]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - kernel/futex.c
Merge branch 'x86-efi-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[karo-tx-linux.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/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
67
68 #include <asm/futex.h>
69
70 #include "locking/rtmutex_common.h"
71
72 /*
73  * READ this before attempting to hack on futexes!
74  *
75  * Basic futex operation and ordering guarantees
76  * =============================================
77  *
78  * The waiter reads the futex value in user space and calls
79  * futex_wait(). This function computes the hash bucket and acquires
80  * the hash bucket lock. After that it reads the futex user space value
81  * again and verifies that the data has not changed. If it has not changed
82  * it enqueues itself into the hash bucket, releases the hash bucket lock
83  * and schedules.
84  *
85  * The waker side modifies the user space value of the futex and calls
86  * futex_wake(). This function computes the hash bucket and acquires the
87  * hash bucket lock. Then it looks for waiters on that futex in the hash
88  * bucket and wakes them.
89  *
90  * In futex wake up scenarios where no tasks are blocked on a futex, taking
91  * the hb spinlock can be avoided and simply return. In order for this
92  * optimization to work, ordering guarantees must exist so that the waiter
93  * being added to the list is acknowledged when the list is concurrently being
94  * checked by the waker, avoiding scenarios like the following:
95  *
96  * CPU 0                               CPU 1
97  * val = *futex;
98  * sys_futex(WAIT, futex, val);
99  *   futex_wait(futex, val);
100  *   uval = *futex;
101  *                                     *futex = newval;
102  *                                     sys_futex(WAKE, futex);
103  *                                       futex_wake(futex);
104  *                                       if (queue_empty())
105  *                                         return;
106  *   if (uval == val)
107  *      lock(hash_bucket(futex));
108  *      queue();
109  *     unlock(hash_bucket(futex));
110  *     schedule();
111  *
112  * This would cause the waiter on CPU 0 to wait forever because it
113  * missed the transition of the user space value from val to newval
114  * and the waker did not find the waiter in the hash bucket queue.
115  *
116  * The correct serialization ensures that a waiter either observes
117  * the changed user space value before blocking or is woken by a
118  * concurrent waker:
119  *
120  * CPU 0                                 CPU 1
121  * val = *futex;
122  * sys_futex(WAIT, futex, val);
123  *   futex_wait(futex, val);
124  *
125  *   waiters++; (a)
126  *   mb(); (A) <-- paired with -.
127  *                              |
128  *   lock(hash_bucket(futex));  |
129  *                              |
130  *   uval = *futex;             |
131  *                              |        *futex = newval;
132  *                              |        sys_futex(WAKE, futex);
133  *                              |          futex_wake(futex);
134  *                              |
135  *                              `------->  mb(); (B)
136  *   if (uval == val)
137  *     queue();
138  *     unlock(hash_bucket(futex));
139  *     schedule();                         if (waiters)
140  *                                           lock(hash_bucket(futex));
141  *   else                                    wake_waiters(futex);
142  *     waiters--; (b)                        unlock(hash_bucket(futex));
143  *
144  * Where (A) orders the waiters increment and the futex value read through
145  * atomic operations (see hb_waiters_inc) and where (B) orders the write
146  * to futex and the waiters read -- this is done by the barriers in
147  * get_futex_key_refs(), through either ihold or atomic_inc, depending on the
148  * futex type.
149  *
150  * This yields the following case (where X:=waiters, Y:=futex):
151  *
152  *      X = Y = 0
153  *
154  *      w[X]=1          w[Y]=1
155  *      MB              MB
156  *      r[Y]=y          r[X]=x
157  *
158  * Which guarantees that x==0 && y==0 is impossible; which translates back into
159  * the guarantee that we cannot both miss the futex variable change and the
160  * enqueue.
161  *
162  * Note that a new waiter is accounted for in (a) even when it is possible that
163  * the wait call can return error, in which case we backtrack from it in (b).
164  * Refer to the comment in queue_lock().
165  *
166  * Similarly, in order to account for waiters being requeued on another
167  * address we always increment the waiters for the destination bucket before
168  * acquiring the lock. It then decrements them again  after releasing it -
169  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170  * will do the additional required waiter count housekeeping. This is done for
171  * double_lock_hb() and double_unlock_hb(), respectively.
172  */
173
174 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175 int __read_mostly futex_cmpxchg_enabled;
176 #endif
177
178 /*
179  * Futex flags used to encode options to functions and preserve them across
180  * restarts.
181  */
182 #define FLAGS_SHARED            0x01
183 #define FLAGS_CLOCKRT           0x02
184 #define FLAGS_HAS_TIMEOUT       0x04
185
186 /*
187  * Priority Inheritance state:
188  */
189 struct futex_pi_state {
190         /*
191          * list of 'owned' pi_state instances - these have to be
192          * cleaned up in do_exit() if the task exits prematurely:
193          */
194         struct list_head list;
195
196         /*
197          * The PI object:
198          */
199         struct rt_mutex pi_mutex;
200
201         struct task_struct *owner;
202         atomic_t refcount;
203
204         union futex_key key;
205 };
206
207 /**
208  * struct futex_q - The hashed futex queue entry, one per waiting task
209  * @list:               priority-sorted list of tasks waiting on this futex
210  * @task:               the task waiting on the futex
211  * @lock_ptr:           the hash bucket lock
212  * @key:                the key the futex is hashed on
213  * @pi_state:           optional priority inheritance state
214  * @rt_waiter:          rt_waiter storage for use with requeue_pi
215  * @requeue_pi_key:     the requeue_pi target futex key
216  * @bitset:             bitset for the optional bitmasked wakeup
217  *
218  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
219  * we can wake only the relevant ones (hashed queues may be shared).
220  *
221  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223  * The order of wakeup is always to make the first condition true, then
224  * the second.
225  *
226  * PI futexes are typically woken before they are removed from the hash list via
227  * the rt_mutex code. See unqueue_me_pi().
228  */
229 struct futex_q {
230         struct plist_node list;
231
232         struct task_struct *task;
233         spinlock_t *lock_ptr;
234         union futex_key key;
235         struct futex_pi_state *pi_state;
236         struct rt_mutex_waiter *rt_waiter;
237         union futex_key *requeue_pi_key;
238         u32 bitset;
239 };
240
241 static const struct futex_q futex_q_init = {
242         /* list gets initialized in queue_me()*/
243         .key = FUTEX_KEY_INIT,
244         .bitset = FUTEX_BITSET_MATCH_ANY
245 };
246
247 /*
248  * Hash buckets are shared by all the futex_keys that hash to the same
249  * location.  Each key may have multiple futex_q structures, one for each task
250  * waiting on a futex.
251  */
252 struct futex_hash_bucket {
253         atomic_t waiters;
254         spinlock_t lock;
255         struct plist_head chain;
256 } ____cacheline_aligned_in_smp;
257
258 static unsigned long __read_mostly futex_hashsize;
259
260 static struct futex_hash_bucket *futex_queues;
261
262 static inline void futex_get_mm(union futex_key *key)
263 {
264         atomic_inc(&key->private.mm->mm_count);
265         /*
266          * Ensure futex_get_mm() implies a full barrier such that
267          * get_futex_key() implies a full barrier. This is relied upon
268          * as full barrier (B), see the ordering comment above.
269          */
270         smp_mb__after_atomic();
271 }
272
273 /*
274  * Reflects a new waiter being added to the waitqueue.
275  */
276 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
277 {
278 #ifdef CONFIG_SMP
279         atomic_inc(&hb->waiters);
280         /*
281          * Full barrier (A), see the ordering comment above.
282          */
283         smp_mb__after_atomic();
284 #endif
285 }
286
287 /*
288  * Reflects a waiter being removed from the waitqueue by wakeup
289  * paths.
290  */
291 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
292 {
293 #ifdef CONFIG_SMP
294         atomic_dec(&hb->waiters);
295 #endif
296 }
297
298 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
299 {
300 #ifdef CONFIG_SMP
301         return atomic_read(&hb->waiters);
302 #else
303         return 1;
304 #endif
305 }
306
307 /*
308  * We hash on the keys returned from get_futex_key (see below).
309  */
310 static struct futex_hash_bucket *hash_futex(union futex_key *key)
311 {
312         u32 hash = jhash2((u32*)&key->both.word,
313                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
314                           key->both.offset);
315         return &futex_queues[hash & (futex_hashsize - 1)];
316 }
317
318 /*
319  * Return 1 if two futex_keys are equal, 0 otherwise.
320  */
321 static inline int match_futex(union futex_key *key1, union futex_key *key2)
322 {
323         return (key1 && key2
324                 && key1->both.word == key2->both.word
325                 && key1->both.ptr == key2->both.ptr
326                 && key1->both.offset == key2->both.offset);
327 }
328
329 /*
330  * Take a reference to the resource addressed by a key.
331  * Can be called while holding spinlocks.
332  *
333  */
334 static void get_futex_key_refs(union futex_key *key)
335 {
336         if (!key->both.ptr)
337                 return;
338
339         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
340         case FUT_OFF_INODE:
341                 ihold(key->shared.inode); /* implies MB (B) */
342                 break;
343         case FUT_OFF_MMSHARED:
344                 futex_get_mm(key); /* implies MB (B) */
345                 break;
346         }
347 }
348
349 /*
350  * Drop a reference to the resource addressed by a key.
351  * The hash bucket spinlock must not be held.
352  */
353 static void drop_futex_key_refs(union futex_key *key)
354 {
355         if (!key->both.ptr) {
356                 /* If we're here then we tried to put a key we failed to get */
357                 WARN_ON_ONCE(1);
358                 return;
359         }
360
361         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
362         case FUT_OFF_INODE:
363                 iput(key->shared.inode);
364                 break;
365         case FUT_OFF_MMSHARED:
366                 mmdrop(key->private.mm);
367                 break;
368         }
369 }
370
371 /**
372  * get_futex_key() - Get parameters which are the keys for a futex
373  * @uaddr:      virtual address of the futex
374  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
375  * @key:        address where result is stored.
376  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
377  *              VERIFY_WRITE)
378  *
379  * Return: a negative error code or 0
380  *
381  * The key words are stored in *key on success.
382  *
383  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
384  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
385  * We can usually work out the index without swapping in the page.
386  *
387  * lock_page() might sleep, the caller should not hold a spinlock.
388  */
389 static int
390 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
391 {
392         unsigned long address = (unsigned long)uaddr;
393         struct mm_struct *mm = current->mm;
394         struct page *page, *page_head;
395         int err, ro = 0;
396
397         /*
398          * The futex address must be "naturally" aligned.
399          */
400         key->both.offset = address % PAGE_SIZE;
401         if (unlikely((address % sizeof(u32)) != 0))
402                 return -EINVAL;
403         address -= key->both.offset;
404
405         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
406                 return -EFAULT;
407
408         /*
409          * PROCESS_PRIVATE futexes are fast.
410          * As the mm cannot disappear under us and the 'key' only needs
411          * virtual address, we dont even have to find the underlying vma.
412          * Note : We do have to check 'uaddr' is a valid user address,
413          *        but access_ok() should be faster than find_vma()
414          */
415         if (!fshared) {
416                 key->private.mm = mm;
417                 key->private.address = address;
418                 get_futex_key_refs(key);  /* implies MB (B) */
419                 return 0;
420         }
421
422 again:
423         err = get_user_pages_fast(address, 1, 1, &page);
424         /*
425          * If write access is not required (eg. FUTEX_WAIT), try
426          * and get read-only access.
427          */
428         if (err == -EFAULT && rw == VERIFY_READ) {
429                 err = get_user_pages_fast(address, 1, 0, &page);
430                 ro = 1;
431         }
432         if (err < 0)
433                 return err;
434         else
435                 err = 0;
436
437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
438         page_head = page;
439         if (unlikely(PageTail(page))) {
440                 put_page(page);
441                 /* serialize against __split_huge_page_splitting() */
442                 local_irq_disable();
443                 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
444                         page_head = compound_head(page);
445                         /*
446                          * page_head is valid pointer but we must pin
447                          * it before taking the PG_lock and/or
448                          * PG_compound_lock. The moment we re-enable
449                          * irqs __split_huge_page_splitting() can
450                          * return and the head page can be freed from
451                          * under us. We can't take the PG_lock and/or
452                          * PG_compound_lock on a page that could be
453                          * freed from under us.
454                          */
455                         if (page != page_head) {
456                                 get_page(page_head);
457                                 put_page(page);
458                         }
459                         local_irq_enable();
460                 } else {
461                         local_irq_enable();
462                         goto again;
463                 }
464         }
465 #else
466         page_head = compound_head(page);
467         if (page != page_head) {
468                 get_page(page_head);
469                 put_page(page);
470         }
471 #endif
472
473         lock_page(page_head);
474
475         /*
476          * If page_head->mapping is NULL, then it cannot be a PageAnon
477          * page; but it might be the ZERO_PAGE or in the gate area or
478          * in a special mapping (all cases which we are happy to fail);
479          * or it may have been a good file page when get_user_pages_fast
480          * found it, but truncated or holepunched or subjected to
481          * invalidate_complete_page2 before we got the page lock (also
482          * cases which we are happy to fail).  And we hold a reference,
483          * so refcount care in invalidate_complete_page's remove_mapping
484          * prevents drop_caches from setting mapping to NULL beneath us.
485          *
486          * The case we do have to guard against is when memory pressure made
487          * shmem_writepage move it from filecache to swapcache beneath us:
488          * an unlikely race, but we do need to retry for page_head->mapping.
489          */
490         if (!page_head->mapping) {
491                 int shmem_swizzled = PageSwapCache(page_head);
492                 unlock_page(page_head);
493                 put_page(page_head);
494                 if (shmem_swizzled)
495                         goto again;
496                 return -EFAULT;
497         }
498
499         /*
500          * Private mappings are handled in a simple way.
501          *
502          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
503          * it's a read-only handle, it's expected that futexes attach to
504          * the object not the particular process.
505          */
506         if (PageAnon(page_head)) {
507                 /*
508                  * A RO anonymous page will never change and thus doesn't make
509                  * sense for futex operations.
510                  */
511                 if (ro) {
512                         err = -EFAULT;
513                         goto out;
514                 }
515
516                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
517                 key->private.mm = mm;
518                 key->private.address = address;
519         } else {
520                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
521                 key->shared.inode = page_head->mapping->host;
522                 key->shared.pgoff = basepage_index(page);
523         }
524
525         get_futex_key_refs(key); /* implies MB (B) */
526
527 out:
528         unlock_page(page_head);
529         put_page(page_head);
530         return err;
531 }
532
533 static inline void put_futex_key(union futex_key *key)
534 {
535         drop_futex_key_refs(key);
536 }
537
538 /**
539  * fault_in_user_writeable() - Fault in user address and verify RW access
540  * @uaddr:      pointer to faulting user space address
541  *
542  * Slow path to fixup the fault we just took in the atomic write
543  * access to @uaddr.
544  *
545  * We have no generic implementation of a non-destructive write to the
546  * user address. We know that we faulted in the atomic pagefault
547  * disabled section so we can as well avoid the #PF overhead by
548  * calling get_user_pages() right away.
549  */
550 static int fault_in_user_writeable(u32 __user *uaddr)
551 {
552         struct mm_struct *mm = current->mm;
553         int ret;
554
555         down_read(&mm->mmap_sem);
556         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
557                                FAULT_FLAG_WRITE);
558         up_read(&mm->mmap_sem);
559
560         return ret < 0 ? ret : 0;
561 }
562
563 /**
564  * futex_top_waiter() - Return the highest priority waiter on a futex
565  * @hb:         the hash bucket the futex_q's reside in
566  * @key:        the futex key (to distinguish it from other futex futex_q's)
567  *
568  * Must be called with the hb lock held.
569  */
570 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
571                                         union futex_key *key)
572 {
573         struct futex_q *this;
574
575         plist_for_each_entry(this, &hb->chain, list) {
576                 if (match_futex(&this->key, key))
577                         return this;
578         }
579         return NULL;
580 }
581
582 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
583                                       u32 uval, u32 newval)
584 {
585         int ret;
586
587         pagefault_disable();
588         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
589         pagefault_enable();
590
591         return ret;
592 }
593
594 static int get_futex_value_locked(u32 *dest, u32 __user *from)
595 {
596         int ret;
597
598         pagefault_disable();
599         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
600         pagefault_enable();
601
602         return ret ? -EFAULT : 0;
603 }
604
605
606 /*
607  * PI code:
608  */
609 static int refill_pi_state_cache(void)
610 {
611         struct futex_pi_state *pi_state;
612
613         if (likely(current->pi_state_cache))
614                 return 0;
615
616         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
617
618         if (!pi_state)
619                 return -ENOMEM;
620
621         INIT_LIST_HEAD(&pi_state->list);
622         /* pi_mutex gets initialized later */
623         pi_state->owner = NULL;
624         atomic_set(&pi_state->refcount, 1);
625         pi_state->key = FUTEX_KEY_INIT;
626
627         current->pi_state_cache = pi_state;
628
629         return 0;
630 }
631
632 static struct futex_pi_state * alloc_pi_state(void)
633 {
634         struct futex_pi_state *pi_state = current->pi_state_cache;
635
636         WARN_ON(!pi_state);
637         current->pi_state_cache = NULL;
638
639         return pi_state;
640 }
641
642 static void free_pi_state(struct futex_pi_state *pi_state)
643 {
644         if (!atomic_dec_and_test(&pi_state->refcount))
645                 return;
646
647         /*
648          * If pi_state->owner is NULL, the owner is most probably dying
649          * and has cleaned up the pi_state already
650          */
651         if (pi_state->owner) {
652                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
653                 list_del_init(&pi_state->list);
654                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
655
656                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
657         }
658
659         if (current->pi_state_cache)
660                 kfree(pi_state);
661         else {
662                 /*
663                  * pi_state->list is already empty.
664                  * clear pi_state->owner.
665                  * refcount is at 0 - put it back to 1.
666                  */
667                 pi_state->owner = NULL;
668                 atomic_set(&pi_state->refcount, 1);
669                 current->pi_state_cache = pi_state;
670         }
671 }
672
673 /*
674  * Look up the task based on what TID userspace gave us.
675  * We dont trust it.
676  */
677 static struct task_struct * futex_find_get_task(pid_t pid)
678 {
679         struct task_struct *p;
680
681         rcu_read_lock();
682         p = find_task_by_vpid(pid);
683         if (p)
684                 get_task_struct(p);
685
686         rcu_read_unlock();
687
688         return p;
689 }
690
691 /*
692  * This task is holding PI mutexes at exit time => bad.
693  * Kernel cleans up PI-state, but userspace is likely hosed.
694  * (Robust-futex cleanup is separate and might save the day for userspace.)
695  */
696 void exit_pi_state_list(struct task_struct *curr)
697 {
698         struct list_head *next, *head = &curr->pi_state_list;
699         struct futex_pi_state *pi_state;
700         struct futex_hash_bucket *hb;
701         union futex_key key = FUTEX_KEY_INIT;
702
703         if (!futex_cmpxchg_enabled)
704                 return;
705         /*
706          * We are a ZOMBIE and nobody can enqueue itself on
707          * pi_state_list anymore, but we have to be careful
708          * versus waiters unqueueing themselves:
709          */
710         raw_spin_lock_irq(&curr->pi_lock);
711         while (!list_empty(head)) {
712
713                 next = head->next;
714                 pi_state = list_entry(next, struct futex_pi_state, list);
715                 key = pi_state->key;
716                 hb = hash_futex(&key);
717                 raw_spin_unlock_irq(&curr->pi_lock);
718
719                 spin_lock(&hb->lock);
720
721                 raw_spin_lock_irq(&curr->pi_lock);
722                 /*
723                  * We dropped the pi-lock, so re-check whether this
724                  * task still owns the PI-state:
725                  */
726                 if (head->next != next) {
727                         spin_unlock(&hb->lock);
728                         continue;
729                 }
730
731                 WARN_ON(pi_state->owner != curr);
732                 WARN_ON(list_empty(&pi_state->list));
733                 list_del_init(&pi_state->list);
734                 pi_state->owner = NULL;
735                 raw_spin_unlock_irq(&curr->pi_lock);
736
737                 rt_mutex_unlock(&pi_state->pi_mutex);
738
739                 spin_unlock(&hb->lock);
740
741                 raw_spin_lock_irq(&curr->pi_lock);
742         }
743         raw_spin_unlock_irq(&curr->pi_lock);
744 }
745
746 /*
747  * We need to check the following states:
748  *
749  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
750  *
751  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
752  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
753  *
754  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
755  *
756  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
757  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
758  *
759  * [6]  Found  | Found    | task      | 0         | 1      | Valid
760  *
761  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
762  *
763  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
764  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
765  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
766  *
767  * [1]  Indicates that the kernel can acquire the futex atomically. We
768  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
769  *
770  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
771  *      thread is found then it indicates that the owner TID has died.
772  *
773  * [3]  Invalid. The waiter is queued on a non PI futex
774  *
775  * [4]  Valid state after exit_robust_list(), which sets the user space
776  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
777  *
778  * [5]  The user space value got manipulated between exit_robust_list()
779  *      and exit_pi_state_list()
780  *
781  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
782  *      the pi_state but cannot access the user space value.
783  *
784  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
785  *
786  * [8]  Owner and user space value match
787  *
788  * [9]  There is no transient state which sets the user space TID to 0
789  *      except exit_robust_list(), but this is indicated by the
790  *      FUTEX_OWNER_DIED bit. See [4]
791  *
792  * [10] There is no transient state which leaves owner and user space
793  *      TID out of sync.
794  */
795
796 /*
797  * Validate that the existing waiter has a pi_state and sanity check
798  * the pi_state against the user space value. If correct, attach to
799  * it.
800  */
801 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
802                               struct futex_pi_state **ps)
803 {
804         pid_t pid = uval & FUTEX_TID_MASK;
805
806         /*
807          * Userspace might have messed up non-PI and PI futexes [3]
808          */
809         if (unlikely(!pi_state))
810                 return -EINVAL;
811
812         WARN_ON(!atomic_read(&pi_state->refcount));
813
814         /*
815          * Handle the owner died case:
816          */
817         if (uval & FUTEX_OWNER_DIED) {
818                 /*
819                  * exit_pi_state_list sets owner to NULL and wakes the
820                  * topmost waiter. The task which acquires the
821                  * pi_state->rt_mutex will fixup owner.
822                  */
823                 if (!pi_state->owner) {
824                         /*
825                          * No pi state owner, but the user space TID
826                          * is not 0. Inconsistent state. [5]
827                          */
828                         if (pid)
829                                 return -EINVAL;
830                         /*
831                          * Take a ref on the state and return success. [4]
832                          */
833                         goto out_state;
834                 }
835
836                 /*
837                  * If TID is 0, then either the dying owner has not
838                  * yet executed exit_pi_state_list() or some waiter
839                  * acquired the rtmutex in the pi state, but did not
840                  * yet fixup the TID in user space.
841                  *
842                  * Take a ref on the state and return success. [6]
843                  */
844                 if (!pid)
845                         goto out_state;
846         } else {
847                 /*
848                  * If the owner died bit is not set, then the pi_state
849                  * must have an owner. [7]
850                  */
851                 if (!pi_state->owner)
852                         return -EINVAL;
853         }
854
855         /*
856          * Bail out if user space manipulated the futex value. If pi
857          * state exists then the owner TID must be the same as the
858          * user space TID. [9/10]
859          */
860         if (pid != task_pid_vnr(pi_state->owner))
861                 return -EINVAL;
862 out_state:
863         atomic_inc(&pi_state->refcount);
864         *ps = pi_state;
865         return 0;
866 }
867
868 /*
869  * Lookup the task for the TID provided from user space and attach to
870  * it after doing proper sanity checks.
871  */
872 static int attach_to_pi_owner(u32 uval, union futex_key *key,
873                               struct futex_pi_state **ps)
874 {
875         pid_t pid = uval & FUTEX_TID_MASK;
876         struct futex_pi_state *pi_state;
877         struct task_struct *p;
878
879         /*
880          * We are the first waiter - try to look up the real owner and attach
881          * the new pi_state to it, but bail out when TID = 0 [1]
882          */
883         if (!pid)
884                 return -ESRCH;
885         p = futex_find_get_task(pid);
886         if (!p)
887                 return -ESRCH;
888
889         if (!p->mm) {
890                 put_task_struct(p);
891                 return -EPERM;
892         }
893
894         /*
895          * We need to look at the task state flags to figure out,
896          * whether the task is exiting. To protect against the do_exit
897          * change of the task flags, we do this protected by
898          * p->pi_lock:
899          */
900         raw_spin_lock_irq(&p->pi_lock);
901         if (unlikely(p->flags & PF_EXITING)) {
902                 /*
903                  * The task is on the way out. When PF_EXITPIDONE is
904                  * set, we know that the task has finished the
905                  * cleanup:
906                  */
907                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
908
909                 raw_spin_unlock_irq(&p->pi_lock);
910                 put_task_struct(p);
911                 return ret;
912         }
913
914         /*
915          * No existing pi state. First waiter. [2]
916          */
917         pi_state = alloc_pi_state();
918
919         /*
920          * Initialize the pi_mutex in locked state and make @p
921          * the owner of it:
922          */
923         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
924
925         /* Store the key for possible exit cleanups: */
926         pi_state->key = *key;
927
928         WARN_ON(!list_empty(&pi_state->list));
929         list_add(&pi_state->list, &p->pi_state_list);
930         pi_state->owner = p;
931         raw_spin_unlock_irq(&p->pi_lock);
932
933         put_task_struct(p);
934
935         *ps = pi_state;
936
937         return 0;
938 }
939
940 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
941                            union futex_key *key, struct futex_pi_state **ps)
942 {
943         struct futex_q *match = futex_top_waiter(hb, key);
944
945         /*
946          * If there is a waiter on that futex, validate it and
947          * attach to the pi_state when the validation succeeds.
948          */
949         if (match)
950                 return attach_to_pi_state(uval, match->pi_state, ps);
951
952         /*
953          * We are the first waiter - try to look up the owner based on
954          * @uval and attach to it.
955          */
956         return attach_to_pi_owner(uval, key, ps);
957 }
958
959 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
960 {
961         u32 uninitialized_var(curval);
962
963         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
964                 return -EFAULT;
965
966         /*If user space value changed, let the caller retry */
967         return curval != uval ? -EAGAIN : 0;
968 }
969
970 /**
971  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
972  * @uaddr:              the pi futex user address
973  * @hb:                 the pi futex hash bucket
974  * @key:                the futex key associated with uaddr and hb
975  * @ps:                 the pi_state pointer where we store the result of the
976  *                      lookup
977  * @task:               the task to perform the atomic lock work for.  This will
978  *                      be "current" except in the case of requeue pi.
979  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
980  *
981  * Return:
982  *  0 - ready to wait;
983  *  1 - acquired the lock;
984  * <0 - error
985  *
986  * The hb->lock and futex_key refs shall be held by the caller.
987  */
988 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
989                                 union futex_key *key,
990                                 struct futex_pi_state **ps,
991                                 struct task_struct *task, int set_waiters)
992 {
993         u32 uval, newval, vpid = task_pid_vnr(task);
994         struct futex_q *match;
995         int ret;
996
997         /*
998          * Read the user space value first so we can validate a few
999          * things before proceeding further.
1000          */
1001         if (get_futex_value_locked(&uval, uaddr))
1002                 return -EFAULT;
1003
1004         /*
1005          * Detect deadlocks.
1006          */
1007         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1008                 return -EDEADLK;
1009
1010         /*
1011          * Lookup existing state first. If it exists, try to attach to
1012          * its pi_state.
1013          */
1014         match = futex_top_waiter(hb, key);
1015         if (match)
1016                 return attach_to_pi_state(uval, match->pi_state, ps);
1017
1018         /*
1019          * No waiter and user TID is 0. We are here because the
1020          * waiters or the owner died bit is set or called from
1021          * requeue_cmp_pi or for whatever reason something took the
1022          * syscall.
1023          */
1024         if (!(uval & FUTEX_TID_MASK)) {
1025                 /*
1026                  * We take over the futex. No other waiters and the user space
1027                  * TID is 0. We preserve the owner died bit.
1028                  */
1029                 newval = uval & FUTEX_OWNER_DIED;
1030                 newval |= vpid;
1031
1032                 /* The futex requeue_pi code can enforce the waiters bit */
1033                 if (set_waiters)
1034                         newval |= FUTEX_WAITERS;
1035
1036                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1037                 /* If the take over worked, return 1 */
1038                 return ret < 0 ? ret : 1;
1039         }
1040
1041         /*
1042          * First waiter. Set the waiters bit before attaching ourself to
1043          * the owner. If owner tries to unlock, it will be forced into
1044          * the kernel and blocked on hb->lock.
1045          */
1046         newval = uval | FUTEX_WAITERS;
1047         ret = lock_pi_update_atomic(uaddr, uval, newval);
1048         if (ret)
1049                 return ret;
1050         /*
1051          * If the update of the user space value succeeded, we try to
1052          * attach to the owner. If that fails, no harm done, we only
1053          * set the FUTEX_WAITERS bit in the user space variable.
1054          */
1055         return attach_to_pi_owner(uval, key, ps);
1056 }
1057
1058 /**
1059  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1060  * @q:  The futex_q to unqueue
1061  *
1062  * The q->lock_ptr must not be NULL and must be held by the caller.
1063  */
1064 static void __unqueue_futex(struct futex_q *q)
1065 {
1066         struct futex_hash_bucket *hb;
1067
1068         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1069             || WARN_ON(plist_node_empty(&q->list)))
1070                 return;
1071
1072         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1073         plist_del(&q->list, &hb->chain);
1074         hb_waiters_dec(hb);
1075 }
1076
1077 /*
1078  * The hash bucket lock must be held when this is called.
1079  * Afterwards, the futex_q must not be accessed.
1080  */
1081 static void wake_futex(struct futex_q *q)
1082 {
1083         struct task_struct *p = q->task;
1084
1085         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1086                 return;
1087
1088         /*
1089          * We set q->lock_ptr = NULL _before_ we wake up the task. If
1090          * a non-futex wake up happens on another CPU then the task
1091          * might exit and p would dereference a non-existing task
1092          * struct. Prevent this by holding a reference on p across the
1093          * wake up.
1094          */
1095         get_task_struct(p);
1096
1097         __unqueue_futex(q);
1098         /*
1099          * The waiting task can free the futex_q as soon as
1100          * q->lock_ptr = NULL is written, without taking any locks. A
1101          * memory barrier is required here to prevent the following
1102          * store to lock_ptr from getting ahead of the plist_del.
1103          */
1104         smp_wmb();
1105         q->lock_ptr = NULL;
1106
1107         wake_up_state(p, TASK_NORMAL);
1108         put_task_struct(p);
1109 }
1110
1111 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1112 {
1113         struct task_struct *new_owner;
1114         struct futex_pi_state *pi_state = this->pi_state;
1115         u32 uninitialized_var(curval), newval;
1116         int ret = 0;
1117
1118         if (!pi_state)
1119                 return -EINVAL;
1120
1121         /*
1122          * If current does not own the pi_state then the futex is
1123          * inconsistent and user space fiddled with the futex value.
1124          */
1125         if (pi_state->owner != current)
1126                 return -EINVAL;
1127
1128         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1129         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1130
1131         /*
1132          * It is possible that the next waiter (the one that brought
1133          * this owner to the kernel) timed out and is no longer
1134          * waiting on the lock.
1135          */
1136         if (!new_owner)
1137                 new_owner = this->task;
1138
1139         /*
1140          * We pass it to the next owner. The WAITERS bit is always
1141          * kept enabled while there is PI state around. We cleanup the
1142          * owner died bit, because we are the owner.
1143          */
1144         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1145
1146         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1147                 ret = -EFAULT;
1148         else if (curval != uval)
1149                 ret = -EINVAL;
1150         if (ret) {
1151                 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1152                 return ret;
1153         }
1154
1155         raw_spin_lock_irq(&pi_state->owner->pi_lock);
1156         WARN_ON(list_empty(&pi_state->list));
1157         list_del_init(&pi_state->list);
1158         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1159
1160         raw_spin_lock_irq(&new_owner->pi_lock);
1161         WARN_ON(!list_empty(&pi_state->list));
1162         list_add(&pi_state->list, &new_owner->pi_state_list);
1163         pi_state->owner = new_owner;
1164         raw_spin_unlock_irq(&new_owner->pi_lock);
1165
1166         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1167         rt_mutex_unlock(&pi_state->pi_mutex);
1168
1169         return 0;
1170 }
1171
1172 /*
1173  * Express the locking dependencies for lockdep:
1174  */
1175 static inline void
1176 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1177 {
1178         if (hb1 <= hb2) {
1179                 spin_lock(&hb1->lock);
1180                 if (hb1 < hb2)
1181                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1182         } else { /* hb1 > hb2 */
1183                 spin_lock(&hb2->lock);
1184                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1185         }
1186 }
1187
1188 static inline void
1189 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1190 {
1191         spin_unlock(&hb1->lock);
1192         if (hb1 != hb2)
1193                 spin_unlock(&hb2->lock);
1194 }
1195
1196 /*
1197  * Wake up waiters matching bitset queued on this futex (uaddr).
1198  */
1199 static int
1200 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1201 {
1202         struct futex_hash_bucket *hb;
1203         struct futex_q *this, *next;
1204         union futex_key key = FUTEX_KEY_INIT;
1205         int ret;
1206
1207         if (!bitset)
1208                 return -EINVAL;
1209
1210         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1211         if (unlikely(ret != 0))
1212                 goto out;
1213
1214         hb = hash_futex(&key);
1215
1216         /* Make sure we really have tasks to wakeup */
1217         if (!hb_waiters_pending(hb))
1218                 goto out_put_key;
1219
1220         spin_lock(&hb->lock);
1221
1222         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1223                 if (match_futex (&this->key, &key)) {
1224                         if (this->pi_state || this->rt_waiter) {
1225                                 ret = -EINVAL;
1226                                 break;
1227                         }
1228
1229                         /* Check if one of the bits is set in both bitsets */
1230                         if (!(this->bitset & bitset))
1231                                 continue;
1232
1233                         wake_futex(this);
1234                         if (++ret >= nr_wake)
1235                                 break;
1236                 }
1237         }
1238
1239         spin_unlock(&hb->lock);
1240 out_put_key:
1241         put_futex_key(&key);
1242 out:
1243         return ret;
1244 }
1245
1246 /*
1247  * Wake up all waiters hashed on the physical page that is mapped
1248  * to this virtual address:
1249  */
1250 static int
1251 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1252               int nr_wake, int nr_wake2, int op)
1253 {
1254         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1255         struct futex_hash_bucket *hb1, *hb2;
1256         struct futex_q *this, *next;
1257         int ret, op_ret;
1258
1259 retry:
1260         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1261         if (unlikely(ret != 0))
1262                 goto out;
1263         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1264         if (unlikely(ret != 0))
1265                 goto out_put_key1;
1266
1267         hb1 = hash_futex(&key1);
1268         hb2 = hash_futex(&key2);
1269
1270 retry_private:
1271         double_lock_hb(hb1, hb2);
1272         op_ret = futex_atomic_op_inuser(op, uaddr2);
1273         if (unlikely(op_ret < 0)) {
1274
1275                 double_unlock_hb(hb1, hb2);
1276
1277 #ifndef CONFIG_MMU
1278                 /*
1279                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1280                  * but we might get them from range checking
1281                  */
1282                 ret = op_ret;
1283                 goto out_put_keys;
1284 #endif
1285
1286                 if (unlikely(op_ret != -EFAULT)) {
1287                         ret = op_ret;
1288                         goto out_put_keys;
1289                 }
1290
1291                 ret = fault_in_user_writeable(uaddr2);
1292                 if (ret)
1293                         goto out_put_keys;
1294
1295                 if (!(flags & FLAGS_SHARED))
1296                         goto retry_private;
1297
1298                 put_futex_key(&key2);
1299                 put_futex_key(&key1);
1300                 goto retry;
1301         }
1302
1303         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1304                 if (match_futex (&this->key, &key1)) {
1305                         if (this->pi_state || this->rt_waiter) {
1306                                 ret = -EINVAL;
1307                                 goto out_unlock;
1308                         }
1309                         wake_futex(this);
1310                         if (++ret >= nr_wake)
1311                                 break;
1312                 }
1313         }
1314
1315         if (op_ret > 0) {
1316                 op_ret = 0;
1317                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1318                         if (match_futex (&this->key, &key2)) {
1319                                 if (this->pi_state || this->rt_waiter) {
1320                                         ret = -EINVAL;
1321                                         goto out_unlock;
1322                                 }
1323                                 wake_futex(this);
1324                                 if (++op_ret >= nr_wake2)
1325                                         break;
1326                         }
1327                 }
1328                 ret += op_ret;
1329         }
1330
1331 out_unlock:
1332         double_unlock_hb(hb1, hb2);
1333 out_put_keys:
1334         put_futex_key(&key2);
1335 out_put_key1:
1336         put_futex_key(&key1);
1337 out:
1338         return ret;
1339 }
1340
1341 /**
1342  * requeue_futex() - Requeue a futex_q from one hb to another
1343  * @q:          the futex_q to requeue
1344  * @hb1:        the source hash_bucket
1345  * @hb2:        the target hash_bucket
1346  * @key2:       the new key for the requeued futex_q
1347  */
1348 static inline
1349 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1350                    struct futex_hash_bucket *hb2, union futex_key *key2)
1351 {
1352
1353         /*
1354          * If key1 and key2 hash to the same bucket, no need to
1355          * requeue.
1356          */
1357         if (likely(&hb1->chain != &hb2->chain)) {
1358                 plist_del(&q->list, &hb1->chain);
1359                 hb_waiters_dec(hb1);
1360                 plist_add(&q->list, &hb2->chain);
1361                 hb_waiters_inc(hb2);
1362                 q->lock_ptr = &hb2->lock;
1363         }
1364         get_futex_key_refs(key2);
1365         q->key = *key2;
1366 }
1367
1368 /**
1369  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1370  * @q:          the futex_q
1371  * @key:        the key of the requeue target futex
1372  * @hb:         the hash_bucket of the requeue target futex
1373  *
1374  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1375  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1376  * to the requeue target futex so the waiter can detect the wakeup on the right
1377  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1378  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1379  * to protect access to the pi_state to fixup the owner later.  Must be called
1380  * with both q->lock_ptr and hb->lock held.
1381  */
1382 static inline
1383 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1384                            struct futex_hash_bucket *hb)
1385 {
1386         get_futex_key_refs(key);
1387         q->key = *key;
1388
1389         __unqueue_futex(q);
1390
1391         WARN_ON(!q->rt_waiter);
1392         q->rt_waiter = NULL;
1393
1394         q->lock_ptr = &hb->lock;
1395
1396         wake_up_state(q->task, TASK_NORMAL);
1397 }
1398
1399 /**
1400  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1401  * @pifutex:            the user address of the to futex
1402  * @hb1:                the from futex hash bucket, must be locked by the caller
1403  * @hb2:                the to futex hash bucket, must be locked by the caller
1404  * @key1:               the from futex key
1405  * @key2:               the to futex key
1406  * @ps:                 address to store the pi_state pointer
1407  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1408  *
1409  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1410  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1411  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1412  * hb1 and hb2 must be held by the caller.
1413  *
1414  * Return:
1415  *  0 - failed to acquire the lock atomically;
1416  * >0 - acquired the lock, return value is vpid of the top_waiter
1417  * <0 - error
1418  */
1419 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1420                                  struct futex_hash_bucket *hb1,
1421                                  struct futex_hash_bucket *hb2,
1422                                  union futex_key *key1, union futex_key *key2,
1423                                  struct futex_pi_state **ps, int set_waiters)
1424 {
1425         struct futex_q *top_waiter = NULL;
1426         u32 curval;
1427         int ret, vpid;
1428
1429         if (get_futex_value_locked(&curval, pifutex))
1430                 return -EFAULT;
1431
1432         /*
1433          * Find the top_waiter and determine if there are additional waiters.
1434          * If the caller intends to requeue more than 1 waiter to pifutex,
1435          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1436          * as we have means to handle the possible fault.  If not, don't set
1437          * the bit unecessarily as it will force the subsequent unlock to enter
1438          * the kernel.
1439          */
1440         top_waiter = futex_top_waiter(hb1, key1);
1441
1442         /* There are no waiters, nothing for us to do. */
1443         if (!top_waiter)
1444                 return 0;
1445
1446         /* Ensure we requeue to the expected futex. */
1447         if (!match_futex(top_waiter->requeue_pi_key, key2))
1448                 return -EINVAL;
1449
1450         /*
1451          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1452          * the contended case or if set_waiters is 1.  The pi_state is returned
1453          * in ps in contended cases.
1454          */
1455         vpid = task_pid_vnr(top_waiter->task);
1456         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1457                                    set_waiters);
1458         if (ret == 1) {
1459                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1460                 return vpid;
1461         }
1462         return ret;
1463 }
1464
1465 /**
1466  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1467  * @uaddr1:     source futex user address
1468  * @flags:      futex flags (FLAGS_SHARED, etc.)
1469  * @uaddr2:     target futex user address
1470  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1471  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1472  * @cmpval:     @uaddr1 expected value (or %NULL)
1473  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1474  *              pi futex (pi to pi requeue is not supported)
1475  *
1476  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1477  * uaddr2 atomically on behalf of the top waiter.
1478  *
1479  * Return:
1480  * >=0 - on success, the number of tasks requeued or woken;
1481  *  <0 - on error
1482  */
1483 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1484                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1485                          u32 *cmpval, int requeue_pi)
1486 {
1487         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1488         int drop_count = 0, task_count = 0, ret;
1489         struct futex_pi_state *pi_state = NULL;
1490         struct futex_hash_bucket *hb1, *hb2;
1491         struct futex_q *this, *next;
1492
1493         if (requeue_pi) {
1494                 /*
1495                  * Requeue PI only works on two distinct uaddrs. This
1496                  * check is only valid for private futexes. See below.
1497                  */
1498                 if (uaddr1 == uaddr2)
1499                         return -EINVAL;
1500
1501                 /*
1502                  * requeue_pi requires a pi_state, try to allocate it now
1503                  * without any locks in case it fails.
1504                  */
1505                 if (refill_pi_state_cache())
1506                         return -ENOMEM;
1507                 /*
1508                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1509                  * + nr_requeue, since it acquires the rt_mutex prior to
1510                  * returning to userspace, so as to not leave the rt_mutex with
1511                  * waiters and no owner.  However, second and third wake-ups
1512                  * cannot be predicted as they involve race conditions with the
1513                  * first wake and a fault while looking up the pi_state.  Both
1514                  * pthread_cond_signal() and pthread_cond_broadcast() should
1515                  * use nr_wake=1.
1516                  */
1517                 if (nr_wake != 1)
1518                         return -EINVAL;
1519         }
1520
1521 retry:
1522         if (pi_state != NULL) {
1523                 /*
1524                  * We will have to lookup the pi_state again, so free this one
1525                  * to keep the accounting correct.
1526                  */
1527                 free_pi_state(pi_state);
1528                 pi_state = NULL;
1529         }
1530
1531         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1532         if (unlikely(ret != 0))
1533                 goto out;
1534         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1535                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1536         if (unlikely(ret != 0))
1537                 goto out_put_key1;
1538
1539         /*
1540          * The check above which compares uaddrs is not sufficient for
1541          * shared futexes. We need to compare the keys:
1542          */
1543         if (requeue_pi && match_futex(&key1, &key2)) {
1544                 ret = -EINVAL;
1545                 goto out_put_keys;
1546         }
1547
1548         hb1 = hash_futex(&key1);
1549         hb2 = hash_futex(&key2);
1550
1551 retry_private:
1552         hb_waiters_inc(hb2);
1553         double_lock_hb(hb1, hb2);
1554
1555         if (likely(cmpval != NULL)) {
1556                 u32 curval;
1557
1558                 ret = get_futex_value_locked(&curval, uaddr1);
1559
1560                 if (unlikely(ret)) {
1561                         double_unlock_hb(hb1, hb2);
1562                         hb_waiters_dec(hb2);
1563
1564                         ret = get_user(curval, uaddr1);
1565                         if (ret)
1566                                 goto out_put_keys;
1567
1568                         if (!(flags & FLAGS_SHARED))
1569                                 goto retry_private;
1570
1571                         put_futex_key(&key2);
1572                         put_futex_key(&key1);
1573                         goto retry;
1574                 }
1575                 if (curval != *cmpval) {
1576                         ret = -EAGAIN;
1577                         goto out_unlock;
1578                 }
1579         }
1580
1581         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1582                 /*
1583                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1584                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1585                  * bit.  We force this here where we are able to easily handle
1586                  * faults rather in the requeue loop below.
1587                  */
1588                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1589                                                  &key2, &pi_state, nr_requeue);
1590
1591                 /*
1592                  * At this point the top_waiter has either taken uaddr2 or is
1593                  * waiting on it.  If the former, then the pi_state will not
1594                  * exist yet, look it up one more time to ensure we have a
1595                  * reference to it. If the lock was taken, ret contains the
1596                  * vpid of the top waiter task.
1597                  */
1598                 if (ret > 0) {
1599                         WARN_ON(pi_state);
1600                         drop_count++;
1601                         task_count++;
1602                         /*
1603                          * If we acquired the lock, then the user
1604                          * space value of uaddr2 should be vpid. It
1605                          * cannot be changed by the top waiter as it
1606                          * is blocked on hb2 lock if it tries to do
1607                          * so. If something fiddled with it behind our
1608                          * back the pi state lookup might unearth
1609                          * it. So we rather use the known value than
1610                          * rereading and handing potential crap to
1611                          * lookup_pi_state.
1612                          */
1613                         ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1614                 }
1615
1616                 switch (ret) {
1617                 case 0:
1618                         break;
1619                 case -EFAULT:
1620                         double_unlock_hb(hb1, hb2);
1621                         hb_waiters_dec(hb2);
1622                         put_futex_key(&key2);
1623                         put_futex_key(&key1);
1624                         ret = fault_in_user_writeable(uaddr2);
1625                         if (!ret)
1626                                 goto retry;
1627                         goto out;
1628                 case -EAGAIN:
1629                         /*
1630                          * Two reasons for this:
1631                          * - Owner is exiting and we just wait for the
1632                          *   exit to complete.
1633                          * - The user space value changed.
1634                          */
1635                         double_unlock_hb(hb1, hb2);
1636                         hb_waiters_dec(hb2);
1637                         put_futex_key(&key2);
1638                         put_futex_key(&key1);
1639                         cond_resched();
1640                         goto retry;
1641                 default:
1642                         goto out_unlock;
1643                 }
1644         }
1645
1646         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1647                 if (task_count - nr_wake >= nr_requeue)
1648                         break;
1649
1650                 if (!match_futex(&this->key, &key1))
1651                         continue;
1652
1653                 /*
1654                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1655                  * be paired with each other and no other futex ops.
1656                  *
1657                  * We should never be requeueing a futex_q with a pi_state,
1658                  * which is awaiting a futex_unlock_pi().
1659                  */
1660                 if ((requeue_pi && !this->rt_waiter) ||
1661                     (!requeue_pi && this->rt_waiter) ||
1662                     this->pi_state) {
1663                         ret = -EINVAL;
1664                         break;
1665                 }
1666
1667                 /*
1668                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1669                  * lock, we already woke the top_waiter.  If not, it will be
1670                  * woken by futex_unlock_pi().
1671                  */
1672                 if (++task_count <= nr_wake && !requeue_pi) {
1673                         wake_futex(this);
1674                         continue;
1675                 }
1676
1677                 /* Ensure we requeue to the expected futex for requeue_pi. */
1678                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1679                         ret = -EINVAL;
1680                         break;
1681                 }
1682
1683                 /*
1684                  * Requeue nr_requeue waiters and possibly one more in the case
1685                  * of requeue_pi if we couldn't acquire the lock atomically.
1686                  */
1687                 if (requeue_pi) {
1688                         /* Prepare the waiter to take the rt_mutex. */
1689                         atomic_inc(&pi_state->refcount);
1690                         this->pi_state = pi_state;
1691                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1692                                                         this->rt_waiter,
1693                                                         this->task);
1694                         if (ret == 1) {
1695                                 /* We got the lock. */
1696                                 requeue_pi_wake_futex(this, &key2, hb2);
1697                                 drop_count++;
1698                                 continue;
1699                         } else if (ret) {
1700                                 /* -EDEADLK */
1701                                 this->pi_state = NULL;
1702                                 free_pi_state(pi_state);
1703                                 goto out_unlock;
1704                         }
1705                 }
1706                 requeue_futex(this, hb1, hb2, &key2);
1707                 drop_count++;
1708         }
1709
1710 out_unlock:
1711         double_unlock_hb(hb1, hb2);
1712         hb_waiters_dec(hb2);
1713
1714         /*
1715          * drop_futex_key_refs() must be called outside the spinlocks. During
1716          * the requeue we moved futex_q's from the hash bucket at key1 to the
1717          * one at key2 and updated their key pointer.  We no longer need to
1718          * hold the references to key1.
1719          */
1720         while (--drop_count >= 0)
1721                 drop_futex_key_refs(&key1);
1722
1723 out_put_keys:
1724         put_futex_key(&key2);
1725 out_put_key1:
1726         put_futex_key(&key1);
1727 out:
1728         if (pi_state != NULL)
1729                 free_pi_state(pi_state);
1730         return ret ? ret : task_count;
1731 }
1732
1733 /* The key must be already stored in q->key. */
1734 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1735         __acquires(&hb->lock)
1736 {
1737         struct futex_hash_bucket *hb;
1738
1739         hb = hash_futex(&q->key);
1740
1741         /*
1742          * Increment the counter before taking the lock so that
1743          * a potential waker won't miss a to-be-slept task that is
1744          * waiting for the spinlock. This is safe as all queue_lock()
1745          * users end up calling queue_me(). Similarly, for housekeeping,
1746          * decrement the counter at queue_unlock() when some error has
1747          * occurred and we don't end up adding the task to the list.
1748          */
1749         hb_waiters_inc(hb);
1750
1751         q->lock_ptr = &hb->lock;
1752
1753         spin_lock(&hb->lock); /* implies MB (A) */
1754         return hb;
1755 }
1756
1757 static inline void
1758 queue_unlock(struct futex_hash_bucket *hb)
1759         __releases(&hb->lock)
1760 {
1761         spin_unlock(&hb->lock);
1762         hb_waiters_dec(hb);
1763 }
1764
1765 /**
1766  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1767  * @q:  The futex_q to enqueue
1768  * @hb: The destination hash bucket
1769  *
1770  * The hb->lock must be held by the caller, and is released here. A call to
1771  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1772  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1773  * or nothing if the unqueue is done as part of the wake process and the unqueue
1774  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1775  * an example).
1776  */
1777 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1778         __releases(&hb->lock)
1779 {
1780         int prio;
1781
1782         /*
1783          * The priority used to register this element is
1784          * - either the real thread-priority for the real-time threads
1785          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1786          * - or MAX_RT_PRIO for non-RT threads.
1787          * Thus, all RT-threads are woken first in priority order, and
1788          * the others are woken last, in FIFO order.
1789          */
1790         prio = min(current->normal_prio, MAX_RT_PRIO);
1791
1792         plist_node_init(&q->list, prio);
1793         plist_add(&q->list, &hb->chain);
1794         q->task = current;
1795         spin_unlock(&hb->lock);
1796 }
1797
1798 /**
1799  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1800  * @q:  The futex_q to unqueue
1801  *
1802  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1803  * be paired with exactly one earlier call to queue_me().
1804  *
1805  * Return:
1806  *   1 - if the futex_q was still queued (and we removed unqueued it);
1807  *   0 - if the futex_q was already removed by the waking thread
1808  */
1809 static int unqueue_me(struct futex_q *q)
1810 {
1811         spinlock_t *lock_ptr;
1812         int ret = 0;
1813
1814         /* In the common case we don't take the spinlock, which is nice. */
1815 retry:
1816         lock_ptr = q->lock_ptr;
1817         barrier();
1818         if (lock_ptr != NULL) {
1819                 spin_lock(lock_ptr);
1820                 /*
1821                  * q->lock_ptr can change between reading it and
1822                  * spin_lock(), causing us to take the wrong lock.  This
1823                  * corrects the race condition.
1824                  *
1825                  * Reasoning goes like this: if we have the wrong lock,
1826                  * q->lock_ptr must have changed (maybe several times)
1827                  * between reading it and the spin_lock().  It can
1828                  * change again after the spin_lock() but only if it was
1829                  * already changed before the spin_lock().  It cannot,
1830                  * however, change back to the original value.  Therefore
1831                  * we can detect whether we acquired the correct lock.
1832                  */
1833                 if (unlikely(lock_ptr != q->lock_ptr)) {
1834                         spin_unlock(lock_ptr);
1835                         goto retry;
1836                 }
1837                 __unqueue_futex(q);
1838
1839                 BUG_ON(q->pi_state);
1840
1841                 spin_unlock(lock_ptr);
1842                 ret = 1;
1843         }
1844
1845         drop_futex_key_refs(&q->key);
1846         return ret;
1847 }
1848
1849 /*
1850  * PI futexes can not be requeued and must remove themself from the
1851  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1852  * and dropped here.
1853  */
1854 static void unqueue_me_pi(struct futex_q *q)
1855         __releases(q->lock_ptr)
1856 {
1857         __unqueue_futex(q);
1858
1859         BUG_ON(!q->pi_state);
1860         free_pi_state(q->pi_state);
1861         q->pi_state = NULL;
1862
1863         spin_unlock(q->lock_ptr);
1864 }
1865
1866 /*
1867  * Fixup the pi_state owner with the new owner.
1868  *
1869  * Must be called with hash bucket lock held and mm->sem held for non
1870  * private futexes.
1871  */
1872 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1873                                 struct task_struct *newowner)
1874 {
1875         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1876         struct futex_pi_state *pi_state = q->pi_state;
1877         struct task_struct *oldowner = pi_state->owner;
1878         u32 uval, uninitialized_var(curval), newval;
1879         int ret;
1880
1881         /* Owner died? */
1882         if (!pi_state->owner)
1883                 newtid |= FUTEX_OWNER_DIED;
1884
1885         /*
1886          * We are here either because we stole the rtmutex from the
1887          * previous highest priority waiter or we are the highest priority
1888          * waiter but failed to get the rtmutex the first time.
1889          * We have to replace the newowner TID in the user space variable.
1890          * This must be atomic as we have to preserve the owner died bit here.
1891          *
1892          * Note: We write the user space value _before_ changing the pi_state
1893          * because we can fault here. Imagine swapped out pages or a fork
1894          * that marked all the anonymous memory readonly for cow.
1895          *
1896          * Modifying pi_state _before_ the user space value would
1897          * leave the pi_state in an inconsistent state when we fault
1898          * here, because we need to drop the hash bucket lock to
1899          * handle the fault. This might be observed in the PID check
1900          * in lookup_pi_state.
1901          */
1902 retry:
1903         if (get_futex_value_locked(&uval, uaddr))
1904                 goto handle_fault;
1905
1906         while (1) {
1907                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1908
1909                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1910                         goto handle_fault;
1911                 if (curval == uval)
1912                         break;
1913                 uval = curval;
1914         }
1915
1916         /*
1917          * We fixed up user space. Now we need to fix the pi_state
1918          * itself.
1919          */
1920         if (pi_state->owner != NULL) {
1921                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1922                 WARN_ON(list_empty(&pi_state->list));
1923                 list_del_init(&pi_state->list);
1924                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1925         }
1926
1927         pi_state->owner = newowner;
1928
1929         raw_spin_lock_irq(&newowner->pi_lock);
1930         WARN_ON(!list_empty(&pi_state->list));
1931         list_add(&pi_state->list, &newowner->pi_state_list);
1932         raw_spin_unlock_irq(&newowner->pi_lock);
1933         return 0;
1934
1935         /*
1936          * To handle the page fault we need to drop the hash bucket
1937          * lock here. That gives the other task (either the highest priority
1938          * waiter itself or the task which stole the rtmutex) the
1939          * chance to try the fixup of the pi_state. So once we are
1940          * back from handling the fault we need to check the pi_state
1941          * after reacquiring the hash bucket lock and before trying to
1942          * do another fixup. When the fixup has been done already we
1943          * simply return.
1944          */
1945 handle_fault:
1946         spin_unlock(q->lock_ptr);
1947
1948         ret = fault_in_user_writeable(uaddr);
1949
1950         spin_lock(q->lock_ptr);
1951
1952         /*
1953          * Check if someone else fixed it for us:
1954          */
1955         if (pi_state->owner != oldowner)
1956                 return 0;
1957
1958         if (ret)
1959                 return ret;
1960
1961         goto retry;
1962 }
1963
1964 static long futex_wait_restart(struct restart_block *restart);
1965
1966 /**
1967  * fixup_owner() - Post lock pi_state and corner case management
1968  * @uaddr:      user address of the futex
1969  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1970  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1971  *
1972  * After attempting to lock an rt_mutex, this function is called to cleanup
1973  * the pi_state owner as well as handle race conditions that may allow us to
1974  * acquire the lock. Must be called with the hb lock held.
1975  *
1976  * Return:
1977  *  1 - success, lock taken;
1978  *  0 - success, lock not taken;
1979  * <0 - on error (-EFAULT)
1980  */
1981 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1982 {
1983         struct task_struct *owner;
1984         int ret = 0;
1985
1986         if (locked) {
1987                 /*
1988                  * Got the lock. We might not be the anticipated owner if we
1989                  * did a lock-steal - fix up the PI-state in that case:
1990                  */
1991                 if (q->pi_state->owner != current)
1992                         ret = fixup_pi_state_owner(uaddr, q, current);
1993                 goto out;
1994         }
1995
1996         /*
1997          * Catch the rare case, where the lock was released when we were on the
1998          * way back before we locked the hash bucket.
1999          */
2000         if (q->pi_state->owner == current) {
2001                 /*
2002                  * Try to get the rt_mutex now. This might fail as some other
2003                  * task acquired the rt_mutex after we removed ourself from the
2004                  * rt_mutex waiters list.
2005                  */
2006                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2007                         locked = 1;
2008                         goto out;
2009                 }
2010
2011                 /*
2012                  * pi_state is incorrect, some other task did a lock steal and
2013                  * we returned due to timeout or signal without taking the
2014                  * rt_mutex. Too late.
2015                  */
2016                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2017                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2018                 if (!owner)
2019                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2020                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2021                 ret = fixup_pi_state_owner(uaddr, q, owner);
2022                 goto out;
2023         }
2024
2025         /*
2026          * Paranoia check. If we did not take the lock, then we should not be
2027          * the owner of the rt_mutex.
2028          */
2029         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2030                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2031                                 "pi-state %p\n", ret,
2032                                 q->pi_state->pi_mutex.owner,
2033                                 q->pi_state->owner);
2034
2035 out:
2036         return ret ? ret : locked;
2037 }
2038
2039 /**
2040  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2041  * @hb:         the futex hash bucket, must be locked by the caller
2042  * @q:          the futex_q to queue up on
2043  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2044  */
2045 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2046                                 struct hrtimer_sleeper *timeout)
2047 {
2048         /*
2049          * The task state is guaranteed to be set before another task can
2050          * wake it. set_current_state() is implemented using set_mb() and
2051          * queue_me() calls spin_unlock() upon completion, both serializing
2052          * access to the hash list and forcing another memory barrier.
2053          */
2054         set_current_state(TASK_INTERRUPTIBLE);
2055         queue_me(q, hb);
2056
2057         /* Arm the timer */
2058         if (timeout) {
2059                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2060                 if (!hrtimer_active(&timeout->timer))
2061                         timeout->task = NULL;
2062         }
2063
2064         /*
2065          * If we have been removed from the hash list, then another task
2066          * has tried to wake us, and we can skip the call to schedule().
2067          */
2068         if (likely(!plist_node_empty(&q->list))) {
2069                 /*
2070                  * If the timer has already expired, current will already be
2071                  * flagged for rescheduling. Only call schedule if there
2072                  * is no timeout, or if it has yet to expire.
2073                  */
2074                 if (!timeout || timeout->task)
2075                         freezable_schedule();
2076         }
2077         __set_current_state(TASK_RUNNING);
2078 }
2079
2080 /**
2081  * futex_wait_setup() - Prepare to wait on a futex
2082  * @uaddr:      the futex userspace address
2083  * @val:        the expected value
2084  * @flags:      futex flags (FLAGS_SHARED, etc.)
2085  * @q:          the associated futex_q
2086  * @hb:         storage for hash_bucket pointer to be returned to caller
2087  *
2088  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2089  * compare it with the expected value.  Handle atomic faults internally.
2090  * Return with the hb lock held and a q.key reference on success, and unlocked
2091  * with no q.key reference on failure.
2092  *
2093  * Return:
2094  *  0 - uaddr contains val and hb has been locked;
2095  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2096  */
2097 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2098                            struct futex_q *q, struct futex_hash_bucket **hb)
2099 {
2100         u32 uval;
2101         int ret;
2102
2103         /*
2104          * Access the page AFTER the hash-bucket is locked.
2105          * Order is important:
2106          *
2107          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2108          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2109          *
2110          * The basic logical guarantee of a futex is that it blocks ONLY
2111          * if cond(var) is known to be true at the time of blocking, for
2112          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2113          * would open a race condition where we could block indefinitely with
2114          * cond(var) false, which would violate the guarantee.
2115          *
2116          * On the other hand, we insert q and release the hash-bucket only
2117          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2118          * absorb a wakeup if *uaddr does not match the desired values
2119          * while the syscall executes.
2120          */
2121 retry:
2122         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2123         if (unlikely(ret != 0))
2124                 return ret;
2125
2126 retry_private:
2127         *hb = queue_lock(q);
2128
2129         ret = get_futex_value_locked(&uval, uaddr);
2130
2131         if (ret) {
2132                 queue_unlock(*hb);
2133
2134                 ret = get_user(uval, uaddr);
2135                 if (ret)
2136                         goto out;
2137
2138                 if (!(flags & FLAGS_SHARED))
2139                         goto retry_private;
2140
2141                 put_futex_key(&q->key);
2142                 goto retry;
2143         }
2144
2145         if (uval != val) {
2146                 queue_unlock(*hb);
2147                 ret = -EWOULDBLOCK;
2148         }
2149
2150 out:
2151         if (ret)
2152                 put_futex_key(&q->key);
2153         return ret;
2154 }
2155
2156 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2157                       ktime_t *abs_time, u32 bitset)
2158 {
2159         struct hrtimer_sleeper timeout, *to = NULL;
2160         struct restart_block *restart;
2161         struct futex_hash_bucket *hb;
2162         struct futex_q q = futex_q_init;
2163         int ret;
2164
2165         if (!bitset)
2166                 return -EINVAL;
2167         q.bitset = bitset;
2168
2169         if (abs_time) {
2170                 to = &timeout;
2171
2172                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2173                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2174                                       HRTIMER_MODE_ABS);
2175                 hrtimer_init_sleeper(to, current);
2176                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2177                                              current->timer_slack_ns);
2178         }
2179
2180 retry:
2181         /*
2182          * Prepare to wait on uaddr. On success, holds hb lock and increments
2183          * q.key refs.
2184          */
2185         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2186         if (ret)
2187                 goto out;
2188
2189         /* queue_me and wait for wakeup, timeout, or a signal. */
2190         futex_wait_queue_me(hb, &q, to);
2191
2192         /* If we were woken (and unqueued), we succeeded, whatever. */
2193         ret = 0;
2194         /* unqueue_me() drops q.key ref */
2195         if (!unqueue_me(&q))
2196                 goto out;
2197         ret = -ETIMEDOUT;
2198         if (to && !to->task)
2199                 goto out;
2200
2201         /*
2202          * We expect signal_pending(current), but we might be the
2203          * victim of a spurious wakeup as well.
2204          */
2205         if (!signal_pending(current))
2206                 goto retry;
2207
2208         ret = -ERESTARTSYS;
2209         if (!abs_time)
2210                 goto out;
2211
2212         restart = &current_thread_info()->restart_block;
2213         restart->fn = futex_wait_restart;
2214         restart->futex.uaddr = uaddr;
2215         restart->futex.val = val;
2216         restart->futex.time = abs_time->tv64;
2217         restart->futex.bitset = bitset;
2218         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2219
2220         ret = -ERESTART_RESTARTBLOCK;
2221
2222 out:
2223         if (to) {
2224                 hrtimer_cancel(&to->timer);
2225                 destroy_hrtimer_on_stack(&to->timer);
2226         }
2227         return ret;
2228 }
2229
2230
2231 static long futex_wait_restart(struct restart_block *restart)
2232 {
2233         u32 __user *uaddr = restart->futex.uaddr;
2234         ktime_t t, *tp = NULL;
2235
2236         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2237                 t.tv64 = restart->futex.time;
2238                 tp = &t;
2239         }
2240         restart->fn = do_no_restart_syscall;
2241
2242         return (long)futex_wait(uaddr, restart->futex.flags,
2243                                 restart->futex.val, tp, restart->futex.bitset);
2244 }
2245
2246
2247 /*
2248  * Userspace tried a 0 -> TID atomic transition of the futex value
2249  * and failed. The kernel side here does the whole locking operation:
2250  * if there are waiters then it will block, it does PI, etc. (Due to
2251  * races the kernel might see a 0 value of the futex too.)
2252  */
2253 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2254                          ktime_t *time, int trylock)
2255 {
2256         struct hrtimer_sleeper timeout, *to = NULL;
2257         struct futex_hash_bucket *hb;
2258         struct futex_q q = futex_q_init;
2259         int res, ret;
2260
2261         if (refill_pi_state_cache())
2262                 return -ENOMEM;
2263
2264         if (time) {
2265                 to = &timeout;
2266                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2267                                       HRTIMER_MODE_ABS);
2268                 hrtimer_init_sleeper(to, current);
2269                 hrtimer_set_expires(&to->timer, *time);
2270         }
2271
2272 retry:
2273         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2274         if (unlikely(ret != 0))
2275                 goto out;
2276
2277 retry_private:
2278         hb = queue_lock(&q);
2279
2280         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2281         if (unlikely(ret)) {
2282                 switch (ret) {
2283                 case 1:
2284                         /* We got the lock. */
2285                         ret = 0;
2286                         goto out_unlock_put_key;
2287                 case -EFAULT:
2288                         goto uaddr_faulted;
2289                 case -EAGAIN:
2290                         /*
2291                          * Two reasons for this:
2292                          * - Task is exiting and we just wait for the
2293                          *   exit to complete.
2294                          * - The user space value changed.
2295                          */
2296                         queue_unlock(hb);
2297                         put_futex_key(&q.key);
2298                         cond_resched();
2299                         goto retry;
2300                 default:
2301                         goto out_unlock_put_key;
2302                 }
2303         }
2304
2305         /*
2306          * Only actually queue now that the atomic ops are done:
2307          */
2308         queue_me(&q, hb);
2309
2310         WARN_ON(!q.pi_state);
2311         /*
2312          * Block on the PI mutex:
2313          */
2314         if (!trylock) {
2315                 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2316         } else {
2317                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2318                 /* Fixup the trylock return value: */
2319                 ret = ret ? 0 : -EWOULDBLOCK;
2320         }
2321
2322         spin_lock(q.lock_ptr);
2323         /*
2324          * Fixup the pi_state owner and possibly acquire the lock if we
2325          * haven't already.
2326          */
2327         res = fixup_owner(uaddr, &q, !ret);
2328         /*
2329          * If fixup_owner() returned an error, proprogate that.  If it acquired
2330          * the lock, clear our -ETIMEDOUT or -EINTR.
2331          */
2332         if (res)
2333                 ret = (res < 0) ? res : 0;
2334
2335         /*
2336          * If fixup_owner() faulted and was unable to handle the fault, unlock
2337          * it and return the fault to userspace.
2338          */
2339         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2340                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2341
2342         /* Unqueue and drop the lock */
2343         unqueue_me_pi(&q);
2344
2345         goto out_put_key;
2346
2347 out_unlock_put_key:
2348         queue_unlock(hb);
2349
2350 out_put_key:
2351         put_futex_key(&q.key);
2352 out:
2353         if (to)
2354                 destroy_hrtimer_on_stack(&to->timer);
2355         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2356
2357 uaddr_faulted:
2358         queue_unlock(hb);
2359
2360         ret = fault_in_user_writeable(uaddr);
2361         if (ret)
2362                 goto out_put_key;
2363
2364         if (!(flags & FLAGS_SHARED))
2365                 goto retry_private;
2366
2367         put_futex_key(&q.key);
2368         goto retry;
2369 }
2370
2371 /*
2372  * Userspace attempted a TID -> 0 atomic transition, and failed.
2373  * This is the in-kernel slowpath: we look up the PI state (if any),
2374  * and do the rt-mutex unlock.
2375  */
2376 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2377 {
2378         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2379         union futex_key key = FUTEX_KEY_INIT;
2380         struct futex_hash_bucket *hb;
2381         struct futex_q *match;
2382         int ret;
2383
2384 retry:
2385         if (get_user(uval, uaddr))
2386                 return -EFAULT;
2387         /*
2388          * We release only a lock we actually own:
2389          */
2390         if ((uval & FUTEX_TID_MASK) != vpid)
2391                 return -EPERM;
2392
2393         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2394         if (ret)
2395                 return ret;
2396
2397         hb = hash_futex(&key);
2398         spin_lock(&hb->lock);
2399
2400         /*
2401          * Check waiters first. We do not trust user space values at
2402          * all and we at least want to know if user space fiddled
2403          * with the futex value instead of blindly unlocking.
2404          */
2405         match = futex_top_waiter(hb, &key);
2406         if (match) {
2407                 ret = wake_futex_pi(uaddr, uval, match);
2408                 /*
2409                  * The atomic access to the futex value generated a
2410                  * pagefault, so retry the user-access and the wakeup:
2411                  */
2412                 if (ret == -EFAULT)
2413                         goto pi_faulted;
2414                 goto out_unlock;
2415         }
2416
2417         /*
2418          * We have no kernel internal state, i.e. no waiters in the
2419          * kernel. Waiters which are about to queue themselves are stuck
2420          * on hb->lock. So we can safely ignore them. We do neither
2421          * preserve the WAITERS bit not the OWNER_DIED one. We are the
2422          * owner.
2423          */
2424         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2425                 goto pi_faulted;
2426
2427         /*
2428          * If uval has changed, let user space handle it.
2429          */
2430         ret = (curval == uval) ? 0 : -EAGAIN;
2431
2432 out_unlock:
2433         spin_unlock(&hb->lock);
2434         put_futex_key(&key);
2435         return ret;
2436
2437 pi_faulted:
2438         spin_unlock(&hb->lock);
2439         put_futex_key(&key);
2440
2441         ret = fault_in_user_writeable(uaddr);
2442         if (!ret)
2443                 goto retry;
2444
2445         return ret;
2446 }
2447
2448 /**
2449  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2450  * @hb:         the hash_bucket futex_q was original enqueued on
2451  * @q:          the futex_q woken while waiting to be requeued
2452  * @key2:       the futex_key of the requeue target futex
2453  * @timeout:    the timeout associated with the wait (NULL if none)
2454  *
2455  * Detect if the task was woken on the initial futex as opposed to the requeue
2456  * target futex.  If so, determine if it was a timeout or a signal that caused
2457  * the wakeup and return the appropriate error code to the caller.  Must be
2458  * called with the hb lock held.
2459  *
2460  * Return:
2461  *  0 = no early wakeup detected;
2462  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2463  */
2464 static inline
2465 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2466                                    struct futex_q *q, union futex_key *key2,
2467                                    struct hrtimer_sleeper *timeout)
2468 {
2469         int ret = 0;
2470
2471         /*
2472          * With the hb lock held, we avoid races while we process the wakeup.
2473          * We only need to hold hb (and not hb2) to ensure atomicity as the
2474          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2475          * It can't be requeued from uaddr2 to something else since we don't
2476          * support a PI aware source futex for requeue.
2477          */
2478         if (!match_futex(&q->key, key2)) {
2479                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2480                 /*
2481                  * We were woken prior to requeue by a timeout or a signal.
2482                  * Unqueue the futex_q and determine which it was.
2483                  */
2484                 plist_del(&q->list, &hb->chain);
2485                 hb_waiters_dec(hb);
2486
2487                 /* Handle spurious wakeups gracefully */
2488                 ret = -EWOULDBLOCK;
2489                 if (timeout && !timeout->task)
2490                         ret = -ETIMEDOUT;
2491                 else if (signal_pending(current))
2492                         ret = -ERESTARTNOINTR;
2493         }
2494         return ret;
2495 }
2496
2497 /**
2498  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2499  * @uaddr:      the futex we initially wait on (non-pi)
2500  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2501  *              the same type, no requeueing from private to shared, etc.
2502  * @val:        the expected value of uaddr
2503  * @abs_time:   absolute timeout
2504  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2505  * @uaddr2:     the pi futex we will take prior to returning to user-space
2506  *
2507  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2508  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2509  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2510  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2511  * without one, the pi logic would not know which task to boost/deboost, if
2512  * there was a need to.
2513  *
2514  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2515  * via the following--
2516  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2517  * 2) wakeup on uaddr2 after a requeue
2518  * 3) signal
2519  * 4) timeout
2520  *
2521  * If 3, cleanup and return -ERESTARTNOINTR.
2522  *
2523  * If 2, we may then block on trying to take the rt_mutex and return via:
2524  * 5) successful lock
2525  * 6) signal
2526  * 7) timeout
2527  * 8) other lock acquisition failure
2528  *
2529  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2530  *
2531  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2532  *
2533  * Return:
2534  *  0 - On success;
2535  * <0 - On error
2536  */
2537 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2538                                  u32 val, ktime_t *abs_time, u32 bitset,
2539                                  u32 __user *uaddr2)
2540 {
2541         struct hrtimer_sleeper timeout, *to = NULL;
2542         struct rt_mutex_waiter rt_waiter;
2543         struct rt_mutex *pi_mutex = NULL;
2544         struct futex_hash_bucket *hb;
2545         union futex_key key2 = FUTEX_KEY_INIT;
2546         struct futex_q q = futex_q_init;
2547         int res, ret;
2548
2549         if (uaddr == uaddr2)
2550                 return -EINVAL;
2551
2552         if (!bitset)
2553                 return -EINVAL;
2554
2555         if (abs_time) {
2556                 to = &timeout;
2557                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2558                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2559                                       HRTIMER_MODE_ABS);
2560                 hrtimer_init_sleeper(to, current);
2561                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2562                                              current->timer_slack_ns);
2563         }
2564
2565         /*
2566          * The waiter is allocated on our stack, manipulated by the requeue
2567          * code while we sleep on uaddr.
2568          */
2569         debug_rt_mutex_init_waiter(&rt_waiter);
2570         RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2571         RB_CLEAR_NODE(&rt_waiter.tree_entry);
2572         rt_waiter.task = NULL;
2573
2574         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2575         if (unlikely(ret != 0))
2576                 goto out;
2577
2578         q.bitset = bitset;
2579         q.rt_waiter = &rt_waiter;
2580         q.requeue_pi_key = &key2;
2581
2582         /*
2583          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2584          * count.
2585          */
2586         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2587         if (ret)
2588                 goto out_key2;
2589
2590         /*
2591          * The check above which compares uaddrs is not sufficient for
2592          * shared futexes. We need to compare the keys:
2593          */
2594         if (match_futex(&q.key, &key2)) {
2595                 ret = -EINVAL;
2596                 goto out_put_keys;
2597         }
2598
2599         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2600         futex_wait_queue_me(hb, &q, to);
2601
2602         spin_lock(&hb->lock);
2603         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2604         spin_unlock(&hb->lock);
2605         if (ret)
2606                 goto out_put_keys;
2607
2608         /*
2609          * In order for us to be here, we know our q.key == key2, and since
2610          * we took the hb->lock above, we also know that futex_requeue() has
2611          * completed and we no longer have to concern ourselves with a wakeup
2612          * race with the atomic proxy lock acquisition by the requeue code. The
2613          * futex_requeue dropped our key1 reference and incremented our key2
2614          * reference count.
2615          */
2616
2617         /* Check if the requeue code acquired the second futex for us. */
2618         if (!q.rt_waiter) {
2619                 /*
2620                  * Got the lock. We might not be the anticipated owner if we
2621                  * did a lock-steal - fix up the PI-state in that case.
2622                  */
2623                 if (q.pi_state && (q.pi_state->owner != current)) {
2624                         spin_lock(q.lock_ptr);
2625                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2626                         spin_unlock(q.lock_ptr);
2627                 }
2628         } else {
2629                 /*
2630                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2631                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2632                  * the pi_state.
2633                  */
2634                 WARN_ON(!q.pi_state);
2635                 pi_mutex = &q.pi_state->pi_mutex;
2636                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2637                 debug_rt_mutex_free_waiter(&rt_waiter);
2638
2639                 spin_lock(q.lock_ptr);
2640                 /*
2641                  * Fixup the pi_state owner and possibly acquire the lock if we
2642                  * haven't already.
2643                  */
2644                 res = fixup_owner(uaddr2, &q, !ret);
2645                 /*
2646                  * If fixup_owner() returned an error, proprogate that.  If it
2647                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2648                  */
2649                 if (res)
2650                         ret = (res < 0) ? res : 0;
2651
2652                 /* Unqueue and drop the lock. */
2653                 unqueue_me_pi(&q);
2654         }
2655
2656         /*
2657          * If fixup_pi_state_owner() faulted and was unable to handle the
2658          * fault, unlock the rt_mutex and return the fault to userspace.
2659          */
2660         if (ret == -EFAULT) {
2661                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2662                         rt_mutex_unlock(pi_mutex);
2663         } else if (ret == -EINTR) {
2664                 /*
2665                  * We've already been requeued, but cannot restart by calling
2666                  * futex_lock_pi() directly. We could restart this syscall, but
2667                  * it would detect that the user space "val" changed and return
2668                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2669                  * -EWOULDBLOCK directly.
2670                  */
2671                 ret = -EWOULDBLOCK;
2672         }
2673
2674 out_put_keys:
2675         put_futex_key(&q.key);
2676 out_key2:
2677         put_futex_key(&key2);
2678
2679 out:
2680         if (to) {
2681                 hrtimer_cancel(&to->timer);
2682                 destroy_hrtimer_on_stack(&to->timer);
2683         }
2684         return ret;
2685 }
2686
2687 /*
2688  * Support for robust futexes: the kernel cleans up held futexes at
2689  * thread exit time.
2690  *
2691  * Implementation: user-space maintains a per-thread list of locks it
2692  * is holding. Upon do_exit(), the kernel carefully walks this list,
2693  * and marks all locks that are owned by this thread with the
2694  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2695  * always manipulated with the lock held, so the list is private and
2696  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2697  * field, to allow the kernel to clean up if the thread dies after
2698  * acquiring the lock, but just before it could have added itself to
2699  * the list. There can only be one such pending lock.
2700  */
2701
2702 /**
2703  * sys_set_robust_list() - Set the robust-futex list head of a task
2704  * @head:       pointer to the list-head
2705  * @len:        length of the list-head, as userspace expects
2706  */
2707 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2708                 size_t, len)
2709 {
2710         if (!futex_cmpxchg_enabled)
2711                 return -ENOSYS;
2712         /*
2713          * The kernel knows only one size for now:
2714          */
2715         if (unlikely(len != sizeof(*head)))
2716                 return -EINVAL;
2717
2718         current->robust_list = head;
2719
2720         return 0;
2721 }
2722
2723 /**
2724  * sys_get_robust_list() - Get the robust-futex list head of a task
2725  * @pid:        pid of the process [zero for current task]
2726  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2727  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2728  */
2729 SYSCALL_DEFINE3(get_robust_list, int, pid,
2730                 struct robust_list_head __user * __user *, head_ptr,
2731                 size_t __user *, len_ptr)
2732 {
2733         struct robust_list_head __user *head;
2734         unsigned long ret;
2735         struct task_struct *p;
2736
2737         if (!futex_cmpxchg_enabled)
2738                 return -ENOSYS;
2739
2740         rcu_read_lock();
2741
2742         ret = -ESRCH;
2743         if (!pid)
2744                 p = current;
2745         else {
2746                 p = find_task_by_vpid(pid);
2747                 if (!p)
2748                         goto err_unlock;
2749         }
2750
2751         ret = -EPERM;
2752         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2753                 goto err_unlock;
2754
2755         head = p->robust_list;
2756         rcu_read_unlock();
2757
2758         if (put_user(sizeof(*head), len_ptr))
2759                 return -EFAULT;
2760         return put_user(head, head_ptr);
2761
2762 err_unlock:
2763         rcu_read_unlock();
2764
2765         return ret;
2766 }
2767
2768 /*
2769  * Process a futex-list entry, check whether it's owned by the
2770  * dying task, and do notification if so:
2771  */
2772 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2773 {
2774         u32 uval, uninitialized_var(nval), mval;
2775
2776 retry:
2777         if (get_user(uval, uaddr))
2778                 return -1;
2779
2780         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2781                 /*
2782                  * Ok, this dying thread is truly holding a futex
2783                  * of interest. Set the OWNER_DIED bit atomically
2784                  * via cmpxchg, and if the value had FUTEX_WAITERS
2785                  * set, wake up a waiter (if any). (We have to do a
2786                  * futex_wake() even if OWNER_DIED is already set -
2787                  * to handle the rare but possible case of recursive
2788                  * thread-death.) The rest of the cleanup is done in
2789                  * userspace.
2790                  */
2791                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2792                 /*
2793                  * We are not holding a lock here, but we want to have
2794                  * the pagefault_disable/enable() protection because
2795                  * we want to handle the fault gracefully. If the
2796                  * access fails we try to fault in the futex with R/W
2797                  * verification via get_user_pages. get_user() above
2798                  * does not guarantee R/W access. If that fails we
2799                  * give up and leave the futex locked.
2800                  */
2801                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2802                         if (fault_in_user_writeable(uaddr))
2803                                 return -1;
2804                         goto retry;
2805                 }
2806                 if (nval != uval)
2807                         goto retry;
2808
2809                 /*
2810                  * Wake robust non-PI futexes here. The wakeup of
2811                  * PI futexes happens in exit_pi_state():
2812                  */
2813                 if (!pi && (uval & FUTEX_WAITERS))
2814                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2815         }
2816         return 0;
2817 }
2818
2819 /*
2820  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2821  */
2822 static inline int fetch_robust_entry(struct robust_list __user **entry,
2823                                      struct robust_list __user * __user *head,
2824                                      unsigned int *pi)
2825 {
2826         unsigned long uentry;
2827
2828         if (get_user(uentry, (unsigned long __user *)head))
2829                 return -EFAULT;
2830
2831         *entry = (void __user *)(uentry & ~1UL);
2832         *pi = uentry & 1;
2833
2834         return 0;
2835 }
2836
2837 /*
2838  * Walk curr->robust_list (very carefully, it's a userspace list!)
2839  * and mark any locks found there dead, and notify any waiters.
2840  *
2841  * We silently return on any sign of list-walking problem.
2842  */
2843 void exit_robust_list(struct task_struct *curr)
2844 {
2845         struct robust_list_head __user *head = curr->robust_list;
2846         struct robust_list __user *entry, *next_entry, *pending;
2847         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2848         unsigned int uninitialized_var(next_pi);
2849         unsigned long futex_offset;
2850         int rc;
2851
2852         if (!futex_cmpxchg_enabled)
2853                 return;
2854
2855         /*
2856          * Fetch the list head (which was registered earlier, via
2857          * sys_set_robust_list()):
2858          */
2859         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2860                 return;
2861         /*
2862          * Fetch the relative futex offset:
2863          */
2864         if (get_user(futex_offset, &head->futex_offset))
2865                 return;
2866         /*
2867          * Fetch any possibly pending lock-add first, and handle it
2868          * if it exists:
2869          */
2870         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2871                 return;
2872
2873         next_entry = NULL;      /* avoid warning with gcc */
2874         while (entry != &head->list) {
2875                 /*
2876                  * Fetch the next entry in the list before calling
2877                  * handle_futex_death:
2878                  */
2879                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2880                 /*
2881                  * A pending lock might already be on the list, so
2882                  * don't process it twice:
2883                  */
2884                 if (entry != pending)
2885                         if (handle_futex_death((void __user *)entry + futex_offset,
2886                                                 curr, pi))
2887                                 return;
2888                 if (rc)
2889                         return;
2890                 entry = next_entry;
2891                 pi = next_pi;
2892                 /*
2893                  * Avoid excessively long or circular lists:
2894                  */
2895                 if (!--limit)
2896                         break;
2897
2898                 cond_resched();
2899         }
2900
2901         if (pending)
2902                 handle_futex_death((void __user *)pending + futex_offset,
2903                                    curr, pip);
2904 }
2905
2906 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2907                 u32 __user *uaddr2, u32 val2, u32 val3)
2908 {
2909         int cmd = op & FUTEX_CMD_MASK;
2910         unsigned int flags = 0;
2911
2912         if (!(op & FUTEX_PRIVATE_FLAG))
2913                 flags |= FLAGS_SHARED;
2914
2915         if (op & FUTEX_CLOCK_REALTIME) {
2916                 flags |= FLAGS_CLOCKRT;
2917                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2918                         return -ENOSYS;
2919         }
2920
2921         switch (cmd) {
2922         case FUTEX_LOCK_PI:
2923         case FUTEX_UNLOCK_PI:
2924         case FUTEX_TRYLOCK_PI:
2925         case FUTEX_WAIT_REQUEUE_PI:
2926         case FUTEX_CMP_REQUEUE_PI:
2927                 if (!futex_cmpxchg_enabled)
2928                         return -ENOSYS;
2929         }
2930
2931         switch (cmd) {
2932         case FUTEX_WAIT:
2933                 val3 = FUTEX_BITSET_MATCH_ANY;
2934         case FUTEX_WAIT_BITSET:
2935                 return futex_wait(uaddr, flags, val, timeout, val3);
2936         case FUTEX_WAKE:
2937                 val3 = FUTEX_BITSET_MATCH_ANY;
2938         case FUTEX_WAKE_BITSET:
2939                 return futex_wake(uaddr, flags, val, val3);
2940         case FUTEX_REQUEUE:
2941                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2942         case FUTEX_CMP_REQUEUE:
2943                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2944         case FUTEX_WAKE_OP:
2945                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2946         case FUTEX_LOCK_PI:
2947                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2948         case FUTEX_UNLOCK_PI:
2949                 return futex_unlock_pi(uaddr, flags);
2950         case FUTEX_TRYLOCK_PI:
2951                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2952         case FUTEX_WAIT_REQUEUE_PI:
2953                 val3 = FUTEX_BITSET_MATCH_ANY;
2954                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2955                                              uaddr2);
2956         case FUTEX_CMP_REQUEUE_PI:
2957                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2958         }
2959         return -ENOSYS;
2960 }
2961
2962
2963 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2964                 struct timespec __user *, utime, u32 __user *, uaddr2,
2965                 u32, val3)
2966 {
2967         struct timespec ts;
2968         ktime_t t, *tp = NULL;
2969         u32 val2 = 0;
2970         int cmd = op & FUTEX_CMD_MASK;
2971
2972         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2973                       cmd == FUTEX_WAIT_BITSET ||
2974                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
2975                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2976                         return -EFAULT;
2977                 if (!timespec_valid(&ts))
2978                         return -EINVAL;
2979
2980                 t = timespec_to_ktime(ts);
2981                 if (cmd == FUTEX_WAIT)
2982                         t = ktime_add_safe(ktime_get(), t);
2983                 tp = &t;
2984         }
2985         /*
2986          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2987          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2988          */
2989         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2990             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2991                 val2 = (u32) (unsigned long) utime;
2992
2993         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2994 }
2995
2996 static void __init futex_detect_cmpxchg(void)
2997 {
2998 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2999         u32 curval;
3000
3001         /*
3002          * This will fail and we want it. Some arch implementations do
3003          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3004          * functionality. We want to know that before we call in any
3005          * of the complex code paths. Also we want to prevent
3006          * registration of robust lists in that case. NULL is
3007          * guaranteed to fault and we get -EFAULT on functional
3008          * implementation, the non-functional ones will return
3009          * -ENOSYS.
3010          */
3011         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3012                 futex_cmpxchg_enabled = 1;
3013 #endif
3014 }
3015
3016 static int __init futex_init(void)
3017 {
3018         unsigned int futex_shift;
3019         unsigned long i;
3020
3021 #if CONFIG_BASE_SMALL
3022         futex_hashsize = 16;
3023 #else
3024         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3025 #endif
3026
3027         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3028                                                futex_hashsize, 0,
3029                                                futex_hashsize < 256 ? HASH_SMALL : 0,
3030                                                &futex_shift, NULL,
3031                                                futex_hashsize, futex_hashsize);
3032         futex_hashsize = 1UL << futex_shift;
3033
3034         futex_detect_cmpxchg();
3035
3036         for (i = 0; i < futex_hashsize; i++) {
3037                 atomic_set(&futex_queues[i].waiters, 0);
3038                 plist_head_init(&futex_queues[i].chain);
3039                 spin_lock_init(&futex_queues[i].lock);
3040         }
3041
3042         return 0;
3043 }
3044 __initcall(futex_init);