Linux 5.2-rc4
[sfrench/cifs-2.6.git] / kernel / futex.c
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
50 #include <linux/fs.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/memblock.h>
70 #include <linux/fault-inject.h>
71 #include <linux/refcount.h>
72
73 #include <asm/futex.h>
74
75 #include "locking/rtmutex_common.h"
76
77 /*
78  * READ this before attempting to hack on futexes!
79  *
80  * Basic futex operation and ordering guarantees
81  * =============================================
82  *
83  * The waiter reads the futex value in user space and calls
84  * futex_wait(). This function computes the hash bucket and acquires
85  * the hash bucket lock. After that it reads the futex user space value
86  * again and verifies that the data has not changed. If it has not changed
87  * it enqueues itself into the hash bucket, releases the hash bucket lock
88  * and schedules.
89  *
90  * The waker side modifies the user space value of the futex and calls
91  * futex_wake(). This function computes the hash bucket and acquires the
92  * hash bucket lock. Then it looks for waiters on that futex in the hash
93  * bucket and wakes them.
94  *
95  * In futex wake up scenarios where no tasks are blocked on a futex, taking
96  * the hb spinlock can be avoided and simply return. In order for this
97  * optimization to work, ordering guarantees must exist so that the waiter
98  * being added to the list is acknowledged when the list is concurrently being
99  * checked by the waker, avoiding scenarios like the following:
100  *
101  * CPU 0                               CPU 1
102  * val = *futex;
103  * sys_futex(WAIT, futex, val);
104  *   futex_wait(futex, val);
105  *   uval = *futex;
106  *                                     *futex = newval;
107  *                                     sys_futex(WAKE, futex);
108  *                                       futex_wake(futex);
109  *                                       if (queue_empty())
110  *                                         return;
111  *   if (uval == val)
112  *      lock(hash_bucket(futex));
113  *      queue();
114  *     unlock(hash_bucket(futex));
115  *     schedule();
116  *
117  * This would cause the waiter on CPU 0 to wait forever because it
118  * missed the transition of the user space value from val to newval
119  * and the waker did not find the waiter in the hash bucket queue.
120  *
121  * The correct serialization ensures that a waiter either observes
122  * the changed user space value before blocking or is woken by a
123  * concurrent waker:
124  *
125  * CPU 0                                 CPU 1
126  * val = *futex;
127  * sys_futex(WAIT, futex, val);
128  *   futex_wait(futex, val);
129  *
130  *   waiters++; (a)
131  *   smp_mb(); (A) <-- paired with -.
132  *                                  |
133  *   lock(hash_bucket(futex));      |
134  *                                  |
135  *   uval = *futex;                 |
136  *                                  |        *futex = newval;
137  *                                  |        sys_futex(WAKE, futex);
138  *                                  |          futex_wake(futex);
139  *                                  |
140  *                                  `--------> smp_mb(); (B)
141  *   if (uval == val)
142  *     queue();
143  *     unlock(hash_bucket(futex));
144  *     schedule();                         if (waiters)
145  *                                           lock(hash_bucket(futex));
146  *   else                                    wake_waiters(futex);
147  *     waiters--; (b)                        unlock(hash_bucket(futex));
148  *
149  * Where (A) orders the waiters increment and the futex value read through
150  * atomic operations (see hb_waiters_inc) and where (B) orders the write
151  * to futex and the waiters read -- this is done by the barriers for both
152  * shared and private futexes in get_futex_key_refs().
153  *
154  * This yields the following case (where X:=waiters, Y:=futex):
155  *
156  *      X = Y = 0
157  *
158  *      w[X]=1          w[Y]=1
159  *      MB              MB
160  *      r[Y]=y          r[X]=x
161  *
162  * Which guarantees that x==0 && y==0 is impossible; which translates back into
163  * the guarantee that we cannot both miss the futex variable change and the
164  * enqueue.
165  *
166  * Note that a new waiter is accounted for in (a) even when it is possible that
167  * the wait call can return error, in which case we backtrack from it in (b).
168  * Refer to the comment in queue_lock().
169  *
170  * Similarly, in order to account for waiters being requeued on another
171  * address we always increment the waiters for the destination bucket before
172  * acquiring the lock. It then decrements them again  after releasing it -
173  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
174  * will do the additional required waiter count housekeeping. This is done for
175  * double_lock_hb() and double_unlock_hb(), respectively.
176  */
177
178 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
179 #define futex_cmpxchg_enabled 1
180 #else
181 static int  __read_mostly futex_cmpxchg_enabled;
182 #endif
183
184 /*
185  * Futex flags used to encode options to functions and preserve them across
186  * restarts.
187  */
188 #ifdef CONFIG_MMU
189 # define FLAGS_SHARED           0x01
190 #else
191 /*
192  * NOMMU does not have per process address space. Let the compiler optimize
193  * code away.
194  */
195 # define FLAGS_SHARED           0x00
196 #endif
197 #define FLAGS_CLOCKRT           0x02
198 #define FLAGS_HAS_TIMEOUT       0x04
199
200 /*
201  * Priority Inheritance state:
202  */
203 struct futex_pi_state {
204         /*
205          * list of 'owned' pi_state instances - these have to be
206          * cleaned up in do_exit() if the task exits prematurely:
207          */
208         struct list_head list;
209
210         /*
211          * The PI object:
212          */
213         struct rt_mutex pi_mutex;
214
215         struct task_struct *owner;
216         refcount_t refcount;
217
218         union futex_key key;
219 } __randomize_layout;
220
221 /**
222  * struct futex_q - The hashed futex queue entry, one per waiting task
223  * @list:               priority-sorted list of tasks waiting on this futex
224  * @task:               the task waiting on the futex
225  * @lock_ptr:           the hash bucket lock
226  * @key:                the key the futex is hashed on
227  * @pi_state:           optional priority inheritance state
228  * @rt_waiter:          rt_waiter storage for use with requeue_pi
229  * @requeue_pi_key:     the requeue_pi target futex key
230  * @bitset:             bitset for the optional bitmasked wakeup
231  *
232  * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
233  * we can wake only the relevant ones (hashed queues may be shared).
234  *
235  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
236  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
237  * The order of wakeup is always to make the first condition true, then
238  * the second.
239  *
240  * PI futexes are typically woken before they are removed from the hash list via
241  * the rt_mutex code. See unqueue_me_pi().
242  */
243 struct futex_q {
244         struct plist_node list;
245
246         struct task_struct *task;
247         spinlock_t *lock_ptr;
248         union futex_key key;
249         struct futex_pi_state *pi_state;
250         struct rt_mutex_waiter *rt_waiter;
251         union futex_key *requeue_pi_key;
252         u32 bitset;
253 } __randomize_layout;
254
255 static const struct futex_q futex_q_init = {
256         /* list gets initialized in queue_me()*/
257         .key = FUTEX_KEY_INIT,
258         .bitset = FUTEX_BITSET_MATCH_ANY
259 };
260
261 /*
262  * Hash buckets are shared by all the futex_keys that hash to the same
263  * location.  Each key may have multiple futex_q structures, one for each task
264  * waiting on a futex.
265  */
266 struct futex_hash_bucket {
267         atomic_t waiters;
268         spinlock_t lock;
269         struct plist_head chain;
270 } ____cacheline_aligned_in_smp;
271
272 /*
273  * The base of the bucket array and its size are always used together
274  * (after initialization only in hash_futex()), so ensure that they
275  * reside in the same cacheline.
276  */
277 static struct {
278         struct futex_hash_bucket *queues;
279         unsigned long            hashsize;
280 } __futex_data __read_mostly __aligned(2*sizeof(long));
281 #define futex_queues   (__futex_data.queues)
282 #define futex_hashsize (__futex_data.hashsize)
283
284
285 /*
286  * Fault injections for futexes.
287  */
288 #ifdef CONFIG_FAIL_FUTEX
289
290 static struct {
291         struct fault_attr attr;
292
293         bool ignore_private;
294 } fail_futex = {
295         .attr = FAULT_ATTR_INITIALIZER,
296         .ignore_private = false,
297 };
298
299 static int __init setup_fail_futex(char *str)
300 {
301         return setup_fault_attr(&fail_futex.attr, str);
302 }
303 __setup("fail_futex=", setup_fail_futex);
304
305 static bool should_fail_futex(bool fshared)
306 {
307         if (fail_futex.ignore_private && !fshared)
308                 return false;
309
310         return should_fail(&fail_futex.attr, 1);
311 }
312
313 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314
315 static int __init fail_futex_debugfs(void)
316 {
317         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
318         struct dentry *dir;
319
320         dir = fault_create_debugfs_attr("fail_futex", NULL,
321                                         &fail_futex.attr);
322         if (IS_ERR(dir))
323                 return PTR_ERR(dir);
324
325         debugfs_create_bool("ignore-private", mode, dir,
326                             &fail_futex.ignore_private);
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 enum futex_access {
482         FUTEX_READ,
483         FUTEX_WRITE
484 };
485
486 /**
487  * get_futex_key() - Get parameters which are the keys for a futex
488  * @uaddr:      virtual address of the futex
489  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
490  * @key:        address where result is stored.
491  * @rw:         mapping needs to be read/write (values: FUTEX_READ,
492  *              FUTEX_WRITE)
493  *
494  * Return: a negative error code or 0
495  *
496  * The key words are stored in @key on success.
497  *
498  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
499  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
500  * We can usually work out the index without swapping in the page.
501  *
502  * lock_page() might sleep, the caller should not hold a spinlock.
503  */
504 static int
505 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
506 {
507         unsigned long address = (unsigned long)uaddr;
508         struct mm_struct *mm = current->mm;
509         struct page *page, *tail;
510         struct address_space *mapping;
511         int err, ro = 0;
512
513         /*
514          * The futex address must be "naturally" aligned.
515          */
516         key->both.offset = address % PAGE_SIZE;
517         if (unlikely((address % sizeof(u32)) != 0))
518                 return -EINVAL;
519         address -= key->both.offset;
520
521         if (unlikely(!access_ok(uaddr, sizeof(u32))))
522                 return -EFAULT;
523
524         if (unlikely(should_fail_futex(fshared)))
525                 return -EFAULT;
526
527         /*
528          * PROCESS_PRIVATE futexes are fast.
529          * As the mm cannot disappear under us and the 'key' only needs
530          * virtual address, we dont even have to find the underlying vma.
531          * Note : We do have to check 'uaddr' is a valid user address,
532          *        but access_ok() should be faster than find_vma()
533          */
534         if (!fshared) {
535                 key->private.mm = mm;
536                 key->private.address = address;
537                 get_futex_key_refs(key);  /* implies smp_mb(); (B) */
538                 return 0;
539         }
540
541 again:
542         /* Ignore any VERIFY_READ mapping (futex common case) */
543         if (unlikely(should_fail_futex(fshared)))
544                 return -EFAULT;
545
546         err = get_user_pages_fast(address, 1, 1, &page);
547         /*
548          * If write access is not required (eg. FUTEX_WAIT), try
549          * and get read-only access.
550          */
551         if (err == -EFAULT && rw == FUTEX_READ) {
552                 err = get_user_pages_fast(address, 1, 0, &page);
553                 ro = 1;
554         }
555         if (err < 0)
556                 return err;
557         else
558                 err = 0;
559
560         /*
561          * The treatment of mapping from this point on is critical. The page
562          * lock protects many things but in this context the page lock
563          * stabilizes mapping, prevents inode freeing in the shared
564          * file-backed region case and guards against movement to swap cache.
565          *
566          * Strictly speaking the page lock is not needed in all cases being
567          * considered here and page lock forces unnecessarily serialization
568          * From this point on, mapping will be re-verified if necessary and
569          * page lock will be acquired only if it is unavoidable
570          *
571          * Mapping checks require the head page for any compound page so the
572          * head page and mapping is looked up now. For anonymous pages, it
573          * does not matter if the page splits in the future as the key is
574          * based on the address. For filesystem-backed pages, the tail is
575          * required as the index of the page determines the key. For
576          * base pages, there is no tail page and tail == page.
577          */
578         tail = page;
579         page = compound_head(page);
580         mapping = READ_ONCE(page->mapping);
581
582         /*
583          * If page->mapping is NULL, then it cannot be a PageAnon
584          * page; but it might be the ZERO_PAGE or in the gate area or
585          * in a special mapping (all cases which we are happy to fail);
586          * or it may have been a good file page when get_user_pages_fast
587          * found it, but truncated or holepunched or subjected to
588          * invalidate_complete_page2 before we got the page lock (also
589          * cases which we are happy to fail).  And we hold a reference,
590          * so refcount care in invalidate_complete_page's remove_mapping
591          * prevents drop_caches from setting mapping to NULL beneath us.
592          *
593          * The case we do have to guard against is when memory pressure made
594          * shmem_writepage move it from filecache to swapcache beneath us:
595          * an unlikely race, but we do need to retry for page->mapping.
596          */
597         if (unlikely(!mapping)) {
598                 int shmem_swizzled;
599
600                 /*
601                  * Page lock is required to identify which special case above
602                  * applies. If this is really a shmem page then the page lock
603                  * will prevent unexpected transitions.
604                  */
605                 lock_page(page);
606                 shmem_swizzled = PageSwapCache(page) || page->mapping;
607                 unlock_page(page);
608                 put_page(page);
609
610                 if (shmem_swizzled)
611                         goto again;
612
613                 return -EFAULT;
614         }
615
616         /*
617          * Private mappings are handled in a simple way.
618          *
619          * If the futex key is stored on an anonymous page, then the associated
620          * object is the mm which is implicitly pinned by the calling process.
621          *
622          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
623          * it's a read-only handle, it's expected that futexes attach to
624          * the object not the particular process.
625          */
626         if (PageAnon(page)) {
627                 /*
628                  * A RO anonymous page will never change and thus doesn't make
629                  * sense for futex operations.
630                  */
631                 if (unlikely(should_fail_futex(fshared)) || ro) {
632                         err = -EFAULT;
633                         goto out;
634                 }
635
636                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
637                 key->private.mm = mm;
638                 key->private.address = address;
639
640                 get_futex_key_refs(key); /* implies smp_mb(); (B) */
641
642         } else {
643                 struct inode *inode;
644
645                 /*
646                  * The associated futex object in this case is the inode and
647                  * the page->mapping must be traversed. Ordinarily this should
648                  * be stabilised under page lock but it's not strictly
649                  * necessary in this case as we just want to pin the inode, not
650                  * update the radix tree or anything like that.
651                  *
652                  * The RCU read lock is taken as the inode is finally freed
653                  * under RCU. If the mapping still matches expectations then the
654                  * mapping->host can be safely accessed as being a valid inode.
655                  */
656                 rcu_read_lock();
657
658                 if (READ_ONCE(page->mapping) != mapping) {
659                         rcu_read_unlock();
660                         put_page(page);
661
662                         goto again;
663                 }
664
665                 inode = READ_ONCE(mapping->host);
666                 if (!inode) {
667                         rcu_read_unlock();
668                         put_page(page);
669
670                         goto again;
671                 }
672
673                 /*
674                  * Take a reference unless it is about to be freed. Previously
675                  * this reference was taken by ihold under the page lock
676                  * pinning the inode in place so i_lock was unnecessary. The
677                  * only way for this check to fail is if the inode was
678                  * truncated in parallel which is almost certainly an
679                  * application bug. In such a case, just retry.
680                  *
681                  * We are not calling into get_futex_key_refs() in file-backed
682                  * cases, therefore a successful atomic_inc return below will
683                  * guarantee that get_futex_key() will still imply smp_mb(); (B).
684                  */
685                 if (!atomic_inc_not_zero(&inode->i_count)) {
686                         rcu_read_unlock();
687                         put_page(page);
688
689                         goto again;
690                 }
691
692                 /* Should be impossible but lets be paranoid for now */
693                 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
694                         err = -EFAULT;
695                         rcu_read_unlock();
696                         iput(inode);
697
698                         goto out;
699                 }
700
701                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
702                 key->shared.inode = inode;
703                 key->shared.pgoff = basepage_index(tail);
704                 rcu_read_unlock();
705         }
706
707 out:
708         put_page(page);
709         return err;
710 }
711
712 static inline void put_futex_key(union futex_key *key)
713 {
714         drop_futex_key_refs(key);
715 }
716
717 /**
718  * fault_in_user_writeable() - Fault in user address and verify RW access
719  * @uaddr:      pointer to faulting user space address
720  *
721  * Slow path to fixup the fault we just took in the atomic write
722  * access to @uaddr.
723  *
724  * We have no generic implementation of a non-destructive write to the
725  * user address. We know that we faulted in the atomic pagefault
726  * disabled section so we can as well avoid the #PF overhead by
727  * calling get_user_pages() right away.
728  */
729 static int fault_in_user_writeable(u32 __user *uaddr)
730 {
731         struct mm_struct *mm = current->mm;
732         int ret;
733
734         down_read(&mm->mmap_sem);
735         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
736                                FAULT_FLAG_WRITE, NULL);
737         up_read(&mm->mmap_sem);
738
739         return ret < 0 ? ret : 0;
740 }
741
742 /**
743  * futex_top_waiter() - Return the highest priority waiter on a futex
744  * @hb:         the hash bucket the futex_q's reside in
745  * @key:        the futex key (to distinguish it from other futex futex_q's)
746  *
747  * Must be called with the hb lock held.
748  */
749 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
750                                         union futex_key *key)
751 {
752         struct futex_q *this;
753
754         plist_for_each_entry(this, &hb->chain, list) {
755                 if (match_futex(&this->key, key))
756                         return this;
757         }
758         return NULL;
759 }
760
761 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
762                                       u32 uval, u32 newval)
763 {
764         int ret;
765
766         pagefault_disable();
767         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
768         pagefault_enable();
769
770         return ret;
771 }
772
773 static int get_futex_value_locked(u32 *dest, u32 __user *from)
774 {
775         int ret;
776
777         pagefault_disable();
778         ret = __get_user(*dest, from);
779         pagefault_enable();
780
781         return ret ? -EFAULT : 0;
782 }
783
784
785 /*
786  * PI code:
787  */
788 static int refill_pi_state_cache(void)
789 {
790         struct futex_pi_state *pi_state;
791
792         if (likely(current->pi_state_cache))
793                 return 0;
794
795         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
796
797         if (!pi_state)
798                 return -ENOMEM;
799
800         INIT_LIST_HEAD(&pi_state->list);
801         /* pi_mutex gets initialized later */
802         pi_state->owner = NULL;
803         refcount_set(&pi_state->refcount, 1);
804         pi_state->key = FUTEX_KEY_INIT;
805
806         current->pi_state_cache = pi_state;
807
808         return 0;
809 }
810
811 static struct futex_pi_state *alloc_pi_state(void)
812 {
813         struct futex_pi_state *pi_state = current->pi_state_cache;
814
815         WARN_ON(!pi_state);
816         current->pi_state_cache = NULL;
817
818         return pi_state;
819 }
820
821 static void get_pi_state(struct futex_pi_state *pi_state)
822 {
823         WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
824 }
825
826 /*
827  * Drops a reference to the pi_state object and frees or caches it
828  * when the last reference is gone.
829  */
830 static void put_pi_state(struct futex_pi_state *pi_state)
831 {
832         if (!pi_state)
833                 return;
834
835         if (!refcount_dec_and_test(&pi_state->refcount))
836                 return;
837
838         /*
839          * If pi_state->owner is NULL, the owner is most probably dying
840          * and has cleaned up the pi_state already
841          */
842         if (pi_state->owner) {
843                 struct task_struct *owner;
844
845                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
846                 owner = pi_state->owner;
847                 if (owner) {
848                         raw_spin_lock(&owner->pi_lock);
849                         list_del_init(&pi_state->list);
850                         raw_spin_unlock(&owner->pi_lock);
851                 }
852                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
853                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
854         }
855
856         if (current->pi_state_cache) {
857                 kfree(pi_state);
858         } else {
859                 /*
860                  * pi_state->list is already empty.
861                  * clear pi_state->owner.
862                  * refcount is at 0 - put it back to 1.
863                  */
864                 pi_state->owner = NULL;
865                 refcount_set(&pi_state->refcount, 1);
866                 current->pi_state_cache = pi_state;
867         }
868 }
869
870 #ifdef CONFIG_FUTEX_PI
871
872 /*
873  * This task is holding PI mutexes at exit time => bad.
874  * Kernel cleans up PI-state, but userspace is likely hosed.
875  * (Robust-futex cleanup is separate and might save the day for userspace.)
876  */
877 void exit_pi_state_list(struct task_struct *curr)
878 {
879         struct list_head *next, *head = &curr->pi_state_list;
880         struct futex_pi_state *pi_state;
881         struct futex_hash_bucket *hb;
882         union futex_key key = FUTEX_KEY_INIT;
883
884         if (!futex_cmpxchg_enabled)
885                 return;
886         /*
887          * We are a ZOMBIE and nobody can enqueue itself on
888          * pi_state_list anymore, but we have to be careful
889          * versus waiters unqueueing themselves:
890          */
891         raw_spin_lock_irq(&curr->pi_lock);
892         while (!list_empty(head)) {
893                 next = head->next;
894                 pi_state = list_entry(next, struct futex_pi_state, list);
895                 key = pi_state->key;
896                 hb = hash_futex(&key);
897
898                 /*
899                  * We can race against put_pi_state() removing itself from the
900                  * list (a waiter going away). put_pi_state() will first
901                  * decrement the reference count and then modify the list, so
902                  * its possible to see the list entry but fail this reference
903                  * acquire.
904                  *
905                  * In that case; drop the locks to let put_pi_state() make
906                  * progress and retry the loop.
907                  */
908                 if (!refcount_inc_not_zero(&pi_state->refcount)) {
909                         raw_spin_unlock_irq(&curr->pi_lock);
910                         cpu_relax();
911                         raw_spin_lock_irq(&curr->pi_lock);
912                         continue;
913                 }
914                 raw_spin_unlock_irq(&curr->pi_lock);
915
916                 spin_lock(&hb->lock);
917                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
918                 raw_spin_lock(&curr->pi_lock);
919                 /*
920                  * We dropped the pi-lock, so re-check whether this
921                  * task still owns the PI-state:
922                  */
923                 if (head->next != next) {
924                         /* retain curr->pi_lock for the loop invariant */
925                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
926                         spin_unlock(&hb->lock);
927                         put_pi_state(pi_state);
928                         continue;
929                 }
930
931                 WARN_ON(pi_state->owner != curr);
932                 WARN_ON(list_empty(&pi_state->list));
933                 list_del_init(&pi_state->list);
934                 pi_state->owner = NULL;
935
936                 raw_spin_unlock(&curr->pi_lock);
937                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
938                 spin_unlock(&hb->lock);
939
940                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
941                 put_pi_state(pi_state);
942
943                 raw_spin_lock_irq(&curr->pi_lock);
944         }
945         raw_spin_unlock_irq(&curr->pi_lock);
946 }
947
948 #endif
949
950 /*
951  * We need to check the following states:
952  *
953  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
954  *
955  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
956  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
957  *
958  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
959  *
960  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
961  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
962  *
963  * [6]  Found  | Found    | task      | 0         | 1      | Valid
964  *
965  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
966  *
967  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
968  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
969  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
970  *
971  * [1]  Indicates that the kernel can acquire the futex atomically. We
972  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
973  *
974  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
975  *      thread is found then it indicates that the owner TID has died.
976  *
977  * [3]  Invalid. The waiter is queued on a non PI futex
978  *
979  * [4]  Valid state after exit_robust_list(), which sets the user space
980  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
981  *
982  * [5]  The user space value got manipulated between exit_robust_list()
983  *      and exit_pi_state_list()
984  *
985  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
986  *      the pi_state but cannot access the user space value.
987  *
988  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
989  *
990  * [8]  Owner and user space value match
991  *
992  * [9]  There is no transient state which sets the user space TID to 0
993  *      except exit_robust_list(), but this is indicated by the
994  *      FUTEX_OWNER_DIED bit. See [4]
995  *
996  * [10] There is no transient state which leaves owner and user space
997  *      TID out of sync.
998  *
999  *
1000  * Serialization and lifetime rules:
1001  *
1002  * hb->lock:
1003  *
1004  *      hb -> futex_q, relation
1005  *      futex_q -> pi_state, relation
1006  *
1007  *      (cannot be raw because hb can contain arbitrary amount
1008  *       of futex_q's)
1009  *
1010  * pi_mutex->wait_lock:
1011  *
1012  *      {uval, pi_state}
1013  *
1014  *      (and pi_mutex 'obviously')
1015  *
1016  * p->pi_lock:
1017  *
1018  *      p->pi_state_list -> pi_state->list, relation
1019  *
1020  * pi_state->refcount:
1021  *
1022  *      pi_state lifetime
1023  *
1024  *
1025  * Lock order:
1026  *
1027  *   hb->lock
1028  *     pi_mutex->wait_lock
1029  *       p->pi_lock
1030  *
1031  */
1032
1033 /*
1034  * Validate that the existing waiter has a pi_state and sanity check
1035  * the pi_state against the user space value. If correct, attach to
1036  * it.
1037  */
1038 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1039                               struct futex_pi_state *pi_state,
1040                               struct futex_pi_state **ps)
1041 {
1042         pid_t pid = uval & FUTEX_TID_MASK;
1043         u32 uval2;
1044         int ret;
1045
1046         /*
1047          * Userspace might have messed up non-PI and PI futexes [3]
1048          */
1049         if (unlikely(!pi_state))
1050                 return -EINVAL;
1051
1052         /*
1053          * We get here with hb->lock held, and having found a
1054          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1055          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1056          * which in turn means that futex_lock_pi() still has a reference on
1057          * our pi_state.
1058          *
1059          * The waiter holding a reference on @pi_state also protects against
1060          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1061          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1062          * free pi_state before we can take a reference ourselves.
1063          */
1064         WARN_ON(!refcount_read(&pi_state->refcount));
1065
1066         /*
1067          * Now that we have a pi_state, we can acquire wait_lock
1068          * and do the state validation.
1069          */
1070         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1071
1072         /*
1073          * Since {uval, pi_state} is serialized by wait_lock, and our current
1074          * uval was read without holding it, it can have changed. Verify it
1075          * still is what we expect it to be, otherwise retry the entire
1076          * operation.
1077          */
1078         if (get_futex_value_locked(&uval2, uaddr))
1079                 goto out_efault;
1080
1081         if (uval != uval2)
1082                 goto out_eagain;
1083
1084         /*
1085          * Handle the owner died case:
1086          */
1087         if (uval & FUTEX_OWNER_DIED) {
1088                 /*
1089                  * exit_pi_state_list sets owner to NULL and wakes the
1090                  * topmost waiter. The task which acquires the
1091                  * pi_state->rt_mutex will fixup owner.
1092                  */
1093                 if (!pi_state->owner) {
1094                         /*
1095                          * No pi state owner, but the user space TID
1096                          * is not 0. Inconsistent state. [5]
1097                          */
1098                         if (pid)
1099                                 goto out_einval;
1100                         /*
1101                          * Take a ref on the state and return success. [4]
1102                          */
1103                         goto out_attach;
1104                 }
1105
1106                 /*
1107                  * If TID is 0, then either the dying owner has not
1108                  * yet executed exit_pi_state_list() or some waiter
1109                  * acquired the rtmutex in the pi state, but did not
1110                  * yet fixup the TID in user space.
1111                  *
1112                  * Take a ref on the state and return success. [6]
1113                  */
1114                 if (!pid)
1115                         goto out_attach;
1116         } else {
1117                 /*
1118                  * If the owner died bit is not set, then the pi_state
1119                  * must have an owner. [7]
1120                  */
1121                 if (!pi_state->owner)
1122                         goto out_einval;
1123         }
1124
1125         /*
1126          * Bail out if user space manipulated the futex value. If pi
1127          * state exists then the owner TID must be the same as the
1128          * user space TID. [9/10]
1129          */
1130         if (pid != task_pid_vnr(pi_state->owner))
1131                 goto out_einval;
1132
1133 out_attach:
1134         get_pi_state(pi_state);
1135         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1136         *ps = pi_state;
1137         return 0;
1138
1139 out_einval:
1140         ret = -EINVAL;
1141         goto out_error;
1142
1143 out_eagain:
1144         ret = -EAGAIN;
1145         goto out_error;
1146
1147 out_efault:
1148         ret = -EFAULT;
1149         goto out_error;
1150
1151 out_error:
1152         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1153         return ret;
1154 }
1155
1156 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1157                             struct task_struct *tsk)
1158 {
1159         u32 uval2;
1160
1161         /*
1162          * If PF_EXITPIDONE is not yet set, then try again.
1163          */
1164         if (tsk && !(tsk->flags & PF_EXITPIDONE))
1165                 return -EAGAIN;
1166
1167         /*
1168          * Reread the user space value to handle the following situation:
1169          *
1170          * CPU0                         CPU1
1171          *
1172          * sys_exit()                   sys_futex()
1173          *  do_exit()                    futex_lock_pi()
1174          *                                futex_lock_pi_atomic()
1175          *   exit_signals(tsk)              No waiters:
1176          *    tsk->flags |= PF_EXITING;     *uaddr == 0x00000PID
1177          *  mm_release(tsk)                 Set waiter bit
1178          *   exit_robust_list(tsk) {        *uaddr = 0x80000PID;
1179          *      Set owner died              attach_to_pi_owner() {
1180          *    *uaddr = 0xC0000000;           tsk = get_task(PID);
1181          *   }                               if (!tsk->flags & PF_EXITING) {
1182          *  ...                                attach();
1183          *  tsk->flags |= PF_EXITPIDONE;     } else {
1184          *                                     if (!(tsk->flags & PF_EXITPIDONE))
1185          *                                       return -EAGAIN;
1186          *                                     return -ESRCH; <--- FAIL
1187          *                                   }
1188          *
1189          * Returning ESRCH unconditionally is wrong here because the
1190          * user space value has been changed by the exiting task.
1191          *
1192          * The same logic applies to the case where the exiting task is
1193          * already gone.
1194          */
1195         if (get_futex_value_locked(&uval2, uaddr))
1196                 return -EFAULT;
1197
1198         /* If the user space value has changed, try again. */
1199         if (uval2 != uval)
1200                 return -EAGAIN;
1201
1202         /*
1203          * The exiting task did not have a robust list, the robust list was
1204          * corrupted or the user space value in *uaddr is simply bogus.
1205          * Give up and tell user space.
1206          */
1207         return -ESRCH;
1208 }
1209
1210 /*
1211  * Lookup the task for the TID provided from user space and attach to
1212  * it after doing proper sanity checks.
1213  */
1214 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1215                               struct futex_pi_state **ps)
1216 {
1217         pid_t pid = uval & FUTEX_TID_MASK;
1218         struct futex_pi_state *pi_state;
1219         struct task_struct *p;
1220
1221         /*
1222          * We are the first waiter - try to look up the real owner and attach
1223          * the new pi_state to it, but bail out when TID = 0 [1]
1224          *
1225          * The !pid check is paranoid. None of the call sites should end up
1226          * with pid == 0, but better safe than sorry. Let the caller retry
1227          */
1228         if (!pid)
1229                 return -EAGAIN;
1230         p = find_get_task_by_vpid(pid);
1231         if (!p)
1232                 return handle_exit_race(uaddr, uval, NULL);
1233
1234         if (unlikely(p->flags & PF_KTHREAD)) {
1235                 put_task_struct(p);
1236                 return -EPERM;
1237         }
1238
1239         /*
1240          * We need to look at the task state flags to figure out,
1241          * whether the task is exiting. To protect against the do_exit
1242          * change of the task flags, we do this protected by
1243          * p->pi_lock:
1244          */
1245         raw_spin_lock_irq(&p->pi_lock);
1246         if (unlikely(p->flags & PF_EXITING)) {
1247                 /*
1248                  * The task is on the way out. When PF_EXITPIDONE is
1249                  * set, we know that the task has finished the
1250                  * cleanup:
1251                  */
1252                 int ret = handle_exit_race(uaddr, uval, p);
1253
1254                 raw_spin_unlock_irq(&p->pi_lock);
1255                 put_task_struct(p);
1256                 return ret;
1257         }
1258
1259         /*
1260          * No existing pi state. First waiter. [2]
1261          *
1262          * This creates pi_state, we have hb->lock held, this means nothing can
1263          * observe this state, wait_lock is irrelevant.
1264          */
1265         pi_state = alloc_pi_state();
1266
1267         /*
1268          * Initialize the pi_mutex in locked state and make @p
1269          * the owner of it:
1270          */
1271         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1272
1273         /* Store the key for possible exit cleanups: */
1274         pi_state->key = *key;
1275
1276         WARN_ON(!list_empty(&pi_state->list));
1277         list_add(&pi_state->list, &p->pi_state_list);
1278         /*
1279          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1280          * because there is no concurrency as the object is not published yet.
1281          */
1282         pi_state->owner = p;
1283         raw_spin_unlock_irq(&p->pi_lock);
1284
1285         put_task_struct(p);
1286
1287         *ps = pi_state;
1288
1289         return 0;
1290 }
1291
1292 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1293                            struct futex_hash_bucket *hb,
1294                            union futex_key *key, struct futex_pi_state **ps)
1295 {
1296         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1297
1298         /*
1299          * If there is a waiter on that futex, validate it and
1300          * attach to the pi_state when the validation succeeds.
1301          */
1302         if (top_waiter)
1303                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1304
1305         /*
1306          * We are the first waiter - try to look up the owner based on
1307          * @uval and attach to it.
1308          */
1309         return attach_to_pi_owner(uaddr, uval, key, ps);
1310 }
1311
1312 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1313 {
1314         u32 uninitialized_var(curval);
1315
1316         if (unlikely(should_fail_futex(true)))
1317                 return -EFAULT;
1318
1319         if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1320                 return -EFAULT;
1321
1322         /* If user space value changed, let the caller retry */
1323         return curval != uval ? -EAGAIN : 0;
1324 }
1325
1326 /**
1327  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1328  * @uaddr:              the pi futex user address
1329  * @hb:                 the pi futex hash bucket
1330  * @key:                the futex key associated with uaddr and hb
1331  * @ps:                 the pi_state pointer where we store the result of the
1332  *                      lookup
1333  * @task:               the task to perform the atomic lock work for.  This will
1334  *                      be "current" except in the case of requeue pi.
1335  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1336  *
1337  * Return:
1338  *  -  0 - ready to wait;
1339  *  -  1 - acquired the lock;
1340  *  - <0 - error
1341  *
1342  * The hb->lock and futex_key refs shall be held by the caller.
1343  */
1344 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1345                                 union futex_key *key,
1346                                 struct futex_pi_state **ps,
1347                                 struct task_struct *task, int set_waiters)
1348 {
1349         u32 uval, newval, vpid = task_pid_vnr(task);
1350         struct futex_q *top_waiter;
1351         int ret;
1352
1353         /*
1354          * Read the user space value first so we can validate a few
1355          * things before proceeding further.
1356          */
1357         if (get_futex_value_locked(&uval, uaddr))
1358                 return -EFAULT;
1359
1360         if (unlikely(should_fail_futex(true)))
1361                 return -EFAULT;
1362
1363         /*
1364          * Detect deadlocks.
1365          */
1366         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1367                 return -EDEADLK;
1368
1369         if ((unlikely(should_fail_futex(true))))
1370                 return -EDEADLK;
1371
1372         /*
1373          * Lookup existing state first. If it exists, try to attach to
1374          * its pi_state.
1375          */
1376         top_waiter = futex_top_waiter(hb, key);
1377         if (top_waiter)
1378                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1379
1380         /*
1381          * No waiter and user TID is 0. We are here because the
1382          * waiters or the owner died bit is set or called from
1383          * requeue_cmp_pi or for whatever reason something took the
1384          * syscall.
1385          */
1386         if (!(uval & FUTEX_TID_MASK)) {
1387                 /*
1388                  * We take over the futex. No other waiters and the user space
1389                  * TID is 0. We preserve the owner died bit.
1390                  */
1391                 newval = uval & FUTEX_OWNER_DIED;
1392                 newval |= vpid;
1393
1394                 /* The futex requeue_pi code can enforce the waiters bit */
1395                 if (set_waiters)
1396                         newval |= FUTEX_WAITERS;
1397
1398                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1399                 /* If the take over worked, return 1 */
1400                 return ret < 0 ? ret : 1;
1401         }
1402
1403         /*
1404          * First waiter. Set the waiters bit before attaching ourself to
1405          * the owner. If owner tries to unlock, it will be forced into
1406          * the kernel and blocked on hb->lock.
1407          */
1408         newval = uval | FUTEX_WAITERS;
1409         ret = lock_pi_update_atomic(uaddr, uval, newval);
1410         if (ret)
1411                 return ret;
1412         /*
1413          * If the update of the user space value succeeded, we try to
1414          * attach to the owner. If that fails, no harm done, we only
1415          * set the FUTEX_WAITERS bit in the user space variable.
1416          */
1417         return attach_to_pi_owner(uaddr, newval, key, ps);
1418 }
1419
1420 /**
1421  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1422  * @q:  The futex_q to unqueue
1423  *
1424  * The q->lock_ptr must not be NULL and must be held by the caller.
1425  */
1426 static void __unqueue_futex(struct futex_q *q)
1427 {
1428         struct futex_hash_bucket *hb;
1429
1430         if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1431                 return;
1432         lockdep_assert_held(q->lock_ptr);
1433
1434         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1435         plist_del(&q->list, &hb->chain);
1436         hb_waiters_dec(hb);
1437 }
1438
1439 /*
1440  * The hash bucket lock must be held when this is called.
1441  * Afterwards, the futex_q must not be accessed. Callers
1442  * must ensure to later call wake_up_q() for the actual
1443  * wakeups to occur.
1444  */
1445 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1446 {
1447         struct task_struct *p = q->task;
1448
1449         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1450                 return;
1451
1452         get_task_struct(p);
1453         __unqueue_futex(q);
1454         /*
1455          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1456          * is written, without taking any locks. This is possible in the event
1457          * of a spurious wakeup, for example. A memory barrier is required here
1458          * to prevent the following store to lock_ptr from getting ahead of the
1459          * plist_del in __unqueue_futex().
1460          */
1461         smp_store_release(&q->lock_ptr, NULL);
1462
1463         /*
1464          * Queue the task for later wakeup for after we've released
1465          * the hb->lock. wake_q_add() grabs reference to p.
1466          */
1467         wake_q_add_safe(wake_q, p);
1468 }
1469
1470 /*
1471  * Caller must hold a reference on @pi_state.
1472  */
1473 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1474 {
1475         u32 uninitialized_var(curval), newval;
1476         struct task_struct *new_owner;
1477         bool postunlock = false;
1478         DEFINE_WAKE_Q(wake_q);
1479         int ret = 0;
1480
1481         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1482         if (WARN_ON_ONCE(!new_owner)) {
1483                 /*
1484                  * As per the comment in futex_unlock_pi() this should not happen.
1485                  *
1486                  * When this happens, give up our locks and try again, giving
1487                  * the futex_lock_pi() instance time to complete, either by
1488                  * waiting on the rtmutex or removing itself from the futex
1489                  * queue.
1490                  */
1491                 ret = -EAGAIN;
1492                 goto out_unlock;
1493         }
1494
1495         /*
1496          * We pass it to the next owner. The WAITERS bit is always kept
1497          * enabled while there is PI state around. We cleanup the owner
1498          * died bit, because we are the owner.
1499          */
1500         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1501
1502         if (unlikely(should_fail_futex(true)))
1503                 ret = -EFAULT;
1504
1505         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1506                 ret = -EFAULT;
1507
1508         } else if (curval != uval) {
1509                 /*
1510                  * If a unconditional UNLOCK_PI operation (user space did not
1511                  * try the TID->0 transition) raced with a waiter setting the
1512                  * FUTEX_WAITERS flag between get_user() and locking the hash
1513                  * bucket lock, retry the operation.
1514                  */
1515                 if ((FUTEX_TID_MASK & curval) == uval)
1516                         ret = -EAGAIN;
1517                 else
1518                         ret = -EINVAL;
1519         }
1520
1521         if (ret)
1522                 goto out_unlock;
1523
1524         /*
1525          * This is a point of no return; once we modify the uval there is no
1526          * going back and subsequent operations must not fail.
1527          */
1528
1529         raw_spin_lock(&pi_state->owner->pi_lock);
1530         WARN_ON(list_empty(&pi_state->list));
1531         list_del_init(&pi_state->list);
1532         raw_spin_unlock(&pi_state->owner->pi_lock);
1533
1534         raw_spin_lock(&new_owner->pi_lock);
1535         WARN_ON(!list_empty(&pi_state->list));
1536         list_add(&pi_state->list, &new_owner->pi_state_list);
1537         pi_state->owner = new_owner;
1538         raw_spin_unlock(&new_owner->pi_lock);
1539
1540         postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1541
1542 out_unlock:
1543         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1544
1545         if (postunlock)
1546                 rt_mutex_postunlock(&wake_q);
1547
1548         return ret;
1549 }
1550
1551 /*
1552  * Express the locking dependencies for lockdep:
1553  */
1554 static inline void
1555 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1556 {
1557         if (hb1 <= hb2) {
1558                 spin_lock(&hb1->lock);
1559                 if (hb1 < hb2)
1560                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1561         } else { /* hb1 > hb2 */
1562                 spin_lock(&hb2->lock);
1563                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1564         }
1565 }
1566
1567 static inline void
1568 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1569 {
1570         spin_unlock(&hb1->lock);
1571         if (hb1 != hb2)
1572                 spin_unlock(&hb2->lock);
1573 }
1574
1575 /*
1576  * Wake up waiters matching bitset queued on this futex (uaddr).
1577  */
1578 static int
1579 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1580 {
1581         struct futex_hash_bucket *hb;
1582         struct futex_q *this, *next;
1583         union futex_key key = FUTEX_KEY_INIT;
1584         int ret;
1585         DEFINE_WAKE_Q(wake_q);
1586
1587         if (!bitset)
1588                 return -EINVAL;
1589
1590         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1591         if (unlikely(ret != 0))
1592                 goto out;
1593
1594         hb = hash_futex(&key);
1595
1596         /* Make sure we really have tasks to wakeup */
1597         if (!hb_waiters_pending(hb))
1598                 goto out_put_key;
1599
1600         spin_lock(&hb->lock);
1601
1602         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1603                 if (match_futex (&this->key, &key)) {
1604                         if (this->pi_state || this->rt_waiter) {
1605                                 ret = -EINVAL;
1606                                 break;
1607                         }
1608
1609                         /* Check if one of the bits is set in both bitsets */
1610                         if (!(this->bitset & bitset))
1611                                 continue;
1612
1613                         mark_wake_futex(&wake_q, this);
1614                         if (++ret >= nr_wake)
1615                                 break;
1616                 }
1617         }
1618
1619         spin_unlock(&hb->lock);
1620         wake_up_q(&wake_q);
1621 out_put_key:
1622         put_futex_key(&key);
1623 out:
1624         return ret;
1625 }
1626
1627 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1628 {
1629         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1630         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1631         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1632         int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1633         int oldval, ret;
1634
1635         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1636                 if (oparg < 0 || oparg > 31) {
1637                         char comm[sizeof(current->comm)];
1638                         /*
1639                          * kill this print and return -EINVAL when userspace
1640                          * is sane again
1641                          */
1642                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1643                                         get_task_comm(comm, current), oparg);
1644                         oparg &= 31;
1645                 }
1646                 oparg = 1 << oparg;
1647         }
1648
1649         if (!access_ok(uaddr, sizeof(u32)))
1650                 return -EFAULT;
1651
1652         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1653         if (ret)
1654                 return ret;
1655
1656         switch (cmp) {
1657         case FUTEX_OP_CMP_EQ:
1658                 return oldval == cmparg;
1659         case FUTEX_OP_CMP_NE:
1660                 return oldval != cmparg;
1661         case FUTEX_OP_CMP_LT:
1662                 return oldval < cmparg;
1663         case FUTEX_OP_CMP_GE:
1664                 return oldval >= cmparg;
1665         case FUTEX_OP_CMP_LE:
1666                 return oldval <= cmparg;
1667         case FUTEX_OP_CMP_GT:
1668                 return oldval > cmparg;
1669         default:
1670                 return -ENOSYS;
1671         }
1672 }
1673
1674 /*
1675  * Wake up all waiters hashed on the physical page that is mapped
1676  * to this virtual address:
1677  */
1678 static int
1679 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1680               int nr_wake, int nr_wake2, int op)
1681 {
1682         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1683         struct futex_hash_bucket *hb1, *hb2;
1684         struct futex_q *this, *next;
1685         int ret, op_ret;
1686         DEFINE_WAKE_Q(wake_q);
1687
1688 retry:
1689         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1690         if (unlikely(ret != 0))
1691                 goto out;
1692         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1693         if (unlikely(ret != 0))
1694                 goto out_put_key1;
1695
1696         hb1 = hash_futex(&key1);
1697         hb2 = hash_futex(&key2);
1698
1699 retry_private:
1700         double_lock_hb(hb1, hb2);
1701         op_ret = futex_atomic_op_inuser(op, uaddr2);
1702         if (unlikely(op_ret < 0)) {
1703
1704                 double_unlock_hb(hb1, hb2);
1705
1706 #ifndef CONFIG_MMU
1707                 /*
1708                  * we don't get EFAULT from MMU faults if we don't have an MMU,
1709                  * but we might get them from range checking
1710                  */
1711                 ret = op_ret;
1712                 goto out_put_keys;
1713 #endif
1714
1715                 if (unlikely(op_ret != -EFAULT)) {
1716                         ret = op_ret;
1717                         goto out_put_keys;
1718                 }
1719
1720                 ret = fault_in_user_writeable(uaddr2);
1721                 if (ret)
1722                         goto out_put_keys;
1723
1724                 if (!(flags & FLAGS_SHARED))
1725                         goto retry_private;
1726
1727                 put_futex_key(&key2);
1728                 put_futex_key(&key1);
1729                 goto retry;
1730         }
1731
1732         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1733                 if (match_futex (&this->key, &key1)) {
1734                         if (this->pi_state || this->rt_waiter) {
1735                                 ret = -EINVAL;
1736                                 goto out_unlock;
1737                         }
1738                         mark_wake_futex(&wake_q, this);
1739                         if (++ret >= nr_wake)
1740                                 break;
1741                 }
1742         }
1743
1744         if (op_ret > 0) {
1745                 op_ret = 0;
1746                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1747                         if (match_futex (&this->key, &key2)) {
1748                                 if (this->pi_state || this->rt_waiter) {
1749                                         ret = -EINVAL;
1750                                         goto out_unlock;
1751                                 }
1752                                 mark_wake_futex(&wake_q, this);
1753                                 if (++op_ret >= nr_wake2)
1754                                         break;
1755                         }
1756                 }
1757                 ret += op_ret;
1758         }
1759
1760 out_unlock:
1761         double_unlock_hb(hb1, hb2);
1762         wake_up_q(&wake_q);
1763 out_put_keys:
1764         put_futex_key(&key2);
1765 out_put_key1:
1766         put_futex_key(&key1);
1767 out:
1768         return ret;
1769 }
1770
1771 /**
1772  * requeue_futex() - Requeue a futex_q from one hb to another
1773  * @q:          the futex_q to requeue
1774  * @hb1:        the source hash_bucket
1775  * @hb2:        the target hash_bucket
1776  * @key2:       the new key for the requeued futex_q
1777  */
1778 static inline
1779 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1780                    struct futex_hash_bucket *hb2, union futex_key *key2)
1781 {
1782
1783         /*
1784          * If key1 and key2 hash to the same bucket, no need to
1785          * requeue.
1786          */
1787         if (likely(&hb1->chain != &hb2->chain)) {
1788                 plist_del(&q->list, &hb1->chain);
1789                 hb_waiters_dec(hb1);
1790                 hb_waiters_inc(hb2);
1791                 plist_add(&q->list, &hb2->chain);
1792                 q->lock_ptr = &hb2->lock;
1793         }
1794         get_futex_key_refs(key2);
1795         q->key = *key2;
1796 }
1797
1798 /**
1799  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1800  * @q:          the futex_q
1801  * @key:        the key of the requeue target futex
1802  * @hb:         the hash_bucket of the requeue target futex
1803  *
1804  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1805  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1806  * to the requeue target futex so the waiter can detect the wakeup on the right
1807  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1808  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1809  * to protect access to the pi_state to fixup the owner later.  Must be called
1810  * with both q->lock_ptr and hb->lock held.
1811  */
1812 static inline
1813 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1814                            struct futex_hash_bucket *hb)
1815 {
1816         get_futex_key_refs(key);
1817         q->key = *key;
1818
1819         __unqueue_futex(q);
1820
1821         WARN_ON(!q->rt_waiter);
1822         q->rt_waiter = NULL;
1823
1824         q->lock_ptr = &hb->lock;
1825
1826         wake_up_state(q->task, TASK_NORMAL);
1827 }
1828
1829 /**
1830  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1831  * @pifutex:            the user address of the to futex
1832  * @hb1:                the from futex hash bucket, must be locked by the caller
1833  * @hb2:                the to futex hash bucket, must be locked by the caller
1834  * @key1:               the from futex key
1835  * @key2:               the to futex key
1836  * @ps:                 address to store the pi_state pointer
1837  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1838  *
1839  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1840  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1841  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1842  * hb1 and hb2 must be held by the caller.
1843  *
1844  * Return:
1845  *  -  0 - failed to acquire the lock atomically;
1846  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1847  *  - <0 - error
1848  */
1849 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1850                                  struct futex_hash_bucket *hb1,
1851                                  struct futex_hash_bucket *hb2,
1852                                  union futex_key *key1, union futex_key *key2,
1853                                  struct futex_pi_state **ps, int set_waiters)
1854 {
1855         struct futex_q *top_waiter = NULL;
1856         u32 curval;
1857         int ret, vpid;
1858
1859         if (get_futex_value_locked(&curval, pifutex))
1860                 return -EFAULT;
1861
1862         if (unlikely(should_fail_futex(true)))
1863                 return -EFAULT;
1864
1865         /*
1866          * Find the top_waiter and determine if there are additional waiters.
1867          * If the caller intends to requeue more than 1 waiter to pifutex,
1868          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1869          * as we have means to handle the possible fault.  If not, don't set
1870          * the bit unecessarily as it will force the subsequent unlock to enter
1871          * the kernel.
1872          */
1873         top_waiter = futex_top_waiter(hb1, key1);
1874
1875         /* There are no waiters, nothing for us to do. */
1876         if (!top_waiter)
1877                 return 0;
1878
1879         /* Ensure we requeue to the expected futex. */
1880         if (!match_futex(top_waiter->requeue_pi_key, key2))
1881                 return -EINVAL;
1882
1883         /*
1884          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1885          * the contended case or if set_waiters is 1.  The pi_state is returned
1886          * in ps in contended cases.
1887          */
1888         vpid = task_pid_vnr(top_waiter->task);
1889         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1890                                    set_waiters);
1891         if (ret == 1) {
1892                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1893                 return vpid;
1894         }
1895         return ret;
1896 }
1897
1898 /**
1899  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1900  * @uaddr1:     source futex user address
1901  * @flags:      futex flags (FLAGS_SHARED, etc.)
1902  * @uaddr2:     target futex user address
1903  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1904  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1905  * @cmpval:     @uaddr1 expected value (or %NULL)
1906  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1907  *              pi futex (pi to pi requeue is not supported)
1908  *
1909  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1910  * uaddr2 atomically on behalf of the top waiter.
1911  *
1912  * Return:
1913  *  - >=0 - on success, the number of tasks requeued or woken;
1914  *  -  <0 - on error
1915  */
1916 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1917                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1918                          u32 *cmpval, int requeue_pi)
1919 {
1920         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1921         int drop_count = 0, task_count = 0, ret;
1922         struct futex_pi_state *pi_state = NULL;
1923         struct futex_hash_bucket *hb1, *hb2;
1924         struct futex_q *this, *next;
1925         DEFINE_WAKE_Q(wake_q);
1926
1927         if (nr_wake < 0 || nr_requeue < 0)
1928                 return -EINVAL;
1929
1930         /*
1931          * When PI not supported: return -ENOSYS if requeue_pi is true,
1932          * consequently the compiler knows requeue_pi is always false past
1933          * this point which will optimize away all the conditional code
1934          * further down.
1935          */
1936         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1937                 return -ENOSYS;
1938
1939         if (requeue_pi) {
1940                 /*
1941                  * Requeue PI only works on two distinct uaddrs. This
1942                  * check is only valid for private futexes. See below.
1943                  */
1944                 if (uaddr1 == uaddr2)
1945                         return -EINVAL;
1946
1947                 /*
1948                  * requeue_pi requires a pi_state, try to allocate it now
1949                  * without any locks in case it fails.
1950                  */
1951                 if (refill_pi_state_cache())
1952                         return -ENOMEM;
1953                 /*
1954                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1955                  * + nr_requeue, since it acquires the rt_mutex prior to
1956                  * returning to userspace, so as to not leave the rt_mutex with
1957                  * waiters and no owner.  However, second and third wake-ups
1958                  * cannot be predicted as they involve race conditions with the
1959                  * first wake and a fault while looking up the pi_state.  Both
1960                  * pthread_cond_signal() and pthread_cond_broadcast() should
1961                  * use nr_wake=1.
1962                  */
1963                 if (nr_wake != 1)
1964                         return -EINVAL;
1965         }
1966
1967 retry:
1968         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1969         if (unlikely(ret != 0))
1970                 goto out;
1971         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1972                             requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1973         if (unlikely(ret != 0))
1974                 goto out_put_key1;
1975
1976         /*
1977          * The check above which compares uaddrs is not sufficient for
1978          * shared futexes. We need to compare the keys:
1979          */
1980         if (requeue_pi && match_futex(&key1, &key2)) {
1981                 ret = -EINVAL;
1982                 goto out_put_keys;
1983         }
1984
1985         hb1 = hash_futex(&key1);
1986         hb2 = hash_futex(&key2);
1987
1988 retry_private:
1989         hb_waiters_inc(hb2);
1990         double_lock_hb(hb1, hb2);
1991
1992         if (likely(cmpval != NULL)) {
1993                 u32 curval;
1994
1995                 ret = get_futex_value_locked(&curval, uaddr1);
1996
1997                 if (unlikely(ret)) {
1998                         double_unlock_hb(hb1, hb2);
1999                         hb_waiters_dec(hb2);
2000
2001                         ret = get_user(curval, uaddr1);
2002                         if (ret)
2003                                 goto out_put_keys;
2004
2005                         if (!(flags & FLAGS_SHARED))
2006                                 goto retry_private;
2007
2008                         put_futex_key(&key2);
2009                         put_futex_key(&key1);
2010                         goto retry;
2011                 }
2012                 if (curval != *cmpval) {
2013                         ret = -EAGAIN;
2014                         goto out_unlock;
2015                 }
2016         }
2017
2018         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2019                 /*
2020                  * Attempt to acquire uaddr2 and wake the top waiter. If we
2021                  * intend to requeue waiters, force setting the FUTEX_WAITERS
2022                  * bit.  We force this here where we are able to easily handle
2023                  * faults rather in the requeue loop below.
2024                  */
2025                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2026                                                  &key2, &pi_state, nr_requeue);
2027
2028                 /*
2029                  * At this point the top_waiter has either taken uaddr2 or is
2030                  * waiting on it.  If the former, then the pi_state will not
2031                  * exist yet, look it up one more time to ensure we have a
2032                  * reference to it. If the lock was taken, ret contains the
2033                  * vpid of the top waiter task.
2034                  * If the lock was not taken, we have pi_state and an initial
2035                  * refcount on it. In case of an error we have nothing.
2036                  */
2037                 if (ret > 0) {
2038                         WARN_ON(pi_state);
2039                         drop_count++;
2040                         task_count++;
2041                         /*
2042                          * If we acquired the lock, then the user space value
2043                          * of uaddr2 should be vpid. It cannot be changed by
2044                          * the top waiter as it is blocked on hb2 lock if it
2045                          * tries to do so. If something fiddled with it behind
2046                          * our back the pi state lookup might unearth it. So
2047                          * we rather use the known value than rereading and
2048                          * handing potential crap to lookup_pi_state.
2049                          *
2050                          * If that call succeeds then we have pi_state and an
2051                          * initial refcount on it.
2052                          */
2053                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2054                 }
2055
2056                 switch (ret) {
2057                 case 0:
2058                         /* We hold a reference on the pi state. */
2059                         break;
2060
2061                         /* If the above failed, then pi_state is NULL */
2062                 case -EFAULT:
2063                         double_unlock_hb(hb1, hb2);
2064                         hb_waiters_dec(hb2);
2065                         put_futex_key(&key2);
2066                         put_futex_key(&key1);
2067                         ret = fault_in_user_writeable(uaddr2);
2068                         if (!ret)
2069                                 goto retry;
2070                         goto out;
2071                 case -EAGAIN:
2072                         /*
2073                          * Two reasons for this:
2074                          * - Owner is exiting and we just wait for the
2075                          *   exit to complete.
2076                          * - The user space value changed.
2077                          */
2078                         double_unlock_hb(hb1, hb2);
2079                         hb_waiters_dec(hb2);
2080                         put_futex_key(&key2);
2081                         put_futex_key(&key1);
2082                         cond_resched();
2083                         goto retry;
2084                 default:
2085                         goto out_unlock;
2086                 }
2087         }
2088
2089         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2090                 if (task_count - nr_wake >= nr_requeue)
2091                         break;
2092
2093                 if (!match_futex(&this->key, &key1))
2094                         continue;
2095
2096                 /*
2097                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2098                  * be paired with each other and no other futex ops.
2099                  *
2100                  * We should never be requeueing a futex_q with a pi_state,
2101                  * which is awaiting a futex_unlock_pi().
2102                  */
2103                 if ((requeue_pi && !this->rt_waiter) ||
2104                     (!requeue_pi && this->rt_waiter) ||
2105                     this->pi_state) {
2106                         ret = -EINVAL;
2107                         break;
2108                 }
2109
2110                 /*
2111                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2112                  * lock, we already woke the top_waiter.  If not, it will be
2113                  * woken by futex_unlock_pi().
2114                  */
2115                 if (++task_count <= nr_wake && !requeue_pi) {
2116                         mark_wake_futex(&wake_q, this);
2117                         continue;
2118                 }
2119
2120                 /* Ensure we requeue to the expected futex for requeue_pi. */
2121                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2122                         ret = -EINVAL;
2123                         break;
2124                 }
2125
2126                 /*
2127                  * Requeue nr_requeue waiters and possibly one more in the case
2128                  * of requeue_pi if we couldn't acquire the lock atomically.
2129                  */
2130                 if (requeue_pi) {
2131                         /*
2132                          * Prepare the waiter to take the rt_mutex. Take a
2133                          * refcount on the pi_state and store the pointer in
2134                          * the futex_q object of the waiter.
2135                          */
2136                         get_pi_state(pi_state);
2137                         this->pi_state = pi_state;
2138                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2139                                                         this->rt_waiter,
2140                                                         this->task);
2141                         if (ret == 1) {
2142                                 /*
2143                                  * We got the lock. We do neither drop the
2144                                  * refcount on pi_state nor clear
2145                                  * this->pi_state because the waiter needs the
2146                                  * pi_state for cleaning up the user space
2147                                  * value. It will drop the refcount after
2148                                  * doing so.
2149                                  */
2150                                 requeue_pi_wake_futex(this, &key2, hb2);
2151                                 drop_count++;
2152                                 continue;
2153                         } else if (ret) {
2154                                 /*
2155                                  * rt_mutex_start_proxy_lock() detected a
2156                                  * potential deadlock when we tried to queue
2157                                  * that waiter. Drop the pi_state reference
2158                                  * which we took above and remove the pointer
2159                                  * to the state from the waiters futex_q
2160                                  * object.
2161                                  */
2162                                 this->pi_state = NULL;
2163                                 put_pi_state(pi_state);
2164                                 /*
2165                                  * We stop queueing more waiters and let user
2166                                  * space deal with the mess.
2167                                  */
2168                                 break;
2169                         }
2170                 }
2171                 requeue_futex(this, hb1, hb2, &key2);
2172                 drop_count++;
2173         }
2174
2175         /*
2176          * We took an extra initial reference to the pi_state either
2177          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2178          * need to drop it here again.
2179          */
2180         put_pi_state(pi_state);
2181
2182 out_unlock:
2183         double_unlock_hb(hb1, hb2);
2184         wake_up_q(&wake_q);
2185         hb_waiters_dec(hb2);
2186
2187         /*
2188          * drop_futex_key_refs() must be called outside the spinlocks. During
2189          * the requeue we moved futex_q's from the hash bucket at key1 to the
2190          * one at key2 and updated their key pointer.  We no longer need to
2191          * hold the references to key1.
2192          */
2193         while (--drop_count >= 0)
2194                 drop_futex_key_refs(&key1);
2195
2196 out_put_keys:
2197         put_futex_key(&key2);
2198 out_put_key1:
2199         put_futex_key(&key1);
2200 out:
2201         return ret ? ret : task_count;
2202 }
2203
2204 /* The key must be already stored in q->key. */
2205 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2206         __acquires(&hb->lock)
2207 {
2208         struct futex_hash_bucket *hb;
2209
2210         hb = hash_futex(&q->key);
2211
2212         /*
2213          * Increment the counter before taking the lock so that
2214          * a potential waker won't miss a to-be-slept task that is
2215          * waiting for the spinlock. This is safe as all queue_lock()
2216          * users end up calling queue_me(). Similarly, for housekeeping,
2217          * decrement the counter at queue_unlock() when some error has
2218          * occurred and we don't end up adding the task to the list.
2219          */
2220         hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2221
2222         q->lock_ptr = &hb->lock;
2223
2224         spin_lock(&hb->lock);
2225         return hb;
2226 }
2227
2228 static inline void
2229 queue_unlock(struct futex_hash_bucket *hb)
2230         __releases(&hb->lock)
2231 {
2232         spin_unlock(&hb->lock);
2233         hb_waiters_dec(hb);
2234 }
2235
2236 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2237 {
2238         int prio;
2239
2240         /*
2241          * The priority used to register this element is
2242          * - either the real thread-priority for the real-time threads
2243          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2244          * - or MAX_RT_PRIO for non-RT threads.
2245          * Thus, all RT-threads are woken first in priority order, and
2246          * the others are woken last, in FIFO order.
2247          */
2248         prio = min(current->normal_prio, MAX_RT_PRIO);
2249
2250         plist_node_init(&q->list, prio);
2251         plist_add(&q->list, &hb->chain);
2252         q->task = current;
2253 }
2254
2255 /**
2256  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2257  * @q:  The futex_q to enqueue
2258  * @hb: The destination hash bucket
2259  *
2260  * The hb->lock must be held by the caller, and is released here. A call to
2261  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2262  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2263  * or nothing if the unqueue is done as part of the wake process and the unqueue
2264  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2265  * an example).
2266  */
2267 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2268         __releases(&hb->lock)
2269 {
2270         __queue_me(q, hb);
2271         spin_unlock(&hb->lock);
2272 }
2273
2274 /**
2275  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2276  * @q:  The futex_q to unqueue
2277  *
2278  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2279  * be paired with exactly one earlier call to queue_me().
2280  *
2281  * Return:
2282  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2283  *  - 0 - if the futex_q was already removed by the waking thread
2284  */
2285 static int unqueue_me(struct futex_q *q)
2286 {
2287         spinlock_t *lock_ptr;
2288         int ret = 0;
2289
2290         /* In the common case we don't take the spinlock, which is nice. */
2291 retry:
2292         /*
2293          * q->lock_ptr can change between this read and the following spin_lock.
2294          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2295          * optimizing lock_ptr out of the logic below.
2296          */
2297         lock_ptr = READ_ONCE(q->lock_ptr);
2298         if (lock_ptr != NULL) {
2299                 spin_lock(lock_ptr);
2300                 /*
2301                  * q->lock_ptr can change between reading it and
2302                  * spin_lock(), causing us to take the wrong lock.  This
2303                  * corrects the race condition.
2304                  *
2305                  * Reasoning goes like this: if we have the wrong lock,
2306                  * q->lock_ptr must have changed (maybe several times)
2307                  * between reading it and the spin_lock().  It can
2308                  * change again after the spin_lock() but only if it was
2309                  * already changed before the spin_lock().  It cannot,
2310                  * however, change back to the original value.  Therefore
2311                  * we can detect whether we acquired the correct lock.
2312                  */
2313                 if (unlikely(lock_ptr != q->lock_ptr)) {
2314                         spin_unlock(lock_ptr);
2315                         goto retry;
2316                 }
2317                 __unqueue_futex(q);
2318
2319                 BUG_ON(q->pi_state);
2320
2321                 spin_unlock(lock_ptr);
2322                 ret = 1;
2323         }
2324
2325         drop_futex_key_refs(&q->key);
2326         return ret;
2327 }
2328
2329 /*
2330  * PI futexes can not be requeued and must remove themself from the
2331  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2332  * and dropped here.
2333  */
2334 static void unqueue_me_pi(struct futex_q *q)
2335         __releases(q->lock_ptr)
2336 {
2337         __unqueue_futex(q);
2338
2339         BUG_ON(!q->pi_state);
2340         put_pi_state(q->pi_state);
2341         q->pi_state = NULL;
2342
2343         spin_unlock(q->lock_ptr);
2344 }
2345
2346 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2347                                 struct task_struct *argowner)
2348 {
2349         struct futex_pi_state *pi_state = q->pi_state;
2350         u32 uval, uninitialized_var(curval), newval;
2351         struct task_struct *oldowner, *newowner;
2352         u32 newtid;
2353         int ret;
2354
2355         lockdep_assert_held(q->lock_ptr);
2356
2357         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2358
2359         oldowner = pi_state->owner;
2360
2361         /*
2362          * We are here because either:
2363          *
2364          *  - we stole the lock and pi_state->owner needs updating to reflect
2365          *    that (@argowner == current),
2366          *
2367          * or:
2368          *
2369          *  - someone stole our lock and we need to fix things to point to the
2370          *    new owner (@argowner == NULL).
2371          *
2372          * Either way, we have to replace the TID in the user space variable.
2373          * This must be atomic as we have to preserve the owner died bit here.
2374          *
2375          * Note: We write the user space value _before_ changing the pi_state
2376          * because we can fault here. Imagine swapped out pages or a fork
2377          * that marked all the anonymous memory readonly for cow.
2378          *
2379          * Modifying pi_state _before_ the user space value would leave the
2380          * pi_state in an inconsistent state when we fault here, because we
2381          * need to drop the locks to handle the fault. This might be observed
2382          * in the PID check in lookup_pi_state.
2383          */
2384 retry:
2385         if (!argowner) {
2386                 if (oldowner != current) {
2387                         /*
2388                          * We raced against a concurrent self; things are
2389                          * already fixed up. Nothing to do.
2390                          */
2391                         ret = 0;
2392                         goto out_unlock;
2393                 }
2394
2395                 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2396                         /* We got the lock after all, nothing to fix. */
2397                         ret = 0;
2398                         goto out_unlock;
2399                 }
2400
2401                 /*
2402                  * Since we just failed the trylock; there must be an owner.
2403                  */
2404                 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2405                 BUG_ON(!newowner);
2406         } else {
2407                 WARN_ON_ONCE(argowner != current);
2408                 if (oldowner == current) {
2409                         /*
2410                          * We raced against a concurrent self; things are
2411                          * already fixed up. Nothing to do.
2412                          */
2413                         ret = 0;
2414                         goto out_unlock;
2415                 }
2416                 newowner = argowner;
2417         }
2418
2419         newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2420         /* Owner died? */
2421         if (!pi_state->owner)
2422                 newtid |= FUTEX_OWNER_DIED;
2423
2424         if (get_futex_value_locked(&uval, uaddr))
2425                 goto handle_fault;
2426
2427         for (;;) {
2428                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2429
2430                 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2431                         goto handle_fault;
2432                 if (curval == uval)
2433                         break;
2434                 uval = curval;
2435         }
2436
2437         /*
2438          * We fixed up user space. Now we need to fix the pi_state
2439          * itself.
2440          */
2441         if (pi_state->owner != NULL) {
2442                 raw_spin_lock(&pi_state->owner->pi_lock);
2443                 WARN_ON(list_empty(&pi_state->list));
2444                 list_del_init(&pi_state->list);
2445                 raw_spin_unlock(&pi_state->owner->pi_lock);
2446         }
2447
2448         pi_state->owner = newowner;
2449
2450         raw_spin_lock(&newowner->pi_lock);
2451         WARN_ON(!list_empty(&pi_state->list));
2452         list_add(&pi_state->list, &newowner->pi_state_list);
2453         raw_spin_unlock(&newowner->pi_lock);
2454         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2455
2456         return 0;
2457
2458         /*
2459          * To handle the page fault we need to drop the locks here. That gives
2460          * the other task (either the highest priority waiter itself or the
2461          * task which stole the rtmutex) the chance to try the fixup of the
2462          * pi_state. So once we are back from handling the fault we need to
2463          * check the pi_state after reacquiring the locks and before trying to
2464          * do another fixup. When the fixup has been done already we simply
2465          * return.
2466          *
2467          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2468          * drop hb->lock since the caller owns the hb -> futex_q relation.
2469          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2470          */
2471 handle_fault:
2472         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2473         spin_unlock(q->lock_ptr);
2474
2475         ret = fault_in_user_writeable(uaddr);
2476
2477         spin_lock(q->lock_ptr);
2478         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2479
2480         /*
2481          * Check if someone else fixed it for us:
2482          */
2483         if (pi_state->owner != oldowner) {
2484                 ret = 0;
2485                 goto out_unlock;
2486         }
2487
2488         if (ret)
2489                 goto out_unlock;
2490
2491         goto retry;
2492
2493 out_unlock:
2494         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2495         return ret;
2496 }
2497
2498 static long futex_wait_restart(struct restart_block *restart);
2499
2500 /**
2501  * fixup_owner() - Post lock pi_state and corner case management
2502  * @uaddr:      user address of the futex
2503  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2504  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2505  *
2506  * After attempting to lock an rt_mutex, this function is called to cleanup
2507  * the pi_state owner as well as handle race conditions that may allow us to
2508  * acquire the lock. Must be called with the hb lock held.
2509  *
2510  * Return:
2511  *  -  1 - success, lock taken;
2512  *  -  0 - success, lock not taken;
2513  *  - <0 - on error (-EFAULT)
2514  */
2515 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2516 {
2517         int ret = 0;
2518
2519         if (locked) {
2520                 /*
2521                  * Got the lock. We might not be the anticipated owner if we
2522                  * did a lock-steal - fix up the PI-state in that case:
2523                  *
2524                  * Speculative pi_state->owner read (we don't hold wait_lock);
2525                  * since we own the lock pi_state->owner == current is the
2526                  * stable state, anything else needs more attention.
2527                  */
2528                 if (q->pi_state->owner != current)
2529                         ret = fixup_pi_state_owner(uaddr, q, current);
2530                 goto out;
2531         }
2532
2533         /*
2534          * If we didn't get the lock; check if anybody stole it from us. In
2535          * that case, we need to fix up the uval to point to them instead of
2536          * us, otherwise bad things happen. [10]
2537          *
2538          * Another speculative read; pi_state->owner == current is unstable
2539          * but needs our attention.
2540          */
2541         if (q->pi_state->owner == current) {
2542                 ret = fixup_pi_state_owner(uaddr, q, NULL);
2543                 goto out;
2544         }
2545
2546         /*
2547          * Paranoia check. If we did not take the lock, then we should not be
2548          * the owner of the rt_mutex.
2549          */
2550         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2551                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2552                                 "pi-state %p\n", ret,
2553                                 q->pi_state->pi_mutex.owner,
2554                                 q->pi_state->owner);
2555         }
2556
2557 out:
2558         return ret ? ret : locked;
2559 }
2560
2561 /**
2562  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2563  * @hb:         the futex hash bucket, must be locked by the caller
2564  * @q:          the futex_q to queue up on
2565  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2566  */
2567 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2568                                 struct hrtimer_sleeper *timeout)
2569 {
2570         /*
2571          * The task state is guaranteed to be set before another task can
2572          * wake it. set_current_state() is implemented using smp_store_mb() and
2573          * queue_me() calls spin_unlock() upon completion, both serializing
2574          * access to the hash list and forcing another memory barrier.
2575          */
2576         set_current_state(TASK_INTERRUPTIBLE);
2577         queue_me(q, hb);
2578
2579         /* Arm the timer */
2580         if (timeout)
2581                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2582
2583         /*
2584          * If we have been removed from the hash list, then another task
2585          * has tried to wake us, and we can skip the call to schedule().
2586          */
2587         if (likely(!plist_node_empty(&q->list))) {
2588                 /*
2589                  * If the timer has already expired, current will already be
2590                  * flagged for rescheduling. Only call schedule if there
2591                  * is no timeout, or if it has yet to expire.
2592                  */
2593                 if (!timeout || timeout->task)
2594                         freezable_schedule();
2595         }
2596         __set_current_state(TASK_RUNNING);
2597 }
2598
2599 /**
2600  * futex_wait_setup() - Prepare to wait on a futex
2601  * @uaddr:      the futex userspace address
2602  * @val:        the expected value
2603  * @flags:      futex flags (FLAGS_SHARED, etc.)
2604  * @q:          the associated futex_q
2605  * @hb:         storage for hash_bucket pointer to be returned to caller
2606  *
2607  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2608  * compare it with the expected value.  Handle atomic faults internally.
2609  * Return with the hb lock held and a q.key reference on success, and unlocked
2610  * with no q.key reference on failure.
2611  *
2612  * Return:
2613  *  -  0 - uaddr contains val and hb has been locked;
2614  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2615  */
2616 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2617                            struct futex_q *q, struct futex_hash_bucket **hb)
2618 {
2619         u32 uval;
2620         int ret;
2621
2622         /*
2623          * Access the page AFTER the hash-bucket is locked.
2624          * Order is important:
2625          *
2626          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2627          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2628          *
2629          * The basic logical guarantee of a futex is that it blocks ONLY
2630          * if cond(var) is known to be true at the time of blocking, for
2631          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2632          * would open a race condition where we could block indefinitely with
2633          * cond(var) false, which would violate the guarantee.
2634          *
2635          * On the other hand, we insert q and release the hash-bucket only
2636          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2637          * absorb a wakeup if *uaddr does not match the desired values
2638          * while the syscall executes.
2639          */
2640 retry:
2641         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2642         if (unlikely(ret != 0))
2643                 return ret;
2644
2645 retry_private:
2646         *hb = queue_lock(q);
2647
2648         ret = get_futex_value_locked(&uval, uaddr);
2649
2650         if (ret) {
2651                 queue_unlock(*hb);
2652
2653                 ret = get_user(uval, uaddr);
2654                 if (ret)
2655                         goto out;
2656
2657                 if (!(flags & FLAGS_SHARED))
2658                         goto retry_private;
2659
2660                 put_futex_key(&q->key);
2661                 goto retry;
2662         }
2663
2664         if (uval != val) {
2665                 queue_unlock(*hb);
2666                 ret = -EWOULDBLOCK;
2667         }
2668
2669 out:
2670         if (ret)
2671                 put_futex_key(&q->key);
2672         return ret;
2673 }
2674
2675 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2676                       ktime_t *abs_time, u32 bitset)
2677 {
2678         struct hrtimer_sleeper timeout, *to = NULL;
2679         struct restart_block *restart;
2680         struct futex_hash_bucket *hb;
2681         struct futex_q q = futex_q_init;
2682         int ret;
2683
2684         if (!bitset)
2685                 return -EINVAL;
2686         q.bitset = bitset;
2687
2688         if (abs_time) {
2689                 to = &timeout;
2690
2691                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2692                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2693                                       HRTIMER_MODE_ABS);
2694                 hrtimer_init_sleeper(to, current);
2695                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2696                                              current->timer_slack_ns);
2697         }
2698
2699 retry:
2700         /*
2701          * Prepare to wait on uaddr. On success, holds hb lock and increments
2702          * q.key refs.
2703          */
2704         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2705         if (ret)
2706                 goto out;
2707
2708         /* queue_me and wait for wakeup, timeout, or a signal. */
2709         futex_wait_queue_me(hb, &q, to);
2710
2711         /* If we were woken (and unqueued), we succeeded, whatever. */
2712         ret = 0;
2713         /* unqueue_me() drops q.key ref */
2714         if (!unqueue_me(&q))
2715                 goto out;
2716         ret = -ETIMEDOUT;
2717         if (to && !to->task)
2718                 goto out;
2719
2720         /*
2721          * We expect signal_pending(current), but we might be the
2722          * victim of a spurious wakeup as well.
2723          */
2724         if (!signal_pending(current))
2725                 goto retry;
2726
2727         ret = -ERESTARTSYS;
2728         if (!abs_time)
2729                 goto out;
2730
2731         restart = &current->restart_block;
2732         restart->fn = futex_wait_restart;
2733         restart->futex.uaddr = uaddr;
2734         restart->futex.val = val;
2735         restart->futex.time = *abs_time;
2736         restart->futex.bitset = bitset;
2737         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2738
2739         ret = -ERESTART_RESTARTBLOCK;
2740
2741 out:
2742         if (to) {
2743                 hrtimer_cancel(&to->timer);
2744                 destroy_hrtimer_on_stack(&to->timer);
2745         }
2746         return ret;
2747 }
2748
2749
2750 static long futex_wait_restart(struct restart_block *restart)
2751 {
2752         u32 __user *uaddr = restart->futex.uaddr;
2753         ktime_t t, *tp = NULL;
2754
2755         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2756                 t = restart->futex.time;
2757                 tp = &t;
2758         }
2759         restart->fn = do_no_restart_syscall;
2760
2761         return (long)futex_wait(uaddr, restart->futex.flags,
2762                                 restart->futex.val, tp, restart->futex.bitset);
2763 }
2764
2765
2766 /*
2767  * Userspace tried a 0 -> TID atomic transition of the futex value
2768  * and failed. The kernel side here does the whole locking operation:
2769  * if there are waiters then it will block as a consequence of relying
2770  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2771  * a 0 value of the futex too.).
2772  *
2773  * Also serves as futex trylock_pi()'ing, and due semantics.
2774  */
2775 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2776                          ktime_t *time, int trylock)
2777 {
2778         struct hrtimer_sleeper timeout, *to = NULL;
2779         struct futex_pi_state *pi_state = NULL;
2780         struct rt_mutex_waiter rt_waiter;
2781         struct futex_hash_bucket *hb;
2782         struct futex_q q = futex_q_init;
2783         int res, ret;
2784
2785         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2786                 return -ENOSYS;
2787
2788         if (refill_pi_state_cache())
2789                 return -ENOMEM;
2790
2791         if (time) {
2792                 to = &timeout;
2793                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2794                                       HRTIMER_MODE_ABS);
2795                 hrtimer_init_sleeper(to, current);
2796                 hrtimer_set_expires(&to->timer, *time);
2797         }
2798
2799 retry:
2800         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2801         if (unlikely(ret != 0))
2802                 goto out;
2803
2804 retry_private:
2805         hb = queue_lock(&q);
2806
2807         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2808         if (unlikely(ret)) {
2809                 /*
2810                  * Atomic work succeeded and we got the lock,
2811                  * or failed. Either way, we do _not_ block.
2812                  */
2813                 switch (ret) {
2814                 case 1:
2815                         /* We got the lock. */
2816                         ret = 0;
2817                         goto out_unlock_put_key;
2818                 case -EFAULT:
2819                         goto uaddr_faulted;
2820                 case -EAGAIN:
2821                         /*
2822                          * Two reasons for this:
2823                          * - Task is exiting and we just wait for the
2824                          *   exit to complete.
2825                          * - The user space value changed.
2826                          */
2827                         queue_unlock(hb);
2828                         put_futex_key(&q.key);
2829                         cond_resched();
2830                         goto retry;
2831                 default:
2832                         goto out_unlock_put_key;
2833                 }
2834         }
2835
2836         WARN_ON(!q.pi_state);
2837
2838         /*
2839          * Only actually queue now that the atomic ops are done:
2840          */
2841         __queue_me(&q, hb);
2842
2843         if (trylock) {
2844                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2845                 /* Fixup the trylock return value: */
2846                 ret = ret ? 0 : -EWOULDBLOCK;
2847                 goto no_block;
2848         }
2849
2850         rt_mutex_init_waiter(&rt_waiter);
2851
2852         /*
2853          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2854          * hold it while doing rt_mutex_start_proxy(), because then it will
2855          * include hb->lock in the blocking chain, even through we'll not in
2856          * fact hold it while blocking. This will lead it to report -EDEADLK
2857          * and BUG when futex_unlock_pi() interleaves with this.
2858          *
2859          * Therefore acquire wait_lock while holding hb->lock, but drop the
2860          * latter before calling __rt_mutex_start_proxy_lock(). This
2861          * interleaves with futex_unlock_pi() -- which does a similar lock
2862          * handoff -- such that the latter can observe the futex_q::pi_state
2863          * before __rt_mutex_start_proxy_lock() is done.
2864          */
2865         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2866         spin_unlock(q.lock_ptr);
2867         /*
2868          * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2869          * such that futex_unlock_pi() is guaranteed to observe the waiter when
2870          * it sees the futex_q::pi_state.
2871          */
2872         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2873         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2874
2875         if (ret) {
2876                 if (ret == 1)
2877                         ret = 0;
2878                 goto cleanup;
2879         }
2880
2881         if (unlikely(to))
2882                 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2883
2884         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2885
2886 cleanup:
2887         spin_lock(q.lock_ptr);
2888         /*
2889          * If we failed to acquire the lock (deadlock/signal/timeout), we must
2890          * first acquire the hb->lock before removing the lock from the
2891          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2892          * lists consistent.
2893          *
2894          * In particular; it is important that futex_unlock_pi() can not
2895          * observe this inconsistency.
2896          */
2897         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2898                 ret = 0;
2899
2900 no_block:
2901         /*
2902          * Fixup the pi_state owner and possibly acquire the lock if we
2903          * haven't already.
2904          */
2905         res = fixup_owner(uaddr, &q, !ret);
2906         /*
2907          * If fixup_owner() returned an error, proprogate that.  If it acquired
2908          * the lock, clear our -ETIMEDOUT or -EINTR.
2909          */
2910         if (res)
2911                 ret = (res < 0) ? res : 0;
2912
2913         /*
2914          * If fixup_owner() faulted and was unable to handle the fault, unlock
2915          * it and return the fault to userspace.
2916          */
2917         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2918                 pi_state = q.pi_state;
2919                 get_pi_state(pi_state);
2920         }
2921
2922         /* Unqueue and drop the lock */
2923         unqueue_me_pi(&q);
2924
2925         if (pi_state) {
2926                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2927                 put_pi_state(pi_state);
2928         }
2929
2930         goto out_put_key;
2931
2932 out_unlock_put_key:
2933         queue_unlock(hb);
2934
2935 out_put_key:
2936         put_futex_key(&q.key);
2937 out:
2938         if (to) {
2939                 hrtimer_cancel(&to->timer);
2940                 destroy_hrtimer_on_stack(&to->timer);
2941         }
2942         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2943
2944 uaddr_faulted:
2945         queue_unlock(hb);
2946
2947         ret = fault_in_user_writeable(uaddr);
2948         if (ret)
2949                 goto out_put_key;
2950
2951         if (!(flags & FLAGS_SHARED))
2952                 goto retry_private;
2953
2954         put_futex_key(&q.key);
2955         goto retry;
2956 }
2957
2958 /*
2959  * Userspace attempted a TID -> 0 atomic transition, and failed.
2960  * This is the in-kernel slowpath: we look up the PI state (if any),
2961  * and do the rt-mutex unlock.
2962  */
2963 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2964 {
2965         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2966         union futex_key key = FUTEX_KEY_INIT;
2967         struct futex_hash_bucket *hb;
2968         struct futex_q *top_waiter;
2969         int ret;
2970
2971         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2972                 return -ENOSYS;
2973
2974 retry:
2975         if (get_user(uval, uaddr))
2976                 return -EFAULT;
2977         /*
2978          * We release only a lock we actually own:
2979          */
2980         if ((uval & FUTEX_TID_MASK) != vpid)
2981                 return -EPERM;
2982
2983         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2984         if (ret)
2985                 return ret;
2986
2987         hb = hash_futex(&key);
2988         spin_lock(&hb->lock);
2989
2990         /*
2991          * Check waiters first. We do not trust user space values at
2992          * all and we at least want to know if user space fiddled
2993          * with the futex value instead of blindly unlocking.
2994          */
2995         top_waiter = futex_top_waiter(hb, &key);
2996         if (top_waiter) {
2997                 struct futex_pi_state *pi_state = top_waiter->pi_state;
2998
2999                 ret = -EINVAL;
3000                 if (!pi_state)
3001                         goto out_unlock;
3002
3003                 /*
3004                  * If current does not own the pi_state then the futex is
3005                  * inconsistent and user space fiddled with the futex value.
3006                  */
3007                 if (pi_state->owner != current)
3008                         goto out_unlock;
3009
3010                 get_pi_state(pi_state);
3011                 /*
3012                  * By taking wait_lock while still holding hb->lock, we ensure
3013                  * there is no point where we hold neither; and therefore
3014                  * wake_futex_pi() must observe a state consistent with what we
3015                  * observed.
3016                  *
3017                  * In particular; this forces __rt_mutex_start_proxy() to
3018                  * complete such that we're guaranteed to observe the
3019                  * rt_waiter. Also see the WARN in wake_futex_pi().
3020                  */
3021                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3022                 spin_unlock(&hb->lock);
3023
3024                 /* drops pi_state->pi_mutex.wait_lock */
3025                 ret = wake_futex_pi(uaddr, uval, pi_state);
3026
3027                 put_pi_state(pi_state);
3028
3029                 /*
3030                  * Success, we're done! No tricky corner cases.
3031                  */
3032                 if (!ret)
3033                         goto out_putkey;
3034                 /*
3035                  * The atomic access to the futex value generated a
3036                  * pagefault, so retry the user-access and the wakeup:
3037                  */
3038                 if (ret == -EFAULT)
3039                         goto pi_faulted;
3040                 /*
3041                  * A unconditional UNLOCK_PI op raced against a waiter
3042                  * setting the FUTEX_WAITERS bit. Try again.
3043                  */
3044                 if (ret == -EAGAIN) {
3045                         put_futex_key(&key);
3046                         goto retry;
3047                 }
3048                 /*
3049                  * wake_futex_pi has detected invalid state. Tell user
3050                  * space.
3051                  */
3052                 goto out_putkey;
3053         }
3054
3055         /*
3056          * We have no kernel internal state, i.e. no waiters in the
3057          * kernel. Waiters which are about to queue themselves are stuck
3058          * on hb->lock. So we can safely ignore them. We do neither
3059          * preserve the WAITERS bit not the OWNER_DIED one. We are the
3060          * owner.
3061          */
3062         if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
3063                 spin_unlock(&hb->lock);
3064                 goto pi_faulted;
3065         }
3066
3067         /*
3068          * If uval has changed, let user space handle it.
3069          */
3070         ret = (curval == uval) ? 0 : -EAGAIN;
3071
3072 out_unlock:
3073         spin_unlock(&hb->lock);
3074 out_putkey:
3075         put_futex_key(&key);
3076         return ret;
3077
3078 pi_faulted:
3079         put_futex_key(&key);
3080
3081         ret = fault_in_user_writeable(uaddr);
3082         if (!ret)
3083                 goto retry;
3084
3085         return ret;
3086 }
3087
3088 /**
3089  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3090  * @hb:         the hash_bucket futex_q was original enqueued on
3091  * @q:          the futex_q woken while waiting to be requeued
3092  * @key2:       the futex_key of the requeue target futex
3093  * @timeout:    the timeout associated with the wait (NULL if none)
3094  *
3095  * Detect if the task was woken on the initial futex as opposed to the requeue
3096  * target futex.  If so, determine if it was a timeout or a signal that caused
3097  * the wakeup and return the appropriate error code to the caller.  Must be
3098  * called with the hb lock held.
3099  *
3100  * Return:
3101  *  -  0 = no early wakeup detected;
3102  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3103  */
3104 static inline
3105 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3106                                    struct futex_q *q, union futex_key *key2,
3107                                    struct hrtimer_sleeper *timeout)
3108 {
3109         int ret = 0;
3110
3111         /*
3112          * With the hb lock held, we avoid races while we process the wakeup.
3113          * We only need to hold hb (and not hb2) to ensure atomicity as the
3114          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3115          * It can't be requeued from uaddr2 to something else since we don't
3116          * support a PI aware source futex for requeue.
3117          */
3118         if (!match_futex(&q->key, key2)) {
3119                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3120                 /*
3121                  * We were woken prior to requeue by a timeout or a signal.
3122                  * Unqueue the futex_q and determine which it was.
3123                  */
3124                 plist_del(&q->list, &hb->chain);
3125                 hb_waiters_dec(hb);
3126
3127                 /* Handle spurious wakeups gracefully */
3128                 ret = -EWOULDBLOCK;
3129                 if (timeout && !timeout->task)
3130                         ret = -ETIMEDOUT;
3131                 else if (signal_pending(current))
3132                         ret = -ERESTARTNOINTR;
3133         }
3134         return ret;
3135 }
3136
3137 /**
3138  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3139  * @uaddr:      the futex we initially wait on (non-pi)
3140  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3141  *              the same type, no requeueing from private to shared, etc.
3142  * @val:        the expected value of uaddr
3143  * @abs_time:   absolute timeout
3144  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3145  * @uaddr2:     the pi futex we will take prior to returning to user-space
3146  *
3147  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3148  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3149  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3150  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3151  * without one, the pi logic would not know which task to boost/deboost, if
3152  * there was a need to.
3153  *
3154  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3155  * via the following--
3156  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3157  * 2) wakeup on uaddr2 after a requeue
3158  * 3) signal
3159  * 4) timeout
3160  *
3161  * If 3, cleanup and return -ERESTARTNOINTR.
3162  *
3163  * If 2, we may then block on trying to take the rt_mutex and return via:
3164  * 5) successful lock
3165  * 6) signal
3166  * 7) timeout
3167  * 8) other lock acquisition failure
3168  *
3169  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3170  *
3171  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3172  *
3173  * Return:
3174  *  -  0 - On success;
3175  *  - <0 - On error
3176  */
3177 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3178                                  u32 val, ktime_t *abs_time, u32 bitset,
3179                                  u32 __user *uaddr2)
3180 {
3181         struct hrtimer_sleeper timeout, *to = NULL;
3182         struct futex_pi_state *pi_state = NULL;
3183         struct rt_mutex_waiter rt_waiter;
3184         struct futex_hash_bucket *hb;
3185         union futex_key key2 = FUTEX_KEY_INIT;
3186         struct futex_q q = futex_q_init;
3187         int res, ret;
3188
3189         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3190                 return -ENOSYS;
3191
3192         if (uaddr == uaddr2)
3193                 return -EINVAL;
3194
3195         if (!bitset)
3196                 return -EINVAL;
3197
3198         if (abs_time) {
3199                 to = &timeout;
3200                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3201                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
3202                                       HRTIMER_MODE_ABS);
3203                 hrtimer_init_sleeper(to, current);
3204                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3205                                              current->timer_slack_ns);
3206         }
3207
3208         /*
3209          * The waiter is allocated on our stack, manipulated by the requeue
3210          * code while we sleep on uaddr.
3211          */
3212         rt_mutex_init_waiter(&rt_waiter);
3213
3214         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3215         if (unlikely(ret != 0))
3216                 goto out;
3217
3218         q.bitset = bitset;
3219         q.rt_waiter = &rt_waiter;
3220         q.requeue_pi_key = &key2;
3221
3222         /*
3223          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3224          * count.
3225          */
3226         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3227         if (ret)
3228                 goto out_key2;
3229
3230         /*
3231          * The check above which compares uaddrs is not sufficient for
3232          * shared futexes. We need to compare the keys:
3233          */
3234         if (match_futex(&q.key, &key2)) {
3235                 queue_unlock(hb);
3236                 ret = -EINVAL;
3237                 goto out_put_keys;
3238         }
3239
3240         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3241         futex_wait_queue_me(hb, &q, to);
3242
3243         spin_lock(&hb->lock);
3244         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3245         spin_unlock(&hb->lock);
3246         if (ret)
3247                 goto out_put_keys;
3248
3249         /*
3250          * In order for us to be here, we know our q.key == key2, and since
3251          * we took the hb->lock above, we also know that futex_requeue() has
3252          * completed and we no longer have to concern ourselves with a wakeup
3253          * race with the atomic proxy lock acquisition by the requeue code. The
3254          * futex_requeue dropped our key1 reference and incremented our key2
3255          * reference count.
3256          */
3257
3258         /* Check if the requeue code acquired the second futex for us. */
3259         if (!q.rt_waiter) {
3260                 /*
3261                  * Got the lock. We might not be the anticipated owner if we
3262                  * did a lock-steal - fix up the PI-state in that case.
3263                  */
3264                 if (q.pi_state && (q.pi_state->owner != current)) {
3265                         spin_lock(q.lock_ptr);
3266                         ret = fixup_pi_state_owner(uaddr2, &q, current);
3267                         if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3268                                 pi_state = q.pi_state;
3269                                 get_pi_state(pi_state);
3270                         }
3271                         /*
3272                          * Drop the reference to the pi state which
3273                          * the requeue_pi() code acquired for us.
3274                          */
3275                         put_pi_state(q.pi_state);
3276                         spin_unlock(q.lock_ptr);
3277                 }
3278         } else {
3279                 struct rt_mutex *pi_mutex;
3280
3281                 /*
3282                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3283                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3284                  * the pi_state.
3285                  */
3286                 WARN_ON(!q.pi_state);
3287                 pi_mutex = &q.pi_state->pi_mutex;
3288                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3289
3290                 spin_lock(q.lock_ptr);
3291                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3292                         ret = 0;
3293
3294                 debug_rt_mutex_free_waiter(&rt_waiter);
3295                 /*
3296                  * Fixup the pi_state owner and possibly acquire the lock if we
3297                  * haven't already.
3298                  */
3299                 res = fixup_owner(uaddr2, &q, !ret);
3300                 /*
3301                  * If fixup_owner() returned an error, proprogate that.  If it
3302                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3303                  */
3304                 if (res)
3305                         ret = (res < 0) ? res : 0;
3306
3307                 /*
3308                  * If fixup_pi_state_owner() faulted and was unable to handle
3309                  * the fault, unlock the rt_mutex and return the fault to
3310                  * userspace.
3311                  */
3312                 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3313                         pi_state = q.pi_state;
3314                         get_pi_state(pi_state);
3315                 }
3316
3317                 /* Unqueue and drop the lock. */
3318                 unqueue_me_pi(&q);
3319         }
3320
3321         if (pi_state) {
3322                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3323                 put_pi_state(pi_state);
3324         }
3325
3326         if (ret == -EINTR) {
3327                 /*
3328                  * We've already been requeued, but cannot restart by calling
3329                  * futex_lock_pi() directly. We could restart this syscall, but
3330                  * it would detect that the user space "val" changed and return
3331                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3332                  * -EWOULDBLOCK directly.
3333                  */
3334                 ret = -EWOULDBLOCK;
3335         }
3336
3337 out_put_keys:
3338         put_futex_key(&q.key);
3339 out_key2:
3340         put_futex_key(&key2);
3341
3342 out:
3343         if (to) {
3344                 hrtimer_cancel(&to->timer);
3345                 destroy_hrtimer_on_stack(&to->timer);
3346         }
3347         return ret;
3348 }
3349
3350 /*
3351  * Support for robust futexes: the kernel cleans up held futexes at
3352  * thread exit time.
3353  *
3354  * Implementation: user-space maintains a per-thread list of locks it
3355  * is holding. Upon do_exit(), the kernel carefully walks this list,
3356  * and marks all locks that are owned by this thread with the
3357  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3358  * always manipulated with the lock held, so the list is private and
3359  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3360  * field, to allow the kernel to clean up if the thread dies after
3361  * acquiring the lock, but just before it could have added itself to
3362  * the list. There can only be one such pending lock.
3363  */
3364
3365 /**
3366  * sys_set_robust_list() - Set the robust-futex list head of a task
3367  * @head:       pointer to the list-head
3368  * @len:        length of the list-head, as userspace expects
3369  */
3370 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3371                 size_t, len)
3372 {
3373         if (!futex_cmpxchg_enabled)
3374                 return -ENOSYS;
3375         /*
3376          * The kernel knows only one size for now:
3377          */
3378         if (unlikely(len != sizeof(*head)))
3379                 return -EINVAL;
3380
3381         current->robust_list = head;
3382
3383         return 0;
3384 }
3385
3386 /**
3387  * sys_get_robust_list() - Get the robust-futex list head of a task
3388  * @pid:        pid of the process [zero for current task]
3389  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
3390  * @len_ptr:    pointer to a length field, the kernel fills in the header size
3391  */
3392 SYSCALL_DEFINE3(get_robust_list, int, pid,
3393                 struct robust_list_head __user * __user *, head_ptr,
3394                 size_t __user *, len_ptr)
3395 {
3396         struct robust_list_head __user *head;
3397         unsigned long ret;
3398         struct task_struct *p;
3399
3400         if (!futex_cmpxchg_enabled)
3401                 return -ENOSYS;
3402
3403         rcu_read_lock();
3404
3405         ret = -ESRCH;
3406         if (!pid)
3407                 p = current;
3408         else {
3409                 p = find_task_by_vpid(pid);
3410                 if (!p)
3411                         goto err_unlock;
3412         }
3413
3414         ret = -EPERM;
3415         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3416                 goto err_unlock;
3417
3418         head = p->robust_list;
3419         rcu_read_unlock();
3420
3421         if (put_user(sizeof(*head), len_ptr))
3422                 return -EFAULT;
3423         return put_user(head, head_ptr);
3424
3425 err_unlock:
3426         rcu_read_unlock();
3427
3428         return ret;
3429 }
3430
3431 /*
3432  * Process a futex-list entry, check whether it's owned by the
3433  * dying task, and do notification if so:
3434  */
3435 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3436 {
3437         u32 uval, uninitialized_var(nval), mval;
3438
3439         /* Futex address must be 32bit aligned */
3440         if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3441                 return -1;
3442
3443 retry:
3444         if (get_user(uval, uaddr))
3445                 return -1;
3446
3447         if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3448                 /*
3449                  * Ok, this dying thread is truly holding a futex
3450                  * of interest. Set the OWNER_DIED bit atomically
3451                  * via cmpxchg, and if the value had FUTEX_WAITERS
3452                  * set, wake up a waiter (if any). (We have to do a
3453                  * futex_wake() even if OWNER_DIED is already set -
3454                  * to handle the rare but possible case of recursive
3455                  * thread-death.) The rest of the cleanup is done in
3456                  * userspace.
3457                  */
3458                 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3459                 /*
3460                  * We are not holding a lock here, but we want to have
3461                  * the pagefault_disable/enable() protection because
3462                  * we want to handle the fault gracefully. If the
3463                  * access fails we try to fault in the futex with R/W
3464                  * verification via get_user_pages. get_user() above
3465                  * does not guarantee R/W access. If that fails we
3466                  * give up and leave the futex locked.
3467                  */
3468                 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3469                         if (fault_in_user_writeable(uaddr))
3470                                 return -1;
3471                         goto retry;
3472                 }
3473                 if (nval != uval)
3474                         goto retry;
3475
3476                 /*
3477                  * Wake robust non-PI futexes here. The wakeup of
3478                  * PI futexes happens in exit_pi_state():
3479                  */
3480                 if (!pi && (uval & FUTEX_WAITERS))
3481                         futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3482         }
3483         return 0;
3484 }
3485
3486 /*
3487  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3488  */
3489 static inline int fetch_robust_entry(struct robust_list __user **entry,
3490                                      struct robust_list __user * __user *head,
3491                                      unsigned int *pi)
3492 {
3493         unsigned long uentry;
3494
3495         if (get_user(uentry, (unsigned long __user *)head))
3496                 return -EFAULT;
3497
3498         *entry = (void __user *)(uentry & ~1UL);
3499         *pi = uentry & 1;
3500
3501         return 0;
3502 }
3503
3504 /*
3505  * Walk curr->robust_list (very carefully, it's a userspace list!)
3506  * and mark any locks found there dead, and notify any waiters.
3507  *
3508  * We silently return on any sign of list-walking problem.
3509  */
3510 void exit_robust_list(struct task_struct *curr)
3511 {
3512         struct robust_list_head __user *head = curr->robust_list;
3513         struct robust_list __user *entry, *next_entry, *pending;
3514         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3515         unsigned int uninitialized_var(next_pi);
3516         unsigned long futex_offset;
3517         int rc;
3518
3519         if (!futex_cmpxchg_enabled)
3520                 return;
3521
3522         /*
3523          * Fetch the list head (which was registered earlier, via
3524          * sys_set_robust_list()):
3525          */
3526         if (fetch_robust_entry(&entry, &head->list.next, &pi))
3527                 return;
3528         /*
3529          * Fetch the relative futex offset:
3530          */
3531         if (get_user(futex_offset, &head->futex_offset))
3532                 return;
3533         /*
3534          * Fetch any possibly pending lock-add first, and handle it
3535          * if it exists:
3536          */
3537         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3538                 return;
3539
3540         next_entry = NULL;      /* avoid warning with gcc */
3541         while (entry != &head->list) {
3542                 /*
3543                  * Fetch the next entry in the list before calling
3544                  * handle_futex_death:
3545                  */
3546                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3547                 /*
3548                  * A pending lock might already be on the list, so
3549                  * don't process it twice:
3550                  */
3551                 if (entry != pending)
3552                         if (handle_futex_death((void __user *)entry + futex_offset,
3553                                                 curr, pi))
3554                                 return;
3555                 if (rc)
3556                         return;
3557                 entry = next_entry;
3558                 pi = next_pi;
3559                 /*
3560                  * Avoid excessively long or circular lists:
3561                  */
3562                 if (!--limit)
3563                         break;
3564
3565                 cond_resched();
3566         }
3567
3568         if (pending)
3569                 handle_futex_death((void __user *)pending + futex_offset,
3570                                    curr, pip);
3571 }
3572
3573 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3574                 u32 __user *uaddr2, u32 val2, u32 val3)
3575 {
3576         int cmd = op & FUTEX_CMD_MASK;
3577         unsigned int flags = 0;
3578
3579         if (!(op & FUTEX_PRIVATE_FLAG))
3580                 flags |= FLAGS_SHARED;
3581
3582         if (op & FUTEX_CLOCK_REALTIME) {
3583                 flags |= FLAGS_CLOCKRT;
3584                 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3585                     cmd != FUTEX_WAIT_REQUEUE_PI)
3586                         return -ENOSYS;
3587         }
3588
3589         switch (cmd) {
3590         case FUTEX_LOCK_PI:
3591         case FUTEX_UNLOCK_PI:
3592         case FUTEX_TRYLOCK_PI:
3593         case FUTEX_WAIT_REQUEUE_PI:
3594         case FUTEX_CMP_REQUEUE_PI:
3595                 if (!futex_cmpxchg_enabled)
3596                         return -ENOSYS;
3597         }
3598
3599         switch (cmd) {
3600         case FUTEX_WAIT:
3601                 val3 = FUTEX_BITSET_MATCH_ANY;
3602                 /* fall through */
3603         case FUTEX_WAIT_BITSET:
3604                 return futex_wait(uaddr, flags, val, timeout, val3);
3605         case FUTEX_WAKE:
3606                 val3 = FUTEX_BITSET_MATCH_ANY;
3607                 /* fall through */
3608         case FUTEX_WAKE_BITSET:
3609                 return futex_wake(uaddr, flags, val, val3);
3610         case FUTEX_REQUEUE:
3611                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3612         case FUTEX_CMP_REQUEUE:
3613                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3614         case FUTEX_WAKE_OP:
3615                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3616         case FUTEX_LOCK_PI:
3617                 return futex_lock_pi(uaddr, flags, timeout, 0);
3618         case FUTEX_UNLOCK_PI:
3619                 return futex_unlock_pi(uaddr, flags);
3620         case FUTEX_TRYLOCK_PI:
3621                 return futex_lock_pi(uaddr, flags, NULL, 1);
3622         case FUTEX_WAIT_REQUEUE_PI:
3623                 val3 = FUTEX_BITSET_MATCH_ANY;
3624                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3625                                              uaddr2);
3626         case FUTEX_CMP_REQUEUE_PI:
3627                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3628         }
3629         return -ENOSYS;
3630 }
3631
3632
3633 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3634                 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3635                 u32, val3)
3636 {
3637         struct timespec64 ts;
3638         ktime_t t, *tp = NULL;
3639         u32 val2 = 0;
3640         int cmd = op & FUTEX_CMD_MASK;
3641
3642         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3643                       cmd == FUTEX_WAIT_BITSET ||
3644                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3645                 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3646                         return -EFAULT;
3647                 if (get_timespec64(&ts, utime))
3648                         return -EFAULT;
3649                 if (!timespec64_valid(&ts))
3650                         return -EINVAL;
3651
3652                 t = timespec64_to_ktime(ts);
3653                 if (cmd == FUTEX_WAIT)
3654                         t = ktime_add_safe(ktime_get(), t);
3655                 tp = &t;
3656         }
3657         /*
3658          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3659          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3660          */
3661         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3662             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3663                 val2 = (u32) (unsigned long) utime;
3664
3665         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3666 }
3667
3668 #ifdef CONFIG_COMPAT
3669 /*
3670  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3671  */
3672 static inline int
3673 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3674                    compat_uptr_t __user *head, unsigned int *pi)
3675 {
3676         if (get_user(*uentry, head))
3677                 return -EFAULT;
3678
3679         *entry = compat_ptr((*uentry) & ~1);
3680         *pi = (unsigned int)(*uentry) & 1;
3681
3682         return 0;
3683 }
3684
3685 static void __user *futex_uaddr(struct robust_list __user *entry,
3686                                 compat_long_t futex_offset)
3687 {
3688         compat_uptr_t base = ptr_to_compat(entry);
3689         void __user *uaddr = compat_ptr(base + futex_offset);
3690
3691         return uaddr;
3692 }
3693
3694 /*
3695  * Walk curr->robust_list (very carefully, it's a userspace list!)
3696  * and mark any locks found there dead, and notify any waiters.
3697  *
3698  * We silently return on any sign of list-walking problem.
3699  */
3700 void compat_exit_robust_list(struct task_struct *curr)
3701 {
3702         struct compat_robust_list_head __user *head = curr->compat_robust_list;
3703         struct robust_list __user *entry, *next_entry, *pending;
3704         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3705         unsigned int uninitialized_var(next_pi);
3706         compat_uptr_t uentry, next_uentry, upending;
3707         compat_long_t futex_offset;
3708         int rc;
3709
3710         if (!futex_cmpxchg_enabled)
3711                 return;
3712
3713         /*
3714          * Fetch the list head (which was registered earlier, via
3715          * sys_set_robust_list()):
3716          */
3717         if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3718                 return;
3719         /*
3720          * Fetch the relative futex offset:
3721          */
3722         if (get_user(futex_offset, &head->futex_offset))
3723                 return;
3724         /*
3725          * Fetch any possibly pending lock-add first, and handle it
3726          * if it exists:
3727          */
3728         if (compat_fetch_robust_entry(&upending, &pending,
3729                                &head->list_op_pending, &pip))
3730                 return;
3731
3732         next_entry = NULL;      /* avoid warning with gcc */