futex: futex_wake_op, do not fail on invalid op
[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/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/wake_q.h>
65 #include <linux/sched/mm.h>
66 #include <linux/hugetlb.h>
67 #include <linux/freezer.h>
68 #include <linux/bootmem.h>
69 #include <linux/fault-inject.h>
70
71 #include <asm/futex.h>
72
73 #include "locking/rtmutex_common.h"
74
75 /*
76  * READ this before attempting to hack on futexes!
77  *
78  * Basic futex operation and ordering guarantees
79  * =============================================
80  *
81  * The waiter reads the futex value in user space and calls
82  * futex_wait(). This function computes the hash bucket and acquires
83  * the hash bucket lock. After that it reads the futex user space value
84  * again and verifies that the data has not changed. If it has not changed
85  * it enqueues itself into the hash bucket, releases the hash bucket lock
86  * and schedules.
87  *
88  * The waker side modifies the user space value of the futex and calls
89  * futex_wake(). This function computes the hash bucket and acquires the
90  * hash bucket lock. Then it looks for waiters on that futex in the hash
91  * bucket and wakes them.
92  *
93  * In futex wake up scenarios where no tasks are blocked on a futex, taking
94  * the hb spinlock can be avoided and simply return. In order for this
95  * optimization to work, ordering guarantees must exist so that the waiter
96  * being added to the list is acknowledged when the list is concurrently being
97  * checked by the waker, avoiding scenarios like the following:
98  *
99  * CPU 0                               CPU 1
100  * val = *futex;
101  * sys_futex(WAIT, futex, val);
102  *   futex_wait(futex, val);
103  *   uval = *futex;
104  *                                     *futex = newval;
105  *                                     sys_futex(WAKE, futex);
106  *                                       futex_wake(futex);
107  *                                       if (queue_empty())
108  *                                         return;
109  *   if (uval == val)
110  *      lock(hash_bucket(futex));
111  *      queue();
112  *     unlock(hash_bucket(futex));
113  *     schedule();
114  *
115  * This would cause the waiter on CPU 0 to wait forever because it
116  * missed the transition of the user space value from val to newval
117  * and the waker did not find the waiter in the hash bucket queue.
118  *
119  * The correct serialization ensures that a waiter either observes
120  * the changed user space value before blocking or is woken by a
121  * concurrent waker:
122  *
123  * CPU 0                                 CPU 1
124  * val = *futex;
125  * sys_futex(WAIT, futex, val);
126  *   futex_wait(futex, val);
127  *
128  *   waiters++; (a)
129  *   smp_mb(); (A) <-- paired with -.
130  *                                  |
131  *   lock(hash_bucket(futex));      |
132  *                                  |
133  *   uval = *futex;                 |
134  *                                  |        *futex = newval;
135  *                                  |        sys_futex(WAKE, futex);
136  *                                  |          futex_wake(futex);
137  *                                  |
138  *                                  `--------> smp_mb(); (B)
139  *   if (uval == val)
140  *     queue();
141  *     unlock(hash_bucket(futex));
142  *     schedule();                         if (waiters)
143  *                                           lock(hash_bucket(futex));
144  *   else                                    wake_waiters(futex);
145  *     waiters--; (b)                        unlock(hash_bucket(futex));
146  *
147  * Where (A) orders the waiters increment and the futex value read through
148  * atomic operations (see hb_waiters_inc) and where (B) orders the write
149  * to futex and the waiters read -- this is done by the barriers for both
150  * shared and private futexes in get_futex_key_refs().
151  *
152  * This yields the following case (where X:=waiters, Y:=futex):
153  *
154  *      X = Y = 0
155  *
156  *      w[X]=1          w[Y]=1
157  *      MB              MB
158  *      r[Y]=y          r[X]=x
159  *
160  * Which guarantees that x==0 && y==0 is impossible; which translates back into
161  * the guarantee that we cannot both miss the futex variable change and the
162  * enqueue.
163  *
164  * Note that a new waiter is accounted for in (a) even when it is possible that
165  * the wait call can return error, in which case we backtrack from it in (b).
166  * Refer to the comment in queue_lock().
167  *
168  * Similarly, in order to account for waiters being requeued on another
169  * address we always increment the waiters for the destination bucket before
170  * acquiring the lock. It then decrements them again  after releasing it -
171  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172  * will do the additional required waiter count housekeeping. This is done for
173  * double_lock_hb() and double_unlock_hb(), respectively.
174  */
175
176 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
177 int __read_mostly futex_cmpxchg_enabled;
178 #endif
179
180 /*
181  * Futex flags used to encode options to functions and preserve them across
182  * restarts.
183  */
184 #ifdef CONFIG_MMU
185 # define FLAGS_SHARED           0x01
186 #else
187 /*
188  * NOMMU does not have per process address space. Let the compiler optimize
189  * code away.
190  */
191 # define FLAGS_SHARED           0x00
192 #endif
193 #define FLAGS_CLOCKRT           0x02
194 #define FLAGS_HAS_TIMEOUT       0x04
195
196 /*
197  * Priority Inheritance state:
198  */
199 struct futex_pi_state {
200         /*
201          * list of 'owned' pi_state instances - these have to be
202          * cleaned up in do_exit() if the task exits prematurely:
203          */
204         struct list_head list;
205
206         /*
207          * The PI object:
208          */
209         struct rt_mutex pi_mutex;
210
211         struct task_struct *owner;
212         atomic_t refcount;
213
214         union futex_key key;
215 } __randomize_layout;
216
217 /**
218  * struct futex_q - The hashed futex queue entry, one per waiting task
219  * @list:               priority-sorted list of tasks waiting on this futex
220  * @task:               the task waiting on the futex
221  * @lock_ptr:           the hash bucket lock
222  * @key:                the key the futex is hashed on
223  * @pi_state:           optional priority inheritance state
224  * @rt_waiter:          rt_waiter storage for use with requeue_pi
225  * @requeue_pi_key:     the requeue_pi target futex key
226  * @bitset:             bitset for the optional bitmasked wakeup
227  *
228  * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
229  * we can wake only the relevant ones (hashed queues may be shared).
230  *
231  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
232  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
233  * The order of wakeup is always to make the first condition true, then
234  * the second.
235  *
236  * PI futexes are typically woken before they are removed from the hash list via
237  * the rt_mutex code. See unqueue_me_pi().
238  */
239 struct futex_q {
240         struct plist_node list;
241
242         struct task_struct *task;
243         spinlock_t *lock_ptr;
244         union futex_key key;
245         struct futex_pi_state *pi_state;
246         struct rt_mutex_waiter *rt_waiter;
247         union futex_key *requeue_pi_key;
248         u32 bitset;
249 } __randomize_layout;
250
251 static const struct futex_q futex_q_init = {
252         /* list gets initialized in queue_me()*/
253         .key = FUTEX_KEY_INIT,
254         .bitset = FUTEX_BITSET_MATCH_ANY
255 };
256
257 /*
258  * Hash buckets are shared by all the futex_keys that hash to the same
259  * location.  Each key may have multiple futex_q structures, one for each task
260  * waiting on a futex.
261  */
262 struct futex_hash_bucket {
263         atomic_t waiters;
264         spinlock_t lock;
265         struct plist_head chain;
266 } ____cacheline_aligned_in_smp;
267
268 /*
269  * The base of the bucket array and its size are always used together
270  * (after initialization only in hash_futex()), so ensure that they
271  * reside in the same cacheline.
272  */
273 static struct {
274         struct futex_hash_bucket *queues;
275         unsigned long            hashsize;
276 } __futex_data __read_mostly __aligned(2*sizeof(long));
277 #define futex_queues   (__futex_data.queues)
278 #define futex_hashsize (__futex_data.hashsize)
279
280
281 /*
282  * Fault injections for futexes.
283  */
284 #ifdef CONFIG_FAIL_FUTEX
285
286 static struct {
287         struct fault_attr attr;
288
289         bool ignore_private;
290 } fail_futex = {
291         .attr = FAULT_ATTR_INITIALIZER,
292         .ignore_private = false,
293 };
294
295 static int __init setup_fail_futex(char *str)
296 {
297         return setup_fault_attr(&fail_futex.attr, str);
298 }
299 __setup("fail_futex=", setup_fail_futex);
300
301 static bool should_fail_futex(bool fshared)
302 {
303         if (fail_futex.ignore_private && !fshared)
304                 return false;
305
306         return should_fail(&fail_futex.attr, 1);
307 }
308
309 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
310
311 static int __init fail_futex_debugfs(void)
312 {
313         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
314         struct dentry *dir;
315
316         dir = fault_create_debugfs_attr("fail_futex", NULL,
317                                         &fail_futex.attr);
318         if (IS_ERR(dir))
319                 return PTR_ERR(dir);
320
321         if (!debugfs_create_bool("ignore-private", mode, dir,
322                                  &fail_futex.ignore_private)) {
323                 debugfs_remove_recursive(dir);
324                 return -ENOMEM;
325         }
326
327         return 0;
328 }
329
330 late_initcall(fail_futex_debugfs);
331
332 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
333
334 #else
335 static inline bool should_fail_futex(bool fshared)
336 {
337         return false;
338 }
339 #endif /* CONFIG_FAIL_FUTEX */
340
341 static inline void futex_get_mm(union futex_key *key)
342 {
343         mmgrab(key->private.mm);
344         /*
345          * Ensure futex_get_mm() implies a full barrier such that
346          * get_futex_key() implies a full barrier. This is relied upon
347          * as smp_mb(); (B), see the ordering comment above.
348          */
349         smp_mb__after_atomic();
350 }
351
352 /*
353  * Reflects a new waiter being added to the waitqueue.
354  */
355 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
356 {
357 #ifdef CONFIG_SMP
358         atomic_inc(&hb->waiters);
359         /*
360          * Full barrier (A), see the ordering comment above.
361          */
362         smp_mb__after_atomic();
363 #endif
364 }
365
366 /*
367  * Reflects a waiter being removed from the waitqueue by wakeup
368  * paths.
369  */
370 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
371 {
372 #ifdef CONFIG_SMP
373         atomic_dec(&hb->waiters);
374 #endif
375 }
376
377 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
378 {
379 #ifdef CONFIG_SMP
380         return atomic_read(&hb->waiters);
381 #else
382         return 1;
383 #endif
384 }
385
386 /**
387  * hash_futex - Return the hash bucket in the global hash
388  * @key:        Pointer to the futex key for which the hash is calculated
389  *
390  * We hash on the keys returned from get_futex_key (see below) and return the
391  * corresponding hash bucket in the global hash.
392  */
393 static struct futex_hash_bucket *hash_futex(union futex_key *key)
394 {
395         u32 hash = jhash2((u32*)&key->both.word,
396                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
397                           key->both.offset);
398         return &futex_queues[hash & (futex_hashsize - 1)];
399 }
400
401
402 /**
403  * match_futex - Check whether two futex keys are equal
404  * @key1:       Pointer to key1
405  * @key2:       Pointer to key2
406  *
407  * Return 1 if two futex_keys are equal, 0 otherwise.
408  */
409 static inline int match_futex(union futex_key *key1, union futex_key *key2)
410 {
411         return (key1 && key2
412                 && key1->both.word == key2->both.word
413                 && key1->both.ptr == key2->both.ptr
414                 && key1->both.offset == key2->both.offset);
415 }
416
417 /*
418  * Take a reference to the resource addressed by a key.
419  * Can be called while holding spinlocks.
420  *
421  */
422 static void get_futex_key_refs(union futex_key *key)
423 {
424         if (!key->both.ptr)
425                 return;
426
427         /*
428          * On MMU less systems futexes are always "private" as there is no per
429          * process address space. We need the smp wmb nevertheless - yes,
430          * arch/blackfin has MMU less SMP ...
431          */
432         if (!IS_ENABLED(CONFIG_MMU)) {
433                 smp_mb(); /* explicit smp_mb(); (B) */
434                 return;
435         }
436
437         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
438         case FUT_OFF_INODE:
439                 ihold(key->shared.inode); /* implies smp_mb(); (B) */
440                 break;
441         case FUT_OFF_MMSHARED:
442                 futex_get_mm(key); /* implies smp_mb(); (B) */
443                 break;
444         default:
445                 /*
446                  * Private futexes do not hold reference on an inode or
447                  * mm, therefore the only purpose of calling get_futex_key_refs
448                  * is because we need the barrier for the lockless waiter check.
449                  */
450                 smp_mb(); /* explicit smp_mb(); (B) */
451         }
452 }
453
454 /*
455  * Drop a reference to the resource addressed by a key.
456  * The hash bucket spinlock must not be held. This is
457  * a no-op for private futexes, see comment in the get
458  * counterpart.
459  */
460 static void drop_futex_key_refs(union futex_key *key)
461 {
462         if (!key->both.ptr) {
463                 /* If we're here then we tried to put a key we failed to get */
464                 WARN_ON_ONCE(1);
465                 return;
466         }
467
468         if (!IS_ENABLED(CONFIG_MMU))
469                 return;
470
471         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
472         case FUT_OFF_INODE:
473                 iput(key->shared.inode);
474                 break;
475         case FUT_OFF_MMSHARED:
476                 mmdrop(key->private.mm);
477                 break;
478         }
479 }
480
481 /**
482  * get_futex_key() - Get parameters which are the keys for a futex
483  * @uaddr:      virtual address of the futex
484  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485  * @key:        address where result is stored.
486  * @rw:         mapping needs to be read/write (values: VERIFY_READ,
487  *              VERIFY_WRITE)
488  *
489  * Return: a negative error code or 0
490  *
491  * The key words are stored in @key on success.
492  *
493  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
494  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
495  * We can usually work out the index without swapping in the page.
496  *
497  * lock_page() might sleep, the caller should not hold a spinlock.
498  */
499 static int
500 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
501 {
502         unsigned long address = (unsigned long)uaddr;
503         struct mm_struct *mm = current->mm;
504         struct page *page, *tail;
505         struct address_space *mapping;
506         int err, ro = 0;
507
508         /*
509          * The futex address must be "naturally" aligned.
510          */
511         key->both.offset = address % PAGE_SIZE;
512         if (unlikely((address % sizeof(u32)) != 0))
513                 return -EINVAL;
514         address -= key->both.offset;
515
516         if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
517                 return -EFAULT;
518
519         if (unlikely(should_fail_futex(fshared)))
520                 return -EFAULT;
521
522         /*
523          * PROCESS_PRIVATE futexes are fast.
524          * As the mm cannot disappear under us and the 'key' only needs
525          * virtual address, we dont even have to find the underlying vma.
526          * Note : We do have to check 'uaddr' is a valid user address,
527          *        but access_ok() should be faster than find_vma()
528          */
529         if (!fshared) {
530                 key->private.mm = mm;
531                 key->private.address = address;
532                 get_futex_key_refs(key);  /* implies smp_mb(); (B) */
533                 return 0;
534         }
535
536 again:
537         /* Ignore any VERIFY_READ mapping (futex common case) */
538         if (unlikely(should_fail_futex(fshared)))
539                 return -EFAULT;
540
541         err = get_user_pages_fast(address, 1, 1, &page);
542         /*
543          * If write access is not required (eg. FUTEX_WAIT), try
544          * and get read-only access.
545          */
546         if (err == -EFAULT && rw == VERIFY_READ) {
547                 err = get_user_pages_fast(address, 1, 0, &page);
548                 ro = 1;
549         }
550         if (err < 0)
551                 return err;
552         else
553                 err = 0;
554
555         /*
556          * The treatment of mapping from this point on is critical. The page
557          * lock protects many things but in this context the page lock
558          * stabilizes mapping, prevents inode freeing in the shared
559          * file-backed region case and guards against movement to swap cache.
560          *
561          * Strictly speaking the page lock is not needed in all cases being
562          * considered here and page lock forces unnecessarily serialization
563          * From this point on, mapping will be re-verified if necessary and
564          * page lock will be acquired only if it is unavoidable
565          *
566          * Mapping checks require the head page for any compound page so the
567          * head page and mapping is looked up now. For anonymous pages, it
568          * does not matter if the page splits in the future as the key is
569          * based on the address. For filesystem-backed pages, the tail is
570          * required as the index of the page determines the key. For
571          * base pages, there is no tail page and tail == page.
572          */
573         tail = page;
574         page = compound_head(page);
575         mapping = READ_ONCE(page->mapping);
576
577         /*
578          * If page->mapping is NULL, then it cannot be a PageAnon
579          * page; but it might be the ZERO_PAGE or in the gate area or
580          * in a special mapping (all cases which we are happy to fail);
581          * or it may have been a good file page when get_user_pages_fast
582          * found it, but truncated or holepunched or subjected to
583          * invalidate_complete_page2 before we got the page lock (also
584          * cases which we are happy to fail).  And we hold a reference,
585          * so refcount care in invalidate_complete_page's remove_mapping
586          * prevents drop_caches from setting mapping to NULL beneath us.
587          *
588          * The case we do have to guard against is when memory pressure made
589          * shmem_writepage move it from filecache to swapcache beneath us:
590          * an unlikely race, but we do need to retry for page->mapping.
591          */
592         if (unlikely(!mapping)) {
593                 int shmem_swizzled;
594
595                 /*
596                  * Page lock is required to identify which special case above
597                  * applies. If this is really a shmem page then the page lock
598                  * will prevent unexpected transitions.
599                  */
600                 lock_page(page);
601                 shmem_swizzled = PageSwapCache(page) || page->mapping;
602                 unlock_page(page);
603                 put_page(page);
604
605                 if (shmem_swizzled)
606                         goto again;
607
608                 return -EFAULT;
609         }
610
611         /*
612          * Private mappings are handled in a simple way.
613          *
614          * If the futex key is stored on an anonymous page, then the associated
615          * object is the mm which is implicitly pinned by the calling process.
616          *
617          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618          * it's a read-only handle, it's expected that futexes attach to
619          * the object not the particular process.
620          */
621         if (PageAnon(page)) {
622                 /*
623                  * A RO anonymous page will never change and thus doesn't make
624                  * sense for futex operations.
625                  */
626                 if (unlikely(should_fail_futex(fshared)) || ro) {
627                         err = -EFAULT;
628                         goto out;
629                 }
630
631                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
632                 key->private.mm = mm;
633                 key->private.address = address;
634
635                 get_futex_key_refs(key); /* implies smp_mb(); (B) */
636
637         } else {
638                 struct inode *inode;
639
640                 /*
641                  * The associated futex object in this case is the inode and
642                  * the page->mapping must be traversed. Ordinarily this should
643                  * be stabilised under page lock but it's not strictly
644                  * necessary in this case as we just want to pin the inode, not
645                  * update the radix tree or anything like that.
646                  *
647                  * The RCU read lock is taken as the inode is finally freed
648                  * under RCU. If the mapping still matches expectations then the
649                  * mapping->host can be safely accessed as being a valid inode.
650                  */
651                 rcu_read_lock();
652
653                 if (READ_ONCE(page->mapping) != mapping) {
654                         rcu_read_unlock();
655                         put_page(page);
656
657                         goto again;
658                 }
659
660                 inode = READ_ONCE(mapping->host);
661                 if (!inode) {
662                         rcu_read_unlock();
663                         put_page(page);
664
665                         goto again;
666                 }
667
668                 /*
669                  * Take a reference unless it is about to be freed. Previously
670                  * this reference was taken by ihold under the page lock
671                  * pinning the inode in place so i_lock was unnecessary. The
672                  * only way for this check to fail is if the inode was
673                  * truncated in parallel which is almost certainly an
674                  * application bug. In such a case, just retry.
675                  *
676                  * We are not calling into get_futex_key_refs() in file-backed
677                  * cases, therefore a successful atomic_inc return below will
678                  * guarantee that get_futex_key() will still imply smp_mb(); (B).
679                  */
680                 if (!atomic_inc_not_zero(&inode->i_count)) {
681                         rcu_read_unlock();
682                         put_page(page);
683
684                         goto again;
685                 }
686
687                 /* Should be impossible but lets be paranoid for now */
688                 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
689                         err = -EFAULT;
690                         rcu_read_unlock();
691                         iput(inode);
692
693                         goto out;
694                 }
695
696                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
697                 key->shared.inode = inode;
698                 key->shared.pgoff = basepage_index(tail);
699                 rcu_read_unlock();
700         }
701
702 out:
703         put_page(page);
704         return err;
705 }
706
707 static inline void put_futex_key(union futex_key *key)
708 {
709         drop_futex_key_refs(key);
710 }
711
712 /**
713  * fault_in_user_writeable() - Fault in user address and verify RW access
714  * @uaddr:      pointer to faulting user space address
715  *
716  * Slow path to fixup the fault we just took in the atomic write
717  * access to @uaddr.
718  *
719  * We have no generic implementation of a non-destructive write to the
720  * user address. We know that we faulted in the atomic pagefault
721  * disabled section so we can as well avoid the #PF overhead by
722  * calling get_user_pages() right away.
723  */
724 static int fault_in_user_writeable(u32 __user *uaddr)
725 {
726         struct mm_struct *mm = current->mm;
727         int ret;
728
729         down_read(&mm->mmap_sem);
730         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
731                                FAULT_FLAG_WRITE, NULL);
732         up_read(&mm->mmap_sem);
733
734         return ret < 0 ? ret : 0;
735 }
736
737 /**
738  * futex_top_waiter() - Return the highest priority waiter on a futex
739  * @hb:         the hash bucket the futex_q's reside in
740  * @key:        the futex key (to distinguish it from other futex futex_q's)
741  *
742  * Must be called with the hb lock held.
743  */
744 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
745                                         union futex_key *key)
746 {
747         struct futex_q *this;
748
749         plist_for_each_entry(this, &hb->chain, list) {
750                 if (match_futex(&this->key, key))
751                         return this;
752         }
753         return NULL;
754 }
755
756 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
757                                       u32 uval, u32 newval)
758 {
759         int ret;
760
761         pagefault_disable();
762         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
763         pagefault_enable();
764
765         return ret;
766 }
767
768 static int get_futex_value_locked(u32 *dest, u32 __user *from)
769 {
770         int ret;
771
772         pagefault_disable();
773         ret = __get_user(*dest, from);
774         pagefault_enable();
775
776         return ret ? -EFAULT : 0;
777 }
778
779
780 /*
781  * PI code:
782  */
783 static int refill_pi_state_cache(void)
784 {
785         struct futex_pi_state *pi_state;
786
787         if (likely(current->pi_state_cache))
788                 return 0;
789
790         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
791
792         if (!pi_state)
793                 return -ENOMEM;
794
795         INIT_LIST_HEAD(&pi_state->list);
796         /* pi_mutex gets initialized later */
797         pi_state->owner = NULL;
798         atomic_set(&pi_state->refcount, 1);
799         pi_state->key = FUTEX_KEY_INIT;
800
801         current->pi_state_cache = pi_state;
802
803         return 0;
804 }
805
806 static struct futex_pi_state *alloc_pi_state(void)
807 {
808         struct futex_pi_state *pi_state = current->pi_state_cache;
809
810         WARN_ON(!pi_state);
811         current->pi_state_cache = NULL;
812
813         return pi_state;
814 }
815
816 static void get_pi_state(struct futex_pi_state *pi_state)
817 {
818         WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
819 }
820
821 /*
822  * Drops a reference to the pi_state object and frees or caches it
823  * when the last reference is gone.
824  */
825 static void put_pi_state(struct futex_pi_state *pi_state)
826 {
827         if (!pi_state)
828                 return;
829
830         if (!atomic_dec_and_test(&pi_state->refcount))
831                 return;
832
833         /*
834          * If pi_state->owner is NULL, the owner is most probably dying
835          * and has cleaned up the pi_state already
836          */
837         if (pi_state->owner) {
838                 struct task_struct *owner;
839
840                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
841                 owner = pi_state->owner;
842                 if (owner) {
843                         raw_spin_lock(&owner->pi_lock);
844                         list_del_init(&pi_state->list);
845                         raw_spin_unlock(&owner->pi_lock);
846                 }
847                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
848                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
849         }
850
851         if (current->pi_state_cache) {
852                 kfree(pi_state);
853         } else {
854                 /*
855                  * pi_state->list is already empty.
856                  * clear pi_state->owner.
857                  * refcount is at 0 - put it back to 1.
858                  */
859                 pi_state->owner = NULL;
860                 atomic_set(&pi_state->refcount, 1);
861                 current->pi_state_cache = pi_state;
862         }
863 }
864
865 /*
866  * Look up the task based on what TID userspace gave us.
867  * We dont trust it.
868  */
869 static struct task_struct *futex_find_get_task(pid_t pid)
870 {
871         struct task_struct *p;
872
873         rcu_read_lock();
874         p = find_task_by_vpid(pid);
875         if (p)
876                 get_task_struct(p);
877
878         rcu_read_unlock();
879
880         return p;
881 }
882
883 #ifdef CONFIG_FUTEX_PI
884
885 /*
886  * This task is holding PI mutexes at exit time => bad.
887  * Kernel cleans up PI-state, but userspace is likely hosed.
888  * (Robust-futex cleanup is separate and might save the day for userspace.)
889  */
890 void exit_pi_state_list(struct task_struct *curr)
891 {
892         struct list_head *next, *head = &curr->pi_state_list;
893         struct futex_pi_state *pi_state;
894         struct futex_hash_bucket *hb;
895         union futex_key key = FUTEX_KEY_INIT;
896
897         if (!futex_cmpxchg_enabled)
898                 return;
899         /*
900          * We are a ZOMBIE and nobody can enqueue itself on
901          * pi_state_list anymore, but we have to be careful
902          * versus waiters unqueueing themselves:
903          */
904         raw_spin_lock_irq(&curr->pi_lock);
905         while (!list_empty(head)) {
906
907                 next = head->next;
908                 pi_state = list_entry(next, struct futex_pi_state, list);
909                 key = pi_state->key;
910                 hb = hash_futex(&key);
911                 raw_spin_unlock_irq(&curr->pi_lock);
912
913                 spin_lock(&hb->lock);
914                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
915                 raw_spin_lock(&curr->pi_lock);
916                 /*
917                  * We dropped the pi-lock, so re-check whether this
918                  * task still owns the PI-state:
919                  */
920                 if (head->next != next) {
921                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
922                         spin_unlock(&hb->lock);
923                         continue;
924                 }
925
926                 WARN_ON(pi_state->owner != curr);
927                 WARN_ON(list_empty(&pi_state->list));
928                 list_del_init(&pi_state->list);
929                 pi_state->owner = NULL;
930                 raw_spin_unlock(&curr->pi_lock);
931
932                 get_pi_state(pi_state);
933                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
934                 spin_unlock(&hb->lock);
935
936                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
937                 put_pi_state(pi_state);
938
939                 raw_spin_lock_irq(&curr->pi_lock);
940         }
941         raw_spin_unlock_irq(&curr->pi_lock);
942 }
943
944 #endif
945
946 /*
947  * We need to check the following states:
948  *
949  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
950  *
951  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
952  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
953  *
954  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
955  *
956  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
957  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
958  *
959  * [6]  Found  | Found    | task      | 0         | 1      | Valid
960  *
961  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
962  *
963  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
964  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
965  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
966  *
967  * [1]  Indicates that the kernel can acquire the futex atomically. We
968  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
969  *
970  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
971  *      thread is found then it indicates that the owner TID has died.
972  *
973  * [3]  Invalid. The waiter is queued on a non PI futex
974  *
975  * [4]  Valid state after exit_robust_list(), which sets the user space
976  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
977  *
978  * [5]  The user space value got manipulated between exit_robust_list()
979  *      and exit_pi_state_list()
980  *
981  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
982  *      the pi_state but cannot access the user space value.
983  *
984  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
985  *
986  * [8]  Owner and user space value match
987  *
988  * [9]  There is no transient state which sets the user space TID to 0
989  *      except exit_robust_list(), but this is indicated by the
990  *      FUTEX_OWNER_DIED bit. See [4]
991  *
992  * [10] There is no transient state which leaves owner and user space
993  *      TID out of sync.
994  *
995  *
996  * Serialization and lifetime rules:
997  *
998  * hb->lock:
999  *
1000  *      hb -> futex_q, relation
1001  *      futex_q -> pi_state, relation
1002  *
1003  *      (cannot be raw because hb can contain arbitrary amount
1004  *       of futex_q's)
1005  *
1006  * pi_mutex->wait_lock:
1007  *
1008  *      {uval, pi_state}
1009  *
1010  *      (and pi_mutex 'obviously')
1011  *
1012  * p->pi_lock:
1013  *
1014  *      p->pi_state_list -> pi_state->list, relation
1015  *
1016  * pi_state->refcount:
1017  *
1018  *      pi_state lifetime
1019  *
1020  *
1021  * Lock order:
1022  *
1023  *   hb->lock
1024  *     pi_mutex->wait_lock
1025  *       p->pi_lock
1026  *
1027  */
1028
1029 /*
1030  * Validate that the existing waiter has a pi_state and sanity check
1031  * the pi_state against the user space value. If correct, attach to
1032  * it.
1033  */
1034 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1035                               struct futex_pi_state *pi_state,
1036                               struct futex_pi_state **ps)
1037 {
1038         pid_t pid = uval & FUTEX_TID_MASK;
1039         u32 uval2;
1040         int ret;
1041
1042         /*
1043          * Userspace might have messed up non-PI and PI futexes [3]
1044          */
1045         if (unlikely(!pi_state))
1046                 return -EINVAL;
1047
1048         /*
1049          * We get here with hb->lock held, and having found a
1050          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1051          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1052          * which in turn means that futex_lock_pi() still has a reference on
1053          * our pi_state.
1054          *
1055          * The waiter holding a reference on @pi_state also protects against
1056          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1057          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1058          * free pi_state before we can take a reference ourselves.
1059          */
1060         WARN_ON(!atomic_read(&pi_state->refcount));
1061
1062         /*
1063          * Now that we have a pi_state, we can acquire wait_lock
1064          * and do the state validation.
1065          */
1066         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1067
1068         /*
1069          * Since {uval, pi_state} is serialized by wait_lock, and our current
1070          * uval was read without holding it, it can have changed. Verify it
1071          * still is what we expect it to be, otherwise retry the entire
1072          * operation.
1073          */
1074         if (get_futex_value_locked(&uval2, uaddr))
1075                 goto out_efault;
1076
1077         if (uval != uval2)
1078                 goto out_eagain;
1079
1080         /*
1081          * Handle the owner died case:
1082          */
1083         if (uval & FUTEX_OWNER_DIED) {
1084                 /*
1085                  * exit_pi_state_list sets owner to NULL and wakes the
1086                  * topmost waiter. The task which acquires the
1087                  * pi_state->rt_mutex will fixup owner.
1088                  */
1089                 if (!pi_state->owner) {
1090                         /*
1091                          * No pi state owner, but the user space TID
1092                          * is not 0. Inconsistent state. [5]
1093                          */
1094                         if (pid)
1095                                 goto out_einval;
1096                         /*
1097                          * Take a ref on the state and return success. [4]
1098                          */
1099                         goto out_attach;
1100                 }
1101
1102                 /*
1103                  * If TID is 0, then either the dying owner has not
1104                  * yet executed exit_pi_state_list() or some waiter
1105                  * acquired the rtmutex in the pi state, but did not
1106                  * yet fixup the TID in user space.
1107                  *
1108                  * Take a ref on the state and return success. [6]
1109                  */
1110                 if (!pid)
1111                         goto out_attach;
1112         } else {
1113                 /*
1114                  * If the owner died bit is not set, then the pi_state
1115                  * must have an owner. [7]
1116                  */
1117                 if (!pi_state->owner)
1118                         goto out_einval;
1119         }
1120
1121         /*
1122          * Bail out if user space manipulated the futex value. If pi
1123          * state exists then the owner TID must be the same as the
1124          * user space TID. [9/10]
1125          */
1126         if (pid != task_pid_vnr(pi_state->owner))
1127                 goto out_einval;
1128
1129 out_attach:
1130         get_pi_state(pi_state);
1131         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1132         *ps = pi_state;
1133         return 0;
1134
1135 out_einval:
1136         ret = -EINVAL;
1137         goto out_error;
1138
1139 out_eagain:
1140         ret = -EAGAIN;
1141         goto out_error;
1142
1143 out_efault:
1144         ret = -EFAULT;
1145         goto out_error;
1146
1147 out_error:
1148         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1149         return ret;
1150 }
1151
1152 /*
1153  * Lookup the task for the TID provided from user space and attach to
1154  * it after doing proper sanity checks.
1155  */
1156 static int attach_to_pi_owner(u32 uval, union futex_key *key,
1157                               struct futex_pi_state **ps)
1158 {
1159         pid_t pid = uval & FUTEX_TID_MASK;
1160         struct futex_pi_state *pi_state;
1161         struct task_struct *p;
1162
1163         /*
1164          * We are the first waiter - try to look up the real owner and attach
1165          * the new pi_state to it, but bail out when TID = 0 [1]
1166          */
1167         if (!pid)
1168                 return -ESRCH;
1169         p = futex_find_get_task(pid);
1170         if (!p)
1171                 return -ESRCH;
1172
1173         if (unlikely(p->flags & PF_KTHREAD)) {
1174                 put_task_struct(p);
1175                 return -EPERM;
1176         }
1177
1178         /*
1179          * We need to look at the task state flags to figure out,
1180          * whether the task is exiting. To protect against the do_exit
1181          * change of the task flags, we do this protected by
1182          * p->pi_lock:
1183          */
1184         raw_spin_lock_irq(&p->pi_lock);
1185         if (unlikely(p->flags & PF_EXITING)) {
1186                 /*
1187                  * The task is on the way out. When PF_EXITPIDONE is
1188                  * set, we know that the task has finished the
1189                  * cleanup:
1190                  */
1191                 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1192
1193                 raw_spin_unlock_irq(&p->pi_lock);
1194                 put_task_struct(p);
1195                 return ret;
1196         }
1197
1198         /*
1199          * No existing pi state. First waiter. [2]
1200          *
1201          * This creates pi_state, we have hb->lock held, this means nothing can
1202          * observe this state, wait_lock is irrelevant.
1203          */
1204         pi_state = alloc_pi_state();
1205
1206         /*
1207          * Initialize the pi_mutex in locked state and make @p
1208          * the owner of it:
1209          */
1210         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1211
1212         /* Store the key for possible exit cleanups: */
1213         pi_state->key = *key;
1214
1215         WARN_ON(!list_empty(&pi_state->list));
1216         list_add(&pi_state->list, &p->pi_state_list);
1217         /*
1218          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1219          * because there is no concurrency as the object is not published yet.
1220          */
1221         pi_state->owner = p;
1222         raw_spin_unlock_irq(&p->pi_lock);
1223
1224         put_task_struct(p);
1225
1226         *ps = pi_state;
1227
1228         return 0;
1229 }
1230
1231 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1232                            struct futex_hash_bucket *hb,
1233                            union futex_key *key, struct futex_pi_state **ps)
1234 {
1235         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1236
1237         /*
1238          * If there is a waiter on that futex, validate it and
1239          * attach to the pi_state when the validation succeeds.
1240          */
1241         if (top_waiter)
1242                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1243
1244         /*
1245          * We are the first waiter - try to look up the owner based on
1246          * @uval and attach to it.
1247          */
1248         return attach_to_pi_owner(uval, key, ps);
1249 }
1250
1251 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1252 {
1253         u32 uninitialized_var(curval);
1254
1255         if (unlikely(should_fail_futex(true)))
1256                 return -EFAULT;
1257
1258         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1259                 return -EFAULT;
1260
1261         /* If user space value changed, let the caller retry */
1262         return curval != uval ? -EAGAIN : 0;
1263 }
1264
1265 /**
1266  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1267  * @uaddr:              the pi futex user address
1268  * @hb:                 the pi futex hash bucket
1269  * @key:                the futex key associated with uaddr and hb
1270  * @ps:                 the pi_state pointer where we store the result of the
1271  *                      lookup
1272  * @task:               the task to perform the atomic lock work for.  This will
1273  *                      be "current" except in the case of requeue pi.
1274  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1275  *
1276  * Return:
1277  *  -  0 - ready to wait;
1278  *  -  1 - acquired the lock;
1279  *  - <0 - error
1280  *
1281  * The hb->lock and futex_key refs shall be held by the caller.
1282  */
1283 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1284                                 union futex_key *key,
1285                                 struct futex_pi_state **ps,
1286                                 struct task_struct *task, int set_waiters)
1287 {
1288         u32 uval, newval, vpid = task_pid_vnr(task);
1289         struct futex_q *top_waiter;
1290         int ret;
1291
1292         /*
1293          * Read the user space value first so we can validate a few
1294          * things before proceeding further.
1295          */
1296         if (get_futex_value_locked(&uval, uaddr))
1297                 return -EFAULT;
1298
1299         if (unlikely(should_fail_futex(true)))
1300                 return -EFAULT;
1301
1302         /*
1303          * Detect deadlocks.
1304          */
1305         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1306                 return -EDEADLK;
1307
1308         if ((unlikely(should_fail_futex(true))))
1309                 return -EDEADLK;
1310
1311         /*
1312          * Lookup existing state first. If it exists, try to attach to
1313          * its pi_state.
1314          */
1315         top_waiter = futex_top_waiter(hb, key);
1316         if (top_waiter)
1317                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1318
1319         /*
1320          * No waiter and user TID is 0. We are here because the
1321          * waiters or the owner died bit is set or called from
1322          * requeue_cmp_pi or for whatever reason something took the
1323          * syscall.
1324          */
1325         if (!(uval & FUTEX_TID_MASK)) {
1326                 /*
1327                  * We take over the futex. No other waiters and the user space
1328                  * TID is 0. We preserve the owner died bit.
1329                  */
1330                 newval = uval & FUTEX_OWNER_DIED;
1331                 newval |= vpid;
1332
1333                 /* The futex requeue_pi code can enforce the waiters bit */
1334                 if (set_waiters)
1335                         newval |= FUTEX_WAITERS;
1336
1337                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1338                 /* If the take over worked, return 1 */
1339                 return ret < 0 ? ret : 1;
1340         }
1341
1342         /*
1343          * First waiter. Set the waiters bit before attaching ourself to
1344          * the owner. If owner tries to unlock, it will be forced into
1345          * the kernel and blocked on hb->lock.
1346          */
1347         newval = uval | FUTEX_WAITERS;
1348         ret = lock_pi_update_atomic(uaddr, uval, newval);
1349         if (ret)
1350                 return ret;
1351         /*
1352          * If the update of the user space value succeeded, we try to
1353          * attach to the owner. If that fails, no harm done, we only
1354          * set the FUTEX_WAITERS bit in the user space variable.
1355          */
1356         return attach_to_pi_owner(uval, key, ps);
1357 }
1358
1359 /**
1360  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1361  * @q:  The futex_q to unqueue
1362  *
1363  * The q->lock_ptr must not be NULL and must be held by the caller.
1364  */
1365 static void __unqueue_futex(struct futex_q *q)
1366 {
1367         struct futex_hash_bucket *hb;
1368
1369         if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1370             || WARN_ON(plist_node_empty(&q->list)))
1371                 return;
1372
1373         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1374         plist_del(&q->list, &hb->chain);
1375         hb_waiters_dec(hb);
1376 }
1377
1378 /*
1379  * The hash bucket lock must be held when this is called.
1380  * Afterwards, the futex_q must not be accessed. Callers
1381  * must ensure to later call wake_up_q() for the actual
1382  * wakeups to occur.
1383  */
1384 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1385 {
1386         struct task_struct *p = q->task;
1387
1388         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1389                 return;
1390
1391         /*
1392          * Queue the task for later wakeup for after we've released
1393          * the hb->lock. wake_q_add() grabs reference to p.
1394          */
1395         wake_q_add(wake_q, p);
1396         __unqueue_futex(q);
1397         /*
1398          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1399          * is written, without taking any locks. This is possible in the event
1400          * of a spurious wakeup, for example. A memory barrier is required here
1401          * to prevent the following store to lock_ptr from getting ahead of the
1402          * plist_del in __unqueue_futex().
1403          */
1404         smp_store_release(&q->lock_ptr, NULL);
1405 }
1406
1407 /*
1408  * Caller must hold a reference on @pi_state.
1409  */
1410 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1411 {
1412         u32 uninitialized_var(curval), newval;
1413         struct task_struct *new_owner;
1414         bool postunlock = false;
1415         DEFINE_WAKE_Q(wake_q);
1416         int ret = 0;
1417
1418         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1419         if (WARN_ON_ONCE(!new_owner)) {
1420                 /*
1421                  * As per the comment in futex_unlock_pi() this should not happen.
1422                  *
1423                  * When this happens, give up our locks and try again, giving
1424                  * the futex_lock_pi() instance time to complete, either by
1425                  * waiting on the rtmutex or removing itself from the futex
1426                  * queue.
1427                  */
1428                 ret = -EAGAIN;
1429                 goto out_unlock;
1430         }
1431
1432         /*
1433          * We pass it to the next owner. The WAITERS bit is always kept
1434          * enabled while there is PI state around. We cleanup the owner
1435          * died bit, because we are the owner.
1436          */
1437         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1438
1439         if (unlikely(should_fail_futex(true)))
1440                 ret = -EFAULT;
1441
1442         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1443                 ret = -EFAULT;
1444
1445         } else if (curval != uval) {
1446                 /*
1447                  * If a unconditional UNLOCK_PI operation (user space did not
1448                  * try the TID->0 transition) raced with a waiter setting the
1449                  * FUTEX_WAITERS flag between get_user() and locking the hash
1450                  * bucket lock, retry the operation.
1451                  */
1452                 if ((FUTEX_TID_MASK & curval) == uval)
1453                         ret = -EAGAIN;
1454                 else
1455                         ret = -EINVAL;
1456         }
1457
1458         if (ret)
1459                 goto out_unlock;
1460
1461         /*
1462          * This is a point of no return; once we modify the uval there is no
1463          * going back and subsequent operations must not fail.
1464          */
1465
1466         raw_spin_lock(&pi_state->owner->pi_lock);
1467         WARN_ON(list_empty(&pi_state->list));
1468         list_del_init(&pi_state->list);
1469         raw_spin_unlock(&pi_state->owner->pi_lock);
1470
1471         raw_spin_lock(&new_owner->pi_lock);
1472         WARN_ON(!list_empty(&pi_state->list));
1473         list_add(&pi_state->list, &new_owner->pi_state_list);
1474         pi_state->owner = new_owner;
1475         raw_spin_unlock(&new_owner->pi_lock);
1476
1477         postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1478
1479 out_unlock:
1480         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1481
1482         if (postunlock)
1483                 rt_mutex_postunlock(&wake_q);
1484
1485         return ret;
1486 }
1487
1488 /*
1489  * Express the locking dependencies for lockdep:
1490  */
1491 static inline void
1492 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1493 {
1494         if (hb1 <= hb2) {
1495                 spin_lock(&hb1->lock);
1496                 if (hb1 < hb2)
1497                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1498         } else { /* hb1 > hb2 */
1499                 spin_lock(&hb2->lock);
1500                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1501         }
1502 }
1503
1504 static inline void
1505 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1506 {
1507         spin_unlock(&hb1->lock);
1508         if (hb1 != hb2)
1509                 spin_unlock(&hb2->lock);
1510 }
1511
1512 /*
1513  * Wake up waiters matching bitset queued on this futex (uaddr).
1514  */
1515 static int
1516 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1517 {
1518         struct futex_hash_bucket *hb;
1519         struct futex_q *this, *next;
1520         union futex_key key = FUTEX_KEY_INIT;
1521         int ret;
1522         DEFINE_WAKE_Q(wake_q);
1523
1524         if (!bitset)
1525                 return -EINVAL;
1526
1527         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1528         if (unlikely(ret != 0))
1529                 goto out;
1530
1531         hb = hash_futex(&key);
1532
1533         /* Make sure we really have tasks to wakeup */
1534         if (!hb_waiters_pending(hb))
1535                 goto out_put_key;
1536
1537         spin_lock(&hb->lock);
1538
1539         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1540                 if (match_futex (&this->key, &key)) {
1541                         if (this->pi_state || this->rt_waiter) {
1542                                 ret = -EINVAL;
1543                                 break;
1544                         }
1545
1546                         /* Check if one of the bits is set in both bitsets */
1547                         if (!(this->bitset & bitset))
1548                                 continue;
1549
1550                         mark_wake_futex(&wake_q, this);
1551                         if (++ret >= nr_wake)
1552                                 break;
1553                 }
1554         }
1555
1556         spin_unlock(&hb->lock);
1557         wake_up_q(&wake_q);
1558 out_put_key:
1559         put_futex_key(&key);
1560 out:
1561         return ret;
1562 }
1563
1564 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1565 {
1566         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1567         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1568         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 12);
1569         int cmparg = sign_extend32(encoded_op & 0x00000fff, 12);
1570         int oldval, ret;
1571
1572         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1573                 if (oparg < 0 || oparg > 31) {
1574                         char comm[sizeof(current->comm)];
1575                         /*
1576                          * kill this print and return -EINVAL when userspace
1577                          * is sane again
1578                          */
1579                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1580                                         get_task_comm(comm, current), oparg);
1581                         oparg &= 31;
1582                 }
1583                 oparg = 1 << oparg;
1584         }
1585
1586         if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1587                 return -EFAULT;
1588
1589         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1590         if (ret)
1591                 return ret;
1592
1593         switch (cmp) {
1594         case FUTEX_OP_CMP_EQ:
1595                 return oldval == cmparg;
1596         case FUTEX_OP_CMP_NE:
1597                 return oldval != cmparg;
1598         case FUTEX_OP_CMP_LT:
1599                 return oldval < cmparg;
1600         case FUTEX_OP_CMP_GE:
1601                 return oldval >= cmparg;
1602         case FUTEX_OP_CMP_LE:
1603                 return oldval <= cmparg;
1604         case FUTEX_OP_CMP_GT:
1605                 return oldval > cmparg;
1606         default:
1607                 return -ENOSYS;
1608         }
1609 }
1610
1611 /*
1612  * Wake up all waiters hashed on the physical page that is mapped
1613  * to this virtual address:
1614  */
1615 static int
1616 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1617               int nr_wake, int nr_wake2, int op)
1618 {
1619         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1620         struct futex_hash_bucket *hb1, *hb2;
1621         struct futex_q *this, *next;
1622         int ret, op_ret;
1623         DEFINE_WAKE_Q(wake_q);
1624
1625 retry:
1626         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1627         if (unlikely(ret != 0))
1628                 goto out;
1629         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1630         if (unlikely(ret != 0))
1631                 goto out_put_key1;
1632
1633         hb1 = hash_futex(&key1);
1634         hb2 = hash_futex(&key2);
1635
1636 retry_private:
1637         double_lock_hb(hb1, hb2);
1638         op_ret = futex_atomic_op_inuser(op, uaddr2);
1639         if (unlikely(op_ret < 0)) {
1640
1641                 double_unlock_hb(hb1, hb2);
1642
1643 #ifndef CONFIG_MMU
1644                 /*
1645                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1646                  * but we might get them from range checking
1647                  */
1648                 ret = op_ret;
1649                 goto out_put_keys;
1650 #endif
1651
1652                 if (unlikely(op_ret != -EFAULT)) {
1653                         ret = op_ret;
1654                         goto out_put_keys;
1655                 }
1656
1657                 ret = fault_in_user_writeable(uaddr2);
1658                 if (ret)
1659                         goto out_put_keys;
1660
1661                 if (!(flags & FLAGS_SHARED))
1662                         goto retry_private;
1663
1664                 put_futex_key(&key2);
1665                 put_futex_key(&key1);
1666                 goto retry;
1667         }
1668
1669         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1670                 if (match_futex (&this->key, &key1)) {
1671                         if (this->pi_state || this->rt_waiter) {
1672                                 ret = -EINVAL;
1673                                 goto out_unlock;
1674                         }
1675                         mark_wake_futex(&wake_q, this);
1676                         if (++ret >= nr_wake)
1677                                 break;
1678                 }
1679         }
1680
1681         if (op_ret > 0) {
1682                 op_ret = 0;
1683                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1684                         if (match_futex (&this->key, &key2)) {
1685                                 if (this->pi_state || this->rt_waiter) {
1686                                         ret = -EINVAL;
1687                                         goto out_unlock;
1688                                 }
1689                                 mark_wake_futex(&wake_q, this);
1690                                 if (++op_ret >= nr_wake2)
1691                                         break;
1692                         }
1693                 }
1694                 ret += op_ret;
1695         }
1696
1697 out_unlock:
1698         double_unlock_hb(hb1, hb2);
1699         wake_up_q(&wake_q);
1700 out_put_keys:
1701         put_futex_key(&key2);
1702 out_put_key1:
1703         put_futex_key(&key1);
1704 out:
1705         return ret;
1706 }
1707
1708 /**
1709  * requeue_futex() - Requeue a futex_q from one hb to another
1710  * @q:          the futex_q to requeue
1711  * @hb1:        the source hash_bucket
1712  * @hb2:        the target hash_bucket
1713  * @key2:       the new key for the requeued futex_q
1714  */
1715 static inline
1716 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1717                    struct futex_hash_bucket *hb2, union futex_key *key2)
1718 {
1719
1720         /*
1721          * If key1 and key2 hash to the same bucket, no need to
1722          * requeue.
1723          */
1724         if (likely(&hb1->chain != &hb2->chain)) {
1725                 plist_del(&q->list, &hb1->chain);
1726                 hb_waiters_dec(hb1);
1727                 hb_waiters_inc(hb2);
1728                 plist_add(&q->list, &hb2->chain);
1729                 q->lock_ptr = &hb2->lock;
1730         }
1731         get_futex_key_refs(key2);
1732         q->key = *key2;
1733 }
1734
1735 /**
1736  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1737  * @q:          the futex_q
1738  * @key:        the key of the requeue target futex
1739  * @hb:         the hash_bucket of the requeue target futex
1740  *
1741  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1742  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1743  * to the requeue target futex so the waiter can detect the wakeup on the right
1744  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1745  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1746  * to protect access to the pi_state to fixup the owner later.  Must be called
1747  * with both q->lock_ptr and hb->lock held.
1748  */
1749 static inline
1750 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1751                            struct futex_hash_bucket *hb)
1752 {
1753         get_futex_key_refs(key);
1754         q->key = *key;
1755
1756         __unqueue_futex(q);
1757
1758         WARN_ON(!q->rt_waiter);
1759         q->rt_waiter = NULL;
1760
1761         q->lock_ptr = &hb->lock;
1762
1763         wake_up_state(q->task, TASK_NORMAL);
1764 }
1765
1766 /**
1767  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1768  * @pifutex:            the user address of the to futex
1769  * @hb1:                the from futex hash bucket, must be locked by the caller
1770  * @hb2:                the to futex hash bucket, must be locked by the caller
1771  * @key1:               the from futex key
1772  * @key2:               the to futex key
1773  * @ps:                 address to store the pi_state pointer
1774  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1775  *
1776  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1777  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1778  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1779  * hb1 and hb2 must be held by the caller.
1780  *
1781  * Return:
1782  *  -  0 - failed to acquire the lock atomically;
1783  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1784  *  - <0 - error
1785  */
1786 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1787                                  struct futex_hash_bucket *hb1,
1788                                  struct futex_hash_bucket *hb2,
1789                                  union futex_key *key1, union futex_key *key2,
1790                                  struct futex_pi_state **ps, int set_waiters)
1791 {
1792         struct futex_q *top_waiter = NULL;
1793         u32 curval;
1794         int ret, vpid;
1795
1796         if (get_futex_value_locked(&curval, pifutex))
1797                 return -EFAULT;
1798
1799         if (unlikely(should_fail_futex(true)))
1800                 return -EFAULT;
1801
1802         /*
1803          * Find the top_waiter and determine if there are additional waiters.
1804          * If the caller intends to requeue more than 1 waiter to pifutex,
1805          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1806          * as we have means to handle the possible fault.  If not, don't set
1807          * the bit unecessarily as it will force the subsequent unlock to enter
1808          * the kernel.
1809          */
1810         top_waiter = futex_top_waiter(hb1, key1);
1811
1812         /* There are no waiters, nothing for us to do. */
1813         if (!top_waiter)
1814                 return 0;
1815
1816         /* Ensure we requeue to the expected futex. */
1817         if (!match_futex(top_waiter->requeue_pi_key, key2))
1818                 return -EINVAL;
1819
1820         /*
1821          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1822          * the contended case or if set_waiters is 1.  The pi_state is returned
1823          * in ps in contended cases.
1824          */
1825         vpid = task_pid_vnr(top_waiter->task);
1826         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1827                                    set_waiters);
1828         if (ret == 1) {
1829                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1830                 return vpid;
1831         }
1832         return ret;
1833 }
1834
1835 /**
1836  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1837  * @uaddr1:     source futex user address
1838  * @flags:      futex flags (FLAGS_SHARED, etc.)
1839  * @uaddr2:     target futex user address
1840  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1841  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1842  * @cmpval:     @uaddr1 expected value (or %NULL)
1843  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1844  *              pi futex (pi to pi requeue is not supported)
1845  *
1846  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1847  * uaddr2 atomically on behalf of the top waiter.
1848  *
1849  * Return:
1850  *  - >=0 - on success, the number of tasks requeued or woken;
1851  *  -  <0 - on error
1852  */
1853 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1854                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1855                          u32 *cmpval, int requeue_pi)
1856 {
1857         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1858         int drop_count = 0, task_count = 0, ret;
1859         struct futex_pi_state *pi_state = NULL;
1860         struct futex_hash_bucket *hb1, *hb2;
1861         struct futex_q *this, *next;
1862         DEFINE_WAKE_Q(wake_q);
1863
1864         /*
1865          * When PI not supported: return -ENOSYS if requeue_pi is true,
1866          * consequently the compiler knows requeue_pi is always false past
1867          * this point which will optimize away all the conditional code
1868          * further down.
1869          */
1870         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1871                 return -ENOSYS;
1872
1873         if (requeue_pi) {
1874                 /*
1875                  * Requeue PI only works on two distinct uaddrs. This
1876                  * check is only valid for private futexes. See below.
1877                  */
1878                 if (uaddr1 == uaddr2)
1879                         return -EINVAL;
1880
1881                 /*
1882                  * requeue_pi requires a pi_state, try to allocate it now
1883                  * without any locks in case it fails.
1884                  */
1885                 if (refill_pi_state_cache())
1886                         return -ENOMEM;
1887                 /*
1888                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1889                  * + nr_requeue, since it acquires the rt_mutex prior to
1890                  * returning to userspace, so as to not leave the rt_mutex with
1891                  * waiters and no owner.  However, second and third wake-ups
1892                  * cannot be predicted as they involve race conditions with the
1893                  * first wake and a fault while looking up the pi_state.  Both
1894                  * pthread_cond_signal() and pthread_cond_broadcast() should
1895                  * use nr_wake=1.
1896                  */
1897                 if (nr_wake != 1)
1898                         return -EINVAL;
1899         }
1900
1901 retry:
1902         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1903         if (unlikely(ret != 0))
1904                 goto out;
1905         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1906                             requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1907         if (unlikely(ret != 0))
1908                 goto out_put_key1;
1909
1910         /*
1911          * The check above which compares uaddrs is not sufficient for
1912          * shared futexes. We need to compare the keys:
1913          */
1914         if (requeue_pi && match_futex(&key1, &key2)) {
1915                 ret = -EINVAL;
1916                 goto out_put_keys;
1917         }
1918
1919         hb1 = hash_futex(&key1);
1920         hb2 = hash_futex(&key2);
1921
1922 retry_private:
1923         hb_waiters_inc(hb2);
1924         double_lock_hb(hb1, hb2);
1925
1926         if (likely(cmpval != NULL)) {
1927                 u32 curval;
1928
1929                 ret = get_futex_value_locked(&curval, uaddr1);
1930
1931                 if (unlikely(ret)) {
1932                         double_unlock_hb(hb1, hb2);
1933                         hb_waiters_dec(hb2);
1934
1935                         ret = get_user(curval, uaddr1);
1936                         if (ret)
1937                                 goto out_put_keys;
1938
1939                         if (!(flags & FLAGS_SHARED))
1940                                 goto retry_private;
1941
1942                         put_futex_key(&key2);
1943                         put_futex_key(&key1);
1944                         goto retry;
1945                 }
1946                 if (curval != *cmpval) {
1947                         ret = -EAGAIN;
1948                         goto out_unlock;
1949                 }
1950         }
1951
1952         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1953                 /*
1954                  * Attempt to acquire uaddr2 and wake the top waiter. If we
1955                  * intend to requeue waiters, force setting the FUTEX_WAITERS
1956                  * bit.  We force this here where we are able to easily handle
1957                  * faults rather in the requeue loop below.
1958                  */
1959                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1960                                                  &key2, &pi_state, nr_requeue);
1961
1962                 /*
1963                  * At this point the top_waiter has either taken uaddr2 or is
1964                  * waiting on it.  If the former, then the pi_state will not
1965                  * exist yet, look it up one more time to ensure we have a
1966                  * reference to it. If the lock was taken, ret contains the
1967                  * vpid of the top waiter task.
1968                  * If the lock was not taken, we have pi_state and an initial
1969                  * refcount on it. In case of an error we have nothing.
1970                  */
1971                 if (ret > 0) {
1972                         WARN_ON(pi_state);
1973                         drop_count++;
1974                         task_count++;
1975                         /*
1976                          * If we acquired the lock, then the user space value
1977                          * of uaddr2 should be vpid. It cannot be changed by
1978                          * the top waiter as it is blocked on hb2 lock if it
1979                          * tries to do so. If something fiddled with it behind
1980                          * our back the pi state lookup might unearth it. So
1981                          * we rather use the known value than rereading and
1982                          * handing potential crap to lookup_pi_state.
1983                          *
1984                          * If that call succeeds then we have pi_state and an
1985                          * initial refcount on it.
1986                          */
1987                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
1988                 }
1989
1990                 switch (ret) {
1991                 case 0:
1992                         /* We hold a reference on the pi state. */
1993                         break;
1994
1995                         /* If the above failed, then pi_state is NULL */
1996                 case -EFAULT:
1997                         double_unlock_hb(hb1, hb2);
1998                         hb_waiters_dec(hb2);
1999                         put_futex_key(&key2);
2000                         put_futex_key(&key1);
2001                         ret = fault_in_user_writeable(uaddr2);
2002                         if (!ret)
2003                                 goto retry;
2004                         goto out;
2005                 case -EAGAIN:
2006                         /*
2007                          * Two reasons for this:
2008                          * - Owner is exiting and we just wait for the
2009                          *   exit to complete.
2010                          * - The user space value changed.
2011                          */
2012                         double_unlock_hb(hb1, hb2);
2013                         hb_waiters_dec(hb2);
2014                         put_futex_key(&key2);
2015                         put_futex_key(&key1);
2016                         cond_resched();
2017                         goto retry;
2018                 default:
2019                         goto out_unlock;
2020                 }
2021         }
2022
2023         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2024                 if (task_count - nr_wake >= nr_requeue)
2025                         break;
2026
2027                 if (!match_futex(&this->key, &key1))
2028                         continue;
2029
2030                 /*
2031                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2032                  * be paired with each other and no other futex ops.
2033                  *
2034                  * We should never be requeueing a futex_q with a pi_state,
2035                  * which is awaiting a futex_unlock_pi().
2036                  */
2037                 if ((requeue_pi && !this->rt_waiter) ||
2038                     (!requeue_pi && this->rt_waiter) ||
2039                     this->pi_state) {
2040                         ret = -EINVAL;
2041                         break;
2042                 }
2043
2044                 /*
2045                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2046                  * lock, we already woke the top_waiter.  If not, it will be
2047                  * woken by futex_unlock_pi().
2048                  */
2049                 if (++task_count <= nr_wake && !requeue_pi) {
2050                         mark_wake_futex(&wake_q, this);
2051                         continue;
2052                 }
2053
2054                 /* Ensure we requeue to the expected futex for requeue_pi. */
2055                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2056                         ret = -EINVAL;
2057                         break;
2058                 }
2059
2060                 /*
2061                  * Requeue nr_requeue waiters and possibly one more in the case
2062                  * of requeue_pi if we couldn't acquire the lock atomically.
2063                  */
2064                 if (requeue_pi) {
2065                         /*
2066                          * Prepare the waiter to take the rt_mutex. Take a
2067                          * refcount on the pi_state and store the pointer in
2068                          * the futex_q object of the waiter.
2069                          */
2070                         get_pi_state(pi_state);
2071                         this->pi_state = pi_state;
2072                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2073                                                         this->rt_waiter,
2074                                                         this->task);
2075                         if (ret == 1) {
2076                                 /*
2077                                  * We got the lock. We do neither drop the
2078                                  * refcount on pi_state nor clear
2079                                  * this->pi_state because the waiter needs the
2080                                  * pi_state for cleaning up the user space
2081                                  * value. It will drop the refcount after
2082                                  * doing so.
2083                                  */
2084                                 requeue_pi_wake_futex(this, &key2, hb2);
2085                                 drop_count++;
2086                                 continue;
2087                         } else if (ret) {
2088                                 /*
2089                                  * rt_mutex_start_proxy_lock() detected a
2090                                  * potential deadlock when we tried to queue
2091                                  * that waiter. Drop the pi_state reference
2092                                  * which we took above and remove the pointer
2093                                  * to the state from the waiters futex_q
2094                                  * object.
2095                                  */
2096                                 this->pi_state = NULL;
2097                                 put_pi_state(pi_state);
2098                                 /*
2099                                  * We stop queueing more waiters and let user
2100                                  * space deal with the mess.
2101                                  */
2102                                 break;
2103                         }
2104                 }
2105                 requeue_futex(this, hb1, hb2, &key2);
2106                 drop_count++;
2107         }
2108
2109         /*
2110          * We took an extra initial reference to the pi_state either
2111          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2112          * need to drop it here again.
2113          */
2114         put_pi_state(pi_state);
2115
2116 out_unlock:
2117         double_unlock_hb(hb1, hb2);
2118         wake_up_q(&wake_q);
2119         hb_waiters_dec(hb2);
2120
2121         /*
2122          * drop_futex_key_refs() must be called outside the spinlocks. During
2123          * the requeue we moved futex_q's from the hash bucket at key1 to the
2124          * one at key2 and updated their key pointer.  We no longer need to
2125          * hold the references to key1.
2126          */
2127         while (--drop_count >= 0)
2128                 drop_futex_key_refs(&key1);
2129
2130 out_put_keys:
2131         put_futex_key(&key2);
2132 out_put_key1:
2133         put_futex_key(&key1);
2134 out:
2135         return ret ? ret : task_count;
2136 }
2137
2138 /* The key must be already stored in q->key. */
2139 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2140         __acquires(&hb->lock)
2141 {
2142         struct futex_hash_bucket *hb;
2143
2144         hb = hash_futex(&q->key);
2145
2146         /*
2147          * Increment the counter before taking the lock so that
2148          * a potential waker won't miss a to-be-slept task that is
2149          * waiting for the spinlock. This is safe as all queue_lock()
2150          * users end up calling queue_me(). Similarly, for housekeeping,
2151          * decrement the counter at queue_unlock() when some error has
2152          * occurred and we don't end up adding the task to the list.
2153          */
2154         hb_waiters_inc(hb);
2155
2156         q->lock_ptr = &hb->lock;
2157
2158         spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2159         return hb;
2160 }
2161
2162 static inline void
2163 queue_unlock(struct futex_hash_bucket *hb)
2164         __releases(&hb->lock)
2165 {
2166         spin_unlock(&hb->lock);
2167         hb_waiters_dec(hb);
2168 }
2169
2170 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2171 {
2172         int prio;
2173
2174         /*
2175          * The priority used to register this element is
2176          * - either the real thread-priority for the real-time threads
2177          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2178          * - or MAX_RT_PRIO for non-RT threads.
2179          * Thus, all RT-threads are woken first in priority order, and
2180          * the others are woken last, in FIFO order.
2181          */
2182         prio = min(current->normal_prio, MAX_RT_PRIO);
2183
2184         plist_node_init(&q->list, prio);
2185         plist_add(&q->list, &hb->chain);
2186         q->task = current;
2187 }
2188
2189 /**
2190  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2191  * @q:  The futex_q to enqueue
2192  * @hb: The destination hash bucket
2193  *
2194  * The hb->lock must be held by the caller, and is released here. A call to
2195  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2196  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2197  * or nothing if the unqueue is done as part of the wake process and the unqueue
2198  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2199  * an example).
2200  */
2201 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2202         __releases(&hb->lock)
2203 {
2204         __queue_me(q, hb);
2205         spin_unlock(&hb->lock);
2206 }
2207
2208 /**
2209  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2210  * @q:  The futex_q to unqueue
2211  *
2212  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2213  * be paired with exactly one earlier call to queue_me().
2214  *
2215  * Return:
2216  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2217  *  - 0 - if the futex_q was already removed by the waking thread
2218  */
2219 static int unqueue_me(struct futex_q *q)
2220 {
2221         spinlock_t *lock_ptr;
2222         int ret = 0;
2223
2224         /* In the common case we don't take the spinlock, which is nice. */
2225 retry:
2226         /*
2227          * q->lock_ptr can change between this read and the following spin_lock.
2228          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2229          * optimizing lock_ptr out of the logic below.
2230          */
2231         lock_ptr = READ_ONCE(q->lock_ptr);
2232         if (lock_ptr != NULL) {
2233                 spin_lock(lock_ptr);
2234                 /*
2235                  * q->lock_ptr can change between reading it and
2236                  * spin_lock(), causing us to take the wrong lock.  This
2237                  * corrects the race condition.
2238                  *
2239                  * Reasoning goes like this: if we have the wrong lock,
2240                  * q->lock_ptr must have changed (maybe several times)
2241                  * between reading it and the spin_lock().  It can
2242                  * change again after the spin_lock() but only if it was
2243                  * already changed before the spin_lock().  It cannot,
2244                  * however, change back to the original value.  Therefore
2245                  * we can detect whether we acquired the correct lock.
2246                  */
2247                 if (unlikely(lock_ptr != q->lock_ptr)) {
2248                         spin_unlock(lock_ptr);
2249                         goto retry;
2250                 }
2251                 __unqueue_futex(q);
2252
2253                 BUG_ON(q->pi_state);
2254
2255                 spin_unlock(lock_ptr);
2256                 ret = 1;
2257         }
2258
2259         drop_futex_key_refs(&q->key);
2260         return ret;
2261 }
2262
2263 /*
2264  * PI futexes can not be requeued and must remove themself from the
2265  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2266  * and dropped here.
2267  */
2268 static void unqueue_me_pi(struct futex_q *q)
2269         __releases(q->lock_ptr)
2270 {
2271         __unqueue_futex(q);
2272
2273         BUG_ON(!q->pi_state);
2274         put_pi_state(q->pi_state);
2275         q->pi_state = NULL;
2276
2277         spin_unlock(q->lock_ptr);
2278 }
2279
2280 /*
2281  * Fixup the pi_state owner with the new owner.
2282  *
2283  * Must be called with hash bucket lock held and mm->sem held for non
2284  * private futexes.
2285  */
2286 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2287                                 struct task_struct *newowner)
2288 {
2289         u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2290         struct futex_pi_state *pi_state = q->pi_state;
2291         u32 uval, uninitialized_var(curval), newval;
2292         struct task_struct *oldowner;
2293         int ret;
2294
2295         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2296
2297         oldowner = pi_state->owner;
2298         /* Owner died? */
2299         if (!pi_state->owner)
2300                 newtid |= FUTEX_OWNER_DIED;
2301
2302         /*
2303          * We are here either because we stole the rtmutex from the
2304          * previous highest priority waiter or we are the highest priority
2305          * waiter but have failed to get the rtmutex the first time.
2306          *
2307          * We have to replace the newowner TID in the user space variable.
2308          * This must be atomic as we have to preserve the owner died bit here.
2309          *
2310          * Note: We write the user space value _before_ changing the pi_state
2311          * because we can fault here. Imagine swapped out pages or a fork
2312          * that marked all the anonymous memory readonly for cow.
2313          *
2314          * Modifying pi_state _before_ the user space value would leave the
2315          * pi_state in an inconsistent state when we fault here, because we
2316          * need to drop the locks to handle the fault. This might be observed
2317          * in the PID check in lookup_pi_state.
2318          */
2319 retry:
2320         if (get_futex_value_locked(&uval, uaddr))
2321                 goto handle_fault;
2322
2323         for (;;) {
2324                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2325
2326                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2327                         goto handle_fault;
2328                 if (curval == uval)
2329                         break;
2330                 uval = curval;
2331         }
2332
2333         /*
2334          * We fixed up user space. Now we need to fix the pi_state
2335          * itself.
2336          */
2337         if (pi_state->owner != NULL) {
2338                 raw_spin_lock(&pi_state->owner->pi_lock);
2339                 WARN_ON(list_empty(&pi_state->list));
2340                 list_del_init(&pi_state->list);
2341                 raw_spin_unlock(&pi_state->owner->pi_lock);
2342         }
2343
2344         pi_state->owner = newowner;
2345
2346         raw_spin_lock(&newowner->pi_lock);
2347         WARN_ON(!list_empty(&pi_state->list));
2348         list_add(&pi_state->list, &newowner->pi_state_list);
2349         raw_spin_unlock(&newowner->pi_lock);
2350         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2351
2352         return 0;
2353
2354         /*
2355          * To handle the page fault we need to drop the locks here. That gives
2356          * the other task (either the highest priority waiter itself or the
2357          * task which stole the rtmutex) the chance to try the fixup of the
2358          * pi_state. So once we are back from handling the fault we need to
2359          * check the pi_state after reacquiring the locks and before trying to
2360          * do another fixup. When the fixup has been done already we simply
2361          * return.
2362          *
2363          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2364          * drop hb->lock since the caller owns the hb -> futex_q relation.
2365          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2366          */
2367 handle_fault:
2368         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2369         spin_unlock(q->lock_ptr);
2370
2371         ret = fault_in_user_writeable(uaddr);
2372
2373         spin_lock(q->lock_ptr);
2374         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2375
2376         /*
2377          * Check if someone else fixed it for us:
2378          */
2379         if (pi_state->owner != oldowner) {
2380                 ret = 0;
2381                 goto out_unlock;
2382         }
2383
2384         if (ret)
2385                 goto out_unlock;
2386
2387         goto retry;
2388
2389 out_unlock:
2390         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2391         return ret;
2392 }
2393
2394 static long futex_wait_restart(struct restart_block *restart);
2395
2396 /**
2397  * fixup_owner() - Post lock pi_state and corner case management
2398  * @uaddr:      user address of the futex
2399  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2400  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2401  *
2402  * After attempting to lock an rt_mutex, this function is called to cleanup
2403  * the pi_state owner as well as handle race conditions that may allow us to
2404  * acquire the lock. Must be called with the hb lock held.
2405  *
2406  * Return:
2407  *  -  1 - success, lock taken;
2408  *  -  0 - success, lock not taken;
2409  *  - <0 - on error (-EFAULT)
2410  */
2411 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2412 {
2413         int ret = 0;
2414
2415         if (locked) {
2416                 /*
2417                  * Got the lock. We might not be the anticipated owner if we
2418                  * did a lock-steal - fix up the PI-state in that case:
2419                  *
2420                  * We can safely read pi_state->owner without holding wait_lock
2421                  * because we now own the rt_mutex, only the owner will attempt
2422                  * to change it.
2423                  */
2424                 if (q->pi_state->owner != current)
2425                         ret = fixup_pi_state_owner(uaddr, q, current);
2426                 goto out;
2427         }
2428
2429         /*
2430          * Paranoia check. If we did not take the lock, then we should not be
2431          * the owner of the rt_mutex.
2432          */
2433         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2434                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2435                                 "pi-state %p\n", ret,
2436                                 q->pi_state->pi_mutex.owner,
2437                                 q->pi_state->owner);
2438         }
2439
2440 out:
2441         return ret ? ret : locked;
2442 }
2443
2444 /**
2445  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2446  * @hb:         the futex hash bucket, must be locked by the caller
2447  * @q:          the futex_q to queue up on
2448  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2449  */
2450 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2451                                 struct hrtimer_sleeper *timeout)
2452 {
2453         /*
2454          * The task state is guaranteed to be set before another task can
2455          * wake it. set_current_state() is implemented using smp_store_mb() and
2456          * queue_me() calls spin_unlock() upon completion, both serializing
2457          * access to the hash list and forcing another memory barrier.
2458          */
2459         set_current_state(TASK_INTERRUPTIBLE);
2460         queue_me(q, hb);
2461
2462         /* Arm the timer */
2463         if (timeout)
2464                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2465
2466         /*
2467          * If we have been removed from the hash list, then another task
2468          * has tried to wake us, and we can skip the call to schedule().
2469          */
2470         if (likely(!plist_node_empty(&q->list))) {
2471                 /*
2472                  * If the timer has already expired, current will already be
2473                  * flagged for rescheduling. Only call schedule if there
2474                  * is no timeout, or if it has yet to expire.
2475                  */
2476                 if (!timeout || timeout->task)
2477                         freezable_schedule();
2478         }
2479         __set_current_state(TASK_RUNNING);
2480 }
2481
2482 /**
2483  * futex_wait_setup() - Prepare to wait on a futex
2484  * @uaddr:      the futex userspace address
2485  * @val:        the expected value
2486  * @flags:      futex flags (FLAGS_SHARED, etc.)
2487  * @q:          the associated futex_q
2488  * @hb:         storage for hash_bucket pointer to be returned to caller
2489  *
2490  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2491  * compare it with the expected value.  Handle atomic faults internally.
2492  * Return with the hb lock held and a q.key reference on success, and unlocked
2493  * with no q.key reference on failure.
2494  *
2495  * Return:
2496  *  -  0 - uaddr contains val and hb has been locked;
2497  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2498  */
2499 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2500                            struct futex_q *q, struct futex_hash_bucket **hb)
2501 {
2502         u32 uval;
2503         int ret;
2504
2505         /*
2506          * Access the page AFTER the hash-bucket is locked.
2507          * Order is important:
2508          *
2509          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2510          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2511          *
2512          * The basic logical guarantee of a futex is that it blocks ONLY
2513          * if cond(var) is known to be true at the time of blocking, for
2514          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2515          * would open a race condition where we could block indefinitely with
2516          * cond(var) false, which would violate the guarantee.
2517          *
2518          * On the other hand, we insert q and release the hash-bucket only
2519          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2520          * absorb a wakeup if *uaddr does not match the desired values
2521          * while the syscall executes.
2522          */
2523 retry:
2524         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2525         if (unlikely(ret != 0))
2526                 return ret;
2527
2528 retry_private:
2529         *hb = queue_lock(q);
2530
2531         ret = get_futex_value_locked(&uval, uaddr);
2532
2533         if (ret) {
2534                 queue_unlock(*hb);
2535
2536                 ret = get_user(uval, uaddr);
2537                 if (ret)
2538                         goto out;
2539
2540                 if (!(flags & FLAGS_SHARED))
2541                         goto retry_private;
2542
2543                 put_futex_key(&q->key);
2544                 goto retry;
2545         }
2546
2547         if (uval != val) {
2548                 queue_unlock(*hb);
2549                 ret = -EWOULDBLOCK;
2550         }
2551
2552 out:
2553         if (ret)
2554                 put_futex_key(&q->key);
2555         return ret;
2556 }
2557
2558 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2559                       ktime_t *abs_time, u32 bitset)
2560 {
2561         struct hrtimer_sleeper timeout, *to = NULL;
2562         struct restart_block *restart;
2563         struct futex_hash_bucket *hb;
2564         struct futex_q q = futex_q_init;
2565         int ret;
2566
2567         if (!bitset)
2568                 return -EINVAL;
2569         q.bitset = bitset;
2570
2571         if (abs_time) {
2572                 to = &timeout;
2573
2574                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2575                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2576                                       HRTIMER_MODE_ABS);
2577                 hrtimer_init_sleeper(to, current);
2578                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2579                                              current->timer_slack_ns);
2580         }
2581
2582 retry:
2583         /*
2584          * Prepare to wait on uaddr. On success, holds hb lock and increments
2585          * q.key refs.
2586          */
2587         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2588         if (ret)
2589                 goto out;
2590
2591         /* queue_me and wait for wakeup, timeout, or a signal. */
2592         futex_wait_queue_me(hb, &q, to);
2593
2594         /* If we were woken (and unqueued), we succeeded, whatever. */
2595         ret = 0;
2596         /* unqueue_me() drops q.key ref */
2597         if (!unqueue_me(&q))
2598                 goto out;
2599         ret = -ETIMEDOUT;
2600         if (to && !to->task)
2601                 goto out;
2602
2603         /*
2604          * We expect signal_pending(current), but we might be the
2605          * victim of a spurious wakeup as well.
2606          */
2607         if (!signal_pending(current))
2608                 goto retry;
2609
2610         ret = -ERESTARTSYS;
2611         if (!abs_time)
2612                 goto out;
2613
2614         restart = &current->restart_block;
2615         restart->fn = futex_wait_restart;
2616         restart->futex.uaddr = uaddr;
2617         restart->futex.val = val;
2618         restart->futex.time = *abs_time;
2619         restart->futex.bitset = bitset;
2620         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2621
2622         ret = -ERESTART_RESTARTBLOCK;
2623
2624 out:
2625         if (to) {
2626                 hrtimer_cancel(&to->timer);
2627                 destroy_hrtimer_on_stack(&to->timer);
2628         }
2629         return ret;
2630 }
2631
2632
2633 static long futex_wait_restart(struct restart_block *restart)
2634 {
2635         u32 __user *uaddr = restart->futex.uaddr;
2636         ktime_t t, *tp = NULL;
2637
2638         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2639                 t = restart->futex.time;
2640                 tp = &t;
2641         }
2642         restart->fn = do_no_restart_syscall;
2643
2644         return (long)futex_wait(uaddr, restart->futex.flags,
2645                                 restart->futex.val, tp, restart->futex.bitset);
2646 }
2647
2648
2649 /*
2650  * Userspace tried a 0 -> TID atomic transition of the futex value
2651  * and failed. The kernel side here does the whole locking operation:
2652  * if there are waiters then it will block as a consequence of relying
2653  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2654  * a 0 value of the futex too.).
2655  *
2656  * Also serves as futex trylock_pi()'ing, and due semantics.
2657  */
2658 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2659                          ktime_t *time, int trylock)
2660 {
2661         struct hrtimer_sleeper timeout, *to = NULL;
2662         struct futex_pi_state *pi_state = NULL;
2663         struct rt_mutex_waiter rt_waiter;
2664         struct futex_hash_bucket *hb;
2665         struct futex_q q = futex_q_init;
2666         int res, ret;
2667
2668         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2669                 return -ENOSYS;
2670
2671         if (refill_pi_state_cache())
2672                 return -ENOMEM;
2673
2674         if (time) {
2675                 to = &timeout;
2676                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2677                                       HRTIMER_MODE_ABS);
2678                 hrtimer_init_sleeper(to, current);
2679                 hrtimer_set_expires(&to->timer, *time);
2680         }
2681
2682 retry:
2683         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2684         if (unlikely(ret != 0))
2685                 goto out;
2686
2687 retry_private:
2688         hb = queue_lock(&q);
2689
2690         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2691         if (unlikely(ret)) {
2692                 /*
2693                  * Atomic work succeeded and we got the lock,
2694                  * or failed. Either way, we do _not_ block.
2695                  */
2696                 switch (ret) {
2697                 case 1:
2698                         /* We got the lock. */
2699                         ret = 0;
2700                         goto out_unlock_put_key;
2701                 case -EFAULT:
2702                         goto uaddr_faulted;
2703                 case -EAGAIN:
2704                         /*
2705                          * Two reasons for this:
2706                          * - Task is exiting and we just wait for the
2707                          *   exit to complete.
2708                          * - The user space value changed.
2709                          */
2710                         queue_unlock(hb);
2711                         put_futex_key(&q.key);
2712                         cond_resched();
2713                         goto retry;
2714                 default:
2715                         goto out_unlock_put_key;
2716                 }
2717         }
2718
2719         WARN_ON(!q.pi_state);
2720
2721         /*
2722          * Only actually queue now that the atomic ops are done:
2723          */
2724         __queue_me(&q, hb);
2725
2726         if (trylock) {
2727                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2728                 /* Fixup the trylock return value: */
2729                 ret = ret ? 0 : -EWOULDBLOCK;
2730                 goto no_block;
2731         }
2732
2733         rt_mutex_init_waiter(&rt_waiter);
2734
2735         /*
2736          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2737          * hold it while doing rt_mutex_start_proxy(), because then it will
2738          * include hb->lock in the blocking chain, even through we'll not in
2739          * fact hold it while blocking. This will lead it to report -EDEADLK
2740          * and BUG when futex_unlock_pi() interleaves with this.
2741          *
2742          * Therefore acquire wait_lock while holding hb->lock, but drop the
2743          * latter before calling rt_mutex_start_proxy_lock(). This still fully
2744          * serializes against futex_unlock_pi() as that does the exact same
2745          * lock handoff sequence.
2746          */
2747         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2748         spin_unlock(q.lock_ptr);
2749         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2750         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2751
2752         if (ret) {
2753                 if (ret == 1)
2754                         ret = 0;
2755
2756                 spin_lock(q.lock_ptr);
2757                 goto no_block;
2758         }
2759
2760
2761         if (unlikely(to))
2762                 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2763
2764         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2765
2766         spin_lock(q.lock_ptr);
2767         /*
2768          * If we failed to acquire the lock (signal/timeout), we must
2769          * first acquire the hb->lock before removing the lock from the
2770          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2771          * wait lists consistent.
2772          *
2773          * In particular; it is important that futex_unlock_pi() can not
2774          * observe this inconsistency.
2775          */
2776         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2777                 ret = 0;
2778
2779 no_block:
2780         /*
2781          * Fixup the pi_state owner and possibly acquire the lock if we
2782          * haven't already.
2783          */
2784         res = fixup_owner(uaddr, &q, !ret);
2785         /*
2786          * If fixup_owner() returned an error, proprogate that.  If it acquired
2787          * the lock, clear our -ETIMEDOUT or -EINTR.
2788          */
2789         if (res)
2790                 ret = (res < 0) ? res : 0;
2791
2792         /*
2793          * If fixup_owner() faulted and was unable to handle the fault, unlock
2794          * it and return the fault to userspace.
2795          */
2796         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2797                 pi_state = q.pi_state;
2798                 get_pi_state(pi_state);
2799         }
2800
2801         /* Unqueue and drop the lock */
2802         unqueue_me_pi(&q);
2803
2804         if (pi_state) {
2805                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2806                 put_pi_state(pi_state);
2807         }
2808
2809         goto out_put_key;
2810
2811 out_unlock_put_key:
2812         queue_unlock(hb);
2813
2814 out_put_key:
2815         put_futex_key(&q.key);
2816 out:
2817         if (to) {
2818                 hrtimer_cancel(&to->timer);
2819                 destroy_hrtimer_on_stack(&to->timer);
2820         }
2821         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2822
2823 uaddr_faulted:
2824         queue_unlock(hb);
2825
2826         ret = fault_in_user_writeable(uaddr);
2827         if (ret)
2828                 goto out_put_key;
2829
2830         if (!(flags & FLAGS_SHARED))
2831                 goto retry_private;
2832
2833         put_futex_key(&q.key);
2834         goto retry;
2835 }
2836
2837 /*
2838  * Userspace attempted a TID -> 0 atomic transition, and failed.
2839  * This is the in-kernel slowpath: we look up the PI state (if any),
2840  * and do the rt-mutex unlock.
2841  */
2842 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2843 {
2844         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2845         union futex_key key = FUTEX_KEY_INIT;
2846         struct futex_hash_bucket *hb;
2847         struct futex_q *top_waiter;
2848         int ret;
2849
2850         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2851                 return -ENOSYS;
2852
2853 retry:
2854         if (get_user(uval, uaddr))
2855                 return -EFAULT;
2856         /*
2857          * We release only a lock we actually own:
2858          */
2859         if ((uval & FUTEX_TID_MASK) != vpid)
2860                 return -EPERM;
2861
2862         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2863         if (ret)
2864                 return ret;
2865
2866         hb = hash_futex(&key);
2867         spin_lock(&hb->lock);
2868
2869         /*
2870          * Check waiters first. We do not trust user space values at
2871          * all and we at least want to know if user space fiddled
2872          * with the futex value instead of blindly unlocking.
2873          */
2874         top_waiter = futex_top_waiter(hb, &key);
2875         if (top_waiter) {
2876                 struct futex_pi_state *pi_state = top_waiter->pi_state;
2877
2878                 ret = -EINVAL;
2879                 if (!pi_state)
2880                         goto out_unlock;
2881
2882                 /*
2883                  * If current does not own the pi_state then the futex is
2884                  * inconsistent and user space fiddled with the futex value.
2885                  */
2886                 if (pi_state->owner != current)
2887                         goto out_unlock;
2888
2889                 get_pi_state(pi_state);
2890                 /*
2891                  * By taking wait_lock while still holding hb->lock, we ensure
2892                  * there is no point where we hold neither; and therefore
2893                  * wake_futex_pi() must observe a state consistent with what we
2894                  * observed.
2895                  */
2896                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2897                 spin_unlock(&hb->lock);
2898
2899                 /* drops pi_state->pi_mutex.wait_lock */
2900                 ret = wake_futex_pi(uaddr, uval, pi_state);
2901
2902                 put_pi_state(pi_state);
2903
2904                 /*
2905                  * Success, we're done! No tricky corner cases.
2906                  */
2907                 if (!ret)
2908                         goto out_putkey;
2909                 /*
2910                  * The atomic access to the futex value generated a
2911                  * pagefault, so retry the user-access and the wakeup:
2912                  */
2913                 if (ret == -EFAULT)
2914                         goto pi_faulted;
2915                 /*
2916                  * A unconditional UNLOCK_PI op raced against a waiter
2917                  * setting the FUTEX_WAITERS bit. Try again.
2918                  */
2919                 if (ret == -EAGAIN) {
2920                         put_futex_key(&key);
2921                         goto retry;
2922                 }
2923                 /*
2924                  * wake_futex_pi has detected invalid state. Tell user
2925                  * space.
2926                  */
2927                 goto out_putkey;
2928         }
2929
2930         /*
2931          * We have no kernel internal state, i.e. no waiters in the
2932          * kernel. Waiters which are about to queue themselves are stuck
2933          * on hb->lock. So we can safely ignore them. We do neither
2934          * preserve the WAITERS bit not the OWNER_DIED one. We are the
2935          * owner.
2936          */
2937         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2938                 spin_unlock(&hb->lock);
2939                 goto pi_faulted;
2940         }
2941
2942         /*
2943          * If uval has changed, let user space handle it.
2944          */
2945         ret = (curval == uval) ? 0 : -EAGAIN;
2946
2947 out_unlock:
2948         spin_unlock(&hb->lock);
2949 out_putkey:
2950         put_futex_key(&key);
2951         return ret;
2952
2953 pi_faulted:
2954         put_futex_key(&key);
2955
2956         ret = fault_in_user_writeable(uaddr);
2957         if (!ret)
2958                 goto retry;
2959
2960         return ret;
2961 }
2962
2963 /**
2964  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2965  * @hb:         the hash_bucket futex_q was original enqueued on
2966  * @q:          the futex_q woken while waiting to be requeued
2967  * @key2:       the futex_key of the requeue target futex
2968  * @timeout:    the timeout associated with the wait (NULL if none)
2969  *
2970  * Detect if the task was woken on the initial futex as opposed to the requeue
2971  * target futex.  If so, determine if it was a timeout or a signal that caused
2972  * the wakeup and return the appropriate error code to the caller.  Must be
2973  * called with the hb lock held.
2974  *
2975  * Return:
2976  *  -  0 = no early wakeup detected;
2977  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
2978  */
2979 static inline
2980 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2981                                    struct futex_q *q, union futex_key *key2,
2982                                    struct hrtimer_sleeper *timeout)
2983 {
2984         int ret = 0;
2985
2986         /*
2987          * With the hb lock held, we avoid races while we process the wakeup.
2988          * We only need to hold hb (and not hb2) to ensure atomicity as the
2989          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2990          * It can't be requeued from uaddr2 to something else since we don't
2991          * support a PI aware source futex for requeue.
2992          */
2993         if (!match_futex(&q->key, key2)) {
2994                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2995                 /*
2996                  * We were woken prior to requeue by a timeout or a signal.
2997                  * Unqueue the futex_q and determine which it was.
2998                  */
2999                 plist_del(&q->list, &hb->chain);
3000                 hb_waiters_dec(hb);
3001
3002                 /* Handle spurious wakeups gracefully */
3003                 ret = -EWOULDBLOCK;
3004                 if (timeout && !timeout->task)
3005                         ret = -ETIMEDOUT;
3006                 else if (signal_pending(current))
3007                         ret = -ERESTARTNOINTR;
3008         }
3009         return ret;
3010 }
3011
3012 /**
3013  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3014  * @uaddr:      the futex we initially wait on (non-pi)
3015  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3016  *              the same type, no requeueing from private to shared, etc.
3017  * @val:        the expected value of uaddr
3018  * @abs_time:   absolute timeout
3019  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3020  * @uaddr2:     the pi futex we will take prior to returning to user-space
3021  *
3022  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3023  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3024  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3025  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3026  * without one, the pi logic would not know which task to boost/deboost, if
3027  * there was a need to.
3028  *
3029  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3030  * via the following--
3031  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3032  * 2) wakeup on uaddr2 after a requeue
3033  * 3) signal
3034  * 4) timeout
3035  *
3036  * If 3, cleanup and return -ERESTARTNOINTR.
3037  *
3038  * If 2, we may then block on trying to take the rt_mutex and return via:
3039  * 5) successful lock
3040  * 6) signal
3041  * 7) timeout
3042  * 8) other lock acquisition failure
3043  *
3044  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3045  *
3046  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3047  *
3048  * Return:
3049  *  -  0 - On success;
3050  *  - <0 - On error
3051  */
3052 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3053                                  u32 val, ktime_t *abs_time, u32 bitset,
3054                                  u32 __user *uaddr2)
3055 {
3056         struct hrtimer_sleeper timeout, *to = NULL;
3057         struct futex_pi_state *pi_state = NULL;
3058         struct rt_mutex_waiter rt_waiter;
3059         struct futex_hash_bucket *hb;
3060         union futex_key key2 = FUTEX_KEY_INIT;
3061         struct futex_q q = futex_q_init;
3062         int res, ret;
3063
3064         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3065                 return -ENOSYS;
3066
3067         if (uaddr == uaddr2)
3068                 return -EINVAL;
3069
3070         if (!bitset)
3071                 return -EINVAL;
3072
3073         if (abs_time) {
3074                 to = &timeout;
3075                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3076                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
3077                                       HRTIMER_MODE_ABS);
3078                 hrtimer_init_sleeper(to, current);
3079                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3080                                              current->timer_slack_ns);
3081         }
3082
3083         /*
3084          * The waiter is allocated on our stack, manipulated by the requeue
3085          * code while we sleep on uaddr.
3086          */
3087         rt_mutex_init_waiter(&rt_waiter);
3088
3089         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3090         if (unlikely(ret != 0))
3091                 goto out;
3092
3093         q.bitset = bitset;
3094         q.rt_waiter = &rt_waiter;
3095         q.requeue_pi_key = &key2;
3096
3097         /*
3098          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3099          * count.
3100          */
3101         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3102         if (ret)
3103                 goto out_key2;
3104
3105         /*
3106          * The check above which compares uaddrs is not sufficient for
3107          * shared futexes. We need to compare the keys:
3108          */
3109         if (match_futex(&q.key, &key2)) {
3110                 queue_unlock(hb);
3111                 ret = -EINVAL;
3112                 goto out_put_keys;
3113         }
3114
3115         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3116         futex_wait_queue_me(hb, &q, to);
3117
3118         spin_lock(&hb->lock);
3119         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3120         spin_unlock(&hb->lock);
3121         if (ret)
3122                 goto out_put_keys;
3123
3124         /*
3125          * In order for us to be here, we know our q.key == key2, and since
3126          * we took the hb->lock above, we also know that futex_requeue() has
3127          * completed and we no longer have to concern ourselves with a wakeup
3128          * race with the atomic proxy lock acquisition by the requeue code. The
3129          * futex_requeue dropped our key1 reference and incremented our key2
3130          * reference count.
3131          */
3132
3133         /* Check if the requeue code acquired the second futex for us. */
3134         if (!q.rt_waiter) {
3135                 /*
3136                  * Got the lock. We might not be the anticipated owner if we
3137                  * did a lock-steal - fix up the PI-state in that case.
3138                  */
3139                 if (q.pi_state && (q.pi_state->owner != current)) {
3140                         spin_lock(q.lock_ptr);
3141                         ret = fixup_pi_state_owner(uaddr2, &q, current);
3142                         if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3143                                 pi_state = q.pi_state;
3144                                 get_pi_state(pi_state);
3145                         }
3146                         /*
3147                          * Drop the reference to the pi state which
3148                          * the requeue_pi() code acquired for us.
3149                          */
3150                         put_pi_state(q.pi_state);
3151                         spin_unlock(q.lock_ptr);
3152                 }
3153         } else {
3154                 struct rt_mutex *pi_mutex;
3155
3156                 /*
3157                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3158                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3159                  * the pi_state.
3160                  */
3161                 WARN_ON(!q.pi_state);
3162                 pi_mutex = &q.pi_state->pi_mutex;
3163                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3164
3165                 spin_lock(q.lock_ptr);
3166                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3167                         ret = 0;
3168
3169                 debug_rt_mutex_free_waiter(&rt_waiter);
3170                 /*
3171                  * Fixup the pi_state owner and possibly acquire the lock if we
3172                  * haven't already.
3173                  */
3174                 res = fixup_owner(uaddr2, &q, !ret);
3175                 /*
3176                  * If fixup_owner() returned an error, proprogate that.  If it
3177                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3178                  */
3179                 if (res)
3180                         ret = (res < 0) ? res : 0;
3181
3182                 /*
3183                  * If fixup_pi_state_owner() faulted and was unable to handle
3184                  * the fault, unlock the rt_mutex and return the fault to
3185                  * userspace.
3186                  */
3187                 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3188                         pi_state = q.pi_state;
3189                         get_pi_state(pi_state);
3190                 }
3191
3192                 /* Unqueue and drop the lock. */
3193                 unqueue_me_pi(&q);
3194         }
3195
3196         if (pi_state) {
3197                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3198                 put_pi_state(pi_state);
3199         }
3200
3201         if (ret == -EINTR) {
3202                 /*
3203                  * We've already been requeued, but cannot restart by calling
3204                  * futex_lock_pi() directly. We could restart this syscall, but
3205                  * it would detect that the user space "val" changed and return
3206                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3207                  * -EWOULDBLOCK directly.
3208                  */
3209                 ret = -EWOULDBLOCK;
3210         }
3211
3212 out_put_keys:
3213         put_futex_key(&q.key);
3214 out_key2:
3215         put_futex_key(&key2);
3216
3217 out:
3218         if (to) {
3219                 hrtimer_cancel(&to->timer);
3220                 destroy_hrtimer_on_stack(&to->timer);
3221         }
3222         return ret;
3223 }
3224
3225 /*
3226  * Support for robust futexes: the kernel cleans up held futexes at
3227  * thread exit time.
3228  *
3229  * Implementation: user-space maintains a per-thread list of locks it
3230  * is holding. Upon do_exit(), the kernel carefully walks this list,
3231  * and marks all locks that are owned by this thread with the
3232  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3233  * always manipulated with the lock held, so the list is private and
3234  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3235  * field, to allow the kernel to clean up if the thread dies after
3236  * acquiring the lock, but just before it could have added itself to
3237  * the list. There can only be one such pending lock.
3238  */
3239
3240 /**
3241  * sys_set_robust_list() - Set the robust-futex list head of a task
3242  * @head:       pointer to the list-head
3243  * @len:        length of the list-head, as userspace expects
3244  */
3245 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3246                 size_t, len)
3247 {
3248         if (!futex_cmpxchg_enabled)
3249                 return -ENOSYS;
3250         /*
3251          * The kernel knows only one size for now:
3252          */
3253         if (unlikely(len != sizeof(*head)))
3254                 return -EINVAL;
3255
3256         current->robust_list = head;
3257
3258         return 0;
3259 }
3260
3261 /**
3262  * sys_get_robust_list() - Get the robust-futex list head of a task
3263  * @pid:        pid of the process [zero for current task]
3264  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
3265  * @len_ptr:    pointer to a length field, the kernel fills in the header size
3266  */
3267 SYSCALL_DEFINE3(get_robust_list, int, pid,
3268                 struct robust_list_head __user * __user *, head_ptr,
3269                 size_t __user *, len_ptr)
3270 {
3271         struct robust_list_head __user *head;
3272         unsigned long ret;
3273         struct task_struct *p;
3274
3275         if (!futex_cmpxchg_enabled)
3276                 return -ENOSYS;
3277
3278         rcu_read_lock();
3279
3280         ret = -ESRCH;
3281         if (!pid)
3282                 p = current;
3283         else {
3284                 p = find_task_by_vpid(pid);
3285                 if (!p)
3286                         goto err_unlock;
3287         }
3288
3289         ret = -EPERM;
3290         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3291                 goto err_unlock;
3292
3293         head = p->robust_list;
3294         rcu_read_unlock();
3295
3296         if (put_user(sizeof(*head), len_ptr))
3297                 return -EFAULT;
3298         return put_user(head, head_ptr);
3299
3300 err_unlock:
3301         rcu_read_unlock();
3302
3303         return ret;
3304 }
3305
3306 /*
3307  * Process a futex-list entry, check whether it's owned by the
3308  * dying task, and do notification if so:
3309  */
3310 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3311 {
3312         u32 uval, uninitialized_var(nval), mval;
3313
3314 retry:
3315         if (get_user(uval, uaddr))
3316                 return -1;
3317
3318         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3319                 /*
3320                  * Ok, this dying thread is truly holding a futex
3321                  * of interest. Set the OWNER_DIED bit atomically
3322                  * via cmpxchg, and if the value had FUTEX_WAITERS
3323                  * set, wake up a waiter (if any). (We have to do a
3324                  * futex_wake() even if OWNER_DIED is already set -
3325                  * to handle the rare but possible case of recursive
3326                  * thread-death.) The rest of the cleanup is done in
3327                  * userspace.