6c498b145a4f325da3731802381586b44b2ce7f7
[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  * We use this hashed waitqueue instead of a normal wait_queue_t, so
94  * we can wake only the relevant ones (hashed queues may be shared).
95  *
96  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
97  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
98  * The order of wakup is always to make the first condition true, then
99  * wake up q->waiter, then make the second condition true.
100  */
101 struct futex_q {
102         struct plist_node list;
103         /* Waiter reference */
104         struct task_struct *task;
105
106         /* Which hash list lock to use: */
107         spinlock_t *lock_ptr;
108
109         /* Key which the futex is hashed on: */
110         union futex_key key;
111
112         /* Optional priority inheritance state: */
113         struct futex_pi_state *pi_state;
114
115         /* rt_waiter storage for requeue_pi: */
116         struct rt_mutex_waiter *rt_waiter;
117
118         /* The expected requeue pi target futex key: */
119         union futex_key *requeue_pi_key;
120
121         /* Bitset for the optional bitmasked wakeup */
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->both.word == key2->both.word
154                 && key1->both.ptr == key2->both.ptr
155                 && key1->both.offset == key2->both.offset);
156 }
157
158 /*
159  * Take a reference to the resource addressed by a key.
160  * Can be called while holding spinlocks.
161  *
162  */
163 static void get_futex_key_refs(union futex_key *key)
164 {
165         if (!key->both.ptr)
166                 return;
167
168         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
169         case FUT_OFF_INODE:
170                 atomic_inc(&key->shared.inode->i_count);
171                 break;
172         case FUT_OFF_MMSHARED:
173                 atomic_inc(&key->private.mm->mm_count);
174                 break;
175         }
176 }
177
178 /*
179  * Drop a reference to the resource addressed by a key.
180  * The hash bucket spinlock must not be held.
181  */
182 static void drop_futex_key_refs(union futex_key *key)
183 {
184         if (!key->both.ptr) {
185                 /* If we're here then we tried to put a key we failed to get */
186                 WARN_ON_ONCE(1);
187                 return;
188         }
189
190         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
191         case FUT_OFF_INODE:
192                 iput(key->shared.inode);
193                 break;
194         case FUT_OFF_MMSHARED:
195                 mmdrop(key->private.mm);
196                 break;
197         }
198 }
199
200 /**
201  * get_futex_key - Get parameters which are the keys for a futex.
202  * @uaddr: virtual address of the futex
203  * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
204  * @key: address where result is stored.
205  * @rw: mapping needs to be read/write (values: VERIFY_READ, VERIFY_WRITE)
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, int rw)
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(rw, 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, rw == VERIFY_WRITE, &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         int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
306                                  1, 1, 0, NULL, NULL);
307         return ret < 0 ? ret : 0;
308 }
309
310 /**
311  * futex_top_waiter() - Return the highest priority waiter on a futex
312  * @hb:     the hash bucket the futex_q's reside in
313  * @key:    the futex key (to distinguish it from other futex futex_q's)
314  *
315  * Must be called with the hb lock held.
316  */
317 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
318                                         union futex_key *key)
319 {
320         struct futex_q *this;
321
322         plist_for_each_entry(this, &hb->chain, list) {
323                 if (match_futex(&this->key, key))
324                         return this;
325         }
326         return NULL;
327 }
328
329 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
330 {
331         u32 curval;
332
333         pagefault_disable();
334         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
335         pagefault_enable();
336
337         return curval;
338 }
339
340 static int get_futex_value_locked(u32 *dest, u32 __user *from)
341 {
342         int ret;
343
344         pagefault_disable();
345         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
346         pagefault_enable();
347
348         return ret ? -EFAULT : 0;
349 }
350
351
352 /*
353  * PI code:
354  */
355 static int refill_pi_state_cache(void)
356 {
357         struct futex_pi_state *pi_state;
358
359         if (likely(current->pi_state_cache))
360                 return 0;
361
362         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
363
364         if (!pi_state)
365                 return -ENOMEM;
366
367         INIT_LIST_HEAD(&pi_state->list);
368         /* pi_mutex gets initialized later */
369         pi_state->owner = NULL;
370         atomic_set(&pi_state->refcount, 1);
371         pi_state->key = FUTEX_KEY_INIT;
372
373         current->pi_state_cache = pi_state;
374
375         return 0;
376 }
377
378 static struct futex_pi_state * alloc_pi_state(void)
379 {
380         struct futex_pi_state *pi_state = current->pi_state_cache;
381
382         WARN_ON(!pi_state);
383         current->pi_state_cache = NULL;
384
385         return pi_state;
386 }
387
388 static void free_pi_state(struct futex_pi_state *pi_state)
389 {
390         if (!atomic_dec_and_test(&pi_state->refcount))
391                 return;
392
393         /*
394          * If pi_state->owner is NULL, the owner is most probably dying
395          * and has cleaned up the pi_state already
396          */
397         if (pi_state->owner) {
398                 spin_lock_irq(&pi_state->owner->pi_lock);
399                 list_del_init(&pi_state->list);
400                 spin_unlock_irq(&pi_state->owner->pi_lock);
401
402                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
403         }
404
405         if (current->pi_state_cache)
406                 kfree(pi_state);
407         else {
408                 /*
409                  * pi_state->list is already empty.
410                  * clear pi_state->owner.
411                  * refcount is at 0 - put it back to 1.
412                  */
413                 pi_state->owner = NULL;
414                 atomic_set(&pi_state->refcount, 1);
415                 current->pi_state_cache = pi_state;
416         }
417 }
418
419 /*
420  * Look up the task based on what TID userspace gave us.
421  * We dont trust it.
422  */
423 static struct task_struct * futex_find_get_task(pid_t pid)
424 {
425         struct task_struct *p;
426         const struct cred *cred = current_cred(), *pcred;
427
428         rcu_read_lock();
429         p = find_task_by_vpid(pid);
430         if (!p) {
431                 p = ERR_PTR(-ESRCH);
432         } else {
433                 pcred = __task_cred(p);
434                 if (cred->euid != pcred->euid &&
435                     cred->euid != pcred->uid)
436                         p = ERR_PTR(-ESRCH);
437                 else
438                         get_task_struct(p);
439         }
440
441         rcu_read_unlock();
442
443         return p;
444 }
445
446 /*
447  * This task is holding PI mutexes at exit time => bad.
448  * Kernel cleans up PI-state, but userspace is likely hosed.
449  * (Robust-futex cleanup is separate and might save the day for userspace.)
450  */
451 void exit_pi_state_list(struct task_struct *curr)
452 {
453         struct list_head *next, *head = &curr->pi_state_list;
454         struct futex_pi_state *pi_state;
455         struct futex_hash_bucket *hb;
456         union futex_key key = FUTEX_KEY_INIT;
457
458         if (!futex_cmpxchg_enabled)
459                 return;
460         /*
461          * We are a ZOMBIE and nobody can enqueue itself on
462          * pi_state_list anymore, but we have to be careful
463          * versus waiters unqueueing themselves:
464          */
465         spin_lock_irq(&curr->pi_lock);
466         while (!list_empty(head)) {
467
468                 next = head->next;
469                 pi_state = list_entry(next, struct futex_pi_state, list);
470                 key = pi_state->key;
471                 hb = hash_futex(&key);
472                 spin_unlock_irq(&curr->pi_lock);
473
474                 spin_lock(&hb->lock);
475
476                 spin_lock_irq(&curr->pi_lock);
477                 /*
478                  * We dropped the pi-lock, so re-check whether this
479                  * task still owns the PI-state:
480                  */
481                 if (head->next != next) {
482                         spin_unlock(&hb->lock);
483                         continue;
484                 }
485
486                 WARN_ON(pi_state->owner != curr);
487                 WARN_ON(list_empty(&pi_state->list));
488                 list_del_init(&pi_state->list);
489                 pi_state->owner = NULL;
490                 spin_unlock_irq(&curr->pi_lock);
491
492                 rt_mutex_unlock(&pi_state->pi_mutex);
493
494                 spin_unlock(&hb->lock);
495
496                 spin_lock_irq(&curr->pi_lock);
497         }
498         spin_unlock_irq(&curr->pi_lock);
499 }
500
501 static int
502 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
503                 union futex_key *key, struct futex_pi_state **ps)
504 {
505         struct futex_pi_state *pi_state = NULL;
506         struct futex_q *this, *next;
507         struct plist_head *head;
508         struct task_struct *p;
509         pid_t pid = uval & FUTEX_TID_MASK;
510
511         head = &hb->chain;
512
513         plist_for_each_entry_safe(this, next, head, list) {
514                 if (match_futex(&this->key, key)) {
515                         /*
516                          * Another waiter already exists - bump up
517                          * the refcount and return its pi_state:
518                          */
519                         pi_state = this->pi_state;
520                         /*
521                          * Userspace might have messed up non PI and PI futexes
522                          */
523                         if (unlikely(!pi_state))
524                                 return -EINVAL;
525
526                         WARN_ON(!atomic_read(&pi_state->refcount));
527                         WARN_ON(pid && pi_state->owner &&
528                                 pi_state->owner->pid != pid);
529
530                         atomic_inc(&pi_state->refcount);
531                         *ps = pi_state;
532
533                         return 0;
534                 }
535         }
536
537         /*
538          * We are the first waiter - try to look up the real owner and attach
539          * the new pi_state to it, but bail out when TID = 0
540          */
541         if (!pid)
542                 return -ESRCH;
543         p = futex_find_get_task(pid);
544         if (IS_ERR(p))
545                 return PTR_ERR(p);
546
547         /*
548          * We need to look at the task state flags to figure out,
549          * whether the task is exiting. To protect against the do_exit
550          * change of the task flags, we do this protected by
551          * p->pi_lock:
552          */
553         spin_lock_irq(&p->pi_lock);
554         if (unlikely(p->flags & PF_EXITING)) {
555                 /*
556                  * The task is on the way out. When PF_EXITPIDONE is
557                  * set, we know that the task has finished the
558                  * cleanup:
559                  */
560                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
561
562                 spin_unlock_irq(&p->pi_lock);
563                 put_task_struct(p);
564                 return ret;
565         }
566
567         pi_state = alloc_pi_state();
568
569         /*
570          * Initialize the pi_mutex in locked state and make 'p'
571          * the owner of it:
572          */
573         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
574
575         /* Store the key for possible exit cleanups: */
576         pi_state->key = *key;
577
578         WARN_ON(!list_empty(&pi_state->list));
579         list_add(&pi_state->list, &p->pi_state_list);
580         pi_state->owner = p;
581         spin_unlock_irq(&p->pi_lock);
582
583         put_task_struct(p);
584
585         *ps = pi_state;
586
587         return 0;
588 }
589
590 /**
591  * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
592  * @uaddr:              the pi futex user address
593  * @hb:                 the pi futex hash bucket
594  * @key:                the futex key associated with uaddr and hb
595  * @ps:                 the pi_state pointer where we store the result of the
596  *                      lookup
597  * @task:               the task to perform the atomic lock work for.  This will
598  *                      be "current" except in the case of requeue pi.
599  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
600  *
601  * Returns:
602  *  0 - ready to wait
603  *  1 - acquired the lock
604  * <0 - error
605  *
606  * The hb->lock and futex_key refs shall be held by the caller.
607  */
608 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
609                                 union futex_key *key,
610                                 struct futex_pi_state **ps,
611                                 struct task_struct *task, int set_waiters)
612 {
613         int lock_taken, ret, ownerdied = 0;
614         u32 uval, newval, curval;
615
616 retry:
617         ret = lock_taken = 0;
618
619         /*
620          * To avoid races, we attempt to take the lock here again
621          * (by doing a 0 -> TID atomic cmpxchg), while holding all
622          * the locks. It will most likely not succeed.
623          */
624         newval = task_pid_vnr(task);
625         if (set_waiters)
626                 newval |= FUTEX_WAITERS;
627
628         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
629
630         if (unlikely(curval == -EFAULT))
631                 return -EFAULT;
632
633         /*
634          * Detect deadlocks.
635          */
636         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
637                 return -EDEADLK;
638
639         /*
640          * Surprise - we got the lock. Just return to userspace:
641          */
642         if (unlikely(!curval))
643                 return 1;
644
645         uval = curval;
646
647         /*
648          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
649          * to wake at the next unlock.
650          */
651         newval = curval | FUTEX_WAITERS;
652
653         /*
654          * There are two cases, where a futex might have no owner (the
655          * owner TID is 0): OWNER_DIED. We take over the futex in this
656          * case. We also do an unconditional take over, when the owner
657          * of the futex died.
658          *
659          * This is safe as we are protected by the hash bucket lock !
660          */
661         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
662                 /* Keep the OWNER_DIED bit */
663                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
664                 ownerdied = 0;
665                 lock_taken = 1;
666         }
667
668         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
669
670         if (unlikely(curval == -EFAULT))
671                 return -EFAULT;
672         if (unlikely(curval != uval))
673                 goto retry;
674
675         /*
676          * We took the lock due to owner died take over.
677          */
678         if (unlikely(lock_taken))
679                 return 1;
680
681         /*
682          * We dont have the lock. Look up the PI state (or create it if
683          * we are the first waiter):
684          */
685         ret = lookup_pi_state(uval, hb, key, ps);
686
687         if (unlikely(ret)) {
688                 switch (ret) {
689                 case -ESRCH:
690                         /*
691                          * No owner found for this futex. Check if the
692                          * OWNER_DIED bit is set to figure out whether
693                          * this is a robust futex or not.
694                          */
695                         if (get_futex_value_locked(&curval, uaddr))
696                                 return -EFAULT;
697
698                         /*
699                          * We simply start over in case of a robust
700                          * futex. The code above will take the futex
701                          * and return happy.
702                          */
703                         if (curval & FUTEX_OWNER_DIED) {
704                                 ownerdied = 1;
705                                 goto retry;
706                         }
707                 default:
708                         break;
709                 }
710         }
711
712         return ret;
713 }
714
715 /*
716  * The hash bucket lock must be held when this is called.
717  * Afterwards, the futex_q must not be accessed.
718  */
719 static void wake_futex(struct futex_q *q)
720 {
721         struct task_struct *p = q->task;
722
723         /*
724          * We set q->lock_ptr = NULL _before_ we wake up the task. If
725          * a non futex wake up happens on another CPU then the task
726          * might exit and p would dereference a non existing task
727          * struct. Prevent this by holding a reference on p across the
728          * wake up.
729          */
730         get_task_struct(p);
731
732         plist_del(&q->list, &q->list.plist);
733         /*
734          * The waiting task can free the futex_q as soon as
735          * q->lock_ptr = NULL is written, without taking any locks. A
736          * memory barrier is required here to prevent the following
737          * store to lock_ptr from getting ahead of the plist_del.
738          */
739         smp_wmb();
740         q->lock_ptr = NULL;
741
742         wake_up_state(p, TASK_NORMAL);
743         put_task_struct(p);
744 }
745
746 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
747 {
748         struct task_struct *new_owner;
749         struct futex_pi_state *pi_state = this->pi_state;
750         u32 curval, newval;
751
752         if (!pi_state)
753                 return -EINVAL;
754
755         spin_lock(&pi_state->pi_mutex.wait_lock);
756         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
757
758         /*
759          * This happens when we have stolen the lock and the original
760          * pending owner did not enqueue itself back on the rt_mutex.
761          * Thats not a tragedy. We know that way, that a lock waiter
762          * is on the fly. We make the futex_q waiter the pending owner.
763          */
764         if (!new_owner)
765                 new_owner = this->task;
766
767         /*
768          * We pass it to the next owner. (The WAITERS bit is always
769          * kept enabled while there is PI state around. We must also
770          * preserve the owner died bit.)
771          */
772         if (!(uval & FUTEX_OWNER_DIED)) {
773                 int ret = 0;
774
775                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
776
777                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
778
779                 if (curval == -EFAULT)
780                         ret = -EFAULT;
781                 else if (curval != uval)
782                         ret = -EINVAL;
783                 if (ret) {
784                         spin_unlock(&pi_state->pi_mutex.wait_lock);
785                         return ret;
786                 }
787         }
788
789         spin_lock_irq(&pi_state->owner->pi_lock);
790         WARN_ON(list_empty(&pi_state->list));
791         list_del_init(&pi_state->list);
792         spin_unlock_irq(&pi_state->owner->pi_lock);
793
794         spin_lock_irq(&new_owner->pi_lock);
795         WARN_ON(!list_empty(&pi_state->list));
796         list_add(&pi_state->list, &new_owner->pi_state_list);
797         pi_state->owner = new_owner;
798         spin_unlock_irq(&new_owner->pi_lock);
799
800         spin_unlock(&pi_state->pi_mutex.wait_lock);
801         rt_mutex_unlock(&pi_state->pi_mutex);
802
803         return 0;
804 }
805
806 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
807 {
808         u32 oldval;
809
810         /*
811          * There is no waiter, so we unlock the futex. The owner died
812          * bit has not to be preserved here. We are the owner:
813          */
814         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
815
816         if (oldval == -EFAULT)
817                 return oldval;
818         if (oldval != uval)
819                 return -EAGAIN;
820
821         return 0;
822 }
823
824 /*
825  * Express the locking dependencies for lockdep:
826  */
827 static inline void
828 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
829 {
830         if (hb1 <= hb2) {
831                 spin_lock(&hb1->lock);
832                 if (hb1 < hb2)
833                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
834         } else { /* hb1 > hb2 */
835                 spin_lock(&hb2->lock);
836                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
837         }
838 }
839
840 static inline void
841 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
842 {
843         spin_unlock(&hb1->lock);
844         if (hb1 != hb2)
845                 spin_unlock(&hb2->lock);
846 }
847
848 /*
849  * Wake up waiters matching bitset queued on this futex (uaddr).
850  */
851 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
852 {
853         struct futex_hash_bucket *hb;
854         struct futex_q *this, *next;
855         struct plist_head *head;
856         union futex_key key = FUTEX_KEY_INIT;
857         int ret;
858
859         if (!bitset)
860                 return -EINVAL;
861
862         ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
863         if (unlikely(ret != 0))
864                 goto out;
865
866         hb = hash_futex(&key);
867         spin_lock(&hb->lock);
868         head = &hb->chain;
869
870         plist_for_each_entry_safe(this, next, head, list) {
871                 if (match_futex (&this->key, &key)) {
872                         if (this->pi_state || this->rt_waiter) {
873                                 ret = -EINVAL;
874                                 break;
875                         }
876
877                         /* Check if one of the bits is set in both bitsets */
878                         if (!(this->bitset & bitset))
879                                 continue;
880
881                         wake_futex(this);
882                         if (++ret >= nr_wake)
883                                 break;
884                 }
885         }
886
887         spin_unlock(&hb->lock);
888         put_futex_key(fshared, &key);
889 out:
890         return ret;
891 }
892
893 /*
894  * Wake up all waiters hashed on the physical page that is mapped
895  * to this virtual address:
896  */
897 static int
898 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
899               int nr_wake, int nr_wake2, int op)
900 {
901         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
902         struct futex_hash_bucket *hb1, *hb2;
903         struct plist_head *head;
904         struct futex_q *this, *next;
905         int ret, op_ret;
906
907 retry:
908         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
909         if (unlikely(ret != 0))
910                 goto out;
911         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
912         if (unlikely(ret != 0))
913                 goto out_put_key1;
914
915         hb1 = hash_futex(&key1);
916         hb2 = hash_futex(&key2);
917
918         double_lock_hb(hb1, hb2);
919 retry_private:
920         op_ret = futex_atomic_op_inuser(op, uaddr2);
921         if (unlikely(op_ret < 0)) {
922
923                 double_unlock_hb(hb1, hb2);
924
925 #ifndef CONFIG_MMU
926                 /*
927                  * we don't get EFAULT from MMU faults if we don't have an MMU,
928                  * but we might get them from range checking
929                  */
930                 ret = op_ret;
931                 goto out_put_keys;
932 #endif
933
934                 if (unlikely(op_ret != -EFAULT)) {
935                         ret = op_ret;
936                         goto out_put_keys;
937                 }
938
939                 ret = fault_in_user_writeable(uaddr2);
940                 if (ret)
941                         goto out_put_keys;
942
943                 if (!fshared)
944                         goto retry_private;
945
946                 put_futex_key(fshared, &key2);
947                 put_futex_key(fshared, &key1);
948                 goto retry;
949         }
950
951         head = &hb1->chain;
952
953         plist_for_each_entry_safe(this, next, head, list) {
954                 if (match_futex (&this->key, &key1)) {
955                         wake_futex(this);
956                         if (++ret >= nr_wake)
957                                 break;
958                 }
959         }
960
961         if (op_ret > 0) {
962                 head = &hb2->chain;
963
964                 op_ret = 0;
965                 plist_for_each_entry_safe(this, next, head, list) {
966                         if (match_futex (&this->key, &key2)) {
967                                 wake_futex(this);
968                                 if (++op_ret >= nr_wake2)
969                                         break;
970                         }
971                 }
972                 ret += op_ret;
973         }
974
975         double_unlock_hb(hb1, hb2);
976 out_put_keys:
977         put_futex_key(fshared, &key2);
978 out_put_key1:
979         put_futex_key(fshared, &key1);
980 out:
981         return ret;
982 }
983
984 /**
985  * requeue_futex() - Requeue a futex_q from one hb to another
986  * @q:          the futex_q to requeue
987  * @hb1:        the source hash_bucket
988  * @hb2:        the target hash_bucket
989  * @key2:       the new key for the requeued futex_q
990  */
991 static inline
992 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
993                    struct futex_hash_bucket *hb2, union futex_key *key2)
994 {
995
996         /*
997          * If key1 and key2 hash to the same bucket, no need to
998          * requeue.
999          */
1000         if (likely(&hb1->chain != &hb2->chain)) {
1001                 plist_del(&q->list, &hb1->chain);
1002                 plist_add(&q->list, &hb2->chain);
1003                 q->lock_ptr = &hb2->lock;
1004 #ifdef CONFIG_DEBUG_PI_LIST
1005                 q->list.plist.lock = &hb2->lock;
1006 #endif
1007         }
1008         get_futex_key_refs(key2);
1009         q->key = *key2;
1010 }
1011
1012 /**
1013  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1014  * q:   the futex_q
1015  * key: the key of the requeue target futex
1016  * hb:  the hash_bucket of the requeue target futex
1017  *
1018  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1019  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1020  * to the requeue target futex so the waiter can detect the wakeup on the right
1021  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1022  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1023  * to protect access to the pi_state to fixup the owner later.  Must be called
1024  * with both q->lock_ptr and hb->lock held.
1025  */
1026 static inline
1027 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1028                            struct futex_hash_bucket *hb)
1029 {
1030         drop_futex_key_refs(&q->key);
1031         get_futex_key_refs(key);
1032         q->key = *key;
1033
1034         WARN_ON(plist_node_empty(&q->list));
1035         plist_del(&q->list, &q->list.plist);
1036
1037         WARN_ON(!q->rt_waiter);
1038         q->rt_waiter = NULL;
1039
1040         q->lock_ptr = &hb->lock;
1041 #ifdef CONFIG_DEBUG_PI_LIST
1042         q->list.plist.lock = &hb->lock;
1043 #endif
1044
1045         wake_up_state(q->task, TASK_NORMAL);
1046 }
1047
1048 /**
1049  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1050  * @pifutex:            the user address of the to futex
1051  * @hb1:                the from futex hash bucket, must be locked by the caller
1052  * @hb2:                the to futex hash bucket, must be locked by the caller
1053  * @key1:               the from futex key
1054  * @key2:               the to futex key
1055  * @ps:                 address to store the pi_state pointer
1056  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1057  *
1058  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1059  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1060  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1061  * hb1 and hb2 must be held by the caller.
1062  *
1063  * Returns:
1064  *  0 - failed to acquire the lock atomicly
1065  *  1 - acquired the lock
1066  * <0 - error
1067  */
1068 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1069                                  struct futex_hash_bucket *hb1,
1070                                  struct futex_hash_bucket *hb2,
1071                                  union futex_key *key1, union futex_key *key2,
1072                                  struct futex_pi_state **ps, int set_waiters)
1073 {
1074         struct futex_q *top_waiter = NULL;
1075         u32 curval;
1076         int ret;
1077
1078         if (get_futex_value_locked(&curval, pifutex))
1079                 return -EFAULT;
1080
1081         /*
1082          * Find the top_waiter and determine if there are additional waiters.
1083          * If the caller intends to requeue more than 1 waiter to pifutex,
1084          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1085          * as we have means to handle the possible fault.  If not, don't set
1086          * the bit unecessarily as it will force the subsequent unlock to enter
1087          * the kernel.
1088          */
1089         top_waiter = futex_top_waiter(hb1, key1);
1090
1091         /* There are no waiters, nothing for us to do. */
1092         if (!top_waiter)
1093                 return 0;
1094
1095         /* Ensure we requeue to the expected futex. */
1096         if (!match_futex(top_waiter->requeue_pi_key, key2))
1097                 return -EINVAL;
1098
1099         /*
1100          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1101          * the contended case or if set_waiters is 1.  The pi_state is returned
1102          * in ps in contended cases.
1103          */
1104         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1105                                    set_waiters);
1106         if (ret == 1)
1107                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1108
1109         return ret;
1110 }
1111
1112 /**
1113  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1114  * uaddr1:      source futex user address
1115  * uaddr2:      target futex user address
1116  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1117  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1118  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1119  *              pi futex (pi to pi requeue is not supported)
1120  *
1121  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1122  * uaddr2 atomically on behalf of the top waiter.
1123  *
1124  * Returns:
1125  * >=0 - on success, the number of tasks requeued or woken
1126  *  <0 - on error
1127  */
1128 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1129                          int nr_wake, int nr_requeue, u32 *cmpval,
1130                          int requeue_pi)
1131 {
1132         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1133         int drop_count = 0, task_count = 0, ret;
1134         struct futex_pi_state *pi_state = NULL;
1135         struct futex_hash_bucket *hb1, *hb2;
1136         struct plist_head *head1;
1137         struct futex_q *this, *next;
1138         u32 curval2;
1139
1140         if (requeue_pi) {
1141                 /*
1142                  * requeue_pi requires a pi_state, try to allocate it now
1143                  * without any locks in case it fails.
1144                  */
1145                 if (refill_pi_state_cache())
1146                         return -ENOMEM;
1147                 /*
1148                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1149                  * + nr_requeue, since it acquires the rt_mutex prior to
1150                  * returning to userspace, so as to not leave the rt_mutex with
1151                  * waiters and no owner.  However, second and third wake-ups
1152                  * cannot be predicted as they involve race conditions with the
1153                  * first wake and a fault while looking up the pi_state.  Both
1154                  * pthread_cond_signal() and pthread_cond_broadcast() should
1155                  * use nr_wake=1.
1156                  */
1157                 if (nr_wake != 1)
1158                         return -EINVAL;
1159         }
1160
1161 retry:
1162         if (pi_state != NULL) {
1163                 /*
1164                  * We will have to lookup the pi_state again, so free this one
1165                  * to keep the accounting correct.
1166                  */
1167                 free_pi_state(pi_state);
1168                 pi_state = NULL;
1169         }
1170
1171         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1172         if (unlikely(ret != 0))
1173                 goto out;
1174         ret = get_futex_key(uaddr2, fshared, &key2,
1175                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1176         if (unlikely(ret != 0))
1177                 goto out_put_key1;
1178
1179         hb1 = hash_futex(&key1);
1180         hb2 = hash_futex(&key2);
1181
1182 retry_private:
1183         double_lock_hb(hb1, hb2);
1184
1185         if (likely(cmpval != NULL)) {
1186                 u32 curval;
1187
1188                 ret = get_futex_value_locked(&curval, uaddr1);
1189
1190                 if (unlikely(ret)) {
1191                         double_unlock_hb(hb1, hb2);
1192
1193                         ret = get_user(curval, uaddr1);
1194                         if (ret)
1195                                 goto out_put_keys;
1196
1197                         if (!fshared)
1198                                 goto retry_private;
1199
1200                         put_futex_key(fshared, &key2);
1201                         put_futex_key(fshared, &key1);
1202                         goto retry;
1203                 }
1204                 if (curval != *cmpval) {
1205                         ret = -EAGAIN;
1206                         goto out_unlock;
1207                 }
1208         }
1209
1210         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1211                 /*
1212                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1213                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1214                  * bit.  We force this here where we are able to easily handle
1215                  * faults rather in the requeue loop below.
1216                  */
1217                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1218                                                  &key2, &pi_state, nr_requeue);
1219
1220                 /*
1221                  * At this point the top_waiter has either taken uaddr2 or is
1222                  * waiting on it.  If the former, then the pi_state will not
1223                  * exist yet, look it up one more time to ensure we have a
1224                  * reference to it.
1225                  */
1226                 if (ret == 1) {
1227                         WARN_ON(pi_state);
1228                         task_count++;
1229                         ret = get_futex_value_locked(&curval2, uaddr2);
1230                         if (!ret)
1231                                 ret = lookup_pi_state(curval2, hb2, &key2,
1232                                                       &pi_state);
1233                 }
1234
1235                 switch (ret) {
1236                 case 0:
1237                         break;
1238                 case -EFAULT:
1239                         double_unlock_hb(hb1, hb2);
1240                         put_futex_key(fshared, &key2);
1241                         put_futex_key(fshared, &key1);
1242                         ret = fault_in_user_writeable(uaddr2);
1243                         if (!ret)
1244                                 goto retry;
1245                         goto out;
1246                 case -EAGAIN:
1247                         /* The owner was exiting, try again. */
1248                         double_unlock_hb(hb1, hb2);
1249                         put_futex_key(fshared, &key2);
1250                         put_futex_key(fshared, &key1);
1251                         cond_resched();
1252                         goto retry;
1253                 default:
1254                         goto out_unlock;
1255                 }
1256         }
1257
1258         head1 = &hb1->chain;
1259         plist_for_each_entry_safe(this, next, head1, list) {
1260                 if (task_count - nr_wake >= nr_requeue)
1261                         break;
1262
1263                 if (!match_futex(&this->key, &key1))
1264                         continue;
1265
1266                 /*
1267                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1268                  * be paired with each other and no other futex ops.
1269                  */
1270                 if ((requeue_pi && !this->rt_waiter) ||
1271                     (!requeue_pi && this->rt_waiter)) {
1272                         ret = -EINVAL;
1273                         break;
1274                 }
1275
1276                 /*
1277                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1278                  * lock, we already woke the top_waiter.  If not, it will be
1279                  * woken by futex_unlock_pi().
1280                  */
1281                 if (++task_count <= nr_wake && !requeue_pi) {
1282                         wake_futex(this);
1283                         continue;
1284                 }
1285
1286                 /* Ensure we requeue to the expected futex for requeue_pi. */
1287                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1288                         ret = -EINVAL;
1289                         break;
1290                 }
1291
1292                 /*
1293                  * Requeue nr_requeue waiters and possibly one more in the case
1294                  * of requeue_pi if we couldn't acquire the lock atomically.
1295                  */
1296                 if (requeue_pi) {
1297                         /* Prepare the waiter to take the rt_mutex. */
1298                         atomic_inc(&pi_state->refcount);
1299                         this->pi_state = pi_state;
1300                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1301                                                         this->rt_waiter,
1302                                                         this->task, 1);
1303                         if (ret == 1) {
1304                                 /* We got the lock. */
1305                                 requeue_pi_wake_futex(this, &key2, hb2);
1306                                 continue;
1307                         } else if (ret) {
1308                                 /* -EDEADLK */
1309                                 this->pi_state = NULL;
1310                                 free_pi_state(pi_state);
1311                                 goto out_unlock;
1312                         }
1313                 }
1314                 requeue_futex(this, hb1, hb2, &key2);
1315                 drop_count++;
1316         }
1317
1318 out_unlock:
1319         double_unlock_hb(hb1, hb2);
1320
1321         /*
1322          * drop_futex_key_refs() must be called outside the spinlocks. During
1323          * the requeue we moved futex_q's from the hash bucket at key1 to the
1324          * one at key2 and updated their key pointer.  We no longer need to
1325          * hold the references to key1.
1326          */
1327         while (--drop_count >= 0)
1328                 drop_futex_key_refs(&key1);
1329
1330 out_put_keys:
1331         put_futex_key(fshared, &key2);
1332 out_put_key1:
1333         put_futex_key(fshared, &key1);
1334 out:
1335         if (pi_state != NULL)
1336                 free_pi_state(pi_state);
1337         return ret ? ret : task_count;
1338 }
1339
1340 /* The key must be already stored in q->key. */
1341 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1342 {
1343         struct futex_hash_bucket *hb;
1344
1345         get_futex_key_refs(&q->key);
1346         hb = hash_futex(&q->key);
1347         q->lock_ptr = &hb->lock;
1348
1349         spin_lock(&hb->lock);
1350         return hb;
1351 }
1352
1353 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1354 {
1355         int prio;
1356
1357         /*
1358          * The priority used to register this element is
1359          * - either the real thread-priority for the real-time threads
1360          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1361          * - or MAX_RT_PRIO for non-RT threads.
1362          * Thus, all RT-threads are woken first in priority order, and
1363          * the others are woken last, in FIFO order.
1364          */
1365         prio = min(current->normal_prio, MAX_RT_PRIO);
1366
1367         plist_node_init(&q->list, prio);
1368 #ifdef CONFIG_DEBUG_PI_LIST
1369         q->list.plist.lock = &hb->lock;
1370 #endif
1371         plist_add(&q->list, &hb->chain);
1372         q->task = current;
1373         spin_unlock(&hb->lock);
1374 }
1375
1376 static inline void
1377 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1378 {
1379         spin_unlock(&hb->lock);
1380         drop_futex_key_refs(&q->key);
1381 }
1382
1383 /*
1384  * queue_me and unqueue_me must be called as a pair, each
1385  * exactly once.  They are called with the hashed spinlock held.
1386  */
1387
1388 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1389 static int unqueue_me(struct futex_q *q)
1390 {
1391         spinlock_t *lock_ptr;
1392         int ret = 0;
1393
1394         /* In the common case we don't take the spinlock, which is nice. */
1395 retry:
1396         lock_ptr = q->lock_ptr;
1397         barrier();
1398         if (lock_ptr != NULL) {
1399                 spin_lock(lock_ptr);
1400                 /*
1401                  * q->lock_ptr can change between reading it and
1402                  * spin_lock(), causing us to take the wrong lock.  This
1403                  * corrects the race condition.
1404                  *
1405                  * Reasoning goes like this: if we have the wrong lock,
1406                  * q->lock_ptr must have changed (maybe several times)
1407                  * between reading it and the spin_lock().  It can
1408                  * change again after the spin_lock() but only if it was
1409                  * already changed before the spin_lock().  It cannot,
1410                  * however, change back to the original value.  Therefore
1411                  * we can detect whether we acquired the correct lock.
1412                  */
1413                 if (unlikely(lock_ptr != q->lock_ptr)) {
1414                         spin_unlock(lock_ptr);
1415                         goto retry;
1416                 }
1417                 WARN_ON(plist_node_empty(&q->list));
1418                 plist_del(&q->list, &q->list.plist);
1419
1420                 BUG_ON(q->pi_state);
1421
1422                 spin_unlock(lock_ptr);
1423                 ret = 1;
1424         }
1425
1426         drop_futex_key_refs(&q->key);
1427         return ret;
1428 }
1429
1430 /*
1431  * PI futexes can not be requeued and must remove themself from the
1432  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1433  * and dropped here.
1434  */
1435 static void unqueue_me_pi(struct futex_q *q)
1436 {
1437         WARN_ON(plist_node_empty(&q->list));
1438         plist_del(&q->list, &q->list.plist);
1439
1440         BUG_ON(!q->pi_state);
1441         free_pi_state(q->pi_state);
1442         q->pi_state = NULL;
1443
1444         spin_unlock(q->lock_ptr);
1445
1446         drop_futex_key_refs(&q->key);
1447 }
1448
1449 /*
1450  * Fixup the pi_state owner with the new owner.
1451  *
1452  * Must be called with hash bucket lock held and mm->sem held for non
1453  * private futexes.
1454  */
1455 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1456                                 struct task_struct *newowner, int fshared)
1457 {
1458         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1459         struct futex_pi_state *pi_state = q->pi_state;
1460         struct task_struct *oldowner = pi_state->owner;
1461         u32 uval, curval, newval;
1462         int ret;
1463
1464         /* Owner died? */
1465         if (!pi_state->owner)
1466                 newtid |= FUTEX_OWNER_DIED;
1467
1468         /*
1469          * We are here either because we stole the rtmutex from the
1470          * pending owner or we are the pending owner which failed to
1471          * get the rtmutex. We have to replace the pending owner TID
1472          * in the user space variable. This must be atomic as we have
1473          * to preserve the owner died bit here.
1474          *
1475          * Note: We write the user space value _before_ changing the pi_state
1476          * because we can fault here. Imagine swapped out pages or a fork
1477          * that marked all the anonymous memory readonly for cow.
1478          *
1479          * Modifying pi_state _before_ the user space value would
1480          * leave the pi_state in an inconsistent state when we fault
1481          * here, because we need to drop the hash bucket lock to
1482          * handle the fault. This might be observed in the PID check
1483          * in lookup_pi_state.
1484          */
1485 retry:
1486         if (get_futex_value_locked(&uval, uaddr))
1487                 goto handle_fault;
1488
1489         while (1) {
1490                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1491
1492                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1493
1494                 if (curval == -EFAULT)
1495                         goto handle_fault;
1496                 if (curval == uval)
1497                         break;
1498                 uval = curval;
1499         }
1500
1501         /*
1502          * We fixed up user space. Now we need to fix the pi_state
1503          * itself.
1504          */
1505         if (pi_state->owner != NULL) {
1506                 spin_lock_irq(&pi_state->owner->pi_lock);
1507                 WARN_ON(list_empty(&pi_state->list));
1508                 list_del_init(&pi_state->list);
1509                 spin_unlock_irq(&pi_state->owner->pi_lock);
1510         }
1511
1512         pi_state->owner = newowner;
1513
1514         spin_lock_irq(&newowner->pi_lock);
1515         WARN_ON(!list_empty(&pi_state->list));
1516         list_add(&pi_state->list, &newowner->pi_state_list);
1517         spin_unlock_irq(&newowner->pi_lock);
1518         return 0;
1519
1520         /*
1521          * To handle the page fault we need to drop the hash bucket
1522          * lock here. That gives the other task (either the pending
1523          * owner itself or the task which stole the rtmutex) the
1524          * chance to try the fixup of the pi_state. So once we are
1525          * back from handling the fault we need to check the pi_state
1526          * after reacquiring the hash bucket lock and before trying to
1527          * do another fixup. When the fixup has been done already we
1528          * simply return.
1529          */
1530 handle_fault:
1531         spin_unlock(q->lock_ptr);
1532
1533         ret = fault_in_user_writeable(uaddr);
1534
1535         spin_lock(q->lock_ptr);
1536
1537         /*
1538          * Check if someone else fixed it for us:
1539          */
1540         if (pi_state->owner != oldowner)
1541                 return 0;
1542
1543         if (ret)
1544                 return ret;
1545
1546         goto retry;
1547 }
1548
1549 /*
1550  * In case we must use restart_block to restart a futex_wait,
1551  * we encode in the 'flags' shared capability
1552  */
1553 #define FLAGS_SHARED            0x01
1554 #define FLAGS_CLOCKRT           0x02
1555 #define FLAGS_HAS_TIMEOUT       0x04
1556
1557 static long futex_wait_restart(struct restart_block *restart);
1558
1559 /**
1560  * fixup_owner() - Post lock pi_state and corner case management
1561  * @uaddr:      user address of the futex
1562  * @fshared:    whether the futex is shared (1) or not (0)
1563  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1564  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1565  *
1566  * After attempting to lock an rt_mutex, this function is called to cleanup
1567  * the pi_state owner as well as handle race conditions that may allow us to
1568  * acquire the lock. Must be called with the hb lock held.
1569  *
1570  * Returns:
1571  *  1 - success, lock taken
1572  *  0 - success, lock not taken
1573  * <0 - on error (-EFAULT)
1574  */
1575 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1576                        int locked)
1577 {
1578         struct task_struct *owner;
1579         int ret = 0;
1580
1581         if (locked) {
1582                 /*
1583                  * Got the lock. We might not be the anticipated owner if we
1584                  * did a lock-steal - fix up the PI-state in that case:
1585                  */
1586                 if (q->pi_state->owner != current)
1587                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1588                 goto out;
1589         }
1590
1591         /*
1592          * Catch the rare case, where the lock was released when we were on the
1593          * way back before we locked the hash bucket.
1594          */
1595         if (q->pi_state->owner == current) {
1596                 /*
1597                  * Try to get the rt_mutex now. This might fail as some other
1598                  * task acquired the rt_mutex after we removed ourself from the
1599                  * rt_mutex waiters list.
1600                  */
1601                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1602                         locked = 1;
1603                         goto out;
1604                 }
1605
1606                 /*
1607                  * pi_state is incorrect, some other task did a lock steal and
1608                  * we returned due to timeout or signal without taking the
1609                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1610                  * locking, as the other task is now blocked on the hash bucket
1611                  * lock. Fix the state up.
1612                  */
1613                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1614                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1615                 goto out;
1616         }
1617
1618         /*
1619          * Paranoia check. If we did not take the lock, then we should not be
1620          * the owner, nor the pending owner, of the rt_mutex.
1621          */
1622         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1623                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1624                                 "pi-state %p\n", ret,
1625                                 q->pi_state->pi_mutex.owner,
1626                                 q->pi_state->owner);
1627
1628 out:
1629         return ret ? ret : locked;
1630 }
1631
1632 /**
1633  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1634  * @hb:         the futex hash bucket, must be locked by the caller
1635  * @q:          the futex_q to queue up on
1636  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1637  */
1638 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1639                                 struct hrtimer_sleeper *timeout)
1640 {
1641         queue_me(q, hb);
1642
1643         /*
1644          * There might have been scheduling since the queue_me(), as we
1645          * cannot hold a spinlock across the get_user() in case it
1646          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1647          * queueing ourselves into the futex hash. This code thus has to
1648          * rely on the futex_wake() code removing us from hash when it
1649          * wakes us up.
1650          */
1651         set_current_state(TASK_INTERRUPTIBLE);
1652
1653         /* Arm the timer */
1654         if (timeout) {
1655                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1656                 if (!hrtimer_active(&timeout->timer))
1657                         timeout->task = NULL;
1658         }
1659
1660         /*
1661          * !plist_node_empty() is safe here without any lock.
1662          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1663          */
1664         if (likely(!plist_node_empty(&q->list))) {
1665                 /*
1666                  * If the timer has already expired, current will already be
1667                  * flagged for rescheduling. Only call schedule if there
1668                  * is no timeout, or if it has yet to expire.
1669                  */
1670                 if (!timeout || timeout->task)
1671                         schedule();
1672         }
1673         __set_current_state(TASK_RUNNING);
1674 }
1675
1676 /**
1677  * futex_wait_setup() - Prepare to wait on a futex
1678  * @uaddr:      the futex userspace address
1679  * @val:        the expected value
1680  * @fshared:    whether the futex is shared (1) or not (0)
1681  * @q:          the associated futex_q
1682  * @hb:         storage for hash_bucket pointer to be returned to caller
1683  *
1684  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1685  * compare it with the expected value.  Handle atomic faults internally.
1686  * Return with the hb lock held and a q.key reference on success, and unlocked
1687  * with no q.key reference on failure.
1688  *
1689  * Returns:
1690  *  0 - uaddr contains val and hb has been locked
1691  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1692  */
1693 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1694                            struct futex_q *q, struct futex_hash_bucket **hb)
1695 {
1696         u32 uval;
1697         int ret;
1698
1699         /*
1700          * Access the page AFTER the hash-bucket is locked.
1701          * Order is important:
1702          *
1703          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1704          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1705          *
1706          * The basic logical guarantee of a futex is that it blocks ONLY
1707          * if cond(var) is known to be true at the time of blocking, for
1708          * any cond.  If we queued after testing *uaddr, that would open
1709          * a race condition where we could block indefinitely with
1710          * cond(var) false, which would violate the guarantee.
1711          *
1712          * A consequence is that futex_wait() can return zero and absorb
1713          * a wakeup when *uaddr != val on entry to the syscall.  This is
1714          * rare, but normal.
1715          */
1716 retry:
1717         q->key = FUTEX_KEY_INIT;
1718         ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1719         if (unlikely(ret != 0))
1720                 return ret;
1721
1722 retry_private:
1723         *hb = queue_lock(q);
1724
1725         ret = get_futex_value_locked(&uval, uaddr);
1726
1727         if (ret) {
1728                 queue_unlock(q, *hb);
1729
1730                 ret = get_user(uval, uaddr);
1731                 if (ret)
1732                         goto out;
1733
1734                 if (!fshared)
1735                         goto retry_private;
1736
1737                 put_futex_key(fshared, &q->key);
1738                 goto retry;
1739         }
1740
1741         if (uval != val) {
1742                 queue_unlock(q, *hb);
1743                 ret = -EWOULDBLOCK;
1744         }
1745
1746 out:
1747         if (ret)
1748                 put_futex_key(fshared, &q->key);
1749         return ret;
1750 }
1751
1752 static int futex_wait(u32 __user *uaddr, int fshared,
1753                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1754 {
1755         struct hrtimer_sleeper timeout, *to = NULL;
1756         struct restart_block *restart;
1757         struct futex_hash_bucket *hb;
1758         struct futex_q q;
1759         int ret;
1760
1761         if (!bitset)
1762                 return -EINVAL;
1763
1764         q.pi_state = NULL;
1765         q.bitset = bitset;
1766         q.rt_waiter = NULL;
1767         q.requeue_pi_key = NULL;
1768
1769         if (abs_time) {
1770                 to = &timeout;
1771
1772                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1773                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1774                 hrtimer_init_sleeper(to, current);
1775                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1776                                              current->timer_slack_ns);
1777         }
1778
1779         /* Prepare to wait on uaddr. */
1780         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1781         if (ret)
1782                 goto out;
1783
1784         /* queue_me and wait for wakeup, timeout, or a signal. */
1785         futex_wait_queue_me(hb, &q, to);
1786
1787         /* If we were woken (and unqueued), we succeeded, whatever. */
1788         ret = 0;
1789         if (!unqueue_me(&q))
1790                 goto out_put_key;
1791         ret = -ETIMEDOUT;
1792         if (to && !to->task)
1793                 goto out_put_key;
1794
1795         /*
1796          * We expect signal_pending(current), but another thread may
1797          * have handled it for us already.
1798          */
1799         ret = -ERESTARTSYS;
1800         if (!abs_time)
1801                 goto out_put_key;
1802
1803         restart = &current_thread_info()->restart_block;
1804         restart->fn = futex_wait_restart;
1805         restart->futex.uaddr = (u32 *)uaddr;
1806         restart->futex.val = val;
1807         restart->futex.time = abs_time->tv64;
1808         restart->futex.bitset = bitset;
1809         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1810
1811         if (fshared)
1812                 restart->futex.flags |= FLAGS_SHARED;
1813         if (clockrt)
1814                 restart->futex.flags |= FLAGS_CLOCKRT;
1815
1816         ret = -ERESTART_RESTARTBLOCK;
1817
1818 out_put_key:
1819         put_futex_key(fshared, &q.key);
1820 out:
1821         if (to) {
1822                 hrtimer_cancel(&to->timer);
1823                 destroy_hrtimer_on_stack(&to->timer);
1824         }
1825         return ret;
1826 }
1827
1828
1829 static long futex_wait_restart(struct restart_block *restart)
1830 {
1831         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1832         int fshared = 0;
1833         ktime_t t, *tp = NULL;
1834
1835         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1836                 t.tv64 = restart->futex.time;
1837                 tp = &t;
1838         }
1839         restart->fn = do_no_restart_syscall;
1840         if (restart->futex.flags & FLAGS_SHARED)
1841                 fshared = 1;
1842         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1843                                 restart->futex.bitset,
1844                                 restart->futex.flags & FLAGS_CLOCKRT);
1845 }
1846
1847
1848 /*
1849  * Userspace tried a 0 -> TID atomic transition of the futex value
1850  * and failed. The kernel side here does the whole locking operation:
1851  * if there are waiters then it will block, it does PI, etc. (Due to
1852  * races the kernel might see a 0 value of the futex too.)
1853  */
1854 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1855                          int detect, ktime_t *time, int trylock)
1856 {
1857         struct hrtimer_sleeper timeout, *to = NULL;
1858         struct futex_hash_bucket *hb;
1859         struct futex_q q;
1860         int res, ret;
1861
1862         if (refill_pi_state_cache())
1863                 return -ENOMEM;
1864
1865         if (time) {
1866                 to = &timeout;
1867                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1868                                       HRTIMER_MODE_ABS);
1869                 hrtimer_init_sleeper(to, current);
1870                 hrtimer_set_expires(&to->timer, *time);
1871         }
1872
1873         q.pi_state = NULL;
1874         q.rt_waiter = NULL;
1875         q.requeue_pi_key = NULL;
1876 retry:
1877         q.key = FUTEX_KEY_INIT;
1878         ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1879         if (unlikely(ret != 0))
1880                 goto out;
1881
1882 retry_private:
1883         hb = queue_lock(&q);
1884
1885         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1886         if (unlikely(ret)) {
1887                 switch (ret) {
1888                 case 1:
1889                         /* We got the lock. */
1890                         ret = 0;
1891                         goto out_unlock_put_key;
1892                 case -EFAULT:
1893                         goto uaddr_faulted;
1894                 case -EAGAIN:
1895                         /*
1896                          * Task is exiting and we just wait for the
1897                          * exit to complete.
1898                          */
1899                         queue_unlock(&q, hb);
1900                         put_futex_key(fshared, &q.key);
1901                         cond_resched();
1902                         goto retry;
1903                 default:
1904                         goto out_unlock_put_key;
1905                 }
1906         }
1907
1908         /*
1909          * Only actually queue now that the atomic ops are done:
1910          */
1911         queue_me(&q, hb);
1912
1913         WARN_ON(!q.pi_state);
1914         /*
1915          * Block on the PI mutex:
1916          */
1917         if (!trylock)
1918                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1919         else {
1920                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1921                 /* Fixup the trylock return value: */
1922                 ret = ret ? 0 : -EWOULDBLOCK;
1923         }
1924
1925         spin_lock(q.lock_ptr);
1926         /*
1927          * Fixup the pi_state owner and possibly acquire the lock if we
1928          * haven't already.
1929          */
1930         res = fixup_owner(uaddr, fshared, &q, !ret);
1931         /*
1932          * If fixup_owner() returned an error, proprogate that.  If it acquired
1933          * the lock, clear our -ETIMEDOUT or -EINTR.
1934          */
1935         if (res)
1936                 ret = (res < 0) ? res : 0;
1937
1938         /*
1939          * If fixup_owner() faulted and was unable to handle the fault, unlock
1940          * it and return the fault to userspace.
1941          */
1942         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1943                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1944
1945         /* Unqueue and drop the lock */
1946         unqueue_me_pi(&q);
1947
1948         goto out;
1949
1950 out_unlock_put_key:
1951         queue_unlock(&q, hb);
1952
1953 out_put_key:
1954         put_futex_key(fshared, &q.key);
1955 out:
1956         if (to)
1957                 destroy_hrtimer_on_stack(&to->timer);
1958         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1959
1960 uaddr_faulted:
1961         queue_unlock(&q, hb);
1962
1963         ret = fault_in_user_writeable(uaddr);
1964         if (ret)
1965                 goto out_put_key;
1966
1967         if (!fshared)
1968                 goto retry_private;
1969
1970         put_futex_key(fshared, &q.key);
1971         goto retry;
1972 }
1973
1974 /*
1975  * Userspace attempted a TID -> 0 atomic transition, and failed.
1976  * This is the in-kernel slowpath: we look up the PI state (if any),
1977  * and do the rt-mutex unlock.
1978  */
1979 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1980 {
1981         struct futex_hash_bucket *hb;
1982         struct futex_q *this, *next;
1983         u32 uval;
1984         struct plist_head *head;
1985         union futex_key key = FUTEX_KEY_INIT;
1986         int ret;
1987
1988 retry:
1989         if (get_user(uval, uaddr))
1990                 return -EFAULT;
1991         /*
1992          * We release only a lock we actually own:
1993          */
1994         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1995                 return -EPERM;
1996
1997         ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
1998         if (unlikely(ret != 0))
1999                 goto out;
2000
2001         hb = hash_futex(&key);
2002         spin_lock(&hb->lock);
2003
2004         /*
2005          * To avoid races, try to do the TID -> 0 atomic transition
2006          * again. If it succeeds then we can return without waking
2007          * anyone else up:
2008          */
2009         if (!(uval & FUTEX_OWNER_DIED))
2010                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2011
2012
2013         if (unlikely(uval == -EFAULT))
2014                 goto pi_faulted;
2015         /*
2016          * Rare case: we managed to release the lock atomically,
2017          * no need to wake anyone else up:
2018          */
2019         if (unlikely(uval == task_pid_vnr(current)))
2020                 goto out_unlock;
2021
2022         /*
2023          * Ok, other tasks may need to be woken up - check waiters
2024          * and do the wakeup if necessary:
2025          */
2026         head = &hb->chain;
2027
2028         plist_for_each_entry_safe(this, next, head, list) {
2029                 if (!match_futex (&this->key, &key))
2030                         continue;
2031                 ret = wake_futex_pi(uaddr, uval, this);
2032                 /*
2033                  * The atomic access to the futex value
2034                  * generated a pagefault, so retry the
2035                  * user-access and the wakeup:
2036                  */
2037                 if (ret == -EFAULT)
2038                         goto pi_faulted;
2039                 goto out_unlock;
2040         }
2041         /*
2042          * No waiters - kernel unlocks the futex:
2043          */
2044         if (!(uval & FUTEX_OWNER_DIED)) {
2045                 ret = unlock_futex_pi(uaddr, uval);
2046                 if (ret == -EFAULT)
2047                         goto pi_faulted;
2048         }
2049
2050 out_unlock:
2051         spin_unlock(&hb->lock);
2052         put_futex_key(fshared, &key);
2053
2054 out:
2055         return ret;
2056
2057 pi_faulted:
2058         spin_unlock(&hb->lock);
2059         put_futex_key(fshared, &key);
2060
2061         ret = fault_in_user_writeable(uaddr);
2062         if (!ret)
2063                 goto retry;
2064
2065         return ret;
2066 }
2067
2068 /**
2069  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2070  * @hb:         the hash_bucket futex_q was original enqueued on
2071  * @q:          the futex_q woken while waiting to be requeued
2072  * @key2:       the futex_key of the requeue target futex
2073  * @timeout:    the timeout associated with the wait (NULL if none)
2074  *
2075  * Detect if the task was woken on the initial futex as opposed to the requeue
2076  * target futex.  If so, determine if it was a timeout or a signal that caused
2077  * the wakeup and return the appropriate error code to the caller.  Must be
2078  * called with the hb lock held.
2079  *
2080  * Returns
2081  *  0 - no early wakeup detected
2082  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2083  */
2084 static inline
2085 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2086                                    struct futex_q *q, union futex_key *key2,
2087                                    struct hrtimer_sleeper *timeout)
2088 {
2089         int ret = 0;
2090
2091         /*
2092          * With the hb lock held, we avoid races while we process the wakeup.
2093          * We only need to hold hb (and not hb2) to ensure atomicity as the
2094          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2095          * It can't be requeued from uaddr2 to something else since we don't
2096          * support a PI aware source futex for requeue.
2097          */
2098         if (!match_futex(&q->key, key2)) {
2099                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2100                 /*
2101                  * We were woken prior to requeue by a timeout or a signal.
2102                  * Unqueue the futex_q and determine which it was.
2103                  */
2104                 plist_del(&q->list, &q->list.plist);
2105                 drop_futex_key_refs(&q->key);
2106
2107                 if (timeout && !timeout->task)
2108                         ret = -ETIMEDOUT;
2109                 else
2110                         ret = -ERESTARTNOINTR;
2111         }
2112         return ret;
2113 }
2114
2115 /**
2116  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2117  * @uaddr:      the futex we initially wait on (non-pi)
2118  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2119  *              the same type, no requeueing from private to shared, etc.
2120  * @val:        the expected value of uaddr
2121  * @abs_time:   absolute timeout
2122  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2123  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2124  * @uaddr2:     the pi futex we will take prior to returning to user-space
2125  *
2126  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2127  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2128  * complete the acquisition of the rt_mutex prior to returning to userspace.
2129  * This ensures the rt_mutex maintains an owner when it has waiters; without
2130  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2131  * need to.
2132  *
2133  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2134  * via the following:
2135  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2136  * 2) wakeup on uaddr2 after a requeue
2137  * 3) signal
2138  * 4) timeout
2139  *
2140  * If 3, cleanup and return -ERESTARTNOINTR.
2141  *
2142  * If 2, we may then block on trying to take the rt_mutex and return via:
2143  * 5) successful lock
2144  * 6) signal
2145  * 7) timeout
2146  * 8) other lock acquisition failure
2147  *
2148  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2149  *
2150  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2151  *
2152  * Returns:
2153  *  0 - On success
2154  * <0 - On error
2155  */
2156 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2157                                  u32 val, ktime_t *abs_time, u32 bitset,
2158                                  int clockrt, u32 __user *uaddr2)
2159 {
2160         struct hrtimer_sleeper timeout, *to = NULL;
2161         struct rt_mutex_waiter rt_waiter;
2162         struct rt_mutex *pi_mutex = NULL;
2163         struct futex_hash_bucket *hb;
2164         union futex_key key2;
2165         struct futex_q q;
2166         int res, ret;
2167
2168         if (!bitset)
2169                 return -EINVAL;
2170
2171         if (abs_time) {
2172                 to = &timeout;
2173                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2174                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2175                 hrtimer_init_sleeper(to, current);
2176                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2177                                              current->timer_slack_ns);
2178         }
2179
2180         /*
2181          * The waiter is allocated on our stack, manipulated by the requeue
2182          * code while we sleep on uaddr.
2183          */
2184         debug_rt_mutex_init_waiter(&rt_waiter);
2185         rt_waiter.task = NULL;
2186
2187         key2 = FUTEX_KEY_INIT;
2188         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2189         if (unlikely(ret != 0))
2190                 goto out;
2191
2192         q.pi_state = NULL;
2193         q.bitset = bitset;
2194         q.rt_waiter = &rt_waiter;
2195         q.requeue_pi_key = &key2;
2196
2197         /* Prepare to wait on uaddr. */
2198         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2199         if (ret)
2200                 goto out_key2;
2201
2202         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2203         futex_wait_queue_me(hb, &q, to);
2204
2205         spin_lock(&hb->lock);
2206         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2207         spin_unlock(&hb->lock);
2208         if (ret)
2209                 goto out_put_keys;
2210
2211         /*
2212          * In order for us to be here, we know our q.key == key2, and since
2213          * we took the hb->lock above, we also know that futex_requeue() has
2214          * completed and we no longer have to concern ourselves with a wakeup
2215          * race with the atomic proxy lock acquition by the requeue code.
2216          */
2217
2218         /* Check if the requeue code acquired the second futex for us. */
2219         if (!q.rt_waiter) {
2220                 /*
2221                  * Got the lock. We might not be the anticipated owner if we
2222                  * did a lock-steal - fix up the PI-state in that case.
2223                  */
2224                 if (q.pi_state && (q.pi_state->owner != current)) {
2225                         spin_lock(q.lock_ptr);
2226                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2227                                                    fshared);
2228                         spin_unlock(q.lock_ptr);
2229                 }
2230         } else {
2231                 /*
2232                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2233                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2234                  * the pi_state.
2235                  */
2236                 WARN_ON(!&q.pi_state);
2237                 pi_mutex = &q.pi_state->pi_mutex;
2238                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2239                 debug_rt_mutex_free_waiter(&rt_waiter);
2240
2241                 spin_lock(q.lock_ptr);
2242                 /*
2243                  * Fixup the pi_state owner and possibly acquire the lock if we
2244                  * haven't already.
2245                  */
2246                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2247                 /*
2248                  * If fixup_owner() returned an error, proprogate that.  If it
2249                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2250                  */
2251                 if (res)
2252                         ret = (res < 0) ? res : 0;
2253
2254                 /* Unqueue and drop the lock. */
2255                 unqueue_me_pi(&q);
2256         }
2257
2258         /*
2259          * If fixup_pi_state_owner() faulted and was unable to handle the
2260          * fault, unlock the rt_mutex and return the fault to userspace.
2261          */
2262         if (ret == -EFAULT) {
2263                 if (rt_mutex_owner(pi_mutex) == current)
2264                         rt_mutex_unlock(pi_mutex);
2265         } else if (ret == -EINTR) {
2266                 /*
2267                  * We've already been requeued, but cannot restart by calling
2268                  * futex_lock_pi() directly. We could restart this syscall, but
2269                  * it would detect that the user space "val" changed and return
2270                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2271                  * -EWOULDBLOCK directly.
2272                  */
2273                 ret = -EWOULDBLOCK;
2274         }
2275
2276 out_put_keys:
2277         put_futex_key(fshared, &q.key);
2278 out_key2:
2279         put_futex_key(fshared, &key2);
2280
2281 out:
2282         if (to) {
2283                 hrtimer_cancel(&to->timer);
2284                 destroy_hrtimer_on_stack(&to->timer);
2285         }
2286         return ret;
2287 }
2288
2289 /*
2290  * Support for robust futexes: the kernel cleans up held futexes at
2291  * thread exit time.
2292  *
2293  * Implementation: user-space maintains a per-thread list of locks it
2294  * is holding. Upon do_exit(), the kernel carefully walks this list,
2295  * and marks all locks that are owned by this thread with the
2296  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2297  * always manipulated with the lock held, so the list is private and
2298  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2299  * field, to allow the kernel to clean up if the thread dies after
2300  * acquiring the lock, but just before it could have added itself to
2301  * the list. There can only be one such pending lock.
2302  */
2303
2304 /**
2305  * sys_set_robust_list - set the robust-futex list head of a task
2306  * @head: pointer to the list-head
2307  * @len: length of the list-head, as userspace expects
2308  */
2309 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2310                 size_t, len)
2311 {
2312         if (!futex_cmpxchg_enabled)
2313                 return -ENOSYS;
2314         /*
2315          * The kernel knows only one size for now:
2316          */
2317         if (unlikely(len != sizeof(*head)))
2318                 return -EINVAL;
2319
2320         current->robust_list = head;
2321
2322         return 0;
2323 }
2324
2325 /**
2326  * sys_get_robust_list - get the robust-futex list head of a task
2327  * @pid: pid of the process [zero for current task]
2328  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2329  * @len_ptr: pointer to a length field, the kernel fills in the header size
2330  */
2331 SYSCALL_DEFINE3(get_robust_list, int, pid,
2332                 struct robust_list_head __user * __user *, head_ptr,
2333                 size_t __user *, len_ptr)
2334 {
2335         struct robust_list_head __user *head;
2336         unsigned long ret;
2337         const struct cred *cred = current_cred(), *pcred;
2338
2339         if (!futex_cmpxchg_enabled)
2340                 return -ENOSYS;
2341
2342         if (!pid)
2343                 head = current->robust_list;
2344         else {
2345                 struct task_struct *p;
2346
2347                 ret = -ESRCH;
2348                 rcu_read_lock();
2349                 p = find_task_by_vpid(pid);
2350                 if (!p)
2351                         goto err_unlock;
2352                 ret = -EPERM;
2353                 pcred = __task_cred(p);
2354                 if (cred->euid != pcred->euid &&
2355                     cred->euid != pcred->uid &&
2356                     !capable(CAP_SYS_PTRACE))
2357                         goto err_unlock;
2358                 head = p->robust_list;
2359                 rcu_read_unlock();
2360         }
2361
2362         if (put_user(sizeof(*head), len_ptr))
2363                 return -EFAULT;
2364         return put_user(head, head_ptr);
2365
2366 err_unlock:
2367         rcu_read_unlock();
2368
2369         return ret;
2370 }
2371
2372 /*
2373  * Process a futex-list entry, check whether it's owned by the
2374  * dying task, and do notification if so:
2375  */
2376 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2377 {
2378         u32 uval, nval, mval;
2379
2380 retry:
2381         if (get_user(uval, uaddr))
2382                 return -1;
2383
2384         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2385                 /*
2386                  * Ok, this dying thread is truly holding a futex
2387                  * of interest. Set the OWNER_DIED bit atomically
2388                  * via cmpxchg, and if the value had FUTEX_WAITERS
2389                  * set, wake up a waiter (if any). (We have to do a
2390                  * futex_wake() even if OWNER_DIED is already set -
2391                  * to handle the rare but possible case of recursive
2392                  * thread-death.) The rest of the cleanup is done in
2393                  * userspace.
2394                  */
2395                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2396                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2397
2398                 if (nval == -EFAULT)
2399                         return -1;
2400
2401                 if (nval != uval)
2402                         goto retry;
2403
2404                 /*
2405                  * Wake robust non-PI futexes here. The wakeup of
2406                  * PI futexes happens in exit_pi_state():
2407                  */
2408                 if (!pi && (uval & FUTEX_WAITERS))
2409                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2410         }
2411         return 0;
2412 }
2413
2414 /*
2415  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2416  */
2417 static inline int fetch_robust_entry(struct robust_list __user **entry,
2418                                      struct robust_list __user * __user *head,
2419                                      int *pi)
2420 {
2421         unsigned long uentry;
2422
2423         if (get_user(uentry, (unsigned long __user *)head))
2424                 return -EFAULT;
2425
2426         *entry = (void __user *)(uentry & ~1UL);
2427         *pi = uentry & 1;
2428
2429         return 0;
2430 }
2431
2432 /*
2433  * Walk curr->robust_list (very carefully, it's a userspace list!)
2434  * and mark any locks found there dead, and notify any waiters.
2435  *
2436  * We silently return on any sign of list-walking problem.
2437  */
2438 void exit_robust_list(struct task_struct *curr)
2439 {
2440         struct robust_list_head __user *head = curr->robust_list;
2441         struct robust_list __user *entry, *next_entry, *pending;
2442         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2443         unsigned long futex_offset;
2444         int rc;
2445
2446         if (!futex_cmpxchg_enabled)
2447                 return;
2448
2449         /*
2450          * Fetch the list head (which was registered earlier, via
2451          * sys_set_robust_list()):
2452          */
2453         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2454                 return;
2455         /*
2456          * Fetch the relative futex offset:
2457          */
2458         if (get_user(futex_offset, &head->futex_offset))
2459                 return;
2460         /*
2461          * Fetch any possibly pending lock-add first, and handle it
2462          * if it exists:
2463          */
2464         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2465                 return;
2466
2467         next_entry = NULL;      /* avoid warning with gcc */
2468         while (entry != &head->list) {
2469                 /*
2470                  * Fetch the next entry in the list before calling
2471                  * handle_futex_death:
2472                  */
2473                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2474                 /*
2475                  * A pending lock might already be on the list, so
2476                  * don't process it twice:
2477                  */
2478                 if (entry != pending)
2479                         if (handle_futex_death((void __user *)entry + futex_offset,
2480                                                 curr, pi))
2481                                 return;
2482                 if (rc)
2483                         return;
2484                 entry = next_entry;
2485                 pi = next_pi;
2486                 /*
2487                  * Avoid excessively long or circular lists:
2488                  */
2489                 if (!--limit)
2490                         break;
2491
2492                 cond_resched();
2493         }
2494
2495         if (pending)
2496                 handle_futex_death((void __user *)pending + futex_offset,
2497                                    curr, pip);
2498 }
2499
2500 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2501                 u32 __user *uaddr2, u32 val2, u32 val3)
2502 {
2503         int clockrt, ret = -ENOSYS;
2504         int cmd = op & FUTEX_CMD_MASK;
2505         int fshared = 0;
2506
2507         if (!(op & FUTEX_PRIVATE_FLAG))
2508                 fshared = 1;
2509
2510         clockrt = op & FUTEX_CLOCK_REALTIME;
2511         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2512                 return -ENOSYS;
2513
2514         switch (cmd) {
2515         case FUTEX_WAIT:
2516                 val3 = FUTEX_BITSET_MATCH_ANY;
2517         case FUTEX_WAIT_BITSET:
2518                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2519                 break;
2520         case FUTEX_WAKE:
2521                 val3 = FUTEX_BITSET_MATCH_ANY;
2522         case FUTEX_WAKE_BITSET:
2523                 ret = futex_wake(uaddr, fshared, val, val3);
2524                 break;
2525         case FUTEX_REQUEUE:
2526                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2527                 break;
2528         case FUTEX_CMP_REQUEUE:
2529                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2530                                     0);
2531                 break;
2532         case FUTEX_WAKE_OP:
2533                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2534                 break;
2535         case FUTEX_LOCK_PI:
2536                 if (futex_cmpxchg_enabled)
2537                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2538                 break;
2539         case FUTEX_UNLOCK_PI:
2540                 if (futex_cmpxchg_enabled)
2541                         ret = futex_unlock_pi(uaddr, fshared);
2542                 break;
2543         case FUTEX_TRYLOCK_PI:
2544                 if (futex_cmpxchg_enabled)
2545                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2546                 break;
2547         case FUTEX_WAIT_REQUEUE_PI:
2548                 val3 = FUTEX_BITSET_MATCH_ANY;
2549                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2550                                             clockrt, uaddr2);
2551                 break;
2552         case FUTEX_CMP_REQUEUE_PI:
2553                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2554                                     1);
2555                 break;
2556         default:
2557                 ret = -ENOSYS;
2558         }
2559         return ret;
2560 }
2561
2562
2563 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2564                 struct timespec __user *, utime, u32 __user *, uaddr2,
2565                 u32, val3)
2566 {
2567         struct timespec ts;
2568         ktime_t t, *tp = NULL;
2569         u32 val2 = 0;
2570         int cmd = op & FUTEX_CMD_MASK;
2571
2572         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2573                       cmd == FUTEX_WAIT_BITSET ||
2574                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2575                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2576                         return -EFAULT;
2577                 if (!timespec_valid(&ts))
2578                         return -EINVAL;
2579
2580                 t = timespec_to_ktime(ts);
2581                 if (cmd == FUTEX_WAIT)
2582                         t = ktime_add_safe(ktime_get(), t);
2583                 tp = &t;
2584         }
2585         /*
2586          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2587          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2588          */
2589         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2590             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2591                 val2 = (u32) (unsigned long) utime;
2592
2593         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2594 }
2595
2596 static int __init futex_init(void)
2597 {
2598         u32 curval;
2599         int i;
2600
2601         /*
2602          * This will fail and we want it. Some arch implementations do
2603          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2604          * functionality. We want to know that before we call in any
2605          * of the complex code paths. Also we want to prevent
2606          * registration of robust lists in that case. NULL is
2607          * guaranteed to fault and we get -EFAULT on functional
2608          * implementation, the non functional ones will return
2609          * -ENOSYS.
2610          */
2611         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2612         if (curval == -EFAULT)
2613                 futex_cmpxchg_enabled = 1;
2614
2615         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2616                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2617                 spin_lock_init(&futex_queues[i].lock);
2618         }
2619
2620         return 0;
2621 }
2622 __initcall(futex_init);