Merge branches 'core-fixes-for-linus', 'x86-fixes-for-linus', 'timers-fixes-for-linus...
[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  * Futex flags used to encode options to functions and preserve them across
73  * restarts.
74  */
75 #define FLAGS_SHARED            0x01
76 #define FLAGS_CLOCKRT           0x02
77 #define FLAGS_HAS_TIMEOUT       0x04
78
79 /*
80  * Priority Inheritance state:
81  */
82 struct futex_pi_state {
83         /*
84          * list of 'owned' pi_state instances - these have to be
85          * cleaned up in do_exit() if the task exits prematurely:
86          */
87         struct list_head list;
88
89         /*
90          * The PI object:
91          */
92         struct rt_mutex pi_mutex;
93
94         struct task_struct *owner;
95         atomic_t refcount;
96
97         union futex_key key;
98 };
99
100 /**
101  * struct futex_q - The hashed futex queue entry, one per waiting task
102  * @list:               priority-sorted list of tasks waiting on this futex
103  * @task:               the task waiting on the futex
104  * @lock_ptr:           the hash bucket lock
105  * @key:                the key the futex is hashed on
106  * @pi_state:           optional priority inheritance state
107  * @rt_waiter:          rt_waiter storage for use with requeue_pi
108  * @requeue_pi_key:     the requeue_pi target futex key
109  * @bitset:             bitset for the optional bitmasked wakeup
110  *
111  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
112  * we can wake only the relevant ones (hashed queues may be shared).
113  *
114  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
115  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
116  * The order of wakeup is always to make the first condition true, then
117  * the second.
118  *
119  * PI futexes are typically woken before they are removed from the hash list via
120  * the rt_mutex code. See unqueue_me_pi().
121  */
122 struct futex_q {
123         struct plist_node list;
124
125         struct task_struct *task;
126         spinlock_t *lock_ptr;
127         union futex_key key;
128         struct futex_pi_state *pi_state;
129         struct rt_mutex_waiter *rt_waiter;
130         union futex_key *requeue_pi_key;
131         u32 bitset;
132 };
133
134 static const struct futex_q futex_q_init = {
135         /* list gets initialized in queue_me()*/
136         .key = FUTEX_KEY_INIT,
137         .bitset = FUTEX_BITSET_MATCH_ANY
138 };
139
140 /*
141  * Hash buckets are shared by all the futex_keys that hash to the same
142  * location.  Each key may have multiple futex_q structures, one for each task
143  * waiting on a futex.
144  */
145 struct futex_hash_bucket {
146         spinlock_t lock;
147         struct plist_head chain;
148 };
149
150 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
151
152 /*
153  * We hash on the keys returned from get_futex_key (see below).
154  */
155 static struct futex_hash_bucket *hash_futex(union futex_key *key)
156 {
157         u32 hash = jhash2((u32*)&key->both.word,
158                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
159                           key->both.offset);
160         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
161 }
162
163 /*
164  * Return 1 if two futex_keys are equal, 0 otherwise.
165  */
166 static inline int match_futex(union futex_key *key1, union futex_key *key2)
167 {
168         return (key1 && key2
169                 && key1->both.word == key2->both.word
170                 && key1->both.ptr == key2->both.ptr
171                 && key1->both.offset == key2->both.offset);
172 }
173
174 /*
175  * Take a reference to the resource addressed by a key.
176  * Can be called while holding spinlocks.
177  *
178  */
179 static void get_futex_key_refs(union futex_key *key)
180 {
181         if (!key->both.ptr)
182                 return;
183
184         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
185         case FUT_OFF_INODE:
186                 ihold(key->shared.inode);
187                 break;
188         case FUT_OFF_MMSHARED:
189                 atomic_inc(&key->private.mm->mm_count);
190                 break;
191         }
192 }
193
194 /*
195  * Drop a reference to the resource addressed by a key.
196  * The hash bucket spinlock must not be held.
197  */
198 static void drop_futex_key_refs(union futex_key *key)
199 {
200         if (!key->both.ptr) {
201                 /* If we're here then we tried to put a key we failed to get */
202                 WARN_ON_ONCE(1);
203                 return;
204         }
205
206         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
207         case FUT_OFF_INODE:
208                 iput(key->shared.inode);
209                 break;
210         case FUT_OFF_MMSHARED:
211                 mmdrop(key->private.mm);
212                 break;
213         }
214 }
215
216 /**
217  * get_futex_key() - Get parameters which are the keys for a futex
218  * @uaddr:      virtual address of the futex
219  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
220  * @key:        address where result is stored.
221  *
222  * Returns a negative error code or 0
223  * The key words are stored in *key on success.
224  *
225  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
226  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
227  * We can usually work out the index without swapping in the page.
228  *
229  * lock_page() might sleep, the caller should not hold a spinlock.
230  */
231 static int
232 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
233 {
234         unsigned long address = (unsigned long)uaddr;
235         struct mm_struct *mm = current->mm;
236         struct page *page, *page_head;
237         int err;
238
239         /*
240          * The futex address must be "naturally" aligned.
241          */
242         key->both.offset = address % PAGE_SIZE;
243         if (unlikely((address % sizeof(u32)) != 0))
244                 return -EINVAL;
245         address -= key->both.offset;
246
247         /*
248          * PROCESS_PRIVATE futexes are fast.
249          * As the mm cannot disappear under us and the 'key' only needs
250          * virtual address, we dont even have to find the underlying vma.
251          * Note : We do have to check 'uaddr' is a valid user address,
252          *        but access_ok() should be faster than find_vma()
253          */
254         if (!fshared) {
255                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
256                         return -EFAULT;
257                 key->private.mm = mm;
258                 key->private.address = address;
259                 get_futex_key_refs(key);
260                 return 0;
261         }
262
263 again:
264         err = get_user_pages_fast(address, 1, 1, &page);
265         if (err < 0)
266                 return err;
267
268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
269         page_head = page;
270         if (unlikely(PageTail(page))) {
271                 put_page(page);
272                 /* serialize against __split_huge_page_splitting() */
273                 local_irq_disable();
274                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
275                         page_head = compound_head(page);
276                         /*
277                          * page_head is valid pointer but we must pin
278                          * it before taking the PG_lock and/or
279                          * PG_compound_lock. The moment we re-enable
280                          * irqs __split_huge_page_splitting() can
281                          * return and the head page can be freed from
282                          * under us. We can't take the PG_lock and/or
283                          * PG_compound_lock on a page that could be
284                          * freed from under us.
285                          */
286                         if (page != page_head) {
287                                 get_page(page_head);
288                                 put_page(page);
289                         }
290                         local_irq_enable();
291                 } else {
292                         local_irq_enable();
293                         goto again;
294                 }
295         }
296 #else
297         page_head = compound_head(page);
298         if (page != page_head) {
299                 get_page(page_head);
300                 put_page(page);
301         }
302 #endif
303
304         lock_page(page_head);
305         if (!page_head->mapping) {
306                 unlock_page(page_head);
307                 put_page(page_head);
308                 goto again;
309         }
310
311         /*
312          * Private mappings are handled in a simple way.
313          *
314          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
315          * it's a read-only handle, it's expected that futexes attach to
316          * the object not the particular process.
317          */
318         if (PageAnon(page_head)) {
319                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
320                 key->private.mm = mm;
321                 key->private.address = address;
322         } else {
323                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
324                 key->shared.inode = page_head->mapping->host;
325                 key->shared.pgoff = page_head->index;
326         }
327
328         get_futex_key_refs(key);
329
330         unlock_page(page_head);
331         put_page(page_head);
332         return 0;
333 }
334
335 static inline void put_futex_key(union futex_key *key)
336 {
337         drop_futex_key_refs(key);
338 }
339
340 /**
341  * fault_in_user_writeable() - Fault in user address and verify RW access
342  * @uaddr:      pointer to faulting user space address
343  *
344  * Slow path to fixup the fault we just took in the atomic write
345  * access to @uaddr.
346  *
347  * We have no generic implementation of a non-destructive write to the
348  * user address. We know that we faulted in the atomic pagefault
349  * disabled section so we can as well avoid the #PF overhead by
350  * calling get_user_pages() right away.
351  */
352 static int fault_in_user_writeable(u32 __user *uaddr)
353 {
354         struct mm_struct *mm = current->mm;
355         int ret;
356
357         down_read(&mm->mmap_sem);
358         ret = get_user_pages(current, mm, (unsigned long)uaddr,
359                              1, 1, 0, NULL, NULL);
360         up_read(&mm->mmap_sem);
361
362         return ret < 0 ? ret : 0;
363 }
364
365 /**
366  * futex_top_waiter() - Return the highest priority waiter on a futex
367  * @hb:         the hash bucket the futex_q's reside in
368  * @key:        the futex key (to distinguish it from other futex futex_q's)
369  *
370  * Must be called with the hb lock held.
371  */
372 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
373                                         union futex_key *key)
374 {
375         struct futex_q *this;
376
377         plist_for_each_entry(this, &hb->chain, list) {
378                 if (match_futex(&this->key, key))
379                         return this;
380         }
381         return NULL;
382 }
383
384 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
385 {
386         u32 curval;
387
388         pagefault_disable();
389         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
390         pagefault_enable();
391
392         return curval;
393 }
394
395 static int get_futex_value_locked(u32 *dest, u32 __user *from)
396 {
397         int ret;
398
399         pagefault_disable();
400         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
401         pagefault_enable();
402
403         return ret ? -EFAULT : 0;
404 }
405
406
407 /*
408  * PI code:
409  */
410 static int refill_pi_state_cache(void)
411 {
412         struct futex_pi_state *pi_state;
413
414         if (likely(current->pi_state_cache))
415                 return 0;
416
417         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
418
419         if (!pi_state)
420                 return -ENOMEM;
421
422         INIT_LIST_HEAD(&pi_state->list);
423         /* pi_mutex gets initialized later */
424         pi_state->owner = NULL;
425         atomic_set(&pi_state->refcount, 1);
426         pi_state->key = FUTEX_KEY_INIT;
427
428         current->pi_state_cache = pi_state;
429
430         return 0;
431 }
432
433 static struct futex_pi_state * alloc_pi_state(void)
434 {
435         struct futex_pi_state *pi_state = current->pi_state_cache;
436
437         WARN_ON(!pi_state);
438         current->pi_state_cache = NULL;
439
440         return pi_state;
441 }
442
443 static void free_pi_state(struct futex_pi_state *pi_state)
444 {
445         if (!atomic_dec_and_test(&pi_state->refcount))
446                 return;
447
448         /*
449          * If pi_state->owner is NULL, the owner is most probably dying
450          * and has cleaned up the pi_state already
451          */
452         if (pi_state->owner) {
453                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
454                 list_del_init(&pi_state->list);
455                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
456
457                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
458         }
459
460         if (current->pi_state_cache)
461                 kfree(pi_state);
462         else {
463                 /*
464                  * pi_state->list is already empty.
465                  * clear pi_state->owner.
466                  * refcount is at 0 - put it back to 1.
467                  */
468                 pi_state->owner = NULL;
469                 atomic_set(&pi_state->refcount, 1);
470                 current->pi_state_cache = pi_state;
471         }
472 }
473
474 /*
475  * Look up the task based on what TID userspace gave us.
476  * We dont trust it.
477  */
478 static struct task_struct * futex_find_get_task(pid_t pid)
479 {
480         struct task_struct *p;
481
482         rcu_read_lock();
483         p = find_task_by_vpid(pid);
484         if (p)
485                 get_task_struct(p);
486
487         rcu_read_unlock();
488
489         return p;
490 }
491
492 /*
493  * This task is holding PI mutexes at exit time => bad.
494  * Kernel cleans up PI-state, but userspace is likely hosed.
495  * (Robust-futex cleanup is separate and might save the day for userspace.)
496  */
497 void exit_pi_state_list(struct task_struct *curr)
498 {
499         struct list_head *next, *head = &curr->pi_state_list;
500         struct futex_pi_state *pi_state;
501         struct futex_hash_bucket *hb;
502         union futex_key key = FUTEX_KEY_INIT;
503
504         if (!futex_cmpxchg_enabled)
505                 return;
506         /*
507          * We are a ZOMBIE and nobody can enqueue itself on
508          * pi_state_list anymore, but we have to be careful
509          * versus waiters unqueueing themselves:
510          */
511         raw_spin_lock_irq(&curr->pi_lock);
512         while (!list_empty(head)) {
513
514                 next = head->next;
515                 pi_state = list_entry(next, struct futex_pi_state, list);
516                 key = pi_state->key;
517                 hb = hash_futex(&key);
518                 raw_spin_unlock_irq(&curr->pi_lock);
519
520                 spin_lock(&hb->lock);
521
522                 raw_spin_lock_irq(&curr->pi_lock);
523                 /*
524                  * We dropped the pi-lock, so re-check whether this
525                  * task still owns the PI-state:
526                  */
527                 if (head->next != next) {
528                         spin_unlock(&hb->lock);
529                         continue;
530                 }
531
532                 WARN_ON(pi_state->owner != curr);
533                 WARN_ON(list_empty(&pi_state->list));
534                 list_del_init(&pi_state->list);
535                 pi_state->owner = NULL;
536                 raw_spin_unlock_irq(&curr->pi_lock);
537
538                 rt_mutex_unlock(&pi_state->pi_mutex);
539
540                 spin_unlock(&hb->lock);
541
542                 raw_spin_lock_irq(&curr->pi_lock);
543         }
544         raw_spin_unlock_irq(&curr->pi_lock);
545 }
546
547 static int
548 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
549                 union futex_key *key, struct futex_pi_state **ps)
550 {
551         struct futex_pi_state *pi_state = NULL;
552         struct futex_q *this, *next;
553         struct plist_head *head;
554         struct task_struct *p;
555         pid_t pid = uval & FUTEX_TID_MASK;
556
557         head = &hb->chain;
558
559         plist_for_each_entry_safe(this, next, head, list) {
560                 if (match_futex(&this->key, key)) {
561                         /*
562                          * Another waiter already exists - bump up
563                          * the refcount and return its pi_state:
564                          */
565                         pi_state = this->pi_state;
566                         /*
567                          * Userspace might have messed up non-PI and PI futexes
568                          */
569                         if (unlikely(!pi_state))
570                                 return -EINVAL;
571
572                         WARN_ON(!atomic_read(&pi_state->refcount));
573
574                         /*
575                          * When pi_state->owner is NULL then the owner died
576                          * and another waiter is on the fly. pi_state->owner
577                          * is fixed up by the task which acquires
578                          * pi_state->rt_mutex.
579                          *
580                          * We do not check for pid == 0 which can happen when
581                          * the owner died and robust_list_exit() cleared the
582                          * TID.
583                          */
584                         if (pid && pi_state->owner) {
585                                 /*
586                                  * Bail out if user space manipulated the
587                                  * futex value.
588                                  */
589                                 if (pid != task_pid_vnr(pi_state->owner))
590                                         return -EINVAL;
591                         }
592
593                         atomic_inc(&pi_state->refcount);
594                         *ps = pi_state;
595
596                         return 0;
597                 }
598         }
599
600         /*
601          * We are the first waiter - try to look up the real owner and attach
602          * the new pi_state to it, but bail out when TID = 0
603          */
604         if (!pid)
605                 return -ESRCH;
606         p = futex_find_get_task(pid);
607         if (!p)
608                 return -ESRCH;
609
610         /*
611          * We need to look at the task state flags to figure out,
612          * whether the task is exiting. To protect against the do_exit
613          * change of the task flags, we do this protected by
614          * p->pi_lock:
615          */
616         raw_spin_lock_irq(&p->pi_lock);
617         if (unlikely(p->flags & PF_EXITING)) {
618                 /*
619                  * The task is on the way out. When PF_EXITPIDONE is
620                  * set, we know that the task has finished the
621                  * cleanup:
622                  */
623                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
624
625                 raw_spin_unlock_irq(&p->pi_lock);
626                 put_task_struct(p);
627                 return ret;
628         }
629
630         pi_state = alloc_pi_state();
631
632         /*
633          * Initialize the pi_mutex in locked state and make 'p'
634          * the owner of it:
635          */
636         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
637
638         /* Store the key for possible exit cleanups: */
639         pi_state->key = *key;
640
641         WARN_ON(!list_empty(&pi_state->list));
642         list_add(&pi_state->list, &p->pi_state_list);
643         pi_state->owner = p;
644         raw_spin_unlock_irq(&p->pi_lock);
645
646         put_task_struct(p);
647
648         *ps = pi_state;
649
650         return 0;
651 }
652
653 /**
654  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
655  * @uaddr:              the pi futex user address
656  * @hb:                 the pi futex hash bucket
657  * @key:                the futex key associated with uaddr and hb
658  * @ps:                 the pi_state pointer where we store the result of the
659  *                      lookup
660  * @task:               the task to perform the atomic lock work for.  This will
661  *                      be "current" except in the case of requeue pi.
662  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
663  *
664  * Returns:
665  *  0 - ready to wait
666  *  1 - acquired the lock
667  * <0 - error
668  *
669  * The hb->lock and futex_key refs shall be held by the caller.
670  */
671 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
672                                 union futex_key *key,
673                                 struct futex_pi_state **ps,
674                                 struct task_struct *task, int set_waiters)
675 {
676         int lock_taken, ret, ownerdied = 0;
677         u32 uval, newval, curval;
678
679 retry:
680         ret = lock_taken = 0;
681
682         /*
683          * To avoid races, we attempt to take the lock here again
684          * (by doing a 0 -> TID atomic cmpxchg), while holding all
685          * the locks. It will most likely not succeed.
686          */
687         newval = task_pid_vnr(task);
688         if (set_waiters)
689                 newval |= FUTEX_WAITERS;
690
691         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
692
693         if (unlikely(curval == -EFAULT))
694                 return -EFAULT;
695
696         /*
697          * Detect deadlocks.
698          */
699         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
700                 return -EDEADLK;
701
702         /*
703          * Surprise - we got the lock. Just return to userspace:
704          */
705         if (unlikely(!curval))
706                 return 1;
707
708         uval = curval;
709
710         /*
711          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
712          * to wake at the next unlock.
713          */
714         newval = curval | FUTEX_WAITERS;
715
716         /*
717          * There are two cases, where a futex might have no owner (the
718          * owner TID is 0): OWNER_DIED. We take over the futex in this
719          * case. We also do an unconditional take over, when the owner
720          * of the futex died.
721          *
722          * This is safe as we are protected by the hash bucket lock !
723          */
724         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
725                 /* Keep the OWNER_DIED bit */
726                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
727                 ownerdied = 0;
728                 lock_taken = 1;
729         }
730
731         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
732
733         if (unlikely(curval == -EFAULT))
734                 return -EFAULT;
735         if (unlikely(curval != uval))
736                 goto retry;
737
738         /*
739          * We took the lock due to owner died take over.
740          */
741         if (unlikely(lock_taken))
742                 return 1;
743
744         /*
745          * We dont have the lock. Look up the PI state (or create it if
746          * we are the first waiter):
747          */
748         ret = lookup_pi_state(uval, hb, key, ps);
749
750         if (unlikely(ret)) {
751                 switch (ret) {
752                 case -ESRCH:
753                         /*
754                          * No owner found for this futex. Check if the
755                          * OWNER_DIED bit is set to figure out whether
756                          * this is a robust futex or not.
757                          */
758                         if (get_futex_value_locked(&curval, uaddr))
759                                 return -EFAULT;
760
761                         /*
762                          * We simply start over in case of a robust
763                          * futex. The code above will take the futex
764                          * and return happy.
765                          */
766                         if (curval & FUTEX_OWNER_DIED) {
767                                 ownerdied = 1;
768                                 goto retry;
769                         }
770                 default:
771                         break;
772                 }
773         }
774
775         return ret;
776 }
777
778 /*
779  * The hash bucket lock must be held when this is called.
780  * Afterwards, the futex_q must not be accessed.
781  */
782 static void wake_futex(struct futex_q *q)
783 {
784         struct task_struct *p = q->task;
785
786         /*
787          * We set q->lock_ptr = NULL _before_ we wake up the task. If
788          * a non-futex wake up happens on another CPU then the task
789          * might exit and p would dereference a non-existing task
790          * struct. Prevent this by holding a reference on p across the
791          * wake up.
792          */
793         get_task_struct(p);
794
795         plist_del(&q->list, &q->list.plist);
796         /*
797          * The waiting task can free the futex_q as soon as
798          * q->lock_ptr = NULL is written, without taking any locks. A
799          * memory barrier is required here to prevent the following
800          * store to lock_ptr from getting ahead of the plist_del.
801          */
802         smp_wmb();
803         q->lock_ptr = NULL;
804
805         wake_up_state(p, TASK_NORMAL);
806         put_task_struct(p);
807 }
808
809 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
810 {
811         struct task_struct *new_owner;
812         struct futex_pi_state *pi_state = this->pi_state;
813         u32 curval, newval;
814
815         if (!pi_state)
816                 return -EINVAL;
817
818         /*
819          * If current does not own the pi_state then the futex is
820          * inconsistent and user space fiddled with the futex value.
821          */
822         if (pi_state->owner != current)
823                 return -EINVAL;
824
825         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
826         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
827
828         /*
829          * It is possible that the next waiter (the one that brought
830          * this owner to the kernel) timed out and is no longer
831          * waiting on the lock.
832          */
833         if (!new_owner)
834                 new_owner = this->task;
835
836         /*
837          * We pass it to the next owner. (The WAITERS bit is always
838          * kept enabled while there is PI state around. We must also
839          * preserve the owner died bit.)
840          */
841         if (!(uval & FUTEX_OWNER_DIED)) {
842                 int ret = 0;
843
844                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
845
846                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
847
848                 if (curval == -EFAULT)
849                         ret = -EFAULT;
850                 else if (curval != uval)
851                         ret = -EINVAL;
852                 if (ret) {
853                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
854                         return ret;
855                 }
856         }
857
858         raw_spin_lock_irq(&pi_state->owner->pi_lock);
859         WARN_ON(list_empty(&pi_state->list));
860         list_del_init(&pi_state->list);
861         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
862
863         raw_spin_lock_irq(&new_owner->pi_lock);
864         WARN_ON(!list_empty(&pi_state->list));
865         list_add(&pi_state->list, &new_owner->pi_state_list);
866         pi_state->owner = new_owner;
867         raw_spin_unlock_irq(&new_owner->pi_lock);
868
869         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
870         rt_mutex_unlock(&pi_state->pi_mutex);
871
872         return 0;
873 }
874
875 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
876 {
877         u32 oldval;
878
879         /*
880          * There is no waiter, so we unlock the futex. The owner died
881          * bit has not to be preserved here. We are the owner:
882          */
883         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
884
885         if (oldval == -EFAULT)
886                 return oldval;
887         if (oldval != uval)
888                 return -EAGAIN;
889
890         return 0;
891 }
892
893 /*
894  * Express the locking dependencies for lockdep:
895  */
896 static inline void
897 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
898 {
899         if (hb1 <= hb2) {
900                 spin_lock(&hb1->lock);
901                 if (hb1 < hb2)
902                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
903         } else { /* hb1 > hb2 */
904                 spin_lock(&hb2->lock);
905                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
906         }
907 }
908
909 static inline void
910 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
911 {
912         spin_unlock(&hb1->lock);
913         if (hb1 != hb2)
914                 spin_unlock(&hb2->lock);
915 }
916
917 /*
918  * Wake up waiters matching bitset queued on this futex (uaddr).
919  */
920 static int
921 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
922 {
923         struct futex_hash_bucket *hb;
924         struct futex_q *this, *next;
925         struct plist_head *head;
926         union futex_key key = FUTEX_KEY_INIT;
927         int ret;
928
929         if (!bitset)
930                 return -EINVAL;
931
932         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
933         if (unlikely(ret != 0))
934                 goto out;
935
936         hb = hash_futex(&key);
937         spin_lock(&hb->lock);
938         head = &hb->chain;
939
940         plist_for_each_entry_safe(this, next, head, list) {
941                 if (match_futex (&this->key, &key)) {
942                         if (this->pi_state || this->rt_waiter) {
943                                 ret = -EINVAL;
944                                 break;
945                         }
946
947                         /* Check if one of the bits is set in both bitsets */
948                         if (!(this->bitset & bitset))
949                                 continue;
950
951                         wake_futex(this);
952                         if (++ret >= nr_wake)
953                                 break;
954                 }
955         }
956
957         spin_unlock(&hb->lock);
958         put_futex_key(&key);
959 out:
960         return ret;
961 }
962
963 /*
964  * Wake up all waiters hashed on the physical page that is mapped
965  * to this virtual address:
966  */
967 static int
968 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
969               int nr_wake, int nr_wake2, int op)
970 {
971         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
972         struct futex_hash_bucket *hb1, *hb2;
973         struct plist_head *head;
974         struct futex_q *this, *next;
975         int ret, op_ret;
976
977 retry:
978         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
979         if (unlikely(ret != 0))
980                 goto out;
981         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
982         if (unlikely(ret != 0))
983                 goto out_put_key1;
984
985         hb1 = hash_futex(&key1);
986         hb2 = hash_futex(&key2);
987
988 retry_private:
989         double_lock_hb(hb1, hb2);
990         op_ret = futex_atomic_op_inuser(op, uaddr2);
991         if (unlikely(op_ret < 0)) {
992
993                 double_unlock_hb(hb1, hb2);
994
995 #ifndef CONFIG_MMU
996                 /*
997                  * we don't get EFAULT from MMU faults if we don't have an MMU,
998                  * but we might get them from range checking
999                  */
1000                 ret = op_ret;
1001                 goto out_put_keys;
1002 #endif
1003
1004                 if (unlikely(op_ret != -EFAULT)) {
1005                         ret = op_ret;
1006                         goto out_put_keys;
1007                 }
1008
1009                 ret = fault_in_user_writeable(uaddr2);
1010                 if (ret)
1011                         goto out_put_keys;
1012
1013                 if (!(flags & FLAGS_SHARED))
1014                         goto retry_private;
1015
1016                 put_futex_key(&key2);
1017                 put_futex_key(&key1);
1018                 goto retry;
1019         }
1020
1021         head = &hb1->chain;
1022
1023         plist_for_each_entry_safe(this, next, head, list) {
1024                 if (match_futex (&this->key, &key1)) {
1025                         wake_futex(this);
1026                         if (++ret >= nr_wake)
1027                                 break;
1028                 }
1029         }
1030
1031         if (op_ret > 0) {
1032                 head = &hb2->chain;
1033
1034                 op_ret = 0;
1035                 plist_for_each_entry_safe(this, next, head, list) {
1036                         if (match_futex (&this->key, &key2)) {
1037                                 wake_futex(this);
1038                                 if (++op_ret >= nr_wake2)
1039                                         break;
1040                         }
1041                 }
1042                 ret += op_ret;
1043         }
1044
1045         double_unlock_hb(hb1, hb2);
1046 out_put_keys:
1047         put_futex_key(&key2);
1048 out_put_key1:
1049         put_futex_key(&key1);
1050 out:
1051         return ret;
1052 }
1053
1054 /**
1055  * requeue_futex() - Requeue a futex_q from one hb to another
1056  * @q:          the futex_q to requeue
1057  * @hb1:        the source hash_bucket
1058  * @hb2:        the target hash_bucket
1059  * @key2:       the new key for the requeued futex_q
1060  */
1061 static inline
1062 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1063                    struct futex_hash_bucket *hb2, union futex_key *key2)
1064 {
1065
1066         /*
1067          * If key1 and key2 hash to the same bucket, no need to
1068          * requeue.
1069          */
1070         if (likely(&hb1->chain != &hb2->chain)) {
1071                 plist_del(&q->list, &hb1->chain);
1072                 plist_add(&q->list, &hb2->chain);
1073                 q->lock_ptr = &hb2->lock;
1074 #ifdef CONFIG_DEBUG_PI_LIST
1075                 q->list.plist.spinlock = &hb2->lock;
1076 #endif
1077         }
1078         get_futex_key_refs(key2);
1079         q->key = *key2;
1080 }
1081
1082 /**
1083  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1084  * @q:          the futex_q
1085  * @key:        the key of the requeue target futex
1086  * @hb:         the hash_bucket of the requeue target futex
1087  *
1088  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1089  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1090  * to the requeue target futex so the waiter can detect the wakeup on the right
1091  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1092  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1093  * to protect access to the pi_state to fixup the owner later.  Must be called
1094  * with both q->lock_ptr and hb->lock held.
1095  */
1096 static inline
1097 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1098                            struct futex_hash_bucket *hb)
1099 {
1100         get_futex_key_refs(key);
1101         q->key = *key;
1102
1103         WARN_ON(plist_node_empty(&q->list));
1104         plist_del(&q->list, &q->list.plist);
1105
1106         WARN_ON(!q->rt_waiter);
1107         q->rt_waiter = NULL;
1108
1109         q->lock_ptr = &hb->lock;
1110 #ifdef CONFIG_DEBUG_PI_LIST
1111         q->list.plist.spinlock = &hb->lock;
1112 #endif
1113
1114         wake_up_state(q->task, TASK_NORMAL);
1115 }
1116
1117 /**
1118  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1119  * @pifutex:            the user address of the to futex
1120  * @hb1:                the from futex hash bucket, must be locked by the caller
1121  * @hb2:                the to futex hash bucket, must be locked by the caller
1122  * @key1:               the from futex key
1123  * @key2:               the to futex key
1124  * @ps:                 address to store the pi_state pointer
1125  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1126  *
1127  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1128  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1129  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1130  * hb1 and hb2 must be held by the caller.
1131  *
1132  * Returns:
1133  *  0 - failed to acquire the lock atomicly
1134  *  1 - acquired the lock
1135  * <0 - error
1136  */
1137 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1138                                  struct futex_hash_bucket *hb1,
1139                                  struct futex_hash_bucket *hb2,
1140                                  union futex_key *key1, union futex_key *key2,
1141                                  struct futex_pi_state **ps, int set_waiters)
1142 {
1143         struct futex_q *top_waiter = NULL;
1144         u32 curval;
1145         int ret;
1146
1147         if (get_futex_value_locked(&curval, pifutex))
1148                 return -EFAULT;
1149
1150         /*
1151          * Find the top_waiter and determine if there are additional waiters.
1152          * If the caller intends to requeue more than 1 waiter to pifutex,
1153          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1154          * as we have means to handle the possible fault.  If not, don't set
1155          * the bit unecessarily as it will force the subsequent unlock to enter
1156          * the kernel.
1157          */
1158         top_waiter = futex_top_waiter(hb1, key1);
1159
1160         /* There are no waiters, nothing for us to do. */
1161         if (!top_waiter)
1162                 return 0;
1163
1164         /* Ensure we requeue to the expected futex. */
1165         if (!match_futex(top_waiter->requeue_pi_key, key2))
1166                 return -EINVAL;
1167
1168         /*
1169          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1170          * the contended case or if set_waiters is 1.  The pi_state is returned
1171          * in ps in contended cases.
1172          */
1173         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1174                                    set_waiters);
1175         if (ret == 1)
1176                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1177
1178         return ret;
1179 }
1180
1181 /**
1182  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1183  * @uaddr1:     source futex user address
1184  * @flags:      futex flags (FLAGS_SHARED, etc.)
1185  * @uaddr2:     target futex user address
1186  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1187  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1188  * @cmpval:     @uaddr1 expected value (or %NULL)
1189  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1190  *              pi futex (pi to pi requeue is not supported)
1191  *
1192  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1193  * uaddr2 atomically on behalf of the top waiter.
1194  *
1195  * Returns:
1196  * >=0 - on success, the number of tasks requeued or woken
1197  *  <0 - on error
1198  */
1199 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1200                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1201                          u32 *cmpval, int requeue_pi)
1202 {
1203         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1204         int drop_count = 0, task_count = 0, ret;
1205         struct futex_pi_state *pi_state = NULL;
1206         struct futex_hash_bucket *hb1, *hb2;
1207         struct plist_head *head1;
1208         struct futex_q *this, *next;
1209         u32 curval2;
1210
1211         if (requeue_pi) {
1212                 /*
1213                  * requeue_pi requires a pi_state, try to allocate it now
1214                  * without any locks in case it fails.
1215                  */
1216                 if (refill_pi_state_cache())
1217                         return -ENOMEM;
1218                 /*
1219                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1220                  * + nr_requeue, since it acquires the rt_mutex prior to
1221                  * returning to userspace, so as to not leave the rt_mutex with
1222                  * waiters and no owner.  However, second and third wake-ups
1223                  * cannot be predicted as they involve race conditions with the
1224                  * first wake and a fault while looking up the pi_state.  Both
1225                  * pthread_cond_signal() and pthread_cond_broadcast() should
1226                  * use nr_wake=1.
1227                  */
1228                 if (nr_wake != 1)
1229                         return -EINVAL;
1230         }
1231
1232 retry:
1233         if (pi_state != NULL) {
1234                 /*
1235                  * We will have to lookup the pi_state again, so free this one
1236                  * to keep the accounting correct.
1237                  */
1238                 free_pi_state(pi_state);
1239                 pi_state = NULL;
1240         }
1241
1242         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
1243         if (unlikely(ret != 0))
1244                 goto out;
1245         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
1246         if (unlikely(ret != 0))
1247                 goto out_put_key1;
1248
1249         hb1 = hash_futex(&key1);
1250         hb2 = hash_futex(&key2);
1251
1252 retry_private:
1253         double_lock_hb(hb1, hb2);
1254
1255         if (likely(cmpval != NULL)) {
1256                 u32 curval;
1257
1258                 ret = get_futex_value_locked(&curval, uaddr1);
1259
1260                 if (unlikely(ret)) {
1261                         double_unlock_hb(hb1, hb2);
1262
1263                         ret = get_user(curval, uaddr1);
1264                         if (ret)
1265                                 goto out_put_keys;
1266
1267                         if (!(flags & FLAGS_SHARED))
1268                                 goto retry_private;
1269
1270                         put_futex_key(&key2);
1271                         put_futex_key(&key1);
1272                         goto retry;
1273                 }
1274                 if (curval != *cmpval) {
1275                         ret = -EAGAIN;
1276                         goto out_unlock;
1277                 }
1278         }
1279
1280         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1281                 /*
1282                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1283                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1284                  * bit.  We force this here where we are able to easily handle
1285                  * faults rather in the requeue loop below.
1286                  */
1287                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1288                                                  &key2, &pi_state, nr_requeue);
1289
1290                 /*
1291                  * At this point the top_waiter has either taken uaddr2 or is
1292                  * waiting on it.  If the former, then the pi_state will not
1293                  * exist yet, look it up one more time to ensure we have a
1294                  * reference to it.
1295                  */
1296                 if (ret == 1) {
1297                         WARN_ON(pi_state);
1298                         drop_count++;
1299                         task_count++;
1300                         ret = get_futex_value_locked(&curval2, uaddr2);
1301                         if (!ret)
1302                                 ret = lookup_pi_state(curval2, hb2, &key2,
1303                                                       &pi_state);
1304                 }
1305
1306                 switch (ret) {
1307                 case 0:
1308                         break;
1309                 case -EFAULT:
1310                         double_unlock_hb(hb1, hb2);
1311                         put_futex_key(&key2);
1312                         put_futex_key(&key1);
1313                         ret = fault_in_user_writeable(uaddr2);
1314                         if (!ret)
1315                                 goto retry;
1316                         goto out;
1317                 case -EAGAIN:
1318                         /* The owner was exiting, try again. */
1319                         double_unlock_hb(hb1, hb2);
1320                         put_futex_key(&key2);
1321                         put_futex_key(&key1);
1322                         cond_resched();
1323                         goto retry;
1324                 default:
1325                         goto out_unlock;
1326                 }
1327         }
1328
1329         head1 = &hb1->chain;
1330         plist_for_each_entry_safe(this, next, head1, list) {
1331                 if (task_count - nr_wake >= nr_requeue)
1332                         break;
1333
1334                 if (!match_futex(&this->key, &key1))
1335                         continue;
1336
1337                 /*
1338                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1339                  * be paired with each other and no other futex ops.
1340                  */
1341                 if ((requeue_pi && !this->rt_waiter) ||
1342                     (!requeue_pi && this->rt_waiter)) {
1343                         ret = -EINVAL;
1344                         break;
1345                 }
1346
1347                 /*
1348                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1349                  * lock, we already woke the top_waiter.  If not, it will be
1350                  * woken by futex_unlock_pi().
1351                  */
1352                 if (++task_count <= nr_wake && !requeue_pi) {
1353                         wake_futex(this);
1354                         continue;
1355                 }
1356
1357                 /* Ensure we requeue to the expected futex for requeue_pi. */
1358                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1359                         ret = -EINVAL;
1360                         break;
1361                 }
1362
1363                 /*
1364                  * Requeue nr_requeue waiters and possibly one more in the case
1365                  * of requeue_pi if we couldn't acquire the lock atomically.
1366                  */
1367                 if (requeue_pi) {
1368                         /* Prepare the waiter to take the rt_mutex. */
1369                         atomic_inc(&pi_state->refcount);
1370                         this->pi_state = pi_state;
1371                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1372                                                         this->rt_waiter,
1373                                                         this->task, 1);
1374                         if (ret == 1) {
1375                                 /* We got the lock. */
1376                                 requeue_pi_wake_futex(this, &key2, hb2);
1377                                 drop_count++;
1378                                 continue;
1379                         } else if (ret) {
1380                                 /* -EDEADLK */
1381                                 this->pi_state = NULL;
1382                                 free_pi_state(pi_state);
1383                                 goto out_unlock;
1384                         }
1385                 }
1386                 requeue_futex(this, hb1, hb2, &key2);
1387                 drop_count++;
1388         }
1389
1390 out_unlock:
1391         double_unlock_hb(hb1, hb2);
1392
1393         /*
1394          * drop_futex_key_refs() must be called outside the spinlocks. During
1395          * the requeue we moved futex_q's from the hash bucket at key1 to the
1396          * one at key2 and updated their key pointer.  We no longer need to
1397          * hold the references to key1.
1398          */
1399         while (--drop_count >= 0)
1400                 drop_futex_key_refs(&key1);
1401
1402 out_put_keys:
1403         put_futex_key(&key2);
1404 out_put_key1:
1405         put_futex_key(&key1);
1406 out:
1407         if (pi_state != NULL)
1408                 free_pi_state(pi_state);
1409         return ret ? ret : task_count;
1410 }
1411
1412 /* The key must be already stored in q->key. */
1413 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1414         __acquires(&hb->lock)
1415 {
1416         struct futex_hash_bucket *hb;
1417
1418         hb = hash_futex(&q->key);
1419         q->lock_ptr = &hb->lock;
1420
1421         spin_lock(&hb->lock);
1422         return hb;
1423 }
1424
1425 static inline void
1426 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1427         __releases(&hb->lock)
1428 {
1429         spin_unlock(&hb->lock);
1430 }
1431
1432 /**
1433  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1434  * @q:  The futex_q to enqueue
1435  * @hb: The destination hash bucket
1436  *
1437  * The hb->lock must be held by the caller, and is released here. A call to
1438  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1439  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1440  * or nothing if the unqueue is done as part of the wake process and the unqueue
1441  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1442  * an example).
1443  */
1444 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1445         __releases(&hb->lock)
1446 {
1447         int prio;
1448
1449         /*
1450          * The priority used to register this element is
1451          * - either the real thread-priority for the real-time threads
1452          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1453          * - or MAX_RT_PRIO for non-RT threads.
1454          * Thus, all RT-threads are woken first in priority order, and
1455          * the others are woken last, in FIFO order.
1456          */
1457         prio = min(current->normal_prio, MAX_RT_PRIO);
1458
1459         plist_node_init(&q->list, prio);
1460 #ifdef CONFIG_DEBUG_PI_LIST
1461         q->list.plist.spinlock = &hb->lock;
1462 #endif
1463         plist_add(&q->list, &hb->chain);
1464         q->task = current;
1465         spin_unlock(&hb->lock);
1466 }
1467
1468 /**
1469  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1470  * @q:  The futex_q to unqueue
1471  *
1472  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1473  * be paired with exactly one earlier call to queue_me().
1474  *
1475  * Returns:
1476  *   1 - if the futex_q was still queued (and we removed unqueued it)
1477  *   0 - if the futex_q was already removed by the waking thread
1478  */
1479 static int unqueue_me(struct futex_q *q)
1480 {
1481         spinlock_t *lock_ptr;
1482         int ret = 0;
1483
1484         /* In the common case we don't take the spinlock, which is nice. */
1485 retry:
1486         lock_ptr = q->lock_ptr;
1487         barrier();
1488         if (lock_ptr != NULL) {
1489                 spin_lock(lock_ptr);
1490                 /*
1491                  * q->lock_ptr can change between reading it and
1492                  * spin_lock(), causing us to take the wrong lock.  This
1493                  * corrects the race condition.
1494                  *
1495                  * Reasoning goes like this: if we have the wrong lock,
1496                  * q->lock_ptr must have changed (maybe several times)
1497                  * between reading it and the spin_lock().  It can
1498                  * change again after the spin_lock() but only if it was
1499                  * already changed before the spin_lock().  It cannot,
1500                  * however, change back to the original value.  Therefore
1501                  * we can detect whether we acquired the correct lock.
1502                  */
1503                 if (unlikely(lock_ptr != q->lock_ptr)) {
1504                         spin_unlock(lock_ptr);
1505                         goto retry;
1506                 }
1507                 WARN_ON(plist_node_empty(&q->list));
1508                 plist_del(&q->list, &q->list.plist);
1509
1510                 BUG_ON(q->pi_state);
1511
1512                 spin_unlock(lock_ptr);
1513                 ret = 1;
1514         }
1515
1516         drop_futex_key_refs(&q->key);
1517         return ret;
1518 }
1519
1520 /*
1521  * PI futexes can not be requeued and must remove themself from the
1522  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1523  * and dropped here.
1524  */
1525 static void unqueue_me_pi(struct futex_q *q)
1526         __releases(q->lock_ptr)
1527 {
1528         WARN_ON(plist_node_empty(&q->list));
1529         plist_del(&q->list, &q->list.plist);
1530
1531         BUG_ON(!q->pi_state);
1532         free_pi_state(q->pi_state);
1533         q->pi_state = NULL;
1534
1535         spin_unlock(q->lock_ptr);
1536 }
1537
1538 /*
1539  * Fixup the pi_state owner with the new owner.
1540  *
1541  * Must be called with hash bucket lock held and mm->sem held for non
1542  * private futexes.
1543  */
1544 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1545                                 struct task_struct *newowner)
1546 {
1547         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1548         struct futex_pi_state *pi_state = q->pi_state;
1549         struct task_struct *oldowner = pi_state->owner;
1550         u32 uval, curval, newval;
1551         int ret;
1552
1553         /* Owner died? */
1554         if (!pi_state->owner)
1555                 newtid |= FUTEX_OWNER_DIED;
1556
1557         /*
1558          * We are here either because we stole the rtmutex from the
1559          * pending owner or we are the pending owner which failed to
1560          * get the rtmutex. We have to replace the pending owner TID
1561          * in the user space variable. This must be atomic as we have
1562          * to preserve the owner died bit here.
1563          *
1564          * Note: We write the user space value _before_ changing the pi_state
1565          * because we can fault here. Imagine swapped out pages or a fork
1566          * that marked all the anonymous memory readonly for cow.
1567          *
1568          * Modifying pi_state _before_ the user space value would
1569          * leave the pi_state in an inconsistent state when we fault
1570          * here, because we need to drop the hash bucket lock to
1571          * handle the fault. This might be observed in the PID check
1572          * in lookup_pi_state.
1573          */
1574 retry:
1575         if (get_futex_value_locked(&uval, uaddr))
1576                 goto handle_fault;
1577
1578         while (1) {
1579                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1580
1581                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1582
1583                 if (curval == -EFAULT)
1584                         goto handle_fault;
1585                 if (curval == uval)
1586                         break;
1587                 uval = curval;
1588         }
1589
1590         /*
1591          * We fixed up user space. Now we need to fix the pi_state
1592          * itself.
1593          */
1594         if (pi_state->owner != NULL) {
1595                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1596                 WARN_ON(list_empty(&pi_state->list));
1597                 list_del_init(&pi_state->list);
1598                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1599         }
1600
1601         pi_state->owner = newowner;
1602
1603         raw_spin_lock_irq(&newowner->pi_lock);
1604         WARN_ON(!list_empty(&pi_state->list));
1605         list_add(&pi_state->list, &newowner->pi_state_list);
1606         raw_spin_unlock_irq(&newowner->pi_lock);
1607         return 0;
1608
1609         /*
1610          * To handle the page fault we need to drop the hash bucket
1611          * lock here. That gives the other task (either the pending
1612          * owner itself or the task which stole the rtmutex) the
1613          * chance to try the fixup of the pi_state. So once we are
1614          * back from handling the fault we need to check the pi_state
1615          * after reacquiring the hash bucket lock and before trying to
1616          * do another fixup. When the fixup has been done already we
1617          * simply return.
1618          */
1619 handle_fault:
1620         spin_unlock(q->lock_ptr);
1621
1622         ret = fault_in_user_writeable(uaddr);
1623
1624         spin_lock(q->lock_ptr);
1625
1626         /*
1627          * Check if someone else fixed it for us:
1628          */
1629         if (pi_state->owner != oldowner)
1630                 return 0;
1631
1632         if (ret)
1633                 return ret;
1634
1635         goto retry;
1636 }
1637
1638 static long futex_wait_restart(struct restart_block *restart);
1639
1640 /**
1641  * fixup_owner() - Post lock pi_state and corner case management
1642  * @uaddr:      user address of the futex
1643  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1644  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1645  *
1646  * After attempting to lock an rt_mutex, this function is called to cleanup
1647  * the pi_state owner as well as handle race conditions that may allow us to
1648  * acquire the lock. Must be called with the hb lock held.
1649  *
1650  * Returns:
1651  *  1 - success, lock taken
1652  *  0 - success, lock not taken
1653  * <0 - on error (-EFAULT)
1654  */
1655 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1656 {
1657         struct task_struct *owner;
1658         int ret = 0;
1659
1660         if (locked) {
1661                 /*
1662                  * Got the lock. We might not be the anticipated owner if we
1663                  * did a lock-steal - fix up the PI-state in that case:
1664                  */
1665                 if (q->pi_state->owner != current)
1666                         ret = fixup_pi_state_owner(uaddr, q, current);
1667                 goto out;
1668         }
1669
1670         /*
1671          * Catch the rare case, where the lock was released when we were on the
1672          * way back before we locked the hash bucket.
1673          */
1674         if (q->pi_state->owner == current) {
1675                 /*
1676                  * Try to get the rt_mutex now. This might fail as some other
1677                  * task acquired the rt_mutex after we removed ourself from the
1678                  * rt_mutex waiters list.
1679                  */
1680                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1681                         locked = 1;
1682                         goto out;
1683                 }
1684
1685                 /*
1686                  * pi_state is incorrect, some other task did a lock steal and
1687                  * we returned due to timeout or signal without taking the
1688                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1689                  * locking, as the other task is now blocked on the hash bucket
1690                  * lock. Fix the state up.
1691                  */
1692                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1693                 ret = fixup_pi_state_owner(uaddr, q, owner);
1694                 goto out;
1695         }
1696
1697         /*
1698          * Paranoia check. If we did not take the lock, then we should not be
1699          * the owner, nor the pending owner, of the rt_mutex.
1700          */
1701         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1702                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1703                                 "pi-state %p\n", ret,
1704                                 q->pi_state->pi_mutex.owner,
1705                                 q->pi_state->owner);
1706
1707 out:
1708         return ret ? ret : locked;
1709 }
1710
1711 /**
1712  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1713  * @hb:         the futex hash bucket, must be locked by the caller
1714  * @q:          the futex_q to queue up on
1715  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1716  */
1717 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1718                                 struct hrtimer_sleeper *timeout)
1719 {
1720         /*
1721          * The task state is guaranteed to be set before another task can
1722          * wake it. set_current_state() is implemented using set_mb() and
1723          * queue_me() calls spin_unlock() upon completion, both serializing
1724          * access to the hash list and forcing another memory barrier.
1725          */
1726         set_current_state(TASK_INTERRUPTIBLE);
1727         queue_me(q, hb);
1728
1729         /* Arm the timer */
1730         if (timeout) {
1731                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1732                 if (!hrtimer_active(&timeout->timer))
1733                         timeout->task = NULL;
1734         }
1735
1736         /*
1737          * If we have been removed from the hash list, then another task
1738          * has tried to wake us, and we can skip the call to schedule().
1739          */
1740         if (likely(!plist_node_empty(&q->list))) {
1741                 /*
1742                  * If the timer has already expired, current will already be
1743                  * flagged for rescheduling. Only call schedule if there
1744                  * is no timeout, or if it has yet to expire.
1745                  */
1746                 if (!timeout || timeout->task)
1747                         schedule();
1748         }
1749         __set_current_state(TASK_RUNNING);
1750 }
1751
1752 /**
1753  * futex_wait_setup() - Prepare to wait on a futex
1754  * @uaddr:      the futex userspace address
1755  * @val:        the expected value
1756  * @flags:      futex flags (FLAGS_SHARED, etc.)
1757  * @q:          the associated futex_q
1758  * @hb:         storage for hash_bucket pointer to be returned to caller
1759  *
1760  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1761  * compare it with the expected value.  Handle atomic faults internally.
1762  * Return with the hb lock held and a q.key reference on success, and unlocked
1763  * with no q.key reference on failure.
1764  *
1765  * Returns:
1766  *  0 - uaddr contains val and hb has been locked
1767  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1768  */
1769 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1770                            struct futex_q *q, struct futex_hash_bucket **hb)
1771 {
1772         u32 uval;
1773         int ret;
1774
1775         /*
1776          * Access the page AFTER the hash-bucket is locked.
1777          * Order is important:
1778          *
1779          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1780          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1781          *
1782          * The basic logical guarantee of a futex is that it blocks ONLY
1783          * if cond(var) is known to be true at the time of blocking, for
1784          * any cond.  If we queued after testing *uaddr, that would open
1785          * a race condition where we could block indefinitely with
1786          * cond(var) false, which would violate the guarantee.
1787          *
1788          * A consequence is that futex_wait() can return zero and absorb
1789          * a wakeup when *uaddr != val on entry to the syscall.  This is
1790          * rare, but normal.
1791          */
1792 retry:
1793         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key);
1794         if (unlikely(ret != 0))
1795                 return ret;
1796
1797 retry_private:
1798         *hb = queue_lock(q);
1799
1800         ret = get_futex_value_locked(&uval, uaddr);
1801
1802         if (ret) {
1803                 queue_unlock(q, *hb);
1804
1805                 ret = get_user(uval, uaddr);
1806                 if (ret)
1807                         goto out;
1808
1809                 if (!(flags & FLAGS_SHARED))
1810                         goto retry_private;
1811
1812                 put_futex_key(&q->key);
1813                 goto retry;
1814         }
1815
1816         if (uval != val) {
1817                 queue_unlock(q, *hb);
1818                 ret = -EWOULDBLOCK;
1819         }
1820
1821 out:
1822         if (ret)
1823                 put_futex_key(&q->key);
1824         return ret;
1825 }
1826
1827 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1828                       ktime_t *abs_time, u32 bitset)
1829 {
1830         struct hrtimer_sleeper timeout, *to = NULL;
1831         struct restart_block *restart;
1832         struct futex_hash_bucket *hb;
1833         struct futex_q q = futex_q_init;
1834         int ret;
1835
1836         if (!bitset)
1837                 return -EINVAL;
1838         q.bitset = bitset;
1839
1840         if (abs_time) {
1841                 to = &timeout;
1842
1843                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1844                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
1845                                       HRTIMER_MODE_ABS);
1846                 hrtimer_init_sleeper(to, current);
1847                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1848                                              current->timer_slack_ns);
1849         }
1850
1851 retry:
1852         /*
1853          * Prepare to wait on uaddr. On success, holds hb lock and increments
1854          * q.key refs.
1855          */
1856         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1857         if (ret)
1858                 goto out;
1859
1860         /* queue_me and wait for wakeup, timeout, or a signal. */
1861         futex_wait_queue_me(hb, &q, to);
1862
1863         /* If we were woken (and unqueued), we succeeded, whatever. */
1864         ret = 0;
1865         /* unqueue_me() drops q.key ref */
1866         if (!unqueue_me(&q))
1867                 goto out;
1868         ret = -ETIMEDOUT;
1869         if (to && !to->task)
1870                 goto out;
1871
1872         /*
1873          * We expect signal_pending(current), but we might be the
1874          * victim of a spurious wakeup as well.
1875          */
1876         if (!signal_pending(current))
1877                 goto retry;
1878
1879         ret = -ERESTARTSYS;
1880         if (!abs_time)
1881                 goto out;
1882
1883         restart = &current_thread_info()->restart_block;
1884         restart->fn = futex_wait_restart;
1885         restart->futex.uaddr = uaddr;
1886         restart->futex.val = val;
1887         restart->futex.time = abs_time->tv64;
1888         restart->futex.bitset = bitset;
1889         restart->futex.flags = flags;
1890
1891         ret = -ERESTART_RESTARTBLOCK;
1892
1893 out:
1894         if (to) {
1895                 hrtimer_cancel(&to->timer);
1896                 destroy_hrtimer_on_stack(&to->timer);
1897         }
1898         return ret;
1899 }
1900
1901
1902 static long futex_wait_restart(struct restart_block *restart)
1903 {
1904         u32 __user *uaddr = restart->futex.uaddr;
1905         ktime_t t, *tp = NULL;
1906
1907         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1908                 t.tv64 = restart->futex.time;
1909                 tp = &t;
1910         }
1911         restart->fn = do_no_restart_syscall;
1912
1913         return (long)futex_wait(uaddr, restart->futex.flags,
1914                                 restart->futex.val, tp, restart->futex.bitset);
1915 }
1916
1917
1918 /*
1919  * Userspace tried a 0 -> TID atomic transition of the futex value
1920  * and failed. The kernel side here does the whole locking operation:
1921  * if there are waiters then it will block, it does PI, etc. (Due to
1922  * races the kernel might see a 0 value of the futex too.)
1923  */
1924 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1925                          ktime_t *time, int trylock)
1926 {
1927         struct hrtimer_sleeper timeout, *to = NULL;
1928         struct futex_hash_bucket *hb;
1929         struct futex_q q = futex_q_init;
1930         int res, ret;
1931
1932         if (refill_pi_state_cache())
1933                 return -ENOMEM;
1934
1935         if (time) {
1936                 to = &timeout;
1937                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1938                                       HRTIMER_MODE_ABS);
1939                 hrtimer_init_sleeper(to, current);
1940                 hrtimer_set_expires(&to->timer, *time);
1941         }
1942
1943 retry:
1944         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key);
1945         if (unlikely(ret != 0))
1946                 goto out;
1947
1948 retry_private:
1949         hb = queue_lock(&q);
1950
1951         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1952         if (unlikely(ret)) {
1953                 switch (ret) {
1954                 case 1:
1955                         /* We got the lock. */
1956                         ret = 0;
1957                         goto out_unlock_put_key;
1958                 case -EFAULT:
1959                         goto uaddr_faulted;
1960                 case -EAGAIN:
1961                         /*
1962                          * Task is exiting and we just wait for the
1963                          * exit to complete.
1964                          */
1965                         queue_unlock(&q, hb);
1966                         put_futex_key(&q.key);
1967                         cond_resched();
1968                         goto retry;
1969                 default:
1970                         goto out_unlock_put_key;
1971                 }
1972         }
1973
1974         /*
1975          * Only actually queue now that the atomic ops are done:
1976          */
1977         queue_me(&q, hb);
1978
1979         WARN_ON(!q.pi_state);
1980         /*
1981          * Block on the PI mutex:
1982          */
1983         if (!trylock)
1984                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1985         else {
1986                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1987                 /* Fixup the trylock return value: */
1988                 ret = ret ? 0 : -EWOULDBLOCK;
1989         }
1990
1991         spin_lock(q.lock_ptr);
1992         /*
1993          * Fixup the pi_state owner and possibly acquire the lock if we
1994          * haven't already.
1995          */
1996         res = fixup_owner(uaddr, &q, !ret);
1997         /*
1998          * If fixup_owner() returned an error, proprogate that.  If it acquired
1999          * the lock, clear our -ETIMEDOUT or -EINTR.
2000          */
2001         if (res)
2002                 ret = (res < 0) ? res : 0;
2003
2004         /*
2005          * If fixup_owner() faulted and was unable to handle the fault, unlock
2006          * it and return the fault to userspace.
2007          */
2008         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2009                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2010
2011         /* Unqueue and drop the lock */
2012         unqueue_me_pi(&q);
2013
2014         goto out_put_key;
2015
2016 out_unlock_put_key:
2017         queue_unlock(&q, hb);
2018
2019 out_put_key:
2020         put_futex_key(&q.key);
2021 out:
2022         if (to)
2023                 destroy_hrtimer_on_stack(&to->timer);
2024         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2025
2026 uaddr_faulted:
2027         queue_unlock(&q, hb);
2028
2029         ret = fault_in_user_writeable(uaddr);
2030         if (ret)
2031                 goto out_put_key;
2032
2033         if (!(flags & FLAGS_SHARED))
2034                 goto retry_private;
2035
2036         put_futex_key(&q.key);
2037         goto retry;
2038 }
2039
2040 /*
2041  * Userspace attempted a TID -> 0 atomic transition, and failed.
2042  * This is the in-kernel slowpath: we look up the PI state (if any),
2043  * and do the rt-mutex unlock.
2044  */
2045 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2046 {
2047         struct futex_hash_bucket *hb;
2048         struct futex_q *this, *next;
2049         u32 uval;
2050         struct plist_head *head;
2051         union futex_key key = FUTEX_KEY_INIT;
2052         int ret;
2053
2054 retry:
2055         if (get_user(uval, uaddr))
2056                 return -EFAULT;
2057         /*
2058          * We release only a lock we actually own:
2059          */
2060         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2061                 return -EPERM;
2062
2063         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
2064         if (unlikely(ret != 0))
2065                 goto out;
2066
2067         hb = hash_futex(&key);
2068         spin_lock(&hb->lock);
2069
2070         /*
2071          * To avoid races, try to do the TID -> 0 atomic transition
2072          * again. If it succeeds then we can return without waking
2073          * anyone else up:
2074          */
2075         if (!(uval & FUTEX_OWNER_DIED))
2076                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2077
2078
2079         if (unlikely(uval == -EFAULT))
2080                 goto pi_faulted;
2081         /*
2082          * Rare case: we managed to release the lock atomically,
2083          * no need to wake anyone else up:
2084          */
2085         if (unlikely(uval == task_pid_vnr(current)))
2086                 goto out_unlock;
2087
2088         /*
2089          * Ok, other tasks may need to be woken up - check waiters
2090          * and do the wakeup if necessary:
2091          */
2092         head = &hb->chain;
2093
2094         plist_for_each_entry_safe(this, next, head, list) {
2095                 if (!match_futex (&this->key, &key))
2096                         continue;
2097                 ret = wake_futex_pi(uaddr, uval, this);
2098                 /*
2099                  * The atomic access to the futex value
2100                  * generated a pagefault, so retry the
2101                  * user-access and the wakeup:
2102                  */
2103                 if (ret == -EFAULT)
2104                         goto pi_faulted;
2105                 goto out_unlock;
2106         }
2107         /*
2108          * No waiters - kernel unlocks the futex:
2109          */
2110         if (!(uval & FUTEX_OWNER_DIED)) {
2111                 ret = unlock_futex_pi(uaddr, uval);
2112                 if (ret == -EFAULT)
2113                         goto pi_faulted;
2114         }
2115
2116 out_unlock:
2117         spin_unlock(&hb->lock);
2118         put_futex_key(&key);
2119
2120 out:
2121         return ret;
2122
2123 pi_faulted:
2124         spin_unlock(&hb->lock);
2125         put_futex_key(&key);
2126
2127         ret = fault_in_user_writeable(uaddr);
2128         if (!ret)
2129                 goto retry;
2130
2131         return ret;
2132 }
2133
2134 /**
2135  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2136  * @hb:         the hash_bucket futex_q was original enqueued on
2137  * @q:          the futex_q woken while waiting to be requeued
2138  * @key2:       the futex_key of the requeue target futex
2139  * @timeout:    the timeout associated with the wait (NULL if none)
2140  *
2141  * Detect if the task was woken on the initial futex as opposed to the requeue
2142  * target futex.  If so, determine if it was a timeout or a signal that caused
2143  * the wakeup and return the appropriate error code to the caller.  Must be
2144  * called with the hb lock held.
2145  *
2146  * Returns
2147  *  0 - no early wakeup detected
2148  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2149  */
2150 static inline
2151 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2152                                    struct futex_q *q, union futex_key *key2,
2153                                    struct hrtimer_sleeper *timeout)
2154 {
2155         int ret = 0;
2156
2157         /*
2158          * With the hb lock held, we avoid races while we process the wakeup.
2159          * We only need to hold hb (and not hb2) to ensure atomicity as the
2160          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2161          * It can't be requeued from uaddr2 to something else since we don't
2162          * support a PI aware source futex for requeue.
2163          */
2164         if (!match_futex(&q->key, key2)) {
2165                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2166                 /*
2167                  * We were woken prior to requeue by a timeout or a signal.
2168                  * Unqueue the futex_q and determine which it was.
2169                  */
2170                 plist_del(&q->list, &q->list.plist);
2171
2172                 /* Handle spurious wakeups gracefully */
2173                 ret = -EWOULDBLOCK;
2174                 if (timeout && !timeout->task)
2175                         ret = -ETIMEDOUT;
2176                 else if (signal_pending(current))
2177                         ret = -ERESTARTNOINTR;
2178         }
2179         return ret;
2180 }
2181
2182 /**
2183  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2184  * @uaddr:      the futex we initially wait on (non-pi)
2185  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2186  *              the same type, no requeueing from private to shared, etc.
2187  * @val:        the expected value of uaddr
2188  * @abs_time:   absolute timeout
2189  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2190  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2191  * @uaddr2:     the pi futex we will take prior to returning to user-space
2192  *
2193  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2194  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2195  * complete the acquisition of the rt_mutex prior to returning to userspace.
2196  * This ensures the rt_mutex maintains an owner when it has waiters; without
2197  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2198  * need to.
2199  *
2200  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2201  * via the following:
2202  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2203  * 2) wakeup on uaddr2 after a requeue
2204  * 3) signal
2205  * 4) timeout
2206  *
2207  * If 3, cleanup and return -ERESTARTNOINTR.
2208  *
2209  * If 2, we may then block on trying to take the rt_mutex and return via:
2210  * 5) successful lock
2211  * 6) signal
2212  * 7) timeout
2213  * 8) other lock acquisition failure
2214  *
2215  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2216  *
2217  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2218  *
2219  * Returns:
2220  *  0 - On success
2221  * <0 - On error
2222  */
2223 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2224                                  u32 val, ktime_t *abs_time, u32 bitset,
2225                                  u32 __user *uaddr2)
2226 {
2227         struct hrtimer_sleeper timeout, *to = NULL;
2228         struct rt_mutex_waiter rt_waiter;
2229         struct rt_mutex *pi_mutex = NULL;
2230         struct futex_hash_bucket *hb;
2231         union futex_key key2 = FUTEX_KEY_INIT;
2232         struct futex_q q = futex_q_init;
2233         int res, ret;
2234
2235         if (!bitset)
2236                 return -EINVAL;
2237
2238         if (abs_time) {
2239                 to = &timeout;
2240                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2241                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2242                                       HRTIMER_MODE_ABS);
2243                 hrtimer_init_sleeper(to, current);
2244                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2245                                              current->timer_slack_ns);
2246         }
2247
2248         /*
2249          * The waiter is allocated on our stack, manipulated by the requeue
2250          * code while we sleep on uaddr.
2251          */
2252         debug_rt_mutex_init_waiter(&rt_waiter);
2253         rt_waiter.task = NULL;
2254
2255         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
2256         if (unlikely(ret != 0))
2257                 goto out;
2258
2259         q.bitset = bitset;
2260         q.rt_waiter = &rt_waiter;
2261         q.requeue_pi_key = &key2;
2262
2263         /*
2264          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2265          * count.
2266          */
2267         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2268         if (ret)
2269                 goto out_key2;
2270
2271         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2272         futex_wait_queue_me(hb, &q, to);
2273
2274         spin_lock(&hb->lock);
2275         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2276         spin_unlock(&hb->lock);
2277         if (ret)
2278                 goto out_put_keys;
2279
2280         /*
2281          * In order for us to be here, we know our q.key == key2, and since
2282          * we took the hb->lock above, we also know that futex_requeue() has
2283          * completed and we no longer have to concern ourselves with a wakeup
2284          * race with the atomic proxy lock acquisition by the requeue code. The
2285          * futex_requeue dropped our key1 reference and incremented our key2
2286          * reference count.
2287          */
2288
2289         /* Check if the requeue code acquired the second futex for us. */
2290         if (!q.rt_waiter) {
2291                 /*
2292                  * Got the lock. We might not be the anticipated owner if we
2293                  * did a lock-steal - fix up the PI-state in that case.
2294                  */
2295                 if (q.pi_state && (q.pi_state->owner != current)) {
2296                         spin_lock(q.lock_ptr);
2297                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2298                         spin_unlock(q.lock_ptr);
2299                 }
2300         } else {
2301                 /*
2302                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2303                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2304                  * the pi_state.
2305                  */
2306                 WARN_ON(!&q.pi_state);
2307                 pi_mutex = &q.pi_state->pi_mutex;
2308                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2309                 debug_rt_mutex_free_waiter(&rt_waiter);
2310
2311                 spin_lock(q.lock_ptr);
2312                 /*
2313                  * Fixup the pi_state owner and possibly acquire the lock if we
2314                  * haven't already.
2315                  */
2316                 res = fixup_owner(uaddr2, &q, !ret);
2317                 /*
2318                  * If fixup_owner() returned an error, proprogate that.  If it
2319                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2320                  */
2321                 if (res)
2322                         ret = (res < 0) ? res : 0;
2323
2324                 /* Unqueue and drop the lock. */
2325                 unqueue_me_pi(&q);
2326         }
2327
2328         /*
2329          * If fixup_pi_state_owner() faulted and was unable to handle the
2330          * fault, unlock the rt_mutex and return the fault to userspace.
2331          */
2332         if (ret == -EFAULT) {
2333                 if (rt_mutex_owner(pi_mutex) == current)
2334                         rt_mutex_unlock(pi_mutex);
2335         } else if (ret == -EINTR) {
2336                 /*
2337                  * We've already been requeued, but cannot restart by calling
2338                  * futex_lock_pi() directly. We could restart this syscall, but
2339                  * it would detect that the user space "val" changed and return
2340                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2341                  * -EWOULDBLOCK directly.
2342                  */
2343                 ret = -EWOULDBLOCK;
2344         }
2345
2346 out_put_keys:
2347         put_futex_key(&q.key);
2348 out_key2:
2349         put_futex_key(&key2);
2350
2351 out:
2352         if (to) {
2353                 hrtimer_cancel(&to->timer);
2354                 destroy_hrtimer_on_stack(&to->timer);
2355         }
2356         return ret;
2357 }
2358
2359 /*
2360  * Support for robust futexes: the kernel cleans up held futexes at
2361  * thread exit time.
2362  *
2363  * Implementation: user-space maintains a per-thread list of locks it
2364  * is holding. Upon do_exit(), the kernel carefully walks this list,
2365  * and marks all locks that are owned by this thread with the
2366  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2367  * always manipulated with the lock held, so the list is private and
2368  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2369  * field, to allow the kernel to clean up if the thread dies after
2370  * acquiring the lock, but just before it could have added itself to
2371  * the list. There can only be one such pending lock.
2372  */
2373
2374 /**
2375  * sys_set_robust_list() - Set the robust-futex list head of a task
2376  * @head:       pointer to the list-head
2377  * @len:        length of the list-head, as userspace expects
2378  */
2379 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2380                 size_t, len)
2381 {
2382         if (!futex_cmpxchg_enabled)
2383                 return -ENOSYS;
2384         /*
2385          * The kernel knows only one size for now:
2386          */
2387         if (unlikely(len != sizeof(*head)))
2388                 return -EINVAL;
2389
2390         current->robust_list = head;
2391
2392         return 0;
2393 }
2394
2395 /**
2396  * sys_get_robust_list() - Get the robust-futex list head of a task
2397  * @pid:        pid of the process [zero for current task]
2398  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2399  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2400  */
2401 SYSCALL_DEFINE3(get_robust_list, int, pid,
2402                 struct robust_list_head __user * __user *, head_ptr,
2403                 size_t __user *, len_ptr)
2404 {
2405         struct robust_list_head __user *head;
2406         unsigned long ret;
2407         const struct cred *cred = current_cred(), *pcred;
2408
2409         if (!futex_cmpxchg_enabled)
2410                 return -ENOSYS;
2411
2412         if (!pid)
2413                 head = current->robust_list;
2414         else {
2415                 struct task_struct *p;
2416
2417                 ret = -ESRCH;
2418                 rcu_read_lock();
2419                 p = find_task_by_vpid(pid);
2420                 if (!p)
2421                         goto err_unlock;
2422                 ret = -EPERM;
2423                 pcred = __task_cred(p);
2424                 if (cred->euid != pcred->euid &&
2425                     cred->euid != pcred->uid &&
2426                     !capable(CAP_SYS_PTRACE))
2427                         goto err_unlock;
2428                 head = p->robust_list;
2429                 rcu_read_unlock();
2430         }
2431
2432         if (put_user(sizeof(*head), len_ptr))
2433                 return -EFAULT;
2434         return put_user(head, head_ptr);
2435
2436 err_unlock:
2437         rcu_read_unlock();
2438
2439         return ret;
2440 }
2441
2442 /*
2443  * Process a futex-list entry, check whether it's owned by the
2444  * dying task, and do notification if so:
2445  */
2446 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2447 {
2448         u32 uval, nval, mval;
2449
2450 retry:
2451         if (get_user(uval, uaddr))
2452                 return -1;
2453
2454         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2455                 /*
2456                  * Ok, this dying thread is truly holding a futex
2457                  * of interest. Set the OWNER_DIED bit atomically
2458                  * via cmpxchg, and if the value had FUTEX_WAITERS
2459                  * set, wake up a waiter (if any). (We have to do a
2460                  * futex_wake() even if OWNER_DIED is already set -
2461                  * to handle the rare but possible case of recursive
2462                  * thread-death.) The rest of the cleanup is done in
2463                  * userspace.
2464                  */
2465                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2466                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2467
2468                 if (nval == -EFAULT)
2469                         return -1;
2470
2471                 if (nval != uval)
2472                         goto retry;
2473
2474                 /*
2475                  * Wake robust non-PI futexes here. The wakeup of
2476                  * PI futexes happens in exit_pi_state():
2477                  */
2478                 if (!pi && (uval & FUTEX_WAITERS))
2479                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2480         }
2481         return 0;
2482 }
2483
2484 /*
2485  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2486  */
2487 static inline int fetch_robust_entry(struct robust_list __user **entry,
2488                                      struct robust_list __user * __user *head,
2489                                      unsigned int *pi)
2490 {
2491         unsigned long uentry;
2492
2493         if (get_user(uentry, (unsigned long __user *)head))
2494                 return -EFAULT;
2495
2496         *entry = (void __user *)(uentry & ~1UL);
2497         *pi = uentry & 1;
2498
2499         return 0;
2500 }
2501
2502 /*
2503  * Walk curr->robust_list (very carefully, it's a userspace list!)
2504  * and mark any locks found there dead, and notify any waiters.
2505  *
2506  * We silently return on any sign of list-walking problem.
2507  */
2508 void exit_robust_list(struct task_struct *curr)
2509 {
2510         struct robust_list_head __user *head = curr->robust_list;
2511         struct robust_list __user *entry, *next_entry, *pending;
2512         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2513         unsigned int uninitialized_var(next_pi);
2514         unsigned long futex_offset;
2515         int rc;
2516
2517         if (!futex_cmpxchg_enabled)
2518                 return;
2519
2520         /*
2521          * Fetch the list head (which was registered earlier, via
2522          * sys_set_robust_list()):
2523          */
2524         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2525                 return;
2526         /*
2527          * Fetch the relative futex offset:
2528          */
2529         if (get_user(futex_offset, &head->futex_offset))
2530                 return;
2531         /*
2532          * Fetch any possibly pending lock-add first, and handle it
2533          * if it exists:
2534          */
2535         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2536                 return;
2537
2538         next_entry = NULL;      /* avoid warning with gcc */
2539         while (entry != &head->list) {
2540                 /*
2541                  * Fetch the next entry in the list before calling
2542                  * handle_futex_death:
2543                  */
2544                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2545                 /*
2546                  * A pending lock might already be on the list, so
2547                  * don't process it twice:
2548                  */
2549                 if (entry != pending)
2550                         if (handle_futex_death((void __user *)entry + futex_offset,
2551                                                 curr, pi))
2552                                 return;
2553                 if (rc)
2554                         return;
2555                 entry = next_entry;
2556                 pi = next_pi;
2557                 /*
2558                  * Avoid excessively long or circular lists:
2559                  */
2560                 if (!--limit)
2561                         break;
2562
2563                 cond_resched();
2564         }
2565
2566         if (pending)
2567                 handle_futex_death((void __user *)pending + futex_offset,
2568                                    curr, pip);
2569 }
2570
2571 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2572                 u32 __user *uaddr2, u32 val2, u32 val3)
2573 {
2574         int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2575         unsigned int flags = 0;
2576
2577         if (!(op & FUTEX_PRIVATE_FLAG))
2578                 flags |= FLAGS_SHARED;
2579
2580         if (op & FUTEX_CLOCK_REALTIME) {
2581                 flags |= FLAGS_CLOCKRT;
2582                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2583                         return -ENOSYS;
2584         }
2585
2586         switch (cmd) {
2587         case FUTEX_WAIT:
2588                 val3 = FUTEX_BITSET_MATCH_ANY;
2589         case FUTEX_WAIT_BITSET:
2590                 ret = futex_wait(uaddr, flags, val, timeout, val3);
2591                 break;
2592         case FUTEX_WAKE:
2593                 val3 = FUTEX_BITSET_MATCH_ANY;
2594         case FUTEX_WAKE_BITSET:
2595                 ret = futex_wake(uaddr, flags, val, val3);
2596                 break;
2597         case FUTEX_REQUEUE:
2598                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2599                 break;
2600         case FUTEX_CMP_REQUEUE:
2601                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2602                 break;
2603         case FUTEX_WAKE_OP:
2604                 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2605                 break;
2606         case FUTEX_LOCK_PI:
2607                 if (futex_cmpxchg_enabled)
2608                         ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2609                 break;
2610         case FUTEX_UNLOCK_PI:
2611                 if (futex_cmpxchg_enabled)
2612                         ret = futex_unlock_pi(uaddr, flags);
2613                 break;
2614         case FUTEX_TRYLOCK_PI:
2615                 if (futex_cmpxchg_enabled)
2616                         ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2617                 break;
2618         case FUTEX_WAIT_REQUEUE_PI:
2619                 val3 = FUTEX_BITSET_MATCH_ANY;
2620                 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2621                                             uaddr2);
2622                 break;
2623         case FUTEX_CMP_REQUEUE_PI:
2624                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2625                 break;
2626         default:
2627                 ret = -ENOSYS;
2628         }
2629         return ret;
2630 }
2631
2632
2633 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2634                 struct timespec __user *, utime, u32 __user *, uaddr2,
2635                 u32, val3)
2636 {
2637         struct timespec ts;
2638         ktime_t t, *tp = NULL;
2639         u32 val2 = 0;
2640         int cmd = op & FUTEX_CMD_MASK;
2641
2642         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2643                       cmd == FUTEX_WAIT_BITSET ||
2644                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2645                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2646                         return -EFAULT;
2647                 if (!timespec_valid(&ts))
2648                         return -EINVAL;
2649
2650                 t = timespec_to_ktime(ts);
2651                 if (cmd == FUTEX_WAIT)
2652                         t = ktime_add_safe(ktime_get(), t);
2653                 tp = &t;
2654         }
2655         /*
2656          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2657          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2658          */
2659         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2660             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2661                 val2 = (u32) (unsigned long) utime;
2662
2663         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2664 }
2665
2666 static int __init futex_init(void)
2667 {
2668         u32 curval;
2669         int i;
2670
2671         /*
2672          * This will fail and we want it. Some arch implementations do
2673          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2674          * functionality. We want to know that before we call in any
2675          * of the complex code paths. Also we want to prevent
2676          * registration of robust lists in that case. NULL is
2677          * guaranteed to fault and we get -EFAULT on functional
2678          * implementation, the non-functional ones will return
2679          * -ENOSYS.
2680          */
2681         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2682         if (curval == -EFAULT)
2683                 futex_cmpxchg_enabled = 1;
2684
2685         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2686                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2687                 spin_lock_init(&futex_queues[i].lock);
2688         }
2689
2690         return 0;
2691 }
2692 __initcall(futex_init);