476603afd1478191fa3c1188dd92f94bcd1b666c
[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         /* Bitset for the optional bitmasked wakeup */
119         u32 bitset;
120 };
121
122 /*
123  * Hash buckets are shared by all the futex_keys that hash to the same
124  * location.  Each key may have multiple futex_q structures, one for each task
125  * waiting on a futex.
126  */
127 struct futex_hash_bucket {
128         spinlock_t lock;
129         struct plist_head chain;
130 };
131
132 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
133
134 /*
135  * We hash on the keys returned from get_futex_key (see below).
136  */
137 static struct futex_hash_bucket *hash_futex(union futex_key *key)
138 {
139         u32 hash = jhash2((u32*)&key->both.word,
140                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
141                           key->both.offset);
142         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
143 }
144
145 /*
146  * Return 1 if two futex_keys are equal, 0 otherwise.
147  */
148 static inline int match_futex(union futex_key *key1, union futex_key *key2)
149 {
150         return (key1->both.word == key2->both.word
151                 && key1->both.ptr == key2->both.ptr
152                 && key1->both.offset == key2->both.offset);
153 }
154
155 /*
156  * Take a reference to the resource addressed by a key.
157  * Can be called while holding spinlocks.
158  *
159  */
160 static void get_futex_key_refs(union futex_key *key)
161 {
162         if (!key->both.ptr)
163                 return;
164
165         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
166         case FUT_OFF_INODE:
167                 atomic_inc(&key->shared.inode->i_count);
168                 break;
169         case FUT_OFF_MMSHARED:
170                 atomic_inc(&key->private.mm->mm_count);
171                 break;
172         }
173 }
174
175 /*
176  * Drop a reference to the resource addressed by a key.
177  * The hash bucket spinlock must not be held.
178  */
179 static void drop_futex_key_refs(union futex_key *key)
180 {
181         if (!key->both.ptr) {
182                 /* If we're here then we tried to put a key we failed to get */
183                 WARN_ON_ONCE(1);
184                 return;
185         }
186
187         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
188         case FUT_OFF_INODE:
189                 iput(key->shared.inode);
190                 break;
191         case FUT_OFF_MMSHARED:
192                 mmdrop(key->private.mm);
193                 break;
194         }
195 }
196
197 /**
198  * get_futex_key - Get parameters which are the keys for a futex.
199  * @uaddr: virtual address of the futex
200  * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
201  * @key: address where result is stored.
202  * @rw: mapping needs to be read/write (values: VERIFY_READ, VERIFY_WRITE)
203  *
204  * Returns a negative error code or 0
205  * The key words are stored in *key on success.
206  *
207  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
208  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
209  * We can usually work out the index without swapping in the page.
210  *
211  * lock_page() might sleep, the caller should not hold a spinlock.
212  */
213 static int
214 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
215 {
216         unsigned long address = (unsigned long)uaddr;
217         struct mm_struct *mm = current->mm;
218         struct page *page;
219         int err;
220
221         /*
222          * The futex address must be "naturally" aligned.
223          */
224         key->both.offset = address % PAGE_SIZE;
225         if (unlikely((address % sizeof(u32)) != 0))
226                 return -EINVAL;
227         address -= key->both.offset;
228
229         /*
230          * PROCESS_PRIVATE futexes are fast.
231          * As the mm cannot disappear under us and the 'key' only needs
232          * virtual address, we dont even have to find the underlying vma.
233          * Note : We do have to check 'uaddr' is a valid user address,
234          *        but access_ok() should be faster than find_vma()
235          */
236         if (!fshared) {
237                 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
238                         return -EFAULT;
239                 key->private.mm = mm;
240                 key->private.address = address;
241                 get_futex_key_refs(key);
242                 return 0;
243         }
244
245 again:
246         err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
247         if (err < 0)
248                 return err;
249
250         lock_page(page);
251         if (!page->mapping) {
252                 unlock_page(page);
253                 put_page(page);
254                 goto again;
255         }
256
257         /*
258          * Private mappings are handled in a simple way.
259          *
260          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
261          * it's a read-only handle, it's expected that futexes attach to
262          * the object not the particular process.
263          */
264         if (PageAnon(page)) {
265                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
266                 key->private.mm = mm;
267                 key->private.address = address;
268         } else {
269                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
270                 key->shared.inode = page->mapping->host;
271                 key->shared.pgoff = page->index;
272         }
273
274         get_futex_key_refs(key);
275
276         unlock_page(page);
277         put_page(page);
278         return 0;
279 }
280
281 static inline
282 void put_futex_key(int fshared, union futex_key *key)
283 {
284         drop_futex_key_refs(key);
285 }
286
287 /**
288  * futex_top_waiter() - Return the highest priority waiter on a futex
289  * @hb:     the hash bucket the futex_q's reside in
290  * @key:    the futex key (to distinguish it from other futex futex_q's)
291  *
292  * Must be called with the hb lock held.
293  */
294 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
295                                         union futex_key *key)
296 {
297         struct futex_q *this;
298
299         plist_for_each_entry(this, &hb->chain, list) {
300                 if (match_futex(&this->key, key))
301                         return this;
302         }
303         return NULL;
304 }
305
306 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
307 {
308         u32 curval;
309
310         pagefault_disable();
311         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
312         pagefault_enable();
313
314         return curval;
315 }
316
317 static int get_futex_value_locked(u32 *dest, u32 __user *from)
318 {
319         int ret;
320
321         pagefault_disable();
322         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
323         pagefault_enable();
324
325         return ret ? -EFAULT : 0;
326 }
327
328
329 /*
330  * PI code:
331  */
332 static int refill_pi_state_cache(void)
333 {
334         struct futex_pi_state *pi_state;
335
336         if (likely(current->pi_state_cache))
337                 return 0;
338
339         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
340
341         if (!pi_state)
342                 return -ENOMEM;
343
344         INIT_LIST_HEAD(&pi_state->list);
345         /* pi_mutex gets initialized later */
346         pi_state->owner = NULL;
347         atomic_set(&pi_state->refcount, 1);
348         pi_state->key = FUTEX_KEY_INIT;
349
350         current->pi_state_cache = pi_state;
351
352         return 0;
353 }
354
355 static struct futex_pi_state * alloc_pi_state(void)
356 {
357         struct futex_pi_state *pi_state = current->pi_state_cache;
358
359         WARN_ON(!pi_state);
360         current->pi_state_cache = NULL;
361
362         return pi_state;
363 }
364
365 static void free_pi_state(struct futex_pi_state *pi_state)
366 {
367         if (!atomic_dec_and_test(&pi_state->refcount))
368                 return;
369
370         /*
371          * If pi_state->owner is NULL, the owner is most probably dying
372          * and has cleaned up the pi_state already
373          */
374         if (pi_state->owner) {
375                 spin_lock_irq(&pi_state->owner->pi_lock);
376                 list_del_init(&pi_state->list);
377                 spin_unlock_irq(&pi_state->owner->pi_lock);
378
379                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
380         }
381
382         if (current->pi_state_cache)
383                 kfree(pi_state);
384         else {
385                 /*
386                  * pi_state->list is already empty.
387                  * clear pi_state->owner.
388                  * refcount is at 0 - put it back to 1.
389                  */
390                 pi_state->owner = NULL;
391                 atomic_set(&pi_state->refcount, 1);
392                 current->pi_state_cache = pi_state;
393         }
394 }
395
396 /*
397  * Look up the task based on what TID userspace gave us.
398  * We dont trust it.
399  */
400 static struct task_struct * futex_find_get_task(pid_t pid)
401 {
402         struct task_struct *p;
403         const struct cred *cred = current_cred(), *pcred;
404
405         rcu_read_lock();
406         p = find_task_by_vpid(pid);
407         if (!p) {
408                 p = ERR_PTR(-ESRCH);
409         } else {
410                 pcred = __task_cred(p);
411                 if (cred->euid != pcred->euid &&
412                     cred->euid != pcred->uid)
413                         p = ERR_PTR(-ESRCH);
414                 else
415                         get_task_struct(p);
416         }
417
418         rcu_read_unlock();
419
420         return p;
421 }
422
423 /*
424  * This task is holding PI mutexes at exit time => bad.
425  * Kernel cleans up PI-state, but userspace is likely hosed.
426  * (Robust-futex cleanup is separate and might save the day for userspace.)
427  */
428 void exit_pi_state_list(struct task_struct *curr)
429 {
430         struct list_head *next, *head = &curr->pi_state_list;
431         struct futex_pi_state *pi_state;
432         struct futex_hash_bucket *hb;
433         union futex_key key = FUTEX_KEY_INIT;
434
435         if (!futex_cmpxchg_enabled)
436                 return;
437         /*
438          * We are a ZOMBIE and nobody can enqueue itself on
439          * pi_state_list anymore, but we have to be careful
440          * versus waiters unqueueing themselves:
441          */
442         spin_lock_irq(&curr->pi_lock);
443         while (!list_empty(head)) {
444
445                 next = head->next;
446                 pi_state = list_entry(next, struct futex_pi_state, list);
447                 key = pi_state->key;
448                 hb = hash_futex(&key);
449                 spin_unlock_irq(&curr->pi_lock);
450
451                 spin_lock(&hb->lock);
452
453                 spin_lock_irq(&curr->pi_lock);
454                 /*
455                  * We dropped the pi-lock, so re-check whether this
456                  * task still owns the PI-state:
457                  */
458                 if (head->next != next) {
459                         spin_unlock(&hb->lock);
460                         continue;
461                 }
462
463                 WARN_ON(pi_state->owner != curr);
464                 WARN_ON(list_empty(&pi_state->list));
465                 list_del_init(&pi_state->list);
466                 pi_state->owner = NULL;
467                 spin_unlock_irq(&curr->pi_lock);
468
469                 rt_mutex_unlock(&pi_state->pi_mutex);
470
471                 spin_unlock(&hb->lock);
472
473                 spin_lock_irq(&curr->pi_lock);
474         }
475         spin_unlock_irq(&curr->pi_lock);
476 }
477
478 static int
479 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
480                 union futex_key *key, struct futex_pi_state **ps)
481 {
482         struct futex_pi_state *pi_state = NULL;
483         struct futex_q *this, *next;
484         struct plist_head *head;
485         struct task_struct *p;
486         pid_t pid = uval & FUTEX_TID_MASK;
487
488         head = &hb->chain;
489
490         plist_for_each_entry_safe(this, next, head, list) {
491                 if (match_futex(&this->key, key)) {
492                         /*
493                          * Another waiter already exists - bump up
494                          * the refcount and return its pi_state:
495                          */
496                         pi_state = this->pi_state;
497                         /*
498                          * Userspace might have messed up non PI and PI futexes
499                          */
500                         if (unlikely(!pi_state))
501                                 return -EINVAL;
502
503                         WARN_ON(!atomic_read(&pi_state->refcount));
504                         WARN_ON(pid && pi_state->owner &&
505                                 pi_state->owner->pid != pid);
506
507                         atomic_inc(&pi_state->refcount);
508                         *ps = pi_state;
509
510                         return 0;
511                 }
512         }
513
514         /*
515          * We are the first waiter - try to look up the real owner and attach
516          * the new pi_state to it, but bail out when TID = 0
517          */
518         if (!pid)
519                 return -ESRCH;
520         p = futex_find_get_task(pid);
521         if (IS_ERR(p))
522                 return PTR_ERR(p);
523
524         /*
525          * We need to look at the task state flags to figure out,
526          * whether the task is exiting. To protect against the do_exit
527          * change of the task flags, we do this protected by
528          * p->pi_lock:
529          */
530         spin_lock_irq(&p->pi_lock);
531         if (unlikely(p->flags & PF_EXITING)) {
532                 /*
533                  * The task is on the way out. When PF_EXITPIDONE is
534                  * set, we know that the task has finished the
535                  * cleanup:
536                  */
537                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
538
539                 spin_unlock_irq(&p->pi_lock);
540                 put_task_struct(p);
541                 return ret;
542         }
543
544         pi_state = alloc_pi_state();
545
546         /*
547          * Initialize the pi_mutex in locked state and make 'p'
548          * the owner of it:
549          */
550         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
551
552         /* Store the key for possible exit cleanups: */
553         pi_state->key = *key;
554
555         WARN_ON(!list_empty(&pi_state->list));
556         list_add(&pi_state->list, &p->pi_state_list);
557         pi_state->owner = p;
558         spin_unlock_irq(&p->pi_lock);
559
560         put_task_struct(p);
561
562         *ps = pi_state;
563
564         return 0;
565 }
566
567 /**
568  * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
569  * @uaddr:              the pi futex user address
570  * @hb:                 the pi futex hash bucket
571  * @key:                the futex key associated with uaddr and hb
572  * @ps:                 the pi_state pointer where we store the result of the
573  *                      lookup
574  * @task:               the task to perform the atomic lock work for.  This will
575  *                      be "current" except in the case of requeue pi.
576  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
577  *
578  * Returns:
579  *  0 - ready to wait
580  *  1 - acquired the lock
581  * <0 - error
582  *
583  * The hb->lock and futex_key refs shall be held by the caller.
584  */
585 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
586                                 union futex_key *key,
587                                 struct futex_pi_state **ps,
588                                 struct task_struct *task, int set_waiters)
589 {
590         int lock_taken, ret, ownerdied = 0;
591         u32 uval, newval, curval;
592
593 retry:
594         ret = lock_taken = 0;
595
596         /*
597          * To avoid races, we attempt to take the lock here again
598          * (by doing a 0 -> TID atomic cmpxchg), while holding all
599          * the locks. It will most likely not succeed.
600          */
601         newval = task_pid_vnr(task);
602         if (set_waiters)
603                 newval |= FUTEX_WAITERS;
604
605         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
606
607         if (unlikely(curval == -EFAULT))
608                 return -EFAULT;
609
610         /*
611          * Detect deadlocks.
612          */
613         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
614                 return -EDEADLK;
615
616         /*
617          * Surprise - we got the lock. Just return to userspace:
618          */
619         if (unlikely(!curval))
620                 return 1;
621
622         uval = curval;
623
624         /*
625          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
626          * to wake at the next unlock.
627          */
628         newval = curval | FUTEX_WAITERS;
629
630         /*
631          * There are two cases, where a futex might have no owner (the
632          * owner TID is 0): OWNER_DIED. We take over the futex in this
633          * case. We also do an unconditional take over, when the owner
634          * of the futex died.
635          *
636          * This is safe as we are protected by the hash bucket lock !
637          */
638         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
639                 /* Keep the OWNER_DIED bit */
640                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
641                 ownerdied = 0;
642                 lock_taken = 1;
643         }
644
645         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
646
647         if (unlikely(curval == -EFAULT))
648                 return -EFAULT;
649         if (unlikely(curval != uval))
650                 goto retry;
651
652         /*
653          * We took the lock due to owner died take over.
654          */
655         if (unlikely(lock_taken))
656                 return 1;
657
658         /*
659          * We dont have the lock. Look up the PI state (or create it if
660          * we are the first waiter):
661          */
662         ret = lookup_pi_state(uval, hb, key, ps);
663
664         if (unlikely(ret)) {
665                 switch (ret) {
666                 case -ESRCH:
667                         /*
668                          * No owner found for this futex. Check if the
669                          * OWNER_DIED bit is set to figure out whether
670                          * this is a robust futex or not.
671                          */
672                         if (get_futex_value_locked(&curval, uaddr))
673                                 return -EFAULT;
674
675                         /*
676                          * We simply start over in case of a robust
677                          * futex. The code above will take the futex
678                          * and return happy.
679                          */
680                         if (curval & FUTEX_OWNER_DIED) {
681                                 ownerdied = 1;
682                                 goto retry;
683                         }
684                 default:
685                         break;
686                 }
687         }
688
689         return ret;
690 }
691
692 /*
693  * The hash bucket lock must be held when this is called.
694  * Afterwards, the futex_q must not be accessed.
695  */
696 static void wake_futex(struct futex_q *q)
697 {
698         struct task_struct *p = q->task;
699
700         /*
701          * We set q->lock_ptr = NULL _before_ we wake up the task. If
702          * a non futex wake up happens on another CPU then the task
703          * might exit and p would dereference a non existing task
704          * struct. Prevent this by holding a reference on p across the
705          * wake up.
706          */
707         get_task_struct(p);
708
709         plist_del(&q->list, &q->list.plist);
710         /*
711          * The waiting task can free the futex_q as soon as
712          * q->lock_ptr = NULL is written, without taking any locks. A
713          * memory barrier is required here to prevent the following
714          * store to lock_ptr from getting ahead of the plist_del.
715          */
716         smp_wmb();
717         q->lock_ptr = NULL;
718
719         wake_up_state(p, TASK_NORMAL);
720         put_task_struct(p);
721 }
722
723 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
724 {
725         struct task_struct *new_owner;
726         struct futex_pi_state *pi_state = this->pi_state;
727         u32 curval, newval;
728
729         if (!pi_state)
730                 return -EINVAL;
731
732         spin_lock(&pi_state->pi_mutex.wait_lock);
733         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
734
735         /*
736          * This happens when we have stolen the lock and the original
737          * pending owner did not enqueue itself back on the rt_mutex.
738          * Thats not a tragedy. We know that way, that a lock waiter
739          * is on the fly. We make the futex_q waiter the pending owner.
740          */
741         if (!new_owner)
742                 new_owner = this->task;
743
744         /*
745          * We pass it to the next owner. (The WAITERS bit is always
746          * kept enabled while there is PI state around. We must also
747          * preserve the owner died bit.)
748          */
749         if (!(uval & FUTEX_OWNER_DIED)) {
750                 int ret = 0;
751
752                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
753
754                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
755
756                 if (curval == -EFAULT)
757                         ret = -EFAULT;
758                 else if (curval != uval)
759                         ret = -EINVAL;
760                 if (ret) {
761                         spin_unlock(&pi_state->pi_mutex.wait_lock);
762                         return ret;
763                 }
764         }
765
766         spin_lock_irq(&pi_state->owner->pi_lock);
767         WARN_ON(list_empty(&pi_state->list));
768         list_del_init(&pi_state->list);
769         spin_unlock_irq(&pi_state->owner->pi_lock);
770
771         spin_lock_irq(&new_owner->pi_lock);
772         WARN_ON(!list_empty(&pi_state->list));
773         list_add(&pi_state->list, &new_owner->pi_state_list);
774         pi_state->owner = new_owner;
775         spin_unlock_irq(&new_owner->pi_lock);
776
777         spin_unlock(&pi_state->pi_mutex.wait_lock);
778         rt_mutex_unlock(&pi_state->pi_mutex);
779
780         return 0;
781 }
782
783 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
784 {
785         u32 oldval;
786
787         /*
788          * There is no waiter, so we unlock the futex. The owner died
789          * bit has not to be preserved here. We are the owner:
790          */
791         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
792
793         if (oldval == -EFAULT)
794                 return oldval;
795         if (oldval != uval)
796                 return -EAGAIN;
797
798         return 0;
799 }
800
801 /*
802  * Express the locking dependencies for lockdep:
803  */
804 static inline void
805 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
806 {
807         if (hb1 <= hb2) {
808                 spin_lock(&hb1->lock);
809                 if (hb1 < hb2)
810                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
811         } else { /* hb1 > hb2 */
812                 spin_lock(&hb2->lock);
813                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
814         }
815 }
816
817 static inline void
818 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
819 {
820         spin_unlock(&hb1->lock);
821         if (hb1 != hb2)
822                 spin_unlock(&hb2->lock);
823 }
824
825 /*
826  * Wake up waiters matching bitset queued on this futex (uaddr).
827  */
828 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
829 {
830         struct futex_hash_bucket *hb;
831         struct futex_q *this, *next;
832         struct plist_head *head;
833         union futex_key key = FUTEX_KEY_INIT;
834         int ret;
835
836         if (!bitset)
837                 return -EINVAL;
838
839         ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
840         if (unlikely(ret != 0))
841                 goto out;
842
843         hb = hash_futex(&key);
844         spin_lock(&hb->lock);
845         head = &hb->chain;
846
847         plist_for_each_entry_safe(this, next, head, list) {
848                 if (match_futex (&this->key, &key)) {
849                         if (this->pi_state || this->rt_waiter) {
850                                 ret = -EINVAL;
851                                 break;
852                         }
853
854                         /* Check if one of the bits is set in both bitsets */
855                         if (!(this->bitset & bitset))
856                                 continue;
857
858                         wake_futex(this);
859                         if (++ret >= nr_wake)
860                                 break;
861                 }
862         }
863
864         spin_unlock(&hb->lock);
865         put_futex_key(fshared, &key);
866 out:
867         return ret;
868 }
869
870 /*
871  * Wake up all waiters hashed on the physical page that is mapped
872  * to this virtual address:
873  */
874 static int
875 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
876               int nr_wake, int nr_wake2, int op)
877 {
878         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
879         struct futex_hash_bucket *hb1, *hb2;
880         struct plist_head *head;
881         struct futex_q *this, *next;
882         int ret, op_ret;
883
884 retry:
885         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
886         if (unlikely(ret != 0))
887                 goto out;
888         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
889         if (unlikely(ret != 0))
890                 goto out_put_key1;
891
892         hb1 = hash_futex(&key1);
893         hb2 = hash_futex(&key2);
894
895         double_lock_hb(hb1, hb2);
896 retry_private:
897         op_ret = futex_atomic_op_inuser(op, uaddr2);
898         if (unlikely(op_ret < 0)) {
899                 u32 dummy;
900
901                 double_unlock_hb(hb1, hb2);
902
903 #ifndef CONFIG_MMU
904                 /*
905                  * we don't get EFAULT from MMU faults if we don't have an MMU,
906                  * but we might get them from range checking
907                  */
908                 ret = op_ret;
909                 goto out_put_keys;
910 #endif
911
912                 if (unlikely(op_ret != -EFAULT)) {
913                         ret = op_ret;
914                         goto out_put_keys;
915                 }
916
917                 ret = get_user(dummy, uaddr2);
918                 if (ret)
919                         goto out_put_keys;
920
921                 if (!fshared)
922                         goto retry_private;
923
924                 put_futex_key(fshared, &key2);
925                 put_futex_key(fshared, &key1);
926                 goto retry;
927         }
928
929         head = &hb1->chain;
930
931         plist_for_each_entry_safe(this, next, head, list) {
932                 if (match_futex (&this->key, &key1)) {
933                         wake_futex(this);
934                         if (++ret >= nr_wake)
935                                 break;
936                 }
937         }
938
939         if (op_ret > 0) {
940                 head = &hb2->chain;
941
942                 op_ret = 0;
943                 plist_for_each_entry_safe(this, next, head, list) {
944                         if (match_futex (&this->key, &key2)) {
945                                 wake_futex(this);
946                                 if (++op_ret >= nr_wake2)
947                                         break;
948                         }
949                 }
950                 ret += op_ret;
951         }
952
953         double_unlock_hb(hb1, hb2);
954 out_put_keys:
955         put_futex_key(fshared, &key2);
956 out_put_key1:
957         put_futex_key(fshared, &key1);
958 out:
959         return ret;
960 }
961
962 /**
963  * requeue_futex() - Requeue a futex_q from one hb to another
964  * @q:          the futex_q to requeue
965  * @hb1:        the source hash_bucket
966  * @hb2:        the target hash_bucket
967  * @key2:       the new key for the requeued futex_q
968  */
969 static inline
970 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
971                    struct futex_hash_bucket *hb2, union futex_key *key2)
972 {
973
974         /*
975          * If key1 and key2 hash to the same bucket, no need to
976          * requeue.
977          */
978         if (likely(&hb1->chain != &hb2->chain)) {
979                 plist_del(&q->list, &hb1->chain);
980                 plist_add(&q->list, &hb2->chain);
981                 q->lock_ptr = &hb2->lock;
982 #ifdef CONFIG_DEBUG_PI_LIST
983                 q->list.plist.lock = &hb2->lock;
984 #endif
985         }
986         get_futex_key_refs(key2);
987         q->key = *key2;
988 }
989
990 /**
991  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
992  * q:   the futex_q
993  * key: the key of the requeue target futex
994  *
995  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
996  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
997  * to the requeue target futex so the waiter can detect the wakeup on the right
998  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
999  * atomic lock acquisition.  Must be called with the q->lock_ptr held.
1000  */
1001 static inline
1002 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key)
1003 {
1004         drop_futex_key_refs(&q->key);
1005         get_futex_key_refs(key);
1006         q->key = *key;
1007
1008         WARN_ON(plist_node_empty(&q->list));
1009         plist_del(&q->list, &q->list.plist);
1010
1011         WARN_ON(!q->rt_waiter);
1012         q->rt_waiter = NULL;
1013
1014         wake_up_state(q->task, TASK_NORMAL);
1015 }
1016
1017 /**
1018  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1019  * @pifutex:            the user address of the to futex
1020  * @hb1:                the from futex hash bucket, must be locked by the caller
1021  * @hb2:                the to futex hash bucket, must be locked by the caller
1022  * @key1:               the from futex key
1023  * @key2:               the to futex key
1024  * @ps:                 address to store the pi_state pointer
1025  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1026  *
1027  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1028  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1029  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1030  * hb1 and hb2 must be held by the caller.
1031  *
1032  * Returns:
1033  *  0 - failed to acquire the lock atomicly
1034  *  1 - acquired the lock
1035  * <0 - error
1036  */
1037 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1038                                  struct futex_hash_bucket *hb1,
1039                                  struct futex_hash_bucket *hb2,
1040                                  union futex_key *key1, union futex_key *key2,
1041                                  struct futex_pi_state **ps, int set_waiters)
1042 {
1043         struct futex_q *top_waiter = NULL;
1044         u32 curval;
1045         int ret;
1046
1047         if (get_futex_value_locked(&curval, pifutex))
1048                 return -EFAULT;
1049
1050         /*
1051          * Find the top_waiter and determine if there are additional waiters.
1052          * If the caller intends to requeue more than 1 waiter to pifutex,
1053          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1054          * as we have means to handle the possible fault.  If not, don't set
1055          * the bit unecessarily as it will force the subsequent unlock to enter
1056          * the kernel.
1057          */
1058         top_waiter = futex_top_waiter(hb1, key1);
1059
1060         /* There are no waiters, nothing for us to do. */
1061         if (!top_waiter)
1062                 return 0;
1063
1064         /*
1065          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1066          * the contended case or if set_waiters is 1.  The pi_state is returned
1067          * in ps in contended cases.
1068          */
1069         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1070                                    set_waiters);
1071         if (ret == 1)
1072                 requeue_pi_wake_futex(top_waiter, key2);
1073
1074         return ret;
1075 }
1076
1077 /**
1078  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1079  * uaddr1:      source futex user address
1080  * uaddr2:      target futex user address
1081  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1082  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1083  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1084  *              pi futex (pi to pi requeue is not supported)
1085  *
1086  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1087  * uaddr2 atomically on behalf of the top waiter.
1088  *
1089  * Returns:
1090  * >=0 - on success, the number of tasks requeued or woken
1091  *  <0 - on error
1092  */
1093 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1094                          int nr_wake, int nr_requeue, u32 *cmpval,
1095                          int requeue_pi)
1096 {
1097         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1098         int drop_count = 0, task_count = 0, ret;
1099         struct futex_pi_state *pi_state = NULL;
1100         struct futex_hash_bucket *hb1, *hb2;
1101         struct plist_head *head1;
1102         struct futex_q *this, *next;
1103         u32 curval2;
1104
1105         if (requeue_pi) {
1106                 /*
1107                  * requeue_pi requires a pi_state, try to allocate it now
1108                  * without any locks in case it fails.
1109                  */
1110                 if (refill_pi_state_cache())
1111                         return -ENOMEM;
1112                 /*
1113                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1114                  * + nr_requeue, since it acquires the rt_mutex prior to
1115                  * returning to userspace, so as to not leave the rt_mutex with
1116                  * waiters and no owner.  However, second and third wake-ups
1117                  * cannot be predicted as they involve race conditions with the
1118                  * first wake and a fault while looking up the pi_state.  Both
1119                  * pthread_cond_signal() and pthread_cond_broadcast() should
1120                  * use nr_wake=1.
1121                  */
1122                 if (nr_wake != 1)
1123                         return -EINVAL;
1124         }
1125
1126 retry:
1127         if (pi_state != NULL) {
1128                 /*
1129                  * We will have to lookup the pi_state again, so free this one
1130                  * to keep the accounting correct.
1131                  */
1132                 free_pi_state(pi_state);
1133                 pi_state = NULL;
1134         }
1135
1136         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1137         if (unlikely(ret != 0))
1138                 goto out;
1139         ret = get_futex_key(uaddr2, fshared, &key2,
1140                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1141         if (unlikely(ret != 0))
1142                 goto out_put_key1;
1143
1144         hb1 = hash_futex(&key1);
1145         hb2 = hash_futex(&key2);
1146
1147 retry_private:
1148         double_lock_hb(hb1, hb2);
1149
1150         if (likely(cmpval != NULL)) {
1151                 u32 curval;
1152
1153                 ret = get_futex_value_locked(&curval, uaddr1);
1154
1155                 if (unlikely(ret)) {
1156                         double_unlock_hb(hb1, hb2);
1157
1158                         ret = get_user(curval, uaddr1);
1159                         if (ret)
1160                                 goto out_put_keys;
1161
1162                         if (!fshared)
1163                                 goto retry_private;
1164
1165                         put_futex_key(fshared, &key2);
1166                         put_futex_key(fshared, &key1);
1167                         goto retry;
1168                 }
1169                 if (curval != *cmpval) {
1170                         ret = -EAGAIN;
1171                         goto out_unlock;
1172                 }
1173         }
1174
1175         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1176                 /*
1177                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1178                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1179                  * bit.  We force this here where we are able to easily handle
1180                  * faults rather in the requeue loop below.
1181                  */
1182                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1183                                                  &key2, &pi_state, nr_requeue);
1184
1185                 /*
1186                  * At this point the top_waiter has either taken uaddr2 or is
1187                  * waiting on it.  If the former, then the pi_state will not
1188                  * exist yet, look it up one more time to ensure we have a
1189                  * reference to it.
1190                  */
1191                 if (ret == 1) {
1192                         WARN_ON(pi_state);
1193                         task_count++;
1194                         ret = get_futex_value_locked(&curval2, uaddr2);
1195                         if (!ret)
1196                                 ret = lookup_pi_state(curval2, hb2, &key2,
1197                                                       &pi_state);
1198                 }
1199
1200                 switch (ret) {
1201                 case 0:
1202                         break;
1203                 case -EFAULT:
1204                         double_unlock_hb(hb1, hb2);
1205                         put_futex_key(fshared, &key2);
1206                         put_futex_key(fshared, &key1);
1207                         ret = get_user(curval2, uaddr2);
1208                         if (!ret)
1209                                 goto retry;
1210                         goto out;
1211                 case -EAGAIN:
1212                         /* The owner was exiting, try again. */
1213                         double_unlock_hb(hb1, hb2);
1214                         put_futex_key(fshared, &key2);
1215                         put_futex_key(fshared, &key1);
1216                         cond_resched();
1217                         goto retry;
1218                 default:
1219                         goto out_unlock;
1220                 }
1221         }
1222
1223         head1 = &hb1->chain;
1224         plist_for_each_entry_safe(this, next, head1, list) {
1225                 if (task_count - nr_wake >= nr_requeue)
1226                         break;
1227
1228                 if (!match_futex(&this->key, &key1))
1229                         continue;
1230
1231                 WARN_ON(!requeue_pi && this->rt_waiter);
1232                 WARN_ON(requeue_pi && !this->rt_waiter);
1233
1234                 /*
1235                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1236                  * lock, we already woke the top_waiter.  If not, it will be
1237                  * woken by futex_unlock_pi().
1238                  */
1239                 if (++task_count <= nr_wake && !requeue_pi) {
1240                         wake_futex(this);
1241                         continue;
1242                 }
1243
1244                 /*
1245                  * Requeue nr_requeue waiters and possibly one more in the case
1246                  * of requeue_pi if we couldn't acquire the lock atomically.
1247                  */
1248                 if (requeue_pi) {
1249                         /* Prepare the waiter to take the rt_mutex. */
1250                         atomic_inc(&pi_state->refcount);
1251                         this->pi_state = pi_state;
1252                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1253                                                         this->rt_waiter,
1254                                                         this->task, 1);
1255                         if (ret == 1) {
1256                                 /* We got the lock. */
1257                                 requeue_pi_wake_futex(this, &key2);
1258                                 continue;
1259                         } else if (ret) {
1260                                 /* -EDEADLK */
1261                                 this->pi_state = NULL;
1262                                 free_pi_state(pi_state);
1263                                 goto out_unlock;
1264                         }
1265                 }
1266                 requeue_futex(this, hb1, hb2, &key2);
1267                 drop_count++;
1268         }
1269
1270 out_unlock:
1271         double_unlock_hb(hb1, hb2);
1272
1273         /*
1274          * drop_futex_key_refs() must be called outside the spinlocks. During
1275          * the requeue we moved futex_q's from the hash bucket at key1 to the
1276          * one at key2 and updated their key pointer.  We no longer need to
1277          * hold the references to key1.
1278          */
1279         while (--drop_count >= 0)
1280                 drop_futex_key_refs(&key1);
1281
1282 out_put_keys:
1283         put_futex_key(fshared, &key2);
1284 out_put_key1:
1285         put_futex_key(fshared, &key1);
1286 out:
1287         if (pi_state != NULL)
1288                 free_pi_state(pi_state);
1289         return ret ? ret : task_count;
1290 }
1291
1292 /* The key must be already stored in q->key. */
1293 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1294 {
1295         struct futex_hash_bucket *hb;
1296
1297         get_futex_key_refs(&q->key);
1298         hb = hash_futex(&q->key);
1299         q->lock_ptr = &hb->lock;
1300
1301         spin_lock(&hb->lock);
1302         return hb;
1303 }
1304
1305 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1306 {
1307         int prio;
1308
1309         /*
1310          * The priority used to register this element is
1311          * - either the real thread-priority for the real-time threads
1312          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1313          * - or MAX_RT_PRIO for non-RT threads.
1314          * Thus, all RT-threads are woken first in priority order, and
1315          * the others are woken last, in FIFO order.
1316          */
1317         prio = min(current->normal_prio, MAX_RT_PRIO);
1318
1319         plist_node_init(&q->list, prio);
1320 #ifdef CONFIG_DEBUG_PI_LIST
1321         q->list.plist.lock = &hb->lock;
1322 #endif
1323         plist_add(&q->list, &hb->chain);
1324         q->task = current;
1325         spin_unlock(&hb->lock);
1326 }
1327
1328 static inline void
1329 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1330 {
1331         spin_unlock(&hb->lock);
1332         drop_futex_key_refs(&q->key);
1333 }
1334
1335 /*
1336  * queue_me and unqueue_me must be called as a pair, each
1337  * exactly once.  They are called with the hashed spinlock held.
1338  */
1339
1340 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1341 static int unqueue_me(struct futex_q *q)
1342 {
1343         spinlock_t *lock_ptr;
1344         int ret = 0;
1345
1346         /* In the common case we don't take the spinlock, which is nice. */
1347 retry:
1348         lock_ptr = q->lock_ptr;
1349         barrier();
1350         if (lock_ptr != NULL) {
1351                 spin_lock(lock_ptr);
1352                 /*
1353                  * q->lock_ptr can change between reading it and
1354                  * spin_lock(), causing us to take the wrong lock.  This
1355                  * corrects the race condition.
1356                  *
1357                  * Reasoning goes like this: if we have the wrong lock,
1358                  * q->lock_ptr must have changed (maybe several times)
1359                  * between reading it and the spin_lock().  It can
1360                  * change again after the spin_lock() but only if it was
1361                  * already changed before the spin_lock().  It cannot,
1362                  * however, change back to the original value.  Therefore
1363                  * we can detect whether we acquired the correct lock.
1364                  */
1365                 if (unlikely(lock_ptr != q->lock_ptr)) {
1366                         spin_unlock(lock_ptr);
1367                         goto retry;
1368                 }
1369                 WARN_ON(plist_node_empty(&q->list));
1370                 plist_del(&q->list, &q->list.plist);
1371
1372                 BUG_ON(q->pi_state);
1373
1374                 spin_unlock(lock_ptr);
1375                 ret = 1;
1376         }
1377
1378         drop_futex_key_refs(&q->key);
1379         return ret;
1380 }
1381
1382 /*
1383  * PI futexes can not be requeued and must remove themself from the
1384  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1385  * and dropped here.
1386  */
1387 static void unqueue_me_pi(struct futex_q *q)
1388 {
1389         WARN_ON(plist_node_empty(&q->list));
1390         plist_del(&q->list, &q->list.plist);
1391
1392         BUG_ON(!q->pi_state);
1393         free_pi_state(q->pi_state);
1394         q->pi_state = NULL;
1395
1396         spin_unlock(q->lock_ptr);
1397
1398         drop_futex_key_refs(&q->key);
1399 }
1400
1401 /*
1402  * Fixup the pi_state owner with the new owner.
1403  *
1404  * Must be called with hash bucket lock held and mm->sem held for non
1405  * private futexes.
1406  */
1407 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1408                                 struct task_struct *newowner, int fshared)
1409 {
1410         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1411         struct futex_pi_state *pi_state = q->pi_state;
1412         struct task_struct *oldowner = pi_state->owner;
1413         u32 uval, curval, newval;
1414         int ret;
1415
1416         /* Owner died? */
1417         if (!pi_state->owner)
1418                 newtid |= FUTEX_OWNER_DIED;
1419
1420         /*
1421          * We are here either because we stole the rtmutex from the
1422          * pending owner or we are the pending owner which failed to
1423          * get the rtmutex. We have to replace the pending owner TID
1424          * in the user space variable. This must be atomic as we have
1425          * to preserve the owner died bit here.
1426          *
1427          * Note: We write the user space value _before_ changing the pi_state
1428          * because we can fault here. Imagine swapped out pages or a fork
1429          * that marked all the anonymous memory readonly for cow.
1430          *
1431          * Modifying pi_state _before_ the user space value would
1432          * leave the pi_state in an inconsistent state when we fault
1433          * here, because we need to drop the hash bucket lock to
1434          * handle the fault. This might be observed in the PID check
1435          * in lookup_pi_state.
1436          */
1437 retry:
1438         if (get_futex_value_locked(&uval, uaddr))
1439                 goto handle_fault;
1440
1441         while (1) {
1442                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1443
1444                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1445
1446                 if (curval == -EFAULT)
1447                         goto handle_fault;
1448                 if (curval == uval)
1449                         break;
1450                 uval = curval;
1451         }
1452
1453         /*
1454          * We fixed up user space. Now we need to fix the pi_state
1455          * itself.
1456          */
1457         if (pi_state->owner != NULL) {
1458                 spin_lock_irq(&pi_state->owner->pi_lock);
1459                 WARN_ON(list_empty(&pi_state->list));
1460                 list_del_init(&pi_state->list);
1461                 spin_unlock_irq(&pi_state->owner->pi_lock);
1462         }
1463
1464         pi_state->owner = newowner;
1465
1466         spin_lock_irq(&newowner->pi_lock);
1467         WARN_ON(!list_empty(&pi_state->list));
1468         list_add(&pi_state->list, &newowner->pi_state_list);
1469         spin_unlock_irq(&newowner->pi_lock);
1470         return 0;
1471
1472         /*
1473          * To handle the page fault we need to drop the hash bucket
1474          * lock here. That gives the other task (either the pending
1475          * owner itself or the task which stole the rtmutex) the
1476          * chance to try the fixup of the pi_state. So once we are
1477          * back from handling the fault we need to check the pi_state
1478          * after reacquiring the hash bucket lock and before trying to
1479          * do another fixup. When the fixup has been done already we
1480          * simply return.
1481          */
1482 handle_fault:
1483         spin_unlock(q->lock_ptr);
1484
1485         ret = get_user(uval, uaddr);
1486
1487         spin_lock(q->lock_ptr);
1488
1489         /*
1490          * Check if someone else fixed it for us:
1491          */
1492         if (pi_state->owner != oldowner)
1493                 return 0;
1494
1495         if (ret)
1496                 return ret;
1497
1498         goto retry;
1499 }
1500
1501 /*
1502  * In case we must use restart_block to restart a futex_wait,
1503  * we encode in the 'flags' shared capability
1504  */
1505 #define FLAGS_SHARED            0x01
1506 #define FLAGS_CLOCKRT           0x02
1507 #define FLAGS_HAS_TIMEOUT       0x04
1508
1509 static long futex_wait_restart(struct restart_block *restart);
1510 static long futex_lock_pi_restart(struct restart_block *restart);
1511
1512 /**
1513  * fixup_owner() - Post lock pi_state and corner case management
1514  * @uaddr:      user address of the futex
1515  * @fshared:    whether the futex is shared (1) or not (0)
1516  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1517  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1518  *
1519  * After attempting to lock an rt_mutex, this function is called to cleanup
1520  * the pi_state owner as well as handle race conditions that may allow us to
1521  * acquire the lock. Must be called with the hb lock held.
1522  *
1523  * Returns:
1524  *  1 - success, lock taken
1525  *  0 - success, lock not taken
1526  * <0 - on error (-EFAULT)
1527  */
1528 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1529                        int locked)
1530 {
1531         struct task_struct *owner;
1532         int ret = 0;
1533
1534         if (locked) {
1535                 /*
1536                  * Got the lock. We might not be the anticipated owner if we
1537                  * did a lock-steal - fix up the PI-state in that case:
1538                  */
1539                 if (q->pi_state->owner != current)
1540                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1541                 goto out;
1542         }
1543
1544         /*
1545          * Catch the rare case, where the lock was released when we were on the
1546          * way back before we locked the hash bucket.
1547          */
1548         if (q->pi_state->owner == current) {
1549                 /*
1550                  * Try to get the rt_mutex now. This might fail as some other
1551                  * task acquired the rt_mutex after we removed ourself from the
1552                  * rt_mutex waiters list.
1553                  */
1554                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1555                         locked = 1;
1556                         goto out;
1557                 }
1558
1559                 /*
1560                  * pi_state is incorrect, some other task did a lock steal and
1561                  * we returned due to timeout or signal without taking the
1562                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1563                  * locking, as the other task is now blocked on the hash bucket
1564                  * lock. Fix the state up.
1565                  */
1566                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1567                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1568                 goto out;
1569         }
1570
1571         /*
1572          * Paranoia check. If we did not take the lock, then we should not be
1573          * the owner, nor the pending owner, of the rt_mutex.
1574          */
1575         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1576                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1577                                 "pi-state %p\n", ret,
1578                                 q->pi_state->pi_mutex.owner,
1579                                 q->pi_state->owner);
1580
1581 out:
1582         return ret ? ret : locked;
1583 }
1584
1585 /**
1586  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1587  * @hb:         the futex hash bucket, must be locked by the caller
1588  * @q:          the futex_q to queue up on
1589  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1590  */
1591 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1592                                 struct hrtimer_sleeper *timeout)
1593 {
1594         queue_me(q, hb);
1595
1596         /*
1597          * There might have been scheduling since the queue_me(), as we
1598          * cannot hold a spinlock across the get_user() in case it
1599          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1600          * queueing ourselves into the futex hash. This code thus has to
1601          * rely on the futex_wake() code removing us from hash when it
1602          * wakes us up.
1603          */
1604         set_current_state(TASK_INTERRUPTIBLE);
1605
1606         /* Arm the timer */
1607         if (timeout) {
1608                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1609                 if (!hrtimer_active(&timeout->timer))
1610                         timeout->task = NULL;
1611         }
1612
1613         /*
1614          * !plist_node_empty() is safe here without any lock.
1615          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1616          */
1617         if (likely(!plist_node_empty(&q->list))) {
1618                 /*
1619                  * If the timer has already expired, current will already be
1620                  * flagged for rescheduling. Only call schedule if there
1621                  * is no timeout, or if it has yet to expire.
1622                  */
1623                 if (!timeout || timeout->task)
1624                         schedule();
1625         }
1626         __set_current_state(TASK_RUNNING);
1627 }
1628
1629 /**
1630  * futex_wait_setup() - Prepare to wait on a futex
1631  * @uaddr:      the futex userspace address
1632  * @val:        the expected value
1633  * @fshared:    whether the futex is shared (1) or not (0)
1634  * @q:          the associated futex_q
1635  * @hb:         storage for hash_bucket pointer to be returned to caller
1636  *
1637  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1638  * compare it with the expected value.  Handle atomic faults internally.
1639  * Return with the hb lock held and a q.key reference on success, and unlocked
1640  * with no q.key reference on failure.
1641  *
1642  * Returns:
1643  *  0 - uaddr contains val and hb has been locked
1644  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1645  */
1646 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1647                            struct futex_q *q, struct futex_hash_bucket **hb)
1648 {
1649         u32 uval;
1650         int ret;
1651
1652         /*
1653          * Access the page AFTER the hash-bucket is locked.
1654          * Order is important:
1655          *
1656          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1657          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1658          *
1659          * The basic logical guarantee of a futex is that it blocks ONLY
1660          * if cond(var) is known to be true at the time of blocking, for
1661          * any cond.  If we queued after testing *uaddr, that would open
1662          * a race condition where we could block indefinitely with
1663          * cond(var) false, which would violate the guarantee.
1664          *
1665          * A consequence is that futex_wait() can return zero and absorb
1666          * a wakeup when *uaddr != val on entry to the syscall.  This is
1667          * rare, but normal.
1668          */
1669 retry:
1670         q->key = FUTEX_KEY_INIT;
1671         ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1672         if (unlikely(ret != 0))
1673                 return ret;
1674
1675 retry_private:
1676         *hb = queue_lock(q);
1677
1678         ret = get_futex_value_locked(&uval, uaddr);
1679
1680         if (ret) {
1681                 queue_unlock(q, *hb);
1682
1683                 ret = get_user(uval, uaddr);
1684                 if (ret)
1685                         goto out;
1686
1687                 if (!fshared)
1688                         goto retry_private;
1689
1690                 put_futex_key(fshared, &q->key);
1691                 goto retry;
1692         }
1693
1694         if (uval != val) {
1695                 queue_unlock(q, *hb);
1696                 ret = -EWOULDBLOCK;
1697         }
1698
1699 out:
1700         if (ret)
1701                 put_futex_key(fshared, &q->key);
1702         return ret;
1703 }
1704
1705 static int futex_wait(u32 __user *uaddr, int fshared,
1706                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1707 {
1708         struct hrtimer_sleeper timeout, *to = NULL;
1709         struct restart_block *restart;
1710         struct futex_hash_bucket *hb;
1711         struct futex_q q;
1712         int ret;
1713
1714         if (!bitset)
1715                 return -EINVAL;
1716
1717         q.pi_state = NULL;
1718         q.bitset = bitset;
1719         q.rt_waiter = NULL;
1720
1721         if (abs_time) {
1722                 to = &timeout;
1723
1724                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1725                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1726                 hrtimer_init_sleeper(to, current);
1727                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1728                                              current->timer_slack_ns);
1729         }
1730
1731         /* Prepare to wait on uaddr. */
1732         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1733         if (ret)
1734                 goto out;
1735
1736         /* queue_me and wait for wakeup, timeout, or a signal. */
1737         futex_wait_queue_me(hb, &q, to);
1738
1739         /* If we were woken (and unqueued), we succeeded, whatever. */
1740         ret = 0;
1741         if (!unqueue_me(&q))
1742                 goto out_put_key;
1743         ret = -ETIMEDOUT;
1744         if (to && !to->task)
1745                 goto out_put_key;
1746
1747         /*
1748          * We expect signal_pending(current), but another thread may
1749          * have handled it for us already.
1750          */
1751         ret = -ERESTARTSYS;
1752         if (!abs_time)
1753                 goto out_put_key;
1754
1755         restart = &current_thread_info()->restart_block;
1756         restart->fn = futex_wait_restart;
1757         restart->futex.uaddr = (u32 *)uaddr;
1758         restart->futex.val = val;
1759         restart->futex.time = abs_time->tv64;
1760         restart->futex.bitset = bitset;
1761         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1762
1763         if (fshared)
1764                 restart->futex.flags |= FLAGS_SHARED;
1765         if (clockrt)
1766                 restart->futex.flags |= FLAGS_CLOCKRT;
1767
1768         ret = -ERESTART_RESTARTBLOCK;
1769
1770 out_put_key:
1771         put_futex_key(fshared, &q.key);
1772 out:
1773         if (to) {
1774                 hrtimer_cancel(&to->timer);
1775                 destroy_hrtimer_on_stack(&to->timer);
1776         }
1777         return ret;
1778 }
1779
1780
1781 static long futex_wait_restart(struct restart_block *restart)
1782 {
1783         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1784         int fshared = 0;
1785         ktime_t t, *tp = NULL;
1786
1787         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1788                 t.tv64 = restart->futex.time;
1789                 tp = &t;
1790         }
1791         restart->fn = do_no_restart_syscall;
1792         if (restart->futex.flags & FLAGS_SHARED)
1793                 fshared = 1;
1794         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1795                                 restart->futex.bitset,
1796                                 restart->futex.flags & FLAGS_CLOCKRT);
1797 }
1798
1799
1800 /*
1801  * Userspace tried a 0 -> TID atomic transition of the futex value
1802  * and failed. The kernel side here does the whole locking operation:
1803  * if there are waiters then it will block, it does PI, etc. (Due to
1804  * races the kernel might see a 0 value of the futex too.)
1805  */
1806 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1807                          int detect, ktime_t *time, int trylock)
1808 {
1809         struct hrtimer_sleeper timeout, *to = NULL;
1810         struct futex_hash_bucket *hb;
1811         u32 uval;
1812         struct futex_q q;
1813         int res, ret;
1814
1815         if (refill_pi_state_cache())
1816                 return -ENOMEM;
1817
1818         if (time) {
1819                 to = &timeout;
1820                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1821                                       HRTIMER_MODE_ABS);
1822                 hrtimer_init_sleeper(to, current);
1823                 hrtimer_set_expires(&to->timer, *time);
1824         }
1825
1826         q.pi_state = NULL;
1827         q.rt_waiter = NULL;
1828 retry:
1829         q.key = FUTEX_KEY_INIT;
1830         ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1831         if (unlikely(ret != 0))
1832                 goto out;
1833
1834 retry_private:
1835         hb = queue_lock(&q);
1836
1837         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1838         if (unlikely(ret)) {
1839                 switch (ret) {
1840                 case 1:
1841                         /* We got the lock. */
1842                         ret = 0;
1843                         goto out_unlock_put_key;
1844                 case -EFAULT:
1845                         goto uaddr_faulted;
1846                 case -EAGAIN:
1847                         /*
1848                          * Task is exiting and we just wait for the
1849                          * exit to complete.
1850                          */
1851                         queue_unlock(&q, hb);
1852                         put_futex_key(fshared, &q.key);
1853                         cond_resched();
1854                         goto retry;
1855                 default:
1856                         goto out_unlock_put_key;
1857                 }
1858         }
1859
1860         /*
1861          * Only actually queue now that the atomic ops are done:
1862          */
1863         queue_me(&q, hb);
1864
1865         WARN_ON(!q.pi_state);
1866         /*
1867          * Block on the PI mutex:
1868          */
1869         if (!trylock)
1870                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1871         else {
1872                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1873                 /* Fixup the trylock return value: */
1874                 ret = ret ? 0 : -EWOULDBLOCK;
1875         }
1876
1877         spin_lock(q.lock_ptr);
1878         /*
1879          * Fixup the pi_state owner and possibly acquire the lock if we
1880          * haven't already.
1881          */
1882         res = fixup_owner(uaddr, fshared, &q, !ret);
1883         /*
1884          * If fixup_owner() returned an error, proprogate that.  If it acquired
1885          * the lock, clear our -ETIMEDOUT or -EINTR.
1886          */
1887         if (res)
1888                 ret = (res < 0) ? res : 0;
1889
1890         /*
1891          * If fixup_owner() faulted and was unable to handle the fault, unlock
1892          * it and return the fault to userspace.
1893          */
1894         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1895                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1896
1897         /* Unqueue and drop the lock */
1898         unqueue_me_pi(&q);
1899
1900         goto out;
1901
1902 out_unlock_put_key:
1903         queue_unlock(&q, hb);
1904
1905 out_put_key:
1906         put_futex_key(fshared, &q.key);
1907 out:
1908         if (to)
1909                 destroy_hrtimer_on_stack(&to->timer);
1910         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1911
1912 uaddr_faulted:
1913         /*
1914          * We have to r/w  *(int __user *)uaddr, and we have to modify it
1915          * atomically.  Therefore, if we continue to fault after get_user()
1916          * below, we need to handle the fault ourselves, while still holding
1917          * the mmap_sem.  This can occur if the uaddr is under contention as
1918          * we have to drop the mmap_sem in order to call get_user().
1919          */
1920         queue_unlock(&q, hb);
1921
1922         ret = get_user(uval, uaddr);
1923         if (ret)
1924                 goto out_put_key;
1925
1926         if (!fshared)
1927                 goto retry_private;
1928
1929         put_futex_key(fshared, &q.key);
1930         goto retry;
1931 }
1932
1933 static long futex_lock_pi_restart(struct restart_block *restart)
1934 {
1935         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1936         ktime_t t, *tp = NULL;
1937         int fshared = restart->futex.flags & FLAGS_SHARED;
1938
1939         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1940                 t.tv64 = restart->futex.time;
1941                 tp = &t;
1942         }
1943         restart->fn = do_no_restart_syscall;
1944
1945         return (long)futex_lock_pi(uaddr, fshared, restart->futex.val, tp, 0);
1946 }
1947
1948 /*
1949  * Userspace attempted a TID -> 0 atomic transition, and failed.
1950  * This is the in-kernel slowpath: we look up the PI state (if any),
1951  * and do the rt-mutex unlock.
1952  */
1953 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
1954 {
1955         struct futex_hash_bucket *hb;
1956         struct futex_q *this, *next;
1957         u32 uval;
1958         struct plist_head *head;
1959         union futex_key key = FUTEX_KEY_INIT;
1960         int ret;
1961
1962 retry:
1963         if (get_user(uval, uaddr))
1964                 return -EFAULT;
1965         /*
1966          * We release only a lock we actually own:
1967          */
1968         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1969                 return -EPERM;
1970
1971         ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
1972         if (unlikely(ret != 0))
1973                 goto out;
1974
1975         hb = hash_futex(&key);
1976         spin_lock(&hb->lock);
1977
1978         /*
1979          * To avoid races, try to do the TID -> 0 atomic transition
1980          * again. If it succeeds then we can return without waking
1981          * anyone else up:
1982          */
1983         if (!(uval & FUTEX_OWNER_DIED))
1984                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1985
1986
1987         if (unlikely(uval == -EFAULT))
1988                 goto pi_faulted;
1989         /*
1990          * Rare case: we managed to release the lock atomically,
1991          * no need to wake anyone else up:
1992          */
1993         if (unlikely(uval == task_pid_vnr(current)))
1994                 goto out_unlock;
1995
1996         /*
1997          * Ok, other tasks may need to be woken up - check waiters
1998          * and do the wakeup if necessary:
1999          */
2000         head = &hb->chain;
2001
2002         plist_for_each_entry_safe(this, next, head, list) {
2003                 if (!match_futex (&this->key, &key))
2004                         continue;
2005                 ret = wake_futex_pi(uaddr, uval, this);
2006                 /*
2007                  * The atomic access to the futex value
2008                  * generated a pagefault, so retry the
2009                  * user-access and the wakeup:
2010                  */
2011                 if (ret == -EFAULT)
2012                         goto pi_faulted;
2013                 goto out_unlock;
2014         }
2015         /*
2016          * No waiters - kernel unlocks the futex:
2017          */
2018         if (!(uval & FUTEX_OWNER_DIED)) {
2019                 ret = unlock_futex_pi(uaddr, uval);
2020                 if (ret == -EFAULT)
2021                         goto pi_faulted;
2022         }
2023
2024 out_unlock:
2025         spin_unlock(&hb->lock);
2026         put_futex_key(fshared, &key);
2027
2028 out:
2029         return ret;
2030
2031 pi_faulted:
2032         /*
2033          * We have to r/w  *(int __user *)uaddr, and we have to modify it
2034          * atomically.  Therefore, if we continue to fault after get_user()
2035          * below, we need to handle the fault ourselves, while still holding
2036          * the mmap_sem.  This can occur if the uaddr is under contention as
2037          * we have to drop the mmap_sem in order to call get_user().
2038          */
2039         spin_unlock(&hb->lock);
2040         put_futex_key(fshared, &key);
2041
2042         ret = get_user(uval, uaddr);
2043         if (!ret)
2044                 goto retry;
2045
2046         return ret;
2047 }
2048
2049 /**
2050  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2051  * @hb:         the hash_bucket futex_q was original enqueued on
2052  * @q:          the futex_q woken while waiting to be requeued
2053  * @key2:       the futex_key of the requeue target futex
2054  * @timeout:    the timeout associated with the wait (NULL if none)
2055  *
2056  * Detect if the task was woken on the initial futex as opposed to the requeue
2057  * target futex.  If so, determine if it was a timeout or a signal that caused
2058  * the wakeup and return the appropriate error code to the caller.  Must be
2059  * called with the hb lock held.
2060  *
2061  * Returns
2062  *  0 - no early wakeup detected
2063  * <0 - -ETIMEDOUT or -ERESTARTSYS (FIXME: or ERESTARTNOINTR?)
2064  */
2065 static inline
2066 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2067                                    struct futex_q *q, union futex_key *key2,
2068                                    struct hrtimer_sleeper *timeout)
2069 {
2070         int ret = 0;
2071
2072         /*
2073          * With the hb lock held, we avoid races while we process the wakeup.
2074          * We only need to hold hb (and not hb2) to ensure atomicity as the
2075          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2076          * It can't be requeued from uaddr2 to something else since we don't
2077          * support a PI aware source futex for requeue.
2078          */
2079         if (!match_futex(&q->key, key2)) {
2080                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2081                 /*
2082                  * We were woken prior to requeue by a timeout or a signal.
2083                  * Unqueue the futex_q and determine which it was.
2084                  */
2085                 plist_del(&q->list, &q->list.plist);
2086                 drop_futex_key_refs(&q->key);
2087
2088                 if (timeout && !timeout->task)
2089                         ret = -ETIMEDOUT;
2090                 else {
2091                         /*
2092                          * We expect signal_pending(current), but another
2093                          * thread may have handled it for us already.
2094                          */
2095                         /* FIXME: ERESTARTSYS or ERESTARTNOINTR?  Do we care if
2096                          * the user specified SA_RESTART or not? */
2097                         ret = -ERESTARTSYS;
2098                 }
2099         }
2100         return ret;
2101 }
2102
2103 /**
2104  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2105  * @uaddr:      the futex we initialyl wait on (non-pi)
2106  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2107  *              the same type, no requeueing from private to shared, etc.
2108  * @val:        the expected value of uaddr
2109  * @abs_time:   absolute timeout
2110  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all.
2111  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2112  * @uaddr2:     the pi futex we will take prior to returning to user-space
2113  *
2114  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2115  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2116  * complete the acquisition of the rt_mutex prior to returning to userspace.
2117  * This ensures the rt_mutex maintains an owner when it has waiters; without
2118  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2119  * need to.
2120  *
2121  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2122  * via the following:
2123  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2124  * 2) wakeup on uaddr2 after a requeue and subsequent unlock
2125  * 3) signal (before or after requeue)
2126  * 4) timeout (before or after requeue)
2127  *
2128  * If 3, we setup a restart_block with futex_wait_requeue_pi() as the function.
2129  *
2130  * If 2, we may then block on trying to take the rt_mutex and return via:
2131  * 5) successful lock
2132  * 6) signal
2133  * 7) timeout
2134  * 8) other lock acquisition failure
2135  *
2136  * If 6, we setup a restart_block with futex_lock_pi() as the function.
2137  *
2138  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2139  *
2140  * Returns:
2141  *  0 - On success
2142  * <0 - On error
2143  */
2144 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2145                                  u32 val, ktime_t *abs_time, u32 bitset,
2146                                  int clockrt, u32 __user *uaddr2)
2147 {
2148         struct hrtimer_sleeper timeout, *to = NULL;
2149         struct rt_mutex_waiter rt_waiter;
2150         struct rt_mutex *pi_mutex = NULL;
2151         struct restart_block *restart;
2152         struct futex_hash_bucket *hb;
2153         union futex_key key2;
2154         struct futex_q q;
2155         int res, ret;
2156         u32 uval;
2157
2158         if (!bitset)
2159                 return -EINVAL;
2160
2161         if (abs_time) {
2162                 to = &timeout;
2163                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2164                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2165                 hrtimer_init_sleeper(to, current);
2166                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2167                                              current->timer_slack_ns);
2168         }
2169
2170         /*
2171          * The waiter is allocated on our stack, manipulated by the requeue
2172          * code while we sleep on uaddr.
2173          */
2174         debug_rt_mutex_init_waiter(&rt_waiter);
2175         rt_waiter.task = NULL;
2176
2177         q.pi_state = NULL;
2178         q.bitset = bitset;
2179         q.rt_waiter = &rt_waiter;
2180
2181         key2 = FUTEX_KEY_INIT;
2182         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2183         if (unlikely(ret != 0))
2184                 goto out;
2185
2186         /* Prepare to wait on uaddr. */
2187         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2188         if (ret) {
2189                 put_futex_key(fshared, &key2);
2190                 goto out;
2191         }
2192
2193         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2194         futex_wait_queue_me(hb, &q, to);
2195
2196         spin_lock(&hb->lock);
2197         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2198         spin_unlock(&hb->lock);
2199         if (ret)
2200                 goto out_put_keys;
2201
2202         /*
2203          * In order for us to be here, we know our q.key == key2, and since
2204          * we took the hb->lock above, we also know that futex_requeue() has
2205          * completed and we no longer have to concern ourselves with a wakeup
2206          * race with the atomic proxy lock acquition by the requeue code.
2207          */
2208
2209         /* Check if the requeue code acquired the second futex for us. */
2210         if (!q.rt_waiter) {
2211                 /*
2212                  * Got the lock. We might not be the anticipated owner if we
2213                  * did a lock-steal - fix up the PI-state in that case.
2214                  */
2215                 if (q.pi_state && (q.pi_state->owner != current)) {
2216                         spin_lock(q.lock_ptr);
2217                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2218                                                    fshared);
2219                         spin_unlock(q.lock_ptr);
2220                 }
2221         } else {
2222                 /*
2223                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2224                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2225                  * the pi_state.
2226                  */
2227                 WARN_ON(!&q.pi_state);
2228                 pi_mutex = &q.pi_state->pi_mutex;
2229                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2230                 debug_rt_mutex_free_waiter(&rt_waiter);
2231
2232                 spin_lock(q.lock_ptr);
2233                 /*
2234                  * Fixup the pi_state owner and possibly acquire the lock if we
2235                  * haven't already.
2236                  */
2237                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2238                 /*
2239                  * If fixup_owner() returned an error, proprogate that.  If it
2240                  * acquired the lock, clear our -ETIMEDOUT or -EINTR.
2241                  */
2242                 if (res)
2243                         ret = (res < 0) ? res : 0;
2244
2245                 /* Unqueue and drop the lock. */
2246                 unqueue_me_pi(&q);
2247         }
2248
2249         /*
2250          * If fixup_pi_state_owner() faulted and was unable to handle the
2251          * fault, unlock the rt_mutex and return the fault to userspace.
2252          */
2253         if (ret == -EFAULT) {
2254                 if (rt_mutex_owner(pi_mutex) == current)
2255                         rt_mutex_unlock(pi_mutex);
2256         } else if (ret == -EINTR) {
2257                 ret = -EFAULT;
2258                 if (get_user(uval, uaddr2))
2259                         goto out_put_keys;
2260
2261                 /*
2262                  * We've already been requeued, so restart by calling
2263                  * futex_lock_pi() directly, rather then returning to this
2264                  * function.
2265                  */
2266                 ret = -ERESTART_RESTARTBLOCK;
2267                 restart = &current_thread_info()->restart_block;
2268                 restart->fn = futex_lock_pi_restart;
2269                 restart->futex.uaddr = (u32 *)uaddr2;
2270                 restart->futex.val = uval;
2271                 restart->futex.flags = 0;
2272                 if (abs_time) {
2273                         restart->futex.flags |= FLAGS_HAS_TIMEOUT;
2274                         restart->futex.time = abs_time->tv64;
2275                 }
2276
2277                 if (fshared)
2278                         restart->futex.flags |= FLAGS_SHARED;
2279                 if (clockrt)
2280                         restart->futex.flags |= FLAGS_CLOCKRT;
2281         }
2282
2283 out_put_keys:
2284         put_futex_key(fshared, &q.key);
2285         put_futex_key(fshared, &key2);
2286
2287 out:
2288         if (to) {
2289                 hrtimer_cancel(&to->timer);
2290                 destroy_hrtimer_on_stack(&to->timer);
2291         }
2292         return ret;
2293 }
2294
2295 /*
2296  * Support for robust futexes: the kernel cleans up held futexes at
2297  * thread exit time.
2298  *
2299  * Implementation: user-space maintains a per-thread list of locks it
2300  * is holding. Upon do_exit(), the kernel carefully walks this list,
2301  * and marks all locks that are owned by this thread with the
2302  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2303  * always manipulated with the lock held, so the list is private and
2304  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2305  * field, to allow the kernel to clean up if the thread dies after
2306  * acquiring the lock, but just before it could have added itself to
2307  * the list. There can only be one such pending lock.
2308  */
2309
2310 /**
2311  * sys_set_robust_list - set the robust-futex list head of a task
2312  * @head: pointer to the list-head
2313  * @len: length of the list-head, as userspace expects
2314  */
2315 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2316                 size_t, len)
2317 {
2318         if (!futex_cmpxchg_enabled)
2319                 return -ENOSYS;
2320         /*
2321          * The kernel knows only one size for now:
2322          */
2323         if (unlikely(len != sizeof(*head)))
2324                 return -EINVAL;
2325
2326         current->robust_list = head;
2327
2328         return 0;
2329 }
2330
2331 /**
2332  * sys_get_robust_list - get the robust-futex list head of a task
2333  * @pid: pid of the process [zero for current task]
2334  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2335  * @len_ptr: pointer to a length field, the kernel fills in the header size
2336  */
2337 SYSCALL_DEFINE3(get_robust_list, int, pid,
2338                 struct robust_list_head __user * __user *, head_ptr,
2339                 size_t __user *, len_ptr)
2340 {
2341         struct robust_list_head __user *head;
2342         unsigned long ret;
2343         const struct cred *cred = current_cred(), *pcred;
2344
2345         if (!futex_cmpxchg_enabled)
2346                 return -ENOSYS;
2347
2348         if (!pid)
2349                 head = current->robust_list;
2350         else {
2351                 struct task_struct *p;
2352
2353                 ret = -ESRCH;
2354                 rcu_read_lock();
2355                 p = find_task_by_vpid(pid);
2356                 if (!p)
2357                         goto err_unlock;
2358                 ret = -EPERM;
2359                 pcred = __task_cred(p);
2360                 if (cred->euid != pcred->euid &&
2361                     cred->euid != pcred->uid &&
2362                     !capable(CAP_SYS_PTRACE))
2363                         goto err_unlock;
2364                 head = p->robust_list;
2365                 rcu_read_unlock();
2366         }
2367
2368         if (put_user(sizeof(*head), len_ptr))
2369                 return -EFAULT;
2370         return put_user(head, head_ptr);
2371
2372 err_unlock:
2373         rcu_read_unlock();
2374
2375         return ret;
2376 }
2377
2378 /*
2379  * Process a futex-list entry, check whether it's owned by the
2380  * dying task, and do notification if so:
2381  */
2382 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2383 {
2384         u32 uval, nval, mval;
2385
2386 retry:
2387         if (get_user(uval, uaddr))
2388                 return -1;
2389
2390         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2391                 /*
2392                  * Ok, this dying thread is truly holding a futex
2393                  * of interest. Set the OWNER_DIED bit atomically
2394                  * via cmpxchg, and if the value had FUTEX_WAITERS
2395                  * set, wake up a waiter (if any). (We have to do a
2396                  * futex_wake() even if OWNER_DIED is already set -
2397                  * to handle the rare but possible case of recursive
2398                  * thread-death.) The rest of the cleanup is done in
2399                  * userspace.
2400                  */
2401                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2402                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2403
2404                 if (nval == -EFAULT)
2405                         return -1;
2406
2407                 if (nval != uval)
2408                         goto retry;
2409
2410                 /*
2411                  * Wake robust non-PI futexes here. The wakeup of
2412                  * PI futexes happens in exit_pi_state():
2413                  */
2414                 if (!pi && (uval & FUTEX_WAITERS))
2415                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2416         }
2417         return 0;
2418 }
2419
2420 /*
2421  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2422  */
2423 static inline int fetch_robust_entry(struct robust_list __user **entry,
2424                                      struct robust_list __user * __user *head,
2425                                      int *pi)
2426 {
2427         unsigned long uentry;
2428
2429         if (get_user(uentry, (unsigned long __user *)head))
2430                 return -EFAULT;
2431
2432         *entry = (void __user *)(uentry & ~1UL);
2433         *pi = uentry & 1;
2434
2435         return 0;
2436 }
2437
2438 /*
2439  * Walk curr->robust_list (very carefully, it's a userspace list!)
2440  * and mark any locks found there dead, and notify any waiters.
2441  *
2442  * We silently return on any sign of list-walking problem.
2443  */
2444 void exit_robust_list(struct task_struct *curr)
2445 {
2446         struct robust_list_head __user *head = curr->robust_list;
2447         struct robust_list __user *entry, *next_entry, *pending;
2448         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2449         unsigned long futex_offset;
2450         int rc;
2451
2452         if (!futex_cmpxchg_enabled)
2453                 return;
2454
2455         /*
2456          * Fetch the list head (which was registered earlier, via
2457          * sys_set_robust_list()):
2458          */
2459         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2460                 return;
2461         /*
2462          * Fetch the relative futex offset:
2463          */
2464         if (get_user(futex_offset, &head->futex_offset))
2465                 return;
2466         /*
2467          * Fetch any possibly pending lock-add first, and handle it
2468          * if it exists:
2469          */
2470         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2471                 return;
2472
2473         next_entry = NULL;      /* avoid warning with gcc */
2474         while (entry != &head->list) {
2475                 /*
2476                  * Fetch the next entry in the list before calling
2477                  * handle_futex_death:
2478                  */
2479                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2480                 /*
2481                  * A pending lock might already be on the list, so
2482                  * don't process it twice:
2483                  */
2484                 if (entry != pending)
2485                         if (handle_futex_death((void __user *)entry + futex_offset,
2486                                                 curr, pi))
2487                                 return;
2488                 if (rc)
2489                         return;
2490                 entry = next_entry;
2491                 pi = next_pi;
2492                 /*
2493                  * Avoid excessively long or circular lists:
2494                  */
2495                 if (!--limit)
2496                         break;
2497
2498                 cond_resched();
2499         }
2500
2501         if (pending)
2502                 handle_futex_death((void __user *)pending + futex_offset,
2503                                    curr, pip);
2504 }
2505
2506 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2507                 u32 __user *uaddr2, u32 val2, u32 val3)
2508 {
2509         int clockrt, ret = -ENOSYS;
2510         int cmd = op & FUTEX_CMD_MASK;
2511         int fshared = 0;
2512
2513         if (!(op & FUTEX_PRIVATE_FLAG))
2514                 fshared = 1;
2515
2516         clockrt = op & FUTEX_CLOCK_REALTIME;
2517         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2518                 return -ENOSYS;
2519
2520         switch (cmd) {
2521         case FUTEX_WAIT:
2522                 val3 = FUTEX_BITSET_MATCH_ANY;
2523         case FUTEX_WAIT_BITSET:
2524                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2525                 break;
2526         case FUTEX_WAKE:
2527                 val3 = FUTEX_BITSET_MATCH_ANY;
2528         case FUTEX_WAKE_BITSET:
2529                 ret = futex_wake(uaddr, fshared, val, val3);
2530                 break;
2531         case FUTEX_REQUEUE:
2532                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2533                 break;
2534         case FUTEX_CMP_REQUEUE:
2535                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2536                                     0);
2537                 break;
2538         case FUTEX_WAKE_OP:
2539                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2540                 break;
2541         case FUTEX_LOCK_PI:
2542                 if (futex_cmpxchg_enabled)
2543                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2544                 break;
2545         case FUTEX_UNLOCK_PI:
2546                 if (futex_cmpxchg_enabled)
2547                         ret = futex_unlock_pi(uaddr, fshared);
2548                 break;
2549         case FUTEX_TRYLOCK_PI:
2550                 if (futex_cmpxchg_enabled)
2551                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2552                 break;
2553         case FUTEX_WAIT_REQUEUE_PI:
2554                 val3 = FUTEX_BITSET_MATCH_ANY;
2555                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2556                                             clockrt, uaddr2);
2557                 break;
2558         case FUTEX_CMP_REQUEUE_PI:
2559                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2560                                     1);
2561                 break;
2562         default:
2563                 ret = -ENOSYS;
2564         }
2565         return ret;
2566 }
2567
2568
2569 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2570                 struct timespec __user *, utime, u32 __user *, uaddr2,
2571                 u32, val3)
2572 {
2573         struct timespec ts;
2574         ktime_t t, *tp = NULL;
2575         u32 val2 = 0;
2576         int cmd = op & FUTEX_CMD_MASK;
2577
2578         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2579                       cmd == FUTEX_WAIT_BITSET ||
2580                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2581                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2582                         return -EFAULT;
2583                 if (!timespec_valid(&ts))
2584                         return -EINVAL;
2585
2586                 t = timespec_to_ktime(ts);
2587                 if (cmd == FUTEX_WAIT)
2588                         t = ktime_add_safe(ktime_get(), t);
2589                 tp = &t;
2590         }
2591         /*
2592          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2593          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2594          */
2595         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2596             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2597                 val2 = (u32) (unsigned long) utime;
2598
2599         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2600 }
2601
2602 static int __init futex_init(void)
2603 {
2604         u32 curval;
2605         int i;
2606
2607         /*
2608          * This will fail and we want it. Some arch implementations do
2609          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2610          * functionality. We want to know that before we call in any
2611          * of the complex code paths. Also we want to prevent
2612          * registration of robust lists in that case. NULL is
2613          * guaranteed to fault and we get -EFAULT on functional
2614          * implementation, the non functional ones will return
2615          * -ENOSYS.
2616          */
2617         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2618         if (curval == -EFAULT)
2619                 futex_cmpxchg_enabled = 1;
2620
2621         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2622                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2623                 spin_lock_init(&futex_queues[i].lock);
2624         }
2625
2626         return 0;
2627 }
2628 __initcall(futex_init);