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