Merge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[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 pi_state_update_owner(struct futex_pi_state *pi_state,
767                                   struct task_struct *new_owner)
768 {
769         struct task_struct *old_owner = pi_state->owner;
770
771         lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
772
773         if (old_owner) {
774                 raw_spin_lock(&old_owner->pi_lock);
775                 WARN_ON(list_empty(&pi_state->list));
776                 list_del_init(&pi_state->list);
777                 raw_spin_unlock(&old_owner->pi_lock);
778         }
779
780         if (new_owner) {
781                 raw_spin_lock(&new_owner->pi_lock);
782                 WARN_ON(!list_empty(&pi_state->list));
783                 list_add(&pi_state->list, &new_owner->pi_state_list);
784                 pi_state->owner = new_owner;
785                 raw_spin_unlock(&new_owner->pi_lock);
786         }
787 }
788
789 static void get_pi_state(struct futex_pi_state *pi_state)
790 {
791         WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
792 }
793
794 /*
795  * Drops a reference to the pi_state object and frees or caches it
796  * when the last reference is gone.
797  */
798 static void put_pi_state(struct futex_pi_state *pi_state)
799 {
800         if (!pi_state)
801                 return;
802
803         if (!refcount_dec_and_test(&pi_state->refcount))
804                 return;
805
806         /*
807          * If pi_state->owner is NULL, the owner is most probably dying
808          * and has cleaned up the pi_state already
809          */
810         if (pi_state->owner) {
811                 unsigned long flags;
812
813                 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
814                 pi_state_update_owner(pi_state, NULL);
815                 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
816                 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
817         }
818
819         if (current->pi_state_cache) {
820                 kfree(pi_state);
821         } else {
822                 /*
823                  * pi_state->list is already empty.
824                  * clear pi_state->owner.
825                  * refcount is at 0 - put it back to 1.
826                  */
827                 pi_state->owner = NULL;
828                 refcount_set(&pi_state->refcount, 1);
829                 current->pi_state_cache = pi_state;
830         }
831 }
832
833 #ifdef CONFIG_FUTEX_PI
834
835 /*
836  * This task is holding PI mutexes at exit time => bad.
837  * Kernel cleans up PI-state, but userspace is likely hosed.
838  * (Robust-futex cleanup is separate and might save the day for userspace.)
839  */
840 static void exit_pi_state_list(struct task_struct *curr)
841 {
842         struct list_head *next, *head = &curr->pi_state_list;
843         struct futex_pi_state *pi_state;
844         struct futex_hash_bucket *hb;
845         union futex_key key = FUTEX_KEY_INIT;
846
847         if (!futex_cmpxchg_enabled)
848                 return;
849         /*
850          * We are a ZOMBIE and nobody can enqueue itself on
851          * pi_state_list anymore, but we have to be careful
852          * versus waiters unqueueing themselves:
853          */
854         raw_spin_lock_irq(&curr->pi_lock);
855         while (!list_empty(head)) {
856                 next = head->next;
857                 pi_state = list_entry(next, struct futex_pi_state, list);
858                 key = pi_state->key;
859                 hb = hash_futex(&key);
860
861                 /*
862                  * We can race against put_pi_state() removing itself from the
863                  * list (a waiter going away). put_pi_state() will first
864                  * decrement the reference count and then modify the list, so
865                  * its possible to see the list entry but fail this reference
866                  * acquire.
867                  *
868                  * In that case; drop the locks to let put_pi_state() make
869                  * progress and retry the loop.
870                  */
871                 if (!refcount_inc_not_zero(&pi_state->refcount)) {
872                         raw_spin_unlock_irq(&curr->pi_lock);
873                         cpu_relax();
874                         raw_spin_lock_irq(&curr->pi_lock);
875                         continue;
876                 }
877                 raw_spin_unlock_irq(&curr->pi_lock);
878
879                 spin_lock(&hb->lock);
880                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
881                 raw_spin_lock(&curr->pi_lock);
882                 /*
883                  * We dropped the pi-lock, so re-check whether this
884                  * task still owns the PI-state:
885                  */
886                 if (head->next != next) {
887                         /* retain curr->pi_lock for the loop invariant */
888                         raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
889                         spin_unlock(&hb->lock);
890                         put_pi_state(pi_state);
891                         continue;
892                 }
893
894                 WARN_ON(pi_state->owner != curr);
895                 WARN_ON(list_empty(&pi_state->list));
896                 list_del_init(&pi_state->list);
897                 pi_state->owner = NULL;
898
899                 raw_spin_unlock(&curr->pi_lock);
900                 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
901                 spin_unlock(&hb->lock);
902
903                 rt_mutex_futex_unlock(&pi_state->pi_mutex);
904                 put_pi_state(pi_state);
905
906                 raw_spin_lock_irq(&curr->pi_lock);
907         }
908         raw_spin_unlock_irq(&curr->pi_lock);
909 }
910 #else
911 static inline void exit_pi_state_list(struct task_struct *curr) { }
912 #endif
913
914 /*
915  * We need to check the following states:
916  *
917  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
918  *
919  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
920  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
921  *
922  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
923  *
924  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
925  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
926  *
927  * [6]  Found  | Found    | task      | 0         | 1      | Valid
928  *
929  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
930  *
931  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
932  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
933  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
934  *
935  * [1]  Indicates that the kernel can acquire the futex atomically. We
936  *      came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
937  *
938  * [2]  Valid, if TID does not belong to a kernel thread. If no matching
939  *      thread is found then it indicates that the owner TID has died.
940  *
941  * [3]  Invalid. The waiter is queued on a non PI futex
942  *
943  * [4]  Valid state after exit_robust_list(), which sets the user space
944  *      value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
945  *
946  * [5]  The user space value got manipulated between exit_robust_list()
947  *      and exit_pi_state_list()
948  *
949  * [6]  Valid state after exit_pi_state_list() which sets the new owner in
950  *      the pi_state but cannot access the user space value.
951  *
952  * [7]  pi_state->owner can only be NULL when the OWNER_DIED bit is set.
953  *
954  * [8]  Owner and user space value match
955  *
956  * [9]  There is no transient state which sets the user space TID to 0
957  *      except exit_robust_list(), but this is indicated by the
958  *      FUTEX_OWNER_DIED bit. See [4]
959  *
960  * [10] There is no transient state which leaves owner and user space
961  *      TID out of sync. Except one error case where the kernel is denied
962  *      write access to the user address, see fixup_pi_state_owner().
963  *
964  *
965  * Serialization and lifetime rules:
966  *
967  * hb->lock:
968  *
969  *      hb -> futex_q, relation
970  *      futex_q -> pi_state, relation
971  *
972  *      (cannot be raw because hb can contain arbitrary amount
973  *       of futex_q's)
974  *
975  * pi_mutex->wait_lock:
976  *
977  *      {uval, pi_state}
978  *
979  *      (and pi_mutex 'obviously')
980  *
981  * p->pi_lock:
982  *
983  *      p->pi_state_list -> pi_state->list, relation
984  *      pi_mutex->owner -> pi_state->owner, relation
985  *
986  * pi_state->refcount:
987  *
988  *      pi_state lifetime
989  *
990  *
991  * Lock order:
992  *
993  *   hb->lock
994  *     pi_mutex->wait_lock
995  *       p->pi_lock
996  *
997  */
998
999 /*
1000  * Validate that the existing waiter has a pi_state and sanity check
1001  * the pi_state against the user space value. If correct, attach to
1002  * it.
1003  */
1004 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1005                               struct futex_pi_state *pi_state,
1006                               struct futex_pi_state **ps)
1007 {
1008         pid_t pid = uval & FUTEX_TID_MASK;
1009         u32 uval2;
1010         int ret;
1011
1012         /*
1013          * Userspace might have messed up non-PI and PI futexes [3]
1014          */
1015         if (unlikely(!pi_state))
1016                 return -EINVAL;
1017
1018         /*
1019          * We get here with hb->lock held, and having found a
1020          * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1021          * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1022          * which in turn means that futex_lock_pi() still has a reference on
1023          * our pi_state.
1024          *
1025          * The waiter holding a reference on @pi_state also protects against
1026          * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1027          * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1028          * free pi_state before we can take a reference ourselves.
1029          */
1030         WARN_ON(!refcount_read(&pi_state->refcount));
1031
1032         /*
1033          * Now that we have a pi_state, we can acquire wait_lock
1034          * and do the state validation.
1035          */
1036         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1037
1038         /*
1039          * Since {uval, pi_state} is serialized by wait_lock, and our current
1040          * uval was read without holding it, it can have changed. Verify it
1041          * still is what we expect it to be, otherwise retry the entire
1042          * operation.
1043          */
1044         if (get_futex_value_locked(&uval2, uaddr))
1045                 goto out_efault;
1046
1047         if (uval != uval2)
1048                 goto out_eagain;
1049
1050         /*
1051          * Handle the owner died case:
1052          */
1053         if (uval & FUTEX_OWNER_DIED) {
1054                 /*
1055                  * exit_pi_state_list sets owner to NULL and wakes the
1056                  * topmost waiter. The task which acquires the
1057                  * pi_state->rt_mutex will fixup owner.
1058                  */
1059                 if (!pi_state->owner) {
1060                         /*
1061                          * No pi state owner, but the user space TID
1062                          * is not 0. Inconsistent state. [5]
1063                          */
1064                         if (pid)
1065                                 goto out_einval;
1066                         /*
1067                          * Take a ref on the state and return success. [4]
1068                          */
1069                         goto out_attach;
1070                 }
1071
1072                 /*
1073                  * If TID is 0, then either the dying owner has not
1074                  * yet executed exit_pi_state_list() or some waiter
1075                  * acquired the rtmutex in the pi state, but did not
1076                  * yet fixup the TID in user space.
1077                  *
1078                  * Take a ref on the state and return success. [6]
1079                  */
1080                 if (!pid)
1081                         goto out_attach;
1082         } else {
1083                 /*
1084                  * If the owner died bit is not set, then the pi_state
1085                  * must have an owner. [7]
1086                  */
1087                 if (!pi_state->owner)
1088                         goto out_einval;
1089         }
1090
1091         /*
1092          * Bail out if user space manipulated the futex value. If pi
1093          * state exists then the owner TID must be the same as the
1094          * user space TID. [9/10]
1095          */
1096         if (pid != task_pid_vnr(pi_state->owner))
1097                 goto out_einval;
1098
1099 out_attach:
1100         get_pi_state(pi_state);
1101         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1102         *ps = pi_state;
1103         return 0;
1104
1105 out_einval:
1106         ret = -EINVAL;
1107         goto out_error;
1108
1109 out_eagain:
1110         ret = -EAGAIN;
1111         goto out_error;
1112
1113 out_efault:
1114         ret = -EFAULT;
1115         goto out_error;
1116
1117 out_error:
1118         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1119         return ret;
1120 }
1121
1122 /**
1123  * wait_for_owner_exiting - Block until the owner has exited
1124  * @ret: owner's current futex lock status
1125  * @exiting:    Pointer to the exiting task
1126  *
1127  * Caller must hold a refcount on @exiting.
1128  */
1129 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1130 {
1131         if (ret != -EBUSY) {
1132                 WARN_ON_ONCE(exiting);
1133                 return;
1134         }
1135
1136         if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1137                 return;
1138
1139         mutex_lock(&exiting->futex_exit_mutex);
1140         /*
1141          * No point in doing state checking here. If the waiter got here
1142          * while the task was in exec()->exec_futex_release() then it can
1143          * have any FUTEX_STATE_* value when the waiter has acquired the
1144          * mutex. OK, if running, EXITING or DEAD if it reached exit()
1145          * already. Highly unlikely and not a problem. Just one more round
1146          * through the futex maze.
1147          */
1148         mutex_unlock(&exiting->futex_exit_mutex);
1149
1150         put_task_struct(exiting);
1151 }
1152
1153 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1154                             struct task_struct *tsk)
1155 {
1156         u32 uval2;
1157
1158         /*
1159          * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1160          * caller that the alleged owner is busy.
1161          */
1162         if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1163                 return -EBUSY;
1164
1165         /*
1166          * Reread the user space value to handle the following situation:
1167          *
1168          * CPU0                         CPU1
1169          *
1170          * sys_exit()                   sys_futex()
1171          *  do_exit()                    futex_lock_pi()
1172          *                                futex_lock_pi_atomic()
1173          *   exit_signals(tsk)              No waiters:
1174          *    tsk->flags |= PF_EXITING;     *uaddr == 0x00000PID
1175          *  mm_release(tsk)                 Set waiter bit
1176          *   exit_robust_list(tsk) {        *uaddr = 0x80000PID;
1177          *      Set owner died              attach_to_pi_owner() {
1178          *    *uaddr = 0xC0000000;           tsk = get_task(PID);
1179          *   }                               if (!tsk->flags & PF_EXITING) {
1180          *  ...                                attach();
1181          *  tsk->futex_state =               } else {
1182          *      FUTEX_STATE_DEAD;              if (tsk->futex_state !=
1183          *                                        FUTEX_STATE_DEAD)
1184          *                                       return -EAGAIN;
1185          *                                     return -ESRCH; <--- FAIL
1186          *                                   }
1187          *
1188          * Returning ESRCH unconditionally is wrong here because the
1189          * user space value has been changed by the exiting task.
1190          *
1191          * The same logic applies to the case where the exiting task is
1192          * already gone.
1193          */
1194         if (get_futex_value_locked(&uval2, uaddr))
1195                 return -EFAULT;
1196
1197         /* If the user space value has changed, try again. */
1198         if (uval2 != uval)
1199                 return -EAGAIN;
1200
1201         /*
1202          * The exiting task did not have a robust list, the robust list was
1203          * corrupted or the user space value in *uaddr is simply bogus.
1204          * Give up and tell user space.
1205          */
1206         return -ESRCH;
1207 }
1208
1209 /*
1210  * Lookup the task for the TID provided from user space and attach to
1211  * it after doing proper sanity checks.
1212  */
1213 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1214                               struct futex_pi_state **ps,
1215                               struct task_struct **exiting)
1216 {
1217         pid_t pid = uval & FUTEX_TID_MASK;
1218         struct futex_pi_state *pi_state;
1219         struct task_struct *p;
1220
1221         /*
1222          * We are the first waiter - try to look up the real owner and attach
1223          * the new pi_state to it, but bail out when TID = 0 [1]
1224          *
1225          * The !pid check is paranoid. None of the call sites should end up
1226          * with pid == 0, but better safe than sorry. Let the caller retry
1227          */
1228         if (!pid)
1229                 return -EAGAIN;
1230         p = find_get_task_by_vpid(pid);
1231         if (!p)
1232                 return handle_exit_race(uaddr, uval, NULL);
1233
1234         if (unlikely(p->flags & PF_KTHREAD)) {
1235                 put_task_struct(p);
1236                 return -EPERM;
1237         }
1238
1239         /*
1240          * We need to look at the task state to figure out, whether the
1241          * task is exiting. To protect against the change of the task state
1242          * in futex_exit_release(), we do this protected by p->pi_lock:
1243          */
1244         raw_spin_lock_irq(&p->pi_lock);
1245         if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1246                 /*
1247                  * The task is on the way out. When the futex state is
1248                  * FUTEX_STATE_DEAD, we know that the task has finished
1249                  * the cleanup:
1250                  */
1251                 int ret = handle_exit_race(uaddr, uval, p);
1252
1253                 raw_spin_unlock_irq(&p->pi_lock);
1254                 /*
1255                  * If the owner task is between FUTEX_STATE_EXITING and
1256                  * FUTEX_STATE_DEAD then store the task pointer and keep
1257                  * the reference on the task struct. The calling code will
1258                  * drop all locks, wait for the task to reach
1259                  * FUTEX_STATE_DEAD and then drop the refcount. This is
1260                  * required to prevent a live lock when the current task
1261                  * preempted the exiting task between the two states.
1262                  */
1263                 if (ret == -EBUSY)
1264                         *exiting = p;
1265                 else
1266                         put_task_struct(p);
1267                 return ret;
1268         }
1269
1270         /*
1271          * No existing pi state. First waiter. [2]
1272          *
1273          * This creates pi_state, we have hb->lock held, this means nothing can
1274          * observe this state, wait_lock is irrelevant.
1275          */
1276         pi_state = alloc_pi_state();
1277
1278         /*
1279          * Initialize the pi_mutex in locked state and make @p
1280          * the owner of it:
1281          */
1282         rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1283
1284         /* Store the key for possible exit cleanups: */
1285         pi_state->key = *key;
1286
1287         WARN_ON(!list_empty(&pi_state->list));
1288         list_add(&pi_state->list, &p->pi_state_list);
1289         /*
1290          * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1291          * because there is no concurrency as the object is not published yet.
1292          */
1293         pi_state->owner = p;
1294         raw_spin_unlock_irq(&p->pi_lock);
1295
1296         put_task_struct(p);
1297
1298         *ps = pi_state;
1299
1300         return 0;
1301 }
1302
1303 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1304                            struct futex_hash_bucket *hb,
1305                            union futex_key *key, struct futex_pi_state **ps,
1306                            struct task_struct **exiting)
1307 {
1308         struct futex_q *top_waiter = futex_top_waiter(hb, key);
1309
1310         /*
1311          * If there is a waiter on that futex, validate it and
1312          * attach to the pi_state when the validation succeeds.
1313          */
1314         if (top_waiter)
1315                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1316
1317         /*
1318          * We are the first waiter - try to look up the owner based on
1319          * @uval and attach to it.
1320          */
1321         return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1322 }
1323
1324 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1325 {
1326         int err;
1327         u32 curval;
1328
1329         if (unlikely(should_fail_futex(true)))
1330                 return -EFAULT;
1331
1332         err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1333         if (unlikely(err))
1334                 return err;
1335
1336         /* If user space value changed, let the caller retry */
1337         return curval != uval ? -EAGAIN : 0;
1338 }
1339
1340 /**
1341  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1342  * @uaddr:              the pi futex user address
1343  * @hb:                 the pi futex hash bucket
1344  * @key:                the futex key associated with uaddr and hb
1345  * @ps:                 the pi_state pointer where we store the result of the
1346  *                      lookup
1347  * @task:               the task to perform the atomic lock work for.  This will
1348  *                      be "current" except in the case of requeue pi.
1349  * @exiting:            Pointer to store the task pointer of the owner task
1350  *                      which is in the middle of exiting
1351  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1352  *
1353  * Return:
1354  *  -  0 - ready to wait;
1355  *  -  1 - acquired the lock;
1356  *  - <0 - error
1357  *
1358  * The hb->lock and futex_key refs shall be held by the caller.
1359  *
1360  * @exiting is only set when the return value is -EBUSY. If so, this holds
1361  * a refcount on the exiting task on return and the caller needs to drop it
1362  * after waiting for the exit to complete.
1363  */
1364 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1365                                 union futex_key *key,
1366                                 struct futex_pi_state **ps,
1367                                 struct task_struct *task,
1368                                 struct task_struct **exiting,
1369                                 int set_waiters)
1370 {
1371         u32 uval, newval, vpid = task_pid_vnr(task);
1372         struct futex_q *top_waiter;
1373         int ret;
1374
1375         /*
1376          * Read the user space value first so we can validate a few
1377          * things before proceeding further.
1378          */
1379         if (get_futex_value_locked(&uval, uaddr))
1380                 return -EFAULT;
1381
1382         if (unlikely(should_fail_futex(true)))
1383                 return -EFAULT;
1384
1385         /*
1386          * Detect deadlocks.
1387          */
1388         if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1389                 return -EDEADLK;
1390
1391         if ((unlikely(should_fail_futex(true))))
1392                 return -EDEADLK;
1393
1394         /*
1395          * Lookup existing state first. If it exists, try to attach to
1396          * its pi_state.
1397          */
1398         top_waiter = futex_top_waiter(hb, key);
1399         if (top_waiter)
1400                 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1401
1402         /*
1403          * No waiter and user TID is 0. We are here because the
1404          * waiters or the owner died bit is set or called from
1405          * requeue_cmp_pi or for whatever reason something took the
1406          * syscall.
1407          */
1408         if (!(uval & FUTEX_TID_MASK)) {
1409                 /*
1410                  * We take over the futex. No other waiters and the user space
1411                  * TID is 0. We preserve the owner died bit.
1412                  */
1413                 newval = uval & FUTEX_OWNER_DIED;
1414                 newval |= vpid;
1415
1416                 /* The futex requeue_pi code can enforce the waiters bit */
1417                 if (set_waiters)
1418                         newval |= FUTEX_WAITERS;
1419
1420                 ret = lock_pi_update_atomic(uaddr, uval, newval);
1421                 /* If the take over worked, return 1 */
1422                 return ret < 0 ? ret : 1;
1423         }
1424
1425         /*
1426          * First waiter. Set the waiters bit before attaching ourself to
1427          * the owner. If owner tries to unlock, it will be forced into
1428          * the kernel and blocked on hb->lock.
1429          */
1430         newval = uval | FUTEX_WAITERS;
1431         ret = lock_pi_update_atomic(uaddr, uval, newval);
1432         if (ret)
1433                 return ret;
1434         /*
1435          * If the update of the user space value succeeded, we try to
1436          * attach to the owner. If that fails, no harm done, we only
1437          * set the FUTEX_WAITERS bit in the user space variable.
1438          */
1439         return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1440 }
1441
1442 /**
1443  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1444  * @q:  The futex_q to unqueue
1445  *
1446  * The q->lock_ptr must not be NULL and must be held by the caller.
1447  */
1448 static void __unqueue_futex(struct futex_q *q)
1449 {
1450         struct futex_hash_bucket *hb;
1451
1452         if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1453                 return;
1454         lockdep_assert_held(q->lock_ptr);
1455
1456         hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1457         plist_del(&q->list, &hb->chain);
1458         hb_waiters_dec(hb);
1459 }
1460
1461 /*
1462  * The hash bucket lock must be held when this is called.
1463  * Afterwards, the futex_q must not be accessed. Callers
1464  * must ensure to later call wake_up_q() for the actual
1465  * wakeups to occur.
1466  */
1467 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1468 {
1469         struct task_struct *p = q->task;
1470
1471         if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1472                 return;
1473
1474         get_task_struct(p);
1475         __unqueue_futex(q);
1476         /*
1477          * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1478          * is written, without taking any locks. This is possible in the event
1479          * of a spurious wakeup, for example. A memory barrier is required here
1480          * to prevent the following store to lock_ptr from getting ahead of the
1481          * plist_del in __unqueue_futex().
1482          */
1483         smp_store_release(&q->lock_ptr, NULL);
1484
1485         /*
1486          * Queue the task for later wakeup for after we've released
1487          * the hb->lock.
1488          */
1489         wake_q_add_safe(wake_q, p);
1490 }
1491
1492 /*
1493  * Caller must hold a reference on @pi_state.
1494  */
1495 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1496 {
1497         u32 curval, newval;
1498         struct rt_mutex_waiter *top_waiter;
1499         struct task_struct *new_owner;
1500         bool postunlock = false;
1501         DEFINE_WAKE_Q(wake_q);
1502         int ret = 0;
1503
1504         top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1505         if (WARN_ON_ONCE(!top_waiter)) {
1506                 /*
1507                  * As per the comment in futex_unlock_pi() this should not happen.
1508                  *
1509                  * When this happens, give up our locks and try again, giving
1510                  * the futex_lock_pi() instance time to complete, either by
1511                  * waiting on the rtmutex or removing itself from the futex
1512                  * queue.
1513                  */
1514                 ret = -EAGAIN;
1515                 goto out_unlock;
1516         }
1517
1518         new_owner = top_waiter->task;
1519
1520         /*
1521          * We pass it to the next owner. The WAITERS bit is always kept
1522          * enabled while there is PI state around. We cleanup the owner
1523          * died bit, because we are the owner.
1524          */
1525         newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1526
1527         if (unlikely(should_fail_futex(true))) {
1528                 ret = -EFAULT;
1529                 goto out_unlock;
1530         }
1531
1532         ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1533         if (!ret && (curval != uval)) {
1534                 /*
1535                  * If a unconditional UNLOCK_PI operation (user space did not
1536                  * try the TID->0 transition) raced with a waiter setting the
1537                  * FUTEX_WAITERS flag between get_user() and locking the hash
1538                  * bucket lock, retry the operation.
1539                  */
1540                 if ((FUTEX_TID_MASK & curval) == uval)
1541                         ret = -EAGAIN;
1542                 else
1543                         ret = -EINVAL;
1544         }
1545
1546         if (!ret) {
1547                 /*
1548                  * This is a point of no return; once we modified the uval
1549                  * there is no going back and subsequent operations must
1550                  * not fail.
1551                  */
1552                 pi_state_update_owner(pi_state, new_owner);
1553                 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1554         }
1555
1556 out_unlock:
1557         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1558
1559         if (postunlock)
1560                 rt_mutex_postunlock(&wake_q);
1561
1562         return ret;
1563 }
1564
1565 /*
1566  * Express the locking dependencies for lockdep:
1567  */
1568 static inline void
1569 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1570 {
1571         if (hb1 <= hb2) {
1572                 spin_lock(&hb1->lock);
1573                 if (hb1 < hb2)
1574                         spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1575         } else { /* hb1 > hb2 */
1576                 spin_lock(&hb2->lock);
1577                 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1578         }
1579 }
1580
1581 static inline void
1582 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1583 {
1584         spin_unlock(&hb1->lock);
1585         if (hb1 != hb2)
1586                 spin_unlock(&hb2->lock);
1587 }
1588
1589 /*
1590  * Wake up waiters matching bitset queued on this futex (uaddr).
1591  */
1592 static int
1593 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1594 {
1595         struct futex_hash_bucket *hb;
1596         struct futex_q *this, *next;
1597         union futex_key key = FUTEX_KEY_INIT;
1598         int ret;
1599         DEFINE_WAKE_Q(wake_q);
1600
1601         if (!bitset)
1602                 return -EINVAL;
1603
1604         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1605         if (unlikely(ret != 0))
1606                 return ret;
1607
1608         hb = hash_futex(&key);
1609
1610         /* Make sure we really have tasks to wakeup */
1611         if (!hb_waiters_pending(hb))
1612                 return ret;
1613
1614         spin_lock(&hb->lock);
1615
1616         plist_for_each_entry_safe(this, next, &hb->chain, list) {
1617                 if (match_futex (&this->key, &key)) {
1618                         if (this->pi_state || this->rt_waiter) {
1619                                 ret = -EINVAL;
1620                                 break;
1621                         }
1622
1623                         /* Check if one of the bits is set in both bitsets */
1624                         if (!(this->bitset & bitset))
1625                                 continue;
1626
1627                         mark_wake_futex(&wake_q, this);
1628                         if (++ret >= nr_wake)
1629                                 break;
1630                 }
1631         }
1632
1633         spin_unlock(&hb->lock);
1634         wake_up_q(&wake_q);
1635         return ret;
1636 }
1637
1638 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1639 {
1640         unsigned int op =         (encoded_op & 0x70000000) >> 28;
1641         unsigned int cmp =        (encoded_op & 0x0f000000) >> 24;
1642         int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1643         int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1644         int oldval, ret;
1645
1646         if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1647                 if (oparg < 0 || oparg > 31) {
1648                         char comm[sizeof(current->comm)];
1649                         /*
1650                          * kill this print and return -EINVAL when userspace
1651                          * is sane again
1652                          */
1653                         pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1654                                         get_task_comm(comm, current), oparg);
1655                         oparg &= 31;
1656                 }
1657                 oparg = 1 << oparg;
1658         }
1659
1660         pagefault_disable();
1661         ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1662         pagefault_enable();
1663         if (ret)
1664                 return ret;
1665
1666         switch (cmp) {
1667         case FUTEX_OP_CMP_EQ:
1668                 return oldval == cmparg;
1669         case FUTEX_OP_CMP_NE:
1670                 return oldval != cmparg;
1671         case FUTEX_OP_CMP_LT:
1672                 return oldval < cmparg;
1673         case FUTEX_OP_CMP_GE:
1674                 return oldval >= cmparg;
1675         case FUTEX_OP_CMP_LE:
1676                 return oldval <= cmparg;
1677         case FUTEX_OP_CMP_GT:
1678                 return oldval > cmparg;
1679         default:
1680                 return -ENOSYS;
1681         }
1682 }
1683
1684 /*
1685  * Wake up all waiters hashed on the physical page that is mapped
1686  * to this virtual address:
1687  */
1688 static int
1689 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1690               int nr_wake, int nr_wake2, int op)
1691 {
1692         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1693         struct futex_hash_bucket *hb1, *hb2;
1694         struct futex_q *this, *next;
1695         int ret, op_ret;
1696         DEFINE_WAKE_Q(wake_q);
1697
1698 retry:
1699         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1700         if (unlikely(ret != 0))
1701                 return ret;
1702         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1703         if (unlikely(ret != 0))
1704                 return ret;
1705
1706         hb1 = hash_futex(&key1);
1707         hb2 = hash_futex(&key2);
1708
1709 retry_private:
1710         double_lock_hb(hb1, hb2);
1711         op_ret = futex_atomic_op_inuser(op, uaddr2);
1712         if (unlikely(op_ret < 0)) {
1713                 double_unlock_hb(hb1, hb2);
1714
1715                 if (!IS_ENABLED(CONFIG_MMU) ||
1716                     unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1717                         /*
1718                          * we don't get EFAULT from MMU faults if we don't have
1719                          * an MMU, but we might get them from range checking
1720                          */
1721                         ret = op_ret;
1722                         return ret;
1723                 }
1724
1725                 if (op_ret == -EFAULT) {
1726                         ret = fault_in_user_writeable(uaddr2);
1727                         if (ret)
1728                                 return ret;
1729                 }
1730
1731                 if (!(flags & FLAGS_SHARED)) {
1732                         cond_resched();
1733                         goto retry_private;
1734                 }
1735
1736                 cond_resched();
1737                 goto retry;
1738         }
1739
1740         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1741                 if (match_futex (&this->key, &key1)) {
1742                         if (this->pi_state || this->rt_waiter) {
1743                                 ret = -EINVAL;
1744                                 goto out_unlock;
1745                         }
1746                         mark_wake_futex(&wake_q, this);
1747                         if (++ret >= nr_wake)
1748                                 break;
1749                 }
1750         }
1751
1752         if (op_ret > 0) {
1753                 op_ret = 0;
1754                 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1755                         if (match_futex (&this->key, &key2)) {
1756                                 if (this->pi_state || this->rt_waiter) {
1757                                         ret = -EINVAL;
1758                                         goto out_unlock;
1759                                 }
1760                                 mark_wake_futex(&wake_q, this);
1761                                 if (++op_ret >= nr_wake2)
1762                                         break;
1763                         }
1764                 }
1765                 ret += op_ret;
1766         }
1767
1768 out_unlock:
1769         double_unlock_hb(hb1, hb2);
1770         wake_up_q(&wake_q);
1771         return ret;
1772 }
1773
1774 /**
1775  * requeue_futex() - Requeue a futex_q from one hb to another
1776  * @q:          the futex_q to requeue
1777  * @hb1:        the source hash_bucket
1778  * @hb2:        the target hash_bucket
1779  * @key2:       the new key for the requeued futex_q
1780  */
1781 static inline
1782 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1783                    struct futex_hash_bucket *hb2, union futex_key *key2)
1784 {
1785
1786         /*
1787          * If key1 and key2 hash to the same bucket, no need to
1788          * requeue.
1789          */
1790         if (likely(&hb1->chain != &hb2->chain)) {
1791                 plist_del(&q->list, &hb1->chain);
1792                 hb_waiters_dec(hb1);
1793                 hb_waiters_inc(hb2);
1794                 plist_add(&q->list, &hb2->chain);
1795                 q->lock_ptr = &hb2->lock;
1796         }
1797         q->key = *key2;
1798 }
1799
1800 /**
1801  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1802  * @q:          the futex_q
1803  * @key:        the key of the requeue target futex
1804  * @hb:         the hash_bucket of the requeue target futex
1805  *
1806  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1807  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1808  * to the requeue target futex so the waiter can detect the wakeup on the right
1809  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1810  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1811  * to protect access to the pi_state to fixup the owner later.  Must be called
1812  * with both q->lock_ptr and hb->lock held.
1813  */
1814 static inline
1815 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1816                            struct futex_hash_bucket *hb)
1817 {
1818         q->key = *key;
1819
1820         __unqueue_futex(q);
1821
1822         WARN_ON(!q->rt_waiter);
1823         q->rt_waiter = NULL;
1824
1825         q->lock_ptr = &hb->lock;
1826
1827         wake_up_state(q->task, TASK_NORMAL);
1828 }
1829
1830 /**
1831  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1832  * @pifutex:            the user address of the to futex
1833  * @hb1:                the from futex hash bucket, must be locked by the caller
1834  * @hb2:                the to futex hash bucket, must be locked by the caller
1835  * @key1:               the from futex key
1836  * @key2:               the to futex key
1837  * @ps:                 address to store the pi_state pointer
1838  * @exiting:            Pointer to store the task pointer of the owner task
1839  *                      which is in the middle of exiting
1840  * @set_waiters:        force setting the FUTEX_WAITERS bit (1) or not (0)
1841  *
1842  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1843  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1844  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1845  * hb1 and hb2 must be held by the caller.
1846  *
1847  * @exiting is only set when the return value is -EBUSY. If so, this holds
1848  * a refcount on the exiting task on return and the caller needs to drop it
1849  * after waiting for the exit to complete.
1850  *
1851  * Return:
1852  *  -  0 - failed to acquire the lock atomically;
1853  *  - >0 - acquired the lock, return value is vpid of the top_waiter
1854  *  - <0 - error
1855  */
1856 static int
1857 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1858                            struct futex_hash_bucket *hb2, union futex_key *key1,
1859                            union futex_key *key2, struct futex_pi_state **ps,
1860                            struct task_struct **exiting, int set_waiters)
1861 {
1862         struct futex_q *top_waiter = NULL;
1863         u32 curval;
1864         int ret, vpid;
1865
1866         if (get_futex_value_locked(&curval, pifutex))
1867                 return -EFAULT;
1868
1869         if (unlikely(should_fail_futex(true)))
1870                 return -EFAULT;
1871
1872         /*
1873          * Find the top_waiter and determine if there are additional waiters.
1874          * If the caller intends to requeue more than 1 waiter to pifutex,
1875          * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1876          * as we have means to handle the possible fault.  If not, don't set
1877          * the bit unecessarily as it will force the subsequent unlock to enter
1878          * the kernel.
1879          */
1880         top_waiter = futex_top_waiter(hb1, key1);
1881
1882         /* There are no waiters, nothing for us to do. */
1883         if (!top_waiter)
1884                 return 0;
1885
1886         /* Ensure we requeue to the expected futex. */
1887         if (!match_futex(top_waiter->requeue_pi_key, key2))
1888                 return -EINVAL;
1889
1890         /*
1891          * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1892          * the contended case or if set_waiters is 1.  The pi_state is returned
1893          * in ps in contended cases.
1894          */
1895         vpid = task_pid_vnr(top_waiter->task);
1896         ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1897                                    exiting, set_waiters);
1898         if (ret == 1) {
1899                 requeue_pi_wake_futex(top_waiter, key2, hb2);
1900                 return vpid;
1901         }
1902         return ret;
1903 }
1904
1905 /**
1906  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1907  * @uaddr1:     source futex user address
1908  * @flags:      futex flags (FLAGS_SHARED, etc.)
1909  * @uaddr2:     target futex user address
1910  * @nr_wake:    number of waiters to wake (must be 1 for requeue_pi)
1911  * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1912  * @cmpval:     @uaddr1 expected value (or %NULL)
1913  * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1914  *              pi futex (pi to pi requeue is not supported)
1915  *
1916  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1917  * uaddr2 atomically on behalf of the top waiter.
1918  *
1919  * Return:
1920  *  - >=0 - on success, the number of tasks requeued or woken;
1921  *  -  <0 - on error
1922  */
1923 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1924                          u32 __user *uaddr2, int nr_wake, int nr_requeue,
1925                          u32 *cmpval, int requeue_pi)
1926 {
1927         union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1928         int task_count = 0, ret;
1929         struct futex_pi_state *pi_state = NULL;
1930         struct futex_hash_bucket *hb1, *hb2;
1931         struct futex_q *this, *next;
1932         DEFINE_WAKE_Q(wake_q);
1933
1934         if (nr_wake < 0 || nr_requeue < 0)
1935                 return -EINVAL;
1936
1937         /*
1938          * When PI not supported: return -ENOSYS if requeue_pi is true,
1939          * consequently the compiler knows requeue_pi is always false past
1940          * this point which will optimize away all the conditional code
1941          * further down.
1942          */
1943         if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1944                 return -ENOSYS;
1945
1946         if (requeue_pi) {
1947                 /*
1948                  * Requeue PI only works on two distinct uaddrs. This
1949                  * check is only valid for private futexes. See below.
1950                  */
1951                 if (uaddr1 == uaddr2)
1952                         return -EINVAL;
1953
1954                 /*
1955                  * requeue_pi requires a pi_state, try to allocate it now
1956                  * without any locks in case it fails.
1957                  */
1958                 if (refill_pi_state_cache())
1959                         return -ENOMEM;
1960                 /*
1961                  * requeue_pi must wake as many tasks as it can, up to nr_wake
1962                  * + nr_requeue, since it acquires the rt_mutex prior to
1963                  * returning to userspace, so as to not leave the rt_mutex with
1964                  * waiters and no owner.  However, second and third wake-ups
1965                  * cannot be predicted as they involve race conditions with the
1966                  * first wake and a fault while looking up the pi_state.  Both
1967                  * pthread_cond_signal() and pthread_cond_broadcast() should
1968                  * use nr_wake=1.
1969                  */
1970                 if (nr_wake != 1)
1971                         return -EINVAL;
1972         }
1973
1974 retry:
1975         ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1976         if (unlikely(ret != 0))
1977                 return ret;
1978         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1979                             requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1980         if (unlikely(ret != 0))
1981                 return ret;
1982
1983         /*
1984          * The check above which compares uaddrs is not sufficient for
1985          * shared futexes. We need to compare the keys:
1986          */
1987         if (requeue_pi && match_futex(&key1, &key2))
1988                 return -EINVAL;
1989
1990         hb1 = hash_futex(&key1);
1991         hb2 = hash_futex(&key2);
1992
1993 retry_private:
1994         hb_waiters_inc(hb2);
1995         double_lock_hb(hb1, hb2);
1996
1997         if (likely(cmpval != NULL)) {
1998                 u32 curval;
1999
2000                 ret = get_futex_value_locked(&curval, uaddr1);
2001
2002                 if (unlikely(ret)) {
2003                         double_unlock_hb(hb1, hb2);
2004                         hb_waiters_dec(hb2);
2005
2006                         ret = get_user(curval, uaddr1);
2007                         if (ret)
2008                                 return ret;
2009
2010                         if (!(flags & FLAGS_SHARED))
2011                                 goto retry_private;
2012
2013                         goto retry;
2014                 }
2015                 if (curval != *cmpval) {
2016                         ret = -EAGAIN;
2017                         goto out_unlock;
2018                 }
2019         }
2020
2021         if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2022                 struct task_struct *exiting = NULL;
2023
2024                 /*
2025                  * Attempt to acquire uaddr2 and wake the top waiter. If we
2026                  * intend to requeue waiters, force setting the FUTEX_WAITERS
2027                  * bit.  We force this here where we are able to easily handle
2028                  * faults rather in the requeue loop below.
2029                  */
2030                 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2031                                                  &key2, &pi_state,
2032                                                  &exiting, nr_requeue);
2033
2034                 /*
2035                  * At this point the top_waiter has either taken uaddr2 or is
2036                  * waiting on it.  If the former, then the pi_state will not
2037                  * exist yet, look it up one more time to ensure we have a
2038                  * reference to it. If the lock was taken, ret contains the
2039                  * vpid of the top waiter task.
2040                  * If the lock was not taken, we have pi_state and an initial
2041                  * refcount on it. In case of an error we have nothing.
2042                  */
2043                 if (ret > 0) {
2044                         WARN_ON(pi_state);
2045                         task_count++;
2046                         /*
2047                          * If we acquired the lock, then the user space value
2048                          * of uaddr2 should be vpid. It cannot be changed by
2049                          * the top waiter as it is blocked on hb2 lock if it
2050                          * tries to do so. If something fiddled with it behind
2051                          * our back the pi state lookup might unearth it. So
2052                          * we rather use the known value than rereading and
2053                          * handing potential crap to lookup_pi_state.
2054                          *
2055                          * If that call succeeds then we have pi_state and an
2056                          * initial refcount on it.
2057                          */
2058                         ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2059                                               &pi_state, &exiting);
2060                 }
2061
2062                 switch (ret) {
2063                 case 0:
2064                         /* We hold a reference on the pi state. */
2065                         break;
2066
2067                         /* If the above failed, then pi_state is NULL */
2068                 case -EFAULT:
2069                         double_unlock_hb(hb1, hb2);
2070                         hb_waiters_dec(hb2);
2071                         ret = fault_in_user_writeable(uaddr2);
2072                         if (!ret)
2073                                 goto retry;
2074                         return ret;
2075                 case -EBUSY:
2076                 case -EAGAIN:
2077                         /*
2078                          * Two reasons for this:
2079                          * - EBUSY: Owner is exiting and we just wait for the
2080                          *   exit to complete.
2081                          * - EAGAIN: The user space value changed.
2082                          */
2083                         double_unlock_hb(hb1, hb2);
2084                         hb_waiters_dec(hb2);
2085                         /*
2086                          * Handle the case where the owner is in the middle of
2087                          * exiting. Wait for the exit to complete otherwise
2088                          * this task might loop forever, aka. live lock.
2089                          */
2090                         wait_for_owner_exiting(ret, exiting);
2091                         cond_resched();
2092                         goto retry;
2093                 default:
2094                         goto out_unlock;
2095                 }
2096         }
2097
2098         plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2099                 if (task_count - nr_wake >= nr_requeue)
2100                         break;
2101
2102                 if (!match_futex(&this->key, &key1))
2103                         continue;
2104
2105                 /*
2106                  * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2107                  * be paired with each other and no other futex ops.
2108                  *
2109                  * We should never be requeueing a futex_q with a pi_state,
2110                  * which is awaiting a futex_unlock_pi().
2111                  */
2112                 if ((requeue_pi && !this->rt_waiter) ||
2113                     (!requeue_pi && this->rt_waiter) ||
2114                     this->pi_state) {
2115                         ret = -EINVAL;
2116                         break;
2117                 }
2118
2119                 /*
2120                  * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2121                  * lock, we already woke the top_waiter.  If not, it will be
2122                  * woken by futex_unlock_pi().
2123                  */
2124                 if (++task_count <= nr_wake && !requeue_pi) {
2125                         mark_wake_futex(&wake_q, this);
2126                         continue;
2127                 }
2128
2129                 /* Ensure we requeue to the expected futex for requeue_pi. */
2130                 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2131                         ret = -EINVAL;
2132                         break;
2133                 }
2134
2135                 /*
2136                  * Requeue nr_requeue waiters and possibly one more in the case
2137                  * of requeue_pi if we couldn't acquire the lock atomically.
2138                  */
2139                 if (requeue_pi) {
2140                         /*
2141                          * Prepare the waiter to take the rt_mutex. Take a
2142                          * refcount on the pi_state and store the pointer in
2143                          * the futex_q object of the waiter.
2144                          */
2145                         get_pi_state(pi_state);
2146                         this->pi_state = pi_state;
2147                         ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2148                                                         this->rt_waiter,
2149                                                         this->task);
2150                         if (ret == 1) {
2151                                 /*
2152                                  * We got the lock. We do neither drop the
2153                                  * refcount on pi_state nor clear
2154                                  * this->pi_state because the waiter needs the
2155                                  * pi_state for cleaning up the user space
2156                                  * value. It will drop the refcount after
2157                                  * doing so.
2158                                  */
2159                                 requeue_pi_wake_futex(this, &key2, hb2);
2160                                 continue;
2161                         } else if (ret) {
2162                                 /*
2163                                  * rt_mutex_start_proxy_lock() detected a
2164                                  * potential deadlock when we tried to queue
2165                                  * that waiter. Drop the pi_state reference
2166                                  * which we took above and remove the pointer
2167                                  * to the state from the waiters futex_q
2168                                  * object.
2169                                  */
2170                                 this->pi_state = NULL;
2171                                 put_pi_state(pi_state);
2172                                 /*
2173                                  * We stop queueing more waiters and let user
2174                                  * space deal with the mess.
2175                                  */
2176                                 break;
2177                         }
2178                 }
2179                 requeue_futex(this, hb1, hb2, &key2);
2180         }
2181
2182         /*
2183          * We took an extra initial reference to the pi_state either
2184          * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2185          * need to drop it here again.
2186          */
2187         put_pi_state(pi_state);
2188
2189 out_unlock:
2190         double_unlock_hb(hb1, hb2);
2191         wake_up_q(&wake_q);
2192         hb_waiters_dec(hb2);
2193         return ret ? ret : task_count;
2194 }
2195
2196 /* The key must be already stored in q->key. */
2197 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2198         __acquires(&hb->lock)
2199 {
2200         struct futex_hash_bucket *hb;
2201
2202         hb = hash_futex(&q->key);
2203
2204         /*
2205          * Increment the counter before taking the lock so that
2206          * a potential waker won't miss a to-be-slept task that is
2207          * waiting for the spinlock. This is safe as all queue_lock()
2208          * users end up calling queue_me(). Similarly, for housekeeping,
2209          * decrement the counter at queue_unlock() when some error has
2210          * occurred and we don't end up adding the task to the list.
2211          */
2212         hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2213
2214         q->lock_ptr = &hb->lock;
2215
2216         spin_lock(&hb->lock);
2217         return hb;
2218 }
2219
2220 static inline void
2221 queue_unlock(struct futex_hash_bucket *hb)
2222         __releases(&hb->lock)
2223 {
2224         spin_unlock(&hb->lock);
2225         hb_waiters_dec(hb);
2226 }
2227
2228 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2229 {
2230         int prio;
2231
2232         /*
2233          * The priority used to register this element is
2234          * - either the real thread-priority for the real-time threads
2235          * (i.e. threads with a priority lower than MAX_RT_PRIO)
2236          * - or MAX_RT_PRIO for non-RT threads.
2237          * Thus, all RT-threads are woken first in priority order, and
2238          * the others are woken last, in FIFO order.
2239          */
2240         prio = min(current->normal_prio, MAX_RT_PRIO);
2241
2242         plist_node_init(&q->list, prio);
2243         plist_add(&q->list, &hb->chain);
2244         q->task = current;
2245 }
2246
2247 /**
2248  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2249  * @q:  The futex_q to enqueue
2250  * @hb: The destination hash bucket
2251  *
2252  * The hb->lock must be held by the caller, and is released here. A call to
2253  * queue_me() is typically paired with exactly one call to unqueue_me().  The
2254  * exceptions involve the PI related operations, which may use unqueue_me_pi()
2255  * or nothing if the unqueue is done as part of the wake process and the unqueue
2256  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2257  * an example).
2258  */
2259 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2260         __releases(&hb->lock)
2261 {
2262         __queue_me(q, hb);
2263         spin_unlock(&hb->lock);
2264 }
2265
2266 /**
2267  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2268  * @q:  The futex_q to unqueue
2269  *
2270  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2271  * be paired with exactly one earlier call to queue_me().
2272  *
2273  * Return:
2274  *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2275  *  - 0 - if the futex_q was already removed by the waking thread
2276  */
2277 static int unqueue_me(struct futex_q *q)
2278 {
2279         spinlock_t *lock_ptr;
2280         int ret = 0;
2281
2282         /* In the common case we don't take the spinlock, which is nice. */
2283 retry:
2284         /*
2285          * q->lock_ptr can change between this read and the following spin_lock.
2286          * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2287          * optimizing lock_ptr out of the logic below.
2288          */
2289         lock_ptr = READ_ONCE(q->lock_ptr);
2290         if (lock_ptr != NULL) {
2291                 spin_lock(lock_ptr);
2292                 /*
2293                  * q->lock_ptr can change between reading it and
2294                  * spin_lock(), causing us to take the wrong lock.  This
2295                  * corrects the race condition.
2296                  *
2297                  * Reasoning goes like this: if we have the wrong lock,
2298                  * q->lock_ptr must have changed (maybe several times)
2299                  * between reading it and the spin_lock().  It can
2300                  * change again after the spin_lock() but only if it was
2301                  * already changed before the spin_lock().  It cannot,
2302                  * however, change back to the original value.  Therefore
2303                  * we can detect whether we acquired the correct lock.
2304                  */
2305                 if (unlikely(lock_ptr != q->lock_ptr)) {
2306                         spin_unlock(lock_ptr);
2307                         goto retry;
2308                 }
2309                 __unqueue_futex(q);
2310
2311                 BUG_ON(q->pi_state);
2312
2313                 spin_unlock(lock_ptr);
2314                 ret = 1;
2315         }
2316
2317         return ret;
2318 }
2319
2320 /*
2321  * PI futexes can not be requeued and must remove themself from the
2322  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
2323  */
2324 static void unqueue_me_pi(struct futex_q *q)
2325 {
2326         __unqueue_futex(q);
2327
2328         BUG_ON(!q->pi_state);
2329         put_pi_state(q->pi_state);
2330         q->pi_state = NULL;
2331 }
2332
2333 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2334                                   struct task_struct *argowner)
2335 {
2336         struct futex_pi_state *pi_state = q->pi_state;
2337         struct task_struct *oldowner, *newowner;
2338         u32 uval, curval, newval, newtid;
2339         int err = 0;
2340
2341         oldowner = pi_state->owner;
2342
2343         /*
2344          * We are here because either:
2345          *
2346          *  - we stole the lock and pi_state->owner needs updating to reflect
2347          *    that (@argowner == current),
2348          *
2349          * or:
2350          *
2351          *  - someone stole our lock and we need to fix things to point to the
2352          *    new owner (@argowner == NULL).
2353          *
2354          * Either way, we have to replace the TID in the user space variable.
2355          * This must be atomic as we have to preserve the owner died bit here.
2356          *
2357          * Note: We write the user space value _before_ changing the pi_state
2358          * because we can fault here. Imagine swapped out pages or a fork
2359          * that marked all the anonymous memory readonly for cow.
2360          *
2361          * Modifying pi_state _before_ the user space value would leave the
2362          * pi_state in an inconsistent state when we fault here, because we
2363          * need to drop the locks to handle the fault. This might be observed
2364          * in the PID check in lookup_pi_state.
2365          */
2366 retry:
2367         if (!argowner) {
2368                 if (oldowner != current) {
2369                         /*
2370                          * We raced against a concurrent self; things are
2371                          * already fixed up. Nothing to do.
2372                          */
2373                         return 0;
2374                 }
2375
2376                 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2377                         /* We got the lock. pi_state is correct. Tell caller. */
2378                         return 1;
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                         return 1;
2406                 }
2407                 newowner = argowner;
2408         }
2409
2410         newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2411         /* Owner died? */
2412         if (!pi_state->owner)
2413                 newtid |= FUTEX_OWNER_DIED;
2414
2415         err = get_futex_value_locked(&uval, uaddr);
2416         if (err)
2417                 goto handle_err;
2418
2419         for (;;) {
2420                 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2421
2422                 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2423                 if (err)
2424                         goto handle_err;
2425
2426                 if (curval == uval)
2427                         break;
2428                 uval = curval;
2429         }
2430
2431         /*
2432          * We fixed up user space. Now we need to fix the pi_state
2433          * itself.
2434          */
2435         pi_state_update_owner(pi_state, newowner);
2436
2437         return argowner == current;
2438
2439         /*
2440          * In order to reschedule or handle a page fault, we need to drop the
2441          * locks here. In the case of a fault, this gives the other task
2442          * (either the highest priority waiter itself or the task which stole
2443          * the rtmutex) the chance to try the fixup of the pi_state. So once we
2444          * are back from handling the fault we need to check the pi_state after
2445          * reacquiring the locks and before trying to do another fixup. When
2446          * the fixup has been done already we simply return.
2447          *
2448          * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2449          * drop hb->lock since the caller owns the hb -> futex_q relation.
2450          * Dropping the pi_mutex->wait_lock requires the state revalidate.
2451          */
2452 handle_err:
2453         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2454         spin_unlock(q->lock_ptr);
2455
2456         switch (err) {
2457         case -EFAULT:
2458                 err = fault_in_user_writeable(uaddr);
2459                 break;
2460
2461         case -EAGAIN:
2462                 cond_resched();
2463                 err = 0;
2464                 break;
2465
2466         default:
2467                 WARN_ON_ONCE(1);
2468                 break;
2469         }
2470
2471         spin_lock(q->lock_ptr);
2472         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2473
2474         /*
2475          * Check if someone else fixed it for us:
2476          */
2477         if (pi_state->owner != oldowner)
2478                 return argowner == current;
2479
2480         /* Retry if err was -EAGAIN or the fault in succeeded */
2481         if (!err)
2482                 goto retry;
2483
2484         /*
2485          * fault_in_user_writeable() failed so user state is immutable. At
2486          * best we can make the kernel state consistent but user state will
2487          * be most likely hosed and any subsequent unlock operation will be
2488          * rejected due to PI futex rule [10].
2489          *
2490          * Ensure that the rtmutex owner is also the pi_state owner despite
2491          * the user space value claiming something different. There is no
2492          * point in unlocking the rtmutex if current is the owner as it
2493          * would need to wait until the next waiter has taken the rtmutex
2494          * to guarantee consistent state. Keep it simple. Userspace asked
2495          * for this wreckaged state.
2496          *
2497          * The rtmutex has an owner - either current or some other
2498          * task. See the EAGAIN loop above.
2499          */
2500         pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2501
2502         return err;
2503 }
2504
2505 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2506                                 struct task_struct *argowner)
2507 {
2508         struct futex_pi_state *pi_state = q->pi_state;
2509         int ret;
2510
2511         lockdep_assert_held(q->lock_ptr);
2512
2513         raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2514         ret = __fixup_pi_state_owner(uaddr, q, argowner);
2515         raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2516         return ret;
2517 }
2518
2519 static long futex_wait_restart(struct restart_block *restart);
2520
2521 /**
2522  * fixup_owner() - Post lock pi_state and corner case management
2523  * @uaddr:      user address of the futex
2524  * @q:          futex_q (contains pi_state and access to the rt_mutex)
2525  * @locked:     if the attempt to take the rt_mutex succeeded (1) or not (0)
2526  *
2527  * After attempting to lock an rt_mutex, this function is called to cleanup
2528  * the pi_state owner as well as handle race conditions that may allow us to
2529  * acquire the lock. Must be called with the hb lock held.
2530  *
2531  * Return:
2532  *  -  1 - success, lock taken;
2533  *  -  0 - success, lock not taken;
2534  *  - <0 - on error (-EFAULT)
2535  */
2536 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2537 {
2538         if (locked) {
2539                 /*
2540                  * Got the lock. We might not be the anticipated owner if we
2541                  * did a lock-steal - fix up the PI-state in that case:
2542                  *
2543                  * Speculative pi_state->owner read (we don't hold wait_lock);
2544                  * since we own the lock pi_state->owner == current is the
2545                  * stable state, anything else needs more attention.
2546                  */
2547                 if (q->pi_state->owner != current)
2548                         return fixup_pi_state_owner(uaddr, q, current);
2549                 return 1;
2550         }
2551
2552         /*
2553          * If we didn't get the lock; check if anybody stole it from us. In
2554          * that case, we need to fix up the uval to point to them instead of
2555          * us, otherwise bad things happen. [10]
2556          *
2557          * Another speculative read; pi_state->owner == current is unstable
2558          * but needs our attention.
2559          */
2560         if (q->pi_state->owner == current)
2561                 return fixup_pi_state_owner(uaddr, q, NULL);
2562
2563         /*
2564          * Paranoia check. If we did not take the lock, then we should not be
2565          * the owner of the rt_mutex. Warn and establish consistent state.
2566          */
2567         if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2568                 return fixup_pi_state_owner(uaddr, q, current);
2569
2570         return 0;
2571 }
2572
2573 /**
2574  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2575  * @hb:         the futex hash bucket, must be locked by the caller
2576  * @q:          the futex_q to queue up on
2577  * @timeout:    the prepared hrtimer_sleeper, or null for no timeout
2578  */
2579 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2580                                 struct hrtimer_sleeper *timeout)
2581 {
2582         /*
2583          * The task state is guaranteed to be set before another task can
2584          * wake it. set_current_state() is implemented using smp_store_mb() and
2585          * queue_me() calls spin_unlock() upon completion, both serializing
2586          * access to the hash list and forcing another memory barrier.
2587          */
2588         set_current_state(TASK_INTERRUPTIBLE);
2589         queue_me(q, hb);
2590
2591         /* Arm the timer */
2592         if (timeout)
2593                 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2594
2595         /*
2596          * If we have been removed from the hash list, then another task
2597          * has tried to wake us, and we can skip the call to schedule().
2598          */
2599         if (likely(!plist_node_empty(&q->list))) {
2600                 /*
2601                  * If the timer has already expired, current will already be
2602                  * flagged for rescheduling. Only call schedule if there
2603                  * is no timeout, or if it has yet to expire.
2604                  */
2605                 if (!timeout || timeout->task)
2606                         freezable_schedule();
2607         }
2608         __set_current_state(TASK_RUNNING);
2609 }
2610
2611 /**
2612  * futex_wait_setup() - Prepare to wait on a futex
2613  * @uaddr:      the futex userspace address
2614  * @val:        the expected value
2615  * @flags:      futex flags (FLAGS_SHARED, etc.)
2616  * @q:          the associated futex_q
2617  * @hb:         storage for hash_bucket pointer to be returned to caller
2618  *
2619  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2620  * compare it with the expected value.  Handle atomic faults internally.
2621  * Return with the hb lock held and a q.key reference on success, and unlocked
2622  * with no q.key reference on failure.
2623  *
2624  * Return:
2625  *  -  0 - uaddr contains val and hb has been locked;
2626  *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2627  */
2628 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2629                            struct futex_q *q, struct futex_hash_bucket **hb)
2630 {
2631         u32 uval;
2632         int ret;
2633
2634         /*
2635          * Access the page AFTER the hash-bucket is locked.
2636          * Order is important:
2637          *
2638          *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2639          *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2640          *
2641          * The basic logical guarantee of a futex is that it blocks ONLY
2642          * if cond(var) is known to be true at the time of blocking, for
2643          * any cond.  If we locked the hash-bucket after testing *uaddr, that
2644          * would open a race condition where we could block indefinitely with
2645          * cond(var) false, which would violate the guarantee.
2646          *
2647          * On the other hand, we insert q and release the hash-bucket only
2648          * after testing *uaddr.  This guarantees that futex_wait() will NOT
2649          * absorb a wakeup if *uaddr does not match the desired values
2650          * while the syscall executes.
2651          */
2652 retry:
2653         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2654         if (unlikely(ret != 0))
2655                 return ret;
2656
2657 retry_private:
2658         *hb = queue_lock(q);
2659
2660         ret = get_futex_value_locked(&uval, uaddr);
2661
2662         if (ret) {
2663                 queue_unlock(*hb);
2664
2665                 ret = get_user(uval, uaddr);
2666                 if (ret)
2667                         return ret;
2668
2669                 if (!(flags & FLAGS_SHARED))
2670                         goto retry_private;
2671
2672                 goto retry;
2673         }
2674
2675         if (uval != val) {
2676                 queue_unlock(*hb);
2677                 ret = -EWOULDBLOCK;
2678         }
2679
2680         return ret;
2681 }
2682
2683 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2684                       ktime_t *abs_time, u32 bitset)
2685 {
2686         struct hrtimer_sleeper timeout, *to;
2687         struct restart_block *restart;
2688         struct futex_hash_bucket *hb;
2689         struct futex_q q = futex_q_init;
2690         int ret;
2691
2692         if (!bitset)
2693                 return -EINVAL;
2694         q.bitset = bitset;
2695
2696         to = futex_setup_timer(abs_time, &timeout, flags,
2697                                current->timer_slack_ns);
2698 retry:
2699         /*
2700          * Prepare to wait on uaddr. On success, holds hb lock and increments
2701          * q.key refs.
2702          */
2703         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2704         if (ret)
2705                 goto out;
2706
2707         /* queue_me and wait for wakeup, timeout, or a signal. */
2708         futex_wait_queue_me(hb, &q, to);
2709
2710         /* If we were woken (and unqueued), we succeeded, whatever. */
2711         ret = 0;
2712         /* unqueue_me() drops q.key ref */
2713         if (!unqueue_me(&q))
2714                 goto out;
2715         ret = -ETIMEDOUT;
2716         if (to && !to->task)
2717                 goto out;
2718
2719         /*
2720          * We expect signal_pending(current), but we might be the
2721          * victim of a spurious wakeup as well.
2722          */
2723         if (!signal_pending(current))
2724                 goto retry;
2725
2726         ret = -ERESTARTSYS;
2727         if (!abs_time)
2728                 goto out;
2729
2730         restart = &current->restart_block;
2731         restart->futex.uaddr = uaddr;
2732         restart->futex.val = val;
2733         restart->futex.time = *abs_time;
2734         restart->futex.bitset = bitset;
2735         restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2736
2737         ret = set_restart_fn(restart, futex_wait_restart);
2738
2739 out:
2740         if (to) {
2741                 hrtimer_cancel(&to->timer);
2742                 destroy_hrtimer_on_stack(&to->timer);
2743         }
2744         return ret;
2745 }
2746
2747
2748 static long futex_wait_restart(struct restart_block *restart)
2749 {
2750         u32 __user *uaddr = restart->futex.uaddr;
2751         ktime_t t, *tp = NULL;
2752
2753         if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2754                 t = restart->futex.time;
2755                 tp = &t;
2756         }
2757         restart->fn = do_no_restart_syscall;
2758
2759         return (long)futex_wait(uaddr, restart->futex.flags,
2760                                 restart->futex.val, tp, restart->futex.bitset);
2761 }
2762
2763
2764 /*
2765  * Userspace tried a 0 -> TID atomic transition of the futex value
2766  * and failed. The kernel side here does the whole locking operation:
2767  * if there are waiters then it will block as a consequence of relying
2768  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2769  * a 0 value of the futex too.).
2770  *
2771  * Also serves as futex trylock_pi()'ing, and due semantics.
2772  */
2773 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2774                          ktime_t *time, int trylock)
2775 {
2776         struct hrtimer_sleeper timeout, *to;
2777         struct task_struct *exiting = NULL;
2778         struct rt_mutex_waiter rt_waiter;
2779         struct futex_hash_bucket *hb;
2780         struct futex_q q = futex_q_init;
2781         int res, ret;
2782
2783         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2784                 return -ENOSYS;
2785
2786         if (refill_pi_state_cache())
2787                 return -ENOMEM;
2788
2789         to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2790
2791 retry:
2792         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2793         if (unlikely(ret != 0))
2794                 goto out;
2795
2796 retry_private:
2797         hb = queue_lock(&q);
2798
2799         ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2800                                    &exiting, 0);
2801         if (unlikely(ret)) {
2802                 /*
2803                  * Atomic work succeeded and we got the lock,
2804                  * or failed. Either way, we do _not_ block.
2805                  */
2806                 switch (ret) {
2807                 case 1:
2808                         /* We got the lock. */
2809                         ret = 0;
2810                         goto out_unlock_put_key;
2811                 case -EFAULT:
2812                         goto uaddr_faulted;
2813                 case -EBUSY:
2814                 case -EAGAIN:
2815                         /*
2816                          * Two reasons for this:
2817                          * - EBUSY: Task is exiting and we just wait for the
2818                          *   exit to complete.
2819                          * - EAGAIN: The user space value changed.
2820                          */
2821                         queue_unlock(hb);
2822                         /*
2823                          * Handle the case where the owner is in the middle of
2824                          * exiting. Wait for the exit to complete otherwise
2825                          * this task might loop forever, aka. live lock.
2826                          */
2827                         wait_for_owner_exiting(ret, exiting);
2828                         cond_resched();
2829                         goto retry;
2830                 default:
2831                         goto out_unlock_put_key;
2832                 }
2833         }
2834
2835         WARN_ON(!q.pi_state);
2836
2837         /*
2838          * Only actually queue now that the atomic ops are done:
2839          */
2840         __queue_me(&q, hb);
2841
2842         if (trylock) {
2843                 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2844                 /* Fixup the trylock return value: */
2845                 ret = ret ? 0 : -EWOULDBLOCK;
2846                 goto no_block;
2847         }
2848
2849         rt_mutex_init_waiter(&rt_waiter);
2850
2851         /*
2852          * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2853          * hold it while doing rt_mutex_start_proxy(), because then it will
2854          * include hb->lock in the blocking chain, even through we'll not in
2855          * fact hold it while blocking. This will lead it to report -EDEADLK
2856          * and BUG when futex_unlock_pi() interleaves with this.
2857          *
2858          * Therefore acquire wait_lock while holding hb->lock, but drop the
2859          * latter before calling __rt_mutex_start_proxy_lock(). This
2860          * interleaves with futex_unlock_pi() -- which does a similar lock
2861          * handoff -- such that the latter can observe the futex_q::pi_state
2862          * before __rt_mutex_start_proxy_lock() is done.
2863          */
2864         raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2865         spin_unlock(q.lock_ptr);
2866         /*
2867          * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2868          * such that futex_unlock_pi() is guaranteed to observe the waiter when
2869          * it sees the futex_q::pi_state.
2870          */
2871         ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2872         raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2873
2874         if (ret) {
2875                 if (ret == 1)
2876                         ret = 0;
2877                 goto cleanup;
2878         }
2879
2880         if (unlikely(to))
2881                 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2882
2883         ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2884
2885 cleanup:
2886         spin_lock(q.lock_ptr);
2887         /*
2888          * If we failed to acquire the lock (deadlock/signal/timeout), we must
2889          * first acquire the hb->lock before removing the lock from the
2890          * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2891          * lists consistent.
2892          *
2893          * In particular; it is important that futex_unlock_pi() can not
2894          * observe this inconsistency.
2895          */
2896         if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2897                 ret = 0;
2898
2899 no_block:
2900         /*
2901          * Fixup the pi_state owner and possibly acquire the lock if we
2902          * haven't already.
2903          */
2904         res = fixup_owner(uaddr, &q, !ret);
2905         /*
2906          * If fixup_owner() returned an error, proprogate that.  If it acquired
2907          * the lock, clear our -ETIMEDOUT or -EINTR.
2908          */
2909         if (res)
2910                 ret = (res < 0) ? res : 0;
2911
2912         unqueue_me_pi(&q);
2913         spin_unlock(q.lock_ptr);
2914         goto out;
2915
2916 out_unlock_put_key:
2917         queue_unlock(hb);
2918
2919 out:
2920         if (to) {
2921                 hrtimer_cancel(&to->timer);
2922                 destroy_hrtimer_on_stack(&to->timer);
2923         }
2924         return ret != -EINTR ? ret : -ERESTARTNOINTR;
2925
2926 uaddr_faulted:
2927         queue_unlock(hb);
2928
2929         ret = fault_in_user_writeable(uaddr);
2930         if (ret)
2931                 goto out;
2932
2933         if (!(flags & FLAGS_SHARED))
2934                 goto retry_private;
2935
2936         goto retry;
2937 }
2938
2939 /*
2940  * Userspace attempted a TID -> 0 atomic transition, and failed.
2941  * This is the in-kernel slowpath: we look up the PI state (if any),
2942  * and do the rt-mutex unlock.
2943  */
2944 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2945 {
2946         u32 curval, uval, vpid = task_pid_vnr(current);
2947         union futex_key key = FUTEX_KEY_INIT;
2948         struct futex_hash_bucket *hb;
2949         struct futex_q *top_waiter;
2950         int ret;
2951
2952         if (!IS_ENABLED(CONFIG_FUTEX_PI))
2953                 return -ENOSYS;
2954
2955 retry:
2956         if (get_user(uval, uaddr))
2957                 return -EFAULT;
2958         /*
2959          * We release only a lock we actually own:
2960          */
2961         if ((uval & FUTEX_TID_MASK) != vpid)
2962                 return -EPERM;
2963
2964         ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2965         if (ret)
2966                 return ret;
2967
2968         hb = hash_futex(&key);
2969         spin_lock(&hb->lock);
2970
2971         /*
2972          * Check waiters first. We do not trust user space values at
2973          * all and we at least want to know if user space fiddled
2974          * with the futex value instead of blindly unlocking.
2975          */
2976         top_waiter = futex_top_waiter(hb, &key);
2977         if (top_waiter) {
2978                 struct futex_pi_state *pi_state = top_waiter->pi_state;
2979
2980                 ret = -EINVAL;
2981                 if (!pi_state)
2982                         goto out_unlock;
2983
2984                 /*
2985                  * If current does not own the pi_state then the futex is
2986                  * inconsistent and user space fiddled with the futex value.
2987                  */
2988                 if (pi_state->owner != current)
2989                         goto out_unlock;
2990
2991                 get_pi_state(pi_state);
2992                 /*
2993                  * By taking wait_lock while still holding hb->lock, we ensure
2994                  * there is no point where we hold neither; and therefore
2995                  * wake_futex_pi() must observe a state consistent with what we
2996                  * observed.
2997                  *
2998                  * In particular; this forces __rt_mutex_start_proxy() to
2999                  * complete such that we're guaranteed to observe the
3000                  * rt_waiter. Also see the WARN in wake_futex_pi().
3001                  */
3002                 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3003                 spin_unlock(&hb->lock);
3004
3005                 /* drops pi_state->pi_mutex.wait_lock */
3006                 ret = wake_futex_pi(uaddr, uval, pi_state);
3007
3008                 put_pi_state(pi_state);
3009
3010                 /*
3011                  * Success, we're done! No tricky corner cases.
3012                  */
3013                 if (!ret)
3014                         return ret;
3015                 /*
3016                  * The atomic access to the futex value generated a
3017                  * pagefault, so retry the user-access and the wakeup:
3018                  */
3019                 if (ret == -EFAULT)
3020                         goto pi_faulted;
3021                 /*
3022                  * A unconditional UNLOCK_PI op raced against a waiter
3023                  * setting the FUTEX_WAITERS bit. Try again.
3024                  */
3025                 if (ret == -EAGAIN)
3026                         goto pi_retry;
3027                 /*
3028                  * wake_futex_pi has detected invalid state. Tell user
3029                  * space.
3030                  */
3031                 return ret;
3032         }
3033
3034         /*
3035          * We have no kernel internal state, i.e. no waiters in the
3036          * kernel. Waiters which are about to queue themselves are stuck
3037          * on hb->lock. So we can safely ignore them. We do neither
3038          * preserve the WAITERS bit not the OWNER_DIED one. We are the
3039          * owner.
3040          */
3041         if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3042                 spin_unlock(&hb->lock);
3043                 switch (ret) {
3044                 case -EFAULT:
3045                         goto pi_faulted;
3046
3047                 case -EAGAIN:
3048                         goto pi_retry;
3049
3050                 default:
3051                         WARN_ON_ONCE(1);
3052                         return ret;
3053                 }
3054         }
3055
3056         /*
3057          * If uval has changed, let user space handle it.
3058          */
3059         ret = (curval == uval) ? 0 : -EAGAIN;
3060
3061 out_unlock:
3062         spin_unlock(&hb->lock);
3063         return ret;
3064
3065 pi_retry:
3066         cond_resched();
3067         goto retry;
3068
3069 pi_faulted:
3070
3071         ret = fault_in_user_writeable(uaddr);
3072         if (!ret)
3073                 goto retry;
3074
3075         return ret;
3076 }
3077
3078 /**
3079  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3080  * @hb:         the hash_bucket futex_q was original enqueued on
3081  * @q:          the futex_q woken while waiting to be requeued
3082  * @key2:       the futex_key of the requeue target futex
3083  * @timeout:    the timeout associated with the wait (NULL if none)
3084  *
3085  * Detect if the task was woken on the initial futex as opposed to the requeue
3086  * target futex.  If so, determine if it was a timeout or a signal that caused
3087  * the wakeup and return the appropriate error code to the caller.  Must be
3088  * called with the hb lock held.
3089  *
3090  * Return:
3091  *  -  0 = no early wakeup detected;
3092  *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3093  */
3094 static inline
3095 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3096                                    struct futex_q *q, union futex_key *key2,
3097                                    struct hrtimer_sleeper *timeout)
3098 {
3099         int ret = 0;
3100
3101         /*
3102          * With the hb lock held, we avoid races while we process the wakeup.
3103          * We only need to hold hb (and not hb2) to ensure atomicity as the
3104          * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3105          * It can't be requeued from uaddr2 to something else since we don't
3106          * support a PI aware source futex for requeue.
3107          */
3108         if (!match_futex(&q->key, key2)) {
3109                 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3110                 /*
3111                  * We were woken prior to requeue by a timeout or a signal.
3112                  * Unqueue the futex_q and determine which it was.
3113                  */
3114                 plist_del(&q->list, &hb->chain);
3115                 hb_waiters_dec(hb);
3116
3117                 /* Handle spurious wakeups gracefully */
3118                 ret = -EWOULDBLOCK;
3119                 if (timeout && !timeout->task)
3120                         ret = -ETIMEDOUT;
3121                 else if (signal_pending(current))
3122                         ret = -ERESTARTNOINTR;
3123         }
3124         return ret;
3125 }
3126
3127 /**
3128  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3129  * @uaddr:      the futex we initially wait on (non-pi)
3130  * @flags:      futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3131  *              the same type, no requeueing from private to shared, etc.
3132  * @val:        the expected value of uaddr
3133  * @abs_time:   absolute timeout
3134  * @bitset:     32 bit wakeup bitset set by userspace, defaults to all
3135  * @uaddr2:     the pi futex we will take prior to returning to user-space
3136  *
3137  * The caller will wait on uaddr and will be requeued by futex_requeue() to
3138  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3139  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3140  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3141  * without one, the pi logic would not know which task to boost/deboost, if
3142  * there was a need to.
3143  *
3144  * We call schedule in futex_wait_queue_me() when we enqueue and return there
3145  * via the following--
3146  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3147  * 2) wakeup on uaddr2 after a requeue
3148  * 3) signal
3149  * 4) timeout
3150  *
3151  * If 3, cleanup and return -ERESTARTNOINTR.
3152  *
3153  * If 2, we may then block on trying to take the rt_mutex and return via:
3154  * 5) successful lock
3155  * 6) signal
3156  * 7) timeout
3157  * 8) other lock acquisition failure
3158  *
3159  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3160  *
3161  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3162  *
3163  * Return:
3164  *  -  0 - On success;
3165  *  - <0 - On error
3166  */
3167 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3168                                  u32 val, ktime_t *abs_time, u32 bitset,
3169                                  u32 __user *uaddr2)
3170 {
3171         struct hrtimer_sleeper timeout, *to;
3172         struct rt_mutex_waiter rt_waiter;
3173         struct futex_hash_bucket *hb;
3174         union futex_key key2 = FUTEX_KEY_INIT;
3175         struct futex_q q = futex_q_init;
3176         int res, ret;
3177
3178         if (!IS_ENABLED(CONFIG_FUTEX_PI))
3179                 return -ENOSYS;
3180
3181         if (uaddr == uaddr2)
3182                 return -EINVAL;
3183
3184         if (!bitset)
3185                 return -EINVAL;
3186
3187         to = futex_setup_timer(abs_time, &timeout, flags,
3188                                current->timer_slack_ns);
3189
3190         /*
3191          * The waiter is allocated on our stack, manipulated by the requeue
3192          * code while we sleep on uaddr.
3193          */
3194         rt_mutex_init_waiter(&rt_waiter);
3195
3196         ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3197         if (unlikely(ret != 0))
3198                 goto out;
3199
3200         q.bitset = bitset;
3201         q.rt_waiter = &rt_waiter;
3202         q.requeue_pi_key = &key2;
3203
3204         /*
3205          * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3206          * count.
3207          */
3208         ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3209         if (ret)
3210                 goto out;
3211
3212         /*
3213          * The check above which compares uaddrs is not sufficient for
3214          * shared futexes. We need to compare the keys:
3215          */
3216         if (match_futex(&q.key, &key2)) {
3217                 queue_unlock(hb);
3218                 ret = -EINVAL;
3219                 goto out;
3220         }
3221
3222         /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3223         futex_wait_queue_me(hb, &q, to);
3224
3225         spin_lock(&hb->lock);
3226         ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3227         spin_unlock(&hb->lock);
3228         if (ret)
3229                 goto out;
3230
3231         /*
3232          * In order for us to be here, we know our q.key == key2, and since
3233          * we took the hb->lock above, we also know that futex_requeue() has
3234          * completed and we no longer have to concern ourselves with a wakeup
3235          * race with the atomic proxy lock acquisition by the requeue code. The
3236          * futex_requeue dropped our key1 reference and incremented our key2
3237          * reference count.
3238          */
3239
3240         /*
3241          * Check if the requeue code acquired the second futex for us and do
3242          * any pertinent fixup.
3243          */
3244         if (!q.rt_waiter) {
3245                 if (q.pi_state && (q.pi_state->owner != current)) {
3246                         spin_lock(q.lock_ptr);
3247                         ret = fixup_owner(uaddr2, &q, true);
3248                         /*
3249                          * Drop the reference to the pi state which
3250                          * the requeue_pi() code acquired for us.
3251                          */
3252                         put_pi_state(q.pi_state);
3253                         spin_unlock(q.lock_ptr);
3254                         /*
3255                          * Adjust the return value. It's either -EFAULT or
3256                          * success (1) but the caller expects 0 for success.
3257                          */
3258                         ret = ret < 0 ? ret : 0;
3259                 }
3260         } else {
3261                 struct rt_mutex *pi_mutex;
3262
3263                 /*
3264                  * We have been woken up by futex_unlock_pi(), a timeout, or a
3265                  * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3266                  * the pi_state.
3267                  */
3268                 WARN_ON(!q.pi_state);
3269                 pi_mutex = &q.pi_state->pi_mutex;
3270                 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3271
3272                 spin_lock(q.lock_ptr);
3273                 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3274                         ret = 0;
3275
3276                 debug_rt_mutex_free_waiter(&rt_waiter);
3277                 /*
3278                  * Fixup the pi_state owner and possibly acquire the lock if we
3279                  * haven't already.
3280                  */
3281                 res = fixup_owner(uaddr2, &q, !ret);
3282                 /*
3283                  * If fixup_owner() returned an error, proprogate that.  If it
3284                  * acquired the lock, clear -ETIMEDOUT or -EINTR.
3285                  */
3286                 if (res)
3287                         ret = (res < 0) ? res : 0;
3288
3289                 unqueue_me_pi(&q);
3290                 spin_unlock(q.lock_ptr);
3291         }
3292
3293         if (ret == -EINTR) {
3294                 /*
3295                  * We've already been requeued, but cannot restart by calling
3296                  * futex_lock_pi() directly. We could restart this syscall, but
3297                  * it would detect that the user space "val" changed and return
3298                  * -EWOULDBLOCK.  Save the overhead of the restart and return
3299                  * -EWOULDBLOCK directly.
3300                  */
3301                 ret = -EWOULDBLOCK;
3302         }
3303
3304 out:
3305         if (to) {
3306                 hrtimer_cancel(&to->timer);
3307                 destroy_hrtimer_on_stack(&to->timer);
3308         }
3309         return ret;
3310 }
3311
3312 /*
3313  * Support for robust futexes: the kernel cleans up held futexes at
3314  * thread exit time.
3315  *
3316  * Implementation: user-space maintains a per-thread list of locks it
3317  * is holding. Upon do_exit(), the kernel carefully walks this list,
3318  * and marks all locks that are owned by this thread with the
3319  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3320  * always manipulated with the lock held, so the list is private and
3321  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3322  * field, to allow the kernel to clean up if the thread dies after
3323  * acquiring the lock, but just before it could have added itself to
3324  * the list. There can only be one such pending lock.
3325  */
3326
3327 /**
3328  * sys_set_robust_list() - Set the robust-futex list head of a task
3329  * @head:       pointer to the list-head
3330  * @len:        length of the list-head, as userspace expects
3331  */
3332 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3333                 size_t, len)
3334 {
3335         if (!futex_cmpxchg_enabled)
3336                 return -ENOSYS;
3337         /*
3338          * The kernel knows only one size for now:
3339          */
3340         if (unlikely(len != sizeof(*head)))
3341                 return -EINVAL;
3342
3343         current->robust_list = head;
3344
3345         return 0;
3346 }
3347
3348 /**
3349  * sys_get_robust_list() - Get the robust-futex list head of a task
3350  * @pid:        pid of the process [zero for current task]
3351  * @head_ptr:   pointer to a list-head pointer, the kernel fills it in
3352  * @len_ptr:    pointer to a length field, the kernel fills in the header size
3353  */
3354 SYSCALL_DEFINE3(get_robust_list, int, pid,
3355                 struct robust_list_head __user * __user *, head_ptr,
3356                 size_t __user *, len_ptr)
3357 {
3358         struct robust_list_head __user *head;
3359         unsigned long ret;
3360         struct task_struct *p;
3361
3362         if (!futex_cmpxchg_enabled)
3363                 return -ENOSYS;
3364
3365         rcu_read_lock();
3366
3367         ret = -ESRCH;
3368         if (!pid)
3369                 p = current;
3370         else {
3371                 p = find_task_by_vpid(pid);
3372                 if (!p)
3373                         goto err_unlock;
3374         }
3375
3376         ret = -EPERM;
3377         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3378                 goto err_unlock;
3379
3380         head = p->robust_list;
3381         rcu_read_unlock();
3382
3383         if (put_user(sizeof(*head), len_ptr))
3384                 return -EFAULT;
3385         return put_user(head, head_ptr);
3386
3387 err_unlock:
3388         rcu_read_unlock();
3389
3390         return ret;
3391 }
3392
3393 /* Constants for the pending_op argument of handle_futex_death */
3394 #define HANDLE_DEATH_PENDING    true
3395 #define HANDLE_DEATH_LIST       false
3396
3397 /*
3398  * Process a futex-list entry, check whether it's owned by the
3399  * dying task, and do notification if so:
3400  */
3401 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3402                               bool pi, bool pending_op)
3403 {
3404         u32 uval, nval, mval;
3405         int err;
3406
3407         /* Futex address must be 32bit aligned */
3408         if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3409                 return -1;
3410
3411 retry:
3412         if (get_user(uval, uaddr))
3413                 return -1;
3414
3415         /*
3416          * Special case for regular (non PI) futexes. The unlock path in
3417          * user space has two race scenarios:
3418          *
3419          * 1. The unlock path releases the user space futex value and
3420          *    before it can execute the futex() syscall to wake up
3421          *    waiters it is killed.
3422          *
3423          * 2. A woken up waiter is killed before it can acquire the
3424          *    futex in user space.
3425          *
3426          * In both cases the TID validation below prevents a wakeup of
3427          * potential waiters which can cause these waiters to block
3428          * forever.
3429          *
3430          * In both cases the following conditions are met:
3431          *
3432          *      1) task->robust_list->list_op_pending != NULL
3433          *         @pending_op == true
3434          *      2) User space futex value == 0
3435          *      3) Regular futex: @pi == false
3436          *
3437          * If these conditions are met, it is safe to attempt waking up a
3438          * potential waiter without touching the user space futex value and
3439          * trying to set the OWNER_DIED bit. The user space futex value is
3440          * uncontended and the rest of the user space mutex state is
3441          * consistent, so a woken waiter will just take over the
3442          * uncontended futex. Setting the OWNER_DIED bit would create
3443          * inconsistent state and malfunction of the user space owner died
3444          * handling.
3445          */
3446         if (pending_op && !pi && !uval) {
3447                 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3448                 return 0;
3449         }
3450
3451         if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3452                 return 0;
3453
3454         /*
3455          * Ok, this dying thread is truly holding a futex
3456          * of interest. Set the OWNER_DIED bit atomically
3457          * via cmpxchg, and if the value had FUTEX_WAITERS
3458          * set, wake up a waiter (if any). (We have to do a
3459          * futex_wake() even if OWNER_DIED is already set -
3460          * to handle the rare but possible case of recursive
3461          * thread-death.) The rest of the cleanup is done in
3462          * userspace.
3463          */
3464         mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3465
3466         /*
3467          * We are not holding a lock here, but we want to have
3468          * the pagefault_disable/enable() protection because
3469          * we want to handle the fault gracefully. If the
3470          * access fails we try to fault in the futex with R/W
3471          * verification via get_user_pages. get_user() above
3472          * does not guarantee R/W access. If that fails we
3473          * give up and leave the futex locked.
3474          */
3475         if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3476                 switch (err) {
3477                 case -EFAULT:
3478                         if (fault_in_user_writeable(uaddr))
3479                                 return -1;
3480                         goto retry;
3481
3482                 case -EAGAIN:
3483                         cond_resched();
3484                         goto retry;
3485
3486                 default:
3487                         WARN_ON_ONCE(1);
3488                         return err;
3489                 }
3490         }
3491
3492         if (nval != uval)
3493                 goto retry;
3494
3495         /*
3496          * Wake robust non-PI futexes here. The wakeup of
3497          * PI futexes happens in exit_pi_state():
3498          */
3499         if (!pi && (uval & FUTEX_WAITERS))
3500                 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3501
3502         return 0;
3503 }
3504
3505 /*
3506  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3507  */
3508 static inline int fetch_robust_entry(struct robust_list __user **entry,
3509                                      struct robust_list __user * __user *head,
3510                                      unsigned int *pi)
3511 {
3512         unsigned long uentry;
3513
3514         if (get_user(uentry, (unsigned long __user *)head))
3515                 return -EFAULT;
3516
3517         *entry = (void __user *)(uentry & ~1UL);
3518         *pi = uentry & 1;
3519
3520         return 0;
3521 }
3522
3523 /*
3524  * Walk curr->robust_list (very carefully, it's a userspace list!)
3525  * and mark any locks found there dead, and notify any waiters.
3526  *
3527  * We silently return on any sign of list-walking problem.
3528  */
3529 static void exit_robust_list(struct task_struct *curr)
3530 {
3531         struct robust_list_head __user *head = curr->robust_list;
3532         struct robust_list __user *entry, *next_entry, *pending;
3533         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3534         unsigned int next_pi;
3535         unsigned long futex_offset;
3536         int rc;
3537
3538         if (!futex_cmpxchg_enabled)
3539                 return;
3540
3541         /*
3542          * Fetch the list head (which was registered earlier, via
3543          * sys_set_robust_list()):
3544          */
3545         if (fetch_robust_entry(&entry, &head->list.next, &pi))
3546                 return;
3547         /*
3548          * Fetch the relative futex offset:
3549          */
3550         if (get_user(futex_offset, &head->futex_offset))
3551                 return;
3552         /*
3553          * Fetch any possibly pending lock-add first, and handle it
3554          * if it exists:
3555          */
3556         if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3557                 return;
3558
3559         next_entry = NULL;      /* avoid warning with gcc */
3560         while (entry != &head->list) {
3561                 /*
3562                  * Fetch the next entry in the list before calling
3563                  * handle_futex_death:
3564                  */
3565                 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3566                 /*
3567                  * A pending lock might already be on the list, so
3568                  * don't process it twice:
3569                  */
3570                 if (entry != pending) {
3571                         if (handle_futex_death((void __user *)entry + futex_offset,
3572                                                 curr, pi, HANDLE_DEATH_LIST))
3573                                 return;
3574                 }
3575                 if (rc)
3576                         return;
3577                 entry = next_entry;
3578                 pi = next_pi;
3579                 /*
3580                  * Avoid excessively long or circular lists:
3581                  */
3582                 if (!--limit)
3583                         break;
3584
3585                 cond_resched();
3586         }
3587
3588         if (pending) {
3589                 handle_futex_death((void __user *)pending + futex_offset,
3590                                    curr, pip, HANDLE_DEATH_PENDING);
3591         }
3592 }
3593
3594 static void futex_cleanup(struct task_struct *tsk)
3595 {
3596         if (unlikely(tsk->robust_list)) {
3597                 exit_robust_list(tsk);
3598                 tsk->robust_list = NULL;
3599         }
3600
3601 #ifdef CONFIG_COMPAT
3602         if (unlikely(tsk->compat_robust_list)) {
3603                 compat_exit_robust_list(tsk);
3604                 tsk->compat_robust_list = NULL;
3605         }
3606 #endif
3607
3608         if (unlikely(!list_empty(&tsk->pi_state_list)))
3609                 exit_pi_state_list(tsk);
3610 }
3611
3612 /**
3613  * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3614  * @tsk:        task to set the state on
3615  *
3616  * Set the futex exit state of the task lockless. The futex waiter code
3617  * observes that state when a task is exiting and loops until the task has
3618  * actually finished the futex cleanup. The worst case for this is that the
3619  * waiter runs through the wait loop until the state becomes visible.
3620  *
3621  * This is called from the recursive fault handling path in do_exit().
3622  *
3623  * This is best effort. Either the futex exit code has run already or
3624  * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3625  * take it over. If not, the problem is pushed back to user space. If the
3626  * futex exit code did not run yet, then an already queued waiter might
3627  * block forever, but there is nothing which can be done about that.
3628  */
3629 void futex_exit_recursive(struct task_struct *tsk)
3630 {
3631         /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3632         if (tsk->futex_state == FUTEX_STATE_EXITING)
3633                 mutex_unlock(&tsk->futex_exit_mutex);
3634         tsk->futex_state = FUTEX_STATE_DEAD;
3635 }
3636
3637 static void futex_cleanup_begin(struct task_struct *tsk)
3638 {
3639         /*
3640          * Prevent various race issues against a concurrent incoming waiter
3641          * including live locks by forcing the waiter to block on
3642          * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3643          * attach_to_pi_owner().
3644          */
3645         mutex_lock(&tsk->futex_exit_mutex);
3646
3647         /*
3648          * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3649          *
3650          * This ensures that all subsequent checks of tsk->futex_state in
3651          * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3652          * tsk->pi_lock held.
3653          *
3654          * It guarantees also that a pi_state which was queued right before
3655          * the state change under tsk->pi_lock by a concurrent waiter must
3656          * be observed in exit_pi_state_list().
3657          */
3658         raw_spin_lock_irq(&tsk->pi_lock);
3659         tsk->futex_state = FUTEX_STATE_EXITING;
3660         raw_spin_unlock_irq(&tsk->pi_lock);
3661 }
3662
3663 static void futex_cleanup_end(struct task_struct *tsk, int state)
3664 {
3665         /*
3666          * Lockless store. The only side effect is that an observer might
3667          * take another loop until it becomes visible.
3668          */
3669         tsk->futex_state = state;
3670         /*
3671          * Drop the exit protection. This unblocks waiters which observed
3672          * FUTEX_STATE_EXITING to reevaluate the state.
3673          */
3674         mutex_unlock(&tsk->futex_exit_mutex);
3675 }
3676
3677 void futex_exec_release(struct task_struct *tsk)
3678 {
3679         /*
3680          * The state handling is done for consistency, but in the case of
3681          * exec() there is no way to prevent futher damage as the PID stays
3682          * the same. But for the unlikely and arguably buggy case that a
3683          * futex is held on exec(), this provides at least as much state
3684          * consistency protection which is possible.
3685          */
3686         futex_cleanup_begin(tsk);
3687         futex_cleanup(tsk);
3688         /*
3689          * Reset the state to FUTEX_STATE_OK. The task is alive and about
3690          * exec a new binary.
3691          */
3692         futex_cleanup_end(tsk, FUTEX_STATE_OK);
3693 }
3694
3695 void futex_exit_release(struct task_struct *tsk)
3696 {
3697         futex_cleanup_begin(tsk);
3698         futex_cleanup(tsk);
3699         futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3700 }
3701
3702 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3703                 u32 __user *uaddr2, u32 val2, u32 val3)
3704 {
3705         int cmd = op & FUTEX_CMD_MASK;
3706         unsigned int flags = 0;
3707
3708         if (!(op & FUTEX_PRIVATE_FLAG))
3709                 flags |= FLAGS_SHARED;
3710
3711         if (op & FUTEX_CLOCK_REALTIME) {
3712                 flags |= FLAGS_CLOCKRT;
3713                 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3714                     cmd != FUTEX_WAIT_REQUEUE_PI)
3715                         return -ENOSYS;
3716         }
3717
3718         switch (cmd) {
3719         case FUTEX_LOCK_PI:
3720         case FUTEX_UNLOCK_PI:
3721         case FUTEX_TRYLOCK_PI:
3722         case FUTEX_WAIT_REQUEUE_PI:
3723         case FUTEX_CMP_REQUEUE_PI:
3724                 if (!futex_cmpxchg_enabled)
3725                         return -ENOSYS;
3726         }
3727
3728         switch (cmd) {
3729         case FUTEX_WAIT:
3730                 val3 = FUTEX_BITSET_MATCH_ANY;
3731                 fallthrough;
3732         case FUTEX_WAIT_BITSET:
3733                 return futex_wait(uaddr, flags, val, timeout, val3);
3734         case FUTEX_WAKE:
3735                 val3 = FUTEX_BITSET_MATCH_ANY;
3736                 fallthrough;
3737         case FUTEX_WAKE_BITSET:
3738                 return futex_wake(uaddr, flags, val, val3);
3739         case FUTEX_REQUEUE:
3740                 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3741         case FUTEX_CMP_REQUEUE:
3742                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3743         case FUTEX_WAKE_OP:
3744                 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3745         case FUTEX_LOCK_PI:
3746                 return futex_lock_pi(uaddr, flags, timeout, 0);
3747         case FUTEX_UNLOCK_PI:
3748                 return futex_unlock_pi(uaddr, flags);
3749         case FUTEX_TRYLOCK_PI:
3750                 return futex_lock_pi(uaddr, flags, NULL, 1);
3751         case FUTEX_WAIT_REQUEUE_PI:
3752                 val3 = FUTEX_BITSET_MATCH_ANY;
3753                 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3754                                              uaddr2);
3755         case FUTEX_CMP_REQUEUE_PI:
3756                 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3757         }
3758         return -ENOSYS;
3759 }
3760
3761
3762 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3763                 const struct __kernel_timespec __user *, utime,
3764                 u32 __user *, uaddr2, u32, val3)
3765 {
3766         struct timespec64 ts;
3767         ktime_t t, *tp = NULL;
3768         u32 val2 = 0;
3769         int cmd = op & FUTEX_CMD_MASK;
3770
3771         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3772                       cmd == FUTEX_WAIT_BITSET ||
3773                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3774                 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3775                         return -EFAULT;
3776                 if (get_timespec64(&ts, utime))
3777                         return -EFAULT;
3778                 if (!timespec64_valid(&ts))
3779                         return -EINVAL;
3780
3781                 t = timespec64_to_ktime(ts);
3782                 if (cmd == FUTEX_WAIT)
3783                         t = ktime_add_safe(ktime_get(), t);
3784                 else if (!(op & FUTEX_CLOCK_REALTIME))
3785                         t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3786                 tp = &t;
3787         }
3788         /*
3789          * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3790          * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3791          */
3792         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3793             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3794                 val2 = (u32) (unsigned long) utime;
3795
3796         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3797 }
3798
3799 #ifdef CONFIG_COMPAT
3800 /*
3801  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3802  */
3803 static inline int
3804 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3805                    compat_uptr_t __user *head, unsigned int *pi)
3806 {
3807         if (get_user(*uentry, head))
3808                 return -EFAULT;
3809
3810         *entry = compat_ptr((*uentry) & ~1);
3811         *pi = (unsigned int)(*uentry) & 1;
3812
3813         return 0;
3814 }
3815
3816 static void __user *futex_uaddr(struct robust_list __user *entry,
3817                                 compat_long_t futex_offset)
3818 {
3819         compat_uptr_t base = ptr_to_compat(entry);
3820         void __user *uaddr = compat_ptr(base + futex_offset);
3821
3822         return uaddr;
3823 }
3824
3825 /*
3826  * Walk curr->robust_list (very carefully, it's a userspace list!)
3827  * and mark any locks found there dead, and notify any waiters.
3828  *
3829  * We silently return on any sign of list-walking problem.
3830  */
3831 static void compat_exit_robust_list(struct task_struct *curr)
3832 {
3833         struct compat_robust_list_head __user *head = curr->compat_robust_list;
3834         struct robust_list __user *entry, *next_entry, *pending;
3835         unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3836         unsigned int next_pi;
3837         compat_uptr_t uentry, next_uentry, upending;
3838         compat_long_t futex_offset;
3839         int rc;
3840
3841         if (!futex_cmpxchg_enabled)
3842                 return;
3843
3844         /*
3845          * Fetch the list head (which was registered earlier, via
3846          * sys_set_robust_list()):
3847          */
3848         if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3849                 return;
3850         /*
3851          * Fetch the relative futex offset:
3852          */
3853         if (get_user(futex_offset, &head->futex_offset))
3854                 return;
3855         /*
3856          * Fetch any possibly pending lock-add first, and handle it
3857          * if it exists:
3858          */
3859         if (compat_fetch_robust_entry(&upending, &pending,
3860                                &head->list_op_pending, &pip))
3861                 return;
3862
3863         next_entry = NULL;      /* avoid warning with gcc */
3864         while (entry != (struct robust_list __user *) &head->list) {
3865                 /*
3866                  * Fetch the next entry in the list before calling
3867                  * handle_futex_death:
3868                  */
3869                 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3870                         (compat_uptr_t __user *)&entry->next, &next_pi);
3871                 /*
3872                  * A pending lock might already be on the list, so
3873                  * dont process it twice:
3874                  */
3875                 if (entry != pending) {
3876                         void __user *uaddr = futex_uaddr(entry, futex_offset);
3877
3878                         if (handle_futex_death(uaddr, curr, pi,
3879                                                HANDLE_DEATH_LIST))
3880                                 return;
3881                 }
3882                 if (rc)
3883                         return;
3884                 uentry = next_uentry;
3885                 entry = next_entry;
3886                 pi = next_pi;
3887                 /*
3888                  * Avoid excessively long or circular lists:
3889                  */
3890                 if (!--limit)
3891                         break;
3892
3893                 cond_resched();
3894         }
3895         if (pending) {
3896                 void __user *uaddr = futex_uaddr(pending, futex_offset);
3897
3898                 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3899         }
3900 }
3901
3902 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3903                 struct compat_robust_list_head __user *, head,
3904                 compat_size_t, len)
3905 {
3906         if (!futex_cmpxchg_enabled)
3907                 return -ENOSYS;
3908
3909         if (unlikely(len != sizeof(*head)))
3910                 return -EINVAL;
3911
3912         current->compat_robust_list = head;
3913
3914         return 0;
3915 }
3916
3917 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3918                         compat_uptr_t __user *, head_ptr,
3919                         compat_size_t __user *, len_ptr)
3920 {
3921         struct compat_robust_list_head __user *head;
3922         unsigned long ret;
3923         struct task_struct *p;
3924
3925         if (!futex_cmpxchg_enabled)
3926                 return -ENOSYS;
3927
3928         rcu_read_lock();
3929
3930         ret = -ESRCH;
3931         if (!pid)
3932                 p = current;
3933         else {
3934                 p = find_task_by_vpid(pid);
3935                 if (!p)
3936                         goto err_unlock;
3937         }
3938
3939         ret = -EPERM;
3940         if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3941                 goto err_unlock;
3942
3943         head = p->compat_robust_list;
3944         rcu_read_unlock();
3945
3946         if (put_user(sizeof(*head), len_ptr))
3947                 return -EFAULT;
3948         return put_user(ptr_to_compat(head), head_ptr);
3949
3950 err_unlock:
3951         rcu_read_unlock();
3952
3953         return ret;
3954 }
3955 #endif /* CONFIG_COMPAT */
3956
3957 #ifdef CONFIG_COMPAT_32BIT_TIME
3958 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3959                 const struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3960                 u32, val3)
3961 {
3962         struct timespec64 ts;
3963         ktime_t t, *tp = NULL;
3964         int val2 = 0;
3965         int cmd = op & FUTEX_CMD_MASK;
3966
3967         if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3968                       cmd == FUTEX_WAIT_BITSET ||
3969                       cmd == FUTEX_WAIT_REQUEUE_PI)) {
3970                 if (get_old_timespec32(&ts, utime))
3971                         return -EFAULT;
3972                 if (!timespec64_valid(&ts))
3973                         return -EINVAL;
3974
3975                 t = timespec64_to_ktime(ts);
3976                 if (cmd == FUTEX_WAIT)
3977                         t = ktime_add_safe(ktime_get(), t);
3978                 else if (!(op & FUTEX_CLOCK_REALTIME))
3979                         t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3980                 tp = &t;
3981         }
3982         if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3983             cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3984                 val2 = (int) (unsigned long) utime;
3985
3986         return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3987 }
3988 #endif /* CONFIG_COMPAT_32BIT_TIME */
3989
3990 static void __init futex_detect_cmpxchg(void)
3991 {
3992 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3993         u32 curval;
3994
3995         /*
3996          * This will fail and we want it. Some arch implementations do
3997          * runtime detection of the futex_atomic_cmpxchg_inatomic()
3998          * functionality. We want to know that before we call in any
3999          * of the complex code paths. Also we want to prevent
4000          * registration of robust lists in that case. NULL is
4001          * guaranteed to fault and we get -EFAULT on functional
4002          * implementation, the non-functional ones will return
4003          * -ENOSYS.
4004          */
4005         if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4006                 futex_cmpxchg_enabled = 1;
4007 #endif
4008 }
4009
4010 static int __init futex_init(void)
4011 {
4012         unsigned int futex_shift;
4013         unsigned long i;
4014
4015 #if CONFIG_BASE_SMALL
4016         futex_hashsize = 16;
4017 #else
4018         futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4019 #endif
4020
4021         futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4022                                                futex_hashsize, 0,
4023                                                futex_hashsize < 256 ? HASH_SMALL : 0,
4024                                                &futex_shift, NULL,
4025                                                futex_hashsize, futex_hashsize);
4026         futex_hashsize = 1UL << futex_shift;
4027
4028         futex_detect_cmpxchg();
4029
4030         for (i = 0; i < futex_hashsize; i++) {
4031                 atomic_set(&futex_queues[i].waiters, 0);
4032                 plist_head_init(&futex_queues[i].chain);
4033                 spin_lock_init(&futex_queues[i].lock);
4034         }
4035
4036         return 0;
4037 }
4038 core_initcall(futex_init);