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