Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
5 *
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 *
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
11 *
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 *
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 *
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 *
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 *
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
30 *
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
33 */
34#include <linux/compat.h>
35#include <linux/slab.h>
36#include <linux/poll.h>
37#include <linux/fs.h>
38#include <linux/file.h>
39#include <linux/jhash.h>
40#include <linux/init.h>
41#include <linux/futex.h>
42#include <linux/mount.h>
43#include <linux/pagemap.h>
44#include <linux/syscalls.h>
45#include <linux/signal.h>
46#include <linux/export.h>
47#include <linux/magic.h>
48#include <linux/pid.h>
49#include <linux/nsproxy.h>
50#include <linux/ptrace.h>
51#include <linux/sched/rt.h>
52#include <linux/sched/wake_q.h>
53#include <linux/sched/mm.h>
54#include <linux/hugetlb.h>
55#include <linux/freezer.h>
56#include <linux/memblock.h>
57#include <linux/fault-inject.h>
58#include <linux/refcount.h>
59
60#include <asm/futex.h>
61
62#include "locking/rtmutex_common.h"
63
64/*
65 * READ this before attempting to hack on futexes!
66 *
67 * Basic futex operation and ordering guarantees
68 * =============================================
69 *
70 * The waiter reads the futex value in user space and calls
71 * futex_wait(). This function computes the hash bucket and acquires
72 * the hash bucket lock. After that it reads the futex user space value
73 * again and verifies that the data has not changed. If it has not changed
74 * it enqueues itself into the hash bucket, releases the hash bucket lock
75 * and schedules.
76 *
77 * The waker side modifies the user space value of the futex and calls
78 * futex_wake(). This function computes the hash bucket and acquires the
79 * hash bucket lock. Then it looks for waiters on that futex in the hash
80 * bucket and wakes them.
81 *
82 * In futex wake up scenarios where no tasks are blocked on a futex, taking
83 * the hb spinlock can be avoided and simply return. In order for this
84 * optimization to work, ordering guarantees must exist so that the waiter
85 * being added to the list is acknowledged when the list is concurrently being
86 * checked by the waker, avoiding scenarios like the following:
87 *
88 * CPU 0 CPU 1
89 * val = *futex;
90 * sys_futex(WAIT, futex, val);
91 * futex_wait(futex, val);
92 * uval = *futex;
93 * *futex = newval;
94 * sys_futex(WAKE, futex);
95 * futex_wake(futex);
96 * if (queue_empty())
97 * return;
98 * if (uval == val)
99 * lock(hash_bucket(futex));
100 * queue();
101 * unlock(hash_bucket(futex));
102 * schedule();
103 *
104 * This would cause the waiter on CPU 0 to wait forever because it
105 * missed the transition of the user space value from val to newval
106 * and the waker did not find the waiter in the hash bucket queue.
107 *
108 * The correct serialization ensures that a waiter either observes
109 * the changed user space value before blocking or is woken by a
110 * concurrent waker:
111 *
112 * CPU 0 CPU 1
113 * val = *futex;
114 * sys_futex(WAIT, futex, val);
115 * futex_wait(futex, val);
116 *
117 * waiters++; (a)
118 * smp_mb(); (A) <-- paired with -.
119 * |
120 * lock(hash_bucket(futex)); |
121 * |
122 * uval = *futex; |
123 * | *futex = newval;
124 * | sys_futex(WAKE, futex);
125 * | futex_wake(futex);
126 * |
127 * `--------> smp_mb(); (B)
128 * if (uval == val)
129 * queue();
130 * unlock(hash_bucket(futex));
131 * schedule(); if (waiters)
132 * lock(hash_bucket(futex));
133 * else wake_waiters(futex);
134 * waiters--; (b) unlock(hash_bucket(futex));
135 *
136 * Where (A) orders the waiters increment and the futex value read through
137 * atomic operations (see hb_waiters_inc) and where (B) orders the write
138 * to futex and the waiters read -- this is done by the barriers for both
139 * shared and private futexes in get_futex_key_refs().
140 *
141 * This yields the following case (where X:=waiters, Y:=futex):
142 *
143 * X = Y = 0
144 *
145 * w[X]=1 w[Y]=1
146 * MB MB
147 * r[Y]=y r[X]=x
148 *
149 * Which guarantees that x==0 && y==0 is impossible; which translates back into
150 * the guarantee that we cannot both miss the futex variable change and the
151 * enqueue.
152 *
153 * Note that a new waiter is accounted for in (a) even when it is possible that
154 * the wait call can return error, in which case we backtrack from it in (b).
155 * Refer to the comment in queue_lock().
156 *
157 * Similarly, in order to account for waiters being requeued on another
158 * address we always increment the waiters for the destination bucket before
159 * acquiring the lock. It then decrements them again after releasing it -
160 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
161 * will do the additional required waiter count housekeeping. This is done for
162 * double_lock_hb() and double_unlock_hb(), respectively.
163 */
164
165#ifdef CONFIG_HAVE_FUTEX_CMPXCHG
166#define futex_cmpxchg_enabled 1
167#else
168static int __read_mostly futex_cmpxchg_enabled;
169#endif
170
171/*
172 * Futex flags used to encode options to functions and preserve them across
173 * restarts.
174 */
175#ifdef CONFIG_MMU
176# define FLAGS_SHARED 0x01
177#else
178/*
179 * NOMMU does not have per process address space. Let the compiler optimize
180 * code away.
181 */
182# define FLAGS_SHARED 0x00
183#endif
184#define FLAGS_CLOCKRT 0x02
185#define FLAGS_HAS_TIMEOUT 0x04
186
187/*
188 * Priority Inheritance state:
189 */
190struct futex_pi_state {
191 /*
192 * list of 'owned' pi_state instances - these have to be
193 * cleaned up in do_exit() if the task exits prematurely:
194 */
195 struct list_head list;
196
197 /*
198 * The PI object:
199 */
200 struct rt_mutex pi_mutex;
201
202 struct task_struct *owner;
203 refcount_t refcount;
204
205 union futex_key key;
206} __randomize_layout;
207
208/**
209 * struct futex_q - The hashed futex queue entry, one per waiting task
210 * @list: priority-sorted list of tasks waiting on this futex
211 * @task: the task waiting on the futex
212 * @lock_ptr: the hash bucket lock
213 * @key: the key the futex is hashed on
214 * @pi_state: optional priority inheritance state
215 * @rt_waiter: rt_waiter storage for use with requeue_pi
216 * @requeue_pi_key: the requeue_pi target futex key
217 * @bitset: bitset for the optional bitmasked wakeup
218 *
219 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
220 * we can wake only the relevant ones (hashed queues may be shared).
221 *
222 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
223 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
224 * The order of wakeup is always to make the first condition true, then
225 * the second.
226 *
227 * PI futexes are typically woken before they are removed from the hash list via
228 * the rt_mutex code. See unqueue_me_pi().
229 */
230struct futex_q {
231 struct plist_node list;
232
233 struct task_struct *task;
234 spinlock_t *lock_ptr;
235 union futex_key key;
236 struct futex_pi_state *pi_state;
237 struct rt_mutex_waiter *rt_waiter;
238 union futex_key *requeue_pi_key;
239 u32 bitset;
240} __randomize_layout;
241
242static const struct futex_q futex_q_init = {
243 /* list gets initialized in queue_me()*/
244 .key = FUTEX_KEY_INIT,
245 .bitset = FUTEX_BITSET_MATCH_ANY
246};
247
248/*
249 * Hash buckets are shared by all the futex_keys that hash to the same
250 * location. Each key may have multiple futex_q structures, one for each task
251 * waiting on a futex.
252 */
253struct futex_hash_bucket {
254 atomic_t waiters;
255 spinlock_t lock;
256 struct plist_head chain;
257} ____cacheline_aligned_in_smp;
258
259/*
260 * The base of the bucket array and its size are always used together
261 * (after initialization only in hash_futex()), so ensure that they
262 * reside in the same cacheline.
263 */
264static struct {
265 struct futex_hash_bucket *queues;
266 unsigned long hashsize;
267} __futex_data __read_mostly __aligned(2*sizeof(long));
268#define futex_queues (__futex_data.queues)
269#define futex_hashsize (__futex_data.hashsize)
270
271
272/*
273 * Fault injections for futexes.
274 */
275#ifdef CONFIG_FAIL_FUTEX
276
277static struct {
278 struct fault_attr attr;
279
280 bool ignore_private;
281} fail_futex = {
282 .attr = FAULT_ATTR_INITIALIZER,
283 .ignore_private = false,
284};
285
286static int __init setup_fail_futex(char *str)
287{
288 return setup_fault_attr(&fail_futex.attr, str);
289}
290__setup("fail_futex=", setup_fail_futex);
291
292static bool should_fail_futex(bool fshared)
293{
294 if (fail_futex.ignore_private && !fshared)
295 return false;
296
297 return should_fail(&fail_futex.attr, 1);
298}
299
300#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
301
302static int __init fail_futex_debugfs(void)
303{
304 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
305 struct dentry *dir;
306
307 dir = fault_create_debugfs_attr("fail_futex", NULL,
308 &fail_futex.attr);
309 if (IS_ERR(dir))
310 return PTR_ERR(dir);
311
312 debugfs_create_bool("ignore-private", mode, dir,
313 &fail_futex.ignore_private);
314 return 0;
315}
316
317late_initcall(fail_futex_debugfs);
318
319#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
320
321#else
322static inline bool should_fail_futex(bool fshared)
323{
324 return false;
325}
326#endif /* CONFIG_FAIL_FUTEX */
327
328static inline void futex_get_mm(union futex_key *key)
329{
330 mmgrab(key->private.mm);
331 /*
332 * Ensure futex_get_mm() implies a full barrier such that
333 * get_futex_key() implies a full barrier. This is relied upon
334 * as smp_mb(); (B), see the ordering comment above.
335 */
336 smp_mb__after_atomic();
337}
338
339/*
340 * Reflects a new waiter being added to the waitqueue.
341 */
342static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
343{
344#ifdef CONFIG_SMP
345 atomic_inc(&hb->waiters);
346 /*
347 * Full barrier (A), see the ordering comment above.
348 */
349 smp_mb__after_atomic();
350#endif
351}
352
353/*
354 * Reflects a waiter being removed from the waitqueue by wakeup
355 * paths.
356 */
357static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
358{
359#ifdef CONFIG_SMP
360 atomic_dec(&hb->waiters);
361#endif
362}
363
364static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
365{
366#ifdef CONFIG_SMP
367 return atomic_read(&hb->waiters);
368#else
369 return 1;
370#endif
371}
372
373/**
374 * hash_futex - Return the hash bucket in the global hash
375 * @key: Pointer to the futex key for which the hash is calculated
376 *
377 * We hash on the keys returned from get_futex_key (see below) and return the
378 * corresponding hash bucket in the global hash.
379 */
380static struct futex_hash_bucket *hash_futex(union futex_key *key)
381{
382 u32 hash = jhash2((u32*)&key->both.word,
383 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
384 key->both.offset);
385 return &futex_queues[hash & (futex_hashsize - 1)];
386}
387
388
389/**
390 * match_futex - Check whether two futex keys are equal
391 * @key1: Pointer to key1
392 * @key2: Pointer to key2
393 *
394 * Return 1 if two futex_keys are equal, 0 otherwise.
395 */
396static inline int match_futex(union futex_key *key1, union futex_key *key2)
397{
398 return (key1 && key2
399 && key1->both.word == key2->both.word
400 && key1->both.ptr == key2->both.ptr
401 && key1->both.offset == key2->both.offset);
402}
403
404/*
405 * Take a reference to the resource addressed by a key.
406 * Can be called while holding spinlocks.
407 *
408 */
409static void get_futex_key_refs(union futex_key *key)
410{
411 if (!key->both.ptr)
412 return;
413
414 /*
415 * On MMU less systems futexes are always "private" as there is no per
416 * process address space. We need the smp wmb nevertheless - yes,
417 * arch/blackfin has MMU less SMP ...
418 */
419 if (!IS_ENABLED(CONFIG_MMU)) {
420 smp_mb(); /* explicit smp_mb(); (B) */
421 return;
422 }
423
424 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
425 case FUT_OFF_INODE:
426 ihold(key->shared.inode); /* implies smp_mb(); (B) */
427 break;
428 case FUT_OFF_MMSHARED:
429 futex_get_mm(key); /* implies smp_mb(); (B) */
430 break;
431 default:
432 /*
433 * Private futexes do not hold reference on an inode or
434 * mm, therefore the only purpose of calling get_futex_key_refs
435 * is because we need the barrier for the lockless waiter check.
436 */
437 smp_mb(); /* explicit smp_mb(); (B) */
438 }
439}
440
441/*
442 * Drop a reference to the resource addressed by a key.
443 * The hash bucket spinlock must not be held. This is
444 * a no-op for private futexes, see comment in the get
445 * counterpart.
446 */
447static void drop_futex_key_refs(union futex_key *key)
448{
449 if (!key->both.ptr) {
450 /* If we're here then we tried to put a key we failed to get */
451 WARN_ON_ONCE(1);
452 return;
453 }
454
455 if (!IS_ENABLED(CONFIG_MMU))
456 return;
457
458 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
459 case FUT_OFF_INODE:
460 iput(key->shared.inode);
461 break;
462 case FUT_OFF_MMSHARED:
463 mmdrop(key->private.mm);
464 break;
465 }
466}
467
468enum futex_access {
469 FUTEX_READ,
470 FUTEX_WRITE
471};
472
473/**
474 * futex_setup_timer - set up the sleeping hrtimer.
475 * @time: ptr to the given timeout value
476 * @timeout: the hrtimer_sleeper structure to be set up
477 * @flags: futex flags
478 * @range_ns: optional range in ns
479 *
480 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
481 * value given
482 */
483static inline struct hrtimer_sleeper *
484futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
485 int flags, u64 range_ns)
486{
487 if (!time)
488 return NULL;
489
490 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
491 CLOCK_REALTIME : CLOCK_MONOTONIC,
492 HRTIMER_MODE_ABS);
493 /*
494 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
495 * effectively the same as calling hrtimer_set_expires().
496 */
497 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
498
499 return timeout;
500}
501
502/**
503 * get_futex_key() - Get parameters which are the keys for a futex
504 * @uaddr: virtual address of the futex
505 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
506 * @key: address where result is stored.
507 * @rw: mapping needs to be read/write (values: FUTEX_READ,
508 * FUTEX_WRITE)
509 *
510 * Return: a negative error code or 0
511 *
512 * The key words are stored in @key on success.
513 *
514 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
515 * offset_within_page). For private mappings, it's (uaddr, current->mm).
516 * We can usually work out the index without swapping in the page.
517 *
518 * lock_page() might sleep, the caller should not hold a spinlock.
519 */
520static int
521get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw)
522{
523 unsigned long address = (unsigned long)uaddr;
524 struct mm_struct *mm = current->mm;
525 struct page *page, *tail;
526 struct address_space *mapping;
527 int err, ro = 0;
528
529 /*
530 * The futex address must be "naturally" aligned.
531 */
532 key->both.offset = address % PAGE_SIZE;
533 if (unlikely((address % sizeof(u32)) != 0))
534 return -EINVAL;
535 address -= key->both.offset;
536
537 if (unlikely(!access_ok(uaddr, sizeof(u32))))
538 return -EFAULT;
539
540 if (unlikely(should_fail_futex(fshared)))
541 return -EFAULT;
542
543 /*
544 * PROCESS_PRIVATE futexes are fast.
545 * As the mm cannot disappear under us and the 'key' only needs
546 * virtual address, we dont even have to find the underlying vma.
547 * Note : We do have to check 'uaddr' is a valid user address,
548 * but access_ok() should be faster than find_vma()
549 */
550 if (!fshared) {
551 key->private.mm = mm;
552 key->private.address = address;
553 get_futex_key_refs(key); /* implies smp_mb(); (B) */
554 return 0;
555 }
556
557again:
558 /* Ignore any VERIFY_READ mapping (futex common case) */
559 if (unlikely(should_fail_futex(fshared)))
560 return -EFAULT;
561
562 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
563 /*
564 * If write access is not required (eg. FUTEX_WAIT), try
565 * and get read-only access.
566 */
567 if (err == -EFAULT && rw == FUTEX_READ) {
568 err = get_user_pages_fast(address, 1, 0, &page);
569 ro = 1;
570 }
571 if (err < 0)
572 return err;
573 else
574 err = 0;
575
576 /*
577 * The treatment of mapping from this point on is critical. The page
578 * lock protects many things but in this context the page lock
579 * stabilizes mapping, prevents inode freeing in the shared
580 * file-backed region case and guards against movement to swap cache.
581 *
582 * Strictly speaking the page lock is not needed in all cases being
583 * considered here and page lock forces unnecessarily serialization
584 * From this point on, mapping will be re-verified if necessary and
585 * page lock will be acquired only if it is unavoidable
586 *
587 * Mapping checks require the head page for any compound page so the
588 * head page and mapping is looked up now. For anonymous pages, it
589 * does not matter if the page splits in the future as the key is
590 * based on the address. For filesystem-backed pages, the tail is
591 * required as the index of the page determines the key. For
592 * base pages, there is no tail page and tail == page.
593 */
594 tail = page;
595 page = compound_head(page);
596 mapping = READ_ONCE(page->mapping);
597
598 /*
599 * If page->mapping is NULL, then it cannot be a PageAnon
600 * page; but it might be the ZERO_PAGE or in the gate area or
601 * in a special mapping (all cases which we are happy to fail);
602 * or it may have been a good file page when get_user_pages_fast
603 * found it, but truncated or holepunched or subjected to
604 * invalidate_complete_page2 before we got the page lock (also
605 * cases which we are happy to fail). And we hold a reference,
606 * so refcount care in invalidate_complete_page's remove_mapping
607 * prevents drop_caches from setting mapping to NULL beneath us.
608 *
609 * The case we do have to guard against is when memory pressure made
610 * shmem_writepage move it from filecache to swapcache beneath us:
611 * an unlikely race, but we do need to retry for page->mapping.
612 */
613 if (unlikely(!mapping)) {
614 int shmem_swizzled;
615
616 /*
617 * Page lock is required to identify which special case above
618 * applies. If this is really a shmem page then the page lock
619 * will prevent unexpected transitions.
620 */
621 lock_page(page);
622 shmem_swizzled = PageSwapCache(page) || page->mapping;
623 unlock_page(page);
624 put_page(page);
625
626 if (shmem_swizzled)
627 goto again;
628
629 return -EFAULT;
630 }
631
632 /*
633 * Private mappings are handled in a simple way.
634 *
635 * If the futex key is stored on an anonymous page, then the associated
636 * object is the mm which is implicitly pinned by the calling process.
637 *
638 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
639 * it's a read-only handle, it's expected that futexes attach to
640 * the object not the particular process.
641 */
642 if (PageAnon(page)) {
643 /*
644 * A RO anonymous page will never change and thus doesn't make
645 * sense for futex operations.
646 */
647 if (unlikely(should_fail_futex(fshared)) || ro) {
648 err = -EFAULT;
649 goto out;
650 }
651
652 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
653 key->private.mm = mm;
654 key->private.address = address;
655
656 get_futex_key_refs(key); /* implies smp_mb(); (B) */
657
658 } else {
659 struct inode *inode;
660
661 /*
662 * The associated futex object in this case is the inode and
663 * the page->mapping must be traversed. Ordinarily this should
664 * be stabilised under page lock but it's not strictly
665 * necessary in this case as we just want to pin the inode, not
666 * update the radix tree or anything like that.
667 *
668 * The RCU read lock is taken as the inode is finally freed
669 * under RCU. If the mapping still matches expectations then the
670 * mapping->host can be safely accessed as being a valid inode.
671 */
672 rcu_read_lock();
673
674 if (READ_ONCE(page->mapping) != mapping) {
675 rcu_read_unlock();
676 put_page(page);
677
678 goto again;
679 }
680
681 inode = READ_ONCE(mapping->host);
682 if (!inode) {
683 rcu_read_unlock();
684 put_page(page);
685
686 goto again;
687 }
688
689 /*
690 * Take a reference unless it is about to be freed. Previously
691 * this reference was taken by ihold under the page lock
692 * pinning the inode in place so i_lock was unnecessary. The
693 * only way for this check to fail is if the inode was
694 * truncated in parallel which is almost certainly an
695 * application bug. In such a case, just retry.
696 *
697 * We are not calling into get_futex_key_refs() in file-backed
698 * cases, therefore a successful atomic_inc return below will
699 * guarantee that get_futex_key() will still imply smp_mb(); (B).
700 */
701 if (!atomic_inc_not_zero(&inode->i_count)) {
702 rcu_read_unlock();
703 put_page(page);
704
705 goto again;
706 }
707
708 /* Should be impossible but lets be paranoid for now */
709 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
710 err = -EFAULT;
711 rcu_read_unlock();
712 iput(inode);
713
714 goto out;
715 }
716
717 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
718 key->shared.inode = inode;
719 key->shared.pgoff = basepage_index(tail);
720 rcu_read_unlock();
721 }
722
723out:
724 put_page(page);
725 return err;
726}
727
728static inline void put_futex_key(union futex_key *key)
729{
730 drop_futex_key_refs(key);
731}
732
733/**
734 * fault_in_user_writeable() - Fault in user address and verify RW access
735 * @uaddr: pointer to faulting user space address
736 *
737 * Slow path to fixup the fault we just took in the atomic write
738 * access to @uaddr.
739 *
740 * We have no generic implementation of a non-destructive write to the
741 * user address. We know that we faulted in the atomic pagefault
742 * disabled section so we can as well avoid the #PF overhead by
743 * calling get_user_pages() right away.
744 */
745static int fault_in_user_writeable(u32 __user *uaddr)
746{
747 struct mm_struct *mm = current->mm;
748 int ret;
749
750 down_read(&mm->mmap_sem);
751 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
752 FAULT_FLAG_WRITE, NULL);
753 up_read(&mm->mmap_sem);
754
755 return ret < 0 ? ret : 0;
756}
757
758/**
759 * futex_top_waiter() - Return the highest priority waiter on a futex
760 * @hb: the hash bucket the futex_q's reside in
761 * @key: the futex key (to distinguish it from other futex futex_q's)
762 *
763 * Must be called with the hb lock held.
764 */
765static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
766 union futex_key *key)
767{
768 struct futex_q *this;
769
770 plist_for_each_entry(this, &hb->chain, list) {
771 if (match_futex(&this->key, key))
772 return this;
773 }
774 return NULL;
775}
776
777static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
778 u32 uval, u32 newval)
779{
780 int ret;
781
782 pagefault_disable();
783 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
784 pagefault_enable();
785
786 return ret;
787}
788
789static int get_futex_value_locked(u32 *dest, u32 __user *from)
790{
791 int ret;
792
793 pagefault_disable();
794 ret = __get_user(*dest, from);
795 pagefault_enable();
796
797 return ret ? -EFAULT : 0;
798}
799
800
801/*
802 * PI code:
803 */
804static int refill_pi_state_cache(void)
805{
806 struct futex_pi_state *pi_state;
807
808 if (likely(current->pi_state_cache))
809 return 0;
810
811 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
812
813 if (!pi_state)
814 return -ENOMEM;
815
816 INIT_LIST_HEAD(&pi_state->list);
817 /* pi_mutex gets initialized later */
818 pi_state->owner = NULL;
819 refcount_set(&pi_state->refcount, 1);
820 pi_state->key = FUTEX_KEY_INIT;
821
822 current->pi_state_cache = pi_state;
823
824 return 0;
825}
826
827static struct futex_pi_state *alloc_pi_state(void)
828{
829 struct futex_pi_state *pi_state = current->pi_state_cache;
830
831 WARN_ON(!pi_state);
832 current->pi_state_cache = NULL;
833
834 return pi_state;
835}
836
837static void get_pi_state(struct futex_pi_state *pi_state)
838{
839 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
840}
841
842/*
843 * Drops a reference to the pi_state object and frees or caches it
844 * when the last reference is gone.
845 */
846static void put_pi_state(struct futex_pi_state *pi_state)
847{
848 if (!pi_state)
849 return;
850
851 if (!refcount_dec_and_test(&pi_state->refcount))
852 return;
853
854 /*
855 * If pi_state->owner is NULL, the owner is most probably dying
856 * and has cleaned up the pi_state already
857 */
858 if (pi_state->owner) {
859 struct task_struct *owner;
860
861 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
862 owner = pi_state->owner;
863 if (owner) {
864 raw_spin_lock(&owner->pi_lock);
865 list_del_init(&pi_state->list);
866 raw_spin_unlock(&owner->pi_lock);
867 }
868 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
869 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
870 }
871
872 if (current->pi_state_cache) {
873 kfree(pi_state);
874 } else {
875 /*
876 * pi_state->list is already empty.
877 * clear pi_state->owner.
878 * refcount is at 0 - put it back to 1.
879 */
880 pi_state->owner = NULL;
881 refcount_set(&pi_state->refcount, 1);
882 current->pi_state_cache = pi_state;
883 }
884}
885
886#ifdef CONFIG_FUTEX_PI
887
888/*
889 * This task is holding PI mutexes at exit time => bad.
890 * Kernel cleans up PI-state, but userspace is likely hosed.
891 * (Robust-futex cleanup is separate and might save the day for userspace.)
892 */
893void exit_pi_state_list(struct task_struct *curr)
894{
895 struct list_head *next, *head = &curr->pi_state_list;
896 struct futex_pi_state *pi_state;
897 struct futex_hash_bucket *hb;
898 union futex_key key = FUTEX_KEY_INIT;
899
900 if (!futex_cmpxchg_enabled)
901 return;
902 /*
903 * We are a ZOMBIE and nobody can enqueue itself on
904 * pi_state_list anymore, but we have to be careful
905 * versus waiters unqueueing themselves:
906 */
907 raw_spin_lock_irq(&curr->pi_lock);
908 while (!list_empty(head)) {
909 next = head->next;
910 pi_state = list_entry(next, struct futex_pi_state, list);
911 key = pi_state->key;
912 hb = hash_futex(&key);
913
914 /*
915 * We can race against put_pi_state() removing itself from the
916 * list (a waiter going away). put_pi_state() will first
917 * decrement the reference count and then modify the list, so
918 * its possible to see the list entry but fail this reference
919 * acquire.
920 *
921 * In that case; drop the locks to let put_pi_state() make
922 * progress and retry the loop.
923 */
924 if (!refcount_inc_not_zero(&pi_state->refcount)) {
925 raw_spin_unlock_irq(&curr->pi_lock);
926 cpu_relax();
927 raw_spin_lock_irq(&curr->pi_lock);
928 continue;
929 }
930 raw_spin_unlock_irq(&curr->pi_lock);
931
932 spin_lock(&hb->lock);
933 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
934 raw_spin_lock(&curr->pi_lock);
935 /*
936 * We dropped the pi-lock, so re-check whether this
937 * task still owns the PI-state:
938 */
939 if (head->next != next) {
940 /* retain curr->pi_lock for the loop invariant */
941 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
942 spin_unlock(&hb->lock);
943 put_pi_state(pi_state);
944 continue;
945 }
946
947 WARN_ON(pi_state->owner != curr);
948 WARN_ON(list_empty(&pi_state->list));
949 list_del_init(&pi_state->list);
950 pi_state->owner = NULL;
951
952 raw_spin_unlock(&curr->pi_lock);
953 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
954 spin_unlock(&hb->lock);
955
956 rt_mutex_futex_unlock(&pi_state->pi_mutex);
957 put_pi_state(pi_state);
958
959 raw_spin_lock_irq(&curr->pi_lock);
960 }
961 raw_spin_unlock_irq(&curr->pi_lock);
962}
963
964#endif
965
966/*
967 * We need to check the following states:
968 *
969 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
970 *
971 * [1] NULL | --- | --- | 0 | 0/1 | Valid
972 * [2] NULL | --- | --- | >0 | 0/1 | Valid
973 *
974 * [3] Found | NULL | -- | Any | 0/1 | Invalid
975 *
976 * [4] Found | Found | NULL | 0 | 1 | Valid
977 * [5] Found | Found | NULL | >0 | 1 | Invalid
978 *
979 * [6] Found | Found | task | 0 | 1 | Valid
980 *
981 * [7] Found | Found | NULL | Any | 0 | Invalid
982 *
983 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
984 * [9] Found | Found | task | 0 | 0 | Invalid
985 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
986 *
987 * [1] Indicates that the kernel can acquire the futex atomically. We
988 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
989 *
990 * [2] Valid, if TID does not belong to a kernel thread. If no matching
991 * thread is found then it indicates that the owner TID has died.
992 *
993 * [3] Invalid. The waiter is queued on a non PI futex
994 *
995 * [4] Valid state after exit_robust_list(), which sets the user space
996 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
997 *
998 * [5] The user space value got manipulated between exit_robust_list()
999 * and exit_pi_state_list()
1000 *
1001 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1002 * the pi_state but cannot access the user space value.
1003 *
1004 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1005 *
1006 * [8] Owner and user space value match
1007 *
1008 * [9] There is no transient state which sets the user space TID to 0
1009 * except exit_robust_list(), but this is indicated by the
1010 * FUTEX_OWNER_DIED bit. See [4]
1011 *
1012 * [10] There is no transient state which leaves owner and user space
1013 * TID out of sync.
1014 *
1015 *
1016 * Serialization and lifetime rules:
1017 *
1018 * hb->lock:
1019 *
1020 * hb -> futex_q, relation
1021 * futex_q -> pi_state, relation
1022 *
1023 * (cannot be raw because hb can contain arbitrary amount
1024 * of futex_q's)
1025 *
1026 * pi_mutex->wait_lock:
1027 *
1028 * {uval, pi_state}
1029 *
1030 * (and pi_mutex 'obviously')
1031 *
1032 * p->pi_lock:
1033 *
1034 * p->pi_state_list -> pi_state->list, relation
1035 *
1036 * pi_state->refcount:
1037 *
1038 * pi_state lifetime
1039 *
1040 *
1041 * Lock order:
1042 *
1043 * hb->lock
1044 * pi_mutex->wait_lock
1045 * p->pi_lock
1046 *
1047 */
1048
1049/*
1050 * Validate that the existing waiter has a pi_state and sanity check
1051 * the pi_state against the user space value. If correct, attach to
1052 * it.
1053 */
1054static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1055 struct futex_pi_state *pi_state,
1056 struct futex_pi_state **ps)
1057{
1058 pid_t pid = uval & FUTEX_TID_MASK;
1059 u32 uval2;
1060 int ret;
1061
1062 /*
1063 * Userspace might have messed up non-PI and PI futexes [3]
1064 */
1065 if (unlikely(!pi_state))
1066 return -EINVAL;
1067
1068 /*
1069 * We get here with hb->lock held, and having found a
1070 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1071 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1072 * which in turn means that futex_lock_pi() still has a reference on
1073 * our pi_state.
1074 *
1075 * The waiter holding a reference on @pi_state also protects against
1076 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1077 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1078 * free pi_state before we can take a reference ourselves.
1079 */
1080 WARN_ON(!refcount_read(&pi_state->refcount));
1081
1082 /*
1083 * Now that we have a pi_state, we can acquire wait_lock
1084 * and do the state validation.
1085 */
1086 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1087
1088 /*
1089 * Since {uval, pi_state} is serialized by wait_lock, and our current
1090 * uval was read without holding it, it can have changed. Verify it
1091 * still is what we expect it to be, otherwise retry the entire
1092 * operation.
1093 */
1094 if (get_futex_value_locked(&uval2, uaddr))
1095 goto out_efault;
1096
1097 if (uval != uval2)
1098 goto out_eagain;
1099
1100 /*
1101 * Handle the owner died case:
1102 */
1103 if (uval & FUTEX_OWNER_DIED) {
1104 /*
1105 * exit_pi_state_list sets owner to NULL and wakes the
1106 * topmost waiter. The task which acquires the
1107 * pi_state->rt_mutex will fixup owner.
1108 */
1109 if (!pi_state->owner) {
1110 /*
1111 * No pi state owner, but the user space TID
1112 * is not 0. Inconsistent state. [5]
1113 */
1114 if (pid)
1115 goto out_einval;
1116 /*
1117 * Take a ref on the state and return success. [4]
1118 */
1119 goto out_attach;
1120 }
1121
1122 /*
1123 * If TID is 0, then either the dying owner has not
1124 * yet executed exit_pi_state_list() or some waiter
1125 * acquired the rtmutex in the pi state, but did not
1126 * yet fixup the TID in user space.
1127 *
1128 * Take a ref on the state and return success. [6]
1129 */
1130 if (!pid)
1131 goto out_attach;
1132 } else {
1133 /*
1134 * If the owner died bit is not set, then the pi_state
1135 * must have an owner. [7]
1136 */
1137 if (!pi_state->owner)
1138 goto out_einval;
1139 }
1140
1141 /*
1142 * Bail out if user space manipulated the futex value. If pi
1143 * state exists then the owner TID must be the same as the
1144 * user space TID. [9/10]
1145 */
1146 if (pid != task_pid_vnr(pi_state->owner))
1147 goto out_einval;
1148
1149out_attach:
1150 get_pi_state(pi_state);
1151 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1152 *ps = pi_state;
1153 return 0;
1154
1155out_einval:
1156 ret = -EINVAL;
1157 goto out_error;
1158
1159out_eagain:
1160 ret = -EAGAIN;
1161 goto out_error;
1162
1163out_efault:
1164 ret = -EFAULT;
1165 goto out_error;
1166
1167out_error:
1168 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1169 return ret;
1170}
1171
1172static int handle_exit_race(u32 __user *uaddr, u32 uval,
1173 struct task_struct *tsk)
1174{
1175 u32 uval2;
1176
1177 /*
1178 * If PF_EXITPIDONE is not yet set, then try again.
1179 */
1180 if (tsk && !(tsk->flags & PF_EXITPIDONE))
1181 return -EAGAIN;
1182
1183 /*
1184 * Reread the user space value to handle the following situation:
1185 *
1186 * CPU0 CPU1
1187 *
1188 * sys_exit() sys_futex()
1189 * do_exit() futex_lock_pi()
1190 * futex_lock_pi_atomic()
1191 * exit_signals(tsk) No waiters:
1192 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1193 * mm_release(tsk) Set waiter bit
1194 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1195 * Set owner died attach_to_pi_owner() {
1196 * *uaddr = 0xC0000000; tsk = get_task(PID);
1197 * } if (!tsk->flags & PF_EXITING) {
1198 * ... attach();
1199 * tsk->flags |= PF_EXITPIDONE; } else {
1200 * if (!(tsk->flags & PF_EXITPIDONE))
1201 * return -EAGAIN;
1202 * return -ESRCH; <--- FAIL
1203 * }
1204 *
1205 * Returning ESRCH unconditionally is wrong here because the
1206 * user space value has been changed by the exiting task.
1207 *
1208 * The same logic applies to the case where the exiting task is
1209 * already gone.
1210 */
1211 if (get_futex_value_locked(&uval2, uaddr))
1212 return -EFAULT;
1213
1214 /* If the user space value has changed, try again. */
1215 if (uval2 != uval)
1216 return -EAGAIN;
1217
1218 /*
1219 * The exiting task did not have a robust list, the robust list was
1220 * corrupted or the user space value in *uaddr is simply bogus.
1221 * Give up and tell user space.
1222 */
1223 return -ESRCH;
1224}
1225
1226/*
1227 * Lookup the task for the TID provided from user space and attach to
1228 * it after doing proper sanity checks.
1229 */
1230static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1231 struct futex_pi_state **ps)
1232{
1233 pid_t pid = uval & FUTEX_TID_MASK;
1234 struct futex_pi_state *pi_state;
1235 struct task_struct *p;
1236
1237 /*
1238 * We are the first waiter - try to look up the real owner and attach
1239 * the new pi_state to it, but bail out when TID = 0 [1]
1240 *
1241 * The !pid check is paranoid. None of the call sites should end up
1242 * with pid == 0, but better safe than sorry. Let the caller retry
1243 */
1244 if (!pid)
1245 return -EAGAIN;
1246 p = find_get_task_by_vpid(pid);
1247 if (!p)
1248 return handle_exit_race(uaddr, uval, NULL);
1249
1250 if (unlikely(p->flags & PF_KTHREAD)) {
1251 put_task_struct(p);
1252 return -EPERM;
1253 }
1254
1255 /*
1256 * We need to look at the task state flags to figure out,
1257 * whether the task is exiting. To protect against the do_exit
1258 * change of the task flags, we do this protected by
1259 * p->pi_lock:
1260 */
1261 raw_spin_lock_irq(&p->pi_lock);
1262 if (unlikely(p->flags & PF_EXITING)) {
1263 /*
1264 * The task is on the way out. When PF_EXITPIDONE is
1265 * set, we know that the task has finished the
1266 * cleanup:
1267 */
1268 int ret = handle_exit_race(uaddr, uval, p);
1269
1270 raw_spin_unlock_irq(&p->pi_lock);
1271 put_task_struct(p);
1272 return ret;
1273 }
1274
1275 /*
1276 * No existing pi state. First waiter. [2]
1277 *
1278 * This creates pi_state, we have hb->lock held, this means nothing can
1279 * observe this state, wait_lock is irrelevant.
1280 */
1281 pi_state = alloc_pi_state();
1282
1283 /*
1284 * Initialize the pi_mutex in locked state and make @p
1285 * the owner of it:
1286 */
1287 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1288
1289 /* Store the key for possible exit cleanups: */
1290 pi_state->key = *key;
1291
1292 WARN_ON(!list_empty(&pi_state->list));
1293 list_add(&pi_state->list, &p->pi_state_list);
1294 /*
1295 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1296 * because there is no concurrency as the object is not published yet.
1297 */
1298 pi_state->owner = p;
1299 raw_spin_unlock_irq(&p->pi_lock);
1300
1301 put_task_struct(p);
1302
1303 *ps = pi_state;
1304
1305 return 0;
1306}
1307
1308static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1309 struct futex_hash_bucket *hb,
1310 union futex_key *key, struct futex_pi_state **ps)
1311{
1312 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1313
1314 /*
1315 * If there is a waiter on that futex, validate it and
1316 * attach to the pi_state when the validation succeeds.
1317 */
1318 if (top_waiter)
1319 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1320
1321 /*
1322 * We are the first waiter - try to look up the owner based on
1323 * @uval and attach to it.
1324 */
1325 return attach_to_pi_owner(uaddr, uval, key, ps);
1326}
1327
1328static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1329{
1330 int err;
1331 u32 uninitialized_var(curval);
1332
1333 if (unlikely(should_fail_futex(true)))
1334 return -EFAULT;
1335
1336 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1337 if (unlikely(err))
1338 return err;
1339
1340 /* If user space value changed, let the caller retry */
1341 return curval != uval ? -EAGAIN : 0;
1342}
1343
1344/**
1345 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1346 * @uaddr: the pi futex user address
1347 * @hb: the pi futex hash bucket
1348 * @key: the futex key associated with uaddr and hb
1349 * @ps: the pi_state pointer where we store the result of the
1350 * lookup
1351 * @task: the task to perform the atomic lock work for. This will
1352 * be "current" except in the case of requeue pi.
1353 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1354 *
1355 * Return:
1356 * - 0 - ready to wait;
1357 * - 1 - acquired the lock;
1358 * - <0 - error
1359 *
1360 * The hb->lock and futex_key refs shall be held by the caller.
1361 */
1362static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1363 union futex_key *key,
1364 struct futex_pi_state **ps,
1365 struct task_struct *task, int set_waiters)
1366{
1367 u32 uval, newval, vpid = task_pid_vnr(task);
1368 struct futex_q *top_waiter;
1369 int ret;
1370
1371 /*
1372 * Read the user space value first so we can validate a few
1373 * things before proceeding further.
1374 */
1375 if (get_futex_value_locked(&uval, uaddr))
1376 return -EFAULT;
1377
1378 if (unlikely(should_fail_futex(true)))
1379 return -EFAULT;
1380
1381 /*
1382 * Detect deadlocks.
1383 */
1384 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1385 return -EDEADLK;
1386
1387 if ((unlikely(should_fail_futex(true))))
1388 return -EDEADLK;
1389
1390 /*
1391 * Lookup existing state first. If it exists, try to attach to
1392 * its pi_state.
1393 */
1394 top_waiter = futex_top_waiter(hb, key);
1395 if (top_waiter)
1396 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1397
1398 /*
1399 * No waiter and user TID is 0. We are here because the
1400 * waiters or the owner died bit is set or called from
1401 * requeue_cmp_pi or for whatever reason something took the
1402 * syscall.
1403 */
1404 if (!(uval & FUTEX_TID_MASK)) {
1405 /*
1406 * We take over the futex. No other waiters and the user space
1407 * TID is 0. We preserve the owner died bit.
1408 */
1409 newval = uval & FUTEX_OWNER_DIED;
1410 newval |= vpid;
1411
1412 /* The futex requeue_pi code can enforce the waiters bit */
1413 if (set_waiters)
1414 newval |= FUTEX_WAITERS;
1415
1416 ret = lock_pi_update_atomic(uaddr, uval, newval);
1417 /* If the take over worked, return 1 */
1418 return ret < 0 ? ret : 1;
1419 }
1420
1421 /*
1422 * First waiter. Set the waiters bit before attaching ourself to
1423 * the owner. If owner tries to unlock, it will be forced into
1424 * the kernel and blocked on hb->lock.
1425 */
1426 newval = uval | FUTEX_WAITERS;
1427 ret = lock_pi_update_atomic(uaddr, uval, newval);
1428 if (ret)
1429 return ret;
1430 /*
1431 * If the update of the user space value succeeded, we try to
1432 * attach to the owner. If that fails, no harm done, we only
1433 * set the FUTEX_WAITERS bit in the user space variable.
1434 */
1435 return attach_to_pi_owner(uaddr, newval, key, ps);
1436}
1437
1438/**
1439 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1440 * @q: The futex_q to unqueue
1441 *
1442 * The q->lock_ptr must not be NULL and must be held by the caller.
1443 */
1444static void __unqueue_futex(struct futex_q *q)
1445{
1446 struct futex_hash_bucket *hb;
1447
1448 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1449 return;
1450 lockdep_assert_held(q->lock_ptr);
1451
1452 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1453 plist_del(&q->list, &hb->chain);
1454 hb_waiters_dec(hb);
1455}
1456
1457/*
1458 * The hash bucket lock must be held when this is called.
1459 * Afterwards, the futex_q must not be accessed. Callers
1460 * must ensure to later call wake_up_q() for the actual
1461 * wakeups to occur.
1462 */
1463static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1464{
1465 struct task_struct *p = q->task;
1466
1467 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1468 return;
1469
1470 get_task_struct(p);
1471 __unqueue_futex(q);
1472 /*
1473 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1474 * is written, without taking any locks. This is possible in the event
1475 * of a spurious wakeup, for example. A memory barrier is required here
1476 * to prevent the following store to lock_ptr from getting ahead of the
1477 * plist_del in __unqueue_futex().
1478 */
1479 smp_store_release(&q->lock_ptr, NULL);
1480
1481 /*
1482 * Queue the task for later wakeup for after we've released
1483 * the hb->lock. wake_q_add() grabs reference to p.
1484 */
1485 wake_q_add_safe(wake_q, p);
1486}
1487
1488/*
1489 * Caller must hold a reference on @pi_state.
1490 */
1491static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1492{
1493 u32 uninitialized_var(curval), newval;
1494 struct task_struct *new_owner;
1495 bool postunlock = false;
1496 DEFINE_WAKE_Q(wake_q);
1497 int ret = 0;
1498
1499 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1500 if (WARN_ON_ONCE(!new_owner)) {
1501 /*
1502 * As per the comment in futex_unlock_pi() this should not happen.
1503 *
1504 * When this happens, give up our locks and try again, giving
1505 * the futex_lock_pi() instance time to complete, either by
1506 * waiting on the rtmutex or removing itself from the futex
1507 * queue.
1508 */
1509 ret = -EAGAIN;
1510 goto out_unlock;
1511 }
1512
1513 /*
1514 * We pass it to the next owner. The WAITERS bit is always kept
1515 * enabled while there is PI state around. We cleanup the owner
1516 * died bit, because we are the owner.
1517 */
1518 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1519
1520 if (unlikely(should_fail_futex(true)))
1521 ret = -EFAULT;
1522
1523 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1524 if (!ret && (curval != uval)) {
1525 /*
1526 * If a unconditional UNLOCK_PI operation (user space did not
1527 * try the TID->0 transition) raced with a waiter setting the
1528 * FUTEX_WAITERS flag between get_user() and locking the hash
1529 * bucket lock, retry the operation.
1530 */
1531 if ((FUTEX_TID_MASK & curval) == uval)
1532 ret = -EAGAIN;
1533 else
1534 ret = -EINVAL;
1535 }
1536
1537 if (ret)
1538 goto out_unlock;
1539
1540 /*
1541 * This is a point of no return; once we modify the uval there is no
1542 * going back and subsequent operations must not fail.
1543 */
1544
1545 raw_spin_lock(&pi_state->owner->pi_lock);
1546 WARN_ON(list_empty(&pi_state->list));
1547 list_del_init(&pi_state->list);
1548 raw_spin_unlock(&pi_state->owner->pi_lock);
1549
1550 raw_spin_lock(&new_owner->pi_lock);
1551 WARN_ON(!list_empty(&pi_state->list));
1552 list_add(&pi_state->list, &new_owner->pi_state_list);
1553 pi_state->owner = new_owner;
1554 raw_spin_unlock(&new_owner->pi_lock);
1555
1556 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1557
1558out_unlock:
1559 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1560
1561 if (postunlock)
1562 rt_mutex_postunlock(&wake_q);
1563
1564 return ret;
1565}
1566
1567/*
1568 * Express the locking dependencies for lockdep:
1569 */
1570static inline void
1571double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1572{
1573 if (hb1 <= hb2) {
1574 spin_lock(&hb1->lock);
1575 if (hb1 < hb2)
1576 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1577 } else { /* hb1 > hb2 */
1578 spin_lock(&hb2->lock);
1579 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1580 }
1581}
1582
1583static inline void
1584double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1585{
1586 spin_unlock(&hb1->lock);
1587 if (hb1 != hb2)
1588 spin_unlock(&hb2->lock);
1589}
1590
1591/*
1592 * Wake up waiters matching bitset queued on this futex (uaddr).
1593 */
1594static int
1595futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1596{
1597 struct futex_hash_bucket *hb;
1598 struct futex_q *this, *next;
1599 union futex_key key = FUTEX_KEY_INIT;
1600 int ret;
1601 DEFINE_WAKE_Q(wake_q);
1602
1603 if (!bitset)
1604 return -EINVAL;
1605
1606 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1607 if (unlikely(ret != 0))
1608 goto out;
1609
1610 hb = hash_futex(&key);
1611
1612 /* Make sure we really have tasks to wakeup */
1613 if (!hb_waiters_pending(hb))
1614 goto out_put_key;
1615
1616 spin_lock(&hb->lock);
1617
1618 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1619 if (match_futex (&this->key, &key)) {
1620 if (this->pi_state || this->rt_waiter) {
1621 ret = -EINVAL;
1622 break;
1623 }
1624
1625 /* Check if one of the bits is set in both bitsets */
1626 if (!(this->bitset & bitset))
1627 continue;
1628
1629 mark_wake_futex(&wake_q, this);
1630 if (++ret >= nr_wake)
1631 break;
1632 }
1633 }
1634
1635 spin_unlock(&hb->lock);
1636 wake_up_q(&wake_q);
1637out_put_key:
1638 put_futex_key(&key);
1639out:
1640 return ret;
1641}
1642
1643static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1644{
1645 unsigned int op = (encoded_op & 0x70000000) >> 28;
1646 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1647 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1648 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1649 int oldval, ret;
1650
1651 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1652 if (oparg < 0 || oparg > 31) {
1653 char comm[sizeof(current->comm)];
1654 /*
1655 * kill this print and return -EINVAL when userspace
1656 * is sane again
1657 */
1658 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1659 get_task_comm(comm, current), oparg);
1660 oparg &= 31;
1661 }
1662 oparg = 1 << oparg;
1663 }
1664
1665 if (!access_ok(uaddr, sizeof(u32)))
1666 return -EFAULT;
1667
1668 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1669 if (ret)
1670 return ret;
1671
1672 switch (cmp) {
1673 case FUTEX_OP_CMP_EQ:
1674 return oldval == cmparg;
1675 case FUTEX_OP_CMP_NE:
1676 return oldval != cmparg;
1677 case FUTEX_OP_CMP_LT:
1678 return oldval < cmparg;
1679 case FUTEX_OP_CMP_GE:
1680 return oldval >= cmparg;
1681 case FUTEX_OP_CMP_LE:
1682 return oldval <= cmparg;
1683 case FUTEX_OP_CMP_GT:
1684 return oldval > cmparg;
1685 default:
1686 return -ENOSYS;
1687 }
1688}
1689
1690/*
1691 * Wake up all waiters hashed on the physical page that is mapped
1692 * to this virtual address:
1693 */
1694static int
1695futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1696 int nr_wake, int nr_wake2, int op)
1697{
1698 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1699 struct futex_hash_bucket *hb1, *hb2;
1700 struct futex_q *this, *next;
1701 int ret, op_ret;
1702 DEFINE_WAKE_Q(wake_q);
1703
1704retry:
1705 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1706 if (unlikely(ret != 0))
1707 goto out;
1708 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1709 if (unlikely(ret != 0))
1710 goto out_put_key1;
1711
1712 hb1 = hash_futex(&key1);
1713 hb2 = hash_futex(&key2);
1714
1715retry_private:
1716 double_lock_hb(hb1, hb2);
1717 op_ret = futex_atomic_op_inuser(op, uaddr2);
1718 if (unlikely(op_ret < 0)) {
1719 double_unlock_hb(hb1, hb2);
1720
1721 if (!IS_ENABLED(CONFIG_MMU) ||
1722 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1723 /*
1724 * we don't get EFAULT from MMU faults if we don't have
1725 * an MMU, but we might get them from range checking
1726 */
1727 ret = op_ret;
1728 goto out_put_keys;
1729 }
1730
1731 if (op_ret == -EFAULT) {
1732 ret = fault_in_user_writeable(uaddr2);
1733 if (ret)
1734 goto out_put_keys;
1735 }
1736
1737 if (!(flags & FLAGS_SHARED)) {
1738 cond_resched();
1739 goto retry_private;
1740 }
1741
1742 put_futex_key(&key2);
1743 put_futex_key(&key1);
1744 cond_resched();
1745 goto retry;
1746 }
1747
1748 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1749 if (match_futex (&this->key, &key1)) {
1750 if (this->pi_state || this->rt_waiter) {
1751 ret = -EINVAL;
1752 goto out_unlock;
1753 }
1754 mark_wake_futex(&wake_q, this);
1755 if (++ret >= nr_wake)
1756 break;
1757 }
1758 }
1759
1760 if (op_ret > 0) {
1761 op_ret = 0;
1762 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1763 if (match_futex (&this->key, &key2)) {
1764 if (this->pi_state || this->rt_waiter) {
1765 ret = -EINVAL;
1766 goto out_unlock;
1767 }
1768 mark_wake_futex(&wake_q, this);
1769 if (++op_ret >= nr_wake2)
1770 break;
1771 }
1772 }
1773 ret += op_ret;
1774 }
1775
1776out_unlock:
1777 double_unlock_hb(hb1, hb2);
1778 wake_up_q(&wake_q);
1779out_put_keys:
1780 put_futex_key(&key2);
1781out_put_key1:
1782 put_futex_key(&key1);
1783out:
1784 return ret;
1785}
1786
1787/**
1788 * requeue_futex() - Requeue a futex_q from one hb to another
1789 * @q: the futex_q to requeue
1790 * @hb1: the source hash_bucket
1791 * @hb2: the target hash_bucket
1792 * @key2: the new key for the requeued futex_q
1793 */
1794static inline
1795void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1796 struct futex_hash_bucket *hb2, union futex_key *key2)
1797{
1798
1799 /*
1800 * If key1 and key2 hash to the same bucket, no need to
1801 * requeue.
1802 */
1803 if (likely(&hb1->chain != &hb2->chain)) {
1804 plist_del(&q->list, &hb1->chain);
1805 hb_waiters_dec(hb1);
1806 hb_waiters_inc(hb2);
1807 plist_add(&q->list, &hb2->chain);
1808 q->lock_ptr = &hb2->lock;
1809 }
1810 get_futex_key_refs(key2);
1811 q->key = *key2;
1812}
1813
1814/**
1815 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1816 * @q: the futex_q
1817 * @key: the key of the requeue target futex
1818 * @hb: the hash_bucket of the requeue target futex
1819 *
1820 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1821 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1822 * to the requeue target futex so the waiter can detect the wakeup on the right
1823 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1824 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1825 * to protect access to the pi_state to fixup the owner later. Must be called
1826 * with both q->lock_ptr and hb->lock held.
1827 */
1828static inline
1829void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1830 struct futex_hash_bucket *hb)
1831{
1832 get_futex_key_refs(key);
1833 q->key = *key;
1834
1835 __unqueue_futex(q);
1836
1837 WARN_ON(!q->rt_waiter);
1838 q->rt_waiter = NULL;
1839
1840 q->lock_ptr = &hb->lock;
1841
1842 wake_up_state(q->task, TASK_NORMAL);
1843}
1844
1845/**
1846 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1847 * @pifutex: the user address of the to futex
1848 * @hb1: the from futex hash bucket, must be locked by the caller
1849 * @hb2: the to futex hash bucket, must be locked by the caller
1850 * @key1: the from futex key
1851 * @key2: the to futex key
1852 * @ps: address to store the pi_state pointer
1853 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1854 *
1855 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1856 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1857 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1858 * hb1 and hb2 must be held by the caller.
1859 *
1860 * Return:
1861 * - 0 - failed to acquire the lock atomically;
1862 * - >0 - acquired the lock, return value is vpid of the top_waiter
1863 * - <0 - error
1864 */
1865static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1866 struct futex_hash_bucket *hb1,
1867 struct futex_hash_bucket *hb2,
1868 union futex_key *key1, union futex_key *key2,
1869 struct futex_pi_state **ps, int set_waiters)
1870{
1871 struct futex_q *top_waiter = NULL;
1872 u32 curval;
1873 int ret, vpid;
1874
1875 if (get_futex_value_locked(&curval, pifutex))
1876 return -EFAULT;
1877
1878 if (unlikely(should_fail_futex(true)))
1879 return -EFAULT;
1880
1881 /*
1882 * Find the top_waiter and determine if there are additional waiters.
1883 * If the caller intends to requeue more than 1 waiter to pifutex,
1884 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1885 * as we have means to handle the possible fault. If not, don't set
1886 * the bit unecessarily as it will force the subsequent unlock to enter
1887 * the kernel.
1888 */
1889 top_waiter = futex_top_waiter(hb1, key1);
1890
1891 /* There are no waiters, nothing for us to do. */
1892 if (!top_waiter)
1893 return 0;
1894
1895 /* Ensure we requeue to the expected futex. */
1896 if (!match_futex(top_waiter->requeue_pi_key, key2))
1897 return -EINVAL;
1898
1899 /*
1900 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1901 * the contended case or if set_waiters is 1. The pi_state is returned
1902 * in ps in contended cases.
1903 */
1904 vpid = task_pid_vnr(top_waiter->task);
1905 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1906 set_waiters);
1907 if (ret == 1) {
1908 requeue_pi_wake_futex(top_waiter, key2, hb2);
1909 return vpid;
1910 }
1911 return ret;
1912}
1913
1914/**
1915 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1916 * @uaddr1: source futex user address
1917 * @flags: futex flags (FLAGS_SHARED, etc.)
1918 * @uaddr2: target futex user address
1919 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1920 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1921 * @cmpval: @uaddr1 expected value (or %NULL)
1922 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1923 * pi futex (pi to pi requeue is not supported)
1924 *
1925 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1926 * uaddr2 atomically on behalf of the top waiter.
1927 *
1928 * Return:
1929 * - >=0 - on success, the number of tasks requeued or woken;
1930 * - <0 - on error
1931 */
1932static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1933 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1934 u32 *cmpval, int requeue_pi)
1935{
1936 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1937 int drop_count = 0, task_count = 0, ret;
1938 struct futex_pi_state *pi_state = NULL;
1939 struct futex_hash_bucket *hb1, *hb2;
1940 struct futex_q *this, *next;
1941 DEFINE_WAKE_Q(wake_q);
1942
1943 if (nr_wake < 0 || nr_requeue < 0)
1944 return -EINVAL;
1945
1946 /*
1947 * When PI not supported: return -ENOSYS if requeue_pi is true,
1948 * consequently the compiler knows requeue_pi is always false past
1949 * this point which will optimize away all the conditional code
1950 * further down.
1951 */
1952 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1953 return -ENOSYS;
1954
1955 if (requeue_pi) {
1956 /*
1957 * Requeue PI only works on two distinct uaddrs. This
1958 * check is only valid for private futexes. See below.
1959 */
1960 if (uaddr1 == uaddr2)
1961 return -EINVAL;
1962
1963 /*
1964 * requeue_pi requires a pi_state, try to allocate it now
1965 * without any locks in case it fails.
1966 */
1967 if (refill_pi_state_cache())
1968 return -ENOMEM;
1969 /*
1970 * requeue_pi must wake as many tasks as it can, up to nr_wake
1971 * + nr_requeue, since it acquires the rt_mutex prior to
1972 * returning to userspace, so as to not leave the rt_mutex with
1973 * waiters and no owner. However, second and third wake-ups
1974 * cannot be predicted as they involve race conditions with the
1975 * first wake and a fault while looking up the pi_state. Both
1976 * pthread_cond_signal() and pthread_cond_broadcast() should
1977 * use nr_wake=1.
1978 */
1979 if (nr_wake != 1)
1980 return -EINVAL;
1981 }
1982
1983retry:
1984 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1985 if (unlikely(ret != 0))
1986 goto out;
1987 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1988 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1989 if (unlikely(ret != 0))
1990 goto out_put_key1;
1991
1992 /*
1993 * The check above which compares uaddrs is not sufficient for
1994 * shared futexes. We need to compare the keys:
1995 */
1996 if (requeue_pi && match_futex(&key1, &key2)) {
1997 ret = -EINVAL;
1998 goto out_put_keys;
1999 }
2000
2001 hb1 = hash_futex(&key1);
2002 hb2 = hash_futex(&key2);
2003
2004retry_private:
2005 hb_waiters_inc(hb2);
2006 double_lock_hb(hb1, hb2);
2007
2008 if (likely(cmpval != NULL)) {
2009 u32 curval;
2010
2011 ret = get_futex_value_locked(&curval, uaddr1);
2012
2013 if (unlikely(ret)) {
2014 double_unlock_hb(hb1, hb2);
2015 hb_waiters_dec(hb2);
2016
2017 ret = get_user(curval, uaddr1);
2018 if (ret)
2019 goto out_put_keys;
2020
2021 if (!(flags & FLAGS_SHARED))
2022 goto retry_private;
2023
2024 put_futex_key(&key2);
2025 put_futex_key(&key1);
2026 goto retry;
2027 }
2028 if (curval != *cmpval) {
2029 ret = -EAGAIN;
2030 goto out_unlock;
2031 }
2032 }
2033
2034 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2035 /*
2036 * Attempt to acquire uaddr2 and wake the top waiter. If we
2037 * intend to requeue waiters, force setting the FUTEX_WAITERS
2038 * bit. We force this here where we are able to easily handle
2039 * faults rather in the requeue loop below.
2040 */
2041 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2042 &key2, &pi_state, nr_requeue);
2043
2044 /*
2045 * At this point the top_waiter has either taken uaddr2 or is
2046 * waiting on it. If the former, then the pi_state will not
2047 * exist yet, look it up one more time to ensure we have a
2048 * reference to it. If the lock was taken, ret contains the
2049 * vpid of the top waiter task.
2050 * If the lock was not taken, we have pi_state and an initial
2051 * refcount on it. In case of an error we have nothing.
2052 */
2053 if (ret > 0) {
2054 WARN_ON(pi_state);
2055 drop_count++;
2056 task_count++;
2057 /*
2058 * If we acquired the lock, then the user space value
2059 * of uaddr2 should be vpid. It cannot be changed by
2060 * the top waiter as it is blocked on hb2 lock if it
2061 * tries to do so. If something fiddled with it behind
2062 * our back the pi state lookup might unearth it. So
2063 * we rather use the known value than rereading and
2064 * handing potential crap to lookup_pi_state.
2065 *
2066 * If that call succeeds then we have pi_state and an
2067 * initial refcount on it.
2068 */
2069 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2070 }
2071
2072 switch (ret) {
2073 case 0:
2074 /* We hold a reference on the pi state. */
2075 break;
2076
2077 /* If the above failed, then pi_state is NULL */
2078 case -EFAULT:
2079 double_unlock_hb(hb1, hb2);
2080 hb_waiters_dec(hb2);
2081 put_futex_key(&key2);
2082 put_futex_key(&key1);
2083 ret = fault_in_user_writeable(uaddr2);
2084 if (!ret)
2085 goto retry;
2086 goto out;
2087 case -EAGAIN:
2088 /*
2089 * Two reasons for this:
2090 * - Owner is exiting and we just wait for the
2091 * exit to complete.
2092 * - The user space value changed.
2093 */
2094 double_unlock_hb(hb1, hb2);
2095 hb_waiters_dec(hb2);
2096 put_futex_key(&key2);
2097 put_futex_key(&key1);
2098 cond_resched();
2099 goto retry;
2100 default:
2101 goto out_unlock;
2102 }
2103 }
2104
2105 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2106 if (task_count - nr_wake >= nr_requeue)
2107 break;
2108
2109 if (!match_futex(&this->key, &key1))
2110 continue;
2111
2112 /*
2113 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2114 * be paired with each other and no other futex ops.
2115 *
2116 * We should never be requeueing a futex_q with a pi_state,
2117 * which is awaiting a futex_unlock_pi().
2118 */
2119 if ((requeue_pi && !this->rt_waiter) ||
2120 (!requeue_pi && this->rt_waiter) ||
2121 this->pi_state) {
2122 ret = -EINVAL;
2123 break;
2124 }
2125
2126 /*
2127 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2128 * lock, we already woke the top_waiter. If not, it will be
2129 * woken by futex_unlock_pi().
2130 */
2131 if (++task_count <= nr_wake && !requeue_pi) {
2132 mark_wake_futex(&wake_q, this);
2133 continue;
2134 }
2135
2136 /* Ensure we requeue to the expected futex for requeue_pi. */
2137 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2138 ret = -EINVAL;
2139 break;
2140 }
2141
2142 /*
2143 * Requeue nr_requeue waiters and possibly one more in the case
2144 * of requeue_pi if we couldn't acquire the lock atomically.
2145 */
2146 if (requeue_pi) {
2147 /*
2148 * Prepare the waiter to take the rt_mutex. Take a
2149 * refcount on the pi_state and store the pointer in
2150 * the futex_q object of the waiter.
2151 */
2152 get_pi_state(pi_state);
2153 this->pi_state = pi_state;
2154 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2155 this->rt_waiter,
2156 this->task);
2157 if (ret == 1) {
2158 /*
2159 * We got the lock. We do neither drop the
2160 * refcount on pi_state nor clear
2161 * this->pi_state because the waiter needs the
2162 * pi_state for cleaning up the user space
2163 * value. It will drop the refcount after
2164 * doing so.
2165 */
2166 requeue_pi_wake_futex(this, &key2, hb2);
2167 drop_count++;
2168 continue;
2169 } else if (ret) {
2170 /*
2171 * rt_mutex_start_proxy_lock() detected a
2172 * potential deadlock when we tried to queue
2173 * that waiter. Drop the pi_state reference
2174 * which we took above and remove the pointer
2175 * to the state from the waiters futex_q
2176 * object.
2177 */
2178 this->pi_state = NULL;
2179 put_pi_state(pi_state);
2180 /*
2181 * We stop queueing more waiters and let user
2182 * space deal with the mess.
2183 */
2184 break;
2185 }
2186 }
2187 requeue_futex(this, hb1, hb2, &key2);
2188 drop_count++;
2189 }
2190
2191 /*
2192 * We took an extra initial reference to the pi_state either
2193 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2194 * need to drop it here again.
2195 */
2196 put_pi_state(pi_state);
2197
2198out_unlock:
2199 double_unlock_hb(hb1, hb2);
2200 wake_up_q(&wake_q);
2201 hb_waiters_dec(hb2);
2202
2203 /*
2204 * drop_futex_key_refs() must be called outside the spinlocks. During
2205 * the requeue we moved futex_q's from the hash bucket at key1 to the
2206 * one at key2 and updated their key pointer. We no longer need to
2207 * hold the references to key1.
2208 */
2209 while (--drop_count >= 0)
2210 drop_futex_key_refs(&key1);
2211
2212out_put_keys:
2213 put_futex_key(&key2);
2214out_put_key1:
2215 put_futex_key(&key1);
2216out:
2217 return ret ? ret : task_count;
2218}
2219
2220/* The key must be already stored in q->key. */
2221static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2222 __acquires(&hb->lock)
2223{
2224 struct futex_hash_bucket *hb;
2225
2226 hb = hash_futex(&q->key);
2227
2228 /*
2229 * Increment the counter before taking the lock so that
2230 * a potential waker won't miss a to-be-slept task that is
2231 * waiting for the spinlock. This is safe as all queue_lock()
2232 * users end up calling queue_me(). Similarly, for housekeeping,
2233 * decrement the counter at queue_unlock() when some error has
2234 * occurred and we don't end up adding the task to the list.
2235 */
2236 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2237
2238 q->lock_ptr = &hb->lock;
2239
2240 spin_lock(&hb->lock);
2241 return hb;
2242}
2243
2244static inline void
2245queue_unlock(struct futex_hash_bucket *hb)
2246 __releases(&hb->lock)
2247{
2248 spin_unlock(&hb->lock);
2249 hb_waiters_dec(hb);
2250}
2251
2252static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2253{
2254 int prio;
2255
2256 /*
2257 * The priority used to register this element is
2258 * - either the real thread-priority for the real-time threads
2259 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2260 * - or MAX_RT_PRIO for non-RT threads.
2261 * Thus, all RT-threads are woken first in priority order, and
2262 * the others are woken last, in FIFO order.
2263 */
2264 prio = min(current->normal_prio, MAX_RT_PRIO);
2265
2266 plist_node_init(&q->list, prio);
2267 plist_add(&q->list, &hb->chain);
2268 q->task = current;
2269}
2270
2271/**
2272 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2273 * @q: The futex_q to enqueue
2274 * @hb: The destination hash bucket
2275 *
2276 * The hb->lock must be held by the caller, and is released here. A call to
2277 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2278 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2279 * or nothing if the unqueue is done as part of the wake process and the unqueue
2280 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2281 * an example).
2282 */
2283static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2284 __releases(&hb->lock)
2285{
2286 __queue_me(q, hb);
2287 spin_unlock(&hb->lock);
2288}
2289
2290/**
2291 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2292 * @q: The futex_q to unqueue
2293 *
2294 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2295 * be paired with exactly one earlier call to queue_me().
2296 *
2297 * Return:
2298 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2299 * - 0 - if the futex_q was already removed by the waking thread
2300 */
2301static int unqueue_me(struct futex_q *q)
2302{
2303 spinlock_t *lock_ptr;
2304 int ret = 0;
2305
2306 /* In the common case we don't take the spinlock, which is nice. */
2307retry:
2308 /*
2309 * q->lock_ptr can change between this read and the following spin_lock.
2310 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2311 * optimizing lock_ptr out of the logic below.
2312 */
2313 lock_ptr = READ_ONCE(q->lock_ptr);
2314 if (lock_ptr != NULL) {
2315 spin_lock(lock_ptr);
2316 /*
2317 * q->lock_ptr can change between reading it and
2318 * spin_lock(), causing us to take the wrong lock. This
2319 * corrects the race condition.
2320 *
2321 * Reasoning goes like this: if we have the wrong lock,
2322 * q->lock_ptr must have changed (maybe several times)
2323 * between reading it and the spin_lock(). It can
2324 * change again after the spin_lock() but only if it was
2325 * already changed before the spin_lock(). It cannot,
2326 * however, change back to the original value. Therefore
2327 * we can detect whether we acquired the correct lock.
2328 */
2329 if (unlikely(lock_ptr != q->lock_ptr)) {
2330 spin_unlock(lock_ptr);
2331 goto retry;
2332 }
2333 __unqueue_futex(q);
2334
2335 BUG_ON(q->pi_state);
2336
2337 spin_unlock(lock_ptr);
2338 ret = 1;
2339 }
2340
2341 drop_futex_key_refs(&q->key);
2342 return ret;
2343}
2344
2345/*
2346 * PI futexes can not be requeued and must remove themself from the
2347 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2348 * and dropped here.
2349 */
2350static void unqueue_me_pi(struct futex_q *q)
2351 __releases(q->lock_ptr)
2352{
2353 __unqueue_futex(q);
2354
2355 BUG_ON(!q->pi_state);
2356 put_pi_state(q->pi_state);
2357 q->pi_state = NULL;
2358
2359 spin_unlock(q->lock_ptr);
2360}
2361
2362static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2363 struct task_struct *argowner)
2364{
2365 struct futex_pi_state *pi_state = q->pi_state;
2366 u32 uval, uninitialized_var(curval), newval;
2367 struct task_struct *oldowner, *newowner;
2368 u32 newtid;
2369 int ret, err = 0;
2370
2371 lockdep_assert_held(q->lock_ptr);
2372
2373 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2374
2375 oldowner = pi_state->owner;
2376
2377 /*
2378 * We are here because either:
2379 *
2380 * - we stole the lock and pi_state->owner needs updating to reflect
2381 * that (@argowner == current),
2382 *
2383 * or:
2384 *
2385 * - someone stole our lock and we need to fix things to point to the
2386 * new owner (@argowner == NULL).
2387 *
2388 * Either way, we have to replace the TID in the user space variable.
2389 * This must be atomic as we have to preserve the owner died bit here.
2390 *
2391 * Note: We write the user space value _before_ changing the pi_state
2392 * because we can fault here. Imagine swapped out pages or a fork
2393 * that marked all the anonymous memory readonly for cow.
2394 *
2395 * Modifying pi_state _before_ the user space value would leave the
2396 * pi_state in an inconsistent state when we fault here, because we
2397 * need to drop the locks to handle the fault. This might be observed
2398 * in the PID check in lookup_pi_state.
2399 */
2400retry:
2401 if (!argowner) {
2402 if (oldowner != current) {
2403 /*
2404 * We raced against a concurrent self; things are
2405 * already fixed up. Nothing to do.
2406 */
2407 ret = 0;
2408 goto out_unlock;
2409 }
2410
2411 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2412 /* We got the lock after all, nothing to fix. */
2413 ret = 0;
2414 goto out_unlock;
2415 }
2416
2417 /*
2418 * Since we just failed the trylock; there must be an owner.
2419 */
2420 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2421 BUG_ON(!newowner);
2422 } else {
2423 WARN_ON_ONCE(argowner != current);
2424 if (oldowner == current) {
2425 /*
2426 * We raced against a concurrent self; things are
2427 * already fixed up. Nothing to do.
2428 */
2429 ret = 0;
2430 goto out_unlock;
2431 }
2432 newowner = argowner;
2433 }
2434
2435 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2436 /* Owner died? */
2437 if (!pi_state->owner)
2438 newtid |= FUTEX_OWNER_DIED;
2439
2440 err = get_futex_value_locked(&uval, uaddr);
2441 if (err)
2442 goto handle_err;
2443
2444 for (;;) {
2445 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2446
2447 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2448 if (err)
2449 goto handle_err;
2450
2451 if (curval == uval)
2452 break;
2453 uval = curval;
2454 }
2455
2456 /*
2457 * We fixed up user space. Now we need to fix the pi_state
2458 * itself.
2459 */
2460 if (pi_state->owner != NULL) {
2461 raw_spin_lock(&pi_state->owner->pi_lock);
2462 WARN_ON(list_empty(&pi_state->list));
2463 list_del_init(&pi_state->list);
2464 raw_spin_unlock(&pi_state->owner->pi_lock);
2465 }
2466
2467 pi_state->owner = newowner;
2468
2469 raw_spin_lock(&newowner->pi_lock);
2470 WARN_ON(!list_empty(&pi_state->list));
2471 list_add(&pi_state->list, &newowner->pi_state_list);
2472 raw_spin_unlock(&newowner->pi_lock);
2473 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2474
2475 return 0;
2476
2477 /*
2478 * In order to reschedule or handle a page fault, we need to drop the
2479 * locks here. In the case of a fault, this gives the other task
2480 * (either the highest priority waiter itself or the task which stole
2481 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2482 * are back from handling the fault we need to check the pi_state after
2483 * reacquiring the locks and before trying to do another fixup. When
2484 * the fixup has been done already we simply return.
2485 *
2486 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2487 * drop hb->lock since the caller owns the hb -> futex_q relation.
2488 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2489 */
2490handle_err:
2491 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2492 spin_unlock(q->lock_ptr);
2493
2494 switch (err) {
2495 case -EFAULT:
2496 ret = fault_in_user_writeable(uaddr);
2497 break;
2498
2499 case -EAGAIN:
2500 cond_resched();
2501 ret = 0;
2502 break;
2503
2504 default:
2505 WARN_ON_ONCE(1);
2506 ret = err;
2507 break;
2508 }
2509
2510 spin_lock(q->lock_ptr);
2511 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2512
2513 /*
2514 * Check if someone else fixed it for us:
2515 */
2516 if (pi_state->owner != oldowner) {
2517 ret = 0;
2518 goto out_unlock;
2519 }
2520
2521 if (ret)
2522 goto out_unlock;
2523
2524 goto retry;
2525
2526out_unlock:
2527 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2528 return ret;
2529}
2530
2531static long futex_wait_restart(struct restart_block *restart);
2532
2533/**
2534 * fixup_owner() - Post lock pi_state and corner case management
2535 * @uaddr: user address of the futex
2536 * @q: futex_q (contains pi_state and access to the rt_mutex)
2537 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2538 *
2539 * After attempting to lock an rt_mutex, this function is called to cleanup
2540 * the pi_state owner as well as handle race conditions that may allow us to
2541 * acquire the lock. Must be called with the hb lock held.
2542 *
2543 * Return:
2544 * - 1 - success, lock taken;
2545 * - 0 - success, lock not taken;
2546 * - <0 - on error (-EFAULT)
2547 */
2548static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2549{
2550 int ret = 0;
2551
2552 if (locked) {
2553 /*
2554 * Got the lock. We might not be the anticipated owner if we
2555 * did a lock-steal - fix up the PI-state in that case:
2556 *
2557 * Speculative pi_state->owner read (we don't hold wait_lock);
2558 * since we own the lock pi_state->owner == current is the
2559 * stable state, anything else needs more attention.
2560 */
2561 if (q->pi_state->owner != current)
2562 ret = fixup_pi_state_owner(uaddr, q, current);
2563 goto out;
2564 }
2565
2566 /*
2567 * If we didn't get the lock; check if anybody stole it from us. In
2568 * that case, we need to fix up the uval to point to them instead of
2569 * us, otherwise bad things happen. [10]
2570 *
2571 * Another speculative read; pi_state->owner == current is unstable
2572 * but needs our attention.
2573 */
2574 if (q->pi_state->owner == current) {
2575 ret = fixup_pi_state_owner(uaddr, q, NULL);
2576 goto out;
2577 }
2578
2579 /*
2580 * Paranoia check. If we did not take the lock, then we should not be
2581 * the owner of the rt_mutex.
2582 */
2583 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2584 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2585 "pi-state %p\n", ret,
2586 q->pi_state->pi_mutex.owner,
2587 q->pi_state->owner);
2588 }
2589
2590out:
2591 return ret ? ret : locked;
2592}
2593
2594/**
2595 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2596 * @hb: the futex hash bucket, must be locked by the caller
2597 * @q: the futex_q to queue up on
2598 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2599 */
2600static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2601 struct hrtimer_sleeper *timeout)
2602{
2603 /*
2604 * The task state is guaranteed to be set before another task can
2605 * wake it. set_current_state() is implemented using smp_store_mb() and
2606 * queue_me() calls spin_unlock() upon completion, both serializing
2607 * access to the hash list and forcing another memory barrier.
2608 */
2609 set_current_state(TASK_INTERRUPTIBLE);
2610 queue_me(q, hb);
2611
2612 /* Arm the timer */
2613 if (timeout)
2614 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2615
2616 /*
2617 * If we have been removed from the hash list, then another task
2618 * has tried to wake us, and we can skip the call to schedule().
2619 */
2620 if (likely(!plist_node_empty(&q->list))) {
2621 /*
2622 * If the timer has already expired, current will already be
2623 * flagged for rescheduling. Only call schedule if there
2624 * is no timeout, or if it has yet to expire.
2625 */
2626 if (!timeout || timeout->task)
2627 freezable_schedule();
2628 }
2629 __set_current_state(TASK_RUNNING);
2630}
2631
2632/**
2633 * futex_wait_setup() - Prepare to wait on a futex
2634 * @uaddr: the futex userspace address
2635 * @val: the expected value
2636 * @flags: futex flags (FLAGS_SHARED, etc.)
2637 * @q: the associated futex_q
2638 * @hb: storage for hash_bucket pointer to be returned to caller
2639 *
2640 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2641 * compare it with the expected value. Handle atomic faults internally.
2642 * Return with the hb lock held and a q.key reference on success, and unlocked
2643 * with no q.key reference on failure.
2644 *
2645 * Return:
2646 * - 0 - uaddr contains val and hb has been locked;
2647 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2648 */
2649static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2650 struct futex_q *q, struct futex_hash_bucket **hb)
2651{
2652 u32 uval;
2653 int ret;
2654
2655 /*
2656 * Access the page AFTER the hash-bucket is locked.
2657 * Order is important:
2658 *
2659 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2660 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2661 *
2662 * The basic logical guarantee of a futex is that it blocks ONLY
2663 * if cond(var) is known to be true at the time of blocking, for
2664 * any cond. If we locked the hash-bucket after testing *uaddr, that
2665 * would open a race condition where we could block indefinitely with
2666 * cond(var) false, which would violate the guarantee.
2667 *
2668 * On the other hand, we insert q and release the hash-bucket only
2669 * after testing *uaddr. This guarantees that futex_wait() will NOT
2670 * absorb a wakeup if *uaddr does not match the desired values
2671 * while the syscall executes.
2672 */
2673retry:
2674 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2675 if (unlikely(ret != 0))
2676 return ret;
2677
2678retry_private:
2679 *hb = queue_lock(q);
2680
2681 ret = get_futex_value_locked(&uval, uaddr);
2682
2683 if (ret) {
2684 queue_unlock(*hb);
2685
2686 ret = get_user(uval, uaddr);
2687 if (ret)
2688 goto out;
2689
2690 if (!(flags & FLAGS_SHARED))
2691 goto retry_private;
2692
2693 put_futex_key(&q->key);
2694 goto retry;
2695 }
2696
2697 if (uval != val) {
2698 queue_unlock(*hb);
2699 ret = -EWOULDBLOCK;
2700 }
2701
2702out:
2703 if (ret)
2704 put_futex_key(&q->key);
2705 return ret;
2706}
2707
2708static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2709 ktime_t *abs_time, u32 bitset)
2710{
2711 struct hrtimer_sleeper timeout, *to;
2712 struct restart_block *restart;
2713 struct futex_hash_bucket *hb;
2714 struct futex_q q = futex_q_init;
2715 int ret;
2716
2717 if (!bitset)
2718 return -EINVAL;
2719 q.bitset = bitset;
2720
2721 to = futex_setup_timer(abs_time, &timeout, flags,
2722 current->timer_slack_ns);
2723retry:
2724 /*
2725 * Prepare to wait on uaddr. On success, holds hb lock and increments
2726 * q.key refs.
2727 */
2728 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2729 if (ret)
2730 goto out;
2731
2732 /* queue_me and wait for wakeup, timeout, or a signal. */
2733 futex_wait_queue_me(hb, &q, to);
2734
2735 /* If we were woken (and unqueued), we succeeded, whatever. */
2736 ret = 0;
2737 /* unqueue_me() drops q.key ref */
2738 if (!unqueue_me(&q))
2739 goto out;
2740 ret = -ETIMEDOUT;
2741 if (to && !to->task)
2742 goto out;
2743
2744 /*
2745 * We expect signal_pending(current), but we might be the
2746 * victim of a spurious wakeup as well.
2747 */
2748 if (!signal_pending(current))
2749 goto retry;
2750
2751 ret = -ERESTARTSYS;
2752 if (!abs_time)
2753 goto out;
2754
2755 restart = ¤t->restart_block;
2756 restart->fn = futex_wait_restart;
2757 restart->futex.uaddr = uaddr;
2758 restart->futex.val = val;
2759 restart->futex.time = *abs_time;
2760 restart->futex.bitset = bitset;
2761 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2762
2763 ret = -ERESTART_RESTARTBLOCK;
2764
2765out:
2766 if (to) {
2767 hrtimer_cancel(&to->timer);
2768 destroy_hrtimer_on_stack(&to->timer);
2769 }
2770 return ret;
2771}
2772
2773
2774static long futex_wait_restart(struct restart_block *restart)
2775{
2776 u32 __user *uaddr = restart->futex.uaddr;
2777 ktime_t t, *tp = NULL;
2778
2779 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2780 t = restart->futex.time;
2781 tp = &t;
2782 }
2783 restart->fn = do_no_restart_syscall;
2784
2785 return (long)futex_wait(uaddr, restart->futex.flags,
2786 restart->futex.val, tp, restart->futex.bitset);
2787}
2788
2789
2790/*
2791 * Userspace tried a 0 -> TID atomic transition of the futex value
2792 * and failed. The kernel side here does the whole locking operation:
2793 * if there are waiters then it will block as a consequence of relying
2794 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2795 * a 0 value of the futex too.).
2796 *
2797 * Also serves as futex trylock_pi()'ing, and due semantics.
2798 */
2799static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2800 ktime_t *time, int trylock)
2801{
2802 struct hrtimer_sleeper timeout, *to;
2803 struct futex_pi_state *pi_state = NULL;
2804 struct rt_mutex_waiter rt_waiter;
2805 struct futex_hash_bucket *hb;
2806 struct futex_q q = futex_q_init;
2807 int res, ret;
2808
2809 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2810 return -ENOSYS;
2811
2812 if (refill_pi_state_cache())
2813 return -ENOMEM;
2814
2815 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2816
2817retry:
2818 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2819 if (unlikely(ret != 0))
2820 goto out;
2821
2822retry_private:
2823 hb = queue_lock(&q);
2824
2825 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2826 if (unlikely(ret)) {
2827 /*
2828 * Atomic work succeeded and we got the lock,
2829 * or failed. Either way, we do _not_ block.
2830 */
2831 switch (ret) {
2832 case 1:
2833 /* We got the lock. */
2834 ret = 0;
2835 goto out_unlock_put_key;
2836 case -EFAULT:
2837 goto uaddr_faulted;
2838 case -EAGAIN:
2839 /*
2840 * Two reasons for this:
2841 * - Task is exiting and we just wait for the
2842 * exit to complete.
2843 * - The user space value changed.
2844 */
2845 queue_unlock(hb);
2846 put_futex_key(&q.key);
2847 cond_resched();
2848 goto retry;
2849 default:
2850 goto out_unlock_put_key;
2851 }
2852 }
2853
2854 WARN_ON(!q.pi_state);
2855
2856 /*
2857 * Only actually queue now that the atomic ops are done:
2858 */
2859 __queue_me(&q, hb);
2860
2861 if (trylock) {
2862 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2863 /* Fixup the trylock return value: */
2864 ret = ret ? 0 : -EWOULDBLOCK;
2865 goto no_block;
2866 }
2867
2868 rt_mutex_init_waiter(&rt_waiter);
2869
2870 /*
2871 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2872 * hold it while doing rt_mutex_start_proxy(), because then it will
2873 * include hb->lock in the blocking chain, even through we'll not in
2874 * fact hold it while blocking. This will lead it to report -EDEADLK
2875 * and BUG when futex_unlock_pi() interleaves with this.
2876 *
2877 * Therefore acquire wait_lock while holding hb->lock, but drop the
2878 * latter before calling __rt_mutex_start_proxy_lock(). This
2879 * interleaves with futex_unlock_pi() -- which does a similar lock
2880 * handoff -- such that the latter can observe the futex_q::pi_state
2881 * before __rt_mutex_start_proxy_lock() is done.
2882 */
2883 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2884 spin_unlock(q.lock_ptr);
2885 /*
2886 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2887 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2888 * it sees the futex_q::pi_state.
2889 */
2890 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2891 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2892
2893 if (ret) {
2894 if (ret == 1)
2895 ret = 0;
2896 goto cleanup;
2897 }
2898
2899 if (unlikely(to))
2900 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2901
2902 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2903
2904cleanup:
2905 spin_lock(q.lock_ptr);
2906 /*
2907 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2908 * first acquire the hb->lock before removing the lock from the
2909 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2910 * lists consistent.
2911 *
2912 * In particular; it is important that futex_unlock_pi() can not
2913 * observe this inconsistency.
2914 */
2915 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2916 ret = 0;
2917
2918no_block:
2919 /*
2920 * Fixup the pi_state owner and possibly acquire the lock if we
2921 * haven't already.
2922 */
2923 res = fixup_owner(uaddr, &q, !ret);
2924 /*
2925 * If fixup_owner() returned an error, proprogate that. If it acquired
2926 * the lock, clear our -ETIMEDOUT or -EINTR.
2927 */
2928 if (res)
2929 ret = (res < 0) ? res : 0;
2930
2931 /*
2932 * If fixup_owner() faulted and was unable to handle the fault, unlock
2933 * it and return the fault to userspace.
2934 */
2935 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2936 pi_state = q.pi_state;
2937 get_pi_state(pi_state);
2938 }
2939
2940 /* Unqueue and drop the lock */
2941 unqueue_me_pi(&q);
2942
2943 if (pi_state) {
2944 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2945 put_pi_state(pi_state);
2946 }
2947
2948 goto out_put_key;
2949
2950out_unlock_put_key:
2951 queue_unlock(hb);
2952
2953out_put_key:
2954 put_futex_key(&q.key);
2955out:
2956 if (to) {
2957 hrtimer_cancel(&to->timer);
2958 destroy_hrtimer_on_stack(&to->timer);
2959 }
2960 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2961
2962uaddr_faulted:
2963 queue_unlock(hb);
2964
2965 ret = fault_in_user_writeable(uaddr);
2966 if (ret)
2967 goto out_put_key;
2968
2969 if (!(flags & FLAGS_SHARED))
2970 goto retry_private;
2971
2972 put_futex_key(&q.key);
2973 goto retry;
2974}
2975
2976/*
2977 * Userspace attempted a TID -> 0 atomic transition, and failed.
2978 * This is the in-kernel slowpath: we look up the PI state (if any),
2979 * and do the rt-mutex unlock.
2980 */
2981static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2982{
2983 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2984 union futex_key key = FUTEX_KEY_INIT;
2985 struct futex_hash_bucket *hb;
2986 struct futex_q *top_waiter;
2987 int ret;
2988
2989 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2990 return -ENOSYS;
2991
2992retry:
2993 if (get_user(uval, uaddr))
2994 return -EFAULT;
2995 /*
2996 * We release only a lock we actually own:
2997 */
2998 if ((uval & FUTEX_TID_MASK) != vpid)
2999 return -EPERM;
3000
3001 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
3002 if (ret)
3003 return ret;
3004
3005 hb = hash_futex(&key);
3006 spin_lock(&hb->lock);
3007
3008 /*
3009 * Check waiters first. We do not trust user space values at
3010 * all and we at least want to know if user space fiddled
3011 * with the futex value instead of blindly unlocking.
3012 */
3013 top_waiter = futex_top_waiter(hb, &key);
3014 if (top_waiter) {
3015 struct futex_pi_state *pi_state = top_waiter->pi_state;
3016
3017 ret = -EINVAL;
3018 if (!pi_state)
3019 goto out_unlock;
3020
3021 /*
3022 * If current does not own the pi_state then the futex is
3023 * inconsistent and user space fiddled with the futex value.
3024 */
3025 if (pi_state->owner != current)
3026 goto out_unlock;
3027
3028 get_pi_state(pi_state);
3029 /*
3030 * By taking wait_lock while still holding hb->lock, we ensure
3031 * there is no point where we hold neither; and therefore
3032 * wake_futex_pi() must observe a state consistent with what we
3033 * observed.
3034 *
3035 * In particular; this forces __rt_mutex_start_proxy() to
3036 * complete such that we're guaranteed to observe the
3037 * rt_waiter. Also see the WARN in wake_futex_pi().
3038 */
3039 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3040 spin_unlock(&hb->lock);
3041
3042 /* drops pi_state->pi_mutex.wait_lock */
3043 ret = wake_futex_pi(uaddr, uval, pi_state);
3044
3045 put_pi_state(pi_state);
3046
3047 /*
3048 * Success, we're done! No tricky corner cases.
3049 */
3050 if (!ret)
3051 goto out_putkey;
3052 /*
3053 * The atomic access to the futex value generated a
3054 * pagefault, so retry the user-access and the wakeup:
3055 */
3056 if (ret == -EFAULT)
3057 goto pi_faulted;
3058 /*
3059 * A unconditional UNLOCK_PI op raced against a waiter
3060 * setting the FUTEX_WAITERS bit. Try again.
3061 */
3062 if (ret == -EAGAIN)
3063 goto pi_retry;
3064 /*
3065 * wake_futex_pi has detected invalid state. Tell user
3066 * space.
3067 */
3068 goto out_putkey;
3069 }
3070
3071 /*
3072 * We have no kernel internal state, i.e. no waiters in the
3073 * kernel. Waiters which are about to queue themselves are stuck
3074 * on hb->lock. So we can safely ignore them. We do neither
3075 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3076 * owner.
3077 */
3078 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3079 spin_unlock(&hb->lock);
3080 switch (ret) {
3081 case -EFAULT:
3082 goto pi_faulted;
3083
3084 case -EAGAIN:
3085 goto pi_retry;
3086
3087 default:
3088 WARN_ON_ONCE(1);
3089 goto out_putkey;
3090 }
3091 }
3092
3093 /*
3094 * If uval has changed, let user space handle it.
3095 */
3096 ret = (curval == uval) ? 0 : -EAGAIN;
3097
3098out_unlock:
3099 spin_unlock(&hb->lock);
3100out_putkey:
3101 put_futex_key(&key);
3102 return ret;
3103
3104pi_retry:
3105 put_futex_key(&key);
3106 cond_resched();
3107 goto retry;
3108
3109pi_faulted:
3110 put_futex_key(&key);
3111
3112 ret = fault_in_user_writeable(uaddr);
3113 if (!ret)
3114 goto retry;
3115
3116 return ret;
3117}
3118
3119/**
3120 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3121 * @hb: the hash_bucket futex_q was original enqueued on
3122 * @q: the futex_q woken while waiting to be requeued
3123 * @key2: the futex_key of the requeue target futex
3124 * @timeout: the timeout associated with the wait (NULL if none)
3125 *
3126 * Detect if the task was woken on the initial futex as opposed to the requeue
3127 * target futex. If so, determine if it was a timeout or a signal that caused
3128 * the wakeup and return the appropriate error code to the caller. Must be
3129 * called with the hb lock held.
3130 *
3131 * Return:
3132 * - 0 = no early wakeup detected;
3133 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3134 */
3135static inline
3136int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3137 struct futex_q *q, union futex_key *key2,
3138 struct hrtimer_sleeper *timeout)
3139{
3140 int ret = 0;
3141
3142 /*
3143 * With the hb lock held, we avoid races while we process the wakeup.
3144 * We only need to hold hb (and not hb2) to ensure atomicity as the
3145 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3146 * It can't be requeued from uaddr2 to something else since we don't
3147 * support a PI aware source futex for requeue.
3148 */
3149 if (!match_futex(&q->key, key2)) {
3150 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3151 /*
3152 * We were woken prior to requeue by a timeout or a signal.
3153 * Unqueue the futex_q and determine which it was.
3154 */
3155 plist_del(&q->list, &hb->chain);
3156 hb_waiters_dec(hb);
3157
3158 /* Handle spurious wakeups gracefully */
3159 ret = -EWOULDBLOCK;
3160 if (timeout && !timeout->task)
3161 ret = -ETIMEDOUT;
3162 else if (signal_pending(current))
3163 ret = -ERESTARTNOINTR;
3164 }
3165 return ret;
3166}
3167
3168/**
3169 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3170 * @uaddr: the futex we initially wait on (non-pi)
3171 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3172 * the same type, no requeueing from private to shared, etc.
3173 * @val: the expected value of uaddr
3174 * @abs_time: absolute timeout
3175 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3176 * @uaddr2: the pi futex we will take prior to returning to user-space
3177 *
3178 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3179 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3180 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3181 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3182 * without one, the pi logic would not know which task to boost/deboost, if
3183 * there was a need to.
3184 *
3185 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3186 * via the following--
3187 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3188 * 2) wakeup on uaddr2 after a requeue
3189 * 3) signal
3190 * 4) timeout
3191 *
3192 * If 3, cleanup and return -ERESTARTNOINTR.
3193 *
3194 * If 2, we may then block on trying to take the rt_mutex and return via:
3195 * 5) successful lock
3196 * 6) signal
3197 * 7) timeout
3198 * 8) other lock acquisition failure
3199 *
3200 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3201 *
3202 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3203 *
3204 * Return:
3205 * - 0 - On success;
3206 * - <0 - On error
3207 */
3208static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3209 u32 val, ktime_t *abs_time, u32 bitset,
3210 u32 __user *uaddr2)
3211{
3212 struct hrtimer_sleeper timeout, *to;
3213 struct futex_pi_state *pi_state = NULL;
3214 struct rt_mutex_waiter rt_waiter;
3215 struct futex_hash_bucket *hb;
3216 union futex_key key2 = FUTEX_KEY_INIT;
3217 struct futex_q q = futex_q_init;
3218 int res, ret;
3219
3220 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3221 return -ENOSYS;
3222
3223 if (uaddr == uaddr2)
3224 return -EINVAL;
3225
3226 if (!bitset)
3227 return -EINVAL;
3228
3229 to = futex_setup_timer(abs_time, &timeout, flags,
3230 current->timer_slack_ns);
3231
3232 /*
3233 * The waiter is allocated on our stack, manipulated by the requeue
3234 * code while we sleep on uaddr.
3235 */
3236 rt_mutex_init_waiter(&rt_waiter);
3237
3238 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3239 if (unlikely(ret != 0))
3240 goto out;
3241
3242 q.bitset = bitset;
3243 q.rt_waiter = &rt_waiter;
3244 q.requeue_pi_key = &key2;
3245
3246 /*
3247 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3248 * count.
3249 */
3250 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3251 if (ret)
3252 goto out_key2;
3253
3254 /*
3255 * The check above which compares uaddrs is not sufficient for
3256 * shared futexes. We need to compare the keys:
3257 */
3258 if (match_futex(&q.key, &key2)) {
3259 queue_unlock(hb);
3260 ret = -EINVAL;
3261 goto out_put_keys;
3262 }
3263
3264 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3265 futex_wait_queue_me(hb, &q, to);
3266
3267 spin_lock(&hb->lock);
3268 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3269 spin_unlock(&hb->lock);
3270 if (ret)
3271 goto out_put_keys;
3272
3273 /*
3274 * In order for us to be here, we know our q.key == key2, and since
3275 * we took the hb->lock above, we also know that futex_requeue() has
3276 * completed and we no longer have to concern ourselves with a wakeup
3277 * race with the atomic proxy lock acquisition by the requeue code. The
3278 * futex_requeue dropped our key1 reference and incremented our key2
3279 * reference count.
3280 */
3281
3282 /* Check if the requeue code acquired the second futex for us. */
3283 if (!q.rt_waiter) {
3284 /*
3285 * Got the lock. We might not be the anticipated owner if we
3286 * did a lock-steal - fix up the PI-state in that case.
3287 */
3288 if (q.pi_state && (q.pi_state->owner != current)) {
3289 spin_lock(q.lock_ptr);
3290 ret = fixup_pi_state_owner(uaddr2, &q, current);
3291 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3292 pi_state = q.pi_state;
3293 get_pi_state(pi_state);
3294 }
3295 /*
3296 * Drop the reference to the pi state which
3297 * the requeue_pi() code acquired for us.
3298 */
3299 put_pi_state(q.pi_state);
3300 spin_unlock(q.lock_ptr);
3301 }
3302 } else {
3303 struct rt_mutex *pi_mutex;
3304
3305 /*
3306 * We have been woken up by futex_unlock_pi(), a timeout, or a
3307 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3308 * the pi_state.
3309 */
3310 WARN_ON(!q.pi_state);
3311 pi_mutex = &q.pi_state->pi_mutex;
3312 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3313
3314 spin_lock(q.lock_ptr);
3315 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3316 ret = 0;
3317
3318 debug_rt_mutex_free_waiter(&rt_waiter);
3319 /*
3320 * Fixup the pi_state owner and possibly acquire the lock if we
3321 * haven't already.
3322 */
3323 res = fixup_owner(uaddr2, &q, !ret);
3324 /*
3325 * If fixup_owner() returned an error, proprogate that. If it
3326 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3327 */
3328 if (res)
3329 ret = (res < 0) ? res : 0;
3330
3331 /*
3332 * If fixup_pi_state_owner() faulted and was unable to handle
3333 * the fault, unlock the rt_mutex and return the fault to
3334 * userspace.
3335 */
3336 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3337 pi_state = q.pi_state;
3338 get_pi_state(pi_state);
3339 }
3340
3341 /* Unqueue and drop the lock. */
3342 unqueue_me_pi(&q);
3343 }
3344
3345 if (pi_state) {
3346 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3347 put_pi_state(pi_state);
3348 }
3349
3350 if (ret == -EINTR) {
3351 /*
3352 * We've already been requeued, but cannot restart by calling
3353 * futex_lock_pi() directly. We could restart this syscall, but
3354 * it would detect that the user space "val" changed and return
3355 * -EWOULDBLOCK. Save the overhead of the restart and return
3356 * -EWOULDBLOCK directly.
3357 */
3358 ret = -EWOULDBLOCK;
3359 }
3360
3361out_put_keys:
3362 put_futex_key(&q.key);
3363out_key2:
3364 put_futex_key(&key2);
3365
3366out:
3367 if (to) {
3368 hrtimer_cancel(&to->timer);
3369 destroy_hrtimer_on_stack(&to->timer);
3370 }
3371 return ret;
3372}
3373
3374/*
3375 * Support for robust futexes: the kernel cleans up held futexes at
3376 * thread exit time.
3377 *
3378 * Implementation: user-space maintains a per-thread list of locks it
3379 * is holding. Upon do_exit(), the kernel carefully walks this list,
3380 * and marks all locks that are owned by this thread with the
3381 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3382 * always manipulated with the lock held, so the list is private and
3383 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3384 * field, to allow the kernel to clean up if the thread dies after
3385 * acquiring the lock, but just before it could have added itself to
3386 * the list. There can only be one such pending lock.
3387 */
3388
3389/**
3390 * sys_set_robust_list() - Set the robust-futex list head of a task
3391 * @head: pointer to the list-head
3392 * @len: length of the list-head, as userspace expects
3393 */
3394SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3395 size_t, len)
3396{
3397 if (!futex_cmpxchg_enabled)
3398 return -ENOSYS;
3399 /*
3400 * The kernel knows only one size for now:
3401 */
3402 if (unlikely(len != sizeof(*head)))
3403 return -EINVAL;
3404
3405 current->robust_list = head;
3406
3407 return 0;
3408}
3409
3410/**
3411 * sys_get_robust_list() - Get the robust-futex list head of a task
3412 * @pid: pid of the process [zero for current task]
3413 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3414 * @len_ptr: pointer to a length field, the kernel fills in the header size
3415 */
3416SYSCALL_DEFINE3(get_robust_list, int, pid,
3417 struct robust_list_head __user * __user *, head_ptr,
3418 size_t __user *, len_ptr)
3419{
3420 struct robust_list_head __user *head;
3421 unsigned long ret;
3422 struct task_struct *p;
3423
3424 if (!futex_cmpxchg_enabled)
3425 return -ENOSYS;
3426
3427 rcu_read_lock();
3428
3429 ret = -ESRCH;
3430 if (!pid)
3431 p = current;
3432 else {
3433 p = find_task_by_vpid(pid);
3434 if (!p)
3435 goto err_unlock;
3436 }
3437
3438 ret = -EPERM;
3439 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3440 goto err_unlock;
3441
3442 head = p->robust_list;
3443 rcu_read_unlock();
3444
3445 if (put_user(sizeof(*head), len_ptr))
3446 return -EFAULT;
3447 return put_user(head, head_ptr);
3448
3449err_unlock:
3450 rcu_read_unlock();
3451
3452 return ret;
3453}
3454
3455/*
3456 * Process a futex-list entry, check whether it's owned by the
3457 * dying task, and do notification if so:
3458 */
3459static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3460{
3461 u32 uval, uninitialized_var(nval), mval;
3462 int err;
3463
3464 /* Futex address must be 32bit aligned */
3465 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3466 return -1;
3467
3468retry:
3469 if (get_user(uval, uaddr))
3470 return -1;
3471
3472 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3473 return 0;
3474
3475 /*
3476 * Ok, this dying thread is truly holding a futex
3477 * of interest. Set the OWNER_DIED bit atomically
3478 * via cmpxchg, and if the value had FUTEX_WAITERS
3479 * set, wake up a waiter (if any). (We have to do a
3480 * futex_wake() even if OWNER_DIED is already set -
3481 * to handle the rare but possible case of recursive
3482 * thread-death.) The rest of the cleanup is done in
3483 * userspace.
3484 */
3485 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3486
3487 /*
3488 * We are not holding a lock here, but we want to have
3489 * the pagefault_disable/enable() protection because
3490 * we want to handle the fault gracefully. If the
3491 * access fails we try to fault in the futex with R/W
3492 * verification via get_user_pages. get_user() above
3493 * does not guarantee R/W access. If that fails we
3494 * give up and leave the futex locked.
3495 */
3496 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3497 switch (err) {
3498 case -EFAULT:
3499 if (fault_in_user_writeable(uaddr))
3500 return -1;
3501 goto retry;
3502
3503 case -EAGAIN:
3504 cond_resched();
3505 goto retry;
3506
3507 default:
3508 WARN_ON_ONCE(1);
3509 return err;
3510 }
3511 }
3512
3513 if (nval != uval)
3514 goto retry;
3515
3516 /*
3517 * Wake robust non-PI futexes here. The wakeup of
3518 * PI futexes happens in exit_pi_state():
3519 */
3520 if (!pi && (uval & FUTEX_WAITERS))
3521 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3522
3523 return 0;
3524}
3525
3526/*
3527 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3528 */
3529static inline int fetch_robust_entry(struct robust_list __user **entry,
3530 struct robust_list __user * __user *head,
3531 unsigned int *pi)
3532{
3533 unsigned long uentry;
3534
3535 if (get_user(uentry, (unsigned long __user *)head))
3536 return -EFAULT;
3537
3538 *entry = (void __user *)(uentry & ~1UL);
3539 *pi = uentry & 1;
3540
3541 return 0;
3542}
3543
3544/*
3545 * Walk curr->robust_list (very carefully, it's a userspace list!)
3546 * and mark any locks found there dead, and notify any waiters.
3547 *
3548 * We silently return on any sign of list-walking problem.
3549 */
3550void exit_robust_list(struct task_struct *curr)
3551{
3552 struct robust_list_head __user *head = curr->robust_list;
3553 struct robust_list __user *entry, *next_entry, *pending;
3554 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3555 unsigned int uninitialized_var(next_pi);
3556 unsigned long futex_offset;
3557 int rc;
3558
3559 if (!futex_cmpxchg_enabled)
3560 return;
3561
3562 /*
3563 * Fetch the list head (which was registered earlier, via
3564 * sys_set_robust_list()):
3565 */
3566 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3567 return;
3568 /*
3569 * Fetch the relative futex offset:
3570 */
3571 if (get_user(futex_offset, &head->futex_offset))
3572 return;
3573 /*
3574 * Fetch any possibly pending lock-add first, and handle it
3575 * if it exists:
3576 */
3577 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3578 return;
3579
3580 next_entry = NULL; /* avoid warning with gcc */
3581 while (entry != &head->list) {
3582 /*
3583 * Fetch the next entry in the list before calling
3584 * handle_futex_death:
3585 */
3586 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3587 /*
3588 * A pending lock might already be on the list, so
3589 * don't process it twice:
3590 */
3591 if (entry != pending)
3592 if (handle_futex_death((void __user *)entry + futex_offset,
3593 curr, pi))
3594 return;
3595 if (rc)
3596 return;
3597 entry = next_entry;
3598 pi = next_pi;
3599 /*
3600 * Avoid excessively long or circular lists:
3601 */
3602 if (!--limit)
3603 break;
3604
3605 cond_resched();
3606 }
3607
3608 if (pending)
3609 handle_futex_death((void __user *)pending + futex_offset,
3610 curr, pip);
3611}
3612
3613long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3614 u32 __user *uaddr2, u32 val2, u32 val3)
3615{
3616 int cmd = op & FUTEX_CMD_MASK;
3617 unsigned int flags = 0;
3618
3619 if (!(op & FUTEX_PRIVATE_FLAG))
3620 flags |= FLAGS_SHARED;
3621
3622 if (op & FUTEX_CLOCK_REALTIME) {
3623 flags |= FLAGS_CLOCKRT;
3624 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3625 cmd != FUTEX_WAIT_REQUEUE_PI)
3626 return -ENOSYS;
3627 }
3628
3629 switch (cmd) {
3630 case FUTEX_LOCK_PI:
3631 case FUTEX_UNLOCK_PI:
3632 case FUTEX_TRYLOCK_PI:
3633 case FUTEX_WAIT_REQUEUE_PI:
3634 case FUTEX_CMP_REQUEUE_PI:
3635 if (!futex_cmpxchg_enabled)
3636 return -ENOSYS;
3637 }
3638
3639 switch (cmd) {
3640 case FUTEX_WAIT:
3641 val3 = FUTEX_BITSET_MATCH_ANY;
3642 /* fall through */
3643 case FUTEX_WAIT_BITSET:
3644 return futex_wait(uaddr, flags, val, timeout, val3);
3645 case FUTEX_WAKE:
3646 val3 = FUTEX_BITSET_MATCH_ANY;
3647 /* fall through */
3648 case FUTEX_WAKE_BITSET:
3649 return futex_wake(uaddr, flags, val, val3);
3650 case FUTEX_REQUEUE:
3651 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3652 case FUTEX_CMP_REQUEUE:
3653 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3654 case FUTEX_WAKE_OP:
3655 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3656 case FUTEX_LOCK_PI:
3657 return futex_lock_pi(uaddr, flags, timeout, 0);
3658 case FUTEX_UNLOCK_PI:
3659 return futex_unlock_pi(uaddr, flags);
3660 case FUTEX_TRYLOCK_PI:
3661 return futex_lock_pi(uaddr, flags, NULL, 1);
3662 case FUTEX_WAIT_REQUEUE_PI:
3663 val3 = FUTEX_BITSET_MATCH_ANY;
3664 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3665 uaddr2);
3666 case FUTEX_CMP_REQUEUE_PI:
3667 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3668 }
3669 return -ENOSYS;
3670}
3671
3672
3673SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3674 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3675 u32, val3)
3676{
3677 struct timespec64 ts;
3678 ktime_t t, *tp = NULL;
3679 u32 val2 = 0;
3680 int cmd = op & FUTEX_CMD_MASK;
3681
3682 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3683 cmd == FUTEX_WAIT_BITSET ||
3684 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3685 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3686 return -EFAULT;
3687 if (get_timespec64(&ts, utime))
3688 return -EFAULT;
3689 if (!timespec64_valid(&ts))
3690 return -EINVAL;
3691
3692 t = timespec64_to_ktime(ts);
3693 if (cmd == FUTEX_WAIT)
3694 t = ktime_add_safe(ktime_get(), t);
3695 tp = &t;
3696 }
3697 /*
3698 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3699 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3700 */
3701 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3702 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3703 val2 = (u32) (unsigned long) utime;
3704
3705 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3706}
3707
3708#ifdef CONFIG_COMPAT
3709/*
3710 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3711 */
3712static inline int
3713compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3714 compat_uptr_t __user *head, unsigned int *pi)
3715{
3716 if (get_user(*uentry, head))
3717 return -EFAULT;
3718
3719 *entry = compat_ptr((*uentry) & ~1);
3720 *pi = (unsigned int)(*uentry) & 1;
3721
3722 return 0;
3723}
3724
3725static void __user *futex_uaddr(struct robust_list __user *entry,
3726 compat_long_t futex_offset)
3727{
3728 compat_uptr_t base = ptr_to_compat(entry);
3729 void __user *uaddr = compat_ptr(base + futex_offset);
3730
3731 return uaddr;
3732}
3733
3734/*
3735 * Walk curr->robust_list (very carefully, it's a userspace list!)
3736 * and mark any locks found there dead, and notify any waiters.
3737 *
3738 * We silently return on any sign of list-walking problem.
3739 */
3740void compat_exit_robust_list(struct task_struct *curr)
3741{
3742 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3743 struct robust_list __user *entry, *next_entry, *pending;
3744 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3745 unsigned int uninitialized_var(next_pi);
3746 compat_uptr_t uentry, next_uentry, upending;
3747 compat_long_t futex_offset;
3748 int rc;
3749
3750 if (!futex_cmpxchg_enabled)
3751 return;
3752
3753 /*
3754 * Fetch the list head (which was registered earlier, via
3755 * sys_set_robust_list()):
3756 */
3757 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3758 return;
3759 /*
3760 * Fetch the relative futex offset:
3761 */
3762 if (get_user(futex_offset, &head->futex_offset))
3763 return;
3764 /*
3765 * Fetch any possibly pending lock-add first, and handle it
3766 * if it exists:
3767 */
3768 if (compat_fetch_robust_entry(&upending, &pending,
3769 &head->list_op_pending, &pip))
3770 return;
3771
3772 next_entry = NULL; /* avoid warning with gcc */
3773 while (entry != (struct robust_list __user *) &head->list) {
3774 /*
3775 * Fetch the next entry in the list before calling
3776 * handle_futex_death:
3777 */
3778 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3779 (compat_uptr_t __user *)&entry->next, &next_pi);
3780 /*
3781 * A pending lock might already be on the list, so
3782 * dont process it twice:
3783 */
3784 if (entry != pending) {
3785 void __user *uaddr = futex_uaddr(entry, futex_offset);
3786
3787 if (handle_futex_death(uaddr, curr, pi))
3788 return;
3789 }
3790 if (rc)
3791 return;
3792 uentry = next_uentry;
3793 entry = next_entry;
3794 pi = next_pi;
3795 /*
3796 * Avoid excessively long or circular lists:
3797 */
3798 if (!--limit)
3799 break;
3800
3801 cond_resched();
3802 }
3803 if (pending) {
3804 void __user *uaddr = futex_uaddr(pending, futex_offset);
3805
3806 handle_futex_death(uaddr, curr, pip);
3807 }
3808}
3809
3810COMPAT_SYSCALL_DEFINE2(set_robust_list,
3811 struct compat_robust_list_head __user *, head,
3812 compat_size_t, len)
3813{
3814 if (!futex_cmpxchg_enabled)
3815 return -ENOSYS;
3816
3817 if (unlikely(len != sizeof(*head)))
3818 return -EINVAL;
3819
3820 current->compat_robust_list = head;
3821
3822 return 0;
3823}
3824
3825COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3826 compat_uptr_t __user *, head_ptr,
3827 compat_size_t __user *, len_ptr)
3828{
3829 struct compat_robust_list_head __user *head;
3830 unsigned long ret;
3831 struct task_struct *p;
3832
3833 if (!futex_cmpxchg_enabled)
3834 return -ENOSYS;
3835
3836 rcu_read_lock();
3837
3838 ret = -ESRCH;
3839 if (!pid)
3840 p = current;
3841 else {
3842 p = find_task_by_vpid(pid);
3843 if (!p)
3844 goto err_unlock;
3845 }
3846
3847 ret = -EPERM;
3848 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3849 goto err_unlock;
3850
3851 head = p->compat_robust_list;
3852 rcu_read_unlock();
3853
3854 if (put_user(sizeof(*head), len_ptr))
3855 return -EFAULT;
3856 return put_user(ptr_to_compat(head), head_ptr);
3857
3858err_unlock:
3859 rcu_read_unlock();
3860
3861 return ret;
3862}
3863#endif /* CONFIG_COMPAT */
3864
3865#ifdef CONFIG_COMPAT_32BIT_TIME
3866SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3867 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3868 u32, val3)
3869{
3870 struct timespec64 ts;
3871 ktime_t t, *tp = NULL;
3872 int val2 = 0;
3873 int cmd = op & FUTEX_CMD_MASK;
3874
3875 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3876 cmd == FUTEX_WAIT_BITSET ||
3877 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3878 if (get_old_timespec32(&ts, utime))
3879 return -EFAULT;
3880 if (!timespec64_valid(&ts))
3881 return -EINVAL;
3882
3883 t = timespec64_to_ktime(ts);
3884 if (cmd == FUTEX_WAIT)
3885 t = ktime_add_safe(ktime_get(), t);
3886 tp = &t;
3887 }
3888 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3889 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3890 val2 = (int) (unsigned long) utime;
3891
3892 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3893}
3894#endif /* CONFIG_COMPAT_32BIT_TIME */
3895
3896static void __init futex_detect_cmpxchg(void)
3897{
3898#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3899 u32 curval;
3900
3901 /*
3902 * This will fail and we want it. Some arch implementations do
3903 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3904 * functionality. We want to know that before we call in any
3905 * of the complex code paths. Also we want to prevent
3906 * registration of robust lists in that case. NULL is
3907 * guaranteed to fault and we get -EFAULT on functional
3908 * implementation, the non-functional ones will return
3909 * -ENOSYS.
3910 */
3911 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3912 futex_cmpxchg_enabled = 1;
3913#endif
3914}
3915
3916static int __init futex_init(void)
3917{
3918 unsigned int futex_shift;
3919 unsigned long i;
3920
3921#if CONFIG_BASE_SMALL
3922 futex_hashsize = 16;
3923#else
3924 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3925#endif
3926
3927 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3928 futex_hashsize, 0,
3929 futex_hashsize < 256 ? HASH_SMALL : 0,
3930 &futex_shift, NULL,
3931 futex_hashsize, futex_hashsize);
3932 futex_hashsize = 1UL << futex_shift;
3933
3934 futex_detect_cmpxchg();
3935
3936 for (i = 0; i < futex_hashsize; i++) {
3937 atomic_set(&futex_queues[i].waiters, 0);
3938 plist_head_init(&futex_queues[i].chain);
3939 spin_lock_init(&futex_queues[i].lock);
3940 }
3941
3942 return 0;
3943}
3944core_initcall(futex_init);
1/*
2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
4 *
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7 *
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
10 *
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14 *
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18 *
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21 *
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
25 *
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
29 *
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
32 *
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
37 *
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
42 *
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
46 */
47#include <linux/slab.h>
48#include <linux/poll.h>
49#include <linux/fs.h>
50#include <linux/file.h>
51#include <linux/jhash.h>
52#include <linux/init.h>
53#include <linux/futex.h>
54#include <linux/mount.h>
55#include <linux/pagemap.h>
56#include <linux/syscalls.h>
57#include <linux/signal.h>
58#include <linux/export.h>
59#include <linux/magic.h>
60#include <linux/pid.h>
61#include <linux/nsproxy.h>
62#include <linux/ptrace.h>
63#include <linux/sched/rt.h>
64#include <linux/sched/wake_q.h>
65#include <linux/sched/mm.h>
66#include <linux/hugetlb.h>
67#include <linux/freezer.h>
68#include <linux/bootmem.h>
69#include <linux/fault-inject.h>
70
71#include <asm/futex.h>
72
73#include "locking/rtmutex_common.h"
74
75/*
76 * READ this before attempting to hack on futexes!
77 *
78 * Basic futex operation and ordering guarantees
79 * =============================================
80 *
81 * The waiter reads the futex value in user space and calls
82 * futex_wait(). This function computes the hash bucket and acquires
83 * the hash bucket lock. After that it reads the futex user space value
84 * again and verifies that the data has not changed. If it has not changed
85 * it enqueues itself into the hash bucket, releases the hash bucket lock
86 * and schedules.
87 *
88 * The waker side modifies the user space value of the futex and calls
89 * futex_wake(). This function computes the hash bucket and acquires the
90 * hash bucket lock. Then it looks for waiters on that futex in the hash
91 * bucket and wakes them.
92 *
93 * In futex wake up scenarios where no tasks are blocked on a futex, taking
94 * the hb spinlock can be avoided and simply return. In order for this
95 * optimization to work, ordering guarantees must exist so that the waiter
96 * being added to the list is acknowledged when the list is concurrently being
97 * checked by the waker, avoiding scenarios like the following:
98 *
99 * CPU 0 CPU 1
100 * val = *futex;
101 * sys_futex(WAIT, futex, val);
102 * futex_wait(futex, val);
103 * uval = *futex;
104 * *futex = newval;
105 * sys_futex(WAKE, futex);
106 * futex_wake(futex);
107 * if (queue_empty())
108 * return;
109 * if (uval == val)
110 * lock(hash_bucket(futex));
111 * queue();
112 * unlock(hash_bucket(futex));
113 * schedule();
114 *
115 * This would cause the waiter on CPU 0 to wait forever because it
116 * missed the transition of the user space value from val to newval
117 * and the waker did not find the waiter in the hash bucket queue.
118 *
119 * The correct serialization ensures that a waiter either observes
120 * the changed user space value before blocking or is woken by a
121 * concurrent waker:
122 *
123 * CPU 0 CPU 1
124 * val = *futex;
125 * sys_futex(WAIT, futex, val);
126 * futex_wait(futex, val);
127 *
128 * waiters++; (a)
129 * smp_mb(); (A) <-- paired with -.
130 * |
131 * lock(hash_bucket(futex)); |
132 * |
133 * uval = *futex; |
134 * | *futex = newval;
135 * | sys_futex(WAKE, futex);
136 * | futex_wake(futex);
137 * |
138 * `--------> smp_mb(); (B)
139 * if (uval == val)
140 * queue();
141 * unlock(hash_bucket(futex));
142 * schedule(); if (waiters)
143 * lock(hash_bucket(futex));
144 * else wake_waiters(futex);
145 * waiters--; (b) unlock(hash_bucket(futex));
146 *
147 * Where (A) orders the waiters increment and the futex value read through
148 * atomic operations (see hb_waiters_inc) and where (B) orders the write
149 * to futex and the waiters read -- this is done by the barriers for both
150 * shared and private futexes in get_futex_key_refs().
151 *
152 * This yields the following case (where X:=waiters, Y:=futex):
153 *
154 * X = Y = 0
155 *
156 * w[X]=1 w[Y]=1
157 * MB MB
158 * r[Y]=y r[X]=x
159 *
160 * Which guarantees that x==0 && y==0 is impossible; which translates back into
161 * the guarantee that we cannot both miss the futex variable change and the
162 * enqueue.
163 *
164 * Note that a new waiter is accounted for in (a) even when it is possible that
165 * the wait call can return error, in which case we backtrack from it in (b).
166 * Refer to the comment in queue_lock().
167 *
168 * Similarly, in order to account for waiters being requeued on another
169 * address we always increment the waiters for the destination bucket before
170 * acquiring the lock. It then decrements them again after releasing it -
171 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172 * will do the additional required waiter count housekeeping. This is done for
173 * double_lock_hb() and double_unlock_hb(), respectively.
174 */
175
176#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
177int __read_mostly futex_cmpxchg_enabled;
178#endif
179
180/*
181 * Futex flags used to encode options to functions and preserve them across
182 * restarts.
183 */
184#ifdef CONFIG_MMU
185# define FLAGS_SHARED 0x01
186#else
187/*
188 * NOMMU does not have per process address space. Let the compiler optimize
189 * code away.
190 */
191# define FLAGS_SHARED 0x00
192#endif
193#define FLAGS_CLOCKRT 0x02
194#define FLAGS_HAS_TIMEOUT 0x04
195
196/*
197 * Priority Inheritance state:
198 */
199struct futex_pi_state {
200 /*
201 * list of 'owned' pi_state instances - these have to be
202 * cleaned up in do_exit() if the task exits prematurely:
203 */
204 struct list_head list;
205
206 /*
207 * The PI object:
208 */
209 struct rt_mutex pi_mutex;
210
211 struct task_struct *owner;
212 atomic_t refcount;
213
214 union futex_key key;
215} __randomize_layout;
216
217/**
218 * struct futex_q - The hashed futex queue entry, one per waiting task
219 * @list: priority-sorted list of tasks waiting on this futex
220 * @task: the task waiting on the futex
221 * @lock_ptr: the hash bucket lock
222 * @key: the key the futex is hashed on
223 * @pi_state: optional priority inheritance state
224 * @rt_waiter: rt_waiter storage for use with requeue_pi
225 * @requeue_pi_key: the requeue_pi target futex key
226 * @bitset: bitset for the optional bitmasked wakeup
227 *
228 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
229 * we can wake only the relevant ones (hashed queues may be shared).
230 *
231 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
232 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
233 * The order of wakeup is always to make the first condition true, then
234 * the second.
235 *
236 * PI futexes are typically woken before they are removed from the hash list via
237 * the rt_mutex code. See unqueue_me_pi().
238 */
239struct futex_q {
240 struct plist_node list;
241
242 struct task_struct *task;
243 spinlock_t *lock_ptr;
244 union futex_key key;
245 struct futex_pi_state *pi_state;
246 struct rt_mutex_waiter *rt_waiter;
247 union futex_key *requeue_pi_key;
248 u32 bitset;
249} __randomize_layout;
250
251static const struct futex_q futex_q_init = {
252 /* list gets initialized in queue_me()*/
253 .key = FUTEX_KEY_INIT,
254 .bitset = FUTEX_BITSET_MATCH_ANY
255};
256
257/*
258 * Hash buckets are shared by all the futex_keys that hash to the same
259 * location. Each key may have multiple futex_q structures, one for each task
260 * waiting on a futex.
261 */
262struct futex_hash_bucket {
263 atomic_t waiters;
264 spinlock_t lock;
265 struct plist_head chain;
266} ____cacheline_aligned_in_smp;
267
268/*
269 * The base of the bucket array and its size are always used together
270 * (after initialization only in hash_futex()), so ensure that they
271 * reside in the same cacheline.
272 */
273static struct {
274 struct futex_hash_bucket *queues;
275 unsigned long hashsize;
276} __futex_data __read_mostly __aligned(2*sizeof(long));
277#define futex_queues (__futex_data.queues)
278#define futex_hashsize (__futex_data.hashsize)
279
280
281/*
282 * Fault injections for futexes.
283 */
284#ifdef CONFIG_FAIL_FUTEX
285
286static struct {
287 struct fault_attr attr;
288
289 bool ignore_private;
290} fail_futex = {
291 .attr = FAULT_ATTR_INITIALIZER,
292 .ignore_private = false,
293};
294
295static int __init setup_fail_futex(char *str)
296{
297 return setup_fault_attr(&fail_futex.attr, str);
298}
299__setup("fail_futex=", setup_fail_futex);
300
301static bool should_fail_futex(bool fshared)
302{
303 if (fail_futex.ignore_private && !fshared)
304 return false;
305
306 return should_fail(&fail_futex.attr, 1);
307}
308
309#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
310
311static int __init fail_futex_debugfs(void)
312{
313 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
314 struct dentry *dir;
315
316 dir = fault_create_debugfs_attr("fail_futex", NULL,
317 &fail_futex.attr);
318 if (IS_ERR(dir))
319 return PTR_ERR(dir);
320
321 if (!debugfs_create_bool("ignore-private", mode, dir,
322 &fail_futex.ignore_private)) {
323 debugfs_remove_recursive(dir);
324 return -ENOMEM;
325 }
326
327 return 0;
328}
329
330late_initcall(fail_futex_debugfs);
331
332#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
333
334#else
335static inline bool should_fail_futex(bool fshared)
336{
337 return false;
338}
339#endif /* CONFIG_FAIL_FUTEX */
340
341static inline void futex_get_mm(union futex_key *key)
342{
343 mmgrab(key->private.mm);
344 /*
345 * Ensure futex_get_mm() implies a full barrier such that
346 * get_futex_key() implies a full barrier. This is relied upon
347 * as smp_mb(); (B), see the ordering comment above.
348 */
349 smp_mb__after_atomic();
350}
351
352/*
353 * Reflects a new waiter being added to the waitqueue.
354 */
355static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
356{
357#ifdef CONFIG_SMP
358 atomic_inc(&hb->waiters);
359 /*
360 * Full barrier (A), see the ordering comment above.
361 */
362 smp_mb__after_atomic();
363#endif
364}
365
366/*
367 * Reflects a waiter being removed from the waitqueue by wakeup
368 * paths.
369 */
370static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
371{
372#ifdef CONFIG_SMP
373 atomic_dec(&hb->waiters);
374#endif
375}
376
377static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
378{
379#ifdef CONFIG_SMP
380 return atomic_read(&hb->waiters);
381#else
382 return 1;
383#endif
384}
385
386/**
387 * hash_futex - Return the hash bucket in the global hash
388 * @key: Pointer to the futex key for which the hash is calculated
389 *
390 * We hash on the keys returned from get_futex_key (see below) and return the
391 * corresponding hash bucket in the global hash.
392 */
393static struct futex_hash_bucket *hash_futex(union futex_key *key)
394{
395 u32 hash = jhash2((u32*)&key->both.word,
396 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
397 key->both.offset);
398 return &futex_queues[hash & (futex_hashsize - 1)];
399}
400
401
402/**
403 * match_futex - Check whether two futex keys are equal
404 * @key1: Pointer to key1
405 * @key2: Pointer to key2
406 *
407 * Return 1 if two futex_keys are equal, 0 otherwise.
408 */
409static inline int match_futex(union futex_key *key1, union futex_key *key2)
410{
411 return (key1 && key2
412 && key1->both.word == key2->both.word
413 && key1->both.ptr == key2->both.ptr
414 && key1->both.offset == key2->both.offset);
415}
416
417/*
418 * Take a reference to the resource addressed by a key.
419 * Can be called while holding spinlocks.
420 *
421 */
422static void get_futex_key_refs(union futex_key *key)
423{
424 if (!key->both.ptr)
425 return;
426
427 /*
428 * On MMU less systems futexes are always "private" as there is no per
429 * process address space. We need the smp wmb nevertheless - yes,
430 * arch/blackfin has MMU less SMP ...
431 */
432 if (!IS_ENABLED(CONFIG_MMU)) {
433 smp_mb(); /* explicit smp_mb(); (B) */
434 return;
435 }
436
437 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
438 case FUT_OFF_INODE:
439 ihold(key->shared.inode); /* implies smp_mb(); (B) */
440 break;
441 case FUT_OFF_MMSHARED:
442 futex_get_mm(key); /* implies smp_mb(); (B) */
443 break;
444 default:
445 /*
446 * Private futexes do not hold reference on an inode or
447 * mm, therefore the only purpose of calling get_futex_key_refs
448 * is because we need the barrier for the lockless waiter check.
449 */
450 smp_mb(); /* explicit smp_mb(); (B) */
451 }
452}
453
454/*
455 * Drop a reference to the resource addressed by a key.
456 * The hash bucket spinlock must not be held. This is
457 * a no-op for private futexes, see comment in the get
458 * counterpart.
459 */
460static void drop_futex_key_refs(union futex_key *key)
461{
462 if (!key->both.ptr) {
463 /* If we're here then we tried to put a key we failed to get */
464 WARN_ON_ONCE(1);
465 return;
466 }
467
468 if (!IS_ENABLED(CONFIG_MMU))
469 return;
470
471 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
472 case FUT_OFF_INODE:
473 iput(key->shared.inode);
474 break;
475 case FUT_OFF_MMSHARED:
476 mmdrop(key->private.mm);
477 break;
478 }
479}
480
481/**
482 * get_futex_key() - Get parameters which are the keys for a futex
483 * @uaddr: virtual address of the futex
484 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485 * @key: address where result is stored.
486 * @rw: mapping needs to be read/write (values: VERIFY_READ,
487 * VERIFY_WRITE)
488 *
489 * Return: a negative error code or 0
490 *
491 * The key words are stored in @key on success.
492 *
493 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
494 * offset_within_page). For private mappings, it's (uaddr, current->mm).
495 * We can usually work out the index without swapping in the page.
496 *
497 * lock_page() might sleep, the caller should not hold a spinlock.
498 */
499static int
500get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
501{
502 unsigned long address = (unsigned long)uaddr;
503 struct mm_struct *mm = current->mm;
504 struct page *page, *tail;
505 struct address_space *mapping;
506 int err, ro = 0;
507
508 /*
509 * The futex address must be "naturally" aligned.
510 */
511 key->both.offset = address % PAGE_SIZE;
512 if (unlikely((address % sizeof(u32)) != 0))
513 return -EINVAL;
514 address -= key->both.offset;
515
516 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
517 return -EFAULT;
518
519 if (unlikely(should_fail_futex(fshared)))
520 return -EFAULT;
521
522 /*
523 * PROCESS_PRIVATE futexes are fast.
524 * As the mm cannot disappear under us and the 'key' only needs
525 * virtual address, we dont even have to find the underlying vma.
526 * Note : We do have to check 'uaddr' is a valid user address,
527 * but access_ok() should be faster than find_vma()
528 */
529 if (!fshared) {
530 key->private.mm = mm;
531 key->private.address = address;
532 get_futex_key_refs(key); /* implies smp_mb(); (B) */
533 return 0;
534 }
535
536again:
537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared)))
539 return -EFAULT;
540
541 err = get_user_pages_fast(address, 1, 1, &page);
542 /*
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
545 */
546 if (err == -EFAULT && rw == VERIFY_READ) {
547 err = get_user_pages_fast(address, 1, 0, &page);
548 ro = 1;
549 }
550 if (err < 0)
551 return err;
552 else
553 err = 0;
554
555 /*
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
560 *
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
565 *
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
572 */
573 tail = page;
574 page = compound_head(page);
575 mapping = READ_ONCE(page->mapping);
576
577 /*
578 * If page->mapping is NULL, then it cannot be a PageAnon
579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
587 *
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
590 * an unlikely race, but we do need to retry for page->mapping.
591 */
592 if (unlikely(!mapping)) {
593 int shmem_swizzled;
594
595 /*
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
599 */
600 lock_page(page);
601 shmem_swizzled = PageSwapCache(page) || page->mapping;
602 unlock_page(page);
603 put_page(page);
604
605 if (shmem_swizzled)
606 goto again;
607
608 return -EFAULT;
609 }
610
611 /*
612 * Private mappings are handled in a simple way.
613 *
614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
616 *
617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
619 * the object not the particular process.
620 */
621 if (PageAnon(page)) {
622 /*
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
625 */
626 if (unlikely(should_fail_futex(fshared)) || ro) {
627 err = -EFAULT;
628 goto out;
629 }
630
631 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
632 key->private.mm = mm;
633 key->private.address = address;
634
635 get_futex_key_refs(key); /* implies smp_mb(); (B) */
636
637 } else {
638 struct inode *inode;
639
640 /*
641 * The associated futex object in this case is the inode and
642 * the page->mapping must be traversed. Ordinarily this should
643 * be stabilised under page lock but it's not strictly
644 * necessary in this case as we just want to pin the inode, not
645 * update the radix tree or anything like that.
646 *
647 * The RCU read lock is taken as the inode is finally freed
648 * under RCU. If the mapping still matches expectations then the
649 * mapping->host can be safely accessed as being a valid inode.
650 */
651 rcu_read_lock();
652
653 if (READ_ONCE(page->mapping) != mapping) {
654 rcu_read_unlock();
655 put_page(page);
656
657 goto again;
658 }
659
660 inode = READ_ONCE(mapping->host);
661 if (!inode) {
662 rcu_read_unlock();
663 put_page(page);
664
665 goto again;
666 }
667
668 /*
669 * Take a reference unless it is about to be freed. Previously
670 * this reference was taken by ihold under the page lock
671 * pinning the inode in place so i_lock was unnecessary. The
672 * only way for this check to fail is if the inode was
673 * truncated in parallel which is almost certainly an
674 * application bug. In such a case, just retry.
675 *
676 * We are not calling into get_futex_key_refs() in file-backed
677 * cases, therefore a successful atomic_inc return below will
678 * guarantee that get_futex_key() will still imply smp_mb(); (B).
679 */
680 if (!atomic_inc_not_zero(&inode->i_count)) {
681 rcu_read_unlock();
682 put_page(page);
683
684 goto again;
685 }
686
687 /* Should be impossible but lets be paranoid for now */
688 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
689 err = -EFAULT;
690 rcu_read_unlock();
691 iput(inode);
692
693 goto out;
694 }
695
696 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
697 key->shared.inode = inode;
698 key->shared.pgoff = basepage_index(tail);
699 rcu_read_unlock();
700 }
701
702out:
703 put_page(page);
704 return err;
705}
706
707static inline void put_futex_key(union futex_key *key)
708{
709 drop_futex_key_refs(key);
710}
711
712/**
713 * fault_in_user_writeable() - Fault in user address and verify RW access
714 * @uaddr: pointer to faulting user space address
715 *
716 * Slow path to fixup the fault we just took in the atomic write
717 * access to @uaddr.
718 *
719 * We have no generic implementation of a non-destructive write to the
720 * user address. We know that we faulted in the atomic pagefault
721 * disabled section so we can as well avoid the #PF overhead by
722 * calling get_user_pages() right away.
723 */
724static int fault_in_user_writeable(u32 __user *uaddr)
725{
726 struct mm_struct *mm = current->mm;
727 int ret;
728
729 down_read(&mm->mmap_sem);
730 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
731 FAULT_FLAG_WRITE, NULL);
732 up_read(&mm->mmap_sem);
733
734 return ret < 0 ? ret : 0;
735}
736
737/**
738 * futex_top_waiter() - Return the highest priority waiter on a futex
739 * @hb: the hash bucket the futex_q's reside in
740 * @key: the futex key (to distinguish it from other futex futex_q's)
741 *
742 * Must be called with the hb lock held.
743 */
744static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
745 union futex_key *key)
746{
747 struct futex_q *this;
748
749 plist_for_each_entry(this, &hb->chain, list) {
750 if (match_futex(&this->key, key))
751 return this;
752 }
753 return NULL;
754}
755
756static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
757 u32 uval, u32 newval)
758{
759 int ret;
760
761 pagefault_disable();
762 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
763 pagefault_enable();
764
765 return ret;
766}
767
768static int get_futex_value_locked(u32 *dest, u32 __user *from)
769{
770 int ret;
771
772 pagefault_disable();
773 ret = __get_user(*dest, from);
774 pagefault_enable();
775
776 return ret ? -EFAULT : 0;
777}
778
779
780/*
781 * PI code:
782 */
783static int refill_pi_state_cache(void)
784{
785 struct futex_pi_state *pi_state;
786
787 if (likely(current->pi_state_cache))
788 return 0;
789
790 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
791
792 if (!pi_state)
793 return -ENOMEM;
794
795 INIT_LIST_HEAD(&pi_state->list);
796 /* pi_mutex gets initialized later */
797 pi_state->owner = NULL;
798 atomic_set(&pi_state->refcount, 1);
799 pi_state->key = FUTEX_KEY_INIT;
800
801 current->pi_state_cache = pi_state;
802
803 return 0;
804}
805
806static struct futex_pi_state *alloc_pi_state(void)
807{
808 struct futex_pi_state *pi_state = current->pi_state_cache;
809
810 WARN_ON(!pi_state);
811 current->pi_state_cache = NULL;
812
813 return pi_state;
814}
815
816static void get_pi_state(struct futex_pi_state *pi_state)
817{
818 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
819}
820
821/*
822 * Drops a reference to the pi_state object and frees or caches it
823 * when the last reference is gone.
824 */
825static void put_pi_state(struct futex_pi_state *pi_state)
826{
827 if (!pi_state)
828 return;
829
830 if (!atomic_dec_and_test(&pi_state->refcount))
831 return;
832
833 /*
834 * If pi_state->owner is NULL, the owner is most probably dying
835 * and has cleaned up the pi_state already
836 */
837 if (pi_state->owner) {
838 struct task_struct *owner;
839
840 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
841 owner = pi_state->owner;
842 if (owner) {
843 raw_spin_lock(&owner->pi_lock);
844 list_del_init(&pi_state->list);
845 raw_spin_unlock(&owner->pi_lock);
846 }
847 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
848 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
849 }
850
851 if (current->pi_state_cache) {
852 kfree(pi_state);
853 } else {
854 /*
855 * pi_state->list is already empty.
856 * clear pi_state->owner.
857 * refcount is at 0 - put it back to 1.
858 */
859 pi_state->owner = NULL;
860 atomic_set(&pi_state->refcount, 1);
861 current->pi_state_cache = pi_state;
862 }
863}
864
865#ifdef CONFIG_FUTEX_PI
866
867/*
868 * This task is holding PI mutexes at exit time => bad.
869 * Kernel cleans up PI-state, but userspace is likely hosed.
870 * (Robust-futex cleanup is separate and might save the day for userspace.)
871 */
872void exit_pi_state_list(struct task_struct *curr)
873{
874 struct list_head *next, *head = &curr->pi_state_list;
875 struct futex_pi_state *pi_state;
876 struct futex_hash_bucket *hb;
877 union futex_key key = FUTEX_KEY_INIT;
878
879 if (!futex_cmpxchg_enabled)
880 return;
881 /*
882 * We are a ZOMBIE and nobody can enqueue itself on
883 * pi_state_list anymore, but we have to be careful
884 * versus waiters unqueueing themselves:
885 */
886 raw_spin_lock_irq(&curr->pi_lock);
887 while (!list_empty(head)) {
888 next = head->next;
889 pi_state = list_entry(next, struct futex_pi_state, list);
890 key = pi_state->key;
891 hb = hash_futex(&key);
892
893 /*
894 * We can race against put_pi_state() removing itself from the
895 * list (a waiter going away). put_pi_state() will first
896 * decrement the reference count and then modify the list, so
897 * its possible to see the list entry but fail this reference
898 * acquire.
899 *
900 * In that case; drop the locks to let put_pi_state() make
901 * progress and retry the loop.
902 */
903 if (!atomic_inc_not_zero(&pi_state->refcount)) {
904 raw_spin_unlock_irq(&curr->pi_lock);
905 cpu_relax();
906 raw_spin_lock_irq(&curr->pi_lock);
907 continue;
908 }
909 raw_spin_unlock_irq(&curr->pi_lock);
910
911 spin_lock(&hb->lock);
912 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
913 raw_spin_lock(&curr->pi_lock);
914 /*
915 * We dropped the pi-lock, so re-check whether this
916 * task still owns the PI-state:
917 */
918 if (head->next != next) {
919 /* retain curr->pi_lock for the loop invariant */
920 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
921 spin_unlock(&hb->lock);
922 put_pi_state(pi_state);
923 continue;
924 }
925
926 WARN_ON(pi_state->owner != curr);
927 WARN_ON(list_empty(&pi_state->list));
928 list_del_init(&pi_state->list);
929 pi_state->owner = NULL;
930
931 raw_spin_unlock(&curr->pi_lock);
932 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
933 spin_unlock(&hb->lock);
934
935 rt_mutex_futex_unlock(&pi_state->pi_mutex);
936 put_pi_state(pi_state);
937
938 raw_spin_lock_irq(&curr->pi_lock);
939 }
940 raw_spin_unlock_irq(&curr->pi_lock);
941}
942
943#endif
944
945/*
946 * We need to check the following states:
947 *
948 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
949 *
950 * [1] NULL | --- | --- | 0 | 0/1 | Valid
951 * [2] NULL | --- | --- | >0 | 0/1 | Valid
952 *
953 * [3] Found | NULL | -- | Any | 0/1 | Invalid
954 *
955 * [4] Found | Found | NULL | 0 | 1 | Valid
956 * [5] Found | Found | NULL | >0 | 1 | Invalid
957 *
958 * [6] Found | Found | task | 0 | 1 | Valid
959 *
960 * [7] Found | Found | NULL | Any | 0 | Invalid
961 *
962 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
963 * [9] Found | Found | task | 0 | 0 | Invalid
964 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
965 *
966 * [1] Indicates that the kernel can acquire the futex atomically. We
967 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
968 *
969 * [2] Valid, if TID does not belong to a kernel thread. If no matching
970 * thread is found then it indicates that the owner TID has died.
971 *
972 * [3] Invalid. The waiter is queued on a non PI futex
973 *
974 * [4] Valid state after exit_robust_list(), which sets the user space
975 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
976 *
977 * [5] The user space value got manipulated between exit_robust_list()
978 * and exit_pi_state_list()
979 *
980 * [6] Valid state after exit_pi_state_list() which sets the new owner in
981 * the pi_state but cannot access the user space value.
982 *
983 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
984 *
985 * [8] Owner and user space value match
986 *
987 * [9] There is no transient state which sets the user space TID to 0
988 * except exit_robust_list(), but this is indicated by the
989 * FUTEX_OWNER_DIED bit. See [4]
990 *
991 * [10] There is no transient state which leaves owner and user space
992 * TID out of sync.
993 *
994 *
995 * Serialization and lifetime rules:
996 *
997 * hb->lock:
998 *
999 * hb -> futex_q, relation
1000 * futex_q -> pi_state, relation
1001 *
1002 * (cannot be raw because hb can contain arbitrary amount
1003 * of futex_q's)
1004 *
1005 * pi_mutex->wait_lock:
1006 *
1007 * {uval, pi_state}
1008 *
1009 * (and pi_mutex 'obviously')
1010 *
1011 * p->pi_lock:
1012 *
1013 * p->pi_state_list -> pi_state->list, relation
1014 *
1015 * pi_state->refcount:
1016 *
1017 * pi_state lifetime
1018 *
1019 *
1020 * Lock order:
1021 *
1022 * hb->lock
1023 * pi_mutex->wait_lock
1024 * p->pi_lock
1025 *
1026 */
1027
1028/*
1029 * Validate that the existing waiter has a pi_state and sanity check
1030 * the pi_state against the user space value. If correct, attach to
1031 * it.
1032 */
1033static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1034 struct futex_pi_state *pi_state,
1035 struct futex_pi_state **ps)
1036{
1037 pid_t pid = uval & FUTEX_TID_MASK;
1038 u32 uval2;
1039 int ret;
1040
1041 /*
1042 * Userspace might have messed up non-PI and PI futexes [3]
1043 */
1044 if (unlikely(!pi_state))
1045 return -EINVAL;
1046
1047 /*
1048 * We get here with hb->lock held, and having found a
1049 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1050 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1051 * which in turn means that futex_lock_pi() still has a reference on
1052 * our pi_state.
1053 *
1054 * The waiter holding a reference on @pi_state also protects against
1055 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1056 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1057 * free pi_state before we can take a reference ourselves.
1058 */
1059 WARN_ON(!atomic_read(&pi_state->refcount));
1060
1061 /*
1062 * Now that we have a pi_state, we can acquire wait_lock
1063 * and do the state validation.
1064 */
1065 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1066
1067 /*
1068 * Since {uval, pi_state} is serialized by wait_lock, and our current
1069 * uval was read without holding it, it can have changed. Verify it
1070 * still is what we expect it to be, otherwise retry the entire
1071 * operation.
1072 */
1073 if (get_futex_value_locked(&uval2, uaddr))
1074 goto out_efault;
1075
1076 if (uval != uval2)
1077 goto out_eagain;
1078
1079 /*
1080 * Handle the owner died case:
1081 */
1082 if (uval & FUTEX_OWNER_DIED) {
1083 /*
1084 * exit_pi_state_list sets owner to NULL and wakes the
1085 * topmost waiter. The task which acquires the
1086 * pi_state->rt_mutex will fixup owner.
1087 */
1088 if (!pi_state->owner) {
1089 /*
1090 * No pi state owner, but the user space TID
1091 * is not 0. Inconsistent state. [5]
1092 */
1093 if (pid)
1094 goto out_einval;
1095 /*
1096 * Take a ref on the state and return success. [4]
1097 */
1098 goto out_attach;
1099 }
1100
1101 /*
1102 * If TID is 0, then either the dying owner has not
1103 * yet executed exit_pi_state_list() or some waiter
1104 * acquired the rtmutex in the pi state, but did not
1105 * yet fixup the TID in user space.
1106 *
1107 * Take a ref on the state and return success. [6]
1108 */
1109 if (!pid)
1110 goto out_attach;
1111 } else {
1112 /*
1113 * If the owner died bit is not set, then the pi_state
1114 * must have an owner. [7]
1115 */
1116 if (!pi_state->owner)
1117 goto out_einval;
1118 }
1119
1120 /*
1121 * Bail out if user space manipulated the futex value. If pi
1122 * state exists then the owner TID must be the same as the
1123 * user space TID. [9/10]
1124 */
1125 if (pid != task_pid_vnr(pi_state->owner))
1126 goto out_einval;
1127
1128out_attach:
1129 get_pi_state(pi_state);
1130 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1131 *ps = pi_state;
1132 return 0;
1133
1134out_einval:
1135 ret = -EINVAL;
1136 goto out_error;
1137
1138out_eagain:
1139 ret = -EAGAIN;
1140 goto out_error;
1141
1142out_efault:
1143 ret = -EFAULT;
1144 goto out_error;
1145
1146out_error:
1147 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1148 return ret;
1149}
1150
1151/*
1152 * Lookup the task for the TID provided from user space and attach to
1153 * it after doing proper sanity checks.
1154 */
1155static int attach_to_pi_owner(u32 uval, union futex_key *key,
1156 struct futex_pi_state **ps)
1157{
1158 pid_t pid = uval & FUTEX_TID_MASK;
1159 struct futex_pi_state *pi_state;
1160 struct task_struct *p;
1161
1162 /*
1163 * We are the first waiter - try to look up the real owner and attach
1164 * the new pi_state to it, but bail out when TID = 0 [1]
1165 */
1166 if (!pid)
1167 return -ESRCH;
1168 p = find_get_task_by_vpid(pid);
1169 if (!p)
1170 return -ESRCH;
1171
1172 if (unlikely(p->flags & PF_KTHREAD)) {
1173 put_task_struct(p);
1174 return -EPERM;
1175 }
1176
1177 /*
1178 * We need to look at the task state flags to figure out,
1179 * whether the task is exiting. To protect against the do_exit
1180 * change of the task flags, we do this protected by
1181 * p->pi_lock:
1182 */
1183 raw_spin_lock_irq(&p->pi_lock);
1184 if (unlikely(p->flags & PF_EXITING)) {
1185 /*
1186 * The task is on the way out. When PF_EXITPIDONE is
1187 * set, we know that the task has finished the
1188 * cleanup:
1189 */
1190 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1191
1192 raw_spin_unlock_irq(&p->pi_lock);
1193 put_task_struct(p);
1194 return ret;
1195 }
1196
1197 /*
1198 * No existing pi state. First waiter. [2]
1199 *
1200 * This creates pi_state, we have hb->lock held, this means nothing can
1201 * observe this state, wait_lock is irrelevant.
1202 */
1203 pi_state = alloc_pi_state();
1204
1205 /*
1206 * Initialize the pi_mutex in locked state and make @p
1207 * the owner of it:
1208 */
1209 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1210
1211 /* Store the key for possible exit cleanups: */
1212 pi_state->key = *key;
1213
1214 WARN_ON(!list_empty(&pi_state->list));
1215 list_add(&pi_state->list, &p->pi_state_list);
1216 /*
1217 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1218 * because there is no concurrency as the object is not published yet.
1219 */
1220 pi_state->owner = p;
1221 raw_spin_unlock_irq(&p->pi_lock);
1222
1223 put_task_struct(p);
1224
1225 *ps = pi_state;
1226
1227 return 0;
1228}
1229
1230static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1231 struct futex_hash_bucket *hb,
1232 union futex_key *key, struct futex_pi_state **ps)
1233{
1234 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1235
1236 /*
1237 * If there is a waiter on that futex, validate it and
1238 * attach to the pi_state when the validation succeeds.
1239 */
1240 if (top_waiter)
1241 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1242
1243 /*
1244 * We are the first waiter - try to look up the owner based on
1245 * @uval and attach to it.
1246 */
1247 return attach_to_pi_owner(uval, key, ps);
1248}
1249
1250static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1251{
1252 u32 uninitialized_var(curval);
1253
1254 if (unlikely(should_fail_futex(true)))
1255 return -EFAULT;
1256
1257 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1258 return -EFAULT;
1259
1260 /* If user space value changed, let the caller retry */
1261 return curval != uval ? -EAGAIN : 0;
1262}
1263
1264/**
1265 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1266 * @uaddr: the pi futex user address
1267 * @hb: the pi futex hash bucket
1268 * @key: the futex key associated with uaddr and hb
1269 * @ps: the pi_state pointer where we store the result of the
1270 * lookup
1271 * @task: the task to perform the atomic lock work for. This will
1272 * be "current" except in the case of requeue pi.
1273 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1274 *
1275 * Return:
1276 * - 0 - ready to wait;
1277 * - 1 - acquired the lock;
1278 * - <0 - error
1279 *
1280 * The hb->lock and futex_key refs shall be held by the caller.
1281 */
1282static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1283 union futex_key *key,
1284 struct futex_pi_state **ps,
1285 struct task_struct *task, int set_waiters)
1286{
1287 u32 uval, newval, vpid = task_pid_vnr(task);
1288 struct futex_q *top_waiter;
1289 int ret;
1290
1291 /*
1292 * Read the user space value first so we can validate a few
1293 * things before proceeding further.
1294 */
1295 if (get_futex_value_locked(&uval, uaddr))
1296 return -EFAULT;
1297
1298 if (unlikely(should_fail_futex(true)))
1299 return -EFAULT;
1300
1301 /*
1302 * Detect deadlocks.
1303 */
1304 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1305 return -EDEADLK;
1306
1307 if ((unlikely(should_fail_futex(true))))
1308 return -EDEADLK;
1309
1310 /*
1311 * Lookup existing state first. If it exists, try to attach to
1312 * its pi_state.
1313 */
1314 top_waiter = futex_top_waiter(hb, key);
1315 if (top_waiter)
1316 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1317
1318 /*
1319 * No waiter and user TID is 0. We are here because the
1320 * waiters or the owner died bit is set or called from
1321 * requeue_cmp_pi or for whatever reason something took the
1322 * syscall.
1323 */
1324 if (!(uval & FUTEX_TID_MASK)) {
1325 /*
1326 * We take over the futex. No other waiters and the user space
1327 * TID is 0. We preserve the owner died bit.
1328 */
1329 newval = uval & FUTEX_OWNER_DIED;
1330 newval |= vpid;
1331
1332 /* The futex requeue_pi code can enforce the waiters bit */
1333 if (set_waiters)
1334 newval |= FUTEX_WAITERS;
1335
1336 ret = lock_pi_update_atomic(uaddr, uval, newval);
1337 /* If the take over worked, return 1 */
1338 return ret < 0 ? ret : 1;
1339 }
1340
1341 /*
1342 * First waiter. Set the waiters bit before attaching ourself to
1343 * the owner. If owner tries to unlock, it will be forced into
1344 * the kernel and blocked on hb->lock.
1345 */
1346 newval = uval | FUTEX_WAITERS;
1347 ret = lock_pi_update_atomic(uaddr, uval, newval);
1348 if (ret)
1349 return ret;
1350 /*
1351 * If the update of the user space value succeeded, we try to
1352 * attach to the owner. If that fails, no harm done, we only
1353 * set the FUTEX_WAITERS bit in the user space variable.
1354 */
1355 return attach_to_pi_owner(uval, key, ps);
1356}
1357
1358/**
1359 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1360 * @q: The futex_q to unqueue
1361 *
1362 * The q->lock_ptr must not be NULL and must be held by the caller.
1363 */
1364static void __unqueue_futex(struct futex_q *q)
1365{
1366 struct futex_hash_bucket *hb;
1367
1368 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1369 || WARN_ON(plist_node_empty(&q->list)))
1370 return;
1371
1372 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1373 plist_del(&q->list, &hb->chain);
1374 hb_waiters_dec(hb);
1375}
1376
1377/*
1378 * The hash bucket lock must be held when this is called.
1379 * Afterwards, the futex_q must not be accessed. Callers
1380 * must ensure to later call wake_up_q() for the actual
1381 * wakeups to occur.
1382 */
1383static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1384{
1385 struct task_struct *p = q->task;
1386
1387 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1388 return;
1389
1390 /*
1391 * Queue the task for later wakeup for after we've released
1392 * the hb->lock. wake_q_add() grabs reference to p.
1393 */
1394 wake_q_add(wake_q, p);
1395 __unqueue_futex(q);
1396 /*
1397 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1398 * is written, without taking any locks. This is possible in the event
1399 * of a spurious wakeup, for example. A memory barrier is required here
1400 * to prevent the following store to lock_ptr from getting ahead of the
1401 * plist_del in __unqueue_futex().
1402 */
1403 smp_store_release(&q->lock_ptr, NULL);
1404}
1405
1406/*
1407 * Caller must hold a reference on @pi_state.
1408 */
1409static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1410{
1411 u32 uninitialized_var(curval), newval;
1412 struct task_struct *new_owner;
1413 bool postunlock = false;
1414 DEFINE_WAKE_Q(wake_q);
1415 int ret = 0;
1416
1417 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1418 if (WARN_ON_ONCE(!new_owner)) {
1419 /*
1420 * As per the comment in futex_unlock_pi() this should not happen.
1421 *
1422 * When this happens, give up our locks and try again, giving
1423 * the futex_lock_pi() instance time to complete, either by
1424 * waiting on the rtmutex or removing itself from the futex
1425 * queue.
1426 */
1427 ret = -EAGAIN;
1428 goto out_unlock;
1429 }
1430
1431 /*
1432 * We pass it to the next owner. The WAITERS bit is always kept
1433 * enabled while there is PI state around. We cleanup the owner
1434 * died bit, because we are the owner.
1435 */
1436 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1437
1438 if (unlikely(should_fail_futex(true)))
1439 ret = -EFAULT;
1440
1441 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1442 ret = -EFAULT;
1443
1444 } else if (curval != uval) {
1445 /*
1446 * If a unconditional UNLOCK_PI operation (user space did not
1447 * try the TID->0 transition) raced with a waiter setting the
1448 * FUTEX_WAITERS flag between get_user() and locking the hash
1449 * bucket lock, retry the operation.
1450 */
1451 if ((FUTEX_TID_MASK & curval) == uval)
1452 ret = -EAGAIN;
1453 else
1454 ret = -EINVAL;
1455 }
1456
1457 if (ret)
1458 goto out_unlock;
1459
1460 /*
1461 * This is a point of no return; once we modify the uval there is no
1462 * going back and subsequent operations must not fail.
1463 */
1464
1465 raw_spin_lock(&pi_state->owner->pi_lock);
1466 WARN_ON(list_empty(&pi_state->list));
1467 list_del_init(&pi_state->list);
1468 raw_spin_unlock(&pi_state->owner->pi_lock);
1469
1470 raw_spin_lock(&new_owner->pi_lock);
1471 WARN_ON(!list_empty(&pi_state->list));
1472 list_add(&pi_state->list, &new_owner->pi_state_list);
1473 pi_state->owner = new_owner;
1474 raw_spin_unlock(&new_owner->pi_lock);
1475
1476 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1477
1478out_unlock:
1479 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1480
1481 if (postunlock)
1482 rt_mutex_postunlock(&wake_q);
1483
1484 return ret;
1485}
1486
1487/*
1488 * Express the locking dependencies for lockdep:
1489 */
1490static inline void
1491double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1492{
1493 if (hb1 <= hb2) {
1494 spin_lock(&hb1->lock);
1495 if (hb1 < hb2)
1496 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1497 } else { /* hb1 > hb2 */
1498 spin_lock(&hb2->lock);
1499 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1500 }
1501}
1502
1503static inline void
1504double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1505{
1506 spin_unlock(&hb1->lock);
1507 if (hb1 != hb2)
1508 spin_unlock(&hb2->lock);
1509}
1510
1511/*
1512 * Wake up waiters matching bitset queued on this futex (uaddr).
1513 */
1514static int
1515futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1516{
1517 struct futex_hash_bucket *hb;
1518 struct futex_q *this, *next;
1519 union futex_key key = FUTEX_KEY_INIT;
1520 int ret;
1521 DEFINE_WAKE_Q(wake_q);
1522
1523 if (!bitset)
1524 return -EINVAL;
1525
1526 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1527 if (unlikely(ret != 0))
1528 goto out;
1529
1530 hb = hash_futex(&key);
1531
1532 /* Make sure we really have tasks to wakeup */
1533 if (!hb_waiters_pending(hb))
1534 goto out_put_key;
1535
1536 spin_lock(&hb->lock);
1537
1538 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1539 if (match_futex (&this->key, &key)) {
1540 if (this->pi_state || this->rt_waiter) {
1541 ret = -EINVAL;
1542 break;
1543 }
1544
1545 /* Check if one of the bits is set in both bitsets */
1546 if (!(this->bitset & bitset))
1547 continue;
1548
1549 mark_wake_futex(&wake_q, this);
1550 if (++ret >= nr_wake)
1551 break;
1552 }
1553 }
1554
1555 spin_unlock(&hb->lock);
1556 wake_up_q(&wake_q);
1557out_put_key:
1558 put_futex_key(&key);
1559out:
1560 return ret;
1561}
1562
1563static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1564{
1565 unsigned int op = (encoded_op & 0x70000000) >> 28;
1566 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1567 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1568 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1569 int oldval, ret;
1570
1571 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1572 if (oparg < 0 || oparg > 31) {
1573 char comm[sizeof(current->comm)];
1574 /*
1575 * kill this print and return -EINVAL when userspace
1576 * is sane again
1577 */
1578 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1579 get_task_comm(comm, current), oparg);
1580 oparg &= 31;
1581 }
1582 oparg = 1 << oparg;
1583 }
1584
1585 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1586 return -EFAULT;
1587
1588 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1589 if (ret)
1590 return ret;
1591
1592 switch (cmp) {
1593 case FUTEX_OP_CMP_EQ:
1594 return oldval == cmparg;
1595 case FUTEX_OP_CMP_NE:
1596 return oldval != cmparg;
1597 case FUTEX_OP_CMP_LT:
1598 return oldval < cmparg;
1599 case FUTEX_OP_CMP_GE:
1600 return oldval >= cmparg;
1601 case FUTEX_OP_CMP_LE:
1602 return oldval <= cmparg;
1603 case FUTEX_OP_CMP_GT:
1604 return oldval > cmparg;
1605 default:
1606 return -ENOSYS;
1607 }
1608}
1609
1610/*
1611 * Wake up all waiters hashed on the physical page that is mapped
1612 * to this virtual address:
1613 */
1614static int
1615futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1616 int nr_wake, int nr_wake2, int op)
1617{
1618 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1619 struct futex_hash_bucket *hb1, *hb2;
1620 struct futex_q *this, *next;
1621 int ret, op_ret;
1622 DEFINE_WAKE_Q(wake_q);
1623
1624retry:
1625 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1626 if (unlikely(ret != 0))
1627 goto out;
1628 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1629 if (unlikely(ret != 0))
1630 goto out_put_key1;
1631
1632 hb1 = hash_futex(&key1);
1633 hb2 = hash_futex(&key2);
1634
1635retry_private:
1636 double_lock_hb(hb1, hb2);
1637 op_ret = futex_atomic_op_inuser(op, uaddr2);
1638 if (unlikely(op_ret < 0)) {
1639
1640 double_unlock_hb(hb1, hb2);
1641
1642#ifndef CONFIG_MMU
1643 /*
1644 * we don't get EFAULT from MMU faults if we don't have an MMU,
1645 * but we might get them from range checking
1646 */
1647 ret = op_ret;
1648 goto out_put_keys;
1649#endif
1650
1651 if (unlikely(op_ret != -EFAULT)) {
1652 ret = op_ret;
1653 goto out_put_keys;
1654 }
1655
1656 ret = fault_in_user_writeable(uaddr2);
1657 if (ret)
1658 goto out_put_keys;
1659
1660 if (!(flags & FLAGS_SHARED))
1661 goto retry_private;
1662
1663 put_futex_key(&key2);
1664 put_futex_key(&key1);
1665 goto retry;
1666 }
1667
1668 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1669 if (match_futex (&this->key, &key1)) {
1670 if (this->pi_state || this->rt_waiter) {
1671 ret = -EINVAL;
1672 goto out_unlock;
1673 }
1674 mark_wake_futex(&wake_q, this);
1675 if (++ret >= nr_wake)
1676 break;
1677 }
1678 }
1679
1680 if (op_ret > 0) {
1681 op_ret = 0;
1682 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1683 if (match_futex (&this->key, &key2)) {
1684 if (this->pi_state || this->rt_waiter) {
1685 ret = -EINVAL;
1686 goto out_unlock;
1687 }
1688 mark_wake_futex(&wake_q, this);
1689 if (++op_ret >= nr_wake2)
1690 break;
1691 }
1692 }
1693 ret += op_ret;
1694 }
1695
1696out_unlock:
1697 double_unlock_hb(hb1, hb2);
1698 wake_up_q(&wake_q);
1699out_put_keys:
1700 put_futex_key(&key2);
1701out_put_key1:
1702 put_futex_key(&key1);
1703out:
1704 return ret;
1705}
1706
1707/**
1708 * requeue_futex() - Requeue a futex_q from one hb to another
1709 * @q: the futex_q to requeue
1710 * @hb1: the source hash_bucket
1711 * @hb2: the target hash_bucket
1712 * @key2: the new key for the requeued futex_q
1713 */
1714static inline
1715void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1716 struct futex_hash_bucket *hb2, union futex_key *key2)
1717{
1718
1719 /*
1720 * If key1 and key2 hash to the same bucket, no need to
1721 * requeue.
1722 */
1723 if (likely(&hb1->chain != &hb2->chain)) {
1724 plist_del(&q->list, &hb1->chain);
1725 hb_waiters_dec(hb1);
1726 hb_waiters_inc(hb2);
1727 plist_add(&q->list, &hb2->chain);
1728 q->lock_ptr = &hb2->lock;
1729 }
1730 get_futex_key_refs(key2);
1731 q->key = *key2;
1732}
1733
1734/**
1735 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1736 * @q: the futex_q
1737 * @key: the key of the requeue target futex
1738 * @hb: the hash_bucket of the requeue target futex
1739 *
1740 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1741 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1742 * to the requeue target futex so the waiter can detect the wakeup on the right
1743 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1744 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1745 * to protect access to the pi_state to fixup the owner later. Must be called
1746 * with both q->lock_ptr and hb->lock held.
1747 */
1748static inline
1749void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1750 struct futex_hash_bucket *hb)
1751{
1752 get_futex_key_refs(key);
1753 q->key = *key;
1754
1755 __unqueue_futex(q);
1756
1757 WARN_ON(!q->rt_waiter);
1758 q->rt_waiter = NULL;
1759
1760 q->lock_ptr = &hb->lock;
1761
1762 wake_up_state(q->task, TASK_NORMAL);
1763}
1764
1765/**
1766 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1767 * @pifutex: the user address of the to futex
1768 * @hb1: the from futex hash bucket, must be locked by the caller
1769 * @hb2: the to futex hash bucket, must be locked by the caller
1770 * @key1: the from futex key
1771 * @key2: the to futex key
1772 * @ps: address to store the pi_state pointer
1773 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1774 *
1775 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1776 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1777 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1778 * hb1 and hb2 must be held by the caller.
1779 *
1780 * Return:
1781 * - 0 - failed to acquire the lock atomically;
1782 * - >0 - acquired the lock, return value is vpid of the top_waiter
1783 * - <0 - error
1784 */
1785static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1786 struct futex_hash_bucket *hb1,
1787 struct futex_hash_bucket *hb2,
1788 union futex_key *key1, union futex_key *key2,
1789 struct futex_pi_state **ps, int set_waiters)
1790{
1791 struct futex_q *top_waiter = NULL;
1792 u32 curval;
1793 int ret, vpid;
1794
1795 if (get_futex_value_locked(&curval, pifutex))
1796 return -EFAULT;
1797
1798 if (unlikely(should_fail_futex(true)))
1799 return -EFAULT;
1800
1801 /*
1802 * Find the top_waiter and determine if there are additional waiters.
1803 * If the caller intends to requeue more than 1 waiter to pifutex,
1804 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1805 * as we have means to handle the possible fault. If not, don't set
1806 * the bit unecessarily as it will force the subsequent unlock to enter
1807 * the kernel.
1808 */
1809 top_waiter = futex_top_waiter(hb1, key1);
1810
1811 /* There are no waiters, nothing for us to do. */
1812 if (!top_waiter)
1813 return 0;
1814
1815 /* Ensure we requeue to the expected futex. */
1816 if (!match_futex(top_waiter->requeue_pi_key, key2))
1817 return -EINVAL;
1818
1819 /*
1820 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1821 * the contended case or if set_waiters is 1. The pi_state is returned
1822 * in ps in contended cases.
1823 */
1824 vpid = task_pid_vnr(top_waiter->task);
1825 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1826 set_waiters);
1827 if (ret == 1) {
1828 requeue_pi_wake_futex(top_waiter, key2, hb2);
1829 return vpid;
1830 }
1831 return ret;
1832}
1833
1834/**
1835 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1836 * @uaddr1: source futex user address
1837 * @flags: futex flags (FLAGS_SHARED, etc.)
1838 * @uaddr2: target futex user address
1839 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1840 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1841 * @cmpval: @uaddr1 expected value (or %NULL)
1842 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1843 * pi futex (pi to pi requeue is not supported)
1844 *
1845 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1846 * uaddr2 atomically on behalf of the top waiter.
1847 *
1848 * Return:
1849 * - >=0 - on success, the number of tasks requeued or woken;
1850 * - <0 - on error
1851 */
1852static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1853 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1854 u32 *cmpval, int requeue_pi)
1855{
1856 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1857 int drop_count = 0, task_count = 0, ret;
1858 struct futex_pi_state *pi_state = NULL;
1859 struct futex_hash_bucket *hb1, *hb2;
1860 struct futex_q *this, *next;
1861 DEFINE_WAKE_Q(wake_q);
1862
1863 if (nr_wake < 0 || nr_requeue < 0)
1864 return -EINVAL;
1865
1866 /*
1867 * When PI not supported: return -ENOSYS if requeue_pi is true,
1868 * consequently the compiler knows requeue_pi is always false past
1869 * this point which will optimize away all the conditional code
1870 * further down.
1871 */
1872 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1873 return -ENOSYS;
1874
1875 if (requeue_pi) {
1876 /*
1877 * Requeue PI only works on two distinct uaddrs. This
1878 * check is only valid for private futexes. See below.
1879 */
1880 if (uaddr1 == uaddr2)
1881 return -EINVAL;
1882
1883 /*
1884 * requeue_pi requires a pi_state, try to allocate it now
1885 * without any locks in case it fails.
1886 */
1887 if (refill_pi_state_cache())
1888 return -ENOMEM;
1889 /*
1890 * requeue_pi must wake as many tasks as it can, up to nr_wake
1891 * + nr_requeue, since it acquires the rt_mutex prior to
1892 * returning to userspace, so as to not leave the rt_mutex with
1893 * waiters and no owner. However, second and third wake-ups
1894 * cannot be predicted as they involve race conditions with the
1895 * first wake and a fault while looking up the pi_state. Both
1896 * pthread_cond_signal() and pthread_cond_broadcast() should
1897 * use nr_wake=1.
1898 */
1899 if (nr_wake != 1)
1900 return -EINVAL;
1901 }
1902
1903retry:
1904 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1905 if (unlikely(ret != 0))
1906 goto out;
1907 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1908 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1909 if (unlikely(ret != 0))
1910 goto out_put_key1;
1911
1912 /*
1913 * The check above which compares uaddrs is not sufficient for
1914 * shared futexes. We need to compare the keys:
1915 */
1916 if (requeue_pi && match_futex(&key1, &key2)) {
1917 ret = -EINVAL;
1918 goto out_put_keys;
1919 }
1920
1921 hb1 = hash_futex(&key1);
1922 hb2 = hash_futex(&key2);
1923
1924retry_private:
1925 hb_waiters_inc(hb2);
1926 double_lock_hb(hb1, hb2);
1927
1928 if (likely(cmpval != NULL)) {
1929 u32 curval;
1930
1931 ret = get_futex_value_locked(&curval, uaddr1);
1932
1933 if (unlikely(ret)) {
1934 double_unlock_hb(hb1, hb2);
1935 hb_waiters_dec(hb2);
1936
1937 ret = get_user(curval, uaddr1);
1938 if (ret)
1939 goto out_put_keys;
1940
1941 if (!(flags & FLAGS_SHARED))
1942 goto retry_private;
1943
1944 put_futex_key(&key2);
1945 put_futex_key(&key1);
1946 goto retry;
1947 }
1948 if (curval != *cmpval) {
1949 ret = -EAGAIN;
1950 goto out_unlock;
1951 }
1952 }
1953
1954 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1955 /*
1956 * Attempt to acquire uaddr2 and wake the top waiter. If we
1957 * intend to requeue waiters, force setting the FUTEX_WAITERS
1958 * bit. We force this here where we are able to easily handle
1959 * faults rather in the requeue loop below.
1960 */
1961 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1962 &key2, &pi_state, nr_requeue);
1963
1964 /*
1965 * At this point the top_waiter has either taken uaddr2 or is
1966 * waiting on it. If the former, then the pi_state will not
1967 * exist yet, look it up one more time to ensure we have a
1968 * reference to it. If the lock was taken, ret contains the
1969 * vpid of the top waiter task.
1970 * If the lock was not taken, we have pi_state and an initial
1971 * refcount on it. In case of an error we have nothing.
1972 */
1973 if (ret > 0) {
1974 WARN_ON(pi_state);
1975 drop_count++;
1976 task_count++;
1977 /*
1978 * If we acquired the lock, then the user space value
1979 * of uaddr2 should be vpid. It cannot be changed by
1980 * the top waiter as it is blocked on hb2 lock if it
1981 * tries to do so. If something fiddled with it behind
1982 * our back the pi state lookup might unearth it. So
1983 * we rather use the known value than rereading and
1984 * handing potential crap to lookup_pi_state.
1985 *
1986 * If that call succeeds then we have pi_state and an
1987 * initial refcount on it.
1988 */
1989 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
1990 }
1991
1992 switch (ret) {
1993 case 0:
1994 /* We hold a reference on the pi state. */
1995 break;
1996
1997 /* If the above failed, then pi_state is NULL */
1998 case -EFAULT:
1999 double_unlock_hb(hb1, hb2);
2000 hb_waiters_dec(hb2);
2001 put_futex_key(&key2);
2002 put_futex_key(&key1);
2003 ret = fault_in_user_writeable(uaddr2);
2004 if (!ret)
2005 goto retry;
2006 goto out;
2007 case -EAGAIN:
2008 /*
2009 * Two reasons for this:
2010 * - Owner is exiting and we just wait for the
2011 * exit to complete.
2012 * - The user space value changed.
2013 */
2014 double_unlock_hb(hb1, hb2);
2015 hb_waiters_dec(hb2);
2016 put_futex_key(&key2);
2017 put_futex_key(&key1);
2018 cond_resched();
2019 goto retry;
2020 default:
2021 goto out_unlock;
2022 }
2023 }
2024
2025 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2026 if (task_count - nr_wake >= nr_requeue)
2027 break;
2028
2029 if (!match_futex(&this->key, &key1))
2030 continue;
2031
2032 /*
2033 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2034 * be paired with each other and no other futex ops.
2035 *
2036 * We should never be requeueing a futex_q with a pi_state,
2037 * which is awaiting a futex_unlock_pi().
2038 */
2039 if ((requeue_pi && !this->rt_waiter) ||
2040 (!requeue_pi && this->rt_waiter) ||
2041 this->pi_state) {
2042 ret = -EINVAL;
2043 break;
2044 }
2045
2046 /*
2047 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2048 * lock, we already woke the top_waiter. If not, it will be
2049 * woken by futex_unlock_pi().
2050 */
2051 if (++task_count <= nr_wake && !requeue_pi) {
2052 mark_wake_futex(&wake_q, this);
2053 continue;
2054 }
2055
2056 /* Ensure we requeue to the expected futex for requeue_pi. */
2057 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2058 ret = -EINVAL;
2059 break;
2060 }
2061
2062 /*
2063 * Requeue nr_requeue waiters and possibly one more in the case
2064 * of requeue_pi if we couldn't acquire the lock atomically.
2065 */
2066 if (requeue_pi) {
2067 /*
2068 * Prepare the waiter to take the rt_mutex. Take a
2069 * refcount on the pi_state and store the pointer in
2070 * the futex_q object of the waiter.
2071 */
2072 get_pi_state(pi_state);
2073 this->pi_state = pi_state;
2074 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2075 this->rt_waiter,
2076 this->task);
2077 if (ret == 1) {
2078 /*
2079 * We got the lock. We do neither drop the
2080 * refcount on pi_state nor clear
2081 * this->pi_state because the waiter needs the
2082 * pi_state for cleaning up the user space
2083 * value. It will drop the refcount after
2084 * doing so.
2085 */
2086 requeue_pi_wake_futex(this, &key2, hb2);
2087 drop_count++;
2088 continue;
2089 } else if (ret) {
2090 /*
2091 * rt_mutex_start_proxy_lock() detected a
2092 * potential deadlock when we tried to queue
2093 * that waiter. Drop the pi_state reference
2094 * which we took above and remove the pointer
2095 * to the state from the waiters futex_q
2096 * object.
2097 */
2098 this->pi_state = NULL;
2099 put_pi_state(pi_state);
2100 /*
2101 * We stop queueing more waiters and let user
2102 * space deal with the mess.
2103 */
2104 break;
2105 }
2106 }
2107 requeue_futex(this, hb1, hb2, &key2);
2108 drop_count++;
2109 }
2110
2111 /*
2112 * We took an extra initial reference to the pi_state either
2113 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2114 * need to drop it here again.
2115 */
2116 put_pi_state(pi_state);
2117
2118out_unlock:
2119 double_unlock_hb(hb1, hb2);
2120 wake_up_q(&wake_q);
2121 hb_waiters_dec(hb2);
2122
2123 /*
2124 * drop_futex_key_refs() must be called outside the spinlocks. During
2125 * the requeue we moved futex_q's from the hash bucket at key1 to the
2126 * one at key2 and updated their key pointer. We no longer need to
2127 * hold the references to key1.
2128 */
2129 while (--drop_count >= 0)
2130 drop_futex_key_refs(&key1);
2131
2132out_put_keys:
2133 put_futex_key(&key2);
2134out_put_key1:
2135 put_futex_key(&key1);
2136out:
2137 return ret ? ret : task_count;
2138}
2139
2140/* The key must be already stored in q->key. */
2141static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2142 __acquires(&hb->lock)
2143{
2144 struct futex_hash_bucket *hb;
2145
2146 hb = hash_futex(&q->key);
2147
2148 /*
2149 * Increment the counter before taking the lock so that
2150 * a potential waker won't miss a to-be-slept task that is
2151 * waiting for the spinlock. This is safe as all queue_lock()
2152 * users end up calling queue_me(). Similarly, for housekeeping,
2153 * decrement the counter at queue_unlock() when some error has
2154 * occurred and we don't end up adding the task to the list.
2155 */
2156 hb_waiters_inc(hb);
2157
2158 q->lock_ptr = &hb->lock;
2159
2160 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2161 return hb;
2162}
2163
2164static inline void
2165queue_unlock(struct futex_hash_bucket *hb)
2166 __releases(&hb->lock)
2167{
2168 spin_unlock(&hb->lock);
2169 hb_waiters_dec(hb);
2170}
2171
2172static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2173{
2174 int prio;
2175
2176 /*
2177 * The priority used to register this element is
2178 * - either the real thread-priority for the real-time threads
2179 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2180 * - or MAX_RT_PRIO for non-RT threads.
2181 * Thus, all RT-threads are woken first in priority order, and
2182 * the others are woken last, in FIFO order.
2183 */
2184 prio = min(current->normal_prio, MAX_RT_PRIO);
2185
2186 plist_node_init(&q->list, prio);
2187 plist_add(&q->list, &hb->chain);
2188 q->task = current;
2189}
2190
2191/**
2192 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2193 * @q: The futex_q to enqueue
2194 * @hb: The destination hash bucket
2195 *
2196 * The hb->lock must be held by the caller, and is released here. A call to
2197 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2198 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2199 * or nothing if the unqueue is done as part of the wake process and the unqueue
2200 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2201 * an example).
2202 */
2203static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2204 __releases(&hb->lock)
2205{
2206 __queue_me(q, hb);
2207 spin_unlock(&hb->lock);
2208}
2209
2210/**
2211 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2212 * @q: The futex_q to unqueue
2213 *
2214 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2215 * be paired with exactly one earlier call to queue_me().
2216 *
2217 * Return:
2218 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2219 * - 0 - if the futex_q was already removed by the waking thread
2220 */
2221static int unqueue_me(struct futex_q *q)
2222{
2223 spinlock_t *lock_ptr;
2224 int ret = 0;
2225
2226 /* In the common case we don't take the spinlock, which is nice. */
2227retry:
2228 /*
2229 * q->lock_ptr can change between this read and the following spin_lock.
2230 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2231 * optimizing lock_ptr out of the logic below.
2232 */
2233 lock_ptr = READ_ONCE(q->lock_ptr);
2234 if (lock_ptr != NULL) {
2235 spin_lock(lock_ptr);
2236 /*
2237 * q->lock_ptr can change between reading it and
2238 * spin_lock(), causing us to take the wrong lock. This
2239 * corrects the race condition.
2240 *
2241 * Reasoning goes like this: if we have the wrong lock,
2242 * q->lock_ptr must have changed (maybe several times)
2243 * between reading it and the spin_lock(). It can
2244 * change again after the spin_lock() but only if it was
2245 * already changed before the spin_lock(). It cannot,
2246 * however, change back to the original value. Therefore
2247 * we can detect whether we acquired the correct lock.
2248 */
2249 if (unlikely(lock_ptr != q->lock_ptr)) {
2250 spin_unlock(lock_ptr);
2251 goto retry;
2252 }
2253 __unqueue_futex(q);
2254
2255 BUG_ON(q->pi_state);
2256
2257 spin_unlock(lock_ptr);
2258 ret = 1;
2259 }
2260
2261 drop_futex_key_refs(&q->key);
2262 return ret;
2263}
2264
2265/*
2266 * PI futexes can not be requeued and must remove themself from the
2267 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2268 * and dropped here.
2269 */
2270static void unqueue_me_pi(struct futex_q *q)
2271 __releases(q->lock_ptr)
2272{
2273 __unqueue_futex(q);
2274
2275 BUG_ON(!q->pi_state);
2276 put_pi_state(q->pi_state);
2277 q->pi_state = NULL;
2278
2279 spin_unlock(q->lock_ptr);
2280}
2281
2282static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2283 struct task_struct *argowner)
2284{
2285 struct futex_pi_state *pi_state = q->pi_state;
2286 u32 uval, uninitialized_var(curval), newval;
2287 struct task_struct *oldowner, *newowner;
2288 u32 newtid;
2289 int ret;
2290
2291 lockdep_assert_held(q->lock_ptr);
2292
2293 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2294
2295 oldowner = pi_state->owner;
2296
2297 /*
2298 * We are here because either:
2299 *
2300 * - we stole the lock and pi_state->owner needs updating to reflect
2301 * that (@argowner == current),
2302 *
2303 * or:
2304 *
2305 * - someone stole our lock and we need to fix things to point to the
2306 * new owner (@argowner == NULL).
2307 *
2308 * Either way, we have to replace the TID in the user space variable.
2309 * This must be atomic as we have to preserve the owner died bit here.
2310 *
2311 * Note: We write the user space value _before_ changing the pi_state
2312 * because we can fault here. Imagine swapped out pages or a fork
2313 * that marked all the anonymous memory readonly for cow.
2314 *
2315 * Modifying pi_state _before_ the user space value would leave the
2316 * pi_state in an inconsistent state when we fault here, because we
2317 * need to drop the locks to handle the fault. This might be observed
2318 * in the PID check in lookup_pi_state.
2319 */
2320retry:
2321 if (!argowner) {
2322 if (oldowner != current) {
2323 /*
2324 * We raced against a concurrent self; things are
2325 * already fixed up. Nothing to do.
2326 */
2327 ret = 0;
2328 goto out_unlock;
2329 }
2330
2331 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2332 /* We got the lock after all, nothing to fix. */
2333 ret = 0;
2334 goto out_unlock;
2335 }
2336
2337 /*
2338 * Since we just failed the trylock; there must be an owner.
2339 */
2340 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2341 BUG_ON(!newowner);
2342 } else {
2343 WARN_ON_ONCE(argowner != current);
2344 if (oldowner == current) {
2345 /*
2346 * We raced against a concurrent self; things are
2347 * already fixed up. Nothing to do.
2348 */
2349 ret = 0;
2350 goto out_unlock;
2351 }
2352 newowner = argowner;
2353 }
2354
2355 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2356 /* Owner died? */
2357 if (!pi_state->owner)
2358 newtid |= FUTEX_OWNER_DIED;
2359
2360 if (get_futex_value_locked(&uval, uaddr))
2361 goto handle_fault;
2362
2363 for (;;) {
2364 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2365
2366 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2367 goto handle_fault;
2368 if (curval == uval)
2369 break;
2370 uval = curval;
2371 }
2372
2373 /*
2374 * We fixed up user space. Now we need to fix the pi_state
2375 * itself.
2376 */
2377 if (pi_state->owner != NULL) {
2378 raw_spin_lock(&pi_state->owner->pi_lock);
2379 WARN_ON(list_empty(&pi_state->list));
2380 list_del_init(&pi_state->list);
2381 raw_spin_unlock(&pi_state->owner->pi_lock);
2382 }
2383
2384 pi_state->owner = newowner;
2385
2386 raw_spin_lock(&newowner->pi_lock);
2387 WARN_ON(!list_empty(&pi_state->list));
2388 list_add(&pi_state->list, &newowner->pi_state_list);
2389 raw_spin_unlock(&newowner->pi_lock);
2390 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2391
2392 return 0;
2393
2394 /*
2395 * To handle the page fault we need to drop the locks here. That gives
2396 * the other task (either the highest priority waiter itself or the
2397 * task which stole the rtmutex) the chance to try the fixup of the
2398 * pi_state. So once we are back from handling the fault we need to
2399 * check the pi_state after reacquiring the locks and before trying to
2400 * do another fixup. When the fixup has been done already we simply
2401 * return.
2402 *
2403 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2404 * drop hb->lock since the caller owns the hb -> futex_q relation.
2405 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2406 */
2407handle_fault:
2408 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2409 spin_unlock(q->lock_ptr);
2410
2411 ret = fault_in_user_writeable(uaddr);
2412
2413 spin_lock(q->lock_ptr);
2414 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2415
2416 /*
2417 * Check if someone else fixed it for us:
2418 */
2419 if (pi_state->owner != oldowner) {
2420 ret = 0;
2421 goto out_unlock;
2422 }
2423
2424 if (ret)
2425 goto out_unlock;
2426
2427 goto retry;
2428
2429out_unlock:
2430 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2431 return ret;
2432}
2433
2434static long futex_wait_restart(struct restart_block *restart);
2435
2436/**
2437 * fixup_owner() - Post lock pi_state and corner case management
2438 * @uaddr: user address of the futex
2439 * @q: futex_q (contains pi_state and access to the rt_mutex)
2440 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2441 *
2442 * After attempting to lock an rt_mutex, this function is called to cleanup
2443 * the pi_state owner as well as handle race conditions that may allow us to
2444 * acquire the lock. Must be called with the hb lock held.
2445 *
2446 * Return:
2447 * - 1 - success, lock taken;
2448 * - 0 - success, lock not taken;
2449 * - <0 - on error (-EFAULT)
2450 */
2451static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2452{
2453 int ret = 0;
2454
2455 if (locked) {
2456 /*
2457 * Got the lock. We might not be the anticipated owner if we
2458 * did a lock-steal - fix up the PI-state in that case:
2459 *
2460 * Speculative pi_state->owner read (we don't hold wait_lock);
2461 * since we own the lock pi_state->owner == current is the
2462 * stable state, anything else needs more attention.
2463 */
2464 if (q->pi_state->owner != current)
2465 ret = fixup_pi_state_owner(uaddr, q, current);
2466 goto out;
2467 }
2468
2469 /*
2470 * If we didn't get the lock; check if anybody stole it from us. In
2471 * that case, we need to fix up the uval to point to them instead of
2472 * us, otherwise bad things happen. [10]
2473 *
2474 * Another speculative read; pi_state->owner == current is unstable
2475 * but needs our attention.
2476 */
2477 if (q->pi_state->owner == current) {
2478 ret = fixup_pi_state_owner(uaddr, q, NULL);
2479 goto out;
2480 }
2481
2482 /*
2483 * Paranoia check. If we did not take the lock, then we should not be
2484 * the owner of the rt_mutex.
2485 */
2486 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2487 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2488 "pi-state %p\n", ret,
2489 q->pi_state->pi_mutex.owner,
2490 q->pi_state->owner);
2491 }
2492
2493out:
2494 return ret ? ret : locked;
2495}
2496
2497/**
2498 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2499 * @hb: the futex hash bucket, must be locked by the caller
2500 * @q: the futex_q to queue up on
2501 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2502 */
2503static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2504 struct hrtimer_sleeper *timeout)
2505{
2506 /*
2507 * The task state is guaranteed to be set before another task can
2508 * wake it. set_current_state() is implemented using smp_store_mb() and
2509 * queue_me() calls spin_unlock() upon completion, both serializing
2510 * access to the hash list and forcing another memory barrier.
2511 */
2512 set_current_state(TASK_INTERRUPTIBLE);
2513 queue_me(q, hb);
2514
2515 /* Arm the timer */
2516 if (timeout)
2517 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2518
2519 /*
2520 * If we have been removed from the hash list, then another task
2521 * has tried to wake us, and we can skip the call to schedule().
2522 */
2523 if (likely(!plist_node_empty(&q->list))) {
2524 /*
2525 * If the timer has already expired, current will already be
2526 * flagged for rescheduling. Only call schedule if there
2527 * is no timeout, or if it has yet to expire.
2528 */
2529 if (!timeout || timeout->task)
2530 freezable_schedule();
2531 }
2532 __set_current_state(TASK_RUNNING);
2533}
2534
2535/**
2536 * futex_wait_setup() - Prepare to wait on a futex
2537 * @uaddr: the futex userspace address
2538 * @val: the expected value
2539 * @flags: futex flags (FLAGS_SHARED, etc.)
2540 * @q: the associated futex_q
2541 * @hb: storage for hash_bucket pointer to be returned to caller
2542 *
2543 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2544 * compare it with the expected value. Handle atomic faults internally.
2545 * Return with the hb lock held and a q.key reference on success, and unlocked
2546 * with no q.key reference on failure.
2547 *
2548 * Return:
2549 * - 0 - uaddr contains val and hb has been locked;
2550 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2551 */
2552static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2553 struct futex_q *q, struct futex_hash_bucket **hb)
2554{
2555 u32 uval;
2556 int ret;
2557
2558 /*
2559 * Access the page AFTER the hash-bucket is locked.
2560 * Order is important:
2561 *
2562 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2563 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2564 *
2565 * The basic logical guarantee of a futex is that it blocks ONLY
2566 * if cond(var) is known to be true at the time of blocking, for
2567 * any cond. If we locked the hash-bucket after testing *uaddr, that
2568 * would open a race condition where we could block indefinitely with
2569 * cond(var) false, which would violate the guarantee.
2570 *
2571 * On the other hand, we insert q and release the hash-bucket only
2572 * after testing *uaddr. This guarantees that futex_wait() will NOT
2573 * absorb a wakeup if *uaddr does not match the desired values
2574 * while the syscall executes.
2575 */
2576retry:
2577 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2578 if (unlikely(ret != 0))
2579 return ret;
2580
2581retry_private:
2582 *hb = queue_lock(q);
2583
2584 ret = get_futex_value_locked(&uval, uaddr);
2585
2586 if (ret) {
2587 queue_unlock(*hb);
2588
2589 ret = get_user(uval, uaddr);
2590 if (ret)
2591 goto out;
2592
2593 if (!(flags & FLAGS_SHARED))
2594 goto retry_private;
2595
2596 put_futex_key(&q->key);
2597 goto retry;
2598 }
2599
2600 if (uval != val) {
2601 queue_unlock(*hb);
2602 ret = -EWOULDBLOCK;
2603 }
2604
2605out:
2606 if (ret)
2607 put_futex_key(&q->key);
2608 return ret;
2609}
2610
2611static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2612 ktime_t *abs_time, u32 bitset)
2613{
2614 struct hrtimer_sleeper timeout, *to = NULL;
2615 struct restart_block *restart;
2616 struct futex_hash_bucket *hb;
2617 struct futex_q q = futex_q_init;
2618 int ret;
2619
2620 if (!bitset)
2621 return -EINVAL;
2622 q.bitset = bitset;
2623
2624 if (abs_time) {
2625 to = &timeout;
2626
2627 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2628 CLOCK_REALTIME : CLOCK_MONOTONIC,
2629 HRTIMER_MODE_ABS);
2630 hrtimer_init_sleeper(to, current);
2631 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2632 current->timer_slack_ns);
2633 }
2634
2635retry:
2636 /*
2637 * Prepare to wait on uaddr. On success, holds hb lock and increments
2638 * q.key refs.
2639 */
2640 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2641 if (ret)
2642 goto out;
2643
2644 /* queue_me and wait for wakeup, timeout, or a signal. */
2645 futex_wait_queue_me(hb, &q, to);
2646
2647 /* If we were woken (and unqueued), we succeeded, whatever. */
2648 ret = 0;
2649 /* unqueue_me() drops q.key ref */
2650 if (!unqueue_me(&q))
2651 goto out;
2652 ret = -ETIMEDOUT;
2653 if (to && !to->task)
2654 goto out;
2655
2656 /*
2657 * We expect signal_pending(current), but we might be the
2658 * victim of a spurious wakeup as well.
2659 */
2660 if (!signal_pending(current))
2661 goto retry;
2662
2663 ret = -ERESTARTSYS;
2664 if (!abs_time)
2665 goto out;
2666
2667 restart = ¤t->restart_block;
2668 restart->fn = futex_wait_restart;
2669 restart->futex.uaddr = uaddr;
2670 restart->futex.val = val;
2671 restart->futex.time = *abs_time;
2672 restart->futex.bitset = bitset;
2673 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2674
2675 ret = -ERESTART_RESTARTBLOCK;
2676
2677out:
2678 if (to) {
2679 hrtimer_cancel(&to->timer);
2680 destroy_hrtimer_on_stack(&to->timer);
2681 }
2682 return ret;
2683}
2684
2685
2686static long futex_wait_restart(struct restart_block *restart)
2687{
2688 u32 __user *uaddr = restart->futex.uaddr;
2689 ktime_t t, *tp = NULL;
2690
2691 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2692 t = restart->futex.time;
2693 tp = &t;
2694 }
2695 restart->fn = do_no_restart_syscall;
2696
2697 return (long)futex_wait(uaddr, restart->futex.flags,
2698 restart->futex.val, tp, restart->futex.bitset);
2699}
2700
2701
2702/*
2703 * Userspace tried a 0 -> TID atomic transition of the futex value
2704 * and failed. The kernel side here does the whole locking operation:
2705 * if there are waiters then it will block as a consequence of relying
2706 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2707 * a 0 value of the futex too.).
2708 *
2709 * Also serves as futex trylock_pi()'ing, and due semantics.
2710 */
2711static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2712 ktime_t *time, int trylock)
2713{
2714 struct hrtimer_sleeper timeout, *to = NULL;
2715 struct futex_pi_state *pi_state = NULL;
2716 struct rt_mutex_waiter rt_waiter;
2717 struct futex_hash_bucket *hb;
2718 struct futex_q q = futex_q_init;
2719 int res, ret;
2720
2721 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2722 return -ENOSYS;
2723
2724 if (refill_pi_state_cache())
2725 return -ENOMEM;
2726
2727 if (time) {
2728 to = &timeout;
2729 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2730 HRTIMER_MODE_ABS);
2731 hrtimer_init_sleeper(to, current);
2732 hrtimer_set_expires(&to->timer, *time);
2733 }
2734
2735retry:
2736 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2737 if (unlikely(ret != 0))
2738 goto out;
2739
2740retry_private:
2741 hb = queue_lock(&q);
2742
2743 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2744 if (unlikely(ret)) {
2745 /*
2746 * Atomic work succeeded and we got the lock,
2747 * or failed. Either way, we do _not_ block.
2748 */
2749 switch (ret) {
2750 case 1:
2751 /* We got the lock. */
2752 ret = 0;
2753 goto out_unlock_put_key;
2754 case -EFAULT:
2755 goto uaddr_faulted;
2756 case -EAGAIN:
2757 /*
2758 * Two reasons for this:
2759 * - Task is exiting and we just wait for the
2760 * exit to complete.
2761 * - The user space value changed.
2762 */
2763 queue_unlock(hb);
2764 put_futex_key(&q.key);
2765 cond_resched();
2766 goto retry;
2767 default:
2768 goto out_unlock_put_key;
2769 }
2770 }
2771
2772 WARN_ON(!q.pi_state);
2773
2774 /*
2775 * Only actually queue now that the atomic ops are done:
2776 */
2777 __queue_me(&q, hb);
2778
2779 if (trylock) {
2780 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2781 /* Fixup the trylock return value: */
2782 ret = ret ? 0 : -EWOULDBLOCK;
2783 goto no_block;
2784 }
2785
2786 rt_mutex_init_waiter(&rt_waiter);
2787
2788 /*
2789 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2790 * hold it while doing rt_mutex_start_proxy(), because then it will
2791 * include hb->lock in the blocking chain, even through we'll not in
2792 * fact hold it while blocking. This will lead it to report -EDEADLK
2793 * and BUG when futex_unlock_pi() interleaves with this.
2794 *
2795 * Therefore acquire wait_lock while holding hb->lock, but drop the
2796 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2797 * serializes against futex_unlock_pi() as that does the exact same
2798 * lock handoff sequence.
2799 */
2800 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2801 spin_unlock(q.lock_ptr);
2802 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2803 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2804
2805 if (ret) {
2806 if (ret == 1)
2807 ret = 0;
2808
2809 spin_lock(q.lock_ptr);
2810 goto no_block;
2811 }
2812
2813
2814 if (unlikely(to))
2815 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2816
2817 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2818
2819 spin_lock(q.lock_ptr);
2820 /*
2821 * If we failed to acquire the lock (signal/timeout), we must
2822 * first acquire the hb->lock before removing the lock from the
2823 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2824 * wait lists consistent.
2825 *
2826 * In particular; it is important that futex_unlock_pi() can not
2827 * observe this inconsistency.
2828 */
2829 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2830 ret = 0;
2831
2832no_block:
2833 /*
2834 * Fixup the pi_state owner and possibly acquire the lock if we
2835 * haven't already.
2836 */
2837 res = fixup_owner(uaddr, &q, !ret);
2838 /*
2839 * If fixup_owner() returned an error, proprogate that. If it acquired
2840 * the lock, clear our -ETIMEDOUT or -EINTR.
2841 */
2842 if (res)
2843 ret = (res < 0) ? res : 0;
2844
2845 /*
2846 * If fixup_owner() faulted and was unable to handle the fault, unlock
2847 * it and return the fault to userspace.
2848 */
2849 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2850 pi_state = q.pi_state;
2851 get_pi_state(pi_state);
2852 }
2853
2854 /* Unqueue and drop the lock */
2855 unqueue_me_pi(&q);
2856
2857 if (pi_state) {
2858 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2859 put_pi_state(pi_state);
2860 }
2861
2862 goto out_put_key;
2863
2864out_unlock_put_key:
2865 queue_unlock(hb);
2866
2867out_put_key:
2868 put_futex_key(&q.key);
2869out:
2870 if (to) {
2871 hrtimer_cancel(&to->timer);
2872 destroy_hrtimer_on_stack(&to->timer);
2873 }
2874 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2875
2876uaddr_faulted:
2877 queue_unlock(hb);
2878
2879 ret = fault_in_user_writeable(uaddr);
2880 if (ret)
2881 goto out_put_key;
2882
2883 if (!(flags & FLAGS_SHARED))
2884 goto retry_private;
2885
2886 put_futex_key(&q.key);
2887 goto retry;
2888}
2889
2890/*
2891 * Userspace attempted a TID -> 0 atomic transition, and failed.
2892 * This is the in-kernel slowpath: we look up the PI state (if any),
2893 * and do the rt-mutex unlock.
2894 */
2895static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2896{
2897 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2898 union futex_key key = FUTEX_KEY_INIT;
2899 struct futex_hash_bucket *hb;
2900 struct futex_q *top_waiter;
2901 int ret;
2902
2903 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2904 return -ENOSYS;
2905
2906retry:
2907 if (get_user(uval, uaddr))
2908 return -EFAULT;
2909 /*
2910 * We release only a lock we actually own:
2911 */
2912 if ((uval & FUTEX_TID_MASK) != vpid)
2913 return -EPERM;
2914
2915 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2916 if (ret)
2917 return ret;
2918
2919 hb = hash_futex(&key);
2920 spin_lock(&hb->lock);
2921
2922 /*
2923 * Check waiters first. We do not trust user space values at
2924 * all and we at least want to know if user space fiddled
2925 * with the futex value instead of blindly unlocking.
2926 */
2927 top_waiter = futex_top_waiter(hb, &key);
2928 if (top_waiter) {
2929 struct futex_pi_state *pi_state = top_waiter->pi_state;
2930
2931 ret = -EINVAL;
2932 if (!pi_state)
2933 goto out_unlock;
2934
2935 /*
2936 * If current does not own the pi_state then the futex is
2937 * inconsistent and user space fiddled with the futex value.
2938 */
2939 if (pi_state->owner != current)
2940 goto out_unlock;
2941
2942 get_pi_state(pi_state);
2943 /*
2944 * By taking wait_lock while still holding hb->lock, we ensure
2945 * there is no point where we hold neither; and therefore
2946 * wake_futex_pi() must observe a state consistent with what we
2947 * observed.
2948 */
2949 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2950 spin_unlock(&hb->lock);
2951
2952 /* drops pi_state->pi_mutex.wait_lock */
2953 ret = wake_futex_pi(uaddr, uval, pi_state);
2954
2955 put_pi_state(pi_state);
2956
2957 /*
2958 * Success, we're done! No tricky corner cases.
2959 */
2960 if (!ret)
2961 goto out_putkey;
2962 /*
2963 * The atomic access to the futex value generated a
2964 * pagefault, so retry the user-access and the wakeup:
2965 */
2966 if (ret == -EFAULT)
2967 goto pi_faulted;
2968 /*
2969 * A unconditional UNLOCK_PI op raced against a waiter
2970 * setting the FUTEX_WAITERS bit. Try again.
2971 */
2972 if (ret == -EAGAIN) {
2973 put_futex_key(&key);
2974 goto retry;
2975 }
2976 /*
2977 * wake_futex_pi has detected invalid state. Tell user
2978 * space.
2979 */
2980 goto out_putkey;
2981 }
2982
2983 /*
2984 * We have no kernel internal state, i.e. no waiters in the
2985 * kernel. Waiters which are about to queue themselves are stuck
2986 * on hb->lock. So we can safely ignore them. We do neither
2987 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2988 * owner.
2989 */
2990 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2991 spin_unlock(&hb->lock);
2992 goto pi_faulted;
2993 }
2994
2995 /*
2996 * If uval has changed, let user space handle it.
2997 */
2998 ret = (curval == uval) ? 0 : -EAGAIN;
2999
3000out_unlock:
3001 spin_unlock(&hb->lock);
3002out_putkey:
3003 put_futex_key(&key);
3004 return ret;
3005
3006pi_faulted:
3007 put_futex_key(&key);
3008
3009 ret = fault_in_user_writeable(uaddr);
3010 if (!ret)
3011 goto retry;
3012
3013 return ret;
3014}
3015
3016/**
3017 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3018 * @hb: the hash_bucket futex_q was original enqueued on
3019 * @q: the futex_q woken while waiting to be requeued
3020 * @key2: the futex_key of the requeue target futex
3021 * @timeout: the timeout associated with the wait (NULL if none)
3022 *
3023 * Detect if the task was woken on the initial futex as opposed to the requeue
3024 * target futex. If so, determine if it was a timeout or a signal that caused
3025 * the wakeup and return the appropriate error code to the caller. Must be
3026 * called with the hb lock held.
3027 *
3028 * Return:
3029 * - 0 = no early wakeup detected;
3030 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3031 */
3032static inline
3033int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3034 struct futex_q *q, union futex_key *key2,
3035 struct hrtimer_sleeper *timeout)
3036{
3037 int ret = 0;
3038
3039 /*
3040 * With the hb lock held, we avoid races while we process the wakeup.
3041 * We only need to hold hb (and not hb2) to ensure atomicity as the
3042 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3043 * It can't be requeued from uaddr2 to something else since we don't
3044 * support a PI aware source futex for requeue.
3045 */
3046 if (!match_futex(&q->key, key2)) {
3047 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3048 /*
3049 * We were woken prior to requeue by a timeout or a signal.
3050 * Unqueue the futex_q and determine which it was.
3051 */
3052 plist_del(&q->list, &hb->chain);
3053 hb_waiters_dec(hb);
3054
3055 /* Handle spurious wakeups gracefully */
3056 ret = -EWOULDBLOCK;
3057 if (timeout && !timeout->task)
3058 ret = -ETIMEDOUT;
3059 else if (signal_pending(current))
3060 ret = -ERESTARTNOINTR;
3061 }
3062 return ret;
3063}
3064
3065/**
3066 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3067 * @uaddr: the futex we initially wait on (non-pi)
3068 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3069 * the same type, no requeueing from private to shared, etc.
3070 * @val: the expected value of uaddr
3071 * @abs_time: absolute timeout
3072 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3073 * @uaddr2: the pi futex we will take prior to returning to user-space
3074 *
3075 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3076 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3077 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3078 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3079 * without one, the pi logic would not know which task to boost/deboost, if
3080 * there was a need to.
3081 *
3082 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3083 * via the following--
3084 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3085 * 2) wakeup on uaddr2 after a requeue
3086 * 3) signal
3087 * 4) timeout
3088 *
3089 * If 3, cleanup and return -ERESTARTNOINTR.
3090 *
3091 * If 2, we may then block on trying to take the rt_mutex and return via:
3092 * 5) successful lock
3093 * 6) signal
3094 * 7) timeout
3095 * 8) other lock acquisition failure
3096 *
3097 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3098 *
3099 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3100 *
3101 * Return:
3102 * - 0 - On success;
3103 * - <0 - On error
3104 */
3105static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3106 u32 val, ktime_t *abs_time, u32 bitset,
3107 u32 __user *uaddr2)
3108{
3109 struct hrtimer_sleeper timeout, *to = NULL;
3110 struct futex_pi_state *pi_state = NULL;
3111 struct rt_mutex_waiter rt_waiter;
3112 struct futex_hash_bucket *hb;
3113 union futex_key key2 = FUTEX_KEY_INIT;
3114 struct futex_q q = futex_q_init;
3115 int res, ret;
3116
3117 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3118 return -ENOSYS;
3119
3120 if (uaddr == uaddr2)
3121 return -EINVAL;
3122
3123 if (!bitset)
3124 return -EINVAL;
3125
3126 if (abs_time) {
3127 to = &timeout;
3128 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3129 CLOCK_REALTIME : CLOCK_MONOTONIC,
3130 HRTIMER_MODE_ABS);
3131 hrtimer_init_sleeper(to, current);
3132 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3133 current->timer_slack_ns);
3134 }
3135
3136 /*
3137 * The waiter is allocated on our stack, manipulated by the requeue
3138 * code while we sleep on uaddr.
3139 */
3140 rt_mutex_init_waiter(&rt_waiter);
3141
3142 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3143 if (unlikely(ret != 0))
3144 goto out;
3145
3146 q.bitset = bitset;
3147 q.rt_waiter = &rt_waiter;
3148 q.requeue_pi_key = &key2;
3149
3150 /*
3151 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3152 * count.
3153 */
3154 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3155 if (ret)
3156 goto out_key2;
3157
3158 /*
3159 * The check above which compares uaddrs is not sufficient for
3160 * shared futexes. We need to compare the keys:
3161 */
3162 if (match_futex(&q.key, &key2)) {
3163 queue_unlock(hb);
3164 ret = -EINVAL;
3165 goto out_put_keys;
3166 }
3167
3168 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3169 futex_wait_queue_me(hb, &q, to);
3170
3171 spin_lock(&hb->lock);
3172 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3173 spin_unlock(&hb->lock);
3174 if (ret)
3175 goto out_put_keys;
3176
3177 /*
3178 * In order for us to be here, we know our q.key == key2, and since
3179 * we took the hb->lock above, we also know that futex_requeue() has
3180 * completed and we no longer have to concern ourselves with a wakeup
3181 * race with the atomic proxy lock acquisition by the requeue code. The
3182 * futex_requeue dropped our key1 reference and incremented our key2
3183 * reference count.
3184 */
3185
3186 /* Check if the requeue code acquired the second futex for us. */
3187 if (!q.rt_waiter) {
3188 /*
3189 * Got the lock. We might not be the anticipated owner if we
3190 * did a lock-steal - fix up the PI-state in that case.
3191 */
3192 if (q.pi_state && (q.pi_state->owner != current)) {
3193 spin_lock(q.lock_ptr);
3194 ret = fixup_pi_state_owner(uaddr2, &q, current);
3195 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3196 pi_state = q.pi_state;
3197 get_pi_state(pi_state);
3198 }
3199 /*
3200 * Drop the reference to the pi state which
3201 * the requeue_pi() code acquired for us.
3202 */
3203 put_pi_state(q.pi_state);
3204 spin_unlock(q.lock_ptr);
3205 }
3206 } else {
3207 struct rt_mutex *pi_mutex;
3208
3209 /*
3210 * We have been woken up by futex_unlock_pi(), a timeout, or a
3211 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3212 * the pi_state.
3213 */
3214 WARN_ON(!q.pi_state);
3215 pi_mutex = &q.pi_state->pi_mutex;
3216 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3217
3218 spin_lock(q.lock_ptr);
3219 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3220 ret = 0;
3221
3222 debug_rt_mutex_free_waiter(&rt_waiter);
3223 /*
3224 * Fixup the pi_state owner and possibly acquire the lock if we
3225 * haven't already.
3226 */
3227 res = fixup_owner(uaddr2, &q, !ret);
3228 /*
3229 * If fixup_owner() returned an error, proprogate that. If it
3230 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3231 */
3232 if (res)
3233 ret = (res < 0) ? res : 0;
3234
3235 /*
3236 * If fixup_pi_state_owner() faulted and was unable to handle
3237 * the fault, unlock the rt_mutex and return the fault to
3238 * userspace.
3239 */
3240 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3241 pi_state = q.pi_state;
3242 get_pi_state(pi_state);
3243 }
3244
3245 /* Unqueue and drop the lock. */
3246 unqueue_me_pi(&q);
3247 }
3248
3249 if (pi_state) {
3250 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3251 put_pi_state(pi_state);
3252 }
3253
3254 if (ret == -EINTR) {
3255 /*
3256 * We've already been requeued, but cannot restart by calling
3257 * futex_lock_pi() directly. We could restart this syscall, but
3258 * it would detect that the user space "val" changed and return
3259 * -EWOULDBLOCK. Save the overhead of the restart and return
3260 * -EWOULDBLOCK directly.
3261 */
3262 ret = -EWOULDBLOCK;
3263 }
3264
3265out_put_keys:
3266 put_futex_key(&q.key);
3267out_key2:
3268 put_futex_key(&key2);
3269
3270out:
3271 if (to) {
3272 hrtimer_cancel(&to->timer);
3273 destroy_hrtimer_on_stack(&to->timer);
3274 }
3275 return ret;
3276}
3277
3278/*
3279 * Support for robust futexes: the kernel cleans up held futexes at
3280 * thread exit time.
3281 *
3282 * Implementation: user-space maintains a per-thread list of locks it
3283 * is holding. Upon do_exit(), the kernel carefully walks this list,
3284 * and marks all locks that are owned by this thread with the
3285 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3286 * always manipulated with the lock held, so the list is private and
3287 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3288 * field, to allow the kernel to clean up if the thread dies after
3289 * acquiring the lock, but just before it could have added itself to
3290 * the list. There can only be one such pending lock.
3291 */
3292
3293/**
3294 * sys_set_robust_list() - Set the robust-futex list head of a task
3295 * @head: pointer to the list-head
3296 * @len: length of the list-head, as userspace expects
3297 */
3298SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3299 size_t, len)
3300{
3301 if (!futex_cmpxchg_enabled)
3302 return -ENOSYS;
3303 /*
3304 * The kernel knows only one size for now:
3305 */
3306 if (unlikely(len != sizeof(*head)))
3307 return -EINVAL;
3308
3309 current->robust_list = head;
3310
3311 return 0;
3312}
3313
3314/**
3315 * sys_get_robust_list() - Get the robust-futex list head of a task
3316 * @pid: pid of the process [zero for current task]
3317 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3318 * @len_ptr: pointer to a length field, the kernel fills in the header size
3319 */
3320SYSCALL_DEFINE3(get_robust_list, int, pid,
3321 struct robust_list_head __user * __user *, head_ptr,
3322 size_t __user *, len_ptr)
3323{
3324 struct robust_list_head __user *head;
3325 unsigned long ret;
3326 struct task_struct *p;
3327
3328 if (!futex_cmpxchg_enabled)
3329 return -ENOSYS;
3330
3331 rcu_read_lock();
3332
3333 ret = -ESRCH;
3334 if (!pid)
3335 p = current;
3336 else {
3337 p = find_task_by_vpid(pid);
3338 if (!p)
3339 goto err_unlock;
3340 }
3341
3342 ret = -EPERM;
3343 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3344 goto err_unlock;
3345
3346 head = p->robust_list;
3347 rcu_read_unlock();
3348
3349 if (put_user(sizeof(*head), len_ptr))
3350 return -EFAULT;
3351 return put_user(head, head_ptr);
3352
3353err_unlock:
3354 rcu_read_unlock();
3355
3356 return ret;
3357}
3358
3359/*
3360 * Process a futex-list entry, check whether it's owned by the
3361 * dying task, and do notification if so:
3362 */
3363int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3364{
3365 u32 uval, uninitialized_var(nval), mval;
3366
3367retry:
3368 if (get_user(uval, uaddr))
3369 return -1;
3370
3371 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3372 /*
3373 * Ok, this dying thread is truly holding a futex
3374 * of interest. Set the OWNER_DIED bit atomically
3375 * via cmpxchg, and if the value had FUTEX_WAITERS
3376 * set, wake up a waiter (if any). (We have to do a
3377 * futex_wake() even if OWNER_DIED is already set -
3378 * to handle the rare but possible case of recursive
3379 * thread-death.) The rest of the cleanup is done in
3380 * userspace.
3381 */
3382 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3383 /*
3384 * We are not holding a lock here, but we want to have
3385 * the pagefault_disable/enable() protection because
3386 * we want to handle the fault gracefully. If the
3387 * access fails we try to fault in the futex with R/W
3388 * verification via get_user_pages. get_user() above
3389 * does not guarantee R/W access. If that fails we
3390 * give up and leave the futex locked.
3391 */
3392 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3393 if (fault_in_user_writeable(uaddr))
3394 return -1;
3395 goto retry;
3396 }
3397 if (nval != uval)
3398 goto retry;
3399
3400 /*
3401 * Wake robust non-PI futexes here. The wakeup of
3402 * PI futexes happens in exit_pi_state():
3403 */
3404 if (!pi && (uval & FUTEX_WAITERS))
3405 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3406 }
3407 return 0;
3408}
3409
3410/*
3411 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3412 */
3413static inline int fetch_robust_entry(struct robust_list __user **entry,
3414 struct robust_list __user * __user *head,
3415 unsigned int *pi)
3416{
3417 unsigned long uentry;
3418
3419 if (get_user(uentry, (unsigned long __user *)head))
3420 return -EFAULT;
3421
3422 *entry = (void __user *)(uentry & ~1UL);
3423 *pi = uentry & 1;
3424
3425 return 0;
3426}
3427
3428/*
3429 * Walk curr->robust_list (very carefully, it's a userspace list!)
3430 * and mark any locks found there dead, and notify any waiters.
3431 *
3432 * We silently return on any sign of list-walking problem.
3433 */
3434void exit_robust_list(struct task_struct *curr)
3435{
3436 struct robust_list_head __user *head = curr->robust_list;
3437 struct robust_list __user *entry, *next_entry, *pending;
3438 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3439 unsigned int uninitialized_var(next_pi);
3440 unsigned long futex_offset;
3441 int rc;
3442
3443 if (!futex_cmpxchg_enabled)
3444 return;
3445
3446 /*
3447 * Fetch the list head (which was registered earlier, via
3448 * sys_set_robust_list()):
3449 */
3450 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3451 return;
3452 /*
3453 * Fetch the relative futex offset:
3454 */
3455 if (get_user(futex_offset, &head->futex_offset))
3456 return;
3457 /*
3458 * Fetch any possibly pending lock-add first, and handle it
3459 * if it exists:
3460 */
3461 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3462 return;
3463
3464 next_entry = NULL; /* avoid warning with gcc */
3465 while (entry != &head->list) {
3466 /*
3467 * Fetch the next entry in the list before calling
3468 * handle_futex_death:
3469 */
3470 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3471 /*
3472 * A pending lock might already be on the list, so
3473 * don't process it twice:
3474 */
3475 if (entry != pending)
3476 if (handle_futex_death((void __user *)entry + futex_offset,
3477 curr, pi))
3478 return;
3479 if (rc)
3480 return;
3481 entry = next_entry;
3482 pi = next_pi;
3483 /*
3484 * Avoid excessively long or circular lists:
3485 */
3486 if (!--limit)
3487 break;
3488
3489 cond_resched();
3490 }
3491
3492 if (pending)
3493 handle_futex_death((void __user *)pending + futex_offset,
3494 curr, pip);
3495}
3496
3497long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3498 u32 __user *uaddr2, u32 val2, u32 val3)
3499{
3500 int cmd = op & FUTEX_CMD_MASK;
3501 unsigned int flags = 0;
3502
3503 if (!(op & FUTEX_PRIVATE_FLAG))
3504 flags |= FLAGS_SHARED;
3505
3506 if (op & FUTEX_CLOCK_REALTIME) {
3507 flags |= FLAGS_CLOCKRT;
3508 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3509 cmd != FUTEX_WAIT_REQUEUE_PI)
3510 return -ENOSYS;
3511 }
3512
3513 switch (cmd) {
3514 case FUTEX_LOCK_PI:
3515 case FUTEX_UNLOCK_PI:
3516 case FUTEX_TRYLOCK_PI:
3517 case FUTEX_WAIT_REQUEUE_PI:
3518 case FUTEX_CMP_REQUEUE_PI:
3519 if (!futex_cmpxchg_enabled)
3520 return -ENOSYS;
3521 }
3522
3523 switch (cmd) {
3524 case FUTEX_WAIT:
3525 val3 = FUTEX_BITSET_MATCH_ANY;
3526 case FUTEX_WAIT_BITSET:
3527 return futex_wait(uaddr, flags, val, timeout, val3);
3528 case FUTEX_WAKE:
3529 val3 = FUTEX_BITSET_MATCH_ANY;
3530 case FUTEX_WAKE_BITSET:
3531 return futex_wake(uaddr, flags, val, val3);
3532 case FUTEX_REQUEUE:
3533 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3534 case FUTEX_CMP_REQUEUE:
3535 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3536 case FUTEX_WAKE_OP:
3537 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3538 case FUTEX_LOCK_PI:
3539 return futex_lock_pi(uaddr, flags, timeout, 0);
3540 case FUTEX_UNLOCK_PI:
3541 return futex_unlock_pi(uaddr, flags);
3542 case FUTEX_TRYLOCK_PI:
3543 return futex_lock_pi(uaddr, flags, NULL, 1);
3544 case FUTEX_WAIT_REQUEUE_PI:
3545 val3 = FUTEX_BITSET_MATCH_ANY;
3546 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3547 uaddr2);
3548 case FUTEX_CMP_REQUEUE_PI:
3549 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3550 }
3551 return -ENOSYS;
3552}
3553
3554
3555SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3556 struct timespec __user *, utime, u32 __user *, uaddr2,
3557 u32, val3)
3558{
3559 struct timespec ts;
3560 ktime_t t, *tp = NULL;
3561 u32 val2 = 0;
3562 int cmd = op & FUTEX_CMD_MASK;
3563
3564 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3565 cmd == FUTEX_WAIT_BITSET ||
3566 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3567 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3568 return -EFAULT;
3569 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3570 return -EFAULT;
3571 if (!timespec_valid(&ts))
3572 return -EINVAL;
3573
3574 t = timespec_to_ktime(ts);
3575 if (cmd == FUTEX_WAIT)
3576 t = ktime_add_safe(ktime_get(), t);
3577 tp = &t;
3578 }
3579 /*
3580 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3581 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3582 */
3583 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3584 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3585 val2 = (u32) (unsigned long) utime;
3586
3587 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3588}
3589
3590static void __init futex_detect_cmpxchg(void)
3591{
3592#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3593 u32 curval;
3594
3595 /*
3596 * This will fail and we want it. Some arch implementations do
3597 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3598 * functionality. We want to know that before we call in any
3599 * of the complex code paths. Also we want to prevent
3600 * registration of robust lists in that case. NULL is
3601 * guaranteed to fault and we get -EFAULT on functional
3602 * implementation, the non-functional ones will return
3603 * -ENOSYS.
3604 */
3605 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3606 futex_cmpxchg_enabled = 1;
3607#endif
3608}
3609
3610static int __init futex_init(void)
3611{
3612 unsigned int futex_shift;
3613 unsigned long i;
3614
3615#if CONFIG_BASE_SMALL
3616 futex_hashsize = 16;
3617#else
3618 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3619#endif
3620
3621 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3622 futex_hashsize, 0,
3623 futex_hashsize < 256 ? HASH_SMALL : 0,
3624 &futex_shift, NULL,
3625 futex_hashsize, futex_hashsize);
3626 futex_hashsize = 1UL << futex_shift;
3627
3628 futex_detect_cmpxchg();
3629
3630 for (i = 0; i < futex_hashsize; i++) {
3631 atomic_set(&futex_queues[i].waiters, 0);
3632 plist_head_init(&futex_queues[i].chain);
3633 spin_lock_init(&futex_queues[i].lock);
3634 }
3635
3636 return 0;
3637}
3638core_initcall(futex_init);