Revert "futex_requeue_pi optimization"
[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  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <asm/futex.h>
56
57 #include "rtmutex_common.h"
58
59 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
60
61 /*
62  * Priority Inheritance state:
63  */
64 struct futex_pi_state {
65         /*
66          * list of 'owned' pi_state instances - these have to be
67          * cleaned up in do_exit() if the task exits prematurely:
68          */
69         struct list_head list;
70
71         /*
72          * The PI object:
73          */
74         struct rt_mutex pi_mutex;
75
76         struct task_struct *owner;
77         atomic_t refcount;
78
79         union futex_key key;
80 };
81
82 /*
83  * We use this hashed waitqueue instead of a normal wait_queue_t, so
84  * we can wake only the relevant ones (hashed queues may be shared).
85  *
86  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
87  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
88  * The order of wakup is always to make the first condition true, then
89  * wake up q->waiters, then make the second condition true.
90  */
91 struct futex_q {
92         struct plist_node list;
93         wait_queue_head_t waiters;
94
95         /* Which hash list lock to use: */
96         spinlock_t *lock_ptr;
97
98         /* Key which the futex is hashed on: */
99         union futex_key key;
100
101         /* For fd, sigio sent using these: */
102         int fd;
103         struct file *filp;
104
105         /* Optional priority inheritance state: */
106         struct futex_pi_state *pi_state;
107         struct task_struct *task;
108 };
109
110 /*
111  * Split the global futex_lock into every hash list lock.
112  */
113 struct futex_hash_bucket {
114         spinlock_t lock;
115         struct plist_head chain;
116 };
117
118 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
119
120 /* Futex-fs vfsmount entry: */
121 static struct vfsmount *futex_mnt;
122
123 /*
124  * We hash on the keys returned from get_futex_key (see below).
125  */
126 static struct futex_hash_bucket *hash_futex(union futex_key *key)
127 {
128         u32 hash = jhash2((u32*)&key->both.word,
129                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
130                           key->both.offset);
131         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
132 }
133
134 /*
135  * Return 1 if two futex_keys are equal, 0 otherwise.
136  */
137 static inline int match_futex(union futex_key *key1, union futex_key *key2)
138 {
139         return (key1->both.word == key2->both.word
140                 && key1->both.ptr == key2->both.ptr
141                 && key1->both.offset == key2->both.offset);
142 }
143
144 /**
145  * get_futex_key - Get parameters which are the keys for a futex.
146  * @uaddr: virtual address of the futex
147  * @shared: NULL for a PROCESS_PRIVATE futex,
148  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
149  * @key: address where result is stored.
150  *
151  * Returns a negative error code or 0
152  * The key words are stored in *key on success.
153  *
154  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
155  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
156  * We can usually work out the index without swapping in the page.
157  *
158  * fshared is NULL for PROCESS_PRIVATE futexes
159  * For other futexes, it points to &current->mm->mmap_sem and
160  * caller must have taken the reader lock. but NOT any spinlocks.
161  */
162 int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
163                   union futex_key *key)
164 {
165         unsigned long address = (unsigned long)uaddr;
166         struct mm_struct *mm = current->mm;
167         struct vm_area_struct *vma;
168         struct page *page;
169         int err;
170
171         /*
172          * The futex address must be "naturally" aligned.
173          */
174         key->both.offset = address % PAGE_SIZE;
175         if (unlikely((address % sizeof(u32)) != 0))
176                 return -EINVAL;
177         address -= key->both.offset;
178
179         /*
180          * PROCESS_PRIVATE futexes are fast.
181          * As the mm cannot disappear under us and the 'key' only needs
182          * virtual address, we dont even have to find the underlying vma.
183          * Note : We do have to check 'uaddr' is a valid user address,
184          *        but access_ok() should be faster than find_vma()
185          */
186         if (!fshared) {
187                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
188                         return -EFAULT;
189                 key->private.mm = mm;
190                 key->private.address = address;
191                 return 0;
192         }
193         /*
194          * The futex is hashed differently depending on whether
195          * it's in a shared or private mapping.  So check vma first.
196          */
197         vma = find_extend_vma(mm, address);
198         if (unlikely(!vma))
199                 return -EFAULT;
200
201         /*
202          * Permissions.
203          */
204         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
205                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
206
207         /*
208          * Private mappings are handled in a simple way.
209          *
210          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
211          * it's a read-only handle, it's expected that futexes attach to
212          * the object not the particular process.  Therefore we use
213          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
214          * mappings of _writable_ handles.
215          */
216         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
217                 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
218                 key->private.mm = mm;
219                 key->private.address = address;
220                 return 0;
221         }
222
223         /*
224          * Linear file mappings are also simple.
225          */
226         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
227         key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
228         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
229                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
230                                      + vma->vm_pgoff);
231                 return 0;
232         }
233
234         /*
235          * We could walk the page table to read the non-linear
236          * pte, and get the page index without fetching the page
237          * from swap.  But that's a lot of code to duplicate here
238          * for a rare case, so we simply fetch the page.
239          */
240         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
241         if (err >= 0) {
242                 key->shared.pgoff =
243                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
244                 put_page(page);
245                 return 0;
246         }
247         return err;
248 }
249 EXPORT_SYMBOL_GPL(get_futex_key);
250
251 /*
252  * Take a reference to the resource addressed by a key.
253  * Can be called while holding spinlocks.
254  *
255  */
256 inline void get_futex_key_refs(union futex_key *key)
257 {
258         if (key->both.ptr == 0)
259                 return;
260         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
261                 case FUT_OFF_INODE:
262                         atomic_inc(&key->shared.inode->i_count);
263                         break;
264                 case FUT_OFF_MMSHARED:
265                         atomic_inc(&key->private.mm->mm_count);
266                         break;
267         }
268 }
269 EXPORT_SYMBOL_GPL(get_futex_key_refs);
270
271 /*
272  * Drop a reference to the resource addressed by a key.
273  * The hash bucket spinlock must not be held.
274  */
275 void drop_futex_key_refs(union futex_key *key)
276 {
277         if (key->both.ptr == 0)
278                 return;
279         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
280                 case FUT_OFF_INODE:
281                         iput(key->shared.inode);
282                         break;
283                 case FUT_OFF_MMSHARED:
284                         mmdrop(key->private.mm);
285                         break;
286         }
287 }
288 EXPORT_SYMBOL_GPL(drop_futex_key_refs);
289
290 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
291 {
292         int ret;
293
294         pagefault_disable();
295         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
296         pagefault_enable();
297
298         return ret ? -EFAULT : 0;
299 }
300
301 /*
302  * Fault handling.
303  * if fshared is non NULL, current->mm->mmap_sem is already held
304  */
305 static int futex_handle_fault(unsigned long address,
306                               struct rw_semaphore *fshared, int attempt)
307 {
308         struct vm_area_struct * vma;
309         struct mm_struct *mm = current->mm;
310         int ret = -EFAULT;
311
312         if (attempt > 2)
313                 return ret;
314
315         if (!fshared)
316                 down_read(&mm->mmap_sem);
317         vma = find_vma(mm, address);
318         if (vma && address >= vma->vm_start &&
319             (vma->vm_flags & VM_WRITE)) {
320                 switch (handle_mm_fault(mm, vma, address, 1)) {
321                 case VM_FAULT_MINOR:
322                         ret = 0;
323                         current->min_flt++;
324                         break;
325                 case VM_FAULT_MAJOR:
326                         ret = 0;
327                         current->maj_flt++;
328                         break;
329                 }
330         }
331         if (!fshared)
332                 up_read(&mm->mmap_sem);
333         return ret;
334 }
335
336 /*
337  * PI code:
338  */
339 static int refill_pi_state_cache(void)
340 {
341         struct futex_pi_state *pi_state;
342
343         if (likely(current->pi_state_cache))
344                 return 0;
345
346         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
347
348         if (!pi_state)
349                 return -ENOMEM;
350
351         INIT_LIST_HEAD(&pi_state->list);
352         /* pi_mutex gets initialized later */
353         pi_state->owner = NULL;
354         atomic_set(&pi_state->refcount, 1);
355
356         current->pi_state_cache = pi_state;
357
358         return 0;
359 }
360
361 static struct futex_pi_state * alloc_pi_state(void)
362 {
363         struct futex_pi_state *pi_state = current->pi_state_cache;
364
365         WARN_ON(!pi_state);
366         current->pi_state_cache = NULL;
367
368         return pi_state;
369 }
370
371 static void free_pi_state(struct futex_pi_state *pi_state)
372 {
373         if (!atomic_dec_and_test(&pi_state->refcount))
374                 return;
375
376         /*
377          * If pi_state->owner is NULL, the owner is most probably dying
378          * and has cleaned up the pi_state already
379          */
380         if (pi_state->owner) {
381                 spin_lock_irq(&pi_state->owner->pi_lock);
382                 list_del_init(&pi_state->list);
383                 spin_unlock_irq(&pi_state->owner->pi_lock);
384
385                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
386         }
387
388         if (current->pi_state_cache)
389                 kfree(pi_state);
390         else {
391                 /*
392                  * pi_state->list is already empty.
393                  * clear pi_state->owner.
394                  * refcount is at 0 - put it back to 1.
395                  */
396                 pi_state->owner = NULL;
397                 atomic_set(&pi_state->refcount, 1);
398                 current->pi_state_cache = pi_state;
399         }
400 }
401
402 /*
403  * Look up the task based on what TID userspace gave us.
404  * We dont trust it.
405  */
406 static struct task_struct * futex_find_get_task(pid_t pid)
407 {
408         struct task_struct *p;
409
410         rcu_read_lock();
411         p = find_task_by_pid(pid);
412         if (!p)
413                 goto out_unlock;
414         if ((current->euid != p->euid) && (current->euid != p->uid)) {
415                 p = NULL;
416                 goto out_unlock;
417         }
418         get_task_struct(p);
419 out_unlock:
420         rcu_read_unlock();
421
422         return p;
423 }
424
425 /*
426  * This task is holding PI mutexes at exit time => bad.
427  * Kernel cleans up PI-state, but userspace is likely hosed.
428  * (Robust-futex cleanup is separate and might save the day for userspace.)
429  */
430 void exit_pi_state_list(struct task_struct *curr)
431 {
432         struct list_head *next, *head = &curr->pi_state_list;
433         struct futex_pi_state *pi_state;
434         struct futex_hash_bucket *hb;
435         union futex_key key;
436
437         /*
438          * We are a ZOMBIE and nobody can enqueue itself on
439          * pi_state_list anymore, but we have to be careful
440          * versus waiters unqueueing themselves:
441          */
442         spin_lock_irq(&curr->pi_lock);
443         while (!list_empty(head)) {
444
445                 next = head->next;
446                 pi_state = list_entry(next, struct futex_pi_state, list);
447                 key = pi_state->key;
448                 hb = hash_futex(&key);
449                 spin_unlock_irq(&curr->pi_lock);
450
451                 spin_lock(&hb->lock);
452
453                 spin_lock_irq(&curr->pi_lock);
454                 /*
455                  * We dropped the pi-lock, so re-check whether this
456                  * task still owns the PI-state:
457                  */
458                 if (head->next != next) {
459                         spin_unlock(&hb->lock);
460                         continue;
461                 }
462
463                 WARN_ON(pi_state->owner != curr);
464                 WARN_ON(list_empty(&pi_state->list));
465                 list_del_init(&pi_state->list);
466                 pi_state->owner = NULL;
467                 spin_unlock_irq(&curr->pi_lock);
468
469                 rt_mutex_unlock(&pi_state->pi_mutex);
470
471                 spin_unlock(&hb->lock);
472
473                 spin_lock_irq(&curr->pi_lock);
474         }
475         spin_unlock_irq(&curr->pi_lock);
476 }
477
478 static int
479 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
480                 union futex_key *key, struct futex_pi_state **ps)
481 {
482         struct futex_pi_state *pi_state = NULL;
483         struct futex_q *this, *next;
484         struct plist_head *head;
485         struct task_struct *p;
486         pid_t pid = uval & FUTEX_TID_MASK;
487
488         head = &hb->chain;
489
490         plist_for_each_entry_safe(this, next, head, list) {
491                 if (match_futex(&this->key, key)) {
492                         /*
493                          * Another waiter already exists - bump up
494                          * the refcount and return its pi_state:
495                          */
496                         pi_state = this->pi_state;
497                         /*
498                          * Userspace might have messed up non PI and PI futexes
499                          */
500                         if (unlikely(!pi_state))
501                                 return -EINVAL;
502
503                         WARN_ON(!atomic_read(&pi_state->refcount));
504                         WARN_ON(pid && pi_state->owner &&
505                                 pi_state->owner->pid != pid);
506
507                         atomic_inc(&pi_state->refcount);
508                         *ps = pi_state;
509
510                         return 0;
511                 }
512         }
513
514         /*
515          * We are the first waiter - try to look up the real owner and attach
516          * the new pi_state to it, but bail out when TID = 0
517          */
518         if (!pid)
519                 return -ESRCH;
520         p = futex_find_get_task(pid);
521         if (IS_ERR(p))
522                 return PTR_ERR(p);
523
524         /*
525          * We need to look at the task state flags to figure out,
526          * whether the task is exiting. To protect against the do_exit
527          * change of the task flags, we do this protected by
528          * p->pi_lock:
529          */
530         spin_lock_irq(&p->pi_lock);
531         if (unlikely(p->flags & PF_EXITING)) {
532                 /*
533                  * The task is on the way out. When PF_EXITPIDONE is
534                  * set, we know that the task has finished the
535                  * cleanup:
536                  */
537                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
538
539                 spin_unlock_irq(&p->pi_lock);
540                 put_task_struct(p);
541                 return ret;
542         }
543
544         pi_state = alloc_pi_state();
545
546         /*
547          * Initialize the pi_mutex in locked state and make 'p'
548          * the owner of it:
549          */
550         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
551
552         /* Store the key for possible exit cleanups: */
553         pi_state->key = *key;
554
555         WARN_ON(!list_empty(&pi_state->list));
556         list_add(&pi_state->list, &p->pi_state_list);
557         pi_state->owner = p;
558         spin_unlock_irq(&p->pi_lock);
559
560         put_task_struct(p);
561
562         *ps = pi_state;
563
564         return 0;
565 }
566
567 /*
568  * The hash bucket lock must be held when this is called.
569  * Afterwards, the futex_q must not be accessed.
570  */
571 static void wake_futex(struct futex_q *q)
572 {
573         plist_del(&q->list, &q->list.plist);
574         if (q->filp)
575                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
576         /*
577          * The lock in wake_up_all() is a crucial memory barrier after the
578          * plist_del() and also before assigning to q->lock_ptr.
579          */
580         wake_up_all(&q->waiters);
581         /*
582          * The waiting task can free the futex_q as soon as this is written,
583          * without taking any locks.  This must come last.
584          *
585          * A memory barrier is required here to prevent the following store
586          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
587          * at the end of wake_up_all() does not prevent this store from
588          * moving.
589          */
590         smp_wmb();
591         q->lock_ptr = NULL;
592 }
593
594 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
595 {
596         struct task_struct *new_owner;
597         struct futex_pi_state *pi_state = this->pi_state;
598         u32 curval, newval;
599
600         if (!pi_state)
601                 return -EINVAL;
602
603         spin_lock(&pi_state->pi_mutex.wait_lock);
604         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
605
606         /*
607          * This happens when we have stolen the lock and the original
608          * pending owner did not enqueue itself back on the rt_mutex.
609          * Thats not a tragedy. We know that way, that a lock waiter
610          * is on the fly. We make the futex_q waiter the pending owner.
611          */
612         if (!new_owner)
613                 new_owner = this->task;
614
615         /*
616          * We pass it to the next owner. (The WAITERS bit is always
617          * kept enabled while there is PI state around. We must also
618          * preserve the owner died bit.)
619          */
620         if (!(uval & FUTEX_OWNER_DIED)) {
621                 int ret = 0;
622
623                 newval = FUTEX_WAITERS | new_owner->pid;
624
625                 pagefault_disable();
626                 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
627                 pagefault_enable();
628
629                 if (curval == -EFAULT)
630                         ret = -EFAULT;
631                 if (curval != uval)
632                         ret = -EINVAL;
633                 if (ret) {
634                         spin_unlock(&pi_state->pi_mutex.wait_lock);
635                         return ret;
636                 }
637         }
638
639         spin_lock_irq(&pi_state->owner->pi_lock);
640         WARN_ON(list_empty(&pi_state->list));
641         list_del_init(&pi_state->list);
642         spin_unlock_irq(&pi_state->owner->pi_lock);
643
644         spin_lock_irq(&new_owner->pi_lock);
645         WARN_ON(!list_empty(&pi_state->list));
646         list_add(&pi_state->list, &new_owner->pi_state_list);
647         pi_state->owner = new_owner;
648         spin_unlock_irq(&new_owner->pi_lock);
649
650         spin_unlock(&pi_state->pi_mutex.wait_lock);
651         rt_mutex_unlock(&pi_state->pi_mutex);
652
653         return 0;
654 }
655
656 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
657 {
658         u32 oldval;
659
660         /*
661          * There is no waiter, so we unlock the futex. The owner died
662          * bit has not to be preserved here. We are the owner:
663          */
664         pagefault_disable();
665         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
666         pagefault_enable();
667
668         if (oldval == -EFAULT)
669                 return oldval;
670         if (oldval != uval)
671                 return -EAGAIN;
672
673         return 0;
674 }
675
676 /*
677  * Express the locking dependencies for lockdep:
678  */
679 static inline void
680 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
681 {
682         if (hb1 <= hb2) {
683                 spin_lock(&hb1->lock);
684                 if (hb1 < hb2)
685                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
686         } else { /* hb1 > hb2 */
687                 spin_lock(&hb2->lock);
688                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
689         }
690 }
691
692 /*
693  * Wake up all waiters hashed on the physical page that is mapped
694  * to this virtual address:
695  */
696 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
697                       int nr_wake)
698 {
699         struct futex_hash_bucket *hb;
700         struct futex_q *this, *next;
701         struct plist_head *head;
702         union futex_key key;
703         int ret;
704
705         if (fshared)
706                 down_read(fshared);
707
708         ret = get_futex_key(uaddr, fshared, &key);
709         if (unlikely(ret != 0))
710                 goto out;
711
712         hb = hash_futex(&key);
713         spin_lock(&hb->lock);
714         head = &hb->chain;
715
716         plist_for_each_entry_safe(this, next, head, list) {
717                 if (match_futex (&this->key, &key)) {
718                         if (this->pi_state) {
719                                 ret = -EINVAL;
720                                 break;
721                         }
722                         wake_futex(this);
723                         if (++ret >= nr_wake)
724                                 break;
725                 }
726         }
727
728         spin_unlock(&hb->lock);
729 out:
730         if (fshared)
731                 up_read(fshared);
732         return ret;
733 }
734
735 /*
736  * Wake up all waiters hashed on the physical page that is mapped
737  * to this virtual address:
738  */
739 static int
740 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
741               u32 __user *uaddr2,
742               int nr_wake, int nr_wake2, int op)
743 {
744         union futex_key key1, key2;
745         struct futex_hash_bucket *hb1, *hb2;
746         struct plist_head *head;
747         struct futex_q *this, *next;
748         int ret, op_ret, attempt = 0;
749
750 retryfull:
751         if (fshared)
752                 down_read(fshared);
753
754         ret = get_futex_key(uaddr1, fshared, &key1);
755         if (unlikely(ret != 0))
756                 goto out;
757         ret = get_futex_key(uaddr2, fshared, &key2);
758         if (unlikely(ret != 0))
759                 goto out;
760
761         hb1 = hash_futex(&key1);
762         hb2 = hash_futex(&key2);
763
764 retry:
765         double_lock_hb(hb1, hb2);
766
767         op_ret = futex_atomic_op_inuser(op, uaddr2);
768         if (unlikely(op_ret < 0)) {
769                 u32 dummy;
770
771                 spin_unlock(&hb1->lock);
772                 if (hb1 != hb2)
773                         spin_unlock(&hb2->lock);
774
775 #ifndef CONFIG_MMU
776                 /*
777                  * we don't get EFAULT from MMU faults if we don't have an MMU,
778                  * but we might get them from range checking
779                  */
780                 ret = op_ret;
781                 goto out;
782 #endif
783
784                 if (unlikely(op_ret != -EFAULT)) {
785                         ret = op_ret;
786                         goto out;
787                 }
788
789                 /*
790                  * futex_atomic_op_inuser needs to both read and write
791                  * *(int __user *)uaddr2, but we can't modify it
792                  * non-atomically.  Therefore, if get_user below is not
793                  * enough, we need to handle the fault ourselves, while
794                  * still holding the mmap_sem.
795                  */
796                 if (attempt++) {
797                         ret = futex_handle_fault((unsigned long)uaddr2,
798                                                 fshared, attempt);
799                         if (ret)
800                                 goto out;
801                         goto retry;
802                 }
803
804                 /*
805                  * If we would have faulted, release mmap_sem,
806                  * fault it in and start all over again.
807                  */
808                 if (fshared)
809                         up_read(fshared);
810
811                 ret = get_user(dummy, uaddr2);
812                 if (ret)
813                         return ret;
814
815                 goto retryfull;
816         }
817
818         head = &hb1->chain;
819
820         plist_for_each_entry_safe(this, next, head, list) {
821                 if (match_futex (&this->key, &key1)) {
822                         wake_futex(this);
823                         if (++ret >= nr_wake)
824                                 break;
825                 }
826         }
827
828         if (op_ret > 0) {
829                 head = &hb2->chain;
830
831                 op_ret = 0;
832                 plist_for_each_entry_safe(this, next, head, list) {
833                         if (match_futex (&this->key, &key2)) {
834                                 wake_futex(this);
835                                 if (++op_ret >= nr_wake2)
836                                         break;
837                         }
838                 }
839                 ret += op_ret;
840         }
841
842         spin_unlock(&hb1->lock);
843         if (hb1 != hb2)
844                 spin_unlock(&hb2->lock);
845 out:
846         if (fshared)
847                 up_read(fshared);
848         return ret;
849 }
850
851 /*
852  * Requeue all waiters hashed on one physical page to another
853  * physical page.
854  */
855 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
856                          u32 __user *uaddr2,
857                          int nr_wake, int nr_requeue, u32 *cmpval)
858 {
859         union futex_key key1, key2;
860         struct futex_hash_bucket *hb1, *hb2;
861         struct plist_head *head1;
862         struct futex_q *this, *next;
863         int ret, drop_count = 0;
864
865  retry:
866         if (fshared)
867                 down_read(fshared);
868
869         ret = get_futex_key(uaddr1, fshared, &key1);
870         if (unlikely(ret != 0))
871                 goto out;
872         ret = get_futex_key(uaddr2, fshared, &key2);
873         if (unlikely(ret != 0))
874                 goto out;
875
876         hb1 = hash_futex(&key1);
877         hb2 = hash_futex(&key2);
878
879         double_lock_hb(hb1, hb2);
880
881         if (likely(cmpval != NULL)) {
882                 u32 curval;
883
884                 ret = get_futex_value_locked(&curval, uaddr1);
885
886                 if (unlikely(ret)) {
887                         spin_unlock(&hb1->lock);
888                         if (hb1 != hb2)
889                                 spin_unlock(&hb2->lock);
890
891                         /*
892                          * If we would have faulted, release mmap_sem, fault
893                          * it in and start all over again.
894                          */
895                         if (fshared)
896                                 up_read(fshared);
897
898                         ret = get_user(curval, uaddr1);
899
900                         if (!ret)
901                                 goto retry;
902
903                         return ret;
904                 }
905                 if (curval != *cmpval) {
906                         ret = -EAGAIN;
907                         goto out_unlock;
908                 }
909         }
910
911         head1 = &hb1->chain;
912         plist_for_each_entry_safe(this, next, head1, list) {
913                 if (!match_futex (&this->key, &key1))
914                         continue;
915                 if (++ret <= nr_wake) {
916                         wake_futex(this);
917                 } else {
918                         /*
919                          * If key1 and key2 hash to the same bucket, no need to
920                          * requeue.
921                          */
922                         if (likely(head1 != &hb2->chain)) {
923                                 plist_del(&this->list, &hb1->chain);
924                                 plist_add(&this->list, &hb2->chain);
925                                 this->lock_ptr = &hb2->lock;
926 #ifdef CONFIG_DEBUG_PI_LIST
927                                 this->list.plist.lock = &hb2->lock;
928 #endif
929                         }
930                         this->key = key2;
931                         get_futex_key_refs(&key2);
932                         drop_count++;
933
934                         if (ret - nr_wake >= nr_requeue)
935                                 break;
936                 }
937         }
938
939 out_unlock:
940         spin_unlock(&hb1->lock);
941         if (hb1 != hb2)
942                 spin_unlock(&hb2->lock);
943
944         /* drop_futex_key_refs() must be called outside the spinlocks. */
945         while (--drop_count >= 0)
946                 drop_futex_key_refs(&key1);
947
948 out:
949         if (fshared)
950                 up_read(fshared);
951         return ret;
952 }
953
954 /* The key must be already stored in q->key. */
955 static inline struct futex_hash_bucket *
956 queue_lock(struct futex_q *q, int fd, struct file *filp)
957 {
958         struct futex_hash_bucket *hb;
959
960         q->fd = fd;
961         q->filp = filp;
962
963         init_waitqueue_head(&q->waiters);
964
965         get_futex_key_refs(&q->key);
966         hb = hash_futex(&q->key);
967         q->lock_ptr = &hb->lock;
968
969         spin_lock(&hb->lock);
970         return hb;
971 }
972
973 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
974 {
975         int prio;
976
977         /*
978          * The priority used to register this element is
979          * - either the real thread-priority for the real-time threads
980          * (i.e. threads with a priority lower than MAX_RT_PRIO)
981          * - or MAX_RT_PRIO for non-RT threads.
982          * Thus, all RT-threads are woken first in priority order, and
983          * the others are woken last, in FIFO order.
984          */
985         prio = min(current->normal_prio, MAX_RT_PRIO);
986
987         plist_node_init(&q->list, prio);
988 #ifdef CONFIG_DEBUG_PI_LIST
989         q->list.plist.lock = &hb->lock;
990 #endif
991         plist_add(&q->list, &hb->chain);
992         q->task = current;
993         spin_unlock(&hb->lock);
994 }
995
996 static inline void
997 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
998 {
999         spin_unlock(&hb->lock);
1000         drop_futex_key_refs(&q->key);
1001 }
1002
1003 /*
1004  * queue_me and unqueue_me must be called as a pair, each
1005  * exactly once.  They are called with the hashed spinlock held.
1006  */
1007
1008 /* The key must be already stored in q->key. */
1009 static void queue_me(struct futex_q *q, int fd, struct file *filp)
1010 {
1011         struct futex_hash_bucket *hb;
1012
1013         hb = queue_lock(q, fd, filp);
1014         __queue_me(q, hb);
1015 }
1016
1017 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1018 static int unqueue_me(struct futex_q *q)
1019 {
1020         spinlock_t *lock_ptr;
1021         int ret = 0;
1022
1023         /* In the common case we don't take the spinlock, which is nice. */
1024  retry:
1025         lock_ptr = q->lock_ptr;
1026         barrier();
1027         if (lock_ptr != 0) {
1028                 spin_lock(lock_ptr);
1029                 /*
1030                  * q->lock_ptr can change between reading it and
1031                  * spin_lock(), causing us to take the wrong lock.  This
1032                  * corrects the race condition.
1033                  *
1034                  * Reasoning goes like this: if we have the wrong lock,
1035                  * q->lock_ptr must have changed (maybe several times)
1036                  * between reading it and the spin_lock().  It can
1037                  * change again after the spin_lock() but only if it was
1038                  * already changed before the spin_lock().  It cannot,
1039                  * however, change back to the original value.  Therefore
1040                  * we can detect whether we acquired the correct lock.
1041                  */
1042                 if (unlikely(lock_ptr != q->lock_ptr)) {
1043                         spin_unlock(lock_ptr);
1044                         goto retry;
1045                 }
1046                 WARN_ON(plist_node_empty(&q->list));
1047                 plist_del(&q->list, &q->list.plist);
1048
1049                 BUG_ON(q->pi_state);
1050
1051                 spin_unlock(lock_ptr);
1052                 ret = 1;
1053         }
1054
1055         drop_futex_key_refs(&q->key);
1056         return ret;
1057 }
1058
1059 /*
1060  * PI futexes can not be requeued and must remove themself from the
1061  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1062  * and dropped here.
1063  */
1064 static void unqueue_me_pi(struct futex_q *q)
1065 {
1066         WARN_ON(plist_node_empty(&q->list));
1067         plist_del(&q->list, &q->list.plist);
1068
1069         BUG_ON(!q->pi_state);
1070         free_pi_state(q->pi_state);
1071         q->pi_state = NULL;
1072
1073         spin_unlock(q->lock_ptr);
1074
1075         drop_futex_key_refs(&q->key);
1076 }
1077
1078 /*
1079  * Fixup the pi_state owner with current.
1080  *
1081  * Must be called with hash bucket lock held and mm->sem held for non
1082  * private futexes.
1083  */
1084 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1085                                 struct task_struct *curr)
1086 {
1087         u32 newtid = curr->pid | FUTEX_WAITERS;
1088         struct futex_pi_state *pi_state = q->pi_state;
1089         u32 uval, curval, newval;
1090         int ret;
1091
1092         /* Owner died? */
1093         if (pi_state->owner != NULL) {
1094                 spin_lock_irq(&pi_state->owner->pi_lock);
1095                 WARN_ON(list_empty(&pi_state->list));
1096                 list_del_init(&pi_state->list);
1097                 spin_unlock_irq(&pi_state->owner->pi_lock);
1098         } else
1099                 newtid |= FUTEX_OWNER_DIED;
1100
1101         pi_state->owner = curr;
1102
1103         spin_lock_irq(&curr->pi_lock);
1104         WARN_ON(!list_empty(&pi_state->list));
1105         list_add(&pi_state->list, &curr->pi_state_list);
1106         spin_unlock_irq(&curr->pi_lock);
1107
1108         /*
1109          * We own it, so we have to replace the pending owner
1110          * TID. This must be atomic as we have preserve the
1111          * owner died bit here.
1112          */
1113         ret = get_futex_value_locked(&uval, uaddr);
1114
1115         while (!ret) {
1116                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1117
1118                 pagefault_disable();
1119                 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1120                                                        uval, newval);
1121                 pagefault_enable();
1122
1123                 if (curval == -EFAULT)
1124                         ret = -EFAULT;
1125                 if (curval == uval)
1126                         break;
1127                 uval = curval;
1128         }
1129         return ret;
1130 }
1131
1132 /*
1133  * In case we must use restart_block to restart a futex_wait,
1134  * we encode in the 'arg3' shared capability
1135  */
1136 #define ARG3_SHARED  1
1137
1138 static long futex_wait_restart(struct restart_block *restart);
1139 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1140                       u32 val, ktime_t *abs_time)
1141 {
1142         struct task_struct *curr = current;
1143         DECLARE_WAITQUEUE(wait, curr);
1144         struct futex_hash_bucket *hb;
1145         struct futex_q q;
1146         u32 uval;
1147         int ret;
1148         struct hrtimer_sleeper t;
1149         int rem = 0;
1150
1151         q.pi_state = NULL;
1152  retry:
1153         if (fshared)
1154                 down_read(fshared);
1155
1156         ret = get_futex_key(uaddr, fshared, &q.key);
1157         if (unlikely(ret != 0))
1158                 goto out_release_sem;
1159
1160         hb = queue_lock(&q, -1, NULL);
1161
1162         /*
1163          * Access the page AFTER the futex is queued.
1164          * Order is important:
1165          *
1166          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1167          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1168          *
1169          * The basic logical guarantee of a futex is that it blocks ONLY
1170          * if cond(var) is known to be true at the time of blocking, for
1171          * any cond.  If we queued after testing *uaddr, that would open
1172          * a race condition where we could block indefinitely with
1173          * cond(var) false, which would violate the guarantee.
1174          *
1175          * A consequence is that futex_wait() can return zero and absorb
1176          * a wakeup when *uaddr != val on entry to the syscall.  This is
1177          * rare, but normal.
1178          *
1179          * for shared futexes, we hold the mmap semaphore, so the mapping
1180          * cannot have changed since we looked it up in get_futex_key.
1181          */
1182         ret = get_futex_value_locked(&uval, uaddr);
1183
1184         if (unlikely(ret)) {
1185                 queue_unlock(&q, hb);
1186
1187                 /*
1188                  * If we would have faulted, release mmap_sem, fault it in and
1189                  * start all over again.
1190                  */
1191                 if (fshared)
1192                         up_read(fshared);
1193
1194                 ret = get_user(uval, uaddr);
1195
1196                 if (!ret)
1197                         goto retry;
1198                 return ret;
1199         }
1200         ret = -EWOULDBLOCK;
1201         if (uval != val)
1202                 goto out_unlock_release_sem;
1203
1204         /* Only actually queue if *uaddr contained val.  */
1205         __queue_me(&q, hb);
1206
1207         /*
1208          * Now the futex is queued and we have checked the data, we
1209          * don't want to hold mmap_sem while we sleep.
1210          */
1211         if (fshared)
1212                 up_read(fshared);
1213
1214         /*
1215          * There might have been scheduling since the queue_me(), as we
1216          * cannot hold a spinlock across the get_user() in case it
1217          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1218          * queueing ourselves into the futex hash.  This code thus has to
1219          * rely on the futex_wake() code removing us from hash when it
1220          * wakes us up.
1221          */
1222
1223         /* add_wait_queue is the barrier after __set_current_state. */
1224         __set_current_state(TASK_INTERRUPTIBLE);
1225         add_wait_queue(&q.waiters, &wait);
1226         /*
1227          * !plist_node_empty() is safe here without any lock.
1228          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1229          */
1230         if (likely(!plist_node_empty(&q.list))) {
1231                 if (!abs_time)
1232                         schedule();
1233                 else {
1234                         hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1235                         hrtimer_init_sleeper(&t, current);
1236                         t.timer.expires = *abs_time;
1237
1238                         hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
1239
1240                         /*
1241                          * the timer could have already expired, in which
1242                          * case current would be flagged for rescheduling.
1243                          * Don't bother calling schedule.
1244                          */
1245                         if (likely(t.task))
1246                                 schedule();
1247
1248                         hrtimer_cancel(&t.timer);
1249
1250                         /* Flag if a timeout occured */
1251                         rem = (t.task == NULL);
1252                 }
1253         }
1254         __set_current_state(TASK_RUNNING);
1255
1256         /*
1257          * NOTE: we don't remove ourselves from the waitqueue because
1258          * we are the only user of it.
1259          */
1260
1261         /* If we were woken (and unqueued), we succeeded, whatever. */
1262         if (!unqueue_me(&q))
1263                 return 0;
1264         if (rem)
1265                 return -ETIMEDOUT;
1266
1267         /*
1268          * We expect signal_pending(current), but another thread may
1269          * have handled it for us already.
1270          */
1271         if (!abs_time)
1272                 return -ERESTARTSYS;
1273         else {
1274                 struct restart_block *restart;
1275                 restart = &current_thread_info()->restart_block;
1276                 restart->fn = futex_wait_restart;
1277                 restart->arg0 = (unsigned long)uaddr;
1278                 restart->arg1 = (unsigned long)val;
1279                 restart->arg2 = (unsigned long)abs_time;
1280                 restart->arg3 = 0;
1281                 if (fshared)
1282                         restart->arg3 |= ARG3_SHARED;
1283                 return -ERESTART_RESTARTBLOCK;
1284         }
1285
1286  out_unlock_release_sem:
1287         queue_unlock(&q, hb);
1288
1289  out_release_sem:
1290         if (fshared)
1291                 up_read(fshared);
1292         return ret;
1293 }
1294
1295
1296 static long futex_wait_restart(struct restart_block *restart)
1297 {
1298         u32 __user *uaddr = (u32 __user *)restart->arg0;
1299         u32 val = (u32)restart->arg1;
1300         ktime_t *abs_time = (ktime_t *)restart->arg2;
1301         struct rw_semaphore *fshared = NULL;
1302
1303         restart->fn = do_no_restart_syscall;
1304         if (restart->arg3 & ARG3_SHARED)
1305                 fshared = &current->mm->mmap_sem;
1306         return (long)futex_wait(uaddr, fshared, val, abs_time);
1307 }
1308
1309
1310 /*
1311  * Userspace tried a 0 -> TID atomic transition of the futex value
1312  * and failed. The kernel side here does the whole locking operation:
1313  * if there are waiters then it will block, it does PI, etc. (Due to
1314  * races the kernel might see a 0 value of the futex too.)
1315  */
1316 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1317                          int detect, ktime_t *time, int trylock)
1318 {
1319         struct hrtimer_sleeper timeout, *to = NULL;
1320         struct task_struct *curr = current;
1321         struct futex_hash_bucket *hb;
1322         u32 uval, newval, curval;
1323         struct futex_q q;
1324         int ret, lock_taken, ownerdied = 0, attempt = 0;
1325
1326         if (refill_pi_state_cache())
1327                 return -ENOMEM;
1328
1329         if (time) {
1330                 to = &timeout;
1331                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1332                 hrtimer_init_sleeper(to, current);
1333                 to->timer.expires = *time;
1334         }
1335
1336         q.pi_state = NULL;
1337  retry:
1338         if (fshared)
1339                 down_read(fshared);
1340
1341         ret = get_futex_key(uaddr, fshared, &q.key);
1342         if (unlikely(ret != 0))
1343                 goto out_release_sem;
1344
1345  retry_unlocked:
1346         hb = queue_lock(&q, -1, NULL);
1347
1348  retry_locked:
1349         ret = lock_taken = 0;
1350
1351         /*
1352          * To avoid races, we attempt to take the lock here again
1353          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1354          * the locks. It will most likely not succeed.
1355          */
1356         newval = current->pid;
1357
1358         pagefault_disable();
1359         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1360         pagefault_enable();
1361
1362         if (unlikely(curval == -EFAULT))
1363                 goto uaddr_faulted;
1364
1365         /*
1366          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1367          * situation and we return success to user space.
1368          */
1369         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1370                 ret = -EDEADLK;
1371                 goto out_unlock_release_sem;
1372         }
1373
1374         /*
1375          * Surprise - we got the lock. Just return to userspace:
1376          */
1377         if (unlikely(!curval))
1378                 goto out_unlock_release_sem;
1379
1380         uval = curval;
1381
1382         /*
1383          * Set the WAITERS flag, so the owner will know it has someone
1384          * to wake at next unlock
1385          */
1386         newval = curval | FUTEX_WAITERS;
1387
1388         /*
1389          * There are two cases, where a futex might have no owner (the
1390          * owner TID is 0): OWNER_DIED. We take over the futex in this
1391          * case. We also do an unconditional take over, when the owner
1392          * of the futex died.
1393          *
1394          * This is safe as we are protected by the hash bucket lock !
1395          */
1396         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1397                 /* Keep the OWNER_DIED bit */
1398                 newval = (curval & ~FUTEX_TID_MASK) | current->pid;
1399                 ownerdied = 0;
1400                 lock_taken = 1;
1401         }
1402
1403         pagefault_disable();
1404         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1405         pagefault_enable();
1406
1407         if (unlikely(curval == -EFAULT))
1408                 goto uaddr_faulted;
1409         if (unlikely(curval != uval))
1410                 goto retry_locked;
1411
1412         /*
1413          * We took the lock due to owner died take over.
1414          */
1415         if (unlikely(lock_taken))
1416                 goto out_unlock_release_sem;
1417
1418         /*
1419          * We dont have the lock. Look up the PI state (or create it if
1420          * we are the first waiter):
1421          */
1422         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1423
1424         if (unlikely(ret)) {
1425                 switch (ret) {
1426
1427                 case -EAGAIN:
1428                         /*
1429                          * Task is exiting and we just wait for the
1430                          * exit to complete.
1431                          */
1432                         queue_unlock(&q, hb);
1433                         if (fshared)
1434                                 up_read(fshared);
1435                         cond_resched();
1436                         goto retry;
1437
1438                 case -ESRCH:
1439                         /*
1440                          * No owner found for this futex. Check if the
1441                          * OWNER_DIED bit is set to figure out whether
1442                          * this is a robust futex or not.
1443                          */
1444                         if (get_futex_value_locked(&curval, uaddr))
1445                                 goto uaddr_faulted;
1446
1447                         /*
1448                          * We simply start over in case of a robust
1449                          * futex. The code above will take the futex
1450                          * and return happy.
1451                          */
1452                         if (curval & FUTEX_OWNER_DIED) {
1453                                 ownerdied = 1;
1454                                 goto retry_locked;
1455                         }
1456                 default:
1457                         goto out_unlock_release_sem;
1458                 }
1459         }
1460
1461         /*
1462          * Only actually queue now that the atomic ops are done:
1463          */
1464         __queue_me(&q, hb);
1465
1466         /*
1467          * Now the futex is queued and we have checked the data, we
1468          * don't want to hold mmap_sem while we sleep.
1469          */
1470         if (fshared)
1471                 up_read(fshared);
1472
1473         WARN_ON(!q.pi_state);
1474         /*
1475          * Block on the PI mutex:
1476          */
1477         if (!trylock)
1478                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1479         else {
1480                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1481                 /* Fixup the trylock return value: */
1482                 ret = ret ? 0 : -EWOULDBLOCK;
1483         }
1484
1485         if (fshared)
1486                 down_read(fshared);
1487         spin_lock(q.lock_ptr);
1488
1489         if (!ret) {
1490                 /*
1491                  * Got the lock. We might not be the anticipated owner
1492                  * if we did a lock-steal - fix up the PI-state in
1493                  * that case:
1494                  */
1495                 if (q.pi_state->owner != curr)
1496                         ret = fixup_pi_state_owner(uaddr, &q, curr);
1497         } else {
1498                 /*
1499                  * Catch the rare case, where the lock was released
1500                  * when we were on the way back before we locked the
1501                  * hash bucket.
1502                  */
1503                 if (q.pi_state->owner == curr &&
1504                     rt_mutex_trylock(&q.pi_state->pi_mutex)) {
1505                         ret = 0;
1506                 } else {
1507                         /*
1508                          * Paranoia check. If we did not take the lock
1509                          * in the trylock above, then we should not be
1510                          * the owner of the rtmutex, neither the real
1511                          * nor the pending one:
1512                          */
1513                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1514                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1515                                        "pi-mutex: %p pi-state %p\n", ret,
1516                                        q.pi_state->pi_mutex.owner,
1517                                        q.pi_state->owner);
1518                 }
1519         }
1520
1521         /* Unqueue and drop the lock */
1522         unqueue_me_pi(&q);
1523         if (fshared)
1524                 up_read(fshared);
1525
1526         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1527
1528  out_unlock_release_sem:
1529         queue_unlock(&q, hb);
1530
1531  out_release_sem:
1532         if (fshared)
1533                 up_read(fshared);
1534         return ret;
1535
1536  uaddr_faulted:
1537         /*
1538          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1539          * non-atomically.  Therefore, if get_user below is not
1540          * enough, we need to handle the fault ourselves, while
1541          * still holding the mmap_sem.
1542          *
1543          * ... and hb->lock. :-) --ANK
1544          */
1545         queue_unlock(&q, hb);
1546
1547         if (attempt++) {
1548                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1549                                          attempt);
1550                 if (ret)
1551                         goto out_release_sem;
1552                 goto retry_unlocked;
1553         }
1554
1555         if (fshared)
1556                 up_read(fshared);
1557
1558         ret = get_user(uval, uaddr);
1559         if (!ret && (uval != -EFAULT))
1560                 goto retry;
1561
1562         return ret;
1563 }
1564
1565 /*
1566  * Userspace attempted a TID -> 0 atomic transition, and failed.
1567  * This is the in-kernel slowpath: we look up the PI state (if any),
1568  * and do the rt-mutex unlock.
1569  */
1570 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1571 {
1572         struct futex_hash_bucket *hb;
1573         struct futex_q *this, *next;
1574         u32 uval;
1575         struct plist_head *head;
1576         union futex_key key;
1577         int ret, attempt = 0;
1578
1579 retry:
1580         if (get_user(uval, uaddr))
1581                 return -EFAULT;
1582         /*
1583          * We release only a lock we actually own:
1584          */
1585         if ((uval & FUTEX_TID_MASK) != current->pid)
1586                 return -EPERM;
1587         /*
1588          * First take all the futex related locks:
1589          */
1590         if (fshared)
1591                 down_read(fshared);
1592
1593         ret = get_futex_key(uaddr, fshared, &key);
1594         if (unlikely(ret != 0))
1595                 goto out;
1596
1597         hb = hash_futex(&key);
1598 retry_unlocked:
1599         spin_lock(&hb->lock);
1600
1601         /*
1602          * To avoid races, try to do the TID -> 0 atomic transition
1603          * again. If it succeeds then we can return without waking
1604          * anyone else up:
1605          */
1606         if (!(uval & FUTEX_OWNER_DIED)) {
1607                 pagefault_disable();
1608                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1609                 pagefault_enable();
1610         }
1611
1612         if (unlikely(uval == -EFAULT))
1613                 goto pi_faulted;
1614         /*
1615          * Rare case: we managed to release the lock atomically,
1616          * no need to wake anyone else up:
1617          */
1618         if (unlikely(uval == current->pid))
1619                 goto out_unlock;
1620
1621         /*
1622          * Ok, other tasks may need to be woken up - check waiters
1623          * and do the wakeup if necessary:
1624          */
1625         head = &hb->chain;
1626
1627         plist_for_each_entry_safe(this, next, head, list) {
1628                 if (!match_futex (&this->key, &key))
1629                         continue;
1630                 ret = wake_futex_pi(uaddr, uval, this);
1631                 /*
1632                  * The atomic access to the futex value
1633                  * generated a pagefault, so retry the
1634                  * user-access and the wakeup:
1635                  */
1636                 if (ret == -EFAULT)
1637                         goto pi_faulted;
1638                 goto out_unlock;
1639         }
1640         /*
1641          * No waiters - kernel unlocks the futex:
1642          */
1643         if (!(uval & FUTEX_OWNER_DIED)) {
1644                 ret = unlock_futex_pi(uaddr, uval);
1645                 if (ret == -EFAULT)
1646                         goto pi_faulted;
1647         }
1648
1649 out_unlock:
1650         spin_unlock(&hb->lock);
1651 out:
1652         if (fshared)
1653                 up_read(fshared);
1654
1655         return ret;
1656
1657 pi_faulted:
1658         /*
1659          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1660          * non-atomically.  Therefore, if get_user below is not
1661          * enough, we need to handle the fault ourselves, while
1662          * still holding the mmap_sem.
1663          *
1664          * ... and hb->lock. --ANK
1665          */
1666         spin_unlock(&hb->lock);
1667
1668         if (attempt++) {
1669                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1670                                          attempt);
1671                 if (ret)
1672                         goto out;
1673                 goto retry_unlocked;
1674         }
1675
1676         if (fshared)
1677                 up_read(fshared);
1678
1679         ret = get_user(uval, uaddr);
1680         if (!ret && (uval != -EFAULT))
1681                 goto retry;
1682
1683         return ret;
1684 }
1685
1686 static int futex_close(struct inode *inode, struct file *filp)
1687 {
1688         struct futex_q *q = filp->private_data;
1689
1690         unqueue_me(q);
1691         kfree(q);
1692
1693         return 0;
1694 }
1695
1696 /* This is one-shot: once it's gone off you need a new fd */
1697 static unsigned int futex_poll(struct file *filp,
1698                                struct poll_table_struct *wait)
1699 {
1700         struct futex_q *q = filp->private_data;
1701         int ret = 0;
1702
1703         poll_wait(filp, &q->waiters, wait);
1704
1705         /*
1706          * plist_node_empty() is safe here without any lock.
1707          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1708          */
1709         if (plist_node_empty(&q->list))
1710                 ret = POLLIN | POLLRDNORM;
1711
1712         return ret;
1713 }
1714
1715 static const struct file_operations futex_fops = {
1716         .release        = futex_close,
1717         .poll           = futex_poll,
1718 };
1719
1720 /*
1721  * Signal allows caller to avoid the race which would occur if they
1722  * set the sigio stuff up afterwards.
1723  */
1724 static int futex_fd(u32 __user *uaddr, int signal)
1725 {
1726         struct futex_q *q;
1727         struct file *filp;
1728         int ret, err;
1729         struct rw_semaphore *fshared;
1730         static unsigned long printk_interval;
1731
1732         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1733                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1734                         "will be removed from the kernel in June 2007\n",
1735                         current->comm);
1736         }
1737
1738         ret = -EINVAL;
1739         if (!valid_signal(signal))
1740                 goto out;
1741
1742         ret = get_unused_fd();
1743         if (ret < 0)
1744                 goto out;
1745         filp = get_empty_filp();
1746         if (!filp) {
1747                 put_unused_fd(ret);
1748                 ret = -ENFILE;
1749                 goto out;
1750         }
1751         filp->f_op = &futex_fops;
1752         filp->f_path.mnt = mntget(futex_mnt);
1753         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1754         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1755
1756         if (signal) {
1757                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1758                 if (err < 0) {
1759                         goto error;
1760                 }
1761                 filp->f_owner.signum = signal;
1762         }
1763
1764         q = kmalloc(sizeof(*q), GFP_KERNEL);
1765         if (!q) {
1766                 err = -ENOMEM;
1767                 goto error;
1768         }
1769         q->pi_state = NULL;
1770
1771         fshared = &current->mm->mmap_sem;
1772         down_read(fshared);
1773         err = get_futex_key(uaddr, fshared, &q->key);
1774
1775         if (unlikely(err != 0)) {
1776                 up_read(fshared);
1777                 kfree(q);
1778                 goto error;
1779         }
1780
1781         /*
1782          * queue_me() must be called before releasing mmap_sem, because
1783          * key->shared.inode needs to be referenced while holding it.
1784          */
1785         filp->private_data = q;
1786
1787         queue_me(q, ret, filp);
1788         up_read(fshared);
1789
1790         /* Now we map fd to filp, so userspace can access it */
1791         fd_install(ret, filp);
1792 out:
1793         return ret;
1794 error:
1795         put_unused_fd(ret);
1796         put_filp(filp);
1797         ret = err;
1798         goto out;
1799 }
1800
1801 /*
1802  * Support for robust futexes: the kernel cleans up held futexes at
1803  * thread exit time.
1804  *
1805  * Implementation: user-space maintains a per-thread list of locks it
1806  * is holding. Upon do_exit(), the kernel carefully walks this list,
1807  * and marks all locks that are owned by this thread with the
1808  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1809  * always manipulated with the lock held, so the list is private and
1810  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1811  * field, to allow the kernel to clean up if the thread dies after
1812  * acquiring the lock, but just before it could have added itself to
1813  * the list. There can only be one such pending lock.
1814  */
1815
1816 /**
1817  * sys_set_robust_list - set the robust-futex list head of a task
1818  * @head: pointer to the list-head
1819  * @len: length of the list-head, as userspace expects
1820  */
1821 asmlinkage long
1822 sys_set_robust_list(struct robust_list_head __user *head,
1823                     size_t len)
1824 {
1825         /*
1826          * The kernel knows only one size for now:
1827          */
1828         if (unlikely(len != sizeof(*head)))
1829                 return -EINVAL;
1830
1831         current->robust_list = head;
1832
1833         return 0;
1834 }
1835
1836 /**
1837  * sys_get_robust_list - get the robust-futex list head of a task
1838  * @pid: pid of the process [zero for current task]
1839  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1840  * @len_ptr: pointer to a length field, the kernel fills in the header size
1841  */
1842 asmlinkage long
1843 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1844                     size_t __user *len_ptr)
1845 {
1846         struct robust_list_head __user *head;
1847         unsigned long ret;
1848
1849         if (!pid)
1850                 head = current->robust_list;
1851         else {
1852                 struct task_struct *p;
1853
1854                 ret = -ESRCH;
1855                 rcu_read_lock();
1856                 p = find_task_by_pid(pid);
1857                 if (!p)
1858                         goto err_unlock;
1859                 ret = -EPERM;
1860                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1861                                 !capable(CAP_SYS_PTRACE))
1862                         goto err_unlock;
1863                 head = p->robust_list;
1864                 rcu_read_unlock();
1865         }
1866
1867         if (put_user(sizeof(*head), len_ptr))
1868                 return -EFAULT;
1869         return put_user(head, head_ptr);
1870
1871 err_unlock:
1872         rcu_read_unlock();
1873
1874         return ret;
1875 }
1876
1877 /*
1878  * Process a futex-list entry, check whether it's owned by the
1879  * dying task, and do notification if so:
1880  */
1881 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1882 {
1883         u32 uval, nval, mval;
1884
1885 retry:
1886         if (get_user(uval, uaddr))
1887                 return -1;
1888
1889         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1890                 /*
1891                  * Ok, this dying thread is truly holding a futex
1892                  * of interest. Set the OWNER_DIED bit atomically
1893                  * via cmpxchg, and if the value had FUTEX_WAITERS
1894                  * set, wake up a waiter (if any). (We have to do a
1895                  * futex_wake() even if OWNER_DIED is already set -
1896                  * to handle the rare but possible case of recursive
1897                  * thread-death.) The rest of the cleanup is done in
1898                  * userspace.
1899                  */
1900                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1901                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1902
1903                 if (nval == -EFAULT)
1904                         return -1;
1905
1906                 if (nval != uval)
1907                         goto retry;
1908
1909                 /*
1910                  * Wake robust non-PI futexes here. The wakeup of
1911                  * PI futexes happens in exit_pi_state():
1912                  */
1913                 if (!pi) {
1914                         if (uval & FUTEX_WAITERS)
1915                                 futex_wake(uaddr, &curr->mm->mmap_sem, 1);
1916                 }
1917         }
1918         return 0;
1919 }
1920
1921 /*
1922  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1923  */
1924 static inline int fetch_robust_entry(struct robust_list __user **entry,
1925                                      struct robust_list __user * __user *head,
1926                                      int *pi)
1927 {
1928         unsigned long uentry;
1929
1930         if (get_user(uentry, (unsigned long __user *)head))
1931                 return -EFAULT;
1932
1933         *entry = (void __user *)(uentry & ~1UL);
1934         *pi = uentry & 1;
1935
1936         return 0;
1937 }
1938
1939 /*
1940  * Walk curr->robust_list (very carefully, it's a userspace list!)
1941  * and mark any locks found there dead, and notify any waiters.
1942  *
1943  * We silently return on any sign of list-walking problem.
1944  */
1945 void exit_robust_list(struct task_struct *curr)
1946 {
1947         struct robust_list_head __user *head = curr->robust_list;
1948         struct robust_list __user *entry, *pending;
1949         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1950         unsigned long futex_offset;
1951
1952         /*
1953          * Fetch the list head (which was registered earlier, via
1954          * sys_set_robust_list()):
1955          */
1956         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1957                 return;
1958         /*
1959          * Fetch the relative futex offset:
1960          */
1961         if (get_user(futex_offset, &head->futex_offset))
1962                 return;
1963         /*
1964          * Fetch any possibly pending lock-add first, and handle it
1965          * if it exists:
1966          */
1967         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1968                 return;
1969
1970         if (pending)
1971                 handle_futex_death((void __user *)pending + futex_offset,
1972                                    curr, pip);
1973
1974         while (entry != &head->list) {
1975                 /*
1976                  * A pending lock might already be on the list, so
1977                  * don't process it twice:
1978                  */
1979                 if (entry != pending)
1980                         if (handle_futex_death((void __user *)entry + futex_offset,
1981                                                 curr, pi))
1982                                 return;
1983                 /*
1984                  * Fetch the next entry in the list:
1985                  */
1986                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1987                         return;
1988                 /*
1989                  * Avoid excessively long or circular lists:
1990                  */
1991                 if (!--limit)
1992                         break;
1993
1994                 cond_resched();
1995         }
1996 }
1997
1998 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1999                 u32 __user *uaddr2, u32 val2, u32 val3)
2000 {
2001         int ret;
2002         int cmd = op & FUTEX_CMD_MASK;
2003         struct rw_semaphore *fshared = NULL;
2004
2005         if (!(op & FUTEX_PRIVATE_FLAG))
2006                 fshared = &current->mm->mmap_sem;
2007
2008         switch (cmd) {
2009         case FUTEX_WAIT:
2010                 ret = futex_wait(uaddr, fshared, val, timeout);
2011                 break;
2012         case FUTEX_WAKE:
2013                 ret = futex_wake(uaddr, fshared, val);
2014                 break;
2015         case FUTEX_FD:
2016                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
2017                 ret = futex_fd(uaddr, val);
2018                 break;
2019         case FUTEX_REQUEUE:
2020                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2021                 break;
2022         case FUTEX_CMP_REQUEUE:
2023                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2024                 break;
2025         case FUTEX_WAKE_OP:
2026                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2027                 break;
2028         case FUTEX_LOCK_PI:
2029                 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2030                 break;
2031         case FUTEX_UNLOCK_PI:
2032                 ret = futex_unlock_pi(uaddr, fshared);
2033                 break;
2034         case FUTEX_TRYLOCK_PI:
2035                 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2036                 break;
2037         default:
2038                 ret = -ENOSYS;
2039         }
2040         return ret;
2041 }
2042
2043
2044 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2045                           struct timespec __user *utime, u32 __user *uaddr2,
2046                           u32 val3)
2047 {
2048         struct timespec ts;
2049         ktime_t t, *tp = NULL;
2050         u32 val2 = 0;
2051         int cmd = op & FUTEX_CMD_MASK;
2052
2053         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
2054                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2055                         return -EFAULT;
2056                 if (!timespec_valid(&ts))
2057                         return -EINVAL;
2058
2059                 t = timespec_to_ktime(ts);
2060                 if (cmd == FUTEX_WAIT)
2061                         t = ktime_add(ktime_get(), t);
2062                 tp = &t;
2063         }
2064         /*
2065          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2066          */
2067         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE)
2068                 val2 = (u32) (unsigned long) utime;
2069
2070         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2071 }
2072
2073 static int futexfs_get_sb(struct file_system_type *fs_type,
2074                           int flags, const char *dev_name, void *data,
2075                           struct vfsmount *mnt)
2076 {
2077         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
2078 }
2079
2080 static struct file_system_type futex_fs_type = {
2081         .name           = "futexfs",
2082         .get_sb         = futexfs_get_sb,
2083         .kill_sb        = kill_anon_super,
2084 };
2085
2086 static int __init init(void)
2087 {
2088         int i = register_filesystem(&futex_fs_type);
2089
2090         if (i)
2091                 return i;
2092
2093         futex_mnt = kern_mount(&futex_fs_type);
2094         if (IS_ERR(futex_mnt)) {
2095                 unregister_filesystem(&futex_fs_type);
2096                 return PTR_ERR(futex_mnt);
2097         }
2098
2099         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2100                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2101                 spin_lock_init(&futex_queues[i].lock);
2102         }
2103         return 0;
2104 }
2105 __initcall(init);