Linux 5.2
[sfrench/cifs-2.6.git] / kernel / futex.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  Fast Userspace Mutexes (which I call "Futexes!").
4  *  (C) Rusty Russell, IBM 2002
5  *
6  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8  *
9  *  Removed page pinning, fix privately mapped COW pages and other cleanups
10  *  (C) Copyright 2003, 2004 Jamie Lokier
11  *
12  *  Robust futex support started by Ingo Molnar
13  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15  *
16  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
17  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19  *
20  *  PRIVATE futexes by Eric Dumazet
21  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22  *
23  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24  *  Copyright (C) IBM Corporation, 2009
25  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
26  *
27  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28  *  enough at me, Linus for the original (flawed) idea, Matthew
29  *  Kirkwood for proof-of-concept implementation.
30  *
31  *  "The futexes are also cursed."
32  *  "But they come in a choice of three flavours!"
33  */
34 #include <linux/compat.h>
35 #include <linux/slab.h>
36 #include <linux/poll.h>
37 #include <linux/fs.h>
38 #include <linux/file.h>
39 #include <linux/jhash.h>
40 #include <linux/init.h>
41 #include <linux/futex.h>
42 #include <linux/mount.h>
43 #include <linux/pagemap.h>
44 #include <linux/syscalls.h>
45 #include <linux/signal.h>
46 #include <linux/export.h>
47 #include <linux/magic.h>
48 #include <linux/pid.h>
49 #include <linux/nsproxy.h>
50 #include <linux/ptrace.h>
51 #include <linux/sched/rt.h>
52 #include <linux/sched/wake_q.h>
53 #include <linux/sched/mm.h>
54 #include <linux/hugetlb.h>
55 #include <linux/freezer.h>
56 #include <linux/memblock.h>
57 #include <linux/fault-inject.h>
58 #include <linux/refcount.h>
59
60 #include <asm/futex.h>
61
62 #include "locking/rtmutex_common.h"
63
64 /*
65  * READ this before attempting to hack on futexes!
66  *
67  * Basic futex operation and ordering guarantees
68  * =============================================
69  *
70  * The waiter reads the futex value in user space and calls
71  * futex_wait(). This function computes the hash bucket and acquires
72  * the hash bucket lock. After that it reads the futex user space value
73  * again and verifies that the data has not changed. If it has not changed
74  * it enqueues itself into the hash bucket, releases the hash bucket lock
75  * and schedules.
76  *
77  * The waker side modifies the user space value of the futex and calls
78  * futex_wake(). This function computes the hash bucket and acquires the
79  * hash bucket lock. Then it looks for waiters on that futex in the hash
80  * bucket and wakes them.
81  *
82  * In futex wake up scenarios where no tasks are blocked on a futex, taking
83  * the hb spinlock can be avoided and simply return. In order for this
84  * optimization to work, ordering guarantees must exist so that the waiter
85  * being added to the list is acknowledged when the list is concurrently being
86  * checked by the waker, avoiding scenarios like the following:
87  *
88  * CPU 0                               CPU 1
89  * val = *futex;
90  * sys_futex(WAIT, futex, val);
91  *   futex_wait(futex, val);
92  *   uval = *futex;
93  *                                     *futex = newval;
94  *                                     sys_futex(WAKE, futex);
95  *                                       futex_wake(futex);
96  *                                       if (queue_empty())
97  *                                         return;
98  *   if (uval == val)
99  *      lock(hash_bucket(futex));
100  *      queue();
101  *     unlock(hash_bucket(futex));
102  *     schedule();
103  *
104  * This would cause the waiter on CPU 0 to wait forever because it
105  * missed the transition of the user space value from val to newval
106  * and the waker did not find the waiter in the hash bucket queue.
107  *
108  * The correct serialization ensures that a waiter either observes
109  * the changed user space value before blocking or is woken by a
110  * concurrent waker:
111  *
112  * CPU 0                                 CPU 1
113  * val = *futex;
114  * sys_futex(WAIT, futex, val);
115  *   futex_wait(futex, val);
116  *
117  *   waiters++; (a)
118  *   smp_mb(); (A) <-- paired with -.
119  *                                  |
120  *   lock(hash_bucket(futex));      |
121  *                                  |
122  *   uval = *futex;                 |
123  *                                  |        *futex = newval;
124  *                                  |        sys_futex(WAKE, futex);
125  *                                  |          futex_wake(futex);
126  *                                  |
127  *                                  `--------> smp_mb(); (B)
128  *   if (uval == val)
129  *     queue();
130  *     unlock(hash_bucket(futex));
131  *     schedule();                         if (waiters)
132  *                                           lock(hash_bucket(futex));
133  *   else                                    wake_waiters(futex);
134  *     waiters--; (b)                        unlock(hash_bucket(futex));
135  *
136  * Where (A) orders the waiters increment and the futex value read through
137  * atomic operations (see hb_waiters_inc) and where (B) orders the write
138  * to futex and the waiters read -- this is done by the barriers for both
139  * shared and private futexes in get_futex_key_refs().
140  *
141  * This yields the following case (where X:=waiters, Y:=futex):
142  *
143  *      X = Y = 0
144  *
145  *      w[X]=1          w[Y]=1
146  *      MB              MB
147  *      r[Y]=y          r[X]=x
148  *
149  * Which guarantees that x==0 && y==0 is impossible; which translates back into
150  * the guarantee that we cannot both miss the futex variable change and the
151  * enqueue.
152  *
153  * Note that a new waiter is accounted for in (a) even when it is possible that
154  * the wait call can return error, in which case we backtrack from it in (b).
155  * Refer to the comment in queue_lock().
156  *
157  * Similarly, in order to account for waiters being requeued on another
158  * address we always increment the waiters for the destination bucket before
159  * acquiring the lock. It then decrements them again  after releasing it -
160  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
161  * will do the additional required waiter count housekeeping. This is done for
162  * double_lock_hb() and double_unlock_hb(), respectively.
163  */
164
165 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
166 #define futex_cmpxchg_enabled 1
167 #else
168 static int  __read_mostly futex_cmpxchg_enabled;
169 #endif
170
171 /*
172  * Futex flags used to encode options to functions and preserve them across
173  * restarts.
174  */
175 #ifdef CONFIG_MMU
176 # define FLAGS_SHARED           0x01
177 #else
178 /*
179  * NOMMU does not have per process address space. Let the compiler optimize
180  * code away.
181  */
182 # define FLAGS_SHARED           0x00
183 #endif
184 #define FLAGS_CLOCKRT           0x02
185 #define FLAGS_HAS_TIMEOUT       0x04
186
187 /*
188  * Priority Inheritance state:
189  */
190 struct futex_pi_state {
191         /*
192          * list of 'owned' pi_state instances - these have to be
193          * cleaned up in do_exit() if the task exits prematurely:
194          */
195         struct list_head list;
196
197         /*
198          * The PI object:
199          */
200         struct rt_mutex pi_mutex;
201
202         struct task_struct *owner;
203         refcount_t refcount;
204
205         union futex_key key;
206 } __randomize_layout;
207
208 /**
209  * struct futex_q - The hashed futex queue entry, one per waiting task
210  * @list:               priority-sorted list of tasks waiting on this futex
211  * @task:               the task waiting on the futex
212  * @lock_ptr:           the hash bucket lock
213  * @key:                the key the futex is hashed on
214  * @pi_state:           optional priority inheritance state
215  * @rt_waiter:          rt_waiter storage for use with requeue_pi
216  * @requeue_pi_key:     the requeue_pi target futex key
217  * @bitset:             bitset for the optional bitmasked wakeup
218  *
219  * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
220  * we can wake only the relevant ones (hashed queues may be shared).
221  *
222  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
223  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
224  * The order of wakeup is always to make the first condition true, then
225  * the second.
226  *
227  * PI futexes are typically woken before they are removed from the hash list via
228  * the rt_mutex code. See unqueue_me_pi().
229  */
230 struct futex_q {
231         struct plist_node list;
232
233         struct task_struct *task;
234         spinlock_t *lock_ptr;
235         union futex_key key;
236         struct futex_pi_state *pi_state;
237         struct rt_mutex_waiter *rt_waiter;
238         union futex_key *requeue_pi_key;
239         u32 bitset;
240 } __randomize_layout;
241
242 static const struct futex_q futex_q_init = {
243         /* list gets initialized in queue_me()*/
244         .key = FUTEX_KEY_INIT,
245         .bitset = FUTEX_BITSET_MATCH_ANY
246 };
247
248 /*
249  * Hash buckets are shared by all the futex_keys that hash to the same
250  * location.  Each key may have multiple futex_q structures, one for each task
251  * waiting on a futex.
252  */
253 struct futex_hash_bucket {
254         atomic_t waiters;
255         spinlock_t lock;
256         struct plist_head chain;
257 } ____cacheline_aligned_in_smp;
258
259 /*
260  * The base of the bucket array and its size are always used together
261  * (after initialization only in hash_futex()), so ensure that they
262  * reside in the same cacheline.
263  */
264 static struct {
265         struct futex_hash_bucket *queues;
266         unsigned long            hashsize;
267 } __futex_data __read_mostly __aligned(2*sizeof(long));
268 #define futex_queues   (__futex_data.queues)
269 #define futex_hashsize (__futex_data.hashsize)
270
271
272 /*
273  * Fault injections for futexes.
274  */
275 #ifdef CONFIG_FAIL_FUTEX
276
277 static struct {
278         struct fault_attr attr;
279
280         bool ignore_private;
281 } fail_futex = {
282         .attr = FAULT_ATTR_INITIALIZER,
283         .ignore_private = false,
284 };
285
286 static int __init setup_fail_futex(char *str)
287 {
288         return setup_fault_attr(&fail_futex.attr, str);
289 }
290 __setup("fail_futex=", setup_fail_futex);
291
292 static bool should_fail_futex(bool fshared)
293 {
294         if (fail_futex.ignore_private && !fshared)
295                 return false;
296
297         return should_fail(&fail_futex.attr, 1);
298 }
299
300 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
301
302 static int __init fail_futex_debugfs(void)
303 {
304         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
305         struct dentry *dir;
306
307         dir = fault_create_debugfs_attr("fail_futex", NULL,
308                                         &fail_futex.attr);
309         if (IS_ERR(dir))
310                 return PTR_ERR(dir);
311
312         debugfs_create_bool("ignore-private", mode, dir,
313                             &fail_futex.ignore_private);
314         return 0;
315 }
316
317 late_initcall(fail_futex_debugfs);
318
319 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
320
321 #else
322 static inline bool should_fail_futex(bool fshared)
323 {
324         return false;
325 }
326 #endif /* CONFIG_FAIL_FUTEX */
327
328 static inline void futex_get_mm(union futex_key *key)
329 {
330         mmgrab(key->private.mm);
331         /*
332          * Ensure futex_get_mm() implies a full barrier such that
333          * get_futex_key() implies a full barrier. This is relied upon
334          * as smp_mb(); (B), see the ordering comment above.
335          */
336         smp_mb__after_atomic();
337 }
338
339 /*
340  * Reflects a new waiter being added to the waitqueue.
341  */
342 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
343 {
344 #ifdef CONFIG_SMP
345         atomic_inc(&hb->waiters);
346         /*
347          * Full barrier (A), see the ordering comment above.
348          */
349         smp_mb__after_atomic();
350 #endif
351 }
352
353 /*
354  * Reflects a waiter being removed from the waitqueue by wakeup
355  * paths.
356  */
357 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
358 {
359 #ifdef CONFIG_SMP
360         atomic_dec(&hb->waiters);
361 #endif
362 }
363
364 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
365 {
366 #ifdef CONFIG_SMP
367         return atomic_read(&hb->waiters);
368 #else
369         return 1;
370 #endif
371 }
372
373 /**
374  * hash_futex - Return the hash bucket in the global hash
375  * @key:        Pointer to the futex key for which the hash is calculated
376  *
377  * We hash on the keys returned from get_futex_key (see below) and return the
378  * corresponding hash bucket in the global hash.
379  */
380 static struct futex_hash_bucket *hash_futex(union futex_key *key)
381 {
382         u32 hash = jhash2((u32*)&key->both.word,
383                           (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
384                           key->both.offset);
385         return &futex_queues[hash & (futex_hashsize - 1)];
386 }
387
388
389 /**
390  * match_futex - Check whether two futex keys are equal
391  * @key1:       Pointer to key1
392  * @key2:       Pointer to key2
393  *
394  * Return 1 if two futex_keys are equal, 0 otherwise.
395  */
396 static inline int match_futex(union futex_key *key1, union futex_key *key2)
397 {
398         return (key1 && key2
399                 && key1->both.word == key2->both.word
400                 && key1->both.ptr == key2->both.ptr
401                 && key1->both.offset == key2->both.offset);
402 }
403
404 /*
405  * Take a reference to the resource addressed by a key.
406  * Can be called while holding spinlocks.
407  *
408  */
409 static void get_futex_key_refs(union futex_key *key)
410 {
411         if (!key->both.ptr)
412                 return;
413
414         /*
415          * On MMU less systems futexes are always "private" as there is no per
416          * process address space. We need the smp wmb nevertheless - yes,
417          * arch/blackfin has MMU less SMP ...
418          */
419         if (!IS_ENABLED(CONFIG_MMU)) {
420                 smp_mb(); /* explicit smp_mb(); (B) */
421                 return;
422         }
423
424         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
425         case FUT_OFF_INODE:
426                 ihold(key->shared.inode); /* implies smp_mb(); (B) */
427                 break;
428         case FUT_OFF_MMSHARED:
429                 futex_get_mm(key); /* implies smp_mb(); (B) */
430                 break;
431         default:
432                 /*
433                  * Private futexes do not hold reference on an inode or
434                  * mm, therefore the only purpose of calling get_futex_key_refs
435                  * is because we need the barrier for the lockless waiter check.
436                  */
437                 smp_mb(); /* explicit smp_mb(); (B) */
438         }
439 }
440
441 /*
442  * Drop a reference to the resource addressed by a key.
443  * The hash bucket spinlock must not be held. This is
444  * a no-op for private futexes, see comment in the get
445  * counterpart.
446  */
447 static void drop_futex_key_refs(union futex_key *key)
448 {
449         if (!key->both.ptr) {
450                 /* If we're here then we tried to put a key we failed to get */
451                 WARN_ON_ONCE(1);
452                 return;
453         }
454
455         if (!IS_ENABLED(CONFIG_MMU))
456                 return;
457
458         switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
459         case FUT_OFF_INODE:
460                 iput(key->shared.inode);
461                 break;
462         case FUT_OFF_MMSHARED:
463                 mmdrop(key->private.mm);
464                 break;
465         }
466 }
467
468 enum futex_access {
469         FUTEX_READ,
470         FUTEX_WRITE
471 };
472
473 /**
474  * get_futex_key() - Get parameters which are the keys for a futex
475  * @uaddr:      virtual address of the futex
476  * @fshared:    0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
477  * @key:        address where result is stored.
478  * @rw:         mapping needs to be read/write (values: FUTEX_READ,
479  *              FUTEX_WRITE)
480  *
481  * Return: a negative error code or 0
482  *
483  * The key words are stored in @key on success.
484  *
485  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
486  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
487  * We can usually work out the index without swapping in the page.
488  *
489  * lock_page() might sleep, the caller should not hold a spinlock.
490  */
491 static int
492 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
493 {
494         unsigned long address = (unsigned long)uaddr;
495         struct mm_struct *mm = current->mm;
496         struct page *page, *tail;
497         struct address_space *mapping;
498         int err, ro = 0;
499
500         /*
501          * The futex address must be "naturally" aligned.
502          */
503         key->both.offset = address % PAGE_SIZE;
504         if (unlikely((address % sizeof(u32)) != 0))
505                 return -EINVAL;
506         address -= key->both.offset;
507
508         if (unlikely(!access_ok(uaddr, sizeof(u32))))
509                 return -EFAULT;
510
511         if (unlikely(should_fail_futex(fshared)))
512                 return -EFAULT;
513
514         /*
515          * PROCESS_PRIVATE futexes are fast.
516          * As the mm cannot disappear under us and the 'key' only needs
517          * virtual address, we dont even have to find the underlying vma.
518          * Note : We do have to check 'uaddr' is a valid user address,
519          *        but access_ok() should be faster than find_vma()
520          */
521         if (!fshared) {
522                 key->private.mm = mm;
523                 key->private.address = address;
524                 get_futex_key_refs(key);  /* implies smp_mb(); (B) */
525                 return 0;
526         }
527
528 again:
529         /* Ignore any VERIFY_READ mapping (futex common case) */
530         if (unlikely(should_fail_futex(fshared)))
531                 return -EFAULT;
532
533         err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
534         /*
535          * If write access is not required (eg. FUTEX_WAIT), try
536          * and get read-only access.
537          */
538         if (err == -EFAULT && rw == FUTEX_READ) {
539                 err = get_user_pages_fast(address, 1, 0, &page);
540                 ro = 1;
541         }
542         if (err < 0)
543                 return err;
544         else
545                 err = 0;
546
547         /*
548          * The treatment of mapping from this point on is critical. The page
549          * lock protects many things but in this context the page lock
550          * stabilizes mapping, prevents inode freeing in the shared
551          * file-backed region case and guards against movement to swap cache.
552          *
553          * Strictly speaking the page lock is not needed in all cases being
554          * considered here and page lock forces unnecessarily serialization
555          * From this point on, mapping will be re-verified if necessary and
556          * page lock will be acquired only if it is unavoidable
557          *
558          * Mapping checks require the head page for any compound page so the
559          * head page and mapping is looked up now. For anonymous pages, it
560          * does not matter if the page splits in the future as the key is
561          * based on the address. For filesystem-backed pages, the tail is
562          * required as the index of the page determines the key. For
563          * base pages, there is no tail page and tail == page.
564          */
565         tail = page;
566         page = compound_head(page);
567         mapping = READ_ONCE(page->mapping);
568
569         /*
570          * If page->mapping is NULL, then it cannot be a PageAnon
571          * page; but it might be the ZERO_PAGE or in the gate area or
572          * in a special mapping (all cases which we are happy to fail);
573          * or it may have been a good file page when get_user_pages_fast
574          * found it, but truncated or holepunched or subjected to
575          * invalidate_complete_page2 before we got the page lock (also
576          * cases which we are happy to fail).  And we hold a reference,
577          * so refcount care in invalidate_complete_page's remove_mapping
578          * prevents drop_caches from setting mapping to NULL beneath us.
579          *
580          * The case we do have to guard against is when memory pressure made
581          * shmem_writepage move it from filecache to swapcache beneath us:
582          * an unlikely race, but we do need to retry for page->mapping.
583          */
584         if (unlikely(!mapping)) {
585                 int shmem_swizzled;
586
587                 /*
588                  * Page lock is required to identify which special case above
589                  * applies. If this is really a shmem page then the page lock
590                  * will prevent unexpected transitions.
591                  */
592                 lock_page(page);
593                 shmem_swizzled = PageSwapCache(page) || page->mapping;
594                 unlock_page(page);
595                 put_page(page);
596
597                 if (shmem_swizzled)
598                         goto again;
599
600                 return -EFAULT;
601         }
602
603         /*
604          * Private mappings are handled in a simple way.
605          *
606          * If the futex key is stored on an anonymous page, then the associated
607          * object is the mm which is implicitly pinned by the calling process.
608          *
609          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
610          * it's a read-only handle, it's expected that futexes attach to
611          * the object not the particular process.
612          */
613         if (PageAnon(page)) {
614                 /*
615                  * A RO anonymous page will never change and thus doesn't make
616                  * sense for futex operations.
617                  */
618                 if (unlikely(should_fail_futex(fshared)) || ro) {
619                         err = -EFAULT;
620                         goto out;
621                 }
622
623                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
624                 key->private.mm = mm;
625                 key->private.address = address;
626
627                 get_futex_key_refs(key); /* implies smp_mb(); (B) */
628
629         } else {
630                 struct inode *inode;
631
632                 /*
633                  * The associated futex object in this case is the inode and
634                  * the page->mapping must be traversed. Ordinarily this should
635                  * be stabilised under page lock but it's not strictly
636                  * necessary in this case as we just want to pin the inode, not
637                  * update the radix tree or anything like that.
638                  *
639                  * The RCU read lock is taken as the inode is finally freed
640                  * under RCU. If the mapping still matches expectations then the
641                  * mapping->host can be safely accessed as being a valid inode.
642                  */
643                 rcu_read_lock();
644
645                 if (READ_ONCE(page->mapping) != mapping) {
646                         rcu_read_unlock();
647                         put_page(page);
648
649                         goto again;
650                 }
651
652                 inode = READ_ONCE(mapping->host);
653                 if (!inode) {
654                         rcu_read_unlock();
655                         put_page(page);
656
657                         goto again;
658                 }
659
660                 /*
661                  * Take a reference unless it is about to be freed. Previously
662                  * this reference was taken by ihold under the page lock
663                  * pinning the inode in place so i_lock was unnecessary. The
664                  * only way for this check to fail is if the inode was
665                  * truncated in parallel which is almost certainly an
666                  * application bug. In such a case, just retry.
667                  *
668                  * We are not calling into get_futex_key_refs() in file-backed
669                  * cases, therefore a successful atomic_inc return below will
670                  * guarantee that get_futex_key() will still imply smp_mb(); (B).
671                  */
672                 if (!atomic_inc_not_zero(&inode->i_count)) {
673                         rcu_read_unlock();
674                         put_page(page);
675
676                         goto again;
677                 }
678
679                 /* Should be impossible but lets be paranoid for now */
680                 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
681                         err = -EFAULT;
682                         rcu_read_unlock();
683                         iput(inode);
684
685                         goto out;
686                 }
687
688                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
689                 key->shared.inode = inode;
690                 key->shared.pgoff = basepage_index(tail);
691                 rcu_read_unlock();
692         }
693
694 out:
695         put_page(page);
696         return err;
697 }
698
699 static inline void put_futex_key(union futex_key *key)
700 {
701         drop_futex_key_refs(key);
702 }
703
704 /**
705  * fault_in_user_writeable() - Fault in user address and verify RW access
706  * @uaddr:      pointer to faulting user space address
707  *
708  * Slow path to fixup the fault we just took in the atomic write
709  * access to @uaddr.
710  *
711  * We have no generic implementation of a non-destructive write to the
712  * user address. We know that we faulted in the atomic pagefault
713  * disabled section so we can as well avoid the #PF overhead by
714  * calling get_user_pages() right away.
715  */
716 static int fault_in_user_writeable(u32 __user *uaddr)
717 {
718         struct mm_struct *mm = current->mm;
719         int ret;
720
721         down_read(&mm->mmap_sem);
722         ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
723                                FAULT_FLAG_WRITE, NULL);
724         up_read(&mm->mmap_sem);
725
726         return ret < 0 ? ret : 0;
727 }
728
729 /**
730  * futex_top_waiter() - Return the highest priority waiter on a futex
731  * @hb:         the hash bucket the futex_q's reside in
732  * @key:        the futex key (to distinguish it from other futex futex_q's)
733  *
734  * Must be called with the hb lock held.
735  */
736 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
737                                         union futex_key *key)
738 {
739         struct futex_q *this;
740
741         plist_for_each_entry(this, &hb->chain, list) {
742                 if (match_futex(&this->key, key))
743                         return this;
744         }
745         return NULL;
746 }
747
748 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
749                                       u32 uval, u32 newval)
750 {
751         int ret;
752
753         pagefault_disable();
754         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
755         pagefault_enable();
756
757         return ret;
758 }
759
760 static int get_futex_value_locked(u32 *dest, u32 __user *from)
761 {
762         int ret;
763
764         pagefault_disable();
765         ret = __get_user(*dest, from);
766         pagefault_enable();
767
768         return ret ? -EFAULT : 0;
769 }
770
771
772 /*
773  * PI code:
774  */
775 static int refill_pi_state_cache(void)
776 {
777         struct futex_pi_state *pi_state;
778
779         if (likely(current->pi_state_cache))
780                 return 0;
781
782         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
783
784         if (!pi_state)
785                 return -ENOMEM;
786
787         INIT_LIST_HEAD(&pi_state->list);
788         /* pi_mutex gets initialized later */
789         pi_state->owner = NULL;
790         refcount_set(&pi_state->refcount, 1);
791         pi_state->key = FUTEX_KEY_INIT;
792
793         current->pi_state_cache = pi_state;
794
795         return 0;
796 }
797
798 static struct futex_pi_state *alloc_pi_state(void)
799 {
800         struct futex_pi_state *pi_state = current->pi_state_cache;
801
802         WARN_ON(!pi_state);
803         current->pi_state_cache = NULL;
804
805         return pi_state;
806 }
807
808 static void get_pi_state(struct futex_pi_state *pi_state)
809 {
810         WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
811 }
812
813 /*
814  * Drops a reference to the pi_state object and frees or caches it
815  * when the last reference is gone.
816  */
817 static void put_pi_state(struct futex_pi_state *pi_state)
818 {
819         if (!pi_state)
820                 return;
821
822         if (!refcount_dec_and_test(&pi_state->refcount))
823                 return;
824
825         /*
826          * If pi_state->owner is NULL, the owner is most probably dying
827          * and has cleaned up the pi_state already
828          */
829         if (pi_state->owner) {
830                 struct task_struct *owner;
831
832                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
833                 owner = pi_state->owner;
834                 if (owner) {
835                         raw_spin_lock(&owner->pi_lock);
836                         list_del_init(&pi_state->list);
837                         raw_spin_unlock(&owner->pi_lock);
838                 }
839                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
840                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
841         }
842
843         if (current->pi_state_cache) {
844                 kfree(pi_state);
845         } else {
846                 /*
847                  * pi_state->list is already empty.
848                  * clear pi_state->owner.
849                  * refcount is at 0 - put it back to 1.
850                  */
851                 pi_state->owner = NULL;
852                 refcount_set(&pi_state->refcount, 1);
853                 current->pi_state_cache = pi_state;
854         }
855 }
856
857 #ifdef CONFIG_FUTEX_PI
858
859 /*
860  * This task is holding PI mutexes at exit time => bad.
861  * Kernel cleans up PI-state, but userspace is likely hosed.
862  * (Robust-futex cleanup is separate and might save the day for userspace.)
863  */
864 void exit_pi_state_list(struct task_struct *curr)
865 {
866         struct list_head *next, *head = &curr->pi_state_list;
867         struct futex_pi_state *pi_state;
868         struct futex_hash_bucket *hb;
869         union futex_key key = FUTEX_KEY_INIT;
870
871         if (!futex_cmpxchg_enabled)
872                 return;
873         /*
874          * We are a ZOMBIE and nobody can enqueue itself on
875          * pi_state_list anymore, but we have to be careful
876          * versus waiters unqueueing themselves:
877          */
878         raw_spin_lock_irq(&curr->pi_lock);
879         while (!list_empty(head)) {
880                 next = head->next;
881                 pi_state = list_entry(next, struct futex_pi_state, list);
882                 key = pi_state->key;
883                 hb = hash_futex(&key);
884
885                 /*
886                  * We can race against put_pi_state() removing itself from the
887                  * list (a waiter going away). put_pi_state() will first
888                  * decrement the reference count and then modify the list, so
889                  * its possible to see the list entry but fail this reference
890                  * acquire.
891                  *
892                  * In that case; drop the locks to let put_pi_state() make
893                  * progress and retry the loop.
894                  */
895                 if (!refcount_inc_not_zero(&pi_state->refcount)) {
896                         raw_spin_unlock_irq(&curr->pi_lock);
897                         cpu_relax();
898                         raw_spin_lock_irq(&curr->pi_lock);
899                         continue;
900                 }
901                 raw_spin_unlock_irq(&curr->pi_lock);
902
903                 spin_lock(&hb->lock);
904                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
905                 raw_spin_lock(&curr->pi_lock);
906                 /*
907                  * We dropped the pi-lock, so re-check whether this
908                  * task still owns the PI-state:
909                  */
910                 if (head->next != next) {
911                         /* retain curr->pi_lock for the loop invariant */
912                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
913                         spin_unlock(&hb->lock);
914                         put_pi_state(pi_state);
915                         continue;
916                 }
917
918                 WARN_ON(pi_state->owner != curr);
919                 WARN_ON(list_empty(&pi_state->list));
920                 list_del_init(&pi_state->list);
921                 pi_state->owner = NULL;
922
923                 raw_spin_unlock(&curr->pi_lock);
924                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
925                 spin_unlock(&hb->lock);
926
927                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
928                 put_pi_state(pi_state);
929
930                 raw_spin_lock_irq(&curr->pi_lock);
931         }
932         raw_spin_unlock_irq(&curr->pi_lock);
933 }
934
935 #endif
936
937 /*
938  * We need to check the following states:
939  *
940  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
941  *
942  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
943  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
944  *
945  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
946  *
947  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
948  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
949  *
950  * [6]  Found  | Found    | task      | 0         | 1      | Valid
951  *
952  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
953  *
954  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
955  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
956  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
957  *
958  * [1]  Indicates that the kernel can acquire the futex atomically. We
959  *      came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
960  *
961  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
962  *      thread is found then it indicates that the owner TID has died.
963  *
964  * [3]  Invalid. The waiter is queued on a non PI futex
965  *
966  * [4]  Valid state after exit_robust_list(), which sets the user space
967  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
968  *
969  * [5]  The user space value got manipulated between exit_robust_list()
970  *      and exit_pi_state_list()
971  *
972  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
973  *      the pi_state but cannot access the user space value.
974  *
975  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
976  *
977  * [8]  Owner and user space value match
978  *
979  * [9]  There is no transient state which sets the user space TID to 0
980  *      except exit_robust_list(), but this is indicated by the
981  *      FUTEX_OWNER_DIED bit. See [4]
982  *
983  * [10] There is no transient state which leaves owner and user space
984  *      TID out of sync.
985  *
986  *
987  * Serialization and lifetime rules:
988  *
989  * hb->lock:
990  *
991  *      hb -> futex_q, relation
992  *      futex_q -> pi_state, relation
993  *
994  *      (cannot be raw because hb can contain arbitrary amount
995  *       of futex_q's)
996  *
997  * pi_mutex->wait_lock:
998  *
999  *      {uval, pi_state}
1000  *
1001  *      (and pi_mutex 'obviously')
1002  *
1003  * p->pi_lock:
1004  *
1005  *      p->pi_state_list -> pi_state->list, relation
1006  *
1007  * pi_state->refcount:
1008  *
1009  *      pi_state lifetime
1010  *
1011  *
1012  * Lock order:
1013  *
1014  *   hb->lock
1015  *     pi_mutex->wait_lock
1016  *       p->pi_lock
1017  *
1018  */
1019
1020 /*
1021  * Validate that the existing waiter has a pi_state and sanity check
1022  * the pi_state against the user space value. If correct, attach to
1023  * it.
1024  */
1025 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1026                               struct futex_pi_state *pi_state,
1027                               struct futex_pi_state **ps)
1028 {
1029         pid_t pid = uval & FUTEX_TID_MASK;
1030         u32 uval2;
1031         int ret;
1032
1033         /*
1034          * Userspace might have messed up non-PI and PI futexes [3]
1035          */
1036         if (unlikely(!pi_state))
1037                 return -EINVAL;
1038
1039         /*
1040          * We get here with hb->lock held, and having found a
1041          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1042          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1043          * which in turn means that futex_lock_pi() still has a reference on
1044          * our pi_state.
1045          *
1046          * The waiter holding a reference on @pi_state also protects against
1047          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1048          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1049          * free pi_state before we can take a reference ourselves.
1050          */
1051         WARN_ON(!refcount_read(&pi_state->refcount));
1052
1053         /*
1054          * Now that we have a pi_state, we can acquire wait_lock
1055          * and do the state validation.
1056          */
1057         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1058
1059         /*
1060          * Since {uval, pi_state} is serialized by wait_lock, and our current
1061          * uval was read without holding it, it can have changed. Verify it
1062          * still is what we expect it to be, otherwise retry the entire
1063          * operation.
1064          */
1065         if (get_futex_value_locked(&uval2, uaddr))
1066                 goto out_efault;
1067
1068         if (uval != uval2)
1069                 goto out_eagain;
1070
1071         /*
1072          * Handle the owner died case:
1073          */
1074         if (uval & FUTEX_OWNER_DIED) {
1075                 /*
1076                  * exit_pi_state_list sets owner to NULL and wakes the
1077                  * topmost waiter. The task which acquires the
1078                  * pi_state->rt_mutex will fixup owner.
1079                  */
1080                 if (!pi_state->owner) {
1081                         /*
1082                          * No pi state owner, but the user space TID
1083                          * is not 0. Inconsistent state. [5]
1084                          */
1085                         if (pid)
1086                                 goto out_einval;
1087                         /*
1088                          * Take a ref on the state and return success. [4]
1089                          */
1090                         goto out_attach;
1091                 }
1092
1093                 /*
1094                  * If TID is 0, then either the dying owner has not
1095                  * yet executed exit_pi_state_list() or some waiter
1096                  * acquired the rtmutex in the pi state, but did not
1097                  * yet fixup the TID in user space.
1098                  *
1099                  * Take a ref on the state and return success. [6]
1100                  */
1101                 if (!pid)
1102                         goto out_attach;
1103         } else {
1104                 /*
1105                  * If the owner died bit is not set, then the pi_state
1106                  * must have an owner. [7]
1107                  */
1108                 if (!pi_state->owner)
1109                         goto out_einval;
1110         }
1111
1112         /*
1113          * Bail out if user space manipulated the futex value. If pi
1114          * state exists then the owner TID must be the same as the
1115          * user space TID. [9/10]
1116          */
1117         if (pid != task_pid_vnr(pi_state->owner))
1118                 goto out_einval;
1119
1120 out_attach:
1121         get_pi_state(pi_state);
1122         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1123         *ps = pi_state;
1124         return 0;
1125
1126 out_einval:
1127         ret = -EINVAL;
1128         goto out_error;
1129
1130 out_eagain:
1131         ret = -EAGAIN;
1132         goto out_error;
1133
1134 out_efault:
1135         ret = -EFAULT;
1136         goto out_error;
1137
1138 out_error:
1139         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1140         return ret;
1141 }
1142
1143 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1144                             struct task_struct *tsk)
1145 {
1146         u32 uval2;
1147
1148         /*
1149          * If PF_EXITPIDONE is not yet set, then try again.
1150          */
1151         if (tsk && !(tsk->flags & PF_EXITPIDONE))
1152                 return -EAGAIN;
1153
1154         /*
1155          * Reread the user space value to handle the following situation:
1156          *
1157          * CPU0                         CPU1
1158          *
1159          * sys_exit()                   sys_futex()
1160          *  do_exit()                    futex_lock_pi()
1161          *                                futex_lock_pi_atomic()
1162          *   exit_signals(tsk)              No waiters:
1163          *    tsk->flags |= PF_EXITING;     *uaddr == 0x00000PID
1164          *  mm_release(tsk)                 Set waiter bit
1165          *   exit_robust_list(tsk) {        *uaddr = 0x80000PID;
1166          *      Set owner died              attach_to_pi_owner() {
1167          *    *uaddr = 0xC0000000;           tsk = get_task(PID);
1168          *   }                               if (!tsk->flags & PF_EXITING) {
1169          *  ...                                attach();
1170          *  tsk->flags |= PF_EXITPIDONE;     } else {
1171          *                                     if (!(tsk->flags & PF_EXITPIDONE))
1172          *                                       return -EAGAIN;
1173          *                                     return -ESRCH; <--- FAIL
1174          *                                   }
1175          *
1176          * Returning ESRCH unconditionally is wrong here because the
1177          * user space value has been changed by the exiting task.
1178          *
1179          * The same logic applies to the case where the exiting task is
1180          * already gone.
1181          */
1182         if (get_futex_value_locked(&uval2, uaddr))
1183                 return -EFAULT;
1184
1185         /* If the user space value has changed, try again. */
1186         if (uval2 != uval)
1187                 return -EAGAIN;
1188
1189         /*
1190          * The exiting task did not have a robust list, the robust list was
1191          * corrupted or the user space value in *uaddr is simply bogus.
1192          * Give up and tell user space.
1193          */
1194         return -ESRCH;
1195 }
1196
1197 /*
1198  * Lookup the task for the TID provided from user space and attach to
1199  * it after doing proper sanity checks.
1200  */
1201 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1202                               struct futex_pi_state **ps)
1203 {
1204         pid_t pid = uval & FUTEX_TID_MASK;
1205         struct futex_pi_state *pi_state;
1206         struct task_struct *p;
1207
1208         /*
1209          * We are the first waiter - try to look up the real owner and attach
1210          * the new pi_state to it, but bail out when TID = 0 [1]
1211          *
1212          * The !pid check is paranoid. None of the call sites should end up
1213          * with pid == 0, but better safe than sorry. Let the caller retry
1214          */
1215         if (!pid)
1216                 return -EAGAIN;
1217         p = find_get_task_by_vpid(pid);
1218         if (!p)
1219                 return handle_exit_race(uaddr, uval, NULL);
1220
1221         if (unlikely(p->flags & PF_KTHREAD)) {
1222                 put_task_struct(p);
1223                 return -EPERM;
1224         }
1225
1226         /*
1227          * We need to look at the task state flags to figure out,
1228          * whether the task is exiting. To protect against the do_exit
1229          * change of the task flags, we do this protected by
1230          * p->pi_lock:
1231          */
1232         raw_spin_lock_irq(&p->pi_lock);
1233         if (unlikely(p->flags & PF_EXITING)) {
1234                 /*
1235                  * The task is on the way out. When PF_EXITPIDONE is
1236                  * set, we know that the task has finished the
1237                  * cleanup:
1238                  */
1239                 int ret = handle_exit_race(uaddr, uval, p);
1240
1241                 raw_spin_unlock_irq(&p->pi_lock);
1242                 put_task_struct(p);
1243                 return ret;
1244         }
1245
1246         /*
1247          * No existing pi state. First waiter. [2]
1248          *
1249          * This creates pi_state, we have hb->lock held, this means nothing can
1250          * observe this state, wait_lock is irrelevant.
1251          */
1252         pi_state = alloc_pi_state();
1253
1254         /*
1255          * Initialize the pi_mutex in locked state and make @p
1256          * the owner of it:
1257          */
1258         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1259
1260         /* Store the key for possible exit cleanups: */
1261         pi_state->key = *key;
1262
1263         WARN_ON(!list_empty(&pi_state->list));
1264         list_add(&pi_state->list, &p->pi_state_list);
1265         /*
1266          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1267          * because there is no concurrency as the object is not published yet.
1268          */
1269         pi_state->owner = p;
1270         raw_spin_unlock_irq(&p->pi_lock);
1271
1272         put_task_struct(p);
1273
1274         *ps = pi_state;
1275
1276         return 0;
1277 }
1278
1279 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1280                            struct futex_hash_bucket *hb,
1281                            union futex_key *key, struct futex_pi_state **ps)
1282 {
1283         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1284
1285         /*
1286          * If there is a waiter on that futex, validate it and
1287          * attach to the pi_state when the validation succeeds.
1288          */
1289         if (top_waiter)
1290                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1291
1292         /*
1293          * We are the first waiter - try to look up the owner based on
1294          * @uval and attach to it.
1295          */
1296         return attach_to_pi_owner(uaddr, uval, key, ps);
1297 }
1298
1299 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1300 {
1301         int err;
1302         u32 uninitialized_var(curval);
1303
1304         if (unlikely(should_fail_futex(true)))
1305                 return -EFAULT;
1306
1307         err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1308         if (unlikely(err))
1309                 return err;
1310
1311         /* If user space value changed, let the caller retry */
1312         return curval != uval ? -EAGAIN : 0;
1313 }
1314
1315 /**
1316  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1317  * @uaddr:              the pi futex user address
1318  * @hb:                 the pi futex hash bucket
1319  * @key:                the futex key associated with uaddr and hb
1320  * @ps:                 the pi_state pointer where we store the result of the
1321  *                      lookup
1322  * @task:               the task to perform the atomic lock work for.  This will
1323  *                      be "current" except in the case of requeue pi.
1324  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1325  *
1326  * Return:
1327  *  -  0 - ready to wait;
1328  *  -  1 - acquired the lock;
1329  *  - <0 - error
1330  *
1331  * The hb->lock and futex_key refs shall be held by the caller.
1332  */
1333 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1334                                 union futex_key *key,
1335                                 struct futex_pi_state **ps,
1336                                 struct task_struct *task, int set_waiters)
1337 {
1338         u32 uval, newval, vpid = task_pid_vnr(task);
1339         struct futex_q *top_waiter;
1340         int ret;
1341
1342         /*
1343          * Read the user space value first so we can validate a few
1344          * things before proceeding further.
1345          */
1346         if (get_futex_value_locked(&uval, uaddr))
1347                 return -EFAULT;
1348
1349         if (unlikely(should_fail_futex(true)))
1350                 return -EFAULT;
1351
1352         /*
1353          * Detect deadlocks.
1354          */
1355         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1356                 return -EDEADLK;
1357
1358         if ((unlikely(should_fail_futex(true))))
1359                 return -EDEADLK;
1360
1361         /*
1362          * Lookup existing state first. If it exists, try to attach to
1363          * its pi_state.
1364          */
1365         top_waiter = futex_top_waiter(hb, key);
1366         if (top_waiter)
1367                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1368
1369         /*
1370          * No waiter and user TID is 0. We are here because the
1371          * waiters or the owner died bit is set or called from
1372          * requeue_cmp_pi or for whatever reason something took the
1373          * syscall.
1374          */
1375         if (!(uval & FUTEX_TID_MASK)) {
1376                 /*
1377                  * We take over the futex. No other waiters and the user space
1378                  * TID is 0. We preserve the owner died bit.
1379                  */
1380                 newval = uval & FUTEX_OWNER_DIED;
1381                 newval |= vpid;
1382
1383                 /* The futex requeue_pi code can enforce the waiters bit */
1384                 if (set_waiters)
1385                         newval |= FUTEX_WAITERS;
1386
1387                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1388                 /* If the take over worked, return 1 */
1389                 return ret < 0 ? ret : 1;
1390         }
1391
1392         /*
1393          * First waiter. Set the waiters bit before attaching ourself to
1394          * the owner. If owner tries to unlock, it will be forced into
1395          * the kernel and blocked on hb->lock.
1396          */
1397         newval = uval | FUTEX_WAITERS;
1398         ret = lock_pi_update_atomic(uaddr, uval, newval);
1399         if (ret)
1400                 return ret;
1401         /*
1402          * If the update of the user space value succeeded, we try to
1403          * attach to the owner. If that fails, no harm done, we only
1404          * set the FUTEX_WAITERS bit in the user space variable.
1405          */
1406         return attach_to_pi_owner(uaddr, newval, key, ps);
1407 }
1408
1409 /**
1410  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1411  * @q:  The futex_q to unqueue
1412  *
1413  * The q->lock_ptr must not be NULL and must be held by the caller.
1414  */
1415 static void __unqueue_futex(struct futex_q *q)
1416 {
1417         struct futex_hash_bucket *hb;
1418
1419         if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1420                 return;
1421         lockdep_assert_held(q->lock_ptr);
1422
1423         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1424         plist_del(&q->list, &hb->chain);
1425         hb_waiters_dec(hb);
1426 }
1427
1428 /*
1429  * The hash bucket lock must be held when this is called.
1430  * Afterwards, the futex_q must not be accessed. Callers
1431  * must ensure to later call wake_up_q() for the actual
1432  * wakeups to occur.
1433  */
1434 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1435 {
1436         struct task_struct *p = q->task;
1437
1438         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1439                 return;
1440
1441         get_task_struct(p);
1442         __unqueue_futex(q);
1443         /*
1444          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1445          * is written, without taking any locks. This is possible in the event
1446          * of a spurious wakeup, for example. A memory barrier is required here
1447          * to prevent the following store to lock_ptr from getting ahead of the
1448          * plist_del in __unqueue_futex().
1449          */
1450         smp_store_release(&q->lock_ptr, NULL);
1451
1452         /*
1453          * Queue the task for later wakeup for after we've released
1454          * the hb->lock. wake_q_add() grabs reference to p.
1455          */
1456         wake_q_add_safe(wake_q, p);
1457 }
1458
1459 /*
1460  * Caller must hold a reference on @pi_state.
1461  */
1462 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1463 {
1464         u32 uninitialized_var(curval), newval;
1465         struct task_struct *new_owner;
1466         bool postunlock = false;
1467         DEFINE_WAKE_Q(wake_q);
1468         int ret = 0;
1469
1470         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1471         if (WARN_ON_ONCE(!new_owner)) {
1472                 /*
1473                  * As per the comment in futex_unlock_pi() this should not happen.
1474                  *
1475                  * When this happens, give up our locks and try again, giving
1476                  * the futex_lock_pi() instance time to complete, either by
1477                  * waiting on the rtmutex or removing itself from the futex
1478                  * queue.
1479                  */
1480                 ret = -EAGAIN;
1481                 goto out_unlock;
1482         }
1483
1484         /*
1485          * We pass it to the next owner. The WAITERS bit is always kept
1486          * enabled while there is PI state around. We cleanup the owner
1487          * died bit, because we are the owner.
1488          */
1489         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1490
1491         if (unlikely(should_fail_futex(true)))
1492                 ret = -EFAULT;
1493
1494         ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1495         if (!ret && (curval != uval)) {
1496                 /*
1497                  * If a unconditional UNLOCK_PI operation (user space did not
1498                  * try the TID->0 transition) raced with a waiter setting the
1499                  * FUTEX_WAITERS flag between get_user() and locking the hash
1500                  * bucket lock, retry the operation.
1501                  */
1502                 if ((FUTEX_TID_MASK & curval) == uval)
1503                         ret = -EAGAIN;
1504                 else
1505                         ret = -EINVAL;
1506         }
1507
1508         if (ret)
1509                 goto out_unlock;
1510
1511         /*
1512          * This is a point of no return; once we modify the uval there is no
1513          * going back and subsequent operations must not fail.
1514          */
1515
1516         raw_spin_lock(&pi_state->owner->pi_lock);
1517         WARN_ON(list_empty(&pi_state->list));
1518         list_del_init(&pi_state->list);
1519         raw_spin_unlock(&pi_state->owner->pi_lock);
1520
1521         raw_spin_lock(&new_owner->pi_lock);
1522         WARN_ON(!list_empty(&pi_state->list));
1523         list_add(&pi_state->list, &new_owner->pi_state_list);
1524         pi_state->owner = new_owner;
1525         raw_spin_unlock(&new_owner->pi_lock);
1526
1527         postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1528
1529 out_unlock:
1530         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1531
1532         if (postunlock)
1533                 rt_mutex_postunlock(&wake_q);
1534
1535         return ret;
1536 }
1537
1538 /*
1539  * Express the locking dependencies for lockdep:
1540  */
1541 static inline void
1542 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1543 {
1544         if (hb1 <= hb2) {
1545                 spin_lock(&hb1->lock);
1546                 if (hb1 < hb2)
1547                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1548         } else { /* hb1 > hb2 */
1549                 spin_lock(&hb2->lock);
1550                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1551         }
1552 }
1553
1554 static inline void
1555 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1556 {
1557         spin_unlock(&hb1->lock);
1558         if (hb1 != hb2)
1559                 spin_unlock(&hb2->lock);
1560 }
1561
1562 /*
1563  * Wake up waiters matching bitset queued on this futex (uaddr).
1564  */
1565 static int
1566 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1567 {
1568         struct futex_hash_bucket *hb;
1569         struct futex_q *this, *next;
1570         union futex_key key = FUTEX_KEY_INIT;
1571         int ret;
1572         DEFINE_WAKE_Q(wake_q);
1573
1574         if (!bitset)
1575                 return -EINVAL;
1576
1577         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1578         if (unlikely(ret != 0))
1579                 goto out;
1580
1581         hb = hash_futex(&key);
1582
1583         /* Make sure we really have tasks to wakeup */
1584         if (!hb_waiters_pending(hb))
1585                 goto out_put_key;
1586
1587         spin_lock(&hb->lock);
1588
1589         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1590                 if (match_futex (&this->key, &key)) {
1591                         if (this->pi_state || this->rt_waiter) {
1592                                 ret = -EINVAL;
1593                                 break;
1594                         }
1595
1596                         /* Check if one of the bits is set in both bitsets */
1597                         if (!(this->bitset & bitset))
1598                                 continue;
1599
1600                         mark_wake_futex(&wake_q, this);
1601                         if (++ret >= nr_wake)
1602                                 break;
1603                 }
1604         }
1605
1606         spin_unlock(&hb->lock);
1607         wake_up_q(&wake_q);
1608 out_put_key:
1609         put_futex_key(&key);
1610 out:
1611         return ret;
1612 }
1613
1614 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1615 {
1616         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1617         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1618         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1619         int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1620         int oldval, ret;
1621
1622         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1623                 if (oparg < 0 || oparg > 31) {
1624                         char comm[sizeof(current->comm)];
1625                         /*
1626                          * kill this print and return -EINVAL when userspace
1627                          * is sane again
1628                          */
1629                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1630                                         get_task_comm(comm, current), oparg);
1631                         oparg &= 31;
1632                 }
1633                 oparg = 1 << oparg;
1634         }
1635
1636         if (!access_ok(uaddr, sizeof(u32)))
1637                 return -EFAULT;
1638
1639         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1640         if (ret)
1641                 return ret;
1642
1643         switch (cmp) {
1644         case FUTEX_OP_CMP_EQ:
1645                 return oldval == cmparg;
1646         case FUTEX_OP_CMP_NE:
1647                 return oldval != cmparg;
1648         case FUTEX_OP_CMP_LT:
1649                 return oldval < cmparg;
1650         case FUTEX_OP_CMP_GE:
1651                 return oldval >= cmparg;
1652         case FUTEX_OP_CMP_LE:
1653                 return oldval <= cmparg;
1654         case FUTEX_OP_CMP_GT:
1655                 return oldval > cmparg;
1656         default:
1657                 return -ENOSYS;
1658         }
1659 }
1660
1661 /*
1662  * Wake up all waiters hashed on the physical page that is mapped
1663  * to this virtual address:
1664  */
1665 static int
1666 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1667               int nr_wake, int nr_wake2, int op)
1668 {
1669         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1670         struct futex_hash_bucket *hb1, *hb2;
1671         struct futex_q *this, *next;
1672         int ret, op_ret;
1673         DEFINE_WAKE_Q(wake_q);
1674
1675 retry:
1676         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1677         if (unlikely(ret != 0))
1678                 goto out;
1679         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1680         if (unlikely(ret != 0))
1681                 goto out_put_key1;
1682
1683         hb1 = hash_futex(&key1);
1684         hb2 = hash_futex(&key2);
1685
1686 retry_private:
1687         double_lock_hb(hb1, hb2);
1688         op_ret = futex_atomic_op_inuser(op, uaddr2);
1689         if (unlikely(op_ret < 0)) {
1690                 double_unlock_hb(hb1, hb2);
1691
1692                 if (!IS_ENABLED(CONFIG_MMU) ||
1693                     unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1694                         /*
1695                          * we don't get EFAULT from MMU faults if we don't have
1696                          * an MMU, but we might get them from range checking
1697                          */
1698                         ret = op_ret;
1699                         goto out_put_keys;
1700                 }
1701
1702                 if (op_ret == -EFAULT) {
1703                         ret = fault_in_user_writeable(uaddr2);
1704                         if (ret)
1705                                 goto out_put_keys;
1706                 }
1707
1708                 if (!(flags & FLAGS_SHARED)) {
1709                         cond_resched();
1710                         goto retry_private;
1711                 }
1712
1713                 put_futex_key(&key2);
1714                 put_futex_key(&key1);
1715                 cond_resched();
1716                 goto retry;
1717         }
1718
1719         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1720                 if (match_futex (&this->key, &key1)) {
1721                         if (this->pi_state || this->rt_waiter) {
1722                                 ret = -EINVAL;
1723                                 goto out_unlock;
1724                         }
1725                         mark_wake_futex(&wake_q, this);
1726                         if (++ret >= nr_wake)
1727                                 break;
1728                 }
1729         }
1730
1731         if (op_ret > 0) {
1732                 op_ret = 0;
1733                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1734                         if (match_futex (&this->key, &key2)) {
1735                                 if (this->pi_state || this->rt_waiter) {
1736                                         ret = -EINVAL;
1737                                         goto out_unlock;
1738                                 }
1739                                 mark_wake_futex(&wake_q, this);
1740                                 if (++op_ret >= nr_wake2)
1741                                         break;
1742                         }
1743                 }
1744                 ret += op_ret;
1745         }
1746
1747 out_unlock:
1748         double_unlock_hb(hb1, hb2);
1749         wake_up_q(&wake_q);
1750 out_put_keys:
1751         put_futex_key(&key2);
1752 out_put_key1:
1753         put_futex_key(&key1);
1754 out:
1755         return ret;
1756 }
1757
1758 /**
1759  * requeue_futex() - Requeue a futex_q from one hb to another
1760  * @q:          the futex_q to requeue
1761  * @hb1:        the source hash_bucket
1762  * @hb2:        the target hash_bucket
1763  * @key2:       the new key for the requeued futex_q
1764  */
1765 static inline
1766 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1767                    struct futex_hash_bucket *hb2, union futex_key *key2)
1768 {
1769
1770         /*
1771          * If key1 and key2 hash to the same bucket, no need to
1772          * requeue.
1773          */
1774         if (likely(&hb1->chain != &hb2->chain)) {
1775                 plist_del(&q->list, &hb1->chain);
1776                 hb_waiters_dec(hb1);
1777                 hb_waiters_inc(hb2);
1778                 plist_add(&q->list, &hb2->chain);
1779                 q->lock_ptr = &hb2->lock;
1780         }
1781         get_futex_key_refs(key2);
1782         q->key = *key2;
1783 }
1784
1785 /**
1786  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1787  * @q:          the futex_q
1788  * @key:        the key of the requeue target futex
1789  * @hb:         the hash_bucket of the requeue target futex
1790  *
1791  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1792  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1793  * to the requeue target futex so the waiter can detect the wakeup on the right
1794  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1795  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1796  * to protect access to the pi_state to fixup the owner later.  Must be called
1797  * with both q->lock_ptr and hb->lock held.
1798  */
1799 static inline
1800 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1801                            struct futex_hash_bucket *hb)
1802 {
1803         get_futex_key_refs(key);
1804         q->key = *key;
1805
1806         __unqueue_futex(q);
1807
1808         WARN_ON(!q->rt_waiter);
1809         q->rt_waiter = NULL;
1810
1811         q->lock_ptr = &hb->lock;
1812
1813         wake_up_state(q->task, TASK_NORMAL);
1814 }
1815
1816 /**
1817  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1818  * @pifutex:            the user address of the to futex
1819  * @hb1:                the from futex hash bucket, must be locked by the caller
1820  * @hb2:                the to futex hash bucket, must be locked by the caller
1821  * @key1:               the from futex key
1822  * @key2:               the to futex key
1823  * @ps:                 address to store the pi_state pointer
1824  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1825  *
1826  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1827  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1828  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1829  * hb1 and hb2 must be held by the caller.
1830  *
1831  * Return:
1832  *  -  0 - failed to acquire the lock atomically;
1833  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1834  *  - <0 - error
1835  */
1836 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1837                                  struct futex_hash_bucket *hb1,
1838                                  struct futex_hash_bucket *hb2,
1839                                  union futex_key *key1, union futex_key *key2,
1840                                  struct futex_pi_state **ps, int set_waiters)
1841 {
1842         struct futex_q *top_waiter = NULL;
1843         u32 curval;
1844         int ret, vpid;
1845
1846         if (get_futex_value_locked(&curval, pifutex))
1847                 return -EFAULT;
1848
1849         if (unlikely(should_fail_futex(true)))
1850                 return -EFAULT;
1851
1852         /*
1853          * Find the top_waiter and determine if there are additional waiters.
1854          * If the caller intends to requeue more than 1 waiter to pifutex,
1855          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1856          * as we have means to handle the possible fault.  If not, don't set
1857          * the bit unecessarily as it will force the subsequent unlock to enter
1858          * the kernel.
1859          */
1860         top_waiter = futex_top_waiter(hb1, key1);
1861
1862         /* There are no waiters, nothing for us to do. */
1863         if (!top_waiter)
1864                 return 0;
1865
1866         /* Ensure we requeue to the expected futex. */
1867         if (!match_futex(top_waiter->requeue_pi_key, key2))
1868                 return -EINVAL;
1869
1870         /*
1871          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1872          * the contended case or if set_waiters is 1.  The pi_state is returned
1873          * in ps in contended cases.
1874          */
1875         vpid = task_pid_vnr(top_waiter->task);
1876         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1877                                    set_waiters);
1878         if (ret == 1) {
1879                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1880                 return vpid;
1881         }
1882         return ret;
1883 }
1884
1885 /**
1886  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1887  * @uaddr1:     source futex user address
1888  * @flags:      futex flags (FLAGS_SHARED, etc.)
1889  * @uaddr2:     target futex user address
1890  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1891  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1892  * @cmpval:     @uaddr1 expected value (or %NULL)
1893  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1894  *              pi futex (pi to pi requeue is not supported)
1895  *
1896  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1897  * uaddr2 atomically on behalf of the top waiter.
1898  *
1899  * Return:
1900  *  - >=0 - on success, the number of tasks requeued or woken;
1901  *  -  <0 - on error
1902  */
1903 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1904                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1905                          u32 *cmpval, int requeue_pi)
1906 {
1907         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1908         int drop_count = 0, task_count = 0, ret;
1909         struct futex_pi_state *pi_state = NULL;
1910         struct futex_hash_bucket *hb1, *hb2;
1911         struct futex_q *this, *next;
1912         DEFINE_WAKE_Q(wake_q);
1913
1914         if (nr_wake < 0 || nr_requeue < 0)
1915                 return -EINVAL;
1916
1917         /*
1918          * When PI not supported: return -ENOSYS if requeue_pi is true,
1919          * consequently the compiler knows requeue_pi is always false past
1920          * this point which will optimize away all the conditional code
1921          * further down.
1922          */
1923         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1924                 return -ENOSYS;
1925
1926         if (requeue_pi) {
1927                 /*
1928                  * Requeue PI only works on two distinct uaddrs. This
1929                  * check is only valid for private futexes. See below.
1930                  */
1931                 if (uaddr1 == uaddr2)
1932                         return -EINVAL;
1933
1934                 /*
1935                  * requeue_pi requires a pi_state, try to allocate it now
1936                  * without any locks in case it fails.
1937                  */
1938                 if (refill_pi_state_cache())
1939                         return -ENOMEM;
1940                 /*
1941                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1942                  * + nr_requeue, since it acquires the rt_mutex prior to
1943                  * returning to userspace, so as to not leave the rt_mutex with
1944                  * waiters and no owner.  However, second and third wake-ups
1945                  * cannot be predicted as they involve race conditions with the
1946                  * first wake and a fault while looking up the pi_state.  Both
1947                  * pthread_cond_signal() and pthread_cond_broadcast() should
1948                  * use nr_wake=1.
1949                  */
1950                 if (nr_wake != 1)
1951                         return -EINVAL;
1952         }
1953
1954 retry:
1955         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1956         if (unlikely(ret != 0))
1957                 goto out;
1958         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1959                             requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1960         if (unlikely(ret != 0))
1961                 goto out_put_key1;
1962
1963         /*
1964          * The check above which compares uaddrs is not sufficient for
1965          * shared futexes. We need to compare the keys:
1966          */
1967         if (requeue_pi && match_futex(&key1, &key2)) {
1968                 ret = -EINVAL;
1969                 goto out_put_keys;
1970         }
1971
1972         hb1 = hash_futex(&key1);
1973         hb2 = hash_futex(&key2);
1974
1975 retry_private:
1976         hb_waiters_inc(hb2);
1977         double_lock_hb(hb1, hb2);
1978
1979         if (likely(cmpval != NULL)) {
1980                 u32 curval;
1981
1982                 ret = get_futex_value_locked(&curval, uaddr1);
1983
1984                 if (unlikely(ret)) {
1985                         double_unlock_hb(hb1, hb2);
1986                         hb_waiters_dec(hb2);
1987
1988                         ret = get_user(curval, uaddr1);
1989                         if (ret)
1990                                 goto out_put_keys;
1991
1992                         if (!(flags & FLAGS_SHARED))
1993                                 goto retry_private;
1994
1995                         put_futex_key(&key2);
1996                         put_futex_key(&key1);
1997                         goto retry;
1998                 }
1999                 if (curval != *cmpval) {
2000                         ret = -EAGAIN;
2001                         goto out_unlock;
2002                 }
2003         }
2004
2005         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2006                 /*
2007                  * Attempt to acquire uaddr2 and wake the top waiter. If we
2008                  * intend to requeue waiters, force setting the FUTEX_WAITERS
2009                  * bit.  We force this here where we are able to easily handle
2010                  * faults rather in the requeue loop below.
2011                  */
2012                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2013                                                  &key2, &pi_state, nr_requeue);
2014
2015                 /*
2016                  * At this point the top_waiter has either taken uaddr2 or is
2017                  * waiting on it.  If the former, then the pi_state will not
2018                  * exist yet, look it up one more time to ensure we have a
2019                  * reference to it. If the lock was taken, ret contains the
2020                  * vpid of the top waiter task.
2021                  * If the lock was not taken, we have pi_state and an initial
2022                  * refcount on it. In case of an error we have nothing.
2023                  */
2024                 if (ret > 0) {
2025                         WARN_ON(pi_state);
2026                         drop_count++;
2027                         task_count++;
2028                         /*
2029                          * If we acquired the lock, then the user space value
2030                          * of uaddr2 should be vpid. It cannot be changed by
2031                          * the top waiter as it is blocked on hb2 lock if it
2032                          * tries to do so. If something fiddled with it behind
2033                          * our back the pi state lookup might unearth it. So
2034                          * we rather use the known value than rereading and
2035                          * handing potential crap to lookup_pi_state.
2036                          *
2037                          * If that call succeeds then we have pi_state and an
2038                          * initial refcount on it.
2039                          */
2040                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2041                 }
2042
2043                 switch (ret) {
2044                 case 0:
2045                         /* We hold a reference on the pi state. */
2046                         break;
2047
2048                         /* If the above failed, then pi_state is NULL */
2049                 case -EFAULT:
2050                         double_unlock_hb(hb1, hb2);
2051                         hb_waiters_dec(hb2);
2052                         put_futex_key(&key2);
2053                         put_futex_key(&key1);
2054                         ret = fault_in_user_writeable(uaddr2);
2055                         if (!ret)
2056                                 goto retry;
2057                         goto out;
2058                 case -EAGAIN:
2059                         /*
2060                          * Two reasons for this:
2061                          * - Owner is exiting and we just wait for the
2062                          *   exit to complete.
2063                          * - The user space value changed.
2064                          */
2065                         double_unlock_hb(hb1, hb2);
2066                         hb_waiters_dec(hb2);
2067                         put_futex_key(&key2);
2068                         put_futex_key(&key1);
2069                         cond_resched();
2070                         goto retry;
2071                 default:
2072                         goto out_unlock;
2073                 }
2074         }
2075
2076         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2077                 if (task_count - nr_wake >= nr_requeue)
2078                         break;
2079
2080                 if (!match_futex(&this->key, &key1))
2081                         continue;
2082
2083                 /*
2084                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2085                  * be paired with each other and no other futex ops.
2086                  *
2087                  * We should never be requeueing a futex_q with a pi_state,
2088                  * which is awaiting a futex_unlock_pi().
2089                  */
2090                 if ((requeue_pi && !this->rt_waiter) ||
2091                     (!requeue_pi && this->rt_waiter) ||
2092                     this->pi_state) {
2093                         ret = -EINVAL;
2094                         break;
2095                 }
2096
2097                 /*
2098                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2099                  * lock, we already woke the top_waiter.  If not, it will be
2100                  * woken by futex_unlock_pi().
2101                  */
2102                 if (++task_count <= nr_wake && !requeue_pi) {
2103                         mark_wake_futex(&wake_q, this);
2104                         continue;
2105                 }
2106
2107                 /* Ensure we requeue to the expected futex for requeue_pi. */
2108                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2109                         ret = -EINVAL;
2110                         break;
2111                 }
2112
2113                 /*
2114                  * Requeue nr_requeue waiters and possibly one more in the case
2115                  * of requeue_pi if we couldn't acquire the lock atomically.
2116                  */
2117                 if (requeue_pi) {
2118                         /*
2119                          * Prepare the waiter to take the rt_mutex. Take a
2120                          * refcount on the pi_state and store the pointer in
2121                          * the futex_q object of the waiter.
2122                          */
2123                         get_pi_state(pi_state);
2124                         this->pi_state = pi_state;
2125                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2126                                                         this->rt_waiter,
2127                                                         this->task);
2128                         if (ret == 1) {
2129                                 /*
2130                                  * We got the lock. We do neither drop the
2131                                  * refcount on pi_state nor clear
2132                                  * this->pi_state because the waiter needs the
2133                                  * pi_state for cleaning up the user space
2134                                  * value. It will drop the refcount after
2135                                  * doing so.
2136                                  */
2137                                 requeue_pi_wake_futex(this, &key2, hb2);
2138                                 drop_count++;
2139                                 continue;
2140                         } else if (ret) {
2141                                 /*
2142                                  * rt_mutex_start_proxy_lock() detected a
2143                                  * potential deadlock when we tried to queue
2144                                  * that waiter. Drop the pi_state reference
2145                                  * which we took above and remove the pointer
2146                                  * to the state from the waiters futex_q
2147                                  * object.
2148                                  */
2149                                 this->pi_state = NULL;
2150                                 put_pi_state(pi_state);
2151                                 /*
2152                                  * We stop queueing more waiters and let user
2153                                  * space deal with the mess.
2154                                  */
2155                                 break;
2156                         }
2157                 }
2158                 requeue_futex(this, hb1, hb2, &key2);
2159                 drop_count++;
2160         }
2161
2162         /*
2163          * We took an extra initial reference to the pi_state either
2164          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2165          * need to drop it here again.
2166          */
2167         put_pi_state(pi_state);
2168
2169 out_unlock:
2170         double_unlock_hb(hb1, hb2);
2171         wake_up_q(&wake_q);
2172         hb_waiters_dec(hb2);
2173
2174         /*
2175          * drop_futex_key_refs() must be called outside the spinlocks. During
2176          * the requeue we moved futex_q's from the hash bucket at key1 to the
2177          * one at key2 and updated their key pointer.  We no longer need to
2178          * hold the references to key1.
2179          */
2180         while (--drop_count >= 0)
2181                 drop_futex_key_refs(&key1);
2182
2183 out_put_keys:
2184         put_futex_key(&key2);
2185 out_put_key1:
2186         put_futex_key(&key1);
2187 out:
2188         return ret ? ret : task_count;
2189 }
2190
2191 /* The key must be already stored in q->key. */
2192 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2193         __acquires(&hb->lock)
2194 {
2195         struct futex_hash_bucket *hb;
2196
2197         hb = hash_futex(&q->key);
2198
2199         /*
2200          * Increment the counter before taking the lock so that
2201          * a potential waker won't miss a to-be-slept task that is
2202          * waiting for the spinlock. This is safe as all queue_lock()
2203          * users end up calling queue_me(). Similarly, for housekeeping,
2204          * decrement the counter at queue_unlock() when some error has
2205          * occurred and we don't end up adding the task to the list.
2206          */
2207         hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2208
2209         q->lock_ptr = &hb->lock;
2210
2211         spin_lock(&hb->lock);
2212         return hb;
2213 }
2214
2215 static inline void
2216 queue_unlock(struct futex_hash_bucket *hb)
2217         __releases(&hb->lock)
2218 {
2219         spin_unlock(&hb->lock);
2220         hb_waiters_dec(hb);
2221 }
2222
2223 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2224 {
2225         int prio;
2226
2227         /*
2228          * The priority used to register this element is
2229          * - either the real thread-priority for the real-time threads
2230          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2231          * - or MAX_RT_PRIO for non-RT threads.
2232          * Thus, all RT-threads are woken first in priority order, and
2233          * the others are woken last, in FIFO order.
2234          */
2235         prio = min(current->normal_prio, MAX_RT_PRIO);
2236
2237         plist_node_init(&q->list, prio);
2238         plist_add(&q->list, &hb->chain);
2239         q->task = current;
2240 }
2241
2242 /**
2243  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2244  * @q:  The futex_q to enqueue
2245  * @hb: The destination hash bucket
2246  *
2247  * The hb->lock must be held by the caller, and is released here. A call to
2248  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2249  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2250  * or nothing if the unqueue is done as part of the wake process and the unqueue
2251  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2252  * an example).
2253  */
2254 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2255         __releases(&hb->lock)
2256 {
2257         __queue_me(q, hb);
2258         spin_unlock(&hb->lock);
2259 }
2260
2261 /**
2262  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2263  * @q:  The futex_q to unqueue
2264  *
2265  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2266  * be paired with exactly one earlier call to queue_me().
2267  *
2268  * Return:
2269  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2270  *  - 0 - if the futex_q was already removed by the waking thread
2271  */
2272 static int unqueue_me(struct futex_q *q)
2273 {
2274         spinlock_t *lock_ptr;
2275         int ret = 0;
2276
2277         /* In the common case we don't take the spinlock, which is nice. */
2278 retry:
2279         /*
2280          * q->lock_ptr can change between this read and the following spin_lock.
2281          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2282          * optimizing lock_ptr out of the logic below.
2283          */
2284         lock_ptr = READ_ONCE(q->lock_ptr);
2285         if (lock_ptr != NULL) {
2286                 spin_lock(lock_ptr);
2287                 /*
2288                  * q->lock_ptr can change between reading it and
2289                  * spin_lock(), causing us to take the wrong lock.  This
2290                  * corrects the race condition.
2291                  *
2292                  * Reasoning goes like this: if we have the wrong lock,
2293                  * q->lock_ptr must have changed (maybe several times)
2294                  * between reading it and the spin_lock().  It can
2295                  * change again after the spin_lock() but only if it was
2296                  * already changed before the spin_lock().  It cannot,
2297                  * however, change back to the original value.  Therefore
2298                  * we can detect whether we acquired the correct lock.
2299                  */
2300                 if (unlikely(lock_ptr != q->lock_ptr)) {
2301                         spin_unlock(lock_ptr);
2302                         goto retry;
2303                 }
2304                 __unqueue_futex(q);
2305
2306                 BUG_ON(q->pi_state);
2307
2308                 spin_unlock(lock_ptr);
2309                 ret = 1;
2310         }
2311
2312         drop_futex_key_refs(&q->key);
2313         return ret;
2314 }
2315
2316 /*
2317  * PI futexes can not be requeued and must remove themself from the
2318  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2319  * and dropped here.
2320  */
2321 static void unqueue_me_pi(struct futex_q *q)
2322         __releases(q->lock_ptr)
2323 {
2324         __unqueue_futex(q);
2325
2326         BUG_ON(!q->pi_state);
2327         put_pi_state(q->pi_state);
2328         q->pi_state = NULL;
2329
2330         spin_unlock(q->lock_ptr);
2331 }
2332
2333 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2334                                 struct task_struct *argowner)
2335 {
2336         struct futex_pi_state *pi_state = q->pi_state;
2337         u32 uval, uninitialized_var(curval), newval;
2338         struct task_struct *oldowner, *newowner;
2339         u32 newtid;
2340         int ret, err = 0;
2341
2342         lockdep_assert_held(q->lock_ptr);
2343
2344         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2345
2346         oldowner = pi_state->owner;
2347
2348         /*
2349          * We are here because either:
2350          *
2351          *  - we stole the lock and pi_state->owner needs updating to reflect
2352          *    that (@argowner == current),
2353          *
2354          * or:
2355          *
2356          *  - someone stole our lock and we need to fix things to point to the
2357          *    new owner (@argowner == NULL).
2358          *
2359          * Either way, we have to replace the TID in the user space variable.
2360          * This must be atomic as we have to preserve the owner died bit here.
2361          *
2362          * Note: We write the user space value _before_ changing the pi_state
2363          * because we can fault here. Imagine swapped out pages or a fork
2364          * that marked all the anonymous memory readonly for cow.
2365          *
2366          * Modifying pi_state _before_ the user space value would leave the
2367          * pi_state in an inconsistent state when we fault here, because we
2368          * need to drop the locks to handle the fault. This might be observed
2369          * in the PID check in lookup_pi_state.
2370          */
2371 retry:
2372         if (!argowner) {
2373                 if (oldowner != current) {
2374                         /*
2375                          * We raced against a concurrent self; things are
2376                          * already fixed up. Nothing to do.
2377                          */
2378                         ret = 0;
2379                         goto out_unlock;
2380                 }
2381
2382                 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2383                         /* We got the lock after all, nothing to fix. */
2384                         ret = 0;
2385                         goto out_unlock;
2386                 }
2387
2388                 /*
2389                  * Since we just failed the trylock; there must be an owner.
2390                  */
2391                 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2392                 BUG_ON(!newowner);
2393         } else {
2394                 WARN_ON_ONCE(argowner != current);
2395                 if (oldowner == current) {
2396                         /*
2397                          * We raced against a concurrent self; things are
2398                          * already fixed up. Nothing to do.
2399                          */
2400                         ret = 0;
2401                         goto out_unlock;
2402                 }
2403                 newowner = argowner;
2404         }
2405
2406         newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2407         /* Owner died? */
2408         if (!pi_state->owner)
2409                 newtid |= FUTEX_OWNER_DIED;
2410
2411         err = get_futex_value_locked(&uval, uaddr);
2412         if (err)
2413                 goto handle_err;
2414
2415         for (;;) {
2416                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2417
2418                 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2419                 if (err)
2420                         goto handle_err;
2421
2422                 if (curval == uval)
2423                         break;
2424                 uval = curval;
2425         }
2426
2427         /*
2428          * We fixed up user space. Now we need to fix the pi_state
2429          * itself.
2430          */
2431         if (pi_state->owner != NULL) {
2432                 raw_spin_lock(&pi_state->owner->pi_lock);
2433                 WARN_ON(list_empty(&pi_state->list));
2434                 list_del_init(&pi_state->list);
2435                 raw_spin_unlock(&pi_state->owner->pi_lock);
2436         }
2437
2438         pi_state->owner = newowner;
2439
2440         raw_spin_lock(&newowner->pi_lock);
2441         WARN_ON(!list_empty(&pi_state->list));
2442         list_add(&pi_state->list, &newowner->pi_state_list);
2443         raw_spin_unlock(&newowner->pi_lock);
2444         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2445
2446         return 0;
2447
2448         /*
2449          * In order to reschedule or handle a page fault, we need to drop the
2450          * locks here. In the case of a fault, this gives the other task
2451          * (either the highest priority waiter itself or the task which stole
2452          * the rtmutex) the chance to try the fixup of the pi_state. So once we
2453          * are back from handling the fault we need to check the pi_state after
2454          * reacquiring the locks and before trying to do another fixup. When
2455          * the fixup has been done already we simply return.
2456          *
2457          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2458          * drop hb->lock since the caller owns the hb -> futex_q relation.
2459          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2460          */
2461 handle_err:
2462         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2463         spin_unlock(q->lock_ptr);
2464
2465         switch (err) {
2466         case -EFAULT:
2467                 ret = fault_in_user_writeable(uaddr);
2468                 break;
2469
2470         case -EAGAIN:
2471                 cond_resched();
2472                 ret = 0;
2473                 break;
2474
2475         default:
2476                 WARN_ON_ONCE(1);
2477                 ret = err;
2478                 break;
2479         }
2480
2481         spin_lock(q->lock_ptr);
2482         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2483
2484         /*
2485          * Check if someone else fixed it for us:
2486          */
2487         if (pi_state->owner != oldowner) {
2488                 ret = 0;
2489                 goto out_unlock;
2490         }
2491
2492         if (ret)
2493                 goto out_unlock;
2494
2495         goto retry;
2496
2497 out_unlock:
2498         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2499         return ret;
2500 }
2501
2502 static long futex_wait_restart(struct restart_block *restart);
2503
2504 /**
2505  * fixup_owner() - Post lock pi_state and corner case management
2506  * @uaddr:      user address of the futex
2507  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2508  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2509  *
2510  * After attempting to lock an rt_mutex, this function is called to cleanup
2511  * the pi_state owner as well as handle race conditions that may allow us to
2512  * acquire the lock. Must be called with the hb lock held.
2513  *
2514  * Return:
2515  *  -  1 - success, lock taken;
2516  *  -  0 - success, lock not taken;
2517  *  - <0 - on error (-EFAULT)
2518  */
2519 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2520 {
2521         int ret = 0;
2522
2523         if (locked) {
2524                 /*
2525                  * Got the lock. We might not be the anticipated owner if we
2526                  * did a lock-steal - fix up the PI-state in that case:
2527                  *
2528                  * Speculative pi_state->owner read (we don't hold wait_lock);
2529                  * since we own the lock pi_state->owner == current is the
2530                  * stable state, anything else needs more attention.
2531                  */
2532                 if (q->pi_state->owner != current)
2533                         ret = fixup_pi_state_owner(uaddr, q, current);
2534                 goto out;
2535         }
2536
2537         /*
2538          * If we didn't get the lock; check if anybody stole it from us. In
2539          * that case, we need to fix up the uval to point to them instead of
2540          * us, otherwise bad things happen. [10]
2541          *
2542          * Another speculative read; pi_state->owner == current is unstable
2543          * but needs our attention.
2544          */
2545         if (q->pi_state->owner == current) {
2546                 ret = fixup_pi_state_owner(uaddr, q, NULL);
2547                 goto out;
2548         }
2549
2550         /*
2551          * Paranoia check. If we did not take the lock, then we should not be
2552          * the owner of the rt_mutex.
2553          */
2554         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2555                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2556                                 "pi-state %p\n", ret,
2557                                 q->pi_state->pi_mutex.owner,
2558                                 q->pi_state->owner);
2559         }
2560
2561 out:
2562         return ret ? ret : locked;
2563 }
2564
2565 /**
2566  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2567  * @hb:         the futex hash bucket, must be locked by the caller
2568  * @q:          the futex_q to queue up on
2569  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2570  */
2571 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2572                                 struct hrtimer_sleeper *timeout)
2573 {
2574         /*
2575          * The task state is guaranteed to be set before another task can
2576          * wake it. set_current_state() is implemented using smp_store_mb() and
2577          * queue_me() calls spin_unlock() upon completion, both serializing
2578          * access to the hash list and forcing another memory barrier.
2579          */
2580         set_current_state(TASK_INTERRUPTIBLE);
2581         queue_me(q, hb);
2582
2583         /* Arm the timer */
2584         if (timeout)
2585                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2586
2587         /*
2588          * If we have been removed from the hash list, then another task
2589          * has tried to wake us, and we can skip the call to schedule().
2590          */
2591         if (likely(!plist_node_empty(&q->list))) {
2592                 /*
2593                  * If the timer has already expired, current will already be
2594                  * flagged for rescheduling. Only call schedule if there
2595                  * is no timeout, or if it has yet to expire.
2596                  */
2597                 if (!timeout || timeout->task)
2598                         freezable_schedule();
2599         }
2600         __set_current_state(TASK_RUNNING);
2601 }
2602
2603 /**
2604  * futex_wait_setup() - Prepare to wait on a futex
2605  * @uaddr:      the futex userspace address
2606  * @val:        the expected value
2607  * @flags:      futex flags (FLAGS_SHARED, etc.)
2608  * @q:          the associated futex_q
2609  * @hb:         storage for hash_bucket pointer to be returned to caller
2610  *
2611  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2612  * compare it with the expected value.  Handle atomic faults internally.
2613  * Return with the hb lock held and a q.key reference on success, and unlocked
2614  * with no q.key reference on failure.
2615  *
2616  * Return:
2617  *  -  0 - uaddr contains val and hb has been locked;
2618  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2619  */
2620 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2621                            struct futex_q *q, struct futex_hash_bucket **hb)
2622 {
2623         u32 uval;
2624         int ret;
2625
2626         /*
2627          * Access the page AFTER the hash-bucket is locked.
2628          * Order is important:
2629          *
2630          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2631          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2632          *
2633          * The basic logical guarantee of a futex is that it blocks ONLY
2634          * if cond(var) is known to be true at the time of blocking, for
2635          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2636          * would open a race condition where we could block indefinitely with
2637          * cond(var) false, which would violate the guarantee.
2638          *
2639          * On the other hand, we insert q and release the hash-bucket only
2640          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2641          * absorb a wakeup if *uaddr does not match the desired values
2642          * while the syscall executes.
2643          */
2644 retry:
2645         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2646         if (unlikely(ret != 0))
2647                 return ret;
2648
2649 retry_private:
2650         *hb = queue_lock(q);
2651
2652         ret = get_futex_value_locked(&uval, uaddr);
2653
2654         if (ret) {
2655                 queue_unlock(*hb);
2656
2657                 ret = get_user(uval, uaddr);
2658                 if (ret)
2659                         goto out;
2660
2661                 if (!(flags & FLAGS_SHARED))
2662                         goto retry_private;
2663
2664                 put_futex_key(&q->key);
2665                 goto retry;
2666         }
2667
2668         if (uval != val) {
2669                 queue_unlock(*hb);
2670                 ret = -EWOULDBLOCK;
2671         }
2672
2673 out:
2674         if (ret)
2675                 put_futex_key(&q->key);
2676         return ret;
2677 }
2678
2679 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2680                       ktime_t *abs_time, u32 bitset)
2681 {
2682         struct hrtimer_sleeper timeout, *to = NULL;
2683         struct restart_block *restart;
2684         struct futex_hash_bucket *hb;
2685         struct futex_q q = futex_q_init;
2686         int ret;
2687
2688         if (!bitset)
2689                 return -EINVAL;
2690         q.bitset = bitset;
2691
2692         if (abs_time) {
2693                 to = &timeout;
2694
2695                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2696                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
2697                                       HRTIMER_MODE_ABS);
2698                 hrtimer_init_sleeper(to, current);
2699                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2700                                              current->timer_slack_ns);
2701         }
2702
2703 retry:
2704         /*
2705          * Prepare to wait on uaddr. On success, holds hb lock and increments
2706          * q.key refs.
2707          */
2708         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2709         if (ret)
2710                 goto out;
2711
2712         /* queue_me and wait for wakeup, timeout, or a signal. */
2713         futex_wait_queue_me(hb, &q, to);
2714
2715         /* If we were woken (and unqueued), we succeeded, whatever. */
2716         ret = 0;
2717         /* unqueue_me() drops q.key ref */
2718         if (!unqueue_me(&q))
2719                 goto out;
2720         ret = -ETIMEDOUT;
2721         if (to && !to->task)
2722                 goto out;
2723
2724         /*
2725          * We expect signal_pending(current), but we might be the
2726          * victim of a spurious wakeup as well.
2727          */
2728         if (!signal_pending(current))
2729                 goto retry;
2730
2731         ret = -ERESTARTSYS;
2732         if (!abs_time)
2733                 goto out;
2734
2735         restart = &current->restart_block;
2736         restart->fn = futex_wait_restart;
2737         restart->futex.uaddr = uaddr;
2738         restart->futex.val = val;
2739         restart->futex.time = *abs_time;
2740         restart->futex.bitset = bitset;
2741         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2742
2743         ret = -ERESTART_RESTARTBLOCK;
2744
2745 out:
2746         if (to) {
2747                 hrtimer_cancel(&to->timer);
2748                 destroy_hrtimer_on_stack(&to->timer);
2749         }
2750         return ret;
2751 }
2752
2753
2754 static long futex_wait_restart(struct restart_block *restart)
2755 {
2756         u32 __user *uaddr = restart->futex.uaddr;
2757         ktime_t t, *tp = NULL;
2758
2759         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2760                 t = restart->futex.time;
2761                 tp = &t;
2762         }
2763         restart->fn = do_no_restart_syscall;
2764
2765         return (long)futex_wait(uaddr, restart->futex.flags,
2766                                 restart->futex.val, tp, restart->futex.bitset);
2767 }
2768
2769
2770 /*
2771  * Userspace tried a 0 -> TID atomic transition of the futex value
2772  * and failed. The kernel side here does the whole locking operation:
2773  * if there are waiters then it will block as a consequence of relying
2774  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2775  * a 0 value of the futex too.).
2776  *
2777  * Also serves as futex trylock_pi()'ing, and due semantics.
2778  */
2779 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2780                          ktime_t *time, int trylock)
2781 {
2782         struct hrtimer_sleeper timeout, *to = NULL;
2783         struct futex_pi_state *pi_state = NULL;
2784         struct rt_mutex_waiter rt_waiter;
2785         struct futex_hash_bucket *hb;
2786         struct futex_q q = futex_q_init;
2787         int res, ret;
2788
2789         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2790                 return -ENOSYS;
2791
2792         if (refill_pi_state_cache())
2793                 return -ENOMEM;
2794
2795         if (time) {
2796                 to = &timeout;
2797                 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2798                                       HRTIMER_MODE_ABS);
2799                 hrtimer_init_sleeper(to, current);
2800                 hrtimer_set_expires(&to->timer, *time);
2801         }
2802
2803 retry:
2804         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2805         if (unlikely(ret != 0))
2806                 goto out;
2807
2808 retry_private:
2809         hb = queue_lock(&q);
2810
2811         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2812         if (unlikely(ret)) {
2813                 /*
2814                  * Atomic work succeeded and we got the lock,
2815                  * or failed. Either way, we do _not_ block.
2816                  */
2817                 switch (ret) {
2818                 case 1:
2819                         /* We got the lock. */
2820                         ret = 0;
2821                         goto out_unlock_put_key;
2822                 case -EFAULT:
2823                         goto uaddr_faulted;
2824                 case -EAGAIN:
2825                         /*
2826                          * Two reasons for this:
2827                          * - Task is exiting and we just wait for the
2828                          *   exit to complete.
2829                          * - The user space value changed.
2830                          */
2831                         queue_unlock(hb);
2832                         put_futex_key(&q.key);
2833                         cond_resched();
2834                         goto retry;
2835                 default:
2836                         goto out_unlock_put_key;
2837                 }
2838         }
2839
2840         WARN_ON(!q.pi_state);
2841
2842         /*
2843          * Only actually queue now that the atomic ops are done:
2844          */
2845         __queue_me(&q, hb);
2846
2847         if (trylock) {
2848                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2849                 /* Fixup the trylock return value: */
2850                 ret = ret ? 0 : -EWOULDBLOCK;
2851                 goto no_block;
2852         }
2853
2854         rt_mutex_init_waiter(&rt_waiter);
2855
2856         /*
2857          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2858          * hold it while doing rt_mutex_start_proxy(), because then it will
2859          * include hb->lock in the blocking chain, even through we'll not in
2860          * fact hold it while blocking. This will lead it to report -EDEADLK
2861          * and BUG when futex_unlock_pi() interleaves with this.
2862          *
2863          * Therefore acquire wait_lock while holding hb->lock, but drop the
2864          * latter before calling __rt_mutex_start_proxy_lock(). This
2865          * interleaves with futex_unlock_pi() -- which does a similar lock
2866          * handoff -- such that the latter can observe the futex_q::pi_state
2867          * before __rt_mutex_start_proxy_lock() is done.
2868          */
2869         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2870         spin_unlock(q.lock_ptr);
2871         /*
2872          * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2873          * such that futex_unlock_pi() is guaranteed to observe the waiter when
2874          * it sees the futex_q::pi_state.
2875          */
2876         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2877         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2878
2879         if (ret) {
2880                 if (ret == 1)
2881                         ret = 0;
2882                 goto cleanup;
2883         }
2884
2885         if (unlikely(to))
2886                 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2887
2888         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2889
2890 cleanup:
2891         spin_lock(q.lock_ptr);
2892         /*
2893          * If we failed to acquire the lock (deadlock/signal/timeout), we must
2894          * first acquire the hb->lock before removing the lock from the
2895          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2896          * lists consistent.
2897          *
2898          * In particular; it is important that futex_unlock_pi() can not
2899          * observe this inconsistency.
2900          */
2901         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2902                 ret = 0;
2903
2904 no_block:
2905         /*
2906          * Fixup the pi_state owner and possibly acquire the lock if we
2907          * haven't already.
2908          */
2909         res = fixup_owner(uaddr, &q, !ret);
2910         /*
2911          * If fixup_owner() returned an error, proprogate that.  If it acquired
2912          * the lock, clear our -ETIMEDOUT or -EINTR.
2913          */
2914         if (res)
2915                 ret = (res < 0) ? res : 0;
2916
2917         /*
2918          * If fixup_owner() faulted and was unable to handle the fault, unlock
2919          * it and return the fault to userspace.
2920          */
2921         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2922                 pi_state = q.pi_state;
2923                 get_pi_state(pi_state);
2924         }
2925
2926         /* Unqueue and drop the lock */
2927         unqueue_me_pi(&q);
2928
2929         if (pi_state) {
2930                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2931                 put_pi_state(pi_state);
2932         }
2933
2934         goto out_put_key;
2935
2936 out_unlock_put_key:
2937         queue_unlock(hb);
2938
2939 out_put_key:
2940         put_futex_key(&q.key);
2941 out:
2942         if (to) {
2943                 hrtimer_cancel(&to->timer);
2944                 destroy_hrtimer_on_stack(&to->timer);
2945         }
2946         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2947
2948 uaddr_faulted:
2949         queue_unlock(hb);
2950
2951         ret = fault_in_user_writeable(uaddr);
2952         if (ret)
2953                 goto out_put_key;
2954
2955         if (!(flags & FLAGS_SHARED))
2956                 goto retry_private;
2957
2958         put_futex_key(&q.key);
2959         goto retry;
2960 }
2961
2962 /*
2963  * Userspace attempted a TID -> 0 atomic transition, and failed.
2964  * This is the in-kernel slowpath: we look up the PI state (if any),
2965  * and do the rt-mutex unlock.
2966  */
2967 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2968 {
2969         u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2970         union futex_key key = FUTEX_KEY_INIT;
2971         struct futex_hash_bucket *hb;
2972         struct futex_q *top_waiter;
2973         int ret;
2974
2975         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2976                 return -ENOSYS;
2977
2978 retry:
2979         if (get_user(uval, uaddr))
2980                 return -EFAULT;
2981         /*
2982          * We release only a lock we actually own:
2983          */
2984         if ((uval & FUTEX_TID_MASK) != vpid)
2985                 return -EPERM;
2986
2987         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2988         if (ret)
2989                 return ret;
2990
2991         hb = hash_futex(&key);
2992         spin_lock(&hb->lock);
2993
2994         /*
2995          * Check waiters first. We do not trust user space values at
2996          * all and we at least want to know if user space fiddled
2997          * with the futex value instead of blindly unlocking.
2998          */
2999         top_waiter = futex_top_waiter(hb, &key);
3000         if (top_waiter) {
3001                 struct futex_pi_state *pi_state = top_waiter->pi_state;
3002
3003                 ret = -EINVAL;
3004                 if (!pi_state)
3005                         goto out_unlock;
3006
3007                 /*
3008                  * If current does not own the pi_state then the futex is
3009                  * inconsistent and user space fiddled with the futex value.
3010                  */
3011                 if (pi_state->owner != current)
3012                         goto out_unlock;
3013
3014                 get_pi_state(pi_state);
3015                 /*
3016                  * By taking wait_lock while still holding hb->lock, we ensure
3017                  * there is no point where we hold neither; and therefore
3018                  * wake_futex_pi() must observe a state consistent with what we
3019                  * observed.
3020                  *
3021                  * In particular; this forces __rt_mutex_start_proxy() to
3022                  * complete such that we're guaranteed to observe the
3023                  * rt_waiter. Also see the WARN in wake_futex_pi().
3024                  */
3025                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3026                 spin_unlock(&hb->lock);
3027
3028                 /* drops pi_state->pi_mutex.wait_lock */
3029                 ret = wake_futex_pi(uaddr, uval, pi_state);
3030
3031                 put_pi_state(pi_state);
3032
3033                 /*
3034                  * Success, we're done! No tricky corner cases.
3035                  */
3036                 if (!ret)
3037                         goto out_putkey;
3038                 /*
3039                  * The atomic access to the futex value generated a
3040                  * pagefault, so retry the user-access and the wakeup:
3041                  */
3042                 if (ret == -EFAULT)
3043                         goto pi_faulted;
3044                 /*
3045                  * A unconditional UNLOCK_PI op raced against a waiter
3046                  * setting the FUTEX_WAITERS bit. Try again.
3047                  */
3048                 if (ret == -EAGAIN)
3049                         goto pi_retry;
3050                 /*
3051                  * wake_futex_pi has detected invalid state. Tell user
3052                  * space.
3053                  */
3054                 goto out_putkey;
3055         }
3056
3057         /*
3058          * We have no kernel internal state, i.e. no waiters in the
3059          * kernel. Waiters which are about to queue themselves are stuck
3060          * on hb->lock. So we can safely ignore them. We do neither
3061          * preserve the WAITERS bit not the OWNER_DIED one. We are the
3062          * owner.
3063          */
3064         if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3065                 spin_unlock(&hb->lock);
3066                 switch (ret) {
3067                 case -EFAULT:
3068                         goto pi_faulted;
3069
3070                 case -EAGAIN:
3071                         goto pi_retry;
3072
3073                 default:
3074                         WARN_ON_ONCE(1);
3075                         goto out_putkey;
3076                 }
3077         }
3078
3079         /*
3080          * If uval has changed, let user space handle it.
3081          */
3082         ret = (curval == uval) ? 0 : -EAGAIN;
3083
3084 out_unlock:
3085         spin_unlock(&hb->lock);
3086 out_putkey:
3087         put_futex_key(&key);
3088         return ret;
3089
3090 pi_retry:
3091         put_futex_key(&key);
3092         cond_resched();
3093         goto retry;
3094
3095 pi_faulted:
3096         put_futex_key(&key);
3097
3098         ret = fault_in_user_writeable(uaddr);
3099         if (!ret)
3100                 goto retry;
3101
3102         return ret;
3103 }
3104
3105 /**
3106  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3107  * @hb:         the hash_bucket futex_q was original enqueued on
3108  * @q:          the futex_q woken while waiting to be requeued
3109  * @key2:       the futex_key of the requeue target futex
3110  * @timeout:    the timeout associated with the wait (NULL if none)
3111  *
3112  * Detect if the task was woken on the initial futex as opposed to the requeue
3113  * target futex.  If so, determine if it was a timeout or a signal that caused
3114  * the wakeup and return the appropriate error code to the caller.  Must be
3115  * called with the hb lock held.
3116  *
3117  * Return:
3118  *  -  0 = no early wakeup detected;
3119  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3120  */
3121 static inline
3122 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3123                                    struct futex_q *q, union futex_key *key2,
3124                                    struct hrtimer_sleeper *timeout)
3125 {
3126         int ret = 0;
3127
3128         /*
3129          * With the hb lock held, we avoid races while we process the wakeup.
3130          * We only need to hold hb (and not hb2) to ensure atomicity as the
3131          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3132          * It can't be requeued from uaddr2 to something else since we don't
3133          * support a PI aware source futex for requeue.
3134          */
3135         if (!match_futex(&q->key, key2)) {
3136                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3137                 /*
3138                  * We were woken prior to requeue by a timeout or a signal.
3139                  * Unqueue the futex_q and determine which it was.
3140                  */
3141                 plist_del(&q->list, &hb->chain);
3142                 hb_waiters_dec(hb);
3143
3144                 /* Handle spurious wakeups gracefully */
3145                 ret = -EWOULDBLOCK;
3146                 if (timeout && !timeout->task)
3147                         ret = -ETIMEDOUT;
3148                 else if (signal_pending(current))
3149                         ret = -ERESTARTNOINTR;
3150         }
3151         return ret;
3152 }
3153
3154 /**
3155  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3156  * @uaddr:      the futex we initially wait on (non-pi)
3157  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3158  *              the same type, no requeueing from private to shared, etc.
3159  * @val:        the expected value of uaddr
3160  * @abs_time:   absolute timeout
3161  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3162  * @uaddr2:     the pi futex we will take prior to returning to user-space
3163  *
3164  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3165  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3166  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3167  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3168  * without one, the pi logic would not know which task to boost/deboost, if
3169  * there was a need to.
3170  *
3171  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3172  * via the following--
3173  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3174  * 2) wakeup on uaddr2 after a requeue
3175  * 3) signal
3176  * 4) timeout
3177  *
3178  * If 3, cleanup and return -ERESTARTNOINTR.
3179  *
3180  * If 2, we may then block on trying to take the rt_mutex and return via:
3181  * 5) successful lock
3182  * 6) signal
3183  * 7) timeout
3184  * 8) other lock acquisition failure
3185  *
3186  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3187  *
3188  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3189  *
3190  * Return:
3191  *  -  0 - On success;
3192  *  - <0 - On error
3193  */
3194 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3195                                  u32 val, ktime_t *abs_time, u32 bitset,
3196                                  u32 __user *uaddr2)
3197 {
3198         struct hrtimer_sleeper timeout, *to = NULL;
3199         struct futex_pi_state *pi_state = NULL;
3200         struct rt_mutex_waiter rt_waiter;
3201         struct futex_hash_bucket *hb;
3202         union futex_key key2 = FUTEX_KEY_INIT;
3203         struct futex_q q = futex_q_init;
3204         int res, ret;
3205
3206         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3207                 return -ENOSYS;
3208
3209         if (uaddr == uaddr2)
3210                 return -EINVAL;
3211
3212         if (!bitset)
3213                 return -EINVAL;
3214
3215         if (abs_time) {
3216                 to = &timeout;
3217                 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3218                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
3219                                       HRTIMER_MODE_ABS);
3220                 hrtimer_init_sleeper(to, current);
3221                 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3222                                              current->timer_slack_ns);
3223         }
3224
3225         /*
3226          * The waiter is allocated on our stack, manipulated by the requeue
3227          * code while we sleep on uaddr.
3228          */
3229         rt_mutex_init_waiter(&rt_waiter);
3230
3231         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3232         if (unlikely(ret != 0))
3233                 goto out;
3234
3235         q.bitset = bitset;
3236         q.rt_waiter = &rt_waiter;
3237         q.requeue_pi_key = &key2;
3238
3239         /*
3240          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3241          * count.
3242          */
3243         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3244         if (ret)
3245                 goto out_key2;
3246
3247         /*
3248          * The check above which compares uaddrs is not sufficient for
3249          * shared futexes. We need to compare the keys:
3250          */
3251         if (match_futex(&q.key, &key2)) {
3252                 queue_unlock(hb);
3253                 ret = -EINVAL;
3254                 goto out_put_keys;
3255         }
3256
3257         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3258         futex_wait_queue_me(hb, &q, to);
3259
3260         spin_lock(&hb->lock);
3261         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3262         spin_unlock(&hb->lock);
3263         if (ret)
3264                 goto out_put_keys;
3265
3266         /*
3267          * In order for us to be here, we know our q.key == key2, and since
3268          * we took the hb->lock above, we also know that futex_requeue() has
3269          * completed and we no longer have to concern ourselves with a wakeup
3270          * race with the atomic proxy lock acquisition by the requeue code. The
3271          * futex_requeue dropped our key1 reference and incremented our key2
3272          * reference count.
3273          */
3274
3275         /* Check if the requeue code acquired the second futex for us. */
3276         if (!q.rt_waiter) {
3277                 /*
3278                  * Got the lock. We might not be the anticipated owner if we
3279                  * did a lock-steal - fix up the PI-state in that case.
3280                  */
3281                 if (q.pi_state && (q.pi_state->owner != current)) {
3282                         spin_lock(q.lock_ptr);
3283                         ret = fixup_pi_state_owner(uaddr2, &q, current);
3284                         if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3285                                 pi_state = q.pi_state;
3286                                 get_pi_state(pi_state);
3287                         }
3288                         /*
3289                          * Drop the reference to the pi state which
3290                          * the requeue_pi() code acquired for us.
3291                          */
3292                         put_pi_state(q.pi_state);
3293                         spin_unlock(q.lock_ptr);
3294                 }
3295         } else {
3296                 struct rt_mutex *pi_mutex;
3297
3298                 /*
3299                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3300                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3301                  * the pi_state.
3302                  */
3303                 WARN_ON(!q.pi_state);
3304                 pi_mutex = &q.pi_state->pi_mutex;
3305                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3306
3307                 spin_lock(q.lock_ptr);
3308                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3309                         ret = 0;
3310
3311                 debug_rt_mutex_free_waiter(&rt_waiter);
3312                 /*
3313                  * Fixup the pi_state owner and possibly acquire the lock if we
3314                  * haven't already.
3315                  */
3316                 res = fixup_owner(uaddr2, &q, !ret);
3317                 /*
3318                  * If fixup_owner() returned an error, proprogate that.  If it
3319                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3320                  */
3321                 if (res)
3322                         ret = (res < 0) ? res : 0;
3323
3324                 /*
3325                  * If fixup_pi_state_owner() faulted and was unable to handle
3326                  * the fault, unlock the rt_mutex and return the fault to
3327                  * userspace.
3328                  */
3329                 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3330                         pi_state = q.pi_state;
3331                         get_pi_state(pi_state);
3332                 }
3333
3334                 /* Unqueue and drop the lock. */
3335                 unqueue_me_pi(&q);
3336         }
3337
3338         if (pi_state) {
3339                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3340                 put_pi_state(pi_state);
3341         }
3342
3343         if (ret == -EINTR) {
3344                 /*
3345                  * We've already been requeued, but cannot restart by calling
3346                  * futex_lock_pi() directly. We could restart this syscall, but
3347                  * it would detect that the user space "val" changed and return
3348                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3349                  * -EWOULDBLOCK directly.
3350                  */
3351                 ret = -EWOULDBLOCK;
3352         }
3353
3354 out_put_keys:
3355         put_futex_key(&q.key);
3356 out_key2:
3357         put_futex_key(&key2);
3358
3359 out:
3360         if (to) {
3361                 hrtimer_cancel(&to->timer);
3362                 destroy_hrtimer_on_stack(&to->timer);
3363         }
3364         return ret;
3365 }
3366
3367 /*
3368  * Support for robust futexes: the kernel cleans up held futexes at
3369  * thread exit time.
3370  *
3371  * Implementation: user-space maintains a per-thread list of locks it
3372  * is holding. Upon do_exit(), the kernel carefully walks this list,
3373  * and marks all locks that are owned by this thread with the
3374  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3375  * always manipulated with the lock held, so the list is private and
3376  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3377  * field, to allow the kernel to clean up if the thread dies after
3378  * acquiring the lock, but just before it could have added itself to
3379  * the list. There can only be one such pending lock.
3380  */
3381
3382 /**
3383  * sys_set_robust_list() - Set the robust-futex list head of a task
3384  * @head:       pointer to the list-head
3385  * @len:        length of the list-head, as userspace expects
3386  */
3387 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3388                 size_t, len)
3389 {
3390         if (!futex_cmpxchg_enabled)
3391                 return -ENOSYS;
3392         /*
3393          * The kernel knows only one size for now:
3394          */
3395         if (unlikely(len != sizeof(*head)))
3396                 return -EINVAL;
3397
3398         current->robust_list = head;
3399
3400         return 0;
3401 }
3402
3403 /**
3404  * sys_get_robust_list() - Get the robust-futex list head of a task
3405  * @pid:        pid of the process [zero for current task]
3406  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
3407  * @len_ptr:    pointer to a length field, the kernel fills in the header size
3408  */
3409 SYSCALL_DEFINE3(get_robust_list, int, pid,
3410                 struct robust_list_head __user * __user *, head_ptr,
3411                 size_t __user *, len_ptr)
3412 {
3413         struct robust_list_head __user *head;
3414         unsigned long ret;
3415         struct task_struct *p;
3416
3417         if (!futex_cmpxchg_enabled)
3418                 return -ENOSYS;
3419
3420         rcu_read_lock();
3421
3422         ret = -ESRCH;
3423         if (!pid)
3424                 p = current;
3425         else {
3426                 p = find_task_by_vpid(pid);
3427                 if (!p)
3428                         goto err_unlock;
3429         }
3430
3431         ret = -EPERM;
3432         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3433                 goto err_unlock;
3434
3435         head = p->robust_list;
3436         rcu_read_unlock();
3437
3438         if (put_user(sizeof(*head), len_ptr))
3439                 return -EFAULT;
3440         return put_user(head, head_ptr);
3441
3442 err_unlock:
3443         rcu_read_unlock();
3444
3445         return ret;
3446 }
3447
3448 /*
3449  * Process a futex-list entry, check whether it's owned by the
3450  * dying task, and do notification if so:
3451  */
3452 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3453 {
3454         u32 uval, uninitialized_var(nval), mval;
3455         int err;
3456
3457         /* Futex address must be 32bit aligned */
3458         if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3459                 return -1;
3460
3461 retry:
3462         if (get_user(uval, uaddr))
3463                 return -1;
3464
3465         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3466                 return 0;
3467
3468         /*
3469          * Ok, this dying thread is truly holding a futex
3470          * of interest. Set the OWNER_DIED bit atomically
3471          * via cmpxchg, and if the value had FUTEX_WAITERS
3472          * set, wake up a waiter (if any). (We have to do a
3473          * futex_wake() even if OWNER_DIED is already set -
3474          * to handle the rare but possible case of recursive
3475          * thread-death.) The rest of the cleanup is done in
3476          * userspace.
3477          */
3478         mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3479
3480         /*
3481          * We are not holding a lock here, but we want to have
3482          * the pagefault_disable/enable() protection because
3483          * we want to handle the fault gracefully. If the
3484          * access fails we try to fault in the futex with R/W
3485          * verification via get_user_pages. get_user() above
3486          * does not guarantee R/W access. If that fails we
3487          * give up and leave the futex locked.
3488          */
3489         if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3490                 switch (err) {
3491                 case -EFAULT:
3492                         if (fault_in_user_writeable(uaddr))
3493                                 return -1;
3494                         goto retry;
3495
3496                 case -EAGAIN:
3497                         cond_resched();
3498                         goto retry;
3499
3500                 default:
3501                         WARN_ON_ONCE(1);
3502                         return err;
3503                 }
3504         }
3505
3506         if (nval != uval)
3507                 goto retry;
3508
3509         /*
3510          * Wake robust non-PI futexes here. The wakeup of
3511          * PI futexes happens in exit_pi_state():
3512          */
3513         if (!pi && (uval & FUTEX_WAITERS))
3514                 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3515
3516         return 0;
3517 }
3518
3519 /*
3520  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3521  */
3522 static inline int fetch_robust_entry(struct robust_list __user **entry,
3523                                      struct robust_list __user * __user *head,
3524                                      unsigned int *pi)
3525 {
3526         unsigned long uentry;
3527
3528         if (get_user(uentry, (unsigned long __user *)head))
3529                 return -EFAULT;
3530
3531         *entry = (void __user *)(uentry & ~1UL);
3532         *pi = uentry & 1;
3533
3534         return 0;
3535 }
3536
3537 /*
3538  * Walk curr->robust_list (very carefully, it's a userspace list!)
3539  * and mark any locks found there dead, and notify any waiters.
3540  *
3541  * We silently return on any sign of list-walking problem.
3542  */
3543 void exit_robust_list(struct task_struct *curr)
3544 {
3545         struct robust_list_head __user *head = curr->robust_list;
3546         struct robust_list __user *entry, *next_entry, *pending;
3547         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3548         unsigned int uninitialized_var(next_pi);
3549         unsigned long futex_offset;
3550         int rc;
3551
3552         if (!futex_cmpxchg_enabled)
3553                 return;
3554
3555         /*
3556          * Fetch the list head (which was registered earlier, via
3557          * sys_set_robust_list()):
3558          */
3559         if (fetch_robust_entry(&entry, &head->list.next, &pi))
3560                 return;
3561         /*
3562          * Fetch the relative futex offset:
3563          */
3564         if (get_user(futex_offset, &head->futex_offset))
3565                 return;
3566         /*
3567          * Fetch any possibly pending lock-add first, and handle it
3568          * if it exists:
3569          */
3570         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3571                 return;
3572
3573         next_entry = NULL;      /* avoid warning with gcc */
3574         while (entry != &head->list) {
3575                 /*
3576                  * Fetch the next entry in the list before calling
3577                  * handle_futex_death:
3578                  */
3579                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3580                 /*
3581                  * A pending lock might already be on the list, so
3582                  * don't process it twice:
3583                  */
3584                 if (entry != pending)
3585                         if (handle_futex_death((void __user *)entry + futex_offset,
3586                                                 curr, pi))
3587                                 return;
3588                 if (rc)
3589                         return;
3590                 entry = next_entry;
3591                 pi = next_pi;
3592                 /*
3593                  * Avoid excessively long or circular lists:
3594                  */
3595                 if (!--limit)
3596                         break;
3597
3598                 cond_resched();
3599         }
3600
3601         if (pending)
3602                 handle_futex_death((void __user *)pending + futex_offset,
3603                                    curr, pip);
3604 }
3605
3606 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3607                 u32 __user *uaddr2, u32 val2, u32 val3)
3608 {
3609         int cmd = op & FUTEX_CMD_MASK;
3610         unsigned int flags = 0;
3611
3612         if (!(op & FUTEX_PRIVATE_FLAG))
3613                 flags |= FLAGS_SHARED;
3614
3615         if (op & FUTEX_CLOCK_REALTIME) {
3616                 flags |= FLAGS_CLOCKRT;
3617                 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3618                     cmd != FUTEX_WAIT_REQUEUE_PI)
3619                         return -ENOSYS;
3620         }
3621
3622         switch (cmd) {
3623         case FUTEX_LOCK_PI:
3624         case FUTEX_UNLOCK_PI:
3625         case FUTEX_TRYLOCK_PI:
3626         case FUTEX_WAIT_REQUEUE_PI:
3627         case FUTEX_CMP_REQUEUE_PI:
3628                 if (!futex_cmpxchg_enabled)
3629                         return -ENOSYS;
3630         }
3631
3632         switch (cmd) {
3633         case FUTEX_WAIT:
3634                 val3 = FUTEX_BITSET_MATCH_ANY;
3635                 /* fall through */
3636         case FUTEX_WAIT_BITSET:
3637                 return futex_wait(uaddr, flags, val, timeout, val3);
3638         case FUTEX_WAKE:
3639                 val3 = FUTEX_BITSET_MATCH_ANY;
3640                 /* fall through */
3641         case FUTEX_WAKE_BITSET:
3642                 return futex_wake(uaddr, flags, val, val3);
3643         case FUTEX_REQUEUE:
3644                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3645         case FUTEX_CMP_REQUEUE:
3646                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3647         case FUTEX_WAKE_OP:
3648                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3649         case FUTEX_LOCK_PI:
3650                 return futex_lock_pi(uaddr, flags, timeout, 0);
3651         case FUTEX_UNLOCK_PI:
3652                 return futex_unlock_pi(uaddr, flags);
3653         case FUTEX_TRYLOCK_PI:
3654                 return futex_lock_pi(uaddr, flags, NULL, 1);
3655         case FUTEX_WAIT_REQUEUE_PI:
3656                 val3 = FUTEX_BITSET_MATCH_ANY;
3657                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3658                                              uaddr2);
3659         case FUTEX_CMP_REQUEUE_PI:
3660                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3661         }
3662         return -ENOSYS;
3663 }
3664
3665
3666 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3667                 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3668                 u32, val3)
3669 {
3670         struct timespec64 ts;
3671         ktime_t t, *tp = NULL;
3672         u32 val2 = 0;
3673         int cmd = op & FUTEX_CMD_MASK;
3674
3675         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3676                       cmd == FUTEX_WAIT_BITSET ||
3677                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3678                 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3679                         return -EFAULT;
3680                 if (get_timespec64(&ts, utime))
3681                         return -EFAULT;
3682                 if (!timespec64_valid(&ts))
3683                         return -EINVAL;
3684
3685                 t = timespec64_to_ktime(ts);
3686                 if (cmd == FUTEX_WAIT)
3687                         t = ktime_add_safe(ktime_get(), t);
3688                 tp = &t;
3689         }
3690         /*
3691          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3692          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3693          */
3694         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3695             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3696                 val2 = (u32) (unsigned long) utime;
3697
3698         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3699 }
3700
3701 #ifdef CONFIG_COMPAT
3702 /*
3703  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3704  */
3705 static inline int
3706 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3707                    compat_uptr_t __user *head, unsigned int *pi)
3708 {
3709         if (get_user(*uentry, head))
3710                 return -EFAULT;
3711
3712         *entry = compat_ptr((*uentry) & ~1);
3713         *pi = (unsigned int)(*uentry) & 1;
3714
3715         return 0;
3716 }
3717
3718 static void __user *futex_uaddr(struct robust_list __user *entry,
3719                                 compat_long_t futex_offset)
3720 {
3721         compat_uptr_t base = ptr_to_compat(entry);
3722         void __user *uaddr = compat_ptr(base + futex_offset);
3723
3724         return uaddr;
3725 }
3726
3727 /*
3728  * Walk curr->robust_list (very carefully, it's a userspace list!)
3729  * and mark any locks found there dead, and notify any waiters.
3730  *
3731  * We silently return on any sign of list-walking problem.
3732  */
3733 void compat_exit_robust_list(struct task_struct *curr)
3734 {
3735         struct compat_robust_list_head __user *head = curr->compat_robust_list;
3736         struct robust_list __user *entry, *next_entry, *pending;
3737         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3738         unsigned int uninitialized_var(next_pi);
3739         compat_uptr_t uentry, next_uentry, upending;
3740         compat_long_t futex_offset;
3741         int rc;
3742
3743         if (!futex_cmpxchg_enabled)
3744                 return;
3745
3746         /*
3747          * Fetch the list head (which was registered earlier, via
3748          * sys_set_robust_list()):
3749          */
3750         if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3751                 return;
3752         /*
3753          * Fetch the relative futex offset:
3754          */
3755         if (get_user(futex_offset, &head->futex_offset))
3756                 return;
3757         /*
3758          * Fetch any possibly pending lock-add first, and handle it
3759          * if it exists:
3760          */
3761         if (compat_fetch_robust_entry(&upending, &pending,
3762                                &head->list_op_pending, &pip))
3763                 return;
3764
3765         next_entry = NULL;      /* avoid warning with gcc */
3766         while (entry != (struct robust_list __user *) &head->list) {
3767                 /*
3768                  * Fetch the next entry in the list before calling
3769                  * handle_futex_death:
3770                  */
3771                 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3772                         (compat_uptr_t __user *)&entry->next, &next_pi);
3773                 /*
3774                  * A pending lock might already be on the list, so
3775                  * dont process it twice:
3776                  */
3777                 if (entry != pending) {
3778                         void __user *uaddr = futex_uaddr(entry, futex_offset);
3779
3780                         if (handle_futex_death(uaddr, curr, pi))
3781                                 return;
3782                 }
3783                 if (rc)
3784                         return;
3785                 uentry = next_uentry;
3786                 entry = next_entry;
3787                 pi = next_pi;
3788                 /*
3789                  * Avoid excessively long or circular lists:
3790                  */
3791                 if (!--limit)
3792                         break;
3793
3794                 cond_resched();
3795         }
3796         if (pending) {
3797                 void __user *uaddr = futex_uaddr(pending, futex_offset);
3798
3799                 handle_futex_death(uaddr, curr, pip);
3800         }
3801 }
3802
3803 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3804                 struct compat_robust_list_head __user *, head,
3805                 compat_size_t, len)
3806 {
3807         if (!futex_cmpxchg_enabled)
3808                 return -ENOSYS;
3809
3810         if (unlikely(len != sizeof(*head)))
3811                 return -EINVAL;
3812
3813         current->compat_robust_list = head;
3814
3815         return 0;
3816 }
3817
3818 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3819                         compat_uptr_t __user *, head_ptr,
3820                         compat_size_t __user *, len_ptr)
3821 {
3822         struct compat_robust_list_head __user *head;
3823         unsigned long ret;
3824         struct task_struct *p;
3825
3826         if (!futex_cmpxchg_enabled)
3827                 return -ENOSYS;
3828
3829         rcu_read_lock();
3830
3831         ret = -ESRCH;
3832         if (!pid)
3833                 p = current;
3834         else {
3835                 p = find_task_by_vpid(pid);
3836                 if (!p)
3837                         goto err_unlock;
3838         }
3839
3840         ret = -EPERM;
3841         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3842                 goto err_unlock;
3843
3844         head = p->compat_robust_list;
3845         rcu_read_unlock();
3846
3847         if (put_user(sizeof(*head), len_ptr))
3848                 return -EFAULT;
3849         return put_user(ptr_to_compat(head), head_ptr);
3850
3851 err_unlock:
3852         rcu_read_unlock();
3853
3854         return ret;
3855 }
3856 #endif /* CONFIG_COMPAT */
3857
3858 #ifdef CONFIG_COMPAT_32BIT_TIME
3859 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3860                 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3861                 u32, val3)
3862 {
3863         struct timespec64 ts;
3864         ktime_t t, *tp = NULL;
3865         int val2 = 0;
3866         int cmd = op & FUTEX_CMD_MASK;
3867
3868         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3869                       cmd == FUTEX_WAIT_BITSET ||
3870                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3871                 if (get_old_timespec32(&ts, utime))
3872                         return -EFAULT;
3873                 if (!timespec64_valid(&ts))
3874                         return -EINVAL;
3875
3876                 t = timespec64_to_ktime(ts);
3877                 if (cmd == FUTEX_WAIT)
3878                         t = ktime_add_safe(ktime_get(), t);
3879                 tp = &t;
3880         }
3881         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3882             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3883                 val2 = (int) (unsigned long) utime;
3884
3885         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3886 }
3887 #endif /* CONFIG_COMPAT_32BIT_TIME */
3888
3889 static void __init futex_detect_cmpxchg(void)
3890 {
3891 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3892         u32 curval;
3893
3894         /*
3895          * This will fail and we want it. Some arch implementations do
3896          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3897          * functionality. We want to know that before we call in any
3898          * of the complex code paths. Also we want to prevent
3899          * registration of robust lists in that case. NULL is
3900          * guaranteed to fault and we get -EFAULT on functional
3901          * implementation, the non-functional ones will return
3902          * -ENOSYS.
3903          */
3904         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3905                 futex_cmpxchg_enabled = 1;
3906 #endif
3907 }
3908
3909 static int __init futex_init(void)
3910 {
3911         unsigned int futex_shift;
3912         unsigned long i;
3913
3914 #if CONFIG_BASE_SMALL
3915         futex_hashsize = 16;
3916 #else
3917         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3918 #endif
3919
3920         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3921                                                futex_hashsize, 0,
3922                                                futex_hashsize < 256 ? HASH_SMALL : 0,
3923                                                &futex_shift, NULL,
3924                                                futex_hashsize, futex_hashsize);
3925         futex_hashsize = 1UL << futex_shift;
3926
3927         futex_detect_cmpxchg();
3928
3929         for (i = 0; i < futex_hashsize; i++) {
3930                 atomic_set(&futex_queues[i].waiters, 0);
3931                 plist_head_init(&futex_queues[i].chain);
3932                 spin_lock_init(&futex_queues[i].lock);
3933         }
3934
3935         return 0;
3936 }
3937 core_initcall(futex_init);