futex: Check for NULL keys in match_futex
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Priority Inheritance state:
73  */
74 struct futex_pi_state {
75         /*
76          * list of 'owned' pi_state instances - these have to be
77          * cleaned up in do_exit() if the task exits prematurely:
78          */
79         struct list_head list;
80
81         /*
82          * The PI object:
83          */
84         struct rt_mutex pi_mutex;
85
86         struct task_struct *owner;
87         atomic_t refcount;
88
89         union futex_key key;
90 };
91
92 /**
93  * struct futex_q - The hashed futex queue entry, one per waiting task
94  * @task:               the task waiting on the futex
95  * @lock_ptr:           the hash bucket lock
96  * @key:                the key the futex is hashed on
97  * @pi_state:           optional priority inheritance state
98  * @rt_waiter:          rt_waiter storage for use with requeue_pi
99  * @requeue_pi_key:     the requeue_pi target futex key
100  * @bitset:             bitset for the optional bitmasked wakeup
101  *
102  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103  * we can wake only the relevant ones (hashed queues may be shared).
104  *
105  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107  * The order of wakup is always to make the first condition true, then
108  * the second.
109  *
110  * PI futexes are typically woken before they are removed from the hash list via
111  * the rt_mutex code. See unqueue_me_pi().
112  */
113 struct futex_q {
114         struct plist_node list;
115
116         struct task_struct *task;
117         spinlock_t *lock_ptr;
118         union futex_key key;
119         struct futex_pi_state *pi_state;
120         struct rt_mutex_waiter *rt_waiter;
121         union futex_key *requeue_pi_key;
122         u32 bitset;
123 };
124
125 /*
126  * Hash buckets are shared by all the futex_keys that hash to the same
127  * location.  Each key may have multiple futex_q structures, one for each task
128  * waiting on a futex.
129  */
130 struct futex_hash_bucket {
131         spinlock_t lock;
132         struct plist_head chain;
133 };
134
135 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
136
137 /*
138  * We hash on the keys returned from get_futex_key (see below).
139  */
140 static struct futex_hash_bucket *hash_futex(union futex_key *key)
141 {
142         u32 hash = jhash2((u32*)&key->both.word,
143                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
144                           key->both.offset);
145         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146 }
147
148 /*
149  * Return 1 if two futex_keys are equal, 0 otherwise.
150  */
151 static inline int match_futex(union futex_key *key1, union futex_key *key2)
152 {
153         return (key1 && key2
154                 && key1->both.word == key2->both.word
155                 && key1->both.ptr == key2->both.ptr
156                 && key1->both.offset == key2->both.offset);
157 }
158
159 /*
160  * Take a reference to the resource addressed by a key.
161  * Can be called while holding spinlocks.
162  *
163  */
164 static void get_futex_key_refs(union futex_key *key)
165 {
166         if (!key->both.ptr)
167                 return;
168
169         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170         case FUT_OFF_INODE:
171                 atomic_inc(&key->shared.inode->i_count);
172                 break;
173         case FUT_OFF_MMSHARED:
174                 atomic_inc(&key->private.mm->mm_count);
175                 break;
176         }
177 }
178
179 /*
180  * Drop a reference to the resource addressed by a key.
181  * The hash bucket spinlock must not be held.
182  */
183 static void drop_futex_key_refs(union futex_key *key)
184 {
185         if (!key->both.ptr) {
186                 /* If we're here then we tried to put a key we failed to get */
187                 WARN_ON_ONCE(1);
188                 return;
189         }
190
191         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192         case FUT_OFF_INODE:
193                 iput(key->shared.inode);
194                 break;
195         case FUT_OFF_MMSHARED:
196                 mmdrop(key->private.mm);
197                 break;
198         }
199 }
200
201 /**
202  * get_futex_key() - Get parameters which are the keys for a futex
203  * @uaddr:      virtual address of the futex
204  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205  * @key:        address where result is stored.
206  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
207  *              VERIFY_WRITE)
208  *
209  * Returns a negative error code or 0
210  * The key words are stored in *key on success.
211  *
212  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
213  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
214  * We can usually work out the index without swapping in the page.
215  *
216  * lock_page() might sleep, the caller should not hold a spinlock.
217  */
218 static int
219 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
220 {
221         unsigned long address = (unsigned long)uaddr;
222         struct mm_struct *mm = current->mm;
223         struct page *page;
224         int err;
225
226         /*
227          * The futex address must be "naturally" aligned.
228          */
229         key->both.offset = address % PAGE_SIZE;
230         if (unlikely((address % sizeof(u32)) != 0))
231                 return -EINVAL;
232         address -= key->both.offset;
233
234         /*
235          * PROCESS_PRIVATE futexes are fast.
236          * As the mm cannot disappear under us and the 'key' only needs
237          * virtual address, we dont even have to find the underlying vma.
238          * Note : We do have to check 'uaddr' is a valid user address,
239          *        but access_ok() should be faster than find_vma()
240          */
241         if (!fshared) {
242                 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
243                         return -EFAULT;
244                 key->private.mm = mm;
245                 key->private.address = address;
246                 get_futex_key_refs(key);
247                 return 0;
248         }
249
250 again:
251         err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
252         if (err < 0)
253                 return err;
254
255         page = compound_head(page);
256         lock_page(page);
257         if (!page->mapping) {
258                 unlock_page(page);
259                 put_page(page);
260                 goto again;
261         }
262
263         /*
264          * Private mappings are handled in a simple way.
265          *
266          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
267          * it's a read-only handle, it's expected that futexes attach to
268          * the object not the particular process.
269          */
270         if (PageAnon(page)) {
271                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
272                 key->private.mm = mm;
273                 key->private.address = address;
274         } else {
275                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
276                 key->shared.inode = page->mapping->host;
277                 key->shared.pgoff = page->index;
278         }
279
280         get_futex_key_refs(key);
281
282         unlock_page(page);
283         put_page(page);
284         return 0;
285 }
286
287 static inline
288 void put_futex_key(int fshared, union futex_key *key)
289 {
290         drop_futex_key_refs(key);
291 }
292
293 /**
294  * fault_in_user_writeable() - Fault in user address and verify RW access
295  * @uaddr:      pointer to faulting user space address
296  *
297  * Slow path to fixup the fault we just took in the atomic write
298  * access to @uaddr.
299  *
300  * We have no generic implementation of a non destructive write to the
301  * user address. We know that we faulted in the atomic pagefault
302  * disabled section so we can as well avoid the #PF overhead by
303  * calling get_user_pages() right away.
304  */
305 static int fault_in_user_writeable(u32 __user *uaddr)
306 {
307         int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
308                                  1, 1, 0, NULL, NULL);
309         return ret < 0 ? ret : 0;
310 }
311
312 /**
313  * futex_top_waiter() - Return the highest priority waiter on a futex
314  * @hb:         the hash bucket the futex_q's reside in
315  * @key:        the futex key (to distinguish it from other futex futex_q's)
316  *
317  * Must be called with the hb lock held.
318  */
319 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
320                                         union futex_key *key)
321 {
322         struct futex_q *this;
323
324         plist_for_each_entry(this, &hb->chain, list) {
325                 if (match_futex(&this->key, key))
326                         return this;
327         }
328         return NULL;
329 }
330
331 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
332 {
333         u32 curval;
334
335         pagefault_disable();
336         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
337         pagefault_enable();
338
339         return curval;
340 }
341
342 static int get_futex_value_locked(u32 *dest, u32 __user *from)
343 {
344         int ret;
345
346         pagefault_disable();
347         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
348         pagefault_enable();
349
350         return ret ? -EFAULT : 0;
351 }
352
353
354 /*
355  * PI code:
356  */
357 static int refill_pi_state_cache(void)
358 {
359         struct futex_pi_state *pi_state;
360
361         if (likely(current->pi_state_cache))
362                 return 0;
363
364         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
365
366         if (!pi_state)
367                 return -ENOMEM;
368
369         INIT_LIST_HEAD(&pi_state->list);
370         /* pi_mutex gets initialized later */
371         pi_state->owner = NULL;
372         atomic_set(&pi_state->refcount, 1);
373         pi_state->key = FUTEX_KEY_INIT;
374
375         current->pi_state_cache = pi_state;
376
377         return 0;
378 }
379
380 static struct futex_pi_state * alloc_pi_state(void)
381 {
382         struct futex_pi_state *pi_state = current->pi_state_cache;
383
384         WARN_ON(!pi_state);
385         current->pi_state_cache = NULL;
386
387         return pi_state;
388 }
389
390 static void free_pi_state(struct futex_pi_state *pi_state)
391 {
392         if (!atomic_dec_and_test(&pi_state->refcount))
393                 return;
394
395         /*
396          * If pi_state->owner is NULL, the owner is most probably dying
397          * and has cleaned up the pi_state already
398          */
399         if (pi_state->owner) {
400                 spin_lock_irq(&pi_state->owner->pi_lock);
401                 list_del_init(&pi_state->list);
402                 spin_unlock_irq(&pi_state->owner->pi_lock);
403
404                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
405         }
406
407         if (current->pi_state_cache)
408                 kfree(pi_state);
409         else {
410                 /*
411                  * pi_state->list is already empty.
412                  * clear pi_state->owner.
413                  * refcount is at 0 - put it back to 1.
414                  */
415                 pi_state->owner = NULL;
416                 atomic_set(&pi_state->refcount, 1);
417                 current->pi_state_cache = pi_state;
418         }
419 }
420
421 /*
422  * Look up the task based on what TID userspace gave us.
423  * We dont trust it.
424  */
425 static struct task_struct * futex_find_get_task(pid_t pid)
426 {
427         struct task_struct *p;
428         const struct cred *cred = current_cred(), *pcred;
429
430         rcu_read_lock();
431         p = find_task_by_vpid(pid);
432         if (!p) {
433                 p = ERR_PTR(-ESRCH);
434         } else {
435                 pcred = __task_cred(p);
436                 if (cred->euid != pcred->euid &&
437                     cred->euid != pcred->uid)
438                         p = ERR_PTR(-ESRCH);
439                 else
440                         get_task_struct(p);
441         }
442
443         rcu_read_unlock();
444
445         return p;
446 }
447
448 /*
449  * This task is holding PI mutexes at exit time => bad.
450  * Kernel cleans up PI-state, but userspace is likely hosed.
451  * (Robust-futex cleanup is separate and might save the day for userspace.)
452  */
453 void exit_pi_state_list(struct task_struct *curr)
454 {
455         struct list_head *next, *head = &curr->pi_state_list;
456         struct futex_pi_state *pi_state;
457         struct futex_hash_bucket *hb;
458         union futex_key key = FUTEX_KEY_INIT;
459
460         if (!futex_cmpxchg_enabled)
461                 return;
462         /*
463          * We are a ZOMBIE and nobody can enqueue itself on
464          * pi_state_list anymore, but we have to be careful
465          * versus waiters unqueueing themselves:
466          */
467         spin_lock_irq(&curr->pi_lock);
468         while (!list_empty(head)) {
469
470                 next = head->next;
471                 pi_state = list_entry(next, struct futex_pi_state, list);
472                 key = pi_state->key;
473                 hb = hash_futex(&key);
474                 spin_unlock_irq(&curr->pi_lock);
475
476                 spin_lock(&hb->lock);
477
478                 spin_lock_irq(&curr->pi_lock);
479                 /*
480                  * We dropped the pi-lock, so re-check whether this
481                  * task still owns the PI-state:
482                  */
483                 if (head->next != next) {
484                         spin_unlock(&hb->lock);
485                         continue;
486                 }
487
488                 WARN_ON(pi_state->owner != curr);
489                 WARN_ON(list_empty(&pi_state->list));
490                 list_del_init(&pi_state->list);
491                 pi_state->owner = NULL;
492                 spin_unlock_irq(&curr->pi_lock);
493
494                 rt_mutex_unlock(&pi_state->pi_mutex);
495
496                 spin_unlock(&hb->lock);
497
498                 spin_lock_irq(&curr->pi_lock);
499         }
500         spin_unlock_irq(&curr->pi_lock);
501 }
502
503 static int
504 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
505                 union futex_key *key, struct futex_pi_state **ps)
506 {
507         struct futex_pi_state *pi_state = NULL;
508         struct futex_q *this, *next;
509         struct plist_head *head;
510         struct task_struct *p;
511         pid_t pid = uval & FUTEX_TID_MASK;
512
513         head = &hb->chain;
514
515         plist_for_each_entry_safe(this, next, head, list) {
516                 if (match_futex(&this->key, key)) {
517                         /*
518                          * Another waiter already exists - bump up
519                          * the refcount and return its pi_state:
520                          */
521                         pi_state = this->pi_state;
522                         /*
523                          * Userspace might have messed up non PI and PI futexes
524                          */
525                         if (unlikely(!pi_state))
526                                 return -EINVAL;
527
528                         WARN_ON(!atomic_read(&pi_state->refcount));
529                         WARN_ON(pid && pi_state->owner &&
530                                 pi_state->owner->pid != pid);
531
532                         atomic_inc(&pi_state->refcount);
533                         *ps = pi_state;
534
535                         return 0;
536                 }
537         }
538
539         /*
540          * We are the first waiter - try to look up the real owner and attach
541          * the new pi_state to it, but bail out when TID = 0
542          */
543         if (!pid)
544                 return -ESRCH;
545         p = futex_find_get_task(pid);
546         if (IS_ERR(p))
547                 return PTR_ERR(p);
548
549         /*
550          * We need to look at the task state flags to figure out,
551          * whether the task is exiting. To protect against the do_exit
552          * change of the task flags, we do this protected by
553          * p->pi_lock:
554          */
555         spin_lock_irq(&p->pi_lock);
556         if (unlikely(p->flags & PF_EXITING)) {
557                 /*
558                  * The task is on the way out. When PF_EXITPIDONE is
559                  * set, we know that the task has finished the
560                  * cleanup:
561                  */
562                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
563
564                 spin_unlock_irq(&p->pi_lock);
565                 put_task_struct(p);
566                 return ret;
567         }
568
569         pi_state = alloc_pi_state();
570
571         /*
572          * Initialize the pi_mutex in locked state and make 'p'
573          * the owner of it:
574          */
575         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
576
577         /* Store the key for possible exit cleanups: */
578         pi_state->key = *key;
579
580         WARN_ON(!list_empty(&pi_state->list));
581         list_add(&pi_state->list, &p->pi_state_list);
582         pi_state->owner = p;
583         spin_unlock_irq(&p->pi_lock);
584
585         put_task_struct(p);
586
587         *ps = pi_state;
588
589         return 0;
590 }
591
592 /**
593  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
594  * @uaddr:              the pi futex user address
595  * @hb:                 the pi futex hash bucket
596  * @key:                the futex key associated with uaddr and hb
597  * @ps:                 the pi_state pointer where we store the result of the
598  *                      lookup
599  * @task:               the task to perform the atomic lock work for.  This will
600  *                      be "current" except in the case of requeue pi.
601  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
602  *
603  * Returns:
604  *  0 - ready to wait
605  *  1 - acquired the lock
606  * <0 - error
607  *
608  * The hb->lock and futex_key refs shall be held by the caller.
609  */
610 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
611                                 union futex_key *key,
612                                 struct futex_pi_state **ps,
613                                 struct task_struct *task, int set_waiters)
614 {
615         int lock_taken, ret, ownerdied = 0;
616         u32 uval, newval, curval;
617
618 retry:
619         ret = lock_taken = 0;
620
621         /*
622          * To avoid races, we attempt to take the lock here again
623          * (by doing a 0 -> TID atomic cmpxchg), while holding all
624          * the locks. It will most likely not succeed.
625          */
626         newval = task_pid_vnr(task);
627         if (set_waiters)
628                 newval |= FUTEX_WAITERS;
629
630         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
631
632         if (unlikely(curval == -EFAULT))
633                 return -EFAULT;
634
635         /*
636          * Detect deadlocks.
637          */
638         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
639                 return -EDEADLK;
640
641         /*
642          * Surprise - we got the lock. Just return to userspace:
643          */
644         if (unlikely(!curval))
645                 return 1;
646
647         uval = curval;
648
649         /*
650          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
651          * to wake at the next unlock.
652          */
653         newval = curval | FUTEX_WAITERS;
654
655         /*
656          * There are two cases, where a futex might have no owner (the
657          * owner TID is 0): OWNER_DIED. We take over the futex in this
658          * case. We also do an unconditional take over, when the owner
659          * of the futex died.
660          *
661          * This is safe as we are protected by the hash bucket lock !
662          */
663         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
664                 /* Keep the OWNER_DIED bit */
665                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
666                 ownerdied = 0;
667                 lock_taken = 1;
668         }
669
670         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
671
672         if (unlikely(curval == -EFAULT))
673                 return -EFAULT;
674         if (unlikely(curval != uval))
675                 goto retry;
676
677         /*
678          * We took the lock due to owner died take over.
679          */
680         if (unlikely(lock_taken))
681                 return 1;
682
683         /*
684          * We dont have the lock. Look up the PI state (or create it if
685          * we are the first waiter):
686          */
687         ret = lookup_pi_state(uval, hb, key, ps);
688
689         if (unlikely(ret)) {
690                 switch (ret) {
691                 case -ESRCH:
692                         /*
693                          * No owner found for this futex. Check if the
694                          * OWNER_DIED bit is set to figure out whether
695                          * this is a robust futex or not.
696                          */
697                         if (get_futex_value_locked(&curval, uaddr))
698                                 return -EFAULT;
699
700                         /*
701                          * We simply start over in case of a robust
702                          * futex. The code above will take the futex
703                          * and return happy.
704                          */
705                         if (curval & FUTEX_OWNER_DIED) {
706                                 ownerdied = 1;
707                                 goto retry;
708                         }
709                 default:
710                         break;
711                 }
712         }
713
714         return ret;
715 }
716
717 /*
718  * The hash bucket lock must be held when this is called.
719  * Afterwards, the futex_q must not be accessed.
720  */
721 static void wake_futex(struct futex_q *q)
722 {
723         struct task_struct *p = q->task;
724
725         /*
726          * We set q->lock_ptr = NULL _before_ we wake up the task. If
727          * a non futex wake up happens on another CPU then the task
728          * might exit and p would dereference a non existing task
729          * struct. Prevent this by holding a reference on p across the
730          * wake up.
731          */
732         get_task_struct(p);
733
734         plist_del(&q->list, &q->list.plist);
735         /*
736          * The waiting task can free the futex_q as soon as
737          * q->lock_ptr = NULL is written, without taking any locks. A
738          * memory barrier is required here to prevent the following
739          * store to lock_ptr from getting ahead of the plist_del.
740          */
741         smp_wmb();
742         q->lock_ptr = NULL;
743
744         wake_up_state(p, TASK_NORMAL);
745         put_task_struct(p);
746 }
747
748 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
749 {
750         struct task_struct *new_owner;
751         struct futex_pi_state *pi_state = this->pi_state;
752         u32 curval, newval;
753
754         if (!pi_state)
755                 return -EINVAL;
756
757         spin_lock(&pi_state->pi_mutex.wait_lock);
758         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
759
760         /*
761          * This happens when we have stolen the lock and the original
762          * pending owner did not enqueue itself back on the rt_mutex.
763          * Thats not a tragedy. We know that way, that a lock waiter
764          * is on the fly. We make the futex_q waiter the pending owner.
765          */
766         if (!new_owner)
767                 new_owner = this->task;
768
769         /*
770          * We pass it to the next owner. (The WAITERS bit is always
771          * kept enabled while there is PI state around. We must also
772          * preserve the owner died bit.)
773          */
774         if (!(uval & FUTEX_OWNER_DIED)) {
775                 int ret = 0;
776
777                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
778
779                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
780
781                 if (curval == -EFAULT)
782                         ret = -EFAULT;
783                 else if (curval != uval)
784                         ret = -EINVAL;
785                 if (ret) {
786                         spin_unlock(&pi_state->pi_mutex.wait_lock);
787                         return ret;
788                 }
789         }
790
791         spin_lock_irq(&pi_state->owner->pi_lock);
792         WARN_ON(list_empty(&pi_state->list));
793         list_del_init(&pi_state->list);
794         spin_unlock_irq(&pi_state->owner->pi_lock);
795
796         spin_lock_irq(&new_owner->pi_lock);
797         WARN_ON(!list_empty(&pi_state->list));
798         list_add(&pi_state->list, &new_owner->pi_state_list);
799         pi_state->owner = new_owner;
800         spin_unlock_irq(&new_owner->pi_lock);
801
802         spin_unlock(&pi_state->pi_mutex.wait_lock);
803         rt_mutex_unlock(&pi_state->pi_mutex);
804
805         return 0;
806 }
807
808 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
809 {
810         u32 oldval;
811
812         /*
813          * There is no waiter, so we unlock the futex. The owner died
814          * bit has not to be preserved here. We are the owner:
815          */
816         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
817
818         if (oldval == -EFAULT)
819                 return oldval;
820         if (oldval != uval)
821                 return -EAGAIN;
822
823         return 0;
824 }
825
826 /*
827  * Express the locking dependencies for lockdep:
828  */
829 static inline void
830 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
831 {
832         if (hb1 <= hb2) {
833                 spin_lock(&hb1->lock);
834                 if (hb1 < hb2)
835                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
836         } else { /* hb1 > hb2 */
837                 spin_lock(&hb2->lock);
838                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
839         }
840 }
841
842 static inline void
843 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
844 {
845         spin_unlock(&hb1->lock);
846         if (hb1 != hb2)
847                 spin_unlock(&hb2->lock);
848 }
849
850 /*
851  * Wake up waiters matching bitset queued on this futex (uaddr).
852  */
853 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
854 {
855         struct futex_hash_bucket *hb;
856         struct futex_q *this, *next;
857         struct plist_head *head;
858         union futex_key key = FUTEX_KEY_INIT;
859         int ret;
860
861         if (!bitset)
862                 return -EINVAL;
863
864         ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
865         if (unlikely(ret != 0))
866                 goto out;
867
868         hb = hash_futex(&key);
869         spin_lock(&hb->lock);
870         head = &hb->chain;
871
872         plist_for_each_entry_safe(this, next, head, list) {
873                 if (match_futex (&this->key, &key)) {
874                         if (this->pi_state || this->rt_waiter) {
875                                 ret = -EINVAL;
876                                 break;
877                         }
878
879                         /* Check if one of the bits is set in both bitsets */
880                         if (!(this->bitset & bitset))
881                                 continue;
882
883                         wake_futex(this);
884                         if (++ret >= nr_wake)
885                                 break;
886                 }
887         }
888
889         spin_unlock(&hb->lock);
890         put_futex_key(fshared, &key);
891 out:
892         return ret;
893 }
894
895 /*
896  * Wake up all waiters hashed on the physical page that is mapped
897  * to this virtual address:
898  */
899 static int
900 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
901               int nr_wake, int nr_wake2, int op)
902 {
903         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
904         struct futex_hash_bucket *hb1, *hb2;
905         struct plist_head *head;
906         struct futex_q *this, *next;
907         int ret, op_ret;
908
909 retry:
910         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
911         if (unlikely(ret != 0))
912                 goto out;
913         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
914         if (unlikely(ret != 0))
915                 goto out_put_key1;
916
917         hb1 = hash_futex(&key1);
918         hb2 = hash_futex(&key2);
919
920 retry_private:
921         double_lock_hb(hb1, hb2);
922         op_ret = futex_atomic_op_inuser(op, uaddr2);
923         if (unlikely(op_ret < 0)) {
924
925                 double_unlock_hb(hb1, hb2);
926
927 #ifndef CONFIG_MMU
928                 /*
929                  * we don't get EFAULT from MMU faults if we don't have an MMU,
930                  * but we might get them from range checking
931                  */
932                 ret = op_ret;
933                 goto out_put_keys;
934 #endif
935
936                 if (unlikely(op_ret != -EFAULT)) {
937                         ret = op_ret;
938                         goto out_put_keys;
939                 }
940
941                 ret = fault_in_user_writeable(uaddr2);
942                 if (ret)
943                         goto out_put_keys;
944
945                 if (!fshared)
946                         goto retry_private;
947
948                 put_futex_key(fshared, &key2);
949                 put_futex_key(fshared, &key1);
950                 goto retry;
951         }
952
953         head = &hb1->chain;
954
955         plist_for_each_entry_safe(this, next, head, list) {
956                 if (match_futex (&this->key, &key1)) {
957                         wake_futex(this);
958                         if (++ret >= nr_wake)
959                                 break;
960                 }
961         }
962
963         if (op_ret > 0) {
964                 head = &hb2->chain;
965
966                 op_ret = 0;
967                 plist_for_each_entry_safe(this, next, head, list) {
968                         if (match_futex (&this->key, &key2)) {
969                                 wake_futex(this);
970                                 if (++op_ret >= nr_wake2)
971                                         break;
972                         }
973                 }
974                 ret += op_ret;
975         }
976
977         double_unlock_hb(hb1, hb2);
978 out_put_keys:
979         put_futex_key(fshared, &key2);
980 out_put_key1:
981         put_futex_key(fshared, &key1);
982 out:
983         return ret;
984 }
985
986 /**
987  * requeue_futex() - Requeue a futex_q from one hb to another
988  * @q:          the futex_q to requeue
989  * @hb1:        the source hash_bucket
990  * @hb2:        the target hash_bucket
991  * @key2:       the new key for the requeued futex_q
992  */
993 static inline
994 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
995                    struct futex_hash_bucket *hb2, union futex_key *key2)
996 {
997
998         /*
999          * If key1 and key2 hash to the same bucket, no need to
1000          * requeue.
1001          */
1002         if (likely(&hb1->chain != &hb2->chain)) {
1003                 plist_del(&q->list, &hb1->chain);
1004                 plist_add(&q->list, &hb2->chain);
1005                 q->lock_ptr = &hb2->lock;
1006 #ifdef CONFIG_DEBUG_PI_LIST
1007                 q->list.plist.lock = &hb2->lock;
1008 #endif
1009         }
1010         get_futex_key_refs(key2);
1011         q->key = *key2;
1012 }
1013
1014 /**
1015  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1016  * @q:          the futex_q
1017  * @key:        the key of the requeue target futex
1018  * @hb:         the hash_bucket of the requeue target futex
1019  *
1020  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1021  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1022  * to the requeue target futex so the waiter can detect the wakeup on the right
1023  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1024  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1025  * to protect access to the pi_state to fixup the owner later.  Must be called
1026  * with both q->lock_ptr and hb->lock held.
1027  */
1028 static inline
1029 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1030                            struct futex_hash_bucket *hb)
1031 {
1032         drop_futex_key_refs(&q->key);
1033         get_futex_key_refs(key);
1034         q->key = *key;
1035
1036         WARN_ON(plist_node_empty(&q->list));
1037         plist_del(&q->list, &q->list.plist);
1038
1039         WARN_ON(!q->rt_waiter);
1040         q->rt_waiter = NULL;
1041
1042         q->lock_ptr = &hb->lock;
1043 #ifdef CONFIG_DEBUG_PI_LIST
1044         q->list.plist.lock = &hb->lock;
1045 #endif
1046
1047         wake_up_state(q->task, TASK_NORMAL);
1048 }
1049
1050 /**
1051  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1052  * @pifutex:            the user address of the to futex
1053  * @hb1:                the from futex hash bucket, must be locked by the caller
1054  * @hb2:                the to futex hash bucket, must be locked by the caller
1055  * @key1:               the from futex key
1056  * @key2:               the to futex key
1057  * @ps:                 address to store the pi_state pointer
1058  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1059  *
1060  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1061  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1062  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1063  * hb1 and hb2 must be held by the caller.
1064  *
1065  * Returns:
1066  *  0 - failed to acquire the lock atomicly
1067  *  1 - acquired the lock
1068  * <0 - error
1069  */
1070 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1071                                  struct futex_hash_bucket *hb1,
1072                                  struct futex_hash_bucket *hb2,
1073                                  union futex_key *key1, union futex_key *key2,
1074                                  struct futex_pi_state **ps, int set_waiters)
1075 {
1076         struct futex_q *top_waiter = NULL;
1077         u32 curval;
1078         int ret;
1079
1080         if (get_futex_value_locked(&curval, pifutex))
1081                 return -EFAULT;
1082
1083         /*
1084          * Find the top_waiter and determine if there are additional waiters.
1085          * If the caller intends to requeue more than 1 waiter to pifutex,
1086          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1087          * as we have means to handle the possible fault.  If not, don't set
1088          * the bit unecessarily as it will force the subsequent unlock to enter
1089          * the kernel.
1090          */
1091         top_waiter = futex_top_waiter(hb1, key1);
1092
1093         /* There are no waiters, nothing for us to do. */
1094         if (!top_waiter)
1095                 return 0;
1096
1097         /* Ensure we requeue to the expected futex. */
1098         if (!match_futex(top_waiter->requeue_pi_key, key2))
1099                 return -EINVAL;
1100
1101         /*
1102          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1103          * the contended case or if set_waiters is 1.  The pi_state is returned
1104          * in ps in contended cases.
1105          */
1106         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1107                                    set_waiters);
1108         if (ret == 1)
1109                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1110
1111         return ret;
1112 }
1113
1114 /**
1115  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1116  * uaddr1:      source futex user address
1117  * uaddr2:      target futex user address
1118  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1119  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1120  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1121  *              pi futex (pi to pi requeue is not supported)
1122  *
1123  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1124  * uaddr2 atomically on behalf of the top waiter.
1125  *
1126  * Returns:
1127  * >=0 - on success, the number of tasks requeued or woken
1128  *  <0 - on error
1129  */
1130 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1131                          int nr_wake, int nr_requeue, u32 *cmpval,
1132                          int requeue_pi)
1133 {
1134         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1135         int drop_count = 0, task_count = 0, ret;
1136         struct futex_pi_state *pi_state = NULL;
1137         struct futex_hash_bucket *hb1, *hb2;
1138         struct plist_head *head1;
1139         struct futex_q *this, *next;
1140         u32 curval2;
1141
1142         if (requeue_pi) {
1143                 /*
1144                  * requeue_pi requires a pi_state, try to allocate it now
1145                  * without any locks in case it fails.
1146                  */
1147                 if (refill_pi_state_cache())
1148                         return -ENOMEM;
1149                 /*
1150                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1151                  * + nr_requeue, since it acquires the rt_mutex prior to
1152                  * returning to userspace, so as to not leave the rt_mutex with
1153                  * waiters and no owner.  However, second and third wake-ups
1154                  * cannot be predicted as they involve race conditions with the
1155                  * first wake and a fault while looking up the pi_state.  Both
1156                  * pthread_cond_signal() and pthread_cond_broadcast() should
1157                  * use nr_wake=1.
1158                  */
1159                 if (nr_wake != 1)
1160                         return -EINVAL;
1161         }
1162
1163 retry:
1164         if (pi_state != NULL) {
1165                 /*
1166                  * We will have to lookup the pi_state again, so free this one
1167                  * to keep the accounting correct.
1168                  */
1169                 free_pi_state(pi_state);
1170                 pi_state = NULL;
1171         }
1172
1173         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1174         if (unlikely(ret != 0))
1175                 goto out;
1176         ret = get_futex_key(uaddr2, fshared, &key2,
1177                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1178         if (unlikely(ret != 0))
1179                 goto out_put_key1;
1180
1181         hb1 = hash_futex(&key1);
1182         hb2 = hash_futex(&key2);
1183
1184 retry_private:
1185         double_lock_hb(hb1, hb2);
1186
1187         if (likely(cmpval != NULL)) {
1188                 u32 curval;
1189
1190                 ret = get_futex_value_locked(&curval, uaddr1);
1191
1192                 if (unlikely(ret)) {
1193                         double_unlock_hb(hb1, hb2);
1194
1195                         ret = get_user(curval, uaddr1);
1196                         if (ret)
1197                                 goto out_put_keys;
1198
1199                         if (!fshared)
1200                                 goto retry_private;
1201
1202                         put_futex_key(fshared, &key2);
1203                         put_futex_key(fshared, &key1);
1204                         goto retry;
1205                 }
1206                 if (curval != *cmpval) {
1207                         ret = -EAGAIN;
1208                         goto out_unlock;
1209                 }
1210         }
1211
1212         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1213                 /*
1214                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1215                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1216                  * bit.  We force this here where we are able to easily handle
1217                  * faults rather in the requeue loop below.
1218                  */
1219                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1220                                                  &key2, &pi_state, nr_requeue);
1221
1222                 /*
1223                  * At this point the top_waiter has either taken uaddr2 or is
1224                  * waiting on it.  If the former, then the pi_state will not
1225                  * exist yet, look it up one more time to ensure we have a
1226                  * reference to it.
1227                  */
1228                 if (ret == 1) {
1229                         WARN_ON(pi_state);
1230                         task_count++;
1231                         ret = get_futex_value_locked(&curval2, uaddr2);
1232                         if (!ret)
1233                                 ret = lookup_pi_state(curval2, hb2, &key2,
1234                                                       &pi_state);
1235                 }
1236
1237                 switch (ret) {
1238                 case 0:
1239                         break;
1240                 case -EFAULT:
1241                         double_unlock_hb(hb1, hb2);
1242                         put_futex_key(fshared, &key2);
1243                         put_futex_key(fshared, &key1);
1244                         ret = fault_in_user_writeable(uaddr2);
1245                         if (!ret)
1246                                 goto retry;
1247                         goto out;
1248                 case -EAGAIN:
1249                         /* The owner was exiting, try again. */
1250                         double_unlock_hb(hb1, hb2);
1251                         put_futex_key(fshared, &key2);
1252                         put_futex_key(fshared, &key1);
1253                         cond_resched();
1254                         goto retry;
1255                 default:
1256                         goto out_unlock;
1257                 }
1258         }
1259
1260         head1 = &hb1->chain;
1261         plist_for_each_entry_safe(this, next, head1, list) {
1262                 if (task_count - nr_wake >= nr_requeue)
1263                         break;
1264
1265                 if (!match_futex(&this->key, &key1))
1266                         continue;
1267
1268                 /*
1269                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1270                  * be paired with each other and no other futex ops.
1271                  */
1272                 if ((requeue_pi && !this->rt_waiter) ||
1273                     (!requeue_pi && this->rt_waiter)) {
1274                         ret = -EINVAL;
1275                         break;
1276                 }
1277
1278                 /*
1279                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1280                  * lock, we already woke the top_waiter.  If not, it will be
1281                  * woken by futex_unlock_pi().
1282                  */
1283                 if (++task_count <= nr_wake && !requeue_pi) {
1284                         wake_futex(this);
1285                         continue;
1286                 }
1287
1288                 /* Ensure we requeue to the expected futex for requeue_pi. */
1289                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1290                         ret = -EINVAL;
1291                         break;
1292                 }
1293
1294                 /*
1295                  * Requeue nr_requeue waiters and possibly one more in the case
1296                  * of requeue_pi if we couldn't acquire the lock atomically.
1297                  */
1298                 if (requeue_pi) {
1299                         /* Prepare the waiter to take the rt_mutex. */
1300                         atomic_inc(&pi_state->refcount);
1301                         this->pi_state = pi_state;
1302                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1303                                                         this->rt_waiter,
1304                                                         this->task, 1);
1305                         if (ret == 1) {
1306                                 /* We got the lock. */
1307                                 requeue_pi_wake_futex(this, &key2, hb2);
1308                                 continue;
1309                         } else if (ret) {
1310                                 /* -EDEADLK */
1311                                 this->pi_state = NULL;
1312                                 free_pi_state(pi_state);
1313                                 goto out_unlock;
1314                         }
1315                 }
1316                 requeue_futex(this, hb1, hb2, &key2);
1317                 drop_count++;
1318         }
1319
1320 out_unlock:
1321         double_unlock_hb(hb1, hb2);
1322
1323         /*
1324          * drop_futex_key_refs() must be called outside the spinlocks. During
1325          * the requeue we moved futex_q's from the hash bucket at key1 to the
1326          * one at key2 and updated their key pointer.  We no longer need to
1327          * hold the references to key1.
1328          */
1329         while (--drop_count >= 0)
1330                 drop_futex_key_refs(&key1);
1331
1332 out_put_keys:
1333         put_futex_key(fshared, &key2);
1334 out_put_key1:
1335         put_futex_key(fshared, &key1);
1336 out:
1337         if (pi_state != NULL)
1338                 free_pi_state(pi_state);
1339         return ret ? ret : task_count;
1340 }
1341
1342 /* The key must be already stored in q->key. */
1343 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1344 {
1345         struct futex_hash_bucket *hb;
1346
1347         get_futex_key_refs(&q->key);
1348         hb = hash_futex(&q->key);
1349         q->lock_ptr = &hb->lock;
1350
1351         spin_lock(&hb->lock);
1352         return hb;
1353 }
1354
1355 static inline void
1356 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1357 {
1358         spin_unlock(&hb->lock);
1359         drop_futex_key_refs(&q->key);
1360 }
1361
1362 /**
1363  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1364  * @q:  The futex_q to enqueue
1365  * @hb: The destination hash bucket
1366  *
1367  * The hb->lock must be held by the caller, and is released here. A call to
1368  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1369  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1370  * or nothing if the unqueue is done as part of the wake process and the unqueue
1371  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1372  * an example).
1373  */
1374 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1375 {
1376         int prio;
1377
1378         /*
1379          * The priority used to register this element is
1380          * - either the real thread-priority for the real-time threads
1381          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1382          * - or MAX_RT_PRIO for non-RT threads.
1383          * Thus, all RT-threads are woken first in priority order, and
1384          * the others are woken last, in FIFO order.
1385          */
1386         prio = min(current->normal_prio, MAX_RT_PRIO);
1387
1388         plist_node_init(&q->list, prio);
1389 #ifdef CONFIG_DEBUG_PI_LIST
1390         q->list.plist.lock = &hb->lock;
1391 #endif
1392         plist_add(&q->list, &hb->chain);
1393         q->task = current;
1394         spin_unlock(&hb->lock);
1395 }
1396
1397 /**
1398  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1399  * @q:  The futex_q to unqueue
1400  *
1401  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1402  * be paired with exactly one earlier call to queue_me().
1403  *
1404  * Returns:
1405  *   1 - if the futex_q was still queued (and we removed unqueued it)
1406  *   0 - if the futex_q was already removed by the waking thread
1407  */
1408 static int unqueue_me(struct futex_q *q)
1409 {
1410         spinlock_t *lock_ptr;
1411         int ret = 0;
1412
1413         /* In the common case we don't take the spinlock, which is nice. */
1414 retry:
1415         lock_ptr = q->lock_ptr;
1416         barrier();
1417         if (lock_ptr != NULL) {
1418                 spin_lock(lock_ptr);
1419                 /*
1420                  * q->lock_ptr can change between reading it and
1421                  * spin_lock(), causing us to take the wrong lock.  This
1422                  * corrects the race condition.
1423                  *
1424                  * Reasoning goes like this: if we have the wrong lock,
1425                  * q->lock_ptr must have changed (maybe several times)
1426                  * between reading it and the spin_lock().  It can
1427                  * change again after the spin_lock() but only if it was
1428                  * already changed before the spin_lock().  It cannot,
1429                  * however, change back to the original value.  Therefore
1430                  * we can detect whether we acquired the correct lock.
1431                  */
1432                 if (unlikely(lock_ptr != q->lock_ptr)) {
1433                         spin_unlock(lock_ptr);
1434                         goto retry;
1435                 }
1436                 WARN_ON(plist_node_empty(&q->list));
1437                 plist_del(&q->list, &q->list.plist);
1438
1439                 BUG_ON(q->pi_state);
1440
1441                 spin_unlock(lock_ptr);
1442                 ret = 1;
1443         }
1444
1445         drop_futex_key_refs(&q->key);
1446         return ret;
1447 }
1448
1449 /*
1450  * PI futexes can not be requeued and must remove themself from the
1451  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1452  * and dropped here.
1453  */
1454 static void unqueue_me_pi(struct futex_q *q)
1455 {
1456         WARN_ON(plist_node_empty(&q->list));
1457         plist_del(&q->list, &q->list.plist);
1458
1459         BUG_ON(!q->pi_state);
1460         free_pi_state(q->pi_state);
1461         q->pi_state = NULL;
1462
1463         spin_unlock(q->lock_ptr);
1464
1465         drop_futex_key_refs(&q->key);
1466 }
1467
1468 /*
1469  * Fixup the pi_state owner with the new owner.
1470  *
1471  * Must be called with hash bucket lock held and mm->sem held for non
1472  * private futexes.
1473  */
1474 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1475                                 struct task_struct *newowner, int fshared)
1476 {
1477         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1478         struct futex_pi_state *pi_state = q->pi_state;
1479         struct task_struct *oldowner = pi_state->owner;
1480         u32 uval, curval, newval;
1481         int ret;
1482
1483         /* Owner died? */
1484         if (!pi_state->owner)
1485                 newtid |= FUTEX_OWNER_DIED;
1486
1487         /*
1488          * We are here either because we stole the rtmutex from the
1489          * pending owner or we are the pending owner which failed to
1490          * get the rtmutex. We have to replace the pending owner TID
1491          * in the user space variable. This must be atomic as we have
1492          * to preserve the owner died bit here.
1493          *
1494          * Note: We write the user space value _before_ changing the pi_state
1495          * because we can fault here. Imagine swapped out pages or a fork
1496          * that marked all the anonymous memory readonly for cow.
1497          *
1498          * Modifying pi_state _before_ the user space value would
1499          * leave the pi_state in an inconsistent state when we fault
1500          * here, because we need to drop the hash bucket lock to
1501          * handle the fault. This might be observed in the PID check
1502          * in lookup_pi_state.
1503          */
1504 retry:
1505         if (get_futex_value_locked(&uval, uaddr))
1506                 goto handle_fault;
1507
1508         while (1) {
1509                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1510
1511                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1512
1513                 if (curval == -EFAULT)
1514                         goto handle_fault;
1515                 if (curval == uval)
1516                         break;
1517                 uval = curval;
1518         }
1519
1520         /*
1521          * We fixed up user space. Now we need to fix the pi_state
1522          * itself.
1523          */
1524         if (pi_state->owner != NULL) {
1525                 spin_lock_irq(&pi_state->owner->pi_lock);
1526                 WARN_ON(list_empty(&pi_state->list));
1527                 list_del_init(&pi_state->list);
1528                 spin_unlock_irq(&pi_state->owner->pi_lock);
1529         }
1530
1531         pi_state->owner = newowner;
1532
1533         spin_lock_irq(&newowner->pi_lock);
1534         WARN_ON(!list_empty(&pi_state->list));
1535         list_add(&pi_state->list, &newowner->pi_state_list);
1536         spin_unlock_irq(&newowner->pi_lock);
1537         return 0;
1538
1539         /*
1540          * To handle the page fault we need to drop the hash bucket
1541          * lock here. That gives the other task (either the pending
1542          * owner itself or the task which stole the rtmutex) the
1543          * chance to try the fixup of the pi_state. So once we are
1544          * back from handling the fault we need to check the pi_state
1545          * after reacquiring the hash bucket lock and before trying to
1546          * do another fixup. When the fixup has been done already we
1547          * simply return.
1548          */
1549 handle_fault:
1550         spin_unlock(q->lock_ptr);
1551
1552         ret = fault_in_user_writeable(uaddr);
1553
1554         spin_lock(q->lock_ptr);
1555
1556         /*
1557          * Check if someone else fixed it for us:
1558          */
1559         if (pi_state->owner != oldowner)
1560                 return 0;
1561
1562         if (ret)
1563                 return ret;
1564
1565         goto retry;
1566 }
1567
1568 /*
1569  * In case we must use restart_block to restart a futex_wait,
1570  * we encode in the 'flags' shared capability
1571  */
1572 #define FLAGS_SHARED            0x01
1573 #define FLAGS_CLOCKRT           0x02
1574 #define FLAGS_HAS_TIMEOUT       0x04
1575
1576 static long futex_wait_restart(struct restart_block *restart);
1577
1578 /**
1579  * fixup_owner() - Post lock pi_state and corner case management
1580  * @uaddr:      user address of the futex
1581  * @fshared:    whether the futex is shared (1) or not (0)
1582  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1583  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1584  *
1585  * After attempting to lock an rt_mutex, this function is called to cleanup
1586  * the pi_state owner as well as handle race conditions that may allow us to
1587  * acquire the lock. Must be called with the hb lock held.
1588  *
1589  * Returns:
1590  *  1 - success, lock taken
1591  *  0 - success, lock not taken
1592  * <0 - on error (-EFAULT)
1593  */
1594 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1595                        int locked)
1596 {
1597         struct task_struct *owner;
1598         int ret = 0;
1599
1600         if (locked) {
1601                 /*
1602                  * Got the lock. We might not be the anticipated owner if we
1603                  * did a lock-steal - fix up the PI-state in that case:
1604                  */
1605                 if (q->pi_state->owner != current)
1606                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1607                 goto out;
1608         }
1609
1610         /*
1611          * Catch the rare case, where the lock was released when we were on the
1612          * way back before we locked the hash bucket.
1613          */
1614         if (q->pi_state->owner == current) {
1615                 /*
1616                  * Try to get the rt_mutex now. This might fail as some other
1617                  * task acquired the rt_mutex after we removed ourself from the
1618                  * rt_mutex waiters list.
1619                  */
1620                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1621                         locked = 1;
1622                         goto out;
1623                 }
1624
1625                 /*
1626                  * pi_state is incorrect, some other task did a lock steal and
1627                  * we returned due to timeout or signal without taking the
1628                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1629                  * locking, as the other task is now blocked on the hash bucket
1630                  * lock. Fix the state up.
1631                  */
1632                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1633                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1634                 goto out;
1635         }
1636
1637         /*
1638          * Paranoia check. If we did not take the lock, then we should not be
1639          * the owner, nor the pending owner, of the rt_mutex.
1640          */
1641         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1642                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1643                                 "pi-state %p\n", ret,
1644                                 q->pi_state->pi_mutex.owner,
1645                                 q->pi_state->owner);
1646
1647 out:
1648         return ret ? ret : locked;
1649 }
1650
1651 /**
1652  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1653  * @hb:         the futex hash bucket, must be locked by the caller
1654  * @q:          the futex_q to queue up on
1655  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1656  */
1657 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1658                                 struct hrtimer_sleeper *timeout)
1659 {
1660         /*
1661          * The task state is guaranteed to be set before another task can
1662          * wake it. set_current_state() is implemented using set_mb() and
1663          * queue_me() calls spin_unlock() upon completion, both serializing
1664          * access to the hash list and forcing another memory barrier.
1665          */
1666         set_current_state(TASK_INTERRUPTIBLE);
1667         queue_me(q, hb);
1668
1669         /* Arm the timer */
1670         if (timeout) {
1671                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1672                 if (!hrtimer_active(&timeout->timer))
1673                         timeout->task = NULL;
1674         }
1675
1676         /*
1677          * If we have been removed from the hash list, then another task
1678          * has tried to wake us, and we can skip the call to schedule().
1679          */
1680         if (likely(!plist_node_empty(&q->list))) {
1681                 /*
1682                  * If the timer has already expired, current will already be
1683                  * flagged for rescheduling. Only call schedule if there
1684                  * is no timeout, or if it has yet to expire.
1685                  */
1686                 if (!timeout || timeout->task)
1687                         schedule();
1688         }
1689         __set_current_state(TASK_RUNNING);
1690 }
1691
1692 /**
1693  * futex_wait_setup() - Prepare to wait on a futex
1694  * @uaddr:      the futex userspace address
1695  * @val:        the expected value
1696  * @fshared:    whether the futex is shared (1) or not (0)
1697  * @q:          the associated futex_q
1698  * @hb:         storage for hash_bucket pointer to be returned to caller
1699  *
1700  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1701  * compare it with the expected value.  Handle atomic faults internally.
1702  * Return with the hb lock held and a q.key reference on success, and unlocked
1703  * with no q.key reference on failure.
1704  *
1705  * Returns:
1706  *  0 - uaddr contains val and hb has been locked
1707  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1708  */
1709 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1710                            struct futex_q *q, struct futex_hash_bucket **hb)
1711 {
1712         u32 uval;
1713         int ret;
1714
1715         /*
1716          * Access the page AFTER the hash-bucket is locked.
1717          * Order is important:
1718          *
1719          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1720          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1721          *
1722          * The basic logical guarantee of a futex is that it blocks ONLY
1723          * if cond(var) is known to be true at the time of blocking, for
1724          * any cond.  If we queued after testing *uaddr, that would open
1725          * a race condition where we could block indefinitely with
1726          * cond(var) false, which would violate the guarantee.
1727          *
1728          * A consequence is that futex_wait() can return zero and absorb
1729          * a wakeup when *uaddr != val on entry to the syscall.  This is
1730          * rare, but normal.
1731          */
1732 retry:
1733         q->key = FUTEX_KEY_INIT;
1734         ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1735         if (unlikely(ret != 0))
1736                 return ret;
1737
1738 retry_private:
1739         *hb = queue_lock(q);
1740
1741         ret = get_futex_value_locked(&uval, uaddr);
1742
1743         if (ret) {
1744                 queue_unlock(q, *hb);
1745
1746                 ret = get_user(uval, uaddr);
1747                 if (ret)
1748                         goto out;
1749
1750                 if (!fshared)
1751                         goto retry_private;
1752
1753                 put_futex_key(fshared, &q->key);
1754                 goto retry;
1755         }
1756
1757         if (uval != val) {
1758                 queue_unlock(q, *hb);
1759                 ret = -EWOULDBLOCK;
1760         }
1761
1762 out:
1763         if (ret)
1764                 put_futex_key(fshared, &q->key);
1765         return ret;
1766 }
1767
1768 static int futex_wait(u32 __user *uaddr, int fshared,
1769                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1770 {
1771         struct hrtimer_sleeper timeout, *to = NULL;
1772         struct restart_block *restart;
1773         struct futex_hash_bucket *hb;
1774         struct futex_q q;
1775         int ret;
1776
1777         if (!bitset)
1778                 return -EINVAL;
1779
1780         q.pi_state = NULL;
1781         q.bitset = bitset;
1782         q.rt_waiter = NULL;
1783         q.requeue_pi_key = NULL;
1784
1785         if (abs_time) {
1786                 to = &timeout;
1787
1788                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1789                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1790                 hrtimer_init_sleeper(to, current);
1791                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1792                                              current->timer_slack_ns);
1793         }
1794
1795 retry:
1796         /* Prepare to wait on uaddr. */
1797         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1798         if (ret)
1799                 goto out;
1800
1801         /* queue_me and wait for wakeup, timeout, or a signal. */
1802         futex_wait_queue_me(hb, &q, to);
1803
1804         /* If we were woken (and unqueued), we succeeded, whatever. */
1805         ret = 0;
1806         if (!unqueue_me(&q))
1807                 goto out_put_key;
1808         ret = -ETIMEDOUT;
1809         if (to && !to->task)
1810                 goto out_put_key;
1811
1812         /*
1813          * We expect signal_pending(current), but we might be the
1814          * victim of a spurious wakeup as well.
1815          */
1816         if (!signal_pending(current)) {
1817                 put_futex_key(fshared, &q.key);
1818                 goto retry;
1819         }
1820
1821         ret = -ERESTARTSYS;
1822         if (!abs_time)
1823                 goto out_put_key;
1824
1825         restart = &current_thread_info()->restart_block;
1826         restart->fn = futex_wait_restart;
1827         restart->futex.uaddr = (u32 *)uaddr;
1828         restart->futex.val = val;
1829         restart->futex.time = abs_time->tv64;
1830         restart->futex.bitset = bitset;
1831         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1832
1833         if (fshared)
1834                 restart->futex.flags |= FLAGS_SHARED;
1835         if (clockrt)
1836                 restart->futex.flags |= FLAGS_CLOCKRT;
1837
1838         ret = -ERESTART_RESTARTBLOCK;
1839
1840 out_put_key:
1841         put_futex_key(fshared, &q.key);
1842 out:
1843         if (to) {
1844                 hrtimer_cancel(&to->timer);
1845                 destroy_hrtimer_on_stack(&to->timer);
1846         }
1847         return ret;
1848 }
1849
1850
1851 static long futex_wait_restart(struct restart_block *restart)
1852 {
1853         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1854         int fshared = 0;
1855         ktime_t t, *tp = NULL;
1856
1857         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1858                 t.tv64 = restart->futex.time;
1859                 tp = &t;
1860         }
1861         restart->fn = do_no_restart_syscall;
1862         if (restart->futex.flags & FLAGS_SHARED)
1863                 fshared = 1;
1864         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1865                                 restart->futex.bitset,
1866                                 restart->futex.flags & FLAGS_CLOCKRT);
1867 }
1868
1869
1870 /*
1871  * Userspace tried a 0 -> TID atomic transition of the futex value
1872  * and failed. The kernel side here does the whole locking operation:
1873  * if there are waiters then it will block, it does PI, etc. (Due to
1874  * races the kernel might see a 0 value of the futex too.)
1875  */
1876 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1877                          int detect, ktime_t *time, int trylock)
1878 {
1879         struct hrtimer_sleeper timeout, *to = NULL;
1880         struct futex_hash_bucket *hb;
1881         struct futex_q q;
1882         int res, ret;
1883
1884         if (refill_pi_state_cache())
1885                 return -ENOMEM;
1886
1887         if (time) {
1888                 to = &timeout;
1889                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1890                                       HRTIMER_MODE_ABS);
1891                 hrtimer_init_sleeper(to, current);
1892                 hrtimer_set_expires(&to->timer, *time);
1893         }
1894
1895         q.pi_state = NULL;
1896         q.rt_waiter = NULL;
1897         q.requeue_pi_key = NULL;
1898 retry:
1899         q.key = FUTEX_KEY_INIT;
1900         ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1901         if (unlikely(ret != 0))
1902                 goto out;
1903
1904 retry_private:
1905         hb = queue_lock(&q);
1906
1907         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1908         if (unlikely(ret)) {
1909                 switch (ret) {
1910                 case 1:
1911                         /* We got the lock. */
1912                         ret = 0;
1913                         goto out_unlock_put_key;
1914                 case -EFAULT:
1915                         goto uaddr_faulted;
1916                 case -EAGAIN:
1917                         /*
1918                          * Task is exiting and we just wait for the
1919                          * exit to complete.
1920                          */
1921                         queue_unlock(&q, hb);
1922                         put_futex_key(fshared, &q.key);
1923                         cond_resched();
1924                         goto retry;
1925                 default:
1926                         goto out_unlock_put_key;
1927                 }
1928         }
1929
1930         /*
1931          * Only actually queue now that the atomic ops are done:
1932          */
1933         queue_me(&q, hb);
1934
1935         WARN_ON(!q.pi_state);
1936         /*
1937          * Block on the PI mutex:
1938          */
1939         if (!trylock)
1940                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1941         else {
1942                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1943                 /* Fixup the trylock return value: */
1944                 ret = ret ? 0 : -EWOULDBLOCK;
1945         }
1946
1947         spin_lock(q.lock_ptr);
1948         /*
1949          * Fixup the pi_state owner and possibly acquire the lock if we
1950          * haven't already.
1951          */
1952         res = fixup_owner(uaddr, fshared, &q, !ret);
1953         /*
1954          * If fixup_owner() returned an error, proprogate that.  If it acquired
1955          * the lock, clear our -ETIMEDOUT or -EINTR.
1956          */
1957         if (res)
1958                 ret = (res < 0) ? res : 0;
1959
1960         /*
1961          * If fixup_owner() faulted and was unable to handle the fault, unlock
1962          * it and return the fault to userspace.
1963          */
1964         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1965                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1966
1967         /* Unqueue and drop the lock */
1968         unqueue_me_pi(&q);
1969
1970         goto out;
1971
1972 out_unlock_put_key:
1973         queue_unlock(&q, hb);
1974
1975 out_put_key:
1976         put_futex_key(fshared, &q.key);
1977 out:
1978         if (to)
1979                 destroy_hrtimer_on_stack(&to->timer);
1980         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1981
1982 uaddr_faulted:
1983         queue_unlock(&q, hb);
1984
1985         ret = fault_in_user_writeable(uaddr);
1986         if (ret)
1987                 goto out_put_key;
1988
1989         if (!fshared)
1990                 goto retry_private;
1991
1992         put_futex_key(fshared, &q.key);
1993         goto retry;
1994 }
1995
1996 /*
1997  * Userspace attempted a TID -> 0 atomic transition, and failed.
1998  * This is the in-kernel slowpath: we look up the PI state (if any),
1999  * and do the rt-mutex unlock.
2000  */
2001 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2002 {
2003         struct futex_hash_bucket *hb;
2004         struct futex_q *this, *next;
2005         u32 uval;
2006         struct plist_head *head;
2007         union futex_key key = FUTEX_KEY_INIT;
2008         int ret;
2009
2010 retry:
2011         if (get_user(uval, uaddr))
2012                 return -EFAULT;
2013         /*
2014          * We release only a lock we actually own:
2015          */
2016         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2017                 return -EPERM;
2018
2019         ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
2020         if (unlikely(ret != 0))
2021                 goto out;
2022
2023         hb = hash_futex(&key);
2024         spin_lock(&hb->lock);
2025
2026         /*
2027          * To avoid races, try to do the TID -> 0 atomic transition
2028          * again. If it succeeds then we can return without waking
2029          * anyone else up:
2030          */
2031         if (!(uval & FUTEX_OWNER_DIED))
2032                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2033
2034
2035         if (unlikely(uval == -EFAULT))
2036                 goto pi_faulted;
2037         /*
2038          * Rare case: we managed to release the lock atomically,
2039          * no need to wake anyone else up:
2040          */
2041         if (unlikely(uval == task_pid_vnr(current)))
2042                 goto out_unlock;
2043
2044         /*
2045          * Ok, other tasks may need to be woken up - check waiters
2046          * and do the wakeup if necessary:
2047          */
2048         head = &hb->chain;
2049
2050         plist_for_each_entry_safe(this, next, head, list) {
2051                 if (!match_futex (&this->key, &key))
2052                         continue;
2053                 ret = wake_futex_pi(uaddr, uval, this);
2054                 /*
2055                  * The atomic access to the futex value
2056                  * generated a pagefault, so retry the
2057                  * user-access and the wakeup:
2058                  */
2059                 if (ret == -EFAULT)
2060                         goto pi_faulted;
2061                 goto out_unlock;
2062         }
2063         /*
2064          * No waiters - kernel unlocks the futex:
2065          */
2066         if (!(uval & FUTEX_OWNER_DIED)) {
2067                 ret = unlock_futex_pi(uaddr, uval);
2068                 if (ret == -EFAULT)
2069                         goto pi_faulted;
2070         }
2071
2072 out_unlock:
2073         spin_unlock(&hb->lock);
2074         put_futex_key(fshared, &key);
2075
2076 out:
2077         return ret;
2078
2079 pi_faulted:
2080         spin_unlock(&hb->lock);
2081         put_futex_key(fshared, &key);
2082
2083         ret = fault_in_user_writeable(uaddr);
2084         if (!ret)
2085                 goto retry;
2086
2087         return ret;
2088 }
2089
2090 /**
2091  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2092  * @hb:         the hash_bucket futex_q was original enqueued on
2093  * @q:          the futex_q woken while waiting to be requeued
2094  * @key2:       the futex_key of the requeue target futex
2095  * @timeout:    the timeout associated with the wait (NULL if none)
2096  *
2097  * Detect if the task was woken on the initial futex as opposed to the requeue
2098  * target futex.  If so, determine if it was a timeout or a signal that caused
2099  * the wakeup and return the appropriate error code to the caller.  Must be
2100  * called with the hb lock held.
2101  *
2102  * Returns
2103  *  0 - no early wakeup detected
2104  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2105  */
2106 static inline
2107 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2108                                    struct futex_q *q, union futex_key *key2,
2109                                    struct hrtimer_sleeper *timeout)
2110 {
2111         int ret = 0;
2112
2113         /*
2114          * With the hb lock held, we avoid races while we process the wakeup.
2115          * We only need to hold hb (and not hb2) to ensure atomicity as the
2116          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2117          * It can't be requeued from uaddr2 to something else since we don't
2118          * support a PI aware source futex for requeue.
2119          */
2120         if (!match_futex(&q->key, key2)) {
2121                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2122                 /*
2123                  * We were woken prior to requeue by a timeout or a signal.
2124                  * Unqueue the futex_q and determine which it was.
2125                  */
2126                 plist_del(&q->list, &q->list.plist);
2127
2128                 /* Handle spurious wakeups gracefully */
2129                 ret = -EAGAIN;
2130                 if (timeout && !timeout->task)
2131                         ret = -ETIMEDOUT;
2132                 else if (signal_pending(current))
2133                         ret = -ERESTARTNOINTR;
2134         }
2135         return ret;
2136 }
2137
2138 /**
2139  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2140  * @uaddr:      the futex we initially wait on (non-pi)
2141  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2142  *              the same type, no requeueing from private to shared, etc.
2143  * @val:        the expected value of uaddr
2144  * @abs_time:   absolute timeout
2145  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2146  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2147  * @uaddr2:     the pi futex we will take prior to returning to user-space
2148  *
2149  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2150  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2151  * complete the acquisition of the rt_mutex prior to returning to userspace.
2152  * This ensures the rt_mutex maintains an owner when it has waiters; without
2153  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2154  * need to.
2155  *
2156  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2157  * via the following:
2158  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2159  * 2) wakeup on uaddr2 after a requeue
2160  * 3) signal
2161  * 4) timeout
2162  *
2163  * If 3, cleanup and return -ERESTARTNOINTR.
2164  *
2165  * If 2, we may then block on trying to take the rt_mutex and return via:
2166  * 5) successful lock
2167  * 6) signal
2168  * 7) timeout
2169  * 8) other lock acquisition failure
2170  *
2171  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2172  *
2173  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2174  *
2175  * Returns:
2176  *  0 - On success
2177  * <0 - On error
2178  */
2179 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2180                                  u32 val, ktime_t *abs_time, u32 bitset,
2181                                  int clockrt, u32 __user *uaddr2)
2182 {
2183         struct hrtimer_sleeper timeout, *to = NULL;
2184         struct rt_mutex_waiter rt_waiter;
2185         struct rt_mutex *pi_mutex = NULL;
2186         struct futex_hash_bucket *hb;
2187         union futex_key key2;
2188         struct futex_q q;
2189         int res, ret;
2190
2191         if (!bitset)
2192                 return -EINVAL;
2193
2194         if (abs_time) {
2195                 to = &timeout;
2196                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2197                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2198                 hrtimer_init_sleeper(to, current);
2199                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2200                                              current->timer_slack_ns);
2201         }
2202
2203         /*
2204          * The waiter is allocated on our stack, manipulated by the requeue
2205          * code while we sleep on uaddr.
2206          */
2207         debug_rt_mutex_init_waiter(&rt_waiter);
2208         rt_waiter.task = NULL;
2209
2210 retry:
2211         key2 = FUTEX_KEY_INIT;
2212         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2213         if (unlikely(ret != 0))
2214                 goto out;
2215
2216         q.pi_state = NULL;
2217         q.bitset = bitset;
2218         q.rt_waiter = &rt_waiter;
2219         q.requeue_pi_key = &key2;
2220
2221         /* Prepare to wait on uaddr. */
2222         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2223         if (ret)
2224                 goto out_key2;
2225
2226         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2227         futex_wait_queue_me(hb, &q, to);
2228
2229         spin_lock(&hb->lock);
2230         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2231         spin_unlock(&hb->lock);
2232         if (ret)
2233                 goto out_put_keys;
2234
2235         /*
2236          * In order for us to be here, we know our q.key == key2, and since
2237          * we took the hb->lock above, we also know that futex_requeue() has
2238          * completed and we no longer have to concern ourselves with a wakeup
2239          * race with the atomic proxy lock acquition by the requeue code.
2240          */
2241
2242         /* Check if the requeue code acquired the second futex for us. */
2243         if (!q.rt_waiter) {
2244                 /*
2245                  * Got the lock. We might not be the anticipated owner if we
2246                  * did a lock-steal - fix up the PI-state in that case.
2247                  */
2248                 if (q.pi_state && (q.pi_state->owner != current)) {
2249                         spin_lock(q.lock_ptr);
2250                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2251                                                    fshared);
2252                         spin_unlock(q.lock_ptr);
2253                 }
2254         } else {
2255                 /*
2256                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2257                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2258                  * the pi_state.
2259                  */
2260                 WARN_ON(!&q.pi_state);
2261                 pi_mutex = &q.pi_state->pi_mutex;
2262                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2263                 debug_rt_mutex_free_waiter(&rt_waiter);
2264
2265                 spin_lock(q.lock_ptr);
2266                 /*
2267                  * Fixup the pi_state owner and possibly acquire the lock if we
2268                  * haven't already.
2269                  */
2270                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2271                 /*
2272                  * If fixup_owner() returned an error, proprogate that.  If it
2273                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2274                  */
2275                 if (res)
2276                         ret = (res < 0) ? res : 0;
2277
2278                 /* Unqueue and drop the lock. */
2279                 unqueue_me_pi(&q);
2280         }
2281
2282         /*
2283          * If fixup_pi_state_owner() faulted and was unable to handle the
2284          * fault, unlock the rt_mutex and return the fault to userspace.
2285          */
2286         if (ret == -EFAULT) {
2287                 if (rt_mutex_owner(pi_mutex) == current)
2288                         rt_mutex_unlock(pi_mutex);
2289         } else if (ret == -EINTR) {
2290                 /*
2291                  * We've already been requeued, but cannot restart by calling
2292                  * futex_lock_pi() directly. We could restart this syscall, but
2293                  * it would detect that the user space "val" changed and return
2294                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2295                  * -EWOULDBLOCK directly.
2296                  */
2297                 ret = -EWOULDBLOCK;
2298         }
2299
2300 out_put_keys:
2301         put_futex_key(fshared, &q.key);
2302 out_key2:
2303         put_futex_key(fshared, &key2);
2304
2305         /* Spurious wakeup ? */
2306         if (ret == -EAGAIN)
2307                 goto retry;
2308 out:
2309         if (to) {
2310                 hrtimer_cancel(&to->timer);
2311                 destroy_hrtimer_on_stack(&to->timer);
2312         }
2313         return ret;
2314 }
2315
2316 /*
2317  * Support for robust futexes: the kernel cleans up held futexes at
2318  * thread exit time.
2319  *
2320  * Implementation: user-space maintains a per-thread list of locks it
2321  * is holding. Upon do_exit(), the kernel carefully walks this list,
2322  * and marks all locks that are owned by this thread with the
2323  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2324  * always manipulated with the lock held, so the list is private and
2325  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2326  * field, to allow the kernel to clean up if the thread dies after
2327  * acquiring the lock, but just before it could have added itself to
2328  * the list. There can only be one such pending lock.
2329  */
2330
2331 /**
2332  * sys_set_robust_list() - Set the robust-futex list head of a task
2333  * @head:       pointer to the list-head
2334  * @len:        length of the list-head, as userspace expects
2335  */
2336 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2337                 size_t, len)
2338 {
2339         if (!futex_cmpxchg_enabled)
2340                 return -ENOSYS;
2341         /*
2342          * The kernel knows only one size for now:
2343          */
2344         if (unlikely(len != sizeof(*head)))
2345                 return -EINVAL;
2346
2347         current->robust_list = head;
2348
2349         return 0;
2350 }
2351
2352 /**
2353  * sys_get_robust_list() - Get the robust-futex list head of a task
2354  * @pid:        pid of the process [zero for current task]
2355  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2356  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2357  */
2358 SYSCALL_DEFINE3(get_robust_list, int, pid,
2359                 struct robust_list_head __user * __user *, head_ptr,
2360                 size_t __user *, len_ptr)
2361 {
2362         struct robust_list_head __user *head;
2363         unsigned long ret;
2364         const struct cred *cred = current_cred(), *pcred;
2365
2366         if (!futex_cmpxchg_enabled)
2367                 return -ENOSYS;
2368
2369         if (!pid)
2370                 head = current->robust_list;
2371         else {
2372                 struct task_struct *p;
2373
2374                 ret = -ESRCH;
2375                 rcu_read_lock();
2376                 p = find_task_by_vpid(pid);
2377                 if (!p)
2378                         goto err_unlock;
2379                 ret = -EPERM;
2380                 pcred = __task_cred(p);
2381                 if (cred->euid != pcred->euid &&
2382                     cred->euid != pcred->uid &&
2383                     !capable(CAP_SYS_PTRACE))
2384                         goto err_unlock;
2385                 head = p->robust_list;
2386                 rcu_read_unlock();
2387         }
2388
2389         if (put_user(sizeof(*head), len_ptr))
2390                 return -EFAULT;
2391         return put_user(head, head_ptr);
2392
2393 err_unlock:
2394         rcu_read_unlock();
2395
2396         return ret;
2397 }
2398
2399 /*
2400  * Process a futex-list entry, check whether it's owned by the
2401  * dying task, and do notification if so:
2402  */
2403 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2404 {
2405         u32 uval, nval, mval;
2406
2407 retry:
2408         if (get_user(uval, uaddr))
2409                 return -1;
2410
2411         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2412                 /*
2413                  * Ok, this dying thread is truly holding a futex
2414                  * of interest. Set the OWNER_DIED bit atomically
2415                  * via cmpxchg, and if the value had FUTEX_WAITERS
2416                  * set, wake up a waiter (if any). (We have to do a
2417                  * futex_wake() even if OWNER_DIED is already set -
2418                  * to handle the rare but possible case of recursive
2419                  * thread-death.) The rest of the cleanup is done in
2420                  * userspace.
2421                  */
2422                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2423                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2424
2425                 if (nval == -EFAULT)
2426                         return -1;
2427
2428                 if (nval != uval)
2429                         goto retry;
2430
2431                 /*
2432                  * Wake robust non-PI futexes here. The wakeup of
2433                  * PI futexes happens in exit_pi_state():
2434                  */
2435                 if (!pi && (uval & FUTEX_WAITERS))
2436                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2437         }
2438         return 0;
2439 }
2440
2441 /*
2442  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2443  */
2444 static inline int fetch_robust_entry(struct robust_list __user **entry,
2445                                      struct robust_list __user * __user *head,
2446                                      int *pi)
2447 {
2448         unsigned long uentry;
2449
2450         if (get_user(uentry, (unsigned long __user *)head))
2451                 return -EFAULT;
2452
2453         *entry = (void __user *)(uentry & ~1UL);
2454         *pi = uentry & 1;
2455
2456         return 0;
2457 }
2458
2459 /*
2460  * Walk curr->robust_list (very carefully, it's a userspace list!)
2461  * and mark any locks found there dead, and notify any waiters.
2462  *
2463  * We silently return on any sign of list-walking problem.
2464  */
2465 void exit_robust_list(struct task_struct *curr)
2466 {
2467         struct robust_list_head __user *head = curr->robust_list;
2468         struct robust_list __user *entry, *next_entry, *pending;
2469         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2470         unsigned long futex_offset;
2471         int rc;
2472
2473         if (!futex_cmpxchg_enabled)
2474                 return;
2475
2476         /*
2477          * Fetch the list head (which was registered earlier, via
2478          * sys_set_robust_list()):
2479          */
2480         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2481                 return;
2482         /*
2483          * Fetch the relative futex offset:
2484          */
2485         if (get_user(futex_offset, &head->futex_offset))
2486                 return;
2487         /*
2488          * Fetch any possibly pending lock-add first, and handle it
2489          * if it exists:
2490          */
2491         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2492                 return;
2493
2494         next_entry = NULL;      /* avoid warning with gcc */
2495         while (entry != &head->list) {
2496                 /*
2497                  * Fetch the next entry in the list before calling
2498                  * handle_futex_death:
2499                  */
2500                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2501                 /*
2502                  * A pending lock might already be on the list, so
2503                  * don't process it twice:
2504                  */
2505                 if (entry != pending)
2506                         if (handle_futex_death((void __user *)entry + futex_offset,
2507                                                 curr, pi))
2508                                 return;
2509                 if (rc)
2510                         return;
2511                 entry = next_entry;
2512                 pi = next_pi;
2513                 /*
2514                  * Avoid excessively long or circular lists:
2515                  */
2516                 if (!--limit)
2517                         break;
2518
2519                 cond_resched();
2520         }
2521
2522         if (pending)
2523                 handle_futex_death((void __user *)pending + futex_offset,
2524                                    curr, pip);
2525 }
2526
2527 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2528                 u32 __user *uaddr2, u32 val2, u32 val3)
2529 {
2530         int clockrt, ret = -ENOSYS;
2531         int cmd = op & FUTEX_CMD_MASK;
2532         int fshared = 0;
2533
2534         if (!(op & FUTEX_PRIVATE_FLAG))
2535                 fshared = 1;
2536
2537         clockrt = op & FUTEX_CLOCK_REALTIME;
2538         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2539                 return -ENOSYS;
2540
2541         switch (cmd) {
2542         case FUTEX_WAIT:
2543                 val3 = FUTEX_BITSET_MATCH_ANY;
2544         case FUTEX_WAIT_BITSET:
2545                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2546                 break;
2547         case FUTEX_WAKE:
2548                 val3 = FUTEX_BITSET_MATCH_ANY;
2549         case FUTEX_WAKE_BITSET:
2550                 ret = futex_wake(uaddr, fshared, val, val3);
2551                 break;
2552         case FUTEX_REQUEUE:
2553                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2554                 break;
2555         case FUTEX_CMP_REQUEUE:
2556                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2557                                     0);
2558                 break;
2559         case FUTEX_WAKE_OP:
2560                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2561                 break;
2562         case FUTEX_LOCK_PI:
2563                 if (futex_cmpxchg_enabled)
2564                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2565                 break;
2566         case FUTEX_UNLOCK_PI:
2567                 if (futex_cmpxchg_enabled)
2568                         ret = futex_unlock_pi(uaddr, fshared);
2569                 break;
2570         case FUTEX_TRYLOCK_PI:
2571                 if (futex_cmpxchg_enabled)
2572                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2573                 break;
2574         case FUTEX_WAIT_REQUEUE_PI:
2575                 val3 = FUTEX_BITSET_MATCH_ANY;
2576                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2577                                             clockrt, uaddr2);
2578                 break;
2579         case FUTEX_CMP_REQUEUE_PI:
2580                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2581                                     1);
2582                 break;
2583         default:
2584                 ret = -ENOSYS;
2585         }
2586         return ret;
2587 }
2588
2589
2590 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2591                 struct timespec __user *, utime, u32 __user *, uaddr2,
2592                 u32, val3)
2593 {
2594         struct timespec ts;
2595         ktime_t t, *tp = NULL;
2596         u32 val2 = 0;
2597         int cmd = op & FUTEX_CMD_MASK;
2598
2599         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2600                       cmd == FUTEX_WAIT_BITSET ||
2601                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2602                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2603                         return -EFAULT;
2604                 if (!timespec_valid(&ts))
2605                         return -EINVAL;
2606
2607                 t = timespec_to_ktime(ts);
2608                 if (cmd == FUTEX_WAIT)
2609                         t = ktime_add_safe(ktime_get(), t);
2610                 tp = &t;
2611         }
2612         /*
2613          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2614          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2615          */
2616         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2617             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2618                 val2 = (u32) (unsigned long) utime;
2619
2620         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2621 }
2622
2623 static int __init futex_init(void)
2624 {
2625         u32 curval;
2626         int i;
2627
2628         /*
2629          * This will fail and we want it. Some arch implementations do
2630          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2631          * functionality. We want to know that before we call in any
2632          * of the complex code paths. Also we want to prevent
2633          * registration of robust lists in that case. NULL is
2634          * guaranteed to fault and we get -EFAULT on functional
2635          * implementation, the non functional ones will return
2636          * -ENOSYS.
2637          */
2638         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2639         if (curval == -EFAULT)
2640                 futex_cmpxchg_enabled = 1;
2641
2642         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2643                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2644                 spin_lock_init(&futex_queues[i].lock);
2645         }
2646
2647         return 0;
2648 }
2649 __initcall(futex_init);