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