cpumask: Fix generate_sched_domains() for UP
[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  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Priority Inheritance state:
73  */
74 struct futex_pi_state {
75         /*
76          * list of 'owned' pi_state instances - these have to be
77          * cleaned up in do_exit() if the task exits prematurely:
78          */
79         struct list_head list;
80
81         /*
82          * The PI object:
83          */
84         struct rt_mutex pi_mutex;
85
86         struct task_struct *owner;
87         atomic_t refcount;
88
89         union futex_key key;
90 };
91
92 /**
93  * struct futex_q - The hashed futex queue entry, one per waiting task
94  * @task:               the task waiting on the futex
95  * @lock_ptr:           the hash bucket lock
96  * @key:                the key the futex is hashed on
97  * @pi_state:           optional priority inheritance state
98  * @rt_waiter:          rt_waiter storage for use with requeue_pi
99  * @requeue_pi_key:     the requeue_pi target futex key
100  * @bitset:             bitset for the optional bitmasked wakeup
101  *
102  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103  * we can wake only the relevant ones (hashed queues may be shared).
104  *
105  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107  * The order of wakup is always to make the first condition true, then
108  * the second.
109  *
110  * PI futexes are typically woken before they are removed from the hash list via
111  * the rt_mutex code. See unqueue_me_pi().
112  */
113 struct futex_q {
114         struct plist_node list;
115
116         struct task_struct *task;
117         spinlock_t *lock_ptr;
118         union futex_key key;
119         struct futex_pi_state *pi_state;
120         struct rt_mutex_waiter *rt_waiter;
121         union futex_key *requeue_pi_key;
122         u32 bitset;
123 };
124
125 /*
126  * Hash buckets are shared by all the futex_keys that hash to the same
127  * location.  Each key may have multiple futex_q structures, one for each task
128  * waiting on a futex.
129  */
130 struct futex_hash_bucket {
131         spinlock_t lock;
132         struct plist_head chain;
133 };
134
135 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
136
137 /*
138  * We hash on the keys returned from get_futex_key (see below).
139  */
140 static struct futex_hash_bucket *hash_futex(union futex_key *key)
141 {
142         u32 hash = jhash2((u32*)&key->both.word,
143                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
144                           key->both.offset);
145         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146 }
147
148 /*
149  * Return 1 if two futex_keys are equal, 0 otherwise.
150  */
151 static inline int match_futex(union futex_key *key1, union futex_key *key2)
152 {
153         return (key1 && key2
154                 && key1->both.word == key2->both.word
155                 && key1->both.ptr == key2->both.ptr
156                 && key1->both.offset == key2->both.offset);
157 }
158
159 /*
160  * Take a reference to the resource addressed by a key.
161  * Can be called while holding spinlocks.
162  *
163  */
164 static void get_futex_key_refs(union futex_key *key)
165 {
166         if (!key->both.ptr)
167                 return;
168
169         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170         case FUT_OFF_INODE:
171                 atomic_inc(&key->shared.inode->i_count);
172                 break;
173         case FUT_OFF_MMSHARED:
174                 atomic_inc(&key->private.mm->mm_count);
175                 break;
176         }
177 }
178
179 /*
180  * Drop a reference to the resource addressed by a key.
181  * The hash bucket spinlock must not be held.
182  */
183 static void drop_futex_key_refs(union futex_key *key)
184 {
185         if (!key->both.ptr) {
186                 /* If we're here then we tried to put a key we failed to get */
187                 WARN_ON_ONCE(1);
188                 return;
189         }
190
191         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192         case FUT_OFF_INODE:
193                 iput(key->shared.inode);
194                 break;
195         case FUT_OFF_MMSHARED:
196                 mmdrop(key->private.mm);
197                 break;
198         }
199 }
200
201 /**
202  * get_futex_key() - Get parameters which are the keys for a futex
203  * @uaddr:      virtual address of the futex
204  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205  * @key:        address where result is stored.
206  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
207  *              VERIFY_WRITE)
208  *
209  * Returns a negative error code or 0
210  * The key words are stored in *key on success.
211  *
212  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
213  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
214  * We can usually work out the index without swapping in the page.
215  *
216  * lock_page() might sleep, the caller should not hold a spinlock.
217  */
218 static int
219 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
220 {
221         unsigned long address = (unsigned long)uaddr;
222         struct mm_struct *mm = current->mm;
223         struct page *page;
224         int err;
225
226         /*
227          * The futex address must be "naturally" aligned.
228          */
229         key->both.offset = address % PAGE_SIZE;
230         if (unlikely((address % sizeof(u32)) != 0))
231                 return -EINVAL;
232         address -= key->both.offset;
233
234         /*
235          * PROCESS_PRIVATE futexes are fast.
236          * As the mm cannot disappear under us and the 'key' only needs
237          * virtual address, we dont even have to find the underlying vma.
238          * Note : We do have to check 'uaddr' is a valid user address,
239          *        but access_ok() should be faster than find_vma()
240          */
241         if (!fshared) {
242                 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
243                         return -EFAULT;
244                 key->private.mm = mm;
245                 key->private.address = address;
246                 get_futex_key_refs(key);
247                 return 0;
248         }
249
250 again:
251         err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
252         if (err < 0)
253                 return err;
254
255         page = compound_head(page);
256         lock_page(page);
257         if (!page->mapping) {
258                 unlock_page(page);
259                 put_page(page);
260                 goto again;
261         }
262
263         /*
264          * Private mappings are handled in a simple way.
265          *
266          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
267          * it's a read-only handle, it's expected that futexes attach to
268          * the object not the particular process.
269          */
270         if (PageAnon(page)) {
271                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
272                 key->private.mm = mm;
273                 key->private.address = address;
274         } else {
275                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
276                 key->shared.inode = page->mapping->host;
277                 key->shared.pgoff = page->index;
278         }
279
280         get_futex_key_refs(key);
281
282         unlock_page(page);
283         put_page(page);
284         return 0;
285 }
286
287 static inline
288 void put_futex_key(int fshared, union futex_key *key)
289 {
290         drop_futex_key_refs(key);
291 }
292
293 /**
294  * fault_in_user_writeable() - Fault in user address and verify RW access
295  * @uaddr:      pointer to faulting user space address
296  *
297  * Slow path to fixup the fault we just took in the atomic write
298  * access to @uaddr.
299  *
300  * We have no generic implementation of a non destructive write to the
301  * user address. We know that we faulted in the atomic pagefault
302  * disabled section so we can as well avoid the #PF overhead by
303  * calling get_user_pages() right away.
304  */
305 static int fault_in_user_writeable(u32 __user *uaddr)
306 {
307         int ret = get_user_pages(current, current->mm, (unsigned long)uaddr,
308                                  1, 1, 0, NULL, NULL);
309         return ret < 0 ? ret : 0;
310 }
311
312 /**
313  * futex_top_waiter() - Return the highest priority waiter on a futex
314  * @hb:         the hash bucket the futex_q's reside in
315  * @key:        the futex key (to distinguish it from other futex futex_q's)
316  *
317  * Must be called with the hb lock held.
318  */
319 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
320                                         union futex_key *key)
321 {
322         struct futex_q *this;
323
324         plist_for_each_entry(this, &hb->chain, list) {
325                 if (match_futex(&this->key, key))
326                         return this;
327         }
328         return NULL;
329 }
330
331 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
332 {
333         u32 curval;
334
335         pagefault_disable();
336         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
337         pagefault_enable();
338
339         return curval;
340 }
341
342 static int get_futex_value_locked(u32 *dest, u32 __user *from)
343 {
344         int ret;
345
346         pagefault_disable();
347         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
348         pagefault_enable();
349
350         return ret ? -EFAULT : 0;
351 }
352
353
354 /*
355  * PI code:
356  */
357 static int refill_pi_state_cache(void)
358 {
359         struct futex_pi_state *pi_state;
360
361         if (likely(current->pi_state_cache))
362                 return 0;
363
364         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
365
366         if (!pi_state)
367                 return -ENOMEM;
368
369         INIT_LIST_HEAD(&pi_state->list);
370         /* pi_mutex gets initialized later */
371         pi_state->owner = NULL;
372         atomic_set(&pi_state->refcount, 1);
373         pi_state->key = FUTEX_KEY_INIT;
374
375         current->pi_state_cache = pi_state;
376
377         return 0;
378 }
379
380 static struct futex_pi_state * alloc_pi_state(void)
381 {
382         struct futex_pi_state *pi_state = current->pi_state_cache;
383
384         WARN_ON(!pi_state);
385         current->pi_state_cache = NULL;
386
387         return pi_state;
388 }
389
390 static void free_pi_state(struct futex_pi_state *pi_state)
391 {
392         if (!atomic_dec_and_test(&pi_state->refcount))
393                 return;
394
395         /*
396          * If pi_state->owner is NULL, the owner is most probably dying
397          * and has cleaned up the pi_state already
398          */
399         if (pi_state->owner) {
400                 spin_lock_irq(&pi_state->owner->pi_lock);
401                 list_del_init(&pi_state->list);
402                 spin_unlock_irq(&pi_state->owner->pi_lock);
403
404                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
405         }
406
407         if (current->pi_state_cache)
408                 kfree(pi_state);
409         else {
410                 /*
411                  * pi_state->list is already empty.
412                  * clear pi_state->owner.
413                  * refcount is at 0 - put it back to 1.
414                  */
415                 pi_state->owner = NULL;
416                 atomic_set(&pi_state->refcount, 1);
417                 current->pi_state_cache = pi_state;
418         }
419 }
420
421 /*
422  * Look up the task based on what TID userspace gave us.
423  * We dont trust it.
424  */
425 static struct task_struct * futex_find_get_task(pid_t pid)
426 {
427         struct task_struct *p;
428         const struct cred *cred = current_cred(), *pcred;
429
430         rcu_read_lock();
431         p = find_task_by_vpid(pid);
432         if (!p) {
433                 p = ERR_PTR(-ESRCH);
434         } else {
435                 pcred = __task_cred(p);
436                 if (cred->euid != pcred->euid &&
437                     cred->euid != pcred->uid)
438                         p = ERR_PTR(-ESRCH);
439                 else
440                         get_task_struct(p);
441         }
442
443         rcu_read_unlock();
444
445         return p;
446 }
447
448 /*
449  * This task is holding PI mutexes at exit time => bad.
450  * Kernel cleans up PI-state, but userspace is likely hosed.
451  * (Robust-futex cleanup is separate and might save the day for userspace.)
452  */
453 void exit_pi_state_list(struct task_struct *curr)
454 {
455         struct list_head *next, *head = &curr->pi_state_list;
456         struct futex_pi_state *pi_state;
457         struct futex_hash_bucket *hb;
458         union futex_key key = FUTEX_KEY_INIT;
459
460         if (!futex_cmpxchg_enabled)
461                 return;
462         /*
463          * We are a ZOMBIE and nobody can enqueue itself on
464          * pi_state_list anymore, but we have to be careful
465          * versus waiters unqueueing themselves:
466          */
467         spin_lock_irq(&curr->pi_lock);
468         while (!list_empty(head)) {
469
470                 next = head->next;
471                 pi_state = list_entry(next, struct futex_pi_state, list);
472                 key = pi_state->key;
473                 hb = hash_futex(&key);
474                 spin_unlock_irq(&curr->pi_lock);
475
476                 spin_lock(&hb->lock);
477
478                 spin_lock_irq(&curr->pi_lock);
479                 /*
480                  * We dropped the pi-lock, so re-check whether this
481                  * task still owns the PI-state:
482                  */
483                 if (head->next != next) {
484                         spin_unlock(&hb->lock);
485                         continue;
486                 }
487
488                 WARN_ON(pi_state->owner != curr);
489                 WARN_ON(list_empty(&pi_state->list));
490                 list_del_init(&pi_state->list);
491                 pi_state->owner = NULL;
492                 spin_unlock_irq(&curr->pi_lock);
493
494                 rt_mutex_unlock(&pi_state->pi_mutex);
495
496                 spin_unlock(&hb->lock);
497
498                 spin_lock_irq(&curr->pi_lock);
499         }
500         spin_unlock_irq(&curr->pi_lock);
501 }
502
503 static int
504 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
505                 union futex_key *key, struct futex_pi_state **ps)
506 {
507         struct futex_pi_state *pi_state = NULL;
508         struct futex_q *this, *next;
509         struct plist_head *head;
510         struct task_struct *p;
511         pid_t pid = uval & FUTEX_TID_MASK;
512
513         head = &hb->chain;
514
515         plist_for_each_entry_safe(this, next, head, list) {
516                 if (match_futex(&this->key, key)) {
517                         /*
518                          * Another waiter already exists - bump up
519                          * the refcount and return its pi_state:
520                          */
521                         pi_state = this->pi_state;
522                         /*
523                          * Userspace might have messed up non PI and PI futexes
524                          */
525                         if (unlikely(!pi_state))
526                                 return -EINVAL;
527
528                         WARN_ON(!atomic_read(&pi_state->refcount));
529                         WARN_ON(pid && pi_state->owner &&
530                                 pi_state->owner->pid != pid);
531
532                         atomic_inc(&pi_state->refcount);
533                         *ps = pi_state;
534
535                         return 0;
536                 }
537         }
538
539         /*
540          * We are the first waiter - try to look up the real owner and attach
541          * the new pi_state to it, but bail out when TID = 0
542          */
543         if (!pid)
544                 return -ESRCH;
545         p = futex_find_get_task(pid);
546         if (IS_ERR(p))
547                 return PTR_ERR(p);
548
549         /*
550          * We need to look at the task state flags to figure out,
551          * whether the task is exiting. To protect against the do_exit
552          * change of the task flags, we do this protected by
553          * p->pi_lock:
554          */
555         spin_lock_irq(&p->pi_lock);
556         if (unlikely(p->flags & PF_EXITING)) {
557                 /*
558                  * The task is on the way out. When PF_EXITPIDONE is
559                  * set, we know that the task has finished the
560                  * cleanup:
561                  */
562                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
563
564                 spin_unlock_irq(&p->pi_lock);
565                 put_task_struct(p);
566                 return ret;
567         }
568
569         pi_state = alloc_pi_state();
570
571         /*
572          * Initialize the pi_mutex in locked state and make 'p'
573          * the owner of it:
574          */
575         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
576
577         /* Store the key for possible exit cleanups: */
578         pi_state->key = *key;
579
580         WARN_ON(!list_empty(&pi_state->list));
581         list_add(&pi_state->list, &p->pi_state_list);
582         pi_state->owner = p;
583         spin_unlock_irq(&p->pi_lock);
584
585         put_task_struct(p);
586
587         *ps = pi_state;
588
589         return 0;
590 }
591
592 /**
593  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
594  * @uaddr:              the pi futex user address
595  * @hb:                 the pi futex hash bucket
596  * @key:                the futex key associated with uaddr and hb
597  * @ps:                 the pi_state pointer where we store the result of the
598  *                      lookup
599  * @task:               the task to perform the atomic lock work for.  This will
600  *                      be "current" except in the case of requeue pi.
601  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
602  *
603  * Returns:
604  *  0 - ready to wait
605  *  1 - acquired the lock
606  * <0 - error
607  *
608  * The hb->lock and futex_key refs shall be held by the caller.
609  */
610 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
611                                 union futex_key *key,
612                                 struct futex_pi_state **ps,
613                                 struct task_struct *task, int set_waiters)
614 {
615         int lock_taken, ret, ownerdied = 0;
616         u32 uval, newval, curval;
617
618 retry:
619         ret = lock_taken = 0;
620
621         /*
622          * To avoid races, we attempt to take the lock here again
623          * (by doing a 0 -> TID atomic cmpxchg), while holding all
624          * the locks. It will most likely not succeed.
625          */
626         newval = task_pid_vnr(task);
627         if (set_waiters)
628                 newval |= FUTEX_WAITERS;
629
630         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
631
632         if (unlikely(curval == -EFAULT))
633                 return -EFAULT;
634
635         /*
636          * Detect deadlocks.
637          */
638         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
639                 return -EDEADLK;
640
641         /*
642          * Surprise - we got the lock. Just return to userspace:
643          */
644         if (unlikely(!curval))
645                 return 1;
646
647         uval = curval;
648
649         /*
650          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
651          * to wake at the next unlock.
652          */
653         newval = curval | FUTEX_WAITERS;
654
655         /*
656          * There are two cases, where a futex might have no owner (the
657          * owner TID is 0): OWNER_DIED. We take over the futex in this
658          * case. We also do an unconditional take over, when the owner
659          * of the futex died.
660          *
661          * This is safe as we are protected by the hash bucket lock !
662          */
663         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
664                 /* Keep the OWNER_DIED bit */
665                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
666                 ownerdied = 0;
667                 lock_taken = 1;
668         }
669
670         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
671
672         if (unlikely(curval == -EFAULT))
673                 return -EFAULT;
674         if (unlikely(curval != uval))
675                 goto retry;
676
677         /*
678          * We took the lock due to owner died take over.
679          */
680         if (unlikely(lock_taken))
681                 return 1;
682
683         /*
684          * We dont have the lock. Look up the PI state (or create it if
685          * we are the first waiter):
686          */
687         ret = lookup_pi_state(uval, hb, key, ps);
688
689         if (unlikely(ret)) {
690                 switch (ret) {
691                 case -ESRCH:
692                         /*
693                          * No owner found for this futex. Check if the
694                          * OWNER_DIED bit is set to figure out whether
695                          * this is a robust futex or not.
696                          */
697                         if (get_futex_value_locked(&curval, uaddr))
698                                 return -EFAULT;
699
700                         /*
701                          * We simply start over in case of a robust
702                          * futex. The code above will take the futex
703                          * and return happy.
704                          */
705                         if (curval & FUTEX_OWNER_DIED) {
706                                 ownerdied = 1;
707                                 goto retry;
708                         }
709                 default:
710                         break;
711                 }
712         }
713
714         return ret;
715 }
716
717 /*
718  * The hash bucket lock must be held when this is called.
719  * Afterwards, the futex_q must not be accessed.
720  */
721 static void wake_futex(struct futex_q *q)
722 {
723         struct task_struct *p = q->task;
724
725         /*
726          * We set q->lock_ptr = NULL _before_ we wake up the task. If
727          * a non futex wake up happens on another CPU then the task
728          * might exit and p would dereference a non existing task
729          * struct. Prevent this by holding a reference on p across the
730          * wake up.
731          */
732         get_task_struct(p);
733
734         plist_del(&q->list, &q->list.plist);
735         /*
736          * The waiting task can free the futex_q as soon as
737          * q->lock_ptr = NULL is written, without taking any locks. A
738          * memory barrier is required here to prevent the following
739          * store to lock_ptr from getting ahead of the plist_del.
740          */
741         smp_wmb();
742         q->lock_ptr = NULL;
743
744         wake_up_state(p, TASK_NORMAL);
745         put_task_struct(p);
746 }
747
748 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
749 {
750         struct task_struct *new_owner;
751         struct futex_pi_state *pi_state = this->pi_state;
752         u32 curval, newval;
753
754         if (!pi_state)
755                 return -EINVAL;
756
757         spin_lock(&pi_state->pi_mutex.wait_lock);
758         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
759
760         /*
761          * This happens when we have stolen the lock and the original
762          * pending owner did not enqueue itself back on the rt_mutex.
763          * Thats not a tragedy. We know that way, that a lock waiter
764          * is on the fly. We make the futex_q waiter the pending owner.
765          */
766         if (!new_owner)
767                 new_owner = this->task;
768
769         /*
770          * We pass it to the next owner. (The WAITERS bit is always
771          * kept enabled while there is PI state around. We must also
772          * preserve the owner died bit.)
773          */
774         if (!(uval & FUTEX_OWNER_DIED)) {
775                 int ret = 0;
776
777                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
778
779                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
780
781                 if (curval == -EFAULT)
782                         ret = -EFAULT;
783                 else if (curval != uval)
784                         ret = -EINVAL;
785                 if (ret) {
786                         spin_unlock(&pi_state->pi_mutex.wait_lock);
787                         return ret;
788                 }
789         }
790
791         spin_lock_irq(&pi_state->owner->pi_lock);
792         WARN_ON(list_empty(&pi_state->list));
793         list_del_init(&pi_state->list);
794         spin_unlock_irq(&pi_state->owner->pi_lock);
795
796         spin_lock_irq(&new_owner->pi_lock);
797         WARN_ON(!list_empty(&pi_state->list));
798         list_add(&pi_state->list, &new_owner->pi_state_list);
799         pi_state->owner = new_owner;
800         spin_unlock_irq(&new_owner->pi_lock);
801
802         spin_unlock(&pi_state->pi_mutex.wait_lock);
803         rt_mutex_unlock(&pi_state->pi_mutex);
804
805         return 0;
806 }
807
808 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
809 {
810         u32 oldval;
811
812         /*
813          * There is no waiter, so we unlock the futex. The owner died
814          * bit has not to be preserved here. We are the owner:
815          */
816         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
817
818         if (oldval == -EFAULT)
819                 return oldval;
820         if (oldval != uval)
821                 return -EAGAIN;
822
823         return 0;
824 }
825
826 /*
827  * Express the locking dependencies for lockdep:
828  */
829 static inline void
830 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
831 {
832         if (hb1 <= hb2) {
833                 spin_lock(&hb1->lock);
834                 if (hb1 < hb2)
835                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
836         } else { /* hb1 > hb2 */
837                 spin_lock(&hb2->lock);
838                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
839         }
840 }
841
842 static inline void
843 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
844 {
845         spin_unlock(&hb1->lock);
846         if (hb1 != hb2)
847                 spin_unlock(&hb2->lock);
848 }
849
850 /*
851  * Wake up waiters matching bitset queued on this futex (uaddr).
852  */
853 static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
854 {
855         struct futex_hash_bucket *hb;
856         struct futex_q *this, *next;
857         struct plist_head *head;
858         union futex_key key = FUTEX_KEY_INIT;
859         int ret;
860
861         if (!bitset)
862                 return -EINVAL;
863
864         ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
865         if (unlikely(ret != 0))
866                 goto out;
867
868         hb = hash_futex(&key);
869         spin_lock(&hb->lock);
870         head = &hb->chain;
871
872         plist_for_each_entry_safe(this, next, head, list) {
873                 if (match_futex (&this->key, &key)) {
874                         if (this->pi_state || this->rt_waiter) {
875                                 ret = -EINVAL;
876                                 break;
877                         }
878
879                         /* Check if one of the bits is set in both bitsets */
880                         if (!(this->bitset & bitset))
881                                 continue;
882
883                         wake_futex(this);
884                         if (++ret >= nr_wake)
885                                 break;
886                 }
887         }
888
889         spin_unlock(&hb->lock);
890         put_futex_key(fshared, &key);
891 out:
892         return ret;
893 }
894
895 /*
896  * Wake up all waiters hashed on the physical page that is mapped
897  * to this virtual address:
898  */
899 static int
900 futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
901               int nr_wake, int nr_wake2, int op)
902 {
903         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
904         struct futex_hash_bucket *hb1, *hb2;
905         struct plist_head *head;
906         struct futex_q *this, *next;
907         int ret, op_ret;
908
909 retry:
910         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
911         if (unlikely(ret != 0))
912                 goto out;
913         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
914         if (unlikely(ret != 0))
915                 goto out_put_key1;
916
917         hb1 = hash_futex(&key1);
918         hb2 = hash_futex(&key2);
919
920 retry_private:
921         double_lock_hb(hb1, hb2);
922         op_ret = futex_atomic_op_inuser(op, uaddr2);
923         if (unlikely(op_ret < 0)) {
924
925                 double_unlock_hb(hb1, hb2);
926
927 #ifndef CONFIG_MMU
928                 /*
929                  * we don't get EFAULT from MMU faults if we don't have an MMU,
930                  * but we might get them from range checking
931                  */
932                 ret = op_ret;
933                 goto out_put_keys;
934 #endif
935
936                 if (unlikely(op_ret != -EFAULT)) {
937                         ret = op_ret;
938                         goto out_put_keys;
939                 }
940
941                 ret = fault_in_user_writeable(uaddr2);
942                 if (ret)
943                         goto out_put_keys;
944
945                 if (!fshared)
946                         goto retry_private;
947
948                 put_futex_key(fshared, &key2);
949                 put_futex_key(fshared, &key1);
950                 goto retry;
951         }
952
953         head = &hb1->chain;
954
955         plist_for_each_entry_safe(this, next, head, list) {
956                 if (match_futex (&this->key, &key1)) {
957                         wake_futex(this);
958                         if (++ret >= nr_wake)
959                                 break;
960                 }
961         }
962
963         if (op_ret > 0) {
964                 head = &hb2->chain;
965
966                 op_ret = 0;
967                 plist_for_each_entry_safe(this, next, head, list) {
968                         if (match_futex (&this->key, &key2)) {
969                                 wake_futex(this);
970                                 if (++op_ret >= nr_wake2)
971                                         break;
972                         }
973                 }
974                 ret += op_ret;
975         }
976
977         double_unlock_hb(hb1, hb2);
978 out_put_keys:
979         put_futex_key(fshared, &key2);
980 out_put_key1:
981         put_futex_key(fshared, &key1);
982 out:
983         return ret;
984 }
985
986 /**
987  * requeue_futex() - Requeue a futex_q from one hb to another
988  * @q:          the futex_q to requeue
989  * @hb1:        the source hash_bucket
990  * @hb2:        the target hash_bucket
991  * @key2:       the new key for the requeued futex_q
992  */
993 static inline
994 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
995                    struct futex_hash_bucket *hb2, union futex_key *key2)
996 {
997
998         /*
999          * If key1 and key2 hash to the same bucket, no need to
1000          * requeue.
1001          */
1002         if (likely(&hb1->chain != &hb2->chain)) {
1003                 plist_del(&q->list, &hb1->chain);
1004                 plist_add(&q->list, &hb2->chain);
1005                 q->lock_ptr = &hb2->lock;
1006 #ifdef CONFIG_DEBUG_PI_LIST
1007                 q->list.plist.lock = &hb2->lock;
1008 #endif
1009         }
1010         get_futex_key_refs(key2);
1011         q->key = *key2;
1012 }
1013
1014 /**
1015  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1016  * @q:          the futex_q
1017  * @key:        the key of the requeue target futex
1018  * @hb:         the hash_bucket of the requeue target futex
1019  *
1020  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1021  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1022  * to the requeue target futex so the waiter can detect the wakeup on the right
1023  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1024  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1025  * to protect access to the pi_state to fixup the owner later.  Must be called
1026  * with both q->lock_ptr and hb->lock held.
1027  */
1028 static inline
1029 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1030                            struct futex_hash_bucket *hb)
1031 {
1032         get_futex_key_refs(key);
1033         q->key = *key;
1034
1035         WARN_ON(plist_node_empty(&q->list));
1036         plist_del(&q->list, &q->list.plist);
1037
1038         WARN_ON(!q->rt_waiter);
1039         q->rt_waiter = NULL;
1040
1041         q->lock_ptr = &hb->lock;
1042 #ifdef CONFIG_DEBUG_PI_LIST
1043         q->list.plist.lock = &hb->lock;
1044 #endif
1045
1046         wake_up_state(q->task, TASK_NORMAL);
1047 }
1048
1049 /**
1050  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1051  * @pifutex:            the user address of the to futex
1052  * @hb1:                the from futex hash bucket, must be locked by the caller
1053  * @hb2:                the to futex hash bucket, must be locked by the caller
1054  * @key1:               the from futex key
1055  * @key2:               the to futex key
1056  * @ps:                 address to store the pi_state pointer
1057  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1058  *
1059  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1060  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1061  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1062  * hb1 and hb2 must be held by the caller.
1063  *
1064  * Returns:
1065  *  0 - failed to acquire the lock atomicly
1066  *  1 - acquired the lock
1067  * <0 - error
1068  */
1069 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1070                                  struct futex_hash_bucket *hb1,
1071                                  struct futex_hash_bucket *hb2,
1072                                  union futex_key *key1, union futex_key *key2,
1073                                  struct futex_pi_state **ps, int set_waiters)
1074 {
1075         struct futex_q *top_waiter = NULL;
1076         u32 curval;
1077         int ret;
1078
1079         if (get_futex_value_locked(&curval, pifutex))
1080                 return -EFAULT;
1081
1082         /*
1083          * Find the top_waiter and determine if there are additional waiters.
1084          * If the caller intends to requeue more than 1 waiter to pifutex,
1085          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1086          * as we have means to handle the possible fault.  If not, don't set
1087          * the bit unecessarily as it will force the subsequent unlock to enter
1088          * the kernel.
1089          */
1090         top_waiter = futex_top_waiter(hb1, key1);
1091
1092         /* There are no waiters, nothing for us to do. */
1093         if (!top_waiter)
1094                 return 0;
1095
1096         /* Ensure we requeue to the expected futex. */
1097         if (!match_futex(top_waiter->requeue_pi_key, key2))
1098                 return -EINVAL;
1099
1100         /*
1101          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1102          * the contended case or if set_waiters is 1.  The pi_state is returned
1103          * in ps in contended cases.
1104          */
1105         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1106                                    set_waiters);
1107         if (ret == 1)
1108                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1109
1110         return ret;
1111 }
1112
1113 /**
1114  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1115  * uaddr1:      source futex user address
1116  * uaddr2:      target futex user address
1117  * nr_wake:     number of waiters to wake (must be 1 for requeue_pi)
1118  * nr_requeue:  number of waiters to requeue (0-INT_MAX)
1119  * requeue_pi:  if we are attempting to requeue from a non-pi futex to a
1120  *              pi futex (pi to pi requeue is not supported)
1121  *
1122  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1123  * uaddr2 atomically on behalf of the top waiter.
1124  *
1125  * Returns:
1126  * >=0 - on success, the number of tasks requeued or woken
1127  *  <0 - on error
1128  */
1129 static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1130                          int nr_wake, int nr_requeue, u32 *cmpval,
1131                          int requeue_pi)
1132 {
1133         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1134         int drop_count = 0, task_count = 0, ret;
1135         struct futex_pi_state *pi_state = NULL;
1136         struct futex_hash_bucket *hb1, *hb2;
1137         struct plist_head *head1;
1138         struct futex_q *this, *next;
1139         u32 curval2;
1140
1141         if (requeue_pi) {
1142                 /*
1143                  * requeue_pi requires a pi_state, try to allocate it now
1144                  * without any locks in case it fails.
1145                  */
1146                 if (refill_pi_state_cache())
1147                         return -ENOMEM;
1148                 /*
1149                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1150                  * + nr_requeue, since it acquires the rt_mutex prior to
1151                  * returning to userspace, so as to not leave the rt_mutex with
1152                  * waiters and no owner.  However, second and third wake-ups
1153                  * cannot be predicted as they involve race conditions with the
1154                  * first wake and a fault while looking up the pi_state.  Both
1155                  * pthread_cond_signal() and pthread_cond_broadcast() should
1156                  * use nr_wake=1.
1157                  */
1158                 if (nr_wake != 1)
1159                         return -EINVAL;
1160         }
1161
1162 retry:
1163         if (pi_state != NULL) {
1164                 /*
1165                  * We will have to lookup the pi_state again, so free this one
1166                  * to keep the accounting correct.
1167                  */
1168                 free_pi_state(pi_state);
1169                 pi_state = NULL;
1170         }
1171
1172         ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
1173         if (unlikely(ret != 0))
1174                 goto out;
1175         ret = get_futex_key(uaddr2, fshared, &key2,
1176                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1177         if (unlikely(ret != 0))
1178                 goto out_put_key1;
1179
1180         hb1 = hash_futex(&key1);
1181         hb2 = hash_futex(&key2);
1182
1183 retry_private:
1184         double_lock_hb(hb1, hb2);
1185
1186         if (likely(cmpval != NULL)) {
1187                 u32 curval;
1188
1189                 ret = get_futex_value_locked(&curval, uaddr1);
1190
1191                 if (unlikely(ret)) {
1192                         double_unlock_hb(hb1, hb2);
1193
1194                         ret = get_user(curval, uaddr1);
1195                         if (ret)
1196                                 goto out_put_keys;
1197
1198                         if (!fshared)
1199                                 goto retry_private;
1200
1201                         put_futex_key(fshared, &key2);
1202                         put_futex_key(fshared, &key1);
1203                         goto retry;
1204                 }
1205                 if (curval != *cmpval) {
1206                         ret = -EAGAIN;
1207                         goto out_unlock;
1208                 }
1209         }
1210
1211         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1212                 /*
1213                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1214                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1215                  * bit.  We force this here where we are able to easily handle
1216                  * faults rather in the requeue loop below.
1217                  */
1218                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1219                                                  &key2, &pi_state, nr_requeue);
1220
1221                 /*
1222                  * At this point the top_waiter has either taken uaddr2 or is
1223                  * waiting on it.  If the former, then the pi_state will not
1224                  * exist yet, look it up one more time to ensure we have a
1225                  * reference to it.
1226                  */
1227                 if (ret == 1) {
1228                         WARN_ON(pi_state);
1229                         drop_count++;
1230                         task_count++;
1231                         ret = get_futex_value_locked(&curval2, uaddr2);
1232                         if (!ret)
1233                                 ret = lookup_pi_state(curval2, hb2, &key2,
1234                                                       &pi_state);
1235                 }
1236
1237                 switch (ret) {
1238                 case 0:
1239                         break;
1240                 case -EFAULT:
1241                         double_unlock_hb(hb1, hb2);
1242                         put_futex_key(fshared, &key2);
1243                         put_futex_key(fshared, &key1);
1244                         ret = fault_in_user_writeable(uaddr2);
1245                         if (!ret)
1246                                 goto retry;
1247                         goto out;
1248                 case -EAGAIN:
1249                         /* The owner was exiting, try again. */
1250                         double_unlock_hb(hb1, hb2);
1251                         put_futex_key(fshared, &key2);
1252                         put_futex_key(fshared, &key1);
1253                         cond_resched();
1254                         goto retry;
1255                 default:
1256                         goto out_unlock;
1257                 }
1258         }
1259
1260         head1 = &hb1->chain;
1261         plist_for_each_entry_safe(this, next, head1, list) {
1262                 if (task_count - nr_wake >= nr_requeue)
1263                         break;
1264
1265                 if (!match_futex(&this->key, &key1))
1266                         continue;
1267
1268                 /*
1269                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1270                  * be paired with each other and no other futex ops.
1271                  */
1272                 if ((requeue_pi && !this->rt_waiter) ||
1273                     (!requeue_pi && this->rt_waiter)) {
1274                         ret = -EINVAL;
1275                         break;
1276                 }
1277
1278                 /*
1279                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1280                  * lock, we already woke the top_waiter.  If not, it will be
1281                  * woken by futex_unlock_pi().
1282                  */
1283                 if (++task_count <= nr_wake && !requeue_pi) {
1284                         wake_futex(this);
1285                         continue;
1286                 }
1287
1288                 /* Ensure we requeue to the expected futex for requeue_pi. */
1289                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1290                         ret = -EINVAL;
1291                         break;
1292                 }
1293
1294                 /*
1295                  * Requeue nr_requeue waiters and possibly one more in the case
1296                  * of requeue_pi if we couldn't acquire the lock atomically.
1297                  */
1298                 if (requeue_pi) {
1299                         /* Prepare the waiter to take the rt_mutex. */
1300                         atomic_inc(&pi_state->refcount);
1301                         this->pi_state = pi_state;
1302                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1303                                                         this->rt_waiter,
1304                                                         this->task, 1);
1305                         if (ret == 1) {
1306                                 /* We got the lock. */
1307                                 requeue_pi_wake_futex(this, &key2, hb2);
1308                                 drop_count++;
1309                                 continue;
1310                         } else if (ret) {
1311                                 /* -EDEADLK */
1312                                 this->pi_state = NULL;
1313                                 free_pi_state(pi_state);
1314                                 goto out_unlock;
1315                         }
1316                 }
1317                 requeue_futex(this, hb1, hb2, &key2);
1318                 drop_count++;
1319         }
1320
1321 out_unlock:
1322         double_unlock_hb(hb1, hb2);
1323
1324         /*
1325          * drop_futex_key_refs() must be called outside the spinlocks. During
1326          * the requeue we moved futex_q's from the hash bucket at key1 to the
1327          * one at key2 and updated their key pointer.  We no longer need to
1328          * hold the references to key1.
1329          */
1330         while (--drop_count >= 0)
1331                 drop_futex_key_refs(&key1);
1332
1333 out_put_keys:
1334         put_futex_key(fshared, &key2);
1335 out_put_key1:
1336         put_futex_key(fshared, &key1);
1337 out:
1338         if (pi_state != NULL)
1339                 free_pi_state(pi_state);
1340         return ret ? ret : task_count;
1341 }
1342
1343 /* The key must be already stored in q->key. */
1344 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1345 {
1346         struct futex_hash_bucket *hb;
1347
1348         get_futex_key_refs(&q->key);
1349         hb = hash_futex(&q->key);
1350         q->lock_ptr = &hb->lock;
1351
1352         spin_lock(&hb->lock);
1353         return hb;
1354 }
1355
1356 static inline void
1357 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1358 {
1359         spin_unlock(&hb->lock);
1360         drop_futex_key_refs(&q->key);
1361 }
1362
1363 /**
1364  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1365  * @q:  The futex_q to enqueue
1366  * @hb: The destination hash bucket
1367  *
1368  * The hb->lock must be held by the caller, and is released here. A call to
1369  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1370  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1371  * or nothing if the unqueue is done as part of the wake process and the unqueue
1372  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1373  * an example).
1374  */
1375 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1376 {
1377         int prio;
1378
1379         /*
1380          * The priority used to register this element is
1381          * - either the real thread-priority for the real-time threads
1382          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1383          * - or MAX_RT_PRIO for non-RT threads.
1384          * Thus, all RT-threads are woken first in priority order, and
1385          * the others are woken last, in FIFO order.
1386          */
1387         prio = min(current->normal_prio, MAX_RT_PRIO);
1388
1389         plist_node_init(&q->list, prio);
1390 #ifdef CONFIG_DEBUG_PI_LIST
1391         q->list.plist.lock = &hb->lock;
1392 #endif
1393         plist_add(&q->list, &hb->chain);
1394         q->task = current;
1395         spin_unlock(&hb->lock);
1396 }
1397
1398 /**
1399  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1400  * @q:  The futex_q to unqueue
1401  *
1402  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1403  * be paired with exactly one earlier call to queue_me().
1404  *
1405  * Returns:
1406  *   1 - if the futex_q was still queued (and we removed unqueued it)
1407  *   0 - if the futex_q was already removed by the waking thread
1408  */
1409 static int unqueue_me(struct futex_q *q)
1410 {
1411         spinlock_t *lock_ptr;
1412         int ret = 0;
1413
1414         /* In the common case we don't take the spinlock, which is nice. */
1415 retry:
1416         lock_ptr = q->lock_ptr;
1417         barrier();
1418         if (lock_ptr != NULL) {
1419                 spin_lock(lock_ptr);
1420                 /*
1421                  * q->lock_ptr can change between reading it and
1422                  * spin_lock(), causing us to take the wrong lock.  This
1423                  * corrects the race condition.
1424                  *
1425                  * Reasoning goes like this: if we have the wrong lock,
1426                  * q->lock_ptr must have changed (maybe several times)
1427                  * between reading it and the spin_lock().  It can
1428                  * change again after the spin_lock() but only if it was
1429                  * already changed before the spin_lock().  It cannot,
1430                  * however, change back to the original value.  Therefore
1431                  * we can detect whether we acquired the correct lock.
1432                  */
1433                 if (unlikely(lock_ptr != q->lock_ptr)) {
1434                         spin_unlock(lock_ptr);
1435                         goto retry;
1436                 }
1437                 WARN_ON(plist_node_empty(&q->list));
1438                 plist_del(&q->list, &q->list.plist);
1439
1440                 BUG_ON(q->pi_state);
1441
1442                 spin_unlock(lock_ptr);
1443                 ret = 1;
1444         }
1445
1446         drop_futex_key_refs(&q->key);
1447         return ret;
1448 }
1449
1450 /*
1451  * PI futexes can not be requeued and must remove themself from the
1452  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1453  * and dropped here.
1454  */
1455 static void unqueue_me_pi(struct futex_q *q)
1456 {
1457         WARN_ON(plist_node_empty(&q->list));
1458         plist_del(&q->list, &q->list.plist);
1459
1460         BUG_ON(!q->pi_state);
1461         free_pi_state(q->pi_state);
1462         q->pi_state = NULL;
1463
1464         spin_unlock(q->lock_ptr);
1465
1466         drop_futex_key_refs(&q->key);
1467 }
1468
1469 /*
1470  * Fixup the pi_state owner with the new owner.
1471  *
1472  * Must be called with hash bucket lock held and mm->sem held for non
1473  * private futexes.
1474  */
1475 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1476                                 struct task_struct *newowner, int fshared)
1477 {
1478         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1479         struct futex_pi_state *pi_state = q->pi_state;
1480         struct task_struct *oldowner = pi_state->owner;
1481         u32 uval, curval, newval;
1482         int ret;
1483
1484         /* Owner died? */
1485         if (!pi_state->owner)
1486                 newtid |= FUTEX_OWNER_DIED;
1487
1488         /*
1489          * We are here either because we stole the rtmutex from the
1490          * pending owner or we are the pending owner which failed to
1491          * get the rtmutex. We have to replace the pending owner TID
1492          * in the user space variable. This must be atomic as we have
1493          * to preserve the owner died bit here.
1494          *
1495          * Note: We write the user space value _before_ changing the pi_state
1496          * because we can fault here. Imagine swapped out pages or a fork
1497          * that marked all the anonymous memory readonly for cow.
1498          *
1499          * Modifying pi_state _before_ the user space value would
1500          * leave the pi_state in an inconsistent state when we fault
1501          * here, because we need to drop the hash bucket lock to
1502          * handle the fault. This might be observed in the PID check
1503          * in lookup_pi_state.
1504          */
1505 retry:
1506         if (get_futex_value_locked(&uval, uaddr))
1507                 goto handle_fault;
1508
1509         while (1) {
1510                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1511
1512                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1513
1514                 if (curval == -EFAULT)
1515                         goto handle_fault;
1516                 if (curval == uval)
1517                         break;
1518                 uval = curval;
1519         }
1520
1521         /*
1522          * We fixed up user space. Now we need to fix the pi_state
1523          * itself.
1524          */
1525         if (pi_state->owner != NULL) {
1526                 spin_lock_irq(&pi_state->owner->pi_lock);
1527                 WARN_ON(list_empty(&pi_state->list));
1528                 list_del_init(&pi_state->list);
1529                 spin_unlock_irq(&pi_state->owner->pi_lock);
1530         }
1531
1532         pi_state->owner = newowner;
1533
1534         spin_lock_irq(&newowner->pi_lock);
1535         WARN_ON(!list_empty(&pi_state->list));
1536         list_add(&pi_state->list, &newowner->pi_state_list);
1537         spin_unlock_irq(&newowner->pi_lock);
1538         return 0;
1539
1540         /*
1541          * To handle the page fault we need to drop the hash bucket
1542          * lock here. That gives the other task (either the pending
1543          * owner itself or the task which stole the rtmutex) the
1544          * chance to try the fixup of the pi_state. So once we are
1545          * back from handling the fault we need to check the pi_state
1546          * after reacquiring the hash bucket lock and before trying to
1547          * do another fixup. When the fixup has been done already we
1548          * simply return.
1549          */
1550 handle_fault:
1551         spin_unlock(q->lock_ptr);
1552
1553         ret = fault_in_user_writeable(uaddr);
1554
1555         spin_lock(q->lock_ptr);
1556
1557         /*
1558          * Check if someone else fixed it for us:
1559          */
1560         if (pi_state->owner != oldowner)
1561                 return 0;
1562
1563         if (ret)
1564                 return ret;
1565
1566         goto retry;
1567 }
1568
1569 /*
1570  * In case we must use restart_block to restart a futex_wait,
1571  * we encode in the 'flags' shared capability
1572  */
1573 #define FLAGS_SHARED            0x01
1574 #define FLAGS_CLOCKRT           0x02
1575 #define FLAGS_HAS_TIMEOUT       0x04
1576
1577 static long futex_wait_restart(struct restart_block *restart);
1578
1579 /**
1580  * fixup_owner() - Post lock pi_state and corner case management
1581  * @uaddr:      user address of the futex
1582  * @fshared:    whether the futex is shared (1) or not (0)
1583  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1584  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1585  *
1586  * After attempting to lock an rt_mutex, this function is called to cleanup
1587  * the pi_state owner as well as handle race conditions that may allow us to
1588  * acquire the lock. Must be called with the hb lock held.
1589  *
1590  * Returns:
1591  *  1 - success, lock taken
1592  *  0 - success, lock not taken
1593  * <0 - on error (-EFAULT)
1594  */
1595 static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1596                        int locked)
1597 {
1598         struct task_struct *owner;
1599         int ret = 0;
1600
1601         if (locked) {
1602                 /*
1603                  * Got the lock. We might not be the anticipated owner if we
1604                  * did a lock-steal - fix up the PI-state in that case:
1605                  */
1606                 if (q->pi_state->owner != current)
1607                         ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1608                 goto out;
1609         }
1610
1611         /*
1612          * Catch the rare case, where the lock was released when we were on the
1613          * way back before we locked the hash bucket.
1614          */
1615         if (q->pi_state->owner == current) {
1616                 /*
1617                  * Try to get the rt_mutex now. This might fail as some other
1618                  * task acquired the rt_mutex after we removed ourself from the
1619                  * rt_mutex waiters list.
1620                  */
1621                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1622                         locked = 1;
1623                         goto out;
1624                 }
1625
1626                 /*
1627                  * pi_state is incorrect, some other task did a lock steal and
1628                  * we returned due to timeout or signal without taking the
1629                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1630                  * locking, as the other task is now blocked on the hash bucket
1631                  * lock. Fix the state up.
1632                  */
1633                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1634                 ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1635                 goto out;
1636         }
1637
1638         /*
1639          * Paranoia check. If we did not take the lock, then we should not be
1640          * the owner, nor the pending owner, of the rt_mutex.
1641          */
1642         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1643                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1644                                 "pi-state %p\n", ret,
1645                                 q->pi_state->pi_mutex.owner,
1646                                 q->pi_state->owner);
1647
1648 out:
1649         return ret ? ret : locked;
1650 }
1651
1652 /**
1653  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1654  * @hb:         the futex hash bucket, must be locked by the caller
1655  * @q:          the futex_q to queue up on
1656  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1657  */
1658 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1659                                 struct hrtimer_sleeper *timeout)
1660 {
1661         /*
1662          * The task state is guaranteed to be set before another task can
1663          * wake it. set_current_state() is implemented using set_mb() and
1664          * queue_me() calls spin_unlock() upon completion, both serializing
1665          * access to the hash list and forcing another memory barrier.
1666          */
1667         set_current_state(TASK_INTERRUPTIBLE);
1668         queue_me(q, hb);
1669
1670         /* Arm the timer */
1671         if (timeout) {
1672                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1673                 if (!hrtimer_active(&timeout->timer))
1674                         timeout->task = NULL;
1675         }
1676
1677         /*
1678          * If we have been removed from the hash list, then another task
1679          * has tried to wake us, and we can skip the call to schedule().
1680          */
1681         if (likely(!plist_node_empty(&q->list))) {
1682                 /*
1683                  * If the timer has already expired, current will already be
1684                  * flagged for rescheduling. Only call schedule if there
1685                  * is no timeout, or if it has yet to expire.
1686                  */
1687                 if (!timeout || timeout->task)
1688                         schedule();
1689         }
1690         __set_current_state(TASK_RUNNING);
1691 }
1692
1693 /**
1694  * futex_wait_setup() - Prepare to wait on a futex
1695  * @uaddr:      the futex userspace address
1696  * @val:        the expected value
1697  * @fshared:    whether the futex is shared (1) or not (0)
1698  * @q:          the associated futex_q
1699  * @hb:         storage for hash_bucket pointer to be returned to caller
1700  *
1701  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1702  * compare it with the expected value.  Handle atomic faults internally.
1703  * Return with the hb lock held and a q.key reference on success, and unlocked
1704  * with no q.key reference on failure.
1705  *
1706  * Returns:
1707  *  0 - uaddr contains val and hb has been locked
1708  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1709  */
1710 static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1711                            struct futex_q *q, struct futex_hash_bucket **hb)
1712 {
1713         u32 uval;
1714         int ret;
1715
1716         /*
1717          * Access the page AFTER the hash-bucket is locked.
1718          * Order is important:
1719          *
1720          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1721          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1722          *
1723          * The basic logical guarantee of a futex is that it blocks ONLY
1724          * if cond(var) is known to be true at the time of blocking, for
1725          * any cond.  If we queued after testing *uaddr, that would open
1726          * a race condition where we could block indefinitely with
1727          * cond(var) false, which would violate the guarantee.
1728          *
1729          * A consequence is that futex_wait() can return zero and absorb
1730          * a wakeup when *uaddr != val on entry to the syscall.  This is
1731          * rare, but normal.
1732          */
1733 retry:
1734         q->key = FUTEX_KEY_INIT;
1735         ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
1736         if (unlikely(ret != 0))
1737                 return ret;
1738
1739 retry_private:
1740         *hb = queue_lock(q);
1741
1742         ret = get_futex_value_locked(&uval, uaddr);
1743
1744         if (ret) {
1745                 queue_unlock(q, *hb);
1746
1747                 ret = get_user(uval, uaddr);
1748                 if (ret)
1749                         goto out;
1750
1751                 if (!fshared)
1752                         goto retry_private;
1753
1754                 put_futex_key(fshared, &q->key);
1755                 goto retry;
1756         }
1757
1758         if (uval != val) {
1759                 queue_unlock(q, *hb);
1760                 ret = -EWOULDBLOCK;
1761         }
1762
1763 out:
1764         if (ret)
1765                 put_futex_key(fshared, &q->key);
1766         return ret;
1767 }
1768
1769 static int futex_wait(u32 __user *uaddr, int fshared,
1770                       u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1771 {
1772         struct hrtimer_sleeper timeout, *to = NULL;
1773         struct restart_block *restart;
1774         struct futex_hash_bucket *hb;
1775         struct futex_q q;
1776         int ret;
1777
1778         if (!bitset)
1779                 return -EINVAL;
1780
1781         q.pi_state = NULL;
1782         q.bitset = bitset;
1783         q.rt_waiter = NULL;
1784         q.requeue_pi_key = NULL;
1785
1786         if (abs_time) {
1787                 to = &timeout;
1788
1789                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1790                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1791                 hrtimer_init_sleeper(to, current);
1792                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1793                                              current->timer_slack_ns);
1794         }
1795
1796 retry:
1797         /* Prepare to wait on uaddr. */
1798         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1799         if (ret)
1800                 goto out;
1801
1802         /* queue_me and wait for wakeup, timeout, or a signal. */
1803         futex_wait_queue_me(hb, &q, to);
1804
1805         /* If we were woken (and unqueued), we succeeded, whatever. */
1806         ret = 0;
1807         if (!unqueue_me(&q))
1808                 goto out_put_key;
1809         ret = -ETIMEDOUT;
1810         if (to && !to->task)
1811                 goto out_put_key;
1812
1813         /*
1814          * We expect signal_pending(current), but we might be the
1815          * victim of a spurious wakeup as well.
1816          */
1817         if (!signal_pending(current)) {
1818                 put_futex_key(fshared, &q.key);
1819                 goto retry;
1820         }
1821
1822         ret = -ERESTARTSYS;
1823         if (!abs_time)
1824                 goto out_put_key;
1825
1826         restart = &current_thread_info()->restart_block;
1827         restart->fn = futex_wait_restart;
1828         restart->futex.uaddr = (u32 *)uaddr;
1829         restart->futex.val = val;
1830         restart->futex.time = abs_time->tv64;
1831         restart->futex.bitset = bitset;
1832         restart->futex.flags = FLAGS_HAS_TIMEOUT;
1833
1834         if (fshared)
1835                 restart->futex.flags |= FLAGS_SHARED;
1836         if (clockrt)
1837                 restart->futex.flags |= FLAGS_CLOCKRT;
1838
1839         ret = -ERESTART_RESTARTBLOCK;
1840
1841 out_put_key:
1842         put_futex_key(fshared, &q.key);
1843 out:
1844         if (to) {
1845                 hrtimer_cancel(&to->timer);
1846                 destroy_hrtimer_on_stack(&to->timer);
1847         }
1848         return ret;
1849 }
1850
1851
1852 static long futex_wait_restart(struct restart_block *restart)
1853 {
1854         u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1855         int fshared = 0;
1856         ktime_t t, *tp = NULL;
1857
1858         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1859                 t.tv64 = restart->futex.time;
1860                 tp = &t;
1861         }
1862         restart->fn = do_no_restart_syscall;
1863         if (restart->futex.flags & FLAGS_SHARED)
1864                 fshared = 1;
1865         return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1866                                 restart->futex.bitset,
1867                                 restart->futex.flags & FLAGS_CLOCKRT);
1868 }
1869
1870
1871 /*
1872  * Userspace tried a 0 -> TID atomic transition of the futex value
1873  * and failed. The kernel side here does the whole locking operation:
1874  * if there are waiters then it will block, it does PI, etc. (Due to
1875  * races the kernel might see a 0 value of the futex too.)
1876  */
1877 static int futex_lock_pi(u32 __user *uaddr, int fshared,
1878                          int detect, ktime_t *time, int trylock)
1879 {
1880         struct hrtimer_sleeper timeout, *to = NULL;
1881         struct futex_hash_bucket *hb;
1882         struct futex_q q;
1883         int res, ret;
1884
1885         if (refill_pi_state_cache())
1886                 return -ENOMEM;
1887
1888         if (time) {
1889                 to = &timeout;
1890                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1891                                       HRTIMER_MODE_ABS);
1892                 hrtimer_init_sleeper(to, current);
1893                 hrtimer_set_expires(&to->timer, *time);
1894         }
1895
1896         q.pi_state = NULL;
1897         q.rt_waiter = NULL;
1898         q.requeue_pi_key = NULL;
1899 retry:
1900         q.key = FUTEX_KEY_INIT;
1901         ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
1902         if (unlikely(ret != 0))
1903                 goto out;
1904
1905 retry_private:
1906         hb = queue_lock(&q);
1907
1908         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1909         if (unlikely(ret)) {
1910                 switch (ret) {
1911                 case 1:
1912                         /* We got the lock. */
1913                         ret = 0;
1914                         goto out_unlock_put_key;
1915                 case -EFAULT:
1916                         goto uaddr_faulted;
1917                 case -EAGAIN:
1918                         /*
1919                          * Task is exiting and we just wait for the
1920                          * exit to complete.
1921                          */
1922                         queue_unlock(&q, hb);
1923                         put_futex_key(fshared, &q.key);
1924                         cond_resched();
1925                         goto retry;
1926                 default:
1927                         goto out_unlock_put_key;
1928                 }
1929         }
1930
1931         /*
1932          * Only actually queue now that the atomic ops are done:
1933          */
1934         queue_me(&q, hb);
1935
1936         WARN_ON(!q.pi_state);
1937         /*
1938          * Block on the PI mutex:
1939          */
1940         if (!trylock)
1941                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1942         else {
1943                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1944                 /* Fixup the trylock return value: */
1945                 ret = ret ? 0 : -EWOULDBLOCK;
1946         }
1947
1948         spin_lock(q.lock_ptr);
1949         /*
1950          * Fixup the pi_state owner and possibly acquire the lock if we
1951          * haven't already.
1952          */
1953         res = fixup_owner(uaddr, fshared, &q, !ret);
1954         /*
1955          * If fixup_owner() returned an error, proprogate that.  If it acquired
1956          * the lock, clear our -ETIMEDOUT or -EINTR.
1957          */
1958         if (res)
1959                 ret = (res < 0) ? res : 0;
1960
1961         /*
1962          * If fixup_owner() faulted and was unable to handle the fault, unlock
1963          * it and return the fault to userspace.
1964          */
1965         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1966                 rt_mutex_unlock(&q.pi_state->pi_mutex);
1967
1968         /* Unqueue and drop the lock */
1969         unqueue_me_pi(&q);
1970
1971         goto out;
1972
1973 out_unlock_put_key:
1974         queue_unlock(&q, hb);
1975
1976 out_put_key:
1977         put_futex_key(fshared, &q.key);
1978 out:
1979         if (to)
1980                 destroy_hrtimer_on_stack(&to->timer);
1981         return ret != -EINTR ? ret : -ERESTARTNOINTR;
1982
1983 uaddr_faulted:
1984         queue_unlock(&q, hb);
1985
1986         ret = fault_in_user_writeable(uaddr);
1987         if (ret)
1988                 goto out_put_key;
1989
1990         if (!fshared)
1991                 goto retry_private;
1992
1993         put_futex_key(fshared, &q.key);
1994         goto retry;
1995 }
1996
1997 /*
1998  * Userspace attempted a TID -> 0 atomic transition, and failed.
1999  * This is the in-kernel slowpath: we look up the PI state (if any),
2000  * and do the rt-mutex unlock.
2001  */
2002 static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2003 {
2004         struct futex_hash_bucket *hb;
2005         struct futex_q *this, *next;
2006         u32 uval;
2007         struct plist_head *head;
2008         union futex_key key = FUTEX_KEY_INIT;
2009         int ret;
2010
2011 retry:
2012         if (get_user(uval, uaddr))
2013                 return -EFAULT;
2014         /*
2015          * We release only a lock we actually own:
2016          */
2017         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2018                 return -EPERM;
2019
2020         ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
2021         if (unlikely(ret != 0))
2022                 goto out;
2023
2024         hb = hash_futex(&key);
2025         spin_lock(&hb->lock);
2026
2027         /*
2028          * To avoid races, try to do the TID -> 0 atomic transition
2029          * again. If it succeeds then we can return without waking
2030          * anyone else up:
2031          */
2032         if (!(uval & FUTEX_OWNER_DIED))
2033                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2034
2035
2036         if (unlikely(uval == -EFAULT))
2037                 goto pi_faulted;
2038         /*
2039          * Rare case: we managed to release the lock atomically,
2040          * no need to wake anyone else up:
2041          */
2042         if (unlikely(uval == task_pid_vnr(current)))
2043                 goto out_unlock;
2044
2045         /*
2046          * Ok, other tasks may need to be woken up - check waiters
2047          * and do the wakeup if necessary:
2048          */
2049         head = &hb->chain;
2050
2051         plist_for_each_entry_safe(this, next, head, list) {
2052                 if (!match_futex (&this->key, &key))
2053                         continue;
2054                 ret = wake_futex_pi(uaddr, uval, this);
2055                 /*
2056                  * The atomic access to the futex value
2057                  * generated a pagefault, so retry the
2058                  * user-access and the wakeup:
2059                  */
2060                 if (ret == -EFAULT)
2061                         goto pi_faulted;
2062                 goto out_unlock;
2063         }
2064         /*
2065          * No waiters - kernel unlocks the futex:
2066          */
2067         if (!(uval & FUTEX_OWNER_DIED)) {
2068                 ret = unlock_futex_pi(uaddr, uval);
2069                 if (ret == -EFAULT)
2070                         goto pi_faulted;
2071         }
2072
2073 out_unlock:
2074         spin_unlock(&hb->lock);
2075         put_futex_key(fshared, &key);
2076
2077 out:
2078         return ret;
2079
2080 pi_faulted:
2081         spin_unlock(&hb->lock);
2082         put_futex_key(fshared, &key);
2083
2084         ret = fault_in_user_writeable(uaddr);
2085         if (!ret)
2086                 goto retry;
2087
2088         return ret;
2089 }
2090
2091 /**
2092  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2093  * @hb:         the hash_bucket futex_q was original enqueued on
2094  * @q:          the futex_q woken while waiting to be requeued
2095  * @key2:       the futex_key of the requeue target futex
2096  * @timeout:    the timeout associated with the wait (NULL if none)
2097  *
2098  * Detect if the task was woken on the initial futex as opposed to the requeue
2099  * target futex.  If so, determine if it was a timeout or a signal that caused
2100  * the wakeup and return the appropriate error code to the caller.  Must be
2101  * called with the hb lock held.
2102  *
2103  * Returns
2104  *  0 - no early wakeup detected
2105  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2106  */
2107 static inline
2108 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2109                                    struct futex_q *q, union futex_key *key2,
2110                                    struct hrtimer_sleeper *timeout)
2111 {
2112         int ret = 0;
2113
2114         /*
2115          * With the hb lock held, we avoid races while we process the wakeup.
2116          * We only need to hold hb (and not hb2) to ensure atomicity as the
2117          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2118          * It can't be requeued from uaddr2 to something else since we don't
2119          * support a PI aware source futex for requeue.
2120          */
2121         if (!match_futex(&q->key, key2)) {
2122                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2123                 /*
2124                  * We were woken prior to requeue by a timeout or a signal.
2125                  * Unqueue the futex_q and determine which it was.
2126                  */
2127                 plist_del(&q->list, &q->list.plist);
2128
2129                 /* Handle spurious wakeups gracefully */
2130                 ret = -EWOULDBLOCK;
2131                 if (timeout && !timeout->task)
2132                         ret = -ETIMEDOUT;
2133                 else if (signal_pending(current))
2134                         ret = -ERESTARTNOINTR;
2135         }
2136         return ret;
2137 }
2138
2139 /**
2140  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2141  * @uaddr:      the futex we initially wait on (non-pi)
2142  * @fshared:    whether the futexes are shared (1) or not (0).  They must be
2143  *              the same type, no requeueing from private to shared, etc.
2144  * @val:        the expected value of uaddr
2145  * @abs_time:   absolute timeout
2146  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2147  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2148  * @uaddr2:     the pi futex we will take prior to returning to user-space
2149  *
2150  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2151  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2152  * complete the acquisition of the rt_mutex prior to returning to userspace.
2153  * This ensures the rt_mutex maintains an owner when it has waiters; without
2154  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2155  * need to.
2156  *
2157  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2158  * via the following:
2159  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2160  * 2) wakeup on uaddr2 after a requeue
2161  * 3) signal
2162  * 4) timeout
2163  *
2164  * If 3, cleanup and return -ERESTARTNOINTR.
2165  *
2166  * If 2, we may then block on trying to take the rt_mutex and return via:
2167  * 5) successful lock
2168  * 6) signal
2169  * 7) timeout
2170  * 8) other lock acquisition failure
2171  *
2172  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2173  *
2174  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2175  *
2176  * Returns:
2177  *  0 - On success
2178  * <0 - On error
2179  */
2180 static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2181                                  u32 val, ktime_t *abs_time, u32 bitset,
2182                                  int clockrt, u32 __user *uaddr2)
2183 {
2184         struct hrtimer_sleeper timeout, *to = NULL;
2185         struct rt_mutex_waiter rt_waiter;
2186         struct rt_mutex *pi_mutex = NULL;
2187         struct futex_hash_bucket *hb;
2188         union futex_key key2;
2189         struct futex_q q;
2190         int res, ret;
2191
2192         if (!bitset)
2193                 return -EINVAL;
2194
2195         if (abs_time) {
2196                 to = &timeout;
2197                 hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2198                                       CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2199                 hrtimer_init_sleeper(to, current);
2200                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2201                                              current->timer_slack_ns);
2202         }
2203
2204         /*
2205          * The waiter is allocated on our stack, manipulated by the requeue
2206          * code while we sleep on uaddr.
2207          */
2208         debug_rt_mutex_init_waiter(&rt_waiter);
2209         rt_waiter.task = NULL;
2210
2211         key2 = FUTEX_KEY_INIT;
2212         ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
2213         if (unlikely(ret != 0))
2214                 goto out;
2215
2216         q.pi_state = NULL;
2217         q.bitset = bitset;
2218         q.rt_waiter = &rt_waiter;
2219         q.requeue_pi_key = &key2;
2220
2221         /* Prepare to wait on uaddr. */
2222         ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2223         if (ret)
2224                 goto out_key2;
2225
2226         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2227         futex_wait_queue_me(hb, &q, to);
2228
2229         spin_lock(&hb->lock);
2230         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2231         spin_unlock(&hb->lock);
2232         if (ret)
2233                 goto out_put_keys;
2234
2235         /*
2236          * In order for us to be here, we know our q.key == key2, and since
2237          * we took the hb->lock above, we also know that futex_requeue() has
2238          * completed and we no longer have to concern ourselves with a wakeup
2239          * race with the atomic proxy lock acquition by the requeue code.
2240          */
2241
2242         /* Check if the requeue code acquired the second futex for us. */
2243         if (!q.rt_waiter) {
2244                 /*
2245                  * Got the lock. We might not be the anticipated owner if we
2246                  * did a lock-steal - fix up the PI-state in that case.
2247                  */
2248                 if (q.pi_state && (q.pi_state->owner != current)) {
2249                         spin_lock(q.lock_ptr);
2250                         ret = fixup_pi_state_owner(uaddr2, &q, current,
2251                                                    fshared);
2252                         spin_unlock(q.lock_ptr);
2253                 }
2254         } else {
2255                 /*
2256                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2257                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2258                  * the pi_state.
2259                  */
2260                 WARN_ON(!&q.pi_state);
2261                 pi_mutex = &q.pi_state->pi_mutex;
2262                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2263                 debug_rt_mutex_free_waiter(&rt_waiter);
2264
2265                 spin_lock(q.lock_ptr);
2266                 /*
2267                  * Fixup the pi_state owner and possibly acquire the lock if we
2268                  * haven't already.
2269                  */
2270                 res = fixup_owner(uaddr2, fshared, &q, !ret);
2271                 /*
2272                  * If fixup_owner() returned an error, proprogate that.  If it
2273                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2274                  */
2275                 if (res)
2276                         ret = (res < 0) ? res : 0;
2277
2278                 /* Unqueue and drop the lock. */
2279                 unqueue_me_pi(&q);
2280         }
2281
2282         /*
2283          * If fixup_pi_state_owner() faulted and was unable to handle the
2284          * fault, unlock the rt_mutex and return the fault to userspace.
2285          */
2286         if (ret == -EFAULT) {
2287                 if (rt_mutex_owner(pi_mutex) == current)
2288                         rt_mutex_unlock(pi_mutex);
2289         } else if (ret == -EINTR) {
2290                 /*
2291                  * We've already been requeued, but cannot restart by calling
2292                  * futex_lock_pi() directly. We could restart this syscall, but
2293                  * it would detect that the user space "val" changed and return
2294                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2295                  * -EWOULDBLOCK directly.
2296                  */
2297                 ret = -EWOULDBLOCK;
2298         }
2299
2300 out_put_keys:
2301         put_futex_key(fshared, &q.key);
2302 out_key2:
2303         put_futex_key(fshared, &key2);
2304
2305 out:
2306         if (to) {
2307                 hrtimer_cancel(&to->timer);
2308                 destroy_hrtimer_on_stack(&to->timer);
2309         }
2310         return ret;
2311 }
2312
2313 /*
2314  * Support for robust futexes: the kernel cleans up held futexes at
2315  * thread exit time.
2316  *
2317  * Implementation: user-space maintains a per-thread list of locks it
2318  * is holding. Upon do_exit(), the kernel carefully walks this list,
2319  * and marks all locks that are owned by this thread with the
2320  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2321  * always manipulated with the lock held, so the list is private and
2322  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2323  * field, to allow the kernel to clean up if the thread dies after
2324  * acquiring the lock, but just before it could have added itself to
2325  * the list. There can only be one such pending lock.
2326  */
2327
2328 /**
2329  * sys_set_robust_list() - Set the robust-futex list head of a task
2330  * @head:       pointer to the list-head
2331  * @len:        length of the list-head, as userspace expects
2332  */
2333 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2334                 size_t, len)
2335 {
2336         if (!futex_cmpxchg_enabled)
2337                 return -ENOSYS;
2338         /*
2339          * The kernel knows only one size for now:
2340          */
2341         if (unlikely(len != sizeof(*head)))
2342                 return -EINVAL;
2343
2344         current->robust_list = head;
2345
2346         return 0;
2347 }
2348
2349 /**
2350  * sys_get_robust_list() - Get the robust-futex list head of a task
2351  * @pid:        pid of the process [zero for current task]
2352  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2353  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2354  */
2355 SYSCALL_DEFINE3(get_robust_list, int, pid,
2356                 struct robust_list_head __user * __user *, head_ptr,
2357                 size_t __user *, len_ptr)
2358 {
2359         struct robust_list_head __user *head;
2360         unsigned long ret;
2361         const struct cred *cred = current_cred(), *pcred;
2362
2363         if (!futex_cmpxchg_enabled)
2364                 return -ENOSYS;
2365
2366         if (!pid)
2367                 head = current->robust_list;
2368         else {
2369                 struct task_struct *p;
2370
2371                 ret = -ESRCH;
2372                 rcu_read_lock();
2373                 p = find_task_by_vpid(pid);
2374                 if (!p)
2375                         goto err_unlock;
2376                 ret = -EPERM;
2377                 pcred = __task_cred(p);
2378                 if (cred->euid != pcred->euid &&
2379                     cred->euid != pcred->uid &&
2380                     !capable(CAP_SYS_PTRACE))
2381                         goto err_unlock;
2382                 head = p->robust_list;
2383                 rcu_read_unlock();
2384         }
2385
2386         if (put_user(sizeof(*head), len_ptr))
2387                 return -EFAULT;
2388         return put_user(head, head_ptr);
2389
2390 err_unlock:
2391         rcu_read_unlock();
2392
2393         return ret;
2394 }
2395
2396 /*
2397  * Process a futex-list entry, check whether it's owned by the
2398  * dying task, and do notification if so:
2399  */
2400 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2401 {
2402         u32 uval, nval, mval;
2403
2404 retry:
2405         if (get_user(uval, uaddr))
2406                 return -1;
2407
2408         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2409                 /*
2410                  * Ok, this dying thread is truly holding a futex
2411                  * of interest. Set the OWNER_DIED bit atomically
2412                  * via cmpxchg, and if the value had FUTEX_WAITERS
2413                  * set, wake up a waiter (if any). (We have to do a
2414                  * futex_wake() even if OWNER_DIED is already set -
2415                  * to handle the rare but possible case of recursive
2416                  * thread-death.) The rest of the cleanup is done in
2417                  * userspace.
2418                  */
2419                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2420                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2421
2422                 if (nval == -EFAULT)
2423                         return -1;
2424
2425                 if (nval != uval)
2426                         goto retry;
2427
2428                 /*
2429                  * Wake robust non-PI futexes here. The wakeup of
2430                  * PI futexes happens in exit_pi_state():
2431                  */
2432                 if (!pi && (uval & FUTEX_WAITERS))
2433                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2434         }
2435         return 0;
2436 }
2437
2438 /*
2439  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2440  */
2441 static inline int fetch_robust_entry(struct robust_list __user **entry,
2442                                      struct robust_list __user * __user *head,
2443                                      int *pi)
2444 {
2445         unsigned long uentry;
2446
2447         if (get_user(uentry, (unsigned long __user *)head))
2448                 return -EFAULT;
2449
2450         *entry = (void __user *)(uentry & ~1UL);
2451         *pi = uentry & 1;
2452
2453         return 0;
2454 }
2455
2456 /*
2457  * Walk curr->robust_list (very carefully, it's a userspace list!)
2458  * and mark any locks found there dead, and notify any waiters.
2459  *
2460  * We silently return on any sign of list-walking problem.
2461  */
2462 void exit_robust_list(struct task_struct *curr)
2463 {
2464         struct robust_list_head __user *head = curr->robust_list;
2465         struct robust_list __user *entry, *next_entry, *pending;
2466         unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2467         unsigned long futex_offset;
2468         int rc;
2469
2470         if (!futex_cmpxchg_enabled)
2471                 return;
2472
2473         /*
2474          * Fetch the list head (which was registered earlier, via
2475          * sys_set_robust_list()):
2476          */
2477         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2478                 return;
2479         /*
2480          * Fetch the relative futex offset:
2481          */
2482         if (get_user(futex_offset, &head->futex_offset))
2483                 return;
2484         /*
2485          * Fetch any possibly pending lock-add first, and handle it
2486          * if it exists:
2487          */
2488         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2489                 return;
2490
2491         next_entry = NULL;      /* avoid warning with gcc */
2492         while (entry != &head->list) {
2493                 /*
2494                  * Fetch the next entry in the list before calling
2495                  * handle_futex_death:
2496                  */
2497                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2498                 /*
2499                  * A pending lock might already be on the list, so
2500                  * don't process it twice:
2501                  */
2502                 if (entry != pending)
2503                         if (handle_futex_death((void __user *)entry + futex_offset,
2504                                                 curr, pi))
2505                                 return;
2506                 if (rc)
2507                         return;
2508                 entry = next_entry;
2509                 pi = next_pi;
2510                 /*
2511                  * Avoid excessively long or circular lists:
2512                  */
2513                 if (!--limit)
2514                         break;
2515
2516                 cond_resched();
2517         }
2518
2519         if (pending)
2520                 handle_futex_death((void __user *)pending + futex_offset,
2521                                    curr, pip);
2522 }
2523
2524 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2525                 u32 __user *uaddr2, u32 val2, u32 val3)
2526 {
2527         int clockrt, ret = -ENOSYS;
2528         int cmd = op & FUTEX_CMD_MASK;
2529         int fshared = 0;
2530
2531         if (!(op & FUTEX_PRIVATE_FLAG))
2532                 fshared = 1;
2533
2534         clockrt = op & FUTEX_CLOCK_REALTIME;
2535         if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2536                 return -ENOSYS;
2537
2538         switch (cmd) {
2539         case FUTEX_WAIT:
2540                 val3 = FUTEX_BITSET_MATCH_ANY;
2541         case FUTEX_WAIT_BITSET:
2542                 ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2543                 break;
2544         case FUTEX_WAKE:
2545                 val3 = FUTEX_BITSET_MATCH_ANY;
2546         case FUTEX_WAKE_BITSET:
2547                 ret = futex_wake(uaddr, fshared, val, val3);
2548                 break;
2549         case FUTEX_REQUEUE:
2550                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2551                 break;
2552         case FUTEX_CMP_REQUEUE:
2553                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2554                                     0);
2555                 break;
2556         case FUTEX_WAKE_OP:
2557                 ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2558                 break;
2559         case FUTEX_LOCK_PI:
2560                 if (futex_cmpxchg_enabled)
2561                         ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2562                 break;
2563         case FUTEX_UNLOCK_PI:
2564                 if (futex_cmpxchg_enabled)
2565                         ret = futex_unlock_pi(uaddr, fshared);
2566                 break;
2567         case FUTEX_TRYLOCK_PI:
2568                 if (futex_cmpxchg_enabled)
2569                         ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2570                 break;
2571         case FUTEX_WAIT_REQUEUE_PI:
2572                 val3 = FUTEX_BITSET_MATCH_ANY;
2573                 ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2574                                             clockrt, uaddr2);
2575                 break;
2576         case FUTEX_CMP_REQUEUE_PI:
2577                 ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2578                                     1);
2579                 break;
2580         default:
2581                 ret = -ENOSYS;
2582         }
2583         return ret;
2584 }
2585
2586
2587 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2588                 struct timespec __user *, utime, u32 __user *, uaddr2,
2589                 u32, val3)
2590 {
2591         struct timespec ts;
2592         ktime_t t, *tp = NULL;
2593         u32 val2 = 0;
2594         int cmd = op & FUTEX_CMD_MASK;
2595
2596         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2597                       cmd == FUTEX_WAIT_BITSET ||
2598                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2599                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2600                         return -EFAULT;
2601                 if (!timespec_valid(&ts))
2602                         return -EINVAL;
2603
2604                 t = timespec_to_ktime(ts);
2605                 if (cmd == FUTEX_WAIT)
2606                         t = ktime_add_safe(ktime_get(), t);
2607                 tp = &t;
2608         }
2609         /*
2610          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2611          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2612          */
2613         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2614             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2615                 val2 = (u32) (unsigned long) utime;
2616
2617         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2618 }
2619
2620 static int __init futex_init(void)
2621 {
2622         u32 curval;
2623         int i;
2624
2625         /*
2626          * This will fail and we want it. Some arch implementations do
2627          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2628          * functionality. We want to know that before we call in any
2629          * of the complex code paths. Also we want to prevent
2630          * registration of robust lists in that case. NULL is
2631          * guaranteed to fault and we get -EFAULT on functional
2632          * implementation, the non functional ones will return
2633          * -ENOSYS.
2634          */
2635         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2636         if (curval == -EFAULT)
2637                 futex_cmpxchg_enabled = 1;
2638
2639         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2640                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2641                 spin_lock_init(&futex_queues[i].lock);
2642         }
2643
2644         return 0;
2645 }
2646 __initcall(futex_init);