Merge branch 'master' into upstream
[linux-2.6] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
20  *  enough at me, Linus for the original (flawed) idea, Matthew
21  *  Kirkwood for proof-of-concept implementation.
22  *
23  *  "The futexes are also cursed."
24  *  "But they come in a choice of three flavours!"
25  *
26  *  This program is free software; you can redistribute it and/or modify
27  *  it under the terms of the GNU General Public License as published by
28  *  the Free Software Foundation; either version 2 of the License, or
29  *  (at your option) any later version.
30  *
31  *  This program is distributed in the hope that it will be useful,
32  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
33  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
34  *  GNU General Public License for more details.
35  *
36  *  You should have received a copy of the GNU General Public License
37  *  along with this program; if not, write to the Free Software
38  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
39  */
40 #include <linux/slab.h>
41 #include <linux/poll.h>
42 #include <linux/fs.h>
43 #include <linux/file.h>
44 #include <linux/jhash.h>
45 #include <linux/init.h>
46 #include <linux/futex.h>
47 #include <linux/mount.h>
48 #include <linux/pagemap.h>
49 #include <linux/syscalls.h>
50 #include <linux/signal.h>
51 #include <asm/futex.h>
52
53 #include "rtmutex_common.h"
54
55 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
56
57 /*
58  * Futexes are matched on equal values of this key.
59  * The key type depends on whether it's a shared or private mapping.
60  * Don't rearrange members without looking at hash_futex().
61  *
62  * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
63  * We set bit 0 to indicate if it's an inode-based key.
64  */
65 union futex_key {
66         struct {
67                 unsigned long pgoff;
68                 struct inode *inode;
69                 int offset;
70         } shared;
71         struct {
72                 unsigned long address;
73                 struct mm_struct *mm;
74                 int offset;
75         } private;
76         struct {
77                 unsigned long word;
78                 void *ptr;
79                 int offset;
80         } both;
81 };
82
83 /*
84  * Priority Inheritance state:
85  */
86 struct futex_pi_state {
87         /*
88          * list of 'owned' pi_state instances - these have to be
89          * cleaned up in do_exit() if the task exits prematurely:
90          */
91         struct list_head list;
92
93         /*
94          * The PI object:
95          */
96         struct rt_mutex pi_mutex;
97
98         struct task_struct *owner;
99         atomic_t refcount;
100
101         union futex_key key;
102 };
103
104 /*
105  * We use this hashed waitqueue instead of a normal wait_queue_t, so
106  * we can wake only the relevant ones (hashed queues may be shared).
107  *
108  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
109  * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0.
110  * The order of wakup is always to make the first condition true, then
111  * wake up q->waiters, then make the second condition true.
112  */
113 struct futex_q {
114         struct list_head list;
115         wait_queue_head_t waiters;
116
117         /* Which hash list lock to use: */
118         spinlock_t *lock_ptr;
119
120         /* Key which the futex is hashed on: */
121         union futex_key key;
122
123         /* For fd, sigio sent using these: */
124         int fd;
125         struct file *filp;
126
127         /* Optional priority inheritance state: */
128         struct futex_pi_state *pi_state;
129         struct task_struct *task;
130 };
131
132 /*
133  * Split the global futex_lock into every hash list lock.
134  */
135 struct futex_hash_bucket {
136        spinlock_t              lock;
137        struct list_head       chain;
138 };
139
140 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
141
142 /* Futex-fs vfsmount entry: */
143 static struct vfsmount *futex_mnt;
144
145 /*
146  * We hash on the keys returned from get_futex_key (see below).
147  */
148 static struct futex_hash_bucket *hash_futex(union futex_key *key)
149 {
150         u32 hash = jhash2((u32*)&key->both.word,
151                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
152                           key->both.offset);
153         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
154 }
155
156 /*
157  * Return 1 if two futex_keys are equal, 0 otherwise.
158  */
159 static inline int match_futex(union futex_key *key1, union futex_key *key2)
160 {
161         return (key1->both.word == key2->both.word
162                 && key1->both.ptr == key2->both.ptr
163                 && key1->both.offset == key2->both.offset);
164 }
165
166 /*
167  * Get parameters which are the keys for a futex.
168  *
169  * For shared mappings, it's (page->index, vma->vm_file->f_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 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                                 goto out;
752                         goto retry;
753                 }
754
755                 /*
756                  * If we would have faulted, release mmap_sem,
757                  * fault it in and start all over again.
758                  */
759                 up_read(&current->mm->mmap_sem);
760
761                 ret = get_user(dummy, uaddr2);
762                 if (ret)
763                         return ret;
764
765                 goto retryfull;
766         }
767
768         head = &hb1->chain;
769
770         list_for_each_entry_safe(this, next, head, list) {
771                 if (match_futex (&this->key, &key1)) {
772                         wake_futex(this);
773                         if (++ret >= nr_wake)
774                                 break;
775                 }
776         }
777
778         if (op_ret > 0) {
779                 head = &hb2->chain;
780
781                 op_ret = 0;
782                 list_for_each_entry_safe(this, next, head, list) {
783                         if (match_futex (&this->key, &key2)) {
784                                 wake_futex(this);
785                                 if (++op_ret >= nr_wake2)
786                                         break;
787                         }
788                 }
789                 ret += op_ret;
790         }
791
792         spin_unlock(&hb1->lock);
793         if (hb1 != hb2)
794                 spin_unlock(&hb2->lock);
795 out:
796         up_read(&current->mm->mmap_sem);
797         return ret;
798 }
799
800 /*
801  * Requeue all waiters hashed on one physical page to another
802  * physical page.
803  */
804 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
805                          int nr_wake, int nr_requeue, u32 *cmpval)
806 {
807         union futex_key key1, key2;
808         struct futex_hash_bucket *hb1, *hb2;
809         struct list_head *head1;
810         struct futex_q *this, *next;
811         int ret, drop_count = 0;
812
813  retry:
814         down_read(&current->mm->mmap_sem);
815
816         ret = get_futex_key(uaddr1, &key1);
817         if (unlikely(ret != 0))
818                 goto out;
819         ret = get_futex_key(uaddr2, &key2);
820         if (unlikely(ret != 0))
821                 goto out;
822
823         hb1 = hash_futex(&key1);
824         hb2 = hash_futex(&key2);
825
826         double_lock_hb(hb1, hb2);
827
828         if (likely(cmpval != NULL)) {
829                 u32 curval;
830
831                 ret = get_futex_value_locked(&curval, uaddr1);
832
833                 if (unlikely(ret)) {
834                         spin_unlock(&hb1->lock);
835                         if (hb1 != hb2)
836                                 spin_unlock(&hb2->lock);
837
838                         /*
839                          * If we would have faulted, release mmap_sem, fault
840                          * it in and start all over again.
841                          */
842                         up_read(&current->mm->mmap_sem);
843
844                         ret = get_user(curval, uaddr1);
845
846                         if (!ret)
847                                 goto retry;
848
849                         return ret;
850                 }
851                 if (curval != *cmpval) {
852                         ret = -EAGAIN;
853                         goto out_unlock;
854                 }
855         }
856
857         head1 = &hb1->chain;
858         list_for_each_entry_safe(this, next, head1, list) {
859                 if (!match_futex (&this->key, &key1))
860                         continue;
861                 if (++ret <= nr_wake) {
862                         wake_futex(this);
863                 } else {
864                         /*
865                          * If key1 and key2 hash to the same bucket, no need to
866                          * requeue.
867                          */
868                         if (likely(head1 != &hb2->chain)) {
869                                 list_move_tail(&this->list, &hb2->chain);
870                                 this->lock_ptr = &hb2->lock;
871                         }
872                         this->key = key2;
873                         get_key_refs(&key2);
874                         drop_count++;
875
876                         if (ret - nr_wake >= nr_requeue)
877                                 break;
878                 }
879         }
880
881 out_unlock:
882         spin_unlock(&hb1->lock);
883         if (hb1 != hb2)
884                 spin_unlock(&hb2->lock);
885
886         /* drop_key_refs() must be called outside the spinlocks. */
887         while (--drop_count >= 0)
888                 drop_key_refs(&key1);
889
890 out:
891         up_read(&current->mm->mmap_sem);
892         return ret;
893 }
894
895 /* The key must be already stored in q->key. */
896 static inline struct futex_hash_bucket *
897 queue_lock(struct futex_q *q, int fd, struct file *filp)
898 {
899         struct futex_hash_bucket *hb;
900
901         q->fd = fd;
902         q->filp = filp;
903
904         init_waitqueue_head(&q->waiters);
905
906         get_key_refs(&q->key);
907         hb = hash_futex(&q->key);
908         q->lock_ptr = &hb->lock;
909
910         spin_lock(&hb->lock);
911         return hb;
912 }
913
914 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
915 {
916         list_add_tail(&q->list, &hb->chain);
917         q->task = current;
918         spin_unlock(&hb->lock);
919 }
920
921 static inline void
922 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
923 {
924         spin_unlock(&hb->lock);
925         drop_key_refs(&q->key);
926 }
927
928 /*
929  * queue_me and unqueue_me must be called as a pair, each
930  * exactly once.  They are called with the hashed spinlock held.
931  */
932
933 /* The key must be already stored in q->key. */
934 static void queue_me(struct futex_q *q, int fd, struct file *filp)
935 {
936         struct futex_hash_bucket *hb;
937
938         hb = queue_lock(q, fd, filp);
939         __queue_me(q, hb);
940 }
941
942 /* Return 1 if we were still queued (ie. 0 means we were woken) */
943 static int unqueue_me(struct futex_q *q)
944 {
945         spinlock_t *lock_ptr;
946         int ret = 0;
947
948         /* In the common case we don't take the spinlock, which is nice. */
949  retry:
950         lock_ptr = q->lock_ptr;
951         barrier();
952         if (lock_ptr != 0) {
953                 spin_lock(lock_ptr);
954                 /*
955                  * q->lock_ptr can change between reading it and
956                  * spin_lock(), causing us to take the wrong lock.  This
957                  * corrects the race condition.
958                  *
959                  * Reasoning goes like this: if we have the wrong lock,
960                  * q->lock_ptr must have changed (maybe several times)
961                  * between reading it and the spin_lock().  It can
962                  * change again after the spin_lock() but only if it was
963                  * already changed before the spin_lock().  It cannot,
964                  * however, change back to the original value.  Therefore
965                  * we can detect whether we acquired the correct lock.
966                  */
967                 if (unlikely(lock_ptr != q->lock_ptr)) {
968                         spin_unlock(lock_ptr);
969                         goto retry;
970                 }
971                 WARN_ON(list_empty(&q->list));
972                 list_del(&q->list);
973
974                 BUG_ON(q->pi_state);
975
976                 spin_unlock(lock_ptr);
977                 ret = 1;
978         }
979
980         drop_key_refs(&q->key);
981         return ret;
982 }
983
984 /*
985  * PI futexes can not be requeued and must remove themself from the
986  * hash bucket. The hash bucket lock is held on entry and dropped here.
987  */
988 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
989 {
990         WARN_ON(list_empty(&q->list));
991         list_del(&q->list);
992
993         BUG_ON(!q->pi_state);
994         free_pi_state(q->pi_state);
995         q->pi_state = NULL;
996
997         spin_unlock(&hb->lock);
998
999         drop_key_refs(&q->key);
1000 }
1001
1002 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
1003 {
1004         struct task_struct *curr = current;
1005         DECLARE_WAITQUEUE(wait, curr);
1006         struct futex_hash_bucket *hb;
1007         struct futex_q q;
1008         u32 uval;
1009         int ret;
1010
1011         q.pi_state = NULL;
1012  retry:
1013         down_read(&curr->mm->mmap_sem);
1014
1015         ret = get_futex_key(uaddr, &q.key);
1016         if (unlikely(ret != 0))
1017                 goto out_release_sem;
1018
1019         hb = queue_lock(&q, -1, NULL);
1020
1021         /*
1022          * Access the page AFTER the futex is queued.
1023          * Order is important:
1024          *
1025          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1026          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1027          *
1028          * The basic logical guarantee of a futex is that it blocks ONLY
1029          * if cond(var) is known to be true at the time of blocking, for
1030          * any cond.  If we queued after testing *uaddr, that would open
1031          * a race condition where we could block indefinitely with
1032          * cond(var) false, which would violate the guarantee.
1033          *
1034          * A consequence is that futex_wait() can return zero and absorb
1035          * a wakeup when *uaddr != val on entry to the syscall.  This is
1036          * rare, but normal.
1037          *
1038          * We hold the mmap semaphore, so the mapping cannot have changed
1039          * since we looked it up in get_futex_key.
1040          */
1041         ret = get_futex_value_locked(&uval, uaddr);
1042
1043         if (unlikely(ret)) {
1044                 queue_unlock(&q, hb);
1045
1046                 /*
1047                  * If we would have faulted, release mmap_sem, fault it in and
1048                  * start all over again.
1049                  */
1050                 up_read(&curr->mm->mmap_sem);
1051
1052                 ret = get_user(uval, uaddr);
1053
1054                 if (!ret)
1055                         goto retry;
1056                 return ret;
1057         }
1058         ret = -EWOULDBLOCK;
1059         if (uval != val)
1060                 goto out_unlock_release_sem;
1061
1062         /* Only actually queue if *uaddr contained val.  */
1063         __queue_me(&q, hb);
1064
1065         /*
1066          * Now the futex is queued and we have checked the data, we
1067          * don't want to hold mmap_sem while we sleep.
1068          */
1069         up_read(&curr->mm->mmap_sem);
1070
1071         /*
1072          * There might have been scheduling since the queue_me(), as we
1073          * cannot hold a spinlock across the get_user() in case it
1074          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1075          * queueing ourselves into the futex hash.  This code thus has to
1076          * rely on the futex_wake() code removing us from hash when it
1077          * wakes us up.
1078          */
1079
1080         /* add_wait_queue is the barrier after __set_current_state. */
1081         __set_current_state(TASK_INTERRUPTIBLE);
1082         add_wait_queue(&q.waiters, &wait);
1083         /*
1084          * !list_empty() is safe here without any lock.
1085          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1086          */
1087         if (likely(!list_empty(&q.list)))
1088                 time = schedule_timeout(time);
1089         __set_current_state(TASK_RUNNING);
1090
1091         /*
1092          * NOTE: we don't remove ourselves from the waitqueue because
1093          * we are the only user of it.
1094          */
1095
1096         /* If we were woken (and unqueued), we succeeded, whatever. */
1097         if (!unqueue_me(&q))
1098                 return 0;
1099         if (time == 0)
1100                 return -ETIMEDOUT;
1101         /*
1102          * We expect signal_pending(current), but another thread may
1103          * have handled it for us already.
1104          */
1105         return -EINTR;
1106
1107  out_unlock_release_sem:
1108         queue_unlock(&q, hb);
1109
1110  out_release_sem:
1111         up_read(&curr->mm->mmap_sem);
1112         return ret;
1113 }
1114
1115 /*
1116  * Userspace tried a 0 -> TID atomic transition of the futex value
1117  * and failed. The kernel side here does the whole locking operation:
1118  * if there are waiters then it will block, it does PI, etc. (Due to
1119  * races the kernel might see a 0 value of the futex too.)
1120  */
1121 static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
1122                             struct hrtimer_sleeper *to)
1123 {
1124         struct task_struct *curr = current;
1125         struct futex_hash_bucket *hb;
1126         u32 uval, newval, curval;
1127         struct futex_q q;
1128         int ret, attempt = 0;
1129
1130         if (refill_pi_state_cache())
1131                 return -ENOMEM;
1132
1133         q.pi_state = NULL;
1134  retry:
1135         down_read(&curr->mm->mmap_sem);
1136
1137         ret = get_futex_key(uaddr, &q.key);
1138         if (unlikely(ret != 0))
1139                 goto out_release_sem;
1140
1141         hb = queue_lock(&q, -1, NULL);
1142
1143  retry_locked:
1144         /*
1145          * To avoid races, we attempt to take the lock here again
1146          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1147          * the locks. It will most likely not succeed.
1148          */
1149         newval = current->pid;
1150
1151         inc_preempt_count();
1152         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1153         dec_preempt_count();
1154
1155         if (unlikely(curval == -EFAULT))
1156                 goto uaddr_faulted;
1157
1158         /* We own the lock already */
1159         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1160                 if (!detect && 0)
1161                         force_sig(SIGKILL, current);
1162                 ret = -EDEADLK;
1163                 goto out_unlock_release_sem;
1164         }
1165
1166         /*
1167          * Surprise - we got the lock. Just return
1168          * to userspace:
1169          */
1170         if (unlikely(!curval))
1171                 goto out_unlock_release_sem;
1172
1173         uval = curval;
1174         newval = uval | FUTEX_WAITERS;
1175
1176         inc_preempt_count();
1177         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1178         dec_preempt_count();
1179
1180         if (unlikely(curval == -EFAULT))
1181                 goto uaddr_faulted;
1182         if (unlikely(curval != uval))
1183                 goto retry_locked;
1184
1185         /*
1186          * We dont have the lock. Look up the PI state (or create it if
1187          * we are the first waiter):
1188          */
1189         ret = lookup_pi_state(uval, hb, &q);
1190
1191         if (unlikely(ret)) {
1192                 /*
1193                  * There were no waiters and the owner task lookup
1194                  * failed. When the OWNER_DIED bit is set, then we
1195                  * know that this is a robust futex and we actually
1196                  * take the lock. This is safe as we are protected by
1197                  * the hash bucket lock. We also set the waiters bit
1198                  * unconditionally here, to simplify glibc handling of
1199                  * multiple tasks racing to acquire the lock and
1200                  * cleanup the problems which were left by the dead
1201                  * owner.
1202                  */
1203                 if (curval & FUTEX_OWNER_DIED) {
1204                         uval = newval;
1205                         newval = current->pid |
1206                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1207
1208                         inc_preempt_count();
1209                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1210                                                                uval, newval);
1211                         dec_preempt_count();
1212
1213                         if (unlikely(curval == -EFAULT))
1214                                 goto uaddr_faulted;
1215                         if (unlikely(curval != uval))
1216                                 goto retry_locked;
1217                         ret = 0;
1218                 }
1219                 goto out_unlock_release_sem;
1220         }
1221
1222         /*
1223          * Only actually queue now that the atomic ops are done:
1224          */
1225         __queue_me(&q, hb);
1226
1227         /*
1228          * Now the futex is queued and we have checked the data, we
1229          * don't want to hold mmap_sem while we sleep.
1230          */
1231         up_read(&curr->mm->mmap_sem);
1232
1233         WARN_ON(!q.pi_state);
1234         /*
1235          * Block on the PI mutex:
1236          */
1237         if (!trylock)
1238                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1239         else {
1240                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1241                 /* Fixup the trylock return value: */
1242                 ret = ret ? 0 : -EWOULDBLOCK;
1243         }
1244
1245         down_read(&curr->mm->mmap_sem);
1246         spin_lock(q.lock_ptr);
1247
1248         /*
1249          * Got the lock. We might not be the anticipated owner if we
1250          * did a lock-steal - fix up the PI-state in that case.
1251          */
1252         if (!ret && q.pi_state->owner != curr) {
1253                 u32 newtid = current->pid | FUTEX_WAITERS;
1254
1255                 /* Owner died? */
1256                 if (q.pi_state->owner != NULL) {
1257                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1258                         WARN_ON(list_empty(&q.pi_state->list));
1259                         list_del_init(&q.pi_state->list);
1260                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1261                 } else
1262                         newtid |= FUTEX_OWNER_DIED;
1263
1264                 q.pi_state->owner = current;
1265
1266                 spin_lock_irq(&current->pi_lock);
1267                 WARN_ON(!list_empty(&q.pi_state->list));
1268                 list_add(&q.pi_state->list, &current->pi_state_list);
1269                 spin_unlock_irq(&current->pi_lock);
1270
1271                 /* Unqueue and drop the lock */
1272                 unqueue_me_pi(&q, hb);
1273                 up_read(&curr->mm->mmap_sem);
1274                 /*
1275                  * We own it, so we have to replace the pending owner
1276                  * TID. This must be atomic as we have preserve the
1277                  * owner died bit here.
1278                  */
1279                 ret = get_user(uval, uaddr);
1280                 while (!ret) {
1281                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1282                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1283                                                                uval, newval);
1284                         if (curval == -EFAULT)
1285                                 ret = -EFAULT;
1286                         if (curval == uval)
1287                                 break;
1288                         uval = curval;
1289                 }
1290         } else {
1291                 /*
1292                  * Catch the rare case, where the lock was released
1293                  * when we were on the way back before we locked
1294                  * the hash bucket.
1295                  */
1296                 if (ret && q.pi_state->owner == curr) {
1297                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1298                                 ret = 0;
1299                 }
1300                 /* Unqueue and drop the lock */
1301                 unqueue_me_pi(&q, hb);
1302                 up_read(&curr->mm->mmap_sem);
1303         }
1304
1305         if (!detect && ret == -EDEADLK && 0)
1306                 force_sig(SIGKILL, current);
1307
1308         return ret;
1309
1310  out_unlock_release_sem:
1311         queue_unlock(&q, hb);
1312
1313  out_release_sem:
1314         up_read(&curr->mm->mmap_sem);
1315         return ret;
1316
1317  uaddr_faulted:
1318         /*
1319          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1320          * non-atomically.  Therefore, if get_user below is not
1321          * enough, we need to handle the fault ourselves, while
1322          * still holding the mmap_sem.
1323          */
1324         if (attempt++) {
1325                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1326                         goto out_unlock_release_sem;
1327
1328                 goto retry_locked;
1329         }
1330
1331         queue_unlock(&q, hb);
1332         up_read(&curr->mm->mmap_sem);
1333
1334         ret = get_user(uval, uaddr);
1335         if (!ret && (uval != -EFAULT))
1336                 goto retry;
1337
1338         return ret;
1339 }
1340
1341 /*
1342  * Restart handler
1343  */
1344 static long futex_lock_pi_restart(struct restart_block *restart)
1345 {
1346         struct hrtimer_sleeper timeout, *to = NULL;
1347         int ret;
1348
1349         restart->fn = do_no_restart_syscall;
1350
1351         if (restart->arg2 || restart->arg3) {
1352                 to = &timeout;
1353                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1354                 hrtimer_init_sleeper(to, current);
1355                 to->timer.expires.tv64 = ((u64)restart->arg1 << 32) |
1356                         (u64) restart->arg0;
1357         }
1358
1359         pr_debug("lock_pi restart: %p, %d (%d)\n",
1360                  (u32 __user *)restart->arg0, current->pid);
1361
1362         ret = do_futex_lock_pi((u32 __user *)restart->arg0, restart->arg1,
1363                                0, to);
1364
1365         if (ret != -EINTR)
1366                 return ret;
1367
1368         restart->fn = futex_lock_pi_restart;
1369
1370         /* The other values are filled in */
1371         return -ERESTART_RESTARTBLOCK;
1372 }
1373
1374 /*
1375  * Called from the syscall entry below.
1376  */
1377 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1378                          long nsec, int trylock)
1379 {
1380         struct hrtimer_sleeper timeout, *to = NULL;
1381         struct restart_block *restart;
1382         int ret;
1383
1384         if (sec != MAX_SCHEDULE_TIMEOUT) {
1385                 to = &timeout;
1386                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1387                 hrtimer_init_sleeper(to, current);
1388                 to->timer.expires = ktime_set(sec, nsec);
1389         }
1390
1391         ret = do_futex_lock_pi(uaddr, detect, trylock, to);
1392
1393         if (ret != -EINTR)
1394                 return ret;
1395
1396         pr_debug("lock_pi interrupted: %p, %d (%d)\n", uaddr, current->pid);
1397
1398         restart = &current_thread_info()->restart_block;
1399         restart->fn = futex_lock_pi_restart;
1400         restart->arg0 = (unsigned long) uaddr;
1401         restart->arg1 = detect;
1402         if (to) {
1403                 restart->arg2 = to->timer.expires.tv64 & 0xFFFFFFFF;
1404                 restart->arg3 = to->timer.expires.tv64 >> 32;
1405         } else
1406                 restart->arg2 = restart->arg3 = 0;
1407
1408         return -ERESTART_RESTARTBLOCK;
1409 }
1410
1411 /*
1412  * Userspace attempted a TID -> 0 atomic transition, and failed.
1413  * This is the in-kernel slowpath: we look up the PI state (if any),
1414  * and do the rt-mutex unlock.
1415  */
1416 static int futex_unlock_pi(u32 __user *uaddr)
1417 {
1418         struct futex_hash_bucket *hb;
1419         struct futex_q *this, *next;
1420         u32 uval;
1421         struct list_head *head;
1422         union futex_key key;
1423         int ret, attempt = 0;
1424
1425 retry:
1426         if (get_user(uval, uaddr))
1427                 return -EFAULT;
1428         /*
1429          * We release only a lock we actually own:
1430          */
1431         if ((uval & FUTEX_TID_MASK) != current->pid)
1432                 return -EPERM;
1433         /*
1434          * First take all the futex related locks:
1435          */
1436         down_read(&current->mm->mmap_sem);
1437
1438         ret = get_futex_key(uaddr, &key);
1439         if (unlikely(ret != 0))
1440                 goto out;
1441
1442         hb = hash_futex(&key);
1443         spin_lock(&hb->lock);
1444
1445 retry_locked:
1446         /*
1447          * To avoid races, try to do the TID -> 0 atomic transition
1448          * again. If it succeeds then we can return without waking
1449          * anyone else up:
1450          */
1451         if (!(uval & FUTEX_OWNER_DIED)) {
1452                 inc_preempt_count();
1453                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1454                 dec_preempt_count();
1455         }
1456
1457         if (unlikely(uval == -EFAULT))
1458                 goto pi_faulted;
1459         /*
1460          * Rare case: we managed to release the lock atomically,
1461          * no need to wake anyone else up:
1462          */
1463         if (unlikely(uval == current->pid))
1464                 goto out_unlock;
1465
1466         /*
1467          * Ok, other tasks may need to be woken up - check waiters
1468          * and do the wakeup if necessary:
1469          */
1470         head = &hb->chain;
1471
1472         list_for_each_entry_safe(this, next, head, list) {
1473                 if (!match_futex (&this->key, &key))
1474                         continue;
1475                 ret = wake_futex_pi(uaddr, uval, this);
1476                 /*
1477                  * The atomic access to the futex value
1478                  * generated a pagefault, so retry the
1479                  * user-access and the wakeup:
1480                  */
1481                 if (ret == -EFAULT)
1482                         goto pi_faulted;
1483                 goto out_unlock;
1484         }
1485         /*
1486          * No waiters - kernel unlocks the futex:
1487          */
1488         if (!(uval & FUTEX_OWNER_DIED)) {
1489                 ret = unlock_futex_pi(uaddr, uval);
1490                 if (ret == -EFAULT)
1491                         goto pi_faulted;
1492         }
1493
1494 out_unlock:
1495         spin_unlock(&hb->lock);
1496 out:
1497         up_read(&current->mm->mmap_sem);
1498
1499         return ret;
1500
1501 pi_faulted:
1502         /*
1503          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1504          * non-atomically.  Therefore, if get_user below is not
1505          * enough, we need to handle the fault ourselves, while
1506          * still holding the mmap_sem.
1507          */
1508         if (attempt++) {
1509                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1510                         goto out_unlock;
1511
1512                 goto retry_locked;
1513         }
1514
1515         spin_unlock(&hb->lock);
1516         up_read(&current->mm->mmap_sem);
1517
1518         ret = get_user(uval, uaddr);
1519         if (!ret && (uval != -EFAULT))
1520                 goto retry;
1521
1522         return ret;
1523 }
1524
1525 static int futex_close(struct inode *inode, struct file *filp)
1526 {
1527         struct futex_q *q = filp->private_data;
1528
1529         unqueue_me(q);
1530         kfree(q);
1531
1532         return 0;
1533 }
1534
1535 /* This is one-shot: once it's gone off you need a new fd */
1536 static unsigned int futex_poll(struct file *filp,
1537                                struct poll_table_struct *wait)
1538 {
1539         struct futex_q *q = filp->private_data;
1540         int ret = 0;
1541
1542         poll_wait(filp, &q->waiters, wait);
1543
1544         /*
1545          * list_empty() is safe here without any lock.
1546          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1547          */
1548         if (list_empty(&q->list))
1549                 ret = POLLIN | POLLRDNORM;
1550
1551         return ret;
1552 }
1553
1554 static struct file_operations futex_fops = {
1555         .release        = futex_close,
1556         .poll           = futex_poll,
1557 };
1558
1559 /*
1560  * Signal allows caller to avoid the race which would occur if they
1561  * set the sigio stuff up afterwards.
1562  */
1563 static int futex_fd(u32 __user *uaddr, int signal)
1564 {
1565         struct futex_q *q;
1566         struct file *filp;
1567         int ret, err;
1568
1569         ret = -EINVAL;
1570         if (!valid_signal(signal))
1571                 goto out;
1572
1573         ret = get_unused_fd();
1574         if (ret < 0)
1575                 goto out;
1576         filp = get_empty_filp();
1577         if (!filp) {
1578                 put_unused_fd(ret);
1579                 ret = -ENFILE;
1580                 goto out;
1581         }
1582         filp->f_op = &futex_fops;
1583         filp->f_vfsmnt = mntget(futex_mnt);
1584         filp->f_dentry = dget(futex_mnt->mnt_root);
1585         filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
1586
1587         if (signal) {
1588                 err = f_setown(filp, current->pid, 1);
1589                 if (err < 0) {
1590                         goto error;
1591                 }
1592                 filp->f_owner.signum = signal;
1593         }
1594
1595         q = kmalloc(sizeof(*q), GFP_KERNEL);
1596         if (!q) {
1597                 err = -ENOMEM;
1598                 goto error;
1599         }
1600         q->pi_state = NULL;
1601
1602         down_read(&current->mm->mmap_sem);
1603         err = get_futex_key(uaddr, &q->key);
1604
1605         if (unlikely(err != 0)) {
1606                 up_read(&current->mm->mmap_sem);
1607                 kfree(q);
1608                 goto error;
1609         }
1610
1611         /*
1612          * queue_me() must be called before releasing mmap_sem, because
1613          * key->shared.inode needs to be referenced while holding it.
1614          */
1615         filp->private_data = q;
1616
1617         queue_me(q, ret, filp);
1618         up_read(&current->mm->mmap_sem);
1619
1620         /* Now we map fd to filp, so userspace can access it */
1621         fd_install(ret, filp);
1622 out:
1623         return ret;
1624 error:
1625         put_unused_fd(ret);
1626         put_filp(filp);
1627         ret = err;
1628         goto out;
1629 }
1630
1631 /*
1632  * Support for robust futexes: the kernel cleans up held futexes at
1633  * thread exit time.
1634  *
1635  * Implementation: user-space maintains a per-thread list of locks it
1636  * is holding. Upon do_exit(), the kernel carefully walks this list,
1637  * and marks all locks that are owned by this thread with the
1638  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1639  * always manipulated with the lock held, so the list is private and
1640  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1641  * field, to allow the kernel to clean up if the thread dies after
1642  * acquiring the lock, but just before it could have added itself to
1643  * the list. There can only be one such pending lock.
1644  */
1645
1646 /**
1647  * sys_set_robust_list - set the robust-futex list head of a task
1648  * @head: pointer to the list-head
1649  * @len: length of the list-head, as userspace expects
1650  */
1651 asmlinkage long
1652 sys_set_robust_list(struct robust_list_head __user *head,
1653                     size_t len)
1654 {
1655         /*
1656          * The kernel knows only one size for now:
1657          */
1658         if (unlikely(len != sizeof(*head)))
1659                 return -EINVAL;
1660
1661         current->robust_list = head;
1662
1663         return 0;
1664 }
1665
1666 /**
1667  * sys_get_robust_list - get the robust-futex list head of a task
1668  * @pid: pid of the process [zero for current task]
1669  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1670  * @len_ptr: pointer to a length field, the kernel fills in the header size
1671  */
1672 asmlinkage long
1673 sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
1674                     size_t __user *len_ptr)
1675 {
1676         struct robust_list_head *head;
1677         unsigned long ret;
1678
1679         if (!pid)
1680                 head = current->robust_list;
1681         else {
1682                 struct task_struct *p;
1683
1684                 ret = -ESRCH;
1685                 read_lock(&tasklist_lock);
1686                 p = find_task_by_pid(pid);
1687                 if (!p)
1688                         goto err_unlock;
1689                 ret = -EPERM;
1690                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1691                                 !capable(CAP_SYS_PTRACE))
1692                         goto err_unlock;
1693                 head = p->robust_list;
1694                 read_unlock(&tasklist_lock);
1695         }
1696
1697         if (put_user(sizeof(*head), len_ptr))
1698                 return -EFAULT;
1699         return put_user(head, head_ptr);
1700
1701 err_unlock:
1702         read_unlock(&tasklist_lock);
1703
1704         return ret;
1705 }
1706
1707 /*
1708  * Process a futex-list entry, check whether it's owned by the
1709  * dying task, and do notification if so:
1710  */
1711 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1712 {
1713         u32 uval, nval, mval;
1714
1715 retry:
1716         if (get_user(uval, uaddr))
1717                 return -1;
1718
1719         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1720                 /*
1721                  * Ok, this dying thread is truly holding a futex
1722                  * of interest. Set the OWNER_DIED bit atomically
1723                  * via cmpxchg, and if the value had FUTEX_WAITERS
1724                  * set, wake up a waiter (if any). (We have to do a
1725                  * futex_wake() even if OWNER_DIED is already set -
1726                  * to handle the rare but possible case of recursive
1727                  * thread-death.) The rest of the cleanup is done in
1728                  * userspace.
1729                  */
1730                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1731                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1732
1733                 if (nval == -EFAULT)
1734                         return -1;
1735
1736                 if (nval != uval)
1737                         goto retry;
1738
1739                 /*
1740                  * Wake robust non-PI futexes here. The wakeup of
1741                  * PI futexes happens in exit_pi_state():
1742                  */
1743                 if (!pi) {
1744                         if (uval & FUTEX_WAITERS)
1745                                 futex_wake(uaddr, 1);
1746                 }
1747         }
1748         return 0;
1749 }
1750
1751 /*
1752  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1753  */
1754 static inline int fetch_robust_entry(struct robust_list __user **entry,
1755                                      struct robust_list __user **head, int *pi)
1756 {
1757         unsigned long uentry;
1758
1759         if (get_user(uentry, (unsigned long *)head))
1760                 return -EFAULT;
1761
1762         *entry = (void *)(uentry & ~1UL);
1763         *pi = uentry & 1;
1764
1765         return 0;
1766 }
1767
1768 /*
1769  * Walk curr->robust_list (very carefully, it's a userspace list!)
1770  * and mark any locks found there dead, and notify any waiters.
1771  *
1772  * We silently return on any sign of list-walking problem.
1773  */
1774 void exit_robust_list(struct task_struct *curr)
1775 {
1776         struct robust_list_head __user *head = curr->robust_list;
1777         struct robust_list __user *entry, *pending;
1778         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1779         unsigned long futex_offset;
1780
1781         /*
1782          * Fetch the list head (which was registered earlier, via
1783          * sys_set_robust_list()):
1784          */
1785         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1786                 return;
1787         /*
1788          * Fetch the relative futex offset:
1789          */
1790         if (get_user(futex_offset, &head->futex_offset))
1791                 return;
1792         /*
1793          * Fetch any possibly pending lock-add first, and handle it
1794          * if it exists:
1795          */
1796         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1797                 return;
1798
1799         if (pending)
1800                 handle_futex_death((void *)pending + futex_offset, curr, pip);
1801
1802         while (entry != &head->list) {
1803                 /*
1804                  * A pending lock might already be on the list, so
1805                  * don't process it twice:
1806                  */
1807                 if (entry != pending)
1808                         if (handle_futex_death((void *)entry + futex_offset,
1809                                                 curr, pi))
1810                                 return;
1811                 /*
1812                  * Fetch the next entry in the list:
1813                  */
1814                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1815                         return;
1816                 /*
1817                  * Avoid excessively long or circular lists:
1818                  */
1819                 if (!--limit)
1820                         break;
1821
1822                 cond_resched();
1823         }
1824 }
1825
1826 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1827                 u32 __user *uaddr2, u32 val2, u32 val3)
1828 {
1829         int ret;
1830
1831         switch (op) {
1832         case FUTEX_WAIT:
1833                 ret = futex_wait(uaddr, val, timeout);
1834                 break;
1835         case FUTEX_WAKE:
1836                 ret = futex_wake(uaddr, val);
1837                 break;
1838         case FUTEX_FD:
1839                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1840                 ret = futex_fd(uaddr, val);
1841                 break;
1842         case FUTEX_REQUEUE:
1843                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1844                 break;
1845         case FUTEX_CMP_REQUEUE:
1846                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1847                 break;
1848         case FUTEX_WAKE_OP:
1849                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1850                 break;
1851         case FUTEX_LOCK_PI:
1852                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1853                 break;
1854         case FUTEX_UNLOCK_PI:
1855                 ret = futex_unlock_pi(uaddr);
1856                 break;
1857         case FUTEX_TRYLOCK_PI:
1858                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1859                 break;
1860         default:
1861                 ret = -ENOSYS;
1862         }
1863         return ret;
1864 }
1865
1866
1867 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1868                           struct timespec __user *utime, u32 __user *uaddr2,
1869                           u32 val3)
1870 {
1871         struct timespec t;
1872         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1873         u32 val2 = 0;
1874
1875         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1876                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1877                         return -EFAULT;
1878                 if (!timespec_valid(&t))
1879                         return -EINVAL;
1880                 if (op == FUTEX_WAIT)
1881                         timeout = timespec_to_jiffies(&t) + 1;
1882                 else {
1883                         timeout = t.tv_sec;
1884                         val2 = t.tv_nsec;
1885                 }
1886         }
1887         /*
1888          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1889          */
1890         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1891                 val2 = (u32) (unsigned long) utime;
1892
1893         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1894 }
1895
1896 static int futexfs_get_sb(struct file_system_type *fs_type,
1897                           int flags, const char *dev_name, void *data,
1898                           struct vfsmount *mnt)
1899 {
1900         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1901 }
1902
1903 static struct file_system_type futex_fs_type = {
1904         .name           = "futexfs",
1905         .get_sb         = futexfs_get_sb,
1906         .kill_sb        = kill_anon_super,
1907 };
1908
1909 static int __init init(void)
1910 {
1911         unsigned int i;
1912
1913         register_filesystem(&futex_fs_type);
1914         futex_mnt = kern_mount(&futex_fs_type);
1915
1916         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1917                 INIT_LIST_HEAD(&futex_queues[i].chain);
1918                 spin_lock_init(&futex_queues[i].lock);
1919         }
1920         return 0;
1921 }
1922 __initcall(init);