futex: avoid race between requeue and wake
[platform/adaptation/renesas_rcar/renesas_kernel.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/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  * Basic futex operation and ordering guarantees:
74  *
75  * The waiter reads the futex value in user space and calls
76  * futex_wait(). This function computes the hash bucket and acquires
77  * the hash bucket lock. After that it reads the futex user space value
78  * again and verifies that the data has not changed. If it has not changed
79  * it enqueues itself into the hash bucket, releases the hash bucket lock
80  * and schedules.
81  *
82  * The waker side modifies the user space value of the futex and calls
83  * futex_wake(). This function computes the hash bucket and acquires the
84  * hash bucket lock. Then it looks for waiters on that futex in the hash
85  * bucket and wakes them.
86  *
87  * In futex wake up scenarios where no tasks are blocked on a futex, taking
88  * the hb spinlock can be avoided and simply return. In order for this
89  * optimization to work, ordering guarantees must exist so that the waiter
90  * being added to the list is acknowledged when the list is concurrently being
91  * checked by the waker, avoiding scenarios like the following:
92  *
93  * CPU 0                               CPU 1
94  * val = *futex;
95  * sys_futex(WAIT, futex, val);
96  *   futex_wait(futex, val);
97  *   uval = *futex;
98  *                                     *futex = newval;
99  *                                     sys_futex(WAKE, futex);
100  *                                       futex_wake(futex);
101  *                                       if (queue_empty())
102  *                                         return;
103  *   if (uval == val)
104  *      lock(hash_bucket(futex));
105  *      queue();
106  *     unlock(hash_bucket(futex));
107  *     schedule();
108  *
109  * This would cause the waiter on CPU 0 to wait forever because it
110  * missed the transition of the user space value from val to newval
111  * and the waker did not find the waiter in the hash bucket queue.
112  *
113  * The correct serialization ensures that a waiter either observes
114  * the changed user space value before blocking or is woken by a
115  * concurrent waker:
116  *
117  * CPU 0                                 CPU 1
118  * val = *futex;
119  * sys_futex(WAIT, futex, val);
120  *   futex_wait(futex, val);
121  *
122  *   waiters++;
123  *   mb(); (A) <-- paired with -.
124  *                              |
125  *   lock(hash_bucket(futex));  |
126  *                              |
127  *   uval = *futex;             |
128  *                              |        *futex = newval;
129  *                              |        sys_futex(WAKE, futex);
130  *                              |          futex_wake(futex);
131  *                              |
132  *                              `------->  mb(); (B)
133  *   if (uval == val)
134  *     queue();
135  *     unlock(hash_bucket(futex));
136  *     schedule();                         if (waiters)
137  *                                           lock(hash_bucket(futex));
138  *                                           wake_waiters(futex);
139  *                                           unlock(hash_bucket(futex));
140  *
141  * Where (A) orders the waiters increment and the futex value read -- this
142  * is guaranteed by the head counter in the hb spinlock; and where (B)
143  * orders the write to futex and the waiters read -- this is done by the
144  * barriers in get_futex_key_refs(), through either ihold or atomic_inc,
145  * depending on the futex type.
146  *
147  * This yields the following case (where X:=waiters, Y:=futex):
148  *
149  *      X = Y = 0
150  *
151  *      w[X]=1          w[Y]=1
152  *      MB              MB
153  *      r[Y]=y          r[X]=x
154  *
155  * Which guarantees that x==0 && y==0 is impossible; which translates back into
156  * the guarantee that we cannot both miss the futex variable change and the
157  * enqueue.
158  */
159
160 int __read_mostly futex_cmpxchg_enabled;
161
162 /*
163  * Futex flags used to encode options to functions and preserve them across
164  * restarts.
165  */
166 #define FLAGS_SHARED            0x01
167 #define FLAGS_CLOCKRT           0x02
168 #define FLAGS_HAS_TIMEOUT       0x04
169
170 /*
171  * Priority Inheritance state:
172  */
173 struct futex_pi_state {
174         /*
175          * list of 'owned' pi_state instances - these have to be
176          * cleaned up in do_exit() if the task exits prematurely:
177          */
178         struct list_head list;
179
180         /*
181          * The PI object:
182          */
183         struct rt_mutex pi_mutex;
184
185         struct task_struct *owner;
186         atomic_t refcount;
187
188         union futex_key key;
189 };
190
191 /**
192  * struct futex_q - The hashed futex queue entry, one per waiting task
193  * @list:               priority-sorted list of tasks waiting on this futex
194  * @task:               the task waiting on the futex
195  * @lock_ptr:           the hash bucket lock
196  * @key:                the key the futex is hashed on
197  * @pi_state:           optional priority inheritance state
198  * @rt_waiter:          rt_waiter storage for use with requeue_pi
199  * @requeue_pi_key:     the requeue_pi target futex key
200  * @bitset:             bitset for the optional bitmasked wakeup
201  *
202  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
203  * we can wake only the relevant ones (hashed queues may be shared).
204  *
205  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
206  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
207  * The order of wakeup is always to make the first condition true, then
208  * the second.
209  *
210  * PI futexes are typically woken before they are removed from the hash list via
211  * the rt_mutex code. See unqueue_me_pi().
212  */
213 struct futex_q {
214         struct plist_node list;
215
216         struct task_struct *task;
217         spinlock_t *lock_ptr;
218         union futex_key key;
219         struct futex_pi_state *pi_state;
220         struct rt_mutex_waiter *rt_waiter;
221         union futex_key *requeue_pi_key;
222         u32 bitset;
223 };
224
225 static const struct futex_q futex_q_init = {
226         /* list gets initialized in queue_me()*/
227         .key = FUTEX_KEY_INIT,
228         .bitset = FUTEX_BITSET_MATCH_ANY
229 };
230
231 /*
232  * Hash buckets are shared by all the futex_keys that hash to the same
233  * location.  Each key may have multiple futex_q structures, one for each task
234  * waiting on a futex.
235  */
236 struct futex_hash_bucket {
237         atomic_t waiters;
238         spinlock_t lock;
239         struct plist_head chain;
240 } ____cacheline_aligned_in_smp;
241
242 static unsigned long __read_mostly futex_hashsize;
243
244 static struct futex_hash_bucket *futex_queues;
245
246 static inline void futex_get_mm(union futex_key *key)
247 {
248         atomic_inc(&key->private.mm->mm_count);
249         /*
250          * Ensure futex_get_mm() implies a full barrier such that
251          * get_futex_key() implies a full barrier. This is relied upon
252          * as full barrier (B), see the ordering comment above.
253          */
254         smp_mb__after_atomic_inc();
255 }
256
257 /*
258  * Reflects a new waiter being added to the waitqueue.
259  */
260 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
261 {
262 #ifdef CONFIG_SMP
263         atomic_inc(&hb->waiters);
264         /*
265          * Full barrier (A), see the ordering comment above.
266          */
267         smp_mb__after_atomic_inc();
268 #endif
269 }
270
271 /*
272  * Reflects a waiter being removed from the waitqueue by wakeup
273  * paths.
274  */
275 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
276 {
277 #ifdef CONFIG_SMP
278         atomic_dec(&hb->waiters);
279 #endif
280 }
281
282 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
283 {
284 #ifdef CONFIG_SMP
285         return atomic_read(&hb->waiters);
286 #else
287         return 1;
288 #endif
289 }
290
291 /*
292  * We hash on the keys returned from get_futex_key (see below).
293  */
294 static struct futex_hash_bucket *hash_futex(union futex_key *key)
295 {
296         u32 hash = jhash2((u32*)&key->both.word,
297                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
298                           key->both.offset);
299         return &futex_queues[hash & (futex_hashsize - 1)];
300 }
301
302 /*
303  * Return 1 if two futex_keys are equal, 0 otherwise.
304  */
305 static inline int match_futex(union futex_key *key1, union futex_key *key2)
306 {
307         return (key1 && key2
308                 && key1->both.word == key2->both.word
309                 && key1->both.ptr == key2->both.ptr
310                 && key1->both.offset == key2->both.offset);
311 }
312
313 /*
314  * Take a reference to the resource addressed by a key.
315  * Can be called while holding spinlocks.
316  *
317  */
318 static void get_futex_key_refs(union futex_key *key)
319 {
320         if (!key->both.ptr)
321                 return;
322
323         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
324         case FUT_OFF_INODE:
325                 ihold(key->shared.inode); /* implies MB (B) */
326                 break;
327         case FUT_OFF_MMSHARED:
328                 futex_get_mm(key); /* implies MB (B) */
329                 break;
330         }
331 }
332
333 /*
334  * Drop a reference to the resource addressed by a key.
335  * The hash bucket spinlock must not be held.
336  */
337 static void drop_futex_key_refs(union futex_key *key)
338 {
339         if (!key->both.ptr) {
340                 /* If we're here then we tried to put a key we failed to get */
341                 WARN_ON_ONCE(1);
342                 return;
343         }
344
345         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
346         case FUT_OFF_INODE:
347                 iput(key->shared.inode);
348                 break;
349         case FUT_OFF_MMSHARED:
350                 mmdrop(key->private.mm);
351                 break;
352         }
353 }
354
355 /**
356  * get_futex_key() - Get parameters which are the keys for a futex
357  * @uaddr:      virtual address of the futex
358  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
359  * @key:        address where result is stored.
360  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
361  *              VERIFY_WRITE)
362  *
363  * Return: a negative error code or 0
364  *
365  * The key words are stored in *key on success.
366  *
367  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
368  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
369  * We can usually work out the index without swapping in the page.
370  *
371  * lock_page() might sleep, the caller should not hold a spinlock.
372  */
373 static int
374 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
375 {
376         unsigned long address = (unsigned long)uaddr;
377         struct mm_struct *mm = current->mm;
378         struct page *page, *page_head;
379         int err, ro = 0;
380
381         /*
382          * The futex address must be "naturally" aligned.
383          */
384         key->both.offset = address % PAGE_SIZE;
385         if (unlikely((address % sizeof(u32)) != 0))
386                 return -EINVAL;
387         address -= key->both.offset;
388
389         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
390                 return -EFAULT;
391
392         /*
393          * PROCESS_PRIVATE futexes are fast.
394          * As the mm cannot disappear under us and the 'key' only needs
395          * virtual address, we dont even have to find the underlying vma.
396          * Note : We do have to check 'uaddr' is a valid user address,
397          *        but access_ok() should be faster than find_vma()
398          */
399         if (!fshared) {
400                 key->private.mm = mm;
401                 key->private.address = address;
402                 get_futex_key_refs(key);  /* implies MB (B) */
403                 return 0;
404         }
405
406 again:
407         err = get_user_pages_fast(address, 1, 1, &page);
408         /*
409          * If write access is not required (eg. FUTEX_WAIT), try
410          * and get read-only access.
411          */
412         if (err == -EFAULT && rw == VERIFY_READ) {
413                 err = get_user_pages_fast(address, 1, 0, &page);
414                 ro = 1;
415         }
416         if (err < 0)
417                 return err;
418         else
419                 err = 0;
420
421 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
422         page_head = page;
423         if (unlikely(PageTail(page))) {
424                 put_page(page);
425                 /* serialize against __split_huge_page_splitting() */
426                 local_irq_disable();
427                 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
428                         page_head = compound_head(page);
429                         /*
430                          * page_head is valid pointer but we must pin
431                          * it before taking the PG_lock and/or
432                          * PG_compound_lock. The moment we re-enable
433                          * irqs __split_huge_page_splitting() can
434                          * return and the head page can be freed from
435                          * under us. We can't take the PG_lock and/or
436                          * PG_compound_lock on a page that could be
437                          * freed from under us.
438                          */
439                         if (page != page_head) {
440                                 get_page(page_head);
441                                 put_page(page);
442                         }
443                         local_irq_enable();
444                 } else {
445                         local_irq_enable();
446                         goto again;
447                 }
448         }
449 #else
450         page_head = compound_head(page);
451         if (page != page_head) {
452                 get_page(page_head);
453                 put_page(page);
454         }
455 #endif
456
457         lock_page(page_head);
458
459         /*
460          * If page_head->mapping is NULL, then it cannot be a PageAnon
461          * page; but it might be the ZERO_PAGE or in the gate area or
462          * in a special mapping (all cases which we are happy to fail);
463          * or it may have been a good file page when get_user_pages_fast
464          * found it, but truncated or holepunched or subjected to
465          * invalidate_complete_page2 before we got the page lock (also
466          * cases which we are happy to fail).  And we hold a reference,
467          * so refcount care in invalidate_complete_page's remove_mapping
468          * prevents drop_caches from setting mapping to NULL beneath us.
469          *
470          * The case we do have to guard against is when memory pressure made
471          * shmem_writepage move it from filecache to swapcache beneath us:
472          * an unlikely race, but we do need to retry for page_head->mapping.
473          */
474         if (!page_head->mapping) {
475                 int shmem_swizzled = PageSwapCache(page_head);
476                 unlock_page(page_head);
477                 put_page(page_head);
478                 if (shmem_swizzled)
479                         goto again;
480                 return -EFAULT;
481         }
482
483         /*
484          * Private mappings are handled in a simple way.
485          *
486          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
487          * it's a read-only handle, it's expected that futexes attach to
488          * the object not the particular process.
489          */
490         if (PageAnon(page_head)) {
491                 /*
492                  * A RO anonymous page will never change and thus doesn't make
493                  * sense for futex operations.
494                  */
495                 if (ro) {
496                         err = -EFAULT;
497                         goto out;
498                 }
499
500                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
501                 key->private.mm = mm;
502                 key->private.address = address;
503         } else {
504                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
505                 key->shared.inode = page_head->mapping->host;
506                 key->shared.pgoff = basepage_index(page);
507         }
508
509         get_futex_key_refs(key); /* implies MB (B) */
510
511 out:
512         unlock_page(page_head);
513         put_page(page_head);
514         return err;
515 }
516
517 static inline void put_futex_key(union futex_key *key)
518 {
519         drop_futex_key_refs(key);
520 }
521
522 /**
523  * fault_in_user_writeable() - Fault in user address and verify RW access
524  * @uaddr:      pointer to faulting user space address
525  *
526  * Slow path to fixup the fault we just took in the atomic write
527  * access to @uaddr.
528  *
529  * We have no generic implementation of a non-destructive write to the
530  * user address. We know that we faulted in the atomic pagefault
531  * disabled section so we can as well avoid the #PF overhead by
532  * calling get_user_pages() right away.
533  */
534 static int fault_in_user_writeable(u32 __user *uaddr)
535 {
536         struct mm_struct *mm = current->mm;
537         int ret;
538
539         down_read(&mm->mmap_sem);
540         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
541                                FAULT_FLAG_WRITE);
542         up_read(&mm->mmap_sem);
543
544         return ret < 0 ? ret : 0;
545 }
546
547 /**
548  * futex_top_waiter() - Return the highest priority waiter on a futex
549  * @hb:         the hash bucket the futex_q's reside in
550  * @key:        the futex key (to distinguish it from other futex futex_q's)
551  *
552  * Must be called with the hb lock held.
553  */
554 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
555                                         union futex_key *key)
556 {
557         struct futex_q *this;
558
559         plist_for_each_entry(this, &hb->chain, list) {
560                 if (match_futex(&this->key, key))
561                         return this;
562         }
563         return NULL;
564 }
565
566 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
567                                       u32 uval, u32 newval)
568 {
569         int ret;
570
571         pagefault_disable();
572         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
573         pagefault_enable();
574
575         return ret;
576 }
577
578 static int get_futex_value_locked(u32 *dest, u32 __user *from)
579 {
580         int ret;
581
582         pagefault_disable();
583         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
584         pagefault_enable();
585
586         return ret ? -EFAULT : 0;
587 }
588
589
590 /*
591  * PI code:
592  */
593 static int refill_pi_state_cache(void)
594 {
595         struct futex_pi_state *pi_state;
596
597         if (likely(current->pi_state_cache))
598                 return 0;
599
600         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
601
602         if (!pi_state)
603                 return -ENOMEM;
604
605         INIT_LIST_HEAD(&pi_state->list);
606         /* pi_mutex gets initialized later */
607         pi_state->owner = NULL;
608         atomic_set(&pi_state->refcount, 1);
609         pi_state->key = FUTEX_KEY_INIT;
610
611         current->pi_state_cache = pi_state;
612
613         return 0;
614 }
615
616 static struct futex_pi_state * alloc_pi_state(void)
617 {
618         struct futex_pi_state *pi_state = current->pi_state_cache;
619
620         WARN_ON(!pi_state);
621         current->pi_state_cache = NULL;
622
623         return pi_state;
624 }
625
626 static void free_pi_state(struct futex_pi_state *pi_state)
627 {
628         if (!atomic_dec_and_test(&pi_state->refcount))
629                 return;
630
631         /*
632          * If pi_state->owner is NULL, the owner is most probably dying
633          * and has cleaned up the pi_state already
634          */
635         if (pi_state->owner) {
636                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
637                 list_del_init(&pi_state->list);
638                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
639
640                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
641         }
642
643         if (current->pi_state_cache)
644                 kfree(pi_state);
645         else {
646                 /*
647                  * pi_state->list is already empty.
648                  * clear pi_state->owner.
649                  * refcount is at 0 - put it back to 1.
650                  */
651                 pi_state->owner = NULL;
652                 atomic_set(&pi_state->refcount, 1);
653                 current->pi_state_cache = pi_state;
654         }
655 }
656
657 /*
658  * Look up the task based on what TID userspace gave us.
659  * We dont trust it.
660  */
661 static struct task_struct * futex_find_get_task(pid_t pid)
662 {
663         struct task_struct *p;
664
665         rcu_read_lock();
666         p = find_task_by_vpid(pid);
667         if (p)
668                 get_task_struct(p);
669
670         rcu_read_unlock();
671
672         return p;
673 }
674
675 /*
676  * This task is holding PI mutexes at exit time => bad.
677  * Kernel cleans up PI-state, but userspace is likely hosed.
678  * (Robust-futex cleanup is separate and might save the day for userspace.)
679  */
680 void exit_pi_state_list(struct task_struct *curr)
681 {
682         struct list_head *next, *head = &curr->pi_state_list;
683         struct futex_pi_state *pi_state;
684         struct futex_hash_bucket *hb;
685         union futex_key key = FUTEX_KEY_INIT;
686
687         if (!futex_cmpxchg_enabled)
688                 return;
689         /*
690          * We are a ZOMBIE and nobody can enqueue itself on
691          * pi_state_list anymore, but we have to be careful
692          * versus waiters unqueueing themselves:
693          */
694         raw_spin_lock_irq(&curr->pi_lock);
695         while (!list_empty(head)) {
696
697                 next = head->next;
698                 pi_state = list_entry(next, struct futex_pi_state, list);
699                 key = pi_state->key;
700                 hb = hash_futex(&key);
701                 raw_spin_unlock_irq(&curr->pi_lock);
702
703                 spin_lock(&hb->lock);
704
705                 raw_spin_lock_irq(&curr->pi_lock);
706                 /*
707                  * We dropped the pi-lock, so re-check whether this
708                  * task still owns the PI-state:
709                  */
710                 if (head->next != next) {
711                         spin_unlock(&hb->lock);
712                         continue;
713                 }
714
715                 WARN_ON(pi_state->owner != curr);
716                 WARN_ON(list_empty(&pi_state->list));
717                 list_del_init(&pi_state->list);
718                 pi_state->owner = NULL;
719                 raw_spin_unlock_irq(&curr->pi_lock);
720
721                 rt_mutex_unlock(&pi_state->pi_mutex);
722
723                 spin_unlock(&hb->lock);
724
725                 raw_spin_lock_irq(&curr->pi_lock);
726         }
727         raw_spin_unlock_irq(&curr->pi_lock);
728 }
729
730 static int
731 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
732                 union futex_key *key, struct futex_pi_state **ps)
733 {
734         struct futex_pi_state *pi_state = NULL;
735         struct futex_q *this, *next;
736         struct task_struct *p;
737         pid_t pid = uval & FUTEX_TID_MASK;
738
739         plist_for_each_entry_safe(this, next, &hb->chain, list) {
740                 if (match_futex(&this->key, key)) {
741                         /*
742                          * Another waiter already exists - bump up
743                          * the refcount and return its pi_state:
744                          */
745                         pi_state = this->pi_state;
746                         /*
747                          * Userspace might have messed up non-PI and PI futexes
748                          */
749                         if (unlikely(!pi_state))
750                                 return -EINVAL;
751
752                         WARN_ON(!atomic_read(&pi_state->refcount));
753
754                         /*
755                          * When pi_state->owner is NULL then the owner died
756                          * and another waiter is on the fly. pi_state->owner
757                          * is fixed up by the task which acquires
758                          * pi_state->rt_mutex.
759                          *
760                          * We do not check for pid == 0 which can happen when
761                          * the owner died and robust_list_exit() cleared the
762                          * TID.
763                          */
764                         if (pid && pi_state->owner) {
765                                 /*
766                                  * Bail out if user space manipulated the
767                                  * futex value.
768                                  */
769                                 if (pid != task_pid_vnr(pi_state->owner))
770                                         return -EINVAL;
771                         }
772
773                         atomic_inc(&pi_state->refcount);
774                         *ps = pi_state;
775
776                         return 0;
777                 }
778         }
779
780         /*
781          * We are the first waiter - try to look up the real owner and attach
782          * the new pi_state to it, but bail out when TID = 0
783          */
784         if (!pid)
785                 return -ESRCH;
786         p = futex_find_get_task(pid);
787         if (!p)
788                 return -ESRCH;
789
790         /*
791          * We need to look at the task state flags to figure out,
792          * whether the task is exiting. To protect against the do_exit
793          * change of the task flags, we do this protected by
794          * p->pi_lock:
795          */
796         raw_spin_lock_irq(&p->pi_lock);
797         if (unlikely(p->flags & PF_EXITING)) {
798                 /*
799                  * The task is on the way out. When PF_EXITPIDONE is
800                  * set, we know that the task has finished the
801                  * cleanup:
802                  */
803                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
804
805                 raw_spin_unlock_irq(&p->pi_lock);
806                 put_task_struct(p);
807                 return ret;
808         }
809
810         pi_state = alloc_pi_state();
811
812         /*
813          * Initialize the pi_mutex in locked state and make 'p'
814          * the owner of it:
815          */
816         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
817
818         /* Store the key for possible exit cleanups: */
819         pi_state->key = *key;
820
821         WARN_ON(!list_empty(&pi_state->list));
822         list_add(&pi_state->list, &p->pi_state_list);
823         pi_state->owner = p;
824         raw_spin_unlock_irq(&p->pi_lock);
825
826         put_task_struct(p);
827
828         *ps = pi_state;
829
830         return 0;
831 }
832
833 /**
834  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
835  * @uaddr:              the pi futex user address
836  * @hb:                 the pi futex hash bucket
837  * @key:                the futex key associated with uaddr and hb
838  * @ps:                 the pi_state pointer where we store the result of the
839  *                      lookup
840  * @task:               the task to perform the atomic lock work for.  This will
841  *                      be "current" except in the case of requeue pi.
842  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
843  *
844  * Return:
845  *  0 - ready to wait;
846  *  1 - acquired the lock;
847  * <0 - error
848  *
849  * The hb->lock and futex_key refs shall be held by the caller.
850  */
851 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
852                                 union futex_key *key,
853                                 struct futex_pi_state **ps,
854                                 struct task_struct *task, int set_waiters)
855 {
856         int lock_taken, ret, force_take = 0;
857         u32 uval, newval, curval, vpid = task_pid_vnr(task);
858
859 retry:
860         ret = lock_taken = 0;
861
862         /*
863          * To avoid races, we attempt to take the lock here again
864          * (by doing a 0 -> TID atomic cmpxchg), while holding all
865          * the locks. It will most likely not succeed.
866          */
867         newval = vpid;
868         if (set_waiters)
869                 newval |= FUTEX_WAITERS;
870
871         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
872                 return -EFAULT;
873
874         /*
875          * Detect deadlocks.
876          */
877         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
878                 return -EDEADLK;
879
880         /*
881          * Surprise - we got the lock. Just return to userspace:
882          */
883         if (unlikely(!curval))
884                 return 1;
885
886         uval = curval;
887
888         /*
889          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
890          * to wake at the next unlock.
891          */
892         newval = curval | FUTEX_WAITERS;
893
894         /*
895          * Should we force take the futex? See below.
896          */
897         if (unlikely(force_take)) {
898                 /*
899                  * Keep the OWNER_DIED and the WAITERS bit and set the
900                  * new TID value.
901                  */
902                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
903                 force_take = 0;
904                 lock_taken = 1;
905         }
906
907         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
908                 return -EFAULT;
909         if (unlikely(curval != uval))
910                 goto retry;
911
912         /*
913          * We took the lock due to forced take over.
914          */
915         if (unlikely(lock_taken))
916                 return 1;
917
918         /*
919          * We dont have the lock. Look up the PI state (or create it if
920          * we are the first waiter):
921          */
922         ret = lookup_pi_state(uval, hb, key, ps);
923
924         if (unlikely(ret)) {
925                 switch (ret) {
926                 case -ESRCH:
927                         /*
928                          * We failed to find an owner for this
929                          * futex. So we have no pi_state to block
930                          * on. This can happen in two cases:
931                          *
932                          * 1) The owner died
933                          * 2) A stale FUTEX_WAITERS bit
934                          *
935                          * Re-read the futex value.
936                          */
937                         if (get_futex_value_locked(&curval, uaddr))
938                                 return -EFAULT;
939
940                         /*
941                          * If the owner died or we have a stale
942                          * WAITERS bit the owner TID in the user space
943                          * futex is 0.
944                          */
945                         if (!(curval & FUTEX_TID_MASK)) {
946                                 force_take = 1;
947                                 goto retry;
948                         }
949                 default:
950                         break;
951                 }
952         }
953
954         return ret;
955 }
956
957 /**
958  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
959  * @q:  The futex_q to unqueue
960  *
961  * The q->lock_ptr must not be NULL and must be held by the caller.
962  */
963 static void __unqueue_futex(struct futex_q *q)
964 {
965         struct futex_hash_bucket *hb;
966
967         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
968             || WARN_ON(plist_node_empty(&q->list)))
969                 return;
970
971         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
972         plist_del(&q->list, &hb->chain);
973         hb_waiters_dec(hb);
974 }
975
976 /*
977  * The hash bucket lock must be held when this is called.
978  * Afterwards, the futex_q must not be accessed.
979  */
980 static void wake_futex(struct futex_q *q)
981 {
982         struct task_struct *p = q->task;
983
984         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
985                 return;
986
987         /*
988          * We set q->lock_ptr = NULL _before_ we wake up the task. If
989          * a non-futex wake up happens on another CPU then the task
990          * might exit and p would dereference a non-existing task
991          * struct. Prevent this by holding a reference on p across the
992          * wake up.
993          */
994         get_task_struct(p);
995
996         __unqueue_futex(q);
997         /*
998          * The waiting task can free the futex_q as soon as
999          * q->lock_ptr = NULL is written, without taking any locks. A
1000          * memory barrier is required here to prevent the following
1001          * store to lock_ptr from getting ahead of the plist_del.
1002          */
1003         smp_wmb();
1004         q->lock_ptr = NULL;
1005
1006         wake_up_state(p, TASK_NORMAL);
1007         put_task_struct(p);
1008 }
1009
1010 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1011 {
1012         struct task_struct *new_owner;
1013         struct futex_pi_state *pi_state = this->pi_state;
1014         u32 uninitialized_var(curval), newval;
1015
1016         if (!pi_state)
1017                 return -EINVAL;
1018
1019         /*
1020          * If current does not own the pi_state then the futex is
1021          * inconsistent and user space fiddled with the futex value.
1022          */
1023         if (pi_state->owner != current)
1024                 return -EINVAL;
1025
1026         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1027         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1028
1029         /*
1030          * It is possible that the next waiter (the one that brought
1031          * this owner to the kernel) timed out and is no longer
1032          * waiting on the lock.
1033          */
1034         if (!new_owner)
1035                 new_owner = this->task;
1036
1037         /*
1038          * We pass it to the next owner. (The WAITERS bit is always
1039          * kept enabled while there is PI state around. We must also
1040          * preserve the owner died bit.)
1041          */
1042         if (!(uval & FUTEX_OWNER_DIED)) {
1043                 int ret = 0;
1044
1045                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1046
1047                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1048                         ret = -EFAULT;
1049                 else if (curval != uval)
1050                         ret = -EINVAL;
1051                 if (ret) {
1052                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1053                         return ret;
1054                 }
1055         }
1056
1057         raw_spin_lock_irq(&pi_state->owner->pi_lock);
1058         WARN_ON(list_empty(&pi_state->list));
1059         list_del_init(&pi_state->list);
1060         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1061
1062         raw_spin_lock_irq(&new_owner->pi_lock);
1063         WARN_ON(!list_empty(&pi_state->list));
1064         list_add(&pi_state->list, &new_owner->pi_state_list);
1065         pi_state->owner = new_owner;
1066         raw_spin_unlock_irq(&new_owner->pi_lock);
1067
1068         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1069         rt_mutex_unlock(&pi_state->pi_mutex);
1070
1071         return 0;
1072 }
1073
1074 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1075 {
1076         u32 uninitialized_var(oldval);
1077
1078         /*
1079          * There is no waiter, so we unlock the futex. The owner died
1080          * bit has not to be preserved here. We are the owner:
1081          */
1082         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1083                 return -EFAULT;
1084         if (oldval != uval)
1085                 return -EAGAIN;
1086
1087         return 0;
1088 }
1089
1090 /*
1091  * Express the locking dependencies for lockdep:
1092  */
1093 static inline void
1094 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1095 {
1096         if (hb1 <= hb2) {
1097                 spin_lock(&hb1->lock);
1098                 if (hb1 < hb2)
1099                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1100         } else { /* hb1 > hb2 */
1101                 spin_lock(&hb2->lock);
1102                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1103         }
1104 }
1105
1106 static inline void
1107 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1108 {
1109         spin_unlock(&hb1->lock);
1110         if (hb1 != hb2)
1111                 spin_unlock(&hb2->lock);
1112 }
1113
1114 /*
1115  * Wake up waiters matching bitset queued on this futex (uaddr).
1116  */
1117 static int
1118 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1119 {
1120         struct futex_hash_bucket *hb;
1121         struct futex_q *this, *next;
1122         union futex_key key = FUTEX_KEY_INIT;
1123         int ret;
1124
1125         if (!bitset)
1126                 return -EINVAL;
1127
1128         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1129         if (unlikely(ret != 0))
1130                 goto out;
1131
1132         hb = hash_futex(&key);
1133
1134         /* Make sure we really have tasks to wakeup */
1135         if (!hb_waiters_pending(hb))
1136                 goto out_put_key;
1137
1138         spin_lock(&hb->lock);
1139
1140         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1141                 if (match_futex (&this->key, &key)) {
1142                         if (this->pi_state || this->rt_waiter) {
1143                                 ret = -EINVAL;
1144                                 break;
1145                         }
1146
1147                         /* Check if one of the bits is set in both bitsets */
1148                         if (!(this->bitset & bitset))
1149                                 continue;
1150
1151                         wake_futex(this);
1152                         if (++ret >= nr_wake)
1153                                 break;
1154                 }
1155         }
1156
1157         spin_unlock(&hb->lock);
1158 out_put_key:
1159         put_futex_key(&key);
1160 out:
1161         return ret;
1162 }
1163
1164 /*
1165  * Wake up all waiters hashed on the physical page that is mapped
1166  * to this virtual address:
1167  */
1168 static int
1169 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1170               int nr_wake, int nr_wake2, int op)
1171 {
1172         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1173         struct futex_hash_bucket *hb1, *hb2;
1174         struct futex_q *this, *next;
1175         int ret, op_ret;
1176
1177 retry:
1178         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1179         if (unlikely(ret != 0))
1180                 goto out;
1181         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1182         if (unlikely(ret != 0))
1183                 goto out_put_key1;
1184
1185         hb1 = hash_futex(&key1);
1186         hb2 = hash_futex(&key2);
1187
1188 retry_private:
1189         double_lock_hb(hb1, hb2);
1190         op_ret = futex_atomic_op_inuser(op, uaddr2);
1191         if (unlikely(op_ret < 0)) {
1192
1193                 double_unlock_hb(hb1, hb2);
1194
1195 #ifndef CONFIG_MMU
1196                 /*
1197                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1198                  * but we might get them from range checking
1199                  */
1200                 ret = op_ret;
1201                 goto out_put_keys;
1202 #endif
1203
1204                 if (unlikely(op_ret != -EFAULT)) {
1205                         ret = op_ret;
1206                         goto out_put_keys;
1207                 }
1208
1209                 ret = fault_in_user_writeable(uaddr2);
1210                 if (ret)
1211                         goto out_put_keys;
1212
1213                 if (!(flags & FLAGS_SHARED))
1214                         goto retry_private;
1215
1216                 put_futex_key(&key2);
1217                 put_futex_key(&key1);
1218                 goto retry;
1219         }
1220
1221         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1222                 if (match_futex (&this->key, &key1)) {
1223                         if (this->pi_state || this->rt_waiter) {
1224                                 ret = -EINVAL;
1225                                 goto out_unlock;
1226                         }
1227                         wake_futex(this);
1228                         if (++ret >= nr_wake)
1229                                 break;
1230                 }
1231         }
1232
1233         if (op_ret > 0) {
1234                 op_ret = 0;
1235                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1236                         if (match_futex (&this->key, &key2)) {
1237                                 if (this->pi_state || this->rt_waiter) {
1238                                         ret = -EINVAL;
1239                                         goto out_unlock;
1240                                 }
1241                                 wake_futex(this);
1242                                 if (++op_ret >= nr_wake2)
1243                                         break;
1244                         }
1245                 }
1246                 ret += op_ret;
1247         }
1248
1249 out_unlock:
1250         double_unlock_hb(hb1, hb2);
1251 out_put_keys:
1252         put_futex_key(&key2);
1253 out_put_key1:
1254         put_futex_key(&key1);
1255 out:
1256         return ret;
1257 }
1258
1259 /**
1260  * requeue_futex() - Requeue a futex_q from one hb to another
1261  * @q:          the futex_q to requeue
1262  * @hb1:        the source hash_bucket
1263  * @hb2:        the target hash_bucket
1264  * @key2:       the new key for the requeued futex_q
1265  */
1266 static inline
1267 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1268                    struct futex_hash_bucket *hb2, union futex_key *key2)
1269 {
1270
1271         /*
1272          * If key1 and key2 hash to the same bucket, no need to
1273          * requeue.
1274          */
1275         if (likely(&hb1->chain != &hb2->chain)) {
1276                 plist_del(&q->list, &hb1->chain);
1277                 hb_waiters_dec(hb1);
1278                 plist_add(&q->list, &hb2->chain);
1279                 hb_waiters_inc(hb2);
1280                 q->lock_ptr = &hb2->lock;
1281         }
1282         get_futex_key_refs(key2);
1283         q->key = *key2;
1284 }
1285
1286 /**
1287  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1288  * @q:          the futex_q
1289  * @key:        the key of the requeue target futex
1290  * @hb:         the hash_bucket of the requeue target futex
1291  *
1292  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1293  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1294  * to the requeue target futex so the waiter can detect the wakeup on the right
1295  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1296  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1297  * to protect access to the pi_state to fixup the owner later.  Must be called
1298  * with both q->lock_ptr and hb->lock held.
1299  */
1300 static inline
1301 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1302                            struct futex_hash_bucket *hb)
1303 {
1304         get_futex_key_refs(key);
1305         q->key = *key;
1306
1307         __unqueue_futex(q);
1308
1309         WARN_ON(!q->rt_waiter);
1310         q->rt_waiter = NULL;
1311
1312         q->lock_ptr = &hb->lock;
1313
1314         wake_up_state(q->task, TASK_NORMAL);
1315 }
1316
1317 /**
1318  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1319  * @pifutex:            the user address of the to futex
1320  * @hb1:                the from futex hash bucket, must be locked by the caller
1321  * @hb2:                the to futex hash bucket, must be locked by the caller
1322  * @key1:               the from futex key
1323  * @key2:               the to futex key
1324  * @ps:                 address to store the pi_state pointer
1325  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1326  *
1327  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1328  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1329  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1330  * hb1 and hb2 must be held by the caller.
1331  *
1332  * Return:
1333  *  0 - failed to acquire the lock atomically;
1334  *  1 - acquired the lock;
1335  * <0 - error
1336  */
1337 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1338                                  struct futex_hash_bucket *hb1,
1339                                  struct futex_hash_bucket *hb2,
1340                                  union futex_key *key1, union futex_key *key2,
1341                                  struct futex_pi_state **ps, int set_waiters)
1342 {
1343         struct futex_q *top_waiter = NULL;
1344         u32 curval;
1345         int ret;
1346
1347         if (get_futex_value_locked(&curval, pifutex))
1348                 return -EFAULT;
1349
1350         /*
1351          * Find the top_waiter and determine if there are additional waiters.
1352          * If the caller intends to requeue more than 1 waiter to pifutex,
1353          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1354          * as we have means to handle the possible fault.  If not, don't set
1355          * the bit unecessarily as it will force the subsequent unlock to enter
1356          * the kernel.
1357          */
1358         top_waiter = futex_top_waiter(hb1, key1);
1359
1360         /* There are no waiters, nothing for us to do. */
1361         if (!top_waiter)
1362                 return 0;
1363
1364         /* Ensure we requeue to the expected futex. */
1365         if (!match_futex(top_waiter->requeue_pi_key, key2))
1366                 return -EINVAL;
1367
1368         /*
1369          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1370          * the contended case or if set_waiters is 1.  The pi_state is returned
1371          * in ps in contended cases.
1372          */
1373         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1374                                    set_waiters);
1375         if (ret == 1)
1376                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1377
1378         return ret;
1379 }
1380
1381 /**
1382  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1383  * @uaddr1:     source futex user address
1384  * @flags:      futex flags (FLAGS_SHARED, etc.)
1385  * @uaddr2:     target futex user address
1386  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1387  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1388  * @cmpval:     @uaddr1 expected value (or %NULL)
1389  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1390  *              pi futex (pi to pi requeue is not supported)
1391  *
1392  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1393  * uaddr2 atomically on behalf of the top waiter.
1394  *
1395  * Return:
1396  * >=0 - on success, the number of tasks requeued or woken;
1397  *  <0 - on error
1398  */
1399 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1400                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1401                          u32 *cmpval, int requeue_pi)
1402 {
1403         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1404         int drop_count = 0, task_count = 0, ret;
1405         struct futex_pi_state *pi_state = NULL;
1406         struct futex_hash_bucket *hb1, *hb2;
1407         struct futex_q *this, *next;
1408         u32 curval2;
1409
1410         if (requeue_pi) {
1411                 /*
1412                  * requeue_pi requires a pi_state, try to allocate it now
1413                  * without any locks in case it fails.
1414                  */
1415                 if (refill_pi_state_cache())
1416                         return -ENOMEM;
1417                 /*
1418                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1419                  * + nr_requeue, since it acquires the rt_mutex prior to
1420                  * returning to userspace, so as to not leave the rt_mutex with
1421                  * waiters and no owner.  However, second and third wake-ups
1422                  * cannot be predicted as they involve race conditions with the
1423                  * first wake and a fault while looking up the pi_state.  Both
1424                  * pthread_cond_signal() and pthread_cond_broadcast() should
1425                  * use nr_wake=1.
1426                  */
1427                 if (nr_wake != 1)
1428                         return -EINVAL;
1429         }
1430
1431 retry:
1432         if (pi_state != NULL) {
1433                 /*
1434                  * We will have to lookup the pi_state again, so free this one
1435                  * to keep the accounting correct.
1436                  */
1437                 free_pi_state(pi_state);
1438                 pi_state = NULL;
1439         }
1440
1441         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1442         if (unlikely(ret != 0))
1443                 goto out;
1444         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1445                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1446         if (unlikely(ret != 0))
1447                 goto out_put_key1;
1448
1449         hb1 = hash_futex(&key1);
1450         hb2 = hash_futex(&key2);
1451
1452 retry_private:
1453         hb_waiters_inc(hb2);
1454         double_lock_hb(hb1, hb2);
1455
1456         if (likely(cmpval != NULL)) {
1457                 u32 curval;
1458
1459                 ret = get_futex_value_locked(&curval, uaddr1);
1460
1461                 if (unlikely(ret)) {
1462                         double_unlock_hb(hb1, hb2);
1463                         hb_waiters_dec(hb2);
1464
1465                         ret = get_user(curval, uaddr1);
1466                         if (ret)
1467                                 goto out_put_keys;
1468
1469                         if (!(flags & FLAGS_SHARED))
1470                                 goto retry_private;
1471
1472                         put_futex_key(&key2);
1473                         put_futex_key(&key1);
1474                         goto retry;
1475                 }
1476                 if (curval != *cmpval) {
1477                         ret = -EAGAIN;
1478                         goto out_unlock;
1479                 }
1480         }
1481
1482         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1483                 /*
1484                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1485                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1486                  * bit.  We force this here where we are able to easily handle
1487                  * faults rather in the requeue loop below.
1488                  */
1489                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1490                                                  &key2, &pi_state, nr_requeue);
1491
1492                 /*
1493                  * At this point the top_waiter has either taken uaddr2 or is
1494                  * waiting on it.  If the former, then the pi_state will not
1495                  * exist yet, look it up one more time to ensure we have a
1496                  * reference to it.
1497                  */
1498                 if (ret == 1) {
1499                         WARN_ON(pi_state);
1500                         drop_count++;
1501                         task_count++;
1502                         ret = get_futex_value_locked(&curval2, uaddr2);
1503                         if (!ret)
1504                                 ret = lookup_pi_state(curval2, hb2, &key2,
1505                                                       &pi_state);
1506                 }
1507
1508                 switch (ret) {
1509                 case 0:
1510                         break;
1511                 case -EFAULT:
1512                         double_unlock_hb(hb1, hb2);
1513                         hb_waiters_dec(hb2);
1514                         put_futex_key(&key2);
1515                         put_futex_key(&key1);
1516                         ret = fault_in_user_writeable(uaddr2);
1517                         if (!ret)
1518                                 goto retry;
1519                         goto out;
1520                 case -EAGAIN:
1521                         /* The owner was exiting, try again. */
1522                         double_unlock_hb(hb1, hb2);
1523                         hb_waiters_dec(hb2);
1524                         put_futex_key(&key2);
1525                         put_futex_key(&key1);
1526                         cond_resched();
1527                         goto retry;
1528                 default:
1529                         goto out_unlock;
1530                 }
1531         }
1532
1533         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1534                 if (task_count - nr_wake >= nr_requeue)
1535                         break;
1536
1537                 if (!match_futex(&this->key, &key1))
1538                         continue;
1539
1540                 /*
1541                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1542                  * be paired with each other and no other futex ops.
1543                  *
1544                  * We should never be requeueing a futex_q with a pi_state,
1545                  * which is awaiting a futex_unlock_pi().
1546                  */
1547                 if ((requeue_pi && !this->rt_waiter) ||
1548                     (!requeue_pi && this->rt_waiter) ||
1549                     this->pi_state) {
1550                         ret = -EINVAL;
1551                         break;
1552                 }
1553
1554                 /*
1555                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1556                  * lock, we already woke the top_waiter.  If not, it will be
1557                  * woken by futex_unlock_pi().
1558                  */
1559                 if (++task_count <= nr_wake && !requeue_pi) {
1560                         wake_futex(this);
1561                         continue;
1562                 }
1563
1564                 /* Ensure we requeue to the expected futex for requeue_pi. */
1565                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1566                         ret = -EINVAL;
1567                         break;
1568                 }
1569
1570                 /*
1571                  * Requeue nr_requeue waiters and possibly one more in the case
1572                  * of requeue_pi if we couldn't acquire the lock atomically.
1573                  */
1574                 if (requeue_pi) {
1575                         /* Prepare the waiter to take the rt_mutex. */
1576                         atomic_inc(&pi_state->refcount);
1577                         this->pi_state = pi_state;
1578                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1579                                                         this->rt_waiter,
1580                                                         this->task, 1);
1581                         if (ret == 1) {
1582                                 /* We got the lock. */
1583                                 requeue_pi_wake_futex(this, &key2, hb2);
1584                                 drop_count++;
1585                                 continue;
1586                         } else if (ret) {
1587                                 /* -EDEADLK */
1588                                 this->pi_state = NULL;
1589                                 free_pi_state(pi_state);
1590                                 goto out_unlock;
1591                         }
1592                 }
1593                 requeue_futex(this, hb1, hb2, &key2);
1594                 drop_count++;
1595         }
1596
1597 out_unlock:
1598         double_unlock_hb(hb1, hb2);
1599         hb_waiters_dec(hb2);
1600
1601         /*
1602          * drop_futex_key_refs() must be called outside the spinlocks. During
1603          * the requeue we moved futex_q's from the hash bucket at key1 to the
1604          * one at key2 and updated their key pointer.  We no longer need to
1605          * hold the references to key1.
1606          */
1607         while (--drop_count >= 0)
1608                 drop_futex_key_refs(&key1);
1609
1610 out_put_keys:
1611         put_futex_key(&key2);
1612 out_put_key1:
1613         put_futex_key(&key1);
1614 out:
1615         if (pi_state != NULL)
1616                 free_pi_state(pi_state);
1617         return ret ? ret : task_count;
1618 }
1619
1620 /* The key must be already stored in q->key. */
1621 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1622         __acquires(&hb->lock)
1623 {
1624         struct futex_hash_bucket *hb;
1625
1626         hb = hash_futex(&q->key);
1627
1628         /*
1629          * Increment the counter before taking the lock so that
1630          * a potential waker won't miss a to-be-slept task that is
1631          * waiting for the spinlock. This is safe as all queue_lock()
1632          * users end up calling queue_me(). Similarly, for housekeeping,
1633          * decrement the counter at queue_unlock() when some error has
1634          * occurred and we don't end up adding the task to the list.
1635          */
1636         hb_waiters_inc(hb);
1637
1638         q->lock_ptr = &hb->lock;
1639
1640         spin_lock(&hb->lock); /* implies MB (A) */
1641         return hb;
1642 }
1643
1644 static inline void
1645 queue_unlock(struct futex_hash_bucket *hb)
1646         __releases(&hb->lock)
1647 {
1648         spin_unlock(&hb->lock);
1649         hb_waiters_dec(hb);
1650 }
1651
1652 /**
1653  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1654  * @q:  The futex_q to enqueue
1655  * @hb: The destination hash bucket
1656  *
1657  * The hb->lock must be held by the caller, and is released here. A call to
1658  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1659  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1660  * or nothing if the unqueue is done as part of the wake process and the unqueue
1661  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1662  * an example).
1663  */
1664 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1665         __releases(&hb->lock)
1666 {
1667         int prio;
1668
1669         /*
1670          * The priority used to register this element is
1671          * - either the real thread-priority for the real-time threads
1672          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1673          * - or MAX_RT_PRIO for non-RT threads.
1674          * Thus, all RT-threads are woken first in priority order, and
1675          * the others are woken last, in FIFO order.
1676          */
1677         prio = min(current->normal_prio, MAX_RT_PRIO);
1678
1679         plist_node_init(&q->list, prio);
1680         plist_add(&q->list, &hb->chain);
1681         q->task = current;
1682         spin_unlock(&hb->lock);
1683 }
1684
1685 /**
1686  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1687  * @q:  The futex_q to unqueue
1688  *
1689  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1690  * be paired with exactly one earlier call to queue_me().
1691  *
1692  * Return:
1693  *   1 - if the futex_q was still queued (and we removed unqueued it);
1694  *   0 - if the futex_q was already removed by the waking thread
1695  */
1696 static int unqueue_me(struct futex_q *q)
1697 {
1698         spinlock_t *lock_ptr;
1699         int ret = 0;
1700
1701         /* In the common case we don't take the spinlock, which is nice. */
1702 retry:
1703         lock_ptr = q->lock_ptr;
1704         barrier();
1705         if (lock_ptr != NULL) {
1706                 spin_lock(lock_ptr);
1707                 /*
1708                  * q->lock_ptr can change between reading it and
1709                  * spin_lock(), causing us to take the wrong lock.  This
1710                  * corrects the race condition.
1711                  *
1712                  * Reasoning goes like this: if we have the wrong lock,
1713                  * q->lock_ptr must have changed (maybe several times)
1714                  * between reading it and the spin_lock().  It can
1715                  * change again after the spin_lock() but only if it was
1716                  * already changed before the spin_lock().  It cannot,
1717                  * however, change back to the original value.  Therefore
1718                  * we can detect whether we acquired the correct lock.
1719                  */
1720                 if (unlikely(lock_ptr != q->lock_ptr)) {
1721                         spin_unlock(lock_ptr);
1722                         goto retry;
1723                 }
1724                 __unqueue_futex(q);
1725
1726                 BUG_ON(q->pi_state);
1727
1728                 spin_unlock(lock_ptr);
1729                 ret = 1;
1730         }
1731
1732         drop_futex_key_refs(&q->key);
1733         return ret;
1734 }
1735
1736 /*
1737  * PI futexes can not be requeued and must remove themself from the
1738  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1739  * and dropped here.
1740  */
1741 static void unqueue_me_pi(struct futex_q *q)
1742         __releases(q->lock_ptr)
1743 {
1744         __unqueue_futex(q);
1745
1746         BUG_ON(!q->pi_state);
1747         free_pi_state(q->pi_state);
1748         q->pi_state = NULL;
1749
1750         spin_unlock(q->lock_ptr);
1751 }
1752
1753 /*
1754  * Fixup the pi_state owner with the new owner.
1755  *
1756  * Must be called with hash bucket lock held and mm->sem held for non
1757  * private futexes.
1758  */
1759 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1760                                 struct task_struct *newowner)
1761 {
1762         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1763         struct futex_pi_state *pi_state = q->pi_state;
1764         struct task_struct *oldowner = pi_state->owner;
1765         u32 uval, uninitialized_var(curval), newval;
1766         int ret;
1767
1768         /* Owner died? */
1769         if (!pi_state->owner)
1770                 newtid |= FUTEX_OWNER_DIED;
1771
1772         /*
1773          * We are here either because we stole the rtmutex from the
1774          * previous highest priority waiter or we are the highest priority
1775          * waiter but failed to get the rtmutex the first time.
1776          * We have to replace the newowner TID in the user space variable.
1777          * This must be atomic as we have to preserve the owner died bit here.
1778          *
1779          * Note: We write the user space value _before_ changing the pi_state
1780          * because we can fault here. Imagine swapped out pages or a fork
1781          * that marked all the anonymous memory readonly for cow.
1782          *
1783          * Modifying pi_state _before_ the user space value would
1784          * leave the pi_state in an inconsistent state when we fault
1785          * here, because we need to drop the hash bucket lock to
1786          * handle the fault. This might be observed in the PID check
1787          * in lookup_pi_state.
1788          */
1789 retry:
1790         if (get_futex_value_locked(&uval, uaddr))
1791                 goto handle_fault;
1792
1793         while (1) {
1794                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1795
1796                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1797                         goto handle_fault;
1798                 if (curval == uval)
1799                         break;
1800                 uval = curval;
1801         }
1802
1803         /*
1804          * We fixed up user space. Now we need to fix the pi_state
1805          * itself.
1806          */
1807         if (pi_state->owner != NULL) {
1808                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1809                 WARN_ON(list_empty(&pi_state->list));
1810                 list_del_init(&pi_state->list);
1811                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1812         }
1813
1814         pi_state->owner = newowner;
1815
1816         raw_spin_lock_irq(&newowner->pi_lock);
1817         WARN_ON(!list_empty(&pi_state->list));
1818         list_add(&pi_state->list, &newowner->pi_state_list);
1819         raw_spin_unlock_irq(&newowner->pi_lock);
1820         return 0;
1821
1822         /*
1823          * To handle the page fault we need to drop the hash bucket
1824          * lock here. That gives the other task (either the highest priority
1825          * waiter itself or the task which stole the rtmutex) the
1826          * chance to try the fixup of the pi_state. So once we are
1827          * back from handling the fault we need to check the pi_state
1828          * after reacquiring the hash bucket lock and before trying to
1829          * do another fixup. When the fixup has been done already we
1830          * simply return.
1831          */
1832 handle_fault:
1833         spin_unlock(q->lock_ptr);
1834
1835         ret = fault_in_user_writeable(uaddr);
1836
1837         spin_lock(q->lock_ptr);
1838
1839         /*
1840          * Check if someone else fixed it for us:
1841          */
1842         if (pi_state->owner != oldowner)
1843                 return 0;
1844
1845         if (ret)
1846                 return ret;
1847
1848         goto retry;
1849 }
1850
1851 static long futex_wait_restart(struct restart_block *restart);
1852
1853 /**
1854  * fixup_owner() - Post lock pi_state and corner case management
1855  * @uaddr:      user address of the futex
1856  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1857  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1858  *
1859  * After attempting to lock an rt_mutex, this function is called to cleanup
1860  * the pi_state owner as well as handle race conditions that may allow us to
1861  * acquire the lock. Must be called with the hb lock held.
1862  *
1863  * Return:
1864  *  1 - success, lock taken;
1865  *  0 - success, lock not taken;
1866  * <0 - on error (-EFAULT)
1867  */
1868 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1869 {
1870         struct task_struct *owner;
1871         int ret = 0;
1872
1873         if (locked) {
1874                 /*
1875                  * Got the lock. We might not be the anticipated owner if we
1876                  * did a lock-steal - fix up the PI-state in that case:
1877                  */
1878                 if (q->pi_state->owner != current)
1879                         ret = fixup_pi_state_owner(uaddr, q, current);
1880                 goto out;
1881         }
1882
1883         /*
1884          * Catch the rare case, where the lock was released when we were on the
1885          * way back before we locked the hash bucket.
1886          */
1887         if (q->pi_state->owner == current) {
1888                 /*
1889                  * Try to get the rt_mutex now. This might fail as some other
1890                  * task acquired the rt_mutex after we removed ourself from the
1891                  * rt_mutex waiters list.
1892                  */
1893                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1894                         locked = 1;
1895                         goto out;
1896                 }
1897
1898                 /*
1899                  * pi_state is incorrect, some other task did a lock steal and
1900                  * we returned due to timeout or signal without taking the
1901                  * rt_mutex. Too late.
1902                  */
1903                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1904                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1905                 if (!owner)
1906                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1907                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1908                 ret = fixup_pi_state_owner(uaddr, q, owner);
1909                 goto out;
1910         }
1911
1912         /*
1913          * Paranoia check. If we did not take the lock, then we should not be
1914          * the owner of the rt_mutex.
1915          */
1916         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1917                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1918                                 "pi-state %p\n", ret,
1919                                 q->pi_state->pi_mutex.owner,
1920                                 q->pi_state->owner);
1921
1922 out:
1923         return ret ? ret : locked;
1924 }
1925
1926 /**
1927  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1928  * @hb:         the futex hash bucket, must be locked by the caller
1929  * @q:          the futex_q to queue up on
1930  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1931  */
1932 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1933                                 struct hrtimer_sleeper *timeout)
1934 {
1935         /*
1936          * The task state is guaranteed to be set before another task can
1937          * wake it. set_current_state() is implemented using set_mb() and
1938          * queue_me() calls spin_unlock() upon completion, both serializing
1939          * access to the hash list and forcing another memory barrier.
1940          */
1941         set_current_state(TASK_INTERRUPTIBLE);
1942         queue_me(q, hb);
1943
1944         /* Arm the timer */
1945         if (timeout) {
1946                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1947                 if (!hrtimer_active(&timeout->timer))
1948                         timeout->task = NULL;
1949         }
1950
1951         /*
1952          * If we have been removed from the hash list, then another task
1953          * has tried to wake us, and we can skip the call to schedule().
1954          */
1955         if (likely(!plist_node_empty(&q->list))) {
1956                 /*
1957                  * If the timer has already expired, current will already be
1958                  * flagged for rescheduling. Only call schedule if there
1959                  * is no timeout, or if it has yet to expire.
1960                  */
1961                 if (!timeout || timeout->task)
1962                         freezable_schedule();
1963         }
1964         __set_current_state(TASK_RUNNING);
1965 }
1966
1967 /**
1968  * futex_wait_setup() - Prepare to wait on a futex
1969  * @uaddr:      the futex userspace address
1970  * @val:        the expected value
1971  * @flags:      futex flags (FLAGS_SHARED, etc.)
1972  * @q:          the associated futex_q
1973  * @hb:         storage for hash_bucket pointer to be returned to caller
1974  *
1975  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1976  * compare it with the expected value.  Handle atomic faults internally.
1977  * Return with the hb lock held and a q.key reference on success, and unlocked
1978  * with no q.key reference on failure.
1979  *
1980  * Return:
1981  *  0 - uaddr contains val and hb has been locked;
1982  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1983  */
1984 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1985                            struct futex_q *q, struct futex_hash_bucket **hb)
1986 {
1987         u32 uval;
1988         int ret;
1989
1990         /*
1991          * Access the page AFTER the hash-bucket is locked.
1992          * Order is important:
1993          *
1994          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1995          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1996          *
1997          * The basic logical guarantee of a futex is that it blocks ONLY
1998          * if cond(var) is known to be true at the time of blocking, for
1999          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2000          * would open a race condition where we could block indefinitely with
2001          * cond(var) false, which would violate the guarantee.
2002          *
2003          * On the other hand, we insert q and release the hash-bucket only
2004          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2005          * absorb a wakeup if *uaddr does not match the desired values
2006          * while the syscall executes.
2007          */
2008 retry:
2009         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2010         if (unlikely(ret != 0))
2011                 return ret;
2012
2013 retry_private:
2014         *hb = queue_lock(q);
2015
2016         ret = get_futex_value_locked(&uval, uaddr);
2017
2018         if (ret) {
2019                 queue_unlock(*hb);
2020
2021                 ret = get_user(uval, uaddr);
2022                 if (ret)
2023                         goto out;
2024
2025                 if (!(flags & FLAGS_SHARED))
2026                         goto retry_private;
2027
2028                 put_futex_key(&q->key);
2029                 goto retry;
2030         }
2031
2032         if (uval != val) {
2033                 queue_unlock(*hb);
2034                 ret = -EWOULDBLOCK;
2035         }
2036
2037 out:
2038         if (ret)
2039                 put_futex_key(&q->key);
2040         return ret;
2041 }
2042
2043 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2044                       ktime_t *abs_time, u32 bitset)
2045 {
2046         struct hrtimer_sleeper timeout, *to = NULL;
2047         struct restart_block *restart;
2048         struct futex_hash_bucket *hb;
2049         struct futex_q q = futex_q_init;
2050         int ret;
2051
2052         if (!bitset)
2053                 return -EINVAL;
2054         q.bitset = bitset;
2055
2056         if (abs_time) {
2057                 to = &timeout;
2058
2059                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2060                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2061                                       HRTIMER_MODE_ABS);
2062                 hrtimer_init_sleeper(to, current);
2063                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2064                                              current->timer_slack_ns);
2065         }
2066
2067 retry:
2068         /*
2069          * Prepare to wait on uaddr. On success, holds hb lock and increments
2070          * q.key refs.
2071          */
2072         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2073         if (ret)
2074                 goto out;
2075
2076         /* queue_me and wait for wakeup, timeout, or a signal. */
2077         futex_wait_queue_me(hb, &q, to);
2078
2079         /* If we were woken (and unqueued), we succeeded, whatever. */
2080         ret = 0;
2081         /* unqueue_me() drops q.key ref */
2082         if (!unqueue_me(&q))
2083                 goto out;
2084         ret = -ETIMEDOUT;
2085         if (to && !to->task)
2086                 goto out;
2087
2088         /*
2089          * We expect signal_pending(current), but we might be the
2090          * victim of a spurious wakeup as well.
2091          */
2092         if (!signal_pending(current))
2093                 goto retry;
2094
2095         ret = -ERESTARTSYS;
2096         if (!abs_time)
2097                 goto out;
2098
2099         restart = &current_thread_info()->restart_block;
2100         restart->fn = futex_wait_restart;
2101         restart->futex.uaddr = uaddr;
2102         restart->futex.val = val;
2103         restart->futex.time = abs_time->tv64;
2104         restart->futex.bitset = bitset;
2105         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2106
2107         ret = -ERESTART_RESTARTBLOCK;
2108
2109 out:
2110         if (to) {
2111                 hrtimer_cancel(&to->timer);
2112                 destroy_hrtimer_on_stack(&to->timer);
2113         }
2114         return ret;
2115 }
2116
2117
2118 static long futex_wait_restart(struct restart_block *restart)
2119 {
2120         u32 __user *uaddr = restart->futex.uaddr;
2121         ktime_t t, *tp = NULL;
2122
2123         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2124                 t.tv64 = restart->futex.time;
2125                 tp = &t;
2126         }
2127         restart->fn = do_no_restart_syscall;
2128
2129         return (long)futex_wait(uaddr, restart->futex.flags,
2130                                 restart->futex.val, tp, restart->futex.bitset);
2131 }
2132
2133
2134 /*
2135  * Userspace tried a 0 -> TID atomic transition of the futex value
2136  * and failed. The kernel side here does the whole locking operation:
2137  * if there are waiters then it will block, it does PI, etc. (Due to
2138  * races the kernel might see a 0 value of the futex too.)
2139  */
2140 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2141                          ktime_t *time, int trylock)
2142 {
2143         struct hrtimer_sleeper timeout, *to = NULL;
2144         struct futex_hash_bucket *hb;
2145         struct futex_q q = futex_q_init;
2146         int res, ret;
2147
2148         if (refill_pi_state_cache())
2149                 return -ENOMEM;
2150
2151         if (time) {
2152                 to = &timeout;
2153                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2154                                       HRTIMER_MODE_ABS);
2155                 hrtimer_init_sleeper(to, current);
2156                 hrtimer_set_expires(&to->timer, *time);
2157         }
2158
2159 retry:
2160         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2161         if (unlikely(ret != 0))
2162                 goto out;
2163
2164 retry_private:
2165         hb = queue_lock(&q);
2166
2167         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2168         if (unlikely(ret)) {
2169                 switch (ret) {
2170                 case 1:
2171                         /* We got the lock. */
2172                         ret = 0;
2173                         goto out_unlock_put_key;
2174                 case -EFAULT:
2175                         goto uaddr_faulted;
2176                 case -EAGAIN:
2177                         /*
2178                          * Task is exiting and we just wait for the
2179                          * exit to complete.
2180                          */
2181                         queue_unlock(hb);
2182                         put_futex_key(&q.key);
2183                         cond_resched();
2184                         goto retry;
2185                 default:
2186                         goto out_unlock_put_key;
2187                 }
2188         }
2189
2190         /*
2191          * Only actually queue now that the atomic ops are done:
2192          */
2193         queue_me(&q, hb);
2194
2195         WARN_ON(!q.pi_state);
2196         /*
2197          * Block on the PI mutex:
2198          */
2199         if (!trylock)
2200                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2201         else {
2202                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2203                 /* Fixup the trylock return value: */
2204                 ret = ret ? 0 : -EWOULDBLOCK;
2205         }
2206
2207         spin_lock(q.lock_ptr);
2208         /*
2209          * Fixup the pi_state owner and possibly acquire the lock if we
2210          * haven't already.
2211          */
2212         res = fixup_owner(uaddr, &q, !ret);
2213         /*
2214          * If fixup_owner() returned an error, proprogate that.  If it acquired
2215          * the lock, clear our -ETIMEDOUT or -EINTR.
2216          */
2217         if (res)
2218                 ret = (res < 0) ? res : 0;
2219
2220         /*
2221          * If fixup_owner() faulted and was unable to handle the fault, unlock
2222          * it and return the fault to userspace.
2223          */
2224         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2225                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2226
2227         /* Unqueue and drop the lock */
2228         unqueue_me_pi(&q);
2229
2230         goto out_put_key;
2231
2232 out_unlock_put_key:
2233         queue_unlock(hb);
2234
2235 out_put_key:
2236         put_futex_key(&q.key);
2237 out:
2238         if (to)
2239                 destroy_hrtimer_on_stack(&to->timer);
2240         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2241
2242 uaddr_faulted:
2243         queue_unlock(hb);
2244
2245         ret = fault_in_user_writeable(uaddr);
2246         if (ret)
2247                 goto out_put_key;
2248
2249         if (!(flags & FLAGS_SHARED))
2250                 goto retry_private;
2251
2252         put_futex_key(&q.key);
2253         goto retry;
2254 }
2255
2256 /*
2257  * Userspace attempted a TID -> 0 atomic transition, and failed.
2258  * This is the in-kernel slowpath: we look up the PI state (if any),
2259  * and do the rt-mutex unlock.
2260  */
2261 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2262 {
2263         struct futex_hash_bucket *hb;
2264         struct futex_q *this, *next;
2265         union futex_key key = FUTEX_KEY_INIT;
2266         u32 uval, vpid = task_pid_vnr(current);
2267         int ret;
2268
2269 retry:
2270         if (get_user(uval, uaddr))
2271                 return -EFAULT;
2272         /*
2273          * We release only a lock we actually own:
2274          */
2275         if ((uval & FUTEX_TID_MASK) != vpid)
2276                 return -EPERM;
2277
2278         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2279         if (unlikely(ret != 0))
2280                 goto out;
2281
2282         hb = hash_futex(&key);
2283         spin_lock(&hb->lock);
2284
2285         /*
2286          * To avoid races, try to do the TID -> 0 atomic transition
2287          * again. If it succeeds then we can return without waking
2288          * anyone else up:
2289          */
2290         if (!(uval & FUTEX_OWNER_DIED) &&
2291             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2292                 goto pi_faulted;
2293         /*
2294          * Rare case: we managed to release the lock atomically,
2295          * no need to wake anyone else up:
2296          */
2297         if (unlikely(uval == vpid))
2298                 goto out_unlock;
2299
2300         /*
2301          * Ok, other tasks may need to be woken up - check waiters
2302          * and do the wakeup if necessary:
2303          */
2304         plist_for_each_entry_safe(this, next, &hb->chain, list) {
2305                 if (!match_futex (&this->key, &key))
2306                         continue;
2307                 ret = wake_futex_pi(uaddr, uval, this);
2308                 /*
2309                  * The atomic access to the futex value
2310                  * generated a pagefault, so retry the
2311                  * user-access and the wakeup:
2312                  */
2313                 if (ret == -EFAULT)
2314                         goto pi_faulted;
2315                 goto out_unlock;
2316         }
2317         /*
2318          * No waiters - kernel unlocks the futex:
2319          */
2320         if (!(uval & FUTEX_OWNER_DIED)) {
2321                 ret = unlock_futex_pi(uaddr, uval);
2322                 if (ret == -EFAULT)
2323                         goto pi_faulted;
2324         }
2325
2326 out_unlock:
2327         spin_unlock(&hb->lock);
2328         put_futex_key(&key);
2329
2330 out:
2331         return ret;
2332
2333 pi_faulted:
2334         spin_unlock(&hb->lock);
2335         put_futex_key(&key);
2336
2337         ret = fault_in_user_writeable(uaddr);
2338         if (!ret)
2339                 goto retry;
2340
2341         return ret;
2342 }
2343
2344 /**
2345  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2346  * @hb:         the hash_bucket futex_q was original enqueued on
2347  * @q:          the futex_q woken while waiting to be requeued
2348  * @key2:       the futex_key of the requeue target futex
2349  * @timeout:    the timeout associated with the wait (NULL if none)
2350  *
2351  * Detect if the task was woken on the initial futex as opposed to the requeue
2352  * target futex.  If so, determine if it was a timeout or a signal that caused
2353  * the wakeup and return the appropriate error code to the caller.  Must be
2354  * called with the hb lock held.
2355  *
2356  * Return:
2357  *  0 = no early wakeup detected;
2358  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2359  */
2360 static inline
2361 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2362                                    struct futex_q *q, union futex_key *key2,
2363                                    struct hrtimer_sleeper *timeout)
2364 {
2365         int ret = 0;
2366
2367         /*
2368          * With the hb lock held, we avoid races while we process the wakeup.
2369          * We only need to hold hb (and not hb2) to ensure atomicity as the
2370          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2371          * It can't be requeued from uaddr2 to something else since we don't
2372          * support a PI aware source futex for requeue.
2373          */
2374         if (!match_futex(&q->key, key2)) {
2375                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2376                 /*
2377                  * We were woken prior to requeue by a timeout or a signal.
2378                  * Unqueue the futex_q and determine which it was.
2379                  */
2380                 plist_del(&q->list, &hb->chain);
2381                 hb_waiters_dec(hb);
2382
2383                 /* Handle spurious wakeups gracefully */
2384                 ret = -EWOULDBLOCK;
2385                 if (timeout && !timeout->task)
2386                         ret = -ETIMEDOUT;
2387                 else if (signal_pending(current))
2388                         ret = -ERESTARTNOINTR;
2389         }
2390         return ret;
2391 }
2392
2393 /**
2394  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2395  * @uaddr:      the futex we initially wait on (non-pi)
2396  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2397  *              the same type, no requeueing from private to shared, etc.
2398  * @val:        the expected value of uaddr
2399  * @abs_time:   absolute timeout
2400  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2401  * @uaddr2:     the pi futex we will take prior to returning to user-space
2402  *
2403  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2404  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2405  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2406  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2407  * without one, the pi logic would not know which task to boost/deboost, if
2408  * there was a need to.
2409  *
2410  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2411  * via the following--
2412  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2413  * 2) wakeup on uaddr2 after a requeue
2414  * 3) signal
2415  * 4) timeout
2416  *
2417  * If 3, cleanup and return -ERESTARTNOINTR.
2418  *
2419  * If 2, we may then block on trying to take the rt_mutex and return via:
2420  * 5) successful lock
2421  * 6) signal
2422  * 7) timeout
2423  * 8) other lock acquisition failure
2424  *
2425  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2426  *
2427  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2428  *
2429  * Return:
2430  *  0 - On success;
2431  * <0 - On error
2432  */
2433 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2434                                  u32 val, ktime_t *abs_time, u32 bitset,
2435                                  u32 __user *uaddr2)
2436 {
2437         struct hrtimer_sleeper timeout, *to = NULL;
2438         struct rt_mutex_waiter rt_waiter;
2439         struct rt_mutex *pi_mutex = NULL;
2440         struct futex_hash_bucket *hb;
2441         union futex_key key2 = FUTEX_KEY_INIT;
2442         struct futex_q q = futex_q_init;
2443         int res, ret;
2444
2445         if (uaddr == uaddr2)
2446                 return -EINVAL;
2447
2448         if (!bitset)
2449                 return -EINVAL;
2450
2451         if (abs_time) {
2452                 to = &timeout;
2453                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2454                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2455                                       HRTIMER_MODE_ABS);
2456                 hrtimer_init_sleeper(to, current);
2457                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2458                                              current->timer_slack_ns);
2459         }
2460
2461         /*
2462          * The waiter is allocated on our stack, manipulated by the requeue
2463          * code while we sleep on uaddr.
2464          */
2465         debug_rt_mutex_init_waiter(&rt_waiter);
2466         RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2467         RB_CLEAR_NODE(&rt_waiter.tree_entry);
2468         rt_waiter.task = NULL;
2469
2470         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2471         if (unlikely(ret != 0))
2472                 goto out;
2473
2474         q.bitset = bitset;
2475         q.rt_waiter = &rt_waiter;
2476         q.requeue_pi_key = &key2;
2477
2478         /*
2479          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2480          * count.
2481          */
2482         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2483         if (ret)
2484                 goto out_key2;
2485
2486         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2487         futex_wait_queue_me(hb, &q, to);
2488
2489         spin_lock(&hb->lock);
2490         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2491         spin_unlock(&hb->lock);
2492         if (ret)
2493                 goto out_put_keys;
2494
2495         /*
2496          * In order for us to be here, we know our q.key == key2, and since
2497          * we took the hb->lock above, we also know that futex_requeue() has
2498          * completed and we no longer have to concern ourselves with a wakeup
2499          * race with the atomic proxy lock acquisition by the requeue code. The
2500          * futex_requeue dropped our key1 reference and incremented our key2
2501          * reference count.
2502          */
2503
2504         /* Check if the requeue code acquired the second futex for us. */
2505         if (!q.rt_waiter) {
2506                 /*
2507                  * Got the lock. We might not be the anticipated owner if we
2508                  * did a lock-steal - fix up the PI-state in that case.
2509                  */
2510                 if (q.pi_state && (q.pi_state->owner != current)) {
2511                         spin_lock(q.lock_ptr);
2512                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2513                         spin_unlock(q.lock_ptr);
2514                 }
2515         } else {
2516                 /*
2517                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2518                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2519                  * the pi_state.
2520                  */
2521                 WARN_ON(!q.pi_state);
2522                 pi_mutex = &q.pi_state->pi_mutex;
2523                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2524                 debug_rt_mutex_free_waiter(&rt_waiter);
2525
2526                 spin_lock(q.lock_ptr);
2527                 /*
2528                  * Fixup the pi_state owner and possibly acquire the lock if we
2529                  * haven't already.
2530                  */
2531                 res = fixup_owner(uaddr2, &q, !ret);
2532                 /*
2533                  * If fixup_owner() returned an error, proprogate that.  If it
2534                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2535                  */
2536                 if (res)
2537                         ret = (res < 0) ? res : 0;
2538
2539                 /* Unqueue and drop the lock. */
2540                 unqueue_me_pi(&q);
2541         }
2542
2543         /*
2544          * If fixup_pi_state_owner() faulted and was unable to handle the
2545          * fault, unlock the rt_mutex and return the fault to userspace.
2546          */
2547         if (ret == -EFAULT) {
2548                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2549                         rt_mutex_unlock(pi_mutex);
2550         } else if (ret == -EINTR) {
2551                 /*
2552                  * We've already been requeued, but cannot restart by calling
2553                  * futex_lock_pi() directly. We could restart this syscall, but
2554                  * it would detect that the user space "val" changed and return
2555                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2556                  * -EWOULDBLOCK directly.
2557                  */
2558                 ret = -EWOULDBLOCK;
2559         }
2560
2561 out_put_keys:
2562         put_futex_key(&q.key);
2563 out_key2:
2564         put_futex_key(&key2);
2565
2566 out:
2567         if (to) {
2568                 hrtimer_cancel(&to->timer);
2569                 destroy_hrtimer_on_stack(&to->timer);
2570         }
2571         return ret;
2572 }
2573
2574 /*
2575  * Support for robust futexes: the kernel cleans up held futexes at
2576  * thread exit time.
2577  *
2578  * Implementation: user-space maintains a per-thread list of locks it
2579  * is holding. Upon do_exit(), the kernel carefully walks this list,
2580  * and marks all locks that are owned by this thread with the
2581  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2582  * always manipulated with the lock held, so the list is private and
2583  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2584  * field, to allow the kernel to clean up if the thread dies after
2585  * acquiring the lock, but just before it could have added itself to
2586  * the list. There can only be one such pending lock.
2587  */
2588
2589 /**
2590  * sys_set_robust_list() - Set the robust-futex list head of a task
2591  * @head:       pointer to the list-head
2592  * @len:        length of the list-head, as userspace expects
2593  */
2594 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2595                 size_t, len)
2596 {
2597         if (!futex_cmpxchg_enabled)
2598                 return -ENOSYS;
2599         /*
2600          * The kernel knows only one size for now:
2601          */
2602         if (unlikely(len != sizeof(*head)))
2603                 return -EINVAL;
2604
2605         current->robust_list = head;
2606
2607         return 0;
2608 }
2609
2610 /**
2611  * sys_get_robust_list() - Get the robust-futex list head of a task
2612  * @pid:        pid of the process [zero for current task]
2613  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2614  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2615  */
2616 SYSCALL_DEFINE3(get_robust_list, int, pid,
2617                 struct robust_list_head __user * __user *, head_ptr,
2618                 size_t __user *, len_ptr)
2619 {
2620         struct robust_list_head __user *head;
2621         unsigned long ret;
2622         struct task_struct *p;
2623
2624         if (!futex_cmpxchg_enabled)
2625                 return -ENOSYS;
2626
2627         rcu_read_lock();
2628
2629         ret = -ESRCH;
2630         if (!pid)
2631                 p = current;
2632         else {
2633                 p = find_task_by_vpid(pid);
2634                 if (!p)
2635                         goto err_unlock;
2636         }
2637
2638         ret = -EPERM;
2639         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2640                 goto err_unlock;
2641
2642         head = p->robust_list;
2643         rcu_read_unlock();
2644
2645         if (put_user(sizeof(*head), len_ptr))
2646                 return -EFAULT;
2647         return put_user(head, head_ptr);
2648
2649 err_unlock:
2650         rcu_read_unlock();
2651
2652         return ret;
2653 }
2654
2655 /*
2656  * Process a futex-list entry, check whether it's owned by the
2657  * dying task, and do notification if so:
2658  */
2659 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2660 {
2661         u32 uval, uninitialized_var(nval), mval;
2662
2663 retry:
2664         if (get_user(uval, uaddr))
2665                 return -1;
2666
2667         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2668                 /*
2669                  * Ok, this dying thread is truly holding a futex
2670                  * of interest. Set the OWNER_DIED bit atomically
2671                  * via cmpxchg, and if the value had FUTEX_WAITERS
2672                  * set, wake up a waiter (if any). (We have to do a
2673                  * futex_wake() even if OWNER_DIED is already set -
2674                  * to handle the rare but possible case of recursive
2675                  * thread-death.) The rest of the cleanup is done in
2676                  * userspace.
2677                  */
2678                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2679                 /*
2680                  * We are not holding a lock here, but we want to have
2681                  * the pagefault_disable/enable() protection because
2682                  * we want to handle the fault gracefully. If the
2683                  * access fails we try to fault in the futex with R/W
2684                  * verification via get_user_pages. get_user() above
2685                  * does not guarantee R/W access. If that fails we
2686                  * give up and leave the futex locked.
2687                  */
2688                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2689                         if (fault_in_user_writeable(uaddr))
2690                                 return -1;
2691                         goto retry;
2692                 }
2693                 if (nval != uval)
2694                         goto retry;
2695
2696                 /*
2697                  * Wake robust non-PI futexes here. The wakeup of
2698                  * PI futexes happens in exit_pi_state():
2699                  */
2700                 if (!pi && (uval & FUTEX_WAITERS))
2701                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2702         }
2703         return 0;
2704 }
2705
2706 /*
2707  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2708  */
2709 static inline int fetch_robust_entry(struct robust_list __user **entry,
2710                                      struct robust_list __user * __user *head,
2711                                      unsigned int *pi)
2712 {
2713         unsigned long uentry;
2714
2715         if (get_user(uentry, (unsigned long __user *)head))
2716                 return -EFAULT;
2717
2718         *entry = (void __user *)(uentry & ~1UL);
2719         *pi = uentry & 1;
2720
2721         return 0;
2722 }
2723
2724 /*
2725  * Walk curr->robust_list (very carefully, it's a userspace list!)
2726  * and mark any locks found there dead, and notify any waiters.
2727  *
2728  * We silently return on any sign of list-walking problem.
2729  */
2730 void exit_robust_list(struct task_struct *curr)
2731 {
2732         struct robust_list_head __user *head = curr->robust_list;
2733         struct robust_list __user *entry, *next_entry, *pending;
2734         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2735         unsigned int uninitialized_var(next_pi);
2736         unsigned long futex_offset;
2737         int rc;
2738
2739         if (!futex_cmpxchg_enabled)
2740                 return;
2741
2742         /*
2743          * Fetch the list head (which was registered earlier, via
2744          * sys_set_robust_list()):
2745          */
2746         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2747                 return;
2748         /*
2749          * Fetch the relative futex offset:
2750          */
2751         if (get_user(futex_offset, &head->futex_offset))
2752                 return;
2753         /*
2754          * Fetch any possibly pending lock-add first, and handle it
2755          * if it exists:
2756          */
2757         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2758                 return;
2759
2760         next_entry = NULL;      /* avoid warning with gcc */
2761         while (entry != &head->list) {
2762                 /*
2763                  * Fetch the next entry in the list before calling
2764                  * handle_futex_death:
2765                  */
2766                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2767                 /*
2768                  * A pending lock might already be on the list, so
2769                  * don't process it twice:
2770                  */
2771                 if (entry != pending)
2772                         if (handle_futex_death((void __user *)entry + futex_offset,
2773                                                 curr, pi))
2774                                 return;
2775                 if (rc)
2776                         return;
2777                 entry = next_entry;
2778                 pi = next_pi;
2779                 /*
2780                  * Avoid excessively long or circular lists:
2781                  */
2782                 if (!--limit)
2783                         break;
2784
2785                 cond_resched();
2786         }
2787
2788         if (pending)
2789                 handle_futex_death((void __user *)pending + futex_offset,
2790                                    curr, pip);
2791 }
2792
2793 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2794                 u32 __user *uaddr2, u32 val2, u32 val3)
2795 {
2796         int cmd = op & FUTEX_CMD_MASK;
2797         unsigned int flags = 0;
2798
2799         if (!(op & FUTEX_PRIVATE_FLAG))
2800                 flags |= FLAGS_SHARED;
2801
2802         if (op & FUTEX_CLOCK_REALTIME) {
2803                 flags |= FLAGS_CLOCKRT;
2804                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2805                         return -ENOSYS;
2806         }
2807
2808         switch (cmd) {
2809         case FUTEX_LOCK_PI:
2810         case FUTEX_UNLOCK_PI:
2811         case FUTEX_TRYLOCK_PI:
2812         case FUTEX_WAIT_REQUEUE_PI:
2813         case FUTEX_CMP_REQUEUE_PI:
2814                 if (!futex_cmpxchg_enabled)
2815                         return -ENOSYS;
2816         }
2817
2818         switch (cmd) {
2819         case FUTEX_WAIT:
2820                 val3 = FUTEX_BITSET_MATCH_ANY;
2821         case FUTEX_WAIT_BITSET:
2822                 return futex_wait(uaddr, flags, val, timeout, val3);
2823         case FUTEX_WAKE:
2824                 val3 = FUTEX_BITSET_MATCH_ANY;
2825         case FUTEX_WAKE_BITSET:
2826                 return futex_wake(uaddr, flags, val, val3);
2827         case FUTEX_REQUEUE:
2828                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2829         case FUTEX_CMP_REQUEUE:
2830                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2831         case FUTEX_WAKE_OP:
2832                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2833         case FUTEX_LOCK_PI:
2834                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2835         case FUTEX_UNLOCK_PI:
2836                 return futex_unlock_pi(uaddr, flags);
2837         case FUTEX_TRYLOCK_PI:
2838                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2839         case FUTEX_WAIT_REQUEUE_PI:
2840                 val3 = FUTEX_BITSET_MATCH_ANY;
2841                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2842                                              uaddr2);
2843         case FUTEX_CMP_REQUEUE_PI:
2844                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2845         }
2846         return -ENOSYS;
2847 }
2848
2849
2850 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2851                 struct timespec __user *, utime, u32 __user *, uaddr2,
2852                 u32, val3)
2853 {
2854         struct timespec ts;
2855         ktime_t t, *tp = NULL;
2856         u32 val2 = 0;
2857         int cmd = op & FUTEX_CMD_MASK;
2858
2859         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2860                       cmd == FUTEX_WAIT_BITSET ||
2861                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2862                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2863                         return -EFAULT;
2864                 if (!timespec_valid(&ts))
2865                         return -EINVAL;
2866
2867                 t = timespec_to_ktime(ts);
2868                 if (cmd == FUTEX_WAIT)
2869                         t = ktime_add_safe(ktime_get(), t);
2870                 tp = &t;
2871         }
2872         /*
2873          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2874          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2875          */
2876         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2877             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2878                 val2 = (u32) (unsigned long) utime;
2879
2880         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2881 }
2882
2883 static int __init futex_init(void)
2884 {
2885         u32 curval;
2886         unsigned int futex_shift;
2887         unsigned long i;
2888
2889 #if CONFIG_BASE_SMALL
2890         futex_hashsize = 16;
2891 #else
2892         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
2893 #endif
2894
2895         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
2896                                                futex_hashsize, 0,
2897                                                futex_hashsize < 256 ? HASH_SMALL : 0,
2898                                                &futex_shift, NULL,
2899                                                futex_hashsize, futex_hashsize);
2900         futex_hashsize = 1UL << futex_shift;
2901         /*
2902          * This will fail and we want it. Some arch implementations do
2903          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2904          * functionality. We want to know that before we call in any
2905          * of the complex code paths. Also we want to prevent
2906          * registration of robust lists in that case. NULL is
2907          * guaranteed to fault and we get -EFAULT on functional
2908          * implementation, the non-functional ones will return
2909          * -ENOSYS.
2910          */
2911         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2912                 futex_cmpxchg_enabled = 1;
2913
2914         for (i = 0; i < futex_hashsize; i++) {
2915                 atomic_set(&futex_queues[i].waiters, 0);
2916                 plist_head_init(&futex_queues[i].chain);
2917                 spin_lock_init(&futex_queues[i].lock);
2918         }
2919
2920         return 0;
2921 }
2922 __initcall(futex_init);