futexes: Remove rw parameter from get_futex_key()
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Priority Inheritance state:
73  */
74 struct futex_pi_state {
75         /*
76          * list of 'owned' pi_state instances - these have to be
77          * cleaned up in do_exit() if the task exits prematurely:
78          */
79         struct list_head list;
80
81         /*
82          * The PI object:
83          */
84         struct rt_mutex pi_mutex;
85
86         struct task_struct *owner;
87         atomic_t refcount;
88
89         union futex_key key;
90 };
91
92 /**
93  * struct futex_q - The hashed futex queue entry, one per waiting task
94  * @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                         WARN_ON(pid && pi_state->owner &&
534                                 pi_state->owner->pid != pid);
535
536                         atomic_inc(&pi_state->refcount);
537                         *ps = pi_state;
538
539                         return 0;
540                 }
541         }
542
543         /*
544          * We are the first waiter - try to look up the real owner and attach
545          * the new pi_state to it, but bail out when TID = 0
546          */
547         if (!pid)
548                 return -ESRCH;
549         p = futex_find_get_task(pid);
550         if (IS_ERR(p))
551                 return PTR_ERR(p);
552
553         /*
554          * We need to look at the task state flags to figure out,
555          * whether the task is exiting. To protect against the do_exit
556          * change of the task flags, we do this protected by
557          * p->pi_lock:
558          */
559         raw_spin_lock_irq(&p->pi_lock);
560         if (unlikely(p->flags & PF_EXITING)) {
561                 /*
562                  * The task is on the way out. When PF_EXITPIDONE is
563                  * set, we know that the task has finished the
564                  * cleanup:
565                  */
566                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
567
568                 raw_spin_unlock_irq(&p->pi_lock);
569                 put_task_struct(p);
570                 return ret;
571         }
572
573         pi_state = alloc_pi_state();
574
575         /*
576          * Initialize the pi_mutex in locked state and make 'p'
577          * the owner of it:
578          */
579         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
580
581         /* Store the key for possible exit cleanups: */
582         pi_state->key = *key;
583
584         WARN_ON(!list_empty(&pi_state->list));
585         list_add(&pi_state->list, &p->pi_state_list);
586         pi_state->owner = p;
587         raw_spin_unlock_irq(&p->pi_lock);
588
589         put_task_struct(p);
590
591         *ps = pi_state;
592
593         return 0;
594 }
595
596 /**
597  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
598  * @uaddr:              the pi futex user address
599  * @hb:                 the pi futex hash bucket
600  * @key:                the futex key associated with uaddr and hb
601  * @ps:                 the pi_state pointer where we store the result of the
602  *                      lookup
603  * @task:               the task to perform the atomic lock work for.  This will
604  *                      be "current" except in the case of requeue pi.
605  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
606  *
607  * Returns:
608  *  0 - ready to wait
609  *  1 - acquired the lock
610  * <0 - error
611  *
612  * The hb->lock and futex_key refs shall be held by the caller.
613  */
614 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
615                                 union futex_key *key,
616                                 struct futex_pi_state **ps,
617                                 struct task_struct *task, int set_waiters)
618 {
619         int lock_taken, ret, ownerdied = 0;
620         u32 uval, newval, curval;
621
622 retry:
623         ret = lock_taken = 0;
624
625         /*
626          * To avoid races, we attempt to take the lock here again
627          * (by doing a 0 -> TID atomic cmpxchg), while holding all
628          * the locks. It will most likely not succeed.
629          */
630         newval = task_pid_vnr(task);
631         if (set_waiters)
632                 newval |= FUTEX_WAITERS;
633
634         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
635
636         if (unlikely(curval == -EFAULT))
637                 return -EFAULT;
638
639         /*
640          * Detect deadlocks.
641          */
642         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
643                 return -EDEADLK;
644
645         /*
646          * Surprise - we got the lock. Just return to userspace:
647          */
648         if (unlikely(!curval))
649                 return 1;
650
651         uval = curval;
652
653         /*
654          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
655          * to wake at the next unlock.
656          */
657         newval = curval | FUTEX_WAITERS;
658
659         /*
660          * There are two cases, where a futex might have no owner (the
661          * owner TID is 0): OWNER_DIED. We take over the futex in this
662          * case. We also do an unconditional take over, when the owner
663          * of the futex died.
664          *
665          * This is safe as we are protected by the hash bucket lock !
666          */
667         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
668                 /* Keep the OWNER_DIED bit */
669                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
670                 ownerdied = 0;
671                 lock_taken = 1;
672         }
673
674         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
675
676         if (unlikely(curval == -EFAULT))
677                 return -EFAULT;
678         if (unlikely(curval != uval))
679                 goto retry;
680
681         /*
682          * We took the lock due to owner died take over.
683          */
684         if (unlikely(lock_taken))
685                 return 1;
686
687         /*
688          * We dont have the lock. Look up the PI state (or create it if
689          * we are the first waiter):
690          */
691         ret = lookup_pi_state(uval, hb, key, ps);
692
693         if (unlikely(ret)) {
694                 switch (ret) {
695                 case -ESRCH:
696                         /*
697                          * No owner found for this futex. Check if the
698                          * OWNER_DIED bit is set to figure out whether
699                          * this is a robust futex or not.
700                          */
701                         if (get_futex_value_locked(&curval, uaddr))
702                                 return -EFAULT;
703
704                         /*
705                          * We simply start over in case of a robust
706                          * futex. The code above will take the futex
707                          * and return happy.
708                          */
709                         if (curval & FUTEX_OWNER_DIED) {
710                                 ownerdied = 1;
711                                 goto retry;
712                         }
713                 default:
714                         break;
715                 }
716         }
717
718         return ret;
719 }
720
721 /*
722  * The hash bucket lock must be held when this is called.
723  * Afterwards, the futex_q must not be accessed.
724  */
725 static void wake_futex(struct futex_q *q)
726 {
727         struct task_struct *p = q->task;
728
729         /*
730          * We set q->lock_ptr = NULL _before_ we wake up the task. If
731          * a non futex wake up happens on another CPU then the task
732          * might exit and p would dereference a non existing task
733          * struct. Prevent this by holding a reference on p across the
734          * wake up.
735          */
736         get_task_struct(p);
737
738         plist_del(&q->list, &q->list.plist);
739         /*
740          * The waiting task can free the futex_q as soon as
741          * q->lock_ptr = NULL is written, without taking any locks. A
742          * memory barrier is required here to prevent the following
743          * store to lock_ptr from getting ahead of the plist_del.
744          */
745         smp_wmb();
746         q->lock_ptr = NULL;
747
748         wake_up_state(p, TASK_NORMAL);
749         put_task_struct(p);
750 }
751
752 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
753 {
754         struct task_struct *new_owner;
755         struct futex_pi_state *pi_state = this->pi_state;
756         u32 curval, newval;
757
758         if (!pi_state)
759                 return -EINVAL;
760
761         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
762         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
763
764         /*
765          * This happens when we have stolen the lock and the original
766          * pending owner did not enqueue itself back on the rt_mutex.
767          * Thats not a tragedy. We know that way, that a lock waiter
768          * is on the fly. We make the futex_q waiter the pending owner.
769          */
770         if (!new_owner)
771                 new_owner = this->task;
772
773         /*
774          * We pass it to the next owner. (The WAITERS bit is always
775          * kept enabled while there is PI state around. We must also
776          * preserve the owner died bit.)
777          */
778         if (!(uval & FUTEX_OWNER_DIED)) {
779                 int ret = 0;
780
781                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
782
783                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
784
785                 if (curval == -EFAULT)
786                         ret = -EFAULT;
787                 else if (curval != uval)
788                         ret = -EINVAL;
789                 if (ret) {
790                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
791                         return ret;
792                 }
793         }
794
795         raw_spin_lock_irq(&pi_state->owner->pi_lock);
796         WARN_ON(list_empty(&pi_state->list));
797         list_del_init(&pi_state->list);
798         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
799
800         raw_spin_lock_irq(&new_owner->pi_lock);
801         WARN_ON(!list_empty(&pi_state->list));
802         list_add(&pi_state->list, &new_owner->pi_state_list);
803         pi_state->owner = new_owner;
804         raw_spin_unlock_irq(&new_owner->pi_lock);
805
806         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
807         rt_mutex_unlock(&pi_state->pi_mutex);
808
809         return 0;
810 }
811
812 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
813 {
814         u32 oldval;
815
816         /*
817          * There is no waiter, so we unlock the futex. The owner died
818          * bit has not to be preserved here. We are the owner:
819          */
820         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
821
822         if (oldval == -EFAULT)
823                 return oldval;
824         if (oldval != uval)
825                 return -EAGAIN;
826
827         return 0;
828 }
829
830 /*
831  * Express the locking dependencies for lockdep:
832  */
833 static inline void
834 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
835 {
836         if (hb1 <= hb2) {
837                 spin_lock(&hb1->lock);
838                 if (hb1 < hb2)
839                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
840         } else { /* hb1 > hb2 */
841                 spin_lock(&hb2->lock);
842                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
843         }
844 }
845
846 static inline void
847 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
848 {
849         spin_unlock(&hb1->lock);
850         if (hb1 != hb2)
851                 spin_unlock(&hb2->lock);
852 }
853
854 /*
855  * Wake up waiters matching bitset queued on this futex (uaddr).
856  */
857 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
858 {
859         struct futex_hash_bucket *hb;
860         struct futex_q *this, *next;
861         struct plist_head *head;
862         union futex_key key = FUTEX_KEY_INIT;
863         int ret;
864
865         if (!bitset)
866                 return -EINVAL;
867
868         ret = get_futex_key(uaddr, fshared, &key);
869         if (unlikely(ret != 0))
870                 goto out;
871
872         hb = hash_futex(&key);
873         spin_lock(&hb->lock);
874         head = &hb->chain;
875
876         plist_for_each_entry_safe(this, next, head, list) {
877                 if (match_futex (&this->key, &key)) {
878                         if (this->pi_state || this->rt_waiter) {
879                                 ret = -EINVAL;
880                                 break;
881                         }
882
883                         /* Check if one of the bits is set in both bitsets */
884                         if (!(this->bitset & bitset))
885                                 continue;
886
887                         wake_futex(this);
888                         if (++ret >= nr_wake)
889                                 break;
890                 }
891         }
892
893         spin_unlock(&hb->lock);
894         put_futex_key(fshared, &key);
895 out:
896         return ret;
897 }
898
899 /*
900  * Wake up all waiters hashed on the physical page that is mapped
901  * to this virtual address:
902  */
903 static int
904 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
905               int nr_wake, int nr_wake2, int op)
906 {
907         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
908         struct futex_hash_bucket *hb1, *hb2;
909         struct plist_head *head;
910         struct futex_q *this, *next;
911         int ret, op_ret;
912
913 retry:
914         ret = get_futex_key(uaddr1, fshared, &key1);
915         if (unlikely(ret != 0))
916                 goto out;
917         ret = get_futex_key(uaddr2, fshared, &key2);
918         if (unlikely(ret != 0))
919                 goto out_put_key1;
920
921         hb1 = hash_futex(&key1);
922         hb2 = hash_futex(&key2);
923
924 retry_private:
925         double_lock_hb(hb1, hb2);
926         op_ret = futex_atomic_op_inuser(op, uaddr2);
927         if (unlikely(op_ret < 0)) {
928
929                 double_unlock_hb(hb1, hb2);
930
931 #ifndef CONFIG_MMU
932                 /*
933                  * we don't get EFAULT from MMU faults if we don't have an MMU,
934                  * but we might get them from range checking
935                  */
936                 ret = op_ret;
937                 goto out_put_keys;
938 #endif
939
940                 if (unlikely(op_ret != -EFAULT)) {
941                         ret = op_ret;
942                         goto out_put_keys;
943                 }
944
945                 ret = fault_in_user_writeable(uaddr2);
946                 if (ret)
947                         goto out_put_keys;
948
949                 if (!fshared)
950                         goto retry_private;
951
952                 put_futex_key(fshared, &key2);
953                 put_futex_key(fshared, &key1);
954                 goto retry;
955         }
956
957         head = &hb1->chain;
958
959         plist_for_each_entry_safe(this, next, head, list) {
960                 if (match_futex (&this->key, &key1)) {
961                         wake_futex(this);
962                         if (++ret >= nr_wake)
963                                 break;
964                 }
965         }
966
967         if (op_ret > 0) {
968                 head = &hb2->chain;
969
970                 op_ret = 0;
971                 plist_for_each_entry_safe(this, next, head, list) {
972                         if (match_futex (&this->key, &key2)) {
973                                 wake_futex(this);
974                                 if (++op_ret >= nr_wake2)
975                                         break;
976                         }
977                 }
978                 ret += op_ret;
979         }
980
981         double_unlock_hb(hb1, hb2);
982 out_put_keys:
983         put_futex_key(fshared, &key2);
984 out_put_key1:
985         put_futex_key(fshared, &key1);
986 out:
987         return ret;
988 }
989
990 /**
991  * requeue_futex() - Requeue a futex_q from one hb to another
992  * @q:          the futex_q to requeue
993  * @hb1:        the source hash_bucket
994  * @hb2:        the target hash_bucket
995  * @key2:       the new key for the requeued futex_q
996  */
997 static inline
998 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
999                    struct futex_hash_bucket *hb2, union futex_key *key2)
1000 {
1001
1002         /*
1003          * If key1 and key2 hash to the same bucket, no need to
1004          * requeue.
1005          */
1006         if (likely(&hb1->chain != &hb2->chain)) {
1007                 plist_del(&q->list, &hb1->chain);
1008                 plist_add(&q->list, &hb2->chain);
1009                 q->lock_ptr = &hb2->lock;
1010 #ifdef CONFIG_DEBUG_PI_LIST
1011                 q->list.plist.spinlock = &hb2->lock;
1012 #endif
1013         }
1014         get_futex_key_refs(key2);
1015         q->key = *key2;
1016 }
1017
1018 /**
1019  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1020  * @q:          the futex_q
1021  * @key:        the key of the requeue target futex
1022  * @hb:         the hash_bucket of the requeue target futex
1023  *
1024  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1025  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1026  * to the requeue target futex so the waiter can detect the wakeup on the right
1027  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1028  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1029  * to protect access to the pi_state to fixup the owner later.  Must be called
1030  * with both q->lock_ptr and hb->lock held.
1031  */
1032 static inline
1033 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1034                            struct futex_hash_bucket *hb)
1035 {
1036         get_futex_key_refs(key);
1037         q->key = *key;
1038
1039         WARN_ON(plist_node_empty(&q->list));
1040         plist_del(&q->list, &q->list.plist);
1041
1042         WARN_ON(!q->rt_waiter);
1043         q->rt_waiter = NULL;
1044
1045         q->lock_ptr = &hb->lock;
1046 #ifdef CONFIG_DEBUG_PI_LIST
1047         q->list.plist.spinlock = &hb->lock;
1048 #endif
1049
1050         wake_up_state(q->task, TASK_NORMAL);
1051 }
1052
1053 /**
1054  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1055  * @pifutex:            the user address of the to futex
1056  * @hb1:                the from futex hash bucket, must be locked by the caller
1057  * @hb2:                the to futex hash bucket, must be locked by the caller
1058  * @key1:               the from futex key
1059  * @key2:               the to futex key
1060  * @ps:                 address to store the pi_state pointer
1061  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1062  *
1063  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1064  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1065  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1066  * hb1 and hb2 must be held by the caller.
1067  *
1068  * Returns:
1069  *  0 - failed to acquire the lock atomicly
1070  *  1 - acquired the lock
1071  * <0 - error
1072  */
1073 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1074                                  struct futex_hash_bucket *hb1,
1075                                  struct futex_hash_bucket *hb2,
1076                                  union futex_key *key1, union futex_key *key2,
1077                                  struct futex_pi_state **ps, int set_waiters)
1078 {
1079         struct futex_q *top_waiter = NULL;
1080         u32 curval;
1081         int ret;
1082
1083         if (get_futex_value_locked(&curval, pifutex))
1084                 return -EFAULT;
1085
1086         /*
1087          * Find the top_waiter and determine if there are additional waiters.
1088          * If the caller intends to requeue more than 1 waiter to pifutex,
1089          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1090          * as we have means to handle the possible fault.  If not, don't set
1091          * the bit unecessarily as it will force the subsequent unlock to enter
1092          * the kernel.
1093          */
1094         top_waiter = futex_top_waiter(hb1, key1);
1095
1096         /* There are no waiters, nothing for us to do. */
1097         if (!top_waiter)
1098                 return 0;
1099
1100         /* Ensure we requeue to the expected futex. */
1101         if (!match_futex(top_waiter->requeue_pi_key, key2))
1102                 return -EINVAL;
1103
1104         /*
1105          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1106          * the contended case or if set_waiters is 1.  The pi_state is returned
1107          * in ps in contended cases.
1108          */
1109         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1110                                    set_waiters);
1111         if (ret == 1)
1112                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1113
1114         return ret;
1115 }
1116
1117 /**
1118  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1119  * uaddr1:      source futex user address
1120  * uaddr2:      target futex user address
1121  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1122  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1123  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1124  *              pi futex (pi to pi requeue is not supported)
1125  *
1126  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1127  * uaddr2 atomically on behalf of the top waiter.
1128  *
1129  * Returns:
1130  * >=0 - on success, the number of tasks requeued or woken
1131  *  <0 - on error
1132  */
1133 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1134                          int nr_wake, int nr_requeue, u32 *cmpval,
1135                          int requeue_pi)
1136 {
1137         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1138         int drop_count = 0, task_count = 0, ret;
1139         struct futex_pi_state *pi_state = NULL;
1140         struct futex_hash_bucket *hb1, *hb2;
1141         struct plist_head *head1;
1142         struct futex_q *this, *next;
1143         u32 curval2;
1144
1145         if (requeue_pi) {
1146                 /*
1147                  * requeue_pi requires a pi_state, try to allocate it now
1148                  * without any locks in case it fails.
1149                  */
1150                 if (refill_pi_state_cache())
1151                         return -ENOMEM;
1152                 /*
1153                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1154                  * + nr_requeue, since it acquires the rt_mutex prior to
1155                  * returning to userspace, so as to not leave the rt_mutex with
1156                  * waiters and no owner.  However, second and third wake-ups
1157                  * cannot be predicted as they involve race conditions with the
1158                  * first wake and a fault while looking up the pi_state.  Both
1159                  * pthread_cond_signal() and pthread_cond_broadcast() should
1160                  * use nr_wake=1.
1161                  */
1162                 if (nr_wake != 1)
1163                         return -EINVAL;
1164         }
1165
1166 retry:
1167         if (pi_state != NULL) {
1168                 /*
1169                  * We will have to lookup the pi_state again, so free this one
1170                  * to keep the accounting correct.
1171                  */
1172                 free_pi_state(pi_state);
1173                 pi_state = NULL;
1174         }
1175
1176         ret = get_futex_key(uaddr1, fshared, &key1);
1177         if (unlikely(ret != 0))
1178                 goto out;
1179         ret = get_futex_key(uaddr2, fshared, &key2);
1180         if (unlikely(ret != 0))
1181                 goto out_put_key1;
1182
1183         hb1 = hash_futex(&key1);
1184         hb2 = hash_futex(&key2);
1185
1186 retry_private:
1187         double_lock_hb(hb1, hb2);
1188
1189         if (likely(cmpval != NULL)) {
1190                 u32 curval;
1191
1192                 ret = get_futex_value_locked(&curval, uaddr1);
1193
1194                 if (unlikely(ret)) {
1195                         double_unlock_hb(hb1, hb2);
1196
1197                         ret = get_user(curval, uaddr1);
1198                         if (ret)
1199                                 goto out_put_keys;
1200
1201                         if (!fshared)
1202                                 goto retry_private;
1203
1204                         put_futex_key(fshared, &key2);
1205                         put_futex_key(fshared, &key1);
1206                         goto retry;
1207                 }
1208                 if (curval != *cmpval) {
1209                         ret = -EAGAIN;
1210                         goto out_unlock;
1211                 }
1212         }
1213
1214         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1215                 /*
1216                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1217                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1218                  * bit.  We force this here where we are able to easily handle
1219                  * faults rather in the requeue loop below.
1220                  */
1221                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1222                                                  &key2, &pi_state, nr_requeue);
1223
1224                 /*
1225                  * At this point the top_waiter has either taken uaddr2 or is
1226                  * waiting on it.  If the former, then the pi_state will not
1227                  * exist yet, look it up one more time to ensure we have a
1228                  * reference to it.
1229                  */
1230                 if (ret == 1) {
1231                         WARN_ON(pi_state);
1232                         drop_count++;
1233                         task_count++;
1234                         ret = get_futex_value_locked(&curval2, uaddr2);
1235                         if (!ret)
1236                                 ret = lookup_pi_state(curval2, hb2, &key2,
1237                                                       &pi_state);
1238                 }
1239
1240                 switch (ret) {
1241                 case 0:
1242                         break;
1243                 case -EFAULT:
1244                         double_unlock_hb(hb1, hb2);
1245                         put_futex_key(fshared, &key2);
1246                         put_futex_key(fshared, &key1);
1247                         ret = fault_in_user_writeable(uaddr2);
1248                         if (!ret)
1249                                 goto retry;
1250                         goto out;
1251                 case -EAGAIN:
1252                         /* The owner was exiting, try again. */
1253                         double_unlock_hb(hb1, hb2);
1254                         put_futex_key(fshared, &key2);
1255                         put_futex_key(fshared, &key1);
1256                         cond_resched();
1257                         goto retry;
1258                 default:
1259                         goto out_unlock;
1260                 }
1261         }
1262
1263         head1 = &hb1->chain;
1264         plist_for_each_entry_safe(this, next, head1, list) {
1265                 if (task_count - nr_wake >= nr_requeue)
1266                         break;
1267
1268                 if (!match_futex(&this->key, &key1))
1269                         continue;
1270
1271                 /*
1272                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1273                  * be paired with each other and no other futex ops.
1274                  */
1275                 if ((requeue_pi && !this->rt_waiter) ||
1276                     (!requeue_pi && this->rt_waiter)) {
1277                         ret = -EINVAL;
1278                         break;
1279                 }
1280
1281                 /*
1282                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1283                  * lock, we already woke the top_waiter.  If not, it will be
1284                  * woken by futex_unlock_pi().
1285                  */
1286                 if (++task_count <= nr_wake && !requeue_pi) {
1287                         wake_futex(this);
1288                         continue;
1289                 }
1290
1291                 /* Ensure we requeue to the expected futex for requeue_pi. */
1292                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1293                         ret = -EINVAL;
1294                         break;
1295                 }
1296
1297                 /*
1298                  * Requeue nr_requeue waiters and possibly one more in the case
1299                  * of requeue_pi if we couldn't acquire the lock atomically.
1300                  */
1301                 if (requeue_pi) {
1302                         /* Prepare the waiter to take the rt_mutex. */
1303                         atomic_inc(&pi_state->refcount);
1304                         this->pi_state = pi_state;
1305                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1306                                                         this->rt_waiter,
1307                                                         this->task, 1);
1308                         if (ret == 1) {
1309                                 /* We got the lock. */
1310                                 requeue_pi_wake_futex(this, &key2, hb2);
1311                                 drop_count++;
1312                                 continue;
1313                         } else if (ret) {
1314                                 /* -EDEADLK */
1315                                 this->pi_state = NULL;
1316                                 free_pi_state(pi_state);
1317                                 goto out_unlock;
1318                         }
1319                 }
1320                 requeue_futex(this, hb1, hb2, &key2);
1321                 drop_count++;
1322         }
1323
1324 out_unlock:
1325         double_unlock_hb(hb1, hb2);
1326
1327         /*
1328          * drop_futex_key_refs() must be called outside the spinlocks. During
1329          * the requeue we moved futex_q's from the hash bucket at key1 to the
1330          * one at key2 and updated their key pointer.  We no longer need to
1331          * hold the references to key1.
1332          */
1333         while (--drop_count >= 0)
1334                 drop_futex_key_refs(&key1);
1335
1336 out_put_keys:
1337         put_futex_key(fshared, &key2);
1338 out_put_key1:
1339         put_futex_key(fshared, &key1);
1340 out:
1341         if (pi_state != NULL)
1342                 free_pi_state(pi_state);
1343         return ret ? ret : task_count;
1344 }
1345
1346 /* The key must be already stored in q->key. */
1347 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1348 {
1349         struct futex_hash_bucket *hb;
1350
1351         get_futex_key_refs(&q->key);
1352         hb = hash_futex(&q->key);
1353         q->lock_ptr = &hb->lock;
1354
1355         spin_lock(&hb->lock);
1356         return hb;
1357 }
1358
1359 static inline void
1360 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1361 {
1362         spin_unlock(&hb->lock);
1363         drop_futex_key_refs(&q->key);
1364 }
1365
1366 /**
1367  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1368  * @q:  The futex_q to enqueue
1369  * @hb: The destination hash bucket
1370  *
1371  * The hb->lock must be held by the caller, and is released here. A call to
1372  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1373  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1374  * or nothing if the unqueue is done as part of the wake process and the unqueue
1375  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1376  * an example).
1377  */
1378 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1379 {
1380         int prio;
1381
1382         /*
1383          * The priority used to register this element is
1384          * - either the real thread-priority for the real-time threads
1385          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1386          * - or MAX_RT_PRIO for non-RT threads.
1387          * Thus, all RT-threads are woken first in priority order, and
1388          * the others are woken last, in FIFO order.
1389          */
1390         prio = min(current->normal_prio, MAX_RT_PRIO);
1391
1392         plist_node_init(&q->list, prio);
1393 #ifdef CONFIG_DEBUG_PI_LIST
1394         q->list.plist.spinlock = &hb->lock;
1395 #endif
1396         plist_add(&q->list, &hb->chain);
1397         q->task = current;
1398         spin_unlock(&hb->lock);
1399 }
1400
1401 /**
1402  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1403  * @q:  The futex_q to unqueue
1404  *
1405  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1406  * be paired with exactly one earlier call to queue_me().
1407  *
1408  * Returns:
1409  *   1 - if the futex_q was still queued (and we removed unqueued it)
1410  *   0 - if the futex_q was already removed by the waking thread
1411  */
1412 static int unqueue_me(struct futex_q *q)
1413 {
1414         spinlock_t *lock_ptr;
1415         int ret = 0;
1416
1417         /* In the common case we don't take the spinlock, which is nice. */
1418 retry:
1419         lock_ptr = q->lock_ptr;
1420         barrier();
1421         if (lock_ptr != NULL) {
1422                 spin_lock(lock_ptr);
1423                 /*
1424                  * q->lock_ptr can change between reading it and
1425                  * spin_lock(), causing us to take the wrong lock.  This
1426                  * corrects the race condition.
1427                  *
1428                  * Reasoning goes like this: if we have the wrong lock,
1429                  * q->lock_ptr must have changed (maybe several times)
1430                  * between reading it and the spin_lock().  It can
1431                  * change again after the spin_lock() but only if it was
1432                  * already changed before the spin_lock().  It cannot,
1433                  * however, change back to the original value.  Therefore
1434                  * we can detect whether we acquired the correct lock.
1435                  */
1436                 if (unlikely(lock_ptr != q->lock_ptr)) {
1437                         spin_unlock(lock_ptr);
1438                         goto retry;
1439                 }
1440                 WARN_ON(plist_node_empty(&q->list));
1441                 plist_del(&q->list, &q->list.plist);
1442
1443                 BUG_ON(q->pi_state);
1444
1445                 spin_unlock(lock_ptr);
1446                 ret = 1;
1447         }
1448
1449         drop_futex_key_refs(&q->key);
1450         return ret;
1451 }
1452
1453 /*
1454  * PI futexes can not be requeued and must remove themself from the
1455  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1456  * and dropped here.
1457  */
1458 static void unqueue_me_pi(struct futex_q *q)
1459 {
1460         WARN_ON(plist_node_empty(&q->list));
1461         plist_del(&q->list, &q->list.plist);
1462
1463         BUG_ON(!q->pi_state);
1464         free_pi_state(q->pi_state);
1465         q->pi_state = NULL;
1466
1467         spin_unlock(q->lock_ptr);
1468
1469         drop_futex_key_refs(&q->key);
1470 }
1471
1472 /*
1473  * Fixup the pi_state owner with the new owner.
1474  *
1475  * Must be called with hash bucket lock held and mm->sem held for non
1476  * private futexes.
1477  */
1478 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1479                                 struct task_struct *newowner, int fshared)
1480 {
1481         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1482         struct futex_pi_state *pi_state = q->pi_state;
1483         struct task_struct *oldowner = pi_state->owner;
1484         u32 uval, curval, newval;
1485         int ret;
1486
1487         /* Owner died? */
1488         if (!pi_state->owner)
1489                 newtid |= FUTEX_OWNER_DIED;
1490
1491         /*
1492          * We are here either because we stole the rtmutex from the
1493          * pending owner or we are the pending owner which failed to
1494          * get the rtmutex. We have to replace the pending owner TID
1495          * in the user space variable. This must be atomic as we have
1496          * to preserve the owner died bit here.
1497          *
1498          * Note: We write the user space value _before_ changing the pi_state
1499          * because we can fault here. Imagine swapped out pages or a fork
1500          * that marked all the anonymous memory readonly for cow.
1501          *
1502          * Modifying pi_state _before_ the user space value would
1503          * leave the pi_state in an inconsistent state when we fault
1504          * here, because we need to drop the hash bucket lock to
1505          * handle the fault. This might be observed in the PID check
1506          * in lookup_pi_state.
1507          */
1508 retry:
1509         if (get_futex_value_locked(&uval, uaddr))
1510                 goto handle_fault;
1511
1512         while (1) {
1513                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1514
1515                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1516
1517                 if (curval == -EFAULT)
1518                         goto handle_fault;
1519                 if (curval == uval)
1520                         break;
1521                 uval = curval;
1522         }
1523
1524         /*
1525          * We fixed up user space. Now we need to fix the pi_state
1526          * itself.
1527          */
1528         if (pi_state->owner != NULL) {
1529                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1530                 WARN_ON(list_empty(&pi_state->list));
1531                 list_del_init(&pi_state->list);
1532                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1533         }
1534
1535         pi_state->owner = newowner;
1536
1537         raw_spin_lock_irq(&newowner->pi_lock);
1538         WARN_ON(!list_empty(&pi_state->list));
1539         list_add(&pi_state->list, &newowner->pi_state_list);
1540         raw_spin_unlock_irq(&newowner->pi_lock);
1541         return 0;
1542
1543         /*
1544          * To handle the page fault we need to drop the hash bucket
1545          * lock here. That gives the other task (either the pending
1546          * owner itself or the task which stole the rtmutex) the
1547          * chance to try the fixup of the pi_state. So once we are
1548          * back from handling the fault we need to check the pi_state
1549          * after reacquiring the hash bucket lock and before trying to
1550          * do another fixup. When the fixup has been done already we
1551          * simply return.
1552          */
1553 handle_fault:
1554         spin_unlock(q->lock_ptr);
1555
1556         ret = fault_in_user_writeable(uaddr);
1557
1558         spin_lock(q->lock_ptr);
1559
1560         /*
1561          * Check if someone else fixed it for us:
1562          */
1563         if (pi_state->owner != oldowner)
1564                 return 0;
1565
1566         if (ret)
1567                 return ret;
1568
1569         goto retry;
1570 }
1571
1572 /*
1573  * In case we must use restart_block to restart a futex_wait,
1574  * we encode in the 'flags' shared capability
1575  */
1576 #define FLAGS_SHARED            0x01
1577 #define FLAGS_CLOCKRT           0x02
1578 #define FLAGS_HAS_TIMEOUT       0x04
1579
1580 static long futex_wait_restart(struct restart_block *restart);
1581
1582 /**
1583  * fixup_owner() - Post lock pi_state and corner case management
1584  * @uaddr:      user address of the futex
1585  * @fshared:    whether the futex is shared (1) or not (0)
1586  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1587  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1588  *
1589  * After attempting to lock an rt_mutex, this function is called to cleanup
1590  * the pi_state owner as well as handle race conditions that may allow us to
1591  * acquire the lock. Must be called with the hb lock held.
1592  *
1593  * Returns:
1594  *  1 - success, lock taken
1595  *  0 - success, lock not taken
1596  * <0 - on error (-EFAULT)
1597  */
1598 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1599                        int locked)
1600 {
1601         struct task_struct *owner;
1602         int ret = 0;
1603
1604         if (locked) {
1605                 /*
1606                  * Got the lock. We might not be the anticipated owner if we
1607                  * did a lock-steal - fix up the PI-state in that case:
1608                  */
1609                 if (q->pi_state->owner != current)
1610                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1611                 goto out;
1612         }
1613
1614         /*
1615          * Catch the rare case, where the lock was released when we were on the
1616          * way back before we locked the hash bucket.
1617          */
1618         if (q->pi_state->owner == current) {
1619                 /*
1620                  * Try to get the rt_mutex now. This might fail as some other
1621                  * task acquired the rt_mutex after we removed ourself from the
1622                  * rt_mutex waiters list.
1623                  */
1624                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1625                         locked = 1;
1626                         goto out;
1627                 }
1628
1629                 /*
1630                  * pi_state is incorrect, some other task did a lock steal and
1631                  * we returned due to timeout or signal without taking the
1632                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1633                  * locking, as the other task is now blocked on the hash bucket
1634                  * lock. Fix the state up.
1635                  */
1636                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1637                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1638                 goto out;
1639         }
1640
1641         /*
1642          * Paranoia check. If we did not take the lock, then we should not be
1643          * the owner, nor the pending owner, of the rt_mutex.
1644          */
1645         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1646                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1647                                 "pi-state %p\n", ret,
1648                                 q->pi_state->pi_mutex.owner,
1649                                 q->pi_state->owner);
1650
1651 out:
1652         return ret ? ret : locked;
1653 }
1654
1655 /**
1656  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1657  * @hb:         the futex hash bucket, must be locked by the caller
1658  * @q:          the futex_q to queue up on
1659  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1660  */
1661 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1662                                 struct hrtimer_sleeper *timeout)
1663 {
1664         /*
1665          * The task state is guaranteed to be set before another task can
1666          * wake it. set_current_state() is implemented using set_mb() and
1667          * queue_me() calls spin_unlock() upon completion, both serializing
1668          * access to the hash list and forcing another memory barrier.
1669          */
1670         set_current_state(TASK_INTERRUPTIBLE);
1671         queue_me(q, hb);
1672
1673         /* Arm the timer */
1674         if (timeout) {
1675                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1676                 if (!hrtimer_active(&timeout->timer))
1677                         timeout->task = NULL;
1678         }
1679
1680         /*
1681          * If we have been removed from the hash list, then another task
1682          * has tried to wake us, and we can skip the call to schedule().
1683          */
1684         if (likely(!plist_node_empty(&q->list))) {
1685                 /*
1686                  * If the timer has already expired, current will already be
1687                  * flagged for rescheduling. Only call schedule if there
1688                  * is no timeout, or if it has yet to expire.
1689                  */
1690                 if (!timeout || timeout->task)
1691                         schedule();
1692         }
1693         __set_current_state(TASK_RUNNING);
1694 }
1695
1696 /**
1697  * futex_wait_setup() - Prepare to wait on a futex
1698  * @uaddr:      the futex userspace address
1699  * @val:        the expected value
1700  * @fshared:    whether the futex is shared (1) or not (0)
1701  * @q:          the associated futex_q
1702  * @hb:         storage for hash_bucket pointer to be returned to caller
1703  *
1704  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1705  * compare it with the expected value.  Handle atomic faults internally.
1706  * Return with the hb lock held and a q.key reference on success, and unlocked
1707  * with no q.key reference on failure.
1708  *
1709  * Returns:
1710  *  0 - uaddr contains val and hb has been locked
1711  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1712  */
1713 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1714                            struct futex_q *q, struct futex_hash_bucket **hb)
1715 {
1716         u32 uval;
1717         int ret;
1718
1719         /*
1720          * Access the page AFTER the hash-bucket is locked.
1721          * Order is important:
1722          *
1723          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1724          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1725          *
1726          * The basic logical guarantee of a futex is that it blocks ONLY
1727          * if cond(var) is known to be true at the time of blocking, for
1728          * any cond.  If we queued after testing *uaddr, that would open
1729          * a race condition where we could block indefinitely with
1730          * cond(var) false, which would violate the guarantee.
1731          *
1732          * A consequence is that futex_wait() can return zero and absorb
1733          * a wakeup when *uaddr != val on entry to the syscall.  This is
1734          * rare, but normal.
1735          */
1736 retry:
1737         q->key = FUTEX_KEY_INIT;
1738         ret = get_futex_key(uaddr, fshared, &q->key);
1739         if (unlikely(ret != 0))
1740                 return ret;
1741
1742 retry_private:
1743         *hb = queue_lock(q);
1744
1745         ret = get_futex_value_locked(&uval, uaddr);
1746
1747         if (ret) {
1748                 queue_unlock(q, *hb);
1749
1750                 ret = get_user(uval, uaddr);
1751                 if (ret)
1752                         goto out;
1753
1754                 if (!fshared)
1755                         goto retry_private;
1756
1757                 put_futex_key(fshared, &q->key);
1758                 goto retry;
1759         }
1760
1761         if (uval != val) {
1762                 queue_unlock(q, *hb);
1763                 ret = -EWOULDBLOCK;
1764         }
1765
1766 out:
1767         if (ret)
1768                 put_futex_key(fshared, &q->key);
1769         return ret;
1770 }
1771
1772 static int futex_wait(u32 __user *uaddr, int fshared,
1773                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1774 {
1775         struct hrtimer_sleeper timeout, *to = NULL;
1776         struct restart_block *restart;
1777         struct futex_hash_bucket *hb;
1778         struct futex_q q;
1779         int ret;
1780
1781         if (!bitset)
1782                 return -EINVAL;
1783
1784         q.pi_state = NULL;
1785         q.bitset = bitset;
1786         q.rt_waiter = NULL;
1787         q.requeue_pi_key = NULL;
1788
1789         if (abs_time) {
1790                 to = &timeout;
1791
1792                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1793                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1794                 hrtimer_init_sleeper(to, current);
1795                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1796                                              current->timer_slack_ns);
1797         }
1798
1799 retry:
1800         /* Prepare to wait on uaddr. */
1801         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1802         if (ret)
1803                 goto out;
1804
1805         /* queue_me and wait for wakeup, timeout, or a signal. */
1806         futex_wait_queue_me(hb, &q, to);
1807
1808         /* If we were woken (and unqueued), we succeeded, whatever. */
1809         ret = 0;
1810         if (!unqueue_me(&q))
1811                 goto out_put_key;
1812         ret = -ETIMEDOUT;
1813         if (to && !to->task)
1814                 goto out_put_key;
1815
1816         /*
1817          * We expect signal_pending(current), but we might be the
1818          * victim of a spurious wakeup as well.
1819          */
1820         if (!signal_pending(current)) {
1821                 put_futex_key(fshared, &q.key);
1822                 goto retry;
1823         }
1824
1825         ret = -ERESTARTSYS;
1826         if (!abs_time)
1827                 goto out_put_key;
1828
1829         restart = &current_thread_info()->restart_block;
1830         restart->fn = futex_wait_restart;
1831         restart->futex.uaddr = (u32 *)uaddr;
1832         restart->futex.val = val;
1833         restart->futex.time = abs_time->tv64;
1834         restart->futex.bitset = bitset;
1835         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1836
1837         if (fshared)
1838                 restart->futex.flags |= FLAGS_SHARED;
1839         if (clockrt)
1840                 restart->futex.flags |= FLAGS_CLOCKRT;
1841
1842         ret = -ERESTART_RESTARTBLOCK;
1843
1844 out_put_key:
1845         put_futex_key(fshared, &q.key);
1846 out:
1847         if (to) {
1848                 hrtimer_cancel(&to->timer);
1849                 destroy_hrtimer_on_stack(&to->timer);
1850         }
1851         return ret;
1852 }
1853
1854
1855 static long futex_wait_restart(struct restart_block *restart)
1856 {
1857         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1858         int fshared = 0;
1859         ktime_t t, *tp = NULL;
1860
1861         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1862                 t.tv64 = restart->futex.time;
1863                 tp = &t;
1864         }
1865         restart->fn = do_no_restart_syscall;
1866         if (restart->futex.flags & FLAGS_SHARED)
1867                 fshared = 1;
1868         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1869                                 restart->futex.bitset,
1870                                 restart->futex.flags & FLAGS_CLOCKRT);
1871 }
1872
1873
1874 /*
1875  * Userspace tried a 0 -> TID atomic transition of the futex value
1876  * and failed. The kernel side here does the whole locking operation:
1877  * if there are waiters then it will block, it does PI, etc. (Due to
1878  * races the kernel might see a 0 value of the futex too.)
1879  */
1880 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1881                          int detect, ktime_t *time, int trylock)
1882 {
1883         struct hrtimer_sleeper timeout, *to = NULL;
1884         struct futex_hash_bucket *hb;
1885         struct futex_q q;
1886         int res, ret;
1887
1888         if (refill_pi_state_cache())
1889                 return -ENOMEM;
1890
1891         if (time) {
1892                 to = &timeout;
1893                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1894                                       HRTIMER_MODE_ABS);
1895                 hrtimer_init_sleeper(to, current);
1896                 hrtimer_set_expires(&to->timer, *time);
1897         }
1898
1899         q.pi_state = NULL;
1900         q.rt_waiter = NULL;
1901         q.requeue_pi_key = NULL;
1902 retry:
1903         q.key = FUTEX_KEY_INIT;
1904         ret = get_futex_key(uaddr, fshared, &q.key);
1905         if (unlikely(ret != 0))
1906                 goto out;
1907
1908 retry_private:
1909         hb = queue_lock(&q);
1910
1911         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1912         if (unlikely(ret)) {
1913                 switch (ret) {
1914                 case 1:
1915                         /* We got the lock. */
1916                         ret = 0;
1917                         goto out_unlock_put_key;
1918                 case -EFAULT:
1919                         goto uaddr_faulted;
1920                 case -EAGAIN:
1921                         /*
1922                          * Task is exiting and we just wait for the
1923                          * exit to complete.
1924                          */
1925                         queue_unlock(&q, hb);
1926                         put_futex_key(fshared, &q.key);
1927                         cond_resched();
1928                         goto retry;
1929                 default:
1930                         goto out_unlock_put_key;
1931                 }
1932         }
1933
1934         /*
1935          * Only actually queue now that the atomic ops are done:
1936          */
1937         queue_me(&q, hb);
1938
1939         WARN_ON(!q.pi_state);
1940         /*
1941          * Block on the PI mutex:
1942          */
1943         if (!trylock)
1944                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1945         else {
1946                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1947                 /* Fixup the trylock return value: */
1948                 ret = ret ? 0 : -EWOULDBLOCK;
1949         }
1950
1951         spin_lock(q.lock_ptr);
1952         /*
1953          * Fixup the pi_state owner and possibly acquire the lock if we
1954          * haven't already.
1955          */
1956         res = fixup_owner(uaddr, fshared, &q, !ret);
1957         /*
1958          * If fixup_owner() returned an error, proprogate that.  If it acquired
1959          * the lock, clear our -ETIMEDOUT or -EINTR.
1960          */
1961         if (res)
1962                 ret = (res < 0) ? res : 0;
1963
1964         /*
1965          * If fixup_owner() faulted and was unable to handle the fault, unlock
1966          * it and return the fault to userspace.
1967          */
1968         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1969                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1970
1971         /* Unqueue and drop the lock */
1972         unqueue_me_pi(&q);
1973
1974         goto out;
1975
1976 out_unlock_put_key:
1977         queue_unlock(&q, hb);
1978
1979 out_put_key:
1980         put_futex_key(fshared, &q.key);
1981 out:
1982         if (to)
1983                 destroy_hrtimer_on_stack(&to->timer);
1984         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1985
1986 uaddr_faulted:
1987         queue_unlock(&q, hb);
1988
1989         ret = fault_in_user_writeable(uaddr);
1990         if (ret)
1991                 goto out_put_key;
1992
1993         if (!fshared)
1994                 goto retry_private;
1995
1996         put_futex_key(fshared, &q.key);
1997         goto retry;
1998 }
1999
2000 /*
2001  * Userspace attempted a TID -> 0 atomic transition, and failed.
2002  * This is the in-kernel slowpath: we look up the PI state (if any),
2003  * and do the rt-mutex unlock.
2004  */
2005 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2006 {
2007         struct futex_hash_bucket *hb;
2008         struct futex_q *this, *next;
2009         u32 uval;
2010         struct plist_head *head;
2011         union futex_key key = FUTEX_KEY_INIT;
2012         int ret;
2013
2014 retry:
2015         if (get_user(uval, uaddr))
2016                 return -EFAULT;
2017         /*
2018          * We release only a lock we actually own:
2019          */
2020         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2021                 return -EPERM;
2022
2023         ret = get_futex_key(uaddr, fshared, &key);
2024         if (unlikely(ret != 0))
2025                 goto out;
2026
2027         hb = hash_futex(&key);
2028         spin_lock(&hb->lock);
2029
2030         /*
2031          * To avoid races, try to do the TID -> 0 atomic transition
2032          * again. If it succeeds then we can return without waking
2033          * anyone else up:
2034          */
2035         if (!(uval & FUTEX_OWNER_DIED))
2036                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2037
2038
2039         if (unlikely(uval == -EFAULT))
2040                 goto pi_faulted;
2041         /*
2042          * Rare case: we managed to release the lock atomically,
2043          * no need to wake anyone else up:
2044          */
2045         if (unlikely(uval == task_pid_vnr(current)))
2046                 goto out_unlock;
2047
2048         /*
2049          * Ok, other tasks may need to be woken up - check waiters
2050          * and do the wakeup if necessary:
2051          */
2052         head = &hb->chain;
2053
2054         plist_for_each_entry_safe(this, next, head, list) {
2055                 if (!match_futex (&this->key, &key))
2056                         continue;
2057                 ret = wake_futex_pi(uaddr, uval, this);
2058                 /*
2059                  * The atomic access to the futex value
2060                  * generated a pagefault, so retry the
2061                  * user-access and the wakeup:
2062                  */
2063                 if (ret == -EFAULT)
2064                         goto pi_faulted;
2065                 goto out_unlock;
2066         }
2067         /*
2068          * No waiters - kernel unlocks the futex:
2069          */
2070         if (!(uval & FUTEX_OWNER_DIED)) {
2071                 ret = unlock_futex_pi(uaddr, uval);
2072                 if (ret == -EFAULT)
2073                         goto pi_faulted;
2074         }
2075
2076 out_unlock:
2077         spin_unlock(&hb->lock);
2078         put_futex_key(fshared, &key);
2079
2080 out:
2081         return ret;
2082
2083 pi_faulted:
2084         spin_unlock(&hb->lock);
2085         put_futex_key(fshared, &key);
2086
2087         ret = fault_in_user_writeable(uaddr);
2088         if (!ret)
2089                 goto retry;
2090
2091         return ret;
2092 }
2093
2094 /**
2095  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2096  * @hb:         the hash_bucket futex_q was original enqueued on
2097  * @q:          the futex_q woken while waiting to be requeued
2098  * @key2:       the futex_key of the requeue target futex
2099  * @timeout:    the timeout associated with the wait (NULL if none)
2100  *
2101  * Detect if the task was woken on the initial futex as opposed to the requeue
2102  * target futex.  If so, determine if it was a timeout or a signal that caused
2103  * the wakeup and return the appropriate error code to the caller.  Must be
2104  * called with the hb lock held.
2105  *
2106  * Returns
2107  *  0 - no early wakeup detected
2108  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2109  */
2110 static inline
2111 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2112                                    struct futex_q *q, union futex_key *key2,
2113                                    struct hrtimer_sleeper *timeout)
2114 {
2115         int ret = 0;
2116
2117         /*
2118          * With the hb lock held, we avoid races while we process the wakeup.
2119          * We only need to hold hb (and not hb2) to ensure atomicity as the
2120          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2121          * It can't be requeued from uaddr2 to something else since we don't
2122          * support a PI aware source futex for requeue.
2123          */
2124         if (!match_futex(&q->key, key2)) {
2125                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2126                 /*
2127                  * We were woken prior to requeue by a timeout or a signal.
2128                  * Unqueue the futex_q and determine which it was.
2129                  */
2130                 plist_del(&q->list, &q->list.plist);
2131
2132                 /* Handle spurious wakeups gracefully */
2133                 ret = -EWOULDBLOCK;
2134                 if (timeout && !timeout->task)
2135                         ret = -ETIMEDOUT;
2136                 else if (signal_pending(current))
2137                         ret = -ERESTARTNOINTR;
2138         }
2139         return ret;
2140 }
2141
2142 /**
2143  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2144  * @uaddr:      the futex we initially wait on (non-pi)
2145  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2146  *              the same type, no requeueing from private to shared, etc.
2147  * @val:        the expected value of uaddr
2148  * @abs_time:   absolute timeout
2149  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2150  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2151  * @uaddr2:     the pi futex we will take prior to returning to user-space
2152  *
2153  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2154  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2155  * complete the acquisition of the rt_mutex prior to returning to userspace.
2156  * This ensures the rt_mutex maintains an owner when it has waiters; without
2157  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2158  * need to.
2159  *
2160  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2161  * via the following:
2162  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2163  * 2) wakeup on uaddr2 after a requeue
2164  * 3) signal
2165  * 4) timeout
2166  *
2167  * If 3, cleanup and return -ERESTARTNOINTR.
2168  *
2169  * If 2, we may then block on trying to take the rt_mutex and return via:
2170  * 5) successful lock
2171  * 6) signal
2172  * 7) timeout
2173  * 8) other lock acquisition failure
2174  *
2175  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2176  *
2177  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2178  *
2179  * Returns:
2180  *  0 - On success
2181  * <0 - On error
2182  */
2183 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2184                                  u32 val, ktime_t *abs_time, u32 bitset,
2185                                  int clockrt, u32 __user *uaddr2)
2186 {
2187         struct hrtimer_sleeper timeout, *to = NULL;
2188         struct rt_mutex_waiter rt_waiter;
2189         struct rt_mutex *pi_mutex = NULL;
2190         struct futex_hash_bucket *hb;
2191         union futex_key key2;
2192         struct futex_q q;
2193         int res, ret;
2194
2195         if (!bitset)
2196                 return -EINVAL;
2197
2198         if (abs_time) {
2199                 to = &timeout;
2200                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2201                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2202                 hrtimer_init_sleeper(to, current);
2203                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2204                                              current->timer_slack_ns);
2205         }
2206
2207         /*
2208          * The waiter is allocated on our stack, manipulated by the requeue
2209          * code while we sleep on uaddr.
2210          */
2211         debug_rt_mutex_init_waiter(&rt_waiter);
2212         rt_waiter.task = NULL;
2213
2214         key2 = FUTEX_KEY_INIT;
2215         ret = get_futex_key(uaddr2, fshared, &key2);
2216         if (unlikely(ret != 0))
2217                 goto out;
2218
2219         q.pi_state = NULL;
2220         q.bitset = bitset;
2221         q.rt_waiter = &rt_waiter;
2222         q.requeue_pi_key = &key2;
2223
2224         /* Prepare to wait on uaddr. */
2225         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2226         if (ret)
2227                 goto out_key2;
2228
2229         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2230         futex_wait_queue_me(hb, &q, to);
2231
2232         spin_lock(&hb->lock);
2233         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2234         spin_unlock(&hb->lock);
2235         if (ret)
2236                 goto out_put_keys;
2237
2238         /*
2239          * In order for us to be here, we know our q.key == key2, and since
2240          * we took the hb->lock above, we also know that futex_requeue() has
2241          * completed and we no longer have to concern ourselves with a wakeup
2242          * race with the atomic proxy lock acquition by the requeue code.
2243          */
2244
2245         /* Check if the requeue code acquired the second futex for us. */
2246         if (!q.rt_waiter) {
2247                 /*
2248                  * Got the lock. We might not be the anticipated owner if we
2249                  * did a lock-steal - fix up the PI-state in that case.
2250                  */
2251                 if (q.pi_state && (q.pi_state->owner != current)) {
2252                         spin_lock(q.lock_ptr);
2253                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2254                                                    fshared);
2255                         spin_unlock(q.lock_ptr);
2256                 }
2257         } else {
2258                 /*
2259                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2260                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2261                  * the pi_state.
2262                  */
2263                 WARN_ON(!&q.pi_state);
2264                 pi_mutex = &q.pi_state->pi_mutex;
2265                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2266                 debug_rt_mutex_free_waiter(&rt_waiter);
2267
2268                 spin_lock(q.lock_ptr);
2269                 /*
2270                  * Fixup the pi_state owner and possibly acquire the lock if we
2271                  * haven't already.
2272                  */
2273                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2274                 /*
2275                  * If fixup_owner() returned an error, proprogate that.  If it
2276                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2277                  */
2278                 if (res)
2279                         ret = (res < 0) ? res : 0;
2280
2281                 /* Unqueue and drop the lock. */
2282                 unqueue_me_pi(&q);
2283         }
2284
2285         /*
2286          * If fixup_pi_state_owner() faulted and was unable to handle the
2287          * fault, unlock the rt_mutex and return the fault to userspace.
2288          */
2289         if (ret == -EFAULT) {
2290                 if (rt_mutex_owner(pi_mutex) == current)
2291                         rt_mutex_unlock(pi_mutex);
2292         } else if (ret == -EINTR) {
2293                 /*
2294                  * We've already been requeued, but cannot restart by calling
2295                  * futex_lock_pi() directly. We could restart this syscall, but
2296                  * it would detect that the user space "val" changed and return
2297                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2298                  * -EWOULDBLOCK directly.
2299                  */
2300                 ret = -EWOULDBLOCK;
2301         }
2302
2303 out_put_keys:
2304         put_futex_key(fshared, &q.key);
2305 out_key2:
2306         put_futex_key(fshared, &key2);
2307
2308 out:
2309         if (to) {
2310                 hrtimer_cancel(&to->timer);
2311                 destroy_hrtimer_on_stack(&to->timer);
2312         }
2313         return ret;
2314 }
2315
2316 /*
2317  * Support for robust futexes: the kernel cleans up held futexes at
2318  * thread exit time.
2319  *
2320  * Implementation: user-space maintains a per-thread list of locks it
2321  * is holding. Upon do_exit(), the kernel carefully walks this list,
2322  * and marks all locks that are owned by this thread with the
2323  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2324  * always manipulated with the lock held, so the list is private and
2325  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2326  * field, to allow the kernel to clean up if the thread dies after
2327  * acquiring the lock, but just before it could have added itself to
2328  * the list. There can only be one such pending lock.
2329  */
2330
2331 /**
2332  * sys_set_robust_list() - Set the robust-futex list head of a task
2333  * @head:       pointer to the list-head
2334  * @len:        length of the list-head, as userspace expects
2335  */
2336 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2337                 size_t, len)
2338 {
2339         if (!futex_cmpxchg_enabled)
2340                 return -ENOSYS;
2341         /*
2342          * The kernel knows only one size for now:
2343          */
2344         if (unlikely(len != sizeof(*head)))
2345                 return -EINVAL;
2346
2347         current->robust_list = head;
2348
2349         return 0;
2350 }
2351
2352 /**
2353  * sys_get_robust_list() - Get the robust-futex list head of a task
2354  * @pid:        pid of the process [zero for current task]
2355  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2356  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2357  */
2358 SYSCALL_DEFINE3(get_robust_list, int, pid,
2359                 struct robust_list_head __user * __user *, head_ptr,
2360                 size_t __user *, len_ptr)
2361 {
2362         struct robust_list_head __user *head;
2363         unsigned long ret;
2364         const struct cred *cred = current_cred(), *pcred;
2365
2366         if (!futex_cmpxchg_enabled)
2367                 return -ENOSYS;
2368
2369         if (!pid)
2370                 head = current->robust_list;
2371         else {
2372                 struct task_struct *p;
2373
2374                 ret = -ESRCH;
2375                 rcu_read_lock();
2376                 p = find_task_by_vpid(pid);
2377                 if (!p)
2378                         goto err_unlock;
2379                 ret = -EPERM;
2380                 pcred = __task_cred(p);
2381                 if (cred->euid != pcred->euid &&
2382                     cred->euid != pcred->uid &&
2383                     !capable(CAP_SYS_PTRACE))
2384                         goto err_unlock;
2385                 head = p->robust_list;
2386                 rcu_read_unlock();
2387         }
2388
2389         if (put_user(sizeof(*head), len_ptr))
2390                 return -EFAULT;
2391         return put_user(head, head_ptr);
2392
2393 err_unlock:
2394         rcu_read_unlock();
2395
2396         return ret;
2397 }
2398
2399 /*
2400  * Process a futex-list entry, check whether it's owned by the
2401  * dying task, and do notification if so:
2402  */
2403 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2404 {
2405         u32 uval, nval, mval;
2406
2407 retry:
2408         if (get_user(uval, uaddr))
2409                 return -1;
2410
2411         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2412                 /*
2413                  * Ok, this dying thread is truly holding a futex
2414                  * of interest. Set the OWNER_DIED bit atomically
2415                  * via cmpxchg, and if the value had FUTEX_WAITERS
2416                  * set, wake up a waiter (if any). (We have to do a
2417                  * futex_wake() even if OWNER_DIED is already set -
2418                  * to handle the rare but possible case of recursive
2419                  * thread-death.) The rest of the cleanup is done in
2420                  * userspace.
2421                  */
2422                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2423                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2424
2425                 if (nval == -EFAULT)
2426                         return -1;
2427
2428                 if (nval != uval)
2429                         goto retry;
2430
2431                 /*
2432                  * Wake robust non-PI futexes here. The wakeup of
2433                  * PI futexes happens in exit_pi_state():
2434                  */
2435                 if (!pi && (uval & FUTEX_WAITERS))
2436                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2437         }
2438         return 0;
2439 }
2440
2441 /*
2442  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2443  */
2444 static inline int fetch_robust_entry(struct robust_list __user **entry,
2445                                      struct robust_list __user * __user *head,
2446                                      int *pi)
2447 {
2448         unsigned long uentry;
2449
2450         if (get_user(uentry, (unsigned long __user *)head))
2451                 return -EFAULT;
2452
2453         *entry = (void __user *)(uentry & ~1UL);
2454         *pi = uentry & 1;
2455
2456         return 0;
2457 }
2458
2459 /*
2460  * Walk curr->robust_list (very carefully, it's a userspace list!)
2461  * and mark any locks found there dead, and notify any waiters.
2462  *
2463  * We silently return on any sign of list-walking problem.
2464  */
2465 void exit_robust_list(struct task_struct *curr)
2466 {
2467         struct robust_list_head __user *head = curr->robust_list;
2468         struct robust_list __user *entry, *next_entry, *pending;
2469         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2470         unsigned long futex_offset;
2471         int rc;
2472
2473         if (!futex_cmpxchg_enabled)
2474                 return;
2475
2476         /*
2477          * Fetch the list head (which was registered earlier, via
2478          * sys_set_robust_list()):
2479          */
2480         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2481                 return;
2482         /*
2483          * Fetch the relative futex offset:
2484          */
2485         if (get_user(futex_offset, &head->futex_offset))
2486                 return;
2487         /*
2488          * Fetch any possibly pending lock-add first, and handle it
2489          * if it exists:
2490          */
2491         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2492                 return;
2493
2494         next_entry = NULL;      /* avoid warning with gcc */
2495         while (entry != &head->list) {
2496                 /*
2497                  * Fetch the next entry in the list before calling
2498                  * handle_futex_death:
2499                  */
2500                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2501                 /*
2502                  * A pending lock might already be on the list, so
2503                  * don't process it twice:
2504                  */
2505                 if (entry != pending)
2506                         if (handle_futex_death((void __user *)entry + futex_offset,
2507                                                 curr, pi))
2508                                 return;
2509                 if (rc)
2510                         return;
2511                 entry = next_entry;
2512                 pi = next_pi;
2513                 /*
2514                  * Avoid excessively long or circular lists:
2515                  */
2516                 if (!--limit)
2517                         break;
2518
2519                 cond_resched();
2520         }
2521
2522         if (pending)
2523                 handle_futex_death((void __user *)pending + futex_offset,
2524                                    curr, pip);
2525 }
2526
2527 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2528                 u32 __user *uaddr2, u32 val2, u32 val3)
2529 {
2530         int clockrt, ret = -ENOSYS;
2531         int cmd = op & FUTEX_CMD_MASK;
2532         int fshared = 0;
2533
2534         if (!(op & FUTEX_PRIVATE_FLAG))
2535                 fshared = 1;
2536
2537         clockrt = op & FUTEX_CLOCK_REALTIME;
2538         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2539                 return -ENOSYS;
2540
2541         switch (cmd) {
2542         case FUTEX_WAIT:
2543                 val3 = FUTEX_BITSET_MATCH_ANY;
2544         case FUTEX_WAIT_BITSET:
2545                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2546                 break;
2547         case FUTEX_WAKE:
2548                 val3 = FUTEX_BITSET_MATCH_ANY;
2549         case FUTEX_WAKE_BITSET:
2550                 ret = futex_wake(uaddr, fshared, val, val3);
2551                 break;
2552         case FUTEX_REQUEUE:
2553                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2554                 break;
2555         case FUTEX_CMP_REQUEUE:
2556                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2557                                     0);
2558                 break;
2559         case FUTEX_WAKE_OP:
2560                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2561                 break;
2562         case FUTEX_LOCK_PI:
2563                 if (futex_cmpxchg_enabled)
2564                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2565                 break;
2566         case FUTEX_UNLOCK_PI:
2567                 if (futex_cmpxchg_enabled)
2568                         ret = futex_unlock_pi(uaddr, fshared);
2569                 break;
2570         case FUTEX_TRYLOCK_PI:
2571                 if (futex_cmpxchg_enabled)
2572                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2573                 break;
2574         case FUTEX_WAIT_REQUEUE_PI:
2575                 val3 = FUTEX_BITSET_MATCH_ANY;
2576                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2577                                             clockrt, uaddr2);
2578                 break;
2579         case FUTEX_CMP_REQUEUE_PI:
2580                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2581                                     1);
2582                 break;
2583         default:
2584                 ret = -ENOSYS;
2585         }
2586         return ret;
2587 }
2588
2589
2590 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2591                 struct timespec __user *, utime, u32 __user *, uaddr2,
2592                 u32, val3)
2593 {
2594         struct timespec ts;
2595         ktime_t t, *tp = NULL;
2596         u32 val2 = 0;
2597         int cmd = op & FUTEX_CMD_MASK;
2598
2599         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2600                       cmd == FUTEX_WAIT_BITSET ||
2601                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2602                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2603                         return -EFAULT;
2604                 if (!timespec_valid(&ts))
2605                         return -EINVAL;
2606
2607                 t = timespec_to_ktime(ts);
2608                 if (cmd == FUTEX_WAIT)
2609                         t = ktime_add_safe(ktime_get(), t);
2610                 tp = &t;
2611         }
2612         /*
2613          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2614          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2615          */
2616         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2617             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2618                 val2 = (u32) (unsigned long) utime;
2619
2620         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2621 }
2622
2623 static int __init futex_init(void)
2624 {
2625         u32 curval;
2626         int i;
2627
2628         /*
2629          * This will fail and we want it. Some arch implementations do
2630          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2631          * functionality. We want to know that before we call in any
2632          * of the complex code paths. Also we want to prevent
2633          * registration of robust lists in that case. NULL is
2634          * guaranteed to fault and we get -EFAULT on functional
2635          * implementation, the non functional ones will return
2636          * -ENOSYS.
2637          */
2638         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2639         if (curval == -EFAULT)
2640                 futex_cmpxchg_enabled = 1;
2641
2642         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2643                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2644                 spin_lock_init(&futex_queues[i].lock);
2645         }
2646
2647         return 0;
2648 }
2649 __initcall(futex_init);