futex: restartable futex_wait
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
20  *  enough at me, Linus for the original (flawed) idea, Matthew
21  *  Kirkwood for proof-of-concept implementation.
22  *
23  *  "The futexes are also cursed."
24  *  "But they come in a choice of three flavours!"
25  *
26  *  This program is free software; you can redistribute it and/or modify
27  *  it under the terms of the GNU General Public License as published by
28  *  the Free Software Foundation; either version 2 of the License, or
29  *  (at your option) any later version.
30  *
31  *  This program is distributed in the hope that it will be useful,
32  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
33  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
34  *  GNU General Public License for more details.
35  *
36  *  You should have received a copy of the GNU General Public License
37  *  along with this program; if not, write to the Free Software
38  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
39  */
40 #include <linux/slab.h>
41 #include <linux/poll.h>
42 #include <linux/fs.h>
43 #include <linux/file.h>
44 #include <linux/jhash.h>
45 #include <linux/init.h>
46 #include <linux/futex.h>
47 #include <linux/mount.h>
48 #include <linux/pagemap.h>
49 #include <linux/syscalls.h>
50 #include <linux/signal.h>
51 #include <linux/module.h>
52 #include <asm/futex.h>
53
54 #include "rtmutex_common.h"
55
56 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
57
58 /*
59  * Priority Inheritance state:
60  */
61 struct futex_pi_state {
62         /*
63          * list of 'owned' pi_state instances - these have to be
64          * cleaned up in do_exit() if the task exits prematurely:
65          */
66         struct list_head list;
67
68         /*
69          * The PI object:
70          */
71         struct rt_mutex pi_mutex;
72
73         struct task_struct *owner;
74         atomic_t refcount;
75
76         union futex_key key;
77 };
78
79 /*
80  * We use this hashed waitqueue instead of a normal wait_queue_t, so
81  * we can wake only the relevant ones (hashed queues may be shared).
82  *
83  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
84  * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0.
85  * The order of wakup is always to make the first condition true, then
86  * wake up q->waiters, then make the second condition true.
87  */
88 struct futex_q {
89         struct list_head list;
90         wait_queue_head_t waiters;
91
92         /* Which hash list lock to use: */
93         spinlock_t *lock_ptr;
94
95         /* Key which the futex is hashed on: */
96         union futex_key key;
97
98         /* For fd, sigio sent using these: */
99         int fd;
100         struct file *filp;
101
102         /* Optional priority inheritance state: */
103         struct futex_pi_state *pi_state;
104         struct task_struct *task;
105 };
106
107 /*
108  * Split the global futex_lock into every hash list lock.
109  */
110 struct futex_hash_bucket {
111        spinlock_t              lock;
112        struct list_head       chain;
113 };
114
115 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
116
117 /* Futex-fs vfsmount entry: */
118 static struct vfsmount *futex_mnt;
119
120 /*
121  * We hash on the keys returned from get_futex_key (see below).
122  */
123 static struct futex_hash_bucket *hash_futex(union futex_key *key)
124 {
125         u32 hash = jhash2((u32*)&key->both.word,
126                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
127                           key->both.offset);
128         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
129 }
130
131 /*
132  * Return 1 if two futex_keys are equal, 0 otherwise.
133  */
134 static inline int match_futex(union futex_key *key1, union futex_key *key2)
135 {
136         return (key1->both.word == key2->both.word
137                 && key1->both.ptr == key2->both.ptr
138                 && key1->both.offset == key2->both.offset);
139 }
140
141 /*
142  * Get parameters which are the keys for a futex.
143  *
144  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
145  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
146  * We can usually work out the index without swapping in the page.
147  *
148  * Returns: 0, or negative error code.
149  * The key words are stored in *key on success.
150  *
151  * Should be called with &current->mm->mmap_sem but NOT any spinlocks.
152  */
153 int get_futex_key(u32 __user *uaddr, union futex_key *key)
154 {
155         unsigned long address = (unsigned long)uaddr;
156         struct mm_struct *mm = current->mm;
157         struct vm_area_struct *vma;
158         struct page *page;
159         int err;
160
161         /*
162          * The futex address must be "naturally" aligned.
163          */
164         key->both.offset = address % PAGE_SIZE;
165         if (unlikely((key->both.offset % sizeof(u32)) != 0))
166                 return -EINVAL;
167         address -= key->both.offset;
168
169         /*
170          * The futex is hashed differently depending on whether
171          * it's in a shared or private mapping.  So check vma first.
172          */
173         vma = find_extend_vma(mm, address);
174         if (unlikely(!vma))
175                 return -EFAULT;
176
177         /*
178          * Permissions.
179          */
180         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
181                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
182
183         /*
184          * Private mappings are handled in a simple way.
185          *
186          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
187          * it's a read-only handle, it's expected that futexes attach to
188          * the object not the particular process.  Therefore we use
189          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
190          * mappings of _writable_ handles.
191          */
192         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
193                 key->private.mm = mm;
194                 key->private.address = address;
195                 return 0;
196         }
197
198         /*
199          * Linear file mappings are also simple.
200          */
201         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
202         key->both.offset++; /* Bit 0 of offset indicates inode-based key. */
203         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
204                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
205                                      + vma->vm_pgoff);
206                 return 0;
207         }
208
209         /*
210          * We could walk the page table to read the non-linear
211          * pte, and get the page index without fetching the page
212          * from swap.  But that's a lot of code to duplicate here
213          * for a rare case, so we simply fetch the page.
214          */
215         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
216         if (err >= 0) {
217                 key->shared.pgoff =
218                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
219                 put_page(page);
220                 return 0;
221         }
222         return err;
223 }
224 EXPORT_SYMBOL_GPL(get_futex_key);
225
226 /*
227  * Take a reference to the resource addressed by a key.
228  * Can be called while holding spinlocks.
229  *
230  * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
231  * function, if it is called at all.  mmap_sem keeps key->shared.inode valid.
232  */
233 inline void get_futex_key_refs(union futex_key *key)
234 {
235         if (key->both.ptr != 0) {
236                 if (key->both.offset & 1)
237                         atomic_inc(&key->shared.inode->i_count);
238                 else
239                         atomic_inc(&key->private.mm->mm_count);
240         }
241 }
242 EXPORT_SYMBOL_GPL(get_futex_key_refs);
243
244 /*
245  * Drop a reference to the resource addressed by a key.
246  * The hash bucket spinlock must not be held.
247  */
248 void drop_futex_key_refs(union futex_key *key)
249 {
250         if (key->both.ptr != 0) {
251                 if (key->both.offset & 1)
252                         iput(key->shared.inode);
253                 else
254                         mmdrop(key->private.mm);
255         }
256 }
257 EXPORT_SYMBOL_GPL(drop_futex_key_refs);
258
259 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
260 {
261         int ret;
262
263         pagefault_disable();
264         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
265         pagefault_enable();
266
267         return ret ? -EFAULT : 0;
268 }
269
270 /*
271  * Fault handling. Called with current->mm->mmap_sem held.
272  */
273 static int futex_handle_fault(unsigned long address, int attempt)
274 {
275         struct vm_area_struct * vma;
276         struct mm_struct *mm = current->mm;
277
278         if (attempt > 2 || !(vma = find_vma(mm, address)) ||
279             vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
280                 return -EFAULT;
281
282         switch (handle_mm_fault(mm, vma, address, 1)) {
283         case VM_FAULT_MINOR:
284                 current->min_flt++;
285                 break;
286         case VM_FAULT_MAJOR:
287                 current->maj_flt++;
288                 break;
289         default:
290                 return -EFAULT;
291         }
292         return 0;
293 }
294
295 /*
296  * PI code:
297  */
298 static int refill_pi_state_cache(void)
299 {
300         struct futex_pi_state *pi_state;
301
302         if (likely(current->pi_state_cache))
303                 return 0;
304
305         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
306
307         if (!pi_state)
308                 return -ENOMEM;
309
310         INIT_LIST_HEAD(&pi_state->list);
311         /* pi_mutex gets initialized later */
312         pi_state->owner = NULL;
313         atomic_set(&pi_state->refcount, 1);
314
315         current->pi_state_cache = pi_state;
316
317         return 0;
318 }
319
320 static struct futex_pi_state * alloc_pi_state(void)
321 {
322         struct futex_pi_state *pi_state = current->pi_state_cache;
323
324         WARN_ON(!pi_state);
325         current->pi_state_cache = NULL;
326
327         return pi_state;
328 }
329
330 static void free_pi_state(struct futex_pi_state *pi_state)
331 {
332         if (!atomic_dec_and_test(&pi_state->refcount))
333                 return;
334
335         /*
336          * If pi_state->owner is NULL, the owner is most probably dying
337          * and has cleaned up the pi_state already
338          */
339         if (pi_state->owner) {
340                 spin_lock_irq(&pi_state->owner->pi_lock);
341                 list_del_init(&pi_state->list);
342                 spin_unlock_irq(&pi_state->owner->pi_lock);
343
344                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
345         }
346
347         if (current->pi_state_cache)
348                 kfree(pi_state);
349         else {
350                 /*
351                  * pi_state->list is already empty.
352                  * clear pi_state->owner.
353                  * refcount is at 0 - put it back to 1.
354                  */
355                 pi_state->owner = NULL;
356                 atomic_set(&pi_state->refcount, 1);
357                 current->pi_state_cache = pi_state;
358         }
359 }
360
361 /*
362  * Look up the task based on what TID userspace gave us.
363  * We dont trust it.
364  */
365 static struct task_struct * futex_find_get_task(pid_t pid)
366 {
367         struct task_struct *p;
368
369         rcu_read_lock();
370         p = find_task_by_pid(pid);
371         if (!p)
372                 goto out_unlock;
373         if ((current->euid != p->euid) && (current->euid != p->uid)) {
374                 p = NULL;
375                 goto out_unlock;
376         }
377         if (p->exit_state != 0) {
378                 p = NULL;
379                 goto out_unlock;
380         }
381         get_task_struct(p);
382 out_unlock:
383         rcu_read_unlock();
384
385         return p;
386 }
387
388 /*
389  * This task is holding PI mutexes at exit time => bad.
390  * Kernel cleans up PI-state, but userspace is likely hosed.
391  * (Robust-futex cleanup is separate and might save the day for userspace.)
392  */
393 void exit_pi_state_list(struct task_struct *curr)
394 {
395         struct list_head *next, *head = &curr->pi_state_list;
396         struct futex_pi_state *pi_state;
397         struct futex_hash_bucket *hb;
398         union futex_key key;
399
400         /*
401          * We are a ZOMBIE and nobody can enqueue itself on
402          * pi_state_list anymore, but we have to be careful
403          * versus waiters unqueueing themselves:
404          */
405         spin_lock_irq(&curr->pi_lock);
406         while (!list_empty(head)) {
407
408                 next = head->next;
409                 pi_state = list_entry(next, struct futex_pi_state, list);
410                 key = pi_state->key;
411                 hb = hash_futex(&key);
412                 spin_unlock_irq(&curr->pi_lock);
413
414                 spin_lock(&hb->lock);
415
416                 spin_lock_irq(&curr->pi_lock);
417                 /*
418                  * We dropped the pi-lock, so re-check whether this
419                  * task still owns the PI-state:
420                  */
421                 if (head->next != next) {
422                         spin_unlock(&hb->lock);
423                         continue;
424                 }
425
426                 WARN_ON(pi_state->owner != curr);
427                 WARN_ON(list_empty(&pi_state->list));
428                 list_del_init(&pi_state->list);
429                 pi_state->owner = NULL;
430                 spin_unlock_irq(&curr->pi_lock);
431
432                 rt_mutex_unlock(&pi_state->pi_mutex);
433
434                 spin_unlock(&hb->lock);
435
436                 spin_lock_irq(&curr->pi_lock);
437         }
438         spin_unlock_irq(&curr->pi_lock);
439 }
440
441 static int
442 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
443 {
444         struct futex_pi_state *pi_state = NULL;
445         struct futex_q *this, *next;
446         struct list_head *head;
447         struct task_struct *p;
448         pid_t pid;
449
450         head = &hb->chain;
451
452         list_for_each_entry_safe(this, next, head, list) {
453                 if (match_futex(&this->key, &me->key)) {
454                         /*
455                          * Another waiter already exists - bump up
456                          * the refcount and return its pi_state:
457                          */
458                         pi_state = this->pi_state;
459                         /*
460                          * Userspace might have messed up non PI and PI futexes
461                          */
462                         if (unlikely(!pi_state))
463                                 return -EINVAL;
464
465                         WARN_ON(!atomic_read(&pi_state->refcount));
466
467                         atomic_inc(&pi_state->refcount);
468                         me->pi_state = pi_state;
469
470                         return 0;
471                 }
472         }
473
474         /*
475          * We are the first waiter - try to look up the real owner and attach
476          * the new pi_state to it, but bail out when the owner died bit is set
477          * and TID = 0:
478          */
479         pid = uval & FUTEX_TID_MASK;
480         if (!pid && (uval & FUTEX_OWNER_DIED))
481                 return -ESRCH;
482         p = futex_find_get_task(pid);
483         if (!p)
484                 return -ESRCH;
485
486         pi_state = alloc_pi_state();
487
488         /*
489          * Initialize the pi_mutex in locked state and make 'p'
490          * the owner of it:
491          */
492         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
493
494         /* Store the key for possible exit cleanups: */
495         pi_state->key = me->key;
496
497         spin_lock_irq(&p->pi_lock);
498         WARN_ON(!list_empty(&pi_state->list));
499         list_add(&pi_state->list, &p->pi_state_list);
500         pi_state->owner = p;
501         spin_unlock_irq(&p->pi_lock);
502
503         put_task_struct(p);
504
505         me->pi_state = pi_state;
506
507         return 0;
508 }
509
510 /*
511  * The hash bucket lock must be held when this is called.
512  * Afterwards, the futex_q must not be accessed.
513  */
514 static void wake_futex(struct futex_q *q)
515 {
516         list_del_init(&q->list);
517         if (q->filp)
518                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
519         /*
520          * The lock in wake_up_all() is a crucial memory barrier after the
521          * list_del_init() and also before assigning to q->lock_ptr.
522          */
523         wake_up_all(&q->waiters);
524         /*
525          * The waiting task can free the futex_q as soon as this is written,
526          * without taking any locks.  This must come last.
527          *
528          * A memory barrier is required here to prevent the following store
529          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
530          * at the end of wake_up_all() does not prevent this store from
531          * moving.
532          */
533         smp_wmb();
534         q->lock_ptr = NULL;
535 }
536
537 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
538 {
539         struct task_struct *new_owner;
540         struct futex_pi_state *pi_state = this->pi_state;
541         u32 curval, newval;
542
543         if (!pi_state)
544                 return -EINVAL;
545
546         spin_lock(&pi_state->pi_mutex.wait_lock);
547         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
548
549         /*
550          * This happens when we have stolen the lock and the original
551          * pending owner did not enqueue itself back on the rt_mutex.
552          * Thats not a tragedy. We know that way, that a lock waiter
553          * is on the fly. We make the futex_q waiter the pending owner.
554          */
555         if (!new_owner)
556                 new_owner = this->task;
557
558         /*
559          * We pass it to the next owner. (The WAITERS bit is always
560          * kept enabled while there is PI state around. We must also
561          * preserve the owner died bit.)
562          */
563         if (!(uval & FUTEX_OWNER_DIED)) {
564                 newval = FUTEX_WAITERS | new_owner->pid;
565
566                 pagefault_disable();
567                 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
568                 pagefault_enable();
569                 if (curval == -EFAULT)
570                         return -EFAULT;
571                 if (curval != uval)
572                         return -EINVAL;
573         }
574
575         spin_lock_irq(&pi_state->owner->pi_lock);
576         WARN_ON(list_empty(&pi_state->list));
577         list_del_init(&pi_state->list);
578         spin_unlock_irq(&pi_state->owner->pi_lock);
579
580         spin_lock_irq(&new_owner->pi_lock);
581         WARN_ON(!list_empty(&pi_state->list));
582         list_add(&pi_state->list, &new_owner->pi_state_list);
583         pi_state->owner = new_owner;
584         spin_unlock_irq(&new_owner->pi_lock);
585
586         spin_unlock(&pi_state->pi_mutex.wait_lock);
587         rt_mutex_unlock(&pi_state->pi_mutex);
588
589         return 0;
590 }
591
592 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
593 {
594         u32 oldval;
595
596         /*
597          * There is no waiter, so we unlock the futex. The owner died
598          * bit has not to be preserved here. We are the owner:
599          */
600         pagefault_disable();
601         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
602         pagefault_enable();
603
604         if (oldval == -EFAULT)
605                 return oldval;
606         if (oldval != uval)
607                 return -EAGAIN;
608
609         return 0;
610 }
611
612 /*
613  * Express the locking dependencies for lockdep:
614  */
615 static inline void
616 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
617 {
618         if (hb1 <= hb2) {
619                 spin_lock(&hb1->lock);
620                 if (hb1 < hb2)
621                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
622         } else { /* hb1 > hb2 */
623                 spin_lock(&hb2->lock);
624                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
625         }
626 }
627
628 /*
629  * Wake up all waiters hashed on the physical page that is mapped
630  * to this virtual address:
631  */
632 static int futex_wake(u32 __user *uaddr, int nr_wake)
633 {
634         struct futex_hash_bucket *hb;
635         struct futex_q *this, *next;
636         struct list_head *head;
637         union futex_key key;
638         int ret;
639
640         down_read(&current->mm->mmap_sem);
641
642         ret = get_futex_key(uaddr, &key);
643         if (unlikely(ret != 0))
644                 goto out;
645
646         hb = hash_futex(&key);
647         spin_lock(&hb->lock);
648         head = &hb->chain;
649
650         list_for_each_entry_safe(this, next, head, list) {
651                 if (match_futex (&this->key, &key)) {
652                         if (this->pi_state) {
653                                 ret = -EINVAL;
654                                 break;
655                         }
656                         wake_futex(this);
657                         if (++ret >= nr_wake)
658                                 break;
659                 }
660         }
661
662         spin_unlock(&hb->lock);
663 out:
664         up_read(&current->mm->mmap_sem);
665         return ret;
666 }
667
668 /*
669  * Wake up all waiters hashed on the physical page that is mapped
670  * to this virtual address:
671  */
672 static int
673 futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
674               int nr_wake, int nr_wake2, int op)
675 {
676         union futex_key key1, key2;
677         struct futex_hash_bucket *hb1, *hb2;
678         struct list_head *head;
679         struct futex_q *this, *next;
680         int ret, op_ret, attempt = 0;
681
682 retryfull:
683         down_read(&current->mm->mmap_sem);
684
685         ret = get_futex_key(uaddr1, &key1);
686         if (unlikely(ret != 0))
687                 goto out;
688         ret = get_futex_key(uaddr2, &key2);
689         if (unlikely(ret != 0))
690                 goto out;
691
692         hb1 = hash_futex(&key1);
693         hb2 = hash_futex(&key2);
694
695 retry:
696         double_lock_hb(hb1, hb2);
697
698         op_ret = futex_atomic_op_inuser(op, uaddr2);
699         if (unlikely(op_ret < 0)) {
700                 u32 dummy;
701
702                 spin_unlock(&hb1->lock);
703                 if (hb1 != hb2)
704                         spin_unlock(&hb2->lock);
705
706 #ifndef CONFIG_MMU
707                 /*
708                  * we don't get EFAULT from MMU faults if we don't have an MMU,
709                  * but we might get them from range checking
710                  */
711                 ret = op_ret;
712                 goto out;
713 #endif
714
715                 if (unlikely(op_ret != -EFAULT)) {
716                         ret = op_ret;
717                         goto out;
718                 }
719
720                 /*
721                  * futex_atomic_op_inuser needs to both read and write
722                  * *(int __user *)uaddr2, but we can't modify it
723                  * non-atomically.  Therefore, if get_user below is not
724                  * enough, we need to handle the fault ourselves, while
725                  * still holding the mmap_sem.
726                  */
727                 if (attempt++) {
728                         if (futex_handle_fault((unsigned long)uaddr2,
729                                                 attempt)) {
730                                 ret = -EFAULT;
731                                 goto out;
732                         }
733                         goto retry;
734                 }
735
736                 /*
737                  * If we would have faulted, release mmap_sem,
738                  * fault it in and start all over again.
739                  */
740                 up_read(&current->mm->mmap_sem);
741
742                 ret = get_user(dummy, uaddr2);
743                 if (ret)
744                         return ret;
745
746                 goto retryfull;
747         }
748
749         head = &hb1->chain;
750
751         list_for_each_entry_safe(this, next, head, list) {
752                 if (match_futex (&this->key, &key1)) {
753                         wake_futex(this);
754                         if (++ret >= nr_wake)
755                                 break;
756                 }
757         }
758
759         if (op_ret > 0) {
760                 head = &hb2->chain;
761
762                 op_ret = 0;
763                 list_for_each_entry_safe(this, next, head, list) {
764                         if (match_futex (&this->key, &key2)) {
765                                 wake_futex(this);
766                                 if (++op_ret >= nr_wake2)
767                                         break;
768                         }
769                 }
770                 ret += op_ret;
771         }
772
773         spin_unlock(&hb1->lock);
774         if (hb1 != hb2)
775                 spin_unlock(&hb2->lock);
776 out:
777         up_read(&current->mm->mmap_sem);
778         return ret;
779 }
780
781 /*
782  * Requeue all waiters hashed on one physical page to another
783  * physical page.
784  */
785 static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
786                          int nr_wake, int nr_requeue, u32 *cmpval)
787 {
788         union futex_key key1, key2;
789         struct futex_hash_bucket *hb1, *hb2;
790         struct list_head *head1;
791         struct futex_q *this, *next;
792         int ret, drop_count = 0;
793
794  retry:
795         down_read(&current->mm->mmap_sem);
796
797         ret = get_futex_key(uaddr1, &key1);
798         if (unlikely(ret != 0))
799                 goto out;
800         ret = get_futex_key(uaddr2, &key2);
801         if (unlikely(ret != 0))
802                 goto out;
803
804         hb1 = hash_futex(&key1);
805         hb2 = hash_futex(&key2);
806
807         double_lock_hb(hb1, hb2);
808
809         if (likely(cmpval != NULL)) {
810                 u32 curval;
811
812                 ret = get_futex_value_locked(&curval, uaddr1);
813
814                 if (unlikely(ret)) {
815                         spin_unlock(&hb1->lock);
816                         if (hb1 != hb2)
817                                 spin_unlock(&hb2->lock);
818
819                         /*
820                          * If we would have faulted, release mmap_sem, fault
821                          * it in and start all over again.
822                          */
823                         up_read(&current->mm->mmap_sem);
824
825                         ret = get_user(curval, uaddr1);
826
827                         if (!ret)
828                                 goto retry;
829
830                         return ret;
831                 }
832                 if (curval != *cmpval) {
833                         ret = -EAGAIN;
834                         goto out_unlock;
835                 }
836         }
837
838         head1 = &hb1->chain;
839         list_for_each_entry_safe(this, next, head1, list) {
840                 if (!match_futex (&this->key, &key1))
841                         continue;
842                 if (++ret <= nr_wake) {
843                         wake_futex(this);
844                 } else {
845                         /*
846                          * If key1 and key2 hash to the same bucket, no need to
847                          * requeue.
848                          */
849                         if (likely(head1 != &hb2->chain)) {
850                                 list_move_tail(&this->list, &hb2->chain);
851                                 this->lock_ptr = &hb2->lock;
852                         }
853                         this->key = key2;
854                         get_futex_key_refs(&key2);
855                         drop_count++;
856
857                         if (ret - nr_wake >= nr_requeue)
858                                 break;
859                 }
860         }
861
862 out_unlock:
863         spin_unlock(&hb1->lock);
864         if (hb1 != hb2)
865                 spin_unlock(&hb2->lock);
866
867         /* drop_futex_key_refs() must be called outside the spinlocks. */
868         while (--drop_count >= 0)
869                 drop_futex_key_refs(&key1);
870
871 out:
872         up_read(&current->mm->mmap_sem);
873         return ret;
874 }
875
876 /* The key must be already stored in q->key. */
877 static inline struct futex_hash_bucket *
878 queue_lock(struct futex_q *q, int fd, struct file *filp)
879 {
880         struct futex_hash_bucket *hb;
881
882         q->fd = fd;
883         q->filp = filp;
884
885         init_waitqueue_head(&q->waiters);
886
887         get_futex_key_refs(&q->key);
888         hb = hash_futex(&q->key);
889         q->lock_ptr = &hb->lock;
890
891         spin_lock(&hb->lock);
892         return hb;
893 }
894
895 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
896 {
897         list_add_tail(&q->list, &hb->chain);
898         q->task = current;
899         spin_unlock(&hb->lock);
900 }
901
902 static inline void
903 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
904 {
905         spin_unlock(&hb->lock);
906         drop_futex_key_refs(&q->key);
907 }
908
909 /*
910  * queue_me and unqueue_me must be called as a pair, each
911  * exactly once.  They are called with the hashed spinlock held.
912  */
913
914 /* The key must be already stored in q->key. */
915 static void queue_me(struct futex_q *q, int fd, struct file *filp)
916 {
917         struct futex_hash_bucket *hb;
918
919         hb = queue_lock(q, fd, filp);
920         __queue_me(q, hb);
921 }
922
923 /* Return 1 if we were still queued (ie. 0 means we were woken) */
924 static int unqueue_me(struct futex_q *q)
925 {
926         spinlock_t *lock_ptr;
927         int ret = 0;
928
929         /* In the common case we don't take the spinlock, which is nice. */
930  retry:
931         lock_ptr = q->lock_ptr;
932         barrier();
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_futex_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_futex_key_refs(&q->key);
981 }
982
983 static long futex_wait_restart(struct restart_block *restart);
984 static int futex_wait_abstime(u32 __user *uaddr, u32 val,
985                         int timed, unsigned long abs_time)
986 {
987         struct task_struct *curr = current;
988         DECLARE_WAITQUEUE(wait, curr);
989         struct futex_hash_bucket *hb;
990         struct futex_q q;
991         unsigned long time_left = 0;
992         u32 uval;
993         int ret;
994
995         q.pi_state = NULL;
996  retry:
997         down_read(&curr->mm->mmap_sem);
998
999         ret = get_futex_key(uaddr, &q.key);
1000         if (unlikely(ret != 0))
1001                 goto out_release_sem;
1002
1003         hb = queue_lock(&q, -1, NULL);
1004
1005         /*
1006          * Access the page AFTER the futex is queued.
1007          * Order is important:
1008          *
1009          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1010          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1011          *
1012          * The basic logical guarantee of a futex is that it blocks ONLY
1013          * if cond(var) is known to be true at the time of blocking, for
1014          * any cond.  If we queued after testing *uaddr, that would open
1015          * a race condition where we could block indefinitely with
1016          * cond(var) false, which would violate the guarantee.
1017          *
1018          * A consequence is that futex_wait() can return zero and absorb
1019          * a wakeup when *uaddr != val on entry to the syscall.  This is
1020          * rare, but normal.
1021          *
1022          * We hold the mmap semaphore, so the mapping cannot have changed
1023          * since we looked it up in get_futex_key.
1024          */
1025         ret = get_futex_value_locked(&uval, uaddr);
1026
1027         if (unlikely(ret)) {
1028                 queue_unlock(&q, hb);
1029
1030                 /*
1031                  * If we would have faulted, release mmap_sem, fault it in and
1032                  * start all over again.
1033                  */
1034                 up_read(&curr->mm->mmap_sem);
1035
1036                 ret = get_user(uval, uaddr);
1037
1038                 if (!ret)
1039                         goto retry;
1040                 return ret;
1041         }
1042         ret = -EWOULDBLOCK;
1043         if (uval != val)
1044                 goto out_unlock_release_sem;
1045
1046         /* Only actually queue if *uaddr contained val.  */
1047         __queue_me(&q, hb);
1048
1049         /*
1050          * Now the futex is queued and we have checked the data, we
1051          * don't want to hold mmap_sem while we sleep.
1052          */
1053         up_read(&curr->mm->mmap_sem);
1054
1055         /*
1056          * There might have been scheduling since the queue_me(), as we
1057          * cannot hold a spinlock across the get_user() in case it
1058          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1059          * queueing ourselves into the futex hash.  This code thus has to
1060          * rely on the futex_wake() code removing us from hash when it
1061          * wakes us up.
1062          */
1063
1064         /* add_wait_queue is the barrier after __set_current_state. */
1065         __set_current_state(TASK_INTERRUPTIBLE);
1066         add_wait_queue(&q.waiters, &wait);
1067         /*
1068          * !list_empty() is safe here without any lock.
1069          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1070          */
1071         time_left = 0;
1072         if (likely(!list_empty(&q.list))) {
1073                 unsigned long rel_time;
1074
1075                 if (timed) {
1076                         unsigned long now = jiffies;
1077                         if (time_after(now, abs_time))
1078                                 rel_time = 0;
1079                         else
1080                                 rel_time = abs_time - now;
1081                 } else
1082                         rel_time = MAX_SCHEDULE_TIMEOUT;
1083
1084                 time_left = schedule_timeout(rel_time);
1085         }
1086         __set_current_state(TASK_RUNNING);
1087
1088         /*
1089          * NOTE: we don't remove ourselves from the waitqueue because
1090          * we are the only user of it.
1091          */
1092
1093         /* If we were woken (and unqueued), we succeeded, whatever. */
1094         if (!unqueue_me(&q))
1095                 return 0;
1096         if (time_left == 0)
1097                 return -ETIMEDOUT;
1098
1099         /*
1100          * We expect signal_pending(current), but another thread may
1101          * have handled it for us already.
1102          */
1103         if (time_left == MAX_SCHEDULE_TIMEOUT)
1104                 return -ERESTARTSYS;
1105         else {
1106                 struct restart_block *restart;
1107                 restart = &current_thread_info()->restart_block;
1108                 restart->fn = futex_wait_restart;
1109                 restart->arg0 = (unsigned long)uaddr;
1110                 restart->arg1 = (unsigned long)val;
1111                 restart->arg2 = (unsigned long)timed;
1112                 restart->arg3 = abs_time;
1113                 return -ERESTART_RESTARTBLOCK;
1114         }
1115
1116  out_unlock_release_sem:
1117         queue_unlock(&q, hb);
1118
1119  out_release_sem:
1120         up_read(&curr->mm->mmap_sem);
1121         return ret;
1122 }
1123
1124 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long rel_time)
1125 {
1126         int timed = (rel_time != MAX_SCHEDULE_TIMEOUT);
1127         return futex_wait_abstime(uaddr, val, timed, jiffies+rel_time);
1128 }
1129
1130 static long futex_wait_restart(struct restart_block *restart)
1131 {
1132         u32 __user *uaddr = (u32 __user *)restart->arg0;
1133         u32 val = (u32)restart->arg1;
1134         int timed = (int)restart->arg2;
1135         unsigned long abs_time = restart->arg3;
1136
1137         restart->fn = do_no_restart_syscall;
1138         return (long)futex_wait_abstime(uaddr, val, timed, abs_time);
1139 }
1140
1141
1142 /*
1143  * Userspace tried a 0 -> TID atomic transition of the futex value
1144  * and failed. The kernel side here does the whole locking operation:
1145  * if there are waiters then it will block, it does PI, etc. (Due to
1146  * races the kernel might see a 0 value of the futex too.)
1147  */
1148 static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
1149                          long nsec, int trylock)
1150 {
1151         struct hrtimer_sleeper timeout, *to = NULL;
1152         struct task_struct *curr = current;
1153         struct futex_hash_bucket *hb;
1154         u32 uval, newval, curval;
1155         struct futex_q q;
1156         int ret, attempt = 0;
1157
1158         if (refill_pi_state_cache())
1159                 return -ENOMEM;
1160
1161         if (sec != MAX_SCHEDULE_TIMEOUT) {
1162                 to = &timeout;
1163                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1164                 hrtimer_init_sleeper(to, current);
1165                 to->timer.expires = ktime_set(sec, nsec);
1166         }
1167
1168         q.pi_state = NULL;
1169  retry:
1170         down_read(&curr->mm->mmap_sem);
1171
1172         ret = get_futex_key(uaddr, &q.key);
1173         if (unlikely(ret != 0))
1174                 goto out_release_sem;
1175
1176         hb = queue_lock(&q, -1, NULL);
1177
1178  retry_locked:
1179         /*
1180          * To avoid races, we attempt to take the lock here again
1181          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1182          * the locks. It will most likely not succeed.
1183          */
1184         newval = current->pid;
1185
1186         pagefault_disable();
1187         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1188         pagefault_enable();
1189
1190         if (unlikely(curval == -EFAULT))
1191                 goto uaddr_faulted;
1192
1193         /* We own the lock already */
1194         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1195                 if (!detect && 0)
1196                         force_sig(SIGKILL, current);
1197                 ret = -EDEADLK;
1198                 goto out_unlock_release_sem;
1199         }
1200
1201         /*
1202          * Surprise - we got the lock. Just return
1203          * to userspace:
1204          */
1205         if (unlikely(!curval))
1206                 goto out_unlock_release_sem;
1207
1208         uval = curval;
1209         newval = uval | FUTEX_WAITERS;
1210
1211         pagefault_disable();
1212         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1213         pagefault_enable();
1214
1215         if (unlikely(curval == -EFAULT))
1216                 goto uaddr_faulted;
1217         if (unlikely(curval != uval))
1218                 goto retry_locked;
1219
1220         /*
1221          * We dont have the lock. Look up the PI state (or create it if
1222          * we are the first waiter):
1223          */
1224         ret = lookup_pi_state(uval, hb, &q);
1225
1226         if (unlikely(ret)) {
1227                 /*
1228                  * There were no waiters and the owner task lookup
1229                  * failed. When the OWNER_DIED bit is set, then we
1230                  * know that this is a robust futex and we actually
1231                  * take the lock. This is safe as we are protected by
1232                  * the hash bucket lock. We also set the waiters bit
1233                  * unconditionally here, to simplify glibc handling of
1234                  * multiple tasks racing to acquire the lock and
1235                  * cleanup the problems which were left by the dead
1236                  * owner.
1237                  */
1238                 if (curval & FUTEX_OWNER_DIED) {
1239                         uval = newval;
1240                         newval = current->pid |
1241                                 FUTEX_OWNER_DIED | FUTEX_WAITERS;
1242
1243                         pagefault_disable();
1244                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1245                                                                uval, newval);
1246                         pagefault_enable();
1247
1248                         if (unlikely(curval == -EFAULT))
1249                                 goto uaddr_faulted;
1250                         if (unlikely(curval != uval))
1251                                 goto retry_locked;
1252                         ret = 0;
1253                 }
1254                 goto out_unlock_release_sem;
1255         }
1256
1257         /*
1258          * Only actually queue now that the atomic ops are done:
1259          */
1260         __queue_me(&q, hb);
1261
1262         /*
1263          * Now the futex is queued and we have checked the data, we
1264          * don't want to hold mmap_sem while we sleep.
1265          */
1266         up_read(&curr->mm->mmap_sem);
1267
1268         WARN_ON(!q.pi_state);
1269         /*
1270          * Block on the PI mutex:
1271          */
1272         if (!trylock)
1273                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1274         else {
1275                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1276                 /* Fixup the trylock return value: */
1277                 ret = ret ? 0 : -EWOULDBLOCK;
1278         }
1279
1280         down_read(&curr->mm->mmap_sem);
1281         spin_lock(q.lock_ptr);
1282
1283         /*
1284          * Got the lock. We might not be the anticipated owner if we
1285          * did a lock-steal - fix up the PI-state in that case.
1286          */
1287         if (!ret && q.pi_state->owner != curr) {
1288                 u32 newtid = current->pid | FUTEX_WAITERS;
1289
1290                 /* Owner died? */
1291                 if (q.pi_state->owner != NULL) {
1292                         spin_lock_irq(&q.pi_state->owner->pi_lock);
1293                         WARN_ON(list_empty(&q.pi_state->list));
1294                         list_del_init(&q.pi_state->list);
1295                         spin_unlock_irq(&q.pi_state->owner->pi_lock);
1296                 } else
1297                         newtid |= FUTEX_OWNER_DIED;
1298
1299                 q.pi_state->owner = current;
1300
1301                 spin_lock_irq(&current->pi_lock);
1302                 WARN_ON(!list_empty(&q.pi_state->list));
1303                 list_add(&q.pi_state->list, &current->pi_state_list);
1304                 spin_unlock_irq(&current->pi_lock);
1305
1306                 /* Unqueue and drop the lock */
1307                 unqueue_me_pi(&q, hb);
1308                 up_read(&curr->mm->mmap_sem);
1309                 /*
1310                  * We own it, so we have to replace the pending owner
1311                  * TID. This must be atomic as we have preserve the
1312                  * owner died bit here.
1313                  */
1314                 ret = get_user(uval, uaddr);
1315                 while (!ret) {
1316                         newval = (uval & FUTEX_OWNER_DIED) | newtid;
1317                         curval = futex_atomic_cmpxchg_inatomic(uaddr,
1318                                                                uval, newval);
1319                         if (curval == -EFAULT)
1320                                 ret = -EFAULT;
1321                         if (curval == uval)
1322                                 break;
1323                         uval = curval;
1324                 }
1325         } else {
1326                 /*
1327                  * Catch the rare case, where the lock was released
1328                  * when we were on the way back before we locked
1329                  * the hash bucket.
1330                  */
1331                 if (ret && q.pi_state->owner == curr) {
1332                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1333                                 ret = 0;
1334                 }
1335                 /* Unqueue and drop the lock */
1336                 unqueue_me_pi(&q, hb);
1337                 up_read(&curr->mm->mmap_sem);
1338         }
1339
1340         if (!detect && ret == -EDEADLK && 0)
1341                 force_sig(SIGKILL, current);
1342
1343         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1344
1345  out_unlock_release_sem:
1346         queue_unlock(&q, hb);
1347
1348  out_release_sem:
1349         up_read(&curr->mm->mmap_sem);
1350         return ret;
1351
1352  uaddr_faulted:
1353         /*
1354          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1355          * non-atomically.  Therefore, if get_user below is not
1356          * enough, we need to handle the fault ourselves, while
1357          * still holding the mmap_sem.
1358          */
1359         if (attempt++) {
1360                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1361                         ret = -EFAULT;
1362                         goto out_unlock_release_sem;
1363                 }
1364                 goto retry_locked;
1365         }
1366
1367         queue_unlock(&q, hb);
1368         up_read(&curr->mm->mmap_sem);
1369
1370         ret = get_user(uval, uaddr);
1371         if (!ret && (uval != -EFAULT))
1372                 goto retry;
1373
1374         return ret;
1375 }
1376
1377 /*
1378  * Userspace attempted a TID -> 0 atomic transition, and failed.
1379  * This is the in-kernel slowpath: we look up the PI state (if any),
1380  * and do the rt-mutex unlock.
1381  */
1382 static int futex_unlock_pi(u32 __user *uaddr)
1383 {
1384         struct futex_hash_bucket *hb;
1385         struct futex_q *this, *next;
1386         u32 uval;
1387         struct list_head *head;
1388         union futex_key key;
1389         int ret, attempt = 0;
1390
1391 retry:
1392         if (get_user(uval, uaddr))
1393                 return -EFAULT;
1394         /*
1395          * We release only a lock we actually own:
1396          */
1397         if ((uval & FUTEX_TID_MASK) != current->pid)
1398                 return -EPERM;
1399         /*
1400          * First take all the futex related locks:
1401          */
1402         down_read(&current->mm->mmap_sem);
1403
1404         ret = get_futex_key(uaddr, &key);
1405         if (unlikely(ret != 0))
1406                 goto out;
1407
1408         hb = hash_futex(&key);
1409         spin_lock(&hb->lock);
1410
1411 retry_locked:
1412         /*
1413          * To avoid races, try to do the TID -> 0 atomic transition
1414          * again. If it succeeds then we can return without waking
1415          * anyone else up:
1416          */
1417         if (!(uval & FUTEX_OWNER_DIED)) {
1418                 pagefault_disable();
1419                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1420                 pagefault_enable();
1421         }
1422
1423         if (unlikely(uval == -EFAULT))
1424                 goto pi_faulted;
1425         /*
1426          * Rare case: we managed to release the lock atomically,
1427          * no need to wake anyone else up:
1428          */
1429         if (unlikely(uval == current->pid))
1430                 goto out_unlock;
1431
1432         /*
1433          * Ok, other tasks may need to be woken up - check waiters
1434          * and do the wakeup if necessary:
1435          */
1436         head = &hb->chain;
1437
1438         list_for_each_entry_safe(this, next, head, list) {
1439                 if (!match_futex (&this->key, &key))
1440                         continue;
1441                 ret = wake_futex_pi(uaddr, uval, this);
1442                 /*
1443                  * The atomic access to the futex value
1444                  * generated a pagefault, so retry the
1445                  * user-access and the wakeup:
1446                  */
1447                 if (ret == -EFAULT)
1448                         goto pi_faulted;
1449                 goto out_unlock;
1450         }
1451         /*
1452          * No waiters - kernel unlocks the futex:
1453          */
1454         if (!(uval & FUTEX_OWNER_DIED)) {
1455                 ret = unlock_futex_pi(uaddr, uval);
1456                 if (ret == -EFAULT)
1457                         goto pi_faulted;
1458         }
1459
1460 out_unlock:
1461         spin_unlock(&hb->lock);
1462 out:
1463         up_read(&current->mm->mmap_sem);
1464
1465         return ret;
1466
1467 pi_faulted:
1468         /*
1469          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1470          * non-atomically.  Therefore, if get_user below is not
1471          * enough, we need to handle the fault ourselves, while
1472          * still holding the mmap_sem.
1473          */
1474         if (attempt++) {
1475                 if (futex_handle_fault((unsigned long)uaddr, attempt)) {
1476                         ret = -EFAULT;
1477                         goto out_unlock;
1478                 }
1479                 goto retry_locked;
1480         }
1481
1482         spin_unlock(&hb->lock);
1483         up_read(&current->mm->mmap_sem);
1484
1485         ret = get_user(uval, uaddr);
1486         if (!ret && (uval != -EFAULT))
1487                 goto retry;
1488
1489         return ret;
1490 }
1491
1492 static int futex_close(struct inode *inode, struct file *filp)
1493 {
1494         struct futex_q *q = filp->private_data;
1495
1496         unqueue_me(q);
1497         kfree(q);
1498
1499         return 0;
1500 }
1501
1502 /* This is one-shot: once it's gone off you need a new fd */
1503 static unsigned int futex_poll(struct file *filp,
1504                                struct poll_table_struct *wait)
1505 {
1506         struct futex_q *q = filp->private_data;
1507         int ret = 0;
1508
1509         poll_wait(filp, &q->waiters, wait);
1510
1511         /*
1512          * list_empty() is safe here without any lock.
1513          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1514          */
1515         if (list_empty(&q->list))
1516                 ret = POLLIN | POLLRDNORM;
1517
1518         return ret;
1519 }
1520
1521 static const struct file_operations futex_fops = {
1522         .release        = futex_close,
1523         .poll           = futex_poll,
1524 };
1525
1526 /*
1527  * Signal allows caller to avoid the race which would occur if they
1528  * set the sigio stuff up afterwards.
1529  */
1530 static int futex_fd(u32 __user *uaddr, int signal)
1531 {
1532         struct futex_q *q;
1533         struct file *filp;
1534         int ret, err;
1535         static unsigned long printk_interval;
1536
1537         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1538                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1539                         "will be removed from the kernel in June 2007\n",
1540                         current->comm);
1541         }
1542
1543         ret = -EINVAL;
1544         if (!valid_signal(signal))
1545                 goto out;
1546
1547         ret = get_unused_fd();
1548         if (ret < 0)
1549                 goto out;
1550         filp = get_empty_filp();
1551         if (!filp) {
1552                 put_unused_fd(ret);
1553                 ret = -ENFILE;
1554                 goto out;
1555         }
1556         filp->f_op = &futex_fops;
1557         filp->f_path.mnt = mntget(futex_mnt);
1558         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1559         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1560
1561         if (signal) {
1562                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1563                 if (err < 0) {
1564                         goto error;
1565                 }
1566                 filp->f_owner.signum = signal;
1567         }
1568
1569         q = kmalloc(sizeof(*q), GFP_KERNEL);
1570         if (!q) {
1571                 err = -ENOMEM;
1572                 goto error;
1573         }
1574         q->pi_state = NULL;
1575
1576         down_read(&current->mm->mmap_sem);
1577         err = get_futex_key(uaddr, &q->key);
1578
1579         if (unlikely(err != 0)) {
1580                 up_read(&current->mm->mmap_sem);
1581                 kfree(q);
1582                 goto error;
1583         }
1584
1585         /*
1586          * queue_me() must be called before releasing mmap_sem, because
1587          * key->shared.inode needs to be referenced while holding it.
1588          */
1589         filp->private_data = q;
1590
1591         queue_me(q, ret, filp);
1592         up_read(&current->mm->mmap_sem);
1593
1594         /* Now we map fd to filp, so userspace can access it */
1595         fd_install(ret, filp);
1596 out:
1597         return ret;
1598 error:
1599         put_unused_fd(ret);
1600         put_filp(filp);
1601         ret = err;
1602         goto out;
1603 }
1604
1605 /*
1606  * Support for robust futexes: the kernel cleans up held futexes at
1607  * thread exit time.
1608  *
1609  * Implementation: user-space maintains a per-thread list of locks it
1610  * is holding. Upon do_exit(), the kernel carefully walks this list,
1611  * and marks all locks that are owned by this thread with the
1612  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1613  * always manipulated with the lock held, so the list is private and
1614  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1615  * field, to allow the kernel to clean up if the thread dies after
1616  * acquiring the lock, but just before it could have added itself to
1617  * the list. There can only be one such pending lock.
1618  */
1619
1620 /**
1621  * sys_set_robust_list - set the robust-futex list head of a task
1622  * @head: pointer to the list-head
1623  * @len: length of the list-head, as userspace expects
1624  */
1625 asmlinkage long
1626 sys_set_robust_list(struct robust_list_head __user *head,
1627                     size_t len)
1628 {
1629         /*
1630          * The kernel knows only one size for now:
1631          */
1632         if (unlikely(len != sizeof(*head)))
1633                 return -EINVAL;
1634
1635         current->robust_list = head;
1636
1637         return 0;
1638 }
1639
1640 /**
1641  * sys_get_robust_list - get the robust-futex list head of a task
1642  * @pid: pid of the process [zero for current task]
1643  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1644  * @len_ptr: pointer to a length field, the kernel fills in the header size
1645  */
1646 asmlinkage long
1647 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1648                     size_t __user *len_ptr)
1649 {
1650         struct robust_list_head __user *head;
1651         unsigned long ret;
1652
1653         if (!pid)
1654                 head = current->robust_list;
1655         else {
1656                 struct task_struct *p;
1657
1658                 ret = -ESRCH;
1659                 rcu_read_lock();
1660                 p = find_task_by_pid(pid);
1661                 if (!p)
1662                         goto err_unlock;
1663                 ret = -EPERM;
1664                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1665                                 !capable(CAP_SYS_PTRACE))
1666                         goto err_unlock;
1667                 head = p->robust_list;
1668                 rcu_read_unlock();
1669         }
1670
1671         if (put_user(sizeof(*head), len_ptr))
1672                 return -EFAULT;
1673         return put_user(head, head_ptr);
1674
1675 err_unlock:
1676         rcu_read_unlock();
1677
1678         return ret;
1679 }
1680
1681 /*
1682  * Process a futex-list entry, check whether it's owned by the
1683  * dying task, and do notification if so:
1684  */
1685 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1686 {
1687         u32 uval, nval, mval;
1688
1689 retry:
1690         if (get_user(uval, uaddr))
1691                 return -1;
1692
1693         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1694                 /*
1695                  * Ok, this dying thread is truly holding a futex
1696                  * of interest. Set the OWNER_DIED bit atomically
1697                  * via cmpxchg, and if the value had FUTEX_WAITERS
1698                  * set, wake up a waiter (if any). (We have to do a
1699                  * futex_wake() even if OWNER_DIED is already set -
1700                  * to handle the rare but possible case of recursive
1701                  * thread-death.) The rest of the cleanup is done in
1702                  * userspace.
1703                  */
1704                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1705                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1706
1707                 if (nval == -EFAULT)
1708                         return -1;
1709
1710                 if (nval != uval)
1711                         goto retry;
1712
1713                 /*
1714                  * Wake robust non-PI futexes here. The wakeup of
1715                  * PI futexes happens in exit_pi_state():
1716                  */
1717                 if (!pi) {
1718                         if (uval & FUTEX_WAITERS)
1719                                 futex_wake(uaddr, 1);
1720                 }
1721         }
1722         return 0;
1723 }
1724
1725 /*
1726  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1727  */
1728 static inline int fetch_robust_entry(struct robust_list __user **entry,
1729                                      struct robust_list __user * __user *head,
1730                                      int *pi)
1731 {
1732         unsigned long uentry;
1733
1734         if (get_user(uentry, (unsigned long __user *)head))
1735                 return -EFAULT;
1736
1737         *entry = (void __user *)(uentry & ~1UL);
1738         *pi = uentry & 1;
1739
1740         return 0;
1741 }
1742
1743 /*
1744  * Walk curr->robust_list (very carefully, it's a userspace list!)
1745  * and mark any locks found there dead, and notify any waiters.
1746  *
1747  * We silently return on any sign of list-walking problem.
1748  */
1749 void exit_robust_list(struct task_struct *curr)
1750 {
1751         struct robust_list_head __user *head = curr->robust_list;
1752         struct robust_list __user *entry, *pending;
1753         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1754         unsigned long futex_offset;
1755
1756         /*
1757          * Fetch the list head (which was registered earlier, via
1758          * sys_set_robust_list()):
1759          */
1760         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1761                 return;
1762         /*
1763          * Fetch the relative futex offset:
1764          */
1765         if (get_user(futex_offset, &head->futex_offset))
1766                 return;
1767         /*
1768          * Fetch any possibly pending lock-add first, and handle it
1769          * if it exists:
1770          */
1771         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1772                 return;
1773
1774         if (pending)
1775                 handle_futex_death((void __user *)pending + futex_offset, curr, pip);
1776
1777         while (entry != &head->list) {
1778                 /*
1779                  * A pending lock might already be on the list, so
1780                  * don't process it twice:
1781                  */
1782                 if (entry != pending)
1783                         if (handle_futex_death((void __user *)entry + futex_offset,
1784                                                 curr, pi))
1785                                 return;
1786                 /*
1787                  * Fetch the next entry in the list:
1788                  */
1789                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1790                         return;
1791                 /*
1792                  * Avoid excessively long or circular lists:
1793                  */
1794                 if (!--limit)
1795                         break;
1796
1797                 cond_resched();
1798         }
1799 }
1800
1801 long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
1802                 u32 __user *uaddr2, u32 val2, u32 val3)
1803 {
1804         int ret;
1805
1806         switch (op) {
1807         case FUTEX_WAIT:
1808                 ret = futex_wait(uaddr, val, timeout);
1809                 break;
1810         case FUTEX_WAKE:
1811                 ret = futex_wake(uaddr, val);
1812                 break;
1813         case FUTEX_FD:
1814                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
1815                 ret = futex_fd(uaddr, val);
1816                 break;
1817         case FUTEX_REQUEUE:
1818                 ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
1819                 break;
1820         case FUTEX_CMP_REQUEUE:
1821                 ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
1822                 break;
1823         case FUTEX_WAKE_OP:
1824                 ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
1825                 break;
1826         case FUTEX_LOCK_PI:
1827                 ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
1828                 break;
1829         case FUTEX_UNLOCK_PI:
1830                 ret = futex_unlock_pi(uaddr);
1831                 break;
1832         case FUTEX_TRYLOCK_PI:
1833                 ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
1834                 break;
1835         default:
1836                 ret = -ENOSYS;
1837         }
1838         return ret;
1839 }
1840
1841
1842 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
1843                           struct timespec __user *utime, u32 __user *uaddr2,
1844                           u32 val3)
1845 {
1846         struct timespec t;
1847         unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
1848         u32 val2 = 0;
1849
1850         if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
1851                 if (copy_from_user(&t, utime, sizeof(t)) != 0)
1852                         return -EFAULT;
1853                 if (!timespec_valid(&t))
1854                         return -EINVAL;
1855                 if (op == FUTEX_WAIT)
1856                         timeout = timespec_to_jiffies(&t) + 1;
1857                 else {
1858                         timeout = t.tv_sec;
1859                         val2 = t.tv_nsec;
1860                 }
1861         }
1862         /*
1863          * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
1864          */
1865         if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
1866                 val2 = (u32) (unsigned long) utime;
1867
1868         return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
1869 }
1870
1871 static int futexfs_get_sb(struct file_system_type *fs_type,
1872                           int flags, const char *dev_name, void *data,
1873                           struct vfsmount *mnt)
1874 {
1875         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
1876 }
1877
1878 static struct file_system_type futex_fs_type = {
1879         .name           = "futexfs",
1880         .get_sb         = futexfs_get_sb,
1881         .kill_sb        = kill_anon_super,
1882 };
1883
1884 static int __init init(void)
1885 {
1886         int i = register_filesystem(&futex_fs_type);
1887
1888         if (i)
1889                 return i;
1890
1891         futex_mnt = kern_mount(&futex_fs_type);
1892         if (IS_ERR(futex_mnt)) {
1893                 unregister_filesystem(&futex_fs_type);
1894                 return PTR_ERR(futex_mnt);
1895         }
1896
1897         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
1898                 INIT_LIST_HEAD(&futex_queues[i].chain);
1899                 spin_lock_init(&futex_queues[i].lock);
1900         }
1901         return 0;
1902 }
1903 __initcall(init);