thp: update futex compound knowledge
[linux-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/module.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62
63 #include <asm/futex.h>
64
65 #include "rtmutex_common.h"
66
67 int __read_mostly futex_cmpxchg_enabled;
68
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71 /*
72  * Futex flags used to encode options to functions and preserve them across
73  * restarts.
74  */
75 #define FLAGS_SHARED            0x01
76 #define FLAGS_CLOCKRT           0x02
77 #define FLAGS_HAS_TIMEOUT       0x04
78
79 /*
80  * Priority Inheritance state:
81  */
82 struct futex_pi_state {
83         /*
84          * list of 'owned' pi_state instances - these have to be
85          * cleaned up in do_exit() if the task exits prematurely:
86          */
87         struct list_head list;
88
89         /*
90          * The PI object:
91          */
92         struct rt_mutex pi_mutex;
93
94         struct task_struct *owner;
95         atomic_t refcount;
96
97         union futex_key key;
98 };
99
100 /**
101  * struct futex_q - The hashed futex queue entry, one per waiting task
102  * @list:               priority-sorted list of tasks waiting on this futex
103  * @task:               the task waiting on the futex
104  * @lock_ptr:           the hash bucket lock
105  * @key:                the key the futex is hashed on
106  * @pi_state:           optional priority inheritance state
107  * @rt_waiter:          rt_waiter storage for use with requeue_pi
108  * @requeue_pi_key:     the requeue_pi target futex key
109  * @bitset:             bitset for the optional bitmasked wakeup
110  *
111  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
112  * we can wake only the relevant ones (hashed queues may be shared).
113  *
114  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
115  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
116  * The order of wakeup is always to make the first condition true, then
117  * the second.
118  *
119  * PI futexes are typically woken before they are removed from the hash list via
120  * the rt_mutex code. See unqueue_me_pi().
121  */
122 struct futex_q {
123         struct plist_node list;
124
125         struct task_struct *task;
126         spinlock_t *lock_ptr;
127         union futex_key key;
128         struct futex_pi_state *pi_state;
129         struct rt_mutex_waiter *rt_waiter;
130         union futex_key *requeue_pi_key;
131         u32 bitset;
132 };
133
134 static const struct futex_q futex_q_init = {
135         /* list gets initialized in queue_me()*/
136         .key = FUTEX_KEY_INIT,
137         .bitset = FUTEX_BITSET_MATCH_ANY
138 };
139
140 /*
141  * Hash buckets are shared by all the futex_keys that hash to the same
142  * location.  Each key may have multiple futex_q structures, one for each task
143  * waiting on a futex.
144  */
145 struct futex_hash_bucket {
146         spinlock_t lock;
147         struct plist_head chain;
148 };
149
150 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
151
152 /*
153  * We hash on the keys returned from get_futex_key (see below).
154  */
155 static struct futex_hash_bucket *hash_futex(union futex_key *key)
156 {
157         u32 hash = jhash2((u32*)&key->both.word,
158                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
159                           key->both.offset);
160         return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
161 }
162
163 /*
164  * Return 1 if two futex_keys are equal, 0 otherwise.
165  */
166 static inline int match_futex(union futex_key *key1, union futex_key *key2)
167 {
168         return (key1 && key2
169                 && key1->both.word == key2->both.word
170                 && key1->both.ptr == key2->both.ptr
171                 && key1->both.offset == key2->both.offset);
172 }
173
174 /*
175  * Take a reference to the resource addressed by a key.
176  * Can be called while holding spinlocks.
177  *
178  */
179 static void get_futex_key_refs(union futex_key *key)
180 {
181         if (!key->both.ptr)
182                 return;
183
184         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
185         case FUT_OFF_INODE:
186                 ihold(key->shared.inode);
187                 break;
188         case FUT_OFF_MMSHARED:
189                 atomic_inc(&key->private.mm->mm_count);
190                 break;
191         }
192 }
193
194 /*
195  * Drop a reference to the resource addressed by a key.
196  * The hash bucket spinlock must not be held.
197  */
198 static void drop_futex_key_refs(union futex_key *key)
199 {
200         if (!key->both.ptr) {
201                 /* If we're here then we tried to put a key we failed to get */
202                 WARN_ON_ONCE(1);
203                 return;
204         }
205
206         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
207         case FUT_OFF_INODE:
208                 iput(key->shared.inode);
209                 break;
210         case FUT_OFF_MMSHARED:
211                 mmdrop(key->private.mm);
212                 break;
213         }
214 }
215
216 /**
217  * get_futex_key() - Get parameters which are the keys for a futex
218  * @uaddr:      virtual address of the futex
219  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
220  * @key:        address where result is stored.
221  *
222  * Returns a negative error code or 0
223  * The key words are stored in *key on success.
224  *
225  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
226  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
227  * We can usually work out the index without swapping in the page.
228  *
229  * lock_page() might sleep, the caller should not hold a spinlock.
230  */
231 static int
232 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
233 {
234         unsigned long address = (unsigned long)uaddr;
235         struct mm_struct *mm = current->mm;
236         struct page *page, *page_head;
237         int err;
238
239         /*
240          * The futex address must be "naturally" aligned.
241          */
242         key->both.offset = address % PAGE_SIZE;
243         if (unlikely((address % sizeof(u32)) != 0))
244                 return -EINVAL;
245         address -= key->both.offset;
246
247         /*
248          * PROCESS_PRIVATE futexes are fast.
249          * As the mm cannot disappear under us and the 'key' only needs
250          * virtual address, we dont even have to find the underlying vma.
251          * Note : We do have to check 'uaddr' is a valid user address,
252          *        but access_ok() should be faster than find_vma()
253          */
254         if (!fshared) {
255                 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
256                         return -EFAULT;
257                 key->private.mm = mm;
258                 key->private.address = address;
259                 get_futex_key_refs(key);
260                 return 0;
261         }
262
263 again:
264         err = get_user_pages_fast(address, 1, 1, &page);
265         if (err < 0)
266                 return err;
267
268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
269         page_head = page;
270         if (unlikely(PageTail(page))) {
271                 put_page(page);
272                 /* serialize against __split_huge_page_splitting() */
273                 local_irq_disable();
274                 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
275                         page_head = compound_head(page);
276                         /*
277                          * page_head is valid pointer but we must pin
278                          * it before taking the PG_lock and/or
279                          * PG_compound_lock. The moment we re-enable
280                          * irqs __split_huge_page_splitting() can
281                          * return and the head page can be freed from
282                          * under us. We can't take the PG_lock and/or
283                          * PG_compound_lock on a page that could be
284                          * freed from under us.
285                          */
286                         if (page != page_head) {
287                                 get_page(page_head);
288                                 put_page(page);
289                         }
290                         local_irq_enable();
291                 } else {
292                         local_irq_enable();
293                         goto again;
294                 }
295         }
296 #else
297         page_head = compound_head(page);
298         if (page != page_head) {
299                 get_page(page_head);
300                 put_page(page);
301         }
302 #endif
303
304         lock_page(page_head);
305         if (!page_head->mapping) {
306                 unlock_page(page_head);
307                 put_page(page_head);
308                 goto again;
309         }
310
311         /*
312          * Private mappings are handled in a simple way.
313          *
314          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
315          * it's a read-only handle, it's expected that futexes attach to
316          * the object not the particular process.
317          */
318         if (PageAnon(page_head)) {
319                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
320                 key->private.mm = mm;
321                 key->private.address = address;
322         } else {
323                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
324                 key->shared.inode = page_head->mapping->host;
325                 key->shared.pgoff = page_head->index;
326         }
327
328         get_futex_key_refs(key);
329
330         unlock_page(page_head);
331         put_page(page_head);
332         return 0;
333 }
334
335 static inline void put_futex_key(union futex_key *key)
336 {
337         drop_futex_key_refs(key);
338 }
339
340 /**
341  * fault_in_user_writeable() - Fault in user address and verify RW access
342  * @uaddr:      pointer to faulting user space address
343  *
344  * Slow path to fixup the fault we just took in the atomic write
345  * access to @uaddr.
346  *
347  * We have no generic implementation of a non-destructive write to the
348  * user address. We know that we faulted in the atomic pagefault
349  * disabled section so we can as well avoid the #PF overhead by
350  * calling get_user_pages() right away.
351  */
352 static int fault_in_user_writeable(u32 __user *uaddr)
353 {
354         struct mm_struct *mm = current->mm;
355         int ret;
356
357         down_read(&mm->mmap_sem);
358         ret = get_user_pages(current, mm, (unsigned long)uaddr,
359                              1, 1, 0, NULL, NULL);
360         up_read(&mm->mmap_sem);
361
362         return ret < 0 ? ret : 0;
363 }
364
365 /**
366  * futex_top_waiter() - Return the highest priority waiter on a futex
367  * @hb:         the hash bucket the futex_q's reside in
368  * @key:        the futex key (to distinguish it from other futex futex_q's)
369  *
370  * Must be called with the hb lock held.
371  */
372 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
373                                         union futex_key *key)
374 {
375         struct futex_q *this;
376
377         plist_for_each_entry(this, &hb->chain, list) {
378                 if (match_futex(&this->key, key))
379                         return this;
380         }
381         return NULL;
382 }
383
384 static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
385 {
386         u32 curval;
387
388         pagefault_disable();
389         curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
390         pagefault_enable();
391
392         return curval;
393 }
394
395 static int get_futex_value_locked(u32 *dest, u32 __user *from)
396 {
397         int ret;
398
399         pagefault_disable();
400         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
401         pagefault_enable();
402
403         return ret ? -EFAULT : 0;
404 }
405
406
407 /*
408  * PI code:
409  */
410 static int refill_pi_state_cache(void)
411 {
412         struct futex_pi_state *pi_state;
413
414         if (likely(current->pi_state_cache))
415                 return 0;
416
417         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
418
419         if (!pi_state)
420                 return -ENOMEM;
421
422         INIT_LIST_HEAD(&pi_state->list);
423         /* pi_mutex gets initialized later */
424         pi_state->owner = NULL;
425         atomic_set(&pi_state->refcount, 1);
426         pi_state->key = FUTEX_KEY_INIT;
427
428         current->pi_state_cache = pi_state;
429
430         return 0;
431 }
432
433 static struct futex_pi_state * alloc_pi_state(void)
434 {
435         struct futex_pi_state *pi_state = current->pi_state_cache;
436
437         WARN_ON(!pi_state);
438         current->pi_state_cache = NULL;
439
440         return pi_state;
441 }
442
443 static void free_pi_state(struct futex_pi_state *pi_state)
444 {
445         if (!atomic_dec_and_test(&pi_state->refcount))
446                 return;
447
448         /*
449          * If pi_state->owner is NULL, the owner is most probably dying
450          * and has cleaned up the pi_state already
451          */
452         if (pi_state->owner) {
453                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
454                 list_del_init(&pi_state->list);
455                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
456
457                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
458         }
459
460         if (current->pi_state_cache)
461                 kfree(pi_state);
462         else {
463                 /*
464                  * pi_state->list is already empty.
465                  * clear pi_state->owner.
466                  * refcount is at 0 - put it back to 1.
467                  */
468                 pi_state->owner = NULL;
469                 atomic_set(&pi_state->refcount, 1);
470                 current->pi_state_cache = pi_state;
471         }
472 }
473
474 /*
475  * Look up the task based on what TID userspace gave us.
476  * We dont trust it.
477  */
478 static struct task_struct * futex_find_get_task(pid_t pid)
479 {
480         struct task_struct *p;
481
482         rcu_read_lock();
483         p = find_task_by_vpid(pid);
484         if (p)
485                 get_task_struct(p);
486
487         rcu_read_unlock();
488
489         return p;
490 }
491
492 /*
493  * This task is holding PI mutexes at exit time => bad.
494  * Kernel cleans up PI-state, but userspace is likely hosed.
495  * (Robust-futex cleanup is separate and might save the day for userspace.)
496  */
497 void exit_pi_state_list(struct task_struct *curr)
498 {
499         struct list_head *next, *head = &curr->pi_state_list;
500         struct futex_pi_state *pi_state;
501         struct futex_hash_bucket *hb;
502         union futex_key key = FUTEX_KEY_INIT;
503
504         if (!futex_cmpxchg_enabled)
505                 return;
506         /*
507          * We are a ZOMBIE and nobody can enqueue itself on
508          * pi_state_list anymore, but we have to be careful
509          * versus waiters unqueueing themselves:
510          */
511         raw_spin_lock_irq(&curr->pi_lock);
512         while (!list_empty(head)) {
513
514                 next = head->next;
515                 pi_state = list_entry(next, struct futex_pi_state, list);
516                 key = pi_state->key;
517                 hb = hash_futex(&key);
518                 raw_spin_unlock_irq(&curr->pi_lock);
519
520                 spin_lock(&hb->lock);
521
522                 raw_spin_lock_irq(&curr->pi_lock);
523                 /*
524                  * We dropped the pi-lock, so re-check whether this
525                  * task still owns the PI-state:
526                  */
527                 if (head->next != next) {
528                         spin_unlock(&hb->lock);
529                         continue;
530                 }
531
532                 WARN_ON(pi_state->owner != curr);
533                 WARN_ON(list_empty(&pi_state->list));
534                 list_del_init(&pi_state->list);
535                 pi_state->owner = NULL;
536                 raw_spin_unlock_irq(&curr->pi_lock);
537
538                 rt_mutex_unlock(&pi_state->pi_mutex);
539
540                 spin_unlock(&hb->lock);
541
542                 raw_spin_lock_irq(&curr->pi_lock);
543         }
544         raw_spin_unlock_irq(&curr->pi_lock);
545 }
546
547 static int
548 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
549                 union futex_key *key, struct futex_pi_state **ps)
550 {
551         struct futex_pi_state *pi_state = NULL;
552         struct futex_q *this, *next;
553         struct plist_head *head;
554         struct task_struct *p;
555         pid_t pid = uval & FUTEX_TID_MASK;
556
557         head = &hb->chain;
558
559         plist_for_each_entry_safe(this, next, head, list) {
560                 if (match_futex(&this->key, key)) {
561                         /*
562                          * Another waiter already exists - bump up
563                          * the refcount and return its pi_state:
564                          */
565                         pi_state = this->pi_state;
566                         /*
567                          * Userspace might have messed up non-PI and PI futexes
568                          */
569                         if (unlikely(!pi_state))
570                                 return -EINVAL;
571
572                         WARN_ON(!atomic_read(&pi_state->refcount));
573
574                         /*
575                          * When pi_state->owner is NULL then the owner died
576                          * and another waiter is on the fly. pi_state->owner
577                          * is fixed up by the task which acquires
578                          * pi_state->rt_mutex.
579                          *
580                          * We do not check for pid == 0 which can happen when
581                          * the owner died and robust_list_exit() cleared the
582                          * TID.
583                          */
584                         if (pid && pi_state->owner) {
585                                 /*
586                                  * Bail out if user space manipulated the
587                                  * futex value.
588                                  */
589                                 if (pid != task_pid_vnr(pi_state->owner))
590                                         return -EINVAL;
591                         }
592
593                         atomic_inc(&pi_state->refcount);
594                         *ps = pi_state;
595
596                         return 0;
597                 }
598         }
599
600         /*
601          * We are the first waiter - try to look up the real owner and attach
602          * the new pi_state to it, but bail out when TID = 0
603          */
604         if (!pid)
605                 return -ESRCH;
606         p = futex_find_get_task(pid);
607         if (!p)
608                 return -ESRCH;
609
610         /*
611          * We need to look at the task state flags to figure out,
612          * whether the task is exiting. To protect against the do_exit
613          * change of the task flags, we do this protected by
614          * p->pi_lock:
615          */
616         raw_spin_lock_irq(&p->pi_lock);
617         if (unlikely(p->flags & PF_EXITING)) {
618                 /*
619                  * The task is on the way out. When PF_EXITPIDONE is
620                  * set, we know that the task has finished the
621                  * cleanup:
622                  */
623                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
624
625                 raw_spin_unlock_irq(&p->pi_lock);
626                 put_task_struct(p);
627                 return ret;
628         }
629
630         pi_state = alloc_pi_state();
631
632         /*
633          * Initialize the pi_mutex in locked state and make 'p'
634          * the owner of it:
635          */
636         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
637
638         /* Store the key for possible exit cleanups: */
639         pi_state->key = *key;
640
641         WARN_ON(!list_empty(&pi_state->list));
642         list_add(&pi_state->list, &p->pi_state_list);
643         pi_state->owner = p;
644         raw_spin_unlock_irq(&p->pi_lock);
645
646         put_task_struct(p);
647
648         *ps = pi_state;
649
650         return 0;
651 }
652
653 /**
654  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
655  * @uaddr:              the pi futex user address
656  * @hb:                 the pi futex hash bucket
657  * @key:                the futex key associated with uaddr and hb
658  * @ps:                 the pi_state pointer where we store the result of the
659  *                      lookup
660  * @task:               the task to perform the atomic lock work for.  This will
661  *                      be "current" except in the case of requeue pi.
662  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
663  *
664  * Returns:
665  *  0 - ready to wait
666  *  1 - acquired the lock
667  * <0 - error
668  *
669  * The hb->lock and futex_key refs shall be held by the caller.
670  */
671 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
672                                 union futex_key *key,
673                                 struct futex_pi_state **ps,
674                                 struct task_struct *task, int set_waiters)
675 {
676         int lock_taken, ret, ownerdied = 0;
677         u32 uval, newval, curval;
678
679 retry:
680         ret = lock_taken = 0;
681
682         /*
683          * To avoid races, we attempt to take the lock here again
684          * (by doing a 0 -> TID atomic cmpxchg), while holding all
685          * the locks. It will most likely not succeed.
686          */
687         newval = task_pid_vnr(task);
688         if (set_waiters)
689                 newval |= FUTEX_WAITERS;
690
691         curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
692
693         if (unlikely(curval == -EFAULT))
694                 return -EFAULT;
695
696         /*
697          * Detect deadlocks.
698          */
699         if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
700                 return -EDEADLK;
701
702         /*
703          * Surprise - we got the lock. Just return to userspace:
704          */
705         if (unlikely(!curval))
706                 return 1;
707
708         uval = curval;
709
710         /*
711          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
712          * to wake at the next unlock.
713          */
714         newval = curval | FUTEX_WAITERS;
715
716         /*
717          * There are two cases, where a futex might have no owner (the
718          * owner TID is 0): OWNER_DIED. We take over the futex in this
719          * case. We also do an unconditional take over, when the owner
720          * of the futex died.
721          *
722          * This is safe as we are protected by the hash bucket lock !
723          */
724         if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
725                 /* Keep the OWNER_DIED bit */
726                 newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
727                 ownerdied = 0;
728                 lock_taken = 1;
729         }
730
731         curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
732
733         if (unlikely(curval == -EFAULT))
734                 return -EFAULT;
735         if (unlikely(curval != uval))
736                 goto retry;
737
738         /*
739          * We took the lock due to owner died take over.
740          */
741         if (unlikely(lock_taken))
742                 return 1;
743
744         /*
745          * We dont have the lock. Look up the PI state (or create it if
746          * we are the first waiter):
747          */
748         ret = lookup_pi_state(uval, hb, key, ps);
749
750         if (unlikely(ret)) {
751                 switch (ret) {
752                 case -ESRCH:
753                         /*
754                          * No owner found for this futex. Check if the
755                          * OWNER_DIED bit is set to figure out whether
756                          * this is a robust futex or not.
757                          */
758                         if (get_futex_value_locked(&curval, uaddr))
759                                 return -EFAULT;
760
761                         /*
762                          * We simply start over in case of a robust
763                          * futex. The code above will take the futex
764                          * and return happy.
765                          */
766                         if (curval & FUTEX_OWNER_DIED) {
767                                 ownerdied = 1;
768                                 goto retry;
769                         }
770                 default:
771                         break;
772                 }
773         }
774
775         return ret;
776 }
777
778 /*
779  * The hash bucket lock must be held when this is called.
780  * Afterwards, the futex_q must not be accessed.
781  */
782 static void wake_futex(struct futex_q *q)
783 {
784         struct task_struct *p = q->task;
785
786         /*
787          * We set q->lock_ptr = NULL _before_ we wake up the task. If
788          * a non-futex wake up happens on another CPU then the task
789          * might exit and p would dereference a non-existing task
790          * struct. Prevent this by holding a reference on p across the
791          * wake up.
792          */
793         get_task_struct(p);
794
795         plist_del(&q->list, &q->list.plist);
796         /*
797          * The waiting task can free the futex_q as soon as
798          * q->lock_ptr = NULL is written, without taking any locks. A
799          * memory barrier is required here to prevent the following
800          * store to lock_ptr from getting ahead of the plist_del.
801          */
802         smp_wmb();
803         q->lock_ptr = NULL;
804
805         wake_up_state(p, TASK_NORMAL);
806         put_task_struct(p);
807 }
808
809 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
810 {
811         struct task_struct *new_owner;
812         struct futex_pi_state *pi_state = this->pi_state;
813         u32 curval, newval;
814
815         if (!pi_state)
816                 return -EINVAL;
817
818         /*
819          * If current does not own the pi_state then the futex is
820          * inconsistent and user space fiddled with the futex value.
821          */
822         if (pi_state->owner != current)
823                 return -EINVAL;
824
825         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
826         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
827
828         /*
829          * This happens when we have stolen the lock and the original
830          * pending owner did not enqueue itself back on the rt_mutex.
831          * Thats not a tragedy. We know that way, that a lock waiter
832          * is on the fly. We make the futex_q waiter the pending owner.
833          */
834         if (!new_owner)
835                 new_owner = this->task;
836
837         /*
838          * We pass it to the next owner. (The WAITERS bit is always
839          * kept enabled while there is PI state around. We must also
840          * preserve the owner died bit.)
841          */
842         if (!(uval & FUTEX_OWNER_DIED)) {
843                 int ret = 0;
844
845                 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
846
847                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
848
849                 if (curval == -EFAULT)
850                         ret = -EFAULT;
851                 else if (curval != uval)
852                         ret = -EINVAL;
853                 if (ret) {
854                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
855                         return ret;
856                 }
857         }
858
859         raw_spin_lock_irq(&pi_state->owner->pi_lock);
860         WARN_ON(list_empty(&pi_state->list));
861         list_del_init(&pi_state->list);
862         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
863
864         raw_spin_lock_irq(&new_owner->pi_lock);
865         WARN_ON(!list_empty(&pi_state->list));
866         list_add(&pi_state->list, &new_owner->pi_state_list);
867         pi_state->owner = new_owner;
868         raw_spin_unlock_irq(&new_owner->pi_lock);
869
870         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
871         rt_mutex_unlock(&pi_state->pi_mutex);
872
873         return 0;
874 }
875
876 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
877 {
878         u32 oldval;
879
880         /*
881          * There is no waiter, so we unlock the futex. The owner died
882          * bit has not to be preserved here. We are the owner:
883          */
884         oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
885
886         if (oldval == -EFAULT)
887                 return oldval;
888         if (oldval != uval)
889                 return -EAGAIN;
890
891         return 0;
892 }
893
894 /*
895  * Express the locking dependencies for lockdep:
896  */
897 static inline void
898 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
899 {
900         if (hb1 <= hb2) {
901                 spin_lock(&hb1->lock);
902                 if (hb1 < hb2)
903                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
904         } else { /* hb1 > hb2 */
905                 spin_lock(&hb2->lock);
906                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
907         }
908 }
909
910 static inline void
911 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
912 {
913         spin_unlock(&hb1->lock);
914         if (hb1 != hb2)
915                 spin_unlock(&hb2->lock);
916 }
917
918 /*
919  * Wake up waiters matching bitset queued on this futex (uaddr).
920  */
921 static int
922 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
923 {
924         struct futex_hash_bucket *hb;
925         struct futex_q *this, *next;
926         struct plist_head *head;
927         union futex_key key = FUTEX_KEY_INIT;
928         int ret;
929
930         if (!bitset)
931                 return -EINVAL;
932
933         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
934         if (unlikely(ret != 0))
935                 goto out;
936
937         hb = hash_futex(&key);
938         spin_lock(&hb->lock);
939         head = &hb->chain;
940
941         plist_for_each_entry_safe(this, next, head, list) {
942                 if (match_futex (&this->key, &key)) {
943                         if (this->pi_state || this->rt_waiter) {
944                                 ret = -EINVAL;
945                                 break;
946                         }
947
948                         /* Check if one of the bits is set in both bitsets */
949                         if (!(this->bitset & bitset))
950                                 continue;
951
952                         wake_futex(this);
953                         if (++ret >= nr_wake)
954                                 break;
955                 }
956         }
957
958         spin_unlock(&hb->lock);
959         put_futex_key(&key);
960 out:
961         return ret;
962 }
963
964 /*
965  * Wake up all waiters hashed on the physical page that is mapped
966  * to this virtual address:
967  */
968 static int
969 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
970               int nr_wake, int nr_wake2, int op)
971 {
972         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
973         struct futex_hash_bucket *hb1, *hb2;
974         struct plist_head *head;
975         struct futex_q *this, *next;
976         int ret, op_ret;
977
978 retry:
979         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
980         if (unlikely(ret != 0))
981                 goto out;
982         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
983         if (unlikely(ret != 0))
984                 goto out_put_key1;
985
986         hb1 = hash_futex(&key1);
987         hb2 = hash_futex(&key2);
988
989 retry_private:
990         double_lock_hb(hb1, hb2);
991         op_ret = futex_atomic_op_inuser(op, uaddr2);
992         if (unlikely(op_ret < 0)) {
993
994                 double_unlock_hb(hb1, hb2);
995
996 #ifndef CONFIG_MMU
997                 /*
998                  * we don't get EFAULT from MMU faults if we don't have an MMU,
999                  * but we might get them from range checking
1000                  */
1001                 ret = op_ret;
1002                 goto out_put_keys;
1003 #endif
1004
1005                 if (unlikely(op_ret != -EFAULT)) {
1006                         ret = op_ret;
1007                         goto out_put_keys;
1008                 }
1009
1010                 ret = fault_in_user_writeable(uaddr2);
1011                 if (ret)
1012                         goto out_put_keys;
1013
1014                 if (!(flags & FLAGS_SHARED))
1015                         goto retry_private;
1016
1017                 put_futex_key(&key2);
1018                 put_futex_key(&key1);
1019                 goto retry;
1020         }
1021
1022         head = &hb1->chain;
1023
1024         plist_for_each_entry_safe(this, next, head, list) {
1025                 if (match_futex (&this->key, &key1)) {
1026                         wake_futex(this);
1027                         if (++ret >= nr_wake)
1028                                 break;
1029                 }
1030         }
1031
1032         if (op_ret > 0) {
1033                 head = &hb2->chain;
1034
1035                 op_ret = 0;
1036                 plist_for_each_entry_safe(this, next, head, list) {
1037                         if (match_futex (&this->key, &key2)) {
1038                                 wake_futex(this);
1039                                 if (++op_ret >= nr_wake2)
1040                                         break;
1041                         }
1042                 }
1043                 ret += op_ret;
1044         }
1045
1046         double_unlock_hb(hb1, hb2);
1047 out_put_keys:
1048         put_futex_key(&key2);
1049 out_put_key1:
1050         put_futex_key(&key1);
1051 out:
1052         return ret;
1053 }
1054
1055 /**
1056  * requeue_futex() - Requeue a futex_q from one hb to another
1057  * @q:          the futex_q to requeue
1058  * @hb1:        the source hash_bucket
1059  * @hb2:        the target hash_bucket
1060  * @key2:       the new key for the requeued futex_q
1061  */
1062 static inline
1063 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1064                    struct futex_hash_bucket *hb2, union futex_key *key2)
1065 {
1066
1067         /*
1068          * If key1 and key2 hash to the same bucket, no need to
1069          * requeue.
1070          */
1071         if (likely(&hb1->chain != &hb2->chain)) {
1072                 plist_del(&q->list, &hb1->chain);
1073                 plist_add(&q->list, &hb2->chain);
1074                 q->lock_ptr = &hb2->lock;
1075 #ifdef CONFIG_DEBUG_PI_LIST
1076                 q->list.plist.spinlock = &hb2->lock;
1077 #endif
1078         }
1079         get_futex_key_refs(key2);
1080         q->key = *key2;
1081 }
1082
1083 /**
1084  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1085  * @q:          the futex_q
1086  * @key:        the key of the requeue target futex
1087  * @hb:         the hash_bucket of the requeue target futex
1088  *
1089  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1090  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1091  * to the requeue target futex so the waiter can detect the wakeup on the right
1092  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1093  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1094  * to protect access to the pi_state to fixup the owner later.  Must be called
1095  * with both q->lock_ptr and hb->lock held.
1096  */
1097 static inline
1098 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1099                            struct futex_hash_bucket *hb)
1100 {
1101         get_futex_key_refs(key);
1102         q->key = *key;
1103
1104         WARN_ON(plist_node_empty(&q->list));
1105         plist_del(&q->list, &q->list.plist);
1106
1107         WARN_ON(!q->rt_waiter);
1108         q->rt_waiter = NULL;
1109
1110         q->lock_ptr = &hb->lock;
1111 #ifdef CONFIG_DEBUG_PI_LIST
1112         q->list.plist.spinlock = &hb->lock;
1113 #endif
1114
1115         wake_up_state(q->task, TASK_NORMAL);
1116 }
1117
1118 /**
1119  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1120  * @pifutex:            the user address of the to futex
1121  * @hb1:                the from futex hash bucket, must be locked by the caller
1122  * @hb2:                the to futex hash bucket, must be locked by the caller
1123  * @key1:               the from futex key
1124  * @key2:               the to futex key
1125  * @ps:                 address to store the pi_state pointer
1126  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1127  *
1128  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1129  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1130  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1131  * hb1 and hb2 must be held by the caller.
1132  *
1133  * Returns:
1134  *  0 - failed to acquire the lock atomicly
1135  *  1 - acquired the lock
1136  * <0 - error
1137  */
1138 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1139                                  struct futex_hash_bucket *hb1,
1140                                  struct futex_hash_bucket *hb2,
1141                                  union futex_key *key1, union futex_key *key2,
1142                                  struct futex_pi_state **ps, int set_waiters)
1143 {
1144         struct futex_q *top_waiter = NULL;
1145         u32 curval;
1146         int ret;
1147
1148         if (get_futex_value_locked(&curval, pifutex))
1149                 return -EFAULT;
1150
1151         /*
1152          * Find the top_waiter and determine if there are additional waiters.
1153          * If the caller intends to requeue more than 1 waiter to pifutex,
1154          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1155          * as we have means to handle the possible fault.  If not, don't set
1156          * the bit unecessarily as it will force the subsequent unlock to enter
1157          * the kernel.
1158          */
1159         top_waiter = futex_top_waiter(hb1, key1);
1160
1161         /* There are no waiters, nothing for us to do. */
1162         if (!top_waiter)
1163                 return 0;
1164
1165         /* Ensure we requeue to the expected futex. */
1166         if (!match_futex(top_waiter->requeue_pi_key, key2))
1167                 return -EINVAL;
1168
1169         /*
1170          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1171          * the contended case or if set_waiters is 1.  The pi_state is returned
1172          * in ps in contended cases.
1173          */
1174         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1175                                    set_waiters);
1176         if (ret == 1)
1177                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1178
1179         return ret;
1180 }
1181
1182 /**
1183  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1184  * @uaddr1:     source futex user address
1185  * @flags:      futex flags (FLAGS_SHARED, etc.)
1186  * @uaddr2:     target futex user address
1187  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1188  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1189  * @cmpval:     @uaddr1 expected value (or %NULL)
1190  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1191  *              pi futex (pi to pi requeue is not supported)
1192  *
1193  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1194  * uaddr2 atomically on behalf of the top waiter.
1195  *
1196  * Returns:
1197  * >=0 - on success, the number of tasks requeued or woken
1198  *  <0 - on error
1199  */
1200 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1201                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1202                          u32 *cmpval, int requeue_pi)
1203 {
1204         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1205         int drop_count = 0, task_count = 0, ret;
1206         struct futex_pi_state *pi_state = NULL;
1207         struct futex_hash_bucket *hb1, *hb2;
1208         struct plist_head *head1;
1209         struct futex_q *this, *next;
1210         u32 curval2;
1211
1212         if (requeue_pi) {
1213                 /*
1214                  * requeue_pi requires a pi_state, try to allocate it now
1215                  * without any locks in case it fails.
1216                  */
1217                 if (refill_pi_state_cache())
1218                         return -ENOMEM;
1219                 /*
1220                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1221                  * + nr_requeue, since it acquires the rt_mutex prior to
1222                  * returning to userspace, so as to not leave the rt_mutex with
1223                  * waiters and no owner.  However, second and third wake-ups
1224                  * cannot be predicted as they involve race conditions with the
1225                  * first wake and a fault while looking up the pi_state.  Both
1226                  * pthread_cond_signal() and pthread_cond_broadcast() should
1227                  * use nr_wake=1.
1228                  */
1229                 if (nr_wake != 1)
1230                         return -EINVAL;
1231         }
1232
1233 retry:
1234         if (pi_state != NULL) {
1235                 /*
1236                  * We will have to lookup the pi_state again, so free this one
1237                  * to keep the accounting correct.
1238                  */
1239                 free_pi_state(pi_state);
1240                 pi_state = NULL;
1241         }
1242
1243         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1);
1244         if (unlikely(ret != 0))
1245                 goto out;
1246         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
1247         if (unlikely(ret != 0))
1248                 goto out_put_key1;
1249
1250         hb1 = hash_futex(&key1);
1251         hb2 = hash_futex(&key2);
1252
1253 retry_private:
1254         double_lock_hb(hb1, hb2);
1255
1256         if (likely(cmpval != NULL)) {
1257                 u32 curval;
1258
1259                 ret = get_futex_value_locked(&curval, uaddr1);
1260
1261                 if (unlikely(ret)) {
1262                         double_unlock_hb(hb1, hb2);
1263
1264                         ret = get_user(curval, uaddr1);
1265                         if (ret)
1266                                 goto out_put_keys;
1267
1268                         if (!(flags & FLAGS_SHARED))
1269                                 goto retry_private;
1270
1271                         put_futex_key(&key2);
1272                         put_futex_key(&key1);
1273                         goto retry;
1274                 }
1275                 if (curval != *cmpval) {
1276                         ret = -EAGAIN;
1277                         goto out_unlock;
1278                 }
1279         }
1280
1281         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1282                 /*
1283                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1284                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1285                  * bit.  We force this here where we are able to easily handle
1286                  * faults rather in the requeue loop below.
1287                  */
1288                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1289                                                  &key2, &pi_state, nr_requeue);
1290
1291                 /*
1292                  * At this point the top_waiter has either taken uaddr2 or is
1293                  * waiting on it.  If the former, then the pi_state will not
1294                  * exist yet, look it up one more time to ensure we have a
1295                  * reference to it.
1296                  */
1297                 if (ret == 1) {
1298                         WARN_ON(pi_state);
1299                         drop_count++;
1300                         task_count++;
1301                         ret = get_futex_value_locked(&curval2, uaddr2);
1302                         if (!ret)
1303                                 ret = lookup_pi_state(curval2, hb2, &key2,
1304                                                       &pi_state);
1305                 }
1306
1307                 switch (ret) {
1308                 case 0:
1309                         break;
1310                 case -EFAULT:
1311                         double_unlock_hb(hb1, hb2);
1312                         put_futex_key(&key2);
1313                         put_futex_key(&key1);
1314                         ret = fault_in_user_writeable(uaddr2);
1315                         if (!ret)
1316                                 goto retry;
1317                         goto out;
1318                 case -EAGAIN:
1319                         /* The owner was exiting, try again. */
1320                         double_unlock_hb(hb1, hb2);
1321                         put_futex_key(&key2);
1322                         put_futex_key(&key1);
1323                         cond_resched();
1324                         goto retry;
1325                 default:
1326                         goto out_unlock;
1327                 }
1328         }
1329
1330         head1 = &hb1->chain;
1331         plist_for_each_entry_safe(this, next, head1, list) {
1332                 if (task_count - nr_wake >= nr_requeue)
1333                         break;
1334
1335                 if (!match_futex(&this->key, &key1))
1336                         continue;
1337
1338                 /*
1339                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1340                  * be paired with each other and no other futex ops.
1341                  */
1342                 if ((requeue_pi && !this->rt_waiter) ||
1343                     (!requeue_pi && this->rt_waiter)) {
1344                         ret = -EINVAL;
1345                         break;
1346                 }
1347
1348                 /*
1349                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1350                  * lock, we already woke the top_waiter.  If not, it will be
1351                  * woken by futex_unlock_pi().
1352                  */
1353                 if (++task_count <= nr_wake && !requeue_pi) {
1354                         wake_futex(this);
1355                         continue;
1356                 }
1357
1358                 /* Ensure we requeue to the expected futex for requeue_pi. */
1359                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1360                         ret = -EINVAL;
1361                         break;
1362                 }
1363
1364                 /*
1365                  * Requeue nr_requeue waiters and possibly one more in the case
1366                  * of requeue_pi if we couldn't acquire the lock atomically.
1367                  */
1368                 if (requeue_pi) {
1369                         /* Prepare the waiter to take the rt_mutex. */
1370                         atomic_inc(&pi_state->refcount);
1371                         this->pi_state = pi_state;
1372                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1373                                                         this->rt_waiter,
1374                                                         this->task, 1);
1375                         if (ret == 1) {
1376                                 /* We got the lock. */
1377                                 requeue_pi_wake_futex(this, &key2, hb2);
1378                                 drop_count++;
1379                                 continue;
1380                         } else if (ret) {
1381                                 /* -EDEADLK */
1382                                 this->pi_state = NULL;
1383                                 free_pi_state(pi_state);
1384                                 goto out_unlock;
1385                         }
1386                 }
1387                 requeue_futex(this, hb1, hb2, &key2);
1388                 drop_count++;
1389         }
1390
1391 out_unlock:
1392         double_unlock_hb(hb1, hb2);
1393
1394         /*
1395          * drop_futex_key_refs() must be called outside the spinlocks. During
1396          * the requeue we moved futex_q's from the hash bucket at key1 to the
1397          * one at key2 and updated their key pointer.  We no longer need to
1398          * hold the references to key1.
1399          */
1400         while (--drop_count >= 0)
1401                 drop_futex_key_refs(&key1);
1402
1403 out_put_keys:
1404         put_futex_key(&key2);
1405 out_put_key1:
1406         put_futex_key(&key1);
1407 out:
1408         if (pi_state != NULL)
1409                 free_pi_state(pi_state);
1410         return ret ? ret : task_count;
1411 }
1412
1413 /* The key must be already stored in q->key. */
1414 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1415         __acquires(&hb->lock)
1416 {
1417         struct futex_hash_bucket *hb;
1418
1419         hb = hash_futex(&q->key);
1420         q->lock_ptr = &hb->lock;
1421
1422         spin_lock(&hb->lock);
1423         return hb;
1424 }
1425
1426 static inline void
1427 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1428         __releases(&hb->lock)
1429 {
1430         spin_unlock(&hb->lock);
1431 }
1432
1433 /**
1434  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1435  * @q:  The futex_q to enqueue
1436  * @hb: The destination hash bucket
1437  *
1438  * The hb->lock must be held by the caller, and is released here. A call to
1439  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1440  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1441  * or nothing if the unqueue is done as part of the wake process and the unqueue
1442  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1443  * an example).
1444  */
1445 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1446         __releases(&hb->lock)
1447 {
1448         int prio;
1449
1450         /*
1451          * The priority used to register this element is
1452          * - either the real thread-priority for the real-time threads
1453          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1454          * - or MAX_RT_PRIO for non-RT threads.
1455          * Thus, all RT-threads are woken first in priority order, and
1456          * the others are woken last, in FIFO order.
1457          */
1458         prio = min(current->normal_prio, MAX_RT_PRIO);
1459
1460         plist_node_init(&q->list, prio);
1461 #ifdef CONFIG_DEBUG_PI_LIST
1462         q->list.plist.spinlock = &hb->lock;
1463 #endif
1464         plist_add(&q->list, &hb->chain);
1465         q->task = current;
1466         spin_unlock(&hb->lock);
1467 }
1468
1469 /**
1470  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1471  * @q:  The futex_q to unqueue
1472  *
1473  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1474  * be paired with exactly one earlier call to queue_me().
1475  *
1476  * Returns:
1477  *   1 - if the futex_q was still queued (and we removed unqueued it)
1478  *   0 - if the futex_q was already removed by the waking thread
1479  */
1480 static int unqueue_me(struct futex_q *q)
1481 {
1482         spinlock_t *lock_ptr;
1483         int ret = 0;
1484
1485         /* In the common case we don't take the spinlock, which is nice. */
1486 retry:
1487         lock_ptr = q->lock_ptr;
1488         barrier();
1489         if (lock_ptr != NULL) {
1490                 spin_lock(lock_ptr);
1491                 /*
1492                  * q->lock_ptr can change between reading it and
1493                  * spin_lock(), causing us to take the wrong lock.  This
1494                  * corrects the race condition.
1495                  *
1496                  * Reasoning goes like this: if we have the wrong lock,
1497                  * q->lock_ptr must have changed (maybe several times)
1498                  * between reading it and the spin_lock().  It can
1499                  * change again after the spin_lock() but only if it was
1500                  * already changed before the spin_lock().  It cannot,
1501                  * however, change back to the original value.  Therefore
1502                  * we can detect whether we acquired the correct lock.
1503                  */
1504                 if (unlikely(lock_ptr != q->lock_ptr)) {
1505                         spin_unlock(lock_ptr);
1506                         goto retry;
1507                 }
1508                 WARN_ON(plist_node_empty(&q->list));
1509                 plist_del(&q->list, &q->list.plist);
1510
1511                 BUG_ON(q->pi_state);
1512
1513                 spin_unlock(lock_ptr);
1514                 ret = 1;
1515         }
1516
1517         drop_futex_key_refs(&q->key);
1518         return ret;
1519 }
1520
1521 /*
1522  * PI futexes can not be requeued and must remove themself from the
1523  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1524  * and dropped here.
1525  */
1526 static void unqueue_me_pi(struct futex_q *q)
1527         __releases(q->lock_ptr)
1528 {
1529         WARN_ON(plist_node_empty(&q->list));
1530         plist_del(&q->list, &q->list.plist);
1531
1532         BUG_ON(!q->pi_state);
1533         free_pi_state(q->pi_state);
1534         q->pi_state = NULL;
1535
1536         spin_unlock(q->lock_ptr);
1537 }
1538
1539 /*
1540  * Fixup the pi_state owner with the new owner.
1541  *
1542  * Must be called with hash bucket lock held and mm->sem held for non
1543  * private futexes.
1544  */
1545 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1546                                 struct task_struct *newowner)
1547 {
1548         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1549         struct futex_pi_state *pi_state = q->pi_state;
1550         struct task_struct *oldowner = pi_state->owner;
1551         u32 uval, curval, newval;
1552         int ret;
1553
1554         /* Owner died? */
1555         if (!pi_state->owner)
1556                 newtid |= FUTEX_OWNER_DIED;
1557
1558         /*
1559          * We are here either because we stole the rtmutex from the
1560          * pending owner or we are the pending owner which failed to
1561          * get the rtmutex. We have to replace the pending owner TID
1562          * in the user space variable. This must be atomic as we have
1563          * to preserve the owner died bit here.
1564          *
1565          * Note: We write the user space value _before_ changing the pi_state
1566          * because we can fault here. Imagine swapped out pages or a fork
1567          * that marked all the anonymous memory readonly for cow.
1568          *
1569          * Modifying pi_state _before_ the user space value would
1570          * leave the pi_state in an inconsistent state when we fault
1571          * here, because we need to drop the hash bucket lock to
1572          * handle the fault. This might be observed in the PID check
1573          * in lookup_pi_state.
1574          */
1575 retry:
1576         if (get_futex_value_locked(&uval, uaddr))
1577                 goto handle_fault;
1578
1579         while (1) {
1580                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1581
1582                 curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1583
1584                 if (curval == -EFAULT)
1585                         goto handle_fault;
1586                 if (curval == uval)
1587                         break;
1588                 uval = curval;
1589         }
1590
1591         /*
1592          * We fixed up user space. Now we need to fix the pi_state
1593          * itself.
1594          */
1595         if (pi_state->owner != NULL) {
1596                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1597                 WARN_ON(list_empty(&pi_state->list));
1598                 list_del_init(&pi_state->list);
1599                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1600         }
1601
1602         pi_state->owner = newowner;
1603
1604         raw_spin_lock_irq(&newowner->pi_lock);
1605         WARN_ON(!list_empty(&pi_state->list));
1606         list_add(&pi_state->list, &newowner->pi_state_list);
1607         raw_spin_unlock_irq(&newowner->pi_lock);
1608         return 0;
1609
1610         /*
1611          * To handle the page fault we need to drop the hash bucket
1612          * lock here. That gives the other task (either the pending
1613          * owner itself or the task which stole the rtmutex) the
1614          * chance to try the fixup of the pi_state. So once we are
1615          * back from handling the fault we need to check the pi_state
1616          * after reacquiring the hash bucket lock and before trying to
1617          * do another fixup. When the fixup has been done already we
1618          * simply return.
1619          */
1620 handle_fault:
1621         spin_unlock(q->lock_ptr);
1622
1623         ret = fault_in_user_writeable(uaddr);
1624
1625         spin_lock(q->lock_ptr);
1626
1627         /*
1628          * Check if someone else fixed it for us:
1629          */
1630         if (pi_state->owner != oldowner)
1631                 return 0;
1632
1633         if (ret)
1634                 return ret;
1635
1636         goto retry;
1637 }
1638
1639 static long futex_wait_restart(struct restart_block *restart);
1640
1641 /**
1642  * fixup_owner() - Post lock pi_state and corner case management
1643  * @uaddr:      user address of the futex
1644  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1645  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1646  *
1647  * After attempting to lock an rt_mutex, this function is called to cleanup
1648  * the pi_state owner as well as handle race conditions that may allow us to
1649  * acquire the lock. Must be called with the hb lock held.
1650  *
1651  * Returns:
1652  *  1 - success, lock taken
1653  *  0 - success, lock not taken
1654  * <0 - on error (-EFAULT)
1655  */
1656 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1657 {
1658         struct task_struct *owner;
1659         int ret = 0;
1660
1661         if (locked) {
1662                 /*
1663                  * Got the lock. We might not be the anticipated owner if we
1664                  * did a lock-steal - fix up the PI-state in that case:
1665                  */
1666                 if (q->pi_state->owner != current)
1667                         ret = fixup_pi_state_owner(uaddr, q, current);
1668                 goto out;
1669         }
1670
1671         /*
1672          * Catch the rare case, where the lock was released when we were on the
1673          * way back before we locked the hash bucket.
1674          */
1675         if (q->pi_state->owner == current) {
1676                 /*
1677                  * Try to get the rt_mutex now. This might fail as some other
1678                  * task acquired the rt_mutex after we removed ourself from the
1679                  * rt_mutex waiters list.
1680                  */
1681                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1682                         locked = 1;
1683                         goto out;
1684                 }
1685
1686                 /*
1687                  * pi_state is incorrect, some other task did a lock steal and
1688                  * we returned due to timeout or signal without taking the
1689                  * rt_mutex. Too late. We can access the rt_mutex_owner without
1690                  * locking, as the other task is now blocked on the hash bucket
1691                  * lock. Fix the state up.
1692                  */
1693                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1694                 ret = fixup_pi_state_owner(uaddr, q, owner);
1695                 goto out;
1696         }
1697
1698         /*
1699          * Paranoia check. If we did not take the lock, then we should not be
1700          * the owner, nor the pending owner, of the rt_mutex.
1701          */
1702         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1703                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1704                                 "pi-state %p\n", ret,
1705                                 q->pi_state->pi_mutex.owner,
1706                                 q->pi_state->owner);
1707
1708 out:
1709         return ret ? ret : locked;
1710 }
1711
1712 /**
1713  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1714  * @hb:         the futex hash bucket, must be locked by the caller
1715  * @q:          the futex_q to queue up on
1716  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
1717  */
1718 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1719                                 struct hrtimer_sleeper *timeout)
1720 {
1721         /*
1722          * The task state is guaranteed to be set before another task can
1723          * wake it. set_current_state() is implemented using set_mb() and
1724          * queue_me() calls spin_unlock() upon completion, both serializing
1725          * access to the hash list and forcing another memory barrier.
1726          */
1727         set_current_state(TASK_INTERRUPTIBLE);
1728         queue_me(q, hb);
1729
1730         /* Arm the timer */
1731         if (timeout) {
1732                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1733                 if (!hrtimer_active(&timeout->timer))
1734                         timeout->task = NULL;
1735         }
1736
1737         /*
1738          * If we have been removed from the hash list, then another task
1739          * has tried to wake us, and we can skip the call to schedule().
1740          */
1741         if (likely(!plist_node_empty(&q->list))) {
1742                 /*
1743                  * If the timer has already expired, current will already be
1744                  * flagged for rescheduling. Only call schedule if there
1745                  * is no timeout, or if it has yet to expire.
1746                  */
1747                 if (!timeout || timeout->task)
1748                         schedule();
1749         }
1750         __set_current_state(TASK_RUNNING);
1751 }
1752
1753 /**
1754  * futex_wait_setup() - Prepare to wait on a futex
1755  * @uaddr:      the futex userspace address
1756  * @val:        the expected value
1757  * @flags:      futex flags (FLAGS_SHARED, etc.)
1758  * @q:          the associated futex_q
1759  * @hb:         storage for hash_bucket pointer to be returned to caller
1760  *
1761  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
1762  * compare it with the expected value.  Handle atomic faults internally.
1763  * Return with the hb lock held and a q.key reference on success, and unlocked
1764  * with no q.key reference on failure.
1765  *
1766  * Returns:
1767  *  0 - uaddr contains val and hb has been locked
1768  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1769  */
1770 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1771                            struct futex_q *q, struct futex_hash_bucket **hb)
1772 {
1773         u32 uval;
1774         int ret;
1775
1776         /*
1777          * Access the page AFTER the hash-bucket is locked.
1778          * Order is important:
1779          *
1780          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1781          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
1782          *
1783          * The basic logical guarantee of a futex is that it blocks ONLY
1784          * if cond(var) is known to be true at the time of blocking, for
1785          * any cond.  If we queued after testing *uaddr, that would open
1786          * a race condition where we could block indefinitely with
1787          * cond(var) false, which would violate the guarantee.
1788          *
1789          * A consequence is that futex_wait() can return zero and absorb
1790          * a wakeup when *uaddr != val on entry to the syscall.  This is
1791          * rare, but normal.
1792          */
1793 retry:
1794         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key);
1795         if (unlikely(ret != 0))
1796                 return ret;
1797
1798 retry_private:
1799         *hb = queue_lock(q);
1800
1801         ret = get_futex_value_locked(&uval, uaddr);
1802
1803         if (ret) {
1804                 queue_unlock(q, *hb);
1805
1806                 ret = get_user(uval, uaddr);
1807                 if (ret)
1808                         goto out;
1809
1810                 if (!(flags & FLAGS_SHARED))
1811                         goto retry_private;
1812
1813                 put_futex_key(&q->key);
1814                 goto retry;
1815         }
1816
1817         if (uval != val) {
1818                 queue_unlock(q, *hb);
1819                 ret = -EWOULDBLOCK;
1820         }
1821
1822 out:
1823         if (ret)
1824                 put_futex_key(&q->key);
1825         return ret;
1826 }
1827
1828 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1829                       ktime_t *abs_time, u32 bitset)
1830 {
1831         struct hrtimer_sleeper timeout, *to = NULL;
1832         struct restart_block *restart;
1833         struct futex_hash_bucket *hb;
1834         struct futex_q q = futex_q_init;
1835         int ret;
1836
1837         if (!bitset)
1838                 return -EINVAL;
1839         q.bitset = bitset;
1840
1841         if (abs_time) {
1842                 to = &timeout;
1843
1844                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1845                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
1846                                       HRTIMER_MODE_ABS);
1847                 hrtimer_init_sleeper(to, current);
1848                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1849                                              current->timer_slack_ns);
1850         }
1851
1852 retry:
1853         /*
1854          * Prepare to wait on uaddr. On success, holds hb lock and increments
1855          * q.key refs.
1856          */
1857         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1858         if (ret)
1859                 goto out;
1860
1861         /* queue_me and wait for wakeup, timeout, or a signal. */
1862         futex_wait_queue_me(hb, &q, to);
1863
1864         /* If we were woken (and unqueued), we succeeded, whatever. */
1865         ret = 0;
1866         /* unqueue_me() drops q.key ref */
1867         if (!unqueue_me(&q))
1868                 goto out;
1869         ret = -ETIMEDOUT;
1870         if (to && !to->task)
1871                 goto out;
1872
1873         /*
1874          * We expect signal_pending(current), but we might be the
1875          * victim of a spurious wakeup as well.
1876          */
1877         if (!signal_pending(current))
1878                 goto retry;
1879
1880         ret = -ERESTARTSYS;
1881         if (!abs_time)
1882                 goto out;
1883
1884         restart = &current_thread_info()->restart_block;
1885         restart->fn = futex_wait_restart;
1886         restart->futex.uaddr = uaddr;
1887         restart->futex.val = val;
1888         restart->futex.time = abs_time->tv64;
1889         restart->futex.bitset = bitset;
1890         restart->futex.flags = flags;
1891
1892         ret = -ERESTART_RESTARTBLOCK;
1893
1894 out:
1895         if (to) {
1896                 hrtimer_cancel(&to->timer);
1897                 destroy_hrtimer_on_stack(&to->timer);
1898         }
1899         return ret;
1900 }
1901
1902
1903 static long futex_wait_restart(struct restart_block *restart)
1904 {
1905         u32 __user *uaddr = restart->futex.uaddr;
1906         ktime_t t, *tp = NULL;
1907
1908         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1909                 t.tv64 = restart->futex.time;
1910                 tp = &t;
1911         }
1912         restart->fn = do_no_restart_syscall;
1913
1914         return (long)futex_wait(uaddr, restart->futex.flags,
1915                                 restart->futex.val, tp, restart->futex.bitset);
1916 }
1917
1918
1919 /*
1920  * Userspace tried a 0 -> TID atomic transition of the futex value
1921  * and failed. The kernel side here does the whole locking operation:
1922  * if there are waiters then it will block, it does PI, etc. (Due to
1923  * races the kernel might see a 0 value of the futex too.)
1924  */
1925 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1926                          ktime_t *time, int trylock)
1927 {
1928         struct hrtimer_sleeper timeout, *to = NULL;
1929         struct futex_hash_bucket *hb;
1930         struct futex_q q = futex_q_init;
1931         int res, ret;
1932
1933         if (refill_pi_state_cache())
1934                 return -ENOMEM;
1935
1936         if (time) {
1937                 to = &timeout;
1938                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1939                                       HRTIMER_MODE_ABS);
1940                 hrtimer_init_sleeper(to, current);
1941                 hrtimer_set_expires(&to->timer, *time);
1942         }
1943
1944 retry:
1945         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key);
1946         if (unlikely(ret != 0))
1947                 goto out;
1948
1949 retry_private:
1950         hb = queue_lock(&q);
1951
1952         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1953         if (unlikely(ret)) {
1954                 switch (ret) {
1955                 case 1:
1956                         /* We got the lock. */
1957                         ret = 0;
1958                         goto out_unlock_put_key;
1959                 case -EFAULT:
1960                         goto uaddr_faulted;
1961                 case -EAGAIN:
1962                         /*
1963                          * Task is exiting and we just wait for the
1964                          * exit to complete.
1965                          */
1966                         queue_unlock(&q, hb);
1967                         put_futex_key(&q.key);
1968                         cond_resched();
1969                         goto retry;
1970                 default:
1971                         goto out_unlock_put_key;
1972                 }
1973         }
1974
1975         /*
1976          * Only actually queue now that the atomic ops are done:
1977          */
1978         queue_me(&q, hb);
1979
1980         WARN_ON(!q.pi_state);
1981         /*
1982          * Block on the PI mutex:
1983          */
1984         if (!trylock)
1985                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1986         else {
1987                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1988                 /* Fixup the trylock return value: */
1989                 ret = ret ? 0 : -EWOULDBLOCK;
1990         }
1991
1992         spin_lock(q.lock_ptr);
1993         /*
1994          * Fixup the pi_state owner and possibly acquire the lock if we
1995          * haven't already.
1996          */
1997         res = fixup_owner(uaddr, &q, !ret);
1998         /*
1999          * If fixup_owner() returned an error, proprogate that.  If it acquired
2000          * the lock, clear our -ETIMEDOUT or -EINTR.
2001          */
2002         if (res)
2003                 ret = (res < 0) ? res : 0;
2004
2005         /*
2006          * If fixup_owner() faulted and was unable to handle the fault, unlock
2007          * it and return the fault to userspace.
2008          */
2009         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2010                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2011
2012         /* Unqueue and drop the lock */
2013         unqueue_me_pi(&q);
2014
2015         goto out_put_key;
2016
2017 out_unlock_put_key:
2018         queue_unlock(&q, hb);
2019
2020 out_put_key:
2021         put_futex_key(&q.key);
2022 out:
2023         if (to)
2024                 destroy_hrtimer_on_stack(&to->timer);
2025         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2026
2027 uaddr_faulted:
2028         queue_unlock(&q, hb);
2029
2030         ret = fault_in_user_writeable(uaddr);
2031         if (ret)
2032                 goto out_put_key;
2033
2034         if (!(flags & FLAGS_SHARED))
2035                 goto retry_private;
2036
2037         put_futex_key(&q.key);
2038         goto retry;
2039 }
2040
2041 /*
2042  * Userspace attempted a TID -> 0 atomic transition, and failed.
2043  * This is the in-kernel slowpath: we look up the PI state (if any),
2044  * and do the rt-mutex unlock.
2045  */
2046 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2047 {
2048         struct futex_hash_bucket *hb;
2049         struct futex_q *this, *next;
2050         u32 uval;
2051         struct plist_head *head;
2052         union futex_key key = FUTEX_KEY_INIT;
2053         int ret;
2054
2055 retry:
2056         if (get_user(uval, uaddr))
2057                 return -EFAULT;
2058         /*
2059          * We release only a lock we actually own:
2060          */
2061         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2062                 return -EPERM;
2063
2064         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key);
2065         if (unlikely(ret != 0))
2066                 goto out;
2067
2068         hb = hash_futex(&key);
2069         spin_lock(&hb->lock);
2070
2071         /*
2072          * To avoid races, try to do the TID -> 0 atomic transition
2073          * again. If it succeeds then we can return without waking
2074          * anyone else up:
2075          */
2076         if (!(uval & FUTEX_OWNER_DIED))
2077                 uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2078
2079
2080         if (unlikely(uval == -EFAULT))
2081                 goto pi_faulted;
2082         /*
2083          * Rare case: we managed to release the lock atomically,
2084          * no need to wake anyone else up:
2085          */
2086         if (unlikely(uval == task_pid_vnr(current)))
2087                 goto out_unlock;
2088
2089         /*
2090          * Ok, other tasks may need to be woken up - check waiters
2091          * and do the wakeup if necessary:
2092          */
2093         head = &hb->chain;
2094
2095         plist_for_each_entry_safe(this, next, head, list) {
2096                 if (!match_futex (&this->key, &key))
2097                         continue;
2098                 ret = wake_futex_pi(uaddr, uval, this);
2099                 /*
2100                  * The atomic access to the futex value
2101                  * generated a pagefault, so retry the
2102                  * user-access and the wakeup:
2103                  */
2104                 if (ret == -EFAULT)
2105                         goto pi_faulted;
2106                 goto out_unlock;
2107         }
2108         /*
2109          * No waiters - kernel unlocks the futex:
2110          */
2111         if (!(uval & FUTEX_OWNER_DIED)) {
2112                 ret = unlock_futex_pi(uaddr, uval);
2113                 if (ret == -EFAULT)
2114                         goto pi_faulted;
2115         }
2116
2117 out_unlock:
2118         spin_unlock(&hb->lock);
2119         put_futex_key(&key);
2120
2121 out:
2122         return ret;
2123
2124 pi_faulted:
2125         spin_unlock(&hb->lock);
2126         put_futex_key(&key);
2127
2128         ret = fault_in_user_writeable(uaddr);
2129         if (!ret)
2130                 goto retry;
2131
2132         return ret;
2133 }
2134
2135 /**
2136  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2137  * @hb:         the hash_bucket futex_q was original enqueued on
2138  * @q:          the futex_q woken while waiting to be requeued
2139  * @key2:       the futex_key of the requeue target futex
2140  * @timeout:    the timeout associated with the wait (NULL if none)
2141  *
2142  * Detect if the task was woken on the initial futex as opposed to the requeue
2143  * target futex.  If so, determine if it was a timeout or a signal that caused
2144  * the wakeup and return the appropriate error code to the caller.  Must be
2145  * called with the hb lock held.
2146  *
2147  * Returns
2148  *  0 - no early wakeup detected
2149  * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2150  */
2151 static inline
2152 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2153                                    struct futex_q *q, union futex_key *key2,
2154                                    struct hrtimer_sleeper *timeout)
2155 {
2156         int ret = 0;
2157
2158         /*
2159          * With the hb lock held, we avoid races while we process the wakeup.
2160          * We only need to hold hb (and not hb2) to ensure atomicity as the
2161          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2162          * It can't be requeued from uaddr2 to something else since we don't
2163          * support a PI aware source futex for requeue.
2164          */
2165         if (!match_futex(&q->key, key2)) {
2166                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2167                 /*
2168                  * We were woken prior to requeue by a timeout or a signal.
2169                  * Unqueue the futex_q and determine which it was.
2170                  */
2171                 plist_del(&q->list, &q->list.plist);
2172
2173                 /* Handle spurious wakeups gracefully */
2174                 ret = -EWOULDBLOCK;
2175                 if (timeout && !timeout->task)
2176                         ret = -ETIMEDOUT;
2177                 else if (signal_pending(current))
2178                         ret = -ERESTARTNOINTR;
2179         }
2180         return ret;
2181 }
2182
2183 /**
2184  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2185  * @uaddr:      the futex we initially wait on (non-pi)
2186  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2187  *              the same type, no requeueing from private to shared, etc.
2188  * @val:        the expected value of uaddr
2189  * @abs_time:   absolute timeout
2190  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2191  * @clockrt:    whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2192  * @uaddr2:     the pi futex we will take prior to returning to user-space
2193  *
2194  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2195  * uaddr2 which must be PI aware.  Normal wakeup will wake on uaddr2 and
2196  * complete the acquisition of the rt_mutex prior to returning to userspace.
2197  * This ensures the rt_mutex maintains an owner when it has waiters; without
2198  * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2199  * need to.
2200  *
2201  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2202  * via the following:
2203  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2204  * 2) wakeup on uaddr2 after a requeue
2205  * 3) signal
2206  * 4) timeout
2207  *
2208  * If 3, cleanup and return -ERESTARTNOINTR.
2209  *
2210  * If 2, we may then block on trying to take the rt_mutex and return via:
2211  * 5) successful lock
2212  * 6) signal
2213  * 7) timeout
2214  * 8) other lock acquisition failure
2215  *
2216  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2217  *
2218  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2219  *
2220  * Returns:
2221  *  0 - On success
2222  * <0 - On error
2223  */
2224 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2225                                  u32 val, ktime_t *abs_time, u32 bitset,
2226                                  u32 __user *uaddr2)
2227 {
2228         struct hrtimer_sleeper timeout, *to = NULL;
2229         struct rt_mutex_waiter rt_waiter;
2230         struct rt_mutex *pi_mutex = NULL;
2231         struct futex_hash_bucket *hb;
2232         union futex_key key2 = FUTEX_KEY_INIT;
2233         struct futex_q q = futex_q_init;
2234         int res, ret;
2235
2236         if (!bitset)
2237                 return -EINVAL;
2238
2239         if (abs_time) {
2240                 to = &timeout;
2241                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2242                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2243                                       HRTIMER_MODE_ABS);
2244                 hrtimer_init_sleeper(to, current);
2245                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2246                                              current->timer_slack_ns);
2247         }
2248
2249         /*
2250          * The waiter is allocated on our stack, manipulated by the requeue
2251          * code while we sleep on uaddr.
2252          */
2253         debug_rt_mutex_init_waiter(&rt_waiter);
2254         rt_waiter.task = NULL;
2255
2256         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2);
2257         if (unlikely(ret != 0))
2258                 goto out;
2259
2260         q.bitset = bitset;
2261         q.rt_waiter = &rt_waiter;
2262         q.requeue_pi_key = &key2;
2263
2264         /*
2265          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2266          * count.
2267          */
2268         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2269         if (ret)
2270                 goto out_key2;
2271
2272         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2273         futex_wait_queue_me(hb, &q, to);
2274
2275         spin_lock(&hb->lock);
2276         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2277         spin_unlock(&hb->lock);
2278         if (ret)
2279                 goto out_put_keys;
2280
2281         /*
2282          * In order for us to be here, we know our q.key == key2, and since
2283          * we took the hb->lock above, we also know that futex_requeue() has
2284          * completed and we no longer have to concern ourselves with a wakeup
2285          * race with the atomic proxy lock acquisition by the requeue code. The
2286          * futex_requeue dropped our key1 reference and incremented our key2
2287          * reference count.
2288          */
2289
2290         /* Check if the requeue code acquired the second futex for us. */
2291         if (!q.rt_waiter) {
2292                 /*
2293                  * Got the lock. We might not be the anticipated owner if we
2294                  * did a lock-steal - fix up the PI-state in that case.
2295                  */
2296                 if (q.pi_state && (q.pi_state->owner != current)) {
2297                         spin_lock(q.lock_ptr);
2298                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2299                         spin_unlock(q.lock_ptr);
2300                 }
2301         } else {
2302                 /*
2303                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2304                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2305                  * the pi_state.
2306                  */
2307                 WARN_ON(!&q.pi_state);
2308                 pi_mutex = &q.pi_state->pi_mutex;
2309                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2310                 debug_rt_mutex_free_waiter(&rt_waiter);
2311
2312                 spin_lock(q.lock_ptr);
2313                 /*
2314                  * Fixup the pi_state owner and possibly acquire the lock if we
2315                  * haven't already.
2316                  */
2317                 res = fixup_owner(uaddr2, &q, !ret);
2318                 /*
2319                  * If fixup_owner() returned an error, proprogate that.  If it
2320                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2321                  */
2322                 if (res)
2323                         ret = (res < 0) ? res : 0;
2324
2325                 /* Unqueue and drop the lock. */
2326                 unqueue_me_pi(&q);
2327         }
2328
2329         /*
2330          * If fixup_pi_state_owner() faulted and was unable to handle the
2331          * fault, unlock the rt_mutex and return the fault to userspace.
2332          */
2333         if (ret == -EFAULT) {
2334                 if (rt_mutex_owner(pi_mutex) == current)
2335                         rt_mutex_unlock(pi_mutex);
2336         } else if (ret == -EINTR) {
2337                 /*
2338                  * We've already been requeued, but cannot restart by calling
2339                  * futex_lock_pi() directly. We could restart this syscall, but
2340                  * it would detect that the user space "val" changed and return
2341                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2342                  * -EWOULDBLOCK directly.
2343                  */
2344                 ret = -EWOULDBLOCK;
2345         }
2346
2347 out_put_keys:
2348         put_futex_key(&q.key);
2349 out_key2:
2350         put_futex_key(&key2);
2351
2352 out:
2353         if (to) {
2354                 hrtimer_cancel(&to->timer);
2355                 destroy_hrtimer_on_stack(&to->timer);
2356         }
2357         return ret;
2358 }
2359
2360 /*
2361  * Support for robust futexes: the kernel cleans up held futexes at
2362  * thread exit time.
2363  *
2364  * Implementation: user-space maintains a per-thread list of locks it
2365  * is holding. Upon do_exit(), the kernel carefully walks this list,
2366  * and marks all locks that are owned by this thread with the
2367  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2368  * always manipulated with the lock held, so the list is private and
2369  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2370  * field, to allow the kernel to clean up if the thread dies after
2371  * acquiring the lock, but just before it could have added itself to
2372  * the list. There can only be one such pending lock.
2373  */
2374
2375 /**
2376  * sys_set_robust_list() - Set the robust-futex list head of a task
2377  * @head:       pointer to the list-head
2378  * @len:        length of the list-head, as userspace expects
2379  */
2380 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2381                 size_t, len)
2382 {
2383         if (!futex_cmpxchg_enabled)
2384                 return -ENOSYS;
2385         /*
2386          * The kernel knows only one size for now:
2387          */
2388         if (unlikely(len != sizeof(*head)))
2389                 return -EINVAL;
2390
2391         current->robust_list = head;
2392
2393         return 0;
2394 }
2395
2396 /**
2397  * sys_get_robust_list() - Get the robust-futex list head of a task
2398  * @pid:        pid of the process [zero for current task]
2399  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2400  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2401  */
2402 SYSCALL_DEFINE3(get_robust_list, int, pid,
2403                 struct robust_list_head __user * __user *, head_ptr,
2404                 size_t __user *, len_ptr)
2405 {
2406         struct robust_list_head __user *head;
2407         unsigned long ret;
2408         const struct cred *cred = current_cred(), *pcred;
2409
2410         if (!futex_cmpxchg_enabled)
2411                 return -ENOSYS;
2412
2413         if (!pid)
2414                 head = current->robust_list;
2415         else {
2416                 struct task_struct *p;
2417
2418                 ret = -ESRCH;
2419                 rcu_read_lock();
2420                 p = find_task_by_vpid(pid);
2421                 if (!p)
2422                         goto err_unlock;
2423                 ret = -EPERM;
2424                 pcred = __task_cred(p);
2425                 if (cred->euid != pcred->euid &&
2426                     cred->euid != pcred->uid &&
2427                     !capable(CAP_SYS_PTRACE))
2428                         goto err_unlock;
2429                 head = p->robust_list;
2430                 rcu_read_unlock();
2431         }
2432
2433         if (put_user(sizeof(*head), len_ptr))
2434                 return -EFAULT;
2435         return put_user(head, head_ptr);
2436
2437 err_unlock:
2438         rcu_read_unlock();
2439
2440         return ret;
2441 }
2442
2443 /*
2444  * Process a futex-list entry, check whether it's owned by the
2445  * dying task, and do notification if so:
2446  */
2447 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2448 {
2449         u32 uval, nval, mval;
2450
2451 retry:
2452         if (get_user(uval, uaddr))
2453                 return -1;
2454
2455         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2456                 /*
2457                  * Ok, this dying thread is truly holding a futex
2458                  * of interest. Set the OWNER_DIED bit atomically
2459                  * via cmpxchg, and if the value had FUTEX_WAITERS
2460                  * set, wake up a waiter (if any). (We have to do a
2461                  * futex_wake() even if OWNER_DIED is already set -
2462                  * to handle the rare but possible case of recursive
2463                  * thread-death.) The rest of the cleanup is done in
2464                  * userspace.
2465                  */
2466                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2467                 nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2468
2469                 if (nval == -EFAULT)
2470                         return -1;
2471
2472                 if (nval != uval)
2473                         goto retry;
2474
2475                 /*
2476                  * Wake robust non-PI futexes here. The wakeup of
2477                  * PI futexes happens in exit_pi_state():
2478                  */
2479                 if (!pi && (uval & FUTEX_WAITERS))
2480                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2481         }
2482         return 0;
2483 }
2484
2485 /*
2486  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2487  */
2488 static inline int fetch_robust_entry(struct robust_list __user **entry,
2489                                      struct robust_list __user * __user *head,
2490                                      unsigned int *pi)
2491 {
2492         unsigned long uentry;
2493
2494         if (get_user(uentry, (unsigned long __user *)head))
2495                 return -EFAULT;
2496
2497         *entry = (void __user *)(uentry & ~1UL);
2498         *pi = uentry & 1;
2499
2500         return 0;
2501 }
2502
2503 /*
2504  * Walk curr->robust_list (very carefully, it's a userspace list!)
2505  * and mark any locks found there dead, and notify any waiters.
2506  *
2507  * We silently return on any sign of list-walking problem.
2508  */
2509 void exit_robust_list(struct task_struct *curr)
2510 {
2511         struct robust_list_head __user *head = curr->robust_list;
2512         struct robust_list __user *entry, *next_entry, *pending;
2513         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2514         unsigned int uninitialized_var(next_pi);
2515         unsigned long futex_offset;
2516         int rc;
2517
2518         if (!futex_cmpxchg_enabled)
2519                 return;
2520
2521         /*
2522          * Fetch the list head (which was registered earlier, via
2523          * sys_set_robust_list()):
2524          */
2525         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2526                 return;
2527         /*
2528          * Fetch the relative futex offset:
2529          */
2530         if (get_user(futex_offset, &head->futex_offset))
2531                 return;
2532         /*
2533          * Fetch any possibly pending lock-add first, and handle it
2534          * if it exists:
2535          */
2536         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2537                 return;
2538
2539         next_entry = NULL;      /* avoid warning with gcc */
2540         while (entry != &head->list) {
2541                 /*
2542                  * Fetch the next entry in the list before calling
2543                  * handle_futex_death:
2544                  */
2545                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2546                 /*
2547                  * A pending lock might already be on the list, so
2548                  * don't process it twice:
2549                  */
2550                 if (entry != pending)
2551                         if (handle_futex_death((void __user *)entry + futex_offset,
2552                                                 curr, pi))
2553                                 return;
2554                 if (rc)
2555                         return;
2556                 entry = next_entry;
2557                 pi = next_pi;
2558                 /*
2559                  * Avoid excessively long or circular lists:
2560                  */
2561                 if (!--limit)
2562                         break;
2563
2564                 cond_resched();
2565         }
2566
2567         if (pending)
2568                 handle_futex_death((void __user *)pending + futex_offset,
2569                                    curr, pip);
2570 }
2571
2572 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2573                 u32 __user *uaddr2, u32 val2, u32 val3)
2574 {
2575         int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2576         unsigned int flags = 0;
2577
2578         if (!(op & FUTEX_PRIVATE_FLAG))
2579                 flags |= FLAGS_SHARED;
2580
2581         if (op & FUTEX_CLOCK_REALTIME) {
2582                 flags |= FLAGS_CLOCKRT;
2583                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2584                         return -ENOSYS;
2585         }
2586
2587         switch (cmd) {
2588         case FUTEX_WAIT:
2589                 val3 = FUTEX_BITSET_MATCH_ANY;
2590         case FUTEX_WAIT_BITSET:
2591                 ret = futex_wait(uaddr, flags, val, timeout, val3);
2592                 break;
2593         case FUTEX_WAKE:
2594                 val3 = FUTEX_BITSET_MATCH_ANY;
2595         case FUTEX_WAKE_BITSET:
2596                 ret = futex_wake(uaddr, flags, val, val3);
2597                 break;
2598         case FUTEX_REQUEUE:
2599                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2600                 break;
2601         case FUTEX_CMP_REQUEUE:
2602                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2603                 break;
2604         case FUTEX_WAKE_OP:
2605                 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2606                 break;
2607         case FUTEX_LOCK_PI:
2608                 if (futex_cmpxchg_enabled)
2609                         ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2610                 break;
2611         case FUTEX_UNLOCK_PI:
2612                 if (futex_cmpxchg_enabled)
2613                         ret = futex_unlock_pi(uaddr, flags);
2614                 break;
2615         case FUTEX_TRYLOCK_PI:
2616                 if (futex_cmpxchg_enabled)
2617                         ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2618                 break;
2619         case FUTEX_WAIT_REQUEUE_PI:
2620                 val3 = FUTEX_BITSET_MATCH_ANY;
2621                 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2622                                             uaddr2);
2623                 break;
2624         case FUTEX_CMP_REQUEUE_PI:
2625                 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2626                 break;
2627         default:
2628                 ret = -ENOSYS;
2629         }
2630         return ret;
2631 }
2632
2633
2634 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2635                 struct timespec __user *, utime, u32 __user *, uaddr2,
2636                 u32, val3)
2637 {
2638         struct timespec ts;
2639         ktime_t t, *tp = NULL;
2640         u32 val2 = 0;
2641         int cmd = op & FUTEX_CMD_MASK;
2642
2643         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2644                       cmd == FUTEX_WAIT_BITSET ||
2645                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2646                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2647                         return -EFAULT;
2648                 if (!timespec_valid(&ts))
2649                         return -EINVAL;
2650
2651                 t = timespec_to_ktime(ts);
2652                 if (cmd == FUTEX_WAIT)
2653                         t = ktime_add_safe(ktime_get(), t);
2654                 tp = &t;
2655         }
2656         /*
2657          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2658          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2659          */
2660         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2661             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2662                 val2 = (u32) (unsigned long) utime;
2663
2664         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2665 }
2666
2667 static int __init futex_init(void)
2668 {
2669         u32 curval;
2670         int i;
2671
2672         /*
2673          * This will fail and we want it. Some arch implementations do
2674          * runtime detection of the futex_atomic_cmpxchg_inatomic()
2675          * functionality. We want to know that before we call in any
2676          * of the complex code paths. Also we want to prevent
2677          * registration of robust lists in that case. NULL is
2678          * guaranteed to fault and we get -EFAULT on functional
2679          * implementation, the non-functional ones will return
2680          * -ENOSYS.
2681          */
2682         curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2683         if (curval == -EFAULT)
2684                 futex_cmpxchg_enabled = 1;
2685
2686         for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2687                 plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2688                 spin_lock_init(&futex_queues[i].lock);
2689         }
2690
2691         return 0;
2692 }
2693 __initcall(futex_init);