ubsan: disable UBSAN_TRAP for all*config
[linux-2.6-microblaze.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/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/syscalls.h>
38 #include <linux/hugetlb.h>
39 #include <linux/freezer.h>
40 #include <linux/memblock.h>
41 #include <linux/fault-inject.h>
42 #include <linux/time_namespace.h>
43
44 #include <asm/futex.h>
45
46 #include "locking/rtmutex_common.h"
47
48 /*
49  * READ this before attempting to hack on futexes!
50  *
51  * Basic futex operation and ordering guarantees
52  * =============================================
53  *
54  * The waiter reads the futex value in user space and calls
55  * futex_wait(). This function computes the hash bucket and acquires
56  * the hash bucket lock. After that it reads the futex user space value
57  * again and verifies that the data has not changed. If it has not changed
58  * it enqueues itself into the hash bucket, releases the hash bucket lock
59  * and schedules.
60  *
61  * The waker side modifies the user space value of the futex and calls
62  * futex_wake(). This function computes the hash bucket and acquires the
63  * hash bucket lock. Then it looks for waiters on that futex in the hash
64  * bucket and wakes them.
65  *
66  * In futex wake up scenarios where no tasks are blocked on a futex, taking
67  * the hb spinlock can be avoided and simply return. In order for this
68  * optimization to work, ordering guarantees must exist so that the waiter
69  * being added to the list is acknowledged when the list is concurrently being
70  * checked by the waker, avoiding scenarios like the following:
71  *
72  * CPU 0                               CPU 1
73  * val = *futex;
74  * sys_futex(WAIT, futex, val);
75  *   futex_wait(futex, val);
76  *   uval = *futex;
77  *                                     *futex = newval;
78  *                                     sys_futex(WAKE, futex);
79  *                                       futex_wake(futex);
80  *                                       if (queue_empty())
81  *                                         return;
82  *   if (uval == val)
83  *      lock(hash_bucket(futex));
84  *      queue();
85  *     unlock(hash_bucket(futex));
86  *     schedule();
87  *
88  * This would cause the waiter on CPU 0 to wait forever because it
89  * missed the transition of the user space value from val to newval
90  * and the waker did not find the waiter in the hash bucket queue.
91  *
92  * The correct serialization ensures that a waiter either observes
93  * the changed user space value before blocking or is woken by a
94  * concurrent waker:
95  *
96  * CPU 0                                 CPU 1
97  * val = *futex;
98  * sys_futex(WAIT, futex, val);
99  *   futex_wait(futex, val);
100  *
101  *   waiters++; (a)
102  *   smp_mb(); (A) <-- paired with -.
103  *                                  |
104  *   lock(hash_bucket(futex));      |
105  *                                  |
106  *   uval = *futex;                 |
107  *                                  |        *futex = newval;
108  *                                  |        sys_futex(WAKE, futex);
109  *                                  |          futex_wake(futex);
110  *                                  |
111  *                                  `--------> smp_mb(); (B)
112  *   if (uval == val)
113  *     queue();
114  *     unlock(hash_bucket(futex));
115  *     schedule();                         if (waiters)
116  *                                           lock(hash_bucket(futex));
117  *   else                                    wake_waiters(futex);
118  *     waiters--; (b)                        unlock(hash_bucket(futex));
119  *
120  * Where (A) orders the waiters increment and the futex value read through
121  * atomic operations (see hb_waiters_inc) and where (B) orders the write
122  * to futex and the waiters read (see hb_waiters_pending()).
123  *
124  * This yields the following case (where X:=waiters, Y:=futex):
125  *
126  *      X = Y = 0
127  *
128  *      w[X]=1          w[Y]=1
129  *      MB              MB
130  *      r[Y]=y          r[X]=x
131  *
132  * Which guarantees that x==0 && y==0 is impossible; which translates back into
133  * the guarantee that we cannot both miss the futex variable change and the
134  * enqueue.
135  *
136  * Note that a new waiter is accounted for in (a) even when it is possible that
137  * the wait call can return error, in which case we backtrack from it in (b).
138  * Refer to the comment in queue_lock().
139  *
140  * Similarly, in order to account for waiters being requeued on another
141  * address we always increment the waiters for the destination bucket before
142  * acquiring the lock. It then decrements them again  after releasing it -
143  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
144  * will do the additional required waiter count housekeeping. This is done for
145  * double_lock_hb() and double_unlock_hb(), respectively.
146  */
147
148 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
149 #define futex_cmpxchg_enabled 1
150 #else
151 static int  __read_mostly futex_cmpxchg_enabled;
152 #endif
153
154 /*
155  * Futex flags used to encode options to functions and preserve them across
156  * restarts.
157  */
158 #ifdef CONFIG_MMU
159 # define FLAGS_SHARED           0x01
160 #else
161 /*
162  * NOMMU does not have per process address space. Let the compiler optimize
163  * code away.
164  */
165 # define FLAGS_SHARED           0x00
166 #endif
167 #define FLAGS_CLOCKRT           0x02
168 #define FLAGS_HAS_TIMEOUT       0x04
169
170 /*
171  * Priority Inheritance state:
172  */
173 struct futex_pi_state {
174         /*
175          * list of 'owned' pi_state instances - these have to be
176          * cleaned up in do_exit() if the task exits prematurely:
177          */
178         struct list_head list;
179
180         /*
181          * The PI object:
182          */
183         struct rt_mutex pi_mutex;
184
185         struct task_struct *owner;
186         refcount_t refcount;
187
188         union futex_key key;
189 } __randomize_layout;
190
191 /**
192  * struct futex_q - The hashed futex queue entry, one per waiting task
193  * @list:               priority-sorted list of tasks waiting on this futex
194  * @task:               the task waiting on the futex
195  * @lock_ptr:           the hash bucket lock
196  * @key:                the key the futex is hashed on
197  * @pi_state:           optional priority inheritance state
198  * @rt_waiter:          rt_waiter storage for use with requeue_pi
199  * @requeue_pi_key:     the requeue_pi target futex key
200  * @bitset:             bitset for the optional bitmasked wakeup
201  *
202  * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
203  * we can wake only the relevant ones (hashed queues may be shared).
204  *
205  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
206  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
207  * The order of wakeup is always to make the first condition true, then
208  * the second.
209  *
210  * PI futexes are typically woken before they are removed from the hash list via
211  * the rt_mutex code. See unqueue_me_pi().
212  */
213 struct futex_q {
214         struct plist_node list;
215
216         struct task_struct *task;
217         spinlock_t *lock_ptr;
218         union futex_key key;
219         struct futex_pi_state *pi_state;
220         struct rt_mutex_waiter *rt_waiter;
221         union futex_key *requeue_pi_key;
222         u32 bitset;
223 } __randomize_layout;
224
225 static const struct futex_q futex_q_init = {
226         /* list gets initialized in queue_me()*/
227         .key = FUTEX_KEY_INIT,
228         .bitset = FUTEX_BITSET_MATCH_ANY
229 };
230
231 /*
232  * Hash buckets are shared by all the futex_keys that hash to the same
233  * location.  Each key may have multiple futex_q structures, one for each task
234  * waiting on a futex.
235  */
236 struct futex_hash_bucket {
237         atomic_t waiters;
238         spinlock_t lock;
239         struct plist_head chain;
240 } ____cacheline_aligned_in_smp;
241
242 /*
243  * The base of the bucket array and its size are always used together
244  * (after initialization only in hash_futex()), so ensure that they
245  * reside in the same cacheline.
246  */
247 static struct {
248         struct futex_hash_bucket *queues;
249         unsigned long            hashsize;
250 } __futex_data __read_mostly __aligned(2*sizeof(long));
251 #define futex_queues   (__futex_data.queues)
252 #define futex_hashsize (__futex_data.hashsize)
253
254
255 /*
256  * Fault injections for futexes.
257  */
258 #ifdef CONFIG_FAIL_FUTEX
259
260 static struct {
261         struct fault_attr attr;
262
263         bool ignore_private;
264 } fail_futex = {
265         .attr = FAULT_ATTR_INITIALIZER,
266         .ignore_private = false,
267 };
268
269 static int __init setup_fail_futex(char *str)
270 {
271         return setup_fault_attr(&fail_futex.attr, str);
272 }
273 __setup("fail_futex=", setup_fail_futex);
274
275 static bool should_fail_futex(bool fshared)
276 {
277         if (fail_futex.ignore_private && !fshared)
278                 return false;
279
280         return should_fail(&fail_futex.attr, 1);
281 }
282
283 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
284
285 static int __init fail_futex_debugfs(void)
286 {
287         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
288         struct dentry *dir;
289
290         dir = fault_create_debugfs_attr("fail_futex", NULL,
291                                         &fail_futex.attr);
292         if (IS_ERR(dir))
293                 return PTR_ERR(dir);
294
295         debugfs_create_bool("ignore-private", mode, dir,
296                             &fail_futex.ignore_private);
297         return 0;
298 }
299
300 late_initcall(fail_futex_debugfs);
301
302 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
303
304 #else
305 static inline bool should_fail_futex(bool fshared)
306 {
307         return false;
308 }
309 #endif /* CONFIG_FAIL_FUTEX */
310
311 #ifdef CONFIG_COMPAT
312 static void compat_exit_robust_list(struct task_struct *curr);
313 #endif
314
315 /*
316  * Reflects a new waiter being added to the waitqueue.
317  */
318 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
319 {
320 #ifdef CONFIG_SMP
321         atomic_inc(&hb->waiters);
322         /*
323          * Full barrier (A), see the ordering comment above.
324          */
325         smp_mb__after_atomic();
326 #endif
327 }
328
329 /*
330  * Reflects a waiter being removed from the waitqueue by wakeup
331  * paths.
332  */
333 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
334 {
335 #ifdef CONFIG_SMP
336         atomic_dec(&hb->waiters);
337 #endif
338 }
339
340 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
341 {
342 #ifdef CONFIG_SMP
343         /*
344          * Full barrier (B), see the ordering comment above.
345          */
346         smp_mb();
347         return atomic_read(&hb->waiters);
348 #else
349         return 1;
350 #endif
351 }
352
353 /**
354  * hash_futex - Return the hash bucket in the global hash
355  * @key:        Pointer to the futex key for which the hash is calculated
356  *
357  * We hash on the keys returned from get_futex_key (see below) and return the
358  * corresponding hash bucket in the global hash.
359  */
360 static struct futex_hash_bucket *hash_futex(union futex_key *key)
361 {
362         u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
363                           key->both.offset);
364
365         return &futex_queues[hash & (futex_hashsize - 1)];
366 }
367
368
369 /**
370  * match_futex - Check whether two futex keys are equal
371  * @key1:       Pointer to key1
372  * @key2:       Pointer to key2
373  *
374  * Return 1 if two futex_keys are equal, 0 otherwise.
375  */
376 static inline int match_futex(union futex_key *key1, union futex_key *key2)
377 {
378         return (key1 && key2
379                 && key1->both.word == key2->both.word
380                 && key1->both.ptr == key2->both.ptr
381                 && key1->both.offset == key2->both.offset);
382 }
383
384 enum futex_access {
385         FUTEX_READ,
386         FUTEX_WRITE
387 };
388
389 /**
390  * futex_setup_timer - set up the sleeping hrtimer.
391  * @time:       ptr to the given timeout value
392  * @timeout:    the hrtimer_sleeper structure to be set up
393  * @flags:      futex flags
394  * @range_ns:   optional range in ns
395  *
396  * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
397  *         value given
398  */
399 static inline struct hrtimer_sleeper *
400 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
401                   int flags, u64 range_ns)
402 {
403         if (!time)
404                 return NULL;
405
406         hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
407                                       CLOCK_REALTIME : CLOCK_MONOTONIC,
408                                       HRTIMER_MODE_ABS);
409         /*
410          * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
411          * effectively the same as calling hrtimer_set_expires().
412          */
413         hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
414
415         return timeout;
416 }
417
418 /*
419  * Generate a machine wide unique identifier for this inode.
420  *
421  * This relies on u64 not wrapping in the life-time of the machine; which with
422  * 1ns resolution means almost 585 years.
423  *
424  * This further relies on the fact that a well formed program will not unmap
425  * the file while it has a (shared) futex waiting on it. This mapping will have
426  * a file reference which pins the mount and inode.
427  *
428  * If for some reason an inode gets evicted and read back in again, it will get
429  * a new sequence number and will _NOT_ match, even though it is the exact same
430  * file.
431  *
432  * It is important that match_futex() will never have a false-positive, esp.
433  * for PI futexes that can mess up the state. The above argues that false-negatives
434  * are only possible for malformed programs.
435  */
436 static u64 get_inode_sequence_number(struct inode *inode)
437 {
438         static atomic64_t i_seq;
439         u64 old;
440
441         /* Does the inode already have a sequence number? */
442         old = atomic64_read(&inode->i_sequence);
443         if (likely(old))
444                 return old;
445
446         for (;;) {
447                 u64 new = atomic64_add_return(1, &i_seq);
448                 if (WARN_ON_ONCE(!new))
449                         continue;
450
451                 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
452                 if (old)
453                         return old;
454                 return new;
455         }
456 }
457
458 /**
459  * get_futex_key() - Get parameters which are the keys for a futex
460  * @uaddr:      virtual address of the futex
461  * @fshared:    false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
462  * @key:        address where result is stored.
463  * @rw:         mapping needs to be read/write (values: FUTEX_READ,
464  *              FUTEX_WRITE)
465  *
466  * Return: a negative error code or 0
467  *
468  * The key words are stored in @key on success.
469  *
470  * For shared mappings (when @fshared), the key is:
471  *
472  *   ( inode->i_sequence, page->index, offset_within_page )
473  *
474  * [ also see get_inode_sequence_number() ]
475  *
476  * For private mappings (or when !@fshared), the key is:
477  *
478  *   ( current->mm, address, 0 )
479  *
480  * This allows (cross process, where applicable) identification of the futex
481  * without keeping the page pinned for the duration of the FUTEX_WAIT.
482  *
483  * lock_page() might sleep, the caller should not hold a spinlock.
484  */
485 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
486                          enum futex_access rw)
487 {
488         unsigned long address = (unsigned long)uaddr;
489         struct mm_struct *mm = current->mm;
490         struct page *page, *tail;
491         struct address_space *mapping;
492         int err, ro = 0;
493
494         /*
495          * The futex address must be "naturally" aligned.
496          */
497         key->both.offset = address % PAGE_SIZE;
498         if (unlikely((address % sizeof(u32)) != 0))
499                 return -EINVAL;
500         address -= key->both.offset;
501
502         if (unlikely(!access_ok(uaddr, sizeof(u32))))
503                 return -EFAULT;
504
505         if (unlikely(should_fail_futex(fshared)))
506                 return -EFAULT;
507
508         /*
509          * PROCESS_PRIVATE futexes are fast.
510          * As the mm cannot disappear under us and the 'key' only needs
511          * virtual address, we dont even have to find the underlying vma.
512          * Note : We do have to check 'uaddr' is a valid user address,
513          *        but access_ok() should be faster than find_vma()
514          */
515         if (!fshared) {
516                 key->private.mm = mm;
517                 key->private.address = address;
518                 return 0;
519         }
520
521 again:
522         /* Ignore any VERIFY_READ mapping (futex common case) */
523         if (unlikely(should_fail_futex(true)))
524                 return -EFAULT;
525
526         err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
527         /*
528          * If write access is not required (eg. FUTEX_WAIT), try
529          * and get read-only access.
530          */
531         if (err == -EFAULT && rw == FUTEX_READ) {
532                 err = get_user_pages_fast(address, 1, 0, &page);
533                 ro = 1;
534         }
535         if (err < 0)
536                 return err;
537         else
538                 err = 0;
539
540         /*
541          * The treatment of mapping from this point on is critical. The page
542          * lock protects many things but in this context the page lock
543          * stabilizes mapping, prevents inode freeing in the shared
544          * file-backed region case and guards against movement to swap cache.
545          *
546          * Strictly speaking the page lock is not needed in all cases being
547          * considered here and page lock forces unnecessarily serialization
548          * From this point on, mapping will be re-verified if necessary and
549          * page lock will be acquired only if it is unavoidable
550          *
551          * Mapping checks require the head page for any compound page so the
552          * head page and mapping is looked up now. For anonymous pages, it
553          * does not matter if the page splits in the future as the key is
554          * based on the address. For filesystem-backed pages, the tail is
555          * required as the index of the page determines the key. For
556          * base pages, there is no tail page and tail == page.
557          */
558         tail = page;
559         page = compound_head(page);
560         mapping = READ_ONCE(page->mapping);
561
562         /*
563          * If page->mapping is NULL, then it cannot be a PageAnon
564          * page; but it might be the ZERO_PAGE or in the gate area or
565          * in a special mapping (all cases which we are happy to fail);
566          * or it may have been a good file page when get_user_pages_fast
567          * found it, but truncated or holepunched or subjected to
568          * invalidate_complete_page2 before we got the page lock (also
569          * cases which we are happy to fail).  And we hold a reference,
570          * so refcount care in invalidate_complete_page's remove_mapping
571          * prevents drop_caches from setting mapping to NULL beneath us.
572          *
573          * The case we do have to guard against is when memory pressure made
574          * shmem_writepage move it from filecache to swapcache beneath us:
575          * an unlikely race, but we do need to retry for page->mapping.
576          */
577         if (unlikely(!mapping)) {
578                 int shmem_swizzled;
579
580                 /*
581                  * Page lock is required to identify which special case above
582                  * applies. If this is really a shmem page then the page lock
583                  * will prevent unexpected transitions.
584                  */
585                 lock_page(page);
586                 shmem_swizzled = PageSwapCache(page) || page->mapping;
587                 unlock_page(page);
588                 put_page(page);
589
590                 if (shmem_swizzled)
591                         goto again;
592
593                 return -EFAULT;
594         }
595
596         /*
597          * Private mappings are handled in a simple way.
598          *
599          * If the futex key is stored on an anonymous page, then the associated
600          * object is the mm which is implicitly pinned by the calling process.
601          *
602          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
603          * it's a read-only handle, it's expected that futexes attach to
604          * the object not the particular process.
605          */
606         if (PageAnon(page)) {
607                 /*
608                  * A RO anonymous page will never change and thus doesn't make
609                  * sense for futex operations.
610                  */
611                 if (unlikely(should_fail_futex(true)) || ro) {
612                         err = -EFAULT;
613                         goto out;
614                 }
615
616                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
617                 key->private.mm = mm;
618                 key->private.address = address;
619
620         } else {
621                 struct inode *inode;
622
623                 /*
624                  * The associated futex object in this case is the inode and
625                  * the page->mapping must be traversed. Ordinarily this should
626                  * be stabilised under page lock but it's not strictly
627                  * necessary in this case as we just want to pin the inode, not
628                  * update the radix tree or anything like that.
629                  *
630                  * The RCU read lock is taken as the inode is finally freed
631                  * under RCU. If the mapping still matches expectations then the
632                  * mapping->host can be safely accessed as being a valid inode.
633                  */
634                 rcu_read_lock();
635
636                 if (READ_ONCE(page->mapping) != mapping) {
637                         rcu_read_unlock();
638                         put_page(page);
639
640                         goto again;
641                 }
642
643                 inode = READ_ONCE(mapping->host);
644                 if (!inode) {
645                         rcu_read_unlock();
646                         put_page(page);
647
648                         goto again;
649                 }
650
651                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
652                 key->shared.i_seq = get_inode_sequence_number(inode);
653                 key->shared.pgoff = basepage_index(tail);
654                 rcu_read_unlock();
655         }
656
657 out:
658         put_page(page);
659         return err;
660 }
661
662 /**
663  * fault_in_user_writeable() - Fault in user address and verify RW access
664  * @uaddr:      pointer to faulting user space address
665  *
666  * Slow path to fixup the fault we just took in the atomic write
667  * access to @uaddr.
668  *
669  * We have no generic implementation of a non-destructive write to the
670  * user address. We know that we faulted in the atomic pagefault
671  * disabled section so we can as well avoid the #PF overhead by
672  * calling get_user_pages() right away.
673  */
674 static int fault_in_user_writeable(u32 __user *uaddr)
675 {
676         struct mm_struct *mm = current->mm;
677         int ret;
678
679         mmap_read_lock(mm);
680         ret = fixup_user_fault(mm, (unsigned long)uaddr,
681                                FAULT_FLAG_WRITE, NULL);
682         mmap_read_unlock(mm);
683
684         return ret < 0 ? ret : 0;
685 }
686
687 /**
688  * futex_top_waiter() - Return the highest priority waiter on a futex
689  * @hb:         the hash bucket the futex_q's reside in
690  * @key:        the futex key (to distinguish it from other futex futex_q's)
691  *
692  * Must be called with the hb lock held.
693  */
694 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
695                                         union futex_key *key)
696 {
697         struct futex_q *this;
698
699         plist_for_each_entry(this, &hb->chain, list) {
700                 if (match_futex(&this->key, key))
701                         return this;
702         }
703         return NULL;
704 }
705
706 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
707                                       u32 uval, u32 newval)
708 {
709         int ret;
710
711         pagefault_disable();
712         ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
713         pagefault_enable();
714
715         return ret;
716 }
717
718 static int get_futex_value_locked(u32 *dest, u32 __user *from)
719 {
720         int ret;
721
722         pagefault_disable();
723         ret = __get_user(*dest, from);
724         pagefault_enable();
725
726         return ret ? -EFAULT : 0;
727 }
728
729
730 /*
731  * PI code:
732  */
733 static int refill_pi_state_cache(void)
734 {
735         struct futex_pi_state *pi_state;
736
737         if (likely(current->pi_state_cache))
738                 return 0;
739
740         pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
741
742         if (!pi_state)
743                 return -ENOMEM;
744
745         INIT_LIST_HEAD(&pi_state->list);
746         /* pi_mutex gets initialized later */
747         pi_state->owner = NULL;
748         refcount_set(&pi_state->refcount, 1);
749         pi_state->key = FUTEX_KEY_INIT;
750
751         current->pi_state_cache = pi_state;
752
753         return 0;
754 }
755
756 static struct futex_pi_state *alloc_pi_state(void)
757 {
758         struct futex_pi_state *pi_state = current->pi_state_cache;
759
760         WARN_ON(!pi_state);
761         current->pi_state_cache = NULL;
762
763         return pi_state;
764 }
765
766 static void get_pi_state(struct futex_pi_state *pi_state)
767 {
768         WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
769 }
770
771 /*
772  * Drops a reference to the pi_state object and frees or caches it
773  * when the last reference is gone.
774  */
775 static void put_pi_state(struct futex_pi_state *pi_state)
776 {
777         if (!pi_state)
778                 return;
779
780         if (!refcount_dec_and_test(&pi_state->refcount))
781                 return;
782
783         /*
784          * If pi_state->owner is NULL, the owner is most probably dying
785          * and has cleaned up the pi_state already
786          */
787         if (pi_state->owner) {
788                 struct task_struct *owner;
789                 unsigned long flags;
790
791                 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
792                 owner = pi_state->owner;
793                 if (owner) {
794                         raw_spin_lock(&owner->pi_lock);
795                         list_del_init(&pi_state->list);
796                         raw_spin_unlock(&owner->pi_lock);
797                 }
798                 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
799                 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
800         }
801
802         if (current->pi_state_cache) {
803                 kfree(pi_state);
804         } else {
805                 /*
806                  * pi_state->list is already empty.
807                  * clear pi_state->owner.
808                  * refcount is at 0 - put it back to 1.
809                  */
810                 pi_state->owner = NULL;
811                 refcount_set(&pi_state->refcount, 1);
812                 current->pi_state_cache = pi_state;
813         }
814 }
815
816 #ifdef CONFIG_FUTEX_PI
817
818 /*
819  * This task is holding PI mutexes at exit time => bad.
820  * Kernel cleans up PI-state, but userspace is likely hosed.
821  * (Robust-futex cleanup is separate and might save the day for userspace.)
822  */
823 static void exit_pi_state_list(struct task_struct *curr)
824 {
825         struct list_head *next, *head = &curr->pi_state_list;
826         struct futex_pi_state *pi_state;
827         struct futex_hash_bucket *hb;
828         union futex_key key = FUTEX_KEY_INIT;
829
830         if (!futex_cmpxchg_enabled)
831                 return;
832         /*
833          * We are a ZOMBIE and nobody can enqueue itself on
834          * pi_state_list anymore, but we have to be careful
835          * versus waiters unqueueing themselves:
836          */
837         raw_spin_lock_irq(&curr->pi_lock);
838         while (!list_empty(head)) {
839                 next = head->next;
840                 pi_state = list_entry(next, struct futex_pi_state, list);
841                 key = pi_state->key;
842                 hb = hash_futex(&key);
843
844                 /*
845                  * We can race against put_pi_state() removing itself from the
846                  * list (a waiter going away). put_pi_state() will first
847                  * decrement the reference count and then modify the list, so
848                  * its possible to see the list entry but fail this reference
849                  * acquire.
850                  *
851                  * In that case; drop the locks to let put_pi_state() make
852                  * progress and retry the loop.
853                  */
854                 if (!refcount_inc_not_zero(&pi_state->refcount)) {
855                         raw_spin_unlock_irq(&curr->pi_lock);
856                         cpu_relax();
857                         raw_spin_lock_irq(&curr->pi_lock);
858                         continue;
859                 }
860                 raw_spin_unlock_irq(&curr->pi_lock);
861
862                 spin_lock(&hb->lock);
863                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
864                 raw_spin_lock(&curr->pi_lock);
865                 /*
866                  * We dropped the pi-lock, so re-check whether this
867                  * task still owns the PI-state:
868                  */
869                 if (head->next != next) {
870                         /* retain curr->pi_lock for the loop invariant */
871                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
872                         spin_unlock(&hb->lock);
873                         put_pi_state(pi_state);
874                         continue;
875                 }
876
877                 WARN_ON(pi_state->owner != curr);
878                 WARN_ON(list_empty(&pi_state->list));
879                 list_del_init(&pi_state->list);
880                 pi_state->owner = NULL;
881
882                 raw_spin_unlock(&curr->pi_lock);
883                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
884                 spin_unlock(&hb->lock);
885
886                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
887                 put_pi_state(pi_state);
888
889                 raw_spin_lock_irq(&curr->pi_lock);
890         }
891         raw_spin_unlock_irq(&curr->pi_lock);
892 }
893 #else
894 static inline void exit_pi_state_list(struct task_struct *curr) { }
895 #endif
896
897 /*
898  * We need to check the following states:
899  *
900  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
901  *
902  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
903  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
904  *
905  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
906  *
907  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
908  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
909  *
910  * [6]  Found  | Found    | task      | 0         | 1      | Valid
911  *
912  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
913  *
914  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
915  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
916  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
917  *
918  * [1]  Indicates that the kernel can acquire the futex atomically. We
919  *      came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
920  *
921  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
922  *      thread is found then it indicates that the owner TID has died.
923  *
924  * [3]  Invalid. The waiter is queued on a non PI futex
925  *
926  * [4]  Valid state after exit_robust_list(), which sets the user space
927  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
928  *
929  * [5]  The user space value got manipulated between exit_robust_list()
930  *      and exit_pi_state_list()
931  *
932  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
933  *      the pi_state but cannot access the user space value.
934  *
935  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
936  *
937  * [8]  Owner and user space value match
938  *
939  * [9]  There is no transient state which sets the user space TID to 0
940  *      except exit_robust_list(), but this is indicated by the
941  *      FUTEX_OWNER_DIED bit. See [4]
942  *
943  * [10] There is no transient state which leaves owner and user space
944  *      TID out of sync.
945  *
946  *
947  * Serialization and lifetime rules:
948  *
949  * hb->lock:
950  *
951  *      hb -> futex_q, relation
952  *      futex_q -> pi_state, relation
953  *
954  *      (cannot be raw because hb can contain arbitrary amount
955  *       of futex_q's)
956  *
957  * pi_mutex->wait_lock:
958  *
959  *      {uval, pi_state}
960  *
961  *      (and pi_mutex 'obviously')
962  *
963  * p->pi_lock:
964  *
965  *      p->pi_state_list -> pi_state->list, relation
966  *
967  * pi_state->refcount:
968  *
969  *      pi_state lifetime
970  *
971  *
972  * Lock order:
973  *
974  *   hb->lock
975  *     pi_mutex->wait_lock
976  *       p->pi_lock
977  *
978  */
979
980 /*
981  * Validate that the existing waiter has a pi_state and sanity check
982  * the pi_state against the user space value. If correct, attach to
983  * it.
984  */
985 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
986                               struct futex_pi_state *pi_state,
987                               struct futex_pi_state **ps)
988 {
989         pid_t pid = uval & FUTEX_TID_MASK;
990         u32 uval2;
991         int ret;
992
993         /*
994          * Userspace might have messed up non-PI and PI futexes [3]
995          */
996         if (unlikely(!pi_state))
997                 return -EINVAL;
998
999         /*
1000          * We get here with hb->lock held, and having found a
1001          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1002          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1003          * which in turn means that futex_lock_pi() still has a reference on
1004          * our pi_state.
1005          *
1006          * The waiter holding a reference on @pi_state also protects against
1007          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1008          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1009          * free pi_state before we can take a reference ourselves.
1010          */
1011         WARN_ON(!refcount_read(&pi_state->refcount));
1012
1013         /*
1014          * Now that we have a pi_state, we can acquire wait_lock
1015          * and do the state validation.
1016          */
1017         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1018
1019         /*
1020          * Since {uval, pi_state} is serialized by wait_lock, and our current
1021          * uval was read without holding it, it can have changed. Verify it
1022          * still is what we expect it to be, otherwise retry the entire
1023          * operation.
1024          */
1025         if (get_futex_value_locked(&uval2, uaddr))
1026                 goto out_efault;
1027
1028         if (uval != uval2)
1029                 goto out_eagain;
1030
1031         /*
1032          * Handle the owner died case:
1033          */
1034         if (uval & FUTEX_OWNER_DIED) {
1035                 /*
1036                  * exit_pi_state_list sets owner to NULL and wakes the
1037                  * topmost waiter. The task which acquires the
1038                  * pi_state->rt_mutex will fixup owner.
1039                  */
1040                 if (!pi_state->owner) {
1041                         /*
1042                          * No pi state owner, but the user space TID
1043                          * is not 0. Inconsistent state. [5]
1044                          */
1045                         if (pid)
1046                                 goto out_einval;
1047                         /*
1048                          * Take a ref on the state and return success. [4]
1049                          */
1050                         goto out_attach;
1051                 }
1052
1053                 /*
1054                  * If TID is 0, then either the dying owner has not
1055                  * yet executed exit_pi_state_list() or some waiter
1056                  * acquired the rtmutex in the pi state, but did not
1057                  * yet fixup the TID in user space.
1058                  *
1059                  * Take a ref on the state and return success. [6]
1060                  */
1061                 if (!pid)
1062                         goto out_attach;
1063         } else {
1064                 /*
1065                  * If the owner died bit is not set, then the pi_state
1066                  * must have an owner. [7]
1067                  */
1068                 if (!pi_state->owner)
1069                         goto out_einval;
1070         }
1071
1072         /*
1073          * Bail out if user space manipulated the futex value. If pi
1074          * state exists then the owner TID must be the same as the
1075          * user space TID. [9/10]
1076          */
1077         if (pid != task_pid_vnr(pi_state->owner))
1078                 goto out_einval;
1079
1080 out_attach:
1081         get_pi_state(pi_state);
1082         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1083         *ps = pi_state;
1084         return 0;
1085
1086 out_einval:
1087         ret = -EINVAL;
1088         goto out_error;
1089
1090 out_eagain:
1091         ret = -EAGAIN;
1092         goto out_error;
1093
1094 out_efault:
1095         ret = -EFAULT;
1096         goto out_error;
1097
1098 out_error:
1099         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1100         return ret;
1101 }
1102
1103 /**
1104  * wait_for_owner_exiting - Block until the owner has exited
1105  * @ret: owner's current futex lock status
1106  * @exiting:    Pointer to the exiting task
1107  *
1108  * Caller must hold a refcount on @exiting.
1109  */
1110 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1111 {
1112         if (ret != -EBUSY) {
1113                 WARN_ON_ONCE(exiting);
1114                 return;
1115         }
1116
1117         if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1118                 return;
1119
1120         mutex_lock(&exiting->futex_exit_mutex);
1121         /*
1122          * No point in doing state checking here. If the waiter got here
1123          * while the task was in exec()->exec_futex_release() then it can
1124          * have any FUTEX_STATE_* value when the waiter has acquired the
1125          * mutex. OK, if running, EXITING or DEAD if it reached exit()
1126          * already. Highly unlikely and not a problem. Just one more round
1127          * through the futex maze.
1128          */
1129         mutex_unlock(&exiting->futex_exit_mutex);
1130
1131         put_task_struct(exiting);
1132 }
1133
1134 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1135                             struct task_struct *tsk)
1136 {
1137         u32 uval2;
1138
1139         /*
1140          * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1141          * caller that the alleged owner is busy.
1142          */
1143         if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1144                 return -EBUSY;
1145
1146         /*
1147          * Reread the user space value to handle the following situation:
1148          *
1149          * CPU0                         CPU1
1150          *
1151          * sys_exit()                   sys_futex()
1152          *  do_exit()                    futex_lock_pi()
1153          *                                futex_lock_pi_atomic()
1154          *   exit_signals(tsk)              No waiters:
1155          *    tsk->flags |= PF_EXITING;     *uaddr == 0x00000PID
1156          *  mm_release(tsk)                 Set waiter bit
1157          *   exit_robust_list(tsk) {        *uaddr = 0x80000PID;
1158          *      Set owner died              attach_to_pi_owner() {
1159          *    *uaddr = 0xC0000000;           tsk = get_task(PID);
1160          *   }                               if (!tsk->flags & PF_EXITING) {
1161          *  ...                                attach();
1162          *  tsk->futex_state =               } else {
1163          *      FUTEX_STATE_DEAD;              if (tsk->futex_state !=
1164          *                                        FUTEX_STATE_DEAD)
1165          *                                       return -EAGAIN;
1166          *                                     return -ESRCH; <--- FAIL
1167          *                                   }
1168          *
1169          * Returning ESRCH unconditionally is wrong here because the
1170          * user space value has been changed by the exiting task.
1171          *
1172          * The same logic applies to the case where the exiting task is
1173          * already gone.
1174          */
1175         if (get_futex_value_locked(&uval2, uaddr))
1176                 return -EFAULT;
1177
1178         /* If the user space value has changed, try again. */
1179         if (uval2 != uval)
1180                 return -EAGAIN;
1181
1182         /*
1183          * The exiting task did not have a robust list, the robust list was
1184          * corrupted or the user space value in *uaddr is simply bogus.
1185          * Give up and tell user space.
1186          */
1187         return -ESRCH;
1188 }
1189
1190 /*
1191  * Lookup the task for the TID provided from user space and attach to
1192  * it after doing proper sanity checks.
1193  */
1194 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1195                               struct futex_pi_state **ps,
1196                               struct task_struct **exiting)
1197 {
1198         pid_t pid = uval & FUTEX_TID_MASK;
1199         struct futex_pi_state *pi_state;
1200         struct task_struct *p;
1201
1202         /*
1203          * We are the first waiter - try to look up the real owner and attach
1204          * the new pi_state to it, but bail out when TID = 0 [1]
1205          *
1206          * The !pid check is paranoid. None of the call sites should end up
1207          * with pid == 0, but better safe than sorry. Let the caller retry
1208          */
1209         if (!pid)
1210                 return -EAGAIN;
1211         p = find_get_task_by_vpid(pid);
1212         if (!p)
1213                 return handle_exit_race(uaddr, uval, NULL);
1214
1215         if (unlikely(p->flags & PF_KTHREAD)) {
1216                 put_task_struct(p);
1217                 return -EPERM;
1218         }
1219
1220         /*
1221          * We need to look at the task state to figure out, whether the
1222          * task is exiting. To protect against the change of the task state
1223          * in futex_exit_release(), we do this protected by p->pi_lock:
1224          */
1225         raw_spin_lock_irq(&p->pi_lock);
1226         if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1227                 /*
1228                  * The task is on the way out. When the futex state is
1229                  * FUTEX_STATE_DEAD, we know that the task has finished
1230                  * the cleanup:
1231                  */
1232                 int ret = handle_exit_race(uaddr, uval, p);
1233
1234                 raw_spin_unlock_irq(&p->pi_lock);
1235                 /*
1236                  * If the owner task is between FUTEX_STATE_EXITING and
1237                  * FUTEX_STATE_DEAD then store the task pointer and keep
1238                  * the reference on the task struct. The calling code will
1239                  * drop all locks, wait for the task to reach
1240                  * FUTEX_STATE_DEAD and then drop the refcount. This is
1241                  * required to prevent a live lock when the current task
1242                  * preempted the exiting task between the two states.
1243                  */
1244                 if (ret == -EBUSY)
1245                         *exiting = p;
1246                 else
1247                         put_task_struct(p);
1248                 return ret;
1249         }
1250
1251         /*
1252          * No existing pi state. First waiter. [2]
1253          *
1254          * This creates pi_state, we have hb->lock held, this means nothing can
1255          * observe this state, wait_lock is irrelevant.
1256          */
1257         pi_state = alloc_pi_state();
1258
1259         /*
1260          * Initialize the pi_mutex in locked state and make @p
1261          * the owner of it:
1262          */
1263         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1264
1265         /* Store the key for possible exit cleanups: */
1266         pi_state->key = *key;
1267
1268         WARN_ON(!list_empty(&pi_state->list));
1269         list_add(&pi_state->list, &p->pi_state_list);
1270         /*
1271          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1272          * because there is no concurrency as the object is not published yet.
1273          */
1274         pi_state->owner = p;
1275         raw_spin_unlock_irq(&p->pi_lock);
1276
1277         put_task_struct(p);
1278
1279         *ps = pi_state;
1280
1281         return 0;
1282 }
1283
1284 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1285                            struct futex_hash_bucket *hb,
1286                            union futex_key *key, struct futex_pi_state **ps,
1287                            struct task_struct **exiting)
1288 {
1289         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1290
1291         /*
1292          * If there is a waiter on that futex, validate it and
1293          * attach to the pi_state when the validation succeeds.
1294          */
1295         if (top_waiter)
1296                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1297
1298         /*
1299          * We are the first waiter - try to look up the owner based on
1300          * @uval and attach to it.
1301          */
1302         return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1303 }
1304
1305 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1306 {
1307         int err;
1308         u32 curval;
1309
1310         if (unlikely(should_fail_futex(true)))
1311                 return -EFAULT;
1312
1313         err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1314         if (unlikely(err))
1315                 return err;
1316
1317         /* If user space value changed, let the caller retry */
1318         return curval != uval ? -EAGAIN : 0;
1319 }
1320
1321 /**
1322  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1323  * @uaddr:              the pi futex user address
1324  * @hb:                 the pi futex hash bucket
1325  * @key:                the futex key associated with uaddr and hb
1326  * @ps:                 the pi_state pointer where we store the result of the
1327  *                      lookup
1328  * @task:               the task to perform the atomic lock work for.  This will
1329  *                      be "current" except in the case of requeue pi.
1330  * @exiting:            Pointer to store the task pointer of the owner task
1331  *                      which is in the middle of exiting
1332  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1333  *
1334  * Return:
1335  *  -  0 - ready to wait;
1336  *  -  1 - acquired the lock;
1337  *  - <0 - error
1338  *
1339  * The hb->lock and futex_key refs shall be held by the caller.
1340  *
1341  * @exiting is only set when the return value is -EBUSY. If so, this holds
1342  * a refcount on the exiting task on return and the caller needs to drop it
1343  * after waiting for the exit to complete.
1344  */
1345 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1346                                 union futex_key *key,
1347                                 struct futex_pi_state **ps,
1348                                 struct task_struct *task,
1349                                 struct task_struct **exiting,
1350                                 int set_waiters)
1351 {
1352         u32 uval, newval, vpid = task_pid_vnr(task);
1353         struct futex_q *top_waiter;
1354         int ret;
1355
1356         /*
1357          * Read the user space value first so we can validate a few
1358          * things before proceeding further.
1359          */
1360         if (get_futex_value_locked(&uval, uaddr))
1361                 return -EFAULT;
1362
1363         if (unlikely(should_fail_futex(true)))
1364                 return -EFAULT;
1365
1366         /*
1367          * Detect deadlocks.
1368          */
1369         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1370                 return -EDEADLK;
1371
1372         if ((unlikely(should_fail_futex(true))))
1373                 return -EDEADLK;
1374
1375         /*
1376          * Lookup existing state first. If it exists, try to attach to
1377          * its pi_state.
1378          */
1379         top_waiter = futex_top_waiter(hb, key);
1380         if (top_waiter)
1381                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1382
1383         /*
1384          * No waiter and user TID is 0. We are here because the
1385          * waiters or the owner died bit is set or called from
1386          * requeue_cmp_pi or for whatever reason something took the
1387          * syscall.
1388          */
1389         if (!(uval & FUTEX_TID_MASK)) {
1390                 /*
1391                  * We take over the futex. No other waiters and the user space
1392                  * TID is 0. We preserve the owner died bit.
1393                  */
1394                 newval = uval & FUTEX_OWNER_DIED;
1395                 newval |= vpid;
1396
1397                 /* The futex requeue_pi code can enforce the waiters bit */
1398                 if (set_waiters)
1399                         newval |= FUTEX_WAITERS;
1400
1401                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1402                 /* If the take over worked, return 1 */
1403                 return ret < 0 ? ret : 1;
1404         }
1405
1406         /*
1407          * First waiter. Set the waiters bit before attaching ourself to
1408          * the owner. If owner tries to unlock, it will be forced into
1409          * the kernel and blocked on hb->lock.
1410          */
1411         newval = uval | FUTEX_WAITERS;
1412         ret = lock_pi_update_atomic(uaddr, uval, newval);
1413         if (ret)
1414                 return ret;
1415         /*
1416          * If the update of the user space value succeeded, we try to
1417          * attach to the owner. If that fails, no harm done, we only
1418          * set the FUTEX_WAITERS bit in the user space variable.
1419          */
1420         return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1421 }
1422
1423 /**
1424  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1425  * @q:  The futex_q to unqueue
1426  *
1427  * The q->lock_ptr must not be NULL and must be held by the caller.
1428  */
1429 static void __unqueue_futex(struct futex_q *q)
1430 {
1431         struct futex_hash_bucket *hb;
1432
1433         if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1434                 return;
1435         lockdep_assert_held(q->lock_ptr);
1436
1437         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1438         plist_del(&q->list, &hb->chain);
1439         hb_waiters_dec(hb);
1440 }
1441
1442 /*
1443  * The hash bucket lock must be held when this is called.
1444  * Afterwards, the futex_q must not be accessed. Callers
1445  * must ensure to later call wake_up_q() for the actual
1446  * wakeups to occur.
1447  */
1448 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1449 {
1450         struct task_struct *p = q->task;
1451
1452         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1453                 return;
1454
1455         get_task_struct(p);
1456         __unqueue_futex(q);
1457         /*
1458          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1459          * is written, without taking any locks. This is possible in the event
1460          * of a spurious wakeup, for example. A memory barrier is required here
1461          * to prevent the following store to lock_ptr from getting ahead of the
1462          * plist_del in __unqueue_futex().
1463          */
1464         smp_store_release(&q->lock_ptr, NULL);
1465
1466         /*
1467          * Queue the task for later wakeup for after we've released
1468          * the hb->lock.
1469          */
1470         wake_q_add_safe(wake_q, p);
1471 }
1472
1473 /*
1474  * Caller must hold a reference on @pi_state.
1475  */
1476 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1477 {
1478         u32 curval, newval;
1479         struct task_struct *new_owner;
1480         bool postunlock = false;
1481         DEFINE_WAKE_Q(wake_q);
1482         int ret = 0;
1483
1484         new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1485         if (WARN_ON_ONCE(!new_owner)) {
1486                 /*
1487                  * As per the comment in futex_unlock_pi() this should not happen.
1488                  *
1489                  * When this happens, give up our locks and try again, giving
1490                  * the futex_lock_pi() instance time to complete, either by
1491                  * waiting on the rtmutex or removing itself from the futex
1492                  * queue.
1493                  */
1494                 ret = -EAGAIN;
1495                 goto out_unlock;
1496         }
1497
1498         /*
1499          * We pass it to the next owner. The WAITERS bit is always kept
1500          * enabled while there is PI state around. We cleanup the owner
1501          * died bit, because we are the owner.
1502          */
1503         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1504
1505         if (unlikely(should_fail_futex(true))) {
1506                 ret = -EFAULT;
1507                 goto out_unlock;
1508         }
1509
1510         ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1511         if (!ret && (curval != uval)) {
1512                 /*
1513                  * If a unconditional UNLOCK_PI operation (user space did not
1514                  * try the TID->0 transition) raced with a waiter setting the
1515                  * FUTEX_WAITERS flag between get_user() and locking the hash
1516                  * bucket lock, retry the operation.
1517                  */
1518                 if ((FUTEX_TID_MASK & curval) == uval)
1519                         ret = -EAGAIN;
1520                 else
1521                         ret = -EINVAL;
1522         }
1523
1524         if (ret)
1525                 goto out_unlock;
1526
1527         /*
1528          * This is a point of no return; once we modify the uval there is no
1529          * going back and subsequent operations must not fail.
1530          */
1531
1532         raw_spin_lock(&pi_state->owner->pi_lock);
1533         WARN_ON(list_empty(&pi_state->list));
1534         list_del_init(&pi_state->list);
1535         raw_spin_unlock(&pi_state->owner->pi_lock);
1536
1537         raw_spin_lock(&new_owner->pi_lock);
1538         WARN_ON(!list_empty(&pi_state->list));
1539         list_add(&pi_state->list, &new_owner->pi_state_list);
1540         pi_state->owner = new_owner;
1541         raw_spin_unlock(&new_owner->pi_lock);
1542
1543         postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1544
1545 out_unlock:
1546         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1547
1548         if (postunlock)
1549                 rt_mutex_postunlock(&wake_q);
1550
1551         return ret;
1552 }
1553
1554 /*
1555  * Express the locking dependencies for lockdep:
1556  */
1557 static inline void
1558 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1559 {
1560         if (hb1 <= hb2) {
1561                 spin_lock(&hb1->lock);
1562                 if (hb1 < hb2)
1563                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1564         } else { /* hb1 > hb2 */
1565                 spin_lock(&hb2->lock);
1566                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1567         }
1568 }
1569
1570 static inline void
1571 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1572 {
1573         spin_unlock(&hb1->lock);
1574         if (hb1 != hb2)
1575                 spin_unlock(&hb2->lock);
1576 }
1577
1578 /*
1579  * Wake up waiters matching bitset queued on this futex (uaddr).
1580  */
1581 static int
1582 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1583 {
1584         struct futex_hash_bucket *hb;
1585         struct futex_q *this, *next;
1586         union futex_key key = FUTEX_KEY_INIT;
1587         int ret;
1588         DEFINE_WAKE_Q(wake_q);
1589
1590         if (!bitset)
1591                 return -EINVAL;
1592
1593         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1594         if (unlikely(ret != 0))
1595                 return ret;
1596
1597         hb = hash_futex(&key);
1598
1599         /* Make sure we really have tasks to wakeup */
1600         if (!hb_waiters_pending(hb))
1601                 return ret;
1602
1603         spin_lock(&hb->lock);
1604
1605         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1606                 if (match_futex (&this->key, &key)) {
1607                         if (this->pi_state || this->rt_waiter) {
1608                                 ret = -EINVAL;
1609                                 break;
1610                         }
1611
1612                         /* Check if one of the bits is set in both bitsets */
1613                         if (!(this->bitset & bitset))
1614                                 continue;
1615
1616                         mark_wake_futex(&wake_q, this);
1617                         if (++ret >= nr_wake)
1618                                 break;
1619                 }
1620         }
1621
1622         spin_unlock(&hb->lock);
1623         wake_up_q(&wake_q);
1624         return ret;
1625 }
1626
1627 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1628 {
1629         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1630         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1631         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1632         int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1633         int oldval, ret;
1634
1635         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1636                 if (oparg < 0 || oparg > 31) {
1637                         char comm[sizeof(current->comm)];
1638                         /*
1639                          * kill this print and return -EINVAL when userspace
1640                          * is sane again
1641                          */
1642                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1643                                         get_task_comm(comm, current), oparg);
1644                         oparg &= 31;
1645                 }
1646                 oparg = 1 << oparg;
1647         }
1648
1649         pagefault_disable();
1650         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1651         pagefault_enable();
1652         if (ret)
1653                 return ret;
1654
1655         switch (cmp) {
1656         case FUTEX_OP_CMP_EQ:
1657                 return oldval == cmparg;
1658         case FUTEX_OP_CMP_NE:
1659                 return oldval != cmparg;
1660         case FUTEX_OP_CMP_LT:
1661                 return oldval < cmparg;
1662         case FUTEX_OP_CMP_GE:
1663                 return oldval >= cmparg;
1664         case FUTEX_OP_CMP_LE:
1665                 return oldval <= cmparg;
1666         case FUTEX_OP_CMP_GT:
1667                 return oldval > cmparg;
1668         default:
1669                 return -ENOSYS;
1670         }
1671 }
1672
1673 /*
1674  * Wake up all waiters hashed on the physical page that is mapped
1675  * to this virtual address:
1676  */
1677 static int
1678 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1679               int nr_wake, int nr_wake2, int op)
1680 {
1681         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1682         struct futex_hash_bucket *hb1, *hb2;
1683         struct futex_q *this, *next;
1684         int ret, op_ret;
1685         DEFINE_WAKE_Q(wake_q);
1686
1687 retry:
1688         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1689         if (unlikely(ret != 0))
1690                 return ret;
1691         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1692         if (unlikely(ret != 0))
1693                 return ret;
1694
1695         hb1 = hash_futex(&key1);
1696         hb2 = hash_futex(&key2);
1697
1698 retry_private:
1699         double_lock_hb(hb1, hb2);
1700         op_ret = futex_atomic_op_inuser(op, uaddr2);
1701         if (unlikely(op_ret < 0)) {
1702                 double_unlock_hb(hb1, hb2);
1703
1704                 if (!IS_ENABLED(CONFIG_MMU) ||
1705                     unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1706                         /*
1707                          * we don't get EFAULT from MMU faults if we don't have
1708                          * an MMU, but we might get them from range checking
1709                          */
1710                         ret = op_ret;
1711                         return ret;
1712                 }
1713
1714                 if (op_ret == -EFAULT) {
1715                         ret = fault_in_user_writeable(uaddr2);
1716                         if (ret)
1717                                 return ret;
1718                 }
1719
1720                 if (!(flags & FLAGS_SHARED)) {
1721                         cond_resched();
1722                         goto retry_private;
1723                 }
1724
1725                 cond_resched();
1726                 goto retry;
1727         }
1728
1729         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1730                 if (match_futex (&this->key, &key1)) {
1731                         if (this->pi_state || this->rt_waiter) {
1732                                 ret = -EINVAL;
1733                                 goto out_unlock;
1734                         }
1735                         mark_wake_futex(&wake_q, this);
1736                         if (++ret >= nr_wake)
1737                                 break;
1738                 }
1739         }
1740
1741         if (op_ret > 0) {
1742                 op_ret = 0;
1743                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1744                         if (match_futex (&this->key, &key2)) {
1745                                 if (this->pi_state || this->rt_waiter) {
1746                                         ret = -EINVAL;
1747                                         goto out_unlock;
1748                                 }
1749                                 mark_wake_futex(&wake_q, this);
1750                                 if (++op_ret >= nr_wake2)
1751                                         break;
1752                         }
1753                 }
1754                 ret += op_ret;
1755         }
1756
1757 out_unlock:
1758         double_unlock_hb(hb1, hb2);
1759         wake_up_q(&wake_q);
1760         return ret;
1761 }
1762
1763 /**
1764  * requeue_futex() - Requeue a futex_q from one hb to another
1765  * @q:          the futex_q to requeue
1766  * @hb1:        the source hash_bucket
1767  * @hb2:        the target hash_bucket
1768  * @key2:       the new key for the requeued futex_q
1769  */
1770 static inline
1771 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1772                    struct futex_hash_bucket *hb2, union futex_key *key2)
1773 {
1774
1775         /*
1776          * If key1 and key2 hash to the same bucket, no need to
1777          * requeue.
1778          */
1779         if (likely(&hb1->chain != &hb2->chain)) {
1780                 plist_del(&q->list, &hb1->chain);
1781                 hb_waiters_dec(hb1);
1782                 hb_waiters_inc(hb2);
1783                 plist_add(&q->list, &hb2->chain);
1784                 q->lock_ptr = &hb2->lock;
1785         }
1786         q->key = *key2;
1787 }
1788
1789 /**
1790  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1791  * @q:          the futex_q
1792  * @key:        the key of the requeue target futex
1793  * @hb:         the hash_bucket of the requeue target futex
1794  *
1795  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1796  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1797  * to the requeue target futex so the waiter can detect the wakeup on the right
1798  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1799  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1800  * to protect access to the pi_state to fixup the owner later.  Must be called
1801  * with both q->lock_ptr and hb->lock held.
1802  */
1803 static inline
1804 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1805                            struct futex_hash_bucket *hb)
1806 {
1807         q->key = *key;
1808
1809         __unqueue_futex(q);
1810
1811         WARN_ON(!q->rt_waiter);
1812         q->rt_waiter = NULL;
1813
1814         q->lock_ptr = &hb->lock;
1815
1816         wake_up_state(q->task, TASK_NORMAL);
1817 }
1818
1819 /**
1820  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1821  * @pifutex:            the user address of the to futex
1822  * @hb1:                the from futex hash bucket, must be locked by the caller
1823  * @hb2:                the to futex hash bucket, must be locked by the caller
1824  * @key1:               the from futex key
1825  * @key2:               the to futex key
1826  * @ps:                 address to store the pi_state pointer
1827  * @exiting:            Pointer to store the task pointer of the owner task
1828  *                      which is in the middle of exiting
1829  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1830  *
1831  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1832  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1833  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1834  * hb1 and hb2 must be held by the caller.
1835  *
1836  * @exiting is only set when the return value is -EBUSY. If so, this holds
1837  * a refcount on the exiting task on return and the caller needs to drop it
1838  * after waiting for the exit to complete.
1839  *
1840  * Return:
1841  *  -  0 - failed to acquire the lock atomically;
1842  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1843  *  - <0 - error
1844  */
1845 static int
1846 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1847                            struct futex_hash_bucket *hb2, union futex_key *key1,
1848                            union futex_key *key2, struct futex_pi_state **ps,
1849                            struct task_struct **exiting, int set_waiters)
1850 {
1851         struct futex_q *top_waiter = NULL;
1852         u32 curval;
1853         int ret, vpid;
1854
1855         if (get_futex_value_locked(&curval, pifutex))
1856                 return -EFAULT;
1857
1858         if (unlikely(should_fail_futex(true)))
1859                 return -EFAULT;
1860
1861         /*
1862          * Find the top_waiter and determine if there are additional waiters.
1863          * If the caller intends to requeue more than 1 waiter to pifutex,
1864          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1865          * as we have means to handle the possible fault.  If not, don't set
1866          * the bit unecessarily as it will force the subsequent unlock to enter
1867          * the kernel.
1868          */
1869         top_waiter = futex_top_waiter(hb1, key1);
1870
1871         /* There are no waiters, nothing for us to do. */
1872         if (!top_waiter)
1873                 return 0;
1874
1875         /* Ensure we requeue to the expected futex. */
1876         if (!match_futex(top_waiter->requeue_pi_key, key2))
1877                 return -EINVAL;
1878
1879         /*
1880          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1881          * the contended case or if set_waiters is 1.  The pi_state is returned
1882          * in ps in contended cases.
1883          */
1884         vpid = task_pid_vnr(top_waiter->task);
1885         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1886                                    exiting, set_waiters);
1887         if (ret == 1) {
1888                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1889                 return vpid;
1890         }
1891         return ret;
1892 }
1893
1894 /**
1895  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1896  * @uaddr1:     source futex user address
1897  * @flags:      futex flags (FLAGS_SHARED, etc.)
1898  * @uaddr2:     target futex user address
1899  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1900  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1901  * @cmpval:     @uaddr1 expected value (or %NULL)
1902  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1903  *              pi futex (pi to pi requeue is not supported)
1904  *
1905  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1906  * uaddr2 atomically on behalf of the top waiter.
1907  *
1908  * Return:
1909  *  - >=0 - on success, the number of tasks requeued or woken;
1910  *  -  <0 - on error
1911  */
1912 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1913                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1914                          u32 *cmpval, int requeue_pi)
1915 {
1916         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1917         int task_count = 0, ret;
1918         struct futex_pi_state *pi_state = NULL;
1919         struct futex_hash_bucket *hb1, *hb2;
1920         struct futex_q *this, *next;
1921         DEFINE_WAKE_Q(wake_q);
1922
1923         if (nr_wake < 0 || nr_requeue < 0)
1924                 return -EINVAL;
1925
1926         /*
1927          * When PI not supported: return -ENOSYS if requeue_pi is true,
1928          * consequently the compiler knows requeue_pi is always false past
1929          * this point which will optimize away all the conditional code
1930          * further down.
1931          */
1932         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1933                 return -ENOSYS;
1934
1935         if (requeue_pi) {
1936                 /*
1937                  * Requeue PI only works on two distinct uaddrs. This
1938                  * check is only valid for private futexes. See below.
1939                  */
1940                 if (uaddr1 == uaddr2)
1941                         return -EINVAL;
1942
1943                 /*
1944                  * requeue_pi requires a pi_state, try to allocate it now
1945                  * without any locks in case it fails.
1946                  */
1947                 if (refill_pi_state_cache())
1948                         return -ENOMEM;
1949                 /*
1950                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1951                  * + nr_requeue, since it acquires the rt_mutex prior to
1952                  * returning to userspace, so as to not leave the rt_mutex with
1953                  * waiters and no owner.  However, second and third wake-ups
1954                  * cannot be predicted as they involve race conditions with the
1955                  * first wake and a fault while looking up the pi_state.  Both
1956                  * pthread_cond_signal() and pthread_cond_broadcast() should
1957                  * use nr_wake=1.
1958                  */
1959                 if (nr_wake != 1)
1960                         return -EINVAL;
1961         }
1962
1963 retry:
1964         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1965         if (unlikely(ret != 0))
1966                 return ret;
1967         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1968                             requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1969         if (unlikely(ret != 0))
1970                 return ret;
1971
1972         /*
1973          * The check above which compares uaddrs is not sufficient for
1974          * shared futexes. We need to compare the keys:
1975          */
1976         if (requeue_pi && match_futex(&key1, &key2))
1977                 return -EINVAL;
1978
1979         hb1 = hash_futex(&key1);
1980         hb2 = hash_futex(&key2);
1981
1982 retry_private:
1983         hb_waiters_inc(hb2);
1984         double_lock_hb(hb1, hb2);
1985
1986         if (likely(cmpval != NULL)) {
1987                 u32 curval;
1988
1989                 ret = get_futex_value_locked(&curval, uaddr1);
1990
1991                 if (unlikely(ret)) {
1992                         double_unlock_hb(hb1, hb2);
1993                         hb_waiters_dec(hb2);
1994
1995                         ret = get_user(curval, uaddr1);
1996                         if (ret)
1997                                 return ret;
1998
1999                         if (!(flags & FLAGS_SHARED))
2000                                 goto retry_private;
2001
2002                         goto retry;
2003                 }
2004                 if (curval != *cmpval) {
2005                         ret = -EAGAIN;
2006                         goto out_unlock;
2007                 }
2008         }
2009
2010         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2011                 struct task_struct *exiting = NULL;
2012
2013                 /*
2014                  * Attempt to acquire uaddr2 and wake the top waiter. If we
2015                  * intend to requeue waiters, force setting the FUTEX_WAITERS
2016                  * bit.  We force this here where we are able to easily handle
2017                  * faults rather in the requeue loop below.
2018                  */
2019                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2020                                                  &key2, &pi_state,
2021                                                  &exiting, nr_requeue);
2022
2023                 /*
2024                  * At this point the top_waiter has either taken uaddr2 or is
2025                  * waiting on it.  If the former, then the pi_state will not
2026                  * exist yet, look it up one more time to ensure we have a
2027                  * reference to it. If the lock was taken, ret contains the
2028                  * vpid of the top waiter task.
2029                  * If the lock was not taken, we have pi_state and an initial
2030                  * refcount on it. In case of an error we have nothing.
2031                  */
2032                 if (ret > 0) {
2033                         WARN_ON(pi_state);
2034                         task_count++;
2035                         /*
2036                          * If we acquired the lock, then the user space value
2037                          * of uaddr2 should be vpid. It cannot be changed by
2038                          * the top waiter as it is blocked on hb2 lock if it
2039                          * tries to do so. If something fiddled with it behind
2040                          * our back the pi state lookup might unearth it. So
2041                          * we rather use the known value than rereading and
2042                          * handing potential crap to lookup_pi_state.
2043                          *
2044                          * If that call succeeds then we have pi_state and an
2045                          * initial refcount on it.
2046                          */
2047                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2048                                               &pi_state, &exiting);
2049                 }
2050
2051                 switch (ret) {
2052                 case 0:
2053                         /* We hold a reference on the pi state. */
2054                         break;
2055
2056                         /* If the above failed, then pi_state is NULL */
2057                 case -EFAULT:
2058                         double_unlock_hb(hb1, hb2);
2059                         hb_waiters_dec(hb2);
2060                         ret = fault_in_user_writeable(uaddr2);
2061                         if (!ret)
2062                                 goto retry;
2063                         return ret;
2064                 case -EBUSY:
2065                 case -EAGAIN:
2066                         /*
2067                          * Two reasons for this:
2068                          * - EBUSY: Owner is exiting and we just wait for the
2069                          *   exit to complete.
2070                          * - EAGAIN: The user space value changed.
2071                          */
2072                         double_unlock_hb(hb1, hb2);
2073                         hb_waiters_dec(hb2);
2074                         /*
2075                          * Handle the case where the owner is in the middle of
2076                          * exiting. Wait for the exit to complete otherwise
2077                          * this task might loop forever, aka. live lock.
2078                          */
2079                         wait_for_owner_exiting(ret, exiting);
2080                         cond_resched();
2081                         goto retry;
2082                 default:
2083                         goto out_unlock;
2084                 }
2085         }
2086
2087         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2088                 if (task_count - nr_wake >= nr_requeue)
2089                         break;
2090
2091                 if (!match_futex(&this->key, &key1))
2092                         continue;
2093
2094                 /*
2095                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2096                  * be paired with each other and no other futex ops.
2097                  *
2098                  * We should never be requeueing a futex_q with a pi_state,
2099                  * which is awaiting a futex_unlock_pi().
2100                  */
2101                 if ((requeue_pi && !this->rt_waiter) ||
2102                     (!requeue_pi && this->rt_waiter) ||
2103                     this->pi_state) {
2104                         ret = -EINVAL;
2105                         break;
2106                 }
2107
2108                 /*
2109                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2110                  * lock, we already woke the top_waiter.  If not, it will be
2111                  * woken by futex_unlock_pi().
2112                  */
2113                 if (++task_count <= nr_wake && !requeue_pi) {
2114                         mark_wake_futex(&wake_q, this);
2115                         continue;
2116                 }
2117
2118                 /* Ensure we requeue to the expected futex for requeue_pi. */
2119                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2120                         ret = -EINVAL;
2121                         break;
2122                 }
2123
2124                 /*
2125                  * Requeue nr_requeue waiters and possibly one more in the case
2126                  * of requeue_pi if we couldn't acquire the lock atomically.
2127                  */
2128                 if (requeue_pi) {
2129                         /*
2130                          * Prepare the waiter to take the rt_mutex. Take a
2131                          * refcount on the pi_state and store the pointer in
2132                          * the futex_q object of the waiter.
2133                          */
2134                         get_pi_state(pi_state);
2135                         this->pi_state = pi_state;
2136                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2137                                                         this->rt_waiter,
2138                                                         this->task);
2139                         if (ret == 1) {
2140                                 /*
2141                                  * We got the lock. We do neither drop the
2142                                  * refcount on pi_state nor clear
2143                                  * this->pi_state because the waiter needs the
2144                                  * pi_state for cleaning up the user space
2145                                  * value. It will drop the refcount after
2146                                  * doing so.
2147                                  */
2148                                 requeue_pi_wake_futex(this, &key2, hb2);
2149                                 continue;
2150                         } else if (ret) {
2151                                 /*
2152                                  * rt_mutex_start_proxy_lock() detected a
2153                                  * potential deadlock when we tried to queue
2154                                  * that waiter. Drop the pi_state reference
2155                                  * which we took above and remove the pointer
2156                                  * to the state from the waiters futex_q
2157                                  * object.
2158                                  */
2159                                 this->pi_state = NULL;
2160                                 put_pi_state(pi_state);
2161                                 /*
2162                                  * We stop queueing more waiters and let user
2163                                  * space deal with the mess.
2164                                  */
2165                                 break;
2166                         }
2167                 }
2168                 requeue_futex(this, hb1, hb2, &key2);
2169         }
2170
2171         /*
2172          * We took an extra initial reference to the pi_state either
2173          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2174          * need to drop it here again.
2175          */
2176         put_pi_state(pi_state);
2177
2178 out_unlock:
2179         double_unlock_hb(hb1, hb2);
2180         wake_up_q(&wake_q);
2181         hb_waiters_dec(hb2);
2182         return ret ? ret : task_count;
2183 }
2184
2185 /* The key must be already stored in q->key. */
2186 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2187         __acquires(&hb->lock)
2188 {
2189         struct futex_hash_bucket *hb;
2190
2191         hb = hash_futex(&q->key);
2192
2193         /*
2194          * Increment the counter before taking the lock so that
2195          * a potential waker won't miss a to-be-slept task that is
2196          * waiting for the spinlock. This is safe as all queue_lock()
2197          * users end up calling queue_me(). Similarly, for housekeeping,
2198          * decrement the counter at queue_unlock() when some error has
2199          * occurred and we don't end up adding the task to the list.
2200          */
2201         hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2202
2203         q->lock_ptr = &hb->lock;
2204
2205         spin_lock(&hb->lock);
2206         return hb;
2207 }
2208
2209 static inline void
2210 queue_unlock(struct futex_hash_bucket *hb)
2211         __releases(&hb->lock)
2212 {
2213         spin_unlock(&hb->lock);
2214         hb_waiters_dec(hb);
2215 }
2216
2217 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2218 {
2219         int prio;
2220
2221         /*
2222          * The priority used to register this element is
2223          * - either the real thread-priority for the real-time threads
2224          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2225          * - or MAX_RT_PRIO for non-RT threads.
2226          * Thus, all RT-threads are woken first in priority order, and
2227          * the others are woken last, in FIFO order.
2228          */
2229         prio = min(current->normal_prio, MAX_RT_PRIO);
2230
2231         plist_node_init(&q->list, prio);
2232         plist_add(&q->list, &hb->chain);
2233         q->task = current;
2234 }
2235
2236 /**
2237  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2238  * @q:  The futex_q to enqueue
2239  * @hb: The destination hash bucket
2240  *
2241  * The hb->lock must be held by the caller, and is released here. A call to
2242  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2243  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2244  * or nothing if the unqueue is done as part of the wake process and the unqueue
2245  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2246  * an example).
2247  */
2248 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2249         __releases(&hb->lock)
2250 {
2251         __queue_me(q, hb);
2252         spin_unlock(&hb->lock);
2253 }
2254
2255 /**
2256  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2257  * @q:  The futex_q to unqueue
2258  *
2259  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2260  * be paired with exactly one earlier call to queue_me().
2261  *
2262  * Return:
2263  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2264  *  - 0 - if the futex_q was already removed by the waking thread
2265  */
2266 static int unqueue_me(struct futex_q *q)
2267 {
2268         spinlock_t *lock_ptr;
2269         int ret = 0;
2270
2271         /* In the common case we don't take the spinlock, which is nice. */
2272 retry:
2273         /*
2274          * q->lock_ptr can change between this read and the following spin_lock.
2275          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2276          * optimizing lock_ptr out of the logic below.
2277          */
2278         lock_ptr = READ_ONCE(q->lock_ptr);
2279         if (lock_ptr != NULL) {
2280                 spin_lock(lock_ptr);
2281                 /*
2282                  * q->lock_ptr can change between reading it and
2283                  * spin_lock(), causing us to take the wrong lock.  This
2284                  * corrects the race condition.
2285                  *
2286                  * Reasoning goes like this: if we have the wrong lock,
2287                  * q->lock_ptr must have changed (maybe several times)
2288                  * between reading it and the spin_lock().  It can
2289                  * change again after the spin_lock() but only if it was
2290                  * already changed before the spin_lock().  It cannot,
2291                  * however, change back to the original value.  Therefore
2292                  * we can detect whether we acquired the correct lock.
2293                  */
2294                 if (unlikely(lock_ptr != q->lock_ptr)) {
2295                         spin_unlock(lock_ptr);
2296                         goto retry;
2297                 }
2298                 __unqueue_futex(q);
2299
2300                 BUG_ON(q->pi_state);
2301
2302                 spin_unlock(lock_ptr);
2303                 ret = 1;
2304         }
2305
2306         return ret;
2307 }
2308
2309 /*
2310  * PI futexes can not be requeued and must remove themself from the
2311  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2312  * and dropped here.
2313  */
2314 static void unqueue_me_pi(struct futex_q *q)
2315         __releases(q->lock_ptr)
2316 {
2317         __unqueue_futex(q);
2318
2319         BUG_ON(!q->pi_state);
2320         put_pi_state(q->pi_state);
2321         q->pi_state = NULL;
2322
2323         spin_unlock(q->lock_ptr);
2324 }
2325
2326 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2327                                 struct task_struct *argowner)
2328 {
2329         struct futex_pi_state *pi_state = q->pi_state;
2330         u32 uval, curval, newval;
2331         struct task_struct *oldowner, *newowner;
2332         u32 newtid;
2333         int ret, err = 0;
2334
2335         lockdep_assert_held(q->lock_ptr);
2336
2337         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2338
2339         oldowner = pi_state->owner;
2340
2341         /*
2342          * We are here because either:
2343          *
2344          *  - we stole the lock and pi_state->owner needs updating to reflect
2345          *    that (@argowner == current),
2346          *
2347          * or:
2348          *
2349          *  - someone stole our lock and we need to fix things to point to the
2350          *    new owner (@argowner == NULL).
2351          *
2352          * Either way, we have to replace the TID in the user space variable.
2353          * This must be atomic as we have to preserve the owner died bit here.
2354          *
2355          * Note: We write the user space value _before_ changing the pi_state
2356          * because we can fault here. Imagine swapped out pages or a fork
2357          * that marked all the anonymous memory readonly for cow.
2358          *
2359          * Modifying pi_state _before_ the user space value would leave the
2360          * pi_state in an inconsistent state when we fault here, because we
2361          * need to drop the locks to handle the fault. This might be observed
2362          * in the PID check in lookup_pi_state.
2363          */
2364 retry:
2365         if (!argowner) {
2366                 if (oldowner != current) {
2367                         /*
2368                          * We raced against a concurrent self; things are
2369                          * already fixed up. Nothing to do.
2370                          */
2371                         ret = 0;
2372                         goto out_unlock;
2373                 }
2374
2375                 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2376                         /* We got the lock after all, nothing to fix. */
2377                         ret = 0;
2378                         goto out_unlock;
2379                 }
2380
2381                 /*
2382                  * The trylock just failed, so either there is an owner or
2383                  * there is a higher priority waiter than this one.
2384                  */
2385                 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2386                 /*
2387                  * If the higher priority waiter has not yet taken over the
2388                  * rtmutex then newowner is NULL. We can't return here with
2389                  * that state because it's inconsistent vs. the user space
2390                  * state. So drop the locks and try again. It's a valid
2391                  * situation and not any different from the other retry
2392                  * conditions.
2393                  */
2394                 if (unlikely(!newowner)) {
2395                         err = -EAGAIN;
2396                         goto handle_err;
2397                 }
2398         } else {
2399                 WARN_ON_ONCE(argowner != current);
2400                 if (oldowner == current) {
2401                         /*
2402                          * We raced against a concurrent self; things are
2403                          * already fixed up. Nothing to do.
2404                          */
2405                         ret = 0;
2406                         goto out_unlock;
2407                 }
2408                 newowner = argowner;
2409         }
2410
2411         newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2412         /* Owner died? */
2413         if (!pi_state->owner)
2414                 newtid |= FUTEX_OWNER_DIED;
2415
2416         err = get_futex_value_locked(&uval, uaddr);
2417         if (err)
2418                 goto handle_err;
2419
2420         for (;;) {
2421                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2422
2423                 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2424                 if (err)
2425                         goto handle_err;
2426
2427                 if (curval == uval)
2428                         break;
2429                 uval = curval;
2430         }
2431
2432         /*
2433          * We fixed up user space. Now we need to fix the pi_state
2434          * itself.
2435          */
2436         if (pi_state->owner != NULL) {
2437                 raw_spin_lock(&pi_state->owner->pi_lock);
2438                 WARN_ON(list_empty(&pi_state->list));
2439                 list_del_init(&pi_state->list);
2440                 raw_spin_unlock(&pi_state->owner->pi_lock);
2441         }
2442
2443         pi_state->owner = newowner;
2444
2445         raw_spin_lock(&newowner->pi_lock);
2446         WARN_ON(!list_empty(&pi_state->list));
2447         list_add(&pi_state->list, &newowner->pi_state_list);
2448         raw_spin_unlock(&newowner->pi_lock);
2449         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2450
2451         return 0;
2452
2453         /*
2454          * In order to reschedule or handle a page fault, we need to drop the
2455          * locks here. In the case of a fault, this gives the other task
2456          * (either the highest priority waiter itself or the task which stole
2457          * the rtmutex) the chance to try the fixup of the pi_state. So once we
2458          * are back from handling the fault we need to check the pi_state after
2459          * reacquiring the locks and before trying to do another fixup. When
2460          * the fixup has been done already we simply return.
2461          *
2462          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2463          * drop hb->lock since the caller owns the hb -> futex_q relation.
2464          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2465          */
2466 handle_err:
2467         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2468         spin_unlock(q->lock_ptr);
2469
2470         switch (err) {
2471         case -EFAULT:
2472                 ret = fault_in_user_writeable(uaddr);
2473                 break;
2474
2475         case -EAGAIN:
2476                 cond_resched();
2477                 ret = 0;
2478                 break;
2479
2480         default:
2481                 WARN_ON_ONCE(1);
2482                 ret = err;
2483                 break;
2484         }
2485
2486         spin_lock(q->lock_ptr);
2487         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2488
2489         /*
2490          * Check if someone else fixed it for us:
2491          */
2492         if (pi_state->owner != oldowner) {
2493                 ret = 0;
2494                 goto out_unlock;
2495         }
2496
2497         if (ret)
2498                 goto out_unlock;
2499
2500         goto retry;
2501
2502 out_unlock:
2503         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2504         return ret;
2505 }
2506
2507 static long futex_wait_restart(struct restart_block *restart);
2508
2509 /**
2510  * fixup_owner() - Post lock pi_state and corner case management
2511  * @uaddr:      user address of the futex
2512  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2513  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2514  *
2515  * After attempting to lock an rt_mutex, this function is called to cleanup
2516  * the pi_state owner as well as handle race conditions that may allow us to
2517  * acquire the lock. Must be called with the hb lock held.
2518  *
2519  * Return:
2520  *  -  1 - success, lock taken;
2521  *  -  0 - success, lock not taken;
2522  *  - <0 - on error (-EFAULT)
2523  */
2524 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2525 {
2526         int ret = 0;
2527
2528         if (locked) {
2529                 /*
2530                  * Got the lock. We might not be the anticipated owner if we
2531                  * did a lock-steal - fix up the PI-state in that case:
2532                  *
2533                  * Speculative pi_state->owner read (we don't hold wait_lock);
2534                  * since we own the lock pi_state->owner == current is the
2535                  * stable state, anything else needs more attention.
2536                  */
2537                 if (q->pi_state->owner != current)
2538                         ret = fixup_pi_state_owner(uaddr, q, current);
2539                 return ret ? ret : locked;
2540         }
2541
2542         /*
2543          * If we didn't get the lock; check if anybody stole it from us. In
2544          * that case, we need to fix up the uval to point to them instead of
2545          * us, otherwise bad things happen. [10]
2546          *
2547          * Another speculative read; pi_state->owner == current is unstable
2548          * but needs our attention.
2549          */
2550         if (q->pi_state->owner == current) {
2551                 ret = fixup_pi_state_owner(uaddr, q, NULL);
2552                 return ret;
2553         }
2554
2555         /*
2556          * Paranoia check. If we did not take the lock, then we should not be
2557          * the owner of the rt_mutex.
2558          */
2559         if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2560                 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2561                                 "pi-state %p\n", ret,
2562                                 q->pi_state->pi_mutex.owner,
2563                                 q->pi_state->owner);
2564         }
2565
2566         return ret;
2567 }
2568
2569 /**
2570  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571  * @hb:         the futex hash bucket, must be locked by the caller
2572  * @q:          the futex_q to queue up on
2573  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2574  */
2575 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2576                                 struct hrtimer_sleeper *timeout)
2577 {
2578         /*
2579          * The task state is guaranteed to be set before another task can
2580          * wake it. set_current_state() is implemented using smp_store_mb() and
2581          * queue_me() calls spin_unlock() upon completion, both serializing
2582          * access to the hash list and forcing another memory barrier.
2583          */
2584         set_current_state(TASK_INTERRUPTIBLE);
2585         queue_me(q, hb);
2586
2587         /* Arm the timer */
2588         if (timeout)
2589                 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2590
2591         /*
2592          * If we have been removed from the hash list, then another task
2593          * has tried to wake us, and we can skip the call to schedule().
2594          */
2595         if (likely(!plist_node_empty(&q->list))) {
2596                 /*
2597                  * If the timer has already expired, current will already be
2598                  * flagged for rescheduling. Only call schedule if there
2599                  * is no timeout, or if it has yet to expire.
2600                  */
2601                 if (!timeout || timeout->task)
2602                         freezable_schedule();
2603         }
2604         __set_current_state(TASK_RUNNING);
2605 }
2606
2607 /**
2608  * futex_wait_setup() - Prepare to wait on a futex
2609  * @uaddr:      the futex userspace address
2610  * @val:        the expected value
2611  * @flags:      futex flags (FLAGS_SHARED, etc.)
2612  * @q:          the associated futex_q
2613  * @hb:         storage for hash_bucket pointer to be returned to caller
2614  *
2615  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2616  * compare it with the expected value.  Handle atomic faults internally.
2617  * Return with the hb lock held and a q.key reference on success, and unlocked
2618  * with no q.key reference on failure.
2619  *
2620  * Return:
2621  *  -  0 - uaddr contains val and hb has been locked;
2622  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2623  */
2624 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2625                            struct futex_q *q, struct futex_hash_bucket **hb)
2626 {
2627         u32 uval;
2628         int ret;
2629
2630         /*
2631          * Access the page AFTER the hash-bucket is locked.
2632          * Order is important:
2633          *
2634          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2635          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2636          *
2637          * The basic logical guarantee of a futex is that it blocks ONLY
2638          * if cond(var) is known to be true at the time of blocking, for
2639          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2640          * would open a race condition where we could block indefinitely with
2641          * cond(var) false, which would violate the guarantee.
2642          *
2643          * On the other hand, we insert q and release the hash-bucket only
2644          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2645          * absorb a wakeup if *uaddr does not match the desired values
2646          * while the syscall executes.
2647          */
2648 retry:
2649         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2650         if (unlikely(ret != 0))
2651                 return ret;
2652
2653 retry_private:
2654         *hb = queue_lock(q);
2655
2656         ret = get_futex_value_locked(&uval, uaddr);
2657
2658         if (ret) {
2659                 queue_unlock(*hb);
2660
2661                 ret = get_user(uval, uaddr);
2662                 if (ret)
2663                         return ret;
2664
2665                 if (!(flags & FLAGS_SHARED))
2666                         goto retry_private;
2667
2668                 goto retry;
2669         }
2670
2671         if (uval != val) {
2672                 queue_unlock(*hb);
2673                 ret = -EWOULDBLOCK;
2674         }
2675
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;
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         to = futex_setup_timer(abs_time, &timeout, flags,
2693                                current->timer_slack_ns);
2694 retry:
2695         /*
2696          * Prepare to wait on uaddr. On success, holds hb lock and increments
2697          * q.key refs.
2698          */
2699         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2700         if (ret)
2701                 goto out;
2702
2703         /* queue_me and wait for wakeup, timeout, or a signal. */
2704         futex_wait_queue_me(hb, &q, to);
2705
2706         /* If we were woken (and unqueued), we succeeded, whatever. */
2707         ret = 0;
2708         /* unqueue_me() drops q.key ref */
2709         if (!unqueue_me(&q))
2710                 goto out;
2711         ret = -ETIMEDOUT;
2712         if (to && !to->task)
2713                 goto out;
2714
2715         /*
2716          * We expect signal_pending(current), but we might be the
2717          * victim of a spurious wakeup as well.
2718          */
2719         if (!signal_pending(current))
2720                 goto retry;
2721
2722         ret = -ERESTARTSYS;
2723         if (!abs_time)
2724                 goto out;
2725
2726         restart = &current->restart_block;
2727         restart->fn = futex_wait_restart;
2728         restart->futex.uaddr = uaddr;
2729         restart->futex.val = val;
2730         restart->futex.time = *abs_time;
2731         restart->futex.bitset = bitset;
2732         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2733
2734         ret = -ERESTART_RESTARTBLOCK;
2735
2736 out:
2737         if (to) {
2738                 hrtimer_cancel(&to->timer);
2739                 destroy_hrtimer_on_stack(&to->timer);
2740         }
2741         return ret;
2742 }
2743
2744
2745 static long futex_wait_restart(struct restart_block *restart)
2746 {
2747         u32 __user *uaddr = restart->futex.uaddr;
2748         ktime_t t, *tp = NULL;
2749
2750         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2751                 t = restart->futex.time;
2752                 tp = &t;
2753         }
2754         restart->fn = do_no_restart_syscall;
2755
2756         return (long)futex_wait(uaddr, restart->futex.flags,
2757                                 restart->futex.val, tp, restart->futex.bitset);
2758 }
2759
2760
2761 /*
2762  * Userspace tried a 0 -> TID atomic transition of the futex value
2763  * and failed. The kernel side here does the whole locking operation:
2764  * if there are waiters then it will block as a consequence of relying
2765  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2766  * a 0 value of the futex too.).
2767  *
2768  * Also serves as futex trylock_pi()'ing, and due semantics.
2769  */
2770 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2771                          ktime_t *time, int trylock)
2772 {
2773         struct hrtimer_sleeper timeout, *to;
2774         struct futex_pi_state *pi_state = NULL;
2775         struct task_struct *exiting = NULL;
2776         struct rt_mutex_waiter rt_waiter;
2777         struct futex_hash_bucket *hb;
2778         struct futex_q q = futex_q_init;
2779         int res, ret;
2780
2781         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2782                 return -ENOSYS;
2783
2784         if (refill_pi_state_cache())
2785                 return -ENOMEM;
2786
2787         to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2788
2789 retry:
2790         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2791         if (unlikely(ret != 0))
2792                 goto out;
2793
2794 retry_private:
2795         hb = queue_lock(&q);
2796
2797         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2798                                    &exiting, 0);
2799         if (unlikely(ret)) {
2800                 /*
2801                  * Atomic work succeeded and we got the lock,
2802                  * or failed. Either way, we do _not_ block.
2803                  */
2804                 switch (ret) {
2805                 case 1:
2806                         /* We got the lock. */
2807                         ret = 0;
2808                         goto out_unlock_put_key;
2809                 case -EFAULT:
2810                         goto uaddr_faulted;
2811                 case -EBUSY:
2812                 case -EAGAIN:
2813                         /*
2814                          * Two reasons for this:
2815                          * - EBUSY: Task is exiting and we just wait for the
2816                          *   exit to complete.
2817                          * - EAGAIN: The user space value changed.
2818                          */
2819                         queue_unlock(hb);
2820                         /*
2821                          * Handle the case where the owner is in the middle of
2822                          * exiting. Wait for the exit to complete otherwise
2823                          * this task might loop forever, aka. live lock.
2824                          */
2825                         wait_for_owner_exiting(ret, exiting);
2826                         cond_resched();
2827                         goto retry;
2828                 default:
2829                         goto out_unlock_put_key;
2830                 }
2831         }
2832
2833         WARN_ON(!q.pi_state);
2834
2835         /*
2836          * Only actually queue now that the atomic ops are done:
2837          */
2838         __queue_me(&q, hb);
2839
2840         if (trylock) {
2841                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2842                 /* Fixup the trylock return value: */
2843                 ret = ret ? 0 : -EWOULDBLOCK;
2844                 goto no_block;
2845         }
2846
2847         rt_mutex_init_waiter(&rt_waiter);
2848
2849         /*
2850          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2851          * hold it while doing rt_mutex_start_proxy(), because then it will
2852          * include hb->lock in the blocking chain, even through we'll not in
2853          * fact hold it while blocking. This will lead it to report -EDEADLK
2854          * and BUG when futex_unlock_pi() interleaves with this.
2855          *
2856          * Therefore acquire wait_lock while holding hb->lock, but drop the
2857          * latter before calling __rt_mutex_start_proxy_lock(). This
2858          * interleaves with futex_unlock_pi() -- which does a similar lock
2859          * handoff -- such that the latter can observe the futex_q::pi_state
2860          * before __rt_mutex_start_proxy_lock() is done.
2861          */
2862         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2863         spin_unlock(q.lock_ptr);
2864         /*
2865          * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2866          * such that futex_unlock_pi() is guaranteed to observe the waiter when
2867          * it sees the futex_q::pi_state.
2868          */
2869         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2870         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2871
2872         if (ret) {
2873                 if (ret == 1)
2874                         ret = 0;
2875                 goto cleanup;
2876         }
2877
2878         if (unlikely(to))
2879                 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2880
2881         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2882
2883 cleanup:
2884         spin_lock(q.lock_ptr);
2885         /*
2886          * If we failed to acquire the lock (deadlock/signal/timeout), we must
2887          * first acquire the hb->lock before removing the lock from the
2888          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2889          * lists consistent.
2890          *
2891          * In particular; it is important that futex_unlock_pi() can not
2892          * observe this inconsistency.
2893          */
2894         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2895                 ret = 0;
2896
2897 no_block:
2898         /*
2899          * Fixup the pi_state owner and possibly acquire the lock if we
2900          * haven't already.
2901          */
2902         res = fixup_owner(uaddr, &q, !ret);
2903         /*
2904          * If fixup_owner() returned an error, proprogate that.  If it acquired
2905          * the lock, clear our -ETIMEDOUT or -EINTR.
2906          */
2907         if (res)
2908                 ret = (res < 0) ? res : 0;
2909
2910         /*
2911          * If fixup_owner() faulted and was unable to handle the fault, unlock
2912          * it and return the fault to userspace.
2913          */
2914         if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2915                 pi_state = q.pi_state;
2916                 get_pi_state(pi_state);
2917         }
2918
2919         /* Unqueue and drop the lock */
2920         unqueue_me_pi(&q);
2921
2922         if (pi_state) {
2923                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2924                 put_pi_state(pi_state);
2925         }
2926
2927         goto out;
2928
2929 out_unlock_put_key:
2930         queue_unlock(hb);
2931
2932 out:
2933         if (to) {
2934                 hrtimer_cancel(&to->timer);
2935                 destroy_hrtimer_on_stack(&to->timer);
2936         }
2937         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2938
2939 uaddr_faulted:
2940         queue_unlock(hb);
2941
2942         ret = fault_in_user_writeable(uaddr);
2943         if (ret)
2944                 goto out;
2945
2946         if (!(flags & FLAGS_SHARED))
2947                 goto retry_private;
2948
2949         goto retry;
2950 }
2951
2952 /*
2953  * Userspace attempted a TID -> 0 atomic transition, and failed.
2954  * This is the in-kernel slowpath: we look up the PI state (if any),
2955  * and do the rt-mutex unlock.
2956  */
2957 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2958 {
2959         u32 curval, uval, vpid = task_pid_vnr(current);
2960         union futex_key key = FUTEX_KEY_INIT;
2961         struct futex_hash_bucket *hb;
2962         struct futex_q *top_waiter;
2963         int ret;
2964
2965         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2966                 return -ENOSYS;
2967
2968 retry:
2969         if (get_user(uval, uaddr))
2970                 return -EFAULT;
2971         /*
2972          * We release only a lock we actually own:
2973          */
2974         if ((uval & FUTEX_TID_MASK) != vpid)
2975                 return -EPERM;
2976
2977         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2978         if (ret)
2979                 return ret;
2980
2981         hb = hash_futex(&key);
2982         spin_lock(&hb->lock);
2983
2984         /*
2985          * Check waiters first. We do not trust user space values at
2986          * all and we at least want to know if user space fiddled
2987          * with the futex value instead of blindly unlocking.
2988          */
2989         top_waiter = futex_top_waiter(hb, &key);
2990         if (top_waiter) {
2991                 struct futex_pi_state *pi_state = top_waiter->pi_state;
2992
2993                 ret = -EINVAL;
2994                 if (!pi_state)
2995                         goto out_unlock;
2996
2997                 /*
2998                  * If current does not own the pi_state then the futex is
2999                  * inconsistent and user space fiddled with the futex value.
3000                  */
3001                 if (pi_state->owner != current)
3002                         goto out_unlock;
3003
3004                 get_pi_state(pi_state);
3005                 /*
3006                  * By taking wait_lock while still holding hb->lock, we ensure
3007                  * there is no point where we hold neither; and therefore
3008                  * wake_futex_pi() must observe a state consistent with what we
3009                  * observed.
3010                  *
3011                  * In particular; this forces __rt_mutex_start_proxy() to
3012                  * complete such that we're guaranteed to observe the
3013                  * rt_waiter. Also see the WARN in wake_futex_pi().
3014                  */
3015                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3016                 spin_unlock(&hb->lock);
3017
3018                 /* drops pi_state->pi_mutex.wait_lock */
3019                 ret = wake_futex_pi(uaddr, uval, pi_state);
3020
3021                 put_pi_state(pi_state);
3022
3023                 /*
3024                  * Success, we're done! No tricky corner cases.
3025                  */
3026                 if (!ret)
3027                         goto out_putkey;
3028                 /*
3029                  * The atomic access to the futex value generated a
3030                  * pagefault, so retry the user-access and the wakeup:
3031                  */
3032                 if (ret == -EFAULT)
3033                         goto pi_faulted;
3034                 /*
3035                  * A unconditional UNLOCK_PI op raced against a waiter
3036                  * setting the FUTEX_WAITERS bit. Try again.
3037                  */
3038                 if (ret == -EAGAIN)
3039                         goto pi_retry;
3040                 /*
3041                  * wake_futex_pi has detected invalid state. Tell user
3042                  * space.
3043                  */
3044                 goto out_putkey;
3045         }
3046
3047         /*
3048          * We have no kernel internal state, i.e. no waiters in the
3049          * kernel. Waiters which are about to queue themselves are stuck
3050          * on hb->lock. So we can safely ignore them. We do neither
3051          * preserve the WAITERS bit not the OWNER_DIED one. We are the
3052          * owner.
3053          */
3054         if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3055                 spin_unlock(&hb->lock);
3056                 switch (ret) {
3057                 case -EFAULT:
3058                         goto pi_faulted;
3059
3060                 case -EAGAIN:
3061                         goto pi_retry;
3062
3063                 default:
3064                         WARN_ON_ONCE(1);
3065                         goto out_putkey;
3066                 }
3067         }
3068
3069         /*
3070          * If uval has changed, let user space handle it.
3071          */
3072         ret = (curval == uval) ? 0 : -EAGAIN;
3073
3074 out_unlock:
3075         spin_unlock(&hb->lock);
3076 out_putkey:
3077         return ret;
3078
3079 pi_retry:
3080         cond_resched();
3081         goto retry;
3082
3083 pi_faulted:
3084
3085         ret = fault_in_user_writeable(uaddr);
3086         if (!ret)
3087                 goto retry;
3088
3089         return ret;
3090 }
3091
3092 /**
3093  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3094  * @hb:         the hash_bucket futex_q was original enqueued on
3095  * @q:          the futex_q woken while waiting to be requeued
3096  * @key2:       the futex_key of the requeue target futex
3097  * @timeout:    the timeout associated with the wait (NULL if none)
3098  *
3099  * Detect if the task was woken on the initial futex as opposed to the requeue
3100  * target futex.  If so, determine if it was a timeout or a signal that caused
3101  * the wakeup and return the appropriate error code to the caller.  Must be
3102  * called with the hb lock held.
3103  *
3104  * Return:
3105  *  -  0 = no early wakeup detected;
3106  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3107  */
3108 static inline
3109 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3110                                    struct futex_q *q, union futex_key *key2,
3111                                    struct hrtimer_sleeper *timeout)
3112 {
3113         int ret = 0;
3114
3115         /*
3116          * With the hb lock held, we avoid races while we process the wakeup.
3117          * We only need to hold hb (and not hb2) to ensure atomicity as the
3118          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3119          * It can't be requeued from uaddr2 to something else since we don't
3120          * support a PI aware source futex for requeue.
3121          */
3122         if (!match_futex(&q->key, key2)) {
3123                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3124                 /*
3125                  * We were woken prior to requeue by a timeout or a signal.
3126                  * Unqueue the futex_q and determine which it was.
3127                  */
3128                 plist_del(&q->list, &hb->chain);
3129                 hb_waiters_dec(hb);
3130
3131                 /* Handle spurious wakeups gracefully */
3132                 ret = -EWOULDBLOCK;
3133                 if (timeout && !timeout->task)
3134                         ret = -ETIMEDOUT;
3135                 else if (signal_pending(current))
3136                         ret = -ERESTARTNOINTR;
3137         }
3138         return ret;
3139 }
3140
3141 /**
3142  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3143  * @uaddr:      the futex we initially wait on (non-pi)
3144  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3145  *              the same type, no requeueing from private to shared, etc.
3146  * @val:        the expected value of uaddr
3147  * @abs_time:   absolute timeout
3148  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3149  * @uaddr2:     the pi futex we will take prior to returning to user-space
3150  *
3151  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3152  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3153  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3154  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3155  * without one, the pi logic would not know which task to boost/deboost, if
3156  * there was a need to.
3157  *
3158  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3159  * via the following--
3160  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3161  * 2) wakeup on uaddr2 after a requeue
3162  * 3) signal
3163  * 4) timeout
3164  *
3165  * If 3, cleanup and return -ERESTARTNOINTR.
3166  *
3167  * If 2, we may then block on trying to take the rt_mutex and return via:
3168  * 5) successful lock
3169  * 6) signal
3170  * 7) timeout
3171  * 8) other lock acquisition failure
3172  *
3173  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3174  *
3175  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3176  *
3177  * Return:
3178  *  -  0 - On success;
3179  *  - <0 - On error
3180  */
3181 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3182                                  u32 val, ktime_t *abs_time, u32 bitset,
3183                                  u32 __user *uaddr2)
3184 {
3185         struct hrtimer_sleeper timeout, *to;
3186         struct futex_pi_state *pi_state = NULL;
3187         struct rt_mutex_waiter rt_waiter;
3188         struct futex_hash_bucket *hb;
3189         union futex_key key2 = FUTEX_KEY_INIT;
3190         struct futex_q q = futex_q_init;
3191         int res, ret;
3192
3193         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3194                 return -ENOSYS;
3195
3196         if (uaddr == uaddr2)
3197                 return -EINVAL;
3198
3199         if (!bitset)
3200                 return -EINVAL;
3201
3202         to = futex_setup_timer(abs_time, &timeout, flags,
3203                                current->timer_slack_ns);
3204
3205         /*
3206          * The waiter is allocated on our stack, manipulated by the requeue
3207          * code while we sleep on uaddr.
3208          */
3209         rt_mutex_init_waiter(&rt_waiter);
3210
3211         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3212         if (unlikely(ret != 0))
3213                 goto out;
3214
3215         q.bitset = bitset;
3216         q.rt_waiter = &rt_waiter;
3217         q.requeue_pi_key = &key2;
3218
3219         /*
3220          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3221          * count.
3222          */
3223         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3224         if (ret)
3225                 goto out;
3226
3227         /*
3228          * The check above which compares uaddrs is not sufficient for
3229          * shared futexes. We need to compare the keys:
3230          */
3231         if (match_futex(&q.key, &key2)) {
3232                 queue_unlock(hb);
3233                 ret = -EINVAL;
3234                 goto out;
3235         }
3236
3237         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3238         futex_wait_queue_me(hb, &q, to);
3239
3240         spin_lock(&hb->lock);
3241         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3242         spin_unlock(&hb->lock);
3243         if (ret)
3244                 goto out;
3245
3246         /*
3247          * In order for us to be here, we know our q.key == key2, and since
3248          * we took the hb->lock above, we also know that futex_requeue() has
3249          * completed and we no longer have to concern ourselves with a wakeup
3250          * race with the atomic proxy lock acquisition by the requeue code. The
3251          * futex_requeue dropped our key1 reference and incremented our key2
3252          * reference count.
3253          */
3254
3255         /* Check if the requeue code acquired the second futex for us. */
3256         if (!q.rt_waiter) {
3257                 /*
3258                  * Got the lock. We might not be the anticipated owner if we
3259                  * did a lock-steal - fix up the PI-state in that case.
3260                  */
3261                 if (q.pi_state && (q.pi_state->owner != current)) {
3262                         spin_lock(q.lock_ptr);
3263                         ret = fixup_pi_state_owner(uaddr2, &q, current);
3264                         if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3265                                 pi_state = q.pi_state;
3266                                 get_pi_state(pi_state);
3267                         }
3268                         /*
3269                          * Drop the reference to the pi state which
3270                          * the requeue_pi() code acquired for us.
3271                          */
3272                         put_pi_state(q.pi_state);
3273                         spin_unlock(q.lock_ptr);
3274                 }
3275         } else {
3276                 struct rt_mutex *pi_mutex;
3277
3278                 /*
3279                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3280                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3281                  * the pi_state.
3282                  */
3283                 WARN_ON(!q.pi_state);
3284                 pi_mutex = &q.pi_state->pi_mutex;
3285                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3286
3287                 spin_lock(q.lock_ptr);
3288                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3289                         ret = 0;
3290
3291                 debug_rt_mutex_free_waiter(&rt_waiter);
3292                 /*
3293                  * Fixup the pi_state owner and possibly acquire the lock if we
3294                  * haven't already.
3295                  */
3296                 res = fixup_owner(uaddr2, &q, !ret);
3297                 /*
3298                  * If fixup_owner() returned an error, proprogate that.  If it
3299                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3300                  */
3301                 if (res)
3302                         ret = (res < 0) ? res : 0;
3303
3304                 /*
3305                  * If fixup_pi_state_owner() faulted and was unable to handle
3306                  * the fault, unlock the rt_mutex and return the fault to
3307                  * userspace.
3308                  */
3309                 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3310                         pi_state = q.pi_state;
3311                         get_pi_state(pi_state);
3312                 }
3313
3314                 /* Unqueue and drop the lock. */
3315                 unqueue_me_pi(&q);
3316         }
3317
3318         if (pi_state) {
3319                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3320                 put_pi_state(pi_state);
3321         }
3322
3323         if (ret == -EINTR) {
3324                 /*
3325                  * We've already been requeued, but cannot restart by calling
3326                  * futex_lock_pi() directly. We could restart this syscall, but
3327                  * it would detect that the user space "val" changed and return
3328                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3329                  * -EWOULDBLOCK directly.
3330                  */
3331                 ret = -EWOULDBLOCK;
3332         }
3333
3334 out:
3335         if (to) {
3336                 hrtimer_cancel(&to->timer);
3337                 destroy_hrtimer_on_stack(&to->timer);
3338         }
3339         return ret;
3340 }
3341
3342 /*
3343  * Support for robust futexes: the kernel cleans up held futexes at
3344  * thread exit time.
3345  *
3346  * Implementation: user-space maintains a per-thread list of locks it
3347  * is holding. Upon do_exit(), the kernel carefully walks this list,
3348  * and marks all locks that are owned by this thread with the
3349  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3350  * always manipulated with the lock held, so the list is private and
3351  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3352  * field, to allow the kernel to clean up if the thread dies after
3353  * acquiring the lock, but just before it could have added itself to
3354  * the list. There can only be one such pending lock.
3355  */
3356
3357 /**
3358  * sys_set_robust_list() - Set the robust-futex list head of a task
3359  * @head:       pointer to the list-head
3360  * @len:        length of the list-head, as userspace expects
3361  */
3362 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3363                 size_t, len)
3364 {
3365         if (!futex_cmpxchg_enabled)
3366                 return -ENOSYS;
3367         /*
3368          * The kernel knows only one size for now:
3369          */
3370         if (unlikely(len != sizeof(*head)))
3371                 return -EINVAL;
3372
3373         current->robust_list = head;
3374
3375         return 0;
3376 }
3377
3378 /**
3379  * sys_get_robust_list() - Get the robust-futex list head of a task
3380  * @pid:        pid of the process [zero for current task]
3381  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
3382  * @len_ptr:    pointer to a length field, the kernel fills in the header size
3383  */
3384 SYSCALL_DEFINE3(get_robust_list, int, pid,
3385                 struct robust_list_head __user * __user *, head_ptr,
3386                 size_t __user *, len_ptr)
3387 {
3388         struct robust_list_head __user *head;
3389         unsigned long ret;
3390         struct task_struct *p;
3391
3392         if (!futex_cmpxchg_enabled)
3393                 return -ENOSYS;
3394
3395         rcu_read_lock();
3396
3397         ret = -ESRCH;
3398         if (!pid)
3399                 p = current;
3400         else {
3401                 p = find_task_by_vpid(pid);
3402                 if (!p)
3403                         goto err_unlock;
3404         }
3405
3406         ret = -EPERM;
3407         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3408                 goto err_unlock;
3409
3410         head = p->robust_list;
3411         rcu_read_unlock();
3412
3413         if (put_user(sizeof(*head), len_ptr))
3414                 return -EFAULT;
3415         return put_user(head, head_ptr);
3416
3417 err_unlock:
3418         rcu_read_unlock();
3419
3420         return ret;
3421 }
3422
3423 /* Constants for the pending_op argument of handle_futex_death */
3424 #define HANDLE_DEATH_PENDING    true
3425 #define HANDLE_DEATH_LIST       false
3426
3427 /*
3428  * Process a futex-list entry, check whether it's owned by the
3429  * dying task, and do notification if so:
3430  */
3431 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3432                               bool pi, bool pending_op)
3433 {
3434         u32 uval, nval, mval;
3435         int err;
3436
3437         /* Futex address must be 32bit aligned */
3438         if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3439                 return -1;
3440
3441 retry:
3442         if (get_user(uval, uaddr))
3443                 return -1;
3444
3445         /*
3446          * Special case for regular (non PI) futexes. The unlock path in
3447          * user space has two race scenarios:
3448          *
3449          * 1. The unlock path releases the user space futex value and
3450          *    before it can execute the futex() syscall to wake up
3451          *    waiters it is killed.
3452          *
3453          * 2. A woken up waiter is killed before it can acquire the
3454          *    futex in user space.
3455          *
3456          * In both cases the TID validation below prevents a wakeup of
3457          * potential waiters which can cause these waiters to block
3458          * forever.
3459          *
3460          * In both cases the following conditions are met:
3461          *
3462          *      1) task->robust_list->list_op_pending != NULL
3463          *         @pending_op == true
3464          *      2) User space futex value == 0
3465          *      3) Regular futex: @pi == false
3466          *
3467          * If these conditions are met, it is safe to attempt waking up a
3468          * potential waiter without touching the user space futex value and
3469          * trying to set the OWNER_DIED bit. The user space futex value is
3470          * uncontended and the rest of the user space mutex state is
3471          * consistent, so a woken waiter will just take over the
3472          * uncontended futex. Setting the OWNER_DIED bit would create
3473          * inconsistent state and malfunction of the user space owner died
3474          * handling.
3475          */
3476         if (pending_op && !pi && !uval) {
3477                 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3478                 return 0;
3479         }
3480
3481         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3482                 return 0;
3483
3484         /*
3485          * Ok, this dying thread is truly holding a futex
3486          * of interest. Set the OWNER_DIED bit atomically
3487          * via cmpxchg, and if the value had FUTEX_WAITERS
3488          * set, wake up a waiter (if any). (We have to do a
3489          * futex_wake() even if OWNER_DIED is already set -
3490          * to handle the rare but possible case of recursive
3491          * thread-death.) The rest of the cleanup is done in
3492          * userspace.
3493          */
3494         mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3495
3496         /*
3497          * We are not holding a lock here, but we want to have
3498          * the pagefault_disable/enable() protection because
3499          * we want to handle the fault gracefully. If the
3500          * access fails we try to fault in the futex with R/W
3501          * verification via get_user_pages. get_user() above
3502          * does not guarantee R/W access. If that fails we
3503          * give up and leave the futex locked.
3504          */
3505         if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3506                 switch (err) {
3507                 case -EFAULT:
3508                         if (fault_in_user_writeable(uaddr))
3509                                 return -1;
3510                         goto retry;
3511
3512                 case -EAGAIN:
3513                         cond_resched();
3514                         goto retry;
3515
3516                 default:
3517                         WARN_ON_ONCE(1);
3518                         return err;
3519                 }
3520         }
3521
3522         if (nval != uval)
3523                 goto retry;
3524
3525         /*
3526          * Wake robust non-PI futexes here. The wakeup of
3527          * PI futexes happens in exit_pi_state():
3528          */
3529         if (!pi && (uval & FUTEX_WAITERS))
3530                 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3531
3532         return 0;
3533 }
3534
3535 /*
3536  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3537  */
3538 static inline int fetch_robust_entry(struct robust_list __user **entry,
3539                                      struct robust_list __user * __user *head,
3540                                      unsigned int *pi)
3541 {
3542         unsigned long uentry;
3543
3544         if (get_user(uentry, (unsigned long __user *)head))
3545                 return -EFAULT;
3546
3547         *entry = (void __user *)(uentry & ~1UL);
3548         *pi = uentry & 1;
3549
3550         return 0;
3551 }
3552
3553 /*
3554  * Walk curr->robust_list (very carefully, it's a userspace list!)
3555  * and mark any locks found there dead, and notify any waiters.
3556  *
3557  * We silently return on any sign of list-walking problem.
3558  */
3559 static void exit_robust_list(struct task_struct *curr)
3560 {
3561         struct robust_list_head __user *head = curr->robust_list;
3562         struct robust_list __user *entry, *next_entry, *pending;
3563         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3564         unsigned int next_pi;
3565         unsigned long futex_offset;
3566         int rc;
3567
3568         if (!futex_cmpxchg_enabled)
3569                 return;
3570
3571         /*
3572          * Fetch the list head (which was registered earlier, via
3573          * sys_set_robust_list()):
3574          */
3575         if (fetch_robust_entry(&entry, &head->list.next, &pi))
3576                 return;
3577         /*
3578          * Fetch the relative futex offset:
3579          */
3580         if (get_user(futex_offset, &head->futex_offset))
3581                 return;
3582         /*
3583          * Fetch any possibly pending lock-add first, and handle it
3584          * if it exists:
3585          */
3586         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3587                 return;
3588
3589         next_entry = NULL;      /* avoid warning with gcc */
3590         while (entry != &head->list) {
3591                 /*
3592                  * Fetch the next entry in the list before calling
3593                  * handle_futex_death:
3594                  */
3595                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3596                 /*
3597                  * A pending lock might already be on the list, so
3598                  * don't process it twice:
3599                  */
3600                 if (entry != pending) {
3601                         if (handle_futex_death((void __user *)entry + futex_offset,
3602                                                 curr, pi, HANDLE_DEATH_LIST))
3603                                 return;
3604                 }
3605                 if (rc)
3606                         return;
3607                 entry = next_entry;
3608                 pi = next_pi;
3609                 /*
3610                  * Avoid excessively long or circular lists:
3611                  */
3612                 if (!--limit)
3613                         break;
3614
3615                 cond_resched();
3616         }
3617
3618         if (pending) {
3619                 handle_futex_death((void __user *)pending + futex_offset,
3620                                    curr, pip, HANDLE_DEATH_PENDING);
3621         }
3622 }
3623
3624 static void futex_cleanup(struct task_struct *tsk)
3625 {
3626         if (unlikely(tsk->robust_list)) {
3627                 exit_robust_list(tsk);
3628                 tsk->robust_list = NULL;
3629         }
3630
3631 #ifdef CONFIG_COMPAT
3632         if (unlikely(tsk->compat_robust_list)) {
3633                 compat_exit_robust_list(tsk);
3634                 tsk->compat_robust_list = NULL;
3635         }
3636 #endif
3637
3638         if (unlikely(!list_empty(&tsk->pi_state_list)))
3639                 exit_pi_state_list(tsk);
3640 }
3641
3642 /**
3643  * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3644  * @tsk:        task to set the state on
3645  *
3646  * Set the futex exit state of the task lockless. The futex waiter code
3647  * observes that state when a task is exiting and loops until the task has
3648  * actually finished the futex cleanup. The worst case for this is that the
3649  * waiter runs through the wait loop until the state becomes visible.
3650  *
3651  * This is called from the recursive fault handling path in do_exit().
3652  *
3653  * This is best effort. Either the futex exit code has run already or
3654  * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3655  * take it over. If not, the problem is pushed back to user space. If the
3656  * futex exit code did not run yet, then an already queued waiter might
3657  * block forever, but there is nothing which can be done about that.
3658  */
3659 void futex_exit_recursive(struct task_struct *tsk)
3660 {
3661         /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3662         if (tsk->futex_state == FUTEX_STATE_EXITING)
3663                 mutex_unlock(&tsk->futex_exit_mutex);
3664         tsk->futex_state = FUTEX_STATE_DEAD;
3665 }
3666
3667 static void futex_cleanup_begin(struct task_struct *tsk)
3668 {
3669         /*
3670          * Prevent various race issues against a concurrent incoming waiter
3671          * including live locks by forcing the waiter to block on
3672          * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3673          * attach_to_pi_owner().
3674          */
3675         mutex_lock(&tsk->futex_exit_mutex);
3676
3677         /*
3678          * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3679          *
3680          * This ensures that all subsequent checks of tsk->futex_state in
3681          * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3682          * tsk->pi_lock held.
3683          *
3684          * It guarantees also that a pi_state which was queued right before
3685          * the state change under tsk->pi_lock by a concurrent waiter must
3686          * be observed in exit_pi_state_list().
3687          */
3688         raw_spin_lock_irq(&tsk->pi_lock);
3689         tsk->futex_state = FUTEX_STATE_EXITING;
3690         raw_spin_unlock_irq(&tsk->pi_lock);
3691 }
3692
3693 static void futex_cleanup_end(struct task_struct *tsk, int state)
3694 {
3695         /*
3696          * Lockless store. The only side effect is that an observer might
3697          * take another loop until it becomes visible.
3698          */
3699         tsk->futex_state = state;
3700         /*
3701          * Drop the exit protection. This unblocks waiters which observed
3702          * FUTEX_STATE_EXITING to reevaluate the state.
3703          */
3704         mutex_unlock(&tsk->futex_exit_mutex);
3705 }
3706
3707 void futex_exec_release(struct task_struct *tsk)
3708 {
3709         /*
3710          * The state handling is done for consistency, but in the case of
3711          * exec() there is no way to prevent futher damage as the PID stays
3712          * the same. But for the unlikely and arguably buggy case that a
3713          * futex is held on exec(), this provides at least as much state
3714          * consistency protection which is possible.
3715          */
3716         futex_cleanup_begin(tsk);
3717         futex_cleanup(tsk);
3718         /*
3719          * Reset the state to FUTEX_STATE_OK. The task is alive and about
3720          * exec a new binary.
3721          */
3722         futex_cleanup_end(tsk, FUTEX_STATE_OK);
3723 }
3724
3725 void futex_exit_release(struct task_struct *tsk)
3726 {
3727         futex_cleanup_begin(tsk);
3728         futex_cleanup(tsk);
3729         futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3730 }
3731
3732 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3733                 u32 __user *uaddr2, u32 val2, u32 val3)
3734 {
3735         int cmd = op & FUTEX_CMD_MASK;
3736         unsigned int flags = 0;
3737
3738         if (!(op & FUTEX_PRIVATE_FLAG))
3739                 flags |= FLAGS_SHARED;
3740
3741         if (op & FUTEX_CLOCK_REALTIME) {
3742                 flags |= FLAGS_CLOCKRT;
3743                 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3744                     cmd != FUTEX_WAIT_REQUEUE_PI)
3745                         return -ENOSYS;
3746         }
3747
3748         switch (cmd) {
3749         case FUTEX_LOCK_PI:
3750         case FUTEX_UNLOCK_PI:
3751         case FUTEX_TRYLOCK_PI:
3752         case FUTEX_WAIT_REQUEUE_PI:
3753         case FUTEX_CMP_REQUEUE_PI:
3754                 if (!futex_cmpxchg_enabled)
3755                         return -ENOSYS;
3756         }
3757
3758         switch (cmd) {
3759         case FUTEX_WAIT:
3760                 val3 = FUTEX_BITSET_MATCH_ANY;
3761                 fallthrough;
3762         case FUTEX_WAIT_BITSET:
3763                 return futex_wait(uaddr, flags, val, timeout, val3);
3764         case FUTEX_WAKE:
3765                 val3 = FUTEX_BITSET_MATCH_ANY;
3766                 fallthrough;
3767         case FUTEX_WAKE_BITSET:
3768                 return futex_wake(uaddr, flags, val, val3);
3769         case FUTEX_REQUEUE:
3770                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3771         case FUTEX_CMP_REQUEUE:
3772                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3773         case FUTEX_WAKE_OP:
3774                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3775         case FUTEX_LOCK_PI:
3776                 return futex_lock_pi(uaddr, flags, timeout, 0);
3777         case FUTEX_UNLOCK_PI:
3778                 return futex_unlock_pi(uaddr, flags);
3779         case FUTEX_TRYLOCK_PI:
3780                 return futex_lock_pi(uaddr, flags, NULL, 1);
3781         case FUTEX_WAIT_REQUEUE_PI:
3782                 val3 = FUTEX_BITSET_MATCH_ANY;
3783                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3784                                              uaddr2);
3785         case FUTEX_CMP_REQUEUE_PI:
3786                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3787         }
3788         return -ENOSYS;
3789 }
3790
3791
3792 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3793                 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3794                 u32, val3)
3795 {
3796         struct timespec64 ts;
3797         ktime_t t, *tp = NULL;
3798         u32 val2 = 0;
3799         int cmd = op & FUTEX_CMD_MASK;
3800
3801         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3802                       cmd == FUTEX_WAIT_BITSET ||
3803                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3804                 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3805                         return -EFAULT;
3806                 if (get_timespec64(&ts, utime))
3807                         return -EFAULT;
3808                 if (!timespec64_valid(&ts))
3809                         return -EINVAL;
3810
3811                 t = timespec64_to_ktime(ts);
3812                 if (cmd == FUTEX_WAIT)
3813                         t = ktime_add_safe(ktime_get(), t);
3814                 else if (!(op & FUTEX_CLOCK_REALTIME))
3815                         t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3816                 tp = &t;
3817         }
3818         /*
3819          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3820          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3821          */
3822         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3823             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3824                 val2 = (u32) (unsigned long) utime;
3825
3826         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3827 }
3828
3829 #ifdef CONFIG_COMPAT
3830 /*
3831  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3832  */
3833 static inline int
3834 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3835                    compat_uptr_t __user *head, unsigned int *pi)
3836 {
3837         if (get_user(*uentry, head))
3838                 return -EFAULT;
3839
3840         *entry = compat_ptr((*uentry) & ~1);
3841         *pi = (unsigned int)(*uentry) & 1;
3842
3843         return 0;
3844 }
3845
3846 static void __user *futex_uaddr(struct robust_list __user *entry,
3847                                 compat_long_t futex_offset)
3848 {
3849         compat_uptr_t base = ptr_to_compat(entry);
3850         void __user *uaddr = compat_ptr(base + futex_offset);
3851
3852         return uaddr;
3853 }
3854
3855 /*
3856  * Walk curr->robust_list (very carefully, it's a userspace list!)
3857  * and mark any locks found there dead, and notify any waiters.
3858  *
3859  * We silently return on any sign of list-walking problem.
3860  */
3861 static void compat_exit_robust_list(struct task_struct *curr)
3862 {
3863         struct compat_robust_list_head __user *head = curr->compat_robust_list;
3864         struct robust_list __user *entry, *next_entry, *pending;
3865         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3866         unsigned int next_pi;
3867         compat_uptr_t uentry, next_uentry, upending;
3868         compat_long_t futex_offset;
3869         int rc;
3870
3871         if (!futex_cmpxchg_enabled)
3872                 return;
3873
3874         /*
3875          * Fetch the list head (which was registered earlier, via
3876          * sys_set_robust_list()):
3877          */
3878         if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3879                 return;
3880         /*
3881          * Fetch the relative futex offset:
3882          */
3883         if (get_user(futex_offset, &head->futex_offset))
3884                 return;
3885         /*
3886          * Fetch any possibly pending lock-add first, and handle it
3887          * if it exists:
3888          */
3889         if (compat_fetch_robust_entry(&upending, &pending,
3890                                &head->list_op_pending, &pip))
3891                 return;
3892
3893         next_entry = NULL;      /* avoid warning with gcc */
3894         while (entry != (struct robust_list __user *) &head->list) {
3895                 /*
3896                  * Fetch the next entry in the list before calling
3897                  * handle_futex_death:
3898                  */
3899                 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3900                         (compat_uptr_t __user *)&entry->next, &next_pi);
3901                 /*
3902                  * A pending lock might already be on the list, so
3903                  * dont process it twice:
3904                  */
3905                 if (entry != pending) {
3906                         void __user *uaddr = futex_uaddr(entry, futex_offset);
3907
3908                         if (handle_futex_death(uaddr, curr, pi,
3909                                                HANDLE_DEATH_LIST))
3910                                 return;
3911                 }
3912                 if (rc)
3913                         return;
3914                 uentry = next_uentry;
3915                 entry = next_entry;
3916                 pi = next_pi;
3917                 /*
3918                  * Avoid excessively long or circular lists:
3919                  */
3920                 if (!--limit)
3921                         break;
3922
3923                 cond_resched();
3924         }
3925         if (pending) {
3926                 void __user *uaddr = futex_uaddr(pending, futex_offset);
3927
3928                 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3929         }
3930 }
3931
3932 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3933                 struct compat_robust_list_head __user *, head,
3934                 compat_size_t, len)
3935 {
3936         if (!futex_cmpxchg_enabled)
3937                 return -ENOSYS;
3938
3939         if (unlikely(len != sizeof(*head)))
3940                 return -EINVAL;
3941
3942         current->compat_robust_list = head;
3943
3944         return 0;
3945 }
3946
3947 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3948                         compat_uptr_t __user *, head_ptr,
3949                         compat_size_t __user *, len_ptr)
3950 {
3951         struct compat_robust_list_head __user *head;
3952         unsigned long ret;
3953         struct task_struct *p;
3954
3955         if (!futex_cmpxchg_enabled)
3956                 return -ENOSYS;
3957
3958         rcu_read_lock();
3959
3960         ret = -ESRCH;
3961         if (!pid)
3962                 p = current;
3963         else {
3964                 p = find_task_by_vpid(pid);
3965                 if (!p)
3966                         goto err_unlock;
3967         }
3968
3969         ret = -EPERM;
3970         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3971                 goto err_unlock;
3972
3973         head = p->compat_robust_list;
3974         rcu_read_unlock();
3975
3976         if (put_user(sizeof(*head), len_ptr))
3977                 return -EFAULT;
3978         return put_user(ptr_to_compat(head), head_ptr);
3979
3980 err_unlock:
3981         rcu_read_unlock();
3982
3983         return ret;
3984 }
3985 #endif /* CONFIG_COMPAT */
3986
3987 #ifdef CONFIG_COMPAT_32BIT_TIME
3988 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3989                 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3990                 u32, val3)
3991 {
3992         struct timespec64 ts;
3993         ktime_t t, *tp = NULL;
3994         int val2 = 0;
3995         int cmd = op & FUTEX_CMD_MASK;
3996
3997         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3998                       cmd == FUTEX_WAIT_BITSET ||
3999                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
4000                 if (get_old_timespec32(&ts, utime))
4001                         return -EFAULT;
4002                 if (!timespec64_valid(&ts))
4003                         return -EINVAL;
4004
4005                 t = timespec64_to_ktime(ts);
4006                 if (cmd == FUTEX_WAIT)
4007                         t = ktime_add_safe(ktime_get(), t);
4008                 else if (!(op & FUTEX_CLOCK_REALTIME))
4009                         t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
4010                 tp = &t;
4011         }
4012         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4013             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4014                 val2 = (int) (unsigned long) utime;
4015
4016         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4017 }
4018 #endif /* CONFIG_COMPAT_32BIT_TIME */
4019
4020 static void __init futex_detect_cmpxchg(void)
4021 {
4022 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4023         u32 curval;
4024
4025         /*
4026          * This will fail and we want it. Some arch implementations do
4027          * runtime detection of the futex_atomic_cmpxchg_inatomic()
4028          * functionality. We want to know that before we call in any
4029          * of the complex code paths. Also we want to prevent
4030          * registration of robust lists in that case. NULL is
4031          * guaranteed to fault and we get -EFAULT on functional
4032          * implementation, the non-functional ones will return
4033          * -ENOSYS.
4034          */
4035         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4036                 futex_cmpxchg_enabled = 1;
4037 #endif
4038 }
4039
4040 static int __init futex_init(void)
4041 {
4042         unsigned int futex_shift;
4043         unsigned long i;
4044
4045 #if CONFIG_BASE_SMALL
4046         futex_hashsize = 16;
4047 #else
4048         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4049 #endif
4050
4051         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4052                                                futex_hashsize, 0,
4053                                                futex_hashsize < 256 ? HASH_SMALL : 0,
4054                                                &futex_shift, NULL,
4055                                                futex_hashsize, futex_hashsize);
4056         futex_hashsize = 1UL << futex_shift;
4057
4058         futex_detect_cmpxchg();
4059
4060         for (i = 0; i < futex_hashsize; i++) {
4061                 atomic_set(&futex_queues[i].waiters, 0);
4062                 plist_head_init(&futex_queues[i].chain);
4063                 spin_lock_init(&futex_queues[i].lock);
4064         }
4065
4066         return 0;
4067 }
4068 core_initcall(futex_init);