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/debugfs.h>
38#include <linux/plist.h>
39#include <linux/gfp.h>
40#include <linux/vmalloc.h>
41#include <linux/memblock.h>
42#include <linux/fault-inject.h>
43#include <linux/slab.h>
44#include <linux/prctl.h>
45#include <linux/mempolicy.h>
46#include <linux/mmap_lock.h>
47
48#include "futex.h"
49#include "../locking/rtmutex_common.h"
50
51/*
52 * The base of the bucket array and its size are always used together
53 * (after initialization only in futex_hash()), so ensure that they
54 * reside in the same cacheline.
55 */
56static struct {
57 unsigned long hashmask;
58 unsigned int hashshift;
59 struct futex_hash_bucket *queues[MAX_NUMNODES];
60} __futex_data __read_mostly __aligned(2*sizeof(long));
61
62#define futex_hashmask (__futex_data.hashmask)
63#define futex_hashshift (__futex_data.hashshift)
64#define futex_queues (__futex_data.queues)
65
66struct futex_private_hash {
67 int state;
68 unsigned int hash_mask;
69 struct rcu_head rcu;
70 void *mm;
71 bool custom;
72 struct futex_hash_bucket queues[];
73};
74
75/*
76 * Fault injections for futexes.
77 */
78#ifdef CONFIG_FAIL_FUTEX
79
80static struct {
81 struct fault_attr attr;
82
83 bool ignore_private;
84} fail_futex = {
85 .attr = FAULT_ATTR_INITIALIZER,
86 .ignore_private = false,
87};
88
89static int __init setup_fail_futex(char *str)
90{
91 return setup_fault_attr(&fail_futex.attr, str);
92}
93__setup("fail_futex=", setup_fail_futex);
94
95bool should_fail_futex(bool fshared)
96{
97 if (fail_futex.ignore_private && !fshared)
98 return false;
99
100 return should_fail(&fail_futex.attr, 1);
101}
102
103#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
104
105static int __init fail_futex_debugfs(void)
106{
107 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
108 struct dentry *dir;
109
110 dir = fault_create_debugfs_attr("fail_futex", NULL,
111 &fail_futex.attr);
112 if (IS_ERR(dir))
113 return PTR_ERR(dir);
114
115 debugfs_create_bool("ignore-private", mode, dir,
116 &fail_futex.ignore_private);
117 return 0;
118}
119
120late_initcall(fail_futex_debugfs);
121
122#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
123
124#endif /* CONFIG_FAIL_FUTEX */
125
126static struct futex_hash_bucket *
127__futex_hash(union futex_key *key, struct futex_private_hash *fph);
128
129#ifdef CONFIG_FUTEX_PRIVATE_HASH
130static bool futex_ref_get(struct futex_private_hash *fph);
131static bool futex_ref_put(struct futex_private_hash *fph);
132static bool futex_ref_is_dead(struct futex_private_hash *fph);
133
134enum { FR_PERCPU = 0, FR_ATOMIC };
135
136static inline bool futex_key_is_private(union futex_key *key)
137{
138 /*
139 * Relies on get_futex_key() to set either bit for shared
140 * futexes -- see comment with union futex_key.
141 */
142 return !(key->both.offset & (FUT_OFF_INODE | FUT_OFF_MMSHARED));
143}
144
145static bool futex_private_hash_get(struct futex_private_hash *fph)
146{
147 return futex_ref_get(fph);
148}
149
150void futex_private_hash_put(struct futex_private_hash *fph)
151{
152 if (futex_ref_put(fph))
153 wake_up_var(var: fph->mm);
154}
155
156/**
157 * futex_hash_get - Get an additional reference for the local hash.
158 * @hb: ptr to the private local hash.
159 *
160 * Obtain an additional reference for the already obtained hash bucket. The
161 * caller must already own an reference.
162 */
163void futex_hash_get(struct futex_hash_bucket *hb)
164{
165 struct futex_private_hash *fph = hb->priv;
166
167 if (!fph)
168 return;
169 WARN_ON_ONCE(!futex_private_hash_get(fph));
170}
171
172void futex_hash_put(struct futex_hash_bucket *hb)
173{
174 struct futex_private_hash *fph = hb->priv;
175
176 if (!fph)
177 return;
178 futex_private_hash_put(fph);
179}
180
181static struct futex_hash_bucket *
182__futex_hash_private(union futex_key *key, struct futex_private_hash *fph)
183{
184 u32 hash;
185
186 if (!futex_key_is_private(key))
187 return NULL;
188
189 if (!fph)
190 fph = rcu_dereference(key->private.mm->futex_phash);
191 if (!fph || !fph->hash_mask)
192 return NULL;
193
194 hash = jhash2(k: (void *)&key->private.address,
195 length: sizeof(key->private.address) / 4,
196 initval: key->both.offset);
197 return &fph->queues[hash & fph->hash_mask];
198}
199
200static void futex_rehash_private(struct futex_private_hash *old,
201 struct futex_private_hash *new)
202{
203 struct futex_hash_bucket *hb_old, *hb_new;
204 unsigned int slots = old->hash_mask + 1;
205 unsigned int i;
206
207 for (i = 0; i < slots; i++) {
208 struct futex_q *this, *tmp;
209
210 hb_old = &old->queues[i];
211
212 spin_lock(lock: &hb_old->lock);
213 plist_for_each_entry_safe(this, tmp, &hb_old->chain, list) {
214
215 plist_del(node: &this->list, head: &hb_old->chain);
216 futex_hb_waiters_dec(hb: hb_old);
217
218 WARN_ON_ONCE(this->lock_ptr != &hb_old->lock);
219
220 hb_new = __futex_hash(key: &this->key, fph: new);
221 futex_hb_waiters_inc(hb: hb_new);
222 /*
223 * The new pointer isn't published yet but an already
224 * moved user can be unqueued due to timeout or signal.
225 */
226 spin_lock_nested(&hb_new->lock, SINGLE_DEPTH_NESTING);
227 plist_add(node: &this->list, head: &hb_new->chain);
228 this->lock_ptr = &hb_new->lock;
229 spin_unlock(lock: &hb_new->lock);
230 }
231 spin_unlock(lock: &hb_old->lock);
232 }
233}
234
235static bool __futex_pivot_hash(struct mm_struct *mm,
236 struct futex_private_hash *new)
237{
238 struct futex_private_hash *fph;
239
240 WARN_ON_ONCE(mm->futex_phash_new);
241
242 fph = rcu_dereference_protected(mm->futex_phash,
243 lockdep_is_held(&mm->futex_hash_lock));
244 if (fph) {
245 if (!futex_ref_is_dead(fph)) {
246 mm->futex_phash_new = new;
247 return false;
248 }
249
250 futex_rehash_private(old: fph, new);
251 }
252 new->state = FR_PERCPU;
253 scoped_guard(rcu) {
254 mm->futex_batches = get_state_synchronize_rcu();
255 rcu_assign_pointer(mm->futex_phash, new);
256 }
257 kvfree_rcu(fph, rcu);
258 return true;
259}
260
261static void futex_pivot_hash(struct mm_struct *mm)
262{
263 scoped_guard(mutex, &mm->futex_hash_lock) {
264 struct futex_private_hash *fph;
265
266 fph = mm->futex_phash_new;
267 if (fph) {
268 mm->futex_phash_new = NULL;
269 __futex_pivot_hash(mm, new: fph);
270 }
271 }
272}
273
274struct futex_private_hash *futex_private_hash(void)
275{
276 struct mm_struct *mm = current->mm;
277 /*
278 * Ideally we don't loop. If there is a replacement in progress
279 * then a new private hash is already prepared and a reference can't be
280 * obtained once the last user dropped it's.
281 * In that case we block on mm_struct::futex_hash_lock and either have
282 * to perform the replacement or wait while someone else is doing the
283 * job. Eitherway, on the second iteration we acquire a reference on the
284 * new private hash or loop again because a new replacement has been
285 * requested.
286 */
287again:
288 scoped_guard(rcu) {
289 struct futex_private_hash *fph;
290
291 fph = rcu_dereference(mm->futex_phash);
292 if (!fph)
293 return NULL;
294
295 if (futex_private_hash_get(fph))
296 return fph;
297 }
298 futex_pivot_hash(mm);
299 goto again;
300}
301
302struct futex_hash_bucket *futex_hash(union futex_key *key)
303{
304 struct futex_private_hash *fph;
305 struct futex_hash_bucket *hb;
306
307again:
308 scoped_guard(rcu) {
309 hb = __futex_hash(key, NULL);
310 fph = hb->priv;
311
312 if (!fph || futex_private_hash_get(fph))
313 return hb;
314 }
315 futex_pivot_hash(mm: key->private.mm);
316 goto again;
317}
318
319#else /* !CONFIG_FUTEX_PRIVATE_HASH */
320
321static struct futex_hash_bucket *
322__futex_hash_private(union futex_key *key, struct futex_private_hash *fph)
323{
324 return NULL;
325}
326
327struct futex_hash_bucket *futex_hash(union futex_key *key)
328{
329 return __futex_hash(key, NULL);
330}
331
332#endif /* CONFIG_FUTEX_PRIVATE_HASH */
333
334#ifdef CONFIG_FUTEX_MPOL
335
336static int __futex_key_to_node(struct mm_struct *mm, unsigned long addr)
337{
338 struct vm_area_struct *vma = vma_lookup(mm, addr);
339 struct mempolicy *mpol;
340 int node = FUTEX_NO_NODE;
341
342 if (!vma)
343 return FUTEX_NO_NODE;
344
345 mpol = vma_policy(vma);
346 if (!mpol)
347 return FUTEX_NO_NODE;
348
349 switch (mpol->mode) {
350 case MPOL_PREFERRED:
351 node = first_node(mpol->nodes);
352 break;
353 case MPOL_PREFERRED_MANY:
354 case MPOL_BIND:
355 if (mpol->home_node != NUMA_NO_NODE)
356 node = mpol->home_node;
357 break;
358 default:
359 break;
360 }
361
362 return node;
363}
364
365static int futex_key_to_node_opt(struct mm_struct *mm, unsigned long addr)
366{
367 int seq, node;
368
369 guard(rcu)();
370
371 if (!mmap_lock_speculate_try_begin(mm, seq: &seq))
372 return -EBUSY;
373
374 node = __futex_key_to_node(mm, addr);
375
376 if (mmap_lock_speculate_retry(mm, seq))
377 return -EAGAIN;
378
379 return node;
380}
381
382static int futex_mpol(struct mm_struct *mm, unsigned long addr)
383{
384 int node;
385
386 node = futex_key_to_node_opt(mm, addr);
387 if (node >= FUTEX_NO_NODE)
388 return node;
389
390 guard(mmap_read_lock)(T: mm);
391 return __futex_key_to_node(mm, addr);
392}
393
394#else /* !CONFIG_FUTEX_MPOL */
395
396static int futex_mpol(struct mm_struct *mm, unsigned long addr)
397{
398 return FUTEX_NO_NODE;
399}
400
401#endif /* CONFIG_FUTEX_MPOL */
402
403/**
404 * __futex_hash - Return the hash bucket
405 * @key: Pointer to the futex key for which the hash is calculated
406 * @fph: Pointer to private hash if known
407 *
408 * We hash on the keys returned from get_futex_key (see below) and return the
409 * corresponding hash bucket.
410 * If the FUTEX is PROCESS_PRIVATE then a per-process hash bucket (from the
411 * private hash) is returned if existing. Otherwise a hash bucket from the
412 * global hash is returned.
413 */
414static struct futex_hash_bucket *
415__futex_hash(union futex_key *key, struct futex_private_hash *fph)
416{
417 int node = key->both.node;
418 u32 hash;
419
420 if (node == FUTEX_NO_NODE) {
421 struct futex_hash_bucket *hb;
422
423 hb = __futex_hash_private(key, fph);
424 if (hb)
425 return hb;
426 }
427
428 hash = jhash2(k: (u32 *)key,
429 offsetof(typeof(*key), both.offset) / sizeof(u32),
430 initval: key->both.offset);
431
432 if (node == FUTEX_NO_NODE) {
433 /*
434 * In case of !FLAGS_NUMA, use some unused hash bits to pick a
435 * node -- this ensures regular futexes are interleaved across
436 * the nodes and avoids having to allocate multiple
437 * hash-tables.
438 *
439 * NOTE: this isn't perfectly uniform, but it is fast and
440 * handles sparse node masks.
441 */
442 node = (hash >> futex_hashshift) % nr_node_ids;
443 if (!node_possible(node)) {
444 node = find_next_bit_wrap(node_possible_map.bits,
445 size: nr_node_ids, offset: node);
446 }
447 }
448
449 return &futex_queues[node][hash & futex_hashmask];
450}
451
452/**
453 * futex_setup_timer - set up the sleeping hrtimer.
454 * @time: ptr to the given timeout value
455 * @timeout: the hrtimer_sleeper structure to be set up
456 * @flags: futex flags
457 * @range_ns: optional range in ns
458 *
459 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
460 * value given
461 */
462struct hrtimer_sleeper *
463futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
464 int flags, u64 range_ns)
465{
466 if (!time)
467 return NULL;
468
469 hrtimer_setup_sleeper_on_stack(sl: timeout,
470 clock_id: (flags & FLAGS_CLOCKRT) ? CLOCK_REALTIME : CLOCK_MONOTONIC,
471 mode: HRTIMER_MODE_ABS);
472 /*
473 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
474 * effectively the same as calling hrtimer_set_expires().
475 */
476 hrtimer_set_expires_range_ns(timer: &timeout->timer, time: *time, delta: range_ns);
477
478 return timeout;
479}
480
481/*
482 * Generate a machine wide unique identifier for this inode.
483 *
484 * This relies on u64 not wrapping in the life-time of the machine; which with
485 * 1ns resolution means almost 585 years.
486 *
487 * This further relies on the fact that a well formed program will not unmap
488 * the file while it has a (shared) futex waiting on it. This mapping will have
489 * a file reference which pins the mount and inode.
490 *
491 * If for some reason an inode gets evicted and read back in again, it will get
492 * a new sequence number and will _NOT_ match, even though it is the exact same
493 * file.
494 *
495 * It is important that futex_match() will never have a false-positive, esp.
496 * for PI futexes that can mess up the state. The above argues that false-negatives
497 * are only possible for malformed programs.
498 */
499static u64 get_inode_sequence_number(struct inode *inode)
500{
501 static atomic64_t i_seq;
502 u64 old;
503
504 /* Does the inode already have a sequence number? */
505 old = atomic64_read(v: &inode->i_sequence);
506 if (likely(old))
507 return old;
508
509 for (;;) {
510 u64 new = atomic64_inc_return(v: &i_seq);
511 if (WARN_ON_ONCE(!new))
512 continue;
513
514 old = 0;
515 if (!atomic64_try_cmpxchg_relaxed(v: &inode->i_sequence, old: &old, new))
516 return old;
517 return new;
518 }
519}
520
521/**
522 * get_futex_key() - Get parameters which are the keys for a futex
523 * @uaddr: virtual address of the futex
524 * @flags: FLAGS_*
525 * @key: address where result is stored.
526 * @rw: mapping needs to be read/write (values: FUTEX_READ,
527 * FUTEX_WRITE)
528 *
529 * Return: a negative error code or 0
530 *
531 * The key words are stored in @key on success.
532 *
533 * For shared mappings (when @fshared), the key is:
534 *
535 * ( inode->i_sequence, page offset within mapping, offset_within_page )
536 *
537 * [ also see get_inode_sequence_number() ]
538 *
539 * For private mappings (or when !@fshared), the key is:
540 *
541 * ( current->mm, address, 0 )
542 *
543 * This allows (cross process, where applicable) identification of the futex
544 * without keeping the page pinned for the duration of the FUTEX_WAIT.
545 *
546 * lock_page() might sleep, the caller should not hold a spinlock.
547 */
548int get_futex_key(u32 __user *uaddr, unsigned int flags, union futex_key *key,
549 enum futex_access rw)
550{
551 unsigned long address = (unsigned long)uaddr;
552 struct mm_struct *mm = current->mm;
553 struct page *page;
554 struct folio *folio;
555 struct address_space *mapping;
556 int node, err, size, ro = 0;
557 bool node_updated = false;
558 bool fshared;
559
560 fshared = flags & FLAGS_SHARED;
561 size = futex_size(flags);
562 if (flags & FLAGS_NUMA)
563 size *= 2;
564
565 /*
566 * The futex address must be "naturally" aligned.
567 */
568 key->both.offset = address % PAGE_SIZE;
569 if (unlikely((address % size) != 0))
570 return -EINVAL;
571 address -= key->both.offset;
572
573 if (unlikely(!access_ok(uaddr, size)))
574 return -EFAULT;
575
576 if (unlikely(should_fail_futex(fshared)))
577 return -EFAULT;
578
579 node = FUTEX_NO_NODE;
580
581 if (flags & FLAGS_NUMA) {
582 u32 __user *naddr = (void *)uaddr + size / 2;
583
584 if (futex_get_value(dest: &node, from: naddr))
585 return -EFAULT;
586
587 if ((node != FUTEX_NO_NODE) &&
588 ((unsigned int)node >= MAX_NUMNODES || !node_possible(node)))
589 return -EINVAL;
590 }
591
592 if (node == FUTEX_NO_NODE && (flags & FLAGS_MPOL)) {
593 node = futex_mpol(mm, addr: address);
594 node_updated = true;
595 }
596
597 if (flags & FLAGS_NUMA) {
598 u32 __user *naddr = (void *)uaddr + size / 2;
599
600 if (node == FUTEX_NO_NODE) {
601 node = numa_node_id();
602 node_updated = true;
603 }
604 if (node_updated && futex_put_value(val: node, to: naddr))
605 return -EFAULT;
606 }
607
608 key->both.node = node;
609
610 /*
611 * PROCESS_PRIVATE futexes are fast.
612 * As the mm cannot disappear under us and the 'key' only needs
613 * virtual address, we dont even have to find the underlying vma.
614 * Note : We do have to check 'uaddr' is a valid user address,
615 * but access_ok() should be faster than find_vma()
616 */
617 if (!fshared) {
618 /*
619 * On no-MMU, shared futexes are treated as private, therefore
620 * we must not include the current process in the key. Since
621 * there is only one address space, the address is a unique key
622 * on its own.
623 */
624 if (IS_ENABLED(CONFIG_MMU))
625 key->private.mm = mm;
626 else
627 key->private.mm = NULL;
628
629 key->private.address = address;
630 return 0;
631 }
632
633again:
634 /* Ignore any VERIFY_READ mapping (futex common case) */
635 if (unlikely(should_fail_futex(true)))
636 return -EFAULT;
637
638 err = get_user_pages_fast(start: address, nr_pages: 1, gup_flags: FOLL_WRITE, pages: &page);
639 /*
640 * If write access is not required (eg. FUTEX_WAIT), try
641 * and get read-only access.
642 */
643 if (err == -EFAULT && rw == FUTEX_READ) {
644 err = get_user_pages_fast(start: address, nr_pages: 1, gup_flags: 0, pages: &page);
645 ro = 1;
646 }
647 if (err < 0)
648 return err;
649 else
650 err = 0;
651
652 /*
653 * The treatment of mapping from this point on is critical. The folio
654 * lock protects many things but in this context the folio lock
655 * stabilizes mapping, prevents inode freeing in the shared
656 * file-backed region case and guards against movement to swap cache.
657 *
658 * Strictly speaking the folio lock is not needed in all cases being
659 * considered here and folio lock forces unnecessarily serialization.
660 * From this point on, mapping will be re-verified if necessary and
661 * folio lock will be acquired only if it is unavoidable
662 *
663 * Mapping checks require the folio so it is looked up now. For
664 * anonymous pages, it does not matter if the folio is split
665 * in the future as the key is based on the address. For
666 * filesystem-backed pages, the precise page is required as the
667 * index of the page determines the key.
668 */
669 folio = page_folio(page);
670 mapping = READ_ONCE(folio->mapping);
671
672 /*
673 * If folio->mapping is NULL, then it cannot be an anonymous
674 * page; but it might be the ZERO_PAGE or in the gate area or
675 * in a special mapping (all cases which we are happy to fail);
676 * or it may have been a good file page when get_user_pages_fast
677 * found it, but truncated or holepunched or subjected to
678 * invalidate_complete_page2 before we got the folio lock (also
679 * cases which we are happy to fail). And we hold a reference,
680 * so refcount care in invalidate_inode_page's remove_mapping
681 * prevents drop_caches from setting mapping to NULL beneath us.
682 *
683 * The case we do have to guard against is when memory pressure made
684 * shmem_writepage move it from filecache to swapcache beneath us:
685 * an unlikely race, but we do need to retry for folio->mapping.
686 */
687 if (unlikely(!mapping)) {
688 int shmem_swizzled;
689
690 /*
691 * Folio lock is required to identify which special case above
692 * applies. If this is really a shmem page then the folio lock
693 * will prevent unexpected transitions.
694 */
695 folio_lock(folio);
696 shmem_swizzled = folio_test_swapcache(folio) || folio->mapping;
697 folio_unlock(folio);
698 folio_put(folio);
699
700 if (shmem_swizzled)
701 goto again;
702
703 return -EFAULT;
704 }
705
706 /*
707 * Private mappings are handled in a simple way.
708 *
709 * If the futex key is stored in anonymous memory, then the associated
710 * object is the mm which is implicitly pinned by the calling process.
711 *
712 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
713 * it's a read-only handle, it's expected that futexes attach to
714 * the object not the particular process.
715 */
716 if (folio_test_anon(folio)) {
717 /*
718 * A RO anonymous page will never change and thus doesn't make
719 * sense for futex operations.
720 */
721 if (unlikely(should_fail_futex(true)) || ro) {
722 err = -EFAULT;
723 goto out;
724 }
725
726 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
727 key->private.mm = mm;
728 key->private.address = address;
729
730 } else {
731 struct inode *inode;
732
733 /*
734 * The associated futex object in this case is the inode and
735 * the folio->mapping must be traversed. Ordinarily this should
736 * be stabilised under folio lock but it's not strictly
737 * necessary in this case as we just want to pin the inode, not
738 * update i_pages or anything like that.
739 *
740 * The RCU read lock is taken as the inode is finally freed
741 * under RCU. If the mapping still matches expectations then the
742 * mapping->host can be safely accessed as being a valid inode.
743 */
744 rcu_read_lock();
745
746 if (READ_ONCE(folio->mapping) != mapping) {
747 rcu_read_unlock();
748 folio_put(folio);
749
750 goto again;
751 }
752
753 inode = READ_ONCE(mapping->host);
754 if (!inode) {
755 rcu_read_unlock();
756 folio_put(folio);
757
758 goto again;
759 }
760
761 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
762 key->shared.i_seq = get_inode_sequence_number(inode);
763 key->shared.pgoff = page_pgoff(folio, page);
764 rcu_read_unlock();
765 }
766
767out:
768 folio_put(folio);
769 return err;
770}
771
772/**
773 * fault_in_user_writeable() - Fault in user address and verify RW access
774 * @uaddr: pointer to faulting user space address
775 *
776 * Slow path to fixup the fault we just took in the atomic write
777 * access to @uaddr.
778 *
779 * We have no generic implementation of a non-destructive write to the
780 * user address. We know that we faulted in the atomic pagefault
781 * disabled section so we can as well avoid the #PF overhead by
782 * calling get_user_pages() right away.
783 */
784int fault_in_user_writeable(u32 __user *uaddr)
785{
786 struct mm_struct *mm = current->mm;
787 int ret;
788
789 mmap_read_lock(mm);
790 ret = fixup_user_fault(mm, address: (unsigned long)uaddr,
791 fault_flags: FAULT_FLAG_WRITE, NULL);
792 mmap_read_unlock(mm);
793
794 return ret < 0 ? ret : 0;
795}
796
797/**
798 * futex_top_waiter() - Return the highest priority waiter on a futex
799 * @hb: the hash bucket the futex_q's reside in
800 * @key: the futex key (to distinguish it from other futex futex_q's)
801 *
802 * Must be called with the hb lock held.
803 */
804struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, union futex_key *key)
805{
806 struct futex_q *this;
807
808 plist_for_each_entry(this, &hb->chain, list) {
809 if (futex_match(key1: &this->key, key2: key))
810 return this;
811 }
812 return NULL;
813}
814
815/**
816 * wait_for_owner_exiting - Block until the owner has exited
817 * @ret: owner's current futex lock status
818 * @exiting: Pointer to the exiting task
819 *
820 * Caller must hold a refcount on @exiting.
821 */
822void wait_for_owner_exiting(int ret, struct task_struct *exiting)
823{
824 if (ret != -EBUSY) {
825 WARN_ON_ONCE(exiting);
826 return;
827 }
828
829 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
830 return;
831
832 mutex_lock(lock: &exiting->futex_exit_mutex);
833 /*
834 * No point in doing state checking here. If the waiter got here
835 * while the task was in exec()->exec_futex_release() then it can
836 * have any FUTEX_STATE_* value when the waiter has acquired the
837 * mutex. OK, if running, EXITING or DEAD if it reached exit()
838 * already. Highly unlikely and not a problem. Just one more round
839 * through the futex maze.
840 */
841 mutex_unlock(lock: &exiting->futex_exit_mutex);
842
843 put_task_struct(t: exiting);
844}
845
846/**
847 * __futex_unqueue() - Remove the futex_q from its futex_hash_bucket
848 * @q: The futex_q to unqueue
849 *
850 * The q->lock_ptr must not be NULL and must be held by the caller.
851 */
852void __futex_unqueue(struct futex_q *q)
853{
854 struct futex_hash_bucket *hb;
855
856 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
857 return;
858 lockdep_assert_held(q->lock_ptr);
859
860 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
861 plist_del(node: &q->list, head: &hb->chain);
862 futex_hb_waiters_dec(hb);
863}
864
865/* The key must be already stored in q->key. */
866void futex_q_lock(struct futex_q *q, struct futex_hash_bucket *hb)
867 __acquires(&hb->lock)
868{
869 /*
870 * Increment the counter before taking the lock so that
871 * a potential waker won't miss a to-be-slept task that is
872 * waiting for the spinlock. This is safe as all futex_q_lock()
873 * users end up calling futex_queue(). Similarly, for housekeeping,
874 * decrement the counter at futex_q_unlock() when some error has
875 * occurred and we don't end up adding the task to the list.
876 */
877 futex_hb_waiters_inc(hb); /* implies smp_mb(); (A) */
878
879 q->lock_ptr = &hb->lock;
880
881 spin_lock(lock: &hb->lock);
882}
883
884void futex_q_unlock(struct futex_hash_bucket *hb)
885 __releases(&hb->lock)
886{
887 futex_hb_waiters_dec(hb);
888 spin_unlock(lock: &hb->lock);
889}
890
891void __futex_queue(struct futex_q *q, struct futex_hash_bucket *hb,
892 struct task_struct *task)
893{
894 int prio;
895
896 /*
897 * The priority used to register this element is
898 * - either the real thread-priority for the real-time threads
899 * (i.e. threads with a priority lower than MAX_RT_PRIO)
900 * - or MAX_RT_PRIO for non-RT threads.
901 * Thus, all RT-threads are woken first in priority order, and
902 * the others are woken last, in FIFO order.
903 */
904 prio = min(current->normal_prio, MAX_RT_PRIO);
905
906 plist_node_init(node: &q->list, prio);
907 plist_add(node: &q->list, head: &hb->chain);
908 q->task = task;
909}
910
911/**
912 * futex_unqueue() - Remove the futex_q from its futex_hash_bucket
913 * @q: The futex_q to unqueue
914 *
915 * The q->lock_ptr must not be held by the caller. A call to futex_unqueue() must
916 * be paired with exactly one earlier call to futex_queue().
917 *
918 * Return:
919 * - 1 - if the futex_q was still queued (and we removed unqueued it);
920 * - 0 - if the futex_q was already removed by the waking thread
921 */
922int futex_unqueue(struct futex_q *q)
923{
924 spinlock_t *lock_ptr;
925 int ret = 0;
926
927 /* RCU so lock_ptr is not going away during locking. */
928 guard(rcu)();
929 /* In the common case we don't take the spinlock, which is nice. */
930retry:
931 /*
932 * q->lock_ptr can change between this read and the following spin_lock.
933 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
934 * optimizing lock_ptr out of the logic below.
935 */
936 lock_ptr = READ_ONCE(q->lock_ptr);
937 if (lock_ptr != NULL) {
938 spin_lock(lock: lock_ptr);
939 /*
940 * q->lock_ptr can change between reading it and
941 * spin_lock(), causing us to take the wrong lock. This
942 * corrects the race condition.
943 *
944 * Reasoning goes like this: if we have the wrong lock,
945 * q->lock_ptr must have changed (maybe several times)
946 * between reading it and the spin_lock(). It can
947 * change again after the spin_lock() but only if it was
948 * already changed before the spin_lock(). It cannot,
949 * however, change back to the original value. Therefore
950 * we can detect whether we acquired the correct lock.
951 */
952 if (unlikely(lock_ptr != q->lock_ptr)) {
953 spin_unlock(lock: lock_ptr);
954 goto retry;
955 }
956 __futex_unqueue(q);
957
958 BUG_ON(q->pi_state);
959
960 spin_unlock(lock: lock_ptr);
961 ret = 1;
962 }
963
964 return ret;
965}
966
967void futex_q_lockptr_lock(struct futex_q *q)
968{
969 spinlock_t *lock_ptr;
970
971 /*
972 * See futex_unqueue() why lock_ptr can change.
973 */
974 guard(rcu)();
975retry:
976 lock_ptr = READ_ONCE(q->lock_ptr);
977 spin_lock(lock: lock_ptr);
978
979 if (unlikely(lock_ptr != q->lock_ptr)) {
980 spin_unlock(lock: lock_ptr);
981 goto retry;
982 }
983}
984
985/*
986 * PI futexes can not be requeued and must remove themselves from the hash
987 * bucket. The hash bucket lock (i.e. lock_ptr) is held.
988 */
989void futex_unqueue_pi(struct futex_q *q)
990{
991 /*
992 * If the lock was not acquired (due to timeout or signal) then the
993 * rt_waiter is removed before futex_q is. If this is observed by
994 * an unlocker after dropping the rtmutex wait lock and before
995 * acquiring the hash bucket lock, then the unlocker dequeues the
996 * futex_q from the hash bucket list to guarantee consistent state
997 * vs. userspace. Therefore the dequeue here must be conditional.
998 */
999 if (!plist_node_empty(node: &q->list))
1000 __futex_unqueue(q);
1001
1002 BUG_ON(!q->pi_state);
1003 put_pi_state(pi_state: q->pi_state);
1004 q->pi_state = NULL;
1005}
1006
1007/* Constants for the pending_op argument of handle_futex_death */
1008#define HANDLE_DEATH_PENDING true
1009#define HANDLE_DEATH_LIST false
1010
1011/*
1012 * Process a futex-list entry, check whether it's owned by the
1013 * dying task, and do notification if so:
1014 */
1015static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
1016 bool pi, bool pending_op)
1017{
1018 u32 uval, nval, mval;
1019 pid_t owner;
1020 int err;
1021
1022 /* Futex address must be 32bit aligned */
1023 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
1024 return -1;
1025
1026retry:
1027 if (get_user(uval, uaddr))
1028 return -1;
1029
1030 /*
1031 * Special case for regular (non PI) futexes. The unlock path in
1032 * user space has two race scenarios:
1033 *
1034 * 1. The unlock path releases the user space futex value and
1035 * before it can execute the futex() syscall to wake up
1036 * waiters it is killed.
1037 *
1038 * 2. A woken up waiter is killed before it can acquire the
1039 * futex in user space.
1040 *
1041 * In the second case, the wake up notification could be generated
1042 * by the unlock path in user space after setting the futex value
1043 * to zero or by the kernel after setting the OWNER_DIED bit below.
1044 *
1045 * In both cases the TID validation below prevents a wakeup of
1046 * potential waiters which can cause these waiters to block
1047 * forever.
1048 *
1049 * In both cases the following conditions are met:
1050 *
1051 * 1) task->robust_list->list_op_pending != NULL
1052 * @pending_op == true
1053 * 2) The owner part of user space futex value == 0
1054 * 3) Regular futex: @pi == false
1055 *
1056 * If these conditions are met, it is safe to attempt waking up a
1057 * potential waiter without touching the user space futex value and
1058 * trying to set the OWNER_DIED bit. If the futex value is zero,
1059 * the rest of the user space mutex state is consistent, so a woken
1060 * waiter will just take over the uncontended futex. Setting the
1061 * OWNER_DIED bit would create inconsistent state and malfunction
1062 * of the user space owner died handling. Otherwise, the OWNER_DIED
1063 * bit is already set, and the woken waiter is expected to deal with
1064 * this.
1065 */
1066 owner = uval & FUTEX_TID_MASK;
1067
1068 if (pending_op && !pi && !owner) {
1069 futex_wake(uaddr, FLAGS_SIZE_32 | FLAGS_SHARED, nr_wake: 1,
1070 FUTEX_BITSET_MATCH_ANY);
1071 return 0;
1072 }
1073
1074 if (owner != task_pid_vnr(tsk: curr))
1075 return 0;
1076
1077 /*
1078 * Ok, this dying thread is truly holding a futex
1079 * of interest. Set the OWNER_DIED bit atomically
1080 * via cmpxchg, and if the value had FUTEX_WAITERS
1081 * set, wake up a waiter (if any). (We have to do a
1082 * futex_wake() even if OWNER_DIED is already set -
1083 * to handle the rare but possible case of recursive
1084 * thread-death.) The rest of the cleanup is done in
1085 * userspace.
1086 */
1087 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
1088
1089 /*
1090 * We are not holding a lock here, but we want to have
1091 * the pagefault_disable/enable() protection because
1092 * we want to handle the fault gracefully. If the
1093 * access fails we try to fault in the futex with R/W
1094 * verification via get_user_pages. get_user() above
1095 * does not guarantee R/W access. If that fails we
1096 * give up and leave the futex locked.
1097 */
1098 if ((err = futex_cmpxchg_value_locked(curval: &nval, uaddr, uval, newval: mval))) {
1099 switch (err) {
1100 case -EFAULT:
1101 if (fault_in_user_writeable(uaddr))
1102 return -1;
1103 goto retry;
1104
1105 case -EAGAIN:
1106 cond_resched();
1107 goto retry;
1108
1109 default:
1110 WARN_ON_ONCE(1);
1111 return err;
1112 }
1113 }
1114
1115 if (nval != uval)
1116 goto retry;
1117
1118 /*
1119 * Wake robust non-PI futexes here. The wakeup of
1120 * PI futexes happens in exit_pi_state():
1121 */
1122 if (!pi && (uval & FUTEX_WAITERS)) {
1123 futex_wake(uaddr, FLAGS_SIZE_32 | FLAGS_SHARED, nr_wake: 1,
1124 FUTEX_BITSET_MATCH_ANY);
1125 }
1126
1127 return 0;
1128}
1129
1130/*
1131 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1132 */
1133static inline int fetch_robust_entry(struct robust_list __user **entry,
1134 struct robust_list __user * __user *head,
1135 unsigned int *pi)
1136{
1137 unsigned long uentry;
1138
1139 if (get_user(uentry, (unsigned long __user *)head))
1140 return -EFAULT;
1141
1142 *entry = (void __user *)(uentry & ~1UL);
1143 *pi = uentry & 1;
1144
1145 return 0;
1146}
1147
1148/*
1149 * Walk curr->robust_list (very carefully, it's a userspace list!)
1150 * and mark any locks found there dead, and notify any waiters.
1151 *
1152 * We silently return on any sign of list-walking problem.
1153 */
1154static void exit_robust_list(struct task_struct *curr)
1155{
1156 struct robust_list_head __user *head = curr->robust_list;
1157 struct robust_list __user *entry, *next_entry, *pending;
1158 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1159 unsigned int next_pi;
1160 unsigned long futex_offset;
1161 int rc;
1162
1163 /*
1164 * Fetch the list head (which was registered earlier, via
1165 * sys_set_robust_list()):
1166 */
1167 if (fetch_robust_entry(entry: &entry, head: &head->list.next, pi: &pi))
1168 return;
1169 /*
1170 * Fetch the relative futex offset:
1171 */
1172 if (get_user(futex_offset, &head->futex_offset))
1173 return;
1174 /*
1175 * Fetch any possibly pending lock-add first, and handle it
1176 * if it exists:
1177 */
1178 if (fetch_robust_entry(entry: &pending, head: &head->list_op_pending, pi: &pip))
1179 return;
1180
1181 next_entry = NULL; /* avoid warning with gcc */
1182 while (entry != &head->list) {
1183 /*
1184 * Fetch the next entry in the list before calling
1185 * handle_futex_death:
1186 */
1187 rc = fetch_robust_entry(entry: &next_entry, head: &entry->next, pi: &next_pi);
1188 /*
1189 * A pending lock might already be on the list, so
1190 * don't process it twice:
1191 */
1192 if (entry != pending) {
1193 if (handle_futex_death(uaddr: (void __user *)entry + futex_offset,
1194 curr, pi, HANDLE_DEATH_LIST))
1195 return;
1196 }
1197 if (rc)
1198 return;
1199 entry = next_entry;
1200 pi = next_pi;
1201 /*
1202 * Avoid excessively long or circular lists:
1203 */
1204 if (!--limit)
1205 break;
1206
1207 cond_resched();
1208 }
1209
1210 if (pending) {
1211 handle_futex_death(uaddr: (void __user *)pending + futex_offset,
1212 curr, pi: pip, HANDLE_DEATH_PENDING);
1213 }
1214}
1215
1216#ifdef CONFIG_COMPAT
1217static void __user *futex_uaddr(struct robust_list __user *entry,
1218 compat_long_t futex_offset)
1219{
1220 compat_uptr_t base = ptr_to_compat(uptr: entry);
1221 void __user *uaddr = compat_ptr(uptr: base + futex_offset);
1222
1223 return uaddr;
1224}
1225
1226/*
1227 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
1228 */
1229static inline int
1230compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
1231 compat_uptr_t __user *head, unsigned int *pi)
1232{
1233 if (get_user(*uentry, head))
1234 return -EFAULT;
1235
1236 *entry = compat_ptr(uptr: (*uentry) & ~1);
1237 *pi = (unsigned int)(*uentry) & 1;
1238
1239 return 0;
1240}
1241
1242/*
1243 * Walk curr->robust_list (very carefully, it's a userspace list!)
1244 * and mark any locks found there dead, and notify any waiters.
1245 *
1246 * We silently return on any sign of list-walking problem.
1247 */
1248static void compat_exit_robust_list(struct task_struct *curr)
1249{
1250 struct compat_robust_list_head __user *head = curr->compat_robust_list;
1251 struct robust_list __user *entry, *next_entry, *pending;
1252 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
1253 unsigned int next_pi;
1254 compat_uptr_t uentry, next_uentry, upending;
1255 compat_long_t futex_offset;
1256 int rc;
1257
1258 /*
1259 * Fetch the list head (which was registered earlier, via
1260 * sys_set_robust_list()):
1261 */
1262 if (compat_fetch_robust_entry(uentry: &uentry, entry: &entry, head: &head->list.next, pi: &pi))
1263 return;
1264 /*
1265 * Fetch the relative futex offset:
1266 */
1267 if (get_user(futex_offset, &head->futex_offset))
1268 return;
1269 /*
1270 * Fetch any possibly pending lock-add first, and handle it
1271 * if it exists:
1272 */
1273 if (compat_fetch_robust_entry(uentry: &upending, entry: &pending,
1274 head: &head->list_op_pending, pi: &pip))
1275 return;
1276
1277 next_entry = NULL; /* avoid warning with gcc */
1278 while (entry != (struct robust_list __user *) &head->list) {
1279 /*
1280 * Fetch the next entry in the list before calling
1281 * handle_futex_death:
1282 */
1283 rc = compat_fetch_robust_entry(uentry: &next_uentry, entry: &next_entry,
1284 head: (compat_uptr_t __user *)&entry->next, pi: &next_pi);
1285 /*
1286 * A pending lock might already be on the list, so
1287 * dont process it twice:
1288 */
1289 if (entry != pending) {
1290 void __user *uaddr = futex_uaddr(entry, futex_offset);
1291
1292 if (handle_futex_death(uaddr, curr, pi,
1293 HANDLE_DEATH_LIST))
1294 return;
1295 }
1296 if (rc)
1297 return;
1298 uentry = next_uentry;
1299 entry = next_entry;
1300 pi = next_pi;
1301 /*
1302 * Avoid excessively long or circular lists:
1303 */
1304 if (!--limit)
1305 break;
1306
1307 cond_resched();
1308 }
1309 if (pending) {
1310 void __user *uaddr = futex_uaddr(entry: pending, futex_offset);
1311
1312 handle_futex_death(uaddr, curr, pi: pip, HANDLE_DEATH_PENDING);
1313 }
1314}
1315#endif
1316
1317#ifdef CONFIG_FUTEX_PI
1318
1319/*
1320 * This task is holding PI mutexes at exit time => bad.
1321 * Kernel cleans up PI-state, but userspace is likely hosed.
1322 * (Robust-futex cleanup is separate and might save the day for userspace.)
1323 */
1324static void exit_pi_state_list(struct task_struct *curr)
1325{
1326 struct list_head *next, *head = &curr->pi_state_list;
1327 struct futex_pi_state *pi_state;
1328 union futex_key key = FUTEX_KEY_INIT;
1329
1330 /*
1331 * The mutex mm_struct::futex_hash_lock might be acquired.
1332 */
1333 might_sleep();
1334 /*
1335 * Ensure the hash remains stable (no resize) during the while loop
1336 * below. The hb pointer is acquired under the pi_lock so we can't block
1337 * on the mutex.
1338 */
1339 WARN_ON(curr != current);
1340 guard(private_hash)();
1341 /*
1342 * We are a ZOMBIE and nobody can enqueue itself on
1343 * pi_state_list anymore, but we have to be careful
1344 * versus waiters unqueueing themselves:
1345 */
1346 raw_spin_lock_irq(&curr->pi_lock);
1347 while (!list_empty(head)) {
1348 next = head->next;
1349 pi_state = list_entry(next, struct futex_pi_state, list);
1350 key = pi_state->key;
1351 if (1) {
1352 CLASS(hb, hb)(key: &key);
1353
1354 /*
1355 * We can race against put_pi_state() removing itself from the
1356 * list (a waiter going away). put_pi_state() will first
1357 * decrement the reference count and then modify the list, so
1358 * its possible to see the list entry but fail this reference
1359 * acquire.
1360 *
1361 * In that case; drop the locks to let put_pi_state() make
1362 * progress and retry the loop.
1363 */
1364 if (!refcount_inc_not_zero(r: &pi_state->refcount)) {
1365 raw_spin_unlock_irq(&curr->pi_lock);
1366 cpu_relax();
1367 raw_spin_lock_irq(&curr->pi_lock);
1368 continue;
1369 }
1370 raw_spin_unlock_irq(&curr->pi_lock);
1371
1372 spin_lock(lock: &hb->lock);
1373 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1374 raw_spin_lock(&curr->pi_lock);
1375 /*
1376 * We dropped the pi-lock, so re-check whether this
1377 * task still owns the PI-state:
1378 */
1379 if (head->next != next) {
1380 /* retain curr->pi_lock for the loop invariant */
1381 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1382 spin_unlock(lock: &hb->lock);
1383 put_pi_state(pi_state);
1384 continue;
1385 }
1386
1387 WARN_ON(pi_state->owner != curr);
1388 WARN_ON(list_empty(&pi_state->list));
1389 list_del_init(entry: &pi_state->list);
1390 pi_state->owner = NULL;
1391
1392 raw_spin_unlock(&curr->pi_lock);
1393 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1394 spin_unlock(lock: &hb->lock);
1395 }
1396
1397 rt_mutex_futex_unlock(lock: &pi_state->pi_mutex);
1398 put_pi_state(pi_state);
1399
1400 raw_spin_lock_irq(&curr->pi_lock);
1401 }
1402 raw_spin_unlock_irq(&curr->pi_lock);
1403}
1404#else
1405static inline void exit_pi_state_list(struct task_struct *curr) { }
1406#endif
1407
1408static void futex_cleanup(struct task_struct *tsk)
1409{
1410 if (unlikely(tsk->robust_list)) {
1411 exit_robust_list(curr: tsk);
1412 tsk->robust_list = NULL;
1413 }
1414
1415#ifdef CONFIG_COMPAT
1416 if (unlikely(tsk->compat_robust_list)) {
1417 compat_exit_robust_list(curr: tsk);
1418 tsk->compat_robust_list = NULL;
1419 }
1420#endif
1421
1422 if (unlikely(!list_empty(&tsk->pi_state_list)))
1423 exit_pi_state_list(curr: tsk);
1424}
1425
1426/**
1427 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
1428 * @tsk: task to set the state on
1429 *
1430 * Set the futex exit state of the task lockless. The futex waiter code
1431 * observes that state when a task is exiting and loops until the task has
1432 * actually finished the futex cleanup. The worst case for this is that the
1433 * waiter runs through the wait loop until the state becomes visible.
1434 *
1435 * This is called from the recursive fault handling path in make_task_dead().
1436 *
1437 * This is best effort. Either the futex exit code has run already or
1438 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
1439 * take it over. If not, the problem is pushed back to user space. If the
1440 * futex exit code did not run yet, then an already queued waiter might
1441 * block forever, but there is nothing which can be done about that.
1442 */
1443void futex_exit_recursive(struct task_struct *tsk)
1444{
1445 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
1446 if (tsk->futex_state == FUTEX_STATE_EXITING)
1447 mutex_unlock(lock: &tsk->futex_exit_mutex);
1448 tsk->futex_state = FUTEX_STATE_DEAD;
1449}
1450
1451static void futex_cleanup_begin(struct task_struct *tsk)
1452{
1453 /*
1454 * Prevent various race issues against a concurrent incoming waiter
1455 * including live locks by forcing the waiter to block on
1456 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
1457 * attach_to_pi_owner().
1458 */
1459 mutex_lock(lock: &tsk->futex_exit_mutex);
1460
1461 /*
1462 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
1463 *
1464 * This ensures that all subsequent checks of tsk->futex_state in
1465 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
1466 * tsk->pi_lock held.
1467 *
1468 * It guarantees also that a pi_state which was queued right before
1469 * the state change under tsk->pi_lock by a concurrent waiter must
1470 * be observed in exit_pi_state_list().
1471 */
1472 raw_spin_lock_irq(&tsk->pi_lock);
1473 tsk->futex_state = FUTEX_STATE_EXITING;
1474 raw_spin_unlock_irq(&tsk->pi_lock);
1475}
1476
1477static void futex_cleanup_end(struct task_struct *tsk, int state)
1478{
1479 /*
1480 * Lockless store. The only side effect is that an observer might
1481 * take another loop until it becomes visible.
1482 */
1483 tsk->futex_state = state;
1484 /*
1485 * Drop the exit protection. This unblocks waiters which observed
1486 * FUTEX_STATE_EXITING to reevaluate the state.
1487 */
1488 mutex_unlock(lock: &tsk->futex_exit_mutex);
1489}
1490
1491void futex_exec_release(struct task_struct *tsk)
1492{
1493 /*
1494 * The state handling is done for consistency, but in the case of
1495 * exec() there is no way to prevent further damage as the PID stays
1496 * the same. But for the unlikely and arguably buggy case that a
1497 * futex is held on exec(), this provides at least as much state
1498 * consistency protection which is possible.
1499 */
1500 futex_cleanup_begin(tsk);
1501 futex_cleanup(tsk);
1502 /*
1503 * Reset the state to FUTEX_STATE_OK. The task is alive and about
1504 * exec a new binary.
1505 */
1506 futex_cleanup_end(tsk, state: FUTEX_STATE_OK);
1507}
1508
1509void futex_exit_release(struct task_struct *tsk)
1510{
1511 futex_cleanup_begin(tsk);
1512 futex_cleanup(tsk);
1513 futex_cleanup_end(tsk, state: FUTEX_STATE_DEAD);
1514}
1515
1516static void futex_hash_bucket_init(struct futex_hash_bucket *fhb,
1517 struct futex_private_hash *fph)
1518{
1519#ifdef CONFIG_FUTEX_PRIVATE_HASH
1520 fhb->priv = fph;
1521#endif
1522 atomic_set(v: &fhb->waiters, i: 0);
1523 plist_head_init(head: &fhb->chain);
1524 spin_lock_init(&fhb->lock);
1525}
1526
1527#define FH_CUSTOM 0x01
1528
1529#ifdef CONFIG_FUTEX_PRIVATE_HASH
1530
1531/*
1532 * futex-ref
1533 *
1534 * Heavily inspired by percpu-rwsem/percpu-refcount; not reusing any of that
1535 * code because it just doesn't fit right.
1536 *
1537 * Dual counter, per-cpu / atomic approach like percpu-refcount, except it
1538 * re-initializes the state automatically, such that the fph swizzle is also a
1539 * transition back to per-cpu.
1540 */
1541
1542static void futex_ref_rcu(struct rcu_head *head);
1543
1544static void __futex_ref_atomic_begin(struct futex_private_hash *fph)
1545{
1546 struct mm_struct *mm = fph->mm;
1547
1548 /*
1549 * The counter we're about to switch to must have fully switched;
1550 * otherwise it would be impossible for it to have reported success
1551 * from futex_ref_is_dead().
1552 */
1553 WARN_ON_ONCE(atomic_long_read(&mm->futex_atomic) != 0);
1554
1555 /*
1556 * Set the atomic to the bias value such that futex_ref_{get,put}()
1557 * will never observe 0. Will be fixed up in __futex_ref_atomic_end()
1558 * when folding in the percpu count.
1559 */
1560 atomic_long_set(v: &mm->futex_atomic, LONG_MAX);
1561 smp_store_release(&fph->state, FR_ATOMIC);
1562
1563 call_rcu_hurry(head: &mm->futex_rcu, func: futex_ref_rcu);
1564}
1565
1566static void __futex_ref_atomic_end(struct futex_private_hash *fph)
1567{
1568 struct mm_struct *mm = fph->mm;
1569 unsigned int count = 0;
1570 long ret;
1571 int cpu;
1572
1573 /*
1574 * Per __futex_ref_atomic_begin() the state of the fph must be ATOMIC
1575 * and per this RCU callback, everybody must now observe this state and
1576 * use the atomic variable.
1577 */
1578 WARN_ON_ONCE(fph->state != FR_ATOMIC);
1579
1580 /*
1581 * Therefore the per-cpu counter is now stable, sum and reset.
1582 */
1583 for_each_possible_cpu(cpu) {
1584 unsigned int *ptr = per_cpu_ptr(mm->futex_ref, cpu);
1585 count += *ptr;
1586 *ptr = 0;
1587 }
1588
1589 /*
1590 * Re-init for the next cycle.
1591 */
1592 this_cpu_inc(*mm->futex_ref); /* 0 -> 1 */
1593
1594 /*
1595 * Add actual count, subtract bias and initial refcount.
1596 *
1597 * The moment this atomic operation happens, futex_ref_is_dead() can
1598 * become true.
1599 */
1600 ret = atomic_long_add_return(i: count - LONG_MAX - 1, v: &mm->futex_atomic);
1601 if (!ret)
1602 wake_up_var(var: mm);
1603
1604 WARN_ON_ONCE(ret < 0);
1605 mmput_async(mm);
1606}
1607
1608static void futex_ref_rcu(struct rcu_head *head)
1609{
1610 struct mm_struct *mm = container_of(head, struct mm_struct, futex_rcu);
1611 struct futex_private_hash *fph = rcu_dereference_raw(mm->futex_phash);
1612
1613 if (fph->state == FR_PERCPU) {
1614 /*
1615 * Per this extra grace-period, everybody must now observe
1616 * fph as the current fph and no previously observed fph's
1617 * are in-flight.
1618 *
1619 * Notably, nobody will now rely on the atomic
1620 * futex_ref_is_dead() state anymore so we can begin the
1621 * migration of the per-cpu counter into the atomic.
1622 */
1623 __futex_ref_atomic_begin(fph);
1624 return;
1625 }
1626
1627 __futex_ref_atomic_end(fph);
1628}
1629
1630/*
1631 * Drop the initial refcount and transition to atomics.
1632 */
1633static void futex_ref_drop(struct futex_private_hash *fph)
1634{
1635 struct mm_struct *mm = fph->mm;
1636
1637 /*
1638 * Can only transition the current fph;
1639 */
1640 WARN_ON_ONCE(rcu_dereference_raw(mm->futex_phash) != fph);
1641 /*
1642 * We enqueue at least one RCU callback. Ensure mm stays if the task
1643 * exits before the transition is completed.
1644 */
1645 mmget(mm);
1646
1647 /*
1648 * In order to avoid the following scenario:
1649 *
1650 * futex_hash() __futex_pivot_hash()
1651 * guard(rcu); guard(mm->futex_hash_lock);
1652 * fph = mm->futex_phash;
1653 * rcu_assign_pointer(&mm->futex_phash, new);
1654 * futex_hash_allocate()
1655 * futex_ref_drop()
1656 * fph->state = FR_ATOMIC;
1657 * atomic_set(, BIAS);
1658 *
1659 * futex_private_hash_get(fph); // OOPS
1660 *
1661 * Where an old fph (which is FR_ATOMIC) and should fail on
1662 * inc_not_zero, will succeed because a new transition is started and
1663 * the atomic is bias'ed away from 0.
1664 *
1665 * There must be at least one full grace-period between publishing a
1666 * new fph and trying to replace it.
1667 */
1668 if (poll_state_synchronize_rcu(oldstate: mm->futex_batches)) {
1669 /*
1670 * There was a grace-period, we can begin now.
1671 */
1672 __futex_ref_atomic_begin(fph);
1673 return;
1674 }
1675
1676 call_rcu_hurry(head: &mm->futex_rcu, func: futex_ref_rcu);
1677}
1678
1679static bool futex_ref_get(struct futex_private_hash *fph)
1680{
1681 struct mm_struct *mm = fph->mm;
1682
1683 guard(rcu)();
1684
1685 if (smp_load_acquire(&fph->state) == FR_PERCPU) {
1686 this_cpu_inc(*mm->futex_ref);
1687 return true;
1688 }
1689
1690 return atomic_long_inc_not_zero(v: &mm->futex_atomic);
1691}
1692
1693static bool futex_ref_put(struct futex_private_hash *fph)
1694{
1695 struct mm_struct *mm = fph->mm;
1696
1697 guard(rcu)();
1698
1699 if (smp_load_acquire(&fph->state) == FR_PERCPU) {
1700 this_cpu_dec(*mm->futex_ref);
1701 return false;
1702 }
1703
1704 return atomic_long_dec_and_test(v: &mm->futex_atomic);
1705}
1706
1707static bool futex_ref_is_dead(struct futex_private_hash *fph)
1708{
1709 struct mm_struct *mm = fph->mm;
1710
1711 guard(rcu)();
1712
1713 if (smp_load_acquire(&fph->state) == FR_PERCPU)
1714 return false;
1715
1716 return atomic_long_read(v: &mm->futex_atomic) == 0;
1717}
1718
1719int futex_mm_init(struct mm_struct *mm)
1720{
1721 mutex_init(&mm->futex_hash_lock);
1722 RCU_INIT_POINTER(mm->futex_phash, NULL);
1723 mm->futex_phash_new = NULL;
1724 /* futex-ref */
1725 mm->futex_ref = NULL;
1726 atomic_long_set(v: &mm->futex_atomic, i: 0);
1727 mm->futex_batches = get_state_synchronize_rcu();
1728 return 0;
1729}
1730
1731void futex_hash_free(struct mm_struct *mm)
1732{
1733 struct futex_private_hash *fph;
1734
1735 free_percpu(pdata: mm->futex_ref);
1736 kvfree(addr: mm->futex_phash_new);
1737 fph = rcu_dereference_raw(mm->futex_phash);
1738 if (fph)
1739 kvfree(addr: fph);
1740}
1741
1742static bool futex_pivot_pending(struct mm_struct *mm)
1743{
1744 struct futex_private_hash *fph;
1745
1746 guard(rcu)();
1747
1748 if (!mm->futex_phash_new)
1749 return true;
1750
1751 fph = rcu_dereference(mm->futex_phash);
1752 return futex_ref_is_dead(fph);
1753}
1754
1755static bool futex_hash_less(struct futex_private_hash *a,
1756 struct futex_private_hash *b)
1757{
1758 /* user provided always wins */
1759 if (!a->custom && b->custom)
1760 return true;
1761 if (a->custom && !b->custom)
1762 return false;
1763
1764 /* zero-sized hash wins */
1765 if (!b->hash_mask)
1766 return true;
1767 if (!a->hash_mask)
1768 return false;
1769
1770 /* keep the biggest */
1771 if (a->hash_mask < b->hash_mask)
1772 return true;
1773 if (a->hash_mask > b->hash_mask)
1774 return false;
1775
1776 return false; /* equal */
1777}
1778
1779static int futex_hash_allocate(unsigned int hash_slots, unsigned int flags)
1780{
1781 struct mm_struct *mm = current->mm;
1782 struct futex_private_hash *fph;
1783 bool custom = flags & FH_CUSTOM;
1784 int i;
1785
1786 if (hash_slots && (hash_slots == 1 || !is_power_of_2(n: hash_slots)))
1787 return -EINVAL;
1788
1789 /*
1790 * Once we've disabled the global hash there is no way back.
1791 */
1792 scoped_guard(rcu) {
1793 fph = rcu_dereference(mm->futex_phash);
1794 if (fph && !fph->hash_mask) {
1795 if (custom)
1796 return -EBUSY;
1797 return 0;
1798 }
1799 }
1800
1801 if (!mm->futex_ref) {
1802 /*
1803 * This will always be allocated by the first thread and
1804 * therefore requires no locking.
1805 */
1806 mm->futex_ref = alloc_percpu(unsigned int);
1807 if (!mm->futex_ref)
1808 return -ENOMEM;
1809 this_cpu_inc(*mm->futex_ref); /* 0 -> 1 */
1810 }
1811
1812 fph = kvzalloc(struct_size(fph, queues, hash_slots),
1813 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1814 if (!fph)
1815 return -ENOMEM;
1816
1817 fph->hash_mask = hash_slots ? hash_slots - 1 : 0;
1818 fph->custom = custom;
1819 fph->mm = mm;
1820
1821 for (i = 0; i < hash_slots; i++)
1822 futex_hash_bucket_init(fhb: &fph->queues[i], fph);
1823
1824 if (custom) {
1825 /*
1826 * Only let prctl() wait / retry; don't unduly delay clone().
1827 */
1828again:
1829 wait_var_event(mm, futex_pivot_pending(mm));
1830 }
1831
1832 scoped_guard(mutex, &mm->futex_hash_lock) {
1833 struct futex_private_hash *free __free(kvfree) = NULL;
1834 struct futex_private_hash *cur, *new;
1835
1836 cur = rcu_dereference_protected(mm->futex_phash,
1837 lockdep_is_held(&mm->futex_hash_lock));
1838 new = mm->futex_phash_new;
1839 mm->futex_phash_new = NULL;
1840
1841 if (fph) {
1842 if (cur && !cur->hash_mask) {
1843 /*
1844 * If two threads simultaneously request the global
1845 * hash then the first one performs the switch,
1846 * the second one returns here.
1847 */
1848 free = fph;
1849 mm->futex_phash_new = new;
1850 return -EBUSY;
1851 }
1852 if (cur && !new) {
1853 /*
1854 * If we have an existing hash, but do not yet have
1855 * allocated a replacement hash, drop the initial
1856 * reference on the existing hash.
1857 */
1858 futex_ref_drop(fph: cur);
1859 }
1860
1861 if (new) {
1862 /*
1863 * Two updates raced; throw out the lesser one.
1864 */
1865 if (futex_hash_less(a: new, b: fph)) {
1866 free = new;
1867 new = fph;
1868 } else {
1869 free = fph;
1870 }
1871 } else {
1872 new = fph;
1873 }
1874 fph = NULL;
1875 }
1876
1877 if (new) {
1878 /*
1879 * Will set mm->futex_phash_new on failure;
1880 * futex_private_hash_get() will try again.
1881 */
1882 if (!__futex_pivot_hash(mm, new) && custom)
1883 goto again;
1884 }
1885 }
1886 return 0;
1887}
1888
1889int futex_hash_allocate_default(void)
1890{
1891 unsigned int threads, buckets, current_buckets = 0;
1892 struct futex_private_hash *fph;
1893
1894 if (!current->mm)
1895 return 0;
1896
1897 scoped_guard(rcu) {
1898 threads = min_t(unsigned int,
1899 get_nr_threads(current),
1900 num_online_cpus());
1901
1902 fph = rcu_dereference(current->mm->futex_phash);
1903 if (fph) {
1904 if (fph->custom)
1905 return 0;
1906
1907 current_buckets = fph->hash_mask + 1;
1908 }
1909 }
1910
1911 /*
1912 * The default allocation will remain within
1913 * 16 <= threads * 4 <= global hash size
1914 */
1915 buckets = roundup_pow_of_two(4 * threads);
1916 buckets = clamp(buckets, 16, futex_hashmask + 1);
1917
1918 if (current_buckets >= buckets)
1919 return 0;
1920
1921 return futex_hash_allocate(hash_slots: buckets, flags: 0);
1922}
1923
1924static int futex_hash_get_slots(void)
1925{
1926 struct futex_private_hash *fph;
1927
1928 guard(rcu)();
1929 fph = rcu_dereference(current->mm->futex_phash);
1930 if (fph && fph->hash_mask)
1931 return fph->hash_mask + 1;
1932 return 0;
1933}
1934
1935#else
1936
1937static int futex_hash_allocate(unsigned int hash_slots, unsigned int flags)
1938{
1939 return -EINVAL;
1940}
1941
1942static int futex_hash_get_slots(void)
1943{
1944 return 0;
1945}
1946
1947#endif
1948
1949int futex_hash_prctl(unsigned long arg2, unsigned long arg3, unsigned long arg4)
1950{
1951 unsigned int flags = FH_CUSTOM;
1952 int ret;
1953
1954 switch (arg2) {
1955 case PR_FUTEX_HASH_SET_SLOTS:
1956 if (arg4)
1957 return -EINVAL;
1958 ret = futex_hash_allocate(hash_slots: arg3, flags);
1959 break;
1960
1961 case PR_FUTEX_HASH_GET_SLOTS:
1962 ret = futex_hash_get_slots();
1963 break;
1964
1965 default:
1966 ret = -EINVAL;
1967 break;
1968 }
1969 return ret;
1970}
1971
1972static int __init futex_init(void)
1973{
1974 unsigned long hashsize, i;
1975 unsigned int order, n;
1976 unsigned long size;
1977
1978#ifdef CONFIG_BASE_SMALL
1979 hashsize = 16;
1980#else
1981 hashsize = 256 * num_possible_cpus();
1982 hashsize /= num_possible_nodes();
1983 hashsize = max(4, hashsize);
1984 hashsize = roundup_pow_of_two(hashsize);
1985#endif
1986 futex_hashshift = ilog2(hashsize);
1987 size = sizeof(struct futex_hash_bucket) * hashsize;
1988 order = get_order(size);
1989
1990 for_each_node(n) {
1991 struct futex_hash_bucket *table;
1992
1993 if (order > MAX_PAGE_ORDER)
1994 table = vmalloc_huge_node(size, GFP_KERNEL, n);
1995 else
1996 table = alloc_pages_exact_nid(n, size, GFP_KERNEL);
1997
1998 BUG_ON(!table);
1999
2000 for (i = 0; i < hashsize; i++)
2001 futex_hash_bucket_init(fhb: &table[i], NULL);
2002
2003 futex_queues[n] = table;
2004 }
2005
2006 futex_hashmask = hashsize - 1;
2007 pr_info("futex hash table entries: %lu (%lu bytes on %d NUMA nodes, total %lu KiB, %s).\n",
2008 hashsize, size, num_possible_nodes(), size * num_possible_nodes() / 1024,
2009 order > MAX_PAGE_ORDER ? "vmalloc" : "linear");
2010 return 0;
2011}
2012core_initcall(futex_init);
2013