Merge branch 'upstream-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/linvil...
[linux-2.6] / 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  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
20  *  enough at me, Linus for the original (flawed) idea, Matthew
21  *  Kirkwood for proof-of-concept implementation.
22  *
23  *  "The futexes are also cursed."
24  *  "But they come in a choice of three flavours!"
25  *
26  *  This program is free software; you can redistribute it and/or modify
27  *  it under the terms of the GNU General Public License as published by
28  *  the Free Software Foundation; either version 2 of the License, or
29  *  (at your option) any later version.
30  *
31  *  This program is distributed in the hope that it will be useful,
32  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
33  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
34  *  GNU General Public License for more details.
35  *
36  *  You should have received a copy of the GNU General Public License
37  *  along with this program; if not, write to the Free Software
38  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
39  */
40 #include <linux/slab.h>
41 #include <linux/poll.h>
42 #include <linux/fs.h>
43 #include <linux/file.h>
44 #include <linux/jhash.h>
45 #include <linux/init.h>
46 #include <linux/futex.h>
47 #include <linux/mount.h>
48 #include <linux/pagemap.h>
49 #include <linux/syscalls.h>
50 #include <linux/signal.h>
51 #include <asm/futex.h>
52
53 #include "rtmutex_common.h"
54
55 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
56
57 /*
58  * Futexes are matched on equal values of this key.
59  * The key type depends on whether it's a shared or private mapping.
60  * Don't rearrange members without looking at hash_futex().
61  *
62  * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
63  * We set bit 0 to indicate if it's an inode-based key.
64  */
65 union futex_key {
66         struct {
67                 unsigned long pgoff;
68                 struct inode *inode;
69                 int offset;
70         } shared;
71         struct {
72                 unsigned long address;
73                 struct mm_struct *mm;
74                 int offset;
75         } private;
76         struct {
77                 unsigned long word;
78                 void *ptr;
79                 int offset;
80         } both;
81 };
82
83 /*
84  * Priority Inheritance state:
85  */
86 struct futex_pi_state {
87         /*
88          * list of 'owned' pi_state instances - these have to be
89          * cleaned up in do_exit() if the task exits prematurely:
90          */
91         struct list_head list;
92
93         /*
94          * The PI object:
95          */
96         struct rt_mutex pi_mutex;
97
98         struct task_struct *owner;
99         atomic_t refcount;
100
101         union futex_key key;
102 };
103
104 /*
105  * We use this hashed waitqueue instead of a normal wait_queue_t, so
106  * we can wake only the relevant ones (hashed queues may be shared).
107  *
108  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
109  * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0.
110  * The order of wakup is always to make the first condition true, then
111  * wake up q->waiters, then make the second condition true.
112  */
113 struct futex_q {
114         struct list_head list;
115         wait_queue_head_t waiters;
116
117         /* Which hash list lock to use: */
118         spinlock_t *lock_ptr;
119
120         /* Key which the futex is hashed on: */
121         union futex_key key;
122
123         /* For fd, sigio sent using these: */
124         int fd;
125         struct file *filp;
126
127         /* Optional priority inheritance state: */
128         struct futex_pi_state *pi_state;
129         struct task_struct *task;
130 };
131
132 /*
133  * Split the global futex_lock into every hash list lock.
134  */
135 struct futex_hash_bucket {
136        spinlock_t              lock;
137        struct list_head       chain;
138 };
139
140 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
141
142 /* Futex-fs vfsmount entry: */
143 static struct vfsmount *futex_mnt;
144
145 /*
146  * We hash on the keys returned from get_futex_key (see below).
147  */
148 static struct futex_hash_bucket *hash_futex(union futex_key *key)
149 {
150         u32 hash = jhash2((u32*)&key->both.word,
151                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
152                           key->both.offset);
153         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
154 }
155
156 /*
157  * Return 1 if two futex_keys are equal, 0 otherwise.
158  */
159 static inline int match_futex(union futex_key *key1, union futex_key *key2)
160 {
161         return (key1->both.word == key2->both.word
162                 && key1->both.ptr == key2->both.ptr
163                 && key1->both.offset == key2->both.offset);
164 }
165
166 /*
167  * Get parameters which are the keys for a futex.
168  *
169  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
170  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
171  * We can usually work out the index without swapping in the page.
172  *
173  * Returns: 0, or negative error code.
174  * The key words are stored in *key on success.
175  *
176  * Should be called with &current->mm->mmap_sem but NOT any spinlocks.
177  */
178 static int get_futex_key(u32 __user *uaddr, union futex_key *key)
179 {
180         unsigned long address = (unsigned long)uaddr;
181         struct mm_struct *mm = current->mm;
182         struct vm_area_struct *vma;
183         struct page *page;
184         int err;
185
186         /*
187          * The futex address must be "naturally" aligned.
188          */
189         key->both.offset = address % PAGE_SIZE;
190         if (unlikely((key->both.offset % sizeof(u32)) != 0))
191                 return -EINVAL;
192         address -= key->both.offset;
193
194         /*
195          * The futex is hashed differently depending on whether
196          * it's in a shared or private mapping.  So check vma first.
197          */
198         vma = find_extend_vma(mm, address);
199         if (unlikely(!vma))
200                 return -EFAULT;
201
202         /*
203          * Permissions.
204          */
205         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
206                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
207
208         /*
209          * Private mappings are handled in a simple way.
210          *
211          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
212          * it's a read-only handle, it's expected that futexes attach to
213          * the object not the particular process.  Therefore we use
214          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
215          * mappings of _writable_ handles.
216          */
217         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
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++; /* Bit 0 of offset indicates 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
250 /*
251  * Take a reference to the resource addressed by a key.
252  * Can be called while holding spinlocks.
253  *
254  * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
255  * function, if it is called at all.  mmap_sem keeps key->shared.inode valid.
256  */
257 static inline void get_key_refs(union futex_key *key)
258 {
259         if (key->both.ptr != 0) {
260                 if (key->both.offset & 1)
261                         atomic_inc(&key->shared.inode->i_count);
262                 else
263                         atomic_inc(&key->private.mm->mm_count);
264         }
265 }
266
267 /*
268  * Drop a reference to the resource addressed by a key.
269  * The hash bucket spinlock must not be held.
270  */
271 static void drop_key_refs(union futex_key *key)
272 {
273         if (key->both.ptr != 0) {
274                 if (key->both.offset & 1)
275                         iput(key->shared.inode);
276                 else
277                         mmdrop(key->private.mm);
278         }
279 }
280
281 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
282 {
283         int ret;
284
285         pagefault_disable();
286         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
287         pagefault_enable();
288
289         return ret ? -EFAULT : 0;
290 }
291
292 /*
293  * Fault handling. Called with current->mm->mmap_sem held.
294  */
295 static int futex_handle_fault(unsigned long address, int attempt)
296 {
297         struct vm_area_struct * vma;
298         struct mm_struct *mm = current->mm;
299
300         if (attempt > 2 || !(vma = find_vma(mm, address)) ||
301             vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
302                 return -EFAULT;
303
304         switch (handle_mm_fault(mm, vma, address, 1)) {
305         case VM_FAULT_MINOR:
306                 current->min_flt++;
307                 break;
308         case VM_FAULT_MAJOR:
309                 current->maj_flt++;
310                 break;
311         default:
312                 return -EFAULT;
313         }
314         return 0;
315 }
316
317 /*
318  * PI code:
319  */
320 static int refill_pi_state_cache(void)
321 {
322         struct futex_pi_state *pi_state;
323
324         if (likely(current->pi_state_cache))
325                 return 0;
326
327         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
328
329         if (!pi_state)
330                 return -ENOMEM;
331
332         INIT_LIST_HEAD(&pi_state->list);
333         /* pi_mutex gets initialized later */
334         pi_state->owner = NULL;
335         atomic_set(&pi_state->refcount, 1);
336
337         current->pi_state_cache = pi_state;
338
339         return 0;
340 }
341
342 static struct futex_pi_state * alloc_pi_state(void)
343 {
344         struct futex_pi_state *pi_state = current->pi_state_cache;
345
346         WARN_ON(!pi_state);
347         current->pi_state_cache = NULL;
348
349         return pi_state;
350 }
351
352 static void free_pi_state(struct futex_pi_state *pi_state)
353 {
354         if (!atomic_dec_and_test(&pi_state->refcount))
355                 return;
356
357         /*
358          * If pi_state->owner is NULL, the owner is most probably dying
359          * and has cleaned up the pi_state already
360          */
361         if (pi_state->owner) {
362                 spin_lock_irq(&pi_state->owner->pi_lock);
363                 list_del_init(&pi_state->list);
364                 spin_unlock_irq(&pi_state->owner->pi_lock);
365
366                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
367         }
368
369         if (current->pi_state_cache)
370                 kfree(pi_state);
371         else {
372                 /*
373                  * pi_state->list is already empty.
374                  * clear pi_state->owner.
375                  * refcount is at 0 - put it back to 1.
376                  */
377                 pi_state->owner = NULL;
378                 atomic_set(&pi_state->refcount, 1);
379                 current->pi_state_cache = pi_state;
380         }
381 }
382
383 /*
384  * Look up the task based on what TID userspace gave us.
385  * We dont trust it.
386  */
387 static struct task_struct * futex_find_get_task(pid_t pid)
388 {
389         struct task_struct *p;
390
391         rcu_read_lock();
392         p = find_task_by_pid(pid);
393         if (!p)
394                 goto out_unlock;
395         if ((current->euid != p->euid) && (current->euid != p->uid)) {
396                 p = NULL;
397                 goto out_unlock;
398         }
399         if (p->exit_state != 0) {
400                 p = NULL;
401                 goto out_unlock;
402         }
403         get_task_struct(p);
404 out_unlock:
405         rcu_read_unlock();
406
407         return p;
408 }
409
410 /*
411  * This task is holding PI mutexes at exit time => bad.
412  * Kernel cleans up PI-state, but userspace is likely hosed.
413  * (Robust-futex cleanup is separate and might save the day for userspace.)
414  */
415 void exit_pi_state_list(struct task_struct *curr)
416 {
417         struct list_head *next, *head = &curr->pi_state_list;
418         struct futex_pi_state *pi_state;
419         struct futex_hash_bucket *hb;
420         union futex_key key;
421
422         /*
423          * We are a ZOMBIE and nobody can enqueue itself on
424          * pi_state_list anymore, but we have to be careful
425          * versus waiters unqueueing themselves:
426          */
427         spin_lock_irq(&curr->pi_lock);
428         while (!list_empty(head)) {
429
430                 next = head->next;
431                 pi_state = list_entry(next, struct futex_pi_state, list);
432                 key = pi_state->key;
433                 hb = hash_futex(&key);
434                 spin_unlock_irq(&curr->pi_lock);
435
436                 spin_lock(&hb->lock);
437
438                 spin_lock_irq(&curr->pi_lock);
439                 /*
440                  * We dropped the pi-lock, so re-check whether this
441                  * task still owns the PI-state:
442                  */
443                 if (head->next != next) {
444                         spin_unlock(&hb->lock);
445                         continue;
446                 }
447
448                 WARN_ON(pi_state->owner != curr);
449                 WARN_ON(list_empty(&pi_state->list));
450                 list_del_init(&pi_state->list);
451                 pi_state->owner = NULL;
452                 spin_unlock_irq(&curr->pi_lock);
453
454                 rt_mutex_unlock(&pi_state->pi_mutex);
455
456                 spin_unlock(&hb->lock);
457
458                 spin_lock_irq(&curr->pi_lock);
459         }
460         spin_unlock_irq(&curr->pi_lock);
461 }
462
463 static int
464 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
465 {
466         struct futex_pi_state *pi_state = NULL;
467         struct futex_q *this, *next;
468         struct list_head *head;
469         struct task_struct *p;
470         pid_t pid;
471
472         head = &hb->chain;
473
474         list_for_each_entry_safe(this, next, head, list) {
475                 if (match_futex(&this->key, &me->key)) {
476                         /*
477                          * Another waiter already exists - bump up
478                          * the refcount and return its pi_state:
479                          */
480                         pi_state = this->pi_state;
481                         /*
482                          * Userspace might have messed up non PI and PI futexes
483                          */
484                         if (unlikely(!pi_state))
485                                 return -EINVAL;
486
487                         WARN_ON(!atomic_read(&pi_state->refcount));
488
489                         atomic_inc(&pi_state->refcount);
490                         me->pi_state = pi_state;
491
492                         return 0;
493                 }
494         }
495
496         /*
497          * We are the first waiter - try to look up the real owner and attach
498          * the new pi_state to it, but bail out when the owner died bit is set
499          * and TID = 0:
500          */
501         pid = uval & FUTEX_TID_MASK;
502         if (!pid && (uval & FUTEX_OWNER_DIED))
503                 return -ESRCH;
504         p = futex_find_get_task(pid);
505         if (!p)
506                 return -ESRCH;
507
508         pi_state = alloc_pi_state();
509
510         /*
511          * Initialize the pi_mutex in locked state and make 'p'
512          * the owner of it:
513          */
514         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
515
516         /* Store the key for possible exit cleanups: */
517         pi_state->key = me->key;
518
519         spin_lock_irq(&p->pi_lock);
520         WARN_ON(!list_empty(&pi_state->list));
521         list_add(&pi_state->list, &p->pi_state_list);
522         pi_state->owner = p;
523         spin_unlock_irq(&p->pi_lock);
524
525         put_task_struct(p);
526
527         me->pi_state = pi_state;
528
529         return 0;
530 }
531
532 /*
533  * The hash bucket lock must be held when this is called.
534  * Afterwards, the futex_q must not be accessed.
535  */
536 static void wake_futex(struct futex_q *q)
537 {
538         list_del_init(&q->list);
539         if (q->filp)
540                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
541         /*
542          * The lock in wake_up_all() is a crucial memory barrier after the
543          * list_del_init() and also before assigning to q->lock_ptr.
544          */
545         wake_up_all(&q->waiters);
546         /*
547          * The waiting task can free the futex_q as soon as this is written,
548          * without taking any locks.  This must come last.
549          *
550          * A memory barrier is required here to prevent the following store
551          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
552          * at the end of wake_up_all() does not prevent this store from
553          * moving.
554          */
555         smp_wmb();
556         q->lock_ptr = NULL;
557 }
558
559 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
560 {
561         struct task_struct *new_owner;
562         struct futex_pi_state *pi_state = this->pi_state;
563         u32 curval, newval;
564
565         if (!pi_state)
566                 return -EINVAL;
567
568         spin_lock(&pi_state->pi_mutex.wait_lock);
569         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
570
571         /*
572          * This happens when we have stolen the lock and the original
573          * pending owner did not enqueue itself back on the rt_mutex.
574          * Thats not a tragedy. We know that way, that a lock waiter
575          * is on the fly. We make the futex_q waiter the pending owner.
576          */
577         if (!new_owner)
578                 new_owner = this->task;
579
580         /*
581          * We pass it to the next owner. (The WAITERS bit is always
582          * kept enabled while there is PI state around. We must also
583          * preserve the owner died bit.)
584          */
585         if (!(uval & FUTEX_OWNER_DIED)) {
586                 newval = FUTEX_WAITERS | new_owner->pid;
587
588                 pagefault_disable();
589                 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
590                 pagefault_enable();
591                 if (curval == -EFAULT)
592                         return -EFAULT;
593                 if (curval != uval)
594                         return -EINVAL;
595         }
596
597         spin_lock_irq(&pi_state->owner->pi_lock);
598         WARN_ON(list_empty(&pi_state->list));
599         list_del_init(&pi_state->list);
600         spin_unlock_irq(&pi_state->owner->pi_lock);
601
602         spin_lock_irq(&new_owner->pi_lock);
603         WARN_ON(!list_empty(&pi_state->list));
604         list_add(&pi_state->list, &new_owner->pi_state_list);
605         pi_state->owner = new_owner;
606         spin_unlock_irq(&new_owner->pi_lock);
607
608         spin_unlock(&pi_state->pi_mutex.wait_lock);
609         rt_mutex_unlock(&pi_state->pi_mutex);
610
611         return 0;
612 }
613
614 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
615 {
616         u32 oldval;
617
618         /*
619          * There is no waiter, so we unlock the futex. The owner died
620          * bit has not to be preserved here. We are the owner:
621          */
622         pagefault_disable();
623         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
624         pagefault_enable();
625
626         if (oldval == -EFAULT)
627                 return oldval;
628         if (oldval != uval)
629                 return -EAGAIN;
630
631         return 0;
632 }
633
634 /*
635  * Express the locking dependencies for lockdep:
636  */
637 static inline void
638 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
639 {
640         if (hb1 <= hb2) {
641                 spin_lock(&hb1->lock);
642                 if (hb1 < hb2)
643                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
644         } else { /* hb1 > hb2 */
645                 spin_lock(&hb2->lock);
646                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
647         }
648 }
649
650 /*
651  * Wake up all waiters hashed on the physical page that is mapped
652  * to this virtual address:
653  */
654 static int futex_wake(u32 __user *uaddr, int nr_wake)
655 {
656         struct futex_hash_bucket *hb;
657         struct futex_q *this, *next;
658         struct list_head *head;
659         union futex_key key;
660         int ret;
661
662         down_read(&current->mm->mmap_sem);
663
664         ret = get_futex_key(uaddr, &key);
665         if (unlikely(ret != 0))
666                 goto out;
667
668         hb = hash_futex(&key);
669         spin_lock(&hb->lock);
670         head = &hb->chain;
671
672         list_for_each_entry_safe(this, next, head, list) {
673                 if (match_futex (&this->key, &key)) {
674                         if (this->pi_state) {
675                                 ret = -EINVAL;
676                                 break;
677                         }
678                         wake_futex(this);
679                         if (++ret >= nr_wake)
680                                 break;
681                 }
682         }
683
684         spin_unlock(&hb->lock);
685 out:
686         up_read(&current->mm->mmap_sem);
687         return ret;
688 }
689
690 /*
691  * Wake up all waiters hashed on the physical page that is mapped
692  * to this virtual address:
693  */
694 static int
695 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
696               int nr_wake, int nr_wake2, int op)
697 {
698         union futex_key key1, key2;
699         struct futex_hash_bucket *hb1, *hb2;
700         struct list_head *head;
701         struct futex_q *this, *next;
702         int ret, op_ret, attempt = 0;
703
704 retryfull:
705         down_read(&current->mm->mmap_sem);
706
707         ret = get_futex_key(uaddr1, &key1);
708         if (unlikely(ret != 0))
709                 goto out;
710         ret = get_futex_key(uaddr2, &key2);
711         if (unlikely(ret != 0))
712                 goto out;
713
714         hb1 = hash_futex(&key1);
715         hb2 = hash_futex(&key2);
716
717 retry:
718         double_lock_hb(hb1, hb2);
719
720         op_ret = futex_atomic_op_inuser(op, uaddr2);
721         if (unlikely(op_ret < 0)) {
722                 u32 dummy;
723
724                 spin_unlock(&hb1->lock);
725                 if (hb1 != hb2)
726                         spin_unlock(&hb2->lock);
727
728 #ifndef CONFIG_MMU
729                 /*
730                  * we don't get EFAULT from MMU faults if we don't have an MMU,
731                  * but we might get them from range checking
732                  */
733                 ret = op_ret;
734                 goto out;
735 #endif
736
737                 if (unlikely(op_ret != -EFAULT)) {
738                         ret = op_ret;
739                         goto out;
740                 }
741
742                 /*
743                  * futex_atomic_op_inuser needs to both read and write
744                  * *(int __user *)uaddr2, but we can't modify it
745                  * non-atomically.  Therefore, if get_user below is not
746                  * enough, we need to handle the fault ourselves, while
747                  * still holding the mmap_sem.
748                  */
749                 if (attempt++) {
750                         if (futex_handle_fault((unsigned long)uaddr2,
751                                                 attempt)) {
752                                 ret = -EFAULT;
753                                 goto out;
754                         }
755                         goto retry;
756                 }
757
758                 /*
759                  * If we would have faulted, release mmap_sem,
760                  * fault it in and start all over again.
761                  */
762                 up_read(&current->mm->mmap_sem);
763
764                 ret = get_user(dummy, uaddr2);
765                 if (ret)
766                         return ret;
767
768                 goto retryfull;
769         }
770
771         head = &hb1->chain;
772
773         list_for_each_entry_safe(this, next, head, list) {
774                 if (match_futex (&this->key, &key1)) {
775                         wake_futex(this);
776                         if (++ret >= nr_wake)
777                                 break;
778                 }
779         }
780
781         if (op_ret > 0) {
782                 head = &hb2->chain;
783
784                 op_ret = 0;
785                 list_for_each_entry_safe(this, next, head, list) {
786                         if (match_futex (&this->key, &key2)) {
787                                 wake_futex(this);
788                                 if (++op_ret >= nr_wake2)
789                                         break;
790                         }
791                 }
792                 ret += op_ret;
793         }
794
795         spin_unlock(&hb1->lock);
796         if (hb1 != hb2)
797                 spin_unlock(&hb2->lock);
798 out:
799         up_read(&current->mm->mmap_sem);
800         return ret;
801 }
802
803 /*
804  * Requeue all waiters hashed on one physical page to another
805  * physical page.
806  */
807 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
808                          int nr_wake, int nr_requeue, u32 *cmpval)
809 {
810         union futex_key key1, key2;
811         struct futex_hash_bucket *hb1, *hb2;
812         struct list_head *head1;
813         struct futex_q *this, *next;
814         int ret, drop_count = 0;
815
816  retry:
817         down_read(&current->mm->mmap_sem);
818
819         ret = get_futex_key(uaddr1, &key1);
820         if (unlikely(ret != 0))
821                 goto out;
822         ret = get_futex_key(uaddr2, &key2);
823         if (unlikely(ret != 0))
824                 goto out;
825
826         hb1 = hash_futex(&key1);
827         hb2 = hash_futex(&key2);
828
829         double_lock_hb(hb1, hb2);
830
831         if (likely(cmpval != NULL)) {
832                 u32 curval;
833
834                 ret = get_futex_value_locked(&curval, uaddr1);
835
836                 if (unlikely(ret)) {
837                         spin_unlock(&hb1->lock);
838                         if (hb1 != hb2)
839                                 spin_unlock(&hb2->lock);
840
841                         /*
842                          * If we would have faulted, release mmap_sem, fault
843                          * it in and start all over again.
844                          */
845                         up_read(&current->mm->mmap_sem);
846
847                         ret = get_user(curval, uaddr1);
848
849                         if (!ret)
850                                 goto retry;
851
852                         return ret;
853                 }
854                 if (curval != *cmpval) {
855                         ret = -EAGAIN;
856                         goto out_unlock;
857                 }
858         }
859
860         head1 = &hb1->chain;
861         list_for_each_entry_safe(this, next, head1, list) {
862                 if (!match_futex (&this->key, &key1))
863                         continue;
864                 if (++ret <= nr_wake) {
865                         wake_futex(this);
866                 } else {
867                         /*
868                          * If key1 and key2 hash to the same bucket, no need to
869                          * requeue.
870                          */
871                         if (likely(head1 != &hb2->chain)) {
872                                 list_move_tail(&this->list, &hb2->chain);
873                                 this->lock_ptr = &hb2->lock;
874                         }
875                         this->key = key2;
876                         get_key_refs(&key2);
877                         drop_count++;
878
879                         if (ret - nr_wake >= nr_requeue)
880                                 break;
881                 }
882         }
883
884 out_unlock:
885         spin_unlock(&hb1->lock);
886         if (hb1 != hb2)
887                 spin_unlock(&hb2->lock);
888
889         /* drop_key_refs() must be called outside the spinlocks. */
890         while (--drop_count >= 0)
891                 drop_key_refs(&key1);
892
893 out:
894         up_read(&current->mm->mmap_sem);
895         return ret;
896 }
897
898 /* The key must be already stored in q->key. */
899 static inline struct futex_hash_bucket *
900 queue_lock(struct futex_q *q, int fd, struct file *filp)
901 {
902         struct futex_hash_bucket *hb;
903
904         q->fd = fd;
905         q->filp = filp;
906
907         init_waitqueue_head(&q->waiters);
908
909         get_key_refs(&q->key);
910         hb = hash_futex(&q->key);
911         q->lock_ptr = &hb->lock;
912
913         spin_lock(&hb->lock);
914         return hb;
915 }
916
917 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
918 {
919         list_add_tail(&q->list, &hb->chain);
920         q->task = current;
921         spin_unlock(&hb->lock);
922 }
923
924 static inline void
925 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
926 {
927         spin_unlock(&hb->lock);
928         drop_key_refs(&q->key);
929 }
930
931 /*
932  * queue_me and unqueue_me must be called as a pair, each
933  * exactly once.  They are called with the hashed spinlock held.
934  */
935
936 /* The key must be already stored in q->key. */
937 static void queue_me(struct futex_q *q, int fd, struct file *filp)
938 {
939         struct futex_hash_bucket *hb;
940
941         hb = queue_lock(q, fd, filp);
942         __queue_me(q, hb);
943 }
944
945 /* Return 1 if we were still queued (ie. 0 means we were woken) */
946 static int unqueue_me(struct futex_q *q)
947 {
948         spinlock_t *lock_ptr;
949         int ret = 0;
950
951         /* In the common case we don't take the spinlock, which is nice. */
952  retry:
953         lock_ptr = q->lock_ptr;
954         barrier();
955         if (lock_ptr != 0) {
956                 spin_lock(lock_ptr);
957                 /*
958                  * q->lock_ptr can change between reading it and
959                  * spin_lock(), causing us to take the wrong lock.  This
960                  * corrects the race condition.
961                  *
962                  * Reasoning goes like this: if we have the wrong lock,
963                  * q->lock_ptr must have changed (maybe several times)
964                  * between reading it and the spin_lock().  It can
965                  * change again after the spin_lock() but only if it was
966                  * already changed before the spin_lock().  It cannot,
967                  * however, change back to the original value.  Therefore
968                  * we can detect whether we acquired the correct lock.
969                  */
970                 if (unlikely(lock_ptr != q->lock_ptr)) {
971                         spin_unlock(lock_ptr);
972                         goto retry;
973                 }
974                 WARN_ON(list_empty(&q->list));
975                 list_del(&q->list);
976
977                 BUG_ON(q->pi_state);
978
979                 spin_unlock(lock_ptr);
980                 ret = 1;
981         }
982
983         drop_key_refs(&q->key);
984         return ret;
985 }
986
987 /*
988  * PI futexes can not be requeued and must remove themself from the
989  * hash bucket. The hash bucket lock is held on entry and dropped here.
990  */
991 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
992 {
993         WARN_ON(list_empty(&q->list));
994         list_del(&q->list);
995
996         BUG_ON(!q->pi_state);
997         free_pi_state(q->pi_state);
998         q->pi_state = NULL;
999
1000         spin_unlock(&hb->lock);
1001
1002         drop_key_refs(&q->key);
1003 }
1004
1005 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
1006 {
1007         struct task_struct *curr = current;
1008         DECLARE_WAITQUEUE(wait, curr);
1009         struct futex_hash_bucket *hb;
1010         struct futex_q q;
1011         u32 uval;
1012         int ret;
1013
1014         q.pi_state = NULL;
1015  retry:
1016         down_read(&curr->mm->mmap_sem);
1017
1018         ret = get_futex_key(uaddr, &q.key);
1019         if (unlikely(ret != 0))
1020                 goto out_release_sem;
1021
1022         hb = queue_lock(&q, -1, NULL);
1023
1024         /*
1025          * Access the page AFTER the futex is queued.
1026          * Order is important:
1027          *
1028          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1029          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1030          *
1031          * The basic logical guarantee of a futex is that it blocks ONLY
1032          * if cond(var) is known to be true at the time of blocking, for
1033          * any cond.  If we queued after testing *uaddr, that would open
1034          * a race condition where we could block indefinitely with
1035          * cond(var) false, which would violate the guarantee.
1036          *
1037          * A consequence is that futex_wait() can return zero and absorb
1038          * a wakeup when *uaddr != val on entry to the syscall.  This is
1039          * rare, but normal.
1040          *
1041          * We hold the mmap semaphore, so the mapping cannot have changed
1042          * since we looked it up in get_futex_key.
1043          */
1044         ret = get_futex_value_locked(&uval, uaddr);
1045
1046         if (unlikely(ret)) {
1047                 queue_unlock(&q, hb);
1048
1049                 /*
1050                  * If we would have faulted, release mmap_sem, fault it in and
1051                  * start all over again.
1052                  */
1053                 up_read(&curr->mm->mmap_sem);
1054
1055                 ret = get_user(uval, uaddr);
1056
1057                 if (!ret)
1058                         goto retry;
1059                 return ret;
1060         }
1061         ret = -EWOULDBLOCK;
1062         if (uval != val)
1063                 goto out_unlock_release_sem;
1064
1065         /* Only actually queue if *uaddr contained val.  */
1066         __queue_me(&q, hb);
1067
1068         /*
1069          * Now the futex is queued and we have checked the data, we
1070          * don't want to hold mmap_sem while we sleep.
1071          */
1072         up_read(&curr->mm->mmap_sem);
1073
1074         /*
1075          * There might have been scheduling since the queue_me(), as we
1076          * cannot hold a spinlock across the get_user() in case it
1077          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1078          * queueing ourselves into the futex hash.  This code thus has to
1079          * rely on the futex_wake() code removing us from hash when it
1080          * wakes us up.
1081          */
1082
1083         /* add_wait_queue is the barrier after __set_current_state. */
1084         __set_current_state(TASK_INTERRUPTIBLE);
1085         add_wait_queue(&q.waiters, &wait);
1086         /*
1087          * !list_empty() is safe here without any lock.
1088          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1089          */
1090         if (likely(!list_empty(&q.list)))
1091                 time = schedule_timeout(time);
1092         __set_current_state(TASK_RUNNING);
1093
1094         /*
1095          * NOTE: we don't remove ourselves from the waitqueue because
1096          * we are the only user of it.
1097          */
1098
1099         /* If we were woken (and unqueued), we succeeded, whatever. */
1100         if (!unqueue_me(&q))
1101                 return 0;
1102         if (time == 0)
1103                 return -ETIMEDOUT;
1104         /*
1105          * We expect signal_pending(current), but another thread may
1106          * have handled it for us already.
1107          */
1108         return -EINTR;
1109
1110  out_unlock_release_sem:
1111         queue_unlock(&q, hb);
1112
1113  out_release_sem:
1114         up_read(&curr->mm->mmap_sem);
1115         return ret;
1116 }
1117
1118 /*
1119  * Userspace tried a 0 -> TID atomic transition of the futex value
1120  * and failed. The kernel side here does the whole locking operation:
1121  * if there are waiters then it will block, it does PI, etc. (Due to
1122  * races the kernel might see a 0 value of the futex too.)
1123  */
1124 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1125                          long nsec, int trylock)
1126 {
1127         struct hrtimer_sleeper timeout, *to = NULL;
1128         struct task_struct *curr = current;
1129         struct futex_hash_bucket *hb;
1130         u32 uval, newval, curval;
1131         struct futex_q q;
1132         int ret, attempt = 0;
1133
1134         if (refill_pi_state_cache())
1135                 return -ENOMEM;
1136
1137         if (sec != MAX_SCHEDULE_TIMEOUT) {
1138                 to = &timeout;
1139                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1140                 hrtimer_init_sleeper(to, current);
1141                 to->timer.expires = ktime_set(sec, nsec);
1142         }
1143
1144         q.pi_state = NULL;
1145  retry:
1146         down_read(&curr->mm->mmap_sem);
1147
1148         ret = get_futex_key(uaddr, &q.key);
1149         if (unlikely(ret != 0))
1150                 goto out_release_sem;
1151
1152         hb = queue_lock(&q, -1, NULL);
1153
1154  retry_locked:
1155         /*
1156          * To avoid races, we attempt to take the lock here again
1157          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1158          * the locks. It will most likely not succeed.
1159          */
1160         newval = current->pid;
1161
1162         pagefault_disable();
1163         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1164         pagefault_enable();
1165
1166         if (unlikely(curval == -EFAULT))
1167                 goto uaddr_faulted;
1168
1169         /* We own the lock already */
1170         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1171                 if (!detect && 0)
1172                         force_sig(SIGKILL, current);
1173                 ret = -EDEADLK;
1174                 goto out_unlock_release_sem;
1175         }
1176
1177         /*
1178          * Surprise - we got the lock. Just return
1179          * to userspace:
1180          */
1181         if (unlikely(!curval))
1182                 goto out_unlock_release_sem;
1183
1184         uval = curval;
1185         newval = uval | FUTEX_WAITERS;
1186
1187         pagefault_disable();
1188         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1189         pagefault_enable();
1190
1191         if (unlikely(curval == -EFAULT))
1192                 goto uaddr_faulted;
1193         if (unlikely(curval != uval))
1194                 goto retry_locked;
1195
1196         /*
1197          * We dont have the lock. Look up the PI state (or create it if
1198          * we are the first waiter):
1199          */
1200         ret = lookup_pi_state(uval, hb, &q);
1201
1202         if (unlikely(ret)) {
1203                 /*
1204                  * There were no waiters and the owner task lookup
1205                  * failed. When the OWNER_DIED bit is set, then we
1206                  * know that this is a robust futex and we actually
1207                  * take the lock. This is safe as we are protected by
1208                  * the hash bucket lock. We also set the waiters bit
1209                  * unconditionally here, to simplify glibc handling of
1210                  * multiple tasks racing to acquire the lock and
1211                  * cleanup the problems which were left by the dead
1212                  * owner.
1213                  */
1214                 if (curval & FUTEX_OWNER_DIED) {
1215                         uval = newval;
1216                         newval = current->pid |
1217                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1218
1219                         pagefault_disable();
1220                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1221                                                                uval, newval);
1222                         pagefault_enable();
1223
1224                         if (unlikely(curval == -EFAULT))
1225                                 goto uaddr_faulted;
1226                         if (unlikely(curval != uval))
1227                                 goto retry_locked;
1228                         ret = 0;
1229                 }
1230                 goto out_unlock_release_sem;
1231         }
1232
1233         /*
1234          * Only actually queue now that the atomic ops are done:
1235          */
1236         __queue_me(&q, hb);
1237
1238         /*
1239          * Now the futex is queued and we have checked the data, we
1240          * don't want to hold mmap_sem while we sleep.
1241          */
1242         up_read(&curr->mm->mmap_sem);
1243
1244         WARN_ON(!q.pi_state);
1245         /*
1246          * Block on the PI mutex:
1247          */
1248         if (!trylock)
1249                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1250         else {
1251                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1252                 /* Fixup the trylock return value: */
1253                 ret = ret ? 0 : -EWOULDBLOCK;
1254         }
1255
1256         down_read(&curr->mm->mmap_sem);
1257         spin_lock(q.lock_ptr);
1258
1259         /*
1260          * Got the lock. We might not be the anticipated owner if we
1261          * did a lock-steal - fix up the PI-state in that case.
1262          */
1263         if (!ret && q.pi_state->owner != curr) {
1264                 u32 newtid = current->pid | FUTEX_WAITERS;
1265
1266                 /* Owner died? */
1267                 if (q.pi_state->owner != NULL) {
1268                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1269                         WARN_ON(list_empty(&q.pi_state->list));
1270                         list_del_init(&q.pi_state->list);
1271                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1272                 } else
1273                         newtid |= FUTEX_OWNER_DIED;
1274
1275                 q.pi_state->owner = current;
1276
1277                 spin_lock_irq(&current->pi_lock);
1278                 WARN_ON(!list_empty(&q.pi_state->list));
1279                 list_add(&q.pi_state->list, &current->pi_state_list);
1280                 spin_unlock_irq(&current->pi_lock);
1281
1282                 /* Unqueue and drop the lock */
1283                 unqueue_me_pi(&q, hb);
1284                 up_read(&curr->mm->mmap_sem);
1285                 /*
1286                  * We own it, so we have to replace the pending owner
1287                  * TID. This must be atomic as we have preserve the
1288                  * owner died bit here.
1289                  */
1290                 ret = get_user(uval, uaddr);
1291                 while (!ret) {
1292                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1293                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1294                                                                uval, newval);
1295                         if (curval == -EFAULT)
1296                                 ret = -EFAULT;
1297                         if (curval == uval)
1298                                 break;
1299                         uval = curval;
1300                 }
1301         } else {
1302                 /*
1303                  * Catch the rare case, where the lock was released
1304                  * when we were on the way back before we locked
1305                  * the hash bucket.
1306                  */
1307                 if (ret && q.pi_state->owner == curr) {
1308                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1309                                 ret = 0;
1310                 }
1311                 /* Unqueue and drop the lock */
1312                 unqueue_me_pi(&q, hb);
1313                 up_read(&curr->mm->mmap_sem);
1314         }
1315
1316         if (!detect && ret == -EDEADLK && 0)
1317                 force_sig(SIGKILL, current);
1318
1319         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1320
1321  out_unlock_release_sem:
1322         queue_unlock(&q, hb);
1323
1324  out_release_sem:
1325         up_read(&curr->mm->mmap_sem);
1326         return ret;
1327
1328  uaddr_faulted:
1329         /*
1330          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1331          * non-atomically.  Therefore, if get_user below is not
1332          * enough, we need to handle the fault ourselves, while
1333          * still holding the mmap_sem.
1334          */
1335         if (attempt++) {
1336                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1337                         ret = -EFAULT;
1338                         goto out_unlock_release_sem;
1339                 }
1340                 goto retry_locked;
1341         }
1342
1343         queue_unlock(&q, hb);
1344         up_read(&curr->mm->mmap_sem);
1345
1346         ret = get_user(uval, uaddr);
1347         if (!ret && (uval != -EFAULT))
1348                 goto retry;
1349
1350         return ret;
1351 }
1352
1353 /*
1354  * Userspace attempted a TID -> 0 atomic transition, and failed.
1355  * This is the in-kernel slowpath: we look up the PI state (if any),
1356  * and do the rt-mutex unlock.
1357  */
1358 static int futex_unlock_pi(u32 __user *uaddr)
1359 {
1360         struct futex_hash_bucket *hb;
1361         struct futex_q *this, *next;
1362         u32 uval;
1363         struct list_head *head;
1364         union futex_key key;
1365         int ret, attempt = 0;
1366
1367 retry:
1368         if (get_user(uval, uaddr))
1369                 return -EFAULT;
1370         /*
1371          * We release only a lock we actually own:
1372          */
1373         if ((uval & FUTEX_TID_MASK) != current->pid)
1374                 return -EPERM;
1375         /*
1376          * First take all the futex related locks:
1377          */
1378         down_read(&current->mm->mmap_sem);
1379
1380         ret = get_futex_key(uaddr, &key);
1381         if (unlikely(ret != 0))
1382                 goto out;
1383
1384         hb = hash_futex(&key);
1385         spin_lock(&hb->lock);
1386
1387 retry_locked:
1388         /*
1389          * To avoid races, try to do the TID -> 0 atomic transition
1390          * again. If it succeeds then we can return without waking
1391          * anyone else up:
1392          */
1393         if (!(uval & FUTEX_OWNER_DIED)) {
1394                 pagefault_disable();
1395                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1396                 pagefault_enable();
1397         }
1398
1399         if (unlikely(uval == -EFAULT))
1400                 goto pi_faulted;
1401         /*
1402          * Rare case: we managed to release the lock atomically,
1403          * no need to wake anyone else up:
1404          */
1405         if (unlikely(uval == current->pid))
1406                 goto out_unlock;
1407
1408         /*
1409          * Ok, other tasks may need to be woken up - check waiters
1410          * and do the wakeup if necessary:
1411          */
1412         head = &hb->chain;
1413
1414         list_for_each_entry_safe(this, next, head, list) {
1415                 if (!match_futex (&this->key, &key))
1416                         continue;
1417                 ret = wake_futex_pi(uaddr, uval, this);
1418                 /*
1419                  * The atomic access to the futex value
1420                  * generated a pagefault, so retry the
1421                  * user-access and the wakeup:
1422                  */
1423                 if (ret == -EFAULT)
1424                         goto pi_faulted;
1425                 goto out_unlock;
1426         }
1427         /*
1428          * No waiters - kernel unlocks the futex:
1429          */
1430         if (!(uval & FUTEX_OWNER_DIED)) {
1431                 ret = unlock_futex_pi(uaddr, uval);
1432                 if (ret == -EFAULT)
1433                         goto pi_faulted;
1434         }
1435
1436 out_unlock:
1437         spin_unlock(&hb->lock);
1438 out:
1439         up_read(&current->mm->mmap_sem);
1440
1441         return ret;
1442
1443 pi_faulted:
1444         /*
1445          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1446          * non-atomically.  Therefore, if get_user below is not
1447          * enough, we need to handle the fault ourselves, while
1448          * still holding the mmap_sem.
1449          */
1450         if (attempt++) {
1451                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1452                         ret = -EFAULT;
1453                         goto out_unlock;
1454                 }
1455                 goto retry_locked;
1456         }
1457
1458         spin_unlock(&hb->lock);
1459         up_read(&current->mm->mmap_sem);
1460
1461         ret = get_user(uval, uaddr);
1462         if (!ret && (uval != -EFAULT))
1463                 goto retry;
1464
1465         return ret;
1466 }
1467
1468 static int futex_close(struct inode *inode, struct file *filp)
1469 {
1470         struct futex_q *q = filp->private_data;
1471
1472         unqueue_me(q);
1473         kfree(q);
1474
1475         return 0;
1476 }
1477
1478 /* This is one-shot: once it's gone off you need a new fd */
1479 static unsigned int futex_poll(struct file *filp,
1480                                struct poll_table_struct *wait)
1481 {
1482         struct futex_q *q = filp->private_data;
1483         int ret = 0;
1484
1485         poll_wait(filp, &q->waiters, wait);
1486
1487         /*
1488          * list_empty() is safe here without any lock.
1489          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1490          */
1491         if (list_empty(&q->list))
1492                 ret = POLLIN | POLLRDNORM;
1493
1494         return ret;
1495 }
1496
1497 static const struct file_operations futex_fops = {
1498         .release        = futex_close,
1499         .poll           = futex_poll,
1500 };
1501
1502 /*
1503  * Signal allows caller to avoid the race which would occur if they
1504  * set the sigio stuff up afterwards.
1505  */
1506 static int futex_fd(u32 __user *uaddr, int signal)
1507 {
1508         struct futex_q *q;
1509         struct file *filp;
1510         int ret, err;
1511         static unsigned long printk_interval;
1512
1513         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1514                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1515                         "will be removed from the kernel in June 2007\n",
1516                         current->comm);
1517         }
1518
1519         ret = -EINVAL;
1520         if (!valid_signal(signal))
1521                 goto out;
1522
1523         ret = get_unused_fd();
1524         if (ret < 0)
1525                 goto out;
1526         filp = get_empty_filp();
1527         if (!filp) {
1528                 put_unused_fd(ret);
1529                 ret = -ENFILE;
1530                 goto out;
1531         }
1532         filp->f_op = &futex_fops;
1533         filp->f_path.mnt = mntget(futex_mnt);
1534         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1535         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1536
1537         if (signal) {
1538                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1539                 if (err < 0) {
1540                         goto error;
1541                 }
1542                 filp->f_owner.signum = signal;
1543         }
1544
1545         q = kmalloc(sizeof(*q), GFP_KERNEL);
1546         if (!q) {
1547                 err = -ENOMEM;
1548                 goto error;
1549         }
1550         q->pi_state = NULL;
1551
1552         down_read(&current->mm->mmap_sem);
1553         err = get_futex_key(uaddr, &q->key);
1554
1555         if (unlikely(err != 0)) {
1556                 up_read(&current->mm->mmap_sem);
1557                 kfree(q);
1558                 goto error;
1559         }
1560
1561         /*
1562          * queue_me() must be called before releasing mmap_sem, because
1563          * key->shared.inode needs to be referenced while holding it.
1564          */
1565         filp->private_data = q;
1566
1567         queue_me(q, ret, filp);
1568         up_read(&current->mm->mmap_sem);
1569
1570         /* Now we map fd to filp, so userspace can access it */
1571         fd_install(ret, filp);
1572 out:
1573         return ret;
1574 error:
1575         put_unused_fd(ret);
1576         put_filp(filp);
1577         ret = err;
1578         goto out;
1579 }
1580
1581 /*
1582  * Support for robust futexes: the kernel cleans up held futexes at
1583  * thread exit time.
1584  *
1585  * Implementation: user-space maintains a per-thread list of locks it
1586  * is holding. Upon do_exit(), the kernel carefully walks this list,
1587  * and marks all locks that are owned by this thread with the
1588  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1589  * always manipulated with the lock held, so the list is private and
1590  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1591  * field, to allow the kernel to clean up if the thread dies after
1592  * acquiring the lock, but just before it could have added itself to
1593  * the list. There can only be one such pending lock.
1594  */
1595
1596 /**
1597  * sys_set_robust_list - set the robust-futex list head of a task
1598  * @head: pointer to the list-head
1599  * @len: length of the list-head, as userspace expects
1600  */
1601 asmlinkage long
1602 sys_set_robust_list(struct robust_list_head __user *head,
1603                     size_t len)
1604 {
1605         /*
1606          * The kernel knows only one size for now:
1607          */
1608         if (unlikely(len != sizeof(*head)))
1609                 return -EINVAL;
1610
1611         current->robust_list = head;
1612
1613         return 0;
1614 }
1615
1616 /**
1617  * sys_get_robust_list - get the robust-futex list head of a task
1618  * @pid: pid of the process [zero for current task]
1619  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1620  * @len_ptr: pointer to a length field, the kernel fills in the header size
1621  */
1622 asmlinkage long
1623 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1624                     size_t __user *len_ptr)
1625 {
1626         struct robust_list_head __user *head;
1627         unsigned long ret;
1628
1629         if (!pid)
1630                 head = current->robust_list;
1631         else {
1632                 struct task_struct *p;
1633
1634                 ret = -ESRCH;
1635                 rcu_read_lock();
1636                 p = find_task_by_pid(pid);
1637                 if (!p)
1638                         goto err_unlock;
1639                 ret = -EPERM;
1640                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1641                                 !capable(CAP_SYS_PTRACE))
1642                         goto err_unlock;
1643                 head = p->robust_list;
1644                 rcu_read_unlock();
1645         }
1646
1647         if (put_user(sizeof(*head), len_ptr))
1648                 return -EFAULT;
1649         return put_user(head, head_ptr);
1650
1651 err_unlock:
1652         rcu_read_unlock();
1653
1654         return ret;
1655 }
1656
1657 /*
1658  * Process a futex-list entry, check whether it's owned by the
1659  * dying task, and do notification if so:
1660  */
1661 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1662 {
1663         u32 uval, nval, mval;
1664
1665 retry:
1666         if (get_user(uval, uaddr))
1667                 return -1;
1668
1669         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1670                 /*
1671                  * Ok, this dying thread is truly holding a futex
1672                  * of interest. Set the OWNER_DIED bit atomically
1673                  * via cmpxchg, and if the value had FUTEX_WAITERS
1674                  * set, wake up a waiter (if any). (We have to do a
1675                  * futex_wake() even if OWNER_DIED is already set -
1676                  * to handle the rare but possible case of recursive
1677                  * thread-death.) The rest of the cleanup is done in
1678                  * userspace.
1679                  */
1680                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1681                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1682
1683                 if (nval == -EFAULT)
1684                         return -1;
1685
1686                 if (nval != uval)
1687                         goto retry;
1688
1689                 /*
1690                  * Wake robust non-PI futexes here. The wakeup of
1691                  * PI futexes happens in exit_pi_state():
1692                  */
1693                 if (!pi) {
1694                         if (uval & FUTEX_WAITERS)
1695                                 futex_wake(uaddr, 1);
1696                 }
1697         }
1698         return 0;
1699 }
1700
1701 /*
1702  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1703  */
1704 static inline int fetch_robust_entry(struct robust_list __user **entry,
1705                                      struct robust_list __user * __user *head,
1706                                      int *pi)
1707 {
1708         unsigned long uentry;
1709
1710         if (get_user(uentry, (unsigned long __user *)head))
1711                 return -EFAULT;
1712
1713         *entry = (void __user *)(uentry & ~1UL);
1714         *pi = uentry & 1;
1715
1716         return 0;
1717 }
1718
1719 /*
1720  * Walk curr->robust_list (very carefully, it's a userspace list!)
1721  * and mark any locks found there dead, and notify any waiters.
1722  *
1723  * We silently return on any sign of list-walking problem.
1724  */
1725 void exit_robust_list(struct task_struct *curr)
1726 {
1727         struct robust_list_head __user *head = curr->robust_list;
1728         struct robust_list __user *entry, *pending;
1729         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1730         unsigned long futex_offset;
1731
1732         /*
1733          * Fetch the list head (which was registered earlier, via
1734          * sys_set_robust_list()):
1735          */
1736         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1737                 return;
1738         /*
1739          * Fetch the relative futex offset:
1740          */
1741         if (get_user(futex_offset, &head->futex_offset))
1742                 return;
1743         /*
1744          * Fetch any possibly pending lock-add first, and handle it
1745          * if it exists:
1746          */
1747         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1748                 return;
1749
1750         if (pending)
1751                 handle_futex_death((void __user *)pending + futex_offset, curr, pip);
1752
1753         while (entry != &head->list) {
1754                 /*
1755                  * A pending lock might already be on the list, so
1756                  * don't process it twice:
1757                  */
1758                 if (entry != pending)
1759                         if (handle_futex_death((void __user *)entry + futex_offset,
1760                                                 curr, pi))
1761                                 return;
1762                 /*
1763                  * Fetch the next entry in the list:
1764                  */
1765                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1766                         return;
1767                 /*
1768                  * Avoid excessively long or circular lists:
1769                  */
1770                 if (!--limit)
1771                         break;
1772
1773                 cond_resched();
1774         }
1775 }
1776
1777 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1778                 u32 __user *uaddr2, u32 val2, u32 val3)
1779 {
1780         int ret;
1781
1782         switch (op) {
1783         case FUTEX_WAIT:
1784                 ret = futex_wait(uaddr, val, timeout);
1785                 break;
1786         case FUTEX_WAKE:
1787                 ret = futex_wake(uaddr, val);
1788                 break;
1789         case FUTEX_FD:
1790                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1791                 ret = futex_fd(uaddr, val);
1792                 break;
1793         case FUTEX_REQUEUE:
1794                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1795                 break;
1796         case FUTEX_CMP_REQUEUE:
1797                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1798                 break;
1799         case FUTEX_WAKE_OP:
1800                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1801                 break;
1802         case FUTEX_LOCK_PI:
1803                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1804                 break;
1805         case FUTEX_UNLOCK_PI:
1806                 ret = futex_unlock_pi(uaddr);
1807                 break;
1808         case FUTEX_TRYLOCK_PI:
1809                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1810                 break;
1811         default:
1812                 ret = -ENOSYS;
1813         }
1814         return ret;
1815 }
1816
1817
1818 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1819                           struct timespec __user *utime, u32 __user *uaddr2,
1820                           u32 val3)
1821 {
1822         struct timespec t;
1823         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1824         u32 val2 = 0;
1825
1826         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1827                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1828                         return -EFAULT;
1829                 if (!timespec_valid(&t))
1830                         return -EINVAL;
1831                 if (op == FUTEX_WAIT)
1832                         timeout = timespec_to_jiffies(&t) + 1;
1833                 else {
1834                         timeout = t.tv_sec;
1835                         val2 = t.tv_nsec;
1836                 }
1837         }
1838         /*
1839          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1840          */
1841         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1842                 val2 = (u32) (unsigned long) utime;
1843
1844         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1845 }
1846
1847 static int futexfs_get_sb(struct file_system_type *fs_type,
1848                           int flags, const char *dev_name, void *data,
1849                           struct vfsmount *mnt)
1850 {
1851         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1852 }
1853
1854 static struct file_system_type futex_fs_type = {
1855         .name           = "futexfs",
1856         .get_sb         = futexfs_get_sb,
1857         .kill_sb        = kill_anon_super,
1858 };
1859
1860 static int __init init(void)
1861 {
1862         int i = register_filesystem(&futex_fs_type);
1863
1864         if (i)
1865                 return i;
1866
1867         futex_mnt = kern_mount(&futex_fs_type);
1868         if (IS_ERR(futex_mnt)) {
1869                 unregister_filesystem(&futex_fs_type);
1870                 return PTR_ERR(futex_mnt);
1871         }
1872
1873         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1874                 INIT_LIST_HEAD(&futex_queues[i].chain);
1875                 spin_lock_init(&futex_queues[i].lock);
1876         }
1877         return 0;
1878 }
1879 __initcall(init);