Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/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 #include <linux/fault-inject.h>
 68 
 69 #include <asm/futex.h>
 70 
 71 #include "locking/rtmutex_common.h"
 72 
 73 /*
 74  * READ this before attempting to hack on futexes!
 75  *
 76  * Basic futex operation and ordering guarantees
 77  * =============================================
 78  *
 79  * The waiter reads the futex value in user space and calls
 80  * futex_wait(). This function computes the hash bucket and acquires
 81  * the hash bucket lock. After that it reads the futex user space value
 82  * again and verifies that the data has not changed. If it has not changed
 83  * it enqueues itself into the hash bucket, releases the hash bucket lock
 84  * and schedules.
 85  *
 86  * The waker side modifies the user space value of the futex and calls
 87  * futex_wake(). This function computes the hash bucket and acquires the
 88  * hash bucket lock. Then it looks for waiters on that futex in the hash
 89  * bucket and wakes them.
 90  *
 91  * In futex wake up scenarios where no tasks are blocked on a futex, taking
 92  * the hb spinlock can be avoided and simply return. In order for this
 93  * optimization to work, ordering guarantees must exist so that the waiter
 94  * being added to the list is acknowledged when the list is concurrently being
 95  * checked by the waker, avoiding scenarios like the following:
 96  *
 97  * CPU 0                               CPU 1
 98  * val = *futex;
 99  * sys_futex(WAIT, futex, val);
100  *   futex_wait(futex, val);
101  *   uval = *futex;
102  *                                     *futex = newval;
103  *                                     sys_futex(WAKE, futex);
104  *                                       futex_wake(futex);
105  *                                       if (queue_empty())
106  *                                         return;
107  *   if (uval == val)
108  *      lock(hash_bucket(futex));
109  *      queue();
110  *     unlock(hash_bucket(futex));
111  *     schedule();
112  *
113  * This would cause the waiter on CPU 0 to wait forever because it
114  * missed the transition of the user space value from val to newval
115  * and the waker did not find the waiter in the hash bucket queue.
116  *
117  * The correct serialization ensures that a waiter either observes
118  * the changed user space value before blocking or is woken by a
119  * concurrent waker:
120  *
121  * CPU 0                                 CPU 1
122  * val = *futex;
123  * sys_futex(WAIT, futex, val);
124  *   futex_wait(futex, val);
125  *
126  *   waiters++; (a)
127  *   mb(); (A) <-- paired with -.
128  *                              |
129  *   lock(hash_bucket(futex));  |
130  *                              |
131  *   uval = *futex;             |
132  *                              |        *futex = newval;
133  *                              |        sys_futex(WAKE, futex);
134  *                              |          futex_wake(futex);
135  *                              |
136  *                              `------->  mb(); (B)
137  *   if (uval == val)
138  *     queue();
139  *     unlock(hash_bucket(futex));
140  *     schedule();                         if (waiters)
141  *                                           lock(hash_bucket(futex));
142  *   else                                    wake_waiters(futex);
143  *     waiters--; (b)                        unlock(hash_bucket(futex));
144  *
145  * Where (A) orders the waiters increment and the futex value read through
146  * atomic operations (see hb_waiters_inc) and where (B) orders the write
147  * to futex and the waiters read -- this is done by the barriers for both
148  * shared and private futexes in get_futex_key_refs().
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 /*
259  * The base of the bucket array and its size are always used together
260  * (after initialization only in hash_futex()), so ensure that they
261  * reside in the same cacheline.
262  */
263 static struct {
264         struct futex_hash_bucket *queues;
265         unsigned long            hashsize;
266 } __futex_data __read_mostly __aligned(2*sizeof(long));
267 #define futex_queues   (__futex_data.queues)
268 #define futex_hashsize (__futex_data.hashsize)
269 
270 
271 /*
272  * Fault injections for futexes.
273  */
274 #ifdef CONFIG_FAIL_FUTEX
275 
276 static struct {
277         struct fault_attr attr;
278 
279         bool ignore_private;
280 } fail_futex = {
281         .attr = FAULT_ATTR_INITIALIZER,
282         .ignore_private = false,
283 };
284 
285 static int __init setup_fail_futex(char *str)
286 {
287         return setup_fault_attr(&fail_futex.attr, str);
288 }
289 __setup("fail_futex=", setup_fail_futex);
290 
291 static bool should_fail_futex(bool fshared)
292 {
293         if (fail_futex.ignore_private && !fshared)
294                 return false;
295 
296         return should_fail(&fail_futex.attr, 1);
297 }
298 
299 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
300 
301 static int __init fail_futex_debugfs(void)
302 {
303         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
304         struct dentry *dir;
305 
306         dir = fault_create_debugfs_attr("fail_futex", NULL,
307                                         &fail_futex.attr);
308         if (IS_ERR(dir))
309                 return PTR_ERR(dir);
310 
311         if (!debugfs_create_bool("ignore-private", mode, dir,
312                                  &fail_futex.ignore_private)) {
313                 debugfs_remove_recursive(dir);
314                 return -ENOMEM;
315         }
316 
317         return 0;
318 }
319 
320 late_initcall(fail_futex_debugfs);
321 
322 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
323 
324 #else
325 static inline bool should_fail_futex(bool fshared)
326 {
327         return false;
328 }
329 #endif /* CONFIG_FAIL_FUTEX */
330 
331 static inline void futex_get_mm(union futex_key *key)
332 {
333         atomic_inc(&key->private.mm->mm_count);
334         /*
335          * Ensure futex_get_mm() implies a full barrier such that
336          * get_futex_key() implies a full barrier. This is relied upon
337          * as full barrier (B), see the ordering comment above.
338          */
339         smp_mb__after_atomic();
340 }
341 
342 /*
343  * Reflects a new waiter being added to the waitqueue.
344  */
345 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
346 {
347 #ifdef CONFIG_SMP
348         atomic_inc(&hb->waiters);
349         /*
350          * Full barrier (A), see the ordering comment above.
351          */
352         smp_mb__after_atomic();
353 #endif
354 }
355 
356 /*
357  * Reflects a waiter being removed from the waitqueue by wakeup
358  * paths.
359  */
360 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
361 {
362 #ifdef CONFIG_SMP
363         atomic_dec(&hb->waiters);
364 #endif
365 }
366 
367 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
368 {
369 #ifdef CONFIG_SMP
370         return atomic_read(&hb->waiters);
371 #else
372         return 1;
373 #endif
374 }
375 
376 /*
377  * We hash on the keys returned from get_futex_key (see below).
378  */
379 static struct futex_hash_bucket *hash_futex(union futex_key *key)
380 {
381         u32 hash = jhash2((u32*)&key->both.word,
382                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
383                           key->both.offset);
384         return &futex_queues[hash & (futex_hashsize - 1)];
385 }
386 
387 /*
388  * Return 1 if two futex_keys are equal, 0 otherwise.
389  */
390 static inline int match_futex(union futex_key *key1, union futex_key *key2)
391 {
392         return (key1 && key2
393                 && key1->both.word == key2->both.word
394                 && key1->both.ptr == key2->both.ptr
395                 && key1->both.offset == key2->both.offset);
396 }
397 
398 /*
399  * Take a reference to the resource addressed by a key.
400  * Can be called while holding spinlocks.
401  *
402  */
403 static void get_futex_key_refs(union futex_key *key)
404 {
405         if (!key->both.ptr)
406                 return;
407 
408         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
409         case FUT_OFF_INODE:
410                 ihold(key->shared.inode); /* implies MB (B) */
411                 break;
412         case FUT_OFF_MMSHARED:
413                 futex_get_mm(key); /* implies MB (B) */
414                 break;
415         default:
416                 /*
417                  * Private futexes do not hold reference on an inode or
418                  * mm, therefore the only purpose of calling get_futex_key_refs
419                  * is because we need the barrier for the lockless waiter check.
420                  */
421                 smp_mb(); /* explicit MB (B) */
422         }
423 }
424 
425 /*
426  * Drop a reference to the resource addressed by a key.
427  * The hash bucket spinlock must not be held. This is
428  * a no-op for private futexes, see comment in the get
429  * counterpart.
430  */
431 static void drop_futex_key_refs(union futex_key *key)
432 {
433         if (!key->both.ptr) {
434                 /* If we're here then we tried to put a key we failed to get */
435                 WARN_ON_ONCE(1);
436                 return;
437         }
438 
439         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
440         case FUT_OFF_INODE:
441                 iput(key->shared.inode);
442                 break;
443         case FUT_OFF_MMSHARED:
444                 mmdrop(key->private.mm);
445                 break;
446         }
447 }
448 
449 /**
450  * get_futex_key() - Get parameters which are the keys for a futex
451  * @uaddr:      virtual address of the futex
452  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
453  * @key:        address where result is stored.
454  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
455  *              VERIFY_WRITE)
456  *
457  * Return: a negative error code or 0
458  *
459  * The key words are stored in *key on success.
460  *
461  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
462  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
463  * We can usually work out the index without swapping in the page.
464  *
465  * lock_page() might sleep, the caller should not hold a spinlock.
466  */
467 static int
468 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
469 {
470         unsigned long address = (unsigned long)uaddr;
471         struct mm_struct *mm = current->mm;
472         struct page *page;
473         struct address_space *mapping;
474         int err, ro = 0;
475 
476         /*
477          * The futex address must be "naturally" aligned.
478          */
479         key->both.offset = address % PAGE_SIZE;
480         if (unlikely((address % sizeof(u32)) != 0))
481                 return -EINVAL;
482         address -= key->both.offset;
483 
484         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
485                 return -EFAULT;
486 
487         if (unlikely(should_fail_futex(fshared)))
488                 return -EFAULT;
489 
490         /*
491          * PROCESS_PRIVATE futexes are fast.
492          * As the mm cannot disappear under us and the 'key' only needs
493          * virtual address, we dont even have to find the underlying vma.
494          * Note : We do have to check 'uaddr' is a valid user address,
495          *        but access_ok() should be faster than find_vma()
496          */
497         if (!fshared) {
498                 key->private.mm = mm;
499                 key->private.address = address;
500                 get_futex_key_refs(key);  /* implies MB (B) */
501                 return 0;
502         }
503 
504 again:
505         /* Ignore any VERIFY_READ mapping (futex common case) */
506         if (unlikely(should_fail_futex(fshared)))
507                 return -EFAULT;
508 
509         err = get_user_pages_fast(address, 1, 1, &page);
510         /*
511          * If write access is not required (eg. FUTEX_WAIT), try
512          * and get read-only access.
513          */
514         if (err == -EFAULT && rw == VERIFY_READ) {
515                 err = get_user_pages_fast(address, 1, 0, &page);
516                 ro = 1;
517         }
518         if (err < 0)
519                 return err;
520         else
521                 err = 0;
522 
523         lock_page(page);
524         /*
525          * If page->mapping is NULL, then it cannot be a PageAnon
526          * page; but it might be the ZERO_PAGE or in the gate area or
527          * in a special mapping (all cases which we are happy to fail);
528          * or it may have been a good file page when get_user_pages_fast
529          * found it, but truncated or holepunched or subjected to
530          * invalidate_complete_page2 before we got the page lock (also
531          * cases which we are happy to fail).  And we hold a reference,
532          * so refcount care in invalidate_complete_page's remove_mapping
533          * prevents drop_caches from setting mapping to NULL beneath us.
534          *
535          * The case we do have to guard against is when memory pressure made
536          * shmem_writepage move it from filecache to swapcache beneath us:
537          * an unlikely race, but we do need to retry for page->mapping.
538          */
539         mapping = compound_head(page)->mapping;
540         if (!mapping) {
541                 int shmem_swizzled = PageSwapCache(page);
542                 unlock_page(page);
543                 put_page(page);
544                 if (shmem_swizzled)
545                         goto again;
546                 return -EFAULT;
547         }
548 
549         /*
550          * Private mappings are handled in a simple way.
551          *
552          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
553          * it's a read-only handle, it's expected that futexes attach to
554          * the object not the particular process.
555          */
556         if (PageAnon(page)) {
557                 /*
558                  * A RO anonymous page will never change and thus doesn't make
559                  * sense for futex operations.
560                  */
561                 if (unlikely(should_fail_futex(fshared)) || ro) {
562                         err = -EFAULT;
563                         goto out;
564                 }
565 
566                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
567                 key->private.mm = mm;
568                 key->private.address = address;
569         } else {
570                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
571                 key->shared.inode = mapping->host;
572                 key->shared.pgoff = basepage_index(page);
573         }
574 
575         get_futex_key_refs(key); /* implies MB (B) */
576 
577 out:
578         unlock_page(page);
579         put_page(page);
580         return err;
581 }
582 
583 static inline void put_futex_key(union futex_key *key)
584 {
585         drop_futex_key_refs(key);
586 }
587 
588 /**
589  * fault_in_user_writeable() - Fault in user address and verify RW access
590  * @uaddr:      pointer to faulting user space address
591  *
592  * Slow path to fixup the fault we just took in the atomic write
593  * access to @uaddr.
594  *
595  * We have no generic implementation of a non-destructive write to the
596  * user address. We know that we faulted in the atomic pagefault
597  * disabled section so we can as well avoid the #PF overhead by
598  * calling get_user_pages() right away.
599  */
600 static int fault_in_user_writeable(u32 __user *uaddr)
601 {
602         struct mm_struct *mm = current->mm;
603         int ret;
604 
605         down_read(&mm->mmap_sem);
606         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
607                                FAULT_FLAG_WRITE, NULL);
608         up_read(&mm->mmap_sem);
609 
610         return ret < 0 ? ret : 0;
611 }
612 
613 /**
614  * futex_top_waiter() - Return the highest priority waiter on a futex
615  * @hb:         the hash bucket the futex_q's reside in
616  * @key:        the futex key (to distinguish it from other futex futex_q's)
617  *
618  * Must be called with the hb lock held.
619  */
620 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
621                                         union futex_key *key)
622 {
623         struct futex_q *this;
624 
625         plist_for_each_entry(this, &hb->chain, list) {
626                 if (match_futex(&this->key, key))
627                         return this;
628         }
629         return NULL;
630 }
631 
632 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
633                                       u32 uval, u32 newval)
634 {
635         int ret;
636 
637         pagefault_disable();
638         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
639         pagefault_enable();
640 
641         return ret;
642 }
643 
644 static int get_futex_value_locked(u32 *dest, u32 __user *from)
645 {
646         int ret;
647 
648         pagefault_disable();
649         ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
650         pagefault_enable();
651 
652         return ret ? -EFAULT : 0;
653 }
654 
655 
656 /*
657  * PI code:
658  */
659 static int refill_pi_state_cache(void)
660 {
661         struct futex_pi_state *pi_state;
662 
663         if (likely(current->pi_state_cache))
664                 return 0;
665 
666         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
667 
668         if (!pi_state)
669                 return -ENOMEM;
670 
671         INIT_LIST_HEAD(&pi_state->list);
672         /* pi_mutex gets initialized later */
673         pi_state->owner = NULL;
674         atomic_set(&pi_state->refcount, 1);
675         pi_state->key = FUTEX_KEY_INIT;
676 
677         current->pi_state_cache = pi_state;
678 
679         return 0;
680 }
681 
682 static struct futex_pi_state * alloc_pi_state(void)
683 {
684         struct futex_pi_state *pi_state = current->pi_state_cache;
685 
686         WARN_ON(!pi_state);
687         current->pi_state_cache = NULL;
688 
689         return pi_state;
690 }
691 
692 /*
693  * Drops a reference to the pi_state object and frees or caches it
694  * when the last reference is gone.
695  *
696  * Must be called with the hb lock held.
697  */
698 static void put_pi_state(struct futex_pi_state *pi_state)
699 {
700         if (!pi_state)
701                 return;
702 
703         if (!atomic_dec_and_test(&pi_state->refcount))
704                 return;
705 
706         /*
707          * If pi_state->owner is NULL, the owner is most probably dying
708          * and has cleaned up the pi_state already
709          */
710         if (pi_state->owner) {
711                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
712                 list_del_init(&pi_state->list);
713                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
714 
715                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
716         }
717 
718         if (current->pi_state_cache)
719                 kfree(pi_state);
720         else {
721                 /*
722                  * pi_state->list is already empty.
723                  * clear pi_state->owner.
724                  * refcount is at 0 - put it back to 1.
725                  */
726                 pi_state->owner = NULL;
727                 atomic_set(&pi_state->refcount, 1);
728                 current->pi_state_cache = pi_state;
729         }
730 }
731 
732 /*
733  * Look up the task based on what TID userspace gave us.
734  * We dont trust it.
735  */
736 static struct task_struct * futex_find_get_task(pid_t pid)
737 {
738         struct task_struct *p;
739 
740         rcu_read_lock();
741         p = find_task_by_vpid(pid);
742         if (p)
743                 get_task_struct(p);
744 
745         rcu_read_unlock();
746 
747         return p;
748 }
749 
750 /*
751  * This task is holding PI mutexes at exit time => bad.
752  * Kernel cleans up PI-state, but userspace is likely hosed.
753  * (Robust-futex cleanup is separate and might save the day for userspace.)
754  */
755 void exit_pi_state_list(struct task_struct *curr)
756 {
757         struct list_head *next, *head = &curr->pi_state_list;
758         struct futex_pi_state *pi_state;
759         struct futex_hash_bucket *hb;
760         union futex_key key = FUTEX_KEY_INIT;
761 
762         if (!futex_cmpxchg_enabled)
763                 return;
764         /*
765          * We are a ZOMBIE and nobody can enqueue itself on
766          * pi_state_list anymore, but we have to be careful
767          * versus waiters unqueueing themselves:
768          */
769         raw_spin_lock_irq(&curr->pi_lock);
770         while (!list_empty(head)) {
771 
772                 next = head->next;
773                 pi_state = list_entry(next, struct futex_pi_state, list);
774                 key = pi_state->key;
775                 hb = hash_futex(&key);
776                 raw_spin_unlock_irq(&curr->pi_lock);
777 
778                 spin_lock(&hb->lock);
779 
780                 raw_spin_lock_irq(&curr->pi_lock);
781                 /*
782                  * We dropped the pi-lock, so re-check whether this
783                  * task still owns the PI-state:
784                  */
785                 if (head->next != next) {
786                         spin_unlock(&hb->lock);
787                         continue;
788                 }
789 
790                 WARN_ON(pi_state->owner != curr);
791                 WARN_ON(list_empty(&pi_state->list));
792                 list_del_init(&pi_state->list);
793                 pi_state->owner = NULL;
794                 raw_spin_unlock_irq(&curr->pi_lock);
795 
796                 rt_mutex_unlock(&pi_state->pi_mutex);
797 
798                 spin_unlock(&hb->lock);
799 
800                 raw_spin_lock_irq(&curr->pi_lock);
801         }
802         raw_spin_unlock_irq(&curr->pi_lock);
803 }
804 
805 /*
806  * We need to check the following states:
807  *
808  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
809  *
810  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
811  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
812  *
813  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
814  *
815  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
816  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
817  *
818  * [6]  Found  | Found    | task      | 0         | 1      | Valid
819  *
820  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
821  *
822  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
823  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
824  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
825  *
826  * [1]  Indicates that the kernel can acquire the futex atomically. We
827  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
828  *
829  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
830  *      thread is found then it indicates that the owner TID has died.
831  *
832  * [3]  Invalid. The waiter is queued on a non PI futex
833  *
834  * [4]  Valid state after exit_robust_list(), which sets the user space
835  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
836  *
837  * [5]  The user space value got manipulated between exit_robust_list()
838  *      and exit_pi_state_list()
839  *
840  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
841  *      the pi_state but cannot access the user space value.
842  *
843  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
844  *
845  * [8]  Owner and user space value match
846  *
847  * [9]  There is no transient state which sets the user space TID to 0
848  *      except exit_robust_list(), but this is indicated by the
849  *      FUTEX_OWNER_DIED bit. See [4]
850  *
851  * [10] There is no transient state which leaves owner and user space
852  *      TID out of sync.
853  */
854 
855 /*
856  * Validate that the existing waiter has a pi_state and sanity check
857  * the pi_state against the user space value. If correct, attach to
858  * it.
859  */
860 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
861                               struct futex_pi_state **ps)
862 {
863         pid_t pid = uval & FUTEX_TID_MASK;
864 
865         /*
866          * Userspace might have messed up non-PI and PI futexes [3]
867          */
868         if (unlikely(!pi_state))
869                 return -EINVAL;
870 
871         WARN_ON(!atomic_read(&pi_state->refcount));
872 
873         /*
874          * Handle the owner died case:
875          */
876         if (uval & FUTEX_OWNER_DIED) {
877                 /*
878                  * exit_pi_state_list sets owner to NULL and wakes the
879                  * topmost waiter. The task which acquires the
880                  * pi_state->rt_mutex will fixup owner.
881                  */
882                 if (!pi_state->owner) {
883                         /*
884                          * No pi state owner, but the user space TID
885                          * is not 0. Inconsistent state. [5]
886                          */
887                         if (pid)
888                                 return -EINVAL;
889                         /*
890                          * Take a ref on the state and return success. [4]
891                          */
892                         goto out_state;
893                 }
894 
895                 /*
896                  * If TID is 0, then either the dying owner has not
897                  * yet executed exit_pi_state_list() or some waiter
898                  * acquired the rtmutex in the pi state, but did not
899                  * yet fixup the TID in user space.
900                  *
901                  * Take a ref on the state and return success. [6]
902                  */
903                 if (!pid)
904                         goto out_state;
905         } else {
906                 /*
907                  * If the owner died bit is not set, then the pi_state
908                  * must have an owner. [7]
909                  */
910                 if (!pi_state->owner)
911                         return -EINVAL;
912         }
913 
914         /*
915          * Bail out if user space manipulated the futex value. If pi
916          * state exists then the owner TID must be the same as the
917          * user space TID. [9/10]
918          */
919         if (pid != task_pid_vnr(pi_state->owner))
920                 return -EINVAL;
921 out_state:
922         atomic_inc(&pi_state->refcount);
923         *ps = pi_state;
924         return 0;
925 }
926 
927 /*
928  * Lookup the task for the TID provided from user space and attach to
929  * it after doing proper sanity checks.
930  */
931 static int attach_to_pi_owner(u32 uval, union futex_key *key,
932                               struct futex_pi_state **ps)
933 {
934         pid_t pid = uval & FUTEX_TID_MASK;
935         struct futex_pi_state *pi_state;
936         struct task_struct *p;
937 
938         /*
939          * We are the first waiter - try to look up the real owner and attach
940          * the new pi_state to it, but bail out when TID = 0 [1]
941          */
942         if (!pid)
943                 return -ESRCH;
944         p = futex_find_get_task(pid);
945         if (!p)
946                 return -ESRCH;
947 
948         if (unlikely(p->flags & PF_KTHREAD)) {
949                 put_task_struct(p);
950                 return -EPERM;
951         }
952 
953         /*
954          * We need to look at the task state flags to figure out,
955          * whether the task is exiting. To protect against the do_exit
956          * change of the task flags, we do this protected by
957          * p->pi_lock:
958          */
959         raw_spin_lock_irq(&p->pi_lock);
960         if (unlikely(p->flags & PF_EXITING)) {
961                 /*
962                  * The task is on the way out. When PF_EXITPIDONE is
963                  * set, we know that the task has finished the
964                  * cleanup:
965                  */
966                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
967 
968                 raw_spin_unlock_irq(&p->pi_lock);
969                 put_task_struct(p);
970                 return ret;
971         }
972 
973         /*
974          * No existing pi state. First waiter. [2]
975          */
976         pi_state = alloc_pi_state();
977 
978         /*
979          * Initialize the pi_mutex in locked state and make @p
980          * the owner of it:
981          */
982         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
983 
984         /* Store the key for possible exit cleanups: */
985         pi_state->key = *key;
986 
987         WARN_ON(!list_empty(&pi_state->list));
988         list_add(&pi_state->list, &p->pi_state_list);
989         pi_state->owner = p;
990         raw_spin_unlock_irq(&p->pi_lock);
991 
992         put_task_struct(p);
993 
994         *ps = pi_state;
995 
996         return 0;
997 }
998 
999 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1000                            union futex_key *key, struct futex_pi_state **ps)
1001 {
1002         struct futex_q *match = futex_top_waiter(hb, key);
1003 
1004         /*
1005          * If there is a waiter on that futex, validate it and
1006          * attach to the pi_state when the validation succeeds.
1007          */
1008         if (match)
1009                 return attach_to_pi_state(uval, match->pi_state, ps);
1010 
1011         /*
1012          * We are the first waiter - try to look up the owner based on
1013          * @uval and attach to it.
1014          */
1015         return attach_to_pi_owner(uval, key, ps);
1016 }
1017 
1018 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1019 {
1020         u32 uninitialized_var(curval);
1021 
1022         if (unlikely(should_fail_futex(true)))
1023                 return -EFAULT;
1024 
1025         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1026                 return -EFAULT;
1027 
1028         /*If user space value changed, let the caller retry */
1029         return curval != uval ? -EAGAIN : 0;
1030 }
1031 
1032 /**
1033  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1034  * @uaddr:              the pi futex user address
1035  * @hb:                 the pi futex hash bucket
1036  * @key:                the futex key associated with uaddr and hb
1037  * @ps:                 the pi_state pointer where we store the result of the
1038  *                      lookup
1039  * @task:               the task to perform the atomic lock work for.  This will
1040  *                      be "current" except in the case of requeue pi.
1041  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1042  *
1043  * Return:
1044  *  0 - ready to wait;
1045  *  1 - acquired the lock;
1046  * <0 - error
1047  *
1048  * The hb->lock and futex_key refs shall be held by the caller.
1049  */
1050 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1051                                 union futex_key *key,
1052                                 struct futex_pi_state **ps,
1053                                 struct task_struct *task, int set_waiters)
1054 {
1055         u32 uval, newval, vpid = task_pid_vnr(task);
1056         struct futex_q *match;
1057         int ret;
1058 
1059         /*
1060          * Read the user space value first so we can validate a few
1061          * things before proceeding further.
1062          */
1063         if (get_futex_value_locked(&uval, uaddr))
1064                 return -EFAULT;
1065 
1066         if (unlikely(should_fail_futex(true)))
1067                 return -EFAULT;
1068 
1069         /*
1070          * Detect deadlocks.
1071          */
1072         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1073                 return -EDEADLK;
1074 
1075         if ((unlikely(should_fail_futex(true))))
1076                 return -EDEADLK;
1077 
1078         /*
1079          * Lookup existing state first. If it exists, try to attach to
1080          * its pi_state.
1081          */
1082         match = futex_top_waiter(hb, key);
1083         if (match)
1084                 return attach_to_pi_state(uval, match->pi_state, ps);
1085 
1086         /*
1087          * No waiter and user TID is 0. We are here because the
1088          * waiters or the owner died bit is set or called from
1089          * requeue_cmp_pi or for whatever reason something took the
1090          * syscall.
1091          */
1092         if (!(uval & FUTEX_TID_MASK)) {
1093                 /*
1094                  * We take over the futex. No other waiters and the user space
1095                  * TID is 0. We preserve the owner died bit.
1096                  */
1097                 newval = uval & FUTEX_OWNER_DIED;
1098                 newval |= vpid;
1099 
1100                 /* The futex requeue_pi code can enforce the waiters bit */
1101                 if (set_waiters)
1102                         newval |= FUTEX_WAITERS;
1103 
1104                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1105                 /* If the take over worked, return 1 */
1106                 return ret < 0 ? ret : 1;
1107         }
1108 
1109         /*
1110          * First waiter. Set the waiters bit before attaching ourself to
1111          * the owner. If owner tries to unlock, it will be forced into
1112          * the kernel and blocked on hb->lock.
1113          */
1114         newval = uval | FUTEX_WAITERS;
1115         ret = lock_pi_update_atomic(uaddr, uval, newval);
1116         if (ret)
1117                 return ret;
1118         /*
1119          * If the update of the user space value succeeded, we try to
1120          * attach to the owner. If that fails, no harm done, we only
1121          * set the FUTEX_WAITERS bit in the user space variable.
1122          */
1123         return attach_to_pi_owner(uval, key, ps);
1124 }
1125 
1126 /**
1127  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1128  * @q:  The futex_q to unqueue
1129  *
1130  * The q->lock_ptr must not be NULL and must be held by the caller.
1131  */
1132 static void __unqueue_futex(struct futex_q *q)
1133 {
1134         struct futex_hash_bucket *hb;
1135 
1136         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1137             || WARN_ON(plist_node_empty(&q->list)))
1138                 return;
1139 
1140         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1141         plist_del(&q->list, &hb->chain);
1142         hb_waiters_dec(hb);
1143 }
1144 
1145 /*
1146  * The hash bucket lock must be held when this is called.
1147  * Afterwards, the futex_q must not be accessed. Callers
1148  * must ensure to later call wake_up_q() for the actual
1149  * wakeups to occur.
1150  */
1151 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1152 {
1153         struct task_struct *p = q->task;
1154 
1155         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1156                 return;
1157 
1158         /*
1159          * Queue the task for later wakeup for after we've released
1160          * the hb->lock. wake_q_add() grabs reference to p.
1161          */
1162         wake_q_add(wake_q, p);
1163         __unqueue_futex(q);
1164         /*
1165          * The waiting task can free the futex_q as soon as
1166          * q->lock_ptr = NULL is written, without taking any locks. A
1167          * memory barrier is required here to prevent the following
1168          * store to lock_ptr from getting ahead of the plist_del.
1169          */
1170         smp_wmb();
1171         q->lock_ptr = NULL;
1172 }
1173 
1174 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1175                          struct futex_hash_bucket *hb)
1176 {
1177         struct task_struct *new_owner;
1178         struct futex_pi_state *pi_state = this->pi_state;
1179         u32 uninitialized_var(curval), newval;
1180         WAKE_Q(wake_q);
1181         bool deboost;
1182         int ret = 0;
1183 
1184         if (!pi_state)
1185                 return -EINVAL;
1186 
1187         /*
1188          * If current does not own the pi_state then the futex is
1189          * inconsistent and user space fiddled with the futex value.
1190          */
1191         if (pi_state->owner != current)
1192                 return -EINVAL;
1193 
1194         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1195         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1196 
1197         /*
1198          * It is possible that the next waiter (the one that brought
1199          * this owner to the kernel) timed out and is no longer
1200          * waiting on the lock.
1201          */
1202         if (!new_owner)
1203                 new_owner = this->task;
1204 
1205         /*
1206          * We pass it to the next owner. The WAITERS bit is always
1207          * kept enabled while there is PI state around. We cleanup the
1208          * owner died bit, because we are the owner.
1209          */
1210         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1211 
1212         if (unlikely(should_fail_futex(true)))
1213                 ret = -EFAULT;
1214 
1215         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1216                 ret = -EFAULT;
1217         else if (curval != uval)
1218                 ret = -EINVAL;
1219         if (ret) {
1220                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1221                 return ret;
1222         }
1223 
1224         raw_spin_lock(&pi_state->owner->pi_lock);
1225         WARN_ON(list_empty(&pi_state->list));
1226         list_del_init(&pi_state->list);
1227         raw_spin_unlock(&pi_state->owner->pi_lock);
1228 
1229         raw_spin_lock(&new_owner->pi_lock);
1230         WARN_ON(!list_empty(&pi_state->list));
1231         list_add(&pi_state->list, &new_owner->pi_state_list);
1232         pi_state->owner = new_owner;
1233         raw_spin_unlock(&new_owner->pi_lock);
1234 
1235         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1236 
1237         deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1238 
1239         /*
1240          * First unlock HB so the waiter does not spin on it once he got woken
1241          * up. Second wake up the waiter before the priority is adjusted. If we
1242          * deboost first (and lose our higher priority), then the task might get
1243          * scheduled away before the wake up can take place.
1244          */
1245         spin_unlock(&hb->lock);
1246         wake_up_q(&wake_q);
1247         if (deboost)
1248                 rt_mutex_adjust_prio(current);
1249 
1250         return 0;
1251 }
1252 
1253 /*
1254  * Express the locking dependencies for lockdep:
1255  */
1256 static inline void
1257 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1258 {
1259         if (hb1 <= hb2) {
1260                 spin_lock(&hb1->lock);
1261                 if (hb1 < hb2)
1262                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1263         } else { /* hb1 > hb2 */
1264                 spin_lock(&hb2->lock);
1265                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1266         }
1267 }
1268 
1269 static inline void
1270 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1271 {
1272         spin_unlock(&hb1->lock);
1273         if (hb1 != hb2)
1274                 spin_unlock(&hb2->lock);
1275 }
1276 
1277 /*
1278  * Wake up waiters matching bitset queued on this futex (uaddr).
1279  */
1280 static int
1281 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1282 {
1283         struct futex_hash_bucket *hb;
1284         struct futex_q *this, *next;
1285         union futex_key key = FUTEX_KEY_INIT;
1286         int ret;
1287         WAKE_Q(wake_q);
1288 
1289         if (!bitset)
1290                 return -EINVAL;
1291 
1292         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1293         if (unlikely(ret != 0))
1294                 goto out;
1295 
1296         hb = hash_futex(&key);
1297 
1298         /* Make sure we really have tasks to wakeup */
1299         if (!hb_waiters_pending(hb))
1300                 goto out_put_key;
1301 
1302         spin_lock(&hb->lock);
1303 
1304         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1305                 if (match_futex (&this->key, &key)) {
1306                         if (this->pi_state || this->rt_waiter) {
1307                                 ret = -EINVAL;
1308                                 break;
1309                         }
1310 
1311                         /* Check if one of the bits is set in both bitsets */
1312                         if (!(this->bitset & bitset))
1313                                 continue;
1314 
1315                         mark_wake_futex(&wake_q, this);
1316                         if (++ret >= nr_wake)
1317                                 break;
1318                 }
1319         }
1320 
1321         spin_unlock(&hb->lock);
1322         wake_up_q(&wake_q);
1323 out_put_key:
1324         put_futex_key(&key);
1325 out:
1326         return ret;
1327 }
1328 
1329 /*
1330  * Wake up all waiters hashed on the physical page that is mapped
1331  * to this virtual address:
1332  */
1333 static int
1334 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1335               int nr_wake, int nr_wake2, int op)
1336 {
1337         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1338         struct futex_hash_bucket *hb1, *hb2;
1339         struct futex_q *this, *next;
1340         int ret, op_ret;
1341         WAKE_Q(wake_q);
1342 
1343 retry:
1344         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1345         if (unlikely(ret != 0))
1346                 goto out;
1347         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1348         if (unlikely(ret != 0))
1349                 goto out_put_key1;
1350 
1351         hb1 = hash_futex(&key1);
1352         hb2 = hash_futex(&key2);
1353 
1354 retry_private:
1355         double_lock_hb(hb1, hb2);
1356         op_ret = futex_atomic_op_inuser(op, uaddr2);
1357         if (unlikely(op_ret < 0)) {
1358 
1359                 double_unlock_hb(hb1, hb2);
1360 
1361 #ifndef CONFIG_MMU
1362                 /*
1363                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1364                  * but we might get them from range checking
1365                  */
1366                 ret = op_ret;
1367                 goto out_put_keys;
1368 #endif
1369 
1370                 if (unlikely(op_ret != -EFAULT)) {
1371                         ret = op_ret;
1372                         goto out_put_keys;
1373                 }
1374 
1375                 ret = fault_in_user_writeable(uaddr2);
1376                 if (ret)
1377                         goto out_put_keys;
1378 
1379                 if (!(flags & FLAGS_SHARED))
1380                         goto retry_private;
1381 
1382                 put_futex_key(&key2);
1383                 put_futex_key(&key1);
1384                 goto retry;
1385         }
1386 
1387         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1388                 if (match_futex (&this->key, &key1)) {
1389                         if (this->pi_state || this->rt_waiter) {
1390                                 ret = -EINVAL;
1391                                 goto out_unlock;
1392                         }
1393                         mark_wake_futex(&wake_q, this);
1394                         if (++ret >= nr_wake)
1395                                 break;
1396                 }
1397         }
1398 
1399         if (op_ret > 0) {
1400                 op_ret = 0;
1401                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1402                         if (match_futex (&this->key, &key2)) {
1403                                 if (this->pi_state || this->rt_waiter) {
1404                                         ret = -EINVAL;
1405                                         goto out_unlock;
1406                                 }
1407                                 mark_wake_futex(&wake_q, this);
1408                                 if (++op_ret >= nr_wake2)
1409                                         break;
1410                         }
1411                 }
1412                 ret += op_ret;
1413         }
1414 
1415 out_unlock:
1416         double_unlock_hb(hb1, hb2);
1417         wake_up_q(&wake_q);
1418 out_put_keys:
1419         put_futex_key(&key2);
1420 out_put_key1:
1421         put_futex_key(&key1);
1422 out:
1423         return ret;
1424 }
1425 
1426 /**
1427  * requeue_futex() - Requeue a futex_q from one hb to another
1428  * @q:          the futex_q to requeue
1429  * @hb1:        the source hash_bucket
1430  * @hb2:        the target hash_bucket
1431  * @key2:       the new key for the requeued futex_q
1432  */
1433 static inline
1434 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1435                    struct futex_hash_bucket *hb2, union futex_key *key2)
1436 {
1437 
1438         /*
1439          * If key1 and key2 hash to the same bucket, no need to
1440          * requeue.
1441          */
1442         if (likely(&hb1->chain != &hb2->chain)) {
1443                 plist_del(&q->list, &hb1->chain);
1444                 hb_waiters_dec(hb1);
1445                 plist_add(&q->list, &hb2->chain);
1446                 hb_waiters_inc(hb2);
1447                 q->lock_ptr = &hb2->lock;
1448         }
1449         get_futex_key_refs(key2);
1450         q->key = *key2;
1451 }
1452 
1453 /**
1454  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1455  * @q:          the futex_q
1456  * @key:        the key of the requeue target futex
1457  * @hb:         the hash_bucket of the requeue target futex
1458  *
1459  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1460  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1461  * to the requeue target futex so the waiter can detect the wakeup on the right
1462  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1463  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1464  * to protect access to the pi_state to fixup the owner later.  Must be called
1465  * with both q->lock_ptr and hb->lock held.
1466  */
1467 static inline
1468 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1469                            struct futex_hash_bucket *hb)
1470 {
1471         get_futex_key_refs(key);
1472         q->key = *key;
1473 
1474         __unqueue_futex(q);
1475 
1476         WARN_ON(!q->rt_waiter);
1477         q->rt_waiter = NULL;
1478 
1479         q->lock_ptr = &hb->lock;
1480 
1481         wake_up_state(q->task, TASK_NORMAL);
1482 }
1483 
1484 /**
1485  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1486  * @pifutex:            the user address of the to futex
1487  * @hb1:                the from futex hash bucket, must be locked by the caller
1488  * @hb2:                the to futex hash bucket, must be locked by the caller
1489  * @key1:               the from futex key
1490  * @key2:               the to futex key
1491  * @ps:                 address to store the pi_state pointer
1492  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1493  *
1494  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1495  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1496  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1497  * hb1 and hb2 must be held by the caller.
1498  *
1499  * Return:
1500  *  0 - failed to acquire the lock atomically;
1501  * >0 - acquired the lock, return value is vpid of the top_waiter
1502  * <0 - error
1503  */
1504 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1505                                  struct futex_hash_bucket *hb1,
1506                                  struct futex_hash_bucket *hb2,
1507                                  union futex_key *key1, union futex_key *key2,
1508                                  struct futex_pi_state **ps, int set_waiters)
1509 {
1510         struct futex_q *top_waiter = NULL;
1511         u32 curval;
1512         int ret, vpid;
1513 
1514         if (get_futex_value_locked(&curval, pifutex))
1515                 return -EFAULT;
1516 
1517         if (unlikely(should_fail_futex(true)))
1518                 return -EFAULT;
1519 
1520         /*
1521          * Find the top_waiter and determine if there are additional waiters.
1522          * If the caller intends to requeue more than 1 waiter to pifutex,
1523          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1524          * as we have means to handle the possible fault.  If not, don't set
1525          * the bit unecessarily as it will force the subsequent unlock to enter
1526          * the kernel.
1527          */
1528         top_waiter = futex_top_waiter(hb1, key1);
1529 
1530         /* There are no waiters, nothing for us to do. */
1531         if (!top_waiter)
1532                 return 0;
1533 
1534         /* Ensure we requeue to the expected futex. */
1535         if (!match_futex(top_waiter->requeue_pi_key, key2))
1536                 return -EINVAL;
1537 
1538         /*
1539          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1540          * the contended case or if set_waiters is 1.  The pi_state is returned
1541          * in ps in contended cases.
1542          */
1543         vpid = task_pid_vnr(top_waiter->task);
1544         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1545                                    set_waiters);
1546         if (ret == 1) {
1547                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1548                 return vpid;
1549         }
1550         return ret;
1551 }
1552 
1553 /**
1554  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1555  * @uaddr1:     source futex user address
1556  * @flags:      futex flags (FLAGS_SHARED, etc.)
1557  * @uaddr2:     target futex user address
1558  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1559  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1560  * @cmpval:     @uaddr1 expected value (or %NULL)
1561  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1562  *              pi futex (pi to pi requeue is not supported)
1563  *
1564  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1565  * uaddr2 atomically on behalf of the top waiter.
1566  *
1567  * Return:
1568  * >=0 - on success, the number of tasks requeued or woken;
1569  *  <0 - on error
1570  */
1571 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1572                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1573                          u32 *cmpval, int requeue_pi)
1574 {
1575         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1576         int drop_count = 0, task_count = 0, ret;
1577         struct futex_pi_state *pi_state = NULL;
1578         struct futex_hash_bucket *hb1, *hb2;
1579         struct futex_q *this, *next;
1580         WAKE_Q(wake_q);
1581 
1582         if (requeue_pi) {
1583                 /*
1584                  * Requeue PI only works on two distinct uaddrs. This
1585                  * check is only valid for private futexes. See below.
1586                  */
1587                 if (uaddr1 == uaddr2)
1588                         return -EINVAL;
1589 
1590                 /*
1591                  * requeue_pi requires a pi_state, try to allocate it now
1592                  * without any locks in case it fails.
1593                  */
1594                 if (refill_pi_state_cache())
1595                         return -ENOMEM;
1596                 /*
1597                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1598                  * + nr_requeue, since it acquires the rt_mutex prior to
1599                  * returning to userspace, so as to not leave the rt_mutex with
1600                  * waiters and no owner.  However, second and third wake-ups
1601                  * cannot be predicted as they involve race conditions with the
1602                  * first wake and a fault while looking up the pi_state.  Both
1603                  * pthread_cond_signal() and pthread_cond_broadcast() should
1604                  * use nr_wake=1.
1605                  */
1606                 if (nr_wake != 1)
1607                         return -EINVAL;
1608         }
1609 
1610 retry:
1611         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1612         if (unlikely(ret != 0))
1613                 goto out;
1614         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1615                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1616         if (unlikely(ret != 0))
1617                 goto out_put_key1;
1618 
1619         /*
1620          * The check above which compares uaddrs is not sufficient for
1621          * shared futexes. We need to compare the keys:
1622          */
1623         if (requeue_pi && match_futex(&key1, &key2)) {
1624                 ret = -EINVAL;
1625                 goto out_put_keys;
1626         }
1627 
1628         hb1 = hash_futex(&key1);
1629         hb2 = hash_futex(&key2);
1630 
1631 retry_private:
1632         hb_waiters_inc(hb2);
1633         double_lock_hb(hb1, hb2);
1634 
1635         if (likely(cmpval != NULL)) {
1636                 u32 curval;
1637 
1638                 ret = get_futex_value_locked(&curval, uaddr1);
1639 
1640                 if (unlikely(ret)) {
1641                         double_unlock_hb(hb1, hb2);
1642                         hb_waiters_dec(hb2);
1643 
1644                         ret = get_user(curval, uaddr1);
1645                         if (ret)
1646                                 goto out_put_keys;
1647 
1648                         if (!(flags & FLAGS_SHARED))
1649                                 goto retry_private;
1650 
1651                         put_futex_key(&key2);
1652                         put_futex_key(&key1);
1653                         goto retry;
1654                 }
1655                 if (curval != *cmpval) {
1656                         ret = -EAGAIN;
1657                         goto out_unlock;
1658                 }
1659         }
1660 
1661         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1662                 /*
1663                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1664                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1665                  * bit.  We force this here where we are able to easily handle
1666                  * faults rather in the requeue loop below.
1667                  */
1668                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1669                                                  &key2, &pi_state, nr_requeue);
1670 
1671                 /*
1672                  * At this point the top_waiter has either taken uaddr2 or is
1673                  * waiting on it.  If the former, then the pi_state will not
1674                  * exist yet, look it up one more time to ensure we have a
1675                  * reference to it. If the lock was taken, ret contains the
1676                  * vpid of the top waiter task.
1677                  * If the lock was not taken, we have pi_state and an initial
1678                  * refcount on it. In case of an error we have nothing.
1679                  */
1680                 if (ret > 0) {
1681                         WARN_ON(pi_state);
1682                         drop_count++;
1683                         task_count++;
1684                         /*
1685                          * If we acquired the lock, then the user space value
1686                          * of uaddr2 should be vpid. It cannot be changed by
1687                          * the top waiter as it is blocked on hb2 lock if it
1688                          * tries to do so. If something fiddled with it behind
1689                          * our back the pi state lookup might unearth it. So
1690                          * we rather use the known value than rereading and
1691                          * handing potential crap to lookup_pi_state.
1692                          *
1693                          * If that call succeeds then we have pi_state and an
1694                          * initial refcount on it.
1695                          */
1696                         ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1697                 }
1698 
1699                 switch (ret) {
1700                 case 0:
1701                         /* We hold a reference on the pi state. */
1702                         break;
1703 
1704                         /* If the above failed, then pi_state is NULL */
1705                 case -EFAULT:
1706                         double_unlock_hb(hb1, hb2);
1707                         hb_waiters_dec(hb2);
1708                         put_futex_key(&key2);
1709                         put_futex_key(&key1);
1710                         ret = fault_in_user_writeable(uaddr2);
1711                         if (!ret)
1712                                 goto retry;
1713                         goto out;
1714                 case -EAGAIN:
1715                         /*
1716                          * Two reasons for this:
1717                          * - Owner is exiting and we just wait for the
1718                          *   exit to complete.
1719                          * - The user space value changed.
1720                          */
1721                         double_unlock_hb(hb1, hb2);
1722                         hb_waiters_dec(hb2);
1723                         put_futex_key(&key2);
1724                         put_futex_key(&key1);
1725                         cond_resched();
1726                         goto retry;
1727                 default:
1728                         goto out_unlock;
1729                 }
1730         }
1731 
1732         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1733                 if (task_count - nr_wake >= nr_requeue)
1734                         break;
1735 
1736                 if (!match_futex(&this->key, &key1))
1737                         continue;
1738 
1739                 /*
1740                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1741                  * be paired with each other and no other futex ops.
1742                  *
1743                  * We should never be requeueing a futex_q with a pi_state,
1744                  * which is awaiting a futex_unlock_pi().
1745                  */
1746                 if ((requeue_pi && !this->rt_waiter) ||
1747                     (!requeue_pi && this->rt_waiter) ||
1748                     this->pi_state) {
1749                         ret = -EINVAL;
1750                         break;
1751                 }
1752 
1753                 /*
1754                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1755                  * lock, we already woke the top_waiter.  If not, it will be
1756                  * woken by futex_unlock_pi().
1757                  */
1758                 if (++task_count <= nr_wake && !requeue_pi) {
1759                         mark_wake_futex(&wake_q, this);
1760                         continue;
1761                 }
1762 
1763                 /* Ensure we requeue to the expected futex for requeue_pi. */
1764                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1765                         ret = -EINVAL;
1766                         break;
1767                 }
1768 
1769                 /*
1770                  * Requeue nr_requeue waiters and possibly one more in the case
1771                  * of requeue_pi if we couldn't acquire the lock atomically.
1772                  */
1773                 if (requeue_pi) {
1774                         /*
1775                          * Prepare the waiter to take the rt_mutex. Take a
1776                          * refcount on the pi_state and store the pointer in
1777                          * the futex_q object of the waiter.
1778                          */
1779                         atomic_inc(&pi_state->refcount);
1780                         this->pi_state = pi_state;
1781                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1782                                                         this->rt_waiter,
1783                                                         this->task);
1784                         if (ret == 1) {
1785                                 /*
1786                                  * We got the lock. We do neither drop the
1787                                  * refcount on pi_state nor clear
1788                                  * this->pi_state because the waiter needs the
1789                                  * pi_state for cleaning up the user space
1790                                  * value. It will drop the refcount after
1791                                  * doing so.
1792                                  */
1793                                 requeue_pi_wake_futex(this, &key2, hb2);
1794                                 drop_count++;
1795                                 continue;
1796                         } else if (ret) {
1797                                 /*
1798                                  * rt_mutex_start_proxy_lock() detected a
1799                                  * potential deadlock when we tried to queue
1800                                  * that waiter. Drop the pi_state reference
1801                                  * which we took above and remove the pointer
1802                                  * to the state from the waiters futex_q
1803                                  * object.
1804                                  */
1805                                 this->pi_state = NULL;
1806                                 put_pi_state(pi_state);
1807                                 /*
1808                                  * We stop queueing more waiters and let user
1809                                  * space deal with the mess.
1810                                  */
1811                                 break;
1812                         }
1813                 }
1814                 requeue_futex(this, hb1, hb2, &key2);
1815                 drop_count++;
1816         }
1817 
1818         /*
1819          * We took an extra initial reference to the pi_state either
1820          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
1821          * need to drop it here again.
1822          */
1823         put_pi_state(pi_state);
1824 
1825 out_unlock:
1826         double_unlock_hb(hb1, hb2);
1827         wake_up_q(&wake_q);
1828         hb_waiters_dec(hb2);
1829 
1830         /*
1831          * drop_futex_key_refs() must be called outside the spinlocks. During
1832          * the requeue we moved futex_q's from the hash bucket at key1 to the
1833          * one at key2 and updated their key pointer.  We no longer need to
1834          * hold the references to key1.
1835          */
1836         while (--drop_count >= 0)
1837                 drop_futex_key_refs(&key1);
1838 
1839 out_put_keys:
1840         put_futex_key(&key2);
1841 out_put_key1:
1842         put_futex_key(&key1);
1843 out:
1844         return ret ? ret : task_count;
1845 }
1846 
1847 /* The key must be already stored in q->key. */
1848 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1849         __acquires(&hb->lock)
1850 {
1851         struct futex_hash_bucket *hb;
1852 
1853         hb = hash_futex(&q->key);
1854 
1855         /*
1856          * Increment the counter before taking the lock so that
1857          * a potential waker won't miss a to-be-slept task that is
1858          * waiting for the spinlock. This is safe as all queue_lock()
1859          * users end up calling queue_me(). Similarly, for housekeeping,
1860          * decrement the counter at queue_unlock() when some error has
1861          * occurred and we don't end up adding the task to the list.
1862          */
1863         hb_waiters_inc(hb);
1864 
1865         q->lock_ptr = &hb->lock;
1866 
1867         spin_lock(&hb->lock); /* implies MB (A) */
1868         return hb;
1869 }
1870 
1871 static inline void
1872 queue_unlock(struct futex_hash_bucket *hb)
1873         __releases(&hb->lock)
1874 {
1875         spin_unlock(&hb->lock);
1876         hb_waiters_dec(hb);
1877 }
1878 
1879 /**
1880  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1881  * @q:  The futex_q to enqueue
1882  * @hb: The destination hash bucket
1883  *
1884  * The hb->lock must be held by the caller, and is released here. A call to
1885  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1886  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1887  * or nothing if the unqueue is done as part of the wake process and the unqueue
1888  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1889  * an example).
1890  */
1891 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1892         __releases(&hb->lock)
1893 {
1894         int prio;
1895 
1896         /*
1897          * The priority used to register this element is
1898          * - either the real thread-priority for the real-time threads
1899          * (i.e. threads with a priority lower than MAX_RT_PRIO)
1900          * - or MAX_RT_PRIO for non-RT threads.
1901          * Thus, all RT-threads are woken first in priority order, and
1902          * the others are woken last, in FIFO order.
1903          */
1904         prio = min(current->normal_prio, MAX_RT_PRIO);
1905 
1906         plist_node_init(&q->list, prio);
1907         plist_add(&q->list, &hb->chain);
1908         q->task = current;
1909         spin_unlock(&hb->lock);
1910 }
1911 
1912 /**
1913  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1914  * @q:  The futex_q to unqueue
1915  *
1916  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1917  * be paired with exactly one earlier call to queue_me().
1918  *
1919  * Return:
1920  *   1 - if the futex_q was still queued (and we removed unqueued it);
1921  *   0 - if the futex_q was already removed by the waking thread
1922  */
1923 static int unqueue_me(struct futex_q *q)
1924 {
1925         spinlock_t *lock_ptr;
1926         int ret = 0;
1927 
1928         /* In the common case we don't take the spinlock, which is nice. */
1929 retry:
1930         lock_ptr = q->lock_ptr;
1931         barrier();
1932         if (lock_ptr != NULL) {
1933                 spin_lock(lock_ptr);
1934                 /*
1935                  * q->lock_ptr can change between reading it and
1936                  * spin_lock(), causing us to take the wrong lock.  This
1937                  * corrects the race condition.
1938                  *
1939                  * Reasoning goes like this: if we have the wrong lock,
1940                  * q->lock_ptr must have changed (maybe several times)
1941                  * between reading it and the spin_lock().  It can
1942                  * change again after the spin_lock() but only if it was
1943                  * already changed before the spin_lock().  It cannot,
1944                  * however, change back to the original value.  Therefore
1945                  * we can detect whether we acquired the correct lock.
1946                  */
1947                 if (unlikely(lock_ptr != q->lock_ptr)) {
1948                         spin_unlock(lock_ptr);
1949                         goto retry;
1950                 }
1951                 __unqueue_futex(q);
1952 
1953                 BUG_ON(q->pi_state);
1954 
1955                 spin_unlock(lock_ptr);
1956                 ret = 1;
1957         }
1958 
1959         drop_futex_key_refs(&q->key);
1960         return ret;
1961 }
1962 
1963 /*
1964  * PI futexes can not be requeued and must remove themself from the
1965  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1966  * and dropped here.
1967  */
1968 static void unqueue_me_pi(struct futex_q *q)
1969         __releases(q->lock_ptr)
1970 {
1971         __unqueue_futex(q);
1972 
1973         BUG_ON(!q->pi_state);
1974         put_pi_state(q->pi_state);
1975         q->pi_state = NULL;
1976 
1977         spin_unlock(q->lock_ptr);
1978 }
1979 
1980 /*
1981  * Fixup the pi_state owner with the new owner.
1982  *
1983  * Must be called with hash bucket lock held and mm->sem held for non
1984  * private futexes.
1985  */
1986 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1987                                 struct task_struct *newowner)
1988 {
1989         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1990         struct futex_pi_state *pi_state = q->pi_state;
1991         struct task_struct *oldowner = pi_state->owner;
1992         u32 uval, uninitialized_var(curval), newval;
1993         int ret;
1994 
1995         /* Owner died? */
1996         if (!pi_state->owner)
1997                 newtid |= FUTEX_OWNER_DIED;
1998 
1999         /*
2000          * We are here either because we stole the rtmutex from the
2001          * previous highest priority waiter or we are the highest priority
2002          * waiter but failed to get the rtmutex the first time.
2003          * We have to replace the newowner TID in the user space variable.
2004          * This must be atomic as we have to preserve the owner died bit here.
2005          *
2006          * Note: We write the user space value _before_ changing the pi_state
2007          * because we can fault here. Imagine swapped out pages or a fork
2008          * that marked all the anonymous memory readonly for cow.
2009          *
2010          * Modifying pi_state _before_ the user space value would
2011          * leave the pi_state in an inconsistent state when we fault
2012          * here, because we need to drop the hash bucket lock to
2013          * handle the fault. This might be observed in the PID check
2014          * in lookup_pi_state.
2015          */
2016 retry:
2017         if (get_futex_value_locked(&uval, uaddr))
2018                 goto handle_fault;
2019 
2020         while (1) {
2021                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2022 
2023                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2024                         goto handle_fault;
2025                 if (curval == uval)
2026                         break;
2027                 uval = curval;
2028         }
2029 
2030         /*
2031          * We fixed up user space. Now we need to fix the pi_state
2032          * itself.
2033          */
2034         if (pi_state->owner != NULL) {
2035                 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2036                 WARN_ON(list_empty(&pi_state->list));
2037                 list_del_init(&pi_state->list);
2038                 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2039         }
2040 
2041         pi_state->owner = newowner;
2042 
2043         raw_spin_lock_irq(&newowner->pi_lock);
2044         WARN_ON(!list_empty(&pi_state->list));
2045         list_add(&pi_state->list, &newowner->pi_state_list);
2046         raw_spin_unlock_irq(&newowner->pi_lock);
2047         return 0;
2048 
2049         /*
2050          * To handle the page fault we need to drop the hash bucket
2051          * lock here. That gives the other task (either the highest priority
2052          * waiter itself or the task which stole the rtmutex) the
2053          * chance to try the fixup of the pi_state. So once we are
2054          * back from handling the fault we need to check the pi_state
2055          * after reacquiring the hash bucket lock and before trying to
2056          * do another fixup. When the fixup has been done already we
2057          * simply return.
2058          */
2059 handle_fault:
2060         spin_unlock(q->lock_ptr);
2061 
2062         ret = fault_in_user_writeable(uaddr);
2063 
2064         spin_lock(q->lock_ptr);
2065 
2066         /*
2067          * Check if someone else fixed it for us:
2068          */
2069         if (pi_state->owner != oldowner)
2070                 return 0;
2071 
2072         if (ret)
2073                 return ret;
2074 
2075         goto retry;
2076 }
2077 
2078 static long futex_wait_restart(struct restart_block *restart);
2079 
2080 /**
2081  * fixup_owner() - Post lock pi_state and corner case management
2082  * @uaddr:      user address of the futex
2083  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2084  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2085  *
2086  * After attempting to lock an rt_mutex, this function is called to cleanup
2087  * the pi_state owner as well as handle race conditions that may allow us to
2088  * acquire the lock. Must be called with the hb lock held.
2089  *
2090  * Return:
2091  *  1 - success, lock taken;
2092  *  0 - success, lock not taken;
2093  * <0 - on error (-EFAULT)
2094  */
2095 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2096 {
2097         struct task_struct *owner;
2098         int ret = 0;
2099 
2100         if (locked) {
2101                 /*
2102                  * Got the lock. We might not be the anticipated owner if we
2103                  * did a lock-steal - fix up the PI-state in that case:
2104                  */
2105                 if (q->pi_state->owner != current)
2106                         ret = fixup_pi_state_owner(uaddr, q, current);
2107                 goto out;
2108         }
2109 
2110         /*
2111          * Catch the rare case, where the lock was released when we were on the
2112          * way back before we locked the hash bucket.
2113          */
2114         if (q->pi_state->owner == current) {
2115                 /*
2116                  * Try to get the rt_mutex now. This might fail as some other
2117                  * task acquired the rt_mutex after we removed ourself from the
2118                  * rt_mutex waiters list.
2119                  */
2120                 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2121                         locked = 1;
2122                         goto out;
2123                 }
2124 
2125                 /*
2126                  * pi_state is incorrect, some other task did a lock steal and
2127                  * we returned due to timeout or signal without taking the
2128                  * rt_mutex. Too late.
2129                  */
2130                 raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2131                 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2132                 if (!owner)
2133                         owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2134                 raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2135                 ret = fixup_pi_state_owner(uaddr, q, owner);
2136                 goto out;
2137         }
2138 
2139         /*
2140          * Paranoia check. If we did not take the lock, then we should not be
2141          * the owner of the rt_mutex.
2142          */
2143         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2144                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2145                                 "pi-state %p\n", ret,
2146                                 q->pi_state->pi_mutex.owner,
2147                                 q->pi_state->owner);
2148 
2149 out:
2150         return ret ? ret : locked;
2151 }
2152 
2153 /**
2154  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2155  * @hb:         the futex hash bucket, must be locked by the caller
2156  * @q:          the futex_q to queue up on
2157  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2158  */
2159 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2160                                 struct hrtimer_sleeper *timeout)
2161 {
2162         /*
2163          * The task state is guaranteed to be set before another task can
2164          * wake it. set_current_state() is implemented using smp_store_mb() and
2165          * queue_me() calls spin_unlock() upon completion, both serializing
2166          * access to the hash list and forcing another memory barrier.
2167          */
2168         set_current_state(TASK_INTERRUPTIBLE);
2169         queue_me(q, hb);
2170 
2171         /* Arm the timer */
2172         if (timeout)
2173                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2174 
2175         /*
2176          * If we have been removed from the hash list, then another task
2177          * has tried to wake us, and we can skip the call to schedule().
2178          */
2179         if (likely(!plist_node_empty(&q->list))) {
2180                 /*
2181                  * If the timer has already expired, current will already be
2182                  * flagged for rescheduling. Only call schedule if there
2183                  * is no timeout, or if it has yet to expire.
2184                  */
2185                 if (!timeout || timeout->task)
2186                         freezable_schedule();
2187         }
2188         __set_current_state(TASK_RUNNING);
2189 }
2190 
2191 /**
2192  * futex_wait_setup() - Prepare to wait on a futex
2193  * @uaddr:      the futex userspace address
2194  * @val:        the expected value
2195  * @flags:      futex flags (FLAGS_SHARED, etc.)
2196  * @q:          the associated futex_q
2197  * @hb:         storage for hash_bucket pointer to be returned to caller
2198  *
2199  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2200  * compare it with the expected value.  Handle atomic faults internally.
2201  * Return with the hb lock held and a q.key reference on success, and unlocked
2202  * with no q.key reference on failure.
2203  *
2204  * Return:
2205  *  0 - uaddr contains val and hb has been locked;
2206  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2207  */
2208 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2209                            struct futex_q *q, struct futex_hash_bucket **hb)
2210 {
2211         u32 uval;
2212         int ret;
2213 
2214         /*
2215          * Access the page AFTER the hash-bucket is locked.
2216          * Order is important:
2217          *
2218          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2219          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2220          *
2221          * The basic logical guarantee of a futex is that it blocks ONLY
2222          * if cond(var) is known to be true at the time of blocking, for
2223          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2224          * would open a race condition where we could block indefinitely with
2225          * cond(var) false, which would violate the guarantee.
2226          *
2227          * On the other hand, we insert q and release the hash-bucket only
2228          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2229          * absorb a wakeup if *uaddr does not match the desired values
2230          * while the syscall executes.
2231          */
2232 retry:
2233         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2234         if (unlikely(ret != 0))
2235                 return ret;
2236 
2237 retry_private:
2238         *hb = queue_lock(q);
2239 
2240         ret = get_futex_value_locked(&uval, uaddr);
2241 
2242         if (ret) {
2243                 queue_unlock(*hb);
2244 
2245                 ret = get_user(uval, uaddr);
2246                 if (ret)
2247                         goto out;
2248 
2249                 if (!(flags & FLAGS_SHARED))
2250                         goto retry_private;
2251 
2252                 put_futex_key(&q->key);
2253                 goto retry;
2254         }
2255 
2256         if (uval != val) {
2257                 queue_unlock(*hb);
2258                 ret = -EWOULDBLOCK;
2259         }
2260 
2261 out:
2262         if (ret)
2263                 put_futex_key(&q->key);
2264         return ret;
2265 }
2266 
2267 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2268                       ktime_t *abs_time, u32 bitset)
2269 {
2270         struct hrtimer_sleeper timeout, *to = NULL;
2271         struct restart_block *restart;
2272         struct futex_hash_bucket *hb;
2273         struct futex_q q = futex_q_init;
2274         int ret;
2275 
2276         if (!bitset)
2277                 return -EINVAL;
2278         q.bitset = bitset;
2279 
2280         if (abs_time) {
2281                 to = &timeout;
2282 
2283                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2284                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2285                                       HRTIMER_MODE_ABS);
2286                 hrtimer_init_sleeper(to, current);
2287                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2288                                              current->timer_slack_ns);
2289         }
2290 
2291 retry:
2292         /*
2293          * Prepare to wait on uaddr. On success, holds hb lock and increments
2294          * q.key refs.
2295          */
2296         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2297         if (ret)
2298                 goto out;
2299 
2300         /* queue_me and wait for wakeup, timeout, or a signal. */
2301         futex_wait_queue_me(hb, &q, to);
2302 
2303         /* If we were woken (and unqueued), we succeeded, whatever. */
2304         ret = 0;
2305         /* unqueue_me() drops q.key ref */
2306         if (!unqueue_me(&q))
2307                 goto out;
2308         ret = -ETIMEDOUT;
2309         if (to && !to->task)
2310                 goto out;
2311 
2312         /*
2313          * We expect signal_pending(current), but we might be the
2314          * victim of a spurious wakeup as well.
2315          */
2316         if (!signal_pending(current))
2317                 goto retry;
2318 
2319         ret = -ERESTARTSYS;
2320         if (!abs_time)
2321                 goto out;
2322 
2323         restart = &current->restart_block;
2324         restart->fn = futex_wait_restart;
2325         restart->futex.uaddr = uaddr;
2326         restart->futex.val = val;
2327         restart->futex.time = abs_time->tv64;
2328         restart->futex.bitset = bitset;
2329         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2330 
2331         ret = -ERESTART_RESTARTBLOCK;
2332 
2333 out:
2334         if (to) {
2335                 hrtimer_cancel(&to->timer);
2336                 destroy_hrtimer_on_stack(&to->timer);
2337         }
2338         return ret;
2339 }
2340 
2341 
2342 static long futex_wait_restart(struct restart_block *restart)
2343 {
2344         u32 __user *uaddr = restart->futex.uaddr;
2345         ktime_t t, *tp = NULL;
2346 
2347         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2348                 t.tv64 = restart->futex.time;
2349                 tp = &t;
2350         }
2351         restart->fn = do_no_restart_syscall;
2352 
2353         return (long)futex_wait(uaddr, restart->futex.flags,
2354                                 restart->futex.val, tp, restart->futex.bitset);
2355 }
2356 
2357 
2358 /*
2359  * Userspace tried a 0 -> TID atomic transition of the futex value
2360  * and failed. The kernel side here does the whole locking operation:
2361  * if there are waiters then it will block as a consequence of relying
2362  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2363  * a 0 value of the futex too.).
2364  *
2365  * Also serves as futex trylock_pi()'ing, and due semantics.
2366  */
2367 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2368                          ktime_t *time, int trylock)
2369 {
2370         struct hrtimer_sleeper timeout, *to = NULL;
2371         struct futex_hash_bucket *hb;
2372         struct futex_q q = futex_q_init;
2373         int res, ret;
2374 
2375         if (refill_pi_state_cache())
2376                 return -ENOMEM;
2377 
2378         if (time) {
2379                 to = &timeout;
2380                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2381                                       HRTIMER_MODE_ABS);
2382                 hrtimer_init_sleeper(to, current);
2383                 hrtimer_set_expires(&to->timer, *time);
2384         }
2385 
2386 retry:
2387         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2388         if (unlikely(ret != 0))
2389                 goto out;
2390 
2391 retry_private:
2392         hb = queue_lock(&q);
2393 
2394         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2395         if (unlikely(ret)) {
2396                 /*
2397                  * Atomic work succeeded and we got the lock,
2398                  * or failed. Either way, we do _not_ block.
2399                  */
2400                 switch (ret) {
2401                 case 1:
2402                         /* We got the lock. */
2403                         ret = 0;
2404                         goto out_unlock_put_key;
2405                 case -EFAULT:
2406                         goto uaddr_faulted;
2407                 case -EAGAIN:
2408                         /*
2409                          * Two reasons for this:
2410                          * - Task is exiting and we just wait for the
2411                          *   exit to complete.
2412                          * - The user space value changed.
2413                          */
2414                         queue_unlock(hb);
2415                         put_futex_key(&q.key);
2416                         cond_resched();
2417                         goto retry;
2418                 default:
2419                         goto out_unlock_put_key;
2420                 }
2421         }
2422 
2423         /*
2424          * Only actually queue now that the atomic ops are done:
2425          */
2426         queue_me(&q, hb);
2427 
2428         WARN_ON(!q.pi_state);
2429         /*
2430          * Block on the PI mutex:
2431          */
2432         if (!trylock) {
2433                 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2434         } else {
2435                 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2436                 /* Fixup the trylock return value: */
2437                 ret = ret ? 0 : -EWOULDBLOCK;
2438         }
2439 
2440         spin_lock(q.lock_ptr);
2441         /*
2442          * Fixup the pi_state owner and possibly acquire the lock if we
2443          * haven't already.
2444          */
2445         res = fixup_owner(uaddr, &q, !ret);
2446         /*
2447          * If fixup_owner() returned an error, proprogate that.  If it acquired
2448          * the lock, clear our -ETIMEDOUT or -EINTR.
2449          */
2450         if (res)
2451                 ret = (res < 0) ? res : 0;
2452 
2453         /*
2454          * If fixup_owner() faulted and was unable to handle the fault, unlock
2455          * it and return the fault to userspace.
2456          */
2457         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2458                 rt_mutex_unlock(&q.pi_state->pi_mutex);
2459 
2460         /* Unqueue and drop the lock */
2461         unqueue_me_pi(&q);
2462 
2463         goto out_put_key;
2464 
2465 out_unlock_put_key:
2466         queue_unlock(hb);
2467 
2468 out_put_key:
2469         put_futex_key(&q.key);
2470 out:
2471         if (to)
2472                 destroy_hrtimer_on_stack(&to->timer);
2473         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2474 
2475 uaddr_faulted:
2476         queue_unlock(hb);
2477 
2478         ret = fault_in_user_writeable(uaddr);
2479         if (ret)
2480                 goto out_put_key;
2481 
2482         if (!(flags & FLAGS_SHARED))
2483                 goto retry_private;
2484 
2485         put_futex_key(&q.key);
2486         goto retry;
2487 }
2488 
2489 /*
2490  * Userspace attempted a TID -> 0 atomic transition, and failed.
2491  * This is the in-kernel slowpath: we look up the PI state (if any),
2492  * and do the rt-mutex unlock.
2493  */
2494 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2495 {
2496         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2497         union futex_key key = FUTEX_KEY_INIT;
2498         struct futex_hash_bucket *hb;
2499         struct futex_q *match;
2500         int ret;
2501 
2502 retry:
2503         if (get_user(uval, uaddr))
2504                 return -EFAULT;
2505         /*
2506          * We release only a lock we actually own:
2507          */
2508         if ((uval & FUTEX_TID_MASK) != vpid)
2509                 return -EPERM;
2510 
2511         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2512         if (ret)
2513                 return ret;
2514 
2515         hb = hash_futex(&key);
2516         spin_lock(&hb->lock);
2517 
2518         /*
2519          * Check waiters first. We do not trust user space values at
2520          * all and we at least want to know if user space fiddled
2521          * with the futex value instead of blindly unlocking.
2522          */
2523         match = futex_top_waiter(hb, &key);
2524         if (match) {
2525                 ret = wake_futex_pi(uaddr, uval, match, hb);
2526                 /*
2527                  * In case of success wake_futex_pi dropped the hash
2528                  * bucket lock.
2529                  */
2530                 if (!ret)
2531                         goto out_putkey;
2532                 /*
2533                  * The atomic access to the futex value generated a
2534                  * pagefault, so retry the user-access and the wakeup:
2535                  */
2536                 if (ret == -EFAULT)
2537                         goto pi_faulted;
2538                 /*
2539                  * wake_futex_pi has detected invalid state. Tell user
2540                  * space.
2541                  */
2542                 goto out_unlock;
2543         }
2544 
2545         /*
2546          * We have no kernel internal state, i.e. no waiters in the
2547          * kernel. Waiters which are about to queue themselves are stuck
2548          * on hb->lock. So we can safely ignore them. We do neither
2549          * preserve the WAITERS bit not the OWNER_DIED one. We are the
2550          * owner.
2551          */
2552         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2553                 goto pi_faulted;
2554 
2555         /*
2556          * If uval has changed, let user space handle it.
2557          */
2558         ret = (curval == uval) ? 0 : -EAGAIN;
2559 
2560 out_unlock:
2561         spin_unlock(&hb->lock);
2562 out_putkey:
2563         put_futex_key(&key);
2564         return ret;
2565 
2566 pi_faulted:
2567         spin_unlock(&hb->lock);
2568         put_futex_key(&key);
2569 
2570         ret = fault_in_user_writeable(uaddr);
2571         if (!ret)
2572                 goto retry;
2573 
2574         return ret;
2575 }
2576 
2577 /**
2578  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2579  * @hb:         the hash_bucket futex_q was original enqueued on
2580  * @q:          the futex_q woken while waiting to be requeued
2581  * @key2:       the futex_key of the requeue target futex
2582  * @timeout:    the timeout associated with the wait (NULL if none)
2583  *
2584  * Detect if the task was woken on the initial futex as opposed to the requeue
2585  * target futex.  If so, determine if it was a timeout or a signal that caused
2586  * the wakeup and return the appropriate error code to the caller.  Must be
2587  * called with the hb lock held.
2588  *
2589  * Return:
2590  *  0 = no early wakeup detected;
2591  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2592  */
2593 static inline
2594 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2595                                    struct futex_q *q, union futex_key *key2,
2596                                    struct hrtimer_sleeper *timeout)
2597 {
2598         int ret = 0;
2599 
2600         /*
2601          * With the hb lock held, we avoid races while we process the wakeup.
2602          * We only need to hold hb (and not hb2) to ensure atomicity as the
2603          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2604          * It can't be requeued from uaddr2 to something else since we don't
2605          * support a PI aware source futex for requeue.
2606          */
2607         if (!match_futex(&q->key, key2)) {
2608                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2609                 /*
2610                  * We were woken prior to requeue by a timeout or a signal.
2611                  * Unqueue the futex_q and determine which it was.
2612                  */
2613                 plist_del(&q->list, &hb->chain);
2614                 hb_waiters_dec(hb);
2615 
2616                 /* Handle spurious wakeups gracefully */
2617                 ret = -EWOULDBLOCK;
2618                 if (timeout && !timeout->task)
2619                         ret = -ETIMEDOUT;
2620                 else if (signal_pending(current))
2621                         ret = -ERESTARTNOINTR;
2622         }
2623         return ret;
2624 }
2625 
2626 /**
2627  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2628  * @uaddr:      the futex we initially wait on (non-pi)
2629  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2630  *              the same type, no requeueing from private to shared, etc.
2631  * @val:        the expected value of uaddr
2632  * @abs_time:   absolute timeout
2633  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
2634  * @uaddr2:     the pi futex we will take prior to returning to user-space
2635  *
2636  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2637  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2638  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2639  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2640  * without one, the pi logic would not know which task to boost/deboost, if
2641  * there was a need to.
2642  *
2643  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2644  * via the following--
2645  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2646  * 2) wakeup on uaddr2 after a requeue
2647  * 3) signal
2648  * 4) timeout
2649  *
2650  * If 3, cleanup and return -ERESTARTNOINTR.
2651  *
2652  * If 2, we may then block on trying to take the rt_mutex and return via:
2653  * 5) successful lock
2654  * 6) signal
2655  * 7) timeout
2656  * 8) other lock acquisition failure
2657  *
2658  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2659  *
2660  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2661  *
2662  * Return:
2663  *  0 - On success;
2664  * <0 - On error
2665  */
2666 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2667                                  u32 val, ktime_t *abs_time, u32 bitset,
2668                                  u32 __user *uaddr2)
2669 {
2670         struct hrtimer_sleeper timeout, *to = NULL;
2671         struct rt_mutex_waiter rt_waiter;
2672         struct rt_mutex *pi_mutex = NULL;
2673         struct futex_hash_bucket *hb;
2674         union futex_key key2 = FUTEX_KEY_INIT;
2675         struct futex_q q = futex_q_init;
2676         int res, ret;
2677 
2678         if (uaddr == uaddr2)
2679                 return -EINVAL;
2680 
2681         if (!bitset)
2682                 return -EINVAL;
2683 
2684         if (abs_time) {
2685                 to = &timeout;
2686                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2687                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2688                                       HRTIMER_MODE_ABS);
2689                 hrtimer_init_sleeper(to, current);
2690                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2691                                              current->timer_slack_ns);
2692         }
2693 
2694         /*
2695          * The waiter is allocated on our stack, manipulated by the requeue
2696          * code while we sleep on uaddr.
2697          */
2698         debug_rt_mutex_init_waiter(&rt_waiter);
2699         RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2700         RB_CLEAR_NODE(&rt_waiter.tree_entry);
2701         rt_waiter.task = NULL;
2702 
2703         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2704         if (unlikely(ret != 0))
2705                 goto out;
2706 
2707         q.bitset = bitset;
2708         q.rt_waiter = &rt_waiter;
2709         q.requeue_pi_key = &key2;
2710 
2711         /*
2712          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2713          * count.
2714          */
2715         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2716         if (ret)
2717                 goto out_key2;
2718 
2719         /*
2720          * The check above which compares uaddrs is not sufficient for
2721          * shared futexes. We need to compare the keys:
2722          */
2723         if (match_futex(&q.key, &key2)) {
2724                 queue_unlock(hb);
2725                 ret = -EINVAL;
2726                 goto out_put_keys;
2727         }
2728 
2729         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2730         futex_wait_queue_me(hb, &q, to);
2731 
2732         spin_lock(&hb->lock);
2733         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2734         spin_unlock(&hb->lock);
2735         if (ret)
2736                 goto out_put_keys;
2737 
2738         /*
2739          * In order for us to be here, we know our q.key == key2, and since
2740          * we took the hb->lock above, we also know that futex_requeue() has
2741          * completed and we no longer have to concern ourselves with a wakeup
2742          * race with the atomic proxy lock acquisition by the requeue code. The
2743          * futex_requeue dropped our key1 reference and incremented our key2
2744          * reference count.
2745          */
2746 
2747         /* Check if the requeue code acquired the second futex for us. */
2748         if (!q.rt_waiter) {
2749                 /*
2750                  * Got the lock. We might not be the anticipated owner if we
2751                  * did a lock-steal - fix up the PI-state in that case.
2752                  */
2753                 if (q.pi_state && (q.pi_state->owner != current)) {
2754                         spin_lock(q.lock_ptr);
2755                         ret = fixup_pi_state_owner(uaddr2, &q, current);
2756                         /*
2757                          * Drop the reference to the pi state which
2758                          * the requeue_pi() code acquired for us.
2759                          */
2760                         put_pi_state(q.pi_state);
2761                         spin_unlock(q.lock_ptr);
2762                 }
2763         } else {
2764                 /*
2765                  * We have been woken up by futex_unlock_pi(), a timeout, or a
2766                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2767                  * the pi_state.
2768                  */
2769                 WARN_ON(!q.pi_state);
2770                 pi_mutex = &q.pi_state->pi_mutex;
2771                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2772                 debug_rt_mutex_free_waiter(&rt_waiter);
2773 
2774                 spin_lock(q.lock_ptr);
2775                 /*
2776                  * Fixup the pi_state owner and possibly acquire the lock if we
2777                  * haven't already.
2778                  */
2779                 res = fixup_owner(uaddr2, &q, !ret);
2780                 /*
2781                  * If fixup_owner() returned an error, proprogate that.  If it
2782                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
2783                  */
2784                 if (res)
2785                         ret = (res < 0) ? res : 0;
2786 
2787                 /* Unqueue and drop the lock. */
2788                 unqueue_me_pi(&q);
2789         }
2790 
2791         /*
2792          * If fixup_pi_state_owner() faulted and was unable to handle the
2793          * fault, unlock the rt_mutex and return the fault to userspace.
2794          */
2795         if (ret == -EFAULT) {
2796                 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2797                         rt_mutex_unlock(pi_mutex);
2798         } else if (ret == -EINTR) {
2799                 /*
2800                  * We've already been requeued, but cannot restart by calling
2801                  * futex_lock_pi() directly. We could restart this syscall, but
2802                  * it would detect that the user space "val" changed and return
2803                  * -EWOULDBLOCK.  Save the overhead of the restart and return
2804                  * -EWOULDBLOCK directly.
2805                  */
2806                 ret = -EWOULDBLOCK;
2807         }
2808 
2809 out_put_keys:
2810         put_futex_key(&q.key);
2811 out_key2:
2812         put_futex_key(&key2);
2813 
2814 out:
2815         if (to) {
2816                 hrtimer_cancel(&to->timer);
2817                 destroy_hrtimer_on_stack(&to->timer);
2818         }
2819         return ret;
2820 }
2821 
2822 /*
2823  * Support for robust futexes: the kernel cleans up held futexes at
2824  * thread exit time.
2825  *
2826  * Implementation: user-space maintains a per-thread list of locks it
2827  * is holding. Upon do_exit(), the kernel carefully walks this list,
2828  * and marks all locks that are owned by this thread with the
2829  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2830  * always manipulated with the lock held, so the list is private and
2831  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2832  * field, to allow the kernel to clean up if the thread dies after
2833  * acquiring the lock, but just before it could have added itself to
2834  * the list. There can only be one such pending lock.
2835  */
2836 
2837 /**
2838  * sys_set_robust_list() - Set the robust-futex list head of a task
2839  * @head:       pointer to the list-head
2840  * @len:        length of the list-head, as userspace expects
2841  */
2842 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2843                 size_t, len)
2844 {
2845         if (!futex_cmpxchg_enabled)
2846                 return -ENOSYS;
2847         /*
2848          * The kernel knows only one size for now:
2849          */
2850         if (unlikely(len != sizeof(*head)))
2851                 return -EINVAL;
2852 
2853         current->robust_list = head;
2854 
2855         return 0;
2856 }
2857 
2858 /**
2859  * sys_get_robust_list() - Get the robust-futex list head of a task
2860  * @pid:        pid of the process [zero for current task]
2861  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
2862  * @len_ptr:    pointer to a length field, the kernel fills in the header size
2863  */
2864 SYSCALL_DEFINE3(get_robust_list, int, pid,
2865                 struct robust_list_head __user * __user *, head_ptr,
2866                 size_t __user *, len_ptr)
2867 {
2868         struct robust_list_head __user *head;
2869         unsigned long ret;
2870         struct task_struct *p;
2871 
2872         if (!futex_cmpxchg_enabled)
2873                 return -ENOSYS;
2874 
2875         rcu_read_lock();
2876 
2877         ret = -ESRCH;
2878         if (!pid)
2879                 p = current;
2880         else {
2881                 p = find_task_by_vpid(pid);
2882                 if (!p)
2883                         goto err_unlock;
2884         }
2885 
2886         ret = -EPERM;
2887         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
2888                 goto err_unlock;
2889 
2890         head = p->robust_list;
2891         rcu_read_unlock();
2892 
2893         if (put_user(sizeof(*head), len_ptr))
2894                 return -EFAULT;
2895         return put_user(head, head_ptr);
2896 
2897 err_unlock:
2898         rcu_read_unlock();
2899 
2900         return ret;
2901 }
2902 
2903 /*
2904  * Process a futex-list entry, check whether it's owned by the
2905  * dying task, and do notification if so:
2906  */
2907 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2908 {
2909         u32 uval, uninitialized_var(nval), mval;
2910 
2911 retry:
2912         if (get_user(uval, uaddr))
2913                 return -1;
2914 
2915         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2916                 /*
2917                  * Ok, this dying thread is truly holding a futex
2918                  * of interest. Set the OWNER_DIED bit atomically
2919                  * via cmpxchg, and if the value had FUTEX_WAITERS
2920                  * set, wake up a waiter (if any). (We have to do a
2921                  * futex_wake() even if OWNER_DIED is already set -
2922                  * to handle the rare but possible case of recursive
2923                  * thread-death.) The rest of the cleanup is done in
2924                  * userspace.
2925                  */
2926                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2927                 /*
2928                  * We are not holding a lock here, but we want to have
2929                  * the pagefault_disable/enable() protection because
2930                  * we want to handle the fault gracefully. If the
2931                  * access fails we try to fault in the futex with R/W
2932                  * verification via get_user_pages. get_user() above
2933                  * does not guarantee R/W access. If that fails we
2934                  * give up and leave the futex locked.
2935                  */
2936                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2937                         if (fault_in_user_writeable(uaddr))
2938                                 return -1;
2939                         goto retry;
2940                 }
2941                 if (nval != uval)
2942                         goto retry;
2943 
2944                 /*
2945                  * Wake robust non-PI futexes here. The wakeup of
2946                  * PI futexes happens in exit_pi_state():
2947                  */
2948                 if (!pi && (uval & FUTEX_WAITERS))
2949                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2950         }
2951         return 0;
2952 }
2953 
2954 /*
2955  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2956  */
2957 static inline int fetch_robust_entry(struct robust_list __user **entry,
2958                                      struct robust_list __user * __user *head,
2959                                      unsigned int *pi)
2960 {
2961         unsigned long uentry;
2962 
2963         if (get_user(uentry, (unsigned long __user *)head))
2964                 return -EFAULT;
2965 
2966         *entry = (void __user *)(uentry & ~1UL);
2967         *pi = uentry & 1;
2968 
2969         return 0;
2970 }
2971 
2972 /*
2973  * Walk curr->robust_list (very carefully, it's a userspace list!)
2974  * and mark any locks found there dead, and notify any waiters.
2975  *
2976  * We silently return on any sign of list-walking problem.
2977  */
2978 void exit_robust_list(struct task_struct *curr)
2979 {
2980         struct robust_list_head __user *head = curr->robust_list;
2981         struct robust_list __user *entry, *next_entry, *pending;
2982         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2983         unsigned int uninitialized_var(next_pi);
2984         unsigned long futex_offset;
2985         int rc;
2986 
2987         if (!futex_cmpxchg_enabled)
2988                 return;
2989 
2990         /*
2991          * Fetch the list head (which was registered earlier, via
2992          * sys_set_robust_list()):
2993          */
2994         if (fetch_robust_entry(&entry, &head->list.next, &pi))
2995                 return;
2996         /*
2997          * Fetch the relative futex offset:
2998          */
2999         if (get_user(futex_offset, &head->futex_offset))
3000                 return;
3001         /*
3002          * Fetch any possibly pending lock-add first, and handle it
3003          * if it exists:
3004          */
3005         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3006                 return;
3007 
3008         next_entry = NULL;      /* avoid warning with gcc */
3009         while (entry != &head->list) {
3010                 /*
3011                  * Fetch the next entry in the list before calling
3012                  * handle_futex_death:
3013                  */
3014                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3015                 /*
3016                  * A pending lock might already be on the list, so
3017                  * don't process it twice:
3018                  */
3019                 if (entry != pending)
3020                         if (handle_futex_death((void __user *)entry + futex_offset,
3021                                                 curr, pi))
3022                                 return;
3023                 if (rc)
3024                         return;
3025                 entry = next_entry;
3026                 pi = next_pi;
3027                 /*
3028                  * Avoid excessively long or circular lists:
3029                  */
3030                 if (!--limit)
3031                         break;
3032 
3033                 cond_resched();
3034         }
3035 
3036         if (pending)
3037                 handle_futex_death((void __user *)pending + futex_offset,
3038                                    curr, pip);
3039 }
3040 
3041 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3042                 u32 __user *uaddr2, u32 val2, u32 val3)
3043 {
3044         int cmd = op & FUTEX_CMD_MASK;
3045         unsigned int flags = 0;
3046 
3047         if (!(op & FUTEX_PRIVATE_FLAG))
3048                 flags |= FLAGS_SHARED;
3049 
3050         if (op & FUTEX_CLOCK_REALTIME) {
3051                 flags |= FLAGS_CLOCKRT;
3052                 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3053                     cmd != FUTEX_WAIT_REQUEUE_PI)
3054                         return -ENOSYS;
3055         }
3056 
3057         switch (cmd) {
3058         case FUTEX_LOCK_PI:
3059         case FUTEX_UNLOCK_PI:
3060         case FUTEX_TRYLOCK_PI:
3061         case FUTEX_WAIT_REQUEUE_PI:
3062         case FUTEX_CMP_REQUEUE_PI:
3063                 if (!futex_cmpxchg_enabled)
3064                         return -ENOSYS;
3065         }
3066 
3067         switch (cmd) {
3068         case FUTEX_WAIT:
3069                 val3 = FUTEX_BITSET_MATCH_ANY;
3070         case FUTEX_WAIT_BITSET:
3071                 return futex_wait(uaddr, flags, val, timeout, val3);
3072         case FUTEX_WAKE:
3073                 val3 = FUTEX_BITSET_MATCH_ANY;
3074         case FUTEX_WAKE_BITSET:
3075                 return futex_wake(uaddr, flags, val, val3);
3076         case FUTEX_REQUEUE:
3077                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3078         case FUTEX_CMP_REQUEUE:
3079                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3080         case FUTEX_WAKE_OP:
3081                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3082         case FUTEX_LOCK_PI:
3083                 return futex_lock_pi(uaddr, flags, timeout, 0);
3084         case FUTEX_UNLOCK_PI:
3085                 return futex_unlock_pi(uaddr, flags);
3086         case FUTEX_TRYLOCK_PI:
3087                 return futex_lock_pi(uaddr, flags, NULL, 1);
3088         case FUTEX_WAIT_REQUEUE_PI:
3089                 val3 = FUTEX_BITSET_MATCH_ANY;
3090                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3091                                              uaddr2);
3092         case FUTEX_CMP_REQUEUE_PI:
3093                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3094         }
3095         return -ENOSYS;
3096 }
3097 
3098 
3099 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3100                 struct timespec __user *, utime, u32 __user *, uaddr2,
3101                 u32, val3)
3102 {
3103         struct timespec ts;
3104         ktime_t t, *tp = NULL;
3105         u32 val2 = 0;
3106         int cmd = op & FUTEX_CMD_MASK;
3107 
3108         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3109                       cmd == FUTEX_WAIT_BITSET ||
3110                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3111                 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3112                         return -EFAULT;
3113                 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3114                         return -EFAULT;
3115                 if (!timespec_valid(&ts))
3116                         return -EINVAL;
3117 
3118                 t = timespec_to_ktime(ts);
3119                 if (cmd == FUTEX_WAIT)
3120                         t = ktime_add_safe(ktime_get(), t);
3121                 tp = &t;
3122         }
3123         /*
3124          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3125          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3126          */
3127         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3128             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3129                 val2 = (u32) (unsigned long) utime;
3130 
3131         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3132 }
3133 
3134 static void __init futex_detect_cmpxchg(void)
3135 {
3136 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3137         u32 curval;
3138 
3139         /*
3140          * This will fail and we want it. Some arch implementations do
3141          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3142          * functionality. We want to know that before we call in any
3143          * of the complex code paths. Also we want to prevent
3144          * registration of robust lists in that case. NULL is
3145          * guaranteed to fault and we get -EFAULT on functional
3146          * implementation, the non-functional ones will return
3147          * -ENOSYS.
3148          */
3149         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3150                 futex_cmpxchg_enabled = 1;
3151 #endif
3152 }
3153 
3154 static int __init futex_init(void)
3155 {
3156         unsigned int futex_shift;
3157         unsigned long i;
3158 
3159 #if CONFIG_BASE_SMALL
3160         futex_hashsize = 16;
3161 #else
3162         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3163 #endif
3164 
3165         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3166                                                futex_hashsize, 0,
3167                                                futex_hashsize < 256 ? HASH_SMALL : 0,
3168                                                &futex_shift, NULL,
3169                                                futex_hashsize, futex_hashsize);
3170         futex_hashsize = 1UL << futex_shift;
3171 
3172         futex_detect_cmpxchg();
3173 
3174         for (i = 0; i < futex_hashsize; i++) {
3175                 atomic_set(&futex_queues[i].waiters, 0);
3176                 plist_head_init(&futex_queues[i].chain);
3177                 spin_lock_init(&futex_queues[i].lock);
3178         }
3179 
3180         return 0;
3181 }
3182 __initcall(futex_init);
3183 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us