userns: user namespaces: convert several capable() calls
[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 int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
385                                       u32 uval, u32 newval)
386 {
387         int ret;
388
389         pagefault_disable();
390         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
391         pagefault_enable();
392
393         return ret;
394 }
395
396 static int get_futex_value_locked(u32 *dest, u32 __user *from)
397 {
398         int ret;
399
400         pagefault_disable();
401         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
402         pagefault_enable();
403
404         return ret ? -EFAULT : 0;
405 }
406
407
408 /*
409  * PI code:
410  */
411 static int refill_pi_state_cache(void)
412 {
413         struct futex_pi_state *pi_state;
414
415         if (likely(current->pi_state_cache))
416                 return 0;
417
418         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
419
420         if (!pi_state)
421                 return -ENOMEM;
422
423         INIT_LIST_HEAD(&pi_state->list);
424         /* pi_mutex gets initialized later */
425         pi_state->owner = NULL;
426         atomic_set(&pi_state->refcount, 1);
427         pi_state->key = FUTEX_KEY_INIT;
428
429         current->pi_state_cache = pi_state;
430
431         return 0;
432 }
433
434 static struct futex_pi_state * alloc_pi_state(void)
435 {
436         struct futex_pi_state *pi_state = current->pi_state_cache;
437
438         WARN_ON(!pi_state);
439         current->pi_state_cache = NULL;
440
441         return pi_state;
442 }
443
444 static void free_pi_state(struct futex_pi_state *pi_state)
445 {
446         if (!atomic_dec_and_test(&pi_state->refcount))
447                 return;
448
449         /*
450          * If pi_state->owner is NULL, the owner is most probably dying
451          * and has cleaned up the pi_state already
452          */
453         if (pi_state->owner) {
454                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
455                 list_del_init(&pi_state->list);
456                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
457
458                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
459         }
460
461         if (current->pi_state_cache)
462                 kfree(pi_state);
463         else {
464                 /*
465                  * pi_state->list is already empty.
466                  * clear pi_state->owner.
467                  * refcount is at 0 - put it back to 1.
468                  */
469                 pi_state->owner = NULL;
470                 atomic_set(&pi_state->refcount, 1);
471                 current->pi_state_cache = pi_state;
472         }
473 }
474
475 /*
476  * Look up the task based on what TID userspace gave us.
477  * We dont trust it.
478  */
479 static struct task_struct * futex_find_get_task(pid_t pid)
480 {
481         struct task_struct *p;
482
483         rcu_read_lock();
484         p = find_task_by_vpid(pid);
485         if (p)
486                 get_task_struct(p);
487
488         rcu_read_unlock();
489
490         return p;
491 }
492
493 /*
494  * This task is holding PI mutexes at exit time => bad.
495  * Kernel cleans up PI-state, but userspace is likely hosed.
496  * (Robust-futex cleanup is separate and might save the day for userspace.)
497  */
498 void exit_pi_state_list(struct task_struct *curr)
499 {
500         struct list_head *next, *head = &curr->pi_state_list;
501         struct futex_pi_state *pi_state;
502         struct futex_hash_bucket *hb;
503         union futex_key key = FUTEX_KEY_INIT;
504
505         if (!futex_cmpxchg_enabled)
506                 return;
507         /*
508          * We are a ZOMBIE and nobody can enqueue itself on
509          * pi_state_list anymore, but we have to be careful
510          * versus waiters unqueueing themselves:
511          */
512         raw_spin_lock_irq(&curr->pi_lock);
513         while (!list_empty(head)) {
514
515                 next = head->next;
516                 pi_state = list_entry(next, struct futex_pi_state, list);
517                 key = pi_state->key;
518                 hb = hash_futex(&key);
519                 raw_spin_unlock_irq(&curr->pi_lock);
520
521                 spin_lock(&hb->lock);
522
523                 raw_spin_lock_irq(&curr->pi_lock);
524                 /*
525                  * We dropped the pi-lock, so re-check whether this
526                  * task still owns the PI-state:
527                  */
528                 if (head->next != next) {
529                         spin_unlock(&hb->lock);
530                         continue;
531                 }
532
533                 WARN_ON(pi_state->owner != curr);
534                 WARN_ON(list_empty(&pi_state->list));
535                 list_del_init(&pi_state->list);
536                 pi_state->owner = NULL;
537                 raw_spin_unlock_irq(&curr->pi_lock);
538
539                 rt_mutex_unlock(&pi_state->pi_mutex);
540
541                 spin_unlock(&hb->lock);
542
543                 raw_spin_lock_irq(&curr->pi_lock);
544         }
545         raw_spin_unlock_irq(&curr->pi_lock);
546 }
547
548 static int
549 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
550                 union futex_key *key, struct futex_pi_state **ps)
551 {
552         struct futex_pi_state *pi_state = NULL;
553         struct futex_q *this, *next;
554         struct plist_head *head;
555         struct task_struct *p;
556         pid_t pid = uval & FUTEX_TID_MASK;
557
558         head = &hb->chain;
559
560         plist_for_each_entry_safe(this, next, head, list) {
561                 if (match_futex(&this->key, key)) {
562                         /*
563                          * Another waiter already exists - bump up
564                          * the refcount and return its pi_state:
565                          */
566                         pi_state = this->pi_state;
567                         /*
568                          * Userspace might have messed up non-PI and PI futexes
569                          */
570                         if (unlikely(!pi_state))
571                                 return -EINVAL;
572
573                         WARN_ON(!atomic_read(&pi_state->refcount));
574
575                         /*
576                          * When pi_state->owner is NULL then the owner died
577                          * and another waiter is on the fly. pi_state->owner
578                          * is fixed up by the task which acquires
579                          * pi_state->rt_mutex.
580                          *
581                          * We do not check for pid == 0 which can happen when
582                          * the owner died and robust_list_exit() cleared the
583                          * TID.
584                          */
585                         if (pid && pi_state->owner) {
586                                 /*
587                                  * Bail out if user space manipulated the
588                                  * futex value.
589                                  */
590                                 if (pid != task_pid_vnr(pi_state->owner))
591                                         return -EINVAL;
592                         }
593
594                         atomic_inc(&pi_state->refcount);
595                         *ps = pi_state;
596
597                         return 0;
598                 }
599         }
600
601         /*
602          * We are the first waiter - try to look up the real owner and attach
603          * the new pi_state to it, but bail out when TID = 0
604          */
605         if (!pid)
606                 return -ESRCH;
607         p = futex_find_get_task(pid);
608         if (!p)
609                 return -ESRCH;
610
611         /*
612          * We need to look at the task state flags to figure out,
613          * whether the task is exiting. To protect against the do_exit
614          * change of the task flags, we do this protected by
615          * p->pi_lock:
616          */
617         raw_spin_lock_irq(&p->pi_lock);
618         if (unlikely(p->flags & PF_EXITING)) {
619                 /*
620                  * The task is on the way out. When PF_EXITPIDONE is
621                  * set, we know that the task has finished the
622                  * cleanup:
623                  */
624                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
625
626                 raw_spin_unlock_irq(&p->pi_lock);
627                 put_task_struct(p);
628                 return ret;
629         }
630
631         pi_state = alloc_pi_state();
632
633         /*
634          * Initialize the pi_mutex in locked state and make 'p'
635          * the owner of it:
636          */
637         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
638
639         /* Store the key for possible exit cleanups: */
640         pi_state->key = *key;
641
642         WARN_ON(!list_empty(&pi_state->list));
643         list_add(&pi_state->list, &p->pi_state_list);
644         pi_state->owner = p;
645         raw_spin_unlock_irq(&p->pi_lock);
646
647         put_task_struct(p);
648
649         *ps = pi_state;
650
651         return 0;
652 }
653
654 /**
655  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
656  * @uaddr:              the pi futex user address
657  * @hb:                 the pi futex hash bucket
658  * @key:                the futex key associated with uaddr and hb
659  * @ps:                 the pi_state pointer where we store the result of the
660  *                      lookup
661  * @task:               the task to perform the atomic lock work for.  This will
662  *                      be "current" except in the case of requeue pi.
663  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
664  *
665  * Returns:
666  *  0 - ready to wait
667  *  1 - acquired the lock
668  * <0 - error
669  *
670  * The hb->lock and futex_key refs shall be held by the caller.
671  */
672 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
673                                 union futex_key *key,
674                                 struct futex_pi_state **ps,
675                                 struct task_struct *task, int set_waiters)
676 {
677         int lock_taken, ret, ownerdied = 0;
678         u32 uval, newval, curval, vpid = task_pid_vnr(task);
679
680 retry:
681         ret = lock_taken = 0;
682
683         /*
684          * To avoid races, we attempt to take the lock here again
685          * (by doing a 0 -> TID atomic cmpxchg), while holding all
686          * the locks. It will most likely not succeed.
687          */
688         newval = vpid;
689         if (set_waiters)
690                 newval |= FUTEX_WAITERS;
691
692         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
693                 return -EFAULT;
694
695         /*
696          * Detect deadlocks.
697          */
698         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
699                 return -EDEADLK;
700
701         /*
702          * Surprise - we got the lock. Just return to userspace:
703          */
704         if (unlikely(!curval))
705                 return 1;
706
707         uval = curval;
708
709         /*
710          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
711          * to wake at the next unlock.
712          */
713         newval = curval | FUTEX_WAITERS;
714
715         /*
716          * There are two cases, where a futex might have no owner (the
717          * owner TID is 0): OWNER_DIED. We take over the futex in this
718          * case. We also do an unconditional take over, when the owner
719          * of the futex died.
720          *
721          * This is safe as we are protected by the hash bucket lock !
722          */
723         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
724                 /* Keep the OWNER_DIED bit */
725                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
726                 ownerdied = 0;
727                 lock_taken = 1;
728         }
729
730         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
731                 return -EFAULT;
732         if (unlikely(curval != uval))
733                 goto retry;
734
735         /*
736          * We took the lock due to owner died take over.
737          */
738         if (unlikely(lock_taken))
739                 return 1;
740
741         /*
742          * We dont have the lock. Look up the PI state (or create it if
743          * we are the first waiter):
744          */
745         ret = lookup_pi_state(uval, hb, key, ps);
746
747         if (unlikely(ret)) {
748                 switch (ret) {
749                 case -ESRCH:
750                         /*
751                          * No owner found for this futex. Check if the
752                          * OWNER_DIED bit is set to figure out whether
753                          * this is a robust futex or not.
754                          */
755                         if (get_futex_value_locked(&curval, uaddr))
756                                 return -EFAULT;
757
758                         /*
759                          * We simply start over in case of a robust
760                          * futex. The code above will take the futex
761                          * and return happy.
762                          */
763                         if (curval & FUTEX_OWNER_DIED) {
764                                 ownerdied = 1;
765                                 goto retry;
766                         }
767                 default:
768                         break;
769                 }
770         }
771
772         return ret;
773 }
774
775 /**
776  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
777  * @q:  The futex_q to unqueue
778  *
779  * The q->lock_ptr must not be NULL and must be held by the caller.
780  */
781 static void __unqueue_futex(struct futex_q *q)
782 {
783         struct futex_hash_bucket *hb;
784
785         if (WARN_ON(!q->lock_ptr || !spin_is_locked(q->lock_ptr)
786                         || plist_node_empty(&q->list)))
787                 return;
788
789         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
790         plist_del(&q->list, &hb->chain);
791 }
792
793 /*
794  * The hash bucket lock must be held when this is called.
795  * Afterwards, the futex_q must not be accessed.
796  */
797 static void wake_futex(struct futex_q *q)
798 {
799         struct task_struct *p = q->task;
800
801         /*
802          * We set q->lock_ptr = NULL _before_ we wake up the task. If
803          * a non-futex wake up happens on another CPU then the task
804          * might exit and p would dereference a non-existing task
805          * struct. Prevent this by holding a reference on p across the
806          * wake up.
807          */
808         get_task_struct(p);
809
810         __unqueue_futex(q);
811         /*
812          * The waiting task can free the futex_q as soon as
813          * q->lock_ptr = NULL is written, without taking any locks. A
814          * memory barrier is required here to prevent the following
815          * store to lock_ptr from getting ahead of the plist_del.
816          */
817         smp_wmb();
818         q->lock_ptr = NULL;
819
820         wake_up_state(p, TASK_NORMAL);
821         put_task_struct(p);
822 }
823
824 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
825 {
826         struct task_struct *new_owner;
827         struct futex_pi_state *pi_state = this->pi_state;
828         u32 curval, newval;
829
830         if (!pi_state)
831                 return -EINVAL;
832
833         /*
834          * If current does not own the pi_state then the futex is
835          * inconsistent and user space fiddled with the futex value.
836          */
837         if (pi_state->owner != current)
838                 return -EINVAL;
839
840         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
841         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
842
843         /*
844          * It is possible that the next waiter (the one that brought
845          * this owner to the kernel) timed out and is no longer
846          * waiting on the lock.
847          */
848         if (!new_owner)
849                 new_owner = this->task;
850
851         /*
852          * We pass it to the next owner. (The WAITERS bit is always
853          * kept enabled while there is PI state around. We must also
854          * preserve the owner died bit.)
855          */
856         if (!(uval & FUTEX_OWNER_DIED)) {
857                 int ret = 0;
858
859                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
860
861                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
862                         ret = -EFAULT;
863                 else if (curval != uval)
864                         ret = -EINVAL;
865                 if (ret) {
866                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
867                         return ret;
868                 }
869         }
870
871         raw_spin_lock_irq(&pi_state->owner->pi_lock);
872         WARN_ON(list_empty(&pi_state->list));
873         list_del_init(&pi_state->list);
874         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
875
876         raw_spin_lock_irq(&new_owner->pi_lock);
877         WARN_ON(!list_empty(&pi_state->list));
878         list_add(&pi_state->list, &new_owner->pi_state_list);
879         pi_state->owner = new_owner;
880         raw_spin_unlock_irq(&new_owner->pi_lock);
881
882         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
883         rt_mutex_unlock(&pi_state->pi_mutex);
884
885         return 0;
886 }
887
888 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
889 {
890         u32 oldval;
891
892         /*
893          * There is no waiter, so we unlock the futex. The owner died
894          * bit has not to be preserved here. We are the owner:
895          */
896         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
897                 return -EFAULT;
898         if (oldval != uval)
899                 return -EAGAIN;
900
901         return 0;
902 }
903
904 /*
905  * Express the locking dependencies for lockdep:
906  */
907 static inline void
908 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
909 {
910         if (hb1 <= hb2) {
911                 spin_lock(&hb1->lock);
912                 if (hb1 < hb2)
913                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
914         } else { /* hb1 > hb2 */
915                 spin_lock(&hb2->lock);
916                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
917         }
918 }
919
920 static inline void
921 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
922 {
923         spin_unlock(&hb1->lock);
924         if (hb1 != hb2)
925                 spin_unlock(&hb2->lock);
926 }
927
928 /*
929  * Wake up waiters matching bitset queued on this futex (uaddr).
930  */
931 static int
932 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
933 {
934         struct futex_hash_bucket *hb;
935         struct futex_q *this, *next;
936         struct plist_head *head;
937         union futex_key key = FUTEX_KEY_INIT;
938         int ret;
939
940         if (!bitset)
941                 return -EINVAL;
942
943         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
944         if (unlikely(ret != 0))
945                 goto out;
946
947         hb = hash_futex(&key);
948         spin_lock(&hb->lock);
949         head = &hb->chain;
950
951         plist_for_each_entry_safe(this, next, head, list) {
952                 if (match_futex (&this->key, &key)) {
953                         if (this->pi_state || this->rt_waiter) {
954                                 ret = -EINVAL;
955                                 break;
956                         }
957
958                         /* Check if one of the bits is set in both bitsets */
959                         if (!(this->bitset & bitset))
960                                 continue;
961
962                         wake_futex(this);
963                         if (++ret >= nr_wake)
964                                 break;
965                 }
966         }
967
968         spin_unlock(&hb->lock);
969         put_futex_key(&key);
970 out:
971         return ret;
972 }
973
974 /*
975  * Wake up all waiters hashed on the physical page that is mapped
976  * to this virtual address:
977  */
978 static int
979 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
980               int nr_wake, int nr_wake2, int op)
981 {
982         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
983         struct futex_hash_bucket *hb1, *hb2;
984         struct plist_head *head;
985         struct futex_q *this, *next;
986         int ret, op_ret;
987
988 retry:
989         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
990         if (unlikely(ret != 0))
991                 goto out;
992         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
993         if (unlikely(ret != 0))
994                 goto out_put_key1;
995
996         hb1 = hash_futex(&key1);
997         hb2 = hash_futex(&key2);
998
999 retry_private:
1000         double_lock_hb(hb1, hb2);
1001         op_ret = futex_atomic_op_inuser(op, uaddr2);
1002         if (unlikely(op_ret < 0)) {
1003
1004                 double_unlock_hb(hb1, hb2);
1005
1006 #ifndef CONFIG_MMU
1007                 /*
1008                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1009                  * but we might get them from range checking
1010                  */
1011                 ret = op_ret;
1012                 goto out_put_keys;
1013 #endif
1014
1015                 if (unlikely(op_ret != -EFAULT)) {
1016                         ret = op_ret;
1017                         goto out_put_keys;
1018                 }
1019
1020                 ret = fault_in_user_writeable(uaddr2);
1021                 if (ret)
1022                         goto out_put_keys;
1023
1024                 if (!(flags & FLAGS_SHARED))
1025                         goto retry_private;
1026
1027                 put_futex_key(&key2);
1028                 put_futex_key(&key1);
1029                 goto retry;
1030         }
1031
1032         head = &hb1->chain;
1033
1034         plist_for_each_entry_safe(this, next, head, list) {
1035                 if (match_futex (&this->key, &key1)) {
1036                         wake_futex(this);
1037                         if (++ret >= nr_wake)
1038                                 break;
1039                 }
1040         }
1041
1042         if (op_ret > 0) {
1043                 head = &hb2->chain;
1044
1045                 op_ret = 0;
1046                 plist_for_each_entry_safe(this, next, head, list) {
1047                         if (match_futex (&this->key, &key2)) {
1048                                 wake_futex(this);
1049                                 if (++op_ret >= nr_wake2)
1050                                         break;
1051                         }
1052                 }
1053                 ret += op_ret;
1054         }
1055
1056         double_unlock_hb(hb1, hb2);
1057 out_put_keys:
1058         put_futex_key(&key2);
1059 out_put_key1:
1060         put_futex_key(&key1);
1061 out:
1062         return ret;
1063 }
1064
1065 /**
1066  * requeue_futex() - Requeue a futex_q from one hb to another
1067  * @q:          the futex_q to requeue
1068  * @hb1:        the source hash_bucket
1069  * @hb2:        the target hash_bucket
1070  * @key2:       the new key for the requeued futex_q
1071  */
1072 static inline
1073 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1074                    struct futex_hash_bucket *hb2, union futex_key *key2)
1075 {
1076
1077         /*
1078          * If key1 and key2 hash to the same bucket, no need to
1079          * requeue.
1080          */
1081         if (likely(&hb1->chain != &hb2->chain)) {
1082                 plist_del(&q->list, &hb1->chain);
1083                 plist_add(&q->list, &hb2->chain);
1084                 q->lock_ptr = &hb2->lock;
1085         }
1086         get_futex_key_refs(key2);
1087         q->key = *key2;
1088 }
1089
1090 /**
1091  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1092  * @q:          the futex_q
1093  * @key:        the key of the requeue target futex
1094  * @hb:         the hash_bucket of the requeue target futex
1095  *
1096  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1097  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1098  * to the requeue target futex so the waiter can detect the wakeup on the right
1099  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1100  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1101  * to protect access to the pi_state to fixup the owner later.  Must be called
1102  * with both q->lock_ptr and hb->lock held.
1103  */
1104 static inline
1105 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1106                            struct futex_hash_bucket *hb)
1107 {
1108         get_futex_key_refs(key);
1109         q->key = *key;
1110
1111         __unqueue_futex(q);
1112
1113         WARN_ON(!q->rt_waiter);
1114         q->rt_waiter = NULL;
1115
1116         q->lock_ptr = &hb->lock;
1117
1118         wake_up_state(q->task, TASK_NORMAL);
1119 }
1120
1121 /**
1122  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1123  * @pifutex:            the user address of the to futex
1124  * @hb1:                the from futex hash bucket, must be locked by the caller
1125  * @hb2:                the to futex hash bucket, must be locked by the caller
1126  * @key1:               the from futex key
1127  * @key2:               the to futex key
1128  * @ps:                 address to store the pi_state pointer
1129  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1130  *
1131  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1132  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1133  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1134  * hb1 and hb2 must be held by the caller.
1135  *
1136  * Returns:
1137  *  0 - failed to acquire the lock atomicly
1138  *  1 - acquired the lock
1139  * <0 - error
1140  */
1141 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1142                                  struct futex_hash_bucket *hb1,
1143                                  struct futex_hash_bucket *hb2,
1144                                  union futex_key *key1, union futex_key *key2,
1145                                  struct futex_pi_state **ps, int set_waiters)
1146 {
1147         struct futex_q *top_waiter = NULL;
1148         u32 curval;
1149         int ret;
1150
1151         if (get_futex_value_locked(&curval, pifutex))
1152                 return -EFAULT;
1153
1154         /*
1155          * Find the top_waiter and determine if there are additional waiters.
1156          * If the caller intends to requeue more than 1 waiter to pifutex,
1157          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1158          * as we have means to handle the possible fault.  If not, don't set
1159          * the bit unecessarily as it will force the subsequent unlock to enter
1160          * the kernel.
1161          */
1162         top_waiter = futex_top_waiter(hb1, key1);
1163
1164         /* There are no waiters, nothing for us to do. */
1165         if (!top_waiter)
1166                 return 0;
1167
1168         /* Ensure we requeue to the expected futex. */
1169         if (!match_futex(top_waiter->requeue_pi_key, key2))
1170                 return -EINVAL;
1171
1172         /*
1173          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1174          * the contended case or if set_waiters is 1.  The pi_state is returned
1175          * in ps in contended cases.
1176          */
1177         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1178                                    set_waiters);
1179         if (ret == 1)
1180                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1181
1182         return ret;
1183 }
1184
1185 /**
1186  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1187  * @uaddr1:     source futex user address
1188  * @flags:      futex flags (FLAGS_SHARED, etc.)
1189  * @uaddr2:     target futex user address
1190  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1191  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1192  * @cmpval:     @uaddr1 expected value (or %NULL)
1193  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1194  *              pi futex (pi to pi requeue is not supported)
1195  *
1196  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1197  * uaddr2 atomically on behalf of the top waiter.
1198  *
1199  * Returns:
1200  * >=0 - on success, the number of tasks requeued or woken
1201  *  <0 - on error
1202  */
1203 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1204                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1205                          u32 *cmpval, int requeue_pi)
1206 {
1207         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1208         int drop_count = 0, task_count = 0, ret;
1209         struct futex_pi_state *pi_state = NULL;
1210         struct futex_hash_bucket *hb1, *hb2;
1211         struct plist_head *head1;
1212         struct futex_q *this, *next;
1213         u32 curval2;
1214
1215         if (requeue_pi) {
1216                 /*
1217                  * requeue_pi requires a pi_state, try to allocate it now
1218                  * without any locks in case it fails.
1219                  */
1220                 if (refill_pi_state_cache())
1221                         return -ENOMEM;
1222                 /*
1223                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1224                  * + nr_requeue, since it acquires the rt_mutex prior to
1225                  * returning to userspace, so as to not leave the rt_mutex with
1226                  * waiters and no owner.  However, second and third wake-ups
1227                  * cannot be predicted as they involve race conditions with the
1228                  * first wake and a fault while looking up the pi_state.  Both
1229                  * pthread_cond_signal() and pthread_cond_broadcast() should
1230                  * use nr_wake=1.
1231                  */
1232                 if (nr_wake != 1)
1233                         return -EINVAL;
1234         }
1235
1236 retry:
1237         if (pi_state != NULL) {
1238                 /*
1239                  * We will have to lookup the pi_state again, so free this one
1240                  * to keep the accounting correct.
1241                  */
1242                 free_pi_state(pi_state);
1243                 pi_state = NULL;
1244         }
1245
1246         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
1247         if (unlikely(ret != 0))
1248                 goto out;
1249         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
1250         if (unlikely(ret != 0))
1251                 goto out_put_key1;
1252
1253         hb1 = hash_futex(&key1);
1254         hb2 = hash_futex(&key2);
1255
1256 retry_private:
1257         double_lock_hb(hb1, hb2);
1258
1259         if (likely(cmpval != NULL)) {
1260                 u32 curval;
1261
1262                 ret = get_futex_value_locked(&curval, uaddr1);
1263
1264                 if (unlikely(ret)) {
1265                         double_unlock_hb(hb1, hb2);
1266
1267                         ret = get_user(curval, uaddr1);
1268                         if (ret)
1269                                 goto out_put_keys;
1270
1271                         if (!(flags & FLAGS_SHARED))
1272                                 goto retry_private;
1273
1274                         put_futex_key(&key2);
1275                         put_futex_key(&key1);
1276                         goto retry;
1277                 }
1278                 if (curval != *cmpval) {
1279                         ret = -EAGAIN;
1280                         goto out_unlock;
1281                 }
1282         }
1283
1284         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1285                 /*
1286                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1287                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1288                  * bit.  We force this here where we are able to easily handle
1289                  * faults rather in the requeue loop below.
1290                  */
1291                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1292                                                  &key2, &pi_state, nr_requeue);
1293
1294                 /*
1295                  * At this point the top_waiter has either taken uaddr2 or is
1296                  * waiting on it.  If the former, then the pi_state will not
1297                  * exist yet, look it up one more time to ensure we have a
1298                  * reference to it.
1299                  */
1300                 if (ret == 1) {
1301                         WARN_ON(pi_state);
1302                         drop_count++;
1303                         task_count++;
1304                         ret = get_futex_value_locked(&curval2, uaddr2);
1305                         if (!ret)
1306                                 ret = lookup_pi_state(curval2, hb2, &key2,
1307                                                       &pi_state);
1308                 }
1309
1310                 switch (ret) {
1311                 case 0:
1312                         break;
1313                 case -EFAULT:
1314                         double_unlock_hb(hb1, hb2);
1315                         put_futex_key(&key2);
1316                         put_futex_key(&key1);
1317                         ret = fault_in_user_writeable(uaddr2);
1318                         if (!ret)
1319                                 goto retry;
1320                         goto out;
1321                 case -EAGAIN:
1322                         /* The owner was exiting, try again. */
1323                         double_unlock_hb(hb1, hb2);
1324                         put_futex_key(&key2);
1325                         put_futex_key(&key1);
1326                         cond_resched();
1327                         goto retry;
1328                 default:
1329                         goto out_unlock;
1330                 }
1331         }
1332
1333         head1 = &hb1->chain;
1334         plist_for_each_entry_safe(this, next, head1, list) {
1335                 if (task_count - nr_wake >= nr_requeue)
1336                         break;
1337
1338                 if (!match_futex(&this->key, &key1))
1339                         continue;
1340
1341                 /*
1342                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1343                  * be paired with each other and no other futex ops.
1344                  */
1345                 if ((requeue_pi && !this->rt_waiter) ||
1346                     (!requeue_pi && this->rt_waiter)) {
1347                         ret = -EINVAL;
1348                         break;
1349                 }
1350
1351                 /*
1352                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1353                  * lock, we already woke the top_waiter.  If not, it will be
1354                  * woken by futex_unlock_pi().
1355                  */
1356                 if (++task_count <= nr_wake && !requeue_pi) {
1357                         wake_futex(this);
1358                         continue;
1359                 }
1360
1361                 /* Ensure we requeue to the expected futex for requeue_pi. */
1362                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1363                         ret = -EINVAL;
1364                         break;
1365                 }
1366
1367                 /*
1368                  * Requeue nr_requeue waiters and possibly one more in the case
1369                  * of requeue_pi if we couldn't acquire the lock atomically.
1370                  */
1371                 if (requeue_pi) {
1372                         /* Prepare the waiter to take the rt_mutex. */
1373                         atomic_inc(&pi_state->refcount);
1374                         this->pi_state = pi_state;
1375                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1376                                                         this->rt_waiter,
1377                                                         this->task, 1);
1378                         if (ret == 1) {
1379                                 /* We got the lock. */
1380                                 requeue_pi_wake_futex(this, &key2, hb2);
1381                                 drop_count++;
1382                                 continue;
1383                         } else if (ret) {
1384                                 /* -EDEADLK */
1385                                 this->pi_state = NULL;
1386                                 free_pi_state(pi_state);
1387                                 goto out_unlock;
1388                         }
1389                 }
1390                 requeue_futex(this, hb1, hb2, &key2);
1391                 drop_count++;
1392         }
1393
1394 out_unlock:
1395         double_unlock_hb(hb1, hb2);
1396
1397         /*
1398          * drop_futex_key_refs() must be called outside the spinlocks. During
1399          * the requeue we moved futex_q's from the hash bucket at key1 to the
1400          * one at key2 and updated their key pointer.  We no longer need to
1401          * hold the references to key1.
1402          */
1403         while (--drop_count >= 0)
1404                 drop_futex_key_refs(&key1);
1405
1406 out_put_keys:
1407         put_futex_key(&key2);
1408 out_put_key1:
1409         put_futex_key(&key1);
1410 out:
1411         if (pi_state != NULL)
1412                 free_pi_state(pi_state);
1413         return ret ? ret : task_count;
1414 }
1415
1416 /* The key must be already stored in q->key. */
1417 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1418         __acquires(&hb->lock)
1419 {
1420         struct futex_hash_bucket *hb;
1421
1422         hb = hash_futex(&q->key);
1423         q->lock_ptr = &hb->lock;
1424
1425         spin_lock(&hb->lock);
1426         return hb;
1427 }
1428
1429 static inline void
1430 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1431         __releases(&hb->lock)
1432 {
1433         spin_unlock(&hb->lock);
1434 }
1435
1436 /**
1437  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1438  * @q:  The futex_q to enqueue
1439  * @hb: The destination hash bucket
1440  *
1441  * The hb->lock must be held by the caller, and is released here. A call to
1442  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1443  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1444  * or nothing if the unqueue is done as part of the wake process and the unqueue
1445  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1446  * an example).
1447  */
1448 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1449         __releases(&hb->lock)
1450 {
1451         int prio;
1452
1453         /*
1454          * The priority used to register this element is
1455          * - either the real thread-priority for the real-time threads
1456          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1457          * - or MAX_RT_PRIO for non-RT threads.
1458          * Thus, all RT-threads are woken first in priority order, and
1459          * the others are woken last, in FIFO order.
1460          */
1461         prio = min(current->normal_prio, MAX_RT_PRIO);
1462
1463         plist_node_init(&q->list, prio);
1464         plist_add(&q->list, &hb->chain);
1465         q->task = current;
1466         spin_unlock(&hb->lock);
1467 }
1468
1469 /**
1470  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1471  * @q:  The futex_q to unqueue
1472  *
1473  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1474  * be paired with exactly one earlier call to queue_me().
1475  *
1476  * Returns:
1477  *   1 - if the futex_q was still queued (and we removed unqueued it)
1478  *   0 - if the futex_q was already removed by the waking thread
1479  */
1480 static int unqueue_me(struct futex_q *q)
1481 {
1482         spinlock_t *lock_ptr;
1483         int ret = 0;
1484
1485         /* In the common case we don't take the spinlock, which is nice. */
1486 retry:
1487         lock_ptr = q->lock_ptr;
1488         barrier();
1489         if (lock_ptr != NULL) {
1490                 spin_lock(lock_ptr);
1491                 /*
1492                  * q->lock_ptr can change between reading it and
1493                  * spin_lock(), causing us to take the wrong lock.  This
1494                  * corrects the race condition.
1495                  *
1496                  * Reasoning goes like this: if we have the wrong lock,
1497                  * q->lock_ptr must have changed (maybe several times)
1498                  * between reading it and the spin_lock().  It can
1499                  * change again after the spin_lock() but only if it was
1500                  * already changed before the spin_lock().  It cannot,
1501                  * however, change back to the original value.  Therefore
1502                  * we can detect whether we acquired the correct lock.
1503                  */
1504                 if (unlikely(lock_ptr != q->lock_ptr)) {
1505                         spin_unlock(lock_ptr);
1506                         goto retry;
1507                 }
1508                 __unqueue_futex(q);
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         __unqueue_futex(q);
1529
1530         BUG_ON(!q->pi_state);
1531         free_pi_state(q->pi_state);
1532         q->pi_state = NULL;
1533
1534         spin_unlock(q->lock_ptr);
1535 }
1536
1537 /*
1538  * Fixup the pi_state owner with the new owner.
1539  *
1540  * Must be called with hash bucket lock held and mm->sem held for non
1541  * private futexes.
1542  */
1543 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1544                                 struct task_struct *newowner)
1545 {
1546         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1547         struct futex_pi_state *pi_state = q->pi_state;
1548         struct task_struct *oldowner = pi_state->owner;
1549         u32 uval, curval, newval;
1550         int ret;
1551
1552         /* Owner died? */
1553         if (!pi_state->owner)
1554                 newtid |= FUTEX_OWNER_DIED;
1555
1556         /*
1557          * We are here either because we stole the rtmutex from the
1558          * previous highest priority waiter or we are the highest priority
1559          * waiter but failed to get the rtmutex the first time.
1560          * We have to replace the newowner TID in the user space variable.
1561          * This must be atomic as we have to preserve the owner died bit here.
1562          *
1563          * Note: We write the user space value _before_ changing the pi_state
1564          * because we can fault here. Imagine swapped out pages or a fork
1565          * that marked all the anonymous memory readonly for cow.
1566          *
1567          * Modifying pi_state _before_ the user space value would
1568          * leave the pi_state in an inconsistent state when we fault
1569          * here, because we need to drop the hash bucket lock to
1570          * handle the fault. This might be observed in the PID check
1571          * in lookup_pi_state.
1572          */
1573 retry:
1574         if (get_futex_value_locked(&uval, uaddr))
1575                 goto handle_fault;
1576
1577         while (1) {
1578                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1579
1580                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1581                         goto handle_fault;
1582                 if (curval == uval)
1583                         break;
1584                 uval = curval;
1585         }
1586
1587         /*
1588          * We fixed up user space. Now we need to fix the pi_state
1589          * itself.
1590          */
1591         if (pi_state->owner != NULL) {
1592                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1593                 WARN_ON(list_empty(&pi_state->list));
1594                 list_del_init(&pi_state->list);
1595                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1596         }
1597
1598         pi_state->owner = newowner;
1599
1600         raw_spin_lock_irq(&newowner->pi_lock);
1601         WARN_ON(!list_empty(&pi_state->list));
1602         list_add(&pi_state->list, &newowner->pi_state_list);
1603         raw_spin_unlock_irq(&newowner->pi_lock);
1604         return 0;
1605
1606         /*
1607          * To handle the page fault we need to drop the hash bucket
1608          * lock here. That gives the other task (either the highest priority
1609          * waiter itself or the task which stole the rtmutex) the
1610          * chance to try the fixup of the pi_state. So once we are
1611          * back from handling the fault we need to check the pi_state
1612          * after reacquiring the hash bucket lock and before trying to
1613          * do another fixup. When the fixup has been done already we
1614          * simply return.
1615          */
1616 handle_fault:
1617         spin_unlock(q->lock_ptr);
1618
1619         ret = fault_in_user_writeable(uaddr);
1620
1621         spin_lock(q->lock_ptr);
1622
1623         /*
1624          * Check if someone else fixed it for us:
1625          */
1626         if (pi_state->owner != oldowner)
1627                 return 0;
1628
1629         if (ret)
1630                 return ret;
1631
1632         goto retry;
1633 }
1634
1635 static long futex_wait_restart(struct restart_block *restart);
1636
1637 /**
1638  * fixup_owner() - Post lock pi_state and corner case management
1639  * @uaddr:      user address of the futex
1640  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1641  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1642  *
1643  * After attempting to lock an rt_mutex, this function is called to cleanup
1644  * the pi_state owner as well as handle race conditions that may allow us to
1645  * acquire the lock. Must be called with the hb lock held.
1646  *
1647  * Returns:
1648  *  1 - success, lock taken
1649  *  0 - success, lock not taken
1650  * <0 - on error (-EFAULT)
1651  */
1652 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1653 {
1654         struct task_struct *owner;
1655         int ret = 0;
1656
1657         if (locked) {
1658                 /*
1659                  * Got the lock. We might not be the anticipated owner if we
1660                  * did a lock-steal - fix up the PI-state in that case:
1661                  */
1662                 if (q->pi_state->owner != current)
1663                         ret = fixup_pi_state_owner(uaddr, q, current);
1664                 goto out;
1665         }
1666
1667         /*
1668          * Catch the rare case, where the lock was released when we were on the
1669          * way back before we locked the hash bucket.
1670          */
1671         if (q->pi_state->owner == current) {
1672                 /*
1673                  * Try to get the rt_mutex now. This might fail as some other
1674                  * task acquired the rt_mutex after we removed ourself from the
1675                  * rt_mutex waiters list.
1676                  */
1677                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1678                         locked = 1;
1679                         goto out;
1680                 }
1681
1682                 /*
1683                  * pi_state is incorrect, some other task did a lock steal and
1684                  * we returned due to timeout or signal without taking the
1685                  * rt_mutex. Too late.
1686                  */
1687                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1688                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1689                 if (!owner)
1690                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1691                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1692                 ret = fixup_pi_state_owner(uaddr, q, owner);
1693                 goto out;
1694         }
1695
1696         /*
1697          * Paranoia check. If we did not take the lock, then we should not be
1698          * the owner of the rt_mutex.
1699          */
1700         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1701                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1702                                 "pi-state %p\n", ret,
1703                                 q->pi_state->pi_mutex.owner,
1704                                 q->pi_state->owner);
1705
1706 out:
1707         return ret ? ret : locked;
1708 }
1709
1710 /**
1711  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1712  * @hb:         the futex hash bucket, must be locked by the caller
1713  * @q:          the futex_q to queue up on
1714  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1715  */
1716 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1717                                 struct hrtimer_sleeper *timeout)
1718 {
1719         /*
1720          * The task state is guaranteed to be set before another task can
1721          * wake it. set_current_state() is implemented using set_mb() and
1722          * queue_me() calls spin_unlock() upon completion, both serializing
1723          * access to the hash list and forcing another memory barrier.
1724          */
1725         set_current_state(TASK_INTERRUPTIBLE);
1726         queue_me(q, hb);
1727
1728         /* Arm the timer */
1729         if (timeout) {
1730                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1731                 if (!hrtimer_active(&timeout->timer))
1732                         timeout->task = NULL;
1733         }
1734
1735         /*
1736          * If we have been removed from the hash list, then another task
1737          * has tried to wake us, and we can skip the call to schedule().
1738          */
1739         if (likely(!plist_node_empty(&q->list))) {
1740                 /*
1741                  * If the timer has already expired, current will already be
1742                  * flagged for rescheduling. Only call schedule if there
1743                  * is no timeout, or if it has yet to expire.
1744                  */
1745                 if (!timeout || timeout->task)
1746                         schedule();
1747         }
1748         __set_current_state(TASK_RUNNING);
1749 }
1750
1751 /**
1752  * futex_wait_setup() - Prepare to wait on a futex
1753  * @uaddr:      the futex userspace address
1754  * @val:        the expected value
1755  * @flags:      futex flags (FLAGS_SHARED, etc.)
1756  * @q:          the associated futex_q
1757  * @hb:         storage for hash_bucket pointer to be returned to caller
1758  *
1759  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1760  * compare it with the expected value.  Handle atomic faults internally.
1761  * Return with the hb lock held and a q.key reference on success, and unlocked
1762  * with no q.key reference on failure.
1763  *
1764  * Returns:
1765  *  0 - uaddr contains val and hb has been locked
1766  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1767  */
1768 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1769                            struct futex_q *q, struct futex_hash_bucket **hb)
1770 {
1771         u32 uval;
1772         int ret;
1773
1774         /*
1775          * Access the page AFTER the hash-bucket is locked.
1776          * Order is important:
1777          *
1778          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1779          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1780          *
1781          * The basic logical guarantee of a futex is that it blocks ONLY
1782          * if cond(var) is known to be true at the time of blocking, for
1783          * any cond.  If we locked the hash-bucket after testing *uaddr, that
1784          * would open a race condition where we could block indefinitely with
1785          * cond(var) false, which would violate the guarantee.
1786          *
1787          * On the other hand, we insert q and release the hash-bucket only
1788          * after testing *uaddr.  This guarantees that futex_wait() will NOT
1789          * absorb a wakeup if *uaddr does not match the desired values
1790          * while the syscall executes.
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         struct plist_head *head;
2050         union futex_key key = FUTEX_KEY_INIT;
2051         u32 uval, vpid = task_pid_vnr(current);
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) != vpid)
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             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2077                 goto pi_faulted;
2078         /*
2079          * Rare case: we managed to release the lock atomically,
2080          * no need to wake anyone else up:
2081          */
2082         if (unlikely(uval == vpid))
2083                 goto out_unlock;
2084
2085         /*
2086          * Ok, other tasks may need to be woken up - check waiters
2087          * and do the wakeup if necessary:
2088          */
2089         head = &hb->chain;
2090
2091         plist_for_each_entry_safe(this, next, head, list) {
2092                 if (!match_futex (&this->key, &key))
2093                         continue;
2094                 ret = wake_futex_pi(uaddr, uval, this);
2095                 /*
2096                  * The atomic access to the futex value
2097                  * generated a pagefault, so retry the
2098                  * user-access and the wakeup:
2099                  */
2100                 if (ret == -EFAULT)
2101                         goto pi_faulted;
2102                 goto out_unlock;
2103         }
2104         /*
2105          * No waiters - kernel unlocks the futex:
2106          */
2107         if (!(uval & FUTEX_OWNER_DIED)) {
2108                 ret = unlock_futex_pi(uaddr, uval);
2109                 if (ret == -EFAULT)
2110                         goto pi_faulted;
2111         }
2112
2113 out_unlock:
2114         spin_unlock(&hb->lock);
2115         put_futex_key(&key);
2116
2117 out:
2118         return ret;
2119
2120 pi_faulted:
2121         spin_unlock(&hb->lock);
2122         put_futex_key(&key);
2123
2124         ret = fault_in_user_writeable(uaddr);
2125         if (!ret)
2126                 goto retry;
2127
2128         return ret;
2129 }
2130
2131 /**
2132  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2133  * @hb:         the hash_bucket futex_q was original enqueued on
2134  * @q:          the futex_q woken while waiting to be requeued
2135  * @key2:       the futex_key of the requeue target futex
2136  * @timeout:    the timeout associated with the wait (NULL if none)
2137  *
2138  * Detect if the task was woken on the initial futex as opposed to the requeue
2139  * target futex.  If so, determine if it was a timeout or a signal that caused
2140  * the wakeup and return the appropriate error code to the caller.  Must be
2141  * called with the hb lock held.
2142  *
2143  * Returns
2144  *  0 - no early wakeup detected
2145  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2146  */
2147 static inline
2148 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2149                                    struct futex_q *q, union futex_key *key2,
2150                                    struct hrtimer_sleeper *timeout)
2151 {
2152         int ret = 0;
2153
2154         /*
2155          * With the hb lock held, we avoid races while we process the wakeup.
2156          * We only need to hold hb (and not hb2) to ensure atomicity as the
2157          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2158          * It can't be requeued from uaddr2 to something else since we don't
2159          * support a PI aware source futex for requeue.
2160          */
2161         if (!match_futex(&q->key, key2)) {
2162                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2163                 /*
2164                  * We were woken prior to requeue by a timeout or a signal.
2165                  * Unqueue the futex_q and determine which it was.
2166                  */
2167                 plist_del(&q->list, &hb->chain);
2168
2169                 /* Handle spurious wakeups gracefully */
2170                 ret = -EWOULDBLOCK;
2171                 if (timeout && !timeout->task)
2172                         ret = -ETIMEDOUT;
2173                 else if (signal_pending(current))
2174                         ret = -ERESTARTNOINTR;
2175         }
2176         return ret;
2177 }
2178
2179 /**
2180  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2181  * @uaddr:      the futex we initially wait on (non-pi)
2182  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2183  *              the same type, no requeueing from private to shared, etc.
2184  * @val:        the expected value of uaddr
2185  * @abs_time:   absolute timeout
2186  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2187  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2188  * @uaddr2:     the pi futex we will take prior to returning to user-space
2189  *
2190  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2191  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2192  * complete the acquisition of the rt_mutex prior to returning to userspace.
2193  * This ensures the rt_mutex maintains an owner when it has waiters; without
2194  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2195  * need to.
2196  *
2197  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2198  * via the following:
2199  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2200  * 2) wakeup on uaddr2 after a requeue
2201  * 3) signal
2202  * 4) timeout
2203  *
2204  * If 3, cleanup and return -ERESTARTNOINTR.
2205  *
2206  * If 2, we may then block on trying to take the rt_mutex and return via:
2207  * 5) successful lock
2208  * 6) signal
2209  * 7) timeout
2210  * 8) other lock acquisition failure
2211  *
2212  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2213  *
2214  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2215  *
2216  * Returns:
2217  *  0 - On success
2218  * <0 - On error
2219  */
2220 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2221                                  u32 val, ktime_t *abs_time, u32 bitset,
2222                                  u32 __user *uaddr2)
2223 {
2224         struct hrtimer_sleeper timeout, *to = NULL;
2225         struct rt_mutex_waiter rt_waiter;
2226         struct rt_mutex *pi_mutex = NULL;
2227         struct futex_hash_bucket *hb;
2228         union futex_key key2 = FUTEX_KEY_INIT;
2229         struct futex_q q = futex_q_init;
2230         int res, ret;
2231
2232         if (!bitset)
2233                 return -EINVAL;
2234
2235         if (abs_time) {
2236                 to = &timeout;
2237                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2238                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2239                                       HRTIMER_MODE_ABS);
2240                 hrtimer_init_sleeper(to, current);
2241                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2242                                              current->timer_slack_ns);
2243         }
2244
2245         /*
2246          * The waiter is allocated on our stack, manipulated by the requeue
2247          * code while we sleep on uaddr.
2248          */
2249         debug_rt_mutex_init_waiter(&rt_waiter);
2250         rt_waiter.task = NULL;
2251
2252         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
2253         if (unlikely(ret != 0))
2254                 goto out;
2255
2256         q.bitset = bitset;
2257         q.rt_waiter = &rt_waiter;
2258         q.requeue_pi_key = &key2;
2259
2260         /*
2261          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2262          * count.
2263          */
2264         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2265         if (ret)
2266                 goto out_key2;
2267
2268         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2269         futex_wait_queue_me(hb, &q, to);
2270
2271         spin_lock(&hb->lock);
2272         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2273         spin_unlock(&hb->lock);
2274         if (ret)
2275                 goto out_put_keys;
2276
2277         /*
2278          * In order for us to be here, we know our q.key == key2, and since
2279          * we took the hb->lock above, we also know that futex_requeue() has
2280          * completed and we no longer have to concern ourselves with a wakeup
2281          * race with the atomic proxy lock acquisition by the requeue code. The
2282          * futex_requeue dropped our key1 reference and incremented our key2
2283          * reference count.
2284          */
2285
2286         /* Check if the requeue code acquired the second futex for us. */
2287         if (!q.rt_waiter) {
2288                 /*
2289                  * Got the lock. We might not be the anticipated owner if we
2290                  * did a lock-steal - fix up the PI-state in that case.
2291                  */
2292                 if (q.pi_state && (q.pi_state->owner != current)) {
2293                         spin_lock(q.lock_ptr);
2294                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2295                         spin_unlock(q.lock_ptr);
2296                 }
2297         } else {
2298                 /*
2299                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2300                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2301                  * the pi_state.
2302                  */
2303                 WARN_ON(!&q.pi_state);
2304                 pi_mutex = &q.pi_state->pi_mutex;
2305                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2306                 debug_rt_mutex_free_waiter(&rt_waiter);
2307
2308                 spin_lock(q.lock_ptr);
2309                 /*
2310                  * Fixup the pi_state owner and possibly acquire the lock if we
2311                  * haven't already.
2312                  */
2313                 res = fixup_owner(uaddr2, &q, !ret);
2314                 /*
2315                  * If fixup_owner() returned an error, proprogate that.  If it
2316                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2317                  */
2318                 if (res)
2319                         ret = (res < 0) ? res : 0;
2320
2321                 /* Unqueue and drop the lock. */
2322                 unqueue_me_pi(&q);
2323         }
2324
2325         /*
2326          * If fixup_pi_state_owner() faulted and was unable to handle the
2327          * fault, unlock the rt_mutex and return the fault to userspace.
2328          */
2329         if (ret == -EFAULT) {
2330                 if (rt_mutex_owner(pi_mutex) == current)
2331                         rt_mutex_unlock(pi_mutex);
2332         } else if (ret == -EINTR) {
2333                 /*
2334                  * We've already been requeued, but cannot restart by calling
2335                  * futex_lock_pi() directly. We could restart this syscall, but
2336                  * it would detect that the user space "val" changed and return
2337                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2338                  * -EWOULDBLOCK directly.
2339                  */
2340                 ret = -EWOULDBLOCK;
2341         }
2342
2343 out_put_keys:
2344         put_futex_key(&q.key);
2345 out_key2:
2346         put_futex_key(&key2);
2347
2348 out:
2349         if (to) {
2350                 hrtimer_cancel(&to->timer);
2351                 destroy_hrtimer_on_stack(&to->timer);
2352         }
2353         return ret;
2354 }
2355
2356 /*
2357  * Support for robust futexes: the kernel cleans up held futexes at
2358  * thread exit time.
2359  *
2360  * Implementation: user-space maintains a per-thread list of locks it
2361  * is holding. Upon do_exit(), the kernel carefully walks this list,
2362  * and marks all locks that are owned by this thread with the
2363  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2364  * always manipulated with the lock held, so the list is private and
2365  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2366  * field, to allow the kernel to clean up if the thread dies after
2367  * acquiring the lock, but just before it could have added itself to
2368  * the list. There can only be one such pending lock.
2369  */
2370
2371 /**
2372  * sys_set_robust_list() - Set the robust-futex list head of a task
2373  * @head:       pointer to the list-head
2374  * @len:        length of the list-head, as userspace expects
2375  */
2376 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2377                 size_t, len)
2378 {
2379         if (!futex_cmpxchg_enabled)
2380                 return -ENOSYS;
2381         /*
2382          * The kernel knows only one size for now:
2383          */
2384         if (unlikely(len != sizeof(*head)))
2385                 return -EINVAL;
2386
2387         current->robust_list = head;
2388
2389         return 0;
2390 }
2391
2392 /**
2393  * sys_get_robust_list() - Get the robust-futex list head of a task
2394  * @pid:        pid of the process [zero for current task]
2395  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2396  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2397  */
2398 SYSCALL_DEFINE3(get_robust_list, int, pid,
2399                 struct robust_list_head __user * __user *, head_ptr,
2400                 size_t __user *, len_ptr)
2401 {
2402         struct robust_list_head __user *head;
2403         unsigned long ret;
2404         const struct cred *cred = current_cred(), *pcred;
2405
2406         if (!futex_cmpxchg_enabled)
2407                 return -ENOSYS;
2408
2409         if (!pid)
2410                 head = current->robust_list;
2411         else {
2412                 struct task_struct *p;
2413
2414                 ret = -ESRCH;
2415                 rcu_read_lock();
2416                 p = find_task_by_vpid(pid);
2417                 if (!p)
2418                         goto err_unlock;
2419                 ret = -EPERM;
2420                 pcred = __task_cred(p);
2421                 /* If victim is in different user_ns, then uids are not
2422                    comparable, so we must have CAP_SYS_PTRACE */
2423                 if (cred->user->user_ns != pcred->user->user_ns) {
2424                         if (!ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2425                                 goto err_unlock;
2426                         goto ok;
2427                 }
2428                 /* If victim is in same user_ns, then uids are comparable */
2429                 if (cred->euid != pcred->euid &&
2430                     cred->euid != pcred->uid &&
2431                     !ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2432                         goto err_unlock;
2433 ok:
2434                 head = p->robust_list;
2435                 rcu_read_unlock();
2436         }
2437
2438         if (put_user(sizeof(*head), len_ptr))
2439                 return -EFAULT;
2440         return put_user(head, head_ptr);
2441
2442 err_unlock:
2443         rcu_read_unlock();
2444
2445         return ret;
2446 }
2447
2448 /*
2449  * Process a futex-list entry, check whether it's owned by the
2450  * dying task, and do notification if so:
2451  */
2452 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2453 {
2454         u32 uval, nval, mval;
2455
2456 retry:
2457         if (get_user(uval, uaddr))
2458                 return -1;
2459
2460         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2461                 /*
2462                  * Ok, this dying thread is truly holding a futex
2463                  * of interest. Set the OWNER_DIED bit atomically
2464                  * via cmpxchg, and if the value had FUTEX_WAITERS
2465                  * set, wake up a waiter (if any). (We have to do a
2466                  * futex_wake() even if OWNER_DIED is already set -
2467                  * to handle the rare but possible case of recursive
2468                  * thread-death.) The rest of the cleanup is done in
2469                  * userspace.
2470                  */
2471                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2472                 /*
2473                  * We are not holding a lock here, but we want to have
2474                  * the pagefault_disable/enable() protection because
2475                  * we want to handle the fault gracefully. If the
2476                  * access fails we try to fault in the futex with R/W
2477                  * verification via get_user_pages. get_user() above
2478                  * does not guarantee R/W access. If that fails we
2479                  * give up and leave the futex locked.
2480                  */
2481                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2482                         if (fault_in_user_writeable(uaddr))
2483                                 return -1;
2484                         goto retry;
2485                 }
2486                 if (nval != uval)
2487                         goto retry;
2488
2489                 /*
2490                  * Wake robust non-PI futexes here. The wakeup of
2491                  * PI futexes happens in exit_pi_state():
2492                  */
2493                 if (!pi && (uval & FUTEX_WAITERS))
2494                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2495         }
2496         return 0;
2497 }
2498
2499 /*
2500  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2501  */
2502 static inline int fetch_robust_entry(struct robust_list __user **entry,
2503                                      struct robust_list __user * __user *head,
2504                                      unsigned int *pi)
2505 {
2506         unsigned long uentry;
2507
2508         if (get_user(uentry, (unsigned long __user *)head))
2509                 return -EFAULT;
2510
2511         *entry = (void __user *)(uentry & ~1UL);
2512         *pi = uentry & 1;
2513
2514         return 0;
2515 }
2516
2517 /*
2518  * Walk curr->robust_list (very carefully, it's a userspace list!)
2519  * and mark any locks found there dead, and notify any waiters.
2520  *
2521  * We silently return on any sign of list-walking problem.
2522  */
2523 void exit_robust_list(struct task_struct *curr)
2524 {
2525         struct robust_list_head __user *head = curr->robust_list;
2526         struct robust_list __user *entry, *next_entry, *pending;
2527         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2528         unsigned int uninitialized_var(next_pi);
2529         unsigned long futex_offset;
2530         int rc;
2531
2532         if (!futex_cmpxchg_enabled)
2533                 return;
2534
2535         /*
2536          * Fetch the list head (which was registered earlier, via
2537          * sys_set_robust_list()):
2538          */
2539         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2540                 return;
2541         /*
2542          * Fetch the relative futex offset:
2543          */
2544         if (get_user(futex_offset, &head->futex_offset))
2545                 return;
2546         /*
2547          * Fetch any possibly pending lock-add first, and handle it
2548          * if it exists:
2549          */
2550         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2551                 return;
2552
2553         next_entry = NULL;      /* avoid warning with gcc */
2554         while (entry != &head->list) {
2555                 /*
2556                  * Fetch the next entry in the list before calling
2557                  * handle_futex_death:
2558                  */
2559                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2560                 /*
2561                  * A pending lock might already be on the list, so
2562                  * don't process it twice:
2563                  */
2564                 if (entry != pending)
2565                         if (handle_futex_death((void __user *)entry + futex_offset,
2566                                                 curr, pi))
2567                                 return;
2568                 if (rc)
2569                         return;
2570                 entry = next_entry;
2571                 pi = next_pi;
2572                 /*
2573                  * Avoid excessively long or circular lists:
2574                  */
2575                 if (!--limit)
2576                         break;
2577
2578                 cond_resched();
2579         }
2580
2581         if (pending)
2582                 handle_futex_death((void __user *)pending + futex_offset,
2583                                    curr, pip);
2584 }
2585
2586 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2587                 u32 __user *uaddr2, u32 val2, u32 val3)
2588 {
2589         int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2590         unsigned int flags = 0;
2591
2592         if (!(op & FUTEX_PRIVATE_FLAG))
2593                 flags |= FLAGS_SHARED;
2594
2595         if (op & FUTEX_CLOCK_REALTIME) {
2596                 flags |= FLAGS_CLOCKRT;
2597                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2598                         return -ENOSYS;
2599         }
2600
2601         switch (cmd) {
2602         case FUTEX_WAIT:
2603                 val3 = FUTEX_BITSET_MATCH_ANY;
2604         case FUTEX_WAIT_BITSET:
2605                 ret = futex_wait(uaddr, flags, val, timeout, val3);
2606                 break;
2607         case FUTEX_WAKE:
2608                 val3 = FUTEX_BITSET_MATCH_ANY;
2609         case FUTEX_WAKE_BITSET:
2610                 ret = futex_wake(uaddr, flags, val, val3);
2611                 break;
2612         case FUTEX_REQUEUE:
2613                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2614                 break;
2615         case FUTEX_CMP_REQUEUE:
2616                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2617                 break;
2618         case FUTEX_WAKE_OP:
2619                 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2620                 break;
2621         case FUTEX_LOCK_PI:
2622                 if (futex_cmpxchg_enabled)
2623                         ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2624                 break;
2625         case FUTEX_UNLOCK_PI:
2626                 if (futex_cmpxchg_enabled)
2627                         ret = futex_unlock_pi(uaddr, flags);
2628                 break;
2629         case FUTEX_TRYLOCK_PI:
2630                 if (futex_cmpxchg_enabled)
2631                         ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2632                 break;
2633         case FUTEX_WAIT_REQUEUE_PI:
2634                 val3 = FUTEX_BITSET_MATCH_ANY;
2635                 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2636                                             uaddr2);
2637                 break;
2638         case FUTEX_CMP_REQUEUE_PI:
2639                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2640                 break;
2641         default:
2642                 ret = -ENOSYS;
2643         }
2644         return ret;
2645 }
2646
2647
2648 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2649                 struct timespec __user *, utime, u32 __user *, uaddr2,
2650                 u32, val3)
2651 {
2652         struct timespec ts;
2653         ktime_t t, *tp = NULL;
2654         u32 val2 = 0;
2655         int cmd = op & FUTEX_CMD_MASK;
2656
2657         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2658                       cmd == FUTEX_WAIT_BITSET ||
2659                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2660                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2661                         return -EFAULT;
2662                 if (!timespec_valid(&ts))
2663                         return -EINVAL;
2664
2665                 t = timespec_to_ktime(ts);
2666                 if (cmd == FUTEX_WAIT)
2667                         t = ktime_add_safe(ktime_get(), t);
2668                 tp = &t;
2669         }
2670         /*
2671          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2672          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2673          */
2674         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2675             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2676                 val2 = (u32) (unsigned long) utime;
2677
2678         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2679 }
2680
2681 static int __init futex_init(void)
2682 {
2683         u32 curval;
2684         int i;
2685
2686         /*
2687          * This will fail and we want it. Some arch implementations do
2688          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2689          * functionality. We want to know that before we call in any
2690          * of the complex code paths. Also we want to prevent
2691          * registration of robust lists in that case. NULL is
2692          * guaranteed to fault and we get -EFAULT on functional
2693          * implementation, the non-functional ones will return
2694          * -ENOSYS.
2695          */
2696         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2697                 futex_cmpxchg_enabled = 1;
2698
2699         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2700                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2701                 spin_lock_init(&futex_queues[i].lock);
2702         }
2703
2704         return 0;
2705 }
2706 __initcall(futex_init);