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