[PATCH] pi-futex: robust-futex exit crash fix
[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         read_lock(&tasklist_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->state == EXIT_ZOMBIE || p->exit_state == EXIT_ZOMBIE) {
401                 p = NULL;
402                 goto out_unlock;
403         }
404         get_task_struct(p);
405 out_unlock:
406         read_unlock(&tasklist_lock);
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
499          * attach the new pi_state to it:
500          */
501         pid = uval & FUTEX_TID_MASK;
502         p = futex_find_get_task(pid);
503         if (!p)
504                 return -ESRCH;
505
506         pi_state = alloc_pi_state();
507
508         /*
509          * Initialize the pi_mutex in locked state and make 'p'
510          * the owner of it:
511          */
512         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
513
514         /* Store the key for possible exit cleanups: */
515         pi_state->key = me->key;
516
517         spin_lock_irq(&p->pi_lock);
518         WARN_ON(!list_empty(&pi_state->list));
519         list_add(&pi_state->list, &p->pi_state_list);
520         pi_state->owner = p;
521         spin_unlock_irq(&p->pi_lock);
522
523         put_task_struct(p);
524
525         me->pi_state = pi_state;
526
527         return 0;
528 }
529
530 /*
531  * The hash bucket lock must be held when this is called.
532  * Afterwards, the futex_q must not be accessed.
533  */
534 static void wake_futex(struct futex_q *q)
535 {
536         list_del_init(&q->list);
537         if (q->filp)
538                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
539         /*
540          * The lock in wake_up_all() is a crucial memory barrier after the
541          * list_del_init() and also before assigning to q->lock_ptr.
542          */
543         wake_up_all(&q->waiters);
544         /*
545          * The waiting task can free the futex_q as soon as this is written,
546          * without taking any locks.  This must come last.
547          *
548          * A memory barrier is required here to prevent the following store
549          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
550          * at the end of wake_up_all() does not prevent this store from
551          * moving.
552          */
553         wmb();
554         q->lock_ptr = NULL;
555 }
556
557 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
558 {
559         struct task_struct *new_owner;
560         struct futex_pi_state *pi_state = this->pi_state;
561         u32 curval, newval;
562
563         if (!pi_state)
564                 return -EINVAL;
565
566         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
567
568         /*
569          * This happens when we have stolen the lock and the original
570          * pending owner did not enqueue itself back on the rt_mutex.
571          * Thats not a tragedy. We know that way, that a lock waiter
572          * is on the fly. We make the futex_q waiter the pending owner.
573          */
574         if (!new_owner)
575                 new_owner = this->task;
576
577         /*
578          * We pass it to the next owner. (The WAITERS bit is always
579          * kept enabled while there is PI state around. We must also
580          * preserve the owner died bit.)
581          */
582         newval = (uval & FUTEX_OWNER_DIED) | FUTEX_WAITERS | new_owner->pid;
583
584         inc_preempt_count();
585         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
586         dec_preempt_count();
587
588         if (curval == -EFAULT)
589                 return -EFAULT;
590         if (curval != uval)
591                 return -EINVAL;
592
593         spin_lock_irq(&pi_state->owner->pi_lock);
594         WARN_ON(list_empty(&pi_state->list));
595         list_del_init(&pi_state->list);
596         spin_unlock_irq(&pi_state->owner->pi_lock);
597
598         spin_lock_irq(&new_owner->pi_lock);
599         WARN_ON(!list_empty(&pi_state->list));
600         list_add(&pi_state->list, &new_owner->pi_state_list);
601         pi_state->owner = new_owner;
602         spin_unlock_irq(&new_owner->pi_lock);
603
604         rt_mutex_unlock(&pi_state->pi_mutex);
605
606         return 0;
607 }
608
609 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
610 {
611         u32 oldval;
612
613         /*
614          * There is no waiter, so we unlock the futex. The owner died
615          * bit has not to be preserved here. We are the owner:
616          */
617         inc_preempt_count();
618         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
619         dec_preempt_count();
620
621         if (oldval == -EFAULT)
622                 return oldval;
623         if (oldval != uval)
624                 return -EAGAIN;
625
626         return 0;
627 }
628
629 /*
630  * Express the locking dependencies for lockdep:
631  */
632 static inline void
633 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
634 {
635         if (hb1 <= hb2) {
636                 spin_lock(&hb1->lock);
637                 if (hb1 < hb2)
638                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
639         } else { /* hb1 > hb2 */
640                 spin_lock(&hb2->lock);
641                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
642         }
643 }
644
645 /*
646  * Wake up all waiters hashed on the physical page that is mapped
647  * to this virtual address:
648  */
649 static int futex_wake(u32 __user *uaddr, int nr_wake)
650 {
651         struct futex_hash_bucket *hb;
652         struct futex_q *this, *next;
653         struct list_head *head;
654         union futex_key key;
655         int ret;
656
657         down_read(&current->mm->mmap_sem);
658
659         ret = get_futex_key(uaddr, &key);
660         if (unlikely(ret != 0))
661                 goto out;
662
663         hb = hash_futex(&key);
664         spin_lock(&hb->lock);
665         head = &hb->chain;
666
667         list_for_each_entry_safe(this, next, head, list) {
668                 if (match_futex (&this->key, &key)) {
669                         if (this->pi_state) {
670                                 ret = -EINVAL;
671                                 break;
672                         }
673                         wake_futex(this);
674                         if (++ret >= nr_wake)
675                                 break;
676                 }
677         }
678
679         spin_unlock(&hb->lock);
680 out:
681         up_read(&current->mm->mmap_sem);
682         return ret;
683 }
684
685 /*
686  * Wake up all waiters hashed on the physical page that is mapped
687  * to this virtual address:
688  */
689 static int
690 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
691               int nr_wake, int nr_wake2, int op)
692 {
693         union futex_key key1, key2;
694         struct futex_hash_bucket *hb1, *hb2;
695         struct list_head *head;
696         struct futex_q *this, *next;
697         int ret, op_ret, attempt = 0;
698
699 retryfull:
700         down_read(&current->mm->mmap_sem);
701
702         ret = get_futex_key(uaddr1, &key1);
703         if (unlikely(ret != 0))
704                 goto out;
705         ret = get_futex_key(uaddr2, &key2);
706         if (unlikely(ret != 0))
707                 goto out;
708
709         hb1 = hash_futex(&key1);
710         hb2 = hash_futex(&key2);
711
712 retry:
713         double_lock_hb(hb1, hb2);
714
715         op_ret = futex_atomic_op_inuser(op, uaddr2);
716         if (unlikely(op_ret < 0)) {
717                 u32 dummy;
718
719                 spin_unlock(&hb1->lock);
720                 if (hb1 != hb2)
721                         spin_unlock(&hb2->lock);
722
723 #ifndef CONFIG_MMU
724                 /*
725                  * we don't get EFAULT from MMU faults if we don't have an MMU,
726                  * but we might get them from range checking
727                  */
728                 ret = op_ret;
729                 goto out;
730 #endif
731
732                 if (unlikely(op_ret != -EFAULT)) {
733                         ret = op_ret;
734                         goto out;
735                 }
736
737                 /*
738                  * futex_atomic_op_inuser needs to both read and write
739                  * *(int __user *)uaddr2, but we can't modify it
740                  * non-atomically.  Therefore, if get_user below is not
741                  * enough, we need to handle the fault ourselves, while
742                  * still holding the mmap_sem.
743                  */
744                 if (attempt++) {
745                         if (futex_handle_fault((unsigned long)uaddr2,
746                                                attempt))
747                                 goto out;
748                         goto retry;
749                 }
750
751                 /*
752                  * If we would have faulted, release mmap_sem,
753                  * fault it in and start all over again.
754                  */
755                 up_read(&current->mm->mmap_sem);
756
757                 ret = get_user(dummy, uaddr2);
758                 if (ret)
759                         return ret;
760
761                 goto retryfull;
762         }
763
764         head = &hb1->chain;
765
766         list_for_each_entry_safe(this, next, head, list) {
767                 if (match_futex (&this->key, &key1)) {
768                         wake_futex(this);
769                         if (++ret >= nr_wake)
770                                 break;
771                 }
772         }
773
774         if (op_ret > 0) {
775                 head = &hb2->chain;
776
777                 op_ret = 0;
778                 list_for_each_entry_safe(this, next, head, list) {
779                         if (match_futex (&this->key, &key2)) {
780                                 wake_futex(this);
781                                 if (++op_ret >= nr_wake2)
782                                         break;
783                         }
784                 }
785                 ret += op_ret;
786         }
787
788         spin_unlock(&hb1->lock);
789         if (hb1 != hb2)
790                 spin_unlock(&hb2->lock);
791 out:
792         up_read(&current->mm->mmap_sem);
793         return ret;
794 }
795
796 /*
797  * Requeue all waiters hashed on one physical page to another
798  * physical page.
799  */
800 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
801                          int nr_wake, int nr_requeue, u32 *cmpval)
802 {
803         union futex_key key1, key2;
804         struct futex_hash_bucket *hb1, *hb2;
805         struct list_head *head1;
806         struct futex_q *this, *next;
807         int ret, drop_count = 0;
808
809  retry:
810         down_read(&current->mm->mmap_sem);
811
812         ret = get_futex_key(uaddr1, &key1);
813         if (unlikely(ret != 0))
814                 goto out;
815         ret = get_futex_key(uaddr2, &key2);
816         if (unlikely(ret != 0))
817                 goto out;
818
819         hb1 = hash_futex(&key1);
820         hb2 = hash_futex(&key2);
821
822         double_lock_hb(hb1, hb2);
823
824         if (likely(cmpval != NULL)) {
825                 u32 curval;
826
827                 ret = get_futex_value_locked(&curval, uaddr1);
828
829                 if (unlikely(ret)) {
830                         spin_unlock(&hb1->lock);
831                         if (hb1 != hb2)
832                                 spin_unlock(&hb2->lock);
833
834                         /*
835                          * If we would have faulted, release mmap_sem, fault
836                          * it in and start all over again.
837                          */
838                         up_read(&current->mm->mmap_sem);
839
840                         ret = get_user(curval, uaddr1);
841
842                         if (!ret)
843                                 goto retry;
844
845                         return ret;
846                 }
847                 if (curval != *cmpval) {
848                         ret = -EAGAIN;
849                         goto out_unlock;
850                 }
851         }
852
853         head1 = &hb1->chain;
854         list_for_each_entry_safe(this, next, head1, list) {
855                 if (!match_futex (&this->key, &key1))
856                         continue;
857                 if (++ret <= nr_wake) {
858                         wake_futex(this);
859                 } else {
860                         /*
861                          * If key1 and key2 hash to the same bucket, no need to
862                          * requeue.
863                          */
864                         if (likely(head1 != &hb2->chain)) {
865                                 list_move_tail(&this->list, &hb2->chain);
866                                 this->lock_ptr = &hb2->lock;
867                         }
868                         this->key = key2;
869                         get_key_refs(&key2);
870                         drop_count++;
871
872                         if (ret - nr_wake >= nr_requeue)
873                                 break;
874                 }
875         }
876
877 out_unlock:
878         spin_unlock(&hb1->lock);
879         if (hb1 != hb2)
880                 spin_unlock(&hb2->lock);
881
882         /* drop_key_refs() must be called outside the spinlocks. */
883         while (--drop_count >= 0)
884                 drop_key_refs(&key1);
885
886 out:
887         up_read(&current->mm->mmap_sem);
888         return ret;
889 }
890
891 /* The key must be already stored in q->key. */
892 static inline struct futex_hash_bucket *
893 queue_lock(struct futex_q *q, int fd, struct file *filp)
894 {
895         struct futex_hash_bucket *hb;
896
897         q->fd = fd;
898         q->filp = filp;
899
900         init_waitqueue_head(&q->waiters);
901
902         get_key_refs(&q->key);
903         hb = hash_futex(&q->key);
904         q->lock_ptr = &hb->lock;
905
906         spin_lock(&hb->lock);
907         return hb;
908 }
909
910 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
911 {
912         list_add_tail(&q->list, &hb->chain);
913         q->task = current;
914         spin_unlock(&hb->lock);
915 }
916
917 static inline void
918 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
919 {
920         spin_unlock(&hb->lock);
921         drop_key_refs(&q->key);
922 }
923
924 /*
925  * queue_me and unqueue_me must be called as a pair, each
926  * exactly once.  They are called with the hashed spinlock held.
927  */
928
929 /* The key must be already stored in q->key. */
930 static void queue_me(struct futex_q *q, int fd, struct file *filp)
931 {
932         struct futex_hash_bucket *hb;
933
934         hb = queue_lock(q, fd, filp);
935         __queue_me(q, hb);
936 }
937
938 /* Return 1 if we were still queued (ie. 0 means we were woken) */
939 static int unqueue_me(struct futex_q *q)
940 {
941         spinlock_t *lock_ptr;
942         int ret = 0;
943
944         /* In the common case we don't take the spinlock, which is nice. */
945  retry:
946         lock_ptr = q->lock_ptr;
947         if (lock_ptr != 0) {
948                 spin_lock(lock_ptr);
949                 /*
950                  * q->lock_ptr can change between reading it and
951                  * spin_lock(), causing us to take the wrong lock.  This
952                  * corrects the race condition.
953                  *
954                  * Reasoning goes like this: if we have the wrong lock,
955                  * q->lock_ptr must have changed (maybe several times)
956                  * between reading it and the spin_lock().  It can
957                  * change again after the spin_lock() but only if it was
958                  * already changed before the spin_lock().  It cannot,
959                  * however, change back to the original value.  Therefore
960                  * we can detect whether we acquired the correct lock.
961                  */
962                 if (unlikely(lock_ptr != q->lock_ptr)) {
963                         spin_unlock(lock_ptr);
964                         goto retry;
965                 }
966                 WARN_ON(list_empty(&q->list));
967                 list_del(&q->list);
968
969                 BUG_ON(q->pi_state);
970
971                 spin_unlock(lock_ptr);
972                 ret = 1;
973         }
974
975         drop_key_refs(&q->key);
976         return ret;
977 }
978
979 /*
980  * PI futexes can not be requeued and must remove themself from the
981  * hash bucket. The hash bucket lock is held on entry and dropped here.
982  */
983 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
984 {
985         WARN_ON(list_empty(&q->list));
986         list_del(&q->list);
987
988         BUG_ON(!q->pi_state);
989         free_pi_state(q->pi_state);
990         q->pi_state = NULL;
991
992         spin_unlock(&hb->lock);
993
994         drop_key_refs(&q->key);
995 }
996
997 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
998 {
999         struct task_struct *curr = current;
1000         DECLARE_WAITQUEUE(wait, curr);
1001         struct futex_hash_bucket *hb;
1002         struct futex_q q;
1003         u32 uval;
1004         int ret;
1005
1006         q.pi_state = NULL;
1007  retry:
1008         down_read(&curr->mm->mmap_sem);
1009
1010         ret = get_futex_key(uaddr, &q.key);
1011         if (unlikely(ret != 0))
1012                 goto out_release_sem;
1013
1014         hb = queue_lock(&q, -1, NULL);
1015
1016         /*
1017          * Access the page AFTER the futex is queued.
1018          * Order is important:
1019          *
1020          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1021          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1022          *
1023          * The basic logical guarantee of a futex is that it blocks ONLY
1024          * if cond(var) is known to be true at the time of blocking, for
1025          * any cond.  If we queued after testing *uaddr, that would open
1026          * a race condition where we could block indefinitely with
1027          * cond(var) false, which would violate the guarantee.
1028          *
1029          * A consequence is that futex_wait() can return zero and absorb
1030          * a wakeup when *uaddr != val on entry to the syscall.  This is
1031          * rare, but normal.
1032          *
1033          * We hold the mmap semaphore, so the mapping cannot have changed
1034          * since we looked it up in get_futex_key.
1035          */
1036         ret = get_futex_value_locked(&uval, uaddr);
1037
1038         if (unlikely(ret)) {
1039                 queue_unlock(&q, hb);
1040
1041                 /*
1042                  * If we would have faulted, release mmap_sem, fault it in and
1043                  * start all over again.
1044                  */
1045                 up_read(&curr->mm->mmap_sem);
1046
1047                 ret = get_user(uval, uaddr);
1048
1049                 if (!ret)
1050                         goto retry;
1051                 return ret;
1052         }
1053         ret = -EWOULDBLOCK;
1054         if (uval != val)
1055                 goto out_unlock_release_sem;
1056
1057         /* Only actually queue if *uaddr contained val.  */
1058         __queue_me(&q, hb);
1059
1060         /*
1061          * Now the futex is queued and we have checked the data, we
1062          * don't want to hold mmap_sem while we sleep.
1063          */
1064         up_read(&curr->mm->mmap_sem);
1065
1066         /*
1067          * There might have been scheduling since the queue_me(), as we
1068          * cannot hold a spinlock across the get_user() in case it
1069          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1070          * queueing ourselves into the futex hash.  This code thus has to
1071          * rely on the futex_wake() code removing us from hash when it
1072          * wakes us up.
1073          */
1074
1075         /* add_wait_queue is the barrier after __set_current_state. */
1076         __set_current_state(TASK_INTERRUPTIBLE);
1077         add_wait_queue(&q.waiters, &wait);
1078         /*
1079          * !list_empty() is safe here without any lock.
1080          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1081          */
1082         if (likely(!list_empty(&q.list)))
1083                 time = schedule_timeout(time);
1084         __set_current_state(TASK_RUNNING);
1085
1086         /*
1087          * NOTE: we don't remove ourselves from the waitqueue because
1088          * we are the only user of it.
1089          */
1090
1091         /* If we were woken (and unqueued), we succeeded, whatever. */
1092         if (!unqueue_me(&q))
1093                 return 0;
1094         if (time == 0)
1095                 return -ETIMEDOUT;
1096         /*
1097          * We expect signal_pending(current), but another thread may
1098          * have handled it for us already.
1099          */
1100         return -EINTR;
1101
1102  out_unlock_release_sem:
1103         queue_unlock(&q, hb);
1104
1105  out_release_sem:
1106         up_read(&curr->mm->mmap_sem);
1107         return ret;
1108 }
1109
1110 /*
1111  * Userspace tried a 0 -> TID atomic transition of the futex value
1112  * and failed. The kernel side here does the whole locking operation:
1113  * if there are waiters then it will block, it does PI, etc. (Due to
1114  * races the kernel might see a 0 value of the futex too.)
1115  */
1116 static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
1117                             struct hrtimer_sleeper *to)
1118 {
1119         struct task_struct *curr = current;
1120         struct futex_hash_bucket *hb;
1121         u32 uval, newval, curval;
1122         struct futex_q q;
1123         int ret, attempt = 0;
1124
1125         if (refill_pi_state_cache())
1126                 return -ENOMEM;
1127
1128         q.pi_state = NULL;
1129  retry:
1130         down_read(&curr->mm->mmap_sem);
1131
1132         ret = get_futex_key(uaddr, &q.key);
1133         if (unlikely(ret != 0))
1134                 goto out_release_sem;
1135
1136         hb = queue_lock(&q, -1, NULL);
1137
1138  retry_locked:
1139         /*
1140          * To avoid races, we attempt to take the lock here again
1141          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1142          * the locks. It will most likely not succeed.
1143          */
1144         newval = current->pid;
1145
1146         inc_preempt_count();
1147         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1148         dec_preempt_count();
1149
1150         if (unlikely(curval == -EFAULT))
1151                 goto uaddr_faulted;
1152
1153         /* We own the lock already */
1154         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1155                 if (!detect && 0)
1156                         force_sig(SIGKILL, current);
1157                 ret = -EDEADLK;
1158                 goto out_unlock_release_sem;
1159         }
1160
1161         /*
1162          * Surprise - we got the lock. Just return
1163          * to userspace:
1164          */
1165         if (unlikely(!curval))
1166                 goto out_unlock_release_sem;
1167
1168         uval = curval;
1169         newval = uval | FUTEX_WAITERS;
1170
1171         inc_preempt_count();
1172         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1173         dec_preempt_count();
1174
1175         if (unlikely(curval == -EFAULT))
1176                 goto uaddr_faulted;
1177         if (unlikely(curval != uval))
1178                 goto retry_locked;
1179
1180         /*
1181          * We dont have the lock. Look up the PI state (or create it if
1182          * we are the first waiter):
1183          */
1184         ret = lookup_pi_state(uval, hb, &q);
1185
1186         if (unlikely(ret)) {
1187                 /*
1188                  * There were no waiters and the owner task lookup
1189                  * failed. When the OWNER_DIED bit is set, then we
1190                  * know that this is a robust futex and we actually
1191                  * take the lock. This is safe as we are protected by
1192                  * the hash bucket lock. We also set the waiters bit
1193                  * unconditionally here, to simplify glibc handling of
1194                  * multiple tasks racing to acquire the lock and
1195                  * cleanup the problems which were left by the dead
1196                  * owner.
1197                  */
1198                 if (curval & FUTEX_OWNER_DIED) {
1199                         uval = newval;
1200                         newval = current->pid |
1201                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1202
1203                         inc_preempt_count();
1204                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1205                                                                uval, newval);
1206                         dec_preempt_count();
1207
1208                         if (unlikely(curval == -EFAULT))
1209                                 goto uaddr_faulted;
1210                         if (unlikely(curval != uval))
1211                                 goto retry_locked;
1212                         ret = 0;
1213                 }
1214                 goto out_unlock_release_sem;
1215         }
1216
1217         /*
1218          * Only actually queue now that the atomic ops are done:
1219          */
1220         __queue_me(&q, hb);
1221
1222         /*
1223          * Now the futex is queued and we have checked the data, we
1224          * don't want to hold mmap_sem while we sleep.
1225          */
1226         up_read(&curr->mm->mmap_sem);
1227
1228         WARN_ON(!q.pi_state);
1229         /*
1230          * Block on the PI mutex:
1231          */
1232         if (!trylock)
1233                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1234         else {
1235                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1236                 /* Fixup the trylock return value: */
1237                 ret = ret ? 0 : -EWOULDBLOCK;
1238         }
1239
1240         down_read(&curr->mm->mmap_sem);
1241         spin_lock(q.lock_ptr);
1242
1243         /*
1244          * Got the lock. We might not be the anticipated owner if we
1245          * did a lock-steal - fix up the PI-state in that case.
1246          */
1247         if (!ret && q.pi_state->owner != curr) {
1248                 u32 newtid = current->pid | FUTEX_WAITERS;
1249
1250                 /* Owner died? */
1251                 if (q.pi_state->owner != NULL) {
1252                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1253                         WARN_ON(list_empty(&q.pi_state->list));
1254                         list_del_init(&q.pi_state->list);
1255                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1256                 } else
1257                         newtid |= FUTEX_OWNER_DIED;
1258
1259                 q.pi_state->owner = current;
1260
1261                 spin_lock_irq(&current->pi_lock);
1262                 WARN_ON(!list_empty(&q.pi_state->list));
1263                 list_add(&q.pi_state->list, &current->pi_state_list);
1264                 spin_unlock_irq(&current->pi_lock);
1265
1266                 /* Unqueue and drop the lock */
1267                 unqueue_me_pi(&q, hb);
1268                 up_read(&curr->mm->mmap_sem);
1269                 /*
1270                  * We own it, so we have to replace the pending owner
1271                  * TID. This must be atomic as we have preserve the
1272                  * owner died bit here.
1273                  */
1274                 ret = get_user(uval, uaddr);
1275                 while (!ret) {
1276                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1277                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1278                                                                uval, newval);
1279                         if (curval == -EFAULT)
1280                                 ret = -EFAULT;
1281                         if (curval == uval)
1282                                 break;
1283                         uval = curval;
1284                 }
1285         } else {
1286                 /*
1287                  * Catch the rare case, where the lock was released
1288                  * when we were on the way back before we locked
1289                  * the hash bucket.
1290                  */
1291                 if (ret && q.pi_state->owner == curr) {
1292                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1293                                 ret = 0;
1294                 }
1295                 /* Unqueue and drop the lock */
1296                 unqueue_me_pi(&q, hb);
1297                 up_read(&curr->mm->mmap_sem);
1298         }
1299
1300         if (!detect && ret == -EDEADLK && 0)
1301                 force_sig(SIGKILL, current);
1302
1303         return ret;
1304
1305  out_unlock_release_sem:
1306         queue_unlock(&q, hb);
1307
1308  out_release_sem:
1309         up_read(&curr->mm->mmap_sem);
1310         return ret;
1311
1312  uaddr_faulted:
1313         /*
1314          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1315          * non-atomically.  Therefore, if get_user below is not
1316          * enough, we need to handle the fault ourselves, while
1317          * still holding the mmap_sem.
1318          */
1319         if (attempt++) {
1320                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1321                         goto out_unlock_release_sem;
1322
1323                 goto retry_locked;
1324         }
1325
1326         queue_unlock(&q, hb);
1327         up_read(&curr->mm->mmap_sem);
1328
1329         ret = get_user(uval, uaddr);
1330         if (!ret && (uval != -EFAULT))
1331                 goto retry;
1332
1333         return ret;
1334 }
1335
1336 /*
1337  * Restart handler
1338  */
1339 static long futex_lock_pi_restart(struct restart_block *restart)
1340 {
1341         struct hrtimer_sleeper timeout, *to = NULL;
1342         int ret;
1343
1344         restart->fn = do_no_restart_syscall;
1345
1346         if (restart->arg2 || restart->arg3) {
1347                 to = &timeout;
1348                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1349                 hrtimer_init_sleeper(to, current);
1350                 to->timer.expires.tv64 = ((u64)restart->arg1 << 32) |
1351                         (u64) restart->arg0;
1352         }
1353
1354         pr_debug("lock_pi restart: %p, %d (%d)\n",
1355                  (u32 __user *)restart->arg0, current->pid);
1356
1357         ret = do_futex_lock_pi((u32 __user *)restart->arg0, restart->arg1,
1358                                0, to);
1359
1360         if (ret != -EINTR)
1361                 return ret;
1362
1363         restart->fn = futex_lock_pi_restart;
1364
1365         /* The other values are filled in */
1366         return -ERESTART_RESTARTBLOCK;
1367 }
1368
1369 /*
1370  * Called from the syscall entry below.
1371  */
1372 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1373                          long nsec, int trylock)
1374 {
1375         struct hrtimer_sleeper timeout, *to = NULL;
1376         struct restart_block *restart;
1377         int ret;
1378
1379         if (sec != MAX_SCHEDULE_TIMEOUT) {
1380                 to = &timeout;
1381                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1382                 hrtimer_init_sleeper(to, current);
1383                 to->timer.expires = ktime_set(sec, nsec);
1384         }
1385
1386         ret = do_futex_lock_pi(uaddr, detect, trylock, to);
1387
1388         if (ret != -EINTR)
1389                 return ret;
1390
1391         pr_debug("lock_pi interrupted: %p, %d (%d)\n", uaddr, current->pid);
1392
1393         restart = &current_thread_info()->restart_block;
1394         restart->fn = futex_lock_pi_restart;
1395         restart->arg0 = (unsigned long) uaddr;
1396         restart->arg1 = detect;
1397         if (to) {
1398                 restart->arg2 = to->timer.expires.tv64 & 0xFFFFFFFF;
1399                 restart->arg3 = to->timer.expires.tv64 >> 32;
1400         } else
1401                 restart->arg2 = restart->arg3 = 0;
1402
1403         return -ERESTART_RESTARTBLOCK;
1404 }
1405
1406 /*
1407  * Userspace attempted a TID -> 0 atomic transition, and failed.
1408  * This is the in-kernel slowpath: we look up the PI state (if any),
1409  * and do the rt-mutex unlock.
1410  */
1411 static int futex_unlock_pi(u32 __user *uaddr)
1412 {
1413         struct futex_hash_bucket *hb;
1414         struct futex_q *this, *next;
1415         u32 uval;
1416         struct list_head *head;
1417         union futex_key key;
1418         int ret, attempt = 0;
1419
1420 retry:
1421         if (get_user(uval, uaddr))
1422                 return -EFAULT;
1423         /*
1424          * We release only a lock we actually own:
1425          */
1426         if ((uval & FUTEX_TID_MASK) != current->pid)
1427                 return -EPERM;
1428         /*
1429          * First take all the futex related locks:
1430          */
1431         down_read(&current->mm->mmap_sem);
1432
1433         ret = get_futex_key(uaddr, &key);
1434         if (unlikely(ret != 0))
1435                 goto out;
1436
1437         hb = hash_futex(&key);
1438         spin_lock(&hb->lock);
1439
1440 retry_locked:
1441         /*
1442          * To avoid races, try to do the TID -> 0 atomic transition
1443          * again. If it succeeds then we can return without waking
1444          * anyone else up:
1445          */
1446         inc_preempt_count();
1447         uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1448         dec_preempt_count();
1449
1450         if (unlikely(uval == -EFAULT))
1451                 goto pi_faulted;
1452         /*
1453          * Rare case: we managed to release the lock atomically,
1454          * no need to wake anyone else up:
1455          */
1456         if (unlikely(uval == current->pid))
1457                 goto out_unlock;
1458
1459         /*
1460          * Ok, other tasks may need to be woken up - check waiters
1461          * and do the wakeup if necessary:
1462          */
1463         head = &hb->chain;
1464
1465         list_for_each_entry_safe(this, next, head, list) {
1466                 if (!match_futex (&this->key, &key))
1467                         continue;
1468                 ret = wake_futex_pi(uaddr, uval, this);
1469                 /*
1470                  * The atomic access to the futex value
1471                  * generated a pagefault, so retry the
1472                  * user-access and the wakeup:
1473                  */
1474                 if (ret == -EFAULT)
1475                         goto pi_faulted;
1476                 goto out_unlock;
1477         }
1478         /*
1479          * No waiters - kernel unlocks the futex:
1480          */
1481         ret = unlock_futex_pi(uaddr, uval);
1482         if (ret == -EFAULT)
1483                 goto pi_faulted;
1484
1485 out_unlock:
1486         spin_unlock(&hb->lock);
1487 out:
1488         up_read(&current->mm->mmap_sem);
1489
1490         return ret;
1491
1492 pi_faulted:
1493         /*
1494          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1495          * non-atomically.  Therefore, if get_user below is not
1496          * enough, we need to handle the fault ourselves, while
1497          * still holding the mmap_sem.
1498          */
1499         if (attempt++) {
1500                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1501                         goto out_unlock;
1502
1503                 goto retry_locked;
1504         }
1505
1506         spin_unlock(&hb->lock);
1507         up_read(&current->mm->mmap_sem);
1508
1509         ret = get_user(uval, uaddr);
1510         if (!ret && (uval != -EFAULT))
1511                 goto retry;
1512
1513         return ret;
1514 }
1515
1516 static int futex_close(struct inode *inode, struct file *filp)
1517 {
1518         struct futex_q *q = filp->private_data;
1519
1520         unqueue_me(q);
1521         kfree(q);
1522
1523         return 0;
1524 }
1525
1526 /* This is one-shot: once it's gone off you need a new fd */
1527 static unsigned int futex_poll(struct file *filp,
1528                                struct poll_table_struct *wait)
1529 {
1530         struct futex_q *q = filp->private_data;
1531         int ret = 0;
1532
1533         poll_wait(filp, &q->waiters, wait);
1534
1535         /*
1536          * list_empty() is safe here without any lock.
1537          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1538          */
1539         if (list_empty(&q->list))
1540                 ret = POLLIN | POLLRDNORM;
1541
1542         return ret;
1543 }
1544
1545 static struct file_operations futex_fops = {
1546         .release        = futex_close,
1547         .poll           = futex_poll,
1548 };
1549
1550 /*
1551  * Signal allows caller to avoid the race which would occur if they
1552  * set the sigio stuff up afterwards.
1553  */
1554 static int futex_fd(u32 __user *uaddr, int signal)
1555 {
1556         struct futex_q *q;
1557         struct file *filp;
1558         int ret, err;
1559
1560         ret = -EINVAL;
1561         if (!valid_signal(signal))
1562                 goto out;
1563
1564         ret = get_unused_fd();
1565         if (ret < 0)
1566                 goto out;
1567         filp = get_empty_filp();
1568         if (!filp) {
1569                 put_unused_fd(ret);
1570                 ret = -ENFILE;
1571                 goto out;
1572         }
1573         filp->f_op = &futex_fops;
1574         filp->f_vfsmnt = mntget(futex_mnt);
1575         filp->f_dentry = dget(futex_mnt->mnt_root);
1576         filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
1577
1578         if (signal) {
1579                 err = f_setown(filp, current->pid, 1);
1580                 if (err < 0) {
1581                         goto error;
1582                 }
1583                 filp->f_owner.signum = signal;
1584         }
1585
1586         q = kmalloc(sizeof(*q), GFP_KERNEL);
1587         if (!q) {
1588                 err = -ENOMEM;
1589                 goto error;
1590         }
1591         q->pi_state = NULL;
1592
1593         down_read(&current->mm->mmap_sem);
1594         err = get_futex_key(uaddr, &q->key);
1595
1596         if (unlikely(err != 0)) {
1597                 up_read(&current->mm->mmap_sem);
1598                 kfree(q);
1599                 goto error;
1600         }
1601
1602         /*
1603          * queue_me() must be called before releasing mmap_sem, because
1604          * key->shared.inode needs to be referenced while holding it.
1605          */
1606         filp->private_data = q;
1607
1608         queue_me(q, ret, filp);
1609         up_read(&current->mm->mmap_sem);
1610
1611         /* Now we map fd to filp, so userspace can access it */
1612         fd_install(ret, filp);
1613 out:
1614         return ret;
1615 error:
1616         put_unused_fd(ret);
1617         put_filp(filp);
1618         ret = err;
1619         goto out;
1620 }
1621
1622 /*
1623  * Support for robust futexes: the kernel cleans up held futexes at
1624  * thread exit time.
1625  *
1626  * Implementation: user-space maintains a per-thread list of locks it
1627  * is holding. Upon do_exit(), the kernel carefully walks this list,
1628  * and marks all locks that are owned by this thread with the
1629  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1630  * always manipulated with the lock held, so the list is private and
1631  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1632  * field, to allow the kernel to clean up if the thread dies after
1633  * acquiring the lock, but just before it could have added itself to
1634  * the list. There can only be one such pending lock.
1635  */
1636
1637 /**
1638  * sys_set_robust_list - set the robust-futex list head of a task
1639  * @head: pointer to the list-head
1640  * @len: length of the list-head, as userspace expects
1641  */
1642 asmlinkage long
1643 sys_set_robust_list(struct robust_list_head __user *head,
1644                     size_t len)
1645 {
1646         /*
1647          * The kernel knows only one size for now:
1648          */
1649         if (unlikely(len != sizeof(*head)))
1650                 return -EINVAL;
1651
1652         current->robust_list = head;
1653
1654         return 0;
1655 }
1656
1657 /**
1658  * sys_get_robust_list - get the robust-futex list head of a task
1659  * @pid: pid of the process [zero for current task]
1660  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1661  * @len_ptr: pointer to a length field, the kernel fills in the header size
1662  */
1663 asmlinkage long
1664 sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
1665                     size_t __user *len_ptr)
1666 {
1667         struct robust_list_head *head;
1668         unsigned long ret;
1669
1670         if (!pid)
1671                 head = current->robust_list;
1672         else {
1673                 struct task_struct *p;
1674
1675                 ret = -ESRCH;
1676                 read_lock(&tasklist_lock);
1677                 p = find_task_by_pid(pid);
1678                 if (!p)
1679                         goto err_unlock;
1680                 ret = -EPERM;
1681                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1682                                 !capable(CAP_SYS_PTRACE))
1683                         goto err_unlock;
1684                 head = p->robust_list;
1685                 read_unlock(&tasklist_lock);
1686         }
1687
1688         if (put_user(sizeof(*head), len_ptr))
1689                 return -EFAULT;
1690         return put_user(head, head_ptr);
1691
1692 err_unlock:
1693         read_unlock(&tasklist_lock);
1694
1695         return ret;
1696 }
1697
1698 /*
1699  * Process a futex-list entry, check whether it's owned by the
1700  * dying task, and do notification if so:
1701  */
1702 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr)
1703 {
1704         u32 uval, nval;
1705
1706 retry:
1707         if (get_user(uval, uaddr))
1708                 return -1;
1709
1710         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1711                 /*
1712                  * Ok, this dying thread is truly holding a futex
1713                  * of interest. Set the OWNER_DIED bit atomically
1714                  * via cmpxchg, and if the value had FUTEX_WAITERS
1715                  * set, wake up a waiter (if any). (We have to do a
1716                  * futex_wake() even if OWNER_DIED is already set -
1717                  * to handle the rare but possible case of recursive
1718                  * thread-death.) The rest of the cleanup is done in
1719                  * userspace.
1720                  */
1721                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval,
1722                                                      uval | FUTEX_OWNER_DIED);
1723                 if (nval == -EFAULT)
1724                         return -1;
1725
1726                 if (nval != uval)
1727                         goto retry;
1728
1729                 if (uval & FUTEX_WAITERS)
1730                         futex_wake(uaddr, 1);
1731         }
1732         return 0;
1733 }
1734
1735 /*
1736  * Walk curr->robust_list (very carefully, it's a userspace list!)
1737  * and mark any locks found there dead, and notify any waiters.
1738  *
1739  * We silently return on any sign of list-walking problem.
1740  */
1741 void exit_robust_list(struct task_struct *curr)
1742 {
1743         struct robust_list_head __user *head = curr->robust_list;
1744         struct robust_list __user *entry, *pending;
1745         unsigned int limit = ROBUST_LIST_LIMIT;
1746         unsigned long futex_offset;
1747
1748         /*
1749          * Fetch the list head (which was registered earlier, via
1750          * sys_set_robust_list()):
1751          */
1752         if (get_user(entry, &head->list.next))
1753                 return;
1754         /*
1755          * Fetch the relative futex offset:
1756          */
1757         if (get_user(futex_offset, &head->futex_offset))
1758                 return;
1759         /*
1760          * Fetch any possibly pending lock-add first, and handle it
1761          * if it exists:
1762          */
1763         if (get_user(pending, &head->list_op_pending))
1764                 return;
1765         if (pending)
1766                 handle_futex_death((void *)pending + futex_offset, curr);
1767
1768         while (entry != &head->list) {
1769                 /*
1770                  * A pending lock might already be on the list, so
1771                  * don't process it twice:
1772                  */
1773                 if (entry != pending)
1774                         if (handle_futex_death((void *)entry + futex_offset,
1775                                                 curr))
1776                                 return;
1777                 /*
1778                  * Fetch the next entry in the list:
1779                  */
1780                 if (get_user(entry, &entry->next))
1781                         return;
1782                 /*
1783                  * Avoid excessively long or circular lists:
1784                  */
1785                 if (!--limit)
1786                         break;
1787
1788                 cond_resched();
1789         }
1790 }
1791
1792 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1793                 u32 __user *uaddr2, u32 val2, u32 val3)
1794 {
1795         int ret;
1796
1797         switch (op) {
1798         case FUTEX_WAIT:
1799                 ret = futex_wait(uaddr, val, timeout);
1800                 break;
1801         case FUTEX_WAKE:
1802                 ret = futex_wake(uaddr, val);
1803                 break;
1804         case FUTEX_FD:
1805                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1806                 ret = futex_fd(uaddr, val);
1807                 break;
1808         case FUTEX_REQUEUE:
1809                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1810                 break;
1811         case FUTEX_CMP_REQUEUE:
1812                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1813                 break;
1814         case FUTEX_WAKE_OP:
1815                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1816                 break;
1817         case FUTEX_LOCK_PI:
1818                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1819                 break;
1820         case FUTEX_UNLOCK_PI:
1821                 ret = futex_unlock_pi(uaddr);
1822                 break;
1823         case FUTEX_TRYLOCK_PI:
1824                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1825                 break;
1826         default:
1827                 ret = -ENOSYS;
1828         }
1829         return ret;
1830 }
1831
1832
1833 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1834                           struct timespec __user *utime, u32 __user *uaddr2,
1835                           u32 val3)
1836 {
1837         struct timespec t;
1838         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1839         u32 val2 = 0;
1840
1841         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1842                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1843                         return -EFAULT;
1844                 if (!timespec_valid(&t))
1845                         return -EINVAL;
1846                 if (op == FUTEX_WAIT)
1847                         timeout = timespec_to_jiffies(&t) + 1;
1848                 else {
1849                         timeout = t.tv_sec;
1850                         val2 = t.tv_nsec;
1851                 }
1852         }
1853         /*
1854          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1855          */
1856         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1857                 val2 = (u32) (unsigned long) utime;
1858
1859         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1860 }
1861
1862 static int futexfs_get_sb(struct file_system_type *fs_type,
1863                           int flags, const char *dev_name, void *data,
1864                           struct vfsmount *mnt)
1865 {
1866         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1867 }
1868
1869 static struct file_system_type futex_fs_type = {
1870         .name           = "futexfs",
1871         .get_sb         = futexfs_get_sb,
1872         .kill_sb        = kill_anon_super,
1873 };
1874
1875 static int __init init(void)
1876 {
1877         unsigned int i;
1878
1879         register_filesystem(&futex_fs_type);
1880         futex_mnt = kern_mount(&futex_fs_type);
1881
1882         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1883                 INIT_LIST_HEAD(&futex_queues[i].chain);
1884                 spin_lock_init(&futex_queues[i].lock);
1885         }
1886         return 0;
1887 }
1888 __initcall(init);