Merge branch 'from-linus' 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 futex_hash_bucket *hb;
419         struct list_head *next, *head = &curr->pi_state_list;
420         struct futex_pi_state *pi_state;
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 themselfs
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                 spin_unlock_irq(&curr->pi_lock);
435
436                 hb = hash_futex(&key);
437                 spin_lock(&hb->lock);
438
439                 spin_lock_irq(&curr->pi_lock);
440                 if (head->next != next) {
441                         spin_unlock(&hb->lock);
442                         continue;
443                 }
444
445                 list_del_init(&pi_state->list);
446
447                 WARN_ON(pi_state->owner != curr);
448
449                 pi_state->owner = NULL;
450                 spin_unlock_irq(&curr->pi_lock);
451
452                 rt_mutex_unlock(&pi_state->pi_mutex);
453
454                 spin_unlock(&hb->lock);
455
456                 spin_lock_irq(&curr->pi_lock);
457         }
458         spin_unlock_irq(&curr->pi_lock);
459 }
460
461 static int
462 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
463 {
464         struct futex_pi_state *pi_state = NULL;
465         struct futex_q *this, *next;
466         struct list_head *head;
467         struct task_struct *p;
468         pid_t pid;
469
470         head = &hb->chain;
471
472         list_for_each_entry_safe(this, next, head, list) {
473                 if (match_futex (&this->key, &me->key)) {
474                         /*
475                          * Another waiter already exists - bump up
476                          * the refcount and return its pi_state:
477                          */
478                         pi_state = this->pi_state;
479                         /*
480                          * Userspace might have messed up non PI and PI futexes
481                          */
482                         if (unlikely(!pi_state))
483                                 return -EINVAL;
484
485                         atomic_inc(&pi_state->refcount);
486                         me->pi_state = pi_state;
487
488                         return 0;
489                 }
490         }
491
492         /*
493          * We are the first waiter - try to look up the real owner and
494          * attach the new pi_state to it:
495          */
496         pid = uval & FUTEX_TID_MASK;
497         p = futex_find_get_task(pid);
498         if (!p)
499                 return -ESRCH;
500
501         pi_state = alloc_pi_state();
502
503         /*
504          * Initialize the pi_mutex in locked state and make 'p'
505          * the owner of it:
506          */
507         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
508
509         /* Store the key for possible exit cleanups: */
510         pi_state->key = me->key;
511
512         spin_lock_irq(&p->pi_lock);
513         list_add(&pi_state->list, &p->pi_state_list);
514         pi_state->owner = p;
515         spin_unlock_irq(&p->pi_lock);
516
517         put_task_struct(p);
518
519         me->pi_state = pi_state;
520
521         return 0;
522 }
523
524 /*
525  * The hash bucket lock must be held when this is called.
526  * Afterwards, the futex_q must not be accessed.
527  */
528 static void wake_futex(struct futex_q *q)
529 {
530         list_del_init(&q->list);
531         if (q->filp)
532                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
533         /*
534          * The lock in wake_up_all() is a crucial memory barrier after the
535          * list_del_init() and also before assigning to q->lock_ptr.
536          */
537         wake_up_all(&q->waiters);
538         /*
539          * The waiting task can free the futex_q as soon as this is written,
540          * without taking any locks.  This must come last.
541          *
542          * A memory barrier is required here to prevent the following store
543          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
544          * at the end of wake_up_all() does not prevent this store from
545          * moving.
546          */
547         wmb();
548         q->lock_ptr = NULL;
549 }
550
551 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
552 {
553         struct task_struct *new_owner;
554         struct futex_pi_state *pi_state = this->pi_state;
555         u32 curval, newval;
556
557         if (!pi_state)
558                 return -EINVAL;
559
560         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
561
562         /*
563          * This happens when we have stolen the lock and the original
564          * pending owner did not enqueue itself back on the rt_mutex.
565          * Thats not a tragedy. We know that way, that a lock waiter
566          * is on the fly. We make the futex_q waiter the pending owner.
567          */
568         if (!new_owner)
569                 new_owner = this->task;
570
571         /*
572          * We pass it to the next owner. (The WAITERS bit is always
573          * kept enabled while there is PI state around. We must also
574          * preserve the owner died bit.)
575          */
576         newval = (uval & FUTEX_OWNER_DIED) | FUTEX_WAITERS | new_owner->pid;
577
578         inc_preempt_count();
579         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
580         dec_preempt_count();
581
582         if (curval == -EFAULT)
583                 return -EFAULT;
584         if (curval != uval)
585                 return -EINVAL;
586
587         list_del_init(&pi_state->owner->pi_state_list);
588         list_add(&pi_state->list, &new_owner->pi_state_list);
589         pi_state->owner = new_owner;
590         rt_mutex_unlock(&pi_state->pi_mutex);
591
592         return 0;
593 }
594
595 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
596 {
597         u32 oldval;
598
599         /*
600          * There is no waiter, so we unlock the futex. The owner died
601          * bit has not to be preserved here. We are the owner:
602          */
603         inc_preempt_count();
604         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
605         dec_preempt_count();
606
607         if (oldval == -EFAULT)
608                 return oldval;
609         if (oldval != uval)
610                 return -EAGAIN;
611
612         return 0;
613 }
614
615 /*
616  * Express the locking dependencies for lockdep:
617  */
618 static inline void
619 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
620 {
621         if (hb1 <= hb2) {
622                 spin_lock(&hb1->lock);
623                 if (hb1 < hb2)
624                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
625         } else { /* hb1 > hb2 */
626                 spin_lock(&hb2->lock);
627                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
628         }
629 }
630
631 /*
632  * Wake up all waiters hashed on the physical page that is mapped
633  * to this virtual address:
634  */
635 static int futex_wake(u32 __user *uaddr, int nr_wake)
636 {
637         struct futex_hash_bucket *hb;
638         struct futex_q *this, *next;
639         struct list_head *head;
640         union futex_key key;
641         int ret;
642
643         down_read(&current->mm->mmap_sem);
644
645         ret = get_futex_key(uaddr, &key);
646         if (unlikely(ret != 0))
647                 goto out;
648
649         hb = hash_futex(&key);
650         spin_lock(&hb->lock);
651         head = &hb->chain;
652
653         list_for_each_entry_safe(this, next, head, list) {
654                 if (match_futex (&this->key, &key)) {
655                         if (this->pi_state) {
656                                 ret = -EINVAL;
657                                 break;
658                         }
659                         wake_futex(this);
660                         if (++ret >= nr_wake)
661                                 break;
662                 }
663         }
664
665         spin_unlock(&hb->lock);
666 out:
667         up_read(&current->mm->mmap_sem);
668         return ret;
669 }
670
671 /*
672  * Wake up all waiters hashed on the physical page that is mapped
673  * to this virtual address:
674  */
675 static int
676 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
677               int nr_wake, int nr_wake2, int op)
678 {
679         union futex_key key1, key2;
680         struct futex_hash_bucket *hb1, *hb2;
681         struct list_head *head;
682         struct futex_q *this, *next;
683         int ret, op_ret, attempt = 0;
684
685 retryfull:
686         down_read(&current->mm->mmap_sem);
687
688         ret = get_futex_key(uaddr1, &key1);
689         if (unlikely(ret != 0))
690                 goto out;
691         ret = get_futex_key(uaddr2, &key2);
692         if (unlikely(ret != 0))
693                 goto out;
694
695         hb1 = hash_futex(&key1);
696         hb2 = hash_futex(&key2);
697
698 retry:
699         double_lock_hb(hb1, hb2);
700
701         op_ret = futex_atomic_op_inuser(op, uaddr2);
702         if (unlikely(op_ret < 0)) {
703                 u32 dummy;
704
705                 spin_unlock(&hb1->lock);
706                 if (hb1 != hb2)
707                         spin_unlock(&hb2->lock);
708
709 #ifndef CONFIG_MMU
710                 /*
711                  * we don't get EFAULT from MMU faults if we don't have an MMU,
712                  * but we might get them from range checking
713                  */
714                 ret = op_ret;
715                 goto out;
716 #endif
717
718                 if (unlikely(op_ret != -EFAULT)) {
719                         ret = op_ret;
720                         goto out;
721                 }
722
723                 /*
724                  * futex_atomic_op_inuser needs to both read and write
725                  * *(int __user *)uaddr2, but we can't modify it
726                  * non-atomically.  Therefore, if get_user below is not
727                  * enough, we need to handle the fault ourselves, while
728                  * still holding the mmap_sem.
729                  */
730                 if (attempt++) {
731                         if (futex_handle_fault((unsigned long)uaddr2,
732                                                attempt))
733                                 goto out;
734                         goto retry;
735                 }
736
737                 /*
738                  * If we would have faulted, release mmap_sem,
739                  * fault it in and start all over again.
740                  */
741                 up_read(&current->mm->mmap_sem);
742
743                 ret = get_user(dummy, uaddr2);
744                 if (ret)
745                         return ret;
746
747                 goto retryfull;
748         }
749
750         head = &hb1->chain;
751
752         list_for_each_entry_safe(this, next, head, list) {
753                 if (match_futex (&this->key, &key1)) {
754                         wake_futex(this);
755                         if (++ret >= nr_wake)
756                                 break;
757                 }
758         }
759
760         if (op_ret > 0) {
761                 head = &hb2->chain;
762
763                 op_ret = 0;
764                 list_for_each_entry_safe(this, next, head, list) {
765                         if (match_futex (&this->key, &key2)) {
766                                 wake_futex(this);
767                                 if (++op_ret >= nr_wake2)
768                                         break;
769                         }
770                 }
771                 ret += op_ret;
772         }
773
774         spin_unlock(&hb1->lock);
775         if (hb1 != hb2)
776                 spin_unlock(&hb2->lock);
777 out:
778         up_read(&current->mm->mmap_sem);
779         return ret;
780 }
781
782 /*
783  * Requeue all waiters hashed on one physical page to another
784  * physical page.
785  */
786 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
787                          int nr_wake, int nr_requeue, u32 *cmpval)
788 {
789         union futex_key key1, key2;
790         struct futex_hash_bucket *hb1, *hb2;
791         struct list_head *head1;
792         struct futex_q *this, *next;
793         int ret, drop_count = 0;
794
795  retry:
796         down_read(&current->mm->mmap_sem);
797
798         ret = get_futex_key(uaddr1, &key1);
799         if (unlikely(ret != 0))
800                 goto out;
801         ret = get_futex_key(uaddr2, &key2);
802         if (unlikely(ret != 0))
803                 goto out;
804
805         hb1 = hash_futex(&key1);
806         hb2 = hash_futex(&key2);
807
808         double_lock_hb(hb1, hb2);
809
810         if (likely(cmpval != NULL)) {
811                 u32 curval;
812
813                 ret = get_futex_value_locked(&curval, uaddr1);
814
815                 if (unlikely(ret)) {
816                         spin_unlock(&hb1->lock);
817                         if (hb1 != hb2)
818                                 spin_unlock(&hb2->lock);
819
820                         /*
821                          * If we would have faulted, release mmap_sem, fault
822                          * it in and start all over again.
823                          */
824                         up_read(&current->mm->mmap_sem);
825
826                         ret = get_user(curval, uaddr1);
827
828                         if (!ret)
829                                 goto retry;
830
831                         return ret;
832                 }
833                 if (curval != *cmpval) {
834                         ret = -EAGAIN;
835                         goto out_unlock;
836                 }
837         }
838
839         head1 = &hb1->chain;
840         list_for_each_entry_safe(this, next, head1, list) {
841                 if (!match_futex (&this->key, &key1))
842                         continue;
843                 if (++ret <= nr_wake) {
844                         wake_futex(this);
845                 } else {
846                         /*
847                          * If key1 and key2 hash to the same bucket, no need to
848                          * requeue.
849                          */
850                         if (likely(head1 != &hb2->chain)) {
851                                 list_move_tail(&this->list, &hb2->chain);
852                                 this->lock_ptr = &hb2->lock;
853                         }
854                         this->key = key2;
855                         get_key_refs(&key2);
856                         drop_count++;
857
858                         if (ret - nr_wake >= nr_requeue)
859                                 break;
860                 }
861         }
862
863 out_unlock:
864         spin_unlock(&hb1->lock);
865         if (hb1 != hb2)
866                 spin_unlock(&hb2->lock);
867
868         /* drop_key_refs() must be called outside the spinlocks. */
869         while (--drop_count >= 0)
870                 drop_key_refs(&key1);
871
872 out:
873         up_read(&current->mm->mmap_sem);
874         return ret;
875 }
876
877 /* The key must be already stored in q->key. */
878 static inline struct futex_hash_bucket *
879 queue_lock(struct futex_q *q, int fd, struct file *filp)
880 {
881         struct futex_hash_bucket *hb;
882
883         q->fd = fd;
884         q->filp = filp;
885
886         init_waitqueue_head(&q->waiters);
887
888         get_key_refs(&q->key);
889         hb = hash_futex(&q->key);
890         q->lock_ptr = &hb->lock;
891
892         spin_lock(&hb->lock);
893         return hb;
894 }
895
896 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
897 {
898         list_add_tail(&q->list, &hb->chain);
899         q->task = current;
900         spin_unlock(&hb->lock);
901 }
902
903 static inline void
904 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
905 {
906         spin_unlock(&hb->lock);
907         drop_key_refs(&q->key);
908 }
909
910 /*
911  * queue_me and unqueue_me must be called as a pair, each
912  * exactly once.  They are called with the hashed spinlock held.
913  */
914
915 /* The key must be already stored in q->key. */
916 static void queue_me(struct futex_q *q, int fd, struct file *filp)
917 {
918         struct futex_hash_bucket *hb;
919
920         hb = queue_lock(q, fd, filp);
921         __queue_me(q, hb);
922 }
923
924 /* Return 1 if we were still queued (ie. 0 means we were woken) */
925 static int unqueue_me(struct futex_q *q)
926 {
927         spinlock_t *lock_ptr;
928         int ret = 0;
929
930         /* In the common case we don't take the spinlock, which is nice. */
931  retry:
932         lock_ptr = q->lock_ptr;
933         if (lock_ptr != 0) {
934                 spin_lock(lock_ptr);
935                 /*
936                  * q->lock_ptr can change between reading it and
937                  * spin_lock(), causing us to take the wrong lock.  This
938                  * corrects the race condition.
939                  *
940                  * Reasoning goes like this: if we have the wrong lock,
941                  * q->lock_ptr must have changed (maybe several times)
942                  * between reading it and the spin_lock().  It can
943                  * change again after the spin_lock() but only if it was
944                  * already changed before the spin_lock().  It cannot,
945                  * however, change back to the original value.  Therefore
946                  * we can detect whether we acquired the correct lock.
947                  */
948                 if (unlikely(lock_ptr != q->lock_ptr)) {
949                         spin_unlock(lock_ptr);
950                         goto retry;
951                 }
952                 WARN_ON(list_empty(&q->list));
953                 list_del(&q->list);
954
955                 BUG_ON(q->pi_state);
956
957                 spin_unlock(lock_ptr);
958                 ret = 1;
959         }
960
961         drop_key_refs(&q->key);
962         return ret;
963 }
964
965 /*
966  * PI futexes can not be requeued and must remove themself from the
967  * hash bucket. The hash bucket lock is held on entry and dropped here.
968  */
969 static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
970 {
971         WARN_ON(list_empty(&q->list));
972         list_del(&q->list);
973
974         BUG_ON(!q->pi_state);
975         free_pi_state(q->pi_state);
976         q->pi_state = NULL;
977
978         spin_unlock(&hb->lock);
979
980         drop_key_refs(&q->key);
981 }
982
983 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
984 {
985         struct task_struct *curr = current;
986         DECLARE_WAITQUEUE(wait, curr);
987         struct futex_hash_bucket *hb;
988         struct futex_q q;
989         u32 uval;
990         int ret;
991
992         q.pi_state = NULL;
993  retry:
994         down_read(&curr->mm->mmap_sem);
995
996         ret = get_futex_key(uaddr, &q.key);
997         if (unlikely(ret != 0))
998                 goto out_release_sem;
999
1000         hb = queue_lock(&q, -1, NULL);
1001
1002         /*
1003          * Access the page AFTER the futex is queued.
1004          * Order is important:
1005          *
1006          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1007          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1008          *
1009          * The basic logical guarantee of a futex is that it blocks ONLY
1010          * if cond(var) is known to be true at the time of blocking, for
1011          * any cond.  If we queued after testing *uaddr, that would open
1012          * a race condition where we could block indefinitely with
1013          * cond(var) false, which would violate the guarantee.
1014          *
1015          * A consequence is that futex_wait() can return zero and absorb
1016          * a wakeup when *uaddr != val on entry to the syscall.  This is
1017          * rare, but normal.
1018          *
1019          * We hold the mmap semaphore, so the mapping cannot have changed
1020          * since we looked it up in get_futex_key.
1021          */
1022         ret = get_futex_value_locked(&uval, uaddr);
1023
1024         if (unlikely(ret)) {
1025                 queue_unlock(&q, hb);
1026
1027                 /*
1028                  * If we would have faulted, release mmap_sem, fault it in and
1029                  * start all over again.
1030                  */
1031                 up_read(&curr->mm->mmap_sem);
1032
1033                 ret = get_user(uval, uaddr);
1034
1035                 if (!ret)
1036                         goto retry;
1037                 return ret;
1038         }
1039         ret = -EWOULDBLOCK;
1040         if (uval != val)
1041                 goto out_unlock_release_sem;
1042
1043         /* Only actually queue if *uaddr contained val.  */
1044         __queue_me(&q, hb);
1045
1046         /*
1047          * Now the futex is queued and we have checked the data, we
1048          * don't want to hold mmap_sem while we sleep.
1049          */
1050         up_read(&curr->mm->mmap_sem);
1051
1052         /*
1053          * There might have been scheduling since the queue_me(), as we
1054          * cannot hold a spinlock across the get_user() in case it
1055          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1056          * queueing ourselves into the futex hash.  This code thus has to
1057          * rely on the futex_wake() code removing us from hash when it
1058          * wakes us up.
1059          */
1060
1061         /* add_wait_queue is the barrier after __set_current_state. */
1062         __set_current_state(TASK_INTERRUPTIBLE);
1063         add_wait_queue(&q.waiters, &wait);
1064         /*
1065          * !list_empty() is safe here without any lock.
1066          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1067          */
1068         if (likely(!list_empty(&q.list)))
1069                 time = schedule_timeout(time);
1070         __set_current_state(TASK_RUNNING);
1071
1072         /*
1073          * NOTE: we don't remove ourselves from the waitqueue because
1074          * we are the only user of it.
1075          */
1076
1077         /* If we were woken (and unqueued), we succeeded, whatever. */
1078         if (!unqueue_me(&q))
1079                 return 0;
1080         if (time == 0)
1081                 return -ETIMEDOUT;
1082         /*
1083          * We expect signal_pending(current), but another thread may
1084          * have handled it for us already.
1085          */
1086         return -EINTR;
1087
1088  out_unlock_release_sem:
1089         queue_unlock(&q, hb);
1090
1091  out_release_sem:
1092         up_read(&curr->mm->mmap_sem);
1093         return ret;
1094 }
1095
1096 /*
1097  * Userspace tried a 0 -> TID atomic transition of the futex value
1098  * and failed. The kernel side here does the whole locking operation:
1099  * if there are waiters then it will block, it does PI, etc. (Due to
1100  * races the kernel might see a 0 value of the futex too.)
1101  */
1102 static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
1103                             struct hrtimer_sleeper *to)
1104 {
1105         struct task_struct *curr = current;
1106         struct futex_hash_bucket *hb;
1107         u32 uval, newval, curval;
1108         struct futex_q q;
1109         int ret, attempt = 0;
1110
1111         if (refill_pi_state_cache())
1112                 return -ENOMEM;
1113
1114         q.pi_state = NULL;
1115  retry:
1116         down_read(&curr->mm->mmap_sem);
1117
1118         ret = get_futex_key(uaddr, &q.key);
1119         if (unlikely(ret != 0))
1120                 goto out_release_sem;
1121
1122         hb = queue_lock(&q, -1, NULL);
1123
1124  retry_locked:
1125         /*
1126          * To avoid races, we attempt to take the lock here again
1127          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1128          * the locks. It will most likely not succeed.
1129          */
1130         newval = current->pid;
1131
1132         inc_preempt_count();
1133         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1134         dec_preempt_count();
1135
1136         if (unlikely(curval == -EFAULT))
1137                 goto uaddr_faulted;
1138
1139         /* We own the lock already */
1140         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1141                 if (!detect && 0)
1142                         force_sig(SIGKILL, current);
1143                 ret = -EDEADLK;
1144                 goto out_unlock_release_sem;
1145         }
1146
1147         /*
1148          * Surprise - we got the lock. Just return
1149          * to userspace:
1150          */
1151         if (unlikely(!curval))
1152                 goto out_unlock_release_sem;
1153
1154         uval = curval;
1155         newval = uval | FUTEX_WAITERS;
1156
1157         inc_preempt_count();
1158         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1159         dec_preempt_count();
1160
1161         if (unlikely(curval == -EFAULT))
1162                 goto uaddr_faulted;
1163         if (unlikely(curval != uval))
1164                 goto retry_locked;
1165
1166         /*
1167          * We dont have the lock. Look up the PI state (or create it if
1168          * we are the first waiter):
1169          */
1170         ret = lookup_pi_state(uval, hb, &q);
1171
1172         if (unlikely(ret)) {
1173                 /*
1174                  * There were no waiters and the owner task lookup
1175                  * failed. When the OWNER_DIED bit is set, then we
1176                  * know that this is a robust futex and we actually
1177                  * take the lock. This is safe as we are protected by
1178                  * the hash bucket lock. We also set the waiters bit
1179                  * unconditionally here, to simplify glibc handling of
1180                  * multiple tasks racing to acquire the lock and
1181                  * cleanup the problems which were left by the dead
1182                  * owner.
1183                  */
1184                 if (curval & FUTEX_OWNER_DIED) {
1185                         uval = newval;
1186                         newval = current->pid |
1187                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1188
1189                         inc_preempt_count();
1190                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1191                                                                uval, newval);
1192                         dec_preempt_count();
1193
1194                         if (unlikely(curval == -EFAULT))
1195                                 goto uaddr_faulted;
1196                         if (unlikely(curval != uval))
1197                                 goto retry_locked;
1198                         ret = 0;
1199                 }
1200                 goto out_unlock_release_sem;
1201         }
1202
1203         /*
1204          * Only actually queue now that the atomic ops are done:
1205          */
1206         __queue_me(&q, hb);
1207
1208         /*
1209          * Now the futex is queued and we have checked the data, we
1210          * don't want to hold mmap_sem while we sleep.
1211          */
1212         up_read(&curr->mm->mmap_sem);
1213
1214         WARN_ON(!q.pi_state);
1215         /*
1216          * Block on the PI mutex:
1217          */
1218         if (!trylock)
1219                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1220         else {
1221                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1222                 /* Fixup the trylock return value: */
1223                 ret = ret ? 0 : -EWOULDBLOCK;
1224         }
1225
1226         down_read(&curr->mm->mmap_sem);
1227         spin_lock(q.lock_ptr);
1228
1229         /*
1230          * Got the lock. We might not be the anticipated owner if we
1231          * did a lock-steal - fix up the PI-state in that case.
1232          */
1233         if (!ret && q.pi_state->owner != curr) {
1234                 u32 newtid = current->pid | FUTEX_WAITERS;
1235
1236                 /* Owner died? */
1237                 if (q.pi_state->owner != NULL) {
1238                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1239                         list_del_init(&q.pi_state->list);
1240                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1241                 } else
1242                         newtid |= FUTEX_OWNER_DIED;
1243
1244                 q.pi_state->owner = current;
1245
1246                 spin_lock_irq(&current->pi_lock);
1247                 list_add(&q.pi_state->list, &current->pi_state_list);
1248                 spin_unlock_irq(&current->pi_lock);
1249
1250                 /* Unqueue and drop the lock */
1251                 unqueue_me_pi(&q, hb);
1252                 up_read(&curr->mm->mmap_sem);
1253                 /*
1254                  * We own it, so we have to replace the pending owner
1255                  * TID. This must be atomic as we have preserve the
1256                  * owner died bit here.
1257                  */
1258                 ret = get_user(uval, uaddr);
1259                 while (!ret) {
1260                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1261                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1262                                                                uval, newval);
1263                         if (curval == -EFAULT)
1264                                 ret = -EFAULT;
1265                         if (curval == uval)
1266                                 break;
1267                         uval = curval;
1268                 }
1269         } else {
1270                 /*
1271                  * Catch the rare case, where the lock was released
1272                  * when we were on the way back before we locked
1273                  * the hash bucket.
1274                  */
1275                 if (ret && q.pi_state->owner == curr) {
1276                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1277                                 ret = 0;
1278                 }
1279                 /* Unqueue and drop the lock */
1280                 unqueue_me_pi(&q, hb);
1281                 up_read(&curr->mm->mmap_sem);
1282         }
1283
1284         if (!detect && ret == -EDEADLK && 0)
1285                 force_sig(SIGKILL, current);
1286
1287         return ret;
1288
1289  out_unlock_release_sem:
1290         queue_unlock(&q, hb);
1291
1292  out_release_sem:
1293         up_read(&curr->mm->mmap_sem);
1294         return ret;
1295
1296  uaddr_faulted:
1297         /*
1298          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1299          * non-atomically.  Therefore, if get_user below is not
1300          * enough, we need to handle the fault ourselves, while
1301          * still holding the mmap_sem.
1302          */
1303         if (attempt++) {
1304                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1305                         goto out_unlock_release_sem;
1306
1307                 goto retry_locked;
1308         }
1309
1310         queue_unlock(&q, hb);
1311         up_read(&curr->mm->mmap_sem);
1312
1313         ret = get_user(uval, uaddr);
1314         if (!ret && (uval != -EFAULT))
1315                 goto retry;
1316
1317         return ret;
1318 }
1319
1320 /*
1321  * Restart handler
1322  */
1323 static long futex_lock_pi_restart(struct restart_block *restart)
1324 {
1325         struct hrtimer_sleeper timeout, *to = NULL;
1326         int ret;
1327
1328         restart->fn = do_no_restart_syscall;
1329
1330         if (restart->arg2 || restart->arg3) {
1331                 to = &timeout;
1332                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1333                 hrtimer_init_sleeper(to, current);
1334                 to->timer.expires.tv64 = ((u64)restart->arg1 << 32) |
1335                         (u64) restart->arg0;
1336         }
1337
1338         pr_debug("lock_pi restart: %p, %d (%d)\n",
1339                  (u32 __user *)restart->arg0, current->pid);
1340
1341         ret = do_futex_lock_pi((u32 __user *)restart->arg0, restart->arg1,
1342                                0, to);
1343
1344         if (ret != -EINTR)
1345                 return ret;
1346
1347         restart->fn = futex_lock_pi_restart;
1348
1349         /* The other values are filled in */
1350         return -ERESTART_RESTARTBLOCK;
1351 }
1352
1353 /*
1354  * Called from the syscall entry below.
1355  */
1356 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1357                          long nsec, int trylock)
1358 {
1359         struct hrtimer_sleeper timeout, *to = NULL;
1360         struct restart_block *restart;
1361         int ret;
1362
1363         if (sec != MAX_SCHEDULE_TIMEOUT) {
1364                 to = &timeout;
1365                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
1366                 hrtimer_init_sleeper(to, current);
1367                 to->timer.expires = ktime_set(sec, nsec);
1368         }
1369
1370         ret = do_futex_lock_pi(uaddr, detect, trylock, to);
1371
1372         if (ret != -EINTR)
1373                 return ret;
1374
1375         pr_debug("lock_pi interrupted: %p, %d (%d)\n", uaddr, current->pid);
1376
1377         restart = &current_thread_info()->restart_block;
1378         restart->fn = futex_lock_pi_restart;
1379         restart->arg0 = (unsigned long) uaddr;
1380         restart->arg1 = detect;
1381         if (to) {
1382                 restart->arg2 = to->timer.expires.tv64 & 0xFFFFFFFF;
1383                 restart->arg3 = to->timer.expires.tv64 >> 32;
1384         } else
1385                 restart->arg2 = restart->arg3 = 0;
1386
1387         return -ERESTART_RESTARTBLOCK;
1388 }
1389
1390 /*
1391  * Userspace attempted a TID -> 0 atomic transition, and failed.
1392  * This is the in-kernel slowpath: we look up the PI state (if any),
1393  * and do the rt-mutex unlock.
1394  */
1395 static int futex_unlock_pi(u32 __user *uaddr)
1396 {
1397         struct futex_hash_bucket *hb;
1398         struct futex_q *this, *next;
1399         u32 uval;
1400         struct list_head *head;
1401         union futex_key key;
1402         int ret, attempt = 0;
1403
1404 retry:
1405         if (get_user(uval, uaddr))
1406                 return -EFAULT;
1407         /*
1408          * We release only a lock we actually own:
1409          */
1410         if ((uval & FUTEX_TID_MASK) != current->pid)
1411                 return -EPERM;
1412         /*
1413          * First take all the futex related locks:
1414          */
1415         down_read(&current->mm->mmap_sem);
1416
1417         ret = get_futex_key(uaddr, &key);
1418         if (unlikely(ret != 0))
1419                 goto out;
1420
1421         hb = hash_futex(&key);
1422         spin_lock(&hb->lock);
1423
1424 retry_locked:
1425         /*
1426          * To avoid races, try to do the TID -> 0 atomic transition
1427          * again. If it succeeds then we can return without waking
1428          * anyone else up:
1429          */
1430         inc_preempt_count();
1431         uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1432         dec_preempt_count();
1433
1434         if (unlikely(uval == -EFAULT))
1435                 goto pi_faulted;
1436         /*
1437          * Rare case: we managed to release the lock atomically,
1438          * no need to wake anyone else up:
1439          */
1440         if (unlikely(uval == current->pid))
1441                 goto out_unlock;
1442
1443         /*
1444          * Ok, other tasks may need to be woken up - check waiters
1445          * and do the wakeup if necessary:
1446          */
1447         head = &hb->chain;
1448
1449         list_for_each_entry_safe(this, next, head, list) {
1450                 if (!match_futex (&this->key, &key))
1451                         continue;
1452                 ret = wake_futex_pi(uaddr, uval, this);
1453                 /*
1454                  * The atomic access to the futex value
1455                  * generated a pagefault, so retry the
1456                  * user-access and the wakeup:
1457                  */
1458                 if (ret == -EFAULT)
1459                         goto pi_faulted;
1460                 goto out_unlock;
1461         }
1462         /*
1463          * No waiters - kernel unlocks the futex:
1464          */
1465         ret = unlock_futex_pi(uaddr, uval);
1466         if (ret == -EFAULT)
1467                 goto pi_faulted;
1468
1469 out_unlock:
1470         spin_unlock(&hb->lock);
1471 out:
1472         up_read(&current->mm->mmap_sem);
1473
1474         return ret;
1475
1476 pi_faulted:
1477         /*
1478          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1479          * non-atomically.  Therefore, if get_user below is not
1480          * enough, we need to handle the fault ourselves, while
1481          * still holding the mmap_sem.
1482          */
1483         if (attempt++) {
1484                 if (futex_handle_fault((unsigned long)uaddr, attempt))
1485                         goto out_unlock;
1486
1487                 goto retry_locked;
1488         }
1489
1490         spin_unlock(&hb->lock);
1491         up_read(&current->mm->mmap_sem);
1492
1493         ret = get_user(uval, uaddr);
1494         if (!ret && (uval != -EFAULT))
1495                 goto retry;
1496
1497         return ret;
1498 }
1499
1500 static int futex_close(struct inode *inode, struct file *filp)
1501 {
1502         struct futex_q *q = filp->private_data;
1503
1504         unqueue_me(q);
1505         kfree(q);
1506
1507         return 0;
1508 }
1509
1510 /* This is one-shot: once it's gone off you need a new fd */
1511 static unsigned int futex_poll(struct file *filp,
1512                                struct poll_table_struct *wait)
1513 {
1514         struct futex_q *q = filp->private_data;
1515         int ret = 0;
1516
1517         poll_wait(filp, &q->waiters, wait);
1518
1519         /*
1520          * list_empty() is safe here without any lock.
1521          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1522          */
1523         if (list_empty(&q->list))
1524                 ret = POLLIN | POLLRDNORM;
1525
1526         return ret;
1527 }
1528
1529 static struct file_operations futex_fops = {
1530         .release        = futex_close,
1531         .poll           = futex_poll,
1532 };
1533
1534 /*
1535  * Signal allows caller to avoid the race which would occur if they
1536  * set the sigio stuff up afterwards.
1537  */
1538 static int futex_fd(u32 __user *uaddr, int signal)
1539 {
1540         struct futex_q *q;
1541         struct file *filp;
1542         int ret, err;
1543
1544         ret = -EINVAL;
1545         if (!valid_signal(signal))
1546                 goto out;
1547
1548         ret = get_unused_fd();
1549         if (ret < 0)
1550                 goto out;
1551         filp = get_empty_filp();
1552         if (!filp) {
1553                 put_unused_fd(ret);
1554                 ret = -ENFILE;
1555                 goto out;
1556         }
1557         filp->f_op = &futex_fops;
1558         filp->f_vfsmnt = mntget(futex_mnt);
1559         filp->f_dentry = dget(futex_mnt->mnt_root);
1560         filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
1561
1562         if (signal) {
1563                 err = f_setown(filp, current->pid, 1);
1564                 if (err < 0) {
1565                         goto error;
1566                 }
1567                 filp->f_owner.signum = signal;
1568         }
1569
1570         q = kmalloc(sizeof(*q), GFP_KERNEL);
1571         if (!q) {
1572                 err = -ENOMEM;
1573                 goto error;
1574         }
1575         q->pi_state = NULL;
1576
1577         down_read(&current->mm->mmap_sem);
1578         err = get_futex_key(uaddr, &q->key);
1579
1580         if (unlikely(err != 0)) {
1581                 up_read(&current->mm->mmap_sem);
1582                 kfree(q);
1583                 goto error;
1584         }
1585
1586         /*
1587          * queue_me() must be called before releasing mmap_sem, because
1588          * key->shared.inode needs to be referenced while holding it.
1589          */
1590         filp->private_data = q;
1591
1592         queue_me(q, ret, filp);
1593         up_read(&current->mm->mmap_sem);
1594
1595         /* Now we map fd to filp, so userspace can access it */
1596         fd_install(ret, filp);
1597 out:
1598         return ret;
1599 error:
1600         put_unused_fd(ret);
1601         put_filp(filp);
1602         ret = err;
1603         goto out;
1604 }
1605
1606 /*
1607  * Support for robust futexes: the kernel cleans up held futexes at
1608  * thread exit time.
1609  *
1610  * Implementation: user-space maintains a per-thread list of locks it
1611  * is holding. Upon do_exit(), the kernel carefully walks this list,
1612  * and marks all locks that are owned by this thread with the
1613  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1614  * always manipulated with the lock held, so the list is private and
1615  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1616  * field, to allow the kernel to clean up if the thread dies after
1617  * acquiring the lock, but just before it could have added itself to
1618  * the list. There can only be one such pending lock.
1619  */
1620
1621 /**
1622  * sys_set_robust_list - set the robust-futex list head of a task
1623  * @head: pointer to the list-head
1624  * @len: length of the list-head, as userspace expects
1625  */
1626 asmlinkage long
1627 sys_set_robust_list(struct robust_list_head __user *head,
1628                     size_t len)
1629 {
1630         /*
1631          * The kernel knows only one size for now:
1632          */
1633         if (unlikely(len != sizeof(*head)))
1634                 return -EINVAL;
1635
1636         current->robust_list = head;
1637
1638         return 0;
1639 }
1640
1641 /**
1642  * sys_get_robust_list - get the robust-futex list head of a task
1643  * @pid: pid of the process [zero for current task]
1644  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1645  * @len_ptr: pointer to a length field, the kernel fills in the header size
1646  */
1647 asmlinkage long
1648 sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
1649                     size_t __user *len_ptr)
1650 {
1651         struct robust_list_head *head;
1652         unsigned long ret;
1653
1654         if (!pid)
1655                 head = current->robust_list;
1656         else {
1657                 struct task_struct *p;
1658
1659                 ret = -ESRCH;
1660                 read_lock(&tasklist_lock);
1661                 p = find_task_by_pid(pid);
1662                 if (!p)
1663                         goto err_unlock;
1664                 ret = -EPERM;
1665                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1666                                 !capable(CAP_SYS_PTRACE))
1667                         goto err_unlock;
1668                 head = p->robust_list;
1669                 read_unlock(&tasklist_lock);
1670         }
1671
1672         if (put_user(sizeof(*head), len_ptr))
1673                 return -EFAULT;
1674         return put_user(head, head_ptr);
1675
1676 err_unlock:
1677         read_unlock(&tasklist_lock);
1678
1679         return ret;
1680 }
1681
1682 /*
1683  * Process a futex-list entry, check whether it's owned by the
1684  * dying task, and do notification if so:
1685  */
1686 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr)
1687 {
1688         u32 uval, nval;
1689
1690 retry:
1691         if (get_user(uval, uaddr))
1692                 return -1;
1693
1694         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1695                 /*
1696                  * Ok, this dying thread is truly holding a futex
1697                  * of interest. Set the OWNER_DIED bit atomically
1698                  * via cmpxchg, and if the value had FUTEX_WAITERS
1699                  * set, wake up a waiter (if any). (We have to do a
1700                  * futex_wake() even if OWNER_DIED is already set -
1701                  * to handle the rare but possible case of recursive
1702                  * thread-death.) The rest of the cleanup is done in
1703                  * userspace.
1704                  */
1705                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval,
1706                                                      uval | FUTEX_OWNER_DIED);
1707                 if (nval == -EFAULT)
1708                         return -1;
1709
1710                 if (nval != uval)
1711                         goto retry;
1712
1713                 if (uval & FUTEX_WAITERS)
1714                         futex_wake(uaddr, 1);
1715         }
1716         return 0;
1717 }
1718
1719 /*
1720  * Walk curr->robust_list (very carefully, it's a userspace list!)
1721  * and mark any locks found there dead, and notify any waiters.
1722  *
1723  * We silently return on any sign of list-walking problem.
1724  */
1725 void exit_robust_list(struct task_struct *curr)
1726 {
1727         struct robust_list_head __user *head = curr->robust_list;
1728         struct robust_list __user *entry, *pending;
1729         unsigned int limit = ROBUST_LIST_LIMIT;
1730         unsigned long futex_offset;
1731
1732         /*
1733          * Fetch the list head (which was registered earlier, via
1734          * sys_set_robust_list()):
1735          */
1736         if (get_user(entry, &head->list.next))
1737                 return;
1738         /*
1739          * Fetch the relative futex offset:
1740          */
1741         if (get_user(futex_offset, &head->futex_offset))
1742                 return;
1743         /*
1744          * Fetch any possibly pending lock-add first, and handle it
1745          * if it exists:
1746          */
1747         if (get_user(pending, &head->list_op_pending))
1748                 return;
1749         if (pending)
1750                 handle_futex_death((void *)pending + futex_offset, curr);
1751
1752         while (entry != &head->list) {
1753                 /*
1754                  * A pending lock might already be on the list, so
1755                  * don't process it twice:
1756                  */
1757                 if (entry != pending)
1758                         if (handle_futex_death((void *)entry + futex_offset,
1759                                                 curr))
1760                                 return;
1761                 /*
1762                  * Fetch the next entry in the list:
1763                  */
1764                 if (get_user(entry, &entry->next))
1765                         return;
1766                 /*
1767                  * Avoid excessively long or circular lists:
1768                  */
1769                 if (!--limit)
1770                         break;
1771
1772                 cond_resched();
1773         }
1774 }
1775
1776 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1777                 u32 __user *uaddr2, u32 val2, u32 val3)
1778 {
1779         int ret;
1780
1781         switch (op) {
1782         case FUTEX_WAIT:
1783                 ret = futex_wait(uaddr, val, timeout);
1784                 break;
1785         case FUTEX_WAKE:
1786                 ret = futex_wake(uaddr, val);
1787                 break;
1788         case FUTEX_FD:
1789                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1790                 ret = futex_fd(uaddr, val);
1791                 break;
1792         case FUTEX_REQUEUE:
1793                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1794                 break;
1795         case FUTEX_CMP_REQUEUE:
1796                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1797                 break;
1798         case FUTEX_WAKE_OP:
1799                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1800                 break;
1801         case FUTEX_LOCK_PI:
1802                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1803                 break;
1804         case FUTEX_UNLOCK_PI:
1805                 ret = futex_unlock_pi(uaddr);
1806                 break;
1807         case FUTEX_TRYLOCK_PI:
1808                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1809                 break;
1810         default:
1811                 ret = -ENOSYS;
1812         }
1813         return ret;
1814 }
1815
1816
1817 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1818                           struct timespec __user *utime, u32 __user *uaddr2,
1819                           u32 val3)
1820 {
1821         struct timespec t;
1822         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1823         u32 val2 = 0;
1824
1825         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1826                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1827                         return -EFAULT;
1828                 if (!timespec_valid(&t))
1829                         return -EINVAL;
1830                 if (op == FUTEX_WAIT)
1831                         timeout = timespec_to_jiffies(&t) + 1;
1832                 else {
1833                         timeout = t.tv_sec;
1834                         val2 = t.tv_nsec;
1835                 }
1836         }
1837         /*
1838          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1839          */
1840         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1841                 val2 = (u32) (unsigned long) utime;
1842
1843         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1844 }
1845
1846 static int futexfs_get_sb(struct file_system_type *fs_type,
1847                           int flags, const char *dev_name, void *data,
1848                           struct vfsmount *mnt)
1849 {
1850         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1851 }
1852
1853 static struct file_system_type futex_fs_type = {
1854         .name           = "futexfs",
1855         .get_sb         = futexfs_get_sb,
1856         .kill_sb        = kill_anon_super,
1857 };
1858
1859 static int __init init(void)
1860 {
1861         unsigned int i;
1862
1863         register_filesystem(&futex_fs_type);
1864         futex_mnt = kern_mount(&futex_fs_type);
1865
1866         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1867                 INIT_LIST_HEAD(&futex_queues[i].chain);
1868                 spin_lock_init(&futex_queues[i].lock);
1869         }
1870         return 0;
1871 }
1872 __initcall(init);