sched: rename idle_type/SCHED_IDLE
[linux-2.6] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  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 <asm/futex.h>
56
57 #include "rtmutex_common.h"
58
59 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
60
61 /*
62  * Priority Inheritance state:
63  */
64 struct futex_pi_state {
65         /*
66          * list of 'owned' pi_state instances - these have to be
67          * cleaned up in do_exit() if the task exits prematurely:
68          */
69         struct list_head list;
70
71         /*
72          * The PI object:
73          */
74         struct rt_mutex pi_mutex;
75
76         struct task_struct *owner;
77         atomic_t refcount;
78
79         union futex_key key;
80 };
81
82 /*
83  * We use this hashed waitqueue instead of a normal wait_queue_t, so
84  * we can wake only the relevant ones (hashed queues may be shared).
85  *
86  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
87  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
88  * The order of wakup is always to make the first condition true, then
89  * wake up q->waiters, then make the second condition true.
90  */
91 struct futex_q {
92         struct plist_node list;
93         wait_queue_head_t waiters;
94
95         /* Which hash list lock to use: */
96         spinlock_t *lock_ptr;
97
98         /* Key which the futex is hashed on: */
99         union futex_key key;
100
101         /* For fd, sigio sent using these: */
102         int fd;
103         struct file *filp;
104
105         /* Optional priority inheritance state: */
106         struct futex_pi_state *pi_state;
107         struct task_struct *task;
108 };
109
110 /*
111  * Split the global futex_lock into every hash list lock.
112  */
113 struct futex_hash_bucket {
114         spinlock_t lock;
115         struct plist_head chain;
116 };
117
118 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
119
120 /* Futex-fs vfsmount entry: */
121 static struct vfsmount *futex_mnt;
122
123 /*
124  * We hash on the keys returned from get_futex_key (see below).
125  */
126 static struct futex_hash_bucket *hash_futex(union futex_key *key)
127 {
128         u32 hash = jhash2((u32*)&key->both.word,
129                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
130                           key->both.offset);
131         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
132 }
133
134 /*
135  * Return 1 if two futex_keys are equal, 0 otherwise.
136  */
137 static inline int match_futex(union futex_key *key1, union futex_key *key2)
138 {
139         return (key1->both.word == key2->both.word
140                 && key1->both.ptr == key2->both.ptr
141                 && key1->both.offset == key2->both.offset);
142 }
143
144 /**
145  * get_futex_key - Get parameters which are the keys for a futex.
146  * @uaddr: virtual address of the futex
147  * @shared: NULL for a PROCESS_PRIVATE futex,
148  *      &current->mm->mmap_sem for a PROCESS_SHARED futex
149  * @key: address where result is stored.
150  *
151  * Returns a negative error code or 0
152  * The key words are stored in *key on success.
153  *
154  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
155  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
156  * We can usually work out the index without swapping in the page.
157  *
158  * fshared is NULL for PROCESS_PRIVATE futexes
159  * For other futexes, it points to &current->mm->mmap_sem and
160  * caller must have taken the reader lock. but NOT any spinlocks.
161  */
162 int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
163                   union futex_key *key)
164 {
165         unsigned long address = (unsigned long)uaddr;
166         struct mm_struct *mm = current->mm;
167         struct vm_area_struct *vma;
168         struct page *page;
169         int err;
170
171         /*
172          * The futex address must be "naturally" aligned.
173          */
174         key->both.offset = address % PAGE_SIZE;
175         if (unlikely((address % sizeof(u32)) != 0))
176                 return -EINVAL;
177         address -= key->both.offset;
178
179         /*
180          * PROCESS_PRIVATE futexes are fast.
181          * As the mm cannot disappear under us and the 'key' only needs
182          * virtual address, we dont even have to find the underlying vma.
183          * Note : We do have to check 'uaddr' is a valid user address,
184          *        but access_ok() should be faster than find_vma()
185          */
186         if (!fshared) {
187                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
188                         return -EFAULT;
189                 key->private.mm = mm;
190                 key->private.address = address;
191                 return 0;
192         }
193         /*
194          * The futex is hashed differently depending on whether
195          * it's in a shared or private mapping.  So check vma first.
196          */
197         vma = find_extend_vma(mm, address);
198         if (unlikely(!vma))
199                 return -EFAULT;
200
201         /*
202          * Permissions.
203          */
204         if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ))
205                 return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES;
206
207         /*
208          * Private mappings are handled in a simple way.
209          *
210          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
211          * it's a read-only handle, it's expected that futexes attach to
212          * the object not the particular process.  Therefore we use
213          * VM_MAYSHARE here, not VM_SHARED which is restricted to shared
214          * mappings of _writable_ handles.
215          */
216         if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
217                 key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
218                 key->private.mm = mm;
219                 key->private.address = address;
220                 return 0;
221         }
222
223         /*
224          * Linear file mappings are also simple.
225          */
226         key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
227         key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
228         if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
229                 key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
230                                      + vma->vm_pgoff);
231                 return 0;
232         }
233
234         /*
235          * We could walk the page table to read the non-linear
236          * pte, and get the page index without fetching the page
237          * from swap.  But that's a lot of code to duplicate here
238          * for a rare case, so we simply fetch the page.
239          */
240         err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL);
241         if (err >= 0) {
242                 key->shared.pgoff =
243                         page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
244                 put_page(page);
245                 return 0;
246         }
247         return err;
248 }
249 EXPORT_SYMBOL_GPL(get_futex_key);
250
251 /*
252  * Take a reference to the resource addressed by a key.
253  * Can be called while holding spinlocks.
254  *
255  */
256 inline void get_futex_key_refs(union futex_key *key)
257 {
258         if (key->both.ptr == 0)
259                 return;
260         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
261                 case FUT_OFF_INODE:
262                         atomic_inc(&key->shared.inode->i_count);
263                         break;
264                 case FUT_OFF_MMSHARED:
265                         atomic_inc(&key->private.mm->mm_count);
266                         break;
267         }
268 }
269 EXPORT_SYMBOL_GPL(get_futex_key_refs);
270
271 /*
272  * Drop a reference to the resource addressed by a key.
273  * The hash bucket spinlock must not be held.
274  */
275 void drop_futex_key_refs(union futex_key *key)
276 {
277         if (key->both.ptr == 0)
278                 return;
279         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
280                 case FUT_OFF_INODE:
281                         iput(key->shared.inode);
282                         break;
283                 case FUT_OFF_MMSHARED:
284                         mmdrop(key->private.mm);
285                         break;
286         }
287 }
288 EXPORT_SYMBOL_GPL(drop_futex_key_refs);
289
290 static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
291 {
292         int ret;
293
294         pagefault_disable();
295         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
296         pagefault_enable();
297
298         return ret ? -EFAULT : 0;
299 }
300
301 /*
302  * Fault handling.
303  * if fshared is non NULL, current->mm->mmap_sem is already held
304  */
305 static int futex_handle_fault(unsigned long address,
306                               struct rw_semaphore *fshared, int attempt)
307 {
308         struct vm_area_struct * vma;
309         struct mm_struct *mm = current->mm;
310         int ret = -EFAULT;
311
312         if (attempt > 2)
313                 return ret;
314
315         if (!fshared)
316                 down_read(&mm->mmap_sem);
317         vma = find_vma(mm, address);
318         if (vma && address >= vma->vm_start &&
319             (vma->vm_flags & VM_WRITE)) {
320                 switch (handle_mm_fault(mm, vma, address, 1)) {
321                 case VM_FAULT_MINOR:
322                         ret = 0;
323                         current->min_flt++;
324                         break;
325                 case VM_FAULT_MAJOR:
326                         ret = 0;
327                         current->maj_flt++;
328                         break;
329                 }
330         }
331         if (!fshared)
332                 up_read(&mm->mmap_sem);
333         return ret;
334 }
335
336 /*
337  * PI code:
338  */
339 static int refill_pi_state_cache(void)
340 {
341         struct futex_pi_state *pi_state;
342
343         if (likely(current->pi_state_cache))
344                 return 0;
345
346         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
347
348         if (!pi_state)
349                 return -ENOMEM;
350
351         INIT_LIST_HEAD(&pi_state->list);
352         /* pi_mutex gets initialized later */
353         pi_state->owner = NULL;
354         atomic_set(&pi_state->refcount, 1);
355
356         current->pi_state_cache = pi_state;
357
358         return 0;
359 }
360
361 static struct futex_pi_state * alloc_pi_state(void)
362 {
363         struct futex_pi_state *pi_state = current->pi_state_cache;
364
365         WARN_ON(!pi_state);
366         current->pi_state_cache = NULL;
367
368         return pi_state;
369 }
370
371 static void free_pi_state(struct futex_pi_state *pi_state)
372 {
373         if (!atomic_dec_and_test(&pi_state->refcount))
374                 return;
375
376         /*
377          * If pi_state->owner is NULL, the owner is most probably dying
378          * and has cleaned up the pi_state already
379          */
380         if (pi_state->owner) {
381                 spin_lock_irq(&pi_state->owner->pi_lock);
382                 list_del_init(&pi_state->list);
383                 spin_unlock_irq(&pi_state->owner->pi_lock);
384
385                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
386         }
387
388         if (current->pi_state_cache)
389                 kfree(pi_state);
390         else {
391                 /*
392                  * pi_state->list is already empty.
393                  * clear pi_state->owner.
394                  * refcount is at 0 - put it back to 1.
395                  */
396                 pi_state->owner = NULL;
397                 atomic_set(&pi_state->refcount, 1);
398                 current->pi_state_cache = pi_state;
399         }
400 }
401
402 /*
403  * Look up the task based on what TID userspace gave us.
404  * We dont trust it.
405  */
406 static struct task_struct * futex_find_get_task(pid_t pid)
407 {
408         struct task_struct *p;
409
410         rcu_read_lock();
411         p = find_task_by_pid(pid);
412
413         if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
414                 p = ERR_PTR(-ESRCH);
415         else
416                 get_task_struct(p);
417
418         rcu_read_unlock();
419
420         return p;
421 }
422
423 /*
424  * This task is holding PI mutexes at exit time => bad.
425  * Kernel cleans up PI-state, but userspace is likely hosed.
426  * (Robust-futex cleanup is separate and might save the day for userspace.)
427  */
428 void exit_pi_state_list(struct task_struct *curr)
429 {
430         struct list_head *next, *head = &curr->pi_state_list;
431         struct futex_pi_state *pi_state;
432         struct futex_hash_bucket *hb;
433         union futex_key key;
434
435         /*
436          * We are a ZOMBIE and nobody can enqueue itself on
437          * pi_state_list anymore, but we have to be careful
438          * versus waiters unqueueing themselves:
439          */
440         spin_lock_irq(&curr->pi_lock);
441         while (!list_empty(head)) {
442
443                 next = head->next;
444                 pi_state = list_entry(next, struct futex_pi_state, list);
445                 key = pi_state->key;
446                 hb = hash_futex(&key);
447                 spin_unlock_irq(&curr->pi_lock);
448
449                 spin_lock(&hb->lock);
450
451                 spin_lock_irq(&curr->pi_lock);
452                 /*
453                  * We dropped the pi-lock, so re-check whether this
454                  * task still owns the PI-state:
455                  */
456                 if (head->next != next) {
457                         spin_unlock(&hb->lock);
458                         continue;
459                 }
460
461                 WARN_ON(pi_state->owner != curr);
462                 WARN_ON(list_empty(&pi_state->list));
463                 list_del_init(&pi_state->list);
464                 pi_state->owner = NULL;
465                 spin_unlock_irq(&curr->pi_lock);
466
467                 rt_mutex_unlock(&pi_state->pi_mutex);
468
469                 spin_unlock(&hb->lock);
470
471                 spin_lock_irq(&curr->pi_lock);
472         }
473         spin_unlock_irq(&curr->pi_lock);
474 }
475
476 static int
477 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
478                 union futex_key *key, struct futex_pi_state **ps)
479 {
480         struct futex_pi_state *pi_state = NULL;
481         struct futex_q *this, *next;
482         struct plist_head *head;
483         struct task_struct *p;
484         pid_t pid = uval & FUTEX_TID_MASK;
485
486         head = &hb->chain;
487
488         plist_for_each_entry_safe(this, next, head, list) {
489                 if (match_futex(&this->key, key)) {
490                         /*
491                          * Another waiter already exists - bump up
492                          * the refcount and return its pi_state:
493                          */
494                         pi_state = this->pi_state;
495                         /*
496                          * Userspace might have messed up non PI and PI futexes
497                          */
498                         if (unlikely(!pi_state))
499                                 return -EINVAL;
500
501                         WARN_ON(!atomic_read(&pi_state->refcount));
502                         WARN_ON(pid && pi_state->owner &&
503                                 pi_state->owner->pid != pid);
504
505                         atomic_inc(&pi_state->refcount);
506                         *ps = pi_state;
507
508                         return 0;
509                 }
510         }
511
512         /*
513          * We are the first waiter - try to look up the real owner and attach
514          * the new pi_state to it, but bail out when TID = 0
515          */
516         if (!pid)
517                 return -ESRCH;
518         p = futex_find_get_task(pid);
519         if (IS_ERR(p))
520                 return PTR_ERR(p);
521
522         /*
523          * We need to look at the task state flags to figure out,
524          * whether the task is exiting. To protect against the do_exit
525          * change of the task flags, we do this protected by
526          * p->pi_lock:
527          */
528         spin_lock_irq(&p->pi_lock);
529         if (unlikely(p->flags & PF_EXITING)) {
530                 /*
531                  * The task is on the way out. When PF_EXITPIDONE is
532                  * set, we know that the task has finished the
533                  * cleanup:
534                  */
535                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
536
537                 spin_unlock_irq(&p->pi_lock);
538                 put_task_struct(p);
539                 return ret;
540         }
541
542         pi_state = alloc_pi_state();
543
544         /*
545          * Initialize the pi_mutex in locked state and make 'p'
546          * the owner of it:
547          */
548         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
549
550         /* Store the key for possible exit cleanups: */
551         pi_state->key = *key;
552
553         WARN_ON(!list_empty(&pi_state->list));
554         list_add(&pi_state->list, &p->pi_state_list);
555         pi_state->owner = p;
556         spin_unlock_irq(&p->pi_lock);
557
558         put_task_struct(p);
559
560         *ps = pi_state;
561
562         return 0;
563 }
564
565 /*
566  * The hash bucket lock must be held when this is called.
567  * Afterwards, the futex_q must not be accessed.
568  */
569 static void wake_futex(struct futex_q *q)
570 {
571         plist_del(&q->list, &q->list.plist);
572         if (q->filp)
573                 send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
574         /*
575          * The lock in wake_up_all() is a crucial memory barrier after the
576          * plist_del() and also before assigning to q->lock_ptr.
577          */
578         wake_up_all(&q->waiters);
579         /*
580          * The waiting task can free the futex_q as soon as this is written,
581          * without taking any locks.  This must come last.
582          *
583          * A memory barrier is required here to prevent the following store
584          * to lock_ptr from getting ahead of the wakeup. Clearing the lock
585          * at the end of wake_up_all() does not prevent this store from
586          * moving.
587          */
588         smp_wmb();
589         q->lock_ptr = NULL;
590 }
591
592 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
593 {
594         struct task_struct *new_owner;
595         struct futex_pi_state *pi_state = this->pi_state;
596         u32 curval, newval;
597
598         if (!pi_state)
599                 return -EINVAL;
600
601         spin_lock(&pi_state->pi_mutex.wait_lock);
602         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
603
604         /*
605          * This happens when we have stolen the lock and the original
606          * pending owner did not enqueue itself back on the rt_mutex.
607          * Thats not a tragedy. We know that way, that a lock waiter
608          * is on the fly. We make the futex_q waiter the pending owner.
609          */
610         if (!new_owner)
611                 new_owner = this->task;
612
613         /*
614          * We pass it to the next owner. (The WAITERS bit is always
615          * kept enabled while there is PI state around. We must also
616          * preserve the owner died bit.)
617          */
618         if (!(uval & FUTEX_OWNER_DIED)) {
619                 int ret = 0;
620
621                 newval = FUTEX_WAITERS | new_owner->pid;
622
623                 pagefault_disable();
624                 curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
625                 pagefault_enable();
626
627                 if (curval == -EFAULT)
628                         ret = -EFAULT;
629                 if (curval != uval)
630                         ret = -EINVAL;
631                 if (ret) {
632                         spin_unlock(&pi_state->pi_mutex.wait_lock);
633                         return ret;
634                 }
635         }
636
637         spin_lock_irq(&pi_state->owner->pi_lock);
638         WARN_ON(list_empty(&pi_state->list));
639         list_del_init(&pi_state->list);
640         spin_unlock_irq(&pi_state->owner->pi_lock);
641
642         spin_lock_irq(&new_owner->pi_lock);
643         WARN_ON(!list_empty(&pi_state->list));
644         list_add(&pi_state->list, &new_owner->pi_state_list);
645         pi_state->owner = new_owner;
646         spin_unlock_irq(&new_owner->pi_lock);
647
648         spin_unlock(&pi_state->pi_mutex.wait_lock);
649         rt_mutex_unlock(&pi_state->pi_mutex);
650
651         return 0;
652 }
653
654 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
655 {
656         u32 oldval;
657
658         /*
659          * There is no waiter, so we unlock the futex. The owner died
660          * bit has not to be preserved here. We are the owner:
661          */
662         pagefault_disable();
663         oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
664         pagefault_enable();
665
666         if (oldval == -EFAULT)
667                 return oldval;
668         if (oldval != uval)
669                 return -EAGAIN;
670
671         return 0;
672 }
673
674 /*
675  * Express the locking dependencies for lockdep:
676  */
677 static inline void
678 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
679 {
680         if (hb1 <= hb2) {
681                 spin_lock(&hb1->lock);
682                 if (hb1 < hb2)
683                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
684         } else { /* hb1 > hb2 */
685                 spin_lock(&hb2->lock);
686                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
687         }
688 }
689
690 /*
691  * Wake up all waiters hashed on the physical page that is mapped
692  * to this virtual address:
693  */
694 static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
695                       int nr_wake)
696 {
697         struct futex_hash_bucket *hb;
698         struct futex_q *this, *next;
699         struct plist_head *head;
700         union futex_key key;
701         int ret;
702
703         if (fshared)
704                 down_read(fshared);
705
706         ret = get_futex_key(uaddr, fshared, &key);
707         if (unlikely(ret != 0))
708                 goto out;
709
710         hb = hash_futex(&key);
711         spin_lock(&hb->lock);
712         head = &hb->chain;
713
714         plist_for_each_entry_safe(this, next, head, list) {
715                 if (match_futex (&this->key, &key)) {
716                         if (this->pi_state) {
717                                 ret = -EINVAL;
718                                 break;
719                         }
720                         wake_futex(this);
721                         if (++ret >= nr_wake)
722                                 break;
723                 }
724         }
725
726         spin_unlock(&hb->lock);
727 out:
728         if (fshared)
729                 up_read(fshared);
730         return ret;
731 }
732
733 /*
734  * Wake up all waiters hashed on the physical page that is mapped
735  * to this virtual address:
736  */
737 static int
738 futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
739               u32 __user *uaddr2,
740               int nr_wake, int nr_wake2, int op)
741 {
742         union futex_key key1, key2;
743         struct futex_hash_bucket *hb1, *hb2;
744         struct plist_head *head;
745         struct futex_q *this, *next;
746         int ret, op_ret, attempt = 0;
747
748 retryfull:
749         if (fshared)
750                 down_read(fshared);
751
752         ret = get_futex_key(uaddr1, fshared, &key1);
753         if (unlikely(ret != 0))
754                 goto out;
755         ret = get_futex_key(uaddr2, fshared, &key2);
756         if (unlikely(ret != 0))
757                 goto out;
758
759         hb1 = hash_futex(&key1);
760         hb2 = hash_futex(&key2);
761
762 retry:
763         double_lock_hb(hb1, hb2);
764
765         op_ret = futex_atomic_op_inuser(op, uaddr2);
766         if (unlikely(op_ret < 0)) {
767                 u32 dummy;
768
769                 spin_unlock(&hb1->lock);
770                 if (hb1 != hb2)
771                         spin_unlock(&hb2->lock);
772
773 #ifndef CONFIG_MMU
774                 /*
775                  * we don't get EFAULT from MMU faults if we don't have an MMU,
776                  * but we might get them from range checking
777                  */
778                 ret = op_ret;
779                 goto out;
780 #endif
781
782                 if (unlikely(op_ret != -EFAULT)) {
783                         ret = op_ret;
784                         goto out;
785                 }
786
787                 /*
788                  * futex_atomic_op_inuser needs to both read and write
789                  * *(int __user *)uaddr2, but we can't modify it
790                  * non-atomically.  Therefore, if get_user below is not
791                  * enough, we need to handle the fault ourselves, while
792                  * still holding the mmap_sem.
793                  */
794                 if (attempt++) {
795                         ret = futex_handle_fault((unsigned long)uaddr2,
796                                                 fshared, attempt);
797                         if (ret)
798                                 goto out;
799                         goto retry;
800                 }
801
802                 /*
803                  * If we would have faulted, release mmap_sem,
804                  * fault it in and start all over again.
805                  */
806                 if (fshared)
807                         up_read(fshared);
808
809                 ret = get_user(dummy, uaddr2);
810                 if (ret)
811                         return ret;
812
813                 goto retryfull;
814         }
815
816         head = &hb1->chain;
817
818         plist_for_each_entry_safe(this, next, head, list) {
819                 if (match_futex (&this->key, &key1)) {
820                         wake_futex(this);
821                         if (++ret >= nr_wake)
822                                 break;
823                 }
824         }
825
826         if (op_ret > 0) {
827                 head = &hb2->chain;
828
829                 op_ret = 0;
830                 plist_for_each_entry_safe(this, next, head, list) {
831                         if (match_futex (&this->key, &key2)) {
832                                 wake_futex(this);
833                                 if (++op_ret >= nr_wake2)
834                                         break;
835                         }
836                 }
837                 ret += op_ret;
838         }
839
840         spin_unlock(&hb1->lock);
841         if (hb1 != hb2)
842                 spin_unlock(&hb2->lock);
843 out:
844         if (fshared)
845                 up_read(fshared);
846         return ret;
847 }
848
849 /*
850  * Requeue all waiters hashed on one physical page to another
851  * physical page.
852  */
853 static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
854                          u32 __user *uaddr2,
855                          int nr_wake, int nr_requeue, u32 *cmpval)
856 {
857         union futex_key key1, key2;
858         struct futex_hash_bucket *hb1, *hb2;
859         struct plist_head *head1;
860         struct futex_q *this, *next;
861         int ret, drop_count = 0;
862
863  retry:
864         if (fshared)
865                 down_read(fshared);
866
867         ret = get_futex_key(uaddr1, fshared, &key1);
868         if (unlikely(ret != 0))
869                 goto out;
870         ret = get_futex_key(uaddr2, fshared, &key2);
871         if (unlikely(ret != 0))
872                 goto out;
873
874         hb1 = hash_futex(&key1);
875         hb2 = hash_futex(&key2);
876
877         double_lock_hb(hb1, hb2);
878
879         if (likely(cmpval != NULL)) {
880                 u32 curval;
881
882                 ret = get_futex_value_locked(&curval, uaddr1);
883
884                 if (unlikely(ret)) {
885                         spin_unlock(&hb1->lock);
886                         if (hb1 != hb2)
887                                 spin_unlock(&hb2->lock);
888
889                         /*
890                          * If we would have faulted, release mmap_sem, fault
891                          * it in and start all over again.
892                          */
893                         if (fshared)
894                                 up_read(fshared);
895
896                         ret = get_user(curval, uaddr1);
897
898                         if (!ret)
899                                 goto retry;
900
901                         return ret;
902                 }
903                 if (curval != *cmpval) {
904                         ret = -EAGAIN;
905                         goto out_unlock;
906                 }
907         }
908
909         head1 = &hb1->chain;
910         plist_for_each_entry_safe(this, next, head1, list) {
911                 if (!match_futex (&this->key, &key1))
912                         continue;
913                 if (++ret <= nr_wake) {
914                         wake_futex(this);
915                 } else {
916                         /*
917                          * If key1 and key2 hash to the same bucket, no need to
918                          * requeue.
919                          */
920                         if (likely(head1 != &hb2->chain)) {
921                                 plist_del(&this->list, &hb1->chain);
922                                 plist_add(&this->list, &hb2->chain);
923                                 this->lock_ptr = &hb2->lock;
924 #ifdef CONFIG_DEBUG_PI_LIST
925                                 this->list.plist.lock = &hb2->lock;
926 #endif
927                         }
928                         this->key = key2;
929                         get_futex_key_refs(&key2);
930                         drop_count++;
931
932                         if (ret - nr_wake >= nr_requeue)
933                                 break;
934                 }
935         }
936
937 out_unlock:
938         spin_unlock(&hb1->lock);
939         if (hb1 != hb2)
940                 spin_unlock(&hb2->lock);
941
942         /* drop_futex_key_refs() must be called outside the spinlocks. */
943         while (--drop_count >= 0)
944                 drop_futex_key_refs(&key1);
945
946 out:
947         if (fshared)
948                 up_read(fshared);
949         return ret;
950 }
951
952 /* The key must be already stored in q->key. */
953 static inline struct futex_hash_bucket *
954 queue_lock(struct futex_q *q, int fd, struct file *filp)
955 {
956         struct futex_hash_bucket *hb;
957
958         q->fd = fd;
959         q->filp = filp;
960
961         init_waitqueue_head(&q->waiters);
962
963         get_futex_key_refs(&q->key);
964         hb = hash_futex(&q->key);
965         q->lock_ptr = &hb->lock;
966
967         spin_lock(&hb->lock);
968         return hb;
969 }
970
971 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
972 {
973         int prio;
974
975         /*
976          * The priority used to register this element is
977          * - either the real thread-priority for the real-time threads
978          * (i.e. threads with a priority lower than MAX_RT_PRIO)
979          * - or MAX_RT_PRIO for non-RT threads.
980          * Thus, all RT-threads are woken first in priority order, and
981          * the others are woken last, in FIFO order.
982          */
983         prio = min(current->normal_prio, MAX_RT_PRIO);
984
985         plist_node_init(&q->list, prio);
986 #ifdef CONFIG_DEBUG_PI_LIST
987         q->list.plist.lock = &hb->lock;
988 #endif
989         plist_add(&q->list, &hb->chain);
990         q->task = current;
991         spin_unlock(&hb->lock);
992 }
993
994 static inline void
995 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
996 {
997         spin_unlock(&hb->lock);
998         drop_futex_key_refs(&q->key);
999 }
1000
1001 /*
1002  * queue_me and unqueue_me must be called as a pair, each
1003  * exactly once.  They are called with the hashed spinlock held.
1004  */
1005
1006 /* The key must be already stored in q->key. */
1007 static void queue_me(struct futex_q *q, int fd, struct file *filp)
1008 {
1009         struct futex_hash_bucket *hb;
1010
1011         hb = queue_lock(q, fd, filp);
1012         __queue_me(q, hb);
1013 }
1014
1015 /* Return 1 if we were still queued (ie. 0 means we were woken) */
1016 static int unqueue_me(struct futex_q *q)
1017 {
1018         spinlock_t *lock_ptr;
1019         int ret = 0;
1020
1021         /* In the common case we don't take the spinlock, which is nice. */
1022  retry:
1023         lock_ptr = q->lock_ptr;
1024         barrier();
1025         if (lock_ptr != 0) {
1026                 spin_lock(lock_ptr);
1027                 /*
1028                  * q->lock_ptr can change between reading it and
1029                  * spin_lock(), causing us to take the wrong lock.  This
1030                  * corrects the race condition.
1031                  *
1032                  * Reasoning goes like this: if we have the wrong lock,
1033                  * q->lock_ptr must have changed (maybe several times)
1034                  * between reading it and the spin_lock().  It can
1035                  * change again after the spin_lock() but only if it was
1036                  * already changed before the spin_lock().  It cannot,
1037                  * however, change back to the original value.  Therefore
1038                  * we can detect whether we acquired the correct lock.
1039                  */
1040                 if (unlikely(lock_ptr != q->lock_ptr)) {
1041                         spin_unlock(lock_ptr);
1042                         goto retry;
1043                 }
1044                 WARN_ON(plist_node_empty(&q->list));
1045                 plist_del(&q->list, &q->list.plist);
1046
1047                 BUG_ON(q->pi_state);
1048
1049                 spin_unlock(lock_ptr);
1050                 ret = 1;
1051         }
1052
1053         drop_futex_key_refs(&q->key);
1054         return ret;
1055 }
1056
1057 /*
1058  * PI futexes can not be requeued and must remove themself from the
1059  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1060  * and dropped here.
1061  */
1062 static void unqueue_me_pi(struct futex_q *q)
1063 {
1064         WARN_ON(plist_node_empty(&q->list));
1065         plist_del(&q->list, &q->list.plist);
1066
1067         BUG_ON(!q->pi_state);
1068         free_pi_state(q->pi_state);
1069         q->pi_state = NULL;
1070
1071         spin_unlock(q->lock_ptr);
1072
1073         drop_futex_key_refs(&q->key);
1074 }
1075
1076 /*
1077  * Fixup the pi_state owner with current.
1078  *
1079  * Must be called with hash bucket lock held and mm->sem held for non
1080  * private futexes.
1081  */
1082 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1083                                 struct task_struct *curr)
1084 {
1085         u32 newtid = curr->pid | FUTEX_WAITERS;
1086         struct futex_pi_state *pi_state = q->pi_state;
1087         u32 uval, curval, newval;
1088         int ret;
1089
1090         /* Owner died? */
1091         if (pi_state->owner != NULL) {
1092                 spin_lock_irq(&pi_state->owner->pi_lock);
1093                 WARN_ON(list_empty(&pi_state->list));
1094                 list_del_init(&pi_state->list);
1095                 spin_unlock_irq(&pi_state->owner->pi_lock);
1096         } else
1097                 newtid |= FUTEX_OWNER_DIED;
1098
1099         pi_state->owner = curr;
1100
1101         spin_lock_irq(&curr->pi_lock);
1102         WARN_ON(!list_empty(&pi_state->list));
1103         list_add(&pi_state->list, &curr->pi_state_list);
1104         spin_unlock_irq(&curr->pi_lock);
1105
1106         /*
1107          * We own it, so we have to replace the pending owner
1108          * TID. This must be atomic as we have preserve the
1109          * owner died bit here.
1110          */
1111         ret = get_futex_value_locked(&uval, uaddr);
1112
1113         while (!ret) {
1114                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1115
1116                 pagefault_disable();
1117                 curval = futex_atomic_cmpxchg_inatomic(uaddr,
1118                                                        uval, newval);
1119                 pagefault_enable();
1120
1121                 if (curval == -EFAULT)
1122                         ret = -EFAULT;
1123                 if (curval == uval)
1124                         break;
1125                 uval = curval;
1126         }
1127         return ret;
1128 }
1129
1130 /*
1131  * In case we must use restart_block to restart a futex_wait,
1132  * we encode in the 'arg3' shared capability
1133  */
1134 #define ARG3_SHARED  1
1135
1136 static long futex_wait_restart(struct restart_block *restart);
1137 static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
1138                       u32 val, ktime_t *abs_time)
1139 {
1140         struct task_struct *curr = current;
1141         DECLARE_WAITQUEUE(wait, curr);
1142         struct futex_hash_bucket *hb;
1143         struct futex_q q;
1144         u32 uval;
1145         int ret;
1146         struct hrtimer_sleeper t;
1147         int rem = 0;
1148
1149         q.pi_state = NULL;
1150  retry:
1151         if (fshared)
1152                 down_read(fshared);
1153
1154         ret = get_futex_key(uaddr, fshared, &q.key);
1155         if (unlikely(ret != 0))
1156                 goto out_release_sem;
1157
1158         hb = queue_lock(&q, -1, NULL);
1159
1160         /*
1161          * Access the page AFTER the futex is queued.
1162          * Order is important:
1163          *
1164          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1165          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1166          *
1167          * The basic logical guarantee of a futex is that it blocks ONLY
1168          * if cond(var) is known to be true at the time of blocking, for
1169          * any cond.  If we queued after testing *uaddr, that would open
1170          * a race condition where we could block indefinitely with
1171          * cond(var) false, which would violate the guarantee.
1172          *
1173          * A consequence is that futex_wait() can return zero and absorb
1174          * a wakeup when *uaddr != val on entry to the syscall.  This is
1175          * rare, but normal.
1176          *
1177          * for shared futexes, we hold the mmap semaphore, so the mapping
1178          * cannot have changed since we looked it up in get_futex_key.
1179          */
1180         ret = get_futex_value_locked(&uval, uaddr);
1181
1182         if (unlikely(ret)) {
1183                 queue_unlock(&q, hb);
1184
1185                 /*
1186                  * If we would have faulted, release mmap_sem, fault it in and
1187                  * start all over again.
1188                  */
1189                 if (fshared)
1190                         up_read(fshared);
1191
1192                 ret = get_user(uval, uaddr);
1193
1194                 if (!ret)
1195                         goto retry;
1196                 return ret;
1197         }
1198         ret = -EWOULDBLOCK;
1199         if (uval != val)
1200                 goto out_unlock_release_sem;
1201
1202         /* Only actually queue if *uaddr contained val.  */
1203         __queue_me(&q, hb);
1204
1205         /*
1206          * Now the futex is queued and we have checked the data, we
1207          * don't want to hold mmap_sem while we sleep.
1208          */
1209         if (fshared)
1210                 up_read(fshared);
1211
1212         /*
1213          * There might have been scheduling since the queue_me(), as we
1214          * cannot hold a spinlock across the get_user() in case it
1215          * faults, and we cannot just set TASK_INTERRUPTIBLE state when
1216          * queueing ourselves into the futex hash.  This code thus has to
1217          * rely on the futex_wake() code removing us from hash when it
1218          * wakes us up.
1219          */
1220
1221         /* add_wait_queue is the barrier after __set_current_state. */
1222         __set_current_state(TASK_INTERRUPTIBLE);
1223         add_wait_queue(&q.waiters, &wait);
1224         /*
1225          * !plist_node_empty() is safe here without any lock.
1226          * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
1227          */
1228         if (likely(!plist_node_empty(&q.list))) {
1229                 if (!abs_time)
1230                         schedule();
1231                 else {
1232                         hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1233                         hrtimer_init_sleeper(&t, current);
1234                         t.timer.expires = *abs_time;
1235
1236                         hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
1237
1238                         /*
1239                          * the timer could have already expired, in which
1240                          * case current would be flagged for rescheduling.
1241                          * Don't bother calling schedule.
1242                          */
1243                         if (likely(t.task))
1244                                 schedule();
1245
1246                         hrtimer_cancel(&t.timer);
1247
1248                         /* Flag if a timeout occured */
1249                         rem = (t.task == NULL);
1250                 }
1251         }
1252         __set_current_state(TASK_RUNNING);
1253
1254         /*
1255          * NOTE: we don't remove ourselves from the waitqueue because
1256          * we are the only user of it.
1257          */
1258
1259         /* If we were woken (and unqueued), we succeeded, whatever. */
1260         if (!unqueue_me(&q))
1261                 return 0;
1262         if (rem)
1263                 return -ETIMEDOUT;
1264
1265         /*
1266          * We expect signal_pending(current), but another thread may
1267          * have handled it for us already.
1268          */
1269         if (!abs_time)
1270                 return -ERESTARTSYS;
1271         else {
1272                 struct restart_block *restart;
1273                 restart = &current_thread_info()->restart_block;
1274                 restart->fn = futex_wait_restart;
1275                 restart->arg0 = (unsigned long)uaddr;
1276                 restart->arg1 = (unsigned long)val;
1277                 restart->arg2 = (unsigned long)abs_time;
1278                 restart->arg3 = 0;
1279                 if (fshared)
1280                         restart->arg3 |= ARG3_SHARED;
1281                 return -ERESTART_RESTARTBLOCK;
1282         }
1283
1284  out_unlock_release_sem:
1285         queue_unlock(&q, hb);
1286
1287  out_release_sem:
1288         if (fshared)
1289                 up_read(fshared);
1290         return ret;
1291 }
1292
1293
1294 static long futex_wait_restart(struct restart_block *restart)
1295 {
1296         u32 __user *uaddr = (u32 __user *)restart->arg0;
1297         u32 val = (u32)restart->arg1;
1298         ktime_t *abs_time = (ktime_t *)restart->arg2;
1299         struct rw_semaphore *fshared = NULL;
1300
1301         restart->fn = do_no_restart_syscall;
1302         if (restart->arg3 & ARG3_SHARED)
1303                 fshared = &current->mm->mmap_sem;
1304         return (long)futex_wait(uaddr, fshared, val, abs_time);
1305 }
1306
1307
1308 /*
1309  * Userspace tried a 0 -> TID atomic transition of the futex value
1310  * and failed. The kernel side here does the whole locking operation:
1311  * if there are waiters then it will block, it does PI, etc. (Due to
1312  * races the kernel might see a 0 value of the futex too.)
1313  */
1314 static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
1315                          int detect, ktime_t *time, int trylock)
1316 {
1317         struct hrtimer_sleeper timeout, *to = NULL;
1318         struct task_struct *curr = current;
1319         struct futex_hash_bucket *hb;
1320         u32 uval, newval, curval;
1321         struct futex_q q;
1322         int ret, lock_taken, ownerdied = 0, attempt = 0;
1323
1324         if (refill_pi_state_cache())
1325                 return -ENOMEM;
1326
1327         if (time) {
1328                 to = &timeout;
1329                 hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
1330                 hrtimer_init_sleeper(to, current);
1331                 to->timer.expires = *time;
1332         }
1333
1334         q.pi_state = NULL;
1335  retry:
1336         if (fshared)
1337                 down_read(fshared);
1338
1339         ret = get_futex_key(uaddr, fshared, &q.key);
1340         if (unlikely(ret != 0))
1341                 goto out_release_sem;
1342
1343  retry_unlocked:
1344         hb = queue_lock(&q, -1, NULL);
1345
1346  retry_locked:
1347         ret = lock_taken = 0;
1348
1349         /*
1350          * To avoid races, we attempt to take the lock here again
1351          * (by doing a 0 -> TID atomic cmpxchg), while holding all
1352          * the locks. It will most likely not succeed.
1353          */
1354         newval = current->pid;
1355
1356         pagefault_disable();
1357         curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
1358         pagefault_enable();
1359
1360         if (unlikely(curval == -EFAULT))
1361                 goto uaddr_faulted;
1362
1363         /*
1364          * Detect deadlocks. In case of REQUEUE_PI this is a valid
1365          * situation and we return success to user space.
1366          */
1367         if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
1368                 ret = -EDEADLK;
1369                 goto out_unlock_release_sem;
1370         }
1371
1372         /*
1373          * Surprise - we got the lock. Just return to userspace:
1374          */
1375         if (unlikely(!curval))
1376                 goto out_unlock_release_sem;
1377
1378         uval = curval;
1379
1380         /*
1381          * Set the WAITERS flag, so the owner will know it has someone
1382          * to wake at next unlock
1383          */
1384         newval = curval | FUTEX_WAITERS;
1385
1386         /*
1387          * There are two cases, where a futex might have no owner (the
1388          * owner TID is 0): OWNER_DIED. We take over the futex in this
1389          * case. We also do an unconditional take over, when the owner
1390          * of the futex died.
1391          *
1392          * This is safe as we are protected by the hash bucket lock !
1393          */
1394         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
1395                 /* Keep the OWNER_DIED bit */
1396                 newval = (curval & ~FUTEX_TID_MASK) | current->pid;
1397                 ownerdied = 0;
1398                 lock_taken = 1;
1399         }
1400
1401         pagefault_disable();
1402         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
1403         pagefault_enable();
1404
1405         if (unlikely(curval == -EFAULT))
1406                 goto uaddr_faulted;
1407         if (unlikely(curval != uval))
1408                 goto retry_locked;
1409
1410         /*
1411          * We took the lock due to owner died take over.
1412          */
1413         if (unlikely(lock_taken))
1414                 goto out_unlock_release_sem;
1415
1416         /*
1417          * We dont have the lock. Look up the PI state (or create it if
1418          * we are the first waiter):
1419          */
1420         ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
1421
1422         if (unlikely(ret)) {
1423                 switch (ret) {
1424
1425                 case -EAGAIN:
1426                         /*
1427                          * Task is exiting and we just wait for the
1428                          * exit to complete.
1429                          */
1430                         queue_unlock(&q, hb);
1431                         if (fshared)
1432                                 up_read(fshared);
1433                         cond_resched();
1434                         goto retry;
1435
1436                 case -ESRCH:
1437                         /*
1438                          * No owner found for this futex. Check if the
1439                          * OWNER_DIED bit is set to figure out whether
1440                          * this is a robust futex or not.
1441                          */
1442                         if (get_futex_value_locked(&curval, uaddr))
1443                                 goto uaddr_faulted;
1444
1445                         /*
1446                          * We simply start over in case of a robust
1447                          * futex. The code above will take the futex
1448                          * and return happy.
1449                          */
1450                         if (curval & FUTEX_OWNER_DIED) {
1451                                 ownerdied = 1;
1452                                 goto retry_locked;
1453                         }
1454                 default:
1455                         goto out_unlock_release_sem;
1456                 }
1457         }
1458
1459         /*
1460          * Only actually queue now that the atomic ops are done:
1461          */
1462         __queue_me(&q, hb);
1463
1464         /*
1465          * Now the futex is queued and we have checked the data, we
1466          * don't want to hold mmap_sem while we sleep.
1467          */
1468         if (fshared)
1469                 up_read(fshared);
1470
1471         WARN_ON(!q.pi_state);
1472         /*
1473          * Block on the PI mutex:
1474          */
1475         if (!trylock)
1476                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1477         else {
1478                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1479                 /* Fixup the trylock return value: */
1480                 ret = ret ? 0 : -EWOULDBLOCK;
1481         }
1482
1483         if (fshared)
1484                 down_read(fshared);
1485         spin_lock(q.lock_ptr);
1486
1487         if (!ret) {
1488                 /*
1489                  * Got the lock. We might not be the anticipated owner
1490                  * if we did a lock-steal - fix up the PI-state in
1491                  * that case:
1492                  */
1493                 if (q.pi_state->owner != curr)
1494                         ret = fixup_pi_state_owner(uaddr, &q, curr);
1495         } else {
1496                 /*
1497                  * Catch the rare case, where the lock was released
1498                  * when we were on the way back before we locked the
1499                  * hash bucket.
1500                  */
1501                 if (q.pi_state->owner == curr &&
1502                     rt_mutex_trylock(&q.pi_state->pi_mutex)) {
1503                         ret = 0;
1504                 } else {
1505                         /*
1506                          * Paranoia check. If we did not take the lock
1507                          * in the trylock above, then we should not be
1508                          * the owner of the rtmutex, neither the real
1509                          * nor the pending one:
1510                          */
1511                         if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
1512                                 printk(KERN_ERR "futex_lock_pi: ret = %d "
1513                                        "pi-mutex: %p pi-state %p\n", ret,
1514                                        q.pi_state->pi_mutex.owner,
1515                                        q.pi_state->owner);
1516                 }
1517         }
1518
1519         /* Unqueue and drop the lock */
1520         unqueue_me_pi(&q);
1521         if (fshared)
1522                 up_read(fshared);
1523
1524         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1525
1526  out_unlock_release_sem:
1527         queue_unlock(&q, hb);
1528
1529  out_release_sem:
1530         if (fshared)
1531                 up_read(fshared);
1532         return ret;
1533
1534  uaddr_faulted:
1535         /*
1536          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1537          * non-atomically.  Therefore, if get_user below is not
1538          * enough, we need to handle the fault ourselves, while
1539          * still holding the mmap_sem.
1540          *
1541          * ... and hb->lock. :-) --ANK
1542          */
1543         queue_unlock(&q, hb);
1544
1545         if (attempt++) {
1546                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1547                                          attempt);
1548                 if (ret)
1549                         goto out_release_sem;
1550                 goto retry_unlocked;
1551         }
1552
1553         if (fshared)
1554                 up_read(fshared);
1555
1556         ret = get_user(uval, uaddr);
1557         if (!ret && (uval != -EFAULT))
1558                 goto retry;
1559
1560         return ret;
1561 }
1562
1563 /*
1564  * Userspace attempted a TID -> 0 atomic transition, and failed.
1565  * This is the in-kernel slowpath: we look up the PI state (if any),
1566  * and do the rt-mutex unlock.
1567  */
1568 static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
1569 {
1570         struct futex_hash_bucket *hb;
1571         struct futex_q *this, *next;
1572         u32 uval;
1573         struct plist_head *head;
1574         union futex_key key;
1575         int ret, attempt = 0;
1576
1577 retry:
1578         if (get_user(uval, uaddr))
1579                 return -EFAULT;
1580         /*
1581          * We release only a lock we actually own:
1582          */
1583         if ((uval & FUTEX_TID_MASK) != current->pid)
1584                 return -EPERM;
1585         /*
1586          * First take all the futex related locks:
1587          */
1588         if (fshared)
1589                 down_read(fshared);
1590
1591         ret = get_futex_key(uaddr, fshared, &key);
1592         if (unlikely(ret != 0))
1593                 goto out;
1594
1595         hb = hash_futex(&key);
1596 retry_unlocked:
1597         spin_lock(&hb->lock);
1598
1599         /*
1600          * To avoid races, try to do the TID -> 0 atomic transition
1601          * again. If it succeeds then we can return without waking
1602          * anyone else up:
1603          */
1604         if (!(uval & FUTEX_OWNER_DIED)) {
1605                 pagefault_disable();
1606                 uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
1607                 pagefault_enable();
1608         }
1609
1610         if (unlikely(uval == -EFAULT))
1611                 goto pi_faulted;
1612         /*
1613          * Rare case: we managed to release the lock atomically,
1614          * no need to wake anyone else up:
1615          */
1616         if (unlikely(uval == current->pid))
1617                 goto out_unlock;
1618
1619         /*
1620          * Ok, other tasks may need to be woken up - check waiters
1621          * and do the wakeup if necessary:
1622          */
1623         head = &hb->chain;
1624
1625         plist_for_each_entry_safe(this, next, head, list) {
1626                 if (!match_futex (&this->key, &key))
1627                         continue;
1628                 ret = wake_futex_pi(uaddr, uval, this);
1629                 /*
1630                  * The atomic access to the futex value
1631                  * generated a pagefault, so retry the
1632                  * user-access and the wakeup:
1633                  */
1634                 if (ret == -EFAULT)
1635                         goto pi_faulted;
1636                 goto out_unlock;
1637         }
1638         /*
1639          * No waiters - kernel unlocks the futex:
1640          */
1641         if (!(uval & FUTEX_OWNER_DIED)) {
1642                 ret = unlock_futex_pi(uaddr, uval);
1643                 if (ret == -EFAULT)
1644                         goto pi_faulted;
1645         }
1646
1647 out_unlock:
1648         spin_unlock(&hb->lock);
1649 out:
1650         if (fshared)
1651                 up_read(fshared);
1652
1653         return ret;
1654
1655 pi_faulted:
1656         /*
1657          * We have to r/w  *(int __user *)uaddr, but we can't modify it
1658          * non-atomically.  Therefore, if get_user below is not
1659          * enough, we need to handle the fault ourselves, while
1660          * still holding the mmap_sem.
1661          *
1662          * ... and hb->lock. --ANK
1663          */
1664         spin_unlock(&hb->lock);
1665
1666         if (attempt++) {
1667                 ret = futex_handle_fault((unsigned long)uaddr, fshared,
1668                                          attempt);
1669                 if (ret)
1670                         goto out;
1671                 goto retry_unlocked;
1672         }
1673
1674         if (fshared)
1675                 up_read(fshared);
1676
1677         ret = get_user(uval, uaddr);
1678         if (!ret && (uval != -EFAULT))
1679                 goto retry;
1680
1681         return ret;
1682 }
1683
1684 static int futex_close(struct inode *inode, struct file *filp)
1685 {
1686         struct futex_q *q = filp->private_data;
1687
1688         unqueue_me(q);
1689         kfree(q);
1690
1691         return 0;
1692 }
1693
1694 /* This is one-shot: once it's gone off you need a new fd */
1695 static unsigned int futex_poll(struct file *filp,
1696                                struct poll_table_struct *wait)
1697 {
1698         struct futex_q *q = filp->private_data;
1699         int ret = 0;
1700
1701         poll_wait(filp, &q->waiters, wait);
1702
1703         /*
1704          * plist_node_empty() is safe here without any lock.
1705          * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
1706          */
1707         if (plist_node_empty(&q->list))
1708                 ret = POLLIN | POLLRDNORM;
1709
1710         return ret;
1711 }
1712
1713 static const struct file_operations futex_fops = {
1714         .release        = futex_close,
1715         .poll           = futex_poll,
1716 };
1717
1718 /*
1719  * Signal allows caller to avoid the race which would occur if they
1720  * set the sigio stuff up afterwards.
1721  */
1722 static int futex_fd(u32 __user *uaddr, int signal)
1723 {
1724         struct futex_q *q;
1725         struct file *filp;
1726         int ret, err;
1727         struct rw_semaphore *fshared;
1728         static unsigned long printk_interval;
1729
1730         if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
1731                 printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
1732                         "will be removed from the kernel in June 2007\n",
1733                         current->comm);
1734         }
1735
1736         ret = -EINVAL;
1737         if (!valid_signal(signal))
1738                 goto out;
1739
1740         ret = get_unused_fd();
1741         if (ret < 0)
1742                 goto out;
1743         filp = get_empty_filp();
1744         if (!filp) {
1745                 put_unused_fd(ret);
1746                 ret = -ENFILE;
1747                 goto out;
1748         }
1749         filp->f_op = &futex_fops;
1750         filp->f_path.mnt = mntget(futex_mnt);
1751         filp->f_path.dentry = dget(futex_mnt->mnt_root);
1752         filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
1753
1754         if (signal) {
1755                 err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
1756                 if (err < 0) {
1757                         goto error;
1758                 }
1759                 filp->f_owner.signum = signal;
1760         }
1761
1762         q = kmalloc(sizeof(*q), GFP_KERNEL);
1763         if (!q) {
1764                 err = -ENOMEM;
1765                 goto error;
1766         }
1767         q->pi_state = NULL;
1768
1769         fshared = &current->mm->mmap_sem;
1770         down_read(fshared);
1771         err = get_futex_key(uaddr, fshared, &q->key);
1772
1773         if (unlikely(err != 0)) {
1774                 up_read(fshared);
1775                 kfree(q);
1776                 goto error;
1777         }
1778
1779         /*
1780          * queue_me() must be called before releasing mmap_sem, because
1781          * key->shared.inode needs to be referenced while holding it.
1782          */
1783         filp->private_data = q;
1784
1785         queue_me(q, ret, filp);
1786         up_read(fshared);
1787
1788         /* Now we map fd to filp, so userspace can access it */
1789         fd_install(ret, filp);
1790 out:
1791         return ret;
1792 error:
1793         put_unused_fd(ret);
1794         put_filp(filp);
1795         ret = err;
1796         goto out;
1797 }
1798
1799 /*
1800  * Support for robust futexes: the kernel cleans up held futexes at
1801  * thread exit time.
1802  *
1803  * Implementation: user-space maintains a per-thread list of locks it
1804  * is holding. Upon do_exit(), the kernel carefully walks this list,
1805  * and marks all locks that are owned by this thread with the
1806  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
1807  * always manipulated with the lock held, so the list is private and
1808  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
1809  * field, to allow the kernel to clean up if the thread dies after
1810  * acquiring the lock, but just before it could have added itself to
1811  * the list. There can only be one such pending lock.
1812  */
1813
1814 /**
1815  * sys_set_robust_list - set the robust-futex list head of a task
1816  * @head: pointer to the list-head
1817  * @len: length of the list-head, as userspace expects
1818  */
1819 asmlinkage long
1820 sys_set_robust_list(struct robust_list_head __user *head,
1821                     size_t len)
1822 {
1823         /*
1824          * The kernel knows only one size for now:
1825          */
1826         if (unlikely(len != sizeof(*head)))
1827                 return -EINVAL;
1828
1829         current->robust_list = head;
1830
1831         return 0;
1832 }
1833
1834 /**
1835  * sys_get_robust_list - get the robust-futex list head of a task
1836  * @pid: pid of the process [zero for current task]
1837  * @head_ptr: pointer to a list-head pointer, the kernel fills it in
1838  * @len_ptr: pointer to a length field, the kernel fills in the header size
1839  */
1840 asmlinkage long
1841 sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
1842                     size_t __user *len_ptr)
1843 {
1844         struct robust_list_head __user *head;
1845         unsigned long ret;
1846
1847         if (!pid)
1848                 head = current->robust_list;
1849         else {
1850                 struct task_struct *p;
1851
1852                 ret = -ESRCH;
1853                 rcu_read_lock();
1854                 p = find_task_by_pid(pid);
1855                 if (!p)
1856                         goto err_unlock;
1857                 ret = -EPERM;
1858                 if ((current->euid != p->euid) && (current->euid != p->uid) &&
1859                                 !capable(CAP_SYS_PTRACE))
1860                         goto err_unlock;
1861                 head = p->robust_list;
1862                 rcu_read_unlock();
1863         }
1864
1865         if (put_user(sizeof(*head), len_ptr))
1866                 return -EFAULT;
1867         return put_user(head, head_ptr);
1868
1869 err_unlock:
1870         rcu_read_unlock();
1871
1872         return ret;
1873 }
1874
1875 /*
1876  * Process a futex-list entry, check whether it's owned by the
1877  * dying task, and do notification if so:
1878  */
1879 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
1880 {
1881         u32 uval, nval, mval;
1882
1883 retry:
1884         if (get_user(uval, uaddr))
1885                 return -1;
1886
1887         if ((uval & FUTEX_TID_MASK) == curr->pid) {
1888                 /*
1889                  * Ok, this dying thread is truly holding a futex
1890                  * of interest. Set the OWNER_DIED bit atomically
1891                  * via cmpxchg, and if the value had FUTEX_WAITERS
1892                  * set, wake up a waiter (if any). (We have to do a
1893                  * futex_wake() even if OWNER_DIED is already set -
1894                  * to handle the rare but possible case of recursive
1895                  * thread-death.) The rest of the cleanup is done in
1896                  * userspace.
1897                  */
1898                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1899                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
1900
1901                 if (nval == -EFAULT)
1902                         return -1;
1903
1904                 if (nval != uval)
1905                         goto retry;
1906
1907                 /*
1908                  * Wake robust non-PI futexes here. The wakeup of
1909                  * PI futexes happens in exit_pi_state():
1910                  */
1911                 if (!pi) {
1912                         if (uval & FUTEX_WAITERS)
1913                                 futex_wake(uaddr, &curr->mm->mmap_sem, 1);
1914                 }
1915         }
1916         return 0;
1917 }
1918
1919 /*
1920  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1921  */
1922 static inline int fetch_robust_entry(struct robust_list __user **entry,
1923                                      struct robust_list __user * __user *head,
1924                                      int *pi)
1925 {
1926         unsigned long uentry;
1927
1928         if (get_user(uentry, (unsigned long __user *)head))
1929                 return -EFAULT;
1930
1931         *entry = (void __user *)(uentry & ~1UL);
1932         *pi = uentry & 1;
1933
1934         return 0;
1935 }
1936
1937 /*
1938  * Walk curr->robust_list (very carefully, it's a userspace list!)
1939  * and mark any locks found there dead, and notify any waiters.
1940  *
1941  * We silently return on any sign of list-walking problem.
1942  */
1943 void exit_robust_list(struct task_struct *curr)
1944 {
1945         struct robust_list_head __user *head = curr->robust_list;
1946         struct robust_list __user *entry, *pending;
1947         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1948         unsigned long futex_offset;
1949
1950         /*
1951          * Fetch the list head (which was registered earlier, via
1952          * sys_set_robust_list()):
1953          */
1954         if (fetch_robust_entry(&entry, &head->list.next, &pi))
1955                 return;
1956         /*
1957          * Fetch the relative futex offset:
1958          */
1959         if (get_user(futex_offset, &head->futex_offset))
1960                 return;
1961         /*
1962          * Fetch any possibly pending lock-add first, and handle it
1963          * if it exists:
1964          */
1965         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
1966                 return;
1967
1968         if (pending)
1969                 handle_futex_death((void __user *)pending + futex_offset,
1970                                    curr, pip);
1971
1972         while (entry != &head->list) {
1973                 /*
1974                  * A pending lock might already be on the list, so
1975                  * don't process it twice:
1976                  */
1977                 if (entry != pending)
1978                         if (handle_futex_death((void __user *)entry + futex_offset,
1979                                                 curr, pi))
1980                                 return;
1981                 /*
1982                  * Fetch the next entry in the list:
1983                  */
1984                 if (fetch_robust_entry(&entry, &entry->next, &pi))
1985                         return;
1986                 /*
1987                  * Avoid excessively long or circular lists:
1988                  */
1989                 if (!--limit)
1990                         break;
1991
1992                 cond_resched();
1993         }
1994 }
1995
1996 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
1997                 u32 __user *uaddr2, u32 val2, u32 val3)
1998 {
1999         int ret;
2000         int cmd = op & FUTEX_CMD_MASK;
2001         struct rw_semaphore *fshared = NULL;
2002
2003         if (!(op & FUTEX_PRIVATE_FLAG))
2004                 fshared = &current->mm->mmap_sem;
2005
2006         switch (cmd) {
2007         case FUTEX_WAIT:
2008                 ret = futex_wait(uaddr, fshared, val, timeout);
2009                 break;
2010         case FUTEX_WAKE:
2011                 ret = futex_wake(uaddr, fshared, val);
2012                 break;
2013         case FUTEX_FD:
2014                 /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
2015                 ret = futex_fd(uaddr, val);
2016                 break;
2017         case FUTEX_REQUEUE:
2018                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
2019                 break;
2020         case FUTEX_CMP_REQUEUE:
2021                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
2022                 break;
2023         case FUTEX_WAKE_OP:
2024                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2025                 break;
2026         case FUTEX_LOCK_PI:
2027                 ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2028                 break;
2029         case FUTEX_UNLOCK_PI:
2030                 ret = futex_unlock_pi(uaddr, fshared);
2031                 break;
2032         case FUTEX_TRYLOCK_PI:
2033                 ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2034                 break;
2035         default:
2036                 ret = -ENOSYS;
2037         }
2038         return ret;
2039 }
2040
2041
2042 asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
2043                           struct timespec __user *utime, u32 __user *uaddr2,
2044                           u32 val3)
2045 {
2046         struct timespec ts;
2047         ktime_t t, *tp = NULL;
2048         u32 val2 = 0;
2049         int cmd = op & FUTEX_CMD_MASK;
2050
2051         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
2052                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2053                         return -EFAULT;
2054                 if (!timespec_valid(&ts))
2055                         return -EINVAL;
2056
2057                 t = timespec_to_ktime(ts);
2058                 if (cmd == FUTEX_WAIT)
2059                         t = ktime_add(ktime_get(), t);
2060                 tp = &t;
2061         }
2062         /*
2063          * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
2064          */
2065         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE)
2066                 val2 = (u32) (unsigned long) utime;
2067
2068         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2069 }
2070
2071 static int futexfs_get_sb(struct file_system_type *fs_type,
2072                           int flags, const char *dev_name, void *data,
2073                           struct vfsmount *mnt)
2074 {
2075         return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
2076 }
2077
2078 static struct file_system_type futex_fs_type = {
2079         .name           = "futexfs",
2080         .get_sb         = futexfs_get_sb,
2081         .kill_sb        = kill_anon_super,
2082 };
2083
2084 static int __init init(void)
2085 {
2086         int i = register_filesystem(&futex_fs_type);
2087
2088         if (i)
2089                 return i;
2090
2091         futex_mnt = kern_mount(&futex_fs_type);
2092         if (IS_ERR(futex_mnt)) {
2093                 unregister_filesystem(&futex_fs_type);
2094                 return PTR_ERR(futex_mnt);
2095         }
2096
2097         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2098                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2099                 spin_lock_init(&futex_queues[i].lock);
2100         }
2101         return 0;
2102 }
2103 __initcall(init);