]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - kernel/futex.c
Merge remote-tracking branches 'spi/fix/pxa2xx', 'spi/fix/qup' and 'spi/fix/sh-sci...
[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 static int
796 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
797                 union futex_key *key, struct futex_pi_state **ps)
798 {
799         struct futex_pi_state *pi_state = NULL;
800         struct futex_q *this, *next;
801         struct task_struct *p;
802         pid_t pid = uval & FUTEX_TID_MASK;
803
804         plist_for_each_entry_safe(this, next, &hb->chain, list) {
805                 if (match_futex(&this->key, key)) {
806                         /*
807                          * Sanity check the waiter before increasing
808                          * the refcount and attaching to it.
809                          */
810                         pi_state = this->pi_state;
811                         /*
812                          * Userspace might have messed up non-PI and
813                          * PI futexes [3]
814                          */
815                         if (unlikely(!pi_state))
816                                 return -EINVAL;
817
818                         WARN_ON(!atomic_read(&pi_state->refcount));
819
820                         /*
821                          * Handle the owner died case:
822                          */
823                         if (uval & FUTEX_OWNER_DIED) {
824                                 /*
825                                  * exit_pi_state_list sets owner to NULL and
826                                  * wakes the topmost waiter. The task which
827                                  * acquires the pi_state->rt_mutex will fixup
828                                  * owner.
829                                  */
830                                 if (!pi_state->owner) {
831                                         /*
832                                          * No pi state owner, but the user
833                                          * space TID is not 0. Inconsistent
834                                          * state. [5]
835                                          */
836                                         if (pid)
837                                                 return -EINVAL;
838                                         /*
839                                          * Take a ref on the state and
840                                          * return. [4]
841                                          */
842                                         goto out_state;
843                                 }
844
845                                 /*
846                                  * If TID is 0, then either the dying owner
847                                  * has not yet executed exit_pi_state_list()
848                                  * or some waiter acquired the rtmutex in the
849                                  * pi state, but did not yet fixup the TID in
850                                  * user space.
851                                  *
852                                  * Take a ref on the state and return. [6]
853                                  */
854                                 if (!pid)
855                                         goto out_state;
856                         } else {
857                                 /*
858                                  * If the owner died bit is not set,
859                                  * then the pi_state must have an
860                                  * owner. [7]
861                                  */
862                                 if (!pi_state->owner)
863                                         return -EINVAL;
864                         }
865
866                         /*
867                          * Bail out if user space manipulated the
868                          * futex value. If pi state exists then the
869                          * owner TID must be the same as the user
870                          * space TID. [9/10]
871                          */
872                         if (pid != task_pid_vnr(pi_state->owner))
873                                 return -EINVAL;
874
875                 out_state:
876                         atomic_inc(&pi_state->refcount);
877                         *ps = pi_state;
878                         return 0;
879                 }
880         }
881
882         /*
883          * We are the first waiter - try to look up the real owner and attach
884          * the new pi_state to it, but bail out when TID = 0 [1]
885          */
886         if (!pid)
887                 return -ESRCH;
888         p = futex_find_get_task(pid);
889         if (!p)
890                 return -ESRCH;
891
892         if (!p->mm) {
893                 put_task_struct(p);
894                 return -EPERM;
895         }
896
897         /*
898          * We need to look at the task state flags to figure out,
899          * whether the task is exiting. To protect against the do_exit
900          * change of the task flags, we do this protected by
901          * p->pi_lock:
902          */
903         raw_spin_lock_irq(&p->pi_lock);
904         if (unlikely(p->flags & PF_EXITING)) {
905                 /*
906                  * The task is on the way out. When PF_EXITPIDONE is
907                  * set, we know that the task has finished the
908                  * cleanup:
909                  */
910                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
911
912                 raw_spin_unlock_irq(&p->pi_lock);
913                 put_task_struct(p);
914                 return ret;
915         }
916
917         /*
918          * No existing pi state. First waiter. [2]
919          */
920         pi_state = alloc_pi_state();
921
922         /*
923          * Initialize the pi_mutex in locked state and make 'p'
924          * the owner of it:
925          */
926         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
927
928         /* Store the key for possible exit cleanups: */
929         pi_state->key = *key;
930
931         WARN_ON(!list_empty(&pi_state->list));
932         list_add(&pi_state->list, &p->pi_state_list);
933         pi_state->owner = p;
934         raw_spin_unlock_irq(&p->pi_lock);
935
936         put_task_struct(p);
937
938         *ps = pi_state;
939
940         return 0;
941 }
942
943 /**
944  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
945  * @uaddr:              the pi futex user address
946  * @hb:                 the pi futex hash bucket
947  * @key:                the futex key associated with uaddr and hb
948  * @ps:                 the pi_state pointer where we store the result of the
949  *                      lookup
950  * @task:               the task to perform the atomic lock work for.  This will
951  *                      be "current" except in the case of requeue pi.
952  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
953  *
954  * Return:
955  *  0 - ready to wait;
956  *  1 - acquired the lock;
957  * <0 - error
958  *
959  * The hb->lock and futex_key refs shall be held by the caller.
960  */
961 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
962                                 union futex_key *key,
963                                 struct futex_pi_state **ps,
964                                 struct task_struct *task, int set_waiters)
965 {
966         int lock_taken, ret, force_take = 0;
967         u32 uval, newval, curval, vpid = task_pid_vnr(task);
968
969 retry:
970         ret = lock_taken = 0;
971
972         /*
973          * To avoid races, we attempt to take the lock here again
974          * (by doing a 0 -> TID atomic cmpxchg), while holding all
975          * the locks. It will most likely not succeed.
976          */
977         newval = vpid;
978         if (set_waiters)
979                 newval |= FUTEX_WAITERS;
980
981         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
982                 return -EFAULT;
983
984         /*
985          * Detect deadlocks.
986          */
987         if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
988                 return -EDEADLK;
989
990         /*
991          * Surprise - we got the lock, but we do not trust user space at all.
992          */
993         if (unlikely(!curval)) {
994                 /*
995                  * We verify whether there is kernel state for this
996                  * futex. If not, we can safely assume, that the 0 ->
997                  * TID transition is correct. If state exists, we do
998                  * not bother to fixup the user space state as it was
999                  * corrupted already.
1000                  */
1001                 return futex_top_waiter(hb, key) ? -EINVAL : 1;
1002         }
1003
1004         uval = curval;
1005
1006         /*
1007          * Set the FUTEX_WAITERS flag, so the owner will know it has someone
1008          * to wake at the next unlock.
1009          */
1010         newval = curval | FUTEX_WAITERS;
1011
1012         /*
1013          * Should we force take the futex? See below.
1014          */
1015         if (unlikely(force_take)) {
1016                 /*
1017                  * Keep the OWNER_DIED and the WAITERS bit and set the
1018                  * new TID value.
1019                  */
1020                 newval = (curval & ~FUTEX_TID_MASK) | vpid;
1021                 force_take = 0;
1022                 lock_taken = 1;
1023         }
1024
1025         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1026                 return -EFAULT;
1027         if (unlikely(curval != uval))
1028                 goto retry;
1029
1030         /*
1031          * We took the lock due to forced take over.
1032          */
1033         if (unlikely(lock_taken))
1034                 return 1;
1035
1036         /*
1037          * We dont have the lock. Look up the PI state (or create it if
1038          * we are the first waiter):
1039          */
1040         ret = lookup_pi_state(uval, hb, key, ps);
1041
1042         if (unlikely(ret)) {
1043                 switch (ret) {
1044                 case -ESRCH:
1045                         /*
1046                          * We failed to find an owner for this
1047                          * futex. So we have no pi_state to block
1048                          * on. This can happen in two cases:
1049                          *
1050                          * 1) The owner died
1051                          * 2) A stale FUTEX_WAITERS bit
1052                          *
1053                          * Re-read the futex value.
1054                          */
1055                         if (get_futex_value_locked(&curval, uaddr))
1056                                 return -EFAULT;
1057
1058                         /*
1059                          * If the owner died or we have a stale
1060                          * WAITERS bit the owner TID in the user space
1061                          * futex is 0.
1062                          */
1063                         if (!(curval & FUTEX_TID_MASK)) {
1064                                 force_take = 1;
1065                                 goto retry;
1066                         }
1067                 default:
1068                         break;
1069                 }
1070         }
1071
1072         return ret;
1073 }
1074
1075 /**
1076  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1077  * @q:  The futex_q to unqueue
1078  *
1079  * The q->lock_ptr must not be NULL and must be held by the caller.
1080  */
1081 static void __unqueue_futex(struct futex_q *q)
1082 {
1083         struct futex_hash_bucket *hb;
1084
1085         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1086             || WARN_ON(plist_node_empty(&q->list)))
1087                 return;
1088
1089         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1090         plist_del(&q->list, &hb->chain);
1091         hb_waiters_dec(hb);
1092 }
1093
1094 /*
1095  * The hash bucket lock must be held when this is called.
1096  * Afterwards, the futex_q must not be accessed.
1097  */
1098 static void wake_futex(struct futex_q *q)
1099 {
1100         struct task_struct *p = q->task;
1101
1102         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1103                 return;
1104
1105         /*
1106          * We set q->lock_ptr = NULL _before_ we wake up the task. If
1107          * a non-futex wake up happens on another CPU then the task
1108          * might exit and p would dereference a non-existing task
1109          * struct. Prevent this by holding a reference on p across the
1110          * wake up.
1111          */
1112         get_task_struct(p);
1113
1114         __unqueue_futex(q);
1115         /*
1116          * The waiting task can free the futex_q as soon as
1117          * q->lock_ptr = NULL is written, without taking any locks. A
1118          * memory barrier is required here to prevent the following
1119          * store to lock_ptr from getting ahead of the plist_del.
1120          */
1121         smp_wmb();
1122         q->lock_ptr = NULL;
1123
1124         wake_up_state(p, TASK_NORMAL);
1125         put_task_struct(p);
1126 }
1127
1128 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1129 {
1130         struct task_struct *new_owner;
1131         struct futex_pi_state *pi_state = this->pi_state;
1132         u32 uninitialized_var(curval), newval;
1133         int ret = 0;
1134
1135         if (!pi_state)
1136                 return -EINVAL;
1137
1138         /*
1139          * If current does not own the pi_state then the futex is
1140          * inconsistent and user space fiddled with the futex value.
1141          */
1142         if (pi_state->owner != current)
1143                 return -EINVAL;
1144
1145         raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1146         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1147
1148         /*
1149          * It is possible that the next waiter (the one that brought
1150          * this owner to the kernel) timed out and is no longer
1151          * waiting on the lock.
1152          */
1153         if (!new_owner)
1154                 new_owner = this->task;
1155
1156         /*
1157          * We pass it to the next owner. The WAITERS bit is always
1158          * kept enabled while there is PI state around. We cleanup the
1159          * owner died bit, because we are the owner.
1160          */
1161         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1162
1163         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1164                 ret = -EFAULT;
1165         else if (curval != uval)
1166                 ret = -EINVAL;
1167         if (ret) {
1168                 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1169                 return ret;
1170         }
1171
1172         raw_spin_lock_irq(&pi_state->owner->pi_lock);
1173         WARN_ON(list_empty(&pi_state->list));
1174         list_del_init(&pi_state->list);
1175         raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1176
1177         raw_spin_lock_irq(&new_owner->pi_lock);
1178         WARN_ON(!list_empty(&pi_state->list));
1179         list_add(&pi_state->list, &new_owner->pi_state_list);
1180         pi_state->owner = new_owner;
1181         raw_spin_unlock_irq(&new_owner->pi_lock);
1182
1183         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1184         rt_mutex_unlock(&pi_state->pi_mutex);
1185
1186         return 0;
1187 }
1188
1189 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1190 {
1191         u32 uninitialized_var(oldval);
1192
1193         /*
1194          * There is no waiter, so we unlock the futex. The owner died
1195          * bit has not to be preserved here. We are the owner:
1196          */
1197         if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1198                 return -EFAULT;
1199         if (oldval != uval)
1200                 return -EAGAIN;
1201
1202         return 0;
1203 }
1204
1205 /*
1206  * Express the locking dependencies for lockdep:
1207  */
1208 static inline void
1209 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1210 {
1211         if (hb1 <= hb2) {
1212                 spin_lock(&hb1->lock);
1213                 if (hb1 < hb2)
1214                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1215         } else { /* hb1 > hb2 */
1216                 spin_lock(&hb2->lock);
1217                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1218         }
1219 }
1220
1221 static inline void
1222 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1223 {
1224         spin_unlock(&hb1->lock);
1225         if (hb1 != hb2)
1226                 spin_unlock(&hb2->lock);
1227 }
1228
1229 /*
1230  * Wake up waiters matching bitset queued on this futex (uaddr).
1231  */
1232 static int
1233 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1234 {
1235         struct futex_hash_bucket *hb;
1236         struct futex_q *this, *next;
1237         union futex_key key = FUTEX_KEY_INIT;
1238         int ret;
1239
1240         if (!bitset)
1241                 return -EINVAL;
1242
1243         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1244         if (unlikely(ret != 0))
1245                 goto out;
1246
1247         hb = hash_futex(&key);
1248
1249         /* Make sure we really have tasks to wakeup */
1250         if (!hb_waiters_pending(hb))
1251                 goto out_put_key;
1252
1253         spin_lock(&hb->lock);
1254
1255         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1256                 if (match_futex (&this->key, &key)) {
1257                         if (this->pi_state || this->rt_waiter) {
1258                                 ret = -EINVAL;
1259                                 break;
1260                         }
1261
1262                         /* Check if one of the bits is set in both bitsets */
1263                         if (!(this->bitset & bitset))
1264                                 continue;
1265
1266                         wake_futex(this);
1267                         if (++ret >= nr_wake)
1268                                 break;
1269                 }
1270         }
1271
1272         spin_unlock(&hb->lock);
1273 out_put_key:
1274         put_futex_key(&key);
1275 out:
1276         return ret;
1277 }
1278
1279 /*
1280  * Wake up all waiters hashed on the physical page that is mapped
1281  * to this virtual address:
1282  */
1283 static int
1284 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1285               int nr_wake, int nr_wake2, int op)
1286 {
1287         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1288         struct futex_hash_bucket *hb1, *hb2;
1289         struct futex_q *this, *next;
1290         int ret, op_ret;
1291
1292 retry:
1293         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1294         if (unlikely(ret != 0))
1295                 goto out;
1296         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1297         if (unlikely(ret != 0))
1298                 goto out_put_key1;
1299
1300         hb1 = hash_futex(&key1);
1301         hb2 = hash_futex(&key2);
1302
1303 retry_private:
1304         double_lock_hb(hb1, hb2);
1305         op_ret = futex_atomic_op_inuser(op, uaddr2);
1306         if (unlikely(op_ret < 0)) {
1307
1308                 double_unlock_hb(hb1, hb2);
1309
1310 #ifndef CONFIG_MMU
1311                 /*
1312                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1313                  * but we might get them from range checking
1314                  */
1315                 ret = op_ret;
1316                 goto out_put_keys;
1317 #endif
1318
1319                 if (unlikely(op_ret != -EFAULT)) {
1320                         ret = op_ret;
1321                         goto out_put_keys;
1322                 }
1323
1324                 ret = fault_in_user_writeable(uaddr2);
1325                 if (ret)
1326                         goto out_put_keys;
1327
1328                 if (!(flags & FLAGS_SHARED))
1329                         goto retry_private;
1330
1331                 put_futex_key(&key2);
1332                 put_futex_key(&key1);
1333                 goto retry;
1334         }
1335
1336         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1337                 if (match_futex (&this->key, &key1)) {
1338                         if (this->pi_state || this->rt_waiter) {
1339                                 ret = -EINVAL;
1340                                 goto out_unlock;
1341                         }
1342                         wake_futex(this);
1343                         if (++ret >= nr_wake)
1344                                 break;
1345                 }
1346         }
1347
1348         if (op_ret > 0) {
1349                 op_ret = 0;
1350                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1351                         if (match_futex (&this->key, &key2)) {
1352                                 if (this->pi_state || this->rt_waiter) {
1353                                         ret = -EINVAL;
1354                                         goto out_unlock;
1355                                 }
1356                                 wake_futex(this);
1357                                 if (++op_ret >= nr_wake2)
1358                                         break;
1359                         }
1360                 }
1361                 ret += op_ret;
1362         }
1363
1364 out_unlock:
1365         double_unlock_hb(hb1, hb2);
1366 out_put_keys:
1367         put_futex_key(&key2);
1368 out_put_key1:
1369         put_futex_key(&key1);
1370 out:
1371         return ret;
1372 }
1373
1374 /**
1375  * requeue_futex() - Requeue a futex_q from one hb to another
1376  * @q:          the futex_q to requeue
1377  * @hb1:        the source hash_bucket
1378  * @hb2:        the target hash_bucket
1379  * @key2:       the new key for the requeued futex_q
1380  */
1381 static inline
1382 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1383                    struct futex_hash_bucket *hb2, union futex_key *key2)
1384 {
1385
1386         /*
1387          * If key1 and key2 hash to the same bucket, no need to
1388          * requeue.
1389          */
1390         if (likely(&hb1->chain != &hb2->chain)) {
1391                 plist_del(&q->list, &hb1->chain);
1392                 hb_waiters_dec(hb1);
1393                 plist_add(&q->list, &hb2->chain);
1394                 hb_waiters_inc(hb2);
1395                 q->lock_ptr = &hb2->lock;
1396         }
1397         get_futex_key_refs(key2);
1398         q->key = *key2;
1399 }
1400
1401 /**
1402  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1403  * @q:          the futex_q
1404  * @key:        the key of the requeue target futex
1405  * @hb:         the hash_bucket of the requeue target futex
1406  *
1407  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1408  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1409  * to the requeue target futex so the waiter can detect the wakeup on the right
1410  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1411  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1412  * to protect access to the pi_state to fixup the owner later.  Must be called
1413  * with both q->lock_ptr and hb->lock held.
1414  */
1415 static inline
1416 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1417                            struct futex_hash_bucket *hb)
1418 {
1419         get_futex_key_refs(key);
1420         q->key = *key;
1421
1422         __unqueue_futex(q);
1423
1424         WARN_ON(!q->rt_waiter);
1425         q->rt_waiter = NULL;
1426
1427         q->lock_ptr = &hb->lock;
1428
1429         wake_up_state(q->task, TASK_NORMAL);
1430 }
1431
1432 /**
1433  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1434  * @pifutex:            the user address of the to futex
1435  * @hb1:                the from futex hash bucket, must be locked by the caller
1436  * @hb2:                the to futex hash bucket, must be locked by the caller
1437  * @key1:               the from futex key
1438  * @key2:               the to futex key
1439  * @ps:                 address to store the pi_state pointer
1440  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1441  *
1442  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1443  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1444  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1445  * hb1 and hb2 must be held by the caller.
1446  *
1447  * Return:
1448  *  0 - failed to acquire the lock atomically;
1449  * >0 - acquired the lock, return value is vpid of the top_waiter
1450  * <0 - error
1451  */
1452 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1453                                  struct futex_hash_bucket *hb1,
1454                                  struct futex_hash_bucket *hb2,
1455                                  union futex_key *key1, union futex_key *key2,
1456                                  struct futex_pi_state **ps, int set_waiters)
1457 {
1458         struct futex_q *top_waiter = NULL;
1459         u32 curval;
1460         int ret, vpid;
1461
1462         if (get_futex_value_locked(&curval, pifutex))
1463                 return -EFAULT;
1464
1465         /*
1466          * Find the top_waiter and determine if there are additional waiters.
1467          * If the caller intends to requeue more than 1 waiter to pifutex,
1468          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1469          * as we have means to handle the possible fault.  If not, don't set
1470          * the bit unecessarily as it will force the subsequent unlock to enter
1471          * the kernel.
1472          */
1473         top_waiter = futex_top_waiter(hb1, key1);
1474
1475         /* There are no waiters, nothing for us to do. */
1476         if (!top_waiter)
1477                 return 0;
1478
1479         /* Ensure we requeue to the expected futex. */
1480         if (!match_futex(top_waiter->requeue_pi_key, key2))
1481                 return -EINVAL;
1482
1483         /*
1484          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1485          * the contended case or if set_waiters is 1.  The pi_state is returned
1486          * in ps in contended cases.
1487          */
1488         vpid = task_pid_vnr(top_waiter->task);
1489         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1490                                    set_waiters);
1491         if (ret == 1) {
1492                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1493                 return vpid;
1494         }
1495         return ret;
1496 }
1497
1498 /**
1499  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1500  * @uaddr1:     source futex user address
1501  * @flags:      futex flags (FLAGS_SHARED, etc.)
1502  * @uaddr2:     target futex user address
1503  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1504  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1505  * @cmpval:     @uaddr1 expected value (or %NULL)
1506  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1507  *              pi futex (pi to pi requeue is not supported)
1508  *
1509  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1510  * uaddr2 atomically on behalf of the top waiter.
1511  *
1512  * Return:
1513  * >=0 - on success, the number of tasks requeued or woken;
1514  *  <0 - on error
1515  */
1516 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1517                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1518                          u32 *cmpval, int requeue_pi)
1519 {
1520         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1521         int drop_count = 0, task_count = 0, ret;
1522         struct futex_pi_state *pi_state = NULL;
1523         struct futex_hash_bucket *hb1, *hb2;
1524         struct futex_q *this, *next;
1525
1526         if (requeue_pi) {
1527                 /*
1528                  * Requeue PI only works on two distinct uaddrs. This
1529                  * check is only valid for private futexes. See below.
1530                  */
1531                 if (uaddr1 == uaddr2)
1532                         return -EINVAL;
1533
1534                 /*
1535                  * requeue_pi requires a pi_state, try to allocate it now
1536                  * without any locks in case it fails.
1537                  */
1538                 if (refill_pi_state_cache())
1539                         return -ENOMEM;
1540                 /*
1541                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1542                  * + nr_requeue, since it acquires the rt_mutex prior to
1543                  * returning to userspace, so as to not leave the rt_mutex with
1544                  * waiters and no owner.  However, second and third wake-ups
1545                  * cannot be predicted as they involve race conditions with the
1546                  * first wake and a fault while looking up the pi_state.  Both
1547                  * pthread_cond_signal() and pthread_cond_broadcast() should
1548                  * use nr_wake=1.
1549                  */
1550                 if (nr_wake != 1)
1551                         return -EINVAL;
1552         }
1553
1554 retry:
1555         if (pi_state != NULL) {
1556                 /*
1557                  * We will have to lookup the pi_state again, so free this one
1558                  * to keep the accounting correct.
1559                  */
1560                 free_pi_state(pi_state);
1561                 pi_state = NULL;
1562         }
1563
1564         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1565         if (unlikely(ret != 0))
1566                 goto out;
1567         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1568                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1569         if (unlikely(ret != 0))
1570                 goto out_put_key1;
1571
1572         /*
1573          * The check above which compares uaddrs is not sufficient for
1574          * shared futexes. We need to compare the keys:
1575          */
1576         if (requeue_pi && match_futex(&key1, &key2)) {
1577                 ret = -EINVAL;
1578                 goto out_put_keys;
1579         }
1580
1581         hb1 = hash_futex(&key1);
1582         hb2 = hash_futex(&key2);
1583
1584 retry_private:
1585         hb_waiters_inc(hb2);
1586         double_lock_hb(hb1, hb2);
1587
1588         if (likely(cmpval != NULL)) {
1589                 u32 curval;
1590
1591                 ret = get_futex_value_locked(&curval, uaddr1);
1592
1593                 if (unlikely(ret)) {
1594                         double_unlock_hb(hb1, hb2);
1595                         hb_waiters_dec(hb2);
1596
1597                         ret = get_user(curval, uaddr1);
1598                         if (ret)
1599                                 goto out_put_keys;
1600
1601                         if (!(flags & FLAGS_SHARED))
1602                                 goto retry_private;
1603
1604                         put_futex_key(&key2);
1605                         put_futex_key(&key1);
1606                         goto retry;
1607                 }
1608                 if (curval != *cmpval) {
1609                         ret = -EAGAIN;
1610                         goto out_unlock;
1611                 }
1612         }
1613
1614         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1615                 /*
1616                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1617                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1618                  * bit.  We force this here where we are able to easily handle
1619                  * faults rather in the requeue loop below.
1620                  */
1621                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1622                                                  &key2, &pi_state, nr_requeue);
1623
1624                 /*
1625                  * At this point the top_waiter has either taken uaddr2 or is
1626                  * waiting on it.  If the former, then the pi_state will not
1627                  * exist yet, look it up one more time to ensure we have a
1628                  * reference to it. If the lock was taken, ret contains the
1629                  * vpid of the top waiter task.
1630                  */
1631                 if (ret > 0) {
1632                         WARN_ON(pi_state);
1633                         drop_count++;
1634                         task_count++;
1635                         /*
1636                          * If we acquired the lock, then the user
1637                          * space value of uaddr2 should be vpid. It
1638                          * cannot be changed by the top waiter as it
1639                          * is blocked on hb2 lock if it tries to do
1640                          * so. If something fiddled with it behind our
1641                          * back the pi state lookup might unearth
1642                          * it. So we rather use the known value than
1643                          * rereading and handing potential crap to
1644                          * lookup_pi_state.
1645                          */
1646                         ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1647                 }
1648
1649                 switch (ret) {
1650                 case 0:
1651                         break;
1652                 case -EFAULT:
1653                         double_unlock_hb(hb1, hb2);
1654                         hb_waiters_dec(hb2);
1655                         put_futex_key(&key2);
1656                         put_futex_key(&key1);
1657                         ret = fault_in_user_writeable(uaddr2);
1658                         if (!ret)
1659                                 goto retry;
1660                         goto out;
1661                 case -EAGAIN:
1662                         /* The owner was exiting, try again. */
1663                         double_unlock_hb(hb1, hb2);
1664                         hb_waiters_dec(hb2);
1665                         put_futex_key(&key2);
1666                         put_futex_key(&key1);
1667                         cond_resched();
1668                         goto retry;
1669                 default:
1670                         goto out_unlock;
1671                 }
1672         }
1673
1674         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1675                 if (task_count - nr_wake >= nr_requeue)
1676                         break;
1677
1678                 if (!match_futex(&this->key, &key1))
1679                         continue;
1680
1681                 /*
1682                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1683                  * be paired with each other and no other futex ops.
1684                  *
1685                  * We should never be requeueing a futex_q with a pi_state,
1686                  * which is awaiting a futex_unlock_pi().
1687                  */
1688                 if ((requeue_pi && !this->rt_waiter) ||
1689                     (!requeue_pi && this->rt_waiter) ||
1690                     this->pi_state) {
1691                         ret = -EINVAL;
1692                         break;
1693                 }
1694
1695                 /*
1696                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1697                  * lock, we already woke the top_waiter.  If not, it will be
1698                  * woken by futex_unlock_pi().
1699                  */
1700                 if (++task_count <= nr_wake && !requeue_pi) {
1701                         wake_futex(this);
1702                         continue;
1703                 }
1704
1705                 /* Ensure we requeue to the expected futex for requeue_pi. */
1706                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1707                         ret = -EINVAL;
1708                         break;
1709                 }
1710
1711                 /*
1712                  * Requeue nr_requeue waiters and possibly one more in the case
1713                  * of requeue_pi if we couldn't acquire the lock atomically.
1714                  */
1715                 if (requeue_pi) {
1716                         /* Prepare the waiter to take the rt_mutex. */
1717                         atomic_inc(&pi_state->refcount);
1718                         this->pi_state = pi_state;
1719                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1720                                                         this->rt_waiter,
1721                                                         this->task, 1);
1722                         if (ret == 1) {
1723                                 /* We got the lock. */
1724                                 requeue_pi_wake_futex(this, &key2, hb2);
1725                                 drop_count++;
1726                                 continue;
1727                         } else if (ret) {
1728                                 /* -EDEADLK */
1729                                 this->pi_state = NULL;
1730                                 free_pi_state(pi_state);
1731                                 goto out_unlock;
1732                         }
1733                 }
1734                 requeue_futex(this, hb1, hb2, &key2);
1735                 drop_count++;
1736         }
1737
1738 out_unlock:
1739         double_unlock_hb(hb1, hb2);
1740         hb_waiters_dec(hb2);
1741
1742         /*
1743          * drop_futex_key_refs() must be called outside the spinlocks. During
1744          * the requeue we moved futex_q's from the hash bucket at key1 to the
1745          * one at key2 and updated their key pointer.  We no longer need to
1746          * hold the references to key1.
1747          */
1748         while (--drop_count >= 0)
1749                 drop_futex_key_refs(&key1);
1750
1751 out_put_keys:
1752         put_futex_key(&key2);
1753 out_put_key1:
1754         put_futex_key(&key1);
1755 out:
1756         if (pi_state != NULL)
1757                 free_pi_state(pi_state);
1758         return ret ? ret : task_count;
1759 }
1760
1761 /* The key must be already stored in q->key. */
1762 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1763         __acquires(&hb->lock)
1764 {
1765         struct futex_hash_bucket *hb;
1766
1767         hb = hash_futex(&q->key);
1768
1769         /*
1770          * Increment the counter before taking the lock so that
1771          * a potential waker won't miss a to-be-slept task that is
1772          * waiting for the spinlock. This is safe as all queue_lock()
1773          * users end up calling queue_me(). Similarly, for housekeeping,
1774          * decrement the counter at queue_unlock() when some error has
1775          * occurred and we don't end up adding the task to the list.
1776          */
1777         hb_waiters_inc(hb);
1778
1779         q->lock_ptr = &hb->lock;
1780
1781         spin_lock(&hb->lock); /* implies MB (A) */
1782         return hb;
1783 }
1784
1785 static inline void
1786 queue_unlock(struct futex_hash_bucket *hb)
1787         __releases(&hb->lock)
1788 {
1789         spin_unlock(&hb->lock);
1790         hb_waiters_dec(hb);
1791 }
1792
1793 /**
1794  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1795  * @q:  The futex_q to enqueue
1796  * @hb: The destination hash bucket
1797  *
1798  * The hb->lock must be held by the caller, and is released here. A call to
1799  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1800  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1801  * or nothing if the unqueue is done as part of the wake process and the unqueue
1802  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1803  * an example).
1804  */
1805 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1806         __releases(&hb->lock)
1807 {
1808         int prio;
1809
1810         /*
1811          * The priority used to register this element is
1812          * - either the real thread-priority for the real-time threads
1813          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1814          * - or MAX_RT_PRIO for non-RT threads.
1815          * Thus, all RT-threads are woken first in priority order, and
1816          * the others are woken last, in FIFO order.
1817          */
1818         prio = min(current->normal_prio, MAX_RT_PRIO);
1819
1820         plist_node_init(&q->list, prio);
1821         plist_add(&q->list, &hb->chain);
1822         q->task = current;
1823         spin_unlock(&hb->lock);
1824 }
1825
1826 /**
1827  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1828  * @q:  The futex_q to unqueue
1829  *
1830  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1831  * be paired with exactly one earlier call to queue_me().
1832  *
1833  * Return:
1834  *   1 - if the futex_q was still queued (and we removed unqueued it);
1835  *   0 - if the futex_q was already removed by the waking thread
1836  */
1837 static int unqueue_me(struct futex_q *q)
1838 {
1839         spinlock_t *lock_ptr;
1840         int ret = 0;
1841
1842         /* In the common case we don't take the spinlock, which is nice. */
1843 retry:
1844         lock_ptr = q->lock_ptr;
1845         barrier();
1846         if (lock_ptr != NULL) {
1847                 spin_lock(lock_ptr);
1848                 /*
1849                  * q->lock_ptr can change between reading it and
1850                  * spin_lock(), causing us to take the wrong lock.  This
1851                  * corrects the race condition.
1852                  *
1853                  * Reasoning goes like this: if we have the wrong lock,
1854                  * q->lock_ptr must have changed (maybe several times)
1855                  * between reading it and the spin_lock().  It can
1856                  * change again after the spin_lock() but only if it was
1857                  * already changed before the spin_lock().  It cannot,
1858                  * however, change back to the original value.  Therefore
1859                  * we can detect whether we acquired the correct lock.
1860                  */
1861                 if (unlikely(lock_ptr != q->lock_ptr)) {
1862                         spin_unlock(lock_ptr);
1863                         goto retry;
1864                 }
1865                 __unqueue_futex(q);
1866
1867                 BUG_ON(q->pi_state);
1868
1869                 spin_unlock(lock_ptr);
1870                 ret = 1;
1871         }
1872
1873         drop_futex_key_refs(&q->key);
1874         return ret;
1875 }
1876
1877 /*
1878  * PI futexes can not be requeued and must remove themself from the
1879  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1880  * and dropped here.
1881  */
1882 static void unqueue_me_pi(struct futex_q *q)
1883         __releases(q->lock_ptr)
1884 {
1885         __unqueue_futex(q);
1886
1887         BUG_ON(!q->pi_state);
1888         free_pi_state(q->pi_state);
1889         q->pi_state = NULL;
1890
1891         spin_unlock(q->lock_ptr);
1892 }
1893
1894 /*
1895  * Fixup the pi_state owner with the new owner.
1896  *
1897  * Must be called with hash bucket lock held and mm->sem held for non
1898  * private futexes.
1899  */
1900 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1901                                 struct task_struct *newowner)
1902 {
1903         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1904         struct futex_pi_state *pi_state = q->pi_state;
1905         struct task_struct *oldowner = pi_state->owner;
1906         u32 uval, uninitialized_var(curval), newval;
1907         int ret;
1908
1909         /* Owner died? */
1910         if (!pi_state->owner)
1911                 newtid |= FUTEX_OWNER_DIED;
1912
1913         /*
1914          * We are here either because we stole the rtmutex from the
1915          * previous highest priority waiter or we are the highest priority
1916          * waiter but failed to get the rtmutex the first time.
1917          * We have to replace the newowner TID in the user space variable.
1918          * This must be atomic as we have to preserve the owner died bit here.
1919          *
1920          * Note: We write the user space value _before_ changing the pi_state
1921          * because we can fault here. Imagine swapped out pages or a fork
1922          * that marked all the anonymous memory readonly for cow.
1923          *
1924          * Modifying pi_state _before_ the user space value would
1925          * leave the pi_state in an inconsistent state when we fault
1926          * here, because we need to drop the hash bucket lock to
1927          * handle the fault. This might be observed in the PID check
1928          * in lookup_pi_state.
1929          */
1930 retry:
1931         if (get_futex_value_locked(&uval, uaddr))
1932                 goto handle_fault;
1933
1934         while (1) {
1935                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1936
1937                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1938                         goto handle_fault;
1939                 if (curval == uval)
1940                         break;
1941                 uval = curval;
1942         }
1943
1944         /*
1945          * We fixed up user space. Now we need to fix the pi_state
1946          * itself.
1947          */
1948         if (pi_state->owner != NULL) {
1949                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1950                 WARN_ON(list_empty(&pi_state->list));
1951                 list_del_init(&pi_state->list);
1952                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1953         }
1954
1955         pi_state->owner = newowner;
1956
1957         raw_spin_lock_irq(&newowner->pi_lock);
1958         WARN_ON(!list_empty(&pi_state->list));
1959         list_add(&pi_state->list, &newowner->pi_state_list);
1960         raw_spin_unlock_irq(&newowner->pi_lock);
1961         return 0;
1962
1963         /*
1964          * To handle the page fault we need to drop the hash bucket
1965          * lock here. That gives the other task (either the highest priority
1966          * waiter itself or the task which stole the rtmutex) the
1967          * chance to try the fixup of the pi_state. So once we are
1968          * back from handling the fault we need to check the pi_state
1969          * after reacquiring the hash bucket lock and before trying to
1970          * do another fixup. When the fixup has been done already we
1971          * simply return.
1972          */
1973 handle_fault:
1974         spin_unlock(q->lock_ptr);
1975
1976         ret = fault_in_user_writeable(uaddr);
1977
1978         spin_lock(q->lock_ptr);
1979
1980         /*
1981          * Check if someone else fixed it for us:
1982          */
1983         if (pi_state->owner != oldowner)
1984                 return 0;
1985
1986         if (ret)
1987                 return ret;
1988
1989         goto retry;
1990 }
1991
1992 static long futex_wait_restart(struct restart_block *restart);
1993
1994 /**
1995  * fixup_owner() - Post lock pi_state and corner case management
1996  * @uaddr:      user address of the futex
1997  * @q:          futex_q (contains pi_state and access to the rt_mutex)
1998  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
1999  *
2000  * After attempting to lock an rt_mutex, this function is called to cleanup
2001  * the pi_state owner as well as handle race conditions that may allow us to
2002  * acquire the lock. Must be called with the hb lock held.
2003  *
2004  * Return:
2005  *  1 - success, lock taken;
2006  *  0 - success, lock not taken;
2007  * <0 - on error (-EFAULT)
2008  */
2009 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2010 {
2011         struct task_struct *owner;
2012         int ret = 0;
2013
2014         if (locked) {
2015                 /*
2016                  * Got the lock. We might not be the anticipated owner if we
2017                  * did a lock-steal - fix up the PI-state in that case:
2018                  */
2019                 if (q->pi_state->owner != current)
2020                         ret = fixup_pi_state_owner(uaddr, q, current);
2021                 goto out;
2022         }
2023
2024         /*
2025          * Catch the rare case, where the lock was released when we were on the
2026          * way back before we locked the hash bucket.
2027          */
2028         if (q->pi_state->owner == current) {
2029                 /*
2030                  * Try to get the rt_mutex now. This might fail as some other
2031                  * task acquired the rt_mutex after we removed ourself from the
2032                  * rt_mutex waiters list.
2033                  */
2034                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2035                         locked = 1;
2036                         goto out;
2037                 }
2038
2039                 /*
2040                  * pi_state is incorrect, some other task did a lock steal and
2041                  * we returned due to timeout or signal without taking the
2042                  * rt_mutex. Too late.
2043                  */
2044                 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2045                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2046                 if (!owner)
2047                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2048                 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2049                 ret = fixup_pi_state_owner(uaddr, q, owner);
2050                 goto out;
2051         }
2052
2053         /*
2054          * Paranoia check. If we did not take the lock, then we should not be
2055          * the owner of the rt_mutex.
2056          */
2057         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2058                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2059                                 "pi-state %p\n", ret,
2060                                 q->pi_state->pi_mutex.owner,
2061                                 q->pi_state->owner);
2062
2063 out:
2064         return ret ? ret : locked;
2065 }
2066
2067 /**
2068  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2069  * @hb:         the futex hash bucket, must be locked by the caller
2070  * @q:          the futex_q to queue up on
2071  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2072  */
2073 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2074                                 struct hrtimer_sleeper *timeout)
2075 {
2076         /*
2077          * The task state is guaranteed to be set before another task can
2078          * wake it. set_current_state() is implemented using set_mb() and
2079          * queue_me() calls spin_unlock() upon completion, both serializing
2080          * access to the hash list and forcing another memory barrier.
2081          */
2082         set_current_state(TASK_INTERRUPTIBLE);
2083         queue_me(q, hb);
2084
2085         /* Arm the timer */
2086         if (timeout) {
2087                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2088                 if (!hrtimer_active(&timeout->timer))
2089                         timeout->task = NULL;
2090         }
2091
2092         /*
2093          * If we have been removed from the hash list, then another task
2094          * has tried to wake us, and we can skip the call to schedule().
2095          */
2096         if (likely(!plist_node_empty(&q->list))) {
2097                 /*
2098                  * If the timer has already expired, current will already be
2099                  * flagged for rescheduling. Only call schedule if there
2100                  * is no timeout, or if it has yet to expire.
2101                  */
2102                 if (!timeout || timeout->task)
2103                         freezable_schedule();
2104         }
2105         __set_current_state(TASK_RUNNING);
2106 }
2107
2108 /**
2109  * futex_wait_setup() - Prepare to wait on a futex
2110  * @uaddr:      the futex userspace address
2111  * @val:        the expected value
2112  * @flags:      futex flags (FLAGS_SHARED, etc.)
2113  * @q:          the associated futex_q
2114  * @hb:         storage for hash_bucket pointer to be returned to caller
2115  *
2116  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2117  * compare it with the expected value.  Handle atomic faults internally.
2118  * Return with the hb lock held and a q.key reference on success, and unlocked
2119  * with no q.key reference on failure.
2120  *
2121  * Return:
2122  *  0 - uaddr contains val and hb has been locked;
2123  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2124  */
2125 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2126                            struct futex_q *q, struct futex_hash_bucket **hb)
2127 {
2128         u32 uval;
2129         int ret;
2130
2131         /*
2132          * Access the page AFTER the hash-bucket is locked.
2133          * Order is important:
2134          *
2135          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2136          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2137          *
2138          * The basic logical guarantee of a futex is that it blocks ONLY
2139          * if cond(var) is known to be true at the time of blocking, for
2140          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2141          * would open a race condition where we could block indefinitely with
2142          * cond(var) false, which would violate the guarantee.
2143          *
2144          * On the other hand, we insert q and release the hash-bucket only
2145          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2146          * absorb a wakeup if *uaddr does not match the desired values
2147          * while the syscall executes.
2148          */
2149 retry:
2150         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2151         if (unlikely(ret != 0))
2152                 return ret;
2153
2154 retry_private:
2155         *hb = queue_lock(q);
2156
2157         ret = get_futex_value_locked(&uval, uaddr);
2158
2159         if (ret) {
2160                 queue_unlock(*hb);
2161
2162                 ret = get_user(uval, uaddr);
2163                 if (ret)
2164                         goto out;
2165
2166                 if (!(flags & FLAGS_SHARED))
2167                         goto retry_private;
2168
2169                 put_futex_key(&q->key);
2170                 goto retry;
2171         }
2172
2173         if (uval != val) {
2174                 queue_unlock(*hb);
2175                 ret = -EWOULDBLOCK;
2176         }
2177
2178 out:
2179         if (ret)
2180                 put_futex_key(&q->key);
2181         return ret;
2182 }
2183
2184 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2185                       ktime_t *abs_time, u32 bitset)
2186 {
2187         struct hrtimer_sleeper timeout, *to = NULL;
2188         struct restart_block *restart;
2189         struct futex_hash_bucket *hb;
2190         struct futex_q q = futex_q_init;
2191         int ret;
2192
2193         if (!bitset)
2194                 return -EINVAL;
2195         q.bitset = bitset;
2196
2197         if (abs_time) {
2198                 to = &timeout;
2199
2200                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2201                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2202                                       HRTIMER_MODE_ABS);
2203                 hrtimer_init_sleeper(to, current);
2204                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2205                                              current->timer_slack_ns);
2206         }
2207
2208 retry:
2209         /*
2210          * Prepare to wait on uaddr. On success, holds hb lock and increments
2211          * q.key refs.
2212          */
2213         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2214         if (ret)
2215                 goto out;
2216
2217         /* queue_me and wait for wakeup, timeout, or a signal. */
2218         futex_wait_queue_me(hb, &q, to);
2219
2220         /* If we were woken (and unqueued), we succeeded, whatever. */
2221         ret = 0;
2222         /* unqueue_me() drops q.key ref */
2223         if (!unqueue_me(&q))
2224                 goto out;
2225         ret = -ETIMEDOUT;
2226         if (to && !to->task)
2227                 goto out;
2228
2229         /*
2230          * We expect signal_pending(current), but we might be the
2231          * victim of a spurious wakeup as well.
2232          */
2233         if (!signal_pending(current))
2234                 goto retry;
2235
2236         ret = -ERESTARTSYS;
2237         if (!abs_time)
2238                 goto out;
2239
2240         restart = &current_thread_info()->restart_block;
2241         restart->fn = futex_wait_restart;
2242         restart->futex.uaddr = uaddr;
2243         restart->futex.val = val;
2244         restart->futex.time = abs_time->tv64;
2245         restart->futex.bitset = bitset;
2246         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2247
2248         ret = -ERESTART_RESTARTBLOCK;
2249
2250 out:
2251         if (to) {
2252                 hrtimer_cancel(&to->timer);
2253                 destroy_hrtimer_on_stack(&to->timer);
2254         }
2255         return ret;
2256 }
2257
2258
2259 static long futex_wait_restart(struct restart_block *restart)
2260 {
2261         u32 __user *uaddr = restart->futex.uaddr;
2262         ktime_t t, *tp = NULL;
2263
2264         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2265                 t.tv64 = restart->futex.time;
2266                 tp = &t;
2267         }
2268         restart->fn = do_no_restart_syscall;
2269
2270         return (long)futex_wait(uaddr, restart->futex.flags,
2271                                 restart->futex.val, tp, restart->futex.bitset);
2272 }
2273
2274
2275 /*
2276  * Userspace tried a 0 -> TID atomic transition of the futex value
2277  * and failed. The kernel side here does the whole locking operation:
2278  * if there are waiters then it will block, it does PI, etc. (Due to
2279  * races the kernel might see a 0 value of the futex too.)
2280  */
2281 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2282                          ktime_t *time, int trylock)
2283 {
2284         struct hrtimer_sleeper timeout, *to = NULL;
2285         struct futex_hash_bucket *hb;
2286         struct futex_q q = futex_q_init;
2287         int res, ret;
2288
2289         if (refill_pi_state_cache())
2290                 return -ENOMEM;
2291
2292         if (time) {
2293                 to = &timeout;
2294                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2295                                       HRTIMER_MODE_ABS);
2296                 hrtimer_init_sleeper(to, current);
2297                 hrtimer_set_expires(&to->timer, *time);
2298         }
2299
2300 retry:
2301         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2302         if (unlikely(ret != 0))
2303                 goto out;
2304
2305 retry_private:
2306         hb = queue_lock(&q);
2307
2308         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2309         if (unlikely(ret)) {
2310                 switch (ret) {
2311                 case 1:
2312                         /* We got the lock. */
2313                         ret = 0;
2314                         goto out_unlock_put_key;
2315                 case -EFAULT:
2316                         goto uaddr_faulted;
2317                 case -EAGAIN:
2318                         /*
2319                          * Task is exiting and we just wait for the
2320                          * exit to complete.
2321                          */
2322                         queue_unlock(hb);
2323                         put_futex_key(&q.key);
2324                         cond_resched();
2325                         goto retry;
2326                 default:
2327                         goto out_unlock_put_key;
2328                 }
2329         }
2330
2331         /*
2332          * Only actually queue now that the atomic ops are done:
2333          */
2334         queue_me(&q, hb);
2335
2336         WARN_ON(!q.pi_state);
2337         /*
2338          * Block on the PI mutex:
2339          */
2340         if (!trylock)
2341                 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2342         else {
2343                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2344                 /* Fixup the trylock return value: */
2345                 ret = ret ? 0 : -EWOULDBLOCK;
2346         }
2347
2348         spin_lock(q.lock_ptr);
2349         /*
2350          * Fixup the pi_state owner and possibly acquire the lock if we
2351          * haven't already.
2352          */
2353         res = fixup_owner(uaddr, &q, !ret);
2354         /*
2355          * If fixup_owner() returned an error, proprogate that.  If it acquired
2356          * the lock, clear our -ETIMEDOUT or -EINTR.
2357          */
2358         if (res)
2359                 ret = (res < 0) ? res : 0;
2360
2361         /*
2362          * If fixup_owner() faulted and was unable to handle the fault, unlock
2363          * it and return the fault to userspace.
2364          */
2365         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2366                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2367
2368         /* Unqueue and drop the lock */
2369         unqueue_me_pi(&q);
2370
2371         goto out_put_key;
2372
2373 out_unlock_put_key:
2374         queue_unlock(hb);
2375
2376 out_put_key:
2377         put_futex_key(&q.key);
2378 out:
2379         if (to)
2380                 destroy_hrtimer_on_stack(&to->timer);
2381         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2382
2383 uaddr_faulted:
2384         queue_unlock(hb);
2385
2386         ret = fault_in_user_writeable(uaddr);
2387         if (ret)
2388                 goto out_put_key;
2389
2390         if (!(flags & FLAGS_SHARED))
2391                 goto retry_private;
2392
2393         put_futex_key(&q.key);
2394         goto retry;
2395 }
2396
2397 /*
2398  * Userspace attempted a TID -> 0 atomic transition, and failed.
2399  * This is the in-kernel slowpath: we look up the PI state (if any),
2400  * and do the rt-mutex unlock.
2401  */
2402 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2403 {
2404         struct futex_hash_bucket *hb;
2405         struct futex_q *this, *next;
2406         union futex_key key = FUTEX_KEY_INIT;
2407         u32 uval, vpid = task_pid_vnr(current);
2408         int ret;
2409
2410 retry:
2411         if (get_user(uval, uaddr))
2412                 return -EFAULT;
2413         /*
2414          * We release only a lock we actually own:
2415          */
2416         if ((uval & FUTEX_TID_MASK) != vpid)
2417                 return -EPERM;
2418
2419         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2420         if (unlikely(ret != 0))
2421                 goto out;
2422
2423         hb = hash_futex(&key);
2424         spin_lock(&hb->lock);
2425
2426         /*
2427          * To avoid races, try to do the TID -> 0 atomic transition
2428          * again. If it succeeds then we can return without waking
2429          * anyone else up. We only try this if neither the waiters nor
2430          * the owner died bit are set.
2431          */
2432         if (!(uval & ~FUTEX_TID_MASK) &&
2433             cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2434                 goto pi_faulted;
2435         /*
2436          * Rare case: we managed to release the lock atomically,
2437          * no need to wake anyone else up:
2438          */
2439         if (unlikely(uval == vpid))
2440                 goto out_unlock;
2441
2442         /*
2443          * Ok, other tasks may need to be woken up - check waiters
2444          * and do the wakeup if necessary:
2445          */
2446         plist_for_each_entry_safe(this, next, &hb->chain, list) {
2447                 if (!match_futex (&this->key, &key))
2448                         continue;
2449                 ret = wake_futex_pi(uaddr, uval, this);
2450                 /*
2451                  * The atomic access to the futex value
2452                  * generated a pagefault, so retry the
2453                  * user-access and the wakeup:
2454                  */
2455                 if (ret == -EFAULT)
2456                         goto pi_faulted;
2457                 goto out_unlock;
2458         }
2459         /*
2460          * No waiters - kernel unlocks the futex:
2461          */
2462         ret = unlock_futex_pi(uaddr, uval);
2463         if (ret == -EFAULT)
2464                 goto pi_faulted;
2465
2466 out_unlock:
2467         spin_unlock(&hb->lock);
2468         put_futex_key(&key);
2469
2470 out:
2471         return ret;
2472
2473 pi_faulted:
2474         spin_unlock(&hb->lock);
2475         put_futex_key(&key);
2476
2477         ret = fault_in_user_writeable(uaddr);
2478         if (!ret)
2479                 goto retry;
2480
2481         return ret;
2482 }
2483
2484 /**
2485  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2486  * @hb:         the hash_bucket futex_q was original enqueued on
2487  * @q:          the futex_q woken while waiting to be requeued
2488  * @key2:       the futex_key of the requeue target futex
2489  * @timeout:    the timeout associated with the wait (NULL if none)
2490  *
2491  * Detect if the task was woken on the initial futex as opposed to the requeue
2492  * target futex.  If so, determine if it was a timeout or a signal that caused
2493  * the wakeup and return the appropriate error code to the caller.  Must be
2494  * called with the hb lock held.
2495  *
2496  * Return:
2497  *  0 = no early wakeup detected;
2498  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2499  */
2500 static inline
2501 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2502                                    struct futex_q *q, union futex_key *key2,
2503                                    struct hrtimer_sleeper *timeout)
2504 {
2505         int ret = 0;
2506
2507         /*
2508          * With the hb lock held, we avoid races while we process the wakeup.
2509          * We only need to hold hb (and not hb2) to ensure atomicity as the
2510          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2511          * It can't be requeued from uaddr2 to something else since we don't
2512          * support a PI aware source futex for requeue.
2513          */
2514         if (!match_futex(&q->key, key2)) {
2515                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2516                 /*
2517                  * We were woken prior to requeue by a timeout or a signal.
2518                  * Unqueue the futex_q and determine which it was.
2519                  */
2520                 plist_del(&q->list, &hb->chain);
2521                 hb_waiters_dec(hb);
2522
2523                 /* Handle spurious wakeups gracefully */
2524                 ret = -EWOULDBLOCK;
2525                 if (timeout && !timeout->task)
2526                         ret = -ETIMEDOUT;
2527                 else if (signal_pending(current))
2528                         ret = -ERESTARTNOINTR;
2529         }
2530         return ret;
2531 }
2532
2533 /**
2534  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2535  * @uaddr:      the futex we initially wait on (non-pi)
2536  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2537  *              the same type, no requeueing from private to shared, etc.
2538  * @val:        the expected value of uaddr
2539  * @abs_time:   absolute timeout
2540  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2541  * @uaddr2:     the pi futex we will take prior to returning to user-space
2542  *
2543  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2544  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2545  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2546  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2547  * without one, the pi logic would not know which task to boost/deboost, if
2548  * there was a need to.
2549  *
2550  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2551  * via the following--
2552  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2553  * 2) wakeup on uaddr2 after a requeue
2554  * 3) signal
2555  * 4) timeout
2556  *
2557  * If 3, cleanup and return -ERESTARTNOINTR.
2558  *
2559  * If 2, we may then block on trying to take the rt_mutex and return via:
2560  * 5) successful lock
2561  * 6) signal
2562  * 7) timeout
2563  * 8) other lock acquisition failure
2564  *
2565  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2566  *
2567  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2568  *
2569  * Return:
2570  *  0 - On success;
2571  * <0 - On error
2572  */
2573 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2574                                  u32 val, ktime_t *abs_time, u32 bitset,
2575                                  u32 __user *uaddr2)
2576 {
2577         struct hrtimer_sleeper timeout, *to = NULL;
2578         struct rt_mutex_waiter rt_waiter;
2579         struct rt_mutex *pi_mutex = NULL;
2580         struct futex_hash_bucket *hb;
2581         union futex_key key2 = FUTEX_KEY_INIT;
2582         struct futex_q q = futex_q_init;
2583         int res, ret;
2584
2585         if (uaddr == uaddr2)
2586                 return -EINVAL;
2587
2588         if (!bitset)
2589                 return -EINVAL;
2590
2591         if (abs_time) {
2592                 to = &timeout;
2593                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2594                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2595                                       HRTIMER_MODE_ABS);
2596                 hrtimer_init_sleeper(to, current);
2597                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2598                                              current->timer_slack_ns);
2599         }
2600
2601         /*
2602          * The waiter is allocated on our stack, manipulated by the requeue
2603          * code while we sleep on uaddr.
2604          */
2605         debug_rt_mutex_init_waiter(&rt_waiter);
2606         RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2607         RB_CLEAR_NODE(&rt_waiter.tree_entry);
2608         rt_waiter.task = NULL;
2609
2610         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2611         if (unlikely(ret != 0))
2612                 goto out;
2613
2614         q.bitset = bitset;
2615         q.rt_waiter = &rt_waiter;
2616         q.requeue_pi_key = &key2;
2617
2618         /*
2619          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2620          * count.
2621          */
2622         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2623         if (ret)
2624                 goto out_key2;
2625
2626         /*
2627          * The check above which compares uaddrs is not sufficient for
2628          * shared futexes. We need to compare the keys:
2629          */
2630         if (match_futex(&q.key, &key2)) {
2631                 ret = -EINVAL;
2632                 goto out_put_keys;
2633         }
2634
2635         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2636         futex_wait_queue_me(hb, &q, to);
2637
2638         spin_lock(&hb->lock);
2639         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2640         spin_unlock(&hb->lock);
2641         if (ret)
2642                 goto out_put_keys;
2643
2644         /*
2645          * In order for us to be here, we know our q.key == key2, and since
2646          * we took the hb->lock above, we also know that futex_requeue() has
2647          * completed and we no longer have to concern ourselves with a wakeup
2648          * race with the atomic proxy lock acquisition by the requeue code. The
2649          * futex_requeue dropped our key1 reference and incremented our key2
2650          * reference count.
2651          */
2652
2653         /* Check if the requeue code acquired the second futex for us. */
2654         if (!q.rt_waiter) {
2655                 /*
2656                  * Got the lock. We might not be the anticipated owner if we
2657                  * did a lock-steal - fix up the PI-state in that case.
2658                  */
2659                 if (q.pi_state && (q.pi_state->owner != current)) {
2660                         spin_lock(q.lock_ptr);
2661                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2662                         spin_unlock(q.lock_ptr);
2663                 }
2664         } else {
2665                 /*
2666                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2667                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2668                  * the pi_state.
2669                  */
2670                 WARN_ON(!q.pi_state);
2671                 pi_mutex = &q.pi_state->pi_mutex;
2672                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2673                 debug_rt_mutex_free_waiter(&rt_waiter);
2674
2675                 spin_lock(q.lock_ptr);
2676                 /*
2677                  * Fixup the pi_state owner and possibly acquire the lock if we
2678                  * haven't already.
2679                  */
2680                 res = fixup_owner(uaddr2, &q, !ret);
2681                 /*
2682                  * If fixup_owner() returned an error, proprogate that.  If it
2683                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2684                  */
2685                 if (res)
2686                         ret = (res < 0) ? res : 0;
2687
2688                 /* Unqueue and drop the lock. */
2689                 unqueue_me_pi(&q);
2690         }
2691
2692         /*
2693          * If fixup_pi_state_owner() faulted and was unable to handle the
2694          * fault, unlock the rt_mutex and return the fault to userspace.
2695          */
2696         if (ret == -EFAULT) {
2697                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2698                         rt_mutex_unlock(pi_mutex);
2699         } else if (ret == -EINTR) {
2700                 /*
2701                  * We've already been requeued, but cannot restart by calling
2702                  * futex_lock_pi() directly. We could restart this syscall, but
2703                  * it would detect that the user space "val" changed and return
2704                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2705                  * -EWOULDBLOCK directly.
2706                  */
2707                 ret = -EWOULDBLOCK;
2708         }
2709
2710 out_put_keys:
2711         put_futex_key(&q.key);
2712 out_key2:
2713         put_futex_key(&key2);
2714
2715 out:
2716         if (to) {
2717                 hrtimer_cancel(&to->timer);
2718                 destroy_hrtimer_on_stack(&to->timer);
2719         }
2720         return ret;
2721 }
2722
2723 /*
2724  * Support for robust futexes: the kernel cleans up held futexes at
2725  * thread exit time.
2726  *
2727  * Implementation: user-space maintains a per-thread list of locks it
2728  * is holding. Upon do_exit(), the kernel carefully walks this list,
2729  * and marks all locks that are owned by this thread with the
2730  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2731  * always manipulated with the lock held, so the list is private and
2732  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2733  * field, to allow the kernel to clean up if the thread dies after
2734  * acquiring the lock, but just before it could have added itself to
2735  * the list. There can only be one such pending lock.
2736  */
2737
2738 /**
2739  * sys_set_robust_list() - Set the robust-futex list head of a task
2740  * @head:       pointer to the list-head
2741  * @len:        length of the list-head, as userspace expects
2742  */
2743 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2744                 size_t, len)
2745 {
2746         if (!futex_cmpxchg_enabled)
2747                 return -ENOSYS;
2748         /*
2749          * The kernel knows only one size for now:
2750          */
2751         if (unlikely(len != sizeof(*head)))
2752                 return -EINVAL;
2753
2754         current->robust_list = head;
2755
2756         return 0;
2757 }
2758
2759 /**
2760  * sys_get_robust_list() - Get the robust-futex list head of a task
2761  * @pid:        pid of the process [zero for current task]
2762  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2763  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2764  */
2765 SYSCALL_DEFINE3(get_robust_list, int, pid,
2766                 struct robust_list_head __user * __user *, head_ptr,
2767                 size_t __user *, len_ptr)
2768 {
2769         struct robust_list_head __user *head;
2770         unsigned long ret;
2771         struct task_struct *p;
2772
2773         if (!futex_cmpxchg_enabled)
2774                 return -ENOSYS;
2775
2776         rcu_read_lock();
2777
2778         ret = -ESRCH;
2779         if (!pid)
2780                 p = current;
2781         else {
2782                 p = find_task_by_vpid(pid);
2783                 if (!p)
2784                         goto err_unlock;
2785         }
2786
2787         ret = -EPERM;
2788         if (!ptrace_may_access(p, PTRACE_MODE_READ))
2789                 goto err_unlock;
2790
2791         head = p->robust_list;
2792         rcu_read_unlock();
2793
2794         if (put_user(sizeof(*head), len_ptr))
2795                 return -EFAULT;
2796         return put_user(head, head_ptr);
2797
2798 err_unlock:
2799         rcu_read_unlock();
2800
2801         return ret;
2802 }
2803
2804 /*
2805  * Process a futex-list entry, check whether it's owned by the
2806  * dying task, and do notification if so:
2807  */
2808 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2809 {
2810         u32 uval, uninitialized_var(nval), mval;
2811
2812 retry:
2813         if (get_user(uval, uaddr))
2814                 return -1;
2815
2816         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2817                 /*
2818                  * Ok, this dying thread is truly holding a futex
2819                  * of interest. Set the OWNER_DIED bit atomically
2820                  * via cmpxchg, and if the value had FUTEX_WAITERS
2821                  * set, wake up a waiter (if any). (We have to do a
2822                  * futex_wake() even if OWNER_DIED is already set -
2823                  * to handle the rare but possible case of recursive
2824                  * thread-death.) The rest of the cleanup is done in
2825                  * userspace.
2826                  */
2827                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2828                 /*
2829                  * We are not holding a lock here, but we want to have
2830                  * the pagefault_disable/enable() protection because
2831                  * we want to handle the fault gracefully. If the
2832                  * access fails we try to fault in the futex with R/W
2833                  * verification via get_user_pages. get_user() above
2834                  * does not guarantee R/W access. If that fails we
2835                  * give up and leave the futex locked.
2836                  */
2837                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2838                         if (fault_in_user_writeable(uaddr))
2839                                 return -1;
2840                         goto retry;
2841                 }
2842                 if (nval != uval)
2843                         goto retry;
2844
2845                 /*
2846                  * Wake robust non-PI futexes here. The wakeup of
2847                  * PI futexes happens in exit_pi_state():
2848                  */
2849                 if (!pi && (uval & FUTEX_WAITERS))
2850                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2851         }
2852         return 0;
2853 }
2854
2855 /*
2856  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2857  */
2858 static inline int fetch_robust_entry(struct robust_list __user **entry,
2859                                      struct robust_list __user * __user *head,
2860                                      unsigned int *pi)
2861 {
2862         unsigned long uentry;
2863
2864         if (get_user(uentry, (unsigned long __user *)head))
2865                 return -EFAULT;
2866
2867         *entry = (void __user *)(uentry & ~1UL);
2868         *pi = uentry & 1;
2869
2870         return 0;
2871 }
2872
2873 /*
2874  * Walk curr->robust_list (very carefully, it's a userspace list!)
2875  * and mark any locks found there dead, and notify any waiters.
2876  *
2877  * We silently return on any sign of list-walking problem.
2878  */
2879 void exit_robust_list(struct task_struct *curr)
2880 {
2881         struct robust_list_head __user *head = curr->robust_list;
2882         struct robust_list __user *entry, *next_entry, *pending;
2883         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2884         unsigned int uninitialized_var(next_pi);
2885         unsigned long futex_offset;
2886         int rc;
2887
2888         if (!futex_cmpxchg_enabled)
2889                 return;
2890
2891         /*
2892          * Fetch the list head (which was registered earlier, via
2893          * sys_set_robust_list()):
2894          */
2895         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2896                 return;
2897         /*
2898          * Fetch the relative futex offset:
2899          */
2900         if (get_user(futex_offset, &head->futex_offset))
2901                 return;
2902         /*
2903          * Fetch any possibly pending lock-add first, and handle it
2904          * if it exists:
2905          */
2906         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2907                 return;
2908
2909         next_entry = NULL;      /* avoid warning with gcc */
2910         while (entry != &head->list) {
2911                 /*
2912                  * Fetch the next entry in the list before calling
2913                  * handle_futex_death:
2914                  */
2915                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2916                 /*
2917                  * A pending lock might already be on the list, so
2918                  * don't process it twice:
2919                  */
2920                 if (entry != pending)
2921                         if (handle_futex_death((void __user *)entry + futex_offset,
2922                                                 curr, pi))
2923                                 return;
2924                 if (rc)
2925                         return;
2926                 entry = next_entry;
2927                 pi = next_pi;
2928                 /*
2929                  * Avoid excessively long or circular lists:
2930                  */
2931                 if (!--limit)
2932                         break;
2933
2934                 cond_resched();
2935         }
2936
2937         if (pending)
2938                 handle_futex_death((void __user *)pending + futex_offset,
2939                                    curr, pip);
2940 }
2941
2942 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2943                 u32 __user *uaddr2, u32 val2, u32 val3)
2944 {
2945         int cmd = op & FUTEX_CMD_MASK;
2946         unsigned int flags = 0;
2947
2948         if (!(op & FUTEX_PRIVATE_FLAG))
2949                 flags |= FLAGS_SHARED;
2950
2951         if (op & FUTEX_CLOCK_REALTIME) {
2952                 flags |= FLAGS_CLOCKRT;
2953                 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2954                         return -ENOSYS;
2955         }
2956
2957         switch (cmd) {
2958         case FUTEX_LOCK_PI:
2959         case FUTEX_UNLOCK_PI:
2960         case FUTEX_TRYLOCK_PI:
2961         case FUTEX_WAIT_REQUEUE_PI:
2962         case FUTEX_CMP_REQUEUE_PI:
2963                 if (!futex_cmpxchg_enabled)
2964                         return -ENOSYS;
2965         }
2966
2967         switch (cmd) {
2968         case FUTEX_WAIT:
2969                 val3 = FUTEX_BITSET_MATCH_ANY;
2970         case FUTEX_WAIT_BITSET:
2971                 return futex_wait(uaddr, flags, val, timeout, val3);
2972         case FUTEX_WAKE:
2973                 val3 = FUTEX_BITSET_MATCH_ANY;
2974         case FUTEX_WAKE_BITSET:
2975                 return futex_wake(uaddr, flags, val, val3);
2976         case FUTEX_REQUEUE:
2977                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2978         case FUTEX_CMP_REQUEUE:
2979                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2980         case FUTEX_WAKE_OP:
2981                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2982         case FUTEX_LOCK_PI:
2983                 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2984         case FUTEX_UNLOCK_PI:
2985                 return futex_unlock_pi(uaddr, flags);
2986         case FUTEX_TRYLOCK_PI:
2987                 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2988         case FUTEX_WAIT_REQUEUE_PI:
2989                 val3 = FUTEX_BITSET_MATCH_ANY;
2990                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2991                                              uaddr2);
2992         case FUTEX_CMP_REQUEUE_PI:
2993                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2994         }
2995         return -ENOSYS;
2996 }
2997
2998
2999 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3000                 struct timespec __user *, utime, u32 __user *, uaddr2,
3001                 u32, val3)
3002 {
3003         struct timespec ts;
3004         ktime_t t, *tp = NULL;
3005         u32 val2 = 0;
3006         int cmd = op & FUTEX_CMD_MASK;
3007
3008         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3009                       cmd == FUTEX_WAIT_BITSET ||
3010                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3011                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3012                         return -EFAULT;
3013                 if (!timespec_valid(&ts))
3014                         return -EINVAL;
3015
3016                 t = timespec_to_ktime(ts);
3017                 if (cmd == FUTEX_WAIT)
3018                         t = ktime_add_safe(ktime_get(), t);
3019                 tp = &t;
3020         }
3021         /*
3022          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3023          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3024          */
3025         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3026             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3027                 val2 = (u32) (unsigned long) utime;
3028
3029         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3030 }
3031
3032 static void __init futex_detect_cmpxchg(void)
3033 {
3034 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3035         u32 curval;
3036
3037         /*
3038          * This will fail and we want it. Some arch implementations do
3039          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3040          * functionality. We want to know that before we call in any
3041          * of the complex code paths. Also we want to prevent
3042          * registration of robust lists in that case. NULL is
3043          * guaranteed to fault and we get -EFAULT on functional
3044          * implementation, the non-functional ones will return
3045          * -ENOSYS.
3046          */
3047         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3048                 futex_cmpxchg_enabled = 1;
3049 #endif
3050 }
3051
3052 static int __init futex_init(void)
3053 {
3054         unsigned int futex_shift;
3055         unsigned long i;
3056
3057 #if CONFIG_BASE_SMALL
3058         futex_hashsize = 16;
3059 #else
3060         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3061 #endif
3062
3063         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3064                                                futex_hashsize, 0,
3065                                                futex_hashsize < 256 ? HASH_SMALL : 0,
3066                                                &futex_shift, NULL,
3067                                                futex_hashsize, futex_hashsize);
3068         futex_hashsize = 1UL << futex_shift;
3069
3070         futex_detect_cmpxchg();
3071
3072         for (i = 0; i < futex_hashsize; i++) {
3073                 atomic_set(&futex_queues[i].waiters, 0);
3074                 plist_head_init(&futex_queues[i].chain);
3075                 spin_lock_init(&futex_queues[i].lock);
3076         }
3077
3078         return 0;
3079 }
3080 __initcall(futex_init);