futex: Prevent stale futex owner when interrupted/timeout
[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  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
23  *  enough at me, Linus for the original (flawed) idea, Matthew
24  *  Kirkwood for proof-of-concept implementation.
25  *
26  *  "The futexes are also cursed."
27  *  "But they come in a choice of three flavours!"
28  *
29  *  This program is free software; you can redistribute it and/or modify
30  *  it under the terms of the GNU General Public License as published by
31  *  the Free Software Foundation; either version 2 of the License, or
32  *  (at your option) any later version.
33  *
34  *  This program is distributed in the hope that it will be useful,
35  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
36  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
37  *  GNU General Public License for more details.
38  *
39  *  You should have received a copy of the GNU General Public License
40  *  along with this program; if not, write to the Free Software
41  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
42  */
43 #include <linux/slab.h>
44 #include <linux/poll.h>
45 #include <linux/fs.h>
46 #include <linux/file.h>
47 #include <linux/jhash.h>
48 #include <linux/init.h>
49 #include <linux/futex.h>
50 #include <linux/mount.h>
51 #include <linux/pagemap.h>
52 #include <linux/syscalls.h>
53 #include <linux/signal.h>
54 #include <linux/module.h>
55 #include <linux/magic.h>
56 #include <linux/pid.h>
57 #include <linux/nsproxy.h>
58
59 #include <asm/futex.h>
60
61 #include "rtmutex_common.h"
62
63 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
64
65 /*
66  * Priority Inheritance state:
67  */
68 struct futex_pi_state {
69         /*
70          * list of 'owned' pi_state instances - these have to be
71          * cleaned up in do_exit() if the task exits prematurely:
72          */
73         struct list_head list;
74
75         /*
76          * The PI object:
77          */
78         struct rt_mutex pi_mutex;
79
80         struct task_struct *owner;
81         atomic_t refcount;
82
83         union futex_key key;
84 };
85
86 /*
87  * We use this hashed waitqueue instead of a normal wait_queue_t, so
88  * we can wake only the relevant ones (hashed queues may be shared).
89  *
90  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
91  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
92  * The order of wakup is always to make the first condition true, then
93  * wake up q->waiters, then make the second condition true.
94  */
95 struct futex_q {
96         struct plist_node list;
97         wait_queue_head_t waiters;
98
99         /* Which hash list lock to use: */
100         spinlock_t *lock_ptr;
101
102         /* Key which the futex is hashed on: */
103         union futex_key key;
104
105         /* For fd, sigio sent using these: */
106         int fd;
107         struct file *filp;
108
109         /* Optional priority inheritance state: */
110         struct futex_pi_state *pi_state;
111         struct task_struct *task;
112 };
113
114 /*
115  * Split the global futex_lock into every hash list lock.
116  */
117 struct futex_hash_bucket {
118         spinlock_t lock;
119         struct plist_head chain;
120 };
121
122 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
123
124 /* Futex-fs vfsmount entry: */
125 static struct vfsmount *futex_mnt;
126
127 /*
128  * Take mm->mmap_sem, when futex is shared
129  */
130 static inline void futex_lock_mm(struct rw_semaphore *fshared)
131 {
132         if (fshared)
133                 down_read(fshared);
134 }
135
136 /*
137  * Release mm->mmap_sem, when the futex is shared
138  */
139 static inline void futex_unlock_mm(struct rw_semaphore *fshared)
140 {
141         if (fshared)
142                 up_read(fshared);
143 }
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_futex_key - Get parameters which are the keys for a futex.
168  * @uaddr: virtual address of the futex
169  * @shared: NULL for a PROCESS_PRIVATE futex,
170  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
171  * @key: address where result is stored.
172  *
173  * Returns a negative error code or 0
174  * The key words are stored in *key on success.
175  *
176  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
177  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
178  * We can usually work out the index without swapping in the page.
179  *
180  * fshared is NULL for PROCESS_PRIVATE futexes
181  * For other futexes, it points to &current->mm->mmap_sem and
182  * caller must have taken the reader lock. but NOT any spinlocks.
183  */
184 static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
185                          union futex_key *key)
186 {
187         unsigned long address = (unsigned long)uaddr;
188         struct mm_struct *mm = current->mm;
189         struct vm_area_struct *vma;
190         struct page *page;
191         int err;
192
193         /*
194          * The futex address must be "naturally" aligned.
195          */
196         key->both.offset = address % PAGE_SIZE;
197         if (unlikely((address % sizeof(u32)) != 0))
198                 return -EINVAL;
199         address -= key->both.offset;
200
201         /*
202          * PROCESS_PRIVATE futexes are fast.
203          * As the mm cannot disappear under us and the 'key' only needs
204          * virtual address, we dont even have to find the underlying vma.
205          * Note : We do have to check 'uaddr' is a valid user address,
206          *        but access_ok() should be faster than find_vma()
207          */
208         if (!fshared) {
209                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
210                         return -EFAULT;
211                 key->private.mm = mm;
212                 key->private.address = address;
213                 return 0;
214         }
215         /*
216          * The futex is hashed differently depending on whether
217          * it's in a shared or private mapping.  So check vma first.
218          */
219         vma = find_extend_vma(mm, address);
220         if (unlikely(!vma))
221                 return -EFAULT;
222
223         /*
224          * Permissions.
225          */
226         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
227                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
228
229         /*
230          * Private mappings are handled in a simple way.
231          *
232          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
233          * it's a read-only handle, it's expected that futexes attach to
234          * the object not the particular process.  Therefore we use
235          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
236          * mappings of _writable_ handles.
237          */
238         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
239                 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
240                 key->private.mm = mm;
241                 key->private.address = address;
242                 return 0;
243         }
244
245         /*
246          * Linear file mappings are also simple.
247          */
248         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
249         key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
250         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
251                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
252                                      + vma->vm_pgoff);
253                 return 0;
254         }
255
256         /*
257          * We could walk the page table to read the non-linear
258          * pte, and get the page index without fetching the page
259          * from swap.  But that's a lot of code to duplicate here
260          * for a rare case, so we simply fetch the page.
261          */
262         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
263         if (err >= 0) {
264                 key->shared.pgoff =
265                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
266                 put_page(page);
267                 return 0;
268         }
269         return err;
270 }
271
272 /*
273  * Take a reference to the resource addressed by a key.
274  * Can be called while holding spinlocks.
275  *
276  */
277 static void get_futex_key_refs(union futex_key *key)
278 {
279         if (key->both.ptr == 0)
280                 return;
281         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
282                 case FUT_OFF_INODE:
283                         atomic_inc(&key->shared.inode->i_count);
284                         break;
285                 case FUT_OFF_MMSHARED:
286                         atomic_inc(&key->private.mm->mm_count);
287                         break;
288         }
289 }
290
291 /*
292  * Drop a reference to the resource addressed by a key.
293  * The hash bucket spinlock must not be held.
294  */
295 static void drop_futex_key_refs(union futex_key *key)
296 {
297         if (!key->both.ptr)
298                 return;
299         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
300                 case FUT_OFF_INODE:
301                         iput(key->shared.inode);
302                         break;
303                 case FUT_OFF_MMSHARED:
304                         mmdrop(key->private.mm);
305                         break;
306         }
307 }
308
309 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
310 {
311         u32 curval;
312
313         pagefault_disable();
314         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
315         pagefault_enable();
316
317         return curval;
318 }
319
320 static int get_futex_value_locked(u32 *dest, u32 __user *from)
321 {
322         int ret;
323
324         pagefault_disable();
325         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
326         pagefault_enable();
327
328         return ret ? -EFAULT : 0;
329 }
330
331 /*
332  * Fault handling.
333  * if fshared is non NULL, current->mm->mmap_sem is already held
334  */
335 static int futex_handle_fault(unsigned long address,
336                               struct rw_semaphore *fshared, int attempt)
337 {
338         struct vm_area_struct * vma;
339         struct mm_struct *mm = current->mm;
340         int ret = -EFAULT;
341
342         if (attempt > 2)
343                 return ret;
344
345         if (!fshared)
346                 down_read(&mm->mmap_sem);
347         vma = find_vma(mm, address);
348         if (vma && address >= vma->vm_start &&
349             (vma->vm_flags & VM_WRITE)) {
350                 int fault;
351                 fault = handle_mm_fault(mm, vma, address, 1);
352                 if (unlikely((fault & VM_FAULT_ERROR))) {
353 #if 0
354                         /* XXX: let's do this when we verify it is OK */
355                         if (ret & VM_FAULT_OOM)
356                                 ret = -ENOMEM;
357 #endif
358                 } else {
359                         ret = 0;
360                         if (fault & VM_FAULT_MAJOR)
361                                 current->maj_flt++;
362                         else
363                                 current->min_flt++;
364                 }
365         }
366         if (!fshared)
367                 up_read(&mm->mmap_sem);
368         return ret;
369 }
370
371 /*
372  * PI code:
373  */
374 static int refill_pi_state_cache(void)
375 {
376         struct futex_pi_state *pi_state;
377
378         if (likely(current->pi_state_cache))
379                 return 0;
380
381         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
382
383         if (!pi_state)
384                 return -ENOMEM;
385
386         INIT_LIST_HEAD(&pi_state->list);
387         /* pi_mutex gets initialized later */
388         pi_state->owner = NULL;
389         atomic_set(&pi_state->refcount, 1);
390
391         current->pi_state_cache = pi_state;
392
393         return 0;
394 }
395
396 static struct futex_pi_state * alloc_pi_state(void)
397 {
398         struct futex_pi_state *pi_state = current->pi_state_cache;
399
400         WARN_ON(!pi_state);
401         current->pi_state_cache = NULL;
402
403         return pi_state;
404 }
405
406 static void free_pi_state(struct futex_pi_state *pi_state)
407 {
408         if (!atomic_dec_and_test(&pi_state->refcount))
409                 return;
410
411         /*
412          * If pi_state->owner is NULL, the owner is most probably dying
413          * and has cleaned up the pi_state already
414          */
415         if (pi_state->owner) {
416                 spin_lock_irq(&pi_state->owner->pi_lock);
417                 list_del_init(&pi_state->list);
418                 spin_unlock_irq(&pi_state->owner->pi_lock);
419
420                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
421         }
422
423         if (current->pi_state_cache)
424                 kfree(pi_state);
425         else {
426                 /*
427                  * pi_state->list is already empty.
428                  * clear pi_state->owner.
429                  * refcount is at 0 - put it back to 1.
430                  */
431                 pi_state->owner = NULL;
432                 atomic_set(&pi_state->refcount, 1);
433                 current->pi_state_cache = pi_state;
434         }
435 }
436
437 /*
438  * Look up the task based on what TID userspace gave us.
439  * We dont trust it.
440  */
441 static struct task_struct * futex_find_get_task(pid_t pid)
442 {
443         struct task_struct *p;
444
445         rcu_read_lock();
446         p = find_task_by_vpid(pid);
447         if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
448                 p = ERR_PTR(-ESRCH);
449         else
450                 get_task_struct(p);
451
452         rcu_read_unlock();
453
454         return p;
455 }
456
457 /*
458  * This task is holding PI mutexes at exit time => bad.
459  * Kernel cleans up PI-state, but userspace is likely hosed.
460  * (Robust-futex cleanup is separate and might save the day for userspace.)
461  */
462 void exit_pi_state_list(struct task_struct *curr)
463 {
464         struct list_head *next, *head = &curr->pi_state_list;
465         struct futex_pi_state *pi_state;
466         struct futex_hash_bucket *hb;
467         union futex_key key;
468
469         /*
470          * We are a ZOMBIE and nobody can enqueue itself on
471          * pi_state_list anymore, but we have to be careful
472          * versus waiters unqueueing themselves:
473          */
474         spin_lock_irq(&curr->pi_lock);
475         while (!list_empty(head)) {
476
477                 next = head->next;
478                 pi_state = list_entry(next, struct futex_pi_state, list);
479                 key = pi_state->key;
480                 hb = hash_futex(&key);
481                 spin_unlock_irq(&curr->pi_lock);
482
483                 spin_lock(&hb->lock);
484
485                 spin_lock_irq(&curr->pi_lock);
486                 /*
487                  * We dropped the pi-lock, so re-check whether this
488                  * task still owns the PI-state:
489                  */
490                 if (head->next != next) {
491                         spin_unlock(&hb->lock);
492                         continue;
493                 }
494
495                 WARN_ON(pi_state->owner != curr);
496                 WARN_ON(list_empty(&pi_state->list));
497                 list_del_init(&pi_state->list);
498                 pi_state->owner = NULL;
499                 spin_unlock_irq(&curr->pi_lock);
500
501                 rt_mutex_unlock(&pi_state->pi_mutex);
502
503                 spin_unlock(&hb->lock);
504
505                 spin_lock_irq(&curr->pi_lock);
506         }
507         spin_unlock_irq(&curr->pi_lock);
508 }
509
510 static int
511 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
512                 union futex_key *key, struct futex_pi_state **ps)
513 {
514         struct futex_pi_state *pi_state = NULL;
515         struct futex_q *this, *next;
516         struct plist_head *head;
517         struct task_struct *p;
518         pid_t pid = uval & FUTEX_TID_MASK;
519
520         head = &hb->chain;
521
522         plist_for_each_entry_safe(this, next, head, list) {
523                 if (match_futex(&this->key, key)) {
524                         /*
525                          * Another waiter already exists - bump up
526                          * the refcount and return its pi_state:
527                          */
528                         pi_state = this->pi_state;
529                         /*
530                          * Userspace might have messed up non PI and PI futexes
531                          */
532                         if (unlikely(!pi_state))
533                                 return -EINVAL;
534
535                         WARN_ON(!atomic_read(&pi_state->refcount));
536                         WARN_ON(pid && pi_state->owner &&
537                                 pi_state->owner->pid != pid);
538
539                         atomic_inc(&pi_state->refcount);
540                         *ps = pi_state;
541
542                         return 0;
543                 }
544         }
545
546         /*
547          * We are the first waiter - try to look up the real owner and attach
548          * the new pi_state to it, but bail out when TID = 0
549          */
550         if (!pid)
551                 return -ESRCH;
552         p = futex_find_get_task(pid);
553         if (IS_ERR(p))
554                 return PTR_ERR(p);
555
556         /*
557          * We need to look at the task state flags to figure out,
558          * whether the task is exiting. To protect against the do_exit
559          * change of the task flags, we do this protected by
560          * p->pi_lock:
561          */
562         spin_lock_irq(&p->pi_lock);
563         if (unlikely(p->flags & PF_EXITING)) {
564                 /*
565                  * The task is on the way out. When PF_EXITPIDONE is
566                  * set, we know that the task has finished the
567                  * cleanup:
568                  */
569                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
570
571                 spin_unlock_irq(&p->pi_lock);
572                 put_task_struct(p);
573                 return ret;
574         }
575
576         pi_state = alloc_pi_state();
577
578         /*
579          * Initialize the pi_mutex in locked state and make 'p'
580          * the owner of it:
581          */
582         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
583
584         /* Store the key for possible exit cleanups: */
585         pi_state->key = *key;
586
587         WARN_ON(!list_empty(&pi_state->list));
588         list_add(&pi_state->list, &p->pi_state_list);
589         pi_state->owner = p;
590         spin_unlock_irq(&p->pi_lock);
591
592         put_task_struct(p);
593
594         *ps = pi_state;
595
596         return 0;
597 }
598
599 /*
600  * The hash bucket lock must be held when this is called.
601  * Afterwards, the futex_q must not be accessed.
602  */
603 static void wake_futex(struct futex_q *q)
604 {
605         plist_del(&q->list, &q->list.plist);
606         if (q->filp)
607                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
608         /*
609          * The lock in wake_up_all() is a crucial memory barrier after the
610          * plist_del() and also before assigning to q->lock_ptr.
611          */
612         wake_up_all(&q->waiters);
613         /*
614          * The waiting task can free the futex_q as soon as this is written,
615          * without taking any locks.  This must come last.
616          *
617          * A memory barrier is required here to prevent the following store
618          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
619          * at the end of wake_up_all() does not prevent this store from
620          * moving.
621          */
622         smp_wmb();
623         q->lock_ptr = NULL;
624 }
625
626 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
627 {
628         struct task_struct *new_owner;
629         struct futex_pi_state *pi_state = this->pi_state;
630         u32 curval, newval;
631
632         if (!pi_state)
633                 return -EINVAL;
634
635         spin_lock(&pi_state->pi_mutex.wait_lock);
636         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
637
638         /*
639          * This happens when we have stolen the lock and the original
640          * pending owner did not enqueue itself back on the rt_mutex.
641          * Thats not a tragedy. We know that way, that a lock waiter
642          * is on the fly. We make the futex_q waiter the pending owner.
643          */
644         if (!new_owner)
645                 new_owner = this->task;
646
647         /*
648          * We pass it to the next owner. (The WAITERS bit is always
649          * kept enabled while there is PI state around. We must also
650          * preserve the owner died bit.)
651          */
652         if (!(uval & FUTEX_OWNER_DIED)) {
653                 int ret = 0;
654
655                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
656
657                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
658
659                 if (curval == -EFAULT)
660                         ret = -EFAULT;
661                 else if (curval != uval)
662                         ret = -EINVAL;
663                 if (ret) {
664                         spin_unlock(&pi_state->pi_mutex.wait_lock);
665                         return ret;
666                 }
667         }
668
669         spin_lock_irq(&pi_state->owner->pi_lock);
670         WARN_ON(list_empty(&pi_state->list));
671         list_del_init(&pi_state->list);
672         spin_unlock_irq(&pi_state->owner->pi_lock);
673
674         spin_lock_irq(&new_owner->pi_lock);
675         WARN_ON(!list_empty(&pi_state->list));
676         list_add(&pi_state->list, &new_owner->pi_state_list);
677         pi_state->owner = new_owner;
678         spin_unlock_irq(&new_owner->pi_lock);
679
680         spin_unlock(&pi_state->pi_mutex.wait_lock);
681         rt_mutex_unlock(&pi_state->pi_mutex);
682
683         return 0;
684 }
685
686 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
687 {
688         u32 oldval;
689
690         /*
691          * There is no waiter, so we unlock the futex. The owner died
692          * bit has not to be preserved here. We are the owner:
693          */
694         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
695
696         if (oldval == -EFAULT)
697                 return oldval;
698         if (oldval != uval)
699                 return -EAGAIN;
700
701         return 0;
702 }
703
704 /*
705  * Express the locking dependencies for lockdep:
706  */
707 static inline void
708 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
709 {
710         if (hb1 <= hb2) {
711                 spin_lock(&hb1->lock);
712                 if (hb1 < hb2)
713                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
714         } else { /* hb1 > hb2 */
715                 spin_lock(&hb2->lock);
716                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
717         }
718 }
719
720 /*
721  * Wake up all waiters hashed on the physical page that is mapped
722  * to this virtual address:
723  */
724 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
725                       int nr_wake)
726 {
727         struct futex_hash_bucket *hb;
728         struct futex_q *this, *next;
729         struct plist_head *head;
730         union futex_key key;
731         int ret;
732
733         futex_lock_mm(fshared);
734
735         ret = get_futex_key(uaddr, fshared, &key);
736         if (unlikely(ret != 0))
737                 goto out;
738
739         hb = hash_futex(&key);
740         spin_lock(&hb->lock);
741         head = &hb->chain;
742
743         plist_for_each_entry_safe(this, next, head, list) {
744                 if (match_futex (&this->key, &key)) {
745                         if (this->pi_state) {
746                                 ret = -EINVAL;
747                                 break;
748                         }
749                         wake_futex(this);
750                         if (++ret >= nr_wake)
751                                 break;
752                 }
753         }
754
755         spin_unlock(&hb->lock);
756 out:
757         futex_unlock_mm(fshared);
758         return ret;
759 }
760
761 /*
762  * Wake up all waiters hashed on the physical page that is mapped
763  * to this virtual address:
764  */
765 static int
766 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
767               u32 __user *uaddr2,
768               int nr_wake, int nr_wake2, int op)
769 {
770         union futex_key key1, key2;
771         struct futex_hash_bucket *hb1, *hb2;
772         struct plist_head *head;
773         struct futex_q *this, *next;
774         int ret, op_ret, attempt = 0;
775
776 retryfull:
777         futex_lock_mm(fshared);
778
779         ret = get_futex_key(uaddr1, fshared, &key1);
780         if (unlikely(ret != 0))
781                 goto out;
782         ret = get_futex_key(uaddr2, fshared, &key2);
783         if (unlikely(ret != 0))
784                 goto out;
785
786         hb1 = hash_futex(&key1);
787         hb2 = hash_futex(&key2);
788
789 retry:
790         double_lock_hb(hb1, hb2);
791
792         op_ret = futex_atomic_op_inuser(op, uaddr2);
793         if (unlikely(op_ret < 0)) {
794                 u32 dummy;
795
796                 spin_unlock(&hb1->lock);
797                 if (hb1 != hb2)
798                         spin_unlock(&hb2->lock);
799
800 #ifndef CONFIG_MMU
801                 /*
802                  * we don't get EFAULT from MMU faults if we don't have an MMU,
803                  * but we might get them from range checking
804                  */
805                 ret = op_ret;
806                 goto out;
807 #endif
808
809                 if (unlikely(op_ret != -EFAULT)) {
810                         ret = op_ret;
811                         goto out;
812                 }
813
814                 /*
815                  * futex_atomic_op_inuser needs to both read and write
816                  * *(int __user *)uaddr2, but we can't modify it
817                  * non-atomically.  Therefore, if get_user below is not
818                  * enough, we need to handle the fault ourselves, while
819                  * still holding the mmap_sem.
820                  */
821                 if (attempt++) {
822                         ret = futex_handle_fault((unsigned long)uaddr2,
823                                                  fshared, attempt);
824                         if (ret)
825                                 goto out;
826                         goto retry;
827                 }
828
829                 /*
830                  * If we would have faulted, release mmap_sem,
831                  * fault it in and start all over again.
832                  */
833                 futex_unlock_mm(fshared);
834
835                 ret = get_user(dummy, uaddr2);
836                 if (ret)
837                         return ret;
838
839                 goto retryfull;
840         }
841
842         head = &hb1->chain;
843
844         plist_for_each_entry_safe(this, next, head, list) {
845                 if (match_futex (&this->key, &key1)) {
846                         wake_futex(this);
847                         if (++ret >= nr_wake)
848                                 break;
849                 }
850         }
851
852         if (op_ret > 0) {
853                 head = &hb2->chain;
854
855                 op_ret = 0;
856                 plist_for_each_entry_safe(this, next, head, list) {
857                         if (match_futex (&this->key, &key2)) {
858                                 wake_futex(this);
859                                 if (++op_ret >= nr_wake2)
860                                         break;
861                         }
862                 }
863                 ret += op_ret;
864         }
865
866         spin_unlock(&hb1->lock);
867         if (hb1 != hb2)
868                 spin_unlock(&hb2->lock);
869 out:
870         futex_unlock_mm(fshared);
871
872         return ret;
873 }
874
875 /*
876  * Requeue all waiters hashed on one physical page to another
877  * physical page.
878  */
879 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
880                          u32 __user *uaddr2,
881                          int nr_wake, int nr_requeue, u32 *cmpval)
882 {
883         union futex_key key1, key2;
884         struct futex_hash_bucket *hb1, *hb2;
885         struct plist_head *head1;
886         struct futex_q *this, *next;
887         int ret, drop_count = 0;
888
889  retry:
890         futex_lock_mm(fshared);
891
892         ret = get_futex_key(uaddr1, fshared, &key1);
893         if (unlikely(ret != 0))
894                 goto out;
895         ret = get_futex_key(uaddr2, fshared, &key2);
896         if (unlikely(ret != 0))
897                 goto out;
898
899         hb1 = hash_futex(&key1);
900         hb2 = hash_futex(&key2);
901
902         double_lock_hb(hb1, hb2);
903
904         if (likely(cmpval != NULL)) {
905                 u32 curval;
906
907                 ret = get_futex_value_locked(&curval, uaddr1);
908
909                 if (unlikely(ret)) {
910                         spin_unlock(&hb1->lock);
911                         if (hb1 != hb2)
912                                 spin_unlock(&hb2->lock);
913
914                         /*
915                          * If we would have faulted, release mmap_sem, fault
916                          * it in and start all over again.
917                          */
918                         futex_unlock_mm(fshared);
919
920                         ret = get_user(curval, uaddr1);
921
922                         if (!ret)
923                                 goto retry;
924
925                         return ret;
926                 }
927                 if (curval != *cmpval) {
928                         ret = -EAGAIN;
929                         goto out_unlock;
930                 }
931         }
932
933         head1 = &hb1->chain;
934         plist_for_each_entry_safe(this, next, head1, list) {
935                 if (!match_futex (&this->key, &key1))
936                         continue;
937                 if (++ret <= nr_wake) {
938                         wake_futex(this);
939                 } else {
940                         /*
941                          * If key1 and key2 hash to the same bucket, no need to
942                          * requeue.
943                          */
944                         if (likely(head1 != &hb2->chain)) {
945                                 plist_del(&this->list, &hb1->chain);
946                                 plist_add(&this->list, &hb2->chain);
947                                 this->lock_ptr = &hb2->lock;
948 #ifdef CONFIG_DEBUG_PI_LIST
949                                 this->list.plist.lock = &hb2->lock;
950 #endif
951                         }
952                         this->key = key2;
953                         get_futex_key_refs(&key2);
954                         drop_count++;
955
956                         if (ret - nr_wake >= nr_requeue)
957                                 break;
958                 }
959         }
960
961 out_unlock:
962         spin_unlock(&hb1->lock);
963         if (hb1 != hb2)
964                 spin_unlock(&hb2->lock);
965
966         /* drop_futex_key_refs() must be called outside the spinlocks. */
967         while (--drop_count >= 0)
968                 drop_futex_key_refs(&key1);
969
970 out:
971         futex_unlock_mm(fshared);
972         return ret;
973 }
974
975 /* The key must be already stored in q->key. */
976 static inline struct futex_hash_bucket *
977 queue_lock(struct futex_q *q, int fd, struct file *filp)
978 {
979         struct futex_hash_bucket *hb;
980
981         q->fd = fd;
982         q->filp = filp;
983
984         init_waitqueue_head(&q->waiters);
985
986         get_futex_key_refs(&q->key);
987         hb = hash_futex(&q->key);
988         q->lock_ptr = &hb->lock;
989
990         spin_lock(&hb->lock);
991         return hb;
992 }
993
994 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
995 {
996         int prio;
997
998         /*
999          * The priority used to register this element is
1000          * - either the real thread-priority for the real-time threads
1001          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1002          * - or MAX_RT_PRIO for non-RT threads.
1003          * Thus, all RT-threads are woken first in priority order, and
1004          * the others are woken last, in FIFO order.
1005          */
1006         prio = min(current->normal_prio, MAX_RT_PRIO);
1007
1008         plist_node_init(&q->list, prio);
1009 #ifdef CONFIG_DEBUG_PI_LIST
1010         q->list.plist.lock = &hb->lock;
1011 #endif
1012         plist_add(&q->list, &hb->chain);
1013         q->task = current;
1014         spin_unlock(&hb->lock);
1015 }
1016
1017 static inline void
1018 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1019 {
1020         spin_unlock(&hb->lock);
1021         drop_futex_key_refs(&q->key);
1022 }
1023
1024 /*
1025  * queue_me and unqueue_me must be called as a pair, each
1026  * exactly once.  They are called with the hashed spinlock held.
1027  */
1028
1029 /* The key must be already stored in q->key. */
1030 static void queue_me(struct futex_q *q, int fd, struct file *filp)
1031 {
1032         struct futex_hash_bucket *hb;
1033
1034         hb = queue_lock(q, fd, filp);
1035         __queue_me(q, hb);
1036 }
1037
1038 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1039 static int unqueue_me(struct futex_q *q)
1040 {
1041         spinlock_t *lock_ptr;
1042         int ret = 0;
1043
1044         /* In the common case we don't take the spinlock, which is nice. */
1045  retry:
1046         lock_ptr = q->lock_ptr;
1047         barrier();
1048         if (lock_ptr != NULL) {
1049                 spin_lock(lock_ptr);
1050                 /*
1051                  * q->lock_ptr can change between reading it and
1052                  * spin_lock(), causing us to take the wrong lock.  This
1053                  * corrects the race condition.
1054                  *
1055                  * Reasoning goes like this: if we have the wrong lock,
1056                  * q->lock_ptr must have changed (maybe several times)
1057                  * between reading it and the spin_lock().  It can
1058                  * change again after the spin_lock() but only if it was
1059                  * already changed before the spin_lock().  It cannot,
1060                  * however, change back to the original value.  Therefore
1061                  * we can detect whether we acquired the correct lock.
1062                  */
1063                 if (unlikely(lock_ptr != q->lock_ptr)) {
1064                         spin_unlock(lock_ptr);
1065                         goto retry;
1066                 }
1067                 WARN_ON(plist_node_empty(&q->list));
1068                 plist_del(&q->list, &q->list.plist);
1069
1070                 BUG_ON(q->pi_state);
1071
1072                 spin_unlock(lock_ptr);
1073                 ret = 1;
1074         }
1075
1076         drop_futex_key_refs(&q->key);
1077         return ret;
1078 }
1079
1080 /*
1081  * PI futexes can not be requeued and must remove themself from the
1082  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1083  * and dropped here.
1084  */
1085 static void unqueue_me_pi(struct futex_q *q)
1086 {
1087         WARN_ON(plist_node_empty(&q->list));
1088         plist_del(&q->list, &q->list.plist);
1089
1090         BUG_ON(!q->pi_state);
1091         free_pi_state(q->pi_state);
1092         q->pi_state = NULL;
1093
1094         spin_unlock(q->lock_ptr);
1095
1096         drop_futex_key_refs(&q->key);
1097 }
1098
1099 /*
1100  * Fixup the pi_state owner with the new owner.
1101  *
1102  * Must be called with hash bucket lock held and mm->sem held for non
1103  * private futexes.
1104  */
1105 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1106                                 struct task_struct *newowner)
1107 {
1108         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1109         struct futex_pi_state *pi_state = q->pi_state;
1110         u32 uval, curval, newval;
1111         int ret;
1112
1113         /* Owner died? */
1114         if (pi_state->owner != NULL) {
1115                 spin_lock_irq(&pi_state->owner->pi_lock);
1116                 WARN_ON(list_empty(&pi_state->list));
1117                 list_del_init(&pi_state->list);
1118                 spin_unlock_irq(&pi_state->owner->pi_lock);
1119         } else
1120                 newtid |= FUTEX_OWNER_DIED;
1121
1122         pi_state->owner = newowner;
1123
1124         spin_lock_irq(&newowner->pi_lock);
1125         WARN_ON(!list_empty(&pi_state->list));
1126         list_add(&pi_state->list, &newowner->pi_state_list);
1127         spin_unlock_irq(&newowner->pi_lock);
1128
1129         /*
1130          * We own it, so we have to replace the pending owner
1131          * TID. This must be atomic as we have preserve the
1132          * owner died bit here.
1133          */
1134         ret = get_futex_value_locked(&uval, uaddr);
1135
1136         while (!ret) {
1137                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1138
1139                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1140
1141                 if (curval == -EFAULT)
1142                         ret = -EFAULT;
1143                 if (curval == uval)
1144                         break;
1145                 uval = curval;
1146         }
1147         return ret;
1148 }
1149
1150 /*
1151  * In case we must use restart_block to restart a futex_wait,
1152  * we encode in the 'flags' shared capability
1153  */
1154 #define FLAGS_SHARED  1
1155
1156 static long futex_wait_restart(struct restart_block *restart);
1157
1158 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1159                       u32 val, ktime_t *abs_time)
1160 {
1161         struct task_struct *curr = current;
1162         DECLARE_WAITQUEUE(wait, curr);
1163         struct futex_hash_bucket *hb;
1164         struct futex_q q;
1165         u32 uval;
1166         int ret;
1167         struct hrtimer_sleeper t;
1168         int rem = 0;
1169
1170         q.pi_state = NULL;
1171  retry:
1172         futex_lock_mm(fshared);
1173
1174         ret = get_futex_key(uaddr, fshared, &q.key);
1175         if (unlikely(ret != 0))
1176                 goto out_release_sem;
1177
1178         hb = queue_lock(&q, -1, NULL);
1179
1180         /*
1181          * Access the page AFTER the futex is queued.
1182          * Order is important:
1183          *
1184          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1185          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1186          *
1187          * The basic logical guarantee of a futex is that it blocks ONLY
1188          * if cond(var) is known to be true at the time of blocking, for
1189          * any cond.  If we queued after testing *uaddr, that would open
1190          * a race condition where we could block indefinitely with
1191          * cond(var) false, which would violate the guarantee.
1192          *
1193          * A consequence is that futex_wait() can return zero and absorb
1194          * a wakeup when *uaddr != val on entry to the syscall.  This is
1195          * rare, but normal.
1196          *
1197          * for shared futexes, we hold the mmap semaphore, so the mapping
1198          * cannot have changed since we looked it up in get_futex_key.
1199          */
1200         ret = get_futex_value_locked(&uval, uaddr);
1201
1202         if (unlikely(ret)) {
1203                 queue_unlock(&q, hb);
1204
1205                 /*
1206                  * If we would have faulted, release mmap_sem, fault it in and
1207                  * start all over again.
1208                  */
1209                 futex_unlock_mm(fshared);
1210
1211                 ret = get_user(uval, uaddr);
1212
1213                 if (!ret)
1214                         goto retry;
1215                 return ret;
1216         }
1217         ret = -EWOULDBLOCK;
1218         if (uval != val)
1219                 goto out_unlock_release_sem;
1220
1221         /* Only actually queue if *uaddr contained val.  */
1222         __queue_me(&q, hb);
1223
1224         /*
1225          * Now the futex is queued and we have checked the data, we
1226          * don't want to hold mmap_sem while we sleep.
1227          */
1228         futex_unlock_mm(fshared);
1229
1230         /*
1231          * There might have been scheduling since the queue_me(), as we
1232          * cannot hold a spinlock across the get_user() in case it
1233          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1234          * queueing ourselves into the futex hash.  This code thus has to
1235          * rely on the futex_wake() code removing us from hash when it
1236          * wakes us up.
1237          */
1238
1239         /* add_wait_queue is the barrier after __set_current_state. */
1240         __set_current_state(TASK_INTERRUPTIBLE);
1241         add_wait_queue(&q.waiters, &wait);
1242         /*
1243          * !plist_node_empty() is safe here without any lock.
1244          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1245          */
1246         if (likely(!plist_node_empty(&q.list))) {
1247                 if (!abs_time)
1248                         schedule();
1249                 else {
1250                         hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1251                         hrtimer_init_sleeper(&t, current);
1252                         t.timer.expires = *abs_time;
1253
1254                         hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
1255
1256                         /*
1257                          * the timer could have already expired, in which
1258                          * case current would be flagged for rescheduling.
1259                          * Don't bother calling schedule.
1260                          */
1261                         if (likely(t.task))
1262                                 schedule();
1263
1264                         hrtimer_cancel(&t.timer);
1265
1266                         /* Flag if a timeout occured */
1267                         rem = (t.task == NULL);
1268                 }
1269         }
1270         __set_current_state(TASK_RUNNING);
1271
1272         /*
1273          * NOTE: we don't remove ourselves from the waitqueue because
1274          * we are the only user of it.
1275          */
1276
1277         /* If we were woken (and unqueued), we succeeded, whatever. */
1278         if (!unqueue_me(&q))
1279                 return 0;
1280         if (rem)
1281                 return -ETIMEDOUT;
1282
1283         /*
1284          * We expect signal_pending(current), but another thread may
1285          * have handled it for us already.
1286          */
1287         if (!abs_time)
1288                 return -ERESTARTSYS;
1289         else {
1290                 struct restart_block *restart;
1291                 restart = &current_thread_info()->restart_block;
1292                 restart->fn = futex_wait_restart;
1293                 restart->futex.uaddr = (u32 *)uaddr;
1294                 restart->futex.val = val;
1295                 restart->futex.time = abs_time->tv64;
1296                 restart->futex.flags = 0;
1297
1298                 if (fshared)
1299                         restart->futex.flags |= FLAGS_SHARED;
1300                 return -ERESTART_RESTARTBLOCK;
1301         }
1302
1303  out_unlock_release_sem:
1304         queue_unlock(&q, hb);
1305
1306  out_release_sem:
1307         futex_unlock_mm(fshared);
1308         return ret;
1309 }
1310
1311
1312 static long futex_wait_restart(struct restart_block *restart)
1313 {
1314         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1315         struct rw_semaphore *fshared = NULL;
1316         ktime_t t;
1317
1318         t.tv64 = restart->futex.time;
1319         restart->fn = do_no_restart_syscall;
1320         if (restart->futex.flags & FLAGS_SHARED)
1321                 fshared = &current->mm->mmap_sem;
1322         return (long)futex_wait(uaddr, fshared, restart->futex.val, &t);
1323 }
1324
1325
1326 /*
1327  * Userspace tried a 0 -> TID atomic transition of the futex value
1328  * and failed. The kernel side here does the whole locking operation:
1329  * if there are waiters then it will block, it does PI, etc. (Due to
1330  * races the kernel might see a 0 value of the futex too.)
1331  */
1332 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1333                          int detect, ktime_t *time, int trylock)
1334 {
1335         struct hrtimer_sleeper timeout, *to = NULL;
1336         struct task_struct *curr = current;
1337         struct futex_hash_bucket *hb;
1338         u32 uval, newval, curval;
1339         struct futex_q q;
1340         int ret, lock_taken, ownerdied = 0, attempt = 0;
1341
1342         if (refill_pi_state_cache())
1343                 return -ENOMEM;
1344
1345         if (time) {
1346                 to = &timeout;
1347                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1348                 hrtimer_init_sleeper(to, current);
1349                 to->timer.expires = *time;
1350         }
1351
1352         q.pi_state = NULL;
1353  retry:
1354         futex_lock_mm(fshared);
1355
1356         ret = get_futex_key(uaddr, fshared, &q.key);
1357         if (unlikely(ret != 0))
1358                 goto out_release_sem;
1359
1360  retry_unlocked:
1361         hb = queue_lock(&q, -1, NULL);
1362
1363  retry_locked:
1364         ret = lock_taken = 0;
1365
1366         /*
1367          * To avoid races, we attempt to take the lock here again
1368          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1369          * the locks. It will most likely not succeed.
1370          */
1371         newval = task_pid_vnr(current);
1372
1373         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
1374
1375         if (unlikely(curval == -EFAULT))
1376                 goto uaddr_faulted;
1377
1378         /*
1379          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1380          * situation and we return success to user space.
1381          */
1382         if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
1383                 ret = -EDEADLK;
1384                 goto out_unlock_release_sem;
1385         }
1386
1387         /*
1388          * Surprise - we got the lock. Just return to userspace:
1389          */
1390         if (unlikely(!curval))
1391                 goto out_unlock_release_sem;
1392
1393         uval = curval;
1394
1395         /*
1396          * Set the WAITERS flag, so the owner will know it has someone
1397          * to wake at next unlock
1398          */
1399         newval = curval | FUTEX_WAITERS;
1400
1401         /*
1402          * There are two cases, where a futex might have no owner (the
1403          * owner TID is 0): OWNER_DIED. We take over the futex in this
1404          * case. We also do an unconditional take over, when the owner
1405          * of the futex died.
1406          *
1407          * This is safe as we are protected by the hash bucket lock !
1408          */
1409         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1410                 /* Keep the OWNER_DIED bit */
1411                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
1412                 ownerdied = 0;
1413                 lock_taken = 1;
1414         }
1415
1416         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1417
1418         if (unlikely(curval == -EFAULT))
1419                 goto uaddr_faulted;
1420         if (unlikely(curval != uval))
1421                 goto retry_locked;
1422
1423         /*
1424          * We took the lock due to owner died take over.
1425          */
1426         if (unlikely(lock_taken))
1427                 goto out_unlock_release_sem;
1428
1429         /*
1430          * We dont have the lock. Look up the PI state (or create it if
1431          * we are the first waiter):
1432          */
1433         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1434
1435         if (unlikely(ret)) {
1436                 switch (ret) {
1437
1438                 case -EAGAIN:
1439                         /*
1440                          * Task is exiting and we just wait for the
1441                          * exit to complete.
1442                          */
1443                         queue_unlock(&q, hb);
1444                         futex_unlock_mm(fshared);
1445                         cond_resched();
1446                         goto retry;
1447
1448                 case -ESRCH:
1449                         /*
1450                          * No owner found for this futex. Check if the
1451                          * OWNER_DIED bit is set to figure out whether
1452                          * this is a robust futex or not.
1453                          */
1454                         if (get_futex_value_locked(&curval, uaddr))
1455                                 goto uaddr_faulted;
1456
1457                         /*
1458                          * We simply start over in case of a robust
1459                          * futex. The code above will take the futex
1460                          * and return happy.
1461                          */
1462                         if (curval & FUTEX_OWNER_DIED) {
1463                                 ownerdied = 1;
1464                                 goto retry_locked;
1465                         }
1466                 default:
1467                         goto out_unlock_release_sem;
1468                 }
1469         }
1470
1471         /*
1472          * Only actually queue now that the atomic ops are done:
1473          */
1474         __queue_me(&q, hb);
1475
1476         /*
1477          * Now the futex is queued and we have checked the data, we
1478          * don't want to hold mmap_sem while we sleep.
1479          */
1480         futex_unlock_mm(fshared);
1481
1482         WARN_ON(!q.pi_state);
1483         /*
1484          * Block on the PI mutex:
1485          */
1486         if (!trylock)
1487                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1488         else {
1489                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1490                 /* Fixup the trylock return value: */
1491                 ret = ret ? 0 : -EWOULDBLOCK;
1492         }
1493
1494         futex_lock_mm(fshared);
1495         spin_lock(q.lock_ptr);
1496
1497         if (!ret) {
1498                 /*
1499                  * Got the lock. We might not be the anticipated owner
1500                  * if we did a lock-steal - fix up the PI-state in
1501                  * that case:
1502                  */
1503                 if (q.pi_state->owner != curr)
1504                         ret = fixup_pi_state_owner(uaddr, &q, curr);
1505         } else {
1506                 /*
1507                  * Catch the rare case, where the lock was released
1508                  * when we were on the way back before we locked the
1509                  * hash bucket.
1510                  */
1511                 if (q.pi_state->owner == curr) {
1512                         /*
1513                          * Try to get the rt_mutex now. This might
1514                          * fail as some other task acquired the
1515                          * rt_mutex after we removed ourself from the
1516                          * rt_mutex waiters list.
1517                          */
1518                         if (rt_mutex_trylock(&q.pi_state->pi_mutex))
1519                                 ret = 0;
1520                         else {
1521                                 /*
1522                                  * pi_state is incorrect, some other
1523                                  * task did a lock steal and we
1524                                  * returned due to timeout or signal
1525                                  * without taking the rt_mutex. Too
1526                                  * late. We can access the
1527                                  * rt_mutex_owner without locking, as
1528                                  * the other task is now blocked on
1529                                  * the hash bucket lock. Fix the state
1530                                  * up.
1531                                  */
1532                                 struct task_struct *owner;
1533                                 int res;
1534
1535                                 owner = rt_mutex_owner(&q.pi_state->pi_mutex);
1536                                 res = fixup_pi_state_owner(uaddr, &q, owner);
1537
1538                                 WARN_ON(rt_mutex_owner(&q.pi_state->pi_mutex) !=
1539                                         owner);
1540
1541                                 /* propagate -EFAULT, if the fixup failed */
1542                                 if (res)
1543                                         ret = res;
1544                         }
1545                 } else {
1546                         /*
1547                          * Paranoia check. If we did not take the lock
1548                          * in the trylock above, then we should not be
1549                          * the owner of the rtmutex, neither the real
1550                          * nor the pending one:
1551                          */
1552                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1553                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1554                                        "pi-mutex: %p pi-state %p\n", ret,
1555                                        q.pi_state->pi_mutex.owner,
1556                                        q.pi_state->owner);
1557                 }
1558         }
1559
1560         /* Unqueue and drop the lock */
1561         unqueue_me_pi(&q);
1562         futex_unlock_mm(fshared);
1563
1564         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1565
1566  out_unlock_release_sem:
1567         queue_unlock(&q, hb);
1568
1569  out_release_sem:
1570         futex_unlock_mm(fshared);
1571         return ret;
1572
1573  uaddr_faulted:
1574         /*
1575          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1576          * non-atomically.  Therefore, if get_user below is not
1577          * enough, we need to handle the fault ourselves, while
1578          * still holding the mmap_sem.
1579          *
1580          * ... and hb->lock. :-) --ANK
1581          */
1582         queue_unlock(&q, hb);
1583
1584         if (attempt++) {
1585                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1586                                          attempt);
1587                 if (ret)
1588                         goto out_release_sem;
1589                 goto retry_unlocked;
1590         }
1591
1592         futex_unlock_mm(fshared);
1593
1594         ret = get_user(uval, uaddr);
1595         if (!ret && (uval != -EFAULT))
1596                 goto retry;
1597
1598         return ret;
1599 }
1600
1601 /*
1602  * Userspace attempted a TID -> 0 atomic transition, and failed.
1603  * This is the in-kernel slowpath: we look up the PI state (if any),
1604  * and do the rt-mutex unlock.
1605  */
1606 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1607 {
1608         struct futex_hash_bucket *hb;
1609         struct futex_q *this, *next;
1610         u32 uval;
1611         struct plist_head *head;
1612         union futex_key key;
1613         int ret, attempt = 0;
1614
1615 retry:
1616         if (get_user(uval, uaddr))
1617                 return -EFAULT;
1618         /*
1619          * We release only a lock we actually own:
1620          */
1621         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
1622                 return -EPERM;
1623         /*
1624          * First take all the futex related locks:
1625          */
1626         futex_lock_mm(fshared);
1627
1628         ret = get_futex_key(uaddr, fshared, &key);
1629         if (unlikely(ret != 0))
1630                 goto out;
1631
1632         hb = hash_futex(&key);
1633 retry_unlocked:
1634         spin_lock(&hb->lock);
1635
1636         /*
1637          * To avoid races, try to do the TID -> 0 atomic transition
1638          * again. If it succeeds then we can return without waking
1639          * anyone else up:
1640          */
1641         if (!(uval & FUTEX_OWNER_DIED))
1642                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
1643
1644
1645         if (unlikely(uval == -EFAULT))
1646                 goto pi_faulted;
1647         /*
1648          * Rare case: we managed to release the lock atomically,
1649          * no need to wake anyone else up:
1650          */
1651         if (unlikely(uval == task_pid_vnr(current)))
1652                 goto out_unlock;
1653
1654         /*
1655          * Ok, other tasks may need to be woken up - check waiters
1656          * and do the wakeup if necessary:
1657          */
1658         head = &hb->chain;
1659
1660         plist_for_each_entry_safe(this, next, head, list) {
1661                 if (!match_futex (&this->key, &key))
1662                         continue;
1663                 ret = wake_futex_pi(uaddr, uval, this);
1664                 /*
1665                  * The atomic access to the futex value
1666                  * generated a pagefault, so retry the
1667                  * user-access and the wakeup:
1668                  */
1669                 if (ret == -EFAULT)
1670                         goto pi_faulted;
1671                 goto out_unlock;
1672         }
1673         /*
1674          * No waiters - kernel unlocks the futex:
1675          */
1676         if (!(uval & FUTEX_OWNER_DIED)) {
1677                 ret = unlock_futex_pi(uaddr, uval);
1678                 if (ret == -EFAULT)
1679                         goto pi_faulted;
1680         }
1681
1682 out_unlock:
1683         spin_unlock(&hb->lock);
1684 out:
1685         futex_unlock_mm(fshared);
1686
1687         return ret;
1688
1689 pi_faulted:
1690         /*
1691          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1692          * non-atomically.  Therefore, if get_user below is not
1693          * enough, we need to handle the fault ourselves, while
1694          * still holding the mmap_sem.
1695          *
1696          * ... and hb->lock. --ANK
1697          */
1698         spin_unlock(&hb->lock);
1699
1700         if (attempt++) {
1701                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1702                                          attempt);
1703                 if (ret)
1704                         goto out;
1705                 uval = 0;
1706                 goto retry_unlocked;
1707         }
1708
1709         futex_unlock_mm(fshared);
1710
1711         ret = get_user(uval, uaddr);
1712         if (!ret && (uval != -EFAULT))
1713                 goto retry;
1714
1715         return ret;
1716 }
1717
1718 static int futex_close(struct inode *inode, struct file *filp)
1719 {
1720         struct futex_q *q = filp->private_data;
1721
1722         unqueue_me(q);
1723         kfree(q);
1724
1725         return 0;
1726 }
1727
1728 /* This is one-shot: once it's gone off you need a new fd */
1729 static unsigned int futex_poll(struct file *filp,
1730                                struct poll_table_struct *wait)
1731 {
1732         struct futex_q *q = filp->private_data;
1733         int ret = 0;
1734
1735         poll_wait(filp, &q->waiters, wait);
1736
1737         /*
1738          * plist_node_empty() is safe here without any lock.
1739          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1740          */
1741         if (plist_node_empty(&q->list))
1742                 ret = POLLIN | POLLRDNORM;
1743
1744         return ret;
1745 }
1746
1747 static const struct file_operations futex_fops = {
1748         .release        = futex_close,
1749         .poll           = futex_poll,
1750 };
1751
1752 /*
1753  * Signal allows caller to avoid the race which would occur if they
1754  * set the sigio stuff up afterwards.
1755  */
1756 static int futex_fd(u32 __user *uaddr, int signal)
1757 {
1758         struct futex_q *q;
1759         struct file *filp;
1760         int ret, err;
1761         struct rw_semaphore *fshared;
1762         static unsigned long printk_interval;
1763
1764         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1765                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1766                        "will be removed from the kernel in June 2007\n",
1767                        current->comm);
1768         }
1769
1770         ret = -EINVAL;
1771         if (!valid_signal(signal))
1772                 goto out;
1773
1774         ret = get_unused_fd();
1775         if (ret < 0)
1776                 goto out;
1777         filp = get_empty_filp();
1778         if (!filp) {
1779                 put_unused_fd(ret);
1780                 ret = -ENFILE;
1781                 goto out;
1782         }
1783         filp->f_op = &futex_fops;
1784         filp->f_path.mnt = mntget(futex_mnt);
1785         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1786         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1787
1788         if (signal) {
1789                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1790                 if (err < 0) {
1791                         goto error;
1792                 }
1793                 filp->f_owner.signum = signal;
1794         }
1795
1796         q = kmalloc(sizeof(*q), GFP_KERNEL);
1797         if (!q) {
1798                 err = -ENOMEM;
1799                 goto error;
1800         }
1801         q->pi_state = NULL;
1802
1803         fshared = &current->mm->mmap_sem;
1804         down_read(fshared);
1805         err = get_futex_key(uaddr, fshared, &q->key);
1806
1807         if (unlikely(err != 0)) {
1808                 up_read(fshared);
1809                 kfree(q);
1810                 goto error;
1811         }
1812
1813         /*
1814          * queue_me() must be called before releasing mmap_sem, because
1815          * key->shared.inode needs to be referenced while holding it.
1816          */
1817         filp->private_data = q;
1818
1819         queue_me(q, ret, filp);
1820         up_read(fshared);
1821
1822         /* Now we map fd to filp, so userspace can access it */
1823         fd_install(ret, filp);
1824 out:
1825         return ret;
1826 error:
1827         put_unused_fd(ret);
1828         put_filp(filp);
1829         ret = err;
1830         goto out;
1831 }
1832
1833 /*
1834  * Support for robust futexes: the kernel cleans up held futexes at
1835  * thread exit time.
1836  *
1837  * Implementation: user-space maintains a per-thread list of locks it
1838  * is holding. Upon do_exit(), the kernel carefully walks this list,
1839  * and marks all locks that are owned by this thread with the
1840  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1841  * always manipulated with the lock held, so the list is private and
1842  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1843  * field, to allow the kernel to clean up if the thread dies after
1844  * acquiring the lock, but just before it could have added itself to
1845  * the list. There can only be one such pending lock.
1846  */
1847
1848 /**
1849  * sys_set_robust_list - set the robust-futex list head of a task
1850  * @head: pointer to the list-head
1851  * @len: length of the list-head, as userspace expects
1852  */
1853 asmlinkage long
1854 sys_set_robust_list(struct robust_list_head __user *head,
1855                     size_t len)
1856 {
1857         /*
1858          * The kernel knows only one size for now:
1859          */
1860         if (unlikely(len != sizeof(*head)))
1861                 return -EINVAL;
1862
1863         current->robust_list = head;
1864
1865         return 0;
1866 }
1867
1868 /**
1869  * sys_get_robust_list - get the robust-futex list head of a task
1870  * @pid: pid of the process [zero for current task]
1871  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1872  * @len_ptr: pointer to a length field, the kernel fills in the header size
1873  */
1874 asmlinkage long
1875 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1876                     size_t __user *len_ptr)
1877 {
1878         struct robust_list_head __user *head;
1879         unsigned long ret;
1880
1881         if (!pid)
1882                 head = current->robust_list;
1883         else {
1884                 struct task_struct *p;
1885
1886                 ret = -ESRCH;
1887                 rcu_read_lock();
1888                 p = find_task_by_vpid(pid);
1889                 if (!p)
1890                         goto err_unlock;
1891                 ret = -EPERM;
1892                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1893                                 !capable(CAP_SYS_PTRACE))
1894                         goto err_unlock;
1895                 head = p->robust_list;
1896                 rcu_read_unlock();
1897         }
1898
1899         if (put_user(sizeof(*head), len_ptr))
1900                 return -EFAULT;
1901         return put_user(head, head_ptr);
1902
1903 err_unlock:
1904         rcu_read_unlock();
1905
1906         return ret;
1907 }
1908
1909 /*
1910  * Process a futex-list entry, check whether it's owned by the
1911  * dying task, and do notification if so:
1912  */
1913 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1914 {
1915         u32 uval, nval, mval;
1916
1917 retry:
1918         if (get_user(uval, uaddr))
1919                 return -1;
1920
1921         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
1922                 /*
1923                  * Ok, this dying thread is truly holding a futex
1924                  * of interest. Set the OWNER_DIED bit atomically
1925                  * via cmpxchg, and if the value had FUTEX_WAITERS
1926                  * set, wake up a waiter (if any). (We have to do a
1927                  * futex_wake() even if OWNER_DIED is already set -
1928                  * to handle the rare but possible case of recursive
1929                  * thread-death.) The rest of the cleanup is done in
1930                  * userspace.
1931                  */
1932                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1933                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1934
1935                 if (nval == -EFAULT)
1936                         return -1;
1937
1938                 if (nval != uval)
1939                         goto retry;
1940
1941                 /*
1942                  * Wake robust non-PI futexes here. The wakeup of
1943                  * PI futexes happens in exit_pi_state():
1944                  */
1945                 if (!pi && (uval & FUTEX_WAITERS))
1946                                 futex_wake(uaddr, &curr->mm->mmap_sem, 1);
1947         }
1948         return 0;
1949 }
1950
1951 /*
1952  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1953  */
1954 static inline int fetch_robust_entry(struct robust_list __user **entry,
1955                                      struct robust_list __user * __user *head,
1956                                      int *pi)
1957 {
1958         unsigned long uentry;
1959
1960         if (get_user(uentry, (unsigned long __user *)head))
1961                 return -EFAULT;
1962
1963         *entry = (void __user *)(uentry & ~1UL);
1964         *pi = uentry & 1;
1965
1966         return 0;
1967 }
1968
1969 /*
1970  * Walk curr->robust_list (very carefully, it's a userspace list!)
1971  * and mark any locks found there dead, and notify any waiters.
1972  *
1973  * We silently return on any sign of list-walking problem.
1974  */
1975 void exit_robust_list(struct task_struct *curr)
1976 {
1977         struct robust_list_head __user *head = curr->robust_list;
1978         struct robust_list __user *entry, *next_entry, *pending;
1979         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
1980         unsigned long futex_offset;
1981         int rc;
1982
1983         /*
1984          * Fetch the list head (which was registered earlier, via
1985          * sys_set_robust_list()):
1986          */
1987         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1988                 return;
1989         /*
1990          * Fetch the relative futex offset:
1991          */
1992         if (get_user(futex_offset, &head->futex_offset))
1993                 return;
1994         /*
1995          * Fetch any possibly pending lock-add first, and handle it
1996          * if it exists:
1997          */
1998         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1999                 return;
2000
2001         next_entry = NULL;      /* avoid warning with gcc */
2002         while (entry != &head->list) {
2003                 /*
2004                  * Fetch the next entry in the list before calling
2005                  * handle_futex_death:
2006                  */
2007                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2008                 /*
2009                  * A pending lock might already be on the list, so
2010                  * don't process it twice:
2011                  */
2012                 if (entry != pending)
2013                         if (handle_futex_death((void __user *)entry + futex_offset,
2014                                                 curr, pi))
2015                                 return;
2016                 if (rc)
2017                         return;
2018                 entry = next_entry;
2019                 pi = next_pi;
2020                 /*
2021                  * Avoid excessively long or circular lists:
2022                  */
2023                 if (!--limit)
2024                         break;
2025
2026                 cond_resched();
2027         }
2028
2029         if (pending)
2030                 handle_futex_death((void __user *)pending + futex_offset,
2031                                    curr, pip);
2032 }
2033
2034 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2035                 u32 __user *uaddr2, u32 val2, u32 val3)
2036 {
2037         int ret;
2038         int cmd = op & FUTEX_CMD_MASK;
2039         struct rw_semaphore *fshared = NULL;
2040
2041         if (!(op & FUTEX_PRIVATE_FLAG))
2042                 fshared = &current->mm->mmap_sem;
2043
2044         switch (cmd) {
2045         case FUTEX_WAIT:
2046                 ret = futex_wait(uaddr, fshared, val, timeout);
2047                 break;
2048         case FUTEX_WAKE:
2049                 ret = futex_wake(uaddr, fshared, val);
2050                 break;
2051         case FUTEX_FD:
2052                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
2053                 ret = futex_fd(uaddr, val);
2054                 break;
2055         case FUTEX_REQUEUE:
2056                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2057                 break;
2058         case FUTEX_CMP_REQUEUE:
2059                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2060                 break;
2061         case FUTEX_WAKE_OP:
2062                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2063                 break;
2064         case FUTEX_LOCK_PI:
2065                 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2066                 break;
2067         case FUTEX_UNLOCK_PI:
2068                 ret = futex_unlock_pi(uaddr, fshared);
2069                 break;
2070         case FUTEX_TRYLOCK_PI:
2071                 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2072                 break;
2073         default:
2074                 ret = -ENOSYS;
2075         }
2076         return ret;
2077 }
2078
2079
2080 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2081                           struct timespec __user *utime, u32 __user *uaddr2,
2082                           u32 val3)
2083 {
2084         struct timespec ts;
2085         ktime_t t, *tp = NULL;
2086         u32 val2 = 0;
2087         int cmd = op & FUTEX_CMD_MASK;
2088
2089         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
2090                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2091                         return -EFAULT;
2092                 if (!timespec_valid(&ts))
2093                         return -EINVAL;
2094
2095                 t = timespec_to_ktime(ts);
2096                 if (cmd == FUTEX_WAIT)
2097                         t = ktime_add(ktime_get(), t);
2098                 tp = &t;
2099         }
2100         /*
2101          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2102          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2103          */
2104         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2105             cmd == FUTEX_WAKE_OP)
2106                 val2 = (u32) (unsigned long) utime;
2107
2108         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2109 }
2110
2111 static int futexfs_get_sb(struct file_system_type *fs_type,
2112                           int flags, const char *dev_name, void *data,
2113                           struct vfsmount *mnt)
2114 {
2115         return get_sb_pseudo(fs_type, "futex", NULL, FUTEXFS_SUPER_MAGIC, mnt);
2116 }
2117
2118 static struct file_system_type futex_fs_type = {
2119         .name           = "futexfs",
2120         .get_sb         = futexfs_get_sb,
2121         .kill_sb        = kill_anon_super,
2122 };
2123
2124 static int __init init(void)
2125 {
2126         int i = register_filesystem(&futex_fs_type);
2127
2128         if (i)
2129                 return i;
2130
2131         futex_mnt = kern_mount(&futex_fs_type);
2132         if (IS_ERR(futex_mnt)) {
2133                 unregister_filesystem(&futex_fs_type);
2134                 return PTR_ERR(futex_mnt);
2135         }
2136
2137         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2138                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2139                 spin_lock_init(&futex_queues[i].lock);
2140         }
2141         return 0;
2142 }
2143 __initcall(init);