Loading...
Note: File does not exist in v6.2.
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
5 *
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 *
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
11 *
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 *
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 *
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 *
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 *
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
30 *
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
33 */
34#include <linux/compat.h>
35#include <linux/jhash.h>
36#include <linux/pagemap.h>
37#include <linux/syscalls.h>
38#include <linux/hugetlb.h>
39#include <linux/freezer.h>
40#include <linux/memblock.h>
41#include <linux/fault-inject.h>
42
43#include <asm/futex.h>
44
45#include "locking/rtmutex_common.h"
46
47/*
48 * READ this before attempting to hack on futexes!
49 *
50 * Basic futex operation and ordering guarantees
51 * =============================================
52 *
53 * The waiter reads the futex value in user space and calls
54 * futex_wait(). This function computes the hash bucket and acquires
55 * the hash bucket lock. After that it reads the futex user space value
56 * again and verifies that the data has not changed. If it has not changed
57 * it enqueues itself into the hash bucket, releases the hash bucket lock
58 * and schedules.
59 *
60 * The waker side modifies the user space value of the futex and calls
61 * futex_wake(). This function computes the hash bucket and acquires the
62 * hash bucket lock. Then it looks for waiters on that futex in the hash
63 * bucket and wakes them.
64 *
65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
66 * the hb spinlock can be avoided and simply return. In order for this
67 * optimization to work, ordering guarantees must exist so that the waiter
68 * being added to the list is acknowledged when the list is concurrently being
69 * checked by the waker, avoiding scenarios like the following:
70 *
71 * CPU 0 CPU 1
72 * val = *futex;
73 * sys_futex(WAIT, futex, val);
74 * futex_wait(futex, val);
75 * uval = *futex;
76 * *futex = newval;
77 * sys_futex(WAKE, futex);
78 * futex_wake(futex);
79 * if (queue_empty())
80 * return;
81 * if (uval == val)
82 * lock(hash_bucket(futex));
83 * queue();
84 * unlock(hash_bucket(futex));
85 * schedule();
86 *
87 * This would cause the waiter on CPU 0 to wait forever because it
88 * missed the transition of the user space value from val to newval
89 * and the waker did not find the waiter in the hash bucket queue.
90 *
91 * The correct serialization ensures that a waiter either observes
92 * the changed user space value before blocking or is woken by a
93 * concurrent waker:
94 *
95 * CPU 0 CPU 1
96 * val = *futex;
97 * sys_futex(WAIT, futex, val);
98 * futex_wait(futex, val);
99 *
100 * waiters++; (a)
101 * smp_mb(); (A) <-- paired with -.
102 * |
103 * lock(hash_bucket(futex)); |
104 * |
105 * uval = *futex; |
106 * | *futex = newval;
107 * | sys_futex(WAKE, futex);
108 * | futex_wake(futex);
109 * |
110 * `--------> smp_mb(); (B)
111 * if (uval == val)
112 * queue();
113 * unlock(hash_bucket(futex));
114 * schedule(); if (waiters)
115 * lock(hash_bucket(futex));
116 * else wake_waiters(futex);
117 * waiters--; (b) unlock(hash_bucket(futex));
118 *
119 * Where (A) orders the waiters increment and the futex value read through
120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
121 * to futex and the waiters read (see hb_waiters_pending()).
122 *
123 * This yields the following case (where X:=waiters, Y:=futex):
124 *
125 * X = Y = 0
126 *
127 * w[X]=1 w[Y]=1
128 * MB MB
129 * r[Y]=y r[X]=x
130 *
131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
132 * the guarantee that we cannot both miss the futex variable change and the
133 * enqueue.
134 *
135 * Note that a new waiter is accounted for in (a) even when it is possible that
136 * the wait call can return error, in which case we backtrack from it in (b).
137 * Refer to the comment in queue_lock().
138 *
139 * Similarly, in order to account for waiters being requeued on another
140 * address we always increment the waiters for the destination bucket before
141 * acquiring the lock. It then decrements them again after releasing it -
142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
143 * will do the additional required waiter count housekeeping. This is done for
144 * double_lock_hb() and double_unlock_hb(), respectively.
145 */
146
147#ifdef CONFIG_HAVE_FUTEX_CMPXCHG
148#define futex_cmpxchg_enabled 1
149#else
150static int __read_mostly futex_cmpxchg_enabled;
151#endif
152
153/*
154 * Futex flags used to encode options to functions and preserve them across
155 * restarts.
156 */
157#ifdef CONFIG_MMU
158# define FLAGS_SHARED 0x01
159#else
160/*
161 * NOMMU does not have per process address space. Let the compiler optimize
162 * code away.
163 */
164# define FLAGS_SHARED 0x00
165#endif
166#define FLAGS_CLOCKRT 0x02
167#define FLAGS_HAS_TIMEOUT 0x04
168
169/*
170 * Priority Inheritance state:
171 */
172struct futex_pi_state {
173 /*
174 * list of 'owned' pi_state instances - these have to be
175 * cleaned up in do_exit() if the task exits prematurely:
176 */
177 struct list_head list;
178
179 /*
180 * The PI object:
181 */
182 struct rt_mutex pi_mutex;
183
184 struct task_struct *owner;
185 refcount_t refcount;
186
187 union futex_key key;
188} __randomize_layout;
189
190/**
191 * struct futex_q - The hashed futex queue entry, one per waiting task
192 * @list: priority-sorted list of tasks waiting on this futex
193 * @task: the task waiting on the futex
194 * @lock_ptr: the hash bucket lock
195 * @key: the key the futex is hashed on
196 * @pi_state: optional priority inheritance state
197 * @rt_waiter: rt_waiter storage for use with requeue_pi
198 * @requeue_pi_key: the requeue_pi target futex key
199 * @bitset: bitset for the optional bitmasked wakeup
200 *
201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
202 * we can wake only the relevant ones (hashed queues may be shared).
203 *
204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
206 * The order of wakeup is always to make the first condition true, then
207 * the second.
208 *
209 * PI futexes are typically woken before they are removed from the hash list via
210 * the rt_mutex code. See unqueue_me_pi().
211 */
212struct futex_q {
213 struct plist_node list;
214
215 struct task_struct *task;
216 spinlock_t *lock_ptr;
217 union futex_key key;
218 struct futex_pi_state *pi_state;
219 struct rt_mutex_waiter *rt_waiter;
220 union futex_key *requeue_pi_key;
221 u32 bitset;
222} __randomize_layout;
223
224static const struct futex_q futex_q_init = {
225 /* list gets initialized in queue_me()*/
226 .key = FUTEX_KEY_INIT,
227 .bitset = FUTEX_BITSET_MATCH_ANY
228};
229
230/*
231 * Hash buckets are shared by all the futex_keys that hash to the same
232 * location. Each key may have multiple futex_q structures, one for each task
233 * waiting on a futex.
234 */
235struct futex_hash_bucket {
236 atomic_t waiters;
237 spinlock_t lock;
238 struct plist_head chain;
239} ____cacheline_aligned_in_smp;
240
241/*
242 * The base of the bucket array and its size are always used together
243 * (after initialization only in hash_futex()), so ensure that they
244 * reside in the same cacheline.
245 */
246static struct {
247 struct futex_hash_bucket *queues;
248 unsigned long hashsize;
249} __futex_data __read_mostly __aligned(2*sizeof(long));
250#define futex_queues (__futex_data.queues)
251#define futex_hashsize (__futex_data.hashsize)
252
253
254/*
255 * Fault injections for futexes.
256 */
257#ifdef CONFIG_FAIL_FUTEX
258
259static struct {
260 struct fault_attr attr;
261
262 bool ignore_private;
263} fail_futex = {
264 .attr = FAULT_ATTR_INITIALIZER,
265 .ignore_private = false,
266};
267
268static int __init setup_fail_futex(char *str)
269{
270 return setup_fault_attr(&fail_futex.attr, str);
271}
272__setup("fail_futex=", setup_fail_futex);
273
274static bool should_fail_futex(bool fshared)
275{
276 if (fail_futex.ignore_private && !fshared)
277 return false;
278
279 return should_fail(&fail_futex.attr, 1);
280}
281
282#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
283
284static int __init fail_futex_debugfs(void)
285{
286 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
287 struct dentry *dir;
288
289 dir = fault_create_debugfs_attr("fail_futex", NULL,
290 &fail_futex.attr);
291 if (IS_ERR(dir))
292 return PTR_ERR(dir);
293
294 debugfs_create_bool("ignore-private", mode, dir,
295 &fail_futex.ignore_private);
296 return 0;
297}
298
299late_initcall(fail_futex_debugfs);
300
301#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
302
303#else
304static inline bool should_fail_futex(bool fshared)
305{
306 return false;
307}
308#endif /* CONFIG_FAIL_FUTEX */
309
310#ifdef CONFIG_COMPAT
311static void compat_exit_robust_list(struct task_struct *curr);
312#else
313static inline void compat_exit_robust_list(struct task_struct *curr) { }
314#endif
315
316/*
317 * Reflects a new waiter being added to the waitqueue.
318 */
319static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
320{
321#ifdef CONFIG_SMP
322 atomic_inc(&hb->waiters);
323 /*
324 * Full barrier (A), see the ordering comment above.
325 */
326 smp_mb__after_atomic();
327#endif
328}
329
330/*
331 * Reflects a waiter being removed from the waitqueue by wakeup
332 * paths.
333 */
334static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
335{
336#ifdef CONFIG_SMP
337 atomic_dec(&hb->waiters);
338#endif
339}
340
341static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
342{
343#ifdef CONFIG_SMP
344 /*
345 * Full barrier (B), see the ordering comment above.
346 */
347 smp_mb();
348 return atomic_read(&hb->waiters);
349#else
350 return 1;
351#endif
352}
353
354/**
355 * hash_futex - Return the hash bucket in the global hash
356 * @key: Pointer to the futex key for which the hash is calculated
357 *
358 * We hash on the keys returned from get_futex_key (see below) and return the
359 * corresponding hash bucket in the global hash.
360 */
361static struct futex_hash_bucket *hash_futex(union futex_key *key)
362{
363 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
364 key->both.offset);
365
366 return &futex_queues[hash & (futex_hashsize - 1)];
367}
368
369
370/**
371 * match_futex - Check whether two futex keys are equal
372 * @key1: Pointer to key1
373 * @key2: Pointer to key2
374 *
375 * Return 1 if two futex_keys are equal, 0 otherwise.
376 */
377static inline int match_futex(union futex_key *key1, union futex_key *key2)
378{
379 return (key1 && key2
380 && key1->both.word == key2->both.word
381 && key1->both.ptr == key2->both.ptr
382 && key1->both.offset == key2->both.offset);
383}
384
385enum futex_access {
386 FUTEX_READ,
387 FUTEX_WRITE
388};
389
390/**
391 * futex_setup_timer - set up the sleeping hrtimer.
392 * @time: ptr to the given timeout value
393 * @timeout: the hrtimer_sleeper structure to be set up
394 * @flags: futex flags
395 * @range_ns: optional range in ns
396 *
397 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
398 * value given
399 */
400static inline struct hrtimer_sleeper *
401futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
402 int flags, u64 range_ns)
403{
404 if (!time)
405 return NULL;
406
407 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
408 CLOCK_REALTIME : CLOCK_MONOTONIC,
409 HRTIMER_MODE_ABS);
410 /*
411 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
412 * effectively the same as calling hrtimer_set_expires().
413 */
414 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
415
416 return timeout;
417}
418
419/*
420 * Generate a machine wide unique identifier for this inode.
421 *
422 * This relies on u64 not wrapping in the life-time of the machine; which with
423 * 1ns resolution means almost 585 years.
424 *
425 * This further relies on the fact that a well formed program will not unmap
426 * the file while it has a (shared) futex waiting on it. This mapping will have
427 * a file reference which pins the mount and inode.
428 *
429 * If for some reason an inode gets evicted and read back in again, it will get
430 * a new sequence number and will _NOT_ match, even though it is the exact same
431 * file.
432 *
433 * It is important that match_futex() will never have a false-positive, esp.
434 * for PI futexes that can mess up the state. The above argues that false-negatives
435 * are only possible for malformed programs.
436 */
437static u64 get_inode_sequence_number(struct inode *inode)
438{
439 static atomic64_t i_seq;
440 u64 old;
441
442 /* Does the inode already have a sequence number? */
443 old = atomic64_read(&inode->i_sequence);
444 if (likely(old))
445 return old;
446
447 for (;;) {
448 u64 new = atomic64_add_return(1, &i_seq);
449 if (WARN_ON_ONCE(!new))
450 continue;
451
452 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
453 if (old)
454 return old;
455 return new;
456 }
457}
458
459/**
460 * get_futex_key() - Get parameters which are the keys for a futex
461 * @uaddr: virtual address of the futex
462 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
463 * @key: address where result is stored.
464 * @rw: mapping needs to be read/write (values: FUTEX_READ,
465 * FUTEX_WRITE)
466 *
467 * Return: a negative error code or 0
468 *
469 * The key words are stored in @key on success.
470 *
471 * For shared mappings (when @fshared), the key is:
472 *
473 * ( inode->i_sequence, page->index, offset_within_page )
474 *
475 * [ also see get_inode_sequence_number() ]
476 *
477 * For private mappings (or when !@fshared), the key is:
478 *
479 * ( current->mm, address, 0 )
480 *
481 * This allows (cross process, where applicable) identification of the futex
482 * without keeping the page pinned for the duration of the FUTEX_WAIT.
483 *
484 * lock_page() might sleep, the caller should not hold a spinlock.
485 */
486static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
487 enum futex_access rw)
488{
489 unsigned long address = (unsigned long)uaddr;
490 struct mm_struct *mm = current->mm;
491 struct page *page, *tail;
492 struct address_space *mapping;
493 int err, ro = 0;
494
495 /*
496 * The futex address must be "naturally" aligned.
497 */
498 key->both.offset = address % PAGE_SIZE;
499 if (unlikely((address % sizeof(u32)) != 0))
500 return -EINVAL;
501 address -= key->both.offset;
502
503 if (unlikely(!access_ok(uaddr, sizeof(u32))))
504 return -EFAULT;
505
506 if (unlikely(should_fail_futex(fshared)))
507 return -EFAULT;
508
509 /*
510 * PROCESS_PRIVATE futexes are fast.
511 * As the mm cannot disappear under us and the 'key' only needs
512 * virtual address, we dont even have to find the underlying vma.
513 * Note : We do have to check 'uaddr' is a valid user address,
514 * but access_ok() should be faster than find_vma()
515 */
516 if (!fshared) {
517 key->private.mm = mm;
518 key->private.address = address;
519 return 0;
520 }
521
522again:
523 /* Ignore any VERIFY_READ mapping (futex common case) */
524 if (unlikely(should_fail_futex(true)))
525 return -EFAULT;
526
527 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
528 /*
529 * If write access is not required (eg. FUTEX_WAIT), try
530 * and get read-only access.
531 */
532 if (err == -EFAULT && rw == FUTEX_READ) {
533 err = get_user_pages_fast(address, 1, 0, &page);
534 ro = 1;
535 }
536 if (err < 0)
537 return err;
538 else
539 err = 0;
540
541 /*
542 * The treatment of mapping from this point on is critical. The page
543 * lock protects many things but in this context the page lock
544 * stabilizes mapping, prevents inode freeing in the shared
545 * file-backed region case and guards against movement to swap cache.
546 *
547 * Strictly speaking the page lock is not needed in all cases being
548 * considered here and page lock forces unnecessarily serialization
549 * From this point on, mapping will be re-verified if necessary and
550 * page lock will be acquired only if it is unavoidable
551 *
552 * Mapping checks require the head page for any compound page so the
553 * head page and mapping is looked up now. For anonymous pages, it
554 * does not matter if the page splits in the future as the key is
555 * based on the address. For filesystem-backed pages, the tail is
556 * required as the index of the page determines the key. For
557 * base pages, there is no tail page and tail == page.
558 */
559 tail = page;
560 page = compound_head(page);
561 mapping = READ_ONCE(page->mapping);
562
563 /*
564 * If page->mapping is NULL, then it cannot be a PageAnon
565 * page; but it might be the ZERO_PAGE or in the gate area or
566 * in a special mapping (all cases which we are happy to fail);
567 * or it may have been a good file page when get_user_pages_fast
568 * found it, but truncated or holepunched or subjected to
569 * invalidate_complete_page2 before we got the page lock (also
570 * cases which we are happy to fail). And we hold a reference,
571 * so refcount care in invalidate_complete_page's remove_mapping
572 * prevents drop_caches from setting mapping to NULL beneath us.
573 *
574 * The case we do have to guard against is when memory pressure made
575 * shmem_writepage move it from filecache to swapcache beneath us:
576 * an unlikely race, but we do need to retry for page->mapping.
577 */
578 if (unlikely(!mapping)) {
579 int shmem_swizzled;
580
581 /*
582 * Page lock is required to identify which special case above
583 * applies. If this is really a shmem page then the page lock
584 * will prevent unexpected transitions.
585 */
586 lock_page(page);
587 shmem_swizzled = PageSwapCache(page) || page->mapping;
588 unlock_page(page);
589 put_page(page);
590
591 if (shmem_swizzled)
592 goto again;
593
594 return -EFAULT;
595 }
596
597 /*
598 * Private mappings are handled in a simple way.
599 *
600 * If the futex key is stored on an anonymous page, then the associated
601 * object is the mm which is implicitly pinned by the calling process.
602 *
603 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
604 * it's a read-only handle, it's expected that futexes attach to
605 * the object not the particular process.
606 */
607 if (PageAnon(page)) {
608 /*
609 * A RO anonymous page will never change and thus doesn't make
610 * sense for futex operations.
611 */
612 if (unlikely(should_fail_futex(true)) || ro) {
613 err = -EFAULT;
614 goto out;
615 }
616
617 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
618 key->private.mm = mm;
619 key->private.address = address;
620
621 } else {
622 struct inode *inode;
623
624 /*
625 * The associated futex object in this case is the inode and
626 * the page->mapping must be traversed. Ordinarily this should
627 * be stabilised under page lock but it's not strictly
628 * necessary in this case as we just want to pin the inode, not
629 * update the radix tree or anything like that.
630 *
631 * The RCU read lock is taken as the inode is finally freed
632 * under RCU. If the mapping still matches expectations then the
633 * mapping->host can be safely accessed as being a valid inode.
634 */
635 rcu_read_lock();
636
637 if (READ_ONCE(page->mapping) != mapping) {
638 rcu_read_unlock();
639 put_page(page);
640
641 goto again;
642 }
643
644 inode = READ_ONCE(mapping->host);
645 if (!inode) {
646 rcu_read_unlock();
647 put_page(page);
648
649 goto again;
650 }
651
652 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
653 key->shared.i_seq = get_inode_sequence_number(inode);
654 key->shared.pgoff = basepage_index(tail);
655 rcu_read_unlock();
656 }
657
658out:
659 put_page(page);
660 return err;
661}
662
663/**
664 * fault_in_user_writeable() - Fault in user address and verify RW access
665 * @uaddr: pointer to faulting user space address
666 *
667 * Slow path to fixup the fault we just took in the atomic write
668 * access to @uaddr.
669 *
670 * We have no generic implementation of a non-destructive write to the
671 * user address. We know that we faulted in the atomic pagefault
672 * disabled section so we can as well avoid the #PF overhead by
673 * calling get_user_pages() right away.
674 */
675static int fault_in_user_writeable(u32 __user *uaddr)
676{
677 struct mm_struct *mm = current->mm;
678 int ret;
679
680 mmap_read_lock(mm);
681 ret = fixup_user_fault(mm, (unsigned long)uaddr,
682 FAULT_FLAG_WRITE, NULL);
683 mmap_read_unlock(mm);
684
685 return ret < 0 ? ret : 0;
686}
687
688/**
689 * futex_top_waiter() - Return the highest priority waiter on a futex
690 * @hb: the hash bucket the futex_q's reside in
691 * @key: the futex key (to distinguish it from other futex futex_q's)
692 *
693 * Must be called with the hb lock held.
694 */
695static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
696 union futex_key *key)
697{
698 struct futex_q *this;
699
700 plist_for_each_entry(this, &hb->chain, list) {
701 if (match_futex(&this->key, key))
702 return this;
703 }
704 return NULL;
705}
706
707static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
708 u32 uval, u32 newval)
709{
710 int ret;
711
712 pagefault_disable();
713 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
714 pagefault_enable();
715
716 return ret;
717}
718
719static int get_futex_value_locked(u32 *dest, u32 __user *from)
720{
721 int ret;
722
723 pagefault_disable();
724 ret = __get_user(*dest, from);
725 pagefault_enable();
726
727 return ret ? -EFAULT : 0;
728}
729
730
731/*
732 * PI code:
733 */
734static int refill_pi_state_cache(void)
735{
736 struct futex_pi_state *pi_state;
737
738 if (likely(current->pi_state_cache))
739 return 0;
740
741 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
742
743 if (!pi_state)
744 return -ENOMEM;
745
746 INIT_LIST_HEAD(&pi_state->list);
747 /* pi_mutex gets initialized later */
748 pi_state->owner = NULL;
749 refcount_set(&pi_state->refcount, 1);
750 pi_state->key = FUTEX_KEY_INIT;
751
752 current->pi_state_cache = pi_state;
753
754 return 0;
755}
756
757static struct futex_pi_state *alloc_pi_state(void)
758{
759 struct futex_pi_state *pi_state = current->pi_state_cache;
760
761 WARN_ON(!pi_state);
762 current->pi_state_cache = NULL;
763
764 return pi_state;
765}
766
767static void get_pi_state(struct futex_pi_state *pi_state)
768{
769 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
770}
771
772/*
773 * Drops a reference to the pi_state object and frees or caches it
774 * when the last reference is gone.
775 */
776static void put_pi_state(struct futex_pi_state *pi_state)
777{
778 if (!pi_state)
779 return;
780
781 if (!refcount_dec_and_test(&pi_state->refcount))
782 return;
783
784 /*
785 * If pi_state->owner is NULL, the owner is most probably dying
786 * and has cleaned up the pi_state already
787 */
788 if (pi_state->owner) {
789 struct task_struct *owner;
790
791 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
792 owner = pi_state->owner;
793 if (owner) {
794 raw_spin_lock(&owner->pi_lock);
795 list_del_init(&pi_state->list);
796 raw_spin_unlock(&owner->pi_lock);
797 }
798 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
799 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
800 }
801
802 if (current->pi_state_cache) {
803 kfree(pi_state);
804 } else {
805 /*
806 * pi_state->list is already empty.
807 * clear pi_state->owner.
808 * refcount is at 0 - put it back to 1.
809 */
810 pi_state->owner = NULL;
811 refcount_set(&pi_state->refcount, 1);
812 current->pi_state_cache = pi_state;
813 }
814}
815
816#ifdef CONFIG_FUTEX_PI
817
818/*
819 * This task is holding PI mutexes at exit time => bad.
820 * Kernel cleans up PI-state, but userspace is likely hosed.
821 * (Robust-futex cleanup is separate and might save the day for userspace.)
822 */
823static void exit_pi_state_list(struct task_struct *curr)
824{
825 struct list_head *next, *head = &curr->pi_state_list;
826 struct futex_pi_state *pi_state;
827 struct futex_hash_bucket *hb;
828 union futex_key key = FUTEX_KEY_INIT;
829
830 if (!futex_cmpxchg_enabled)
831 return;
832 /*
833 * We are a ZOMBIE and nobody can enqueue itself on
834 * pi_state_list anymore, but we have to be careful
835 * versus waiters unqueueing themselves:
836 */
837 raw_spin_lock_irq(&curr->pi_lock);
838 while (!list_empty(head)) {
839 next = head->next;
840 pi_state = list_entry(next, struct futex_pi_state, list);
841 key = pi_state->key;
842 hb = hash_futex(&key);
843
844 /*
845 * We can race against put_pi_state() removing itself from the
846 * list (a waiter going away). put_pi_state() will first
847 * decrement the reference count and then modify the list, so
848 * its possible to see the list entry but fail this reference
849 * acquire.
850 *
851 * In that case; drop the locks to let put_pi_state() make
852 * progress and retry the loop.
853 */
854 if (!refcount_inc_not_zero(&pi_state->refcount)) {
855 raw_spin_unlock_irq(&curr->pi_lock);
856 cpu_relax();
857 raw_spin_lock_irq(&curr->pi_lock);
858 continue;
859 }
860 raw_spin_unlock_irq(&curr->pi_lock);
861
862 spin_lock(&hb->lock);
863 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
864 raw_spin_lock(&curr->pi_lock);
865 /*
866 * We dropped the pi-lock, so re-check whether this
867 * task still owns the PI-state:
868 */
869 if (head->next != next) {
870 /* retain curr->pi_lock for the loop invariant */
871 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
872 spin_unlock(&hb->lock);
873 put_pi_state(pi_state);
874 continue;
875 }
876
877 WARN_ON(pi_state->owner != curr);
878 WARN_ON(list_empty(&pi_state->list));
879 list_del_init(&pi_state->list);
880 pi_state->owner = NULL;
881
882 raw_spin_unlock(&curr->pi_lock);
883 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
884 spin_unlock(&hb->lock);
885
886 rt_mutex_futex_unlock(&pi_state->pi_mutex);
887 put_pi_state(pi_state);
888
889 raw_spin_lock_irq(&curr->pi_lock);
890 }
891 raw_spin_unlock_irq(&curr->pi_lock);
892}
893#else
894static inline void exit_pi_state_list(struct task_struct *curr) { }
895#endif
896
897/*
898 * We need to check the following states:
899 *
900 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
901 *
902 * [1] NULL | --- | --- | 0 | 0/1 | Valid
903 * [2] NULL | --- | --- | >0 | 0/1 | Valid
904 *
905 * [3] Found | NULL | -- | Any | 0/1 | Invalid
906 *
907 * [4] Found | Found | NULL | 0 | 1 | Valid
908 * [5] Found | Found | NULL | >0 | 1 | Invalid
909 *
910 * [6] Found | Found | task | 0 | 1 | Valid
911 *
912 * [7] Found | Found | NULL | Any | 0 | Invalid
913 *
914 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
915 * [9] Found | Found | task | 0 | 0 | Invalid
916 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
917 *
918 * [1] Indicates that the kernel can acquire the futex atomically. We
919 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
920 *
921 * [2] Valid, if TID does not belong to a kernel thread. If no matching
922 * thread is found then it indicates that the owner TID has died.
923 *
924 * [3] Invalid. The waiter is queued on a non PI futex
925 *
926 * [4] Valid state after exit_robust_list(), which sets the user space
927 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
928 *
929 * [5] The user space value got manipulated between exit_robust_list()
930 * and exit_pi_state_list()
931 *
932 * [6] Valid state after exit_pi_state_list() which sets the new owner in
933 * the pi_state but cannot access the user space value.
934 *
935 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
936 *
937 * [8] Owner and user space value match
938 *
939 * [9] There is no transient state which sets the user space TID to 0
940 * except exit_robust_list(), but this is indicated by the
941 * FUTEX_OWNER_DIED bit. See [4]
942 *
943 * [10] There is no transient state which leaves owner and user space
944 * TID out of sync.
945 *
946 *
947 * Serialization and lifetime rules:
948 *
949 * hb->lock:
950 *
951 * hb -> futex_q, relation
952 * futex_q -> pi_state, relation
953 *
954 * (cannot be raw because hb can contain arbitrary amount
955 * of futex_q's)
956 *
957 * pi_mutex->wait_lock:
958 *
959 * {uval, pi_state}
960 *
961 * (and pi_mutex 'obviously')
962 *
963 * p->pi_lock:
964 *
965 * p->pi_state_list -> pi_state->list, relation
966 *
967 * pi_state->refcount:
968 *
969 * pi_state lifetime
970 *
971 *
972 * Lock order:
973 *
974 * hb->lock
975 * pi_mutex->wait_lock
976 * p->pi_lock
977 *
978 */
979
980/*
981 * Validate that the existing waiter has a pi_state and sanity check
982 * the pi_state against the user space value. If correct, attach to
983 * it.
984 */
985static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
986 struct futex_pi_state *pi_state,
987 struct futex_pi_state **ps)
988{
989 pid_t pid = uval & FUTEX_TID_MASK;
990 u32 uval2;
991 int ret;
992
993 /*
994 * Userspace might have messed up non-PI and PI futexes [3]
995 */
996 if (unlikely(!pi_state))
997 return -EINVAL;
998
999 /*
1000 * We get here with hb->lock held, and having found a
1001 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1002 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1003 * which in turn means that futex_lock_pi() still has a reference on
1004 * our pi_state.
1005 *
1006 * The waiter holding a reference on @pi_state also protects against
1007 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1008 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1009 * free pi_state before we can take a reference ourselves.
1010 */
1011 WARN_ON(!refcount_read(&pi_state->refcount));
1012
1013 /*
1014 * Now that we have a pi_state, we can acquire wait_lock
1015 * and do the state validation.
1016 */
1017 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1018
1019 /*
1020 * Since {uval, pi_state} is serialized by wait_lock, and our current
1021 * uval was read without holding it, it can have changed. Verify it
1022 * still is what we expect it to be, otherwise retry the entire
1023 * operation.
1024 */
1025 if (get_futex_value_locked(&uval2, uaddr))
1026 goto out_efault;
1027
1028 if (uval != uval2)
1029 goto out_eagain;
1030
1031 /*
1032 * Handle the owner died case:
1033 */
1034 if (uval & FUTEX_OWNER_DIED) {
1035 /*
1036 * exit_pi_state_list sets owner to NULL and wakes the
1037 * topmost waiter. The task which acquires the
1038 * pi_state->rt_mutex will fixup owner.
1039 */
1040 if (!pi_state->owner) {
1041 /*
1042 * No pi state owner, but the user space TID
1043 * is not 0. Inconsistent state. [5]
1044 */
1045 if (pid)
1046 goto out_einval;
1047 /*
1048 * Take a ref on the state and return success. [4]
1049 */
1050 goto out_attach;
1051 }
1052
1053 /*
1054 * If TID is 0, then either the dying owner has not
1055 * yet executed exit_pi_state_list() or some waiter
1056 * acquired the rtmutex in the pi state, but did not
1057 * yet fixup the TID in user space.
1058 *
1059 * Take a ref on the state and return success. [6]
1060 */
1061 if (!pid)
1062 goto out_attach;
1063 } else {
1064 /*
1065 * If the owner died bit is not set, then the pi_state
1066 * must have an owner. [7]
1067 */
1068 if (!pi_state->owner)
1069 goto out_einval;
1070 }
1071
1072 /*
1073 * Bail out if user space manipulated the futex value. If pi
1074 * state exists then the owner TID must be the same as the
1075 * user space TID. [9/10]
1076 */
1077 if (pid != task_pid_vnr(pi_state->owner))
1078 goto out_einval;
1079
1080out_attach:
1081 get_pi_state(pi_state);
1082 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1083 *ps = pi_state;
1084 return 0;
1085
1086out_einval:
1087 ret = -EINVAL;
1088 goto out_error;
1089
1090out_eagain:
1091 ret = -EAGAIN;
1092 goto out_error;
1093
1094out_efault:
1095 ret = -EFAULT;
1096 goto out_error;
1097
1098out_error:
1099 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1100 return ret;
1101}
1102
1103/**
1104 * wait_for_owner_exiting - Block until the owner has exited
1105 * @ret: owner's current futex lock status
1106 * @exiting: Pointer to the exiting task
1107 *
1108 * Caller must hold a refcount on @exiting.
1109 */
1110static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1111{
1112 if (ret != -EBUSY) {
1113 WARN_ON_ONCE(exiting);
1114 return;
1115 }
1116
1117 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1118 return;
1119
1120 mutex_lock(&exiting->futex_exit_mutex);
1121 /*
1122 * No point in doing state checking here. If the waiter got here
1123 * while the task was in exec()->exec_futex_release() then it can
1124 * have any FUTEX_STATE_* value when the waiter has acquired the
1125 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1126 * already. Highly unlikely and not a problem. Just one more round
1127 * through the futex maze.
1128 */
1129 mutex_unlock(&exiting->futex_exit_mutex);
1130
1131 put_task_struct(exiting);
1132}
1133
1134static int handle_exit_race(u32 __user *uaddr, u32 uval,
1135 struct task_struct *tsk)
1136{
1137 u32 uval2;
1138
1139 /*
1140 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1141 * caller that the alleged owner is busy.
1142 */
1143 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1144 return -EBUSY;
1145
1146 /*
1147 * Reread the user space value to handle the following situation:
1148 *
1149 * CPU0 CPU1
1150 *
1151 * sys_exit() sys_futex()
1152 * do_exit() futex_lock_pi()
1153 * futex_lock_pi_atomic()
1154 * exit_signals(tsk) No waiters:
1155 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1156 * mm_release(tsk) Set waiter bit
1157 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1158 * Set owner died attach_to_pi_owner() {
1159 * *uaddr = 0xC0000000; tsk = get_task(PID);
1160 * } if (!tsk->flags & PF_EXITING) {
1161 * ... attach();
1162 * tsk->futex_state = } else {
1163 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1164 * FUTEX_STATE_DEAD)
1165 * return -EAGAIN;
1166 * return -ESRCH; <--- FAIL
1167 * }
1168 *
1169 * Returning ESRCH unconditionally is wrong here because the
1170 * user space value has been changed by the exiting task.
1171 *
1172 * The same logic applies to the case where the exiting task is
1173 * already gone.
1174 */
1175 if (get_futex_value_locked(&uval2, uaddr))
1176 return -EFAULT;
1177
1178 /* If the user space value has changed, try again. */
1179 if (uval2 != uval)
1180 return -EAGAIN;
1181
1182 /*
1183 * The exiting task did not have a robust list, the robust list was
1184 * corrupted or the user space value in *uaddr is simply bogus.
1185 * Give up and tell user space.
1186 */
1187 return -ESRCH;
1188}
1189
1190/*
1191 * Lookup the task for the TID provided from user space and attach to
1192 * it after doing proper sanity checks.
1193 */
1194static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1195 struct futex_pi_state **ps,
1196 struct task_struct **exiting)
1197{
1198 pid_t pid = uval & FUTEX_TID_MASK;
1199 struct futex_pi_state *pi_state;
1200 struct task_struct *p;
1201
1202 /*
1203 * We are the first waiter - try to look up the real owner and attach
1204 * the new pi_state to it, but bail out when TID = 0 [1]
1205 *
1206 * The !pid check is paranoid. None of the call sites should end up
1207 * with pid == 0, but better safe than sorry. Let the caller retry
1208 */
1209 if (!pid)
1210 return -EAGAIN;
1211 p = find_get_task_by_vpid(pid);
1212 if (!p)
1213 return handle_exit_race(uaddr, uval, NULL);
1214
1215 if (unlikely(p->flags & PF_KTHREAD)) {
1216 put_task_struct(p);
1217 return -EPERM;
1218 }
1219
1220 /*
1221 * We need to look at the task state to figure out, whether the
1222 * task is exiting. To protect against the change of the task state
1223 * in futex_exit_release(), we do this protected by p->pi_lock:
1224 */
1225 raw_spin_lock_irq(&p->pi_lock);
1226 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1227 /*
1228 * The task is on the way out. When the futex state is
1229 * FUTEX_STATE_DEAD, we know that the task has finished
1230 * the cleanup:
1231 */
1232 int ret = handle_exit_race(uaddr, uval, p);
1233
1234 raw_spin_unlock_irq(&p->pi_lock);
1235 /*
1236 * If the owner task is between FUTEX_STATE_EXITING and
1237 * FUTEX_STATE_DEAD then store the task pointer and keep
1238 * the reference on the task struct. The calling code will
1239 * drop all locks, wait for the task to reach
1240 * FUTEX_STATE_DEAD and then drop the refcount. This is
1241 * required to prevent a live lock when the current task
1242 * preempted the exiting task between the two states.
1243 */
1244 if (ret == -EBUSY)
1245 *exiting = p;
1246 else
1247 put_task_struct(p);
1248 return ret;
1249 }
1250
1251 /*
1252 * No existing pi state. First waiter. [2]
1253 *
1254 * This creates pi_state, we have hb->lock held, this means nothing can
1255 * observe this state, wait_lock is irrelevant.
1256 */
1257 pi_state = alloc_pi_state();
1258
1259 /*
1260 * Initialize the pi_mutex in locked state and make @p
1261 * the owner of it:
1262 */
1263 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1264
1265 /* Store the key for possible exit cleanups: */
1266 pi_state->key = *key;
1267
1268 WARN_ON(!list_empty(&pi_state->list));
1269 list_add(&pi_state->list, &p->pi_state_list);
1270 /*
1271 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1272 * because there is no concurrency as the object is not published yet.
1273 */
1274 pi_state->owner = p;
1275 raw_spin_unlock_irq(&p->pi_lock);
1276
1277 put_task_struct(p);
1278
1279 *ps = pi_state;
1280
1281 return 0;
1282}
1283
1284static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1285 struct futex_hash_bucket *hb,
1286 union futex_key *key, struct futex_pi_state **ps,
1287 struct task_struct **exiting)
1288{
1289 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1290
1291 /*
1292 * If there is a waiter on that futex, validate it and
1293 * attach to the pi_state when the validation succeeds.
1294 */
1295 if (top_waiter)
1296 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1297
1298 /*
1299 * We are the first waiter - try to look up the owner based on
1300 * @uval and attach to it.
1301 */
1302 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1303}
1304
1305static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1306{
1307 int err;
1308 u32 curval;
1309
1310 if (unlikely(should_fail_futex(true)))
1311 return -EFAULT;
1312
1313 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1314 if (unlikely(err))
1315 return err;
1316
1317 /* If user space value changed, let the caller retry */
1318 return curval != uval ? -EAGAIN : 0;
1319}
1320
1321/**
1322 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1323 * @uaddr: the pi futex user address
1324 * @hb: the pi futex hash bucket
1325 * @key: the futex key associated with uaddr and hb
1326 * @ps: the pi_state pointer where we store the result of the
1327 * lookup
1328 * @task: the task to perform the atomic lock work for. This will
1329 * be "current" except in the case of requeue pi.
1330 * @exiting: Pointer to store the task pointer of the owner task
1331 * which is in the middle of exiting
1332 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1333 *
1334 * Return:
1335 * - 0 - ready to wait;
1336 * - 1 - acquired the lock;
1337 * - <0 - error
1338 *
1339 * The hb->lock and futex_key refs shall be held by the caller.
1340 *
1341 * @exiting is only set when the return value is -EBUSY. If so, this holds
1342 * a refcount on the exiting task on return and the caller needs to drop it
1343 * after waiting for the exit to complete.
1344 */
1345static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1346 union futex_key *key,
1347 struct futex_pi_state **ps,
1348 struct task_struct *task,
1349 struct task_struct **exiting,
1350 int set_waiters)
1351{
1352 u32 uval, newval, vpid = task_pid_vnr(task);
1353 struct futex_q *top_waiter;
1354 int ret;
1355
1356 /*
1357 * Read the user space value first so we can validate a few
1358 * things before proceeding further.
1359 */
1360 if (get_futex_value_locked(&uval, uaddr))
1361 return -EFAULT;
1362
1363 if (unlikely(should_fail_futex(true)))
1364 return -EFAULT;
1365
1366 /*
1367 * Detect deadlocks.
1368 */
1369 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1370 return -EDEADLK;
1371
1372 if ((unlikely(should_fail_futex(true))))
1373 return -EDEADLK;
1374
1375 /*
1376 * Lookup existing state first. If it exists, try to attach to
1377 * its pi_state.
1378 */
1379 top_waiter = futex_top_waiter(hb, key);
1380 if (top_waiter)
1381 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1382
1383 /*
1384 * No waiter and user TID is 0. We are here because the
1385 * waiters or the owner died bit is set or called from
1386 * requeue_cmp_pi or for whatever reason something took the
1387 * syscall.
1388 */
1389 if (!(uval & FUTEX_TID_MASK)) {
1390 /*
1391 * We take over the futex. No other waiters and the user space
1392 * TID is 0. We preserve the owner died bit.
1393 */
1394 newval = uval & FUTEX_OWNER_DIED;
1395 newval |= vpid;
1396
1397 /* The futex requeue_pi code can enforce the waiters bit */
1398 if (set_waiters)
1399 newval |= FUTEX_WAITERS;
1400
1401 ret = lock_pi_update_atomic(uaddr, uval, newval);
1402 /* If the take over worked, return 1 */
1403 return ret < 0 ? ret : 1;
1404 }
1405
1406 /*
1407 * First waiter. Set the waiters bit before attaching ourself to
1408 * the owner. If owner tries to unlock, it will be forced into
1409 * the kernel and blocked on hb->lock.
1410 */
1411 newval = uval | FUTEX_WAITERS;
1412 ret = lock_pi_update_atomic(uaddr, uval, newval);
1413 if (ret)
1414 return ret;
1415 /*
1416 * If the update of the user space value succeeded, we try to
1417 * attach to the owner. If that fails, no harm done, we only
1418 * set the FUTEX_WAITERS bit in the user space variable.
1419 */
1420 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1421}
1422
1423/**
1424 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1425 * @q: The futex_q to unqueue
1426 *
1427 * The q->lock_ptr must not be NULL and must be held by the caller.
1428 */
1429static void __unqueue_futex(struct futex_q *q)
1430{
1431 struct futex_hash_bucket *hb;
1432
1433 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1434 return;
1435 lockdep_assert_held(q->lock_ptr);
1436
1437 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1438 plist_del(&q->list, &hb->chain);
1439 hb_waiters_dec(hb);
1440}
1441
1442/*
1443 * The hash bucket lock must be held when this is called.
1444 * Afterwards, the futex_q must not be accessed. Callers
1445 * must ensure to later call wake_up_q() for the actual
1446 * wakeups to occur.
1447 */
1448static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1449{
1450 struct task_struct *p = q->task;
1451
1452 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1453 return;
1454
1455 get_task_struct(p);
1456 __unqueue_futex(q);
1457 /*
1458 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1459 * is written, without taking any locks. This is possible in the event
1460 * of a spurious wakeup, for example. A memory barrier is required here
1461 * to prevent the following store to lock_ptr from getting ahead of the
1462 * plist_del in __unqueue_futex().
1463 */
1464 smp_store_release(&q->lock_ptr, NULL);
1465
1466 /*
1467 * Queue the task for later wakeup for after we've released
1468 * the hb->lock.
1469 */
1470 wake_q_add_safe(wake_q, p);
1471}
1472
1473/*
1474 * Caller must hold a reference on @pi_state.
1475 */
1476static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1477{
1478 u32 curval, newval;
1479 struct task_struct *new_owner;
1480 bool postunlock = false;
1481 DEFINE_WAKE_Q(wake_q);
1482 int ret = 0;
1483
1484 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1485 if (WARN_ON_ONCE(!new_owner)) {
1486 /*
1487 * As per the comment in futex_unlock_pi() this should not happen.
1488 *
1489 * When this happens, give up our locks and try again, giving
1490 * the futex_lock_pi() instance time to complete, either by
1491 * waiting on the rtmutex or removing itself from the futex
1492 * queue.
1493 */
1494 ret = -EAGAIN;
1495 goto out_unlock;
1496 }
1497
1498 /*
1499 * We pass it to the next owner. The WAITERS bit is always kept
1500 * enabled while there is PI state around. We cleanup the owner
1501 * died bit, because we are the owner.
1502 */
1503 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1504
1505 if (unlikely(should_fail_futex(true)))
1506 ret = -EFAULT;
1507
1508 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1509 if (!ret && (curval != uval)) {
1510 /*
1511 * If a unconditional UNLOCK_PI operation (user space did not
1512 * try the TID->0 transition) raced with a waiter setting the
1513 * FUTEX_WAITERS flag between get_user() and locking the hash
1514 * bucket lock, retry the operation.
1515 */
1516 if ((FUTEX_TID_MASK & curval) == uval)
1517 ret = -EAGAIN;
1518 else
1519 ret = -EINVAL;
1520 }
1521
1522 if (ret)
1523 goto out_unlock;
1524
1525 /*
1526 * This is a point of no return; once we modify the uval there is no
1527 * going back and subsequent operations must not fail.
1528 */
1529
1530 raw_spin_lock(&pi_state->owner->pi_lock);
1531 WARN_ON(list_empty(&pi_state->list));
1532 list_del_init(&pi_state->list);
1533 raw_spin_unlock(&pi_state->owner->pi_lock);
1534
1535 raw_spin_lock(&new_owner->pi_lock);
1536 WARN_ON(!list_empty(&pi_state->list));
1537 list_add(&pi_state->list, &new_owner->pi_state_list);
1538 pi_state->owner = new_owner;
1539 raw_spin_unlock(&new_owner->pi_lock);
1540
1541 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1542
1543out_unlock:
1544 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1545
1546 if (postunlock)
1547 rt_mutex_postunlock(&wake_q);
1548
1549 return ret;
1550}
1551
1552/*
1553 * Express the locking dependencies for lockdep:
1554 */
1555static inline void
1556double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1557{
1558 if (hb1 <= hb2) {
1559 spin_lock(&hb1->lock);
1560 if (hb1 < hb2)
1561 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1562 } else { /* hb1 > hb2 */
1563 spin_lock(&hb2->lock);
1564 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1565 }
1566}
1567
1568static inline void
1569double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1570{
1571 spin_unlock(&hb1->lock);
1572 if (hb1 != hb2)
1573 spin_unlock(&hb2->lock);
1574}
1575
1576/*
1577 * Wake up waiters matching bitset queued on this futex (uaddr).
1578 */
1579static int
1580futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1581{
1582 struct futex_hash_bucket *hb;
1583 struct futex_q *this, *next;
1584 union futex_key key = FUTEX_KEY_INIT;
1585 int ret;
1586 DEFINE_WAKE_Q(wake_q);
1587
1588 if (!bitset)
1589 return -EINVAL;
1590
1591 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1592 if (unlikely(ret != 0))
1593 return ret;
1594
1595 hb = hash_futex(&key);
1596
1597 /* Make sure we really have tasks to wakeup */
1598 if (!hb_waiters_pending(hb))
1599 return ret;
1600
1601 spin_lock(&hb->lock);
1602
1603 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1604 if (match_futex (&this->key, &key)) {
1605 if (this->pi_state || this->rt_waiter) {
1606 ret = -EINVAL;
1607 break;
1608 }
1609
1610 /* Check if one of the bits is set in both bitsets */
1611 if (!(this->bitset & bitset))
1612 continue;
1613
1614 mark_wake_futex(&wake_q, this);
1615 if (++ret >= nr_wake)
1616 break;
1617 }
1618 }
1619
1620 spin_unlock(&hb->lock);
1621 wake_up_q(&wake_q);
1622 return ret;
1623}
1624
1625static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1626{
1627 unsigned int op = (encoded_op & 0x70000000) >> 28;
1628 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1629 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1630 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1631 int oldval, ret;
1632
1633 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1634 if (oparg < 0 || oparg > 31) {
1635 char comm[sizeof(current->comm)];
1636 /*
1637 * kill this print and return -EINVAL when userspace
1638 * is sane again
1639 */
1640 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1641 get_task_comm(comm, current), oparg);
1642 oparg &= 31;
1643 }
1644 oparg = 1 << oparg;
1645 }
1646
1647 pagefault_disable();
1648 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1649 pagefault_enable();
1650 if (ret)
1651 return ret;
1652
1653 switch (cmp) {
1654 case FUTEX_OP_CMP_EQ:
1655 return oldval == cmparg;
1656 case FUTEX_OP_CMP_NE:
1657 return oldval != cmparg;
1658 case FUTEX_OP_CMP_LT:
1659 return oldval < cmparg;
1660 case FUTEX_OP_CMP_GE:
1661 return oldval >= cmparg;
1662 case FUTEX_OP_CMP_LE:
1663 return oldval <= cmparg;
1664 case FUTEX_OP_CMP_GT:
1665 return oldval > cmparg;
1666 default:
1667 return -ENOSYS;
1668 }
1669}
1670
1671/*
1672 * Wake up all waiters hashed on the physical page that is mapped
1673 * to this virtual address:
1674 */
1675static int
1676futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1677 int nr_wake, int nr_wake2, int op)
1678{
1679 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1680 struct futex_hash_bucket *hb1, *hb2;
1681 struct futex_q *this, *next;
1682 int ret, op_ret;
1683 DEFINE_WAKE_Q(wake_q);
1684
1685retry:
1686 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1687 if (unlikely(ret != 0))
1688 return ret;
1689 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1690 if (unlikely(ret != 0))
1691 return ret;
1692
1693 hb1 = hash_futex(&key1);
1694 hb2 = hash_futex(&key2);
1695
1696retry_private:
1697 double_lock_hb(hb1, hb2);
1698 op_ret = futex_atomic_op_inuser(op, uaddr2);
1699 if (unlikely(op_ret < 0)) {
1700 double_unlock_hb(hb1, hb2);
1701
1702 if (!IS_ENABLED(CONFIG_MMU) ||
1703 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1704 /*
1705 * we don't get EFAULT from MMU faults if we don't have
1706 * an MMU, but we might get them from range checking
1707 */
1708 ret = op_ret;
1709 return ret;
1710 }
1711
1712 if (op_ret == -EFAULT) {
1713 ret = fault_in_user_writeable(uaddr2);
1714 if (ret)
1715 return ret;
1716 }
1717
1718 if (!(flags & FLAGS_SHARED)) {
1719 cond_resched();
1720 goto retry_private;
1721 }
1722
1723 cond_resched();
1724 goto retry;
1725 }
1726
1727 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1728 if (match_futex (&this->key, &key1)) {
1729 if (this->pi_state || this->rt_waiter) {
1730 ret = -EINVAL;
1731 goto out_unlock;
1732 }
1733 mark_wake_futex(&wake_q, this);
1734 if (++ret >= nr_wake)
1735 break;
1736 }
1737 }
1738
1739 if (op_ret > 0) {
1740 op_ret = 0;
1741 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1742 if (match_futex (&this->key, &key2)) {
1743 if (this->pi_state || this->rt_waiter) {
1744 ret = -EINVAL;
1745 goto out_unlock;
1746 }
1747 mark_wake_futex(&wake_q, this);
1748 if (++op_ret >= nr_wake2)
1749 break;
1750 }
1751 }
1752 ret += op_ret;
1753 }
1754
1755out_unlock:
1756 double_unlock_hb(hb1, hb2);
1757 wake_up_q(&wake_q);
1758 return ret;
1759}
1760
1761/**
1762 * requeue_futex() - Requeue a futex_q from one hb to another
1763 * @q: the futex_q to requeue
1764 * @hb1: the source hash_bucket
1765 * @hb2: the target hash_bucket
1766 * @key2: the new key for the requeued futex_q
1767 */
1768static inline
1769void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1770 struct futex_hash_bucket *hb2, union futex_key *key2)
1771{
1772
1773 /*
1774 * If key1 and key2 hash to the same bucket, no need to
1775 * requeue.
1776 */
1777 if (likely(&hb1->chain != &hb2->chain)) {
1778 plist_del(&q->list, &hb1->chain);
1779 hb_waiters_dec(hb1);
1780 hb_waiters_inc(hb2);
1781 plist_add(&q->list, &hb2->chain);
1782 q->lock_ptr = &hb2->lock;
1783 }
1784 q->key = *key2;
1785}
1786
1787/**
1788 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1789 * @q: the futex_q
1790 * @key: the key of the requeue target futex
1791 * @hb: the hash_bucket of the requeue target futex
1792 *
1793 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1794 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1795 * to the requeue target futex so the waiter can detect the wakeup on the right
1796 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1797 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1798 * to protect access to the pi_state to fixup the owner later. Must be called
1799 * with both q->lock_ptr and hb->lock held.
1800 */
1801static inline
1802void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1803 struct futex_hash_bucket *hb)
1804{
1805 q->key = *key;
1806
1807 __unqueue_futex(q);
1808
1809 WARN_ON(!q->rt_waiter);
1810 q->rt_waiter = NULL;
1811
1812 q->lock_ptr = &hb->lock;
1813
1814 wake_up_state(q->task, TASK_NORMAL);
1815}
1816
1817/**
1818 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1819 * @pifutex: the user address of the to futex
1820 * @hb1: the from futex hash bucket, must be locked by the caller
1821 * @hb2: the to futex hash bucket, must be locked by the caller
1822 * @key1: the from futex key
1823 * @key2: the to futex key
1824 * @ps: address to store the pi_state pointer
1825 * @exiting: Pointer to store the task pointer of the owner task
1826 * which is in the middle of exiting
1827 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1828 *
1829 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1830 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1831 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1832 * hb1 and hb2 must be held by the caller.
1833 *
1834 * @exiting is only set when the return value is -EBUSY. If so, this holds
1835 * a refcount on the exiting task on return and the caller needs to drop it
1836 * after waiting for the exit to complete.
1837 *
1838 * Return:
1839 * - 0 - failed to acquire the lock atomically;
1840 * - >0 - acquired the lock, return value is vpid of the top_waiter
1841 * - <0 - error
1842 */
1843static int
1844futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1845 struct futex_hash_bucket *hb2, union futex_key *key1,
1846 union futex_key *key2, struct futex_pi_state **ps,
1847 struct task_struct **exiting, int set_waiters)
1848{
1849 struct futex_q *top_waiter = NULL;
1850 u32 curval;
1851 int ret, vpid;
1852
1853 if (get_futex_value_locked(&curval, pifutex))
1854 return -EFAULT;
1855
1856 if (unlikely(should_fail_futex(true)))
1857 return -EFAULT;
1858
1859 /*
1860 * Find the top_waiter and determine if there are additional waiters.
1861 * If the caller intends to requeue more than 1 waiter to pifutex,
1862 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1863 * as we have means to handle the possible fault. If not, don't set
1864 * the bit unecessarily as it will force the subsequent unlock to enter
1865 * the kernel.
1866 */
1867 top_waiter = futex_top_waiter(hb1, key1);
1868
1869 /* There are no waiters, nothing for us to do. */
1870 if (!top_waiter)
1871 return 0;
1872
1873 /* Ensure we requeue to the expected futex. */
1874 if (!match_futex(top_waiter->requeue_pi_key, key2))
1875 return -EINVAL;
1876
1877 /*
1878 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1879 * the contended case or if set_waiters is 1. The pi_state is returned
1880 * in ps in contended cases.
1881 */
1882 vpid = task_pid_vnr(top_waiter->task);
1883 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1884 exiting, set_waiters);
1885 if (ret == 1) {
1886 requeue_pi_wake_futex(top_waiter, key2, hb2);
1887 return vpid;
1888 }
1889 return ret;
1890}
1891
1892/**
1893 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1894 * @uaddr1: source futex user address
1895 * @flags: futex flags (FLAGS_SHARED, etc.)
1896 * @uaddr2: target futex user address
1897 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1898 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1899 * @cmpval: @uaddr1 expected value (or %NULL)
1900 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1901 * pi futex (pi to pi requeue is not supported)
1902 *
1903 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1904 * uaddr2 atomically on behalf of the top waiter.
1905 *
1906 * Return:
1907 * - >=0 - on success, the number of tasks requeued or woken;
1908 * - <0 - on error
1909 */
1910static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1911 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1912 u32 *cmpval, int requeue_pi)
1913{
1914 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1915 int task_count = 0, ret;
1916 struct futex_pi_state *pi_state = NULL;
1917 struct futex_hash_bucket *hb1, *hb2;
1918 struct futex_q *this, *next;
1919 DEFINE_WAKE_Q(wake_q);
1920
1921 if (nr_wake < 0 || nr_requeue < 0)
1922 return -EINVAL;
1923
1924 /*
1925 * When PI not supported: return -ENOSYS if requeue_pi is true,
1926 * consequently the compiler knows requeue_pi is always false past
1927 * this point which will optimize away all the conditional code
1928 * further down.
1929 */
1930 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1931 return -ENOSYS;
1932
1933 if (requeue_pi) {
1934 /*
1935 * Requeue PI only works on two distinct uaddrs. This
1936 * check is only valid for private futexes. See below.
1937 */
1938 if (uaddr1 == uaddr2)
1939 return -EINVAL;
1940
1941 /*
1942 * requeue_pi requires a pi_state, try to allocate it now
1943 * without any locks in case it fails.
1944 */
1945 if (refill_pi_state_cache())
1946 return -ENOMEM;
1947 /*
1948 * requeue_pi must wake as many tasks as it can, up to nr_wake
1949 * + nr_requeue, since it acquires the rt_mutex prior to
1950 * returning to userspace, so as to not leave the rt_mutex with
1951 * waiters and no owner. However, second and third wake-ups
1952 * cannot be predicted as they involve race conditions with the
1953 * first wake and a fault while looking up the pi_state. Both
1954 * pthread_cond_signal() and pthread_cond_broadcast() should
1955 * use nr_wake=1.
1956 */
1957 if (nr_wake != 1)
1958 return -EINVAL;
1959 }
1960
1961retry:
1962 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1963 if (unlikely(ret != 0))
1964 return ret;
1965 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1966 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1967 if (unlikely(ret != 0))
1968 return ret;
1969
1970 /*
1971 * The check above which compares uaddrs is not sufficient for
1972 * shared futexes. We need to compare the keys:
1973 */
1974 if (requeue_pi && match_futex(&key1, &key2))
1975 return -EINVAL;
1976
1977 hb1 = hash_futex(&key1);
1978 hb2 = hash_futex(&key2);
1979
1980retry_private:
1981 hb_waiters_inc(hb2);
1982 double_lock_hb(hb1, hb2);
1983
1984 if (likely(cmpval != NULL)) {
1985 u32 curval;
1986
1987 ret = get_futex_value_locked(&curval, uaddr1);
1988
1989 if (unlikely(ret)) {
1990 double_unlock_hb(hb1, hb2);
1991 hb_waiters_dec(hb2);
1992
1993 ret = get_user(curval, uaddr1);
1994 if (ret)
1995 return ret;
1996
1997 if (!(flags & FLAGS_SHARED))
1998 goto retry_private;
1999
2000 goto retry;
2001 }
2002 if (curval != *cmpval) {
2003 ret = -EAGAIN;
2004 goto out_unlock;
2005 }
2006 }
2007
2008 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2009 struct task_struct *exiting = NULL;
2010
2011 /*
2012 * Attempt to acquire uaddr2 and wake the top waiter. If we
2013 * intend to requeue waiters, force setting the FUTEX_WAITERS
2014 * bit. We force this here where we are able to easily handle
2015 * faults rather in the requeue loop below.
2016 */
2017 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2018 &key2, &pi_state,
2019 &exiting, nr_requeue);
2020
2021 /*
2022 * At this point the top_waiter has either taken uaddr2 or is
2023 * waiting on it. If the former, then the pi_state will not
2024 * exist yet, look it up one more time to ensure we have a
2025 * reference to it. If the lock was taken, ret contains the
2026 * vpid of the top waiter task.
2027 * If the lock was not taken, we have pi_state and an initial
2028 * refcount on it. In case of an error we have nothing.
2029 */
2030 if (ret > 0) {
2031 WARN_ON(pi_state);
2032 task_count++;
2033 /*
2034 * If we acquired the lock, then the user space value
2035 * of uaddr2 should be vpid. It cannot be changed by
2036 * the top waiter as it is blocked on hb2 lock if it
2037 * tries to do so. If something fiddled with it behind
2038 * our back the pi state lookup might unearth it. So
2039 * we rather use the known value than rereading and
2040 * handing potential crap to lookup_pi_state.
2041 *
2042 * If that call succeeds then we have pi_state and an
2043 * initial refcount on it.
2044 */
2045 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2046 &pi_state, &exiting);
2047 }
2048
2049 switch (ret) {
2050 case 0:
2051 /* We hold a reference on the pi state. */
2052 break;
2053
2054 /* If the above failed, then pi_state is NULL */
2055 case -EFAULT:
2056 double_unlock_hb(hb1, hb2);
2057 hb_waiters_dec(hb2);
2058 ret = fault_in_user_writeable(uaddr2);
2059 if (!ret)
2060 goto retry;
2061 return ret;
2062 case -EBUSY:
2063 case -EAGAIN:
2064 /*
2065 * Two reasons for this:
2066 * - EBUSY: Owner is exiting and we just wait for the
2067 * exit to complete.
2068 * - EAGAIN: The user space value changed.
2069 */
2070 double_unlock_hb(hb1, hb2);
2071 hb_waiters_dec(hb2);
2072 /*
2073 * Handle the case where the owner is in the middle of
2074 * exiting. Wait for the exit to complete otherwise
2075 * this task might loop forever, aka. live lock.
2076 */
2077 wait_for_owner_exiting(ret, exiting);
2078 cond_resched();
2079 goto retry;
2080 default:
2081 goto out_unlock;
2082 }
2083 }
2084
2085 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2086 if (task_count - nr_wake >= nr_requeue)
2087 break;
2088
2089 if (!match_futex(&this->key, &key1))
2090 continue;
2091
2092 /*
2093 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2094 * be paired with each other and no other futex ops.
2095 *
2096 * We should never be requeueing a futex_q with a pi_state,
2097 * which is awaiting a futex_unlock_pi().
2098 */
2099 if ((requeue_pi && !this->rt_waiter) ||
2100 (!requeue_pi && this->rt_waiter) ||
2101 this->pi_state) {
2102 ret = -EINVAL;
2103 break;
2104 }
2105
2106 /*
2107 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2108 * lock, we already woke the top_waiter. If not, it will be
2109 * woken by futex_unlock_pi().
2110 */
2111 if (++task_count <= nr_wake && !requeue_pi) {
2112 mark_wake_futex(&wake_q, this);
2113 continue;
2114 }
2115
2116 /* Ensure we requeue to the expected futex for requeue_pi. */
2117 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2118 ret = -EINVAL;
2119 break;
2120 }
2121
2122 /*
2123 * Requeue nr_requeue waiters and possibly one more in the case
2124 * of requeue_pi if we couldn't acquire the lock atomically.
2125 */
2126 if (requeue_pi) {
2127 /*
2128 * Prepare the waiter to take the rt_mutex. Take a
2129 * refcount on the pi_state and store the pointer in
2130 * the futex_q object of the waiter.
2131 */
2132 get_pi_state(pi_state);
2133 this->pi_state = pi_state;
2134 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2135 this->rt_waiter,
2136 this->task);
2137 if (ret == 1) {
2138 /*
2139 * We got the lock. We do neither drop the
2140 * refcount on pi_state nor clear
2141 * this->pi_state because the waiter needs the
2142 * pi_state for cleaning up the user space
2143 * value. It will drop the refcount after
2144 * doing so.
2145 */
2146 requeue_pi_wake_futex(this, &key2, hb2);
2147 continue;
2148 } else if (ret) {
2149 /*
2150 * rt_mutex_start_proxy_lock() detected a
2151 * potential deadlock when we tried to queue
2152 * that waiter. Drop the pi_state reference
2153 * which we took above and remove the pointer
2154 * to the state from the waiters futex_q
2155 * object.
2156 */
2157 this->pi_state = NULL;
2158 put_pi_state(pi_state);
2159 /*
2160 * We stop queueing more waiters and let user
2161 * space deal with the mess.
2162 */
2163 break;
2164 }
2165 }
2166 requeue_futex(this, hb1, hb2, &key2);
2167 }
2168
2169 /*
2170 * We took an extra initial reference to the pi_state either
2171 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2172 * need to drop it here again.
2173 */
2174 put_pi_state(pi_state);
2175
2176out_unlock:
2177 double_unlock_hb(hb1, hb2);
2178 wake_up_q(&wake_q);
2179 hb_waiters_dec(hb2);
2180 return ret ? ret : task_count;
2181}
2182
2183/* The key must be already stored in q->key. */
2184static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2185 __acquires(&hb->lock)
2186{
2187 struct futex_hash_bucket *hb;
2188
2189 hb = hash_futex(&q->key);
2190
2191 /*
2192 * Increment the counter before taking the lock so that
2193 * a potential waker won't miss a to-be-slept task that is
2194 * waiting for the spinlock. This is safe as all queue_lock()
2195 * users end up calling queue_me(). Similarly, for housekeeping,
2196 * decrement the counter at queue_unlock() when some error has
2197 * occurred and we don't end up adding the task to the list.
2198 */
2199 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2200
2201 q->lock_ptr = &hb->lock;
2202
2203 spin_lock(&hb->lock);
2204 return hb;
2205}
2206
2207static inline void
2208queue_unlock(struct futex_hash_bucket *hb)
2209 __releases(&hb->lock)
2210{
2211 spin_unlock(&hb->lock);
2212 hb_waiters_dec(hb);
2213}
2214
2215static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2216{
2217 int prio;
2218
2219 /*
2220 * The priority used to register this element is
2221 * - either the real thread-priority for the real-time threads
2222 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2223 * - or MAX_RT_PRIO for non-RT threads.
2224 * Thus, all RT-threads are woken first in priority order, and
2225 * the others are woken last, in FIFO order.
2226 */
2227 prio = min(current->normal_prio, MAX_RT_PRIO);
2228
2229 plist_node_init(&q->list, prio);
2230 plist_add(&q->list, &hb->chain);
2231 q->task = current;
2232}
2233
2234/**
2235 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2236 * @q: The futex_q to enqueue
2237 * @hb: The destination hash bucket
2238 *
2239 * The hb->lock must be held by the caller, and is released here. A call to
2240 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2241 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2242 * or nothing if the unqueue is done as part of the wake process and the unqueue
2243 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2244 * an example).
2245 */
2246static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2247 __releases(&hb->lock)
2248{
2249 __queue_me(q, hb);
2250 spin_unlock(&hb->lock);
2251}
2252
2253/**
2254 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2255 * @q: The futex_q to unqueue
2256 *
2257 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2258 * be paired with exactly one earlier call to queue_me().
2259 *
2260 * Return:
2261 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2262 * - 0 - if the futex_q was already removed by the waking thread
2263 */
2264static int unqueue_me(struct futex_q *q)
2265{
2266 spinlock_t *lock_ptr;
2267 int ret = 0;
2268
2269 /* In the common case we don't take the spinlock, which is nice. */
2270retry:
2271 /*
2272 * q->lock_ptr can change between this read and the following spin_lock.
2273 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2274 * optimizing lock_ptr out of the logic below.
2275 */
2276 lock_ptr = READ_ONCE(q->lock_ptr);
2277 if (lock_ptr != NULL) {
2278 spin_lock(lock_ptr);
2279 /*
2280 * q->lock_ptr can change between reading it and
2281 * spin_lock(), causing us to take the wrong lock. This
2282 * corrects the race condition.
2283 *
2284 * Reasoning goes like this: if we have the wrong lock,
2285 * q->lock_ptr must have changed (maybe several times)
2286 * between reading it and the spin_lock(). It can
2287 * change again after the spin_lock() but only if it was
2288 * already changed before the spin_lock(). It cannot,
2289 * however, change back to the original value. Therefore
2290 * we can detect whether we acquired the correct lock.
2291 */
2292 if (unlikely(lock_ptr != q->lock_ptr)) {
2293 spin_unlock(lock_ptr);
2294 goto retry;
2295 }
2296 __unqueue_futex(q);
2297
2298 BUG_ON(q->pi_state);
2299
2300 spin_unlock(lock_ptr);
2301 ret = 1;
2302 }
2303
2304 return ret;
2305}
2306
2307/*
2308 * PI futexes can not be requeued and must remove themself from the
2309 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2310 * and dropped here.
2311 */
2312static void unqueue_me_pi(struct futex_q *q)
2313 __releases(q->lock_ptr)
2314{
2315 __unqueue_futex(q);
2316
2317 BUG_ON(!q->pi_state);
2318 put_pi_state(q->pi_state);
2319 q->pi_state = NULL;
2320
2321 spin_unlock(q->lock_ptr);
2322}
2323
2324static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2325 struct task_struct *argowner)
2326{
2327 struct futex_pi_state *pi_state = q->pi_state;
2328 u32 uval, curval, newval;
2329 struct task_struct *oldowner, *newowner;
2330 u32 newtid;
2331 int ret, err = 0;
2332
2333 lockdep_assert_held(q->lock_ptr);
2334
2335 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2336
2337 oldowner = pi_state->owner;
2338
2339 /*
2340 * We are here because either:
2341 *
2342 * - we stole the lock and pi_state->owner needs updating to reflect
2343 * that (@argowner == current),
2344 *
2345 * or:
2346 *
2347 * - someone stole our lock and we need to fix things to point to the
2348 * new owner (@argowner == NULL).
2349 *
2350 * Either way, we have to replace the TID in the user space variable.
2351 * This must be atomic as we have to preserve the owner died bit here.
2352 *
2353 * Note: We write the user space value _before_ changing the pi_state
2354 * because we can fault here. Imagine swapped out pages or a fork
2355 * that marked all the anonymous memory readonly for cow.
2356 *
2357 * Modifying pi_state _before_ the user space value would leave the
2358 * pi_state in an inconsistent state when we fault here, because we
2359 * need to drop the locks to handle the fault. This might be observed
2360 * in the PID check in lookup_pi_state.
2361 */
2362retry:
2363 if (!argowner) {
2364 if (oldowner != current) {
2365 /*
2366 * We raced against a concurrent self; things are
2367 * already fixed up. Nothing to do.
2368 */
2369 ret = 0;
2370 goto out_unlock;
2371 }
2372
2373 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2374 /* We got the lock after all, nothing to fix. */
2375 ret = 0;
2376 goto out_unlock;
2377 }
2378
2379 /*
2380 * Since we just failed the trylock; there must be an owner.
2381 */
2382 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2383 BUG_ON(!newowner);
2384 } else {
2385 WARN_ON_ONCE(argowner != current);
2386 if (oldowner == current) {
2387 /*
2388 * We raced against a concurrent self; things are
2389 * already fixed up. Nothing to do.
2390 */
2391 ret = 0;
2392 goto out_unlock;
2393 }
2394 newowner = argowner;
2395 }
2396
2397 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2398 /* Owner died? */
2399 if (!pi_state->owner)
2400 newtid |= FUTEX_OWNER_DIED;
2401
2402 err = get_futex_value_locked(&uval, uaddr);
2403 if (err)
2404 goto handle_err;
2405
2406 for (;;) {
2407 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2408
2409 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2410 if (err)
2411 goto handle_err;
2412
2413 if (curval == uval)
2414 break;
2415 uval = curval;
2416 }
2417
2418 /*
2419 * We fixed up user space. Now we need to fix the pi_state
2420 * itself.
2421 */
2422 if (pi_state->owner != NULL) {
2423 raw_spin_lock(&pi_state->owner->pi_lock);
2424 WARN_ON(list_empty(&pi_state->list));
2425 list_del_init(&pi_state->list);
2426 raw_spin_unlock(&pi_state->owner->pi_lock);
2427 }
2428
2429 pi_state->owner = newowner;
2430
2431 raw_spin_lock(&newowner->pi_lock);
2432 WARN_ON(!list_empty(&pi_state->list));
2433 list_add(&pi_state->list, &newowner->pi_state_list);
2434 raw_spin_unlock(&newowner->pi_lock);
2435 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2436
2437 return 0;
2438
2439 /*
2440 * In order to reschedule or handle a page fault, we need to drop the
2441 * locks here. In the case of a fault, this gives the other task
2442 * (either the highest priority waiter itself or the task which stole
2443 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2444 * are back from handling the fault we need to check the pi_state after
2445 * reacquiring the locks and before trying to do another fixup. When
2446 * the fixup has been done already we simply return.
2447 *
2448 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2449 * drop hb->lock since the caller owns the hb -> futex_q relation.
2450 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2451 */
2452handle_err:
2453 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2454 spin_unlock(q->lock_ptr);
2455
2456 switch (err) {
2457 case -EFAULT:
2458 ret = fault_in_user_writeable(uaddr);
2459 break;
2460
2461 case -EAGAIN:
2462 cond_resched();
2463 ret = 0;
2464 break;
2465
2466 default:
2467 WARN_ON_ONCE(1);
2468 ret = err;
2469 break;
2470 }
2471
2472 spin_lock(q->lock_ptr);
2473 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2474
2475 /*
2476 * Check if someone else fixed it for us:
2477 */
2478 if (pi_state->owner != oldowner) {
2479 ret = 0;
2480 goto out_unlock;
2481 }
2482
2483 if (ret)
2484 goto out_unlock;
2485
2486 goto retry;
2487
2488out_unlock:
2489 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2490 return ret;
2491}
2492
2493static long futex_wait_restart(struct restart_block *restart);
2494
2495/**
2496 * fixup_owner() - Post lock pi_state and corner case management
2497 * @uaddr: user address of the futex
2498 * @q: futex_q (contains pi_state and access to the rt_mutex)
2499 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2500 *
2501 * After attempting to lock an rt_mutex, this function is called to cleanup
2502 * the pi_state owner as well as handle race conditions that may allow us to
2503 * acquire the lock. Must be called with the hb lock held.
2504 *
2505 * Return:
2506 * - 1 - success, lock taken;
2507 * - 0 - success, lock not taken;
2508 * - <0 - on error (-EFAULT)
2509 */
2510static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2511{
2512 int ret = 0;
2513
2514 if (locked) {
2515 /*
2516 * Got the lock. We might not be the anticipated owner if we
2517 * did a lock-steal - fix up the PI-state in that case:
2518 *
2519 * Speculative pi_state->owner read (we don't hold wait_lock);
2520 * since we own the lock pi_state->owner == current is the
2521 * stable state, anything else needs more attention.
2522 */
2523 if (q->pi_state->owner != current)
2524 ret = fixup_pi_state_owner(uaddr, q, current);
2525 return ret ? ret : locked;
2526 }
2527
2528 /*
2529 * If we didn't get the lock; check if anybody stole it from us. In
2530 * that case, we need to fix up the uval to point to them instead of
2531 * us, otherwise bad things happen. [10]
2532 *
2533 * Another speculative read; pi_state->owner == current is unstable
2534 * but needs our attention.
2535 */
2536 if (q->pi_state->owner == current) {
2537 ret = fixup_pi_state_owner(uaddr, q, NULL);
2538 return ret;
2539 }
2540
2541 /*
2542 * Paranoia check. If we did not take the lock, then we should not be
2543 * the owner of the rt_mutex.
2544 */
2545 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2546 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2547 "pi-state %p\n", ret,
2548 q->pi_state->pi_mutex.owner,
2549 q->pi_state->owner);
2550 }
2551
2552 return ret;
2553}
2554
2555/**
2556 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2557 * @hb: the futex hash bucket, must be locked by the caller
2558 * @q: the futex_q to queue up on
2559 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2560 */
2561static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2562 struct hrtimer_sleeper *timeout)
2563{
2564 /*
2565 * The task state is guaranteed to be set before another task can
2566 * wake it. set_current_state() is implemented using smp_store_mb() and
2567 * queue_me() calls spin_unlock() upon completion, both serializing
2568 * access to the hash list and forcing another memory barrier.
2569 */
2570 set_current_state(TASK_INTERRUPTIBLE);
2571 queue_me(q, hb);
2572
2573 /* Arm the timer */
2574 if (timeout)
2575 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2576
2577 /*
2578 * If we have been removed from the hash list, then another task
2579 * has tried to wake us, and we can skip the call to schedule().
2580 */
2581 if (likely(!plist_node_empty(&q->list))) {
2582 /*
2583 * If the timer has already expired, current will already be
2584 * flagged for rescheduling. Only call schedule if there
2585 * is no timeout, or if it has yet to expire.
2586 */
2587 if (!timeout || timeout->task)
2588 freezable_schedule();
2589 }
2590 __set_current_state(TASK_RUNNING);
2591}
2592
2593/**
2594 * futex_wait_setup() - Prepare to wait on a futex
2595 * @uaddr: the futex userspace address
2596 * @val: the expected value
2597 * @flags: futex flags (FLAGS_SHARED, etc.)
2598 * @q: the associated futex_q
2599 * @hb: storage for hash_bucket pointer to be returned to caller
2600 *
2601 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2602 * compare it with the expected value. Handle atomic faults internally.
2603 * Return with the hb lock held and a q.key reference on success, and unlocked
2604 * with no q.key reference on failure.
2605 *
2606 * Return:
2607 * - 0 - uaddr contains val and hb has been locked;
2608 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2609 */
2610static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2611 struct futex_q *q, struct futex_hash_bucket **hb)
2612{
2613 u32 uval;
2614 int ret;
2615
2616 /*
2617 * Access the page AFTER the hash-bucket is locked.
2618 * Order is important:
2619 *
2620 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2621 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2622 *
2623 * The basic logical guarantee of a futex is that it blocks ONLY
2624 * if cond(var) is known to be true at the time of blocking, for
2625 * any cond. If we locked the hash-bucket after testing *uaddr, that
2626 * would open a race condition where we could block indefinitely with
2627 * cond(var) false, which would violate the guarantee.
2628 *
2629 * On the other hand, we insert q and release the hash-bucket only
2630 * after testing *uaddr. This guarantees that futex_wait() will NOT
2631 * absorb a wakeup if *uaddr does not match the desired values
2632 * while the syscall executes.
2633 */
2634retry:
2635 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2636 if (unlikely(ret != 0))
2637 return ret;
2638
2639retry_private:
2640 *hb = queue_lock(q);
2641
2642 ret = get_futex_value_locked(&uval, uaddr);
2643
2644 if (ret) {
2645 queue_unlock(*hb);
2646
2647 ret = get_user(uval, uaddr);
2648 if (ret)
2649 return ret;
2650
2651 if (!(flags & FLAGS_SHARED))
2652 goto retry_private;
2653
2654 goto retry;
2655 }
2656
2657 if (uval != val) {
2658 queue_unlock(*hb);
2659 ret = -EWOULDBLOCK;
2660 }
2661
2662 return ret;
2663}
2664
2665static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2666 ktime_t *abs_time, u32 bitset)
2667{
2668 struct hrtimer_sleeper timeout, *to;
2669 struct restart_block *restart;
2670 struct futex_hash_bucket *hb;
2671 struct futex_q q = futex_q_init;
2672 int ret;
2673
2674 if (!bitset)
2675 return -EINVAL;
2676 q.bitset = bitset;
2677
2678 to = futex_setup_timer(abs_time, &timeout, flags,
2679 current->timer_slack_ns);
2680retry:
2681 /*
2682 * Prepare to wait on uaddr. On success, holds hb lock and increments
2683 * q.key refs.
2684 */
2685 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2686 if (ret)
2687 goto out;
2688
2689 /* queue_me and wait for wakeup, timeout, or a signal. */
2690 futex_wait_queue_me(hb, &q, to);
2691
2692 /* If we were woken (and unqueued), we succeeded, whatever. */
2693 ret = 0;
2694 /* unqueue_me() drops q.key ref */
2695 if (!unqueue_me(&q))
2696 goto out;
2697 ret = -ETIMEDOUT;
2698 if (to && !to->task)
2699 goto out;
2700
2701 /*
2702 * We expect signal_pending(current), but we might be the
2703 * victim of a spurious wakeup as well.
2704 */
2705 if (!signal_pending(current))
2706 goto retry;
2707
2708 ret = -ERESTARTSYS;
2709 if (!abs_time)
2710 goto out;
2711
2712 restart = ¤t->restart_block;
2713 restart->fn = futex_wait_restart;
2714 restart->futex.uaddr = uaddr;
2715 restart->futex.val = val;
2716 restart->futex.time = *abs_time;
2717 restart->futex.bitset = bitset;
2718 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2719
2720 ret = -ERESTART_RESTARTBLOCK;
2721
2722out:
2723 if (to) {
2724 hrtimer_cancel(&to->timer);
2725 destroy_hrtimer_on_stack(&to->timer);
2726 }
2727 return ret;
2728}
2729
2730
2731static long futex_wait_restart(struct restart_block *restart)
2732{
2733 u32 __user *uaddr = restart->futex.uaddr;
2734 ktime_t t, *tp = NULL;
2735
2736 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2737 t = restart->futex.time;
2738 tp = &t;
2739 }
2740 restart->fn = do_no_restart_syscall;
2741
2742 return (long)futex_wait(uaddr, restart->futex.flags,
2743 restart->futex.val, tp, restart->futex.bitset);
2744}
2745
2746
2747/*
2748 * Userspace tried a 0 -> TID atomic transition of the futex value
2749 * and failed. The kernel side here does the whole locking operation:
2750 * if there are waiters then it will block as a consequence of relying
2751 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2752 * a 0 value of the futex too.).
2753 *
2754 * Also serves as futex trylock_pi()'ing, and due semantics.
2755 */
2756static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2757 ktime_t *time, int trylock)
2758{
2759 struct hrtimer_sleeper timeout, *to;
2760 struct futex_pi_state *pi_state = NULL;
2761 struct task_struct *exiting = NULL;
2762 struct rt_mutex_waiter rt_waiter;
2763 struct futex_hash_bucket *hb;
2764 struct futex_q q = futex_q_init;
2765 int res, ret;
2766
2767 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2768 return -ENOSYS;
2769
2770 if (refill_pi_state_cache())
2771 return -ENOMEM;
2772
2773 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2774
2775retry:
2776 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2777 if (unlikely(ret != 0))
2778 goto out;
2779
2780retry_private:
2781 hb = queue_lock(&q);
2782
2783 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2784 &exiting, 0);
2785 if (unlikely(ret)) {
2786 /*
2787 * Atomic work succeeded and we got the lock,
2788 * or failed. Either way, we do _not_ block.
2789 */
2790 switch (ret) {
2791 case 1:
2792 /* We got the lock. */
2793 ret = 0;
2794 goto out_unlock_put_key;
2795 case -EFAULT:
2796 goto uaddr_faulted;
2797 case -EBUSY:
2798 case -EAGAIN:
2799 /*
2800 * Two reasons for this:
2801 * - EBUSY: Task is exiting and we just wait for the
2802 * exit to complete.
2803 * - EAGAIN: The user space value changed.
2804 */
2805 queue_unlock(hb);
2806 /*
2807 * Handle the case where the owner is in the middle of
2808 * exiting. Wait for the exit to complete otherwise
2809 * this task might loop forever, aka. live lock.
2810 */
2811 wait_for_owner_exiting(ret, exiting);
2812 cond_resched();
2813 goto retry;
2814 default:
2815 goto out_unlock_put_key;
2816 }
2817 }
2818
2819 WARN_ON(!q.pi_state);
2820
2821 /*
2822 * Only actually queue now that the atomic ops are done:
2823 */
2824 __queue_me(&q, hb);
2825
2826 if (trylock) {
2827 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2828 /* Fixup the trylock return value: */
2829 ret = ret ? 0 : -EWOULDBLOCK;
2830 goto no_block;
2831 }
2832
2833 rt_mutex_init_waiter(&rt_waiter);
2834
2835 /*
2836 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2837 * hold it while doing rt_mutex_start_proxy(), because then it will
2838 * include hb->lock in the blocking chain, even through we'll not in
2839 * fact hold it while blocking. This will lead it to report -EDEADLK
2840 * and BUG when futex_unlock_pi() interleaves with this.
2841 *
2842 * Therefore acquire wait_lock while holding hb->lock, but drop the
2843 * latter before calling __rt_mutex_start_proxy_lock(). This
2844 * interleaves with futex_unlock_pi() -- which does a similar lock
2845 * handoff -- such that the latter can observe the futex_q::pi_state
2846 * before __rt_mutex_start_proxy_lock() is done.
2847 */
2848 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2849 spin_unlock(q.lock_ptr);
2850 /*
2851 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2852 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2853 * it sees the futex_q::pi_state.
2854 */
2855 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2856 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2857
2858 if (ret) {
2859 if (ret == 1)
2860 ret = 0;
2861 goto cleanup;
2862 }
2863
2864 if (unlikely(to))
2865 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2866
2867 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2868
2869cleanup:
2870 spin_lock(q.lock_ptr);
2871 /*
2872 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2873 * first acquire the hb->lock before removing the lock from the
2874 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2875 * lists consistent.
2876 *
2877 * In particular; it is important that futex_unlock_pi() can not
2878 * observe this inconsistency.
2879 */
2880 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2881 ret = 0;
2882
2883no_block:
2884 /*
2885 * Fixup the pi_state owner and possibly acquire the lock if we
2886 * haven't already.
2887 */
2888 res = fixup_owner(uaddr, &q, !ret);
2889 /*
2890 * If fixup_owner() returned an error, proprogate that. If it acquired
2891 * the lock, clear our -ETIMEDOUT or -EINTR.
2892 */
2893 if (res)
2894 ret = (res < 0) ? res : 0;
2895
2896 /*
2897 * If fixup_owner() faulted and was unable to handle the fault, unlock
2898 * it and return the fault to userspace.
2899 */
2900 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2901 pi_state = q.pi_state;
2902 get_pi_state(pi_state);
2903 }
2904
2905 /* Unqueue and drop the lock */
2906 unqueue_me_pi(&q);
2907
2908 if (pi_state) {
2909 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2910 put_pi_state(pi_state);
2911 }
2912
2913 goto out;
2914
2915out_unlock_put_key:
2916 queue_unlock(hb);
2917
2918out:
2919 if (to) {
2920 hrtimer_cancel(&to->timer);
2921 destroy_hrtimer_on_stack(&to->timer);
2922 }
2923 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2924
2925uaddr_faulted:
2926 queue_unlock(hb);
2927
2928 ret = fault_in_user_writeable(uaddr);
2929 if (ret)
2930 goto out;
2931
2932 if (!(flags & FLAGS_SHARED))
2933 goto retry_private;
2934
2935 goto retry;
2936}
2937
2938/*
2939 * Userspace attempted a TID -> 0 atomic transition, and failed.
2940 * This is the in-kernel slowpath: we look up the PI state (if any),
2941 * and do the rt-mutex unlock.
2942 */
2943static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2944{
2945 u32 curval, uval, vpid = task_pid_vnr(current);
2946 union futex_key key = FUTEX_KEY_INIT;
2947 struct futex_hash_bucket *hb;
2948 struct futex_q *top_waiter;
2949 int ret;
2950
2951 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2952 return -ENOSYS;
2953
2954retry:
2955 if (get_user(uval, uaddr))
2956 return -EFAULT;
2957 /*
2958 * We release only a lock we actually own:
2959 */
2960 if ((uval & FUTEX_TID_MASK) != vpid)
2961 return -EPERM;
2962
2963 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2964 if (ret)
2965 return ret;
2966
2967 hb = hash_futex(&key);
2968 spin_lock(&hb->lock);
2969
2970 /*
2971 * Check waiters first. We do not trust user space values at
2972 * all and we at least want to know if user space fiddled
2973 * with the futex value instead of blindly unlocking.
2974 */
2975 top_waiter = futex_top_waiter(hb, &key);
2976 if (top_waiter) {
2977 struct futex_pi_state *pi_state = top_waiter->pi_state;
2978
2979 ret = -EINVAL;
2980 if (!pi_state)
2981 goto out_unlock;
2982
2983 /*
2984 * If current does not own the pi_state then the futex is
2985 * inconsistent and user space fiddled with the futex value.
2986 */
2987 if (pi_state->owner != current)
2988 goto out_unlock;
2989
2990 get_pi_state(pi_state);
2991 /*
2992 * By taking wait_lock while still holding hb->lock, we ensure
2993 * there is no point where we hold neither; and therefore
2994 * wake_futex_pi() must observe a state consistent with what we
2995 * observed.
2996 *
2997 * In particular; this forces __rt_mutex_start_proxy() to
2998 * complete such that we're guaranteed to observe the
2999 * rt_waiter. Also see the WARN in wake_futex_pi().
3000 */
3001 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3002 spin_unlock(&hb->lock);
3003
3004 /* drops pi_state->pi_mutex.wait_lock */
3005 ret = wake_futex_pi(uaddr, uval, pi_state);
3006
3007 put_pi_state(pi_state);
3008
3009 /*
3010 * Success, we're done! No tricky corner cases.
3011 */
3012 if (!ret)
3013 goto out_putkey;
3014 /*
3015 * The atomic access to the futex value generated a
3016 * pagefault, so retry the user-access and the wakeup:
3017 */
3018 if (ret == -EFAULT)
3019 goto pi_faulted;
3020 /*
3021 * A unconditional UNLOCK_PI op raced against a waiter
3022 * setting the FUTEX_WAITERS bit. Try again.
3023 */
3024 if (ret == -EAGAIN)
3025 goto pi_retry;
3026 /*
3027 * wake_futex_pi has detected invalid state. Tell user
3028 * space.
3029 */
3030 goto out_putkey;
3031 }
3032
3033 /*
3034 * We have no kernel internal state, i.e. no waiters in the
3035 * kernel. Waiters which are about to queue themselves are stuck
3036 * on hb->lock. So we can safely ignore them. We do neither
3037 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3038 * owner.
3039 */
3040 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3041 spin_unlock(&hb->lock);
3042 switch (ret) {
3043 case -EFAULT:
3044 goto pi_faulted;
3045
3046 case -EAGAIN:
3047 goto pi_retry;
3048
3049 default:
3050 WARN_ON_ONCE(1);
3051 goto out_putkey;
3052 }
3053 }
3054
3055 /*
3056 * If uval has changed, let user space handle it.
3057 */
3058 ret = (curval == uval) ? 0 : -EAGAIN;
3059
3060out_unlock:
3061 spin_unlock(&hb->lock);
3062out_putkey:
3063 return ret;
3064
3065pi_retry:
3066 cond_resched();
3067 goto retry;
3068
3069pi_faulted:
3070
3071 ret = fault_in_user_writeable(uaddr);
3072 if (!ret)
3073 goto retry;
3074
3075 return ret;
3076}
3077
3078/**
3079 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3080 * @hb: the hash_bucket futex_q was original enqueued on
3081 * @q: the futex_q woken while waiting to be requeued
3082 * @key2: the futex_key of the requeue target futex
3083 * @timeout: the timeout associated with the wait (NULL if none)
3084 *
3085 * Detect if the task was woken on the initial futex as opposed to the requeue
3086 * target futex. If so, determine if it was a timeout or a signal that caused
3087 * the wakeup and return the appropriate error code to the caller. Must be
3088 * called with the hb lock held.
3089 *
3090 * Return:
3091 * - 0 = no early wakeup detected;
3092 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3093 */
3094static inline
3095int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3096 struct futex_q *q, union futex_key *key2,
3097 struct hrtimer_sleeper *timeout)
3098{
3099 int ret = 0;
3100
3101 /*
3102 * With the hb lock held, we avoid races while we process the wakeup.
3103 * We only need to hold hb (and not hb2) to ensure atomicity as the
3104 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3105 * It can't be requeued from uaddr2 to something else since we don't
3106 * support a PI aware source futex for requeue.
3107 */
3108 if (!match_futex(&q->key, key2)) {
3109 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3110 /*
3111 * We were woken prior to requeue by a timeout or a signal.
3112 * Unqueue the futex_q and determine which it was.
3113 */
3114 plist_del(&q->list, &hb->chain);
3115 hb_waiters_dec(hb);
3116
3117 /* Handle spurious wakeups gracefully */
3118 ret = -EWOULDBLOCK;
3119 if (timeout && !timeout->task)
3120 ret = -ETIMEDOUT;
3121 else if (signal_pending(current))
3122 ret = -ERESTARTNOINTR;
3123 }
3124 return ret;
3125}
3126
3127/**
3128 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3129 * @uaddr: the futex we initially wait on (non-pi)
3130 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3131 * the same type, no requeueing from private to shared, etc.
3132 * @val: the expected value of uaddr
3133 * @abs_time: absolute timeout
3134 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3135 * @uaddr2: the pi futex we will take prior to returning to user-space
3136 *
3137 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3138 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3139 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3140 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3141 * without one, the pi logic would not know which task to boost/deboost, if
3142 * there was a need to.
3143 *
3144 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3145 * via the following--
3146 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3147 * 2) wakeup on uaddr2 after a requeue
3148 * 3) signal
3149 * 4) timeout
3150 *
3151 * If 3, cleanup and return -ERESTARTNOINTR.
3152 *
3153 * If 2, we may then block on trying to take the rt_mutex and return via:
3154 * 5) successful lock
3155 * 6) signal
3156 * 7) timeout
3157 * 8) other lock acquisition failure
3158 *
3159 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3160 *
3161 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3162 *
3163 * Return:
3164 * - 0 - On success;
3165 * - <0 - On error
3166 */
3167static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3168 u32 val, ktime_t *abs_time, u32 bitset,
3169 u32 __user *uaddr2)
3170{
3171 struct hrtimer_sleeper timeout, *to;
3172 struct futex_pi_state *pi_state = NULL;
3173 struct rt_mutex_waiter rt_waiter;
3174 struct futex_hash_bucket *hb;
3175 union futex_key key2 = FUTEX_KEY_INIT;
3176 struct futex_q q = futex_q_init;
3177 int res, ret;
3178
3179 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3180 return -ENOSYS;
3181
3182 if (uaddr == uaddr2)
3183 return -EINVAL;
3184
3185 if (!bitset)
3186 return -EINVAL;
3187
3188 to = futex_setup_timer(abs_time, &timeout, flags,
3189 current->timer_slack_ns);
3190
3191 /*
3192 * The waiter is allocated on our stack, manipulated by the requeue
3193 * code while we sleep on uaddr.
3194 */
3195 rt_mutex_init_waiter(&rt_waiter);
3196
3197 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3198 if (unlikely(ret != 0))
3199 goto out;
3200
3201 q.bitset = bitset;
3202 q.rt_waiter = &rt_waiter;
3203 q.requeue_pi_key = &key2;
3204
3205 /*
3206 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3207 * count.
3208 */
3209 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3210 if (ret)
3211 goto out;
3212
3213 /*
3214 * The check above which compares uaddrs is not sufficient for
3215 * shared futexes. We need to compare the keys:
3216 */
3217 if (match_futex(&q.key, &key2)) {
3218 queue_unlock(hb);
3219 ret = -EINVAL;
3220 goto out;
3221 }
3222
3223 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3224 futex_wait_queue_me(hb, &q, to);
3225
3226 spin_lock(&hb->lock);
3227 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3228 spin_unlock(&hb->lock);
3229 if (ret)
3230 goto out;
3231
3232 /*
3233 * In order for us to be here, we know our q.key == key2, and since
3234 * we took the hb->lock above, we also know that futex_requeue() has
3235 * completed and we no longer have to concern ourselves with a wakeup
3236 * race with the atomic proxy lock acquisition by the requeue code. The
3237 * futex_requeue dropped our key1 reference and incremented our key2
3238 * reference count.
3239 */
3240
3241 /* Check if the requeue code acquired the second futex for us. */
3242 if (!q.rt_waiter) {
3243 /*
3244 * Got the lock. We might not be the anticipated owner if we
3245 * did a lock-steal - fix up the PI-state in that case.
3246 */
3247 if (q.pi_state && (q.pi_state->owner != current)) {
3248 spin_lock(q.lock_ptr);
3249 ret = fixup_pi_state_owner(uaddr2, &q, current);
3250 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3251 pi_state = q.pi_state;
3252 get_pi_state(pi_state);
3253 }
3254 /*
3255 * Drop the reference to the pi state which
3256 * the requeue_pi() code acquired for us.
3257 */
3258 put_pi_state(q.pi_state);
3259 spin_unlock(q.lock_ptr);
3260 }
3261 } else {
3262 struct rt_mutex *pi_mutex;
3263
3264 /*
3265 * We have been woken up by futex_unlock_pi(), a timeout, or a
3266 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3267 * the pi_state.
3268 */
3269 WARN_ON(!q.pi_state);
3270 pi_mutex = &q.pi_state->pi_mutex;
3271 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3272
3273 spin_lock(q.lock_ptr);
3274 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3275 ret = 0;
3276
3277 debug_rt_mutex_free_waiter(&rt_waiter);
3278 /*
3279 * Fixup the pi_state owner and possibly acquire the lock if we
3280 * haven't already.
3281 */
3282 res = fixup_owner(uaddr2, &q, !ret);
3283 /*
3284 * If fixup_owner() returned an error, proprogate that. If it
3285 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3286 */
3287 if (res)
3288 ret = (res < 0) ? res : 0;
3289
3290 /*
3291 * If fixup_pi_state_owner() faulted and was unable to handle
3292 * the fault, unlock the rt_mutex and return the fault to
3293 * userspace.
3294 */
3295 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3296 pi_state = q.pi_state;
3297 get_pi_state(pi_state);
3298 }
3299
3300 /* Unqueue and drop the lock. */
3301 unqueue_me_pi(&q);
3302 }
3303
3304 if (pi_state) {
3305 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3306 put_pi_state(pi_state);
3307 }
3308
3309 if (ret == -EINTR) {
3310 /*
3311 * We've already been requeued, but cannot restart by calling
3312 * futex_lock_pi() directly. We could restart this syscall, but
3313 * it would detect that the user space "val" changed and return
3314 * -EWOULDBLOCK. Save the overhead of the restart and return
3315 * -EWOULDBLOCK directly.
3316 */
3317 ret = -EWOULDBLOCK;
3318 }
3319
3320out:
3321 if (to) {
3322 hrtimer_cancel(&to->timer);
3323 destroy_hrtimer_on_stack(&to->timer);
3324 }
3325 return ret;
3326}
3327
3328/*
3329 * Support for robust futexes: the kernel cleans up held futexes at
3330 * thread exit time.
3331 *
3332 * Implementation: user-space maintains a per-thread list of locks it
3333 * is holding. Upon do_exit(), the kernel carefully walks this list,
3334 * and marks all locks that are owned by this thread with the
3335 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3336 * always manipulated with the lock held, so the list is private and
3337 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3338 * field, to allow the kernel to clean up if the thread dies after
3339 * acquiring the lock, but just before it could have added itself to
3340 * the list. There can only be one such pending lock.
3341 */
3342
3343/**
3344 * sys_set_robust_list() - Set the robust-futex list head of a task
3345 * @head: pointer to the list-head
3346 * @len: length of the list-head, as userspace expects
3347 */
3348SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3349 size_t, len)
3350{
3351 if (!futex_cmpxchg_enabled)
3352 return -ENOSYS;
3353 /*
3354 * The kernel knows only one size for now:
3355 */
3356 if (unlikely(len != sizeof(*head)))
3357 return -EINVAL;
3358
3359 current->robust_list = head;
3360
3361 return 0;
3362}
3363
3364/**
3365 * sys_get_robust_list() - Get the robust-futex list head of a task
3366 * @pid: pid of the process [zero for current task]
3367 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3368 * @len_ptr: pointer to a length field, the kernel fills in the header size
3369 */
3370SYSCALL_DEFINE3(get_robust_list, int, pid,
3371 struct robust_list_head __user * __user *, head_ptr,
3372 size_t __user *, len_ptr)
3373{
3374 struct robust_list_head __user *head;
3375 unsigned long ret;
3376 struct task_struct *p;
3377
3378 if (!futex_cmpxchg_enabled)
3379 return -ENOSYS;
3380
3381 rcu_read_lock();
3382
3383 ret = -ESRCH;
3384 if (!pid)
3385 p = current;
3386 else {
3387 p = find_task_by_vpid(pid);
3388 if (!p)
3389 goto err_unlock;
3390 }
3391
3392 ret = -EPERM;
3393 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3394 goto err_unlock;
3395
3396 head = p->robust_list;
3397 rcu_read_unlock();
3398
3399 if (put_user(sizeof(*head), len_ptr))
3400 return -EFAULT;
3401 return put_user(head, head_ptr);
3402
3403err_unlock:
3404 rcu_read_unlock();
3405
3406 return ret;
3407}
3408
3409/* Constants for the pending_op argument of handle_futex_death */
3410#define HANDLE_DEATH_PENDING true
3411#define HANDLE_DEATH_LIST false
3412
3413/*
3414 * Process a futex-list entry, check whether it's owned by the
3415 * dying task, and do notification if so:
3416 */
3417static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3418 bool pi, bool pending_op)
3419{
3420 u32 uval, nval, mval;
3421 int err;
3422
3423 /* Futex address must be 32bit aligned */
3424 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3425 return -1;
3426
3427retry:
3428 if (get_user(uval, uaddr))
3429 return -1;
3430
3431 /*
3432 * Special case for regular (non PI) futexes. The unlock path in
3433 * user space has two race scenarios:
3434 *
3435 * 1. The unlock path releases the user space futex value and
3436 * before it can execute the futex() syscall to wake up
3437 * waiters it is killed.
3438 *
3439 * 2. A woken up waiter is killed before it can acquire the
3440 * futex in user space.
3441 *
3442 * In both cases the TID validation below prevents a wakeup of
3443 * potential waiters which can cause these waiters to block
3444 * forever.
3445 *
3446 * In both cases the following conditions are met:
3447 *
3448 * 1) task->robust_list->list_op_pending != NULL
3449 * @pending_op == true
3450 * 2) User space futex value == 0
3451 * 3) Regular futex: @pi == false
3452 *
3453 * If these conditions are met, it is safe to attempt waking up a
3454 * potential waiter without touching the user space futex value and
3455 * trying to set the OWNER_DIED bit. The user space futex value is
3456 * uncontended and the rest of the user space mutex state is
3457 * consistent, so a woken waiter will just take over the
3458 * uncontended futex. Setting the OWNER_DIED bit would create
3459 * inconsistent state and malfunction of the user space owner died
3460 * handling.
3461 */
3462 if (pending_op && !pi && !uval) {
3463 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3464 return 0;
3465 }
3466
3467 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3468 return 0;
3469
3470 /*
3471 * Ok, this dying thread is truly holding a futex
3472 * of interest. Set the OWNER_DIED bit atomically
3473 * via cmpxchg, and if the value had FUTEX_WAITERS
3474 * set, wake up a waiter (if any). (We have to do a
3475 * futex_wake() even if OWNER_DIED is already set -
3476 * to handle the rare but possible case of recursive
3477 * thread-death.) The rest of the cleanup is done in
3478 * userspace.
3479 */
3480 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3481
3482 /*
3483 * We are not holding a lock here, but we want to have
3484 * the pagefault_disable/enable() protection because
3485 * we want to handle the fault gracefully. If the
3486 * access fails we try to fault in the futex with R/W
3487 * verification via get_user_pages. get_user() above
3488 * does not guarantee R/W access. If that fails we
3489 * give up and leave the futex locked.
3490 */
3491 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3492 switch (err) {
3493 case -EFAULT:
3494 if (fault_in_user_writeable(uaddr))
3495 return -1;
3496 goto retry;
3497
3498 case -EAGAIN:
3499 cond_resched();
3500 goto retry;
3501
3502 default:
3503 WARN_ON_ONCE(1);
3504 return err;
3505 }
3506 }
3507
3508 if (nval != uval)
3509 goto retry;
3510
3511 /*
3512 * Wake robust non-PI futexes here. The wakeup of
3513 * PI futexes happens in exit_pi_state():
3514 */
3515 if (!pi && (uval & FUTEX_WAITERS))
3516 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3517
3518 return 0;
3519}
3520
3521/*
3522 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3523 */
3524static inline int fetch_robust_entry(struct robust_list __user **entry,
3525 struct robust_list __user * __user *head,
3526 unsigned int *pi)
3527{
3528 unsigned long uentry;
3529
3530 if (get_user(uentry, (unsigned long __user *)head))
3531 return -EFAULT;
3532
3533 *entry = (void __user *)(uentry & ~1UL);
3534 *pi = uentry & 1;
3535
3536 return 0;
3537}
3538
3539/*
3540 * Walk curr->robust_list (very carefully, it's a userspace list!)
3541 * and mark any locks found there dead, and notify any waiters.
3542 *
3543 * We silently return on any sign of list-walking problem.
3544 */
3545static void exit_robust_list(struct task_struct *curr)
3546{
3547 struct robust_list_head __user *head = curr->robust_list;
3548 struct robust_list __user *entry, *next_entry, *pending;
3549 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3550 unsigned int next_pi;
3551 unsigned long futex_offset;
3552 int rc;
3553
3554 if (!futex_cmpxchg_enabled)
3555 return;
3556
3557 /*
3558 * Fetch the list head (which was registered earlier, via
3559 * sys_set_robust_list()):
3560 */
3561 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3562 return;
3563 /*
3564 * Fetch the relative futex offset:
3565 */
3566 if (get_user(futex_offset, &head->futex_offset))
3567 return;
3568 /*
3569 * Fetch any possibly pending lock-add first, and handle it
3570 * if it exists:
3571 */
3572 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3573 return;
3574
3575 next_entry = NULL; /* avoid warning with gcc */
3576 while (entry != &head->list) {
3577 /*
3578 * Fetch the next entry in the list before calling
3579 * handle_futex_death:
3580 */
3581 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3582 /*
3583 * A pending lock might already be on the list, so
3584 * don't process it twice:
3585 */
3586 if (entry != pending) {
3587 if (handle_futex_death((void __user *)entry + futex_offset,
3588 curr, pi, HANDLE_DEATH_LIST))
3589 return;
3590 }
3591 if (rc)
3592 return;
3593 entry = next_entry;
3594 pi = next_pi;
3595 /*
3596 * Avoid excessively long or circular lists:
3597 */
3598 if (!--limit)
3599 break;
3600
3601 cond_resched();
3602 }
3603
3604 if (pending) {
3605 handle_futex_death((void __user *)pending + futex_offset,
3606 curr, pip, HANDLE_DEATH_PENDING);
3607 }
3608}
3609
3610static void futex_cleanup(struct task_struct *tsk)
3611{
3612 if (unlikely(tsk->robust_list)) {
3613 exit_robust_list(tsk);
3614 tsk->robust_list = NULL;
3615 }
3616
3617#ifdef CONFIG_COMPAT
3618 if (unlikely(tsk->compat_robust_list)) {
3619 compat_exit_robust_list(tsk);
3620 tsk->compat_robust_list = NULL;
3621 }
3622#endif
3623
3624 if (unlikely(!list_empty(&tsk->pi_state_list)))
3625 exit_pi_state_list(tsk);
3626}
3627
3628/**
3629 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3630 * @tsk: task to set the state on
3631 *
3632 * Set the futex exit state of the task lockless. The futex waiter code
3633 * observes that state when a task is exiting and loops until the task has
3634 * actually finished the futex cleanup. The worst case for this is that the
3635 * waiter runs through the wait loop until the state becomes visible.
3636 *
3637 * This is called from the recursive fault handling path in do_exit().
3638 *
3639 * This is best effort. Either the futex exit code has run already or
3640 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3641 * take it over. If not, the problem is pushed back to user space. If the
3642 * futex exit code did not run yet, then an already queued waiter might
3643 * block forever, but there is nothing which can be done about that.
3644 */
3645void futex_exit_recursive(struct task_struct *tsk)
3646{
3647 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3648 if (tsk->futex_state == FUTEX_STATE_EXITING)
3649 mutex_unlock(&tsk->futex_exit_mutex);
3650 tsk->futex_state = FUTEX_STATE_DEAD;
3651}
3652
3653static void futex_cleanup_begin(struct task_struct *tsk)
3654{
3655 /*
3656 * Prevent various race issues against a concurrent incoming waiter
3657 * including live locks by forcing the waiter to block on
3658 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3659 * attach_to_pi_owner().
3660 */
3661 mutex_lock(&tsk->futex_exit_mutex);
3662
3663 /*
3664 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3665 *
3666 * This ensures that all subsequent checks of tsk->futex_state in
3667 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3668 * tsk->pi_lock held.
3669 *
3670 * It guarantees also that a pi_state which was queued right before
3671 * the state change under tsk->pi_lock by a concurrent waiter must
3672 * be observed in exit_pi_state_list().
3673 */
3674 raw_spin_lock_irq(&tsk->pi_lock);
3675 tsk->futex_state = FUTEX_STATE_EXITING;
3676 raw_spin_unlock_irq(&tsk->pi_lock);
3677}
3678
3679static void futex_cleanup_end(struct task_struct *tsk, int state)
3680{
3681 /*
3682 * Lockless store. The only side effect is that an observer might
3683 * take another loop until it becomes visible.
3684 */
3685 tsk->futex_state = state;
3686 /*
3687 * Drop the exit protection. This unblocks waiters which observed
3688 * FUTEX_STATE_EXITING to reevaluate the state.
3689 */
3690 mutex_unlock(&tsk->futex_exit_mutex);
3691}
3692
3693void futex_exec_release(struct task_struct *tsk)
3694{
3695 /*
3696 * The state handling is done for consistency, but in the case of
3697 * exec() there is no way to prevent futher damage as the PID stays
3698 * the same. But for the unlikely and arguably buggy case that a
3699 * futex is held on exec(), this provides at least as much state
3700 * consistency protection which is possible.
3701 */
3702 futex_cleanup_begin(tsk);
3703 futex_cleanup(tsk);
3704 /*
3705 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3706 * exec a new binary.
3707 */
3708 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3709}
3710
3711void futex_exit_release(struct task_struct *tsk)
3712{
3713 futex_cleanup_begin(tsk);
3714 futex_cleanup(tsk);
3715 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3716}
3717
3718long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3719 u32 __user *uaddr2, u32 val2, u32 val3)
3720{
3721 int cmd = op & FUTEX_CMD_MASK;
3722 unsigned int flags = 0;
3723
3724 if (!(op & FUTEX_PRIVATE_FLAG))
3725 flags |= FLAGS_SHARED;
3726
3727 if (op & FUTEX_CLOCK_REALTIME) {
3728 flags |= FLAGS_CLOCKRT;
3729 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3730 cmd != FUTEX_WAIT_REQUEUE_PI)
3731 return -ENOSYS;
3732 }
3733
3734 switch (cmd) {
3735 case FUTEX_LOCK_PI:
3736 case FUTEX_UNLOCK_PI:
3737 case FUTEX_TRYLOCK_PI:
3738 case FUTEX_WAIT_REQUEUE_PI:
3739 case FUTEX_CMP_REQUEUE_PI:
3740 if (!futex_cmpxchg_enabled)
3741 return -ENOSYS;
3742 }
3743
3744 switch (cmd) {
3745 case FUTEX_WAIT:
3746 val3 = FUTEX_BITSET_MATCH_ANY;
3747 fallthrough;
3748 case FUTEX_WAIT_BITSET:
3749 return futex_wait(uaddr, flags, val, timeout, val3);
3750 case FUTEX_WAKE:
3751 val3 = FUTEX_BITSET_MATCH_ANY;
3752 fallthrough;
3753 case FUTEX_WAKE_BITSET:
3754 return futex_wake(uaddr, flags, val, val3);
3755 case FUTEX_REQUEUE:
3756 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3757 case FUTEX_CMP_REQUEUE:
3758 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3759 case FUTEX_WAKE_OP:
3760 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3761 case FUTEX_LOCK_PI:
3762 return futex_lock_pi(uaddr, flags, timeout, 0);
3763 case FUTEX_UNLOCK_PI:
3764 return futex_unlock_pi(uaddr, flags);
3765 case FUTEX_TRYLOCK_PI:
3766 return futex_lock_pi(uaddr, flags, NULL, 1);
3767 case FUTEX_WAIT_REQUEUE_PI:
3768 val3 = FUTEX_BITSET_MATCH_ANY;
3769 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3770 uaddr2);
3771 case FUTEX_CMP_REQUEUE_PI:
3772 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3773 }
3774 return -ENOSYS;
3775}
3776
3777
3778SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3779 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3780 u32, val3)
3781{
3782 struct timespec64 ts;
3783 ktime_t t, *tp = NULL;
3784 u32 val2 = 0;
3785 int cmd = op & FUTEX_CMD_MASK;
3786
3787 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3788 cmd == FUTEX_WAIT_BITSET ||
3789 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3790 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3791 return -EFAULT;
3792 if (get_timespec64(&ts, utime))
3793 return -EFAULT;
3794 if (!timespec64_valid(&ts))
3795 return -EINVAL;
3796
3797 t = timespec64_to_ktime(ts);
3798 if (cmd == FUTEX_WAIT)
3799 t = ktime_add_safe(ktime_get(), t);
3800 tp = &t;
3801 }
3802 /*
3803 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3804 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3805 */
3806 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3807 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3808 val2 = (u32) (unsigned long) utime;
3809
3810 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3811}
3812
3813#ifdef CONFIG_COMPAT
3814/*
3815 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3816 */
3817static inline int
3818compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3819 compat_uptr_t __user *head, unsigned int *pi)
3820{
3821 if (get_user(*uentry, head))
3822 return -EFAULT;
3823
3824 *entry = compat_ptr((*uentry) & ~1);
3825 *pi = (unsigned int)(*uentry) & 1;
3826
3827 return 0;
3828}
3829
3830static void __user *futex_uaddr(struct robust_list __user *entry,
3831 compat_long_t futex_offset)
3832{
3833 compat_uptr_t base = ptr_to_compat(entry);
3834 void __user *uaddr = compat_ptr(base + futex_offset);
3835
3836 return uaddr;
3837}
3838
3839/*
3840 * Walk curr->robust_list (very carefully, it's a userspace list!)
3841 * and mark any locks found there dead, and notify any waiters.
3842 *
3843 * We silently return on any sign of list-walking problem.
3844 */
3845static void compat_exit_robust_list(struct task_struct *curr)
3846{
3847 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3848 struct robust_list __user *entry, *next_entry, *pending;
3849 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3850 unsigned int next_pi;
3851 compat_uptr_t uentry, next_uentry, upending;
3852 compat_long_t futex_offset;
3853 int rc;
3854
3855 if (!futex_cmpxchg_enabled)
3856 return;
3857
3858 /*
3859 * Fetch the list head (which was registered earlier, via
3860 * sys_set_robust_list()):
3861 */
3862 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3863 return;
3864 /*
3865 * Fetch the relative futex offset:
3866 */
3867 if (get_user(futex_offset, &head->futex_offset))
3868 return;
3869 /*
3870 * Fetch any possibly pending lock-add first, and handle it
3871 * if it exists:
3872 */
3873 if (compat_fetch_robust_entry(&upending, &pending,
3874 &head->list_op_pending, &pip))
3875 return;
3876
3877 next_entry = NULL; /* avoid warning with gcc */
3878 while (entry != (struct robust_list __user *) &head->list) {
3879 /*
3880 * Fetch the next entry in the list before calling
3881 * handle_futex_death:
3882 */
3883 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3884 (compat_uptr_t __user *)&entry->next, &next_pi);
3885 /*
3886 * A pending lock might already be on the list, so
3887 * dont process it twice:
3888 */
3889 if (entry != pending) {
3890 void __user *uaddr = futex_uaddr(entry, futex_offset);
3891
3892 if (handle_futex_death(uaddr, curr, pi,
3893 HANDLE_DEATH_LIST))
3894 return;
3895 }
3896 if (rc)
3897 return;
3898 uentry = next_uentry;
3899 entry = next_entry;
3900 pi = next_pi;
3901 /*
3902 * Avoid excessively long or circular lists:
3903 */
3904 if (!--limit)
3905 break;
3906
3907 cond_resched();
3908 }
3909 if (pending) {
3910 void __user *uaddr = futex_uaddr(pending, futex_offset);
3911
3912 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3913 }
3914}
3915
3916COMPAT_SYSCALL_DEFINE2(set_robust_list,
3917 struct compat_robust_list_head __user *, head,
3918 compat_size_t, len)
3919{
3920 if (!futex_cmpxchg_enabled)
3921 return -ENOSYS;
3922
3923 if (unlikely(len != sizeof(*head)))
3924 return -EINVAL;
3925
3926 current->compat_robust_list = head;
3927
3928 return 0;
3929}
3930
3931COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3932 compat_uptr_t __user *, head_ptr,
3933 compat_size_t __user *, len_ptr)
3934{
3935 struct compat_robust_list_head __user *head;
3936 unsigned long ret;
3937 struct task_struct *p;
3938
3939 if (!futex_cmpxchg_enabled)
3940 return -ENOSYS;
3941
3942 rcu_read_lock();
3943
3944 ret = -ESRCH;
3945 if (!pid)
3946 p = current;
3947 else {
3948 p = find_task_by_vpid(pid);
3949 if (!p)
3950 goto err_unlock;
3951 }
3952
3953 ret = -EPERM;
3954 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3955 goto err_unlock;
3956
3957 head = p->compat_robust_list;
3958 rcu_read_unlock();
3959
3960 if (put_user(sizeof(*head), len_ptr))
3961 return -EFAULT;
3962 return put_user(ptr_to_compat(head), head_ptr);
3963
3964err_unlock:
3965 rcu_read_unlock();
3966
3967 return ret;
3968}
3969#endif /* CONFIG_COMPAT */
3970
3971#ifdef CONFIG_COMPAT_32BIT_TIME
3972SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3973 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3974 u32, val3)
3975{
3976 struct timespec64 ts;
3977 ktime_t t, *tp = NULL;
3978 int val2 = 0;
3979 int cmd = op & FUTEX_CMD_MASK;
3980
3981 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3982 cmd == FUTEX_WAIT_BITSET ||
3983 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3984 if (get_old_timespec32(&ts, utime))
3985 return -EFAULT;
3986 if (!timespec64_valid(&ts))
3987 return -EINVAL;
3988
3989 t = timespec64_to_ktime(ts);
3990 if (cmd == FUTEX_WAIT)
3991 t = ktime_add_safe(ktime_get(), t);
3992 tp = &t;
3993 }
3994 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3995 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3996 val2 = (int) (unsigned long) utime;
3997
3998 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3999}
4000#endif /* CONFIG_COMPAT_32BIT_TIME */
4001
4002static void __init futex_detect_cmpxchg(void)
4003{
4004#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4005 u32 curval;
4006
4007 /*
4008 * This will fail and we want it. Some arch implementations do
4009 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4010 * functionality. We want to know that before we call in any
4011 * of the complex code paths. Also we want to prevent
4012 * registration of robust lists in that case. NULL is
4013 * guaranteed to fault and we get -EFAULT on functional
4014 * implementation, the non-functional ones will return
4015 * -ENOSYS.
4016 */
4017 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4018 futex_cmpxchg_enabled = 1;
4019#endif
4020}
4021
4022static int __init futex_init(void)
4023{
4024 unsigned int futex_shift;
4025 unsigned long i;
4026
4027#if CONFIG_BASE_SMALL
4028 futex_hashsize = 16;
4029#else
4030 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4031#endif
4032
4033 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4034 futex_hashsize, 0,
4035 futex_hashsize < 256 ? HASH_SMALL : 0,
4036 &futex_shift, NULL,
4037 futex_hashsize, futex_hashsize);
4038 futex_hashsize = 1UL << futex_shift;
4039
4040 futex_detect_cmpxchg();
4041
4042 for (i = 0; i < futex_hashsize; i++) {
4043 atomic_set(&futex_queues[i].waiters, 0);
4044 plist_head_init(&futex_queues[i].chain);
4045 spin_lock_init(&futex_queues[i].lock);
4046 }
4047
4048 return 0;
4049}
4050core_initcall(futex_init);