y2038: futex: Add support for __kernel_timespec
[sfrench/cifs-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
50 #include <linux/fs.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/memblock.h>
70 #include <linux/fault-inject.h>
71
72 #include <asm/futex.h>
73
74 #include "locking/rtmutex_common.h"
75
76 /*
77  * READ this before attempting to hack on futexes!
78  *
79  * Basic futex operation and ordering guarantees
80  * =============================================
81  *
82  * The waiter reads the futex value in user space and calls
83  * futex_wait(). This function computes the hash bucket and acquires
84  * the hash bucket lock. After that it reads the futex user space value
85  * again and verifies that the data has not changed. If it has not changed
86  * it enqueues itself into the hash bucket, releases the hash bucket lock
87  * and schedules.
88  *
89  * The waker side modifies the user space value of the futex and calls
90  * futex_wake(). This function computes the hash bucket and acquires the
91  * hash bucket lock. Then it looks for waiters on that futex in the hash
92  * bucket and wakes them.
93  *
94  * In futex wake up scenarios where no tasks are blocked on a futex, taking
95  * the hb spinlock can be avoided and simply return. In order for this
96  * optimization to work, ordering guarantees must exist so that the waiter
97  * being added to the list is acknowledged when the list is concurrently being
98  * checked by the waker, avoiding scenarios like the following:
99  *
100  * CPU 0                               CPU 1
101  * val = *futex;
102  * sys_futex(WAIT, futex, val);
103  *   futex_wait(futex, val);
104  *   uval = *futex;
105  *                                     *futex = newval;
106  *                                     sys_futex(WAKE, futex);
107  *                                       futex_wake(futex);
108  *                                       if (queue_empty())
109  *                                         return;
110  *   if (uval == val)
111  *      lock(hash_bucket(futex));
112  *      queue();
113  *     unlock(hash_bucket(futex));
114  *     schedule();
115  *
116  * This would cause the waiter on CPU 0 to wait forever because it
117  * missed the transition of the user space value from val to newval
118  * and the waker did not find the waiter in the hash bucket queue.
119  *
120  * The correct serialization ensures that a waiter either observes
121  * the changed user space value before blocking or is woken by a
122  * concurrent waker:
123  *
124  * CPU 0                                 CPU 1
125  * val = *futex;
126  * sys_futex(WAIT, futex, val);
127  *   futex_wait(futex, val);
128  *
129  *   waiters++; (a)
130  *   smp_mb(); (A) <-- paired with -.
131  *                                  |
132  *   lock(hash_bucket(futex));      |
133  *                                  |
134  *   uval = *futex;                 |
135  *                                  |        *futex = newval;
136  *                                  |        sys_futex(WAKE, futex);
137  *                                  |          futex_wake(futex);
138  *                                  |
139  *                                  `--------> smp_mb(); (B)
140  *   if (uval == val)
141  *     queue();
142  *     unlock(hash_bucket(futex));
143  *     schedule();                         if (waiters)
144  *                                           lock(hash_bucket(futex));
145  *   else                                    wake_waiters(futex);
146  *     waiters--; (b)                        unlock(hash_bucket(futex));
147  *
148  * Where (A) orders the waiters increment and the futex value read through
149  * atomic operations (see hb_waiters_inc) and where (B) orders the write
150  * to futex and the waiters read -- this is done by the barriers for both
151  * shared and private futexes in get_futex_key_refs().
152  *
153  * This yields the following case (where X:=waiters, Y:=futex):
154  *
155  *      X = Y = 0
156  *
157  *      w[X]=1          w[Y]=1
158  *      MB              MB
159  *      r[Y]=y          r[X]=x
160  *
161  * Which guarantees that x==0 && y==0 is impossible; which translates back into
162  * the guarantee that we cannot both miss the futex variable change and the
163  * enqueue.
164  *
165  * Note that a new waiter is accounted for in (a) even when it is possible that
166  * the wait call can return error, in which case we backtrack from it in (b).
167  * Refer to the comment in queue_lock().
168  *
169  * Similarly, in order to account for waiters being requeued on another
170  * address we always increment the waiters for the destination bucket before
171  * acquiring the lock. It then decrements them again  after releasing it -
172  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173  * will do the additional required waiter count housekeeping. This is done for
174  * double_lock_hb() and double_unlock_hb(), respectively.
175  */
176
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
179 #else
180 static int  __read_mostly futex_cmpxchg_enabled;
181 #endif
182
183 /*
184  * Futex flags used to encode options to functions and preserve them across
185  * restarts.
186  */
187 #ifdef CONFIG_MMU
188 # define FLAGS_SHARED           0x01
189 #else
190 /*
191  * NOMMU does not have per process address space. Let the compiler optimize
192  * code away.
193  */
194 # define FLAGS_SHARED           0x00
195 #endif
196 #define FLAGS_CLOCKRT           0x02
197 #define FLAGS_HAS_TIMEOUT       0x04
198
199 /*
200  * Priority Inheritance state:
201  */
202 struct futex_pi_state {
203         /*
204          * list of 'owned' pi_state instances - these have to be
205          * cleaned up in do_exit() if the task exits prematurely:
206          */
207         struct list_head list;
208
209         /*
210          * The PI object:
211          */
212         struct rt_mutex pi_mutex;
213
214         struct task_struct *owner;
215         atomic_t refcount;
216
217         union futex_key key;
218 } __randomize_layout;
219
220 /**
221  * struct futex_q - The hashed futex queue entry, one per waiting task
222  * @list:               priority-sorted list of tasks waiting on this futex
223  * @task:               the task waiting on the futex
224  * @lock_ptr:           the hash bucket lock
225  * @key:                the key the futex is hashed on
226  * @pi_state:           optional priority inheritance state
227  * @rt_waiter:          rt_waiter storage for use with requeue_pi
228  * @requeue_pi_key:     the requeue_pi target futex key
229  * @bitset:             bitset for the optional bitmasked wakeup
230  *
231  * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232  * we can wake only the relevant ones (hashed queues may be shared).
233  *
234  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236  * The order of wakeup is always to make the first condition true, then
237  * the second.
238  *
239  * PI futexes are typically woken before they are removed from the hash list via
240  * the rt_mutex code. See unqueue_me_pi().
241  */
242 struct futex_q {
243         struct plist_node list;
244
245         struct task_struct *task;
246         spinlock_t *lock_ptr;
247         union futex_key key;
248         struct futex_pi_state *pi_state;
249         struct rt_mutex_waiter *rt_waiter;
250         union futex_key *requeue_pi_key;
251         u32 bitset;
252 } __randomize_layout;
253
254 static const struct futex_q futex_q_init = {
255         /* list gets initialized in queue_me()*/
256         .key = FUTEX_KEY_INIT,
257         .bitset = FUTEX_BITSET_MATCH_ANY
258 };
259
260 /*
261  * Hash buckets are shared by all the futex_keys that hash to the same
262  * location.  Each key may have multiple futex_q structures, one for each task
263  * waiting on a futex.
264  */
265 struct futex_hash_bucket {
266         atomic_t waiters;
267         spinlock_t lock;
268         struct plist_head chain;
269 } ____cacheline_aligned_in_smp;
270
271 /*
272  * The base of the bucket array and its size are always used together
273  * (after initialization only in hash_futex()), so ensure that they
274  * reside in the same cacheline.
275  */
276 static struct {
277         struct futex_hash_bucket *queues;
278         unsigned long            hashsize;
279 } __futex_data __read_mostly __aligned(2*sizeof(long));
280 #define futex_queues   (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
282
283
284 /*
285  * Fault injections for futexes.
286  */
287 #ifdef CONFIG_FAIL_FUTEX
288
289 static struct {
290         struct fault_attr attr;
291
292         bool ignore_private;
293 } fail_futex = {
294         .attr = FAULT_ATTR_INITIALIZER,
295         .ignore_private = false,
296 };
297
298 static int __init setup_fail_futex(char *str)
299 {
300         return setup_fault_attr(&fail_futex.attr, str);
301 }
302 __setup("fail_futex=", setup_fail_futex);
303
304 static bool should_fail_futex(bool fshared)
305 {
306         if (fail_futex.ignore_private && !fshared)
307                 return false;
308
309         return should_fail(&fail_futex.attr, 1);
310 }
311
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
313
314 static int __init fail_futex_debugfs(void)
315 {
316         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
317         struct dentry *dir;
318
319         dir = fault_create_debugfs_attr("fail_futex", NULL,
320                                         &fail_futex.attr);
321         if (IS_ERR(dir))
322                 return PTR_ERR(dir);
323
324         if (!debugfs_create_bool("ignore-private", mode, dir,
325                                  &fail_futex.ignore_private)) {
326                 debugfs_remove_recursive(dir);
327                 return -ENOMEM;
328         }
329
330         return 0;
331 }
332
333 late_initcall(fail_futex_debugfs);
334
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
336
337 #else
338 static inline bool should_fail_futex(bool fshared)
339 {
340         return false;
341 }
342 #endif /* CONFIG_FAIL_FUTEX */
343
344 static inline void futex_get_mm(union futex_key *key)
345 {
346         mmgrab(key->private.mm);
347         /*
348          * Ensure futex_get_mm() implies a full barrier such that
349          * get_futex_key() implies a full barrier. This is relied upon
350          * as smp_mb(); (B), see the ordering comment above.
351          */
352         smp_mb__after_atomic();
353 }
354
355 /*
356  * Reflects a new waiter being added to the waitqueue.
357  */
358 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
359 {
360 #ifdef CONFIG_SMP
361         atomic_inc(&hb->waiters);
362         /*
363          * Full barrier (A), see the ordering comment above.
364          */
365         smp_mb__after_atomic();
366 #endif
367 }
368
369 /*
370  * Reflects a waiter being removed from the waitqueue by wakeup
371  * paths.
372  */
373 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
374 {
375 #ifdef CONFIG_SMP
376         atomic_dec(&hb->waiters);
377 #endif
378 }
379
380 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
381 {
382 #ifdef CONFIG_SMP
383         return atomic_read(&hb->waiters);
384 #else
385         return 1;
386 #endif
387 }
388
389 /**
390  * hash_futex - Return the hash bucket in the global hash
391  * @key:        Pointer to the futex key for which the hash is calculated
392  *
393  * We hash on the keys returned from get_futex_key (see below) and return the
394  * corresponding hash bucket in the global hash.
395  */
396 static struct futex_hash_bucket *hash_futex(union futex_key *key)
397 {
398         u32 hash = jhash2((u32*)&key->both.word,
399                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
400                           key->both.offset);
401         return &futex_queues[hash & (futex_hashsize - 1)];
402 }
403
404
405 /**
406  * match_futex - Check whether two futex keys are equal
407  * @key1:       Pointer to key1
408  * @key2:       Pointer to key2
409  *
410  * Return 1 if two futex_keys are equal, 0 otherwise.
411  */
412 static inline int match_futex(union futex_key *key1, union futex_key *key2)
413 {
414         return (key1 && key2
415                 && key1->both.word == key2->both.word
416                 && key1->both.ptr == key2->both.ptr
417                 && key1->both.offset == key2->both.offset);
418 }
419
420 /*
421  * Take a reference to the resource addressed by a key.
422  * Can be called while holding spinlocks.
423  *
424  */
425 static void get_futex_key_refs(union futex_key *key)
426 {
427         if (!key->both.ptr)
428                 return;
429
430         /*
431          * On MMU less systems futexes are always "private" as there is no per
432          * process address space. We need the smp wmb nevertheless - yes,
433          * arch/blackfin has MMU less SMP ...
434          */
435         if (!IS_ENABLED(CONFIG_MMU)) {
436                 smp_mb(); /* explicit smp_mb(); (B) */
437                 return;
438         }
439
440         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
441         case FUT_OFF_INODE:
442                 ihold(key->shared.inode); /* implies smp_mb(); (B) */
443                 break;
444         case FUT_OFF_MMSHARED:
445                 futex_get_mm(key); /* implies smp_mb(); (B) */
446                 break;
447         default:
448                 /*
449                  * Private futexes do not hold reference on an inode or
450                  * mm, therefore the only purpose of calling get_futex_key_refs
451                  * is because we need the barrier for the lockless waiter check.
452                  */
453                 smp_mb(); /* explicit smp_mb(); (B) */
454         }
455 }
456
457 /*
458  * Drop a reference to the resource addressed by a key.
459  * The hash bucket spinlock must not be held. This is
460  * a no-op for private futexes, see comment in the get
461  * counterpart.
462  */
463 static void drop_futex_key_refs(union futex_key *key)
464 {
465         if (!key->both.ptr) {
466                 /* If we're here then we tried to put a key we failed to get */
467                 WARN_ON_ONCE(1);
468                 return;
469         }
470
471         if (!IS_ENABLED(CONFIG_MMU))
472                 return;
473
474         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
475         case FUT_OFF_INODE:
476                 iput(key->shared.inode);
477                 break;
478         case FUT_OFF_MMSHARED:
479                 mmdrop(key->private.mm);
480                 break;
481         }
482 }
483
484 /**
485  * get_futex_key() - Get parameters which are the keys for a futex
486  * @uaddr:      virtual address of the futex
487  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
488  * @key:        address where result is stored.
489  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
490  *              VERIFY_WRITE)
491  *
492  * Return: a negative error code or 0
493  *
494  * The key words are stored in @key on success.
495  *
496  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
497  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
498  * We can usually work out the index without swapping in the page.
499  *
500  * lock_page() might sleep, the caller should not hold a spinlock.
501  */
502 static int
503 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
504 {
505         unsigned long address = (unsigned long)uaddr;
506         struct mm_struct *mm = current->mm;
507         struct page *page, *tail;
508         struct address_space *mapping;
509         int err, ro = 0;
510
511         /*
512          * The futex address must be "naturally" aligned.
513          */
514         key->both.offset = address % PAGE_SIZE;
515         if (unlikely((address % sizeof(u32)) != 0))
516                 return -EINVAL;
517         address -= key->both.offset;
518
519         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
520                 return -EFAULT;
521
522         if (unlikely(should_fail_futex(fshared)))
523                 return -EFAULT;
524
525         /*
526          * PROCESS_PRIVATE futexes are fast.
527          * As the mm cannot disappear under us and the 'key' only needs
528          * virtual address, we dont even have to find the underlying vma.
529          * Note : We do have to check 'uaddr' is a valid user address,
530          *        but access_ok() should be faster than find_vma()
531          */
532         if (!fshared) {
533                 key->private.mm = mm;
534                 key->private.address = address;
535                 get_futex_key_refs(key);  /* implies smp_mb(); (B) */
536                 return 0;
537         }
538
539 again:
540         /* Ignore any VERIFY_READ mapping (futex common case) */
541         if (unlikely(should_fail_futex(fshared)))
542                 return -EFAULT;
543
544         err = get_user_pages_fast(address, 1, 1, &page);
545         /*
546          * If write access is not required (eg. FUTEX_WAIT), try
547          * and get read-only access.
548          */
549         if (err == -EFAULT && rw == VERIFY_READ) {
550                 err = get_user_pages_fast(address, 1, 0, &page);
551                 ro = 1;
552         }
553         if (err < 0)
554                 return err;
555         else
556                 err = 0;
557
558         /*
559          * The treatment of mapping from this point on is critical. The page
560          * lock protects many things but in this context the page lock
561          * stabilizes mapping, prevents inode freeing in the shared
562          * file-backed region case and guards against movement to swap cache.
563          *
564          * Strictly speaking the page lock is not needed in all cases being
565          * considered here and page lock forces unnecessarily serialization
566          * From this point on, mapping will be re-verified if necessary and
567          * page lock will be acquired only if it is unavoidable
568          *
569          * Mapping checks require the head page for any compound page so the
570          * head page and mapping is looked up now. For anonymous pages, it
571          * does not matter if the page splits in the future as the key is
572          * based on the address. For filesystem-backed pages, the tail is
573          * required as the index of the page determines the key. For
574          * base pages, there is no tail page and tail == page.
575          */
576         tail = page;
577         page = compound_head(page);
578         mapping = READ_ONCE(page->mapping);
579
580         /*
581          * If page->mapping is NULL, then it cannot be a PageAnon
582          * page; but it might be the ZERO_PAGE or in the gate area or
583          * in a special mapping (all cases which we are happy to fail);
584          * or it may have been a good file page when get_user_pages_fast
585          * found it, but truncated or holepunched or subjected to
586          * invalidate_complete_page2 before we got the page lock (also
587          * cases which we are happy to fail).  And we hold a reference,
588          * so refcount care in invalidate_complete_page's remove_mapping
589          * prevents drop_caches from setting mapping to NULL beneath us.
590          *
591          * The case we do have to guard against is when memory pressure made
592          * shmem_writepage move it from filecache to swapcache beneath us:
593          * an unlikely race, but we do need to retry for page->mapping.
594          */
595         if (unlikely(!mapping)) {
596                 int shmem_swizzled;
597
598                 /*
599                  * Page lock is required to identify which special case above
600                  * applies. If this is really a shmem page then the page lock
601                  * will prevent unexpected transitions.
602                  */
603                 lock_page(page);
604                 shmem_swizzled = PageSwapCache(page) || page->mapping;
605                 unlock_page(page);
606                 put_page(page);
607
608                 if (shmem_swizzled)
609                         goto again;
610
611                 return -EFAULT;
612         }
613
614         /*
615          * Private mappings are handled in a simple way.
616          *
617          * If the futex key is stored on an anonymous page, then the associated
618          * object is the mm which is implicitly pinned by the calling process.
619          *
620          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
621          * it's a read-only handle, it's expected that futexes attach to
622          * the object not the particular process.
623          */
624         if (PageAnon(page)) {
625                 /*
626                  * A RO anonymous page will never change and thus doesn't make
627                  * sense for futex operations.
628                  */
629                 if (unlikely(should_fail_futex(fshared)) || ro) {
630                         err = -EFAULT;
631                         goto out;
632                 }
633
634                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
635                 key->private.mm = mm;
636                 key->private.address = address;
637
638                 get_futex_key_refs(key); /* implies smp_mb(); (B) */
639
640         } else {
641                 struct inode *inode;
642
643                 /*
644                  * The associated futex object in this case is the inode and
645                  * the page->mapping must be traversed. Ordinarily this should
646                  * be stabilised under page lock but it's not strictly
647                  * necessary in this case as we just want to pin the inode, not
648                  * update the radix tree or anything like that.
649                  *
650                  * The RCU read lock is taken as the inode is finally freed
651                  * under RCU. If the mapping still matches expectations then the
652                  * mapping->host can be safely accessed as being a valid inode.
653                  */
654                 rcu_read_lock();
655
656                 if (READ_ONCE(page->mapping) != mapping) {
657                         rcu_read_unlock();
658                         put_page(page);
659
660                         goto again;
661                 }
662
663                 inode = READ_ONCE(mapping->host);
664                 if (!inode) {
665                         rcu_read_unlock();
666                         put_page(page);
667
668                         goto again;
669                 }
670
671                 /*
672                  * Take a reference unless it is about to be freed. Previously
673                  * this reference was taken by ihold under the page lock
674                  * pinning the inode in place so i_lock was unnecessary. The
675                  * only way for this check to fail is if the inode was
676                  * truncated in parallel which is almost certainly an
677                  * application bug. In such a case, just retry.
678                  *
679                  * We are not calling into get_futex_key_refs() in file-backed
680                  * cases, therefore a successful atomic_inc return below will
681                  * guarantee that get_futex_key() will still imply smp_mb(); (B).
682                  */
683                 if (!atomic_inc_not_zero(&inode->i_count)) {
684                         rcu_read_unlock();
685                         put_page(page);
686
687                         goto again;
688                 }
689
690                 /* Should be impossible but lets be paranoid for now */
691                 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
692                         err = -EFAULT;
693                         rcu_read_unlock();
694                         iput(inode);
695
696                         goto out;
697                 }
698
699                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
700                 key->shared.inode = inode;
701                 key->shared.pgoff = basepage_index(tail);
702                 rcu_read_unlock();
703         }
704
705 out:
706         put_page(page);
707         return err;
708 }
709
710 static inline void put_futex_key(union futex_key *key)
711 {
712         drop_futex_key_refs(key);
713 }
714
715 /**
716  * fault_in_user_writeable() - Fault in user address and verify RW access
717  * @uaddr:      pointer to faulting user space address
718  *
719  * Slow path to fixup the fault we just took in the atomic write
720  * access to @uaddr.
721  *
722  * We have no generic implementation of a non-destructive write to the
723  * user address. We know that we faulted in the atomic pagefault
724  * disabled section so we can as well avoid the #PF overhead by
725  * calling get_user_pages() right away.
726  */
727 static int fault_in_user_writeable(u32 __user *uaddr)
728 {
729         struct mm_struct *mm = current->mm;
730         int ret;
731
732         down_read(&mm->mmap_sem);
733         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
734                                FAULT_FLAG_WRITE, NULL);
735         up_read(&mm->mmap_sem);
736
737         return ret < 0 ? ret : 0;
738 }
739
740 /**
741  * futex_top_waiter() - Return the highest priority waiter on a futex
742  * @hb:         the hash bucket the futex_q's reside in
743  * @key:        the futex key (to distinguish it from other futex futex_q's)
744  *
745  * Must be called with the hb lock held.
746  */
747 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
748                                         union futex_key *key)
749 {
750         struct futex_q *this;
751
752         plist_for_each_entry(this, &hb->chain, list) {
753                 if (match_futex(&this->key, key))
754                         return this;
755         }
756         return NULL;
757 }
758
759 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
760                                       u32 uval, u32 newval)
761 {
762         int ret;
763
764         pagefault_disable();
765         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
766         pagefault_enable();
767
768         return ret;
769 }
770
771 static int get_futex_value_locked(u32 *dest, u32 __user *from)
772 {
773         int ret;
774
775         pagefault_disable();
776         ret = __get_user(*dest, from);
777         pagefault_enable();
778
779         return ret ? -EFAULT : 0;
780 }
781
782
783 /*
784  * PI code:
785  */
786 static int refill_pi_state_cache(void)
787 {
788         struct futex_pi_state *pi_state;
789
790         if (likely(current->pi_state_cache))
791                 return 0;
792
793         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
794
795         if (!pi_state)
796                 return -ENOMEM;
797
798         INIT_LIST_HEAD(&pi_state->list);
799         /* pi_mutex gets initialized later */
800         pi_state->owner = NULL;
801         atomic_set(&pi_state->refcount, 1);
802         pi_state->key = FUTEX_KEY_INIT;
803
804         current->pi_state_cache = pi_state;
805
806         return 0;
807 }
808
809 static struct futex_pi_state *alloc_pi_state(void)
810 {
811         struct futex_pi_state *pi_state = current->pi_state_cache;
812
813         WARN_ON(!pi_state);
814         current->pi_state_cache = NULL;
815
816         return pi_state;
817 }
818
819 static void get_pi_state(struct futex_pi_state *pi_state)
820 {
821         WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
822 }
823
824 /*
825  * Drops a reference to the pi_state object and frees or caches it
826  * when the last reference is gone.
827  */
828 static void put_pi_state(struct futex_pi_state *pi_state)
829 {
830         if (!pi_state)
831                 return;
832
833         if (!atomic_dec_and_test(&pi_state->refcount))
834                 return;
835
836         /*
837          * If pi_state->owner is NULL, the owner is most probably dying
838          * and has cleaned up the pi_state already
839          */
840         if (pi_state->owner) {
841                 struct task_struct *owner;
842
843                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
844                 owner = pi_state->owner;
845                 if (owner) {
846                         raw_spin_lock(&owner->pi_lock);
847                         list_del_init(&pi_state->list);
848                         raw_spin_unlock(&owner->pi_lock);
849                 }
850                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
851                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
852         }
853
854         if (current->pi_state_cache) {
855                 kfree(pi_state);
856         } else {
857                 /*
858                  * pi_state->list is already empty.
859                  * clear pi_state->owner.
860                  * refcount is at 0 - put it back to 1.
861                  */
862                 pi_state->owner = NULL;
863                 atomic_set(&pi_state->refcount, 1);
864                 current->pi_state_cache = pi_state;
865         }
866 }
867
868 #ifdef CONFIG_FUTEX_PI
869
870 /*
871  * This task is holding PI mutexes at exit time => bad.
872  * Kernel cleans up PI-state, but userspace is likely hosed.
873  * (Robust-futex cleanup is separate and might save the day for userspace.)
874  */
875 void exit_pi_state_list(struct task_struct *curr)
876 {
877         struct list_head *next, *head = &curr->pi_state_list;
878         struct futex_pi_state *pi_state;
879         struct futex_hash_bucket *hb;
880         union futex_key key = FUTEX_KEY_INIT;
881
882         if (!futex_cmpxchg_enabled)
883                 return;
884         /*
885          * We are a ZOMBIE and nobody can enqueue itself on
886          * pi_state_list anymore, but we have to be careful
887          * versus waiters unqueueing themselves:
888          */
889         raw_spin_lock_irq(&curr->pi_lock);
890         while (!list_empty(head)) {
891                 next = head->next;
892                 pi_state = list_entry(next, struct futex_pi_state, list);
893                 key = pi_state->key;
894                 hb = hash_futex(&key);
895
896                 /*
897                  * We can race against put_pi_state() removing itself from the
898                  * list (a waiter going away). put_pi_state() will first
899                  * decrement the reference count and then modify the list, so
900                  * its possible to see the list entry but fail this reference
901                  * acquire.
902                  *
903                  * In that case; drop the locks to let put_pi_state() make
904                  * progress and retry the loop.
905                  */
906                 if (!atomic_inc_not_zero(&pi_state->refcount)) {
907                         raw_spin_unlock_irq(&curr->pi_lock);
908                         cpu_relax();
909                         raw_spin_lock_irq(&curr->pi_lock);
910                         continue;
911                 }
912                 raw_spin_unlock_irq(&curr->pi_lock);
913
914                 spin_lock(&hb->lock);
915                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
916                 raw_spin_lock(&curr->pi_lock);
917                 /*
918                  * We dropped the pi-lock, so re-check whether this
919                  * task still owns the PI-state:
920                  */
921                 if (head->next != next) {
922                         /* retain curr->pi_lock for the loop invariant */
923                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
924                         spin_unlock(&hb->lock);
925                         put_pi_state(pi_state);
926                         continue;
927                 }
928
929                 WARN_ON(pi_state->owner != curr);
930                 WARN_ON(list_empty(&pi_state->list));
931                 list_del_init(&pi_state->list);
932                 pi_state->owner = NULL;
933
934                 raw_spin_unlock(&curr->pi_lock);
935                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
936                 spin_unlock(&hb->lock);
937
938                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
939                 put_pi_state(pi_state);
940
941                 raw_spin_lock_irq(&curr->pi_lock);
942         }
943         raw_spin_unlock_irq(&curr->pi_lock);
944 }
945
946 #endif
947
948 /*
949  * We need to check the following states:
950  *
951  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
952  *
953  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
954  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
955  *
956  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
957  *
958  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
959  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
960  *
961  * [6]  Found  | Found    | task      | 0         | 1      | Valid
962  *
963  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
964  *
965  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
966  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
967  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
968  *
969  * [1]  Indicates that the kernel can acquire the futex atomically. We
970  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
971  *
972  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
973  *      thread is found then it indicates that the owner TID has died.
974  *
975  * [3]  Invalid. The waiter is queued on a non PI futex
976  *
977  * [4]  Valid state after exit_robust_list(), which sets the user space
978  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
979  *
980  * [5]  The user space value got manipulated between exit_robust_list()
981  *      and exit_pi_state_list()
982  *
983  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
984  *      the pi_state but cannot access the user space value.
985  *
986  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
987  *
988  * [8]  Owner and user space value match
989  *
990  * [9]  There is no transient state which sets the user space TID to 0
991  *      except exit_robust_list(), but this is indicated by the
992  *      FUTEX_OWNER_DIED bit. See [4]
993  *
994  * [10] There is no transient state which leaves owner and user space
995  *      TID out of sync.
996  *
997  *
998  * Serialization and lifetime rules:
999  *
1000  * hb->lock:
1001  *
1002  *      hb -> futex_q, relation
1003  *      futex_q -> pi_state, relation
1004  *
1005  *      (cannot be raw because hb can contain arbitrary amount
1006  *       of futex_q's)
1007  *
1008  * pi_mutex->wait_lock:
1009  *
1010  *      {uval, pi_state}
1011  *
1012  *      (and pi_mutex 'obviously')
1013  *
1014  * p->pi_lock:
1015  *
1016  *      p->pi_state_list -> pi_state->list, relation
1017  *
1018  * pi_state->refcount:
1019  *
1020  *      pi_state lifetime
1021  *
1022  *
1023  * Lock order:
1024  *
1025  *   hb->lock
1026  *     pi_mutex->wait_lock
1027  *       p->pi_lock
1028  *
1029  */
1030
1031 /*
1032  * Validate that the existing waiter has a pi_state and sanity check
1033  * the pi_state against the user space value. If correct, attach to
1034  * it.
1035  */
1036 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1037                               struct futex_pi_state *pi_state,
1038                               struct futex_pi_state **ps)
1039 {
1040         pid_t pid = uval & FUTEX_TID_MASK;
1041         u32 uval2;
1042         int ret;
1043
1044         /*
1045          * Userspace might have messed up non-PI and PI futexes [3]
1046          */
1047         if (unlikely(!pi_state))
1048                 return -EINVAL;
1049
1050         /*
1051          * We get here with hb->lock held, and having found a
1052          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1053          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1054          * which in turn means that futex_lock_pi() still has a reference on
1055          * our pi_state.
1056          *
1057          * The waiter holding a reference on @pi_state also protects against
1058          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1059          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1060          * free pi_state before we can take a reference ourselves.
1061          */
1062         WARN_ON(!atomic_read(&pi_state->refcount));
1063
1064         /*
1065          * Now that we have a pi_state, we can acquire wait_lock
1066          * and do the state validation.
1067          */
1068         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1069
1070         /*
1071          * Since {uval, pi_state} is serialized by wait_lock, and our current
1072          * uval was read without holding it, it can have changed. Verify it
1073          * still is what we expect it to be, otherwise retry the entire
1074          * operation.
1075          */
1076         if (get_futex_value_locked(&uval2, uaddr))
1077                 goto out_efault;
1078
1079         if (uval != uval2)
1080                 goto out_eagain;
1081
1082         /*
1083          * Handle the owner died case:
1084          */
1085         if (uval & FUTEX_OWNER_DIED) {
1086                 /*
1087                  * exit_pi_state_list sets owner to NULL and wakes the
1088                  * topmost waiter. The task which acquires the
1089                  * pi_state->rt_mutex will fixup owner.
1090                  */
1091                 if (!pi_state->owner) {
1092                         /*
1093                          * No pi state owner, but the user space TID
1094                          * is not 0. Inconsistent state. [5]
1095                          */
1096                         if (pid)
1097                                 goto out_einval;
1098                         /*
1099                          * Take a ref on the state and return success. [4]
1100                          */
1101                         goto out_attach;
1102                 }
1103
1104                 /*
1105                  * If TID is 0, then either the dying owner has not
1106                  * yet executed exit_pi_state_list() or some waiter
1107                  * acquired the rtmutex in the pi state, but did not
1108                  * yet fixup the TID in user space.
1109                  *
1110                  * Take a ref on the state and return success. [6]
1111                  */
1112                 if (!pid)
1113                         goto out_attach;
1114         } else {
1115                 /*
1116                  * If the owner died bit is not set, then the pi_state
1117                  * must have an owner. [7]
1118                  */
1119                 if (!pi_state->owner)
1120                         goto out_einval;
1121         }
1122
1123         /*
1124          * Bail out if user space manipulated the futex value. If pi
1125          * state exists then the owner TID must be the same as the
1126          * user space TID. [9/10]
1127          */
1128         if (pid != task_pid_vnr(pi_state->owner))
1129                 goto out_einval;
1130
1131 out_attach:
1132         get_pi_state(pi_state);
1133         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1134         *ps = pi_state;
1135         return 0;
1136
1137 out_einval:
1138         ret = -EINVAL;
1139         goto out_error;
1140
1141 out_eagain:
1142         ret = -EAGAIN;
1143         goto out_error;
1144
1145 out_efault:
1146         ret = -EFAULT;
1147         goto out_error;
1148
1149 out_error:
1150         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1151         return ret;
1152 }
1153
1154 /*
1155  * Lookup the task for the TID provided from user space and attach to
1156  * it after doing proper sanity checks.
1157  */
1158 static int attach_to_pi_owner(u32 uval, union futex_key *key,
1159                               struct futex_pi_state **ps)
1160 {
1161         pid_t pid = uval & FUTEX_TID_MASK;
1162         struct futex_pi_state *pi_state;
1163         struct task_struct *p;
1164
1165         /*
1166          * We are the first waiter - try to look up the real owner and attach
1167          * the new pi_state to it, but bail out when TID = 0 [1]
1168          */
1169         if (!pid)
1170                 return -ESRCH;
1171         p = find_get_task_by_vpid(pid);
1172         if (!p)
1173                 return -ESRCH;
1174
1175         if (unlikely(p->flags & PF_KTHREAD)) {
1176                 put_task_struct(p);
1177                 return -EPERM;
1178         }
1179
1180         /*
1181          * We need to look at the task state flags to figure out,
1182          * whether the task is exiting. To protect against the do_exit
1183          * change of the task flags, we do this protected by
1184          * p->pi_lock:
1185          */
1186         raw_spin_lock_irq(&p->pi_lock);
1187         if (unlikely(p->flags & PF_EXITING)) {
1188                 /*
1189                  * The task is on the way out. When PF_EXITPIDONE is
1190                  * set, we know that the task has finished the
1191                  * cleanup:
1192                  */
1193                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1194
1195                 raw_spin_unlock_irq(&p->pi_lock);
1196                 put_task_struct(p);
1197                 return ret;
1198         }
1199
1200         /*
1201          * No existing pi state. First waiter. [2]
1202          *
1203          * This creates pi_state, we have hb->lock held, this means nothing can
1204          * observe this state, wait_lock is irrelevant.
1205          */
1206         pi_state = alloc_pi_state();
1207
1208         /*
1209          * Initialize the pi_mutex in locked state and make @p
1210          * the owner of it:
1211          */
1212         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1213
1214         /* Store the key for possible exit cleanups: */
1215         pi_state->key = *key;
1216
1217         WARN_ON(!list_empty(&pi_state->list));
1218         list_add(&pi_state->list, &p->pi_state_list);
1219         /*
1220          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1221          * because there is no concurrency as the object is not published yet.
1222          */
1223         pi_state->owner = p;
1224         raw_spin_unlock_irq(&p->pi_lock);
1225
1226         put_task_struct(p);
1227
1228         *ps = pi_state;
1229
1230         return 0;
1231 }
1232
1233 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1234                            struct futex_hash_bucket *hb,
1235                            union futex_key *key, struct futex_pi_state **ps)
1236 {
1237         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1238
1239         /*
1240          * If there is a waiter on that futex, validate it and
1241          * attach to the pi_state when the validation succeeds.
1242          */
1243         if (top_waiter)
1244                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1245
1246         /*
1247          * We are the first waiter - try to look up the owner based on
1248          * @uval and attach to it.
1249          */
1250         return attach_to_pi_owner(uval, key, ps);
1251 }
1252
1253 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1254 {
1255         u32 uninitialized_var(curval);
1256
1257         if (unlikely(should_fail_futex(true)))
1258                 return -EFAULT;
1259
1260         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1261                 return -EFAULT;
1262
1263         /* If user space value changed, let the caller retry */
1264         return curval != uval ? -EAGAIN : 0;
1265 }
1266
1267 /**
1268  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1269  * @uaddr:              the pi futex user address
1270  * @hb:                 the pi futex hash bucket
1271  * @key:                the futex key associated with uaddr and hb
1272  * @ps:                 the pi_state pointer where we store the result of the
1273  *                      lookup
1274  * @task:               the task to perform the atomic lock work for.  This will
1275  *                      be "current" except in the case of requeue pi.
1276  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1277  *
1278  * Return:
1279  *  -  0 - ready to wait;
1280  *  -  1 - acquired the lock;
1281  *  - <0 - error
1282  *
1283  * The hb->lock and futex_key refs shall be held by the caller.
1284  */
1285 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1286                                 union futex_key *key,
1287                                 struct futex_pi_state **ps,
1288                                 struct task_struct *task, int set_waiters)
1289 {
1290         u32 uval, newval, vpid = task_pid_vnr(task);
1291         struct futex_q *top_waiter;
1292         int ret;
1293
1294         /*
1295          * Read the user space value first so we can validate a few
1296          * things before proceeding further.
1297          */
1298         if (get_futex_value_locked(&uval, uaddr))
1299                 return -EFAULT;
1300
1301         if (unlikely(should_fail_futex(true)))
1302                 return -EFAULT;
1303
1304         /*
1305          * Detect deadlocks.
1306          */
1307         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1308                 return -EDEADLK;
1309
1310         if ((unlikely(should_fail_futex(true))))
1311                 return -EDEADLK;
1312
1313         /*
1314          * Lookup existing state first. If it exists, try to attach to
1315          * its pi_state.
1316          */
1317         top_waiter = futex_top_waiter(hb, key);
1318         if (top_waiter)
1319                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1320
1321         /*
1322          * No waiter and user TID is 0. We are here because the
1323          * waiters or the owner died bit is set or called from
1324          * requeue_cmp_pi or for whatever reason something took the
1325          * syscall.
1326          */
1327         if (!(uval & FUTEX_TID_MASK)) {
1328                 /*
1329                  * We take over the futex. No other waiters and the user space
1330                  * TID is 0. We preserve the owner died bit.
1331                  */
1332                 newval = uval & FUTEX_OWNER_DIED;
1333                 newval |= vpid;
1334
1335                 /* The futex requeue_pi code can enforce the waiters bit */
1336                 if (set_waiters)
1337                         newval |= FUTEX_WAITERS;
1338
1339                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1340                 /* If the take over worked, return 1 */
1341                 return ret < 0 ? ret : 1;
1342         }
1343
1344         /*
1345          * First waiter. Set the waiters bit before attaching ourself to
1346          * the owner. If owner tries to unlock, it will be forced into
1347          * the kernel and blocked on hb->lock.
1348          */
1349         newval = uval | FUTEX_WAITERS;
1350         ret = lock_pi_update_atomic(uaddr, uval, newval);
1351         if (ret)
1352                 return ret;
1353         /*
1354          * If the update of the user space value succeeded, we try to
1355          * attach to the owner. If that fails, no harm done, we only
1356          * set the FUTEX_WAITERS bit in the user space variable.
1357          */
1358         return attach_to_pi_owner(uval, key, ps);
1359 }
1360
1361 /**
1362  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1363  * @q:  The futex_q to unqueue
1364  *
1365  * The q->lock_ptr must not be NULL and must be held by the caller.
1366  */
1367 static void __unqueue_futex(struct futex_q *q)
1368 {
1369         struct futex_hash_bucket *hb;
1370
1371         if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1372                 return;
1373         lockdep_assert_held(q->lock_ptr);
1374
1375         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1376         plist_del(&q->list, &hb->chain);
1377         hb_waiters_dec(hb);
1378 }
1379
1380 /*
1381  * The hash bucket lock must be held when this is called.
1382  * Afterwards, the futex_q must not be accessed. Callers
1383  * must ensure to later call wake_up_q() for the actual
1384  * wakeups to occur.
1385  */
1386 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1387 {
1388         struct task_struct *p = q->task;
1389
1390         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1391                 return;
1392
1393         /*
1394          * Queue the task for later wakeup for after we've released
1395          * the hb->lock. wake_q_add() grabs reference to p.
1396          */
1397         wake_q_add(wake_q, p);
1398         __unqueue_futex(q);
1399         /*
1400          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1401          * is written, without taking any locks. This is possible in the event
1402          * of a spurious wakeup, for example. A memory barrier is required here
1403          * to prevent the following store to lock_ptr from getting ahead of the
1404          * plist_del in __unqueue_futex().
1405          */
1406         smp_store_release(&q->lock_ptr, NULL);
1407 }
1408
1409 /*
1410  * Caller must hold a reference on @pi_state.
1411  */
1412 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1413 {
1414         u32 uninitialized_var(curval), newval;
1415         struct task_struct *new_owner;
1416         bool postunlock = false;
1417         DEFINE_WAKE_Q(wake_q);
1418         int ret = 0;
1419
1420         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1421         if (WARN_ON_ONCE(!new_owner)) {
1422                 /*
1423                  * As per the comment in futex_unlock_pi() this should not happen.
1424                  *
1425                  * When this happens, give up our locks and try again, giving
1426                  * the futex_lock_pi() instance time to complete, either by
1427                  * waiting on the rtmutex or removing itself from the futex
1428                  * queue.
1429                  */
1430                 ret = -EAGAIN;
1431                 goto out_unlock;
1432         }
1433
1434         /*
1435          * We pass it to the next owner. The WAITERS bit is always kept
1436          * enabled while there is PI state around. We cleanup the owner
1437          * died bit, because we are the owner.
1438          */
1439         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1440
1441         if (unlikely(should_fail_futex(true)))
1442                 ret = -EFAULT;
1443
1444         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1445                 ret = -EFAULT;
1446
1447         } else if (curval != uval) {
1448                 /*
1449                  * If a unconditional UNLOCK_PI operation (user space did not
1450                  * try the TID->0 transition) raced with a waiter setting the
1451                  * FUTEX_WAITERS flag between get_user() and locking the hash
1452                  * bucket lock, retry the operation.
1453                  */
1454                 if ((FUTEX_TID_MASK & curval) == uval)
1455                         ret = -EAGAIN;
1456                 else
1457                         ret = -EINVAL;
1458         }
1459
1460         if (ret)
1461                 goto out_unlock;
1462
1463         /*
1464          * This is a point of no return; once we modify the uval there is no
1465          * going back and subsequent operations must not fail.
1466          */
1467
1468         raw_spin_lock(&pi_state->owner->pi_lock);
1469         WARN_ON(list_empty(&pi_state->list));
1470         list_del_init(&pi_state->list);
1471         raw_spin_unlock(&pi_state->owner->pi_lock);
1472
1473         raw_spin_lock(&new_owner->pi_lock);
1474         WARN_ON(!list_empty(&pi_state->list));
1475         list_add(&pi_state->list, &new_owner->pi_state_list);
1476         pi_state->owner = new_owner;
1477         raw_spin_unlock(&new_owner->pi_lock);
1478
1479         postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1480
1481 out_unlock:
1482         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1483
1484         if (postunlock)
1485                 rt_mutex_postunlock(&wake_q);
1486
1487         return ret;
1488 }
1489
1490 /*
1491  * Express the locking dependencies for lockdep:
1492  */
1493 static inline void
1494 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1495 {
1496         if (hb1 <= hb2) {
1497                 spin_lock(&hb1->lock);
1498                 if (hb1 < hb2)
1499                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1500         } else { /* hb1 > hb2 */
1501                 spin_lock(&hb2->lock);
1502                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1503         }
1504 }
1505
1506 static inline void
1507 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1508 {
1509         spin_unlock(&hb1->lock);
1510         if (hb1 != hb2)
1511                 spin_unlock(&hb2->lock);
1512 }
1513
1514 /*
1515  * Wake up waiters matching bitset queued on this futex (uaddr).
1516  */
1517 static int
1518 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1519 {
1520         struct futex_hash_bucket *hb;
1521         struct futex_q *this, *next;
1522         union futex_key key = FUTEX_KEY_INIT;
1523         int ret;
1524         DEFINE_WAKE_Q(wake_q);
1525
1526         if (!bitset)
1527                 return -EINVAL;
1528
1529         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1530         if (unlikely(ret != 0))
1531                 goto out;
1532
1533         hb = hash_futex(&key);
1534
1535         /* Make sure we really have tasks to wakeup */
1536         if (!hb_waiters_pending(hb))
1537                 goto out_put_key;
1538
1539         spin_lock(&hb->lock);
1540
1541         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1542                 if (match_futex (&this->key, &key)) {
1543                         if (this->pi_state || this->rt_waiter) {
1544                                 ret = -EINVAL;
1545                                 break;
1546                         }
1547
1548                         /* Check if one of the bits is set in both bitsets */
1549                         if (!(this->bitset & bitset))
1550                                 continue;
1551
1552                         mark_wake_futex(&wake_q, this);
1553                         if (++ret >= nr_wake)
1554                                 break;
1555                 }
1556         }
1557
1558         spin_unlock(&hb->lock);
1559         wake_up_q(&wake_q);
1560 out_put_key:
1561         put_futex_key(&key);
1562 out:
1563         return ret;
1564 }
1565
1566 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1567 {
1568         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1569         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1570         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1571         int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1572         int oldval, ret;
1573
1574         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1575                 if (oparg < 0 || oparg > 31) {
1576                         char comm[sizeof(current->comm)];
1577                         /*
1578                          * kill this print and return -EINVAL when userspace
1579                          * is sane again
1580                          */
1581                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1582                                         get_task_comm(comm, current), oparg);
1583                         oparg &= 31;
1584                 }
1585                 oparg = 1 << oparg;
1586         }
1587
1588         if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1589                 return -EFAULT;
1590
1591         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1592         if (ret)
1593                 return ret;
1594
1595         switch (cmp) {
1596         case FUTEX_OP_CMP_EQ:
1597                 return oldval == cmparg;
1598         case FUTEX_OP_CMP_NE:
1599                 return oldval != cmparg;
1600         case FUTEX_OP_CMP_LT:
1601                 return oldval < cmparg;
1602         case FUTEX_OP_CMP_GE:
1603                 return oldval >= cmparg;
1604         case FUTEX_OP_CMP_LE:
1605                 return oldval <= cmparg;
1606         case FUTEX_OP_CMP_GT:
1607                 return oldval > cmparg;
1608         default:
1609                 return -ENOSYS;
1610         }
1611 }
1612
1613 /*
1614  * Wake up all waiters hashed on the physical page that is mapped
1615  * to this virtual address:
1616  */
1617 static int
1618 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1619               int nr_wake, int nr_wake2, int op)
1620 {
1621         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1622         struct futex_hash_bucket *hb1, *hb2;
1623         struct futex_q *this, *next;
1624         int ret, op_ret;
1625         DEFINE_WAKE_Q(wake_q);
1626
1627 retry:
1628         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1629         if (unlikely(ret != 0))
1630                 goto out;
1631         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1632         if (unlikely(ret != 0))
1633                 goto out_put_key1;
1634
1635         hb1 = hash_futex(&key1);
1636         hb2 = hash_futex(&key2);
1637
1638 retry_private:
1639         double_lock_hb(hb1, hb2);
1640         op_ret = futex_atomic_op_inuser(op, uaddr2);
1641         if (unlikely(op_ret < 0)) {
1642
1643                 double_unlock_hb(hb1, hb2);
1644
1645 #ifndef CONFIG_MMU
1646                 /*
1647                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1648                  * but we might get them from range checking
1649                  */
1650                 ret = op_ret;
1651                 goto out_put_keys;
1652 #endif
1653
1654                 if (unlikely(op_ret != -EFAULT)) {
1655                         ret = op_ret;
1656                         goto out_put_keys;
1657                 }
1658
1659                 ret = fault_in_user_writeable(uaddr2);
1660                 if (ret)
1661                         goto out_put_keys;
1662
1663                 if (!(flags & FLAGS_SHARED))
1664                         goto retry_private;
1665
1666                 put_futex_key(&key2);
1667                 put_futex_key(&key1);
1668                 goto retry;
1669         }
1670
1671         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1672                 if (match_futex (&this->key, &key1)) {
1673                         if (this->pi_state || this->rt_waiter) {
1674                                 ret = -EINVAL;
1675                                 goto out_unlock;
1676                         }
1677                         mark_wake_futex(&wake_q, this);
1678                         if (++ret >= nr_wake)
1679                                 break;
1680                 }
1681         }
1682
1683         if (op_ret > 0) {
1684                 op_ret = 0;
1685                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1686                         if (match_futex (&this->key, &key2)) {
1687                                 if (this->pi_state || this->rt_waiter) {
1688                                         ret = -EINVAL;
1689                                         goto out_unlock;
1690                                 }
1691                                 mark_wake_futex(&wake_q, this);
1692                                 if (++op_ret >= nr_wake2)
1693                                         break;
1694                         }
1695                 }
1696                 ret += op_ret;
1697         }
1698
1699 out_unlock:
1700         double_unlock_hb(hb1, hb2);
1701         wake_up_q(&wake_q);
1702 out_put_keys:
1703         put_futex_key(&key2);
1704 out_put_key1:
1705         put_futex_key(&key1);
1706 out:
1707         return ret;
1708 }
1709
1710 /**
1711  * requeue_futex() - Requeue a futex_q from one hb to another
1712  * @q:          the futex_q to requeue
1713  * @hb1:        the source hash_bucket
1714  * @hb2:        the target hash_bucket
1715  * @key2:       the new key for the requeued futex_q
1716  */
1717 static inline
1718 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1719                    struct futex_hash_bucket *hb2, union futex_key *key2)
1720 {
1721
1722         /*
1723          * If key1 and key2 hash to the same bucket, no need to
1724          * requeue.
1725          */
1726         if (likely(&hb1->chain != &hb2->chain)) {
1727                 plist_del(&q->list, &hb1->chain);
1728                 hb_waiters_dec(hb1);
1729                 hb_waiters_inc(hb2);
1730                 plist_add(&q->list, &hb2->chain);
1731                 q->lock_ptr = &hb2->lock;
1732         }
1733         get_futex_key_refs(key2);
1734         q->key = *key2;
1735 }
1736
1737 /**
1738  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1739  * @q:          the futex_q
1740  * @key:        the key of the requeue target futex
1741  * @hb:         the hash_bucket of the requeue target futex
1742  *
1743  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1744  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1745  * to the requeue target futex so the waiter can detect the wakeup on the right
1746  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1747  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1748  * to protect access to the pi_state to fixup the owner later.  Must be called
1749  * with both q->lock_ptr and hb->lock held.
1750  */
1751 static inline
1752 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1753                            struct futex_hash_bucket *hb)
1754 {
1755         get_futex_key_refs(key);
1756         q->key = *key;
1757
1758         __unqueue_futex(q);
1759
1760         WARN_ON(!q->rt_waiter);
1761         q->rt_waiter = NULL;
1762
1763         q->lock_ptr = &hb->lock;
1764
1765         wake_up_state(q->task, TASK_NORMAL);
1766 }
1767
1768 /**
1769  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1770  * @pifutex:            the user address of the to futex
1771  * @hb1:                the from futex hash bucket, must be locked by the caller
1772  * @hb2:                the to futex hash bucket, must be locked by the caller
1773  * @key1:               the from futex key
1774  * @key2:               the to futex key
1775  * @ps:                 address to store the pi_state pointer
1776  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1777  *
1778  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1779  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1780  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1781  * hb1 and hb2 must be held by the caller.
1782  *
1783  * Return:
1784  *  -  0 - failed to acquire the lock atomically;
1785  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1786  *  - <0 - error
1787  */
1788 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1789                                  struct futex_hash_bucket *hb1,
1790                                  struct futex_hash_bucket *hb2,
1791                                  union futex_key *key1, union futex_key *key2,
1792                                  struct futex_pi_state **ps, int set_waiters)
1793 {
1794         struct futex_q *top_waiter = NULL;
1795         u32 curval;
1796         int ret, vpid;
1797
1798         if (get_futex_value_locked(&curval, pifutex))
1799                 return -EFAULT;
1800
1801         if (unlikely(should_fail_futex(true)))
1802                 return -EFAULT;
1803
1804         /*
1805          * Find the top_waiter and determine if there are additional waiters.
1806          * If the caller intends to requeue more than 1 waiter to pifutex,
1807          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1808          * as we have means to handle the possible fault.  If not, don't set
1809          * the bit unecessarily as it will force the subsequent unlock to enter
1810          * the kernel.
1811          */
1812         top_waiter = futex_top_waiter(hb1, key1);
1813
1814         /* There are no waiters, nothing for us to do. */
1815         if (!top_waiter)
1816                 return 0;
1817
1818         /* Ensure we requeue to the expected futex. */
1819         if (!match_futex(top_waiter->requeue_pi_key, key2))
1820                 return -EINVAL;
1821
1822         /*
1823          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1824          * the contended case or if set_waiters is 1.  The pi_state is returned
1825          * in ps in contended cases.
1826          */
1827         vpid = task_pid_vnr(top_waiter->task);
1828         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1829                                    set_waiters);
1830         if (ret == 1) {
1831                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1832                 return vpid;
1833         }
1834         return ret;
1835 }
1836
1837 /**
1838  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1839  * @uaddr1:     source futex user address
1840  * @flags:      futex flags (FLAGS_SHARED, etc.)
1841  * @uaddr2:     target futex user address
1842  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1843  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1844  * @cmpval:     @uaddr1 expected value (or %NULL)
1845  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1846  *              pi futex (pi to pi requeue is not supported)
1847  *
1848  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1849  * uaddr2 atomically on behalf of the top waiter.
1850  *
1851  * Return:
1852  *  - >=0 - on success, the number of tasks requeued or woken;
1853  *  -  <0 - on error
1854  */
1855 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1856                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1857                          u32 *cmpval, int requeue_pi)
1858 {
1859         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1860         int drop_count = 0, task_count = 0, ret;
1861         struct futex_pi_state *pi_state = NULL;
1862         struct futex_hash_bucket *hb1, *hb2;
1863         struct futex_q *this, *next;
1864         DEFINE_WAKE_Q(wake_q);
1865
1866         if (nr_wake < 0 || nr_requeue < 0)
1867                 return -EINVAL;
1868
1869         /*
1870          * When PI not supported: return -ENOSYS if requeue_pi is true,
1871          * consequently the compiler knows requeue_pi is always false past
1872          * this point which will optimize away all the conditional code
1873          * further down.
1874          */
1875         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1876                 return -ENOSYS;
1877
1878         if (requeue_pi) {
1879                 /*
1880                  * Requeue PI only works on two distinct uaddrs. This
1881                  * check is only valid for private futexes. See below.
1882                  */
1883                 if (uaddr1 == uaddr2)
1884                         return -EINVAL;
1885
1886                 /*
1887                  * requeue_pi requires a pi_state, try to allocate it now
1888                  * without any locks in case it fails.
1889                  */
1890                 if (refill_pi_state_cache())
1891                         return -ENOMEM;
1892                 /*
1893                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1894                  * + nr_requeue, since it acquires the rt_mutex prior to
1895                  * returning to userspace, so as to not leave the rt_mutex with
1896                  * waiters and no owner.  However, second and third wake-ups
1897                  * cannot be predicted as they involve race conditions with the
1898                  * first wake and a fault while looking up the pi_state.  Both
1899                  * pthread_cond_signal() and pthread_cond_broadcast() should
1900                  * use nr_wake=1.
1901                  */
1902                 if (nr_wake != 1)
1903                         return -EINVAL;
1904         }
1905
1906 retry:
1907         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1908         if (unlikely(ret != 0))
1909                 goto out;
1910         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1911                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1912         if (unlikely(ret != 0))
1913                 goto out_put_key1;
1914
1915         /*
1916          * The check above which compares uaddrs is not sufficient for
1917          * shared futexes. We need to compare the keys:
1918          */
1919         if (requeue_pi && match_futex(&key1, &key2)) {
1920                 ret = -EINVAL;
1921                 goto out_put_keys;
1922         }
1923
1924         hb1 = hash_futex(&key1);
1925         hb2 = hash_futex(&key2);
1926
1927 retry_private:
1928         hb_waiters_inc(hb2);
1929         double_lock_hb(hb1, hb2);
1930
1931         if (likely(cmpval != NULL)) {
1932                 u32 curval;
1933
1934                 ret = get_futex_value_locked(&curval, uaddr1);
1935
1936                 if (unlikely(ret)) {
1937                         double_unlock_hb(hb1, hb2);
1938                         hb_waiters_dec(hb2);
1939
1940                         ret = get_user(curval, uaddr1);
1941                         if (ret)
1942                                 goto out_put_keys;
1943
1944                         if (!(flags & FLAGS_SHARED))
1945                                 goto retry_private;
1946
1947                         put_futex_key(&key2);
1948                         put_futex_key(&key1);
1949                         goto retry;
1950                 }
1951                 if (curval != *cmpval) {
1952                         ret = -EAGAIN;
1953                         goto out_unlock;
1954                 }
1955         }
1956
1957         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1958                 /*
1959                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1960                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1961                  * bit.  We force this here where we are able to easily handle
1962                  * faults rather in the requeue loop below.
1963                  */
1964                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1965                                                  &key2, &pi_state, nr_requeue);
1966
1967                 /*
1968                  * At this point the top_waiter has either taken uaddr2 or is
1969                  * waiting on it.  If the former, then the pi_state will not
1970                  * exist yet, look it up one more time to ensure we have a
1971                  * reference to it. If the lock was taken, ret contains the
1972                  * vpid of the top waiter task.
1973                  * If the lock was not taken, we have pi_state and an initial
1974                  * refcount on it. In case of an error we have nothing.
1975                  */
1976                 if (ret > 0) {
1977                         WARN_ON(pi_state);
1978                         drop_count++;
1979                         task_count++;
1980                         /*
1981                          * If we acquired the lock, then the user space value
1982                          * of uaddr2 should be vpid. It cannot be changed by
1983                          * the top waiter as it is blocked on hb2 lock if it
1984                          * tries to do so. If something fiddled with it behind
1985                          * our back the pi state lookup might unearth it. So
1986                          * we rather use the known value than rereading and
1987                          * handing potential crap to lookup_pi_state.
1988                          *
1989                          * If that call succeeds then we have pi_state and an
1990                          * initial refcount on it.
1991                          */
1992                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
1993                 }
1994
1995                 switch (ret) {
1996                 case 0:
1997                         /* We hold a reference on the pi state. */
1998                         break;
1999
2000                         /* If the above failed, then pi_state is NULL */
2001                 case -EFAULT:
2002                         double_unlock_hb(hb1, hb2);
2003                         hb_waiters_dec(hb2);
2004                         put_futex_key(&key2);
2005                         put_futex_key(&key1);
2006                         ret = fault_in_user_writeable(uaddr2);
2007                         if (!ret)
2008                                 goto retry;
2009                         goto out;
2010                 case -EAGAIN:
2011                         /*
2012                          * Two reasons for this:
2013                          * - Owner is exiting and we just wait for the
2014                          *   exit to complete.
2015                          * - The user space value changed.
2016                          */
2017                         double_unlock_hb(hb1, hb2);
2018                         hb_waiters_dec(hb2);
2019                         put_futex_key(&key2);
2020                         put_futex_key(&key1);
2021                         cond_resched();
2022                         goto retry;
2023                 default:
2024                         goto out_unlock;
2025                 }
2026         }
2027
2028         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2029                 if (task_count - nr_wake >= nr_requeue)
2030                         break;
2031
2032                 if (!match_futex(&this->key, &key1))
2033                         continue;
2034
2035                 /*
2036                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2037                  * be paired with each other and no other futex ops.
2038                  *
2039                  * We should never be requeueing a futex_q with a pi_state,
2040                  * which is awaiting a futex_unlock_pi().
2041                  */
2042                 if ((requeue_pi && !this->rt_waiter) ||
2043                     (!requeue_pi && this->rt_waiter) ||
2044                     this->pi_state) {
2045                         ret = -EINVAL;
2046                         break;
2047                 }
2048
2049                 /*
2050                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2051                  * lock, we already woke the top_waiter.  If not, it will be
2052                  * woken by futex_unlock_pi().
2053                  */
2054                 if (++task_count <= nr_wake && !requeue_pi) {
2055                         mark_wake_futex(&wake_q, this);
2056                         continue;
2057                 }
2058
2059                 /* Ensure we requeue to the expected futex for requeue_pi. */
2060                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2061                         ret = -EINVAL;
2062                         break;
2063                 }
2064
2065                 /*
2066                  * Requeue nr_requeue waiters and possibly one more in the case
2067                  * of requeue_pi if we couldn't acquire the lock atomically.
2068                  */
2069                 if (requeue_pi) {
2070                         /*
2071                          * Prepare the waiter to take the rt_mutex. Take a
2072                          * refcount on the pi_state and store the pointer in
2073                          * the futex_q object of the waiter.
2074                          */
2075                         get_pi_state(pi_state);
2076                         this->pi_state = pi_state;
2077                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2078                                                         this->rt_waiter,
2079                                                         this->task);
2080                         if (ret == 1) {
2081                                 /*
2082                                  * We got the lock. We do neither drop the
2083                                  * refcount on pi_state nor clear
2084                                  * this->pi_state because the waiter needs the
2085                                  * pi_state for cleaning up the user space
2086                                  * value. It will drop the refcount after
2087                                  * doing so.
2088                                  */
2089                                 requeue_pi_wake_futex(this, &key2, hb2);
2090                                 drop_count++;
2091                                 continue;
2092                         } else if (ret) {
2093                                 /*
2094                                  * rt_mutex_start_proxy_lock() detected a
2095                                  * potential deadlock when we tried to queue
2096                                  * that waiter. Drop the pi_state reference
2097                                  * which we took above and remove the pointer
2098                                  * to the state from the waiters futex_q
2099                                  * object.
2100                                  */
2101                                 this->pi_state = NULL;
2102                                 put_pi_state(pi_state);
2103                                 /*
2104                                  * We stop queueing more waiters and let user
2105                                  * space deal with the mess.
2106                                  */
2107                                 break;
2108                         }
2109                 }
2110                 requeue_futex(this, hb1, hb2, &key2);
2111                 drop_count++;
2112         }
2113
2114         /*
2115          * We took an extra initial reference to the pi_state either
2116          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2117          * need to drop it here again.
2118          */
2119         put_pi_state(pi_state);
2120
2121 out_unlock:
2122         double_unlock_hb(hb1, hb2);
2123         wake_up_q(&wake_q);
2124         hb_waiters_dec(hb2);
2125
2126         /*
2127          * drop_futex_key_refs() must be called outside the spinlocks. During
2128          * the requeue we moved futex_q's from the hash bucket at key1 to the
2129          * one at key2 and updated their key pointer.  We no longer need to
2130          * hold the references to key1.
2131          */
2132         while (--drop_count >= 0)
2133                 drop_futex_key_refs(&key1);
2134
2135 out_put_keys:
2136         put_futex_key(&key2);
2137 out_put_key1:
2138         put_futex_key(&key1);
2139 out:
2140         return ret ? ret : task_count;
2141 }
2142
2143 /* The key must be already stored in q->key. */
2144 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2145         __acquires(&hb->lock)
2146 {
2147         struct futex_hash_bucket *hb;
2148
2149         hb = hash_futex(&q->key);
2150
2151         /*
2152          * Increment the counter before taking the lock so that
2153          * a potential waker won't miss a to-be-slept task that is
2154          * waiting for the spinlock. This is safe as all queue_lock()
2155          * users end up calling queue_me(). Similarly, for housekeeping,
2156          * decrement the counter at queue_unlock() when some error has
2157          * occurred and we don't end up adding the task to the list.
2158          */
2159         hb_waiters_inc(hb);
2160
2161         q->lock_ptr = &hb->lock;
2162
2163         spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2164         return hb;
2165 }
2166
2167 static inline void
2168 queue_unlock(struct futex_hash_bucket *hb)
2169         __releases(&hb->lock)
2170 {
2171         spin_unlock(&hb->lock);
2172         hb_waiters_dec(hb);
2173 }
2174
2175 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2176 {
2177         int prio;
2178
2179         /*
2180          * The priority used to register this element is
2181          * - either the real thread-priority for the real-time threads
2182          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2183          * - or MAX_RT_PRIO for non-RT threads.
2184          * Thus, all RT-threads are woken first in priority order, and
2185          * the others are woken last, in FIFO order.
2186          */
2187         prio = min(current->normal_prio, MAX_RT_PRIO);
2188
2189         plist_node_init(&q->list, prio);
2190         plist_add(&q->list, &hb->chain);
2191         q->task = current;
2192 }
2193
2194 /**
2195  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2196  * @q:  The futex_q to enqueue
2197  * @hb: The destination hash bucket
2198  *
2199  * The hb->lock must be held by the caller, and is released here. A call to
2200  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2201  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2202  * or nothing if the unqueue is done as part of the wake process and the unqueue
2203  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2204  * an example).
2205  */
2206 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2207         __releases(&hb->lock)
2208 {
2209         __queue_me(q, hb);
2210         spin_unlock(&hb->lock);
2211 }
2212
2213 /**
2214  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2215  * @q:  The futex_q to unqueue
2216  *
2217  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2218  * be paired with exactly one earlier call to queue_me().
2219  *
2220  * Return:
2221  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2222  *  - 0 - if the futex_q was already removed by the waking thread
2223  */
2224 static int unqueue_me(struct futex_q *q)
2225 {
2226         spinlock_t *lock_ptr;
2227         int ret = 0;
2228
2229         /* In the common case we don't take the spinlock, which is nice. */
2230 retry:
2231         /*
2232          * q->lock_ptr can change between this read and the following spin_lock.
2233          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2234          * optimizing lock_ptr out of the logic below.
2235          */
2236         lock_ptr = READ_ONCE(q->lock_ptr);
2237         if (lock_ptr != NULL) {
2238                 spin_lock(lock_ptr);
2239                 /*
2240                  * q->lock_ptr can change between reading it and
2241                  * spin_lock(), causing us to take the wrong lock.  This
2242                  * corrects the race condition.
2243                  *
2244                  * Reasoning goes like this: if we have the wrong lock,
2245                  * q->lock_ptr must have changed (maybe several times)
2246                  * between reading it and the spin_lock().  It can
2247                  * change again after the spin_lock() but only if it was
2248                  * already changed before the spin_lock().  It cannot,
2249                  * however, change back to the original value.  Therefore
2250                  * we can detect whether we acquired the correct lock.
2251                  */
2252                 if (unlikely(lock_ptr != q->lock_ptr)) {
2253                         spin_unlock(lock_ptr);
2254                         goto retry;
2255                 }
2256                 __unqueue_futex(q);
2257
2258                 BUG_ON(q->pi_state);
2259
2260                 spin_unlock(lock_ptr);
2261                 ret = 1;
2262         }
2263
2264         drop_futex_key_refs(&q->key);
2265         return ret;
2266 }
2267
2268 /*
2269  * PI futexes can not be requeued and must remove themself from the
2270  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2271  * and dropped here.
2272  */
2273 static void unqueue_me_pi(struct futex_q *q)
2274         __releases(q->lock_ptr)
2275 {
2276         __unqueue_futex(q);
2277
2278         BUG_ON(!q->pi_state);
2279         put_pi_state(q->pi_state);
2280         q->pi_state = NULL;
2281
2282         spin_unlock(q->lock_ptr);
2283 }
2284
2285 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2286                                 struct task_struct *argowner)
2287 {
2288         struct futex_pi_state *pi_state = q->pi_state;
2289         u32 uval, uninitialized_var(curval), newval;
2290         struct task_struct *oldowner, *newowner;
2291         u32 newtid;
2292         int ret;
2293
2294         lockdep_assert_held(q->lock_ptr);
2295
2296         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2297
2298         oldowner = pi_state->owner;
2299
2300         /*
2301          * We are here because either:
2302          *
2303          *  - we stole the lock and pi_state->owner needs updating to reflect
2304          *    that (@argowner == current),
2305          *
2306          * or:
2307          *
2308          *  - someone stole our lock and we need to fix things to point to the
2309          *    new owner (@argowner == NULL).
2310          *
2311          * Either way, we have to replace the TID in the user space variable.
2312          * This must be atomic as we have to preserve the owner died bit here.
2313          *
2314          * Note: We write the user space value _before_ changing the pi_state
2315          * because we can fault here. Imagine swapped out pages or a fork
2316          * that marked all the anonymous memory readonly for cow.
2317          *
2318          * Modifying pi_state _before_ the user space value would leave the
2319          * pi_state in an inconsistent state when we fault here, because we
2320          * need to drop the locks to handle the fault. This might be observed
2321          * in the PID check in lookup_pi_state.
2322          */
2323 retry:
2324         if (!argowner) {
2325                 if (oldowner != current) {
2326                         /*
2327                          * We raced against a concurrent self; things are
2328                          * already fixed up. Nothing to do.
2329                          */
2330                         ret = 0;
2331                         goto out_unlock;
2332                 }
2333
2334                 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2335                         /* We got the lock after all, nothing to fix. */
2336                         ret = 0;
2337                         goto out_unlock;
2338                 }
2339
2340                 /*
2341                  * Since we just failed the trylock; there must be an owner.
2342                  */
2343                 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2344                 BUG_ON(!newowner);
2345         } else {
2346                 WARN_ON_ONCE(argowner != current);
2347                 if (oldowner == current) {
2348                         /*
2349                          * We raced against a concurrent self; things are
2350                          * already fixed up. Nothing to do.
2351                          */
2352                         ret = 0;
2353                         goto out_unlock;
2354                 }
2355                 newowner = argowner;
2356         }
2357
2358         newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2359         /* Owner died? */
2360         if (!pi_state->owner)
2361                 newtid |= FUTEX_OWNER_DIED;
2362
2363         if (get_futex_value_locked(&uval, uaddr))
2364                 goto handle_fault;
2365
2366         for (;;) {
2367                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2368
2369                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2370                         goto handle_fault;
2371                 if (curval == uval)
2372                         break;
2373                 uval = curval;
2374         }
2375
2376         /*
2377          * We fixed up user space. Now we need to fix the pi_state
2378          * itself.
2379          */
2380         if (pi_state->owner != NULL) {
2381                 raw_spin_lock(&pi_state->owner->pi_lock);
2382                 WARN_ON(list_empty(&pi_state->list));
2383                 list_del_init(&pi_state->list);
2384                 raw_spin_unlock(&pi_state->owner->pi_lock);
2385         }
2386
2387         pi_state->owner = newowner;
2388
2389         raw_spin_lock(&newowner->pi_lock);
2390         WARN_ON(!list_empty(&pi_state->list));
2391         list_add(&pi_state->list, &newowner->pi_state_list);
2392         raw_spin_unlock(&newowner->pi_lock);
2393         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2394
2395         return 0;
2396
2397         /*
2398          * To handle the page fault we need to drop the locks here. That gives
2399          * the other task (either the highest priority waiter itself or the
2400          * task which stole the rtmutex) the chance to try the fixup of the
2401          * pi_state. So once we are back from handling the fault we need to
2402          * check the pi_state after reacquiring the locks and before trying to
2403          * do another fixup. When the fixup has been done already we simply
2404          * return.
2405          *
2406          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2407          * drop hb->lock since the caller owns the hb -> futex_q relation.
2408          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2409          */
2410 handle_fault:
2411         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2412         spin_unlock(q->lock_ptr);
2413
2414         ret = fault_in_user_writeable(uaddr);
2415
2416         spin_lock(q->lock_ptr);
2417         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2418
2419         /*
2420          * Check if someone else fixed it for us:
2421          */
2422         if (pi_state->owner != oldowner) {
2423                 ret = 0;
2424                 goto out_unlock;
2425         }
2426
2427         if (ret)
2428                 goto out_unlock;
2429
2430         goto retry;
2431
2432 out_unlock:
2433         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2434         return ret;
2435 }
2436
2437 static long futex_wait_restart(struct restart_block *restart);
2438
2439 /**
2440  * fixup_owner() - Post lock pi_state and corner case management
2441  * @uaddr:      user address of the futex
2442  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2443  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2444  *
2445  * After attempting to lock an rt_mutex, this function is called to cleanup
2446  * the pi_state owner as well as handle race conditions that may allow us to
2447  * acquire the lock. Must be called with the hb lock held.
2448  *
2449  * Return:
2450  *  -  1 - success, lock taken;
2451  *  -  0 - success, lock not taken;
2452  *  - <0 - on error (-EFAULT)
2453  */
2454 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2455 {
2456         int ret = 0;
2457
2458         if (locked) {
2459                 /*
2460                  * Got the lock. We might not be the anticipated owner if we
2461                  * did a lock-steal - fix up the PI-state in that case:
2462                  *
2463                  * Speculative pi_state->owner read (we don't hold wait_lock);
2464                  * since we own the lock pi_state->owner == current is the
2465                  * stable state, anything else needs more attention.
2466                  */
2467                 if (q->pi_state->owner != current)
2468                         ret = fixup_pi_state_owner(uaddr, q, current);
2469                 goto out;
2470         }
2471
2472         /*
2473          * If we didn't get the lock; check if anybody stole it from us. In
2474          * that case, we need to fix up the uval to point to them instead of
2475          * us, otherwise bad things happen. [10]
2476          *
2477          * Another speculative read; pi_state->owner == current is unstable
2478          * but needs our attention.
2479          */
2480         if (q->pi_state->owner == current) {
2481                 ret = fixup_pi_state_owner(uaddr, q, NULL);
2482                 goto out;
2483         }
2484
2485         /*
2486          * Paranoia check. If we did not take the lock, then we should not be
2487          * the owner of the rt_mutex.
2488          */
2489         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2490                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2491                                 "pi-state %p\n", ret,
2492                                 q->pi_state->pi_mutex.owner,
2493                                 q->pi_state->owner);
2494         }
2495
2496 out:
2497         return ret ? ret : locked;
2498 }
2499
2500 /**
2501  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2502  * @hb:         the futex hash bucket, must be locked by the caller
2503  * @q:          the futex_q to queue up on
2504  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2505  */
2506 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2507                                 struct hrtimer_sleeper *timeout)
2508 {
2509         /*
2510          * The task state is guaranteed to be set before another task can
2511          * wake it. set_current_state() is implemented using smp_store_mb() and
2512          * queue_me() calls spin_unlock() upon completion, both serializing
2513          * access to the hash list and forcing another memory barrier.
2514          */
2515         set_current_state(TASK_INTERRUPTIBLE);
2516         queue_me(q, hb);
2517
2518         /* Arm the timer */
2519         if (timeout)
2520                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2521
2522         /*
2523          * If we have been removed from the hash list, then another task
2524          * has tried to wake us, and we can skip the call to schedule().
2525          */
2526         if (likely(!plist_node_empty(&q->list))) {
2527                 /*
2528                  * If the timer has already expired, current will already be
2529                  * flagged for rescheduling. Only call schedule if there
2530                  * is no timeout, or if it has yet to expire.
2531                  */
2532                 if (!timeout || timeout->task)
2533                         freezable_schedule();
2534         }
2535         __set_current_state(TASK_RUNNING);
2536 }
2537
2538 /**
2539  * futex_wait_setup() - Prepare to wait on a futex
2540  * @uaddr:      the futex userspace address
2541  * @val:        the expected value
2542  * @flags:      futex flags (FLAGS_SHARED, etc.)
2543  * @q:          the associated futex_q
2544  * @hb:         storage for hash_bucket pointer to be returned to caller
2545  *
2546  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2547  * compare it with the expected value.  Handle atomic faults internally.
2548  * Return with the hb lock held and a q.key reference on success, and unlocked
2549  * with no q.key reference on failure.
2550  *
2551  * Return:
2552  *  -  0 - uaddr contains val and hb has been locked;
2553  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2554  */
2555 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2556                            struct futex_q *q, struct futex_hash_bucket **hb)
2557 {
2558         u32 uval;
2559         int ret;
2560
2561         /*
2562          * Access the page AFTER the hash-bucket is locked.
2563          * Order is important:
2564          *
2565          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2566          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2567          *
2568          * The basic logical guarantee of a futex is that it blocks ONLY
2569          * if cond(var) is known to be true at the time of blocking, for
2570          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2571          * would open a race condition where we could block indefinitely with
2572          * cond(var) false, which would violate the guarantee.
2573          *
2574          * On the other hand, we insert q and release the hash-bucket only
2575          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2576          * absorb a wakeup if *uaddr does not match the desired values
2577          * while the syscall executes.
2578          */
2579 retry:
2580         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2581         if (unlikely(ret != 0))
2582                 return ret;
2583
2584 retry_private:
2585         *hb = queue_lock(q);
2586
2587         ret = get_futex_value_locked(&uval, uaddr);
2588
2589         if (ret) {
2590                 queue_unlock(*hb);
2591
2592                 ret = get_user(uval, uaddr);
2593                 if (ret)
2594                         goto out;
2595
2596                 if (!(flags & FLAGS_SHARED))
2597                         goto retry_private;
2598
2599                 put_futex_key(&q->key);
2600                 goto retry;
2601         }
2602
2603         if (uval != val) {
2604                 queue_unlock(*hb);
2605                 ret = -EWOULDBLOCK;
2606         }
2607
2608 out:
2609         if (ret)
2610                 put_futex_key(&q->key);
2611         return ret;
2612 }
2613
2614 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2615                       ktime_t *abs_time, u32 bitset)
2616 {
2617         struct hrtimer_sleeper timeout, *to = NULL;
2618         struct restart_block *restart;
2619         struct futex_hash_bucket *hb;
2620         struct futex_q q = futex_q_init;
2621         int ret;
2622
2623         if (!bitset)
2624                 return -EINVAL;
2625         q.bitset = bitset;
2626
2627         if (abs_time) {
2628                 to = &timeout;
2629
2630                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2631                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2632                                       HRTIMER_MODE_ABS);
2633                 hrtimer_init_sleeper(to, current);
2634                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2635                                              current->timer_slack_ns);
2636         }
2637
2638 retry:
2639         /*
2640          * Prepare to wait on uaddr. On success, holds hb lock and increments
2641          * q.key refs.
2642          */
2643         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2644         if (ret)
2645                 goto out;
2646
2647         /* queue_me and wait for wakeup, timeout, or a signal. */
2648         futex_wait_queue_me(hb, &q, to);
2649
2650         /* If we were woken (and unqueued), we succeeded, whatever. */
2651         ret = 0;
2652         /* unqueue_me() drops q.key ref */
2653         if (!unqueue_me(&q))
2654                 goto out;
2655         ret = -ETIMEDOUT;
2656         if (to && !to->task)
2657                 goto out;
2658
2659         /*
2660          * We expect signal_pending(current), but we might be the
2661          * victim of a spurious wakeup as well.
2662          */
2663         if (!signal_pending(current))
2664                 goto retry;
2665
2666         ret = -ERESTARTSYS;
2667         if (!abs_time)
2668                 goto out;
2669
2670         restart = &current->restart_block;
2671         restart->fn = futex_wait_restart;
2672         restart->futex.uaddr = uaddr;
2673         restart->futex.val = val;
2674         restart->futex.time = *abs_time;
2675         restart->futex.bitset = bitset;
2676         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2677
2678         ret = -ERESTART_RESTARTBLOCK;
2679
2680 out:
2681         if (to) {
2682                 hrtimer_cancel(&to->timer);
2683                 destroy_hrtimer_on_stack(&to->timer);
2684         }
2685         return ret;
2686 }
2687
2688
2689 static long futex_wait_restart(struct restart_block *restart)
2690 {
2691         u32 __user *uaddr = restart->futex.uaddr;
2692         ktime_t t, *tp = NULL;
2693
2694         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2695                 t = restart->futex.time;
2696                 tp = &t;
2697         }
2698         restart->fn = do_no_restart_syscall;
2699
2700         return (long)futex_wait(uaddr, restart->futex.flags,
2701                                 restart->futex.val, tp, restart->futex.bitset);
2702 }
2703
2704
2705 /*
2706  * Userspace tried a 0 -> TID atomic transition of the futex value
2707  * and failed. The kernel side here does the whole locking operation:
2708  * if there are waiters then it will block as a consequence of relying
2709  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2710  * a 0 value of the futex too.).
2711  *
2712  * Also serves as futex trylock_pi()'ing, and due semantics.
2713  */
2714 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2715                          ktime_t *time, int trylock)
2716 {
2717         struct hrtimer_sleeper timeout, *to = NULL;
2718         struct futex_pi_state *pi_state = NULL;
2719         struct rt_mutex_waiter rt_waiter;
2720         struct futex_hash_bucket *hb;
2721         struct futex_q q = futex_q_init;
2722         int res, ret;
2723
2724         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2725                 return -ENOSYS;
2726
2727         if (refill_pi_state_cache())
2728                 return -ENOMEM;
2729
2730         if (time) {
2731                 to = &timeout;
2732                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2733                                       HRTIMER_MODE_ABS);
2734                 hrtimer_init_sleeper(to, current);
2735                 hrtimer_set_expires(&to->timer, *time);
2736         }
2737
2738 retry:
2739         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2740         if (unlikely(ret != 0))
2741                 goto out;
2742
2743 retry_private:
2744         hb = queue_lock(&q);
2745
2746         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2747         if (unlikely(ret)) {
2748                 /*
2749                  * Atomic work succeeded and we got the lock,
2750                  * or failed. Either way, we do _not_ block.
2751                  */
2752                 switch (ret) {
2753                 case 1:
2754                         /* We got the lock. */
2755                         ret = 0;
2756                         goto out_unlock_put_key;
2757                 case -EFAULT:
2758                         goto uaddr_faulted;
2759                 case -EAGAIN:
2760                         /*
2761                          * Two reasons for this:
2762                          * - Task is exiting and we just wait for the
2763                          *   exit to complete.
2764                          * - The user space value changed.
2765                          */
2766                         queue_unlock(hb);
2767                         put_futex_key(&q.key);
2768                         cond_resched();
2769                         goto retry;
2770                 default:
2771                         goto out_unlock_put_key;
2772                 }
2773         }
2774
2775         WARN_ON(!q.pi_state);
2776
2777         /*
2778          * Only actually queue now that the atomic ops are done:
2779          */
2780         __queue_me(&q, hb);
2781
2782         if (trylock) {
2783                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2784                 /* Fixup the trylock return value: */
2785                 ret = ret ? 0 : -EWOULDBLOCK;
2786                 goto no_block;
2787         }
2788
2789         rt_mutex_init_waiter(&rt_waiter);
2790
2791         /*
2792          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2793          * hold it while doing rt_mutex_start_proxy(), because then it will
2794          * include hb->lock in the blocking chain, even through we'll not in
2795          * fact hold it while blocking. This will lead it to report -EDEADLK
2796          * and BUG when futex_unlock_pi() interleaves with this.
2797          *
2798          * Therefore acquire wait_lock while holding hb->lock, but drop the
2799          * latter before calling rt_mutex_start_proxy_lock(). This still fully
2800          * serializes against futex_unlock_pi() as that does the exact same
2801          * lock handoff sequence.
2802          */
2803         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2804         spin_unlock(q.lock_ptr);
2805         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2806         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2807
2808         if (ret) {
2809                 if (ret == 1)
2810                         ret = 0;
2811
2812                 spin_lock(q.lock_ptr);
2813                 goto no_block;
2814         }
2815
2816
2817         if (unlikely(to))
2818                 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2819
2820         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2821
2822         spin_lock(q.lock_ptr);
2823         /*
2824          * If we failed to acquire the lock (signal/timeout), we must
2825          * first acquire the hb->lock before removing the lock from the
2826          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2827          * wait lists consistent.
2828          *
2829          * In particular; it is important that futex_unlock_pi() can not
2830          * observe this inconsistency.
2831          */
2832         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2833                 ret = 0;
2834
2835 no_block:
2836         /*
2837          * Fixup the pi_state owner and possibly acquire the lock if we
2838          * haven't already.
2839          */
2840         res = fixup_owner(uaddr, &q, !ret);
2841         /*
2842          * If fixup_owner() returned an error, proprogate that.  If it acquired
2843          * the lock, clear our -ETIMEDOUT or -EINTR.
2844          */
2845         if (res)
2846                 ret = (res < 0) ? res : 0;
2847
2848         /*
2849          * If fixup_owner() faulted and was unable to handle the fault, unlock
2850          * it and return the fault to userspace.
2851          */
2852         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2853                 pi_state = q.pi_state;
2854                 get_pi_state(pi_state);
2855         }
2856
2857         /* Unqueue and drop the lock */
2858         unqueue_me_pi(&q);
2859
2860         if (pi_state) {
2861                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2862                 put_pi_state(pi_state);
2863         }
2864
2865         goto out_put_key;
2866
2867 out_unlock_put_key:
2868         queue_unlock(hb);
2869
2870 out_put_key:
2871         put_futex_key(&q.key);
2872 out:
2873         if (to) {
2874                 hrtimer_cancel(&to->timer);
2875                 destroy_hrtimer_on_stack(&to->timer);
2876         }
2877         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2878
2879 uaddr_faulted:
2880         queue_unlock(hb);
2881
2882         ret = fault_in_user_writeable(uaddr);
2883         if (ret)
2884                 goto out_put_key;
2885
2886         if (!(flags & FLAGS_SHARED))
2887                 goto retry_private;
2888
2889         put_futex_key(&q.key);
2890         goto retry;
2891 }
2892
2893 /*
2894  * Userspace attempted a TID -> 0 atomic transition, and failed.
2895  * This is the in-kernel slowpath: we look up the PI state (if any),
2896  * and do the rt-mutex unlock.
2897  */
2898 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2899 {
2900         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2901         union futex_key key = FUTEX_KEY_INIT;
2902         struct futex_hash_bucket *hb;
2903         struct futex_q *top_waiter;
2904         int ret;
2905
2906         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2907                 return -ENOSYS;
2908
2909 retry:
2910         if (get_user(uval, uaddr))
2911                 return -EFAULT;
2912         /*
2913          * We release only a lock we actually own:
2914          */
2915         if ((uval & FUTEX_TID_MASK) != vpid)
2916                 return -EPERM;
2917
2918         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2919         if (ret)
2920                 return ret;
2921
2922         hb = hash_futex(&key);
2923         spin_lock(&hb->lock);
2924
2925         /*
2926          * Check waiters first. We do not trust user space values at
2927          * all and we at least want to know if user space fiddled
2928          * with the futex value instead of blindly unlocking.
2929          */
2930         top_waiter = futex_top_waiter(hb, &key);
2931         if (top_waiter) {
2932                 struct futex_pi_state *pi_state = top_waiter->pi_state;
2933
2934                 ret = -EINVAL;
2935                 if (!pi_state)
2936                         goto out_unlock;
2937
2938                 /*
2939                  * If current does not own the pi_state then the futex is
2940                  * inconsistent and user space fiddled with the futex value.
2941                  */
2942                 if (pi_state->owner != current)
2943                         goto out_unlock;
2944
2945                 get_pi_state(pi_state);
2946                 /*
2947                  * By taking wait_lock while still holding hb->lock, we ensure
2948                  * there is no point where we hold neither; and therefore
2949                  * wake_futex_pi() must observe a state consistent with what we
2950                  * observed.
2951                  */
2952                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2953                 spin_unlock(&hb->lock);
2954
2955                 /* drops pi_state->pi_mutex.wait_lock */
2956                 ret = wake_futex_pi(uaddr, uval, pi_state);
2957
2958                 put_pi_state(pi_state);
2959
2960                 /*
2961                  * Success, we're done! No tricky corner cases.
2962                  */
2963                 if (!ret)
2964                         goto out_putkey;
2965                 /*
2966                  * The atomic access to the futex value generated a
2967                  * pagefault, so retry the user-access and the wakeup:
2968                  */
2969                 if (ret == -EFAULT)
2970                         goto pi_faulted;
2971                 /*
2972                  * A unconditional UNLOCK_PI op raced against a waiter
2973                  * setting the FUTEX_WAITERS bit. Try again.
2974                  */
2975                 if (ret == -EAGAIN) {
2976                         put_futex_key(&key);
2977                         goto retry;
2978                 }
2979                 /*
2980                  * wake_futex_pi has detected invalid state. Tell user
2981                  * space.
2982                  */
2983                 goto out_putkey;
2984         }
2985
2986         /*
2987          * We have no kernel internal state, i.e. no waiters in the
2988          * kernel. Waiters which are about to queue themselves are stuck
2989          * on hb->lock. So we can safely ignore them. We do neither
2990          * preserve the WAITERS bit not the OWNER_DIED one. We are the
2991          * owner.
2992          */
2993         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2994                 spin_unlock(&hb->lock);
2995                 goto pi_faulted;
2996         }
2997
2998         /*
2999          * If uval has changed, let user space handle it.
3000          */
3001         ret = (curval == uval) ? 0 : -EAGAIN;
3002
3003 out_unlock:
3004         spin_unlock(&hb->lock);
3005 out_putkey:
3006         put_futex_key(&key);
3007         return ret;
3008
3009 pi_faulted:
3010         put_futex_key(&key);
3011
3012         ret = fault_in_user_writeable(uaddr);
3013         if (!ret)
3014                 goto retry;
3015
3016         return ret;
3017 }
3018
3019 /**
3020  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3021  * @hb:         the hash_bucket futex_q was original enqueued on
3022  * @q:          the futex_q woken while waiting to be requeued
3023  * @key2:       the futex_key of the requeue target futex
3024  * @timeout:    the timeout associated with the wait (NULL if none)
3025  *
3026  * Detect if the task was woken on the initial futex as opposed to the requeue
3027  * target futex.  If so, determine if it was a timeout or a signal that caused
3028  * the wakeup and return the appropriate error code to the caller.  Must be
3029  * called with the hb lock held.
3030  *
3031  * Return:
3032  *  -  0 = no early wakeup detected;
3033  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3034  */
3035 static inline
3036 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3037                                    struct futex_q *q, union futex_key *key2,
3038                                    struct hrtimer_sleeper *timeout)
3039 {
3040         int ret = 0;
3041
3042         /*
3043          * With the hb lock held, we avoid races while we process the wakeup.
3044          * We only need to hold hb (and not hb2) to ensure atomicity as the
3045          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3046          * It can't be requeued from uaddr2 to something else since we don't
3047          * support a PI aware source futex for requeue.
3048          */
3049         if (!match_futex(&q->key, key2)) {
3050                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3051                 /*
3052                  * We were woken prior to requeue by a timeout or a signal.
3053                  * Unqueue the futex_q and determine which it was.
3054                  */
3055                 plist_del(&q->list, &hb->chain);
3056                 hb_waiters_dec(hb);
3057
3058                 /* Handle spurious wakeups gracefully */
3059                 ret = -EWOULDBLOCK;
3060                 if (timeout && !timeout->task)
3061                         ret = -ETIMEDOUT;
3062                 else if (signal_pending(current))
3063                         ret = -ERESTARTNOINTR;
3064         }
3065         return ret;
3066 }
3067
3068 /**
3069  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3070  * @uaddr:      the futex we initially wait on (non-pi)
3071  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3072  *              the same type, no requeueing from private to shared, etc.
3073  * @val:        the expected value of uaddr
3074  * @abs_time:   absolute timeout
3075  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3076  * @uaddr2:     the pi futex we will take prior to returning to user-space
3077  *
3078  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3079  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3080  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3081  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3082  * without one, the pi logic would not know which task to boost/deboost, if
3083  * there was a need to.
3084  *
3085  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3086  * via the following--
3087  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3088  * 2) wakeup on uaddr2 after a requeue
3089  * 3) signal
3090  * 4) timeout
3091  *
3092  * If 3, cleanup and return -ERESTARTNOINTR.
3093  *
3094  * If 2, we may then block on trying to take the rt_mutex and return via:
3095  * 5) successful lock
3096  * 6) signal
3097  * 7) timeout
3098  * 8) other lock acquisition failure
3099  *
3100  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3101  *
3102  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3103  *
3104  * Return:
3105  *  -  0 - On success;
3106  *  - <0 - On error
3107  */
3108 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3109                                  u32 val, ktime_t *abs_time, u32 bitset,
3110                                  u32 __user *uaddr2)
3111 {
3112         struct hrtimer_sleeper timeout, *to = NULL;
3113         struct futex_pi_state *pi_state = NULL;
3114         struct rt_mutex_waiter rt_waiter;
3115         struct futex_hash_bucket *hb;
3116         union futex_key key2 = FUTEX_KEY_INIT;
3117         struct futex_q q = futex_q_init;
3118         int res, ret;
3119
3120         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3121                 return -ENOSYS;
3122
3123         if (uaddr == uaddr2)
3124                 return -EINVAL;
3125
3126         if (!bitset)
3127                 return -EINVAL;
3128
3129         if (abs_time) {
3130                 to = &timeout;
3131                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3132                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
3133                                       HRTIMER_MODE_ABS);
3134                 hrtimer_init_sleeper(to, current);
3135                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3136                                              current->timer_slack_ns);
3137         }
3138
3139         /*
3140          * The waiter is allocated on our stack, manipulated by the requeue
3141          * code while we sleep on uaddr.
3142          */
3143         rt_mutex_init_waiter(&rt_waiter);
3144
3145         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3146         if (unlikely(ret != 0))
3147                 goto out;
3148
3149         q.bitset = bitset;
3150         q.rt_waiter = &rt_waiter;
3151         q.requeue_pi_key = &key2;
3152
3153         /*
3154          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3155          * count.
3156          */
3157         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3158         if (ret)
3159                 goto out_key2;
3160
3161         /*
3162          * The check above which compares uaddrs is not sufficient for
3163          * shared futexes. We need to compare the keys:
3164          */
3165         if (match_futex(&q.key, &key2)) {
3166                 queue_unlock(hb);
3167                 ret = -EINVAL;
3168                 goto out_put_keys;
3169         }
3170
3171         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3172         futex_wait_queue_me(hb, &q, to);
3173
3174         spin_lock(&hb->lock);
3175         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3176         spin_unlock(&hb->lock);
3177         if (ret)
3178                 goto out_put_keys;
3179
3180         /*
3181          * In order for us to be here, we know our q.key == key2, and since
3182          * we took the hb->lock above, we also know that futex_requeue() has
3183          * completed and we no longer have to concern ourselves with a wakeup
3184          * race with the atomic proxy lock acquisition by the requeue code. The
3185          * futex_requeue dropped our key1 reference and incremented our key2
3186          * reference count.
3187          */
3188
3189         /* Check if the requeue code acquired the second futex for us. */
3190         if (!q.rt_waiter) {
3191                 /*
3192                  * Got the lock. We might not be the anticipated owner if we
3193                  * did a lock-steal - fix up the PI-state in that case.
3194                  */
3195                 if (q.pi_state && (q.pi_state->owner != current)) {
3196                         spin_lock(q.lock_ptr);
3197                         ret = fixup_pi_state_owner(uaddr2, &q, current);
3198                         if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3199                                 pi_state = q.pi_state;
3200                                 get_pi_state(pi_state);
3201                         }
3202                         /*
3203                          * Drop the reference to the pi state which
3204                          * the requeue_pi() code acquired for us.
3205                          */
3206                         put_pi_state(q.pi_state);
3207                         spin_unlock(q.lock_ptr);
3208                 }
3209         } else {
3210                 struct rt_mutex *pi_mutex;
3211
3212                 /*
3213                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3214                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3215                  * the pi_state.
3216                  */
3217                 WARN_ON(!q.pi_state);
3218                 pi_mutex = &q.pi_state->pi_mutex;
3219                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3220
3221                 spin_lock(q.lock_ptr);
3222                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3223                         ret = 0;
3224
3225                 debug_rt_mutex_free_waiter(&rt_waiter);
3226                 /*
3227                  * Fixup the pi_state owner and possibly acquire the lock if we
3228                  * haven't already.
3229                  */
3230                 res = fixup_owner(uaddr2, &q, !ret);
3231                 /*
3232                  * If fixup_owner() returned an error, proprogate that.  If it
3233                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3234                  */
3235                 if (res)
3236                         ret = (res < 0) ? res : 0;
3237
3238                 /*
3239                  * If fixup_pi_state_owner() faulted and was unable to handle
3240                  * the fault, unlock the rt_mutex and return the fault to
3241                  * userspace.
3242                  */
3243                 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3244                         pi_state = q.pi_state;
3245                         get_pi_state(pi_state);
3246                 }
3247
3248                 /* Unqueue and drop the lock. */
3249                 unqueue_me_pi(&q);
3250         }
3251
3252         if (pi_state) {
3253                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3254                 put_pi_state(pi_state);
3255         }
3256
3257         if (ret == -EINTR) {
3258                 /*
3259                  * We've already been requeued, but cannot restart by calling
3260                  * futex_lock_pi() directly. We could restart this syscall, but
3261                  * it would detect that the user space "val" changed and return
3262                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3263                  * -EWOULDBLOCK directly.
3264                  */
3265                 ret = -EWOULDBLOCK;
3266         }
3267
3268 out_put_keys:
3269         put_futex_key(&q.key);
3270 out_key2:
3271         put_futex_key(&key2);
3272
3273 out:
3274         if (to) {
3275                 hrtimer_cancel(&to->timer);
3276                 destroy_hrtimer_on_stack(&to->timer);
3277         }
3278         return ret;
3279 }
3280
3281 /*
3282  * Support for robust futexes: the kernel cleans up held futexes at
3283  * thread exit time.
3284  *
3285  * Implementation: user-space maintains a per-thread list of locks it
3286  * is holding. Upon do_exit(), the kernel carefully walks this list,
3287  * and marks all locks that are owned by this thread with the
3288  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3289  * always manipulated with the lock held, so the list is private and
3290  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3291  * field, to allow the kernel to clean up if the thread dies after
3292  * acquiring the lock, but just before it could have added itself to
3293  * the list. There can only be one such pending lock.
3294  */
3295
3296 /**
3297  * sys_set_robust_list() - Set the robust-futex list head of a task
3298  * @head:       pointer to the list-head
3299  * @len:        length of the list-head, as userspace expects
3300  */
3301 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3302                 size_t, len)
3303 {
3304         if (!futex_cmpxchg_enabled)
3305                 return -ENOSYS;
3306         /*
3307          * The kernel knows only one size for now:
3308          */
3309         if (unlikely(len != sizeof(*head)))
3310                 return -EINVAL;
3311
3312         current->robust_list = head;