[PATCH] futex: init error check
[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  *  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_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_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         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         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
569
570         /*
571          * This happens when we have stolen the lock and the original
572          * pending owner did not enqueue itself back on the rt_mutex.
573          * Thats not a tragedy. We know that way, that a lock waiter
574          * is on the fly. We make the futex_q waiter the pending owner.
575          */
576         if (!new_owner)
577                 new_owner = this->task;
578
579         /*
580          * We pass it to the next owner. (The WAITERS bit is always
581          * kept enabled while there is PI state around. We must also
582          * preserve the owner died bit.)
583          */
584         if (!(uval & FUTEX_OWNER_DIED)) {
585                 newval = FUTEX_WAITERS | new_owner->pid;
586
587                 pagefault_disable();
588                 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
589                 pagefault_enable();
590                 if (curval == -EFAULT)
591                         return -EFAULT;
592                 if (curval != uval)
593                         return -EINVAL;
594         }
595
596         spin_lock_irq(&pi_state->owner->pi_lock);
597         WARN_ON(list_empty(&pi_state->list));
598         list_del_init(&pi_state->list);
599         spin_unlock_irq(&pi_state->owner->pi_lock);
600
601         spin_lock_irq(&new_owner->pi_lock);
602         WARN_ON(!list_empty(&pi_state->list));
603         list_add(&pi_state->list, &new_owner->pi_state_list);
604         pi_state->owner = new_owner;
605         spin_unlock_irq(&new_owner->pi_lock);
606
607         rt_mutex_unlock(&pi_state->pi_mutex);
608
609         return 0;
610 }
611
612 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
613 {
614         u32 oldval;
615
616         /*
617          * There is no waiter, so we unlock the futex. The owner died
618          * bit has not to be preserved here. We are the owner:
619          */
620         pagefault_disable();
621         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
622         pagefault_enable();
623
624         if (oldval == -EFAULT)
625                 return oldval;
626         if (oldval != uval)
627                 return -EAGAIN;
628
629         return 0;
630 }
631
632 /*
633  * Express the locking dependencies for lockdep:
634  */
635 static inline void
636 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
637 {
638         if (hb1 <= hb2) {
639                 spin_lock(&hb1->lock);
640                 if (hb1 < hb2)
641                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
642         } else { /* hb1 > hb2 */
643                 spin_lock(&hb2->lock);
644                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
645         }
646 }
647
648 /*
649  * Wake up all waiters hashed on the physical page that is mapped
650  * to this virtual address:
651  */
652 static int futex_wake(u32 __user *uaddr, int nr_wake)
653 {
654         struct futex_hash_bucket *hb;
655         struct futex_q *this, *next;
656         struct list_head *head;
657         union futex_key key;
658         int ret;
659
660         down_read(&current->mm->mmap_sem);
661
662         ret = get_futex_key(uaddr, &key);
663         if (unlikely(ret != 0))
664                 goto out;
665
666         hb = hash_futex(&key);
667         spin_lock(&hb->lock);
668         head = &hb->chain;
669
670         list_for_each_entry_safe(this, next, head, list) {
671                 if (match_futex (&this->key, &key)) {
672                         if (this->pi_state) {
673                                 ret = -EINVAL;
674                                 break;
675                         }
676                         wake_futex(this);
677                         if (++ret >= nr_wake)
678                                 break;
679                 }
680         }
681
682         spin_unlock(&hb->lock);
683 out:
684         up_read(&current->mm->mmap_sem);
685         return ret;
686 }
687
688 /*
689  * Wake up all waiters hashed on the physical page that is mapped
690  * to this virtual address:
691  */
692 static int
693 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
694               int nr_wake, int nr_wake2, int op)
695 {
696         union futex_key key1, key2;
697         struct futex_hash_bucket *hb1, *hb2;
698         struct list_head *head;
699         struct futex_q *this, *next;
700         int ret, op_ret, attempt = 0;
701
702 retryfull:
703         down_read(&current->mm->mmap_sem);
704
705         ret = get_futex_key(uaddr1, &key1);
706         if (unlikely(ret != 0))
707                 goto out;
708         ret = get_futex_key(uaddr2, &key2);
709         if (unlikely(ret != 0))
710                 goto out;
711
712         hb1 = hash_futex(&key1);
713         hb2 = hash_futex(&key2);
714
715 retry:
716         double_lock_hb(hb1, hb2);
717
718         op_ret = futex_atomic_op_inuser(op, uaddr2);
719         if (unlikely(op_ret < 0)) {
720                 u32 dummy;
721
722                 spin_unlock(&hb1->lock);
723                 if (hb1 != hb2)
724                         spin_unlock(&hb2->lock);
725
726 #ifndef CONFIG_MMU
727                 /*
728                  * we don't get EFAULT from MMU faults if we don't have an MMU,
729                  * but we might get them from range checking
730                  */
731                 ret = op_ret;
732                 goto out;
733 #endif
734
735                 if (unlikely(op_ret != -EFAULT)) {
736                         ret = op_ret;
737                         goto out;
738                 }
739
740                 /*
741                  * futex_atomic_op_inuser needs to both read and write
742                  * *(int __user *)uaddr2, but we can't modify it
743                  * non-atomically.  Therefore, if get_user below is not
744                  * enough, we need to handle the fault ourselves, while
745                  * still holding the mmap_sem.
746                  */
747                 if (attempt++) {
748                         if (futex_handle_fault((unsigned long)uaddr2,
749                                                 attempt)) {
750                                 ret = -EFAULT;
751                                 goto out;
752                         }
753                         goto retry;
754                 }
755
756                 /*
757                  * If we would have faulted, release mmap_sem,
758                  * fault it in and start all over again.
759                  */
760                 up_read(&current->mm->mmap_sem);
761
762                 ret = get_user(dummy, uaddr2);
763                 if (ret)
764                         return ret;
765
766                 goto retryfull;
767         }
768
769         head = &hb1->chain;
770
771         list_for_each_entry_safe(this, next, head, list) {
772                 if (match_futex (&this->key, &key1)) {
773                         wake_futex(this);
774                         if (++ret >= nr_wake)
775                                 break;
776                 }
777         }
778
779         if (op_ret > 0) {
780                 head = &hb2->chain;
781
782                 op_ret = 0;
783                 list_for_each_entry_safe(this, next, head, list) {
784                         if (match_futex (&this->key, &key2)) {
785                                 wake_futex(this);
786                                 if (++op_ret >= nr_wake2)
787                                         break;
788                         }
789                 }
790                 ret += op_ret;
791         }
792
793         spin_unlock(&hb1->lock);
794         if (hb1 != hb2)
795                 spin_unlock(&hb2->lock);
796 out:
797         up_read(&current->mm->mmap_sem);
798         return ret;
799 }
800
801 /*
802  * Requeue all waiters hashed on one physical page to another
803  * physical page.
804  */
805 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
806                          int nr_wake, int nr_requeue, u32 *cmpval)
807 {
808         union futex_key key1, key2;
809         struct futex_hash_bucket *hb1, *hb2;
810         struct list_head *head1;
811         struct futex_q *this, *next;
812         int ret, drop_count = 0;
813
814  retry:
815         down_read(&current->mm->mmap_sem);
816
817         ret = get_futex_key(uaddr1, &key1);
818         if (unlikely(ret != 0))
819                 goto out;
820         ret = get_futex_key(uaddr2, &key2);
821         if (unlikely(ret != 0))
822                 goto out;
823
824         hb1 = hash_futex(&key1);
825         hb2 = hash_futex(&key2);
826
827         double_lock_hb(hb1, hb2);
828
829         if (likely(cmpval != NULL)) {
830                 u32 curval;
831
832                 ret = get_futex_value_locked(&curval, uaddr1);
833
834                 if (unlikely(ret)) {
835                         spin_unlock(&hb1->lock);
836                         if (hb1 != hb2)
837                                 spin_unlock(&hb2->lock);
838
839                         /*
840                          * If we would have faulted, release mmap_sem, fault
841                          * it in and start all over again.
842                          */
843                         up_read(&current->mm->mmap_sem);
844
845                         ret = get_user(curval, uaddr1);
846
847                         if (!ret)
848                                 goto retry;
849
850                         return ret;
851                 }
852                 if (curval != *cmpval) {
853                         ret = -EAGAIN;
854                         goto out_unlock;
855                 }
856         }
857
858         head1 = &hb1->chain;
859         list_for_each_entry_safe(this, next, head1, list) {
860                 if (!match_futex (&this->key, &key1))
861                         continue;
862                 if (++ret <= nr_wake) {
863                         wake_futex(this);
864                 } else {
865                         /*
866                          * If key1 and key2 hash to the same bucket, no need to
867                          * requeue.
868                          */
869                         if (likely(head1 != &hb2->chain)) {
870                                 list_move_tail(&this->list, &hb2->chain);
871                                 this->lock_ptr = &hb2->lock;
872                         }
873                         this->key = key2;
874                         get_key_refs(&key2);
875                         drop_count++;
876
877                         if (ret - nr_wake >= nr_requeue)
878                                 break;
879                 }
880         }
881
882 out_unlock:
883         spin_unlock(&hb1->lock);
884         if (hb1 != hb2)
885                 spin_unlock(&hb2->lock);
886
887         /* drop_key_refs() must be called outside the spinlocks. */
888         while (--drop_count >= 0)
889                 drop_key_refs(&key1);
890
891 out:
892         up_read(&current->mm->mmap_sem);
893         return ret;
894 }
895
896 /* The key must be already stored in q->key. */
897 static inline struct futex_hash_bucket *
898 queue_lock(struct futex_q *q, int fd, struct file *filp)
899 {
900         struct futex_hash_bucket *hb;
901
902         q->fd = fd;
903         q->filp = filp;
904
905         init_waitqueue_head(&q->waiters);
906
907         get_key_refs(&q->key);
908         hb = hash_futex(&q->key);
909         q->lock_ptr = &hb->lock;
910
911         spin_lock(&hb->lock);
912         return hb;
913 }
914
915 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
916 {
917         list_add_tail(&q->list, &hb->chain);
918         q->task = current;
919         spin_unlock(&hb->lock);
920 }
921
922 static inline void
923 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
924 {
925         spin_unlock(&hb->lock);
926         drop_key_refs(&q->key);
927 }
928
929 /*
930  * queue_me and unqueue_me must be called as a pair, each
931  * exactly once.  They are called with the hashed spinlock held.
932  */
933
934 /* The key must be already stored in q->key. */
935 static void queue_me(struct futex_q *q, int fd, struct file *filp)
936 {
937         struct futex_hash_bucket *hb;
938
939         hb = queue_lock(q, fd, filp);
940         __queue_me(q, hb);
941 }
942
943 /* Return 1 if we were still queued (ie. 0 means we were woken) */
944 static int unqueue_me(struct futex_q *q)
945 {
946         spinlock_t *lock_ptr;
947         int ret = 0;
948
949         /* In the common case we don't take the spinlock, which is nice. */
950  retry:
951         lock_ptr = q->lock_ptr;
952         barrier();
953         if (lock_ptr != 0) {
954                 spin_lock(lock_ptr);
955                 /*
956                  * q->lock_ptr can change between reading it and
957                  * spin_lock(), causing us to take the wrong lock.  This
958                  * corrects the race condition.
959                  *
960                  * Reasoning goes like this: if we have the wrong lock,
961                  * q->lock_ptr must have changed (maybe several times)
962                  * between reading it and the spin_lock().  It can
963                  * change again after the spin_lock() but only if it was
964                  * already changed before the spin_lock().  It cannot,
965                  * however, change back to the original value.  Therefore
966                  * we can detect whether we acquired the correct lock.
967                  */
968                 if (unlikely(lock_ptr != q->lock_ptr)) {
969                         spin_unlock(lock_ptr);
970                         goto retry;
971                 }
972                 WARN_ON(list_empty(&q->list));
973                 list_del(&q->list);
974
975                 BUG_ON(q->pi_state);
976
977                 spin_unlock(lock_ptr);
978                 ret = 1;
979         }
980
981         drop_key_refs(&q->key);
982         return ret;
983 }
984
985 /*
986  * PI futexes can not be requeued and must remove themself from the
987  * hash bucket. The hash bucket lock is held on entry and dropped here.
988  */
989 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
990 {
991         WARN_ON(list_empty(&q->list));
992         list_del(&q->list);
993
994         BUG_ON(!q->pi_state);
995         free_pi_state(q->pi_state);
996         q->pi_state = NULL;
997
998         spin_unlock(&hb->lock);
999
1000         drop_key_refs(&q->key);
1001 }
1002
1003 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
1004 {
1005         struct task_struct *curr = current;
1006         DECLARE_WAITQUEUE(wait, curr);
1007         struct futex_hash_bucket *hb;
1008         struct futex_q q;
1009         u32 uval;
1010         int ret;
1011
1012         q.pi_state = NULL;
1013  retry:
1014         down_read(&curr->mm->mmap_sem);
1015
1016         ret = get_futex_key(uaddr, &q.key);
1017         if (unlikely(ret != 0))
1018                 goto out_release_sem;
1019
1020         hb = queue_lock(&q, -1, NULL);
1021
1022         /*
1023          * Access the page AFTER the futex is queued.
1024          * Order is important:
1025          *
1026          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1027          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1028          *
1029          * The basic logical guarantee of a futex is that it blocks ONLY
1030          * if cond(var) is known to be true at the time of blocking, for
1031          * any cond.  If we queued after testing *uaddr, that would open
1032          * a race condition where we could block indefinitely with
1033          * cond(var) false, which would violate the guarantee.
1034          *
1035          * A consequence is that futex_wait() can return zero and absorb
1036          * a wakeup when *uaddr != val on entry to the syscall.  This is
1037          * rare, but normal.
1038          *
1039          * We hold the mmap semaphore, so the mapping cannot have changed
1040          * since we looked it up in get_futex_key.
1041          */
1042         ret = get_futex_value_locked(&uval, uaddr);
1043
1044         if (unlikely(ret)) {
1045                 queue_unlock(&q, hb);
1046
1047                 /*
1048                  * If we would have faulted, release mmap_sem, fault it in and
1049                  * start all over again.
1050                  */
1051                 up_read(&curr->mm->mmap_sem);
1052
1053                 ret = get_user(uval, uaddr);
1054
1055                 if (!ret)
1056                         goto retry;
1057                 return ret;
1058         }
1059         ret = -EWOULDBLOCK;
1060         if (uval != val)
1061                 goto out_unlock_release_sem;
1062
1063         /* Only actually queue if *uaddr contained val.  */
1064         __queue_me(&q, hb);
1065
1066         /*
1067          * Now the futex is queued and we have checked the data, we
1068          * don't want to hold mmap_sem while we sleep.
1069          */
1070         up_read(&curr->mm->mmap_sem);
1071
1072         /*
1073          * There might have been scheduling since the queue_me(), as we
1074          * cannot hold a spinlock across the get_user() in case it
1075          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1076          * queueing ourselves into the futex hash.  This code thus has to
1077          * rely on the futex_wake() code removing us from hash when it
1078          * wakes us up.
1079          */
1080
1081         /* add_wait_queue is the barrier after __set_current_state. */
1082         __set_current_state(TASK_INTERRUPTIBLE);
1083         add_wait_queue(&q.waiters, &wait);
1084         /*
1085          * !list_empty() is safe here without any lock.
1086          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1087          */
1088         if (likely(!list_empty(&q.list)))
1089                 time = schedule_timeout(time);
1090         __set_current_state(TASK_RUNNING);
1091
1092         /*
1093          * NOTE: we don't remove ourselves from the waitqueue because
1094          * we are the only user of it.
1095          */
1096
1097         /* If we were woken (and unqueued), we succeeded, whatever. */
1098         if (!unqueue_me(&q))
1099                 return 0;
1100         if (time == 0)
1101                 return -ETIMEDOUT;
1102         /*
1103          * We expect signal_pending(current), but another thread may
1104          * have handled it for us already.
1105          */
1106         return -EINTR;
1107
1108  out_unlock_release_sem:
1109         queue_unlock(&q, hb);
1110
1111  out_release_sem:
1112         up_read(&curr->mm->mmap_sem);
1113         return ret;
1114 }
1115
1116 /*
1117  * Userspace tried a 0 -> TID atomic transition of the futex value
1118  * and failed. The kernel side here does the whole locking operation:
1119  * if there are waiters then it will block, it does PI, etc. (Due to
1120  * races the kernel might see a 0 value of the futex too.)
1121  */
1122 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1123                          long nsec, int trylock)
1124 {
1125         struct hrtimer_sleeper timeout, *to = NULL;
1126         struct task_struct *curr = current;
1127         struct futex_hash_bucket *hb;
1128         u32 uval, newval, curval;
1129         struct futex_q q;
1130         int ret, attempt = 0;
1131
1132         if (refill_pi_state_cache())
1133                 return -ENOMEM;
1134
1135         if (sec != MAX_SCHEDULE_TIMEOUT) {
1136                 to = &timeout;
1137                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1138                 hrtimer_init_sleeper(to, current);
1139                 to->timer.expires = ktime_set(sec, nsec);
1140         }
1141
1142         q.pi_state = NULL;
1143  retry:
1144         down_read(&curr->mm->mmap_sem);
1145
1146         ret = get_futex_key(uaddr, &q.key);
1147         if (unlikely(ret != 0))
1148                 goto out_release_sem;
1149
1150         hb = queue_lock(&q, -1, NULL);
1151
1152  retry_locked:
1153         /*
1154          * To avoid races, we attempt to take the lock here again
1155          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1156          * the locks. It will most likely not succeed.
1157          */
1158         newval = current->pid;
1159
1160         pagefault_disable();
1161         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1162         pagefault_enable();
1163
1164         if (unlikely(curval == -EFAULT))
1165                 goto uaddr_faulted;
1166
1167         /* We own the lock already */
1168         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1169                 if (!detect && 0)
1170                         force_sig(SIGKILL, current);
1171                 ret = -EDEADLK;
1172                 goto out_unlock_release_sem;
1173         }
1174
1175         /*
1176          * Surprise - we got the lock. Just return
1177          * to userspace:
1178          */
1179         if (unlikely(!curval))
1180                 goto out_unlock_release_sem;
1181
1182         uval = curval;
1183         newval = uval | FUTEX_WAITERS;
1184
1185         pagefault_disable();
1186         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1187         pagefault_enable();
1188
1189         if (unlikely(curval == -EFAULT))
1190                 goto uaddr_faulted;
1191         if (unlikely(curval != uval))
1192                 goto retry_locked;
1193
1194         /*
1195          * We dont have the lock. Look up the PI state (or create it if
1196          * we are the first waiter):
1197          */
1198         ret = lookup_pi_state(uval, hb, &q);
1199
1200         if (unlikely(ret)) {
1201                 /*
1202                  * There were no waiters and the owner task lookup
1203                  * failed. When the OWNER_DIED bit is set, then we
1204                  * know that this is a robust futex and we actually
1205                  * take the lock. This is safe as we are protected by
1206                  * the hash bucket lock. We also set the waiters bit
1207                  * unconditionally here, to simplify glibc handling of
1208                  * multiple tasks racing to acquire the lock and
1209                  * cleanup the problems which were left by the dead
1210                  * owner.
1211                  */
1212                 if (curval & FUTEX_OWNER_DIED) {
1213                         uval = newval;
1214                         newval = current->pid |
1215                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1216
1217                         pagefault_disable();
1218                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1219                                                                uval, newval);
1220                         pagefault_enable();
1221
1222                         if (unlikely(curval == -EFAULT))
1223                                 goto uaddr_faulted;
1224                         if (unlikely(curval != uval))
1225                                 goto retry_locked;
1226                         ret = 0;
1227                 }
1228                 goto out_unlock_release_sem;
1229         }
1230
1231         /*
1232          * Only actually queue now that the atomic ops are done:
1233          */
1234         __queue_me(&q, hb);
1235
1236         /*
1237          * Now the futex is queued and we have checked the data, we
1238          * don't want to hold mmap_sem while we sleep.
1239          */
1240         up_read(&curr->mm->mmap_sem);
1241
1242         WARN_ON(!q.pi_state);
1243         /*
1244          * Block on the PI mutex:
1245          */
1246         if (!trylock)
1247                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1248         else {
1249                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1250                 /* Fixup the trylock return value: */
1251                 ret = ret ? 0 : -EWOULDBLOCK;
1252         }
1253
1254         down_read(&curr->mm->mmap_sem);
1255         spin_lock(q.lock_ptr);
1256
1257         /*
1258          * Got the lock. We might not be the anticipated owner if we
1259          * did a lock-steal - fix up the PI-state in that case.
1260          */
1261         if (!ret && q.pi_state->owner != curr) {
1262                 u32 newtid = current->pid | FUTEX_WAITERS;
1263
1264                 /* Owner died? */
1265                 if (q.pi_state->owner != NULL) {
1266                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1267                         WARN_ON(list_empty(&q.pi_state->list));
1268                         list_del_init(&q.pi_state->list);
1269                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1270                 } else
1271                         newtid |= FUTEX_OWNER_DIED;
1272
1273                 q.pi_state->owner = current;
1274
1275                 spin_lock_irq(&current->pi_lock);
1276                 WARN_ON(!list_empty(&q.pi_state->list));
1277                 list_add(&q.pi_state->list, &current->pi_state_list);
1278                 spin_unlock_irq(&current->pi_lock);
1279
1280                 /* Unqueue and drop the lock */
1281                 unqueue_me_pi(&q, hb);
1282                 up_read(&curr->mm->mmap_sem);
1283                 /*
1284                  * We own it, so we have to replace the pending owner
1285                  * TID. This must be atomic as we have preserve the
1286                  * owner died bit here.
1287                  */
1288                 ret = get_user(uval, uaddr);
1289                 while (!ret) {
1290                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1291                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1292                                                                uval, newval);
1293                         if (curval == -EFAULT)
1294                                 ret = -EFAULT;
1295                         if (curval == uval)
1296                                 break;
1297                         uval = curval;
1298                 }
1299         } else {
1300                 /*
1301                  * Catch the rare case, where the lock was released
1302                  * when we were on the way back before we locked
1303                  * the hash bucket.
1304                  */
1305                 if (ret && q.pi_state->owner == curr) {
1306                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1307                                 ret = 0;
1308                 }
1309                 /* Unqueue and drop the lock */
1310                 unqueue_me_pi(&q, hb);
1311                 up_read(&curr->mm->mmap_sem);
1312         }
1313
1314         if (!detect && ret == -EDEADLK && 0)
1315                 force_sig(SIGKILL, current);
1316
1317         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1318
1319  out_unlock_release_sem:
1320         queue_unlock(&q, hb);
1321
1322  out_release_sem:
1323         up_read(&curr->mm->mmap_sem);
1324         return ret;
1325
1326  uaddr_faulted:
1327         /*
1328          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1329          * non-atomically.  Therefore, if get_user below is not
1330          * enough, we need to handle the fault ourselves, while
1331          * still holding the mmap_sem.
1332          */
1333         if (attempt++) {
1334                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1335                         ret = -EFAULT;
1336                         goto out_unlock_release_sem;
1337                 }
1338                 goto retry_locked;
1339         }
1340
1341         queue_unlock(&q, hb);
1342         up_read(&curr->mm->mmap_sem);
1343
1344         ret = get_user(uval, uaddr);
1345         if (!ret && (uval != -EFAULT))
1346                 goto retry;
1347
1348         return ret;
1349 }
1350
1351 /*
1352  * Userspace attempted a TID -> 0 atomic transition, and failed.
1353  * This is the in-kernel slowpath: we look up the PI state (if any),
1354  * and do the rt-mutex unlock.
1355  */
1356 static int futex_unlock_pi(u32 __user *uaddr)
1357 {
1358         struct futex_hash_bucket *hb;
1359         struct futex_q *this, *next;
1360         u32 uval;
1361         struct list_head *head;
1362         union futex_key key;
1363         int ret, attempt = 0;
1364
1365 retry:
1366         if (get_user(uval, uaddr))
1367                 return -EFAULT;
1368         /*
1369          * We release only a lock we actually own:
1370          */
1371         if ((uval & FUTEX_TID_MASK) != current->pid)
1372                 return -EPERM;
1373         /*
1374          * First take all the futex related locks:
1375          */
1376         down_read(&current->mm->mmap_sem);
1377
1378         ret = get_futex_key(uaddr, &key);
1379         if (unlikely(ret != 0))
1380                 goto out;
1381
1382         hb = hash_futex(&key);
1383         spin_lock(&hb->lock);
1384
1385 retry_locked:
1386         /*
1387          * To avoid races, try to do the TID -> 0 atomic transition
1388          * again. If it succeeds then we can return without waking
1389          * anyone else up:
1390          */
1391         if (!(uval & FUTEX_OWNER_DIED)) {
1392                 pagefault_disable();
1393                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1394                 pagefault_enable();
1395         }
1396
1397         if (unlikely(uval == -EFAULT))
1398                 goto pi_faulted;
1399         /*
1400          * Rare case: we managed to release the lock atomically,
1401          * no need to wake anyone else up:
1402          */
1403         if (unlikely(uval == current->pid))
1404                 goto out_unlock;
1405
1406         /*
1407          * Ok, other tasks may need to be woken up - check waiters
1408          * and do the wakeup if necessary:
1409          */
1410         head = &hb->chain;
1411
1412         list_for_each_entry_safe(this, next, head, list) {
1413                 if (!match_futex (&this->key, &key))
1414                         continue;
1415                 ret = wake_futex_pi(uaddr, uval, this);
1416                 /*
1417                  * The atomic access to the futex value
1418                  * generated a pagefault, so retry the
1419                  * user-access and the wakeup:
1420                  */
1421                 if (ret == -EFAULT)
1422                         goto pi_faulted;
1423                 goto out_unlock;
1424         }
1425         /*
1426          * No waiters - kernel unlocks the futex:
1427          */
1428         if (!(uval & FUTEX_OWNER_DIED)) {
1429                 ret = unlock_futex_pi(uaddr, uval);
1430                 if (ret == -EFAULT)
1431                         goto pi_faulted;
1432         }
1433
1434 out_unlock:
1435         spin_unlock(&hb->lock);
1436 out:
1437         up_read(&current->mm->mmap_sem);
1438
1439         return ret;
1440
1441 pi_faulted:
1442         /*
1443          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1444          * non-atomically.  Therefore, if get_user below is not
1445          * enough, we need to handle the fault ourselves, while
1446          * still holding the mmap_sem.
1447          */
1448         if (attempt++) {
1449                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1450                         ret = -EFAULT;
1451                         goto out_unlock;
1452                 }
1453                 goto retry_locked;
1454         }
1455
1456         spin_unlock(&hb->lock);
1457         up_read(&current->mm->mmap_sem);
1458
1459         ret = get_user(uval, uaddr);
1460         if (!ret && (uval != -EFAULT))
1461                 goto retry;
1462
1463         return ret;
1464 }
1465
1466 static int futex_close(struct inode *inode, struct file *filp)
1467 {
1468         struct futex_q *q = filp->private_data;
1469
1470         unqueue_me(q);
1471         kfree(q);
1472
1473         return 0;
1474 }
1475
1476 /* This is one-shot: once it's gone off you need a new fd */
1477 static unsigned int futex_poll(struct file *filp,
1478                                struct poll_table_struct *wait)
1479 {
1480         struct futex_q *q = filp->private_data;
1481         int ret = 0;
1482
1483         poll_wait(filp, &q->waiters, wait);
1484
1485         /*
1486          * list_empty() is safe here without any lock.
1487          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1488          */
1489         if (list_empty(&q->list))
1490                 ret = POLLIN | POLLRDNORM;
1491
1492         return ret;
1493 }
1494
1495 static struct file_operations futex_fops = {
1496         .release        = futex_close,
1497         .poll           = futex_poll,
1498 };
1499
1500 /*
1501  * Signal allows caller to avoid the race which would occur if they
1502  * set the sigio stuff up afterwards.
1503  */
1504 static int futex_fd(u32 __user *uaddr, int signal)
1505 {
1506         struct futex_q *q;
1507         struct file *filp;
1508         int ret, err;
1509         static unsigned long printk_interval;
1510
1511         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1512                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1513                         "will be removed from the kernel in June 2007\n",
1514                         current->comm);
1515         }
1516
1517         ret = -EINVAL;
1518         if (!valid_signal(signal))
1519                 goto out;
1520
1521         ret = get_unused_fd();
1522         if (ret < 0)
1523                 goto out;
1524         filp = get_empty_filp();
1525         if (!filp) {
1526                 put_unused_fd(ret);
1527                 ret = -ENFILE;
1528                 goto out;
1529         }
1530         filp->f_op = &futex_fops;
1531         filp->f_vfsmnt = mntget(futex_mnt);
1532         filp->f_dentry = dget(futex_mnt->mnt_root);
1533         filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
1534
1535         if (signal) {
1536                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1537                 if (err < 0) {
1538                         goto error;
1539                 }
1540                 filp->f_owner.signum = signal;
1541         }
1542
1543         q = kmalloc(sizeof(*q), GFP_KERNEL);
1544         if (!q) {
1545                 err = -ENOMEM;
1546                 goto error;
1547         }
1548         q->pi_state = NULL;
1549
1550         down_read(&current->mm->mmap_sem);
1551         err = get_futex_key(uaddr, &q->key);
1552
1553         if (unlikely(err != 0)) {
1554                 up_read(&current->mm->mmap_sem);
1555                 kfree(q);
1556                 goto error;
1557         }
1558
1559         /*
1560          * queue_me() must be called before releasing mmap_sem, because
1561          * key->shared.inode needs to be referenced while holding it.
1562          */
1563         filp->private_data = q;
1564
1565         queue_me(q, ret, filp);
1566         up_read(&current->mm->mmap_sem);
1567
1568         /* Now we map fd to filp, so userspace can access it */
1569         fd_install(ret, filp);
1570 out:
1571         return ret;
1572 error:
1573         put_unused_fd(ret);
1574         put_filp(filp);
1575         ret = err;
1576         goto out;
1577 }
1578
1579 /*
1580  * Support for robust futexes: the kernel cleans up held futexes at
1581  * thread exit time.
1582  *
1583  * Implementation: user-space maintains a per-thread list of locks it
1584  * is holding. Upon do_exit(), the kernel carefully walks this list,
1585  * and marks all locks that are owned by this thread with the
1586  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1587  * always manipulated with the lock held, so the list is private and
1588  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1589  * field, to allow the kernel to clean up if the thread dies after
1590  * acquiring the lock, but just before it could have added itself to
1591  * the list. There can only be one such pending lock.
1592  */
1593
1594 /**
1595  * sys_set_robust_list - set the robust-futex list head of a task
1596  * @head: pointer to the list-head
1597  * @len: length of the list-head, as userspace expects
1598  */
1599 asmlinkage long
1600 sys_set_robust_list(struct robust_list_head __user *head,
1601                     size_t len)
1602 {
1603         /*
1604          * The kernel knows only one size for now:
1605          */
1606         if (unlikely(len != sizeof(*head)))
1607                 return -EINVAL;
1608
1609         current->robust_list = head;
1610
1611         return 0;
1612 }
1613
1614 /**
1615  * sys_get_robust_list - get the robust-futex list head of a task
1616  * @pid: pid of the process [zero for current task]
1617  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1618  * @len_ptr: pointer to a length field, the kernel fills in the header size
1619  */
1620 asmlinkage long
1621 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1622                     size_t __user *len_ptr)
1623 {
1624         struct robust_list_head __user *head;
1625         unsigned long ret;
1626
1627         if (!pid)
1628                 head = current->robust_list;
1629         else {
1630                 struct task_struct *p;
1631
1632                 ret = -ESRCH;
1633                 rcu_read_lock();
1634                 p = find_task_by_pid(pid);
1635                 if (!p)
1636                         goto err_unlock;
1637                 ret = -EPERM;
1638                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1639                                 !capable(CAP_SYS_PTRACE))
1640                         goto err_unlock;
1641                 head = p->robust_list;
1642                 rcu_read_unlock();
1643         }
1644
1645         if (put_user(sizeof(*head), len_ptr))
1646                 return -EFAULT;
1647         return put_user(head, head_ptr);
1648
1649 err_unlock:
1650         rcu_read_unlock();
1651
1652         return ret;
1653 }
1654
1655 /*
1656  * Process a futex-list entry, check whether it's owned by the
1657  * dying task, and do notification if so:
1658  */
1659 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1660 {
1661         u32 uval, nval, mval;
1662
1663 retry:
1664         if (get_user(uval, uaddr))
1665                 return -1;
1666
1667         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1668                 /*
1669                  * Ok, this dying thread is truly holding a futex
1670                  * of interest. Set the OWNER_DIED bit atomically
1671                  * via cmpxchg, and if the value had FUTEX_WAITERS
1672                  * set, wake up a waiter (if any). (We have to do a
1673                  * futex_wake() even if OWNER_DIED is already set -
1674                  * to handle the rare but possible case of recursive
1675                  * thread-death.) The rest of the cleanup is done in
1676                  * userspace.
1677                  */
1678                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1679                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1680
1681                 if (nval == -EFAULT)
1682                         return -1;
1683
1684                 if (nval != uval)
1685                         goto retry;
1686
1687                 /*
1688                  * Wake robust non-PI futexes here. The wakeup of
1689                  * PI futexes happens in exit_pi_state():
1690                  */
1691                 if (!pi) {
1692                         if (uval & FUTEX_WAITERS)
1693                                 futex_wake(uaddr, 1);
1694                 }
1695         }
1696         return 0;
1697 }
1698
1699 /*
1700  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1701  */
1702 static inline int fetch_robust_entry(struct robust_list __user **entry,
1703                                      struct robust_list __user * __user *head,
1704                                      int *pi)
1705 {
1706         unsigned long uentry;
1707
1708         if (get_user(uentry, (unsigned long __user *)head))
1709                 return -EFAULT;
1710
1711         *entry = (void __user *)(uentry & ~1UL);
1712         *pi = uentry & 1;
1713
1714         return 0;
1715 }
1716
1717 /*
1718  * Walk curr->robust_list (very carefully, it's a userspace list!)
1719  * and mark any locks found there dead, and notify any waiters.
1720  *
1721  * We silently return on any sign of list-walking problem.
1722  */
1723 void exit_robust_list(struct task_struct *curr)
1724 {
1725         struct robust_list_head __user *head = curr->robust_list;
1726         struct robust_list __user *entry, *pending;
1727         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1728         unsigned long futex_offset;
1729
1730         /*
1731          * Fetch the list head (which was registered earlier, via
1732          * sys_set_robust_list()):
1733          */
1734         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1735                 return;
1736         /*
1737          * Fetch the relative futex offset:
1738          */
1739         if (get_user(futex_offset, &head->futex_offset))
1740                 return;
1741         /*
1742          * Fetch any possibly pending lock-add first, and handle it
1743          * if it exists:
1744          */
1745         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1746                 return;
1747
1748         if (pending)
1749                 handle_futex_death((void __user *)pending + futex_offset, curr, pip);
1750
1751         while (entry != &head->list) {
1752                 /*
1753                  * A pending lock might already be on the list, so
1754                  * don't process it twice:
1755                  */
1756                 if (entry != pending)
1757                         if (handle_futex_death((void __user *)entry + futex_offset,
1758                                                 curr, pi))
1759                                 return;
1760                 /*
1761                  * Fetch the next entry in the list:
1762                  */
1763                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1764                         return;
1765                 /*
1766                  * Avoid excessively long or circular lists:
1767                  */
1768                 if (!--limit)
1769                         break;
1770
1771                 cond_resched();
1772         }
1773 }
1774
1775 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1776                 u32 __user *uaddr2, u32 val2, u32 val3)
1777 {
1778         int ret;
1779
1780         switch (op) {
1781         case FUTEX_WAIT:
1782                 ret = futex_wait(uaddr, val, timeout);
1783                 break;
1784         case FUTEX_WAKE:
1785                 ret = futex_wake(uaddr, val);
1786                 break;
1787         case FUTEX_FD:
1788                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1789                 ret = futex_fd(uaddr, val);
1790                 break;
1791         case FUTEX_REQUEUE:
1792                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1793                 break;
1794         case FUTEX_CMP_REQUEUE:
1795                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1796                 break;
1797         case FUTEX_WAKE_OP:
1798                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1799                 break;
1800         case FUTEX_LOCK_PI:
1801                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1802                 break;
1803         case FUTEX_UNLOCK_PI:
1804                 ret = futex_unlock_pi(uaddr);
1805                 break;
1806         case FUTEX_TRYLOCK_PI:
1807                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1808                 break;
1809         default:
1810                 ret = -ENOSYS;
1811         }
1812         return ret;
1813 }
1814
1815
1816 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1817                           struct timespec __user *utime, u32 __user *uaddr2,
1818                           u32 val3)
1819 {
1820         struct timespec t;
1821         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1822         u32 val2 = 0;
1823
1824         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1825                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1826                         return -EFAULT;
1827                 if (!timespec_valid(&t))
1828                         return -EINVAL;
1829                 if (op == FUTEX_WAIT)
1830                         timeout = timespec_to_jiffies(&t) + 1;
1831                 else {
1832                         timeout = t.tv_sec;
1833                         val2 = t.tv_nsec;
1834                 }
1835         }
1836         /*
1837          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1838          */
1839         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1840                 val2 = (u32) (unsigned long) utime;
1841
1842         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1843 }
1844
1845 static int futexfs_get_sb(struct file_system_type *fs_type,
1846                           int flags, const char *dev_name, void *data,
1847                           struct vfsmount *mnt)
1848 {
1849         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1850 }
1851
1852 static struct file_system_type futex_fs_type = {
1853         .name           = "futexfs",
1854         .get_sb         = futexfs_get_sb,
1855         .kill_sb        = kill_anon_super,
1856 };
1857
1858 static int __init init(void)
1859 {
1860         int i = register_filesystem(&futex_fs_type);
1861
1862         if (i)
1863                 return i;
1864
1865         futex_mnt = kern_mount(&futex_fs_type);
1866         if (IS_ERR(futex_mnt)) {
1867                 unregister_filesystem(&futex_fs_type);
1868                 return PTR_ERR(futex_mnt);
1869         }
1870
1871         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1872                 INIT_LIST_HEAD(&futex_queues[i].chain);
1873                 spin_lock_init(&futex_queues[i].lock);
1874         }
1875         return 0;
1876 }
1877 __initcall(init);