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