
2709         } else {
2710                 /*
2711                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2712                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2713                  * the pi_state.
2714                  */
2715                 WARN_ON(!q.pi_state);
2716                 pi_mutex = &q.pi_state->pi_mutex;
2717                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2718                 debug_rt_mutex_free_waiter(&rt_waiter);
2719
2720                 spin_lock(&hb2->lock);
2721                 BUG_ON(&hb2->lock != q.lock_ptr);
2722                 /*
2723                  * Fixup the pi_state owner and possibly acquire the lock if we
2724                  * haven't already.
2725                  */
2726                 res = fixup_owner(uaddr2, &q, !ret);
2727                 /*
2728                  * If fixup_owner() returned an error, proprogate that.  If it
2729                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2730                  */
2731                 if (res)
2732                         ret = (res < 0) ? res : 0;
2733
2734                 /* Unqueue and drop the lock. */
2735                 unqueue_me_pi(&q);
2736         }
2737
2738         /*
2739          * If fixup_pi_state_owner() faulted and was unable to handle the
2740          * fault, unlock the rt_mutex and return the fault to userspace.
2741          */
2742         if (ret == -EFAULT) {
2743                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2744                         rt_mutex_unlock(pi_mutex);
2745         } else if (ret == -EINTR) {
2746                 /*
2747                  * We've already been requeued, but cannot restart by calling
2748                  * futex_lock_pi() directly. We could restart this syscall, but
2749                  * it would detect that the user space "val" changed and return
2750                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2751                  * -EWOULDBLOCK directly.
2752                  */
2753                 ret = -EWOULDBLOCK;
2754         }
2755
2756 out_put_keys:
2757         put_futex_key(&q.key);
2758 out_key2:
2759         put_futex_key(&key2);
2760
2761 out:
2762         if (to) {
2763                 hrtimer_cancel(&to->timer);
2764                 destroy_hrtimer_on_stack(&to->timer);
2765         }
2766         return ret;
2767 }
2768
2769 /*
2770  * Support for robust futexes: the kernel cleans up held futexes at
2771  * thread exit time.
2772  *
2773  * Implementation: user-space maintains a per-thread list of locks it
2774  * is holding. Upon do_exit(), the kernel carefully walks this list,
2775  * and marks all locks that are owned by this thread with the
2776  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2777  * always manipulated with the lock held, so the list is private and
2778  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2779  * field, to allow the kernel to clean up if the thread dies after
2780  * acquiring the lock, but just before it could have added itself to
2781  * the list. There can only be one such pending lock.
2782  */
2783
2784 /**
2785  * sys_set_robust_list() - Set the robust-futex list head of a task
2786  * @head:       pointer to the list-head
2787  * @len:        length of the list-head, as userspace expects
2788  */
2789 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2790                 size_t, len)
2791 {
2792         if (!futex_cmpxchg_enabled)
2793                 return -ENOSYS;
2794         /*
2795          * The kernel knows only one size for now:
2796          */
2797         if (unlikely(len != sizeof(*head)))
2798                 return -EINVAL;
2799
2800         current->robust_list = head;
2801
2802         return 0;
2803 }
2804
2805 /**
2806  * sys_get_robust_list() - Get the robust-futex list head of a task
2807  * @pid:        pid of the process [zero for current task]
2808  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2809  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2810  */
2811 SYSCALL_DEFINE3(get_robust_list, int, pid,
2812                 struct robust_list_head __user * __user *, head_ptr,
2813                 size_t __user *, len_ptr)
2814 {
2815         struct robust_list_head __user *head;
2816         unsigned long ret;
2817         struct task_struct *p;
2818
2819         if (!futex_cmpxchg_enabled)
2820                 return -ENOSYS;
2821
2822         rcu_read_lock();
2823
2824         ret = -ESRCH;
2825         if (!pid)
2826                 p = current;
2827         else {
2828                 p = find_task_by_vpid(pid);
2829                 if (!p)
2830                         goto err_unlock;
2831         }
2832
2833         ret = -EPERM;
2834         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2835                 goto err_unlock;
2836
2837         head = p->robust_list;
2838         rcu_read_unlock();
2839
2840         if (put_user(sizeof(*head), len_ptr))
2841                 return -EFAULT;
2842         return put_user(head, head_ptr);
2843
2844 err_unlock:
2845         rcu_read_unlock();
2846
2847         return ret;
2848 }
2849
2850 /*
2851  * Process a futex-list entry, check whether it's owned by the
2852  * dying task, and do notification if so:
2853  */
2854 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2855 {
2856         u32 uval, uninitialized_var(nval), mval;
2857
2858 retry:
2859         if (get_user(uval, uaddr))
2860                 return -1;
2861
2862         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2863                 /*
2864                  * Ok, this dying thread is truly holding a futex
2865                  * of interest. Set the OWNER_DIED bit atomically
2866                  * via cmpxchg, and if the value had FUTEX_WAITERS
2867                  * set, wake up a waiter (if any). (We have to do a
2868                  * futex_wake() even if OWNER_DIED is already set -
2869                  * to handle the rare but possible case of recursive
2870                  * thread-death.) The rest of the cleanup is done in
2871                  * userspace.
2872                  */
2873                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2874                 /*
2875                  * We are not holding a lock here, but we want to have
2876                  * the pagefault_disable/enable() protection because
2877                  * we want to handle the fault gracefully. If the
2878                  * access fails we try to fault in the futex with R/W
2879                  * verification via get_user_pages. get_user() above
2880                  * does not guarantee R/W access. If that fails we
2881                  * give up and leave the futex locked.
2882                  */
2883                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2884                         if (fault_in_user_writeable(uaddr))
2885                                 return -1;
2886                         goto retry;
2887                 }
2888                 if (nval != uval)
2889                         goto retry;
2890
2891                 /*
2892                  * Wake robust non-PI futexes here. The wakeup of
2893                  * PI futexes happens in exit_pi_state():
2894                  */
2895                 if (!pi && (uval & FUTEX_WAITERS))
2896                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2897         }
2898         return 0;
2899 }
2900
2901 /*
2902  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2903  */
2904 static inline int fetch_robust_entry(struct robust_list __user **entry,
2905                                      struct robust_list __user * __user *head,
2906                                      unsigned int *pi)
2907 {
2908         unsigned long uentry;
2909
2910         if (get_user(uentry, (unsigned long __user *)head))
2911                 return -EFAULT;
2912
2913         *entry = (void __user *)(uentry & ~1UL);
2914         *pi = uentry & 1;
2915
2916         return 0;
2917 }
2918
2919 /*
2920  * Walk curr->robust_list (very carefully, it's a userspace list!)
2921  * and mark any locks found there dead, and notify any waiters.
2922  *
2923  * We silently return on any sign of list-walking problem.
2924  */
2925 void exit_robust_list(struct task_struct *curr)
2926 {
2927         struct robust_list_head __user *head = curr->robust_list;
2928         struct robust_list __user *entry, *next_entry, *pending;
2929         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2930         unsigned int uninitialized_var(next_pi);
2931         unsigned long futex_offset;
2932         int rc;
2933
2934         if (!futex_cmpxchg_enabled)
2935                 return;
2936
2937         /*
2938          * Fetch the list head (which was registered earlier, via
2939          * sys_set_robust_list()):
2940          */
2941         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2942                 return;
2943         /*
2944          * Fetch the relative futex offset:
2945          */
2946         if (get_user(futex_offset, &head->futex_offset))
2947                 return;
2948         /*
2949          * Fetch any possibly pending lock-add first, and handle it
2950          * if it exists:
2951          */
2952         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2953                 return;
2954
2955         next_entry = NULL;      /* avoid warning with gcc */
2956         while (entry != &head->list) {
2957                 /*
2958                  * Fetch the next entry in the list before calling
2959                  * handle_futex_death:
2960                  */
2961                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2962                 /*
2963                  * A pending lock might already be on the list, so
2964                  * don't process it twice:
2965                  */
2966                 if (entry != pending)
2967                         if (handle_futex_death((void __user *)entry + futex_offset,
2968                                                 curr, pi))
2969                                 return;
2970                 if (rc)
2971                         return;
2972                 entry = next_entry;
2973                 pi = next_pi;
2974                 /*
2975                  * Avoid excessively long or circular lists:
2976                  */
2977                 if (!--limit)
2978                         break;
2979
2980                 cond_resched();
2981         }
2982
2983         if (pending)
2984                 handle_futex_death((void __user *)pending + futex_offset,
2985                                    curr, pip);
2986 }
2987
2988 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2989                 u32 __user *uaddr2, u32 val2, u32 val3)
2990 {
2991         int cmd = op & FUTEX_CMD_MASK;
2992         unsigned int flags = 0;
2993
2994         if (!(op & FUTEX_PRIVATE_FLAG))
2995                 flags |= FLAGS_SHARED;
2996
2997         if (op & FUTEX_CLOCK_REALTIME) {
2998                 flags |= FLAGS_CLOCKRT;
2999                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3000                         return -ENOSYS;
3001         }
3002
3003         switch (cmd) {
3004         case FUTEX_LOCK_PI:
3005         case FUTEX_UNLOCK_PI:
3006         case FUTEX_TRYLOCK_PI:
3007         case FUTEX_WAIT_REQUEUE_PI:
3008         case FUTEX_CMP_REQUEUE_PI:
3009                 if (!futex_cmpxchg_enabled)
3010                         return -ENOSYS;
3011         }
3012
3013         switch (cmd) {
3014         case FUTEX_WAIT:
3015                 val3 = FUTEX_BITSET_MATCH_ANY;
3016         case FUTEX_WAIT_BITSET:
3017                 return futex_wait(uaddr, flags, val, timeout, val3);
3018         case FUTEX_WAKE:
3019                 val3 = FUTEX_BITSET_MATCH_ANY;
3020         case FUTEX_WAKE_BITSET:
3021                 return futex_wake(uaddr, flags, val, val3);
3022         case FUTEX_REQUEUE:
3023                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3024         case FUTEX_CMP_REQUEUE:
3025                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3026         case FUTEX_WAKE_OP:
3027                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3028         case FUTEX_LOCK_PI:
3029                 return futex_lock_pi(uaddr, flags, timeout, 0);
3030         case FUTEX_UNLOCK_PI:
3031                 return futex_unlock_pi(uaddr, flags);
3032         case FUTEX_TRYLOCK_PI:
3033                 return futex_lock_pi(uaddr, flags, NULL, 1);
3034         case FUTEX_WAIT_REQUEUE_PI:
3035                 val3 = FUTEX_BITSET_MATCH_ANY;
3036                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3037                                              uaddr2);
3038         case FUTEX_CMP_REQUEUE_PI:
3039                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3040         }
3041         return -ENOSYS;
3042 }
3043
3044
3045 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3046                 struct timespec __user *, utime, u32 __user *, uaddr2,
3047                 u32, val3)
3048 {
3049         struct timespec ts;
3050         ktime_t t, *tp = NULL;
3051         u32 val2 = 0;
3052         int cmd = op & FUTEX_CMD_MASK;
3053
3054         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3055                       cmd == FUTEX_WAIT_BITSET ||
3056                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3057                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3058                         return -EFAULT;
3059                 if (!timespec_valid(&ts))
3060                         return -EINVAL;
3061
3062                 t = timespec_to_ktime(ts);
3063                 if (cmd == FUTEX_WAIT)
3064                         t = ktime_add_safe(ktime_get(), t);
3065                 tp = &t;
3066         }
3067         /*
3068          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3069          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3070          */
3071         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3072             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3073                 val2 = (u32) (unsigned long) utime;
3074
3075         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3076 }
3077
3078 static void __init futex_detect_cmpxchg(void)
3079 {
3080 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3081         u32 curval;
3082
3083         /*
3084          * This will fail and we want it. Some arch implementations do
3085          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3086          * functionality. We want to know that before we call in any
3087          * of the complex code paths. Also we want to prevent
3088          * registration of robust lists in that case. NULL is
3089          * guaranteed to fault and we get -EFAULT on functional
3090          * implementation, the non-functional ones will return
3091          * -ENOSYS.
3092          */
3093         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3094                 futex_cmpxchg_enabled = 1;
3095 #endif
3096 }
3097
3098 static int __init futex_init(void)
3099 {
3100         unsigned int futex_shift;
3101         unsigned long i;
3102
3103 #if CONFIG_BASE_SMALL
3104         futex_hashsize = 16;
3105 #else
3106         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3107 #endif
3108
3109         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3110                                                futex_hashsize, 0,
3111                                                futex_hashsize < 256 ? HASH_SMALL : 0,
3112                                                &futex_shift, NULL,
3113                                                futex_hashsize, futex_hashsize);
3114         futex_hashsize = 1UL << futex_shift;
3115
3116         futex_detect_cmpxchg();
3117
3118         for (i = 0; i < futex_hashsize; i++) {
3119                 atomic_set(&futex_queues[i].waiters, 0);
3120                 plist_head_init(&futex_queues[i].chain);
3121                 spin_lock_init(&futex_queues[i].lock);
3122         }
3123
3124         return 0;
3125 }
3126 __initcall(futex_init);