Loading...
1/*
2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
4 *
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7 *
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
10 *
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14 *
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18 *
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21 *
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
25 *
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
29 *
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
32 *
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
37 *
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
42 *
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
46 */
47#include <linux/slab.h>
48#include <linux/poll.h>
49#include <linux/fs.h>
50#include <linux/file.h>
51#include <linux/jhash.h>
52#include <linux/init.h>
53#include <linux/futex.h>
54#include <linux/mount.h>
55#include <linux/pagemap.h>
56#include <linux/syscalls.h>
57#include <linux/signal.h>
58#include <linux/export.h>
59#include <linux/magic.h>
60#include <linux/pid.h>
61#include <linux/nsproxy.h>
62#include <linux/ptrace.h>
63#include <linux/sched/rt.h>
64#include <linux/hugetlb.h>
65#include <linux/freezer.h>
66#include <linux/bootmem.h>
67#include <linux/fault-inject.h>
68
69#include <asm/futex.h>
70
71#include "locking/rtmutex_common.h"
72
73/*
74 * READ this before attempting to hack on futexes!
75 *
76 * Basic futex operation and ordering guarantees
77 * =============================================
78 *
79 * The waiter reads the futex value in user space and calls
80 * futex_wait(). This function computes the hash bucket and acquires
81 * the hash bucket lock. After that it reads the futex user space value
82 * again and verifies that the data has not changed. If it has not changed
83 * it enqueues itself into the hash bucket, releases the hash bucket lock
84 * and schedules.
85 *
86 * The waker side modifies the user space value of the futex and calls
87 * futex_wake(). This function computes the hash bucket and acquires the
88 * hash bucket lock. Then it looks for waiters on that futex in the hash
89 * bucket and wakes them.
90 *
91 * In futex wake up scenarios where no tasks are blocked on a futex, taking
92 * the hb spinlock can be avoided and simply return. In order for this
93 * optimization to work, ordering guarantees must exist so that the waiter
94 * being added to the list is acknowledged when the list is concurrently being
95 * checked by the waker, avoiding scenarios like the following:
96 *
97 * CPU 0 CPU 1
98 * val = *futex;
99 * sys_futex(WAIT, futex, val);
100 * futex_wait(futex, val);
101 * uval = *futex;
102 * *futex = newval;
103 * sys_futex(WAKE, futex);
104 * futex_wake(futex);
105 * if (queue_empty())
106 * return;
107 * if (uval == val)
108 * lock(hash_bucket(futex));
109 * queue();
110 * unlock(hash_bucket(futex));
111 * schedule();
112 *
113 * This would cause the waiter on CPU 0 to wait forever because it
114 * missed the transition of the user space value from val to newval
115 * and the waker did not find the waiter in the hash bucket queue.
116 *
117 * The correct serialization ensures that a waiter either observes
118 * the changed user space value before blocking or is woken by a
119 * concurrent waker:
120 *
121 * CPU 0 CPU 1
122 * val = *futex;
123 * sys_futex(WAIT, futex, val);
124 * futex_wait(futex, val);
125 *
126 * waiters++; (a)
127 * smp_mb(); (A) <-- paired with -.
128 * |
129 * lock(hash_bucket(futex)); |
130 * |
131 * uval = *futex; |
132 * | *futex = newval;
133 * | sys_futex(WAKE, futex);
134 * | futex_wake(futex);
135 * |
136 * `--------> smp_mb(); (B)
137 * if (uval == val)
138 * queue();
139 * unlock(hash_bucket(futex));
140 * schedule(); if (waiters)
141 * lock(hash_bucket(futex));
142 * else wake_waiters(futex);
143 * waiters--; (b) unlock(hash_bucket(futex));
144 *
145 * Where (A) orders the waiters increment and the futex value read through
146 * atomic operations (see hb_waiters_inc) and where (B) orders the write
147 * to futex and the waiters read -- this is done by the barriers for both
148 * shared and private futexes in get_futex_key_refs().
149 *
150 * This yields the following case (where X:=waiters, Y:=futex):
151 *
152 * X = Y = 0
153 *
154 * w[X]=1 w[Y]=1
155 * MB MB
156 * r[Y]=y r[X]=x
157 *
158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
159 * the guarantee that we cannot both miss the futex variable change and the
160 * enqueue.
161 *
162 * Note that a new waiter is accounted for in (a) even when it is possible that
163 * the wait call can return error, in which case we backtrack from it in (b).
164 * Refer to the comment in queue_lock().
165 *
166 * Similarly, in order to account for waiters being requeued on another
167 * address we always increment the waiters for the destination bucket before
168 * acquiring the lock. It then decrements them again after releasing it -
169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170 * will do the additional required waiter count housekeeping. This is done for
171 * double_lock_hb() and double_unlock_hb(), respectively.
172 */
173
174#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175int __read_mostly futex_cmpxchg_enabled;
176#endif
177
178/*
179 * Futex flags used to encode options to functions and preserve them across
180 * restarts.
181 */
182#ifdef CONFIG_MMU
183# define FLAGS_SHARED 0x01
184#else
185/*
186 * NOMMU does not have per process address space. Let the compiler optimize
187 * code away.
188 */
189# define FLAGS_SHARED 0x00
190#endif
191#define FLAGS_CLOCKRT 0x02
192#define FLAGS_HAS_TIMEOUT 0x04
193
194/*
195 * Priority Inheritance state:
196 */
197struct futex_pi_state {
198 /*
199 * list of 'owned' pi_state instances - these have to be
200 * cleaned up in do_exit() if the task exits prematurely:
201 */
202 struct list_head list;
203
204 /*
205 * The PI object:
206 */
207 struct rt_mutex pi_mutex;
208
209 struct task_struct *owner;
210 atomic_t refcount;
211
212 union futex_key key;
213};
214
215/**
216 * struct futex_q - The hashed futex queue entry, one per waiting task
217 * @list: priority-sorted list of tasks waiting on this futex
218 * @task: the task waiting on the futex
219 * @lock_ptr: the hash bucket lock
220 * @key: the key the futex is hashed on
221 * @pi_state: optional priority inheritance state
222 * @rt_waiter: rt_waiter storage for use with requeue_pi
223 * @requeue_pi_key: the requeue_pi target futex key
224 * @bitset: bitset for the optional bitmasked wakeup
225 *
226 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
227 * we can wake only the relevant ones (hashed queues may be shared).
228 *
229 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
230 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
231 * The order of wakeup is always to make the first condition true, then
232 * the second.
233 *
234 * PI futexes are typically woken before they are removed from the hash list via
235 * the rt_mutex code. See unqueue_me_pi().
236 */
237struct futex_q {
238 struct plist_node list;
239
240 struct task_struct *task;
241 spinlock_t *lock_ptr;
242 union futex_key key;
243 struct futex_pi_state *pi_state;
244 struct rt_mutex_waiter *rt_waiter;
245 union futex_key *requeue_pi_key;
246 u32 bitset;
247};
248
249static const struct futex_q futex_q_init = {
250 /* list gets initialized in queue_me()*/
251 .key = FUTEX_KEY_INIT,
252 .bitset = FUTEX_BITSET_MATCH_ANY
253};
254
255/*
256 * Hash buckets are shared by all the futex_keys that hash to the same
257 * location. Each key may have multiple futex_q structures, one for each task
258 * waiting on a futex.
259 */
260struct futex_hash_bucket {
261 atomic_t waiters;
262 spinlock_t lock;
263 struct plist_head chain;
264} ____cacheline_aligned_in_smp;
265
266/*
267 * The base of the bucket array and its size are always used together
268 * (after initialization only in hash_futex()), so ensure that they
269 * reside in the same cacheline.
270 */
271static struct {
272 struct futex_hash_bucket *queues;
273 unsigned long hashsize;
274} __futex_data __read_mostly __aligned(2*sizeof(long));
275#define futex_queues (__futex_data.queues)
276#define futex_hashsize (__futex_data.hashsize)
277
278
279/*
280 * Fault injections for futexes.
281 */
282#ifdef CONFIG_FAIL_FUTEX
283
284static struct {
285 struct fault_attr attr;
286
287 bool ignore_private;
288} fail_futex = {
289 .attr = FAULT_ATTR_INITIALIZER,
290 .ignore_private = false,
291};
292
293static int __init setup_fail_futex(char *str)
294{
295 return setup_fault_attr(&fail_futex.attr, str);
296}
297__setup("fail_futex=", setup_fail_futex);
298
299static bool should_fail_futex(bool fshared)
300{
301 if (fail_futex.ignore_private && !fshared)
302 return false;
303
304 return should_fail(&fail_futex.attr, 1);
305}
306
307#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
308
309static int __init fail_futex_debugfs(void)
310{
311 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
312 struct dentry *dir;
313
314 dir = fault_create_debugfs_attr("fail_futex", NULL,
315 &fail_futex.attr);
316 if (IS_ERR(dir))
317 return PTR_ERR(dir);
318
319 if (!debugfs_create_bool("ignore-private", mode, dir,
320 &fail_futex.ignore_private)) {
321 debugfs_remove_recursive(dir);
322 return -ENOMEM;
323 }
324
325 return 0;
326}
327
328late_initcall(fail_futex_debugfs);
329
330#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
331
332#else
333static inline bool should_fail_futex(bool fshared)
334{
335 return false;
336}
337#endif /* CONFIG_FAIL_FUTEX */
338
339static inline void futex_get_mm(union futex_key *key)
340{
341 atomic_inc(&key->private.mm->mm_count);
342 /*
343 * Ensure futex_get_mm() implies a full barrier such that
344 * get_futex_key() implies a full barrier. This is relied upon
345 * as smp_mb(); (B), see the ordering comment above.
346 */
347 smp_mb__after_atomic();
348}
349
350/*
351 * Reflects a new waiter being added to the waitqueue.
352 */
353static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
354{
355#ifdef CONFIG_SMP
356 atomic_inc(&hb->waiters);
357 /*
358 * Full barrier (A), see the ordering comment above.
359 */
360 smp_mb__after_atomic();
361#endif
362}
363
364/*
365 * Reflects a waiter being removed from the waitqueue by wakeup
366 * paths.
367 */
368static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
369{
370#ifdef CONFIG_SMP
371 atomic_dec(&hb->waiters);
372#endif
373}
374
375static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
376{
377#ifdef CONFIG_SMP
378 return atomic_read(&hb->waiters);
379#else
380 return 1;
381#endif
382}
383
384/**
385 * hash_futex - Return the hash bucket in the global hash
386 * @key: Pointer to the futex key for which the hash is calculated
387 *
388 * We hash on the keys returned from get_futex_key (see below) and return the
389 * corresponding hash bucket in the global hash.
390 */
391static struct futex_hash_bucket *hash_futex(union futex_key *key)
392{
393 u32 hash = jhash2((u32*)&key->both.word,
394 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
395 key->both.offset);
396 return &futex_queues[hash & (futex_hashsize - 1)];
397}
398
399
400/**
401 * match_futex - Check whether two futex keys are equal
402 * @key1: Pointer to key1
403 * @key2: Pointer to key2
404 *
405 * Return 1 if two futex_keys are equal, 0 otherwise.
406 */
407static inline int match_futex(union futex_key *key1, union futex_key *key2)
408{
409 return (key1 && key2
410 && key1->both.word == key2->both.word
411 && key1->both.ptr == key2->both.ptr
412 && key1->both.offset == key2->both.offset);
413}
414
415/*
416 * Take a reference to the resource addressed by a key.
417 * Can be called while holding spinlocks.
418 *
419 */
420static void get_futex_key_refs(union futex_key *key)
421{
422 if (!key->both.ptr)
423 return;
424
425 /*
426 * On MMU less systems futexes are always "private" as there is no per
427 * process address space. We need the smp wmb nevertheless - yes,
428 * arch/blackfin has MMU less SMP ...
429 */
430 if (!IS_ENABLED(CONFIG_MMU)) {
431 smp_mb(); /* explicit smp_mb(); (B) */
432 return;
433 }
434
435 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
436 case FUT_OFF_INODE:
437 ihold(key->shared.inode); /* implies smp_mb(); (B) */
438 break;
439 case FUT_OFF_MMSHARED:
440 futex_get_mm(key); /* implies smp_mb(); (B) */
441 break;
442 default:
443 /*
444 * Private futexes do not hold reference on an inode or
445 * mm, therefore the only purpose of calling get_futex_key_refs
446 * is because we need the barrier for the lockless waiter check.
447 */
448 smp_mb(); /* explicit smp_mb(); (B) */
449 }
450}
451
452/*
453 * Drop a reference to the resource addressed by a key.
454 * The hash bucket spinlock must not be held. This is
455 * a no-op for private futexes, see comment in the get
456 * counterpart.
457 */
458static void drop_futex_key_refs(union futex_key *key)
459{
460 if (!key->both.ptr) {
461 /* If we're here then we tried to put a key we failed to get */
462 WARN_ON_ONCE(1);
463 return;
464 }
465
466 if (!IS_ENABLED(CONFIG_MMU))
467 return;
468
469 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
470 case FUT_OFF_INODE:
471 iput(key->shared.inode);
472 break;
473 case FUT_OFF_MMSHARED:
474 mmdrop(key->private.mm);
475 break;
476 }
477}
478
479/**
480 * get_futex_key() - Get parameters which are the keys for a futex
481 * @uaddr: virtual address of the futex
482 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
483 * @key: address where result is stored.
484 * @rw: mapping needs to be read/write (values: VERIFY_READ,
485 * VERIFY_WRITE)
486 *
487 * Return: a negative error code or 0
488 *
489 * The key words are stored in *key on success.
490 *
491 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
492 * offset_within_page). For private mappings, it's (uaddr, current->mm).
493 * We can usually work out the index without swapping in the page.
494 *
495 * lock_page() might sleep, the caller should not hold a spinlock.
496 */
497static int
498get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
499{
500 unsigned long address = (unsigned long)uaddr;
501 struct mm_struct *mm = current->mm;
502 struct page *page, *tail;
503 struct address_space *mapping;
504 int err, ro = 0;
505
506 /*
507 * The futex address must be "naturally" aligned.
508 */
509 key->both.offset = address % PAGE_SIZE;
510 if (unlikely((address % sizeof(u32)) != 0))
511 return -EINVAL;
512 address -= key->both.offset;
513
514 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
515 return -EFAULT;
516
517 if (unlikely(should_fail_futex(fshared)))
518 return -EFAULT;
519
520 /*
521 * PROCESS_PRIVATE futexes are fast.
522 * As the mm cannot disappear under us and the 'key' only needs
523 * virtual address, we dont even have to find the underlying vma.
524 * Note : We do have to check 'uaddr' is a valid user address,
525 * but access_ok() should be faster than find_vma()
526 */
527 if (!fshared) {
528 key->private.mm = mm;
529 key->private.address = address;
530 get_futex_key_refs(key); /* implies smp_mb(); (B) */
531 return 0;
532 }
533
534again:
535 /* Ignore any VERIFY_READ mapping (futex common case) */
536 if (unlikely(should_fail_futex(fshared)))
537 return -EFAULT;
538
539 err = get_user_pages_fast(address, 1, 1, &page);
540 /*
541 * If write access is not required (eg. FUTEX_WAIT), try
542 * and get read-only access.
543 */
544 if (err == -EFAULT && rw == VERIFY_READ) {
545 err = get_user_pages_fast(address, 1, 0, &page);
546 ro = 1;
547 }
548 if (err < 0)
549 return err;
550 else
551 err = 0;
552
553 /*
554 * The treatment of mapping from this point on is critical. The page
555 * lock protects many things but in this context the page lock
556 * stabilizes mapping, prevents inode freeing in the shared
557 * file-backed region case and guards against movement to swap cache.
558 *
559 * Strictly speaking the page lock is not needed in all cases being
560 * considered here and page lock forces unnecessarily serialization
561 * From this point on, mapping will be re-verified if necessary and
562 * page lock will be acquired only if it is unavoidable
563 *
564 * Mapping checks require the head page for any compound page so the
565 * head page and mapping is looked up now. For anonymous pages, it
566 * does not matter if the page splits in the future as the key is
567 * based on the address. For filesystem-backed pages, the tail is
568 * required as the index of the page determines the key. For
569 * base pages, there is no tail page and tail == page.
570 */
571 tail = page;
572 page = compound_head(page);
573 mapping = READ_ONCE(page->mapping);
574
575 /*
576 * If page->mapping is NULL, then it cannot be a PageAnon
577 * page; but it might be the ZERO_PAGE or in the gate area or
578 * in a special mapping (all cases which we are happy to fail);
579 * or it may have been a good file page when get_user_pages_fast
580 * found it, but truncated or holepunched or subjected to
581 * invalidate_complete_page2 before we got the page lock (also
582 * cases which we are happy to fail). And we hold a reference,
583 * so refcount care in invalidate_complete_page's remove_mapping
584 * prevents drop_caches from setting mapping to NULL beneath us.
585 *
586 * The case we do have to guard against is when memory pressure made
587 * shmem_writepage move it from filecache to swapcache beneath us:
588 * an unlikely race, but we do need to retry for page->mapping.
589 */
590 if (unlikely(!mapping)) {
591 int shmem_swizzled;
592
593 /*
594 * Page lock is required to identify which special case above
595 * applies. If this is really a shmem page then the page lock
596 * will prevent unexpected transitions.
597 */
598 lock_page(page);
599 shmem_swizzled = PageSwapCache(page) || page->mapping;
600 unlock_page(page);
601 put_page(page);
602
603 if (shmem_swizzled)
604 goto again;
605
606 return -EFAULT;
607 }
608
609 /*
610 * Private mappings are handled in a simple way.
611 *
612 * If the futex key is stored on an anonymous page, then the associated
613 * object is the mm which is implicitly pinned by the calling process.
614 *
615 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
616 * it's a read-only handle, it's expected that futexes attach to
617 * the object not the particular process.
618 */
619 if (PageAnon(page)) {
620 /*
621 * A RO anonymous page will never change and thus doesn't make
622 * sense for futex operations.
623 */
624 if (unlikely(should_fail_futex(fshared)) || ro) {
625 err = -EFAULT;
626 goto out;
627 }
628
629 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
630 key->private.mm = mm;
631 key->private.address = address;
632
633 get_futex_key_refs(key); /* implies smp_mb(); (B) */
634
635 } else {
636 struct inode *inode;
637
638 /*
639 * The associated futex object in this case is the inode and
640 * the page->mapping must be traversed. Ordinarily this should
641 * be stabilised under page lock but it's not strictly
642 * necessary in this case as we just want to pin the inode, not
643 * update the radix tree or anything like that.
644 *
645 * The RCU read lock is taken as the inode is finally freed
646 * under RCU. If the mapping still matches expectations then the
647 * mapping->host can be safely accessed as being a valid inode.
648 */
649 rcu_read_lock();
650
651 if (READ_ONCE(page->mapping) != mapping) {
652 rcu_read_unlock();
653 put_page(page);
654
655 goto again;
656 }
657
658 inode = READ_ONCE(mapping->host);
659 if (!inode) {
660 rcu_read_unlock();
661 put_page(page);
662
663 goto again;
664 }
665
666 /*
667 * Take a reference unless it is about to be freed. Previously
668 * this reference was taken by ihold under the page lock
669 * pinning the inode in place so i_lock was unnecessary. The
670 * only way for this check to fail is if the inode was
671 * truncated in parallel so warn for now if this happens.
672 *
673 * We are not calling into get_futex_key_refs() in file-backed
674 * cases, therefore a successful atomic_inc return below will
675 * guarantee that get_futex_key() will still imply smp_mb(); (B).
676 */
677 if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
678 rcu_read_unlock();
679 put_page(page);
680
681 goto again;
682 }
683
684 /* Should be impossible but lets be paranoid for now */
685 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
686 err = -EFAULT;
687 rcu_read_unlock();
688 iput(inode);
689
690 goto out;
691 }
692
693 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
694 key->shared.inode = inode;
695 key->shared.pgoff = basepage_index(tail);
696 rcu_read_unlock();
697 }
698
699out:
700 put_page(page);
701 return err;
702}
703
704static inline void put_futex_key(union futex_key *key)
705{
706 drop_futex_key_refs(key);
707}
708
709/**
710 * fault_in_user_writeable() - Fault in user address and verify RW access
711 * @uaddr: pointer to faulting user space address
712 *
713 * Slow path to fixup the fault we just took in the atomic write
714 * access to @uaddr.
715 *
716 * We have no generic implementation of a non-destructive write to the
717 * user address. We know that we faulted in the atomic pagefault
718 * disabled section so we can as well avoid the #PF overhead by
719 * calling get_user_pages() right away.
720 */
721static int fault_in_user_writeable(u32 __user *uaddr)
722{
723 struct mm_struct *mm = current->mm;
724 int ret;
725
726 down_read(&mm->mmap_sem);
727 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
728 FAULT_FLAG_WRITE, NULL);
729 up_read(&mm->mmap_sem);
730
731 return ret < 0 ? ret : 0;
732}
733
734/**
735 * futex_top_waiter() - Return the highest priority waiter on a futex
736 * @hb: the hash bucket the futex_q's reside in
737 * @key: the futex key (to distinguish it from other futex futex_q's)
738 *
739 * Must be called with the hb lock held.
740 */
741static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
742 union futex_key *key)
743{
744 struct futex_q *this;
745
746 plist_for_each_entry(this, &hb->chain, list) {
747 if (match_futex(&this->key, key))
748 return this;
749 }
750 return NULL;
751}
752
753static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
754 u32 uval, u32 newval)
755{
756 int ret;
757
758 pagefault_disable();
759 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
760 pagefault_enable();
761
762 return ret;
763}
764
765static int get_futex_value_locked(u32 *dest, u32 __user *from)
766{
767 int ret;
768
769 pagefault_disable();
770 ret = __get_user(*dest, from);
771 pagefault_enable();
772
773 return ret ? -EFAULT : 0;
774}
775
776
777/*
778 * PI code:
779 */
780static int refill_pi_state_cache(void)
781{
782 struct futex_pi_state *pi_state;
783
784 if (likely(current->pi_state_cache))
785 return 0;
786
787 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
788
789 if (!pi_state)
790 return -ENOMEM;
791
792 INIT_LIST_HEAD(&pi_state->list);
793 /* pi_mutex gets initialized later */
794 pi_state->owner = NULL;
795 atomic_set(&pi_state->refcount, 1);
796 pi_state->key = FUTEX_KEY_INIT;
797
798 current->pi_state_cache = pi_state;
799
800 return 0;
801}
802
803static struct futex_pi_state * alloc_pi_state(void)
804{
805 struct futex_pi_state *pi_state = current->pi_state_cache;
806
807 WARN_ON(!pi_state);
808 current->pi_state_cache = NULL;
809
810 return pi_state;
811}
812
813/*
814 * Drops a reference to the pi_state object and frees or caches it
815 * when the last reference is gone.
816 *
817 * Must be called with the hb lock held.
818 */
819static void put_pi_state(struct futex_pi_state *pi_state)
820{
821 if (!pi_state)
822 return;
823
824 if (!atomic_dec_and_test(&pi_state->refcount))
825 return;
826
827 /*
828 * If pi_state->owner is NULL, the owner is most probably dying
829 * and has cleaned up the pi_state already
830 */
831 if (pi_state->owner) {
832 raw_spin_lock_irq(&pi_state->owner->pi_lock);
833 list_del_init(&pi_state->list);
834 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
835
836 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
837 }
838
839 if (current->pi_state_cache)
840 kfree(pi_state);
841 else {
842 /*
843 * pi_state->list is already empty.
844 * clear pi_state->owner.
845 * refcount is at 0 - put it back to 1.
846 */
847 pi_state->owner = NULL;
848 atomic_set(&pi_state->refcount, 1);
849 current->pi_state_cache = pi_state;
850 }
851}
852
853/*
854 * Look up the task based on what TID userspace gave us.
855 * We dont trust it.
856 */
857static struct task_struct * futex_find_get_task(pid_t pid)
858{
859 struct task_struct *p;
860
861 rcu_read_lock();
862 p = find_task_by_vpid(pid);
863 if (p)
864 get_task_struct(p);
865
866 rcu_read_unlock();
867
868 return p;
869}
870
871/*
872 * This task is holding PI mutexes at exit time => bad.
873 * Kernel cleans up PI-state, but userspace is likely hosed.
874 * (Robust-futex cleanup is separate and might save the day for userspace.)
875 */
876void exit_pi_state_list(struct task_struct *curr)
877{
878 struct list_head *next, *head = &curr->pi_state_list;
879 struct futex_pi_state *pi_state;
880 struct futex_hash_bucket *hb;
881 union futex_key key = FUTEX_KEY_INIT;
882
883 if (!futex_cmpxchg_enabled)
884 return;
885 /*
886 * We are a ZOMBIE and nobody can enqueue itself on
887 * pi_state_list anymore, but we have to be careful
888 * versus waiters unqueueing themselves:
889 */
890 raw_spin_lock_irq(&curr->pi_lock);
891 while (!list_empty(head)) {
892
893 next = head->next;
894 pi_state = list_entry(next, struct futex_pi_state, list);
895 key = pi_state->key;
896 hb = hash_futex(&key);
897 raw_spin_unlock_irq(&curr->pi_lock);
898
899 spin_lock(&hb->lock);
900
901 raw_spin_lock_irq(&curr->pi_lock);
902 /*
903 * We dropped the pi-lock, so re-check whether this
904 * task still owns the PI-state:
905 */
906 if (head->next != next) {
907 spin_unlock(&hb->lock);
908 continue;
909 }
910
911 WARN_ON(pi_state->owner != curr);
912 WARN_ON(list_empty(&pi_state->list));
913 list_del_init(&pi_state->list);
914 pi_state->owner = NULL;
915 raw_spin_unlock_irq(&curr->pi_lock);
916
917 rt_mutex_unlock(&pi_state->pi_mutex);
918
919 spin_unlock(&hb->lock);
920
921 raw_spin_lock_irq(&curr->pi_lock);
922 }
923 raw_spin_unlock_irq(&curr->pi_lock);
924}
925
926/*
927 * We need to check the following states:
928 *
929 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
930 *
931 * [1] NULL | --- | --- | 0 | 0/1 | Valid
932 * [2] NULL | --- | --- | >0 | 0/1 | Valid
933 *
934 * [3] Found | NULL | -- | Any | 0/1 | Invalid
935 *
936 * [4] Found | Found | NULL | 0 | 1 | Valid
937 * [5] Found | Found | NULL | >0 | 1 | Invalid
938 *
939 * [6] Found | Found | task | 0 | 1 | Valid
940 *
941 * [7] Found | Found | NULL | Any | 0 | Invalid
942 *
943 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
944 * [9] Found | Found | task | 0 | 0 | Invalid
945 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
946 *
947 * [1] Indicates that the kernel can acquire the futex atomically. We
948 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
949 *
950 * [2] Valid, if TID does not belong to a kernel thread. If no matching
951 * thread is found then it indicates that the owner TID has died.
952 *
953 * [3] Invalid. The waiter is queued on a non PI futex
954 *
955 * [4] Valid state after exit_robust_list(), which sets the user space
956 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
957 *
958 * [5] The user space value got manipulated between exit_robust_list()
959 * and exit_pi_state_list()
960 *
961 * [6] Valid state after exit_pi_state_list() which sets the new owner in
962 * the pi_state but cannot access the user space value.
963 *
964 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
965 *
966 * [8] Owner and user space value match
967 *
968 * [9] There is no transient state which sets the user space TID to 0
969 * except exit_robust_list(), but this is indicated by the
970 * FUTEX_OWNER_DIED bit. See [4]
971 *
972 * [10] There is no transient state which leaves owner and user space
973 * TID out of sync.
974 */
975
976/*
977 * Validate that the existing waiter has a pi_state and sanity check
978 * the pi_state against the user space value. If correct, attach to
979 * it.
980 */
981static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
982 struct futex_pi_state **ps)
983{
984 pid_t pid = uval & FUTEX_TID_MASK;
985
986 /*
987 * Userspace might have messed up non-PI and PI futexes [3]
988 */
989 if (unlikely(!pi_state))
990 return -EINVAL;
991
992 WARN_ON(!atomic_read(&pi_state->refcount));
993
994 /*
995 * Handle the owner died case:
996 */
997 if (uval & FUTEX_OWNER_DIED) {
998 /*
999 * exit_pi_state_list sets owner to NULL and wakes the
1000 * topmost waiter. The task which acquires the
1001 * pi_state->rt_mutex will fixup owner.
1002 */
1003 if (!pi_state->owner) {
1004 /*
1005 * No pi state owner, but the user space TID
1006 * is not 0. Inconsistent state. [5]
1007 */
1008 if (pid)
1009 return -EINVAL;
1010 /*
1011 * Take a ref on the state and return success. [4]
1012 */
1013 goto out_state;
1014 }
1015
1016 /*
1017 * If TID is 0, then either the dying owner has not
1018 * yet executed exit_pi_state_list() or some waiter
1019 * acquired the rtmutex in the pi state, but did not
1020 * yet fixup the TID in user space.
1021 *
1022 * Take a ref on the state and return success. [6]
1023 */
1024 if (!pid)
1025 goto out_state;
1026 } else {
1027 /*
1028 * If the owner died bit is not set, then the pi_state
1029 * must have an owner. [7]
1030 */
1031 if (!pi_state->owner)
1032 return -EINVAL;
1033 }
1034
1035 /*
1036 * Bail out if user space manipulated the futex value. If pi
1037 * state exists then the owner TID must be the same as the
1038 * user space TID. [9/10]
1039 */
1040 if (pid != task_pid_vnr(pi_state->owner))
1041 return -EINVAL;
1042out_state:
1043 atomic_inc(&pi_state->refcount);
1044 *ps = pi_state;
1045 return 0;
1046}
1047
1048/*
1049 * Lookup the task for the TID provided from user space and attach to
1050 * it after doing proper sanity checks.
1051 */
1052static int attach_to_pi_owner(u32 uval, union futex_key *key,
1053 struct futex_pi_state **ps)
1054{
1055 pid_t pid = uval & FUTEX_TID_MASK;
1056 struct futex_pi_state *pi_state;
1057 struct task_struct *p;
1058
1059 /*
1060 * We are the first waiter - try to look up the real owner and attach
1061 * the new pi_state to it, but bail out when TID = 0 [1]
1062 */
1063 if (!pid)
1064 return -ESRCH;
1065 p = futex_find_get_task(pid);
1066 if (!p)
1067 return -ESRCH;
1068
1069 if (unlikely(p->flags & PF_KTHREAD)) {
1070 put_task_struct(p);
1071 return -EPERM;
1072 }
1073
1074 /*
1075 * We need to look at the task state flags to figure out,
1076 * whether the task is exiting. To protect against the do_exit
1077 * change of the task flags, we do this protected by
1078 * p->pi_lock:
1079 */
1080 raw_spin_lock_irq(&p->pi_lock);
1081 if (unlikely(p->flags & PF_EXITING)) {
1082 /*
1083 * The task is on the way out. When PF_EXITPIDONE is
1084 * set, we know that the task has finished the
1085 * cleanup:
1086 */
1087 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1088
1089 raw_spin_unlock_irq(&p->pi_lock);
1090 put_task_struct(p);
1091 return ret;
1092 }
1093
1094 /*
1095 * No existing pi state. First waiter. [2]
1096 */
1097 pi_state = alloc_pi_state();
1098
1099 /*
1100 * Initialize the pi_mutex in locked state and make @p
1101 * the owner of it:
1102 */
1103 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1104
1105 /* Store the key for possible exit cleanups: */
1106 pi_state->key = *key;
1107
1108 WARN_ON(!list_empty(&pi_state->list));
1109 list_add(&pi_state->list, &p->pi_state_list);
1110 pi_state->owner = p;
1111 raw_spin_unlock_irq(&p->pi_lock);
1112
1113 put_task_struct(p);
1114
1115 *ps = pi_state;
1116
1117 return 0;
1118}
1119
1120static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1121 union futex_key *key, struct futex_pi_state **ps)
1122{
1123 struct futex_q *match = futex_top_waiter(hb, key);
1124
1125 /*
1126 * If there is a waiter on that futex, validate it and
1127 * attach to the pi_state when the validation succeeds.
1128 */
1129 if (match)
1130 return attach_to_pi_state(uval, match->pi_state, ps);
1131
1132 /*
1133 * We are the first waiter - try to look up the owner based on
1134 * @uval and attach to it.
1135 */
1136 return attach_to_pi_owner(uval, key, ps);
1137}
1138
1139static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1140{
1141 u32 uninitialized_var(curval);
1142
1143 if (unlikely(should_fail_futex(true)))
1144 return -EFAULT;
1145
1146 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1147 return -EFAULT;
1148
1149 /*If user space value changed, let the caller retry */
1150 return curval != uval ? -EAGAIN : 0;
1151}
1152
1153/**
1154 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1155 * @uaddr: the pi futex user address
1156 * @hb: the pi futex hash bucket
1157 * @key: the futex key associated with uaddr and hb
1158 * @ps: the pi_state pointer where we store the result of the
1159 * lookup
1160 * @task: the task to perform the atomic lock work for. This will
1161 * be "current" except in the case of requeue pi.
1162 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1163 *
1164 * Return:
1165 * 0 - ready to wait;
1166 * 1 - acquired the lock;
1167 * <0 - error
1168 *
1169 * The hb->lock and futex_key refs shall be held by the caller.
1170 */
1171static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1172 union futex_key *key,
1173 struct futex_pi_state **ps,
1174 struct task_struct *task, int set_waiters)
1175{
1176 u32 uval, newval, vpid = task_pid_vnr(task);
1177 struct futex_q *match;
1178 int ret;
1179
1180 /*
1181 * Read the user space value first so we can validate a few
1182 * things before proceeding further.
1183 */
1184 if (get_futex_value_locked(&uval, uaddr))
1185 return -EFAULT;
1186
1187 if (unlikely(should_fail_futex(true)))
1188 return -EFAULT;
1189
1190 /*
1191 * Detect deadlocks.
1192 */
1193 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1194 return -EDEADLK;
1195
1196 if ((unlikely(should_fail_futex(true))))
1197 return -EDEADLK;
1198
1199 /*
1200 * Lookup existing state first. If it exists, try to attach to
1201 * its pi_state.
1202 */
1203 match = futex_top_waiter(hb, key);
1204 if (match)
1205 return attach_to_pi_state(uval, match->pi_state, ps);
1206
1207 /*
1208 * No waiter and user TID is 0. We are here because the
1209 * waiters or the owner died bit is set or called from
1210 * requeue_cmp_pi or for whatever reason something took the
1211 * syscall.
1212 */
1213 if (!(uval & FUTEX_TID_MASK)) {
1214 /*
1215 * We take over the futex. No other waiters and the user space
1216 * TID is 0. We preserve the owner died bit.
1217 */
1218 newval = uval & FUTEX_OWNER_DIED;
1219 newval |= vpid;
1220
1221 /* The futex requeue_pi code can enforce the waiters bit */
1222 if (set_waiters)
1223 newval |= FUTEX_WAITERS;
1224
1225 ret = lock_pi_update_atomic(uaddr, uval, newval);
1226 /* If the take over worked, return 1 */
1227 return ret < 0 ? ret : 1;
1228 }
1229
1230 /*
1231 * First waiter. Set the waiters bit before attaching ourself to
1232 * the owner. If owner tries to unlock, it will be forced into
1233 * the kernel and blocked on hb->lock.
1234 */
1235 newval = uval | FUTEX_WAITERS;
1236 ret = lock_pi_update_atomic(uaddr, uval, newval);
1237 if (ret)
1238 return ret;
1239 /*
1240 * If the update of the user space value succeeded, we try to
1241 * attach to the owner. If that fails, no harm done, we only
1242 * set the FUTEX_WAITERS bit in the user space variable.
1243 */
1244 return attach_to_pi_owner(uval, key, ps);
1245}
1246
1247/**
1248 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1249 * @q: The futex_q to unqueue
1250 *
1251 * The q->lock_ptr must not be NULL and must be held by the caller.
1252 */
1253static void __unqueue_futex(struct futex_q *q)
1254{
1255 struct futex_hash_bucket *hb;
1256
1257 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1258 || WARN_ON(plist_node_empty(&q->list)))
1259 return;
1260
1261 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1262 plist_del(&q->list, &hb->chain);
1263 hb_waiters_dec(hb);
1264}
1265
1266/*
1267 * The hash bucket lock must be held when this is called.
1268 * Afterwards, the futex_q must not be accessed. Callers
1269 * must ensure to later call wake_up_q() for the actual
1270 * wakeups to occur.
1271 */
1272static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1273{
1274 struct task_struct *p = q->task;
1275
1276 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1277 return;
1278
1279 /*
1280 * Queue the task for later wakeup for after we've released
1281 * the hb->lock. wake_q_add() grabs reference to p.
1282 */
1283 wake_q_add(wake_q, p);
1284 __unqueue_futex(q);
1285 /*
1286 * The waiting task can free the futex_q as soon as
1287 * q->lock_ptr = NULL is written, without taking any locks. A
1288 * memory barrier is required here to prevent the following
1289 * store to lock_ptr from getting ahead of the plist_del.
1290 */
1291 smp_wmb();
1292 q->lock_ptr = NULL;
1293}
1294
1295static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1296 struct futex_hash_bucket *hb)
1297{
1298 struct task_struct *new_owner;
1299 struct futex_pi_state *pi_state = this->pi_state;
1300 u32 uninitialized_var(curval), newval;
1301 DEFINE_WAKE_Q(wake_q);
1302 bool deboost;
1303 int ret = 0;
1304
1305 if (!pi_state)
1306 return -EINVAL;
1307
1308 /*
1309 * If current does not own the pi_state then the futex is
1310 * inconsistent and user space fiddled with the futex value.
1311 */
1312 if (pi_state->owner != current)
1313 return -EINVAL;
1314
1315 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1316 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1317
1318 /*
1319 * It is possible that the next waiter (the one that brought
1320 * this owner to the kernel) timed out and is no longer
1321 * waiting on the lock.
1322 */
1323 if (!new_owner)
1324 new_owner = this->task;
1325
1326 /*
1327 * We pass it to the next owner. The WAITERS bit is always
1328 * kept enabled while there is PI state around. We cleanup the
1329 * owner died bit, because we are the owner.
1330 */
1331 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1332
1333 if (unlikely(should_fail_futex(true)))
1334 ret = -EFAULT;
1335
1336 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1337 ret = -EFAULT;
1338 } else if (curval != uval) {
1339 /*
1340 * If a unconditional UNLOCK_PI operation (user space did not
1341 * try the TID->0 transition) raced with a waiter setting the
1342 * FUTEX_WAITERS flag between get_user() and locking the hash
1343 * bucket lock, retry the operation.
1344 */
1345 if ((FUTEX_TID_MASK & curval) == uval)
1346 ret = -EAGAIN;
1347 else
1348 ret = -EINVAL;
1349 }
1350 if (ret) {
1351 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1352 return ret;
1353 }
1354
1355 raw_spin_lock(&pi_state->owner->pi_lock);
1356 WARN_ON(list_empty(&pi_state->list));
1357 list_del_init(&pi_state->list);
1358 raw_spin_unlock(&pi_state->owner->pi_lock);
1359
1360 raw_spin_lock(&new_owner->pi_lock);
1361 WARN_ON(!list_empty(&pi_state->list));
1362 list_add(&pi_state->list, &new_owner->pi_state_list);
1363 pi_state->owner = new_owner;
1364 raw_spin_unlock(&new_owner->pi_lock);
1365
1366 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1367
1368 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1369
1370 /*
1371 * First unlock HB so the waiter does not spin on it once he got woken
1372 * up. Second wake up the waiter before the priority is adjusted. If we
1373 * deboost first (and lose our higher priority), then the task might get
1374 * scheduled away before the wake up can take place.
1375 */
1376 spin_unlock(&hb->lock);
1377 wake_up_q(&wake_q);
1378 if (deboost)
1379 rt_mutex_adjust_prio(current);
1380
1381 return 0;
1382}
1383
1384/*
1385 * Express the locking dependencies for lockdep:
1386 */
1387static inline void
1388double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1389{
1390 if (hb1 <= hb2) {
1391 spin_lock(&hb1->lock);
1392 if (hb1 < hb2)
1393 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1394 } else { /* hb1 > hb2 */
1395 spin_lock(&hb2->lock);
1396 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1397 }
1398}
1399
1400static inline void
1401double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1402{
1403 spin_unlock(&hb1->lock);
1404 if (hb1 != hb2)
1405 spin_unlock(&hb2->lock);
1406}
1407
1408/*
1409 * Wake up waiters matching bitset queued on this futex (uaddr).
1410 */
1411static int
1412futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1413{
1414 struct futex_hash_bucket *hb;
1415 struct futex_q *this, *next;
1416 union futex_key key = FUTEX_KEY_INIT;
1417 int ret;
1418 DEFINE_WAKE_Q(wake_q);
1419
1420 if (!bitset)
1421 return -EINVAL;
1422
1423 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1424 if (unlikely(ret != 0))
1425 goto out;
1426
1427 hb = hash_futex(&key);
1428
1429 /* Make sure we really have tasks to wakeup */
1430 if (!hb_waiters_pending(hb))
1431 goto out_put_key;
1432
1433 spin_lock(&hb->lock);
1434
1435 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1436 if (match_futex (&this->key, &key)) {
1437 if (this->pi_state || this->rt_waiter) {
1438 ret = -EINVAL;
1439 break;
1440 }
1441
1442 /* Check if one of the bits is set in both bitsets */
1443 if (!(this->bitset & bitset))
1444 continue;
1445
1446 mark_wake_futex(&wake_q, this);
1447 if (++ret >= nr_wake)
1448 break;
1449 }
1450 }
1451
1452 spin_unlock(&hb->lock);
1453 wake_up_q(&wake_q);
1454out_put_key:
1455 put_futex_key(&key);
1456out:
1457 return ret;
1458}
1459
1460/*
1461 * Wake up all waiters hashed on the physical page that is mapped
1462 * to this virtual address:
1463 */
1464static int
1465futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1466 int nr_wake, int nr_wake2, int op)
1467{
1468 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1469 struct futex_hash_bucket *hb1, *hb2;
1470 struct futex_q *this, *next;
1471 int ret, op_ret;
1472 DEFINE_WAKE_Q(wake_q);
1473
1474retry:
1475 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1476 if (unlikely(ret != 0))
1477 goto out;
1478 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1479 if (unlikely(ret != 0))
1480 goto out_put_key1;
1481
1482 hb1 = hash_futex(&key1);
1483 hb2 = hash_futex(&key2);
1484
1485retry_private:
1486 double_lock_hb(hb1, hb2);
1487 op_ret = futex_atomic_op_inuser(op, uaddr2);
1488 if (unlikely(op_ret < 0)) {
1489
1490 double_unlock_hb(hb1, hb2);
1491
1492#ifndef CONFIG_MMU
1493 /*
1494 * we don't get EFAULT from MMU faults if we don't have an MMU,
1495 * but we might get them from range checking
1496 */
1497 ret = op_ret;
1498 goto out_put_keys;
1499#endif
1500
1501 if (unlikely(op_ret != -EFAULT)) {
1502 ret = op_ret;
1503 goto out_put_keys;
1504 }
1505
1506 ret = fault_in_user_writeable(uaddr2);
1507 if (ret)
1508 goto out_put_keys;
1509
1510 if (!(flags & FLAGS_SHARED))
1511 goto retry_private;
1512
1513 put_futex_key(&key2);
1514 put_futex_key(&key1);
1515 goto retry;
1516 }
1517
1518 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1519 if (match_futex (&this->key, &key1)) {
1520 if (this->pi_state || this->rt_waiter) {
1521 ret = -EINVAL;
1522 goto out_unlock;
1523 }
1524 mark_wake_futex(&wake_q, this);
1525 if (++ret >= nr_wake)
1526 break;
1527 }
1528 }
1529
1530 if (op_ret > 0) {
1531 op_ret = 0;
1532 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1533 if (match_futex (&this->key, &key2)) {
1534 if (this->pi_state || this->rt_waiter) {
1535 ret = -EINVAL;
1536 goto out_unlock;
1537 }
1538 mark_wake_futex(&wake_q, this);
1539 if (++op_ret >= nr_wake2)
1540 break;
1541 }
1542 }
1543 ret += op_ret;
1544 }
1545
1546out_unlock:
1547 double_unlock_hb(hb1, hb2);
1548 wake_up_q(&wake_q);
1549out_put_keys:
1550 put_futex_key(&key2);
1551out_put_key1:
1552 put_futex_key(&key1);
1553out:
1554 return ret;
1555}
1556
1557/**
1558 * requeue_futex() - Requeue a futex_q from one hb to another
1559 * @q: the futex_q to requeue
1560 * @hb1: the source hash_bucket
1561 * @hb2: the target hash_bucket
1562 * @key2: the new key for the requeued futex_q
1563 */
1564static inline
1565void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1566 struct futex_hash_bucket *hb2, union futex_key *key2)
1567{
1568
1569 /*
1570 * If key1 and key2 hash to the same bucket, no need to
1571 * requeue.
1572 */
1573 if (likely(&hb1->chain != &hb2->chain)) {
1574 plist_del(&q->list, &hb1->chain);
1575 hb_waiters_dec(hb1);
1576 hb_waiters_inc(hb2);
1577 plist_add(&q->list, &hb2->chain);
1578 q->lock_ptr = &hb2->lock;
1579 }
1580 get_futex_key_refs(key2);
1581 q->key = *key2;
1582}
1583
1584/**
1585 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1586 * @q: the futex_q
1587 * @key: the key of the requeue target futex
1588 * @hb: the hash_bucket of the requeue target futex
1589 *
1590 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1591 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1592 * to the requeue target futex so the waiter can detect the wakeup on the right
1593 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1594 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1595 * to protect access to the pi_state to fixup the owner later. Must be called
1596 * with both q->lock_ptr and hb->lock held.
1597 */
1598static inline
1599void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1600 struct futex_hash_bucket *hb)
1601{
1602 get_futex_key_refs(key);
1603 q->key = *key;
1604
1605 __unqueue_futex(q);
1606
1607 WARN_ON(!q->rt_waiter);
1608 q->rt_waiter = NULL;
1609
1610 q->lock_ptr = &hb->lock;
1611
1612 wake_up_state(q->task, TASK_NORMAL);
1613}
1614
1615/**
1616 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1617 * @pifutex: the user address of the to futex
1618 * @hb1: the from futex hash bucket, must be locked by the caller
1619 * @hb2: the to futex hash bucket, must be locked by the caller
1620 * @key1: the from futex key
1621 * @key2: the to futex key
1622 * @ps: address to store the pi_state pointer
1623 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1624 *
1625 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1626 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1627 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1628 * hb1 and hb2 must be held by the caller.
1629 *
1630 * Return:
1631 * 0 - failed to acquire the lock atomically;
1632 * >0 - acquired the lock, return value is vpid of the top_waiter
1633 * <0 - error
1634 */
1635static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1636 struct futex_hash_bucket *hb1,
1637 struct futex_hash_bucket *hb2,
1638 union futex_key *key1, union futex_key *key2,
1639 struct futex_pi_state **ps, int set_waiters)
1640{
1641 struct futex_q *top_waiter = NULL;
1642 u32 curval;
1643 int ret, vpid;
1644
1645 if (get_futex_value_locked(&curval, pifutex))
1646 return -EFAULT;
1647
1648 if (unlikely(should_fail_futex(true)))
1649 return -EFAULT;
1650
1651 /*
1652 * Find the top_waiter and determine if there are additional waiters.
1653 * If the caller intends to requeue more than 1 waiter to pifutex,
1654 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1655 * as we have means to handle the possible fault. If not, don't set
1656 * the bit unecessarily as it will force the subsequent unlock to enter
1657 * the kernel.
1658 */
1659 top_waiter = futex_top_waiter(hb1, key1);
1660
1661 /* There are no waiters, nothing for us to do. */
1662 if (!top_waiter)
1663 return 0;
1664
1665 /* Ensure we requeue to the expected futex. */
1666 if (!match_futex(top_waiter->requeue_pi_key, key2))
1667 return -EINVAL;
1668
1669 /*
1670 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1671 * the contended case or if set_waiters is 1. The pi_state is returned
1672 * in ps in contended cases.
1673 */
1674 vpid = task_pid_vnr(top_waiter->task);
1675 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1676 set_waiters);
1677 if (ret == 1) {
1678 requeue_pi_wake_futex(top_waiter, key2, hb2);
1679 return vpid;
1680 }
1681 return ret;
1682}
1683
1684/**
1685 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1686 * @uaddr1: source futex user address
1687 * @flags: futex flags (FLAGS_SHARED, etc.)
1688 * @uaddr2: target futex user address
1689 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1690 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1691 * @cmpval: @uaddr1 expected value (or %NULL)
1692 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1693 * pi futex (pi to pi requeue is not supported)
1694 *
1695 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1696 * uaddr2 atomically on behalf of the top waiter.
1697 *
1698 * Return:
1699 * >=0 - on success, the number of tasks requeued or woken;
1700 * <0 - on error
1701 */
1702static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1703 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1704 u32 *cmpval, int requeue_pi)
1705{
1706 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1707 int drop_count = 0, task_count = 0, ret;
1708 struct futex_pi_state *pi_state = NULL;
1709 struct futex_hash_bucket *hb1, *hb2;
1710 struct futex_q *this, *next;
1711 DEFINE_WAKE_Q(wake_q);
1712
1713 if (requeue_pi) {
1714 /*
1715 * Requeue PI only works on two distinct uaddrs. This
1716 * check is only valid for private futexes. See below.
1717 */
1718 if (uaddr1 == uaddr2)
1719 return -EINVAL;
1720
1721 /*
1722 * requeue_pi requires a pi_state, try to allocate it now
1723 * without any locks in case it fails.
1724 */
1725 if (refill_pi_state_cache())
1726 return -ENOMEM;
1727 /*
1728 * requeue_pi must wake as many tasks as it can, up to nr_wake
1729 * + nr_requeue, since it acquires the rt_mutex prior to
1730 * returning to userspace, so as to not leave the rt_mutex with
1731 * waiters and no owner. However, second and third wake-ups
1732 * cannot be predicted as they involve race conditions with the
1733 * first wake and a fault while looking up the pi_state. Both
1734 * pthread_cond_signal() and pthread_cond_broadcast() should
1735 * use nr_wake=1.
1736 */
1737 if (nr_wake != 1)
1738 return -EINVAL;
1739 }
1740
1741retry:
1742 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1743 if (unlikely(ret != 0))
1744 goto out;
1745 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1746 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1747 if (unlikely(ret != 0))
1748 goto out_put_key1;
1749
1750 /*
1751 * The check above which compares uaddrs is not sufficient for
1752 * shared futexes. We need to compare the keys:
1753 */
1754 if (requeue_pi && match_futex(&key1, &key2)) {
1755 ret = -EINVAL;
1756 goto out_put_keys;
1757 }
1758
1759 hb1 = hash_futex(&key1);
1760 hb2 = hash_futex(&key2);
1761
1762retry_private:
1763 hb_waiters_inc(hb2);
1764 double_lock_hb(hb1, hb2);
1765
1766 if (likely(cmpval != NULL)) {
1767 u32 curval;
1768
1769 ret = get_futex_value_locked(&curval, uaddr1);
1770
1771 if (unlikely(ret)) {
1772 double_unlock_hb(hb1, hb2);
1773 hb_waiters_dec(hb2);
1774
1775 ret = get_user(curval, uaddr1);
1776 if (ret)
1777 goto out_put_keys;
1778
1779 if (!(flags & FLAGS_SHARED))
1780 goto retry_private;
1781
1782 put_futex_key(&key2);
1783 put_futex_key(&key1);
1784 goto retry;
1785 }
1786 if (curval != *cmpval) {
1787 ret = -EAGAIN;
1788 goto out_unlock;
1789 }
1790 }
1791
1792 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1793 /*
1794 * Attempt to acquire uaddr2 and wake the top waiter. If we
1795 * intend to requeue waiters, force setting the FUTEX_WAITERS
1796 * bit. We force this here where we are able to easily handle
1797 * faults rather in the requeue loop below.
1798 */
1799 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1800 &key2, &pi_state, nr_requeue);
1801
1802 /*
1803 * At this point the top_waiter has either taken uaddr2 or is
1804 * waiting on it. If the former, then the pi_state will not
1805 * exist yet, look it up one more time to ensure we have a
1806 * reference to it. If the lock was taken, ret contains the
1807 * vpid of the top waiter task.
1808 * If the lock was not taken, we have pi_state and an initial
1809 * refcount on it. In case of an error we have nothing.
1810 */
1811 if (ret > 0) {
1812 WARN_ON(pi_state);
1813 drop_count++;
1814 task_count++;
1815 /*
1816 * If we acquired the lock, then the user space value
1817 * of uaddr2 should be vpid. It cannot be changed by
1818 * the top waiter as it is blocked on hb2 lock if it
1819 * tries to do so. If something fiddled with it behind
1820 * our back the pi state lookup might unearth it. So
1821 * we rather use the known value than rereading and
1822 * handing potential crap to lookup_pi_state.
1823 *
1824 * If that call succeeds then we have pi_state and an
1825 * initial refcount on it.
1826 */
1827 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1828 }
1829
1830 switch (ret) {
1831 case 0:
1832 /* We hold a reference on the pi state. */
1833 break;
1834
1835 /* If the above failed, then pi_state is NULL */
1836 case -EFAULT:
1837 double_unlock_hb(hb1, hb2);
1838 hb_waiters_dec(hb2);
1839 put_futex_key(&key2);
1840 put_futex_key(&key1);
1841 ret = fault_in_user_writeable(uaddr2);
1842 if (!ret)
1843 goto retry;
1844 goto out;
1845 case -EAGAIN:
1846 /*
1847 * Two reasons for this:
1848 * - Owner is exiting and we just wait for the
1849 * exit to complete.
1850 * - The user space value changed.
1851 */
1852 double_unlock_hb(hb1, hb2);
1853 hb_waiters_dec(hb2);
1854 put_futex_key(&key2);
1855 put_futex_key(&key1);
1856 cond_resched();
1857 goto retry;
1858 default:
1859 goto out_unlock;
1860 }
1861 }
1862
1863 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1864 if (task_count - nr_wake >= nr_requeue)
1865 break;
1866
1867 if (!match_futex(&this->key, &key1))
1868 continue;
1869
1870 /*
1871 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1872 * be paired with each other and no other futex ops.
1873 *
1874 * We should never be requeueing a futex_q with a pi_state,
1875 * which is awaiting a futex_unlock_pi().
1876 */
1877 if ((requeue_pi && !this->rt_waiter) ||
1878 (!requeue_pi && this->rt_waiter) ||
1879 this->pi_state) {
1880 ret = -EINVAL;
1881 break;
1882 }
1883
1884 /*
1885 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1886 * lock, we already woke the top_waiter. If not, it will be
1887 * woken by futex_unlock_pi().
1888 */
1889 if (++task_count <= nr_wake && !requeue_pi) {
1890 mark_wake_futex(&wake_q, this);
1891 continue;
1892 }
1893
1894 /* Ensure we requeue to the expected futex for requeue_pi. */
1895 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1896 ret = -EINVAL;
1897 break;
1898 }
1899
1900 /*
1901 * Requeue nr_requeue waiters and possibly one more in the case
1902 * of requeue_pi if we couldn't acquire the lock atomically.
1903 */
1904 if (requeue_pi) {
1905 /*
1906 * Prepare the waiter to take the rt_mutex. Take a
1907 * refcount on the pi_state and store the pointer in
1908 * the futex_q object of the waiter.
1909 */
1910 atomic_inc(&pi_state->refcount);
1911 this->pi_state = pi_state;
1912 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1913 this->rt_waiter,
1914 this->task);
1915 if (ret == 1) {
1916 /*
1917 * We got the lock. We do neither drop the
1918 * refcount on pi_state nor clear
1919 * this->pi_state because the waiter needs the
1920 * pi_state for cleaning up the user space
1921 * value. It will drop the refcount after
1922 * doing so.
1923 */
1924 requeue_pi_wake_futex(this, &key2, hb2);
1925 drop_count++;
1926 continue;
1927 } else if (ret) {
1928 /*
1929 * rt_mutex_start_proxy_lock() detected a
1930 * potential deadlock when we tried to queue
1931 * that waiter. Drop the pi_state reference
1932 * which we took above and remove the pointer
1933 * to the state from the waiters futex_q
1934 * object.
1935 */
1936 this->pi_state = NULL;
1937 put_pi_state(pi_state);
1938 /*
1939 * We stop queueing more waiters and let user
1940 * space deal with the mess.
1941 */
1942 break;
1943 }
1944 }
1945 requeue_futex(this, hb1, hb2, &key2);
1946 drop_count++;
1947 }
1948
1949 /*
1950 * We took an extra initial reference to the pi_state either
1951 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
1952 * need to drop it here again.
1953 */
1954 put_pi_state(pi_state);
1955
1956out_unlock:
1957 double_unlock_hb(hb1, hb2);
1958 wake_up_q(&wake_q);
1959 hb_waiters_dec(hb2);
1960
1961 /*
1962 * drop_futex_key_refs() must be called outside the spinlocks. During
1963 * the requeue we moved futex_q's from the hash bucket at key1 to the
1964 * one at key2 and updated their key pointer. We no longer need to
1965 * hold the references to key1.
1966 */
1967 while (--drop_count >= 0)
1968 drop_futex_key_refs(&key1);
1969
1970out_put_keys:
1971 put_futex_key(&key2);
1972out_put_key1:
1973 put_futex_key(&key1);
1974out:
1975 return ret ? ret : task_count;
1976}
1977
1978/* The key must be already stored in q->key. */
1979static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1980 __acquires(&hb->lock)
1981{
1982 struct futex_hash_bucket *hb;
1983
1984 hb = hash_futex(&q->key);
1985
1986 /*
1987 * Increment the counter before taking the lock so that
1988 * a potential waker won't miss a to-be-slept task that is
1989 * waiting for the spinlock. This is safe as all queue_lock()
1990 * users end up calling queue_me(). Similarly, for housekeeping,
1991 * decrement the counter at queue_unlock() when some error has
1992 * occurred and we don't end up adding the task to the list.
1993 */
1994 hb_waiters_inc(hb);
1995
1996 q->lock_ptr = &hb->lock;
1997
1998 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
1999 return hb;
2000}
2001
2002static inline void
2003queue_unlock(struct futex_hash_bucket *hb)
2004 __releases(&hb->lock)
2005{
2006 spin_unlock(&hb->lock);
2007 hb_waiters_dec(hb);
2008}
2009
2010/**
2011 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2012 * @q: The futex_q to enqueue
2013 * @hb: The destination hash bucket
2014 *
2015 * The hb->lock must be held by the caller, and is released here. A call to
2016 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2017 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2018 * or nothing if the unqueue is done as part of the wake process and the unqueue
2019 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2020 * an example).
2021 */
2022static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2023 __releases(&hb->lock)
2024{
2025 int prio;
2026
2027 /*
2028 * The priority used to register this element is
2029 * - either the real thread-priority for the real-time threads
2030 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2031 * - or MAX_RT_PRIO for non-RT threads.
2032 * Thus, all RT-threads are woken first in priority order, and
2033 * the others are woken last, in FIFO order.
2034 */
2035 prio = min(current->normal_prio, MAX_RT_PRIO);
2036
2037 plist_node_init(&q->list, prio);
2038 plist_add(&q->list, &hb->chain);
2039 q->task = current;
2040 spin_unlock(&hb->lock);
2041}
2042
2043/**
2044 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2045 * @q: The futex_q to unqueue
2046 *
2047 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2048 * be paired with exactly one earlier call to queue_me().
2049 *
2050 * Return:
2051 * 1 - if the futex_q was still queued (and we removed unqueued it);
2052 * 0 - if the futex_q was already removed by the waking thread
2053 */
2054static int unqueue_me(struct futex_q *q)
2055{
2056 spinlock_t *lock_ptr;
2057 int ret = 0;
2058
2059 /* In the common case we don't take the spinlock, which is nice. */
2060retry:
2061 /*
2062 * q->lock_ptr can change between this read and the following spin_lock.
2063 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2064 * optimizing lock_ptr out of the logic below.
2065 */
2066 lock_ptr = READ_ONCE(q->lock_ptr);
2067 if (lock_ptr != NULL) {
2068 spin_lock(lock_ptr);
2069 /*
2070 * q->lock_ptr can change between reading it and
2071 * spin_lock(), causing us to take the wrong lock. This
2072 * corrects the race condition.
2073 *
2074 * Reasoning goes like this: if we have the wrong lock,
2075 * q->lock_ptr must have changed (maybe several times)
2076 * between reading it and the spin_lock(). It can
2077 * change again after the spin_lock() but only if it was
2078 * already changed before the spin_lock(). It cannot,
2079 * however, change back to the original value. Therefore
2080 * we can detect whether we acquired the correct lock.
2081 */
2082 if (unlikely(lock_ptr != q->lock_ptr)) {
2083 spin_unlock(lock_ptr);
2084 goto retry;
2085 }
2086 __unqueue_futex(q);
2087
2088 BUG_ON(q->pi_state);
2089
2090 spin_unlock(lock_ptr);
2091 ret = 1;
2092 }
2093
2094 drop_futex_key_refs(&q->key);
2095 return ret;
2096}
2097
2098/*
2099 * PI futexes can not be requeued and must remove themself from the
2100 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2101 * and dropped here.
2102 */
2103static void unqueue_me_pi(struct futex_q *q)
2104 __releases(q->lock_ptr)
2105{
2106 __unqueue_futex(q);
2107
2108 BUG_ON(!q->pi_state);
2109 put_pi_state(q->pi_state);
2110 q->pi_state = NULL;
2111
2112 spin_unlock(q->lock_ptr);
2113}
2114
2115/*
2116 * Fixup the pi_state owner with the new owner.
2117 *
2118 * Must be called with hash bucket lock held and mm->sem held for non
2119 * private futexes.
2120 */
2121static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2122 struct task_struct *newowner)
2123{
2124 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2125 struct futex_pi_state *pi_state = q->pi_state;
2126 struct task_struct *oldowner = pi_state->owner;
2127 u32 uval, uninitialized_var(curval), newval;
2128 int ret;
2129
2130 /* Owner died? */
2131 if (!pi_state->owner)
2132 newtid |= FUTEX_OWNER_DIED;
2133
2134 /*
2135 * We are here either because we stole the rtmutex from the
2136 * previous highest priority waiter or we are the highest priority
2137 * waiter but failed to get the rtmutex the first time.
2138 * We have to replace the newowner TID in the user space variable.
2139 * This must be atomic as we have to preserve the owner died bit here.
2140 *
2141 * Note: We write the user space value _before_ changing the pi_state
2142 * because we can fault here. Imagine swapped out pages or a fork
2143 * that marked all the anonymous memory readonly for cow.
2144 *
2145 * Modifying pi_state _before_ the user space value would
2146 * leave the pi_state in an inconsistent state when we fault
2147 * here, because we need to drop the hash bucket lock to
2148 * handle the fault. This might be observed in the PID check
2149 * in lookup_pi_state.
2150 */
2151retry:
2152 if (get_futex_value_locked(&uval, uaddr))
2153 goto handle_fault;
2154
2155 while (1) {
2156 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2157
2158 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2159 goto handle_fault;
2160 if (curval == uval)
2161 break;
2162 uval = curval;
2163 }
2164
2165 /*
2166 * We fixed up user space. Now we need to fix the pi_state
2167 * itself.
2168 */
2169 if (pi_state->owner != NULL) {
2170 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2171 WARN_ON(list_empty(&pi_state->list));
2172 list_del_init(&pi_state->list);
2173 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2174 }
2175
2176 pi_state->owner = newowner;
2177
2178 raw_spin_lock_irq(&newowner->pi_lock);
2179 WARN_ON(!list_empty(&pi_state->list));
2180 list_add(&pi_state->list, &newowner->pi_state_list);
2181 raw_spin_unlock_irq(&newowner->pi_lock);
2182 return 0;
2183
2184 /*
2185 * To handle the page fault we need to drop the hash bucket
2186 * lock here. That gives the other task (either the highest priority
2187 * waiter itself or the task which stole the rtmutex) the
2188 * chance to try the fixup of the pi_state. So once we are
2189 * back from handling the fault we need to check the pi_state
2190 * after reacquiring the hash bucket lock and before trying to
2191 * do another fixup. When the fixup has been done already we
2192 * simply return.
2193 */
2194handle_fault:
2195 spin_unlock(q->lock_ptr);
2196
2197 ret = fault_in_user_writeable(uaddr);
2198
2199 spin_lock(q->lock_ptr);
2200
2201 /*
2202 * Check if someone else fixed it for us:
2203 */
2204 if (pi_state->owner != oldowner)
2205 return 0;
2206
2207 if (ret)
2208 return ret;
2209
2210 goto retry;
2211}
2212
2213static long futex_wait_restart(struct restart_block *restart);
2214
2215/**
2216 * fixup_owner() - Post lock pi_state and corner case management
2217 * @uaddr: user address of the futex
2218 * @q: futex_q (contains pi_state and access to the rt_mutex)
2219 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2220 *
2221 * After attempting to lock an rt_mutex, this function is called to cleanup
2222 * the pi_state owner as well as handle race conditions that may allow us to
2223 * acquire the lock. Must be called with the hb lock held.
2224 *
2225 * Return:
2226 * 1 - success, lock taken;
2227 * 0 - success, lock not taken;
2228 * <0 - on error (-EFAULT)
2229 */
2230static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2231{
2232 struct task_struct *owner;
2233 int ret = 0;
2234
2235 if (locked) {
2236 /*
2237 * Got the lock. We might not be the anticipated owner if we
2238 * did a lock-steal - fix up the PI-state in that case:
2239 */
2240 if (q->pi_state->owner != current)
2241 ret = fixup_pi_state_owner(uaddr, q, current);
2242 goto out;
2243 }
2244
2245 /*
2246 * Catch the rare case, where the lock was released when we were on the
2247 * way back before we locked the hash bucket.
2248 */
2249 if (q->pi_state->owner == current) {
2250 /*
2251 * Try to get the rt_mutex now. This might fail as some other
2252 * task acquired the rt_mutex after we removed ourself from the
2253 * rt_mutex waiters list.
2254 */
2255 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2256 locked = 1;
2257 goto out;
2258 }
2259
2260 /*
2261 * pi_state is incorrect, some other task did a lock steal and
2262 * we returned due to timeout or signal without taking the
2263 * rt_mutex. Too late.
2264 */
2265 raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2266 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2267 if (!owner)
2268 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2269 raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2270 ret = fixup_pi_state_owner(uaddr, q, owner);
2271 goto out;
2272 }
2273
2274 /*
2275 * Paranoia check. If we did not take the lock, then we should not be
2276 * the owner of the rt_mutex.
2277 */
2278 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2279 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2280 "pi-state %p\n", ret,
2281 q->pi_state->pi_mutex.owner,
2282 q->pi_state->owner);
2283
2284out:
2285 return ret ? ret : locked;
2286}
2287
2288/**
2289 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2290 * @hb: the futex hash bucket, must be locked by the caller
2291 * @q: the futex_q to queue up on
2292 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2293 */
2294static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2295 struct hrtimer_sleeper *timeout)
2296{
2297 /*
2298 * The task state is guaranteed to be set before another task can
2299 * wake it. set_current_state() is implemented using smp_store_mb() and
2300 * queue_me() calls spin_unlock() upon completion, both serializing
2301 * access to the hash list and forcing another memory barrier.
2302 */
2303 set_current_state(TASK_INTERRUPTIBLE);
2304 queue_me(q, hb);
2305
2306 /* Arm the timer */
2307 if (timeout)
2308 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2309
2310 /*
2311 * If we have been removed from the hash list, then another task
2312 * has tried to wake us, and we can skip the call to schedule().
2313 */
2314 if (likely(!plist_node_empty(&q->list))) {
2315 /*
2316 * If the timer has already expired, current will already be
2317 * flagged for rescheduling. Only call schedule if there
2318 * is no timeout, or if it has yet to expire.
2319 */
2320 if (!timeout || timeout->task)
2321 freezable_schedule();
2322 }
2323 __set_current_state(TASK_RUNNING);
2324}
2325
2326/**
2327 * futex_wait_setup() - Prepare to wait on a futex
2328 * @uaddr: the futex userspace address
2329 * @val: the expected value
2330 * @flags: futex flags (FLAGS_SHARED, etc.)
2331 * @q: the associated futex_q
2332 * @hb: storage for hash_bucket pointer to be returned to caller
2333 *
2334 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2335 * compare it with the expected value. Handle atomic faults internally.
2336 * Return with the hb lock held and a q.key reference on success, and unlocked
2337 * with no q.key reference on failure.
2338 *
2339 * Return:
2340 * 0 - uaddr contains val and hb has been locked;
2341 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2342 */
2343static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2344 struct futex_q *q, struct futex_hash_bucket **hb)
2345{
2346 u32 uval;
2347 int ret;
2348
2349 /*
2350 * Access the page AFTER the hash-bucket is locked.
2351 * Order is important:
2352 *
2353 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2354 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2355 *
2356 * The basic logical guarantee of a futex is that it blocks ONLY
2357 * if cond(var) is known to be true at the time of blocking, for
2358 * any cond. If we locked the hash-bucket after testing *uaddr, that
2359 * would open a race condition where we could block indefinitely with
2360 * cond(var) false, which would violate the guarantee.
2361 *
2362 * On the other hand, we insert q and release the hash-bucket only
2363 * after testing *uaddr. This guarantees that futex_wait() will NOT
2364 * absorb a wakeup if *uaddr does not match the desired values
2365 * while the syscall executes.
2366 */
2367retry:
2368 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2369 if (unlikely(ret != 0))
2370 return ret;
2371
2372retry_private:
2373 *hb = queue_lock(q);
2374
2375 ret = get_futex_value_locked(&uval, uaddr);
2376
2377 if (ret) {
2378 queue_unlock(*hb);
2379
2380 ret = get_user(uval, uaddr);
2381 if (ret)
2382 goto out;
2383
2384 if (!(flags & FLAGS_SHARED))
2385 goto retry_private;
2386
2387 put_futex_key(&q->key);
2388 goto retry;
2389 }
2390
2391 if (uval != val) {
2392 queue_unlock(*hb);
2393 ret = -EWOULDBLOCK;
2394 }
2395
2396out:
2397 if (ret)
2398 put_futex_key(&q->key);
2399 return ret;
2400}
2401
2402static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2403 ktime_t *abs_time, u32 bitset)
2404{
2405 struct hrtimer_sleeper timeout, *to = NULL;
2406 struct restart_block *restart;
2407 struct futex_hash_bucket *hb;
2408 struct futex_q q = futex_q_init;
2409 int ret;
2410
2411 if (!bitset)
2412 return -EINVAL;
2413 q.bitset = bitset;
2414
2415 if (abs_time) {
2416 to = &timeout;
2417
2418 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2419 CLOCK_REALTIME : CLOCK_MONOTONIC,
2420 HRTIMER_MODE_ABS);
2421 hrtimer_init_sleeper(to, current);
2422 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2423 current->timer_slack_ns);
2424 }
2425
2426retry:
2427 /*
2428 * Prepare to wait on uaddr. On success, holds hb lock and increments
2429 * q.key refs.
2430 */
2431 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2432 if (ret)
2433 goto out;
2434
2435 /* queue_me and wait for wakeup, timeout, or a signal. */
2436 futex_wait_queue_me(hb, &q, to);
2437
2438 /* If we were woken (and unqueued), we succeeded, whatever. */
2439 ret = 0;
2440 /* unqueue_me() drops q.key ref */
2441 if (!unqueue_me(&q))
2442 goto out;
2443 ret = -ETIMEDOUT;
2444 if (to && !to->task)
2445 goto out;
2446
2447 /*
2448 * We expect signal_pending(current), but we might be the
2449 * victim of a spurious wakeup as well.
2450 */
2451 if (!signal_pending(current))
2452 goto retry;
2453
2454 ret = -ERESTARTSYS;
2455 if (!abs_time)
2456 goto out;
2457
2458 restart = ¤t->restart_block;
2459 restart->fn = futex_wait_restart;
2460 restart->futex.uaddr = uaddr;
2461 restart->futex.val = val;
2462 restart->futex.time = *abs_time;
2463 restart->futex.bitset = bitset;
2464 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2465
2466 ret = -ERESTART_RESTARTBLOCK;
2467
2468out:
2469 if (to) {
2470 hrtimer_cancel(&to->timer);
2471 destroy_hrtimer_on_stack(&to->timer);
2472 }
2473 return ret;
2474}
2475
2476
2477static long futex_wait_restart(struct restart_block *restart)
2478{
2479 u32 __user *uaddr = restart->futex.uaddr;
2480 ktime_t t, *tp = NULL;
2481
2482 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2483 t = restart->futex.time;
2484 tp = &t;
2485 }
2486 restart->fn = do_no_restart_syscall;
2487
2488 return (long)futex_wait(uaddr, restart->futex.flags,
2489 restart->futex.val, tp, restart->futex.bitset);
2490}
2491
2492
2493/*
2494 * Userspace tried a 0 -> TID atomic transition of the futex value
2495 * and failed. The kernel side here does the whole locking operation:
2496 * if there are waiters then it will block as a consequence of relying
2497 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2498 * a 0 value of the futex too.).
2499 *
2500 * Also serves as futex trylock_pi()'ing, and due semantics.
2501 */
2502static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2503 ktime_t *time, int trylock)
2504{
2505 struct hrtimer_sleeper timeout, *to = NULL;
2506 struct futex_hash_bucket *hb;
2507 struct futex_q q = futex_q_init;
2508 int res, ret;
2509
2510 if (refill_pi_state_cache())
2511 return -ENOMEM;
2512
2513 if (time) {
2514 to = &timeout;
2515 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2516 HRTIMER_MODE_ABS);
2517 hrtimer_init_sleeper(to, current);
2518 hrtimer_set_expires(&to->timer, *time);
2519 }
2520
2521retry:
2522 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2523 if (unlikely(ret != 0))
2524 goto out;
2525
2526retry_private:
2527 hb = queue_lock(&q);
2528
2529 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2530 if (unlikely(ret)) {
2531 /*
2532 * Atomic work succeeded and we got the lock,
2533 * or failed. Either way, we do _not_ block.
2534 */
2535 switch (ret) {
2536 case 1:
2537 /* We got the lock. */
2538 ret = 0;
2539 goto out_unlock_put_key;
2540 case -EFAULT:
2541 goto uaddr_faulted;
2542 case -EAGAIN:
2543 /*
2544 * Two reasons for this:
2545 * - Task is exiting and we just wait for the
2546 * exit to complete.
2547 * - The user space value changed.
2548 */
2549 queue_unlock(hb);
2550 put_futex_key(&q.key);
2551 cond_resched();
2552 goto retry;
2553 default:
2554 goto out_unlock_put_key;
2555 }
2556 }
2557
2558 /*
2559 * Only actually queue now that the atomic ops are done:
2560 */
2561 queue_me(&q, hb);
2562
2563 WARN_ON(!q.pi_state);
2564 /*
2565 * Block on the PI mutex:
2566 */
2567 if (!trylock) {
2568 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2569 } else {
2570 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2571 /* Fixup the trylock return value: */
2572 ret = ret ? 0 : -EWOULDBLOCK;
2573 }
2574
2575 spin_lock(q.lock_ptr);
2576 /*
2577 * Fixup the pi_state owner and possibly acquire the lock if we
2578 * haven't already.
2579 */
2580 res = fixup_owner(uaddr, &q, !ret);
2581 /*
2582 * If fixup_owner() returned an error, proprogate that. If it acquired
2583 * the lock, clear our -ETIMEDOUT or -EINTR.
2584 */
2585 if (res)
2586 ret = (res < 0) ? res : 0;
2587
2588 /*
2589 * If fixup_owner() faulted and was unable to handle the fault, unlock
2590 * it and return the fault to userspace.
2591 */
2592 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2593 rt_mutex_unlock(&q.pi_state->pi_mutex);
2594
2595 /* Unqueue and drop the lock */
2596 unqueue_me_pi(&q);
2597
2598 goto out_put_key;
2599
2600out_unlock_put_key:
2601 queue_unlock(hb);
2602
2603out_put_key:
2604 put_futex_key(&q.key);
2605out:
2606 if (to)
2607 destroy_hrtimer_on_stack(&to->timer);
2608 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2609
2610uaddr_faulted:
2611 queue_unlock(hb);
2612
2613 ret = fault_in_user_writeable(uaddr);
2614 if (ret)
2615 goto out_put_key;
2616
2617 if (!(flags & FLAGS_SHARED))
2618 goto retry_private;
2619
2620 put_futex_key(&q.key);
2621 goto retry;
2622}
2623
2624/*
2625 * Userspace attempted a TID -> 0 atomic transition, and failed.
2626 * This is the in-kernel slowpath: we look up the PI state (if any),
2627 * and do the rt-mutex unlock.
2628 */
2629static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2630{
2631 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2632 union futex_key key = FUTEX_KEY_INIT;
2633 struct futex_hash_bucket *hb;
2634 struct futex_q *match;
2635 int ret;
2636
2637retry:
2638 if (get_user(uval, uaddr))
2639 return -EFAULT;
2640 /*
2641 * We release only a lock we actually own:
2642 */
2643 if ((uval & FUTEX_TID_MASK) != vpid)
2644 return -EPERM;
2645
2646 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2647 if (ret)
2648 return ret;
2649
2650 hb = hash_futex(&key);
2651 spin_lock(&hb->lock);
2652
2653 /*
2654 * Check waiters first. We do not trust user space values at
2655 * all and we at least want to know if user space fiddled
2656 * with the futex value instead of blindly unlocking.
2657 */
2658 match = futex_top_waiter(hb, &key);
2659 if (match) {
2660 ret = wake_futex_pi(uaddr, uval, match, hb);
2661 /*
2662 * In case of success wake_futex_pi dropped the hash
2663 * bucket lock.
2664 */
2665 if (!ret)
2666 goto out_putkey;
2667 /*
2668 * The atomic access to the futex value generated a
2669 * pagefault, so retry the user-access and the wakeup:
2670 */
2671 if (ret == -EFAULT)
2672 goto pi_faulted;
2673 /*
2674 * A unconditional UNLOCK_PI op raced against a waiter
2675 * setting the FUTEX_WAITERS bit. Try again.
2676 */
2677 if (ret == -EAGAIN) {
2678 spin_unlock(&hb->lock);
2679 put_futex_key(&key);
2680 goto retry;
2681 }
2682 /*
2683 * wake_futex_pi has detected invalid state. Tell user
2684 * space.
2685 */
2686 goto out_unlock;
2687 }
2688
2689 /*
2690 * We have no kernel internal state, i.e. no waiters in the
2691 * kernel. Waiters which are about to queue themselves are stuck
2692 * on hb->lock. So we can safely ignore them. We do neither
2693 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2694 * owner.
2695 */
2696 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2697 goto pi_faulted;
2698
2699 /*
2700 * If uval has changed, let user space handle it.
2701 */
2702 ret = (curval == uval) ? 0 : -EAGAIN;
2703
2704out_unlock:
2705 spin_unlock(&hb->lock);
2706out_putkey:
2707 put_futex_key(&key);
2708 return ret;
2709
2710pi_faulted:
2711 spin_unlock(&hb->lock);
2712 put_futex_key(&key);
2713
2714 ret = fault_in_user_writeable(uaddr);
2715 if (!ret)
2716 goto retry;
2717
2718 return ret;
2719}
2720
2721/**
2722 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2723 * @hb: the hash_bucket futex_q was original enqueued on
2724 * @q: the futex_q woken while waiting to be requeued
2725 * @key2: the futex_key of the requeue target futex
2726 * @timeout: the timeout associated with the wait (NULL if none)
2727 *
2728 * Detect if the task was woken on the initial futex as opposed to the requeue
2729 * target futex. If so, determine if it was a timeout or a signal that caused
2730 * the wakeup and return the appropriate error code to the caller. Must be
2731 * called with the hb lock held.
2732 *
2733 * Return:
2734 * 0 = no early wakeup detected;
2735 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2736 */
2737static inline
2738int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2739 struct futex_q *q, union futex_key *key2,
2740 struct hrtimer_sleeper *timeout)
2741{
2742 int ret = 0;
2743
2744 /*
2745 * With the hb lock held, we avoid races while we process the wakeup.
2746 * We only need to hold hb (and not hb2) to ensure atomicity as the
2747 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2748 * It can't be requeued from uaddr2 to something else since we don't
2749 * support a PI aware source futex for requeue.
2750 */
2751 if (!match_futex(&q->key, key2)) {
2752 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2753 /*
2754 * We were woken prior to requeue by a timeout or a signal.
2755 * Unqueue the futex_q and determine which it was.
2756 */
2757 plist_del(&q->list, &hb->chain);
2758 hb_waiters_dec(hb);
2759
2760 /* Handle spurious wakeups gracefully */
2761 ret = -EWOULDBLOCK;
2762 if (timeout && !timeout->task)
2763 ret = -ETIMEDOUT;
2764 else if (signal_pending(current))
2765 ret = -ERESTARTNOINTR;
2766 }
2767 return ret;
2768}
2769
2770/**
2771 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2772 * @uaddr: the futex we initially wait on (non-pi)
2773 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2774 * the same type, no requeueing from private to shared, etc.
2775 * @val: the expected value of uaddr
2776 * @abs_time: absolute timeout
2777 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2778 * @uaddr2: the pi futex we will take prior to returning to user-space
2779 *
2780 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2781 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2782 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2783 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2784 * without one, the pi logic would not know which task to boost/deboost, if
2785 * there was a need to.
2786 *
2787 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2788 * via the following--
2789 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2790 * 2) wakeup on uaddr2 after a requeue
2791 * 3) signal
2792 * 4) timeout
2793 *
2794 * If 3, cleanup and return -ERESTARTNOINTR.
2795 *
2796 * If 2, we may then block on trying to take the rt_mutex and return via:
2797 * 5) successful lock
2798 * 6) signal
2799 * 7) timeout
2800 * 8) other lock acquisition failure
2801 *
2802 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2803 *
2804 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2805 *
2806 * Return:
2807 * 0 - On success;
2808 * <0 - On error
2809 */
2810static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2811 u32 val, ktime_t *abs_time, u32 bitset,
2812 u32 __user *uaddr2)
2813{
2814 struct hrtimer_sleeper timeout, *to = NULL;
2815 struct rt_mutex_waiter rt_waiter;
2816 struct futex_hash_bucket *hb;
2817 union futex_key key2 = FUTEX_KEY_INIT;
2818 struct futex_q q = futex_q_init;
2819 int res, ret;
2820
2821 if (uaddr == uaddr2)
2822 return -EINVAL;
2823
2824 if (!bitset)
2825 return -EINVAL;
2826
2827 if (abs_time) {
2828 to = &timeout;
2829 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2830 CLOCK_REALTIME : CLOCK_MONOTONIC,
2831 HRTIMER_MODE_ABS);
2832 hrtimer_init_sleeper(to, current);
2833 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2834 current->timer_slack_ns);
2835 }
2836
2837 /*
2838 * The waiter is allocated on our stack, manipulated by the requeue
2839 * code while we sleep on uaddr.
2840 */
2841 debug_rt_mutex_init_waiter(&rt_waiter);
2842 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2843 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2844 rt_waiter.task = NULL;
2845
2846 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2847 if (unlikely(ret != 0))
2848 goto out;
2849
2850 q.bitset = bitset;
2851 q.rt_waiter = &rt_waiter;
2852 q.requeue_pi_key = &key2;
2853
2854 /*
2855 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2856 * count.
2857 */
2858 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2859 if (ret)
2860 goto out_key2;
2861
2862 /*
2863 * The check above which compares uaddrs is not sufficient for
2864 * shared futexes. We need to compare the keys:
2865 */
2866 if (match_futex(&q.key, &key2)) {
2867 queue_unlock(hb);
2868 ret = -EINVAL;
2869 goto out_put_keys;
2870 }
2871
2872 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2873 futex_wait_queue_me(hb, &q, to);
2874
2875 spin_lock(&hb->lock);
2876 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2877 spin_unlock(&hb->lock);
2878 if (ret)
2879 goto out_put_keys;
2880
2881 /*
2882 * In order for us to be here, we know our q.key == key2, and since
2883 * we took the hb->lock above, we also know that futex_requeue() has
2884 * completed and we no longer have to concern ourselves with a wakeup
2885 * race with the atomic proxy lock acquisition by the requeue code. The
2886 * futex_requeue dropped our key1 reference and incremented our key2
2887 * reference count.
2888 */
2889
2890 /* Check if the requeue code acquired the second futex for us. */
2891 if (!q.rt_waiter) {
2892 /*
2893 * Got the lock. We might not be the anticipated owner if we
2894 * did a lock-steal - fix up the PI-state in that case.
2895 */
2896 if (q.pi_state && (q.pi_state->owner != current)) {
2897 spin_lock(q.lock_ptr);
2898 ret = fixup_pi_state_owner(uaddr2, &q, current);
2899 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current)
2900 rt_mutex_unlock(&q.pi_state->pi_mutex);
2901 /*
2902 * Drop the reference to the pi state which
2903 * the requeue_pi() code acquired for us.
2904 */
2905 put_pi_state(q.pi_state);
2906 spin_unlock(q.lock_ptr);
2907 }
2908 } else {
2909 struct rt_mutex *pi_mutex;
2910
2911 /*
2912 * We have been woken up by futex_unlock_pi(), a timeout, or a
2913 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2914 * the pi_state.
2915 */
2916 WARN_ON(!q.pi_state);
2917 pi_mutex = &q.pi_state->pi_mutex;
2918 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2919 debug_rt_mutex_free_waiter(&rt_waiter);
2920
2921 spin_lock(q.lock_ptr);
2922 /*
2923 * Fixup the pi_state owner and possibly acquire the lock if we
2924 * haven't already.
2925 */
2926 res = fixup_owner(uaddr2, &q, !ret);
2927 /*
2928 * If fixup_owner() returned an error, proprogate that. If it
2929 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2930 */
2931 if (res)
2932 ret = (res < 0) ? res : 0;
2933
2934 /*
2935 * If fixup_pi_state_owner() faulted and was unable to handle
2936 * the fault, unlock the rt_mutex and return the fault to
2937 * userspace.
2938 */
2939 if (ret && rt_mutex_owner(pi_mutex) == current)
2940 rt_mutex_unlock(pi_mutex);
2941
2942 /* Unqueue and drop the lock. */
2943 unqueue_me_pi(&q);
2944 }
2945
2946 if (ret == -EINTR) {
2947 /*
2948 * We've already been requeued, but cannot restart by calling
2949 * futex_lock_pi() directly. We could restart this syscall, but
2950 * it would detect that the user space "val" changed and return
2951 * -EWOULDBLOCK. Save the overhead of the restart and return
2952 * -EWOULDBLOCK directly.
2953 */
2954 ret = -EWOULDBLOCK;
2955 }
2956
2957out_put_keys:
2958 put_futex_key(&q.key);
2959out_key2:
2960 put_futex_key(&key2);
2961
2962out:
2963 if (to) {
2964 hrtimer_cancel(&to->timer);
2965 destroy_hrtimer_on_stack(&to->timer);
2966 }
2967 return ret;
2968}
2969
2970/*
2971 * Support for robust futexes: the kernel cleans up held futexes at
2972 * thread exit time.
2973 *
2974 * Implementation: user-space maintains a per-thread list of locks it
2975 * is holding. Upon do_exit(), the kernel carefully walks this list,
2976 * and marks all locks that are owned by this thread with the
2977 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2978 * always manipulated with the lock held, so the list is private and
2979 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2980 * field, to allow the kernel to clean up if the thread dies after
2981 * acquiring the lock, but just before it could have added itself to
2982 * the list. There can only be one such pending lock.
2983 */
2984
2985/**
2986 * sys_set_robust_list() - Set the robust-futex list head of a task
2987 * @head: pointer to the list-head
2988 * @len: length of the list-head, as userspace expects
2989 */
2990SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2991 size_t, len)
2992{
2993 if (!futex_cmpxchg_enabled)
2994 return -ENOSYS;
2995 /*
2996 * The kernel knows only one size for now:
2997 */
2998 if (unlikely(len != sizeof(*head)))
2999 return -EINVAL;
3000
3001 current->robust_list = head;
3002
3003 return 0;
3004}
3005
3006/**
3007 * sys_get_robust_list() - Get the robust-futex list head of a task
3008 * @pid: pid of the process [zero for current task]
3009 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3010 * @len_ptr: pointer to a length field, the kernel fills in the header size
3011 */
3012SYSCALL_DEFINE3(get_robust_list, int, pid,
3013 struct robust_list_head __user * __user *, head_ptr,
3014 size_t __user *, len_ptr)
3015{
3016 struct robust_list_head __user *head;
3017 unsigned long ret;
3018 struct task_struct *p;
3019
3020 if (!futex_cmpxchg_enabled)
3021 return -ENOSYS;
3022
3023 rcu_read_lock();
3024
3025 ret = -ESRCH;
3026 if (!pid)
3027 p = current;
3028 else {
3029 p = find_task_by_vpid(pid);
3030 if (!p)
3031 goto err_unlock;
3032 }
3033
3034 ret = -EPERM;
3035 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3036 goto err_unlock;
3037
3038 head = p->robust_list;
3039 rcu_read_unlock();
3040
3041 if (put_user(sizeof(*head), len_ptr))
3042 return -EFAULT;
3043 return put_user(head, head_ptr);
3044
3045err_unlock:
3046 rcu_read_unlock();
3047
3048 return ret;
3049}
3050
3051/*
3052 * Process a futex-list entry, check whether it's owned by the
3053 * dying task, and do notification if so:
3054 */
3055int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3056{
3057 u32 uval, uninitialized_var(nval), mval;
3058
3059retry:
3060 if (get_user(uval, uaddr))
3061 return -1;
3062
3063 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3064 /*
3065 * Ok, this dying thread is truly holding a futex
3066 * of interest. Set the OWNER_DIED bit atomically
3067 * via cmpxchg, and if the value had FUTEX_WAITERS
3068 * set, wake up a waiter (if any). (We have to do a
3069 * futex_wake() even if OWNER_DIED is already set -
3070 * to handle the rare but possible case of recursive
3071 * thread-death.) The rest of the cleanup is done in
3072 * userspace.
3073 */
3074 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3075 /*
3076 * We are not holding a lock here, but we want to have
3077 * the pagefault_disable/enable() protection because
3078 * we want to handle the fault gracefully. If the
3079 * access fails we try to fault in the futex with R/W
3080 * verification via get_user_pages. get_user() above
3081 * does not guarantee R/W access. If that fails we
3082 * give up and leave the futex locked.
3083 */
3084 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3085 if (fault_in_user_writeable(uaddr))
3086 return -1;
3087 goto retry;
3088 }
3089 if (nval != uval)
3090 goto retry;
3091
3092 /*
3093 * Wake robust non-PI futexes here. The wakeup of
3094 * PI futexes happens in exit_pi_state():
3095 */
3096 if (!pi && (uval & FUTEX_WAITERS))
3097 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3098 }
3099 return 0;
3100}
3101
3102/*
3103 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3104 */
3105static inline int fetch_robust_entry(struct robust_list __user **entry,
3106 struct robust_list __user * __user *head,
3107 unsigned int *pi)
3108{
3109 unsigned long uentry;
3110
3111 if (get_user(uentry, (unsigned long __user *)head))
3112 return -EFAULT;
3113
3114 *entry = (void __user *)(uentry & ~1UL);
3115 *pi = uentry & 1;
3116
3117 return 0;
3118}
3119
3120/*
3121 * Walk curr->robust_list (very carefully, it's a userspace list!)
3122 * and mark any locks found there dead, and notify any waiters.
3123 *
3124 * We silently return on any sign of list-walking problem.
3125 */
3126void exit_robust_list(struct task_struct *curr)
3127{
3128 struct robust_list_head __user *head = curr->robust_list;
3129 struct robust_list __user *entry, *next_entry, *pending;
3130 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3131 unsigned int uninitialized_var(next_pi);
3132 unsigned long futex_offset;
3133 int rc;
3134
3135 if (!futex_cmpxchg_enabled)
3136 return;
3137
3138 /*
3139 * Fetch the list head (which was registered earlier, via
3140 * sys_set_robust_list()):
3141 */
3142 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3143 return;
3144 /*
3145 * Fetch the relative futex offset:
3146 */
3147 if (get_user(futex_offset, &head->futex_offset))
3148 return;
3149 /*
3150 * Fetch any possibly pending lock-add first, and handle it
3151 * if it exists:
3152 */
3153 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3154 return;
3155
3156 next_entry = NULL; /* avoid warning with gcc */
3157 while (entry != &head->list) {
3158 /*
3159 * Fetch the next entry in the list before calling
3160 * handle_futex_death:
3161 */
3162 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3163 /*
3164 * A pending lock might already be on the list, so
3165 * don't process it twice:
3166 */
3167 if (entry != pending)
3168 if (handle_futex_death((void __user *)entry + futex_offset,
3169 curr, pi))
3170 return;
3171 if (rc)
3172 return;
3173 entry = next_entry;
3174 pi = next_pi;
3175 /*
3176 * Avoid excessively long or circular lists:
3177 */
3178 if (!--limit)
3179 break;
3180
3181 cond_resched();
3182 }
3183
3184 if (pending)
3185 handle_futex_death((void __user *)pending + futex_offset,
3186 curr, pip);
3187}
3188
3189long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3190 u32 __user *uaddr2, u32 val2, u32 val3)
3191{
3192 int cmd = op & FUTEX_CMD_MASK;
3193 unsigned int flags = 0;
3194
3195 if (!(op & FUTEX_PRIVATE_FLAG))
3196 flags |= FLAGS_SHARED;
3197
3198 if (op & FUTEX_CLOCK_REALTIME) {
3199 flags |= FLAGS_CLOCKRT;
3200 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3201 cmd != FUTEX_WAIT_REQUEUE_PI)
3202 return -ENOSYS;
3203 }
3204
3205 switch (cmd) {
3206 case FUTEX_LOCK_PI:
3207 case FUTEX_UNLOCK_PI:
3208 case FUTEX_TRYLOCK_PI:
3209 case FUTEX_WAIT_REQUEUE_PI:
3210 case FUTEX_CMP_REQUEUE_PI:
3211 if (!futex_cmpxchg_enabled)
3212 return -ENOSYS;
3213 }
3214
3215 switch (cmd) {
3216 case FUTEX_WAIT:
3217 val3 = FUTEX_BITSET_MATCH_ANY;
3218 case FUTEX_WAIT_BITSET:
3219 return futex_wait(uaddr, flags, val, timeout, val3);
3220 case FUTEX_WAKE:
3221 val3 = FUTEX_BITSET_MATCH_ANY;
3222 case FUTEX_WAKE_BITSET:
3223 return futex_wake(uaddr, flags, val, val3);
3224 case FUTEX_REQUEUE:
3225 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3226 case FUTEX_CMP_REQUEUE:
3227 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3228 case FUTEX_WAKE_OP:
3229 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3230 case FUTEX_LOCK_PI:
3231 return futex_lock_pi(uaddr, flags, timeout, 0);
3232 case FUTEX_UNLOCK_PI:
3233 return futex_unlock_pi(uaddr, flags);
3234 case FUTEX_TRYLOCK_PI:
3235 return futex_lock_pi(uaddr, flags, NULL, 1);
3236 case FUTEX_WAIT_REQUEUE_PI:
3237 val3 = FUTEX_BITSET_MATCH_ANY;
3238 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3239 uaddr2);
3240 case FUTEX_CMP_REQUEUE_PI:
3241 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3242 }
3243 return -ENOSYS;
3244}
3245
3246
3247SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3248 struct timespec __user *, utime, u32 __user *, uaddr2,
3249 u32, val3)
3250{
3251 struct timespec ts;
3252 ktime_t t, *tp = NULL;
3253 u32 val2 = 0;
3254 int cmd = op & FUTEX_CMD_MASK;
3255
3256 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3257 cmd == FUTEX_WAIT_BITSET ||
3258 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3259 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3260 return -EFAULT;
3261 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3262 return -EFAULT;
3263 if (!timespec_valid(&ts))
3264 return -EINVAL;
3265
3266 t = timespec_to_ktime(ts);
3267 if (cmd == FUTEX_WAIT)
3268 t = ktime_add_safe(ktime_get(), t);
3269 tp = &t;
3270 }
3271 /*
3272 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3273 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3274 */
3275 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3276 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3277 val2 = (u32) (unsigned long) utime;
3278
3279 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3280}
3281
3282static void __init futex_detect_cmpxchg(void)
3283{
3284#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3285 u32 curval;
3286
3287 /*
3288 * This will fail and we want it. Some arch implementations do
3289 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3290 * functionality. We want to know that before we call in any
3291 * of the complex code paths. Also we want to prevent
3292 * registration of robust lists in that case. NULL is
3293 * guaranteed to fault and we get -EFAULT on functional
3294 * implementation, the non-functional ones will return
3295 * -ENOSYS.
3296 */
3297 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3298 futex_cmpxchg_enabled = 1;
3299#endif
3300}
3301
3302static int __init futex_init(void)
3303{
3304 unsigned int futex_shift;
3305 unsigned long i;
3306
3307#if CONFIG_BASE_SMALL
3308 futex_hashsize = 16;
3309#else
3310 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3311#endif
3312
3313 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3314 futex_hashsize, 0,
3315 futex_hashsize < 256 ? HASH_SMALL : 0,
3316 &futex_shift, NULL,
3317 futex_hashsize, futex_hashsize);
3318 futex_hashsize = 1UL << futex_shift;
3319
3320 futex_detect_cmpxchg();
3321
3322 for (i = 0; i < futex_hashsize; i++) {
3323 atomic_set(&futex_queues[i].waiters, 0);
3324 plist_head_init(&futex_queues[i].chain);
3325 spin_lock_init(&futex_queues[i].lock);
3326 }
3327
3328 return 0;
3329}
3330core_initcall(futex_init);
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/freezer.h>
39#include <linux/memblock.h>
40#include <linux/fault-inject.h>
41#include <linux/time_namespace.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#endif
313
314/*
315 * Reflects a new waiter being added to the waitqueue.
316 */
317static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
318{
319#ifdef CONFIG_SMP
320 atomic_inc(&hb->waiters);
321 /*
322 * Full barrier (A), see the ordering comment above.
323 */
324 smp_mb__after_atomic();
325#endif
326}
327
328/*
329 * Reflects a waiter being removed from the waitqueue by wakeup
330 * paths.
331 */
332static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
333{
334#ifdef CONFIG_SMP
335 atomic_dec(&hb->waiters);
336#endif
337}
338
339static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
340{
341#ifdef CONFIG_SMP
342 /*
343 * Full barrier (B), see the ordering comment above.
344 */
345 smp_mb();
346 return atomic_read(&hb->waiters);
347#else
348 return 1;
349#endif
350}
351
352/**
353 * hash_futex - Return the hash bucket in the global hash
354 * @key: Pointer to the futex key for which the hash is calculated
355 *
356 * We hash on the keys returned from get_futex_key (see below) and return the
357 * corresponding hash bucket in the global hash.
358 */
359static struct futex_hash_bucket *hash_futex(union futex_key *key)
360{
361 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
362 key->both.offset);
363
364 return &futex_queues[hash & (futex_hashsize - 1)];
365}
366
367
368/**
369 * match_futex - Check whether two futex keys are equal
370 * @key1: Pointer to key1
371 * @key2: Pointer to key2
372 *
373 * Return 1 if two futex_keys are equal, 0 otherwise.
374 */
375static inline int match_futex(union futex_key *key1, union futex_key *key2)
376{
377 return (key1 && key2
378 && key1->both.word == key2->both.word
379 && key1->both.ptr == key2->both.ptr
380 && key1->both.offset == key2->both.offset);
381}
382
383enum futex_access {
384 FUTEX_READ,
385 FUTEX_WRITE
386};
387
388/**
389 * futex_setup_timer - set up the sleeping hrtimer.
390 * @time: ptr to the given timeout value
391 * @timeout: the hrtimer_sleeper structure to be set up
392 * @flags: futex flags
393 * @range_ns: optional range in ns
394 *
395 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
396 * value given
397 */
398static inline struct hrtimer_sleeper *
399futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
400 int flags, u64 range_ns)
401{
402 if (!time)
403 return NULL;
404
405 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
406 CLOCK_REALTIME : CLOCK_MONOTONIC,
407 HRTIMER_MODE_ABS);
408 /*
409 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
410 * effectively the same as calling hrtimer_set_expires().
411 */
412 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
413
414 return timeout;
415}
416
417/*
418 * Generate a machine wide unique identifier for this inode.
419 *
420 * This relies on u64 not wrapping in the life-time of the machine; which with
421 * 1ns resolution means almost 585 years.
422 *
423 * This further relies on the fact that a well formed program will not unmap
424 * the file while it has a (shared) futex waiting on it. This mapping will have
425 * a file reference which pins the mount and inode.
426 *
427 * If for some reason an inode gets evicted and read back in again, it will get
428 * a new sequence number and will _NOT_ match, even though it is the exact same
429 * file.
430 *
431 * It is important that match_futex() will never have a false-positive, esp.
432 * for PI futexes that can mess up the state. The above argues that false-negatives
433 * are only possible for malformed programs.
434 */
435static u64 get_inode_sequence_number(struct inode *inode)
436{
437 static atomic64_t i_seq;
438 u64 old;
439
440 /* Does the inode already have a sequence number? */
441 old = atomic64_read(&inode->i_sequence);
442 if (likely(old))
443 return old;
444
445 for (;;) {
446 u64 new = atomic64_add_return(1, &i_seq);
447 if (WARN_ON_ONCE(!new))
448 continue;
449
450 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
451 if (old)
452 return old;
453 return new;
454 }
455}
456
457/**
458 * get_futex_key() - Get parameters which are the keys for a futex
459 * @uaddr: virtual address of the futex
460 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
461 * @key: address where result is stored.
462 * @rw: mapping needs to be read/write (values: FUTEX_READ,
463 * FUTEX_WRITE)
464 *
465 * Return: a negative error code or 0
466 *
467 * The key words are stored in @key on success.
468 *
469 * For shared mappings (when @fshared), the key is:
470 *
471 * ( inode->i_sequence, page->index, offset_within_page )
472 *
473 * [ also see get_inode_sequence_number() ]
474 *
475 * For private mappings (or when !@fshared), the key is:
476 *
477 * ( current->mm, address, 0 )
478 *
479 * This allows (cross process, where applicable) identification of the futex
480 * without keeping the page pinned for the duration of the FUTEX_WAIT.
481 *
482 * lock_page() might sleep, the caller should not hold a spinlock.
483 */
484static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
485 enum futex_access rw)
486{
487 unsigned long address = (unsigned long)uaddr;
488 struct mm_struct *mm = current->mm;
489 struct page *page, *tail;
490 struct address_space *mapping;
491 int err, ro = 0;
492
493 /*
494 * The futex address must be "naturally" aligned.
495 */
496 key->both.offset = address % PAGE_SIZE;
497 if (unlikely((address % sizeof(u32)) != 0))
498 return -EINVAL;
499 address -= key->both.offset;
500
501 if (unlikely(!access_ok(uaddr, sizeof(u32))))
502 return -EFAULT;
503
504 if (unlikely(should_fail_futex(fshared)))
505 return -EFAULT;
506
507 /*
508 * PROCESS_PRIVATE futexes are fast.
509 * As the mm cannot disappear under us and the 'key' only needs
510 * virtual address, we dont even have to find the underlying vma.
511 * Note : We do have to check 'uaddr' is a valid user address,
512 * but access_ok() should be faster than find_vma()
513 */
514 if (!fshared) {
515 key->private.mm = mm;
516 key->private.address = address;
517 return 0;
518 }
519
520again:
521 /* Ignore any VERIFY_READ mapping (futex common case) */
522 if (unlikely(should_fail_futex(true)))
523 return -EFAULT;
524
525 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
526 /*
527 * If write access is not required (eg. FUTEX_WAIT), try
528 * and get read-only access.
529 */
530 if (err == -EFAULT && rw == FUTEX_READ) {
531 err = get_user_pages_fast(address, 1, 0, &page);
532 ro = 1;
533 }
534 if (err < 0)
535 return err;
536 else
537 err = 0;
538
539 /*
540 * The treatment of mapping from this point on is critical. The page
541 * lock protects many things but in this context the page lock
542 * stabilizes mapping, prevents inode freeing in the shared
543 * file-backed region case and guards against movement to swap cache.
544 *
545 * Strictly speaking the page lock is not needed in all cases being
546 * considered here and page lock forces unnecessarily serialization
547 * From this point on, mapping will be re-verified if necessary and
548 * page lock will be acquired only if it is unavoidable
549 *
550 * Mapping checks require the head page for any compound page so the
551 * head page and mapping is looked up now. For anonymous pages, it
552 * does not matter if the page splits in the future as the key is
553 * based on the address. For filesystem-backed pages, the tail is
554 * required as the index of the page determines the key. For
555 * base pages, there is no tail page and tail == page.
556 */
557 tail = page;
558 page = compound_head(page);
559 mapping = READ_ONCE(page->mapping);
560
561 /*
562 * If page->mapping is NULL, then it cannot be a PageAnon
563 * page; but it might be the ZERO_PAGE or in the gate area or
564 * in a special mapping (all cases which we are happy to fail);
565 * or it may have been a good file page when get_user_pages_fast
566 * found it, but truncated or holepunched or subjected to
567 * invalidate_complete_page2 before we got the page lock (also
568 * cases which we are happy to fail). And we hold a reference,
569 * so refcount care in invalidate_complete_page's remove_mapping
570 * prevents drop_caches from setting mapping to NULL beneath us.
571 *
572 * The case we do have to guard against is when memory pressure made
573 * shmem_writepage move it from filecache to swapcache beneath us:
574 * an unlikely race, but we do need to retry for page->mapping.
575 */
576 if (unlikely(!mapping)) {
577 int shmem_swizzled;
578
579 /*
580 * Page lock is required to identify which special case above
581 * applies. If this is really a shmem page then the page lock
582 * will prevent unexpected transitions.
583 */
584 lock_page(page);
585 shmem_swizzled = PageSwapCache(page) || page->mapping;
586 unlock_page(page);
587 put_page(page);
588
589 if (shmem_swizzled)
590 goto again;
591
592 return -EFAULT;
593 }
594
595 /*
596 * Private mappings are handled in a simple way.
597 *
598 * If the futex key is stored on an anonymous page, then the associated
599 * object is the mm which is implicitly pinned by the calling process.
600 *
601 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
602 * it's a read-only handle, it's expected that futexes attach to
603 * the object not the particular process.
604 */
605 if (PageAnon(page)) {
606 /*
607 * A RO anonymous page will never change and thus doesn't make
608 * sense for futex operations.
609 */
610 if (unlikely(should_fail_futex(true)) || ro) {
611 err = -EFAULT;
612 goto out;
613 }
614
615 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
616 key->private.mm = mm;
617 key->private.address = address;
618
619 } else {
620 struct inode *inode;
621
622 /*
623 * The associated futex object in this case is the inode and
624 * the page->mapping must be traversed. Ordinarily this should
625 * be stabilised under page lock but it's not strictly
626 * necessary in this case as we just want to pin the inode, not
627 * update the radix tree or anything like that.
628 *
629 * The RCU read lock is taken as the inode is finally freed
630 * under RCU. If the mapping still matches expectations then the
631 * mapping->host can be safely accessed as being a valid inode.
632 */
633 rcu_read_lock();
634
635 if (READ_ONCE(page->mapping) != mapping) {
636 rcu_read_unlock();
637 put_page(page);
638
639 goto again;
640 }
641
642 inode = READ_ONCE(mapping->host);
643 if (!inode) {
644 rcu_read_unlock();
645 put_page(page);
646
647 goto again;
648 }
649
650 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
651 key->shared.i_seq = get_inode_sequence_number(inode);
652 key->shared.pgoff = page_to_pgoff(tail);
653 rcu_read_unlock();
654 }
655
656out:
657 put_page(page);
658 return err;
659}
660
661/**
662 * fault_in_user_writeable() - Fault in user address and verify RW access
663 * @uaddr: pointer to faulting user space address
664 *
665 * Slow path to fixup the fault we just took in the atomic write
666 * access to @uaddr.
667 *
668 * We have no generic implementation of a non-destructive write to the
669 * user address. We know that we faulted in the atomic pagefault
670 * disabled section so we can as well avoid the #PF overhead by
671 * calling get_user_pages() right away.
672 */
673static int fault_in_user_writeable(u32 __user *uaddr)
674{
675 struct mm_struct *mm = current->mm;
676 int ret;
677
678 mmap_read_lock(mm);
679 ret = fixup_user_fault(mm, (unsigned long)uaddr,
680 FAULT_FLAG_WRITE, NULL);
681 mmap_read_unlock(mm);
682
683 return ret < 0 ? ret : 0;
684}
685
686/**
687 * futex_top_waiter() - Return the highest priority waiter on a futex
688 * @hb: the hash bucket the futex_q's reside in
689 * @key: the futex key (to distinguish it from other futex futex_q's)
690 *
691 * Must be called with the hb lock held.
692 */
693static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
694 union futex_key *key)
695{
696 struct futex_q *this;
697
698 plist_for_each_entry(this, &hb->chain, list) {
699 if (match_futex(&this->key, key))
700 return this;
701 }
702 return NULL;
703}
704
705static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
706 u32 uval, u32 newval)
707{
708 int ret;
709
710 pagefault_disable();
711 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
712 pagefault_enable();
713
714 return ret;
715}
716
717static int get_futex_value_locked(u32 *dest, u32 __user *from)
718{
719 int ret;
720
721 pagefault_disable();
722 ret = __get_user(*dest, from);
723 pagefault_enable();
724
725 return ret ? -EFAULT : 0;
726}
727
728
729/*
730 * PI code:
731 */
732static int refill_pi_state_cache(void)
733{
734 struct futex_pi_state *pi_state;
735
736 if (likely(current->pi_state_cache))
737 return 0;
738
739 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
740
741 if (!pi_state)
742 return -ENOMEM;
743
744 INIT_LIST_HEAD(&pi_state->list);
745 /* pi_mutex gets initialized later */
746 pi_state->owner = NULL;
747 refcount_set(&pi_state->refcount, 1);
748 pi_state->key = FUTEX_KEY_INIT;
749
750 current->pi_state_cache = pi_state;
751
752 return 0;
753}
754
755static struct futex_pi_state *alloc_pi_state(void)
756{
757 struct futex_pi_state *pi_state = current->pi_state_cache;
758
759 WARN_ON(!pi_state);
760 current->pi_state_cache = NULL;
761
762 return pi_state;
763}
764
765static void pi_state_update_owner(struct futex_pi_state *pi_state,
766 struct task_struct *new_owner)
767{
768 struct task_struct *old_owner = pi_state->owner;
769
770 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
771
772 if (old_owner) {
773 raw_spin_lock(&old_owner->pi_lock);
774 WARN_ON(list_empty(&pi_state->list));
775 list_del_init(&pi_state->list);
776 raw_spin_unlock(&old_owner->pi_lock);
777 }
778
779 if (new_owner) {
780 raw_spin_lock(&new_owner->pi_lock);
781 WARN_ON(!list_empty(&pi_state->list));
782 list_add(&pi_state->list, &new_owner->pi_state_list);
783 pi_state->owner = new_owner;
784 raw_spin_unlock(&new_owner->pi_lock);
785 }
786}
787
788static void get_pi_state(struct futex_pi_state *pi_state)
789{
790 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
791}
792
793/*
794 * Drops a reference to the pi_state object and frees or caches it
795 * when the last reference is gone.
796 */
797static void put_pi_state(struct futex_pi_state *pi_state)
798{
799 if (!pi_state)
800 return;
801
802 if (!refcount_dec_and_test(&pi_state->refcount))
803 return;
804
805 /*
806 * If pi_state->owner is NULL, the owner is most probably dying
807 * and has cleaned up the pi_state already
808 */
809 if (pi_state->owner) {
810 unsigned long flags;
811
812 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
813 pi_state_update_owner(pi_state, NULL);
814 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
815 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
816 }
817
818 if (current->pi_state_cache) {
819 kfree(pi_state);
820 } else {
821 /*
822 * pi_state->list is already empty.
823 * clear pi_state->owner.
824 * refcount is at 0 - put it back to 1.
825 */
826 pi_state->owner = NULL;
827 refcount_set(&pi_state->refcount, 1);
828 current->pi_state_cache = pi_state;
829 }
830}
831
832#ifdef CONFIG_FUTEX_PI
833
834/*
835 * This task is holding PI mutexes at exit time => bad.
836 * Kernel cleans up PI-state, but userspace is likely hosed.
837 * (Robust-futex cleanup is separate and might save the day for userspace.)
838 */
839static void exit_pi_state_list(struct task_struct *curr)
840{
841 struct list_head *next, *head = &curr->pi_state_list;
842 struct futex_pi_state *pi_state;
843 struct futex_hash_bucket *hb;
844 union futex_key key = FUTEX_KEY_INIT;
845
846 if (!futex_cmpxchg_enabled)
847 return;
848 /*
849 * We are a ZOMBIE and nobody can enqueue itself on
850 * pi_state_list anymore, but we have to be careful
851 * versus waiters unqueueing themselves:
852 */
853 raw_spin_lock_irq(&curr->pi_lock);
854 while (!list_empty(head)) {
855 next = head->next;
856 pi_state = list_entry(next, struct futex_pi_state, list);
857 key = pi_state->key;
858 hb = hash_futex(&key);
859
860 /*
861 * We can race against put_pi_state() removing itself from the
862 * list (a waiter going away). put_pi_state() will first
863 * decrement the reference count and then modify the list, so
864 * its possible to see the list entry but fail this reference
865 * acquire.
866 *
867 * In that case; drop the locks to let put_pi_state() make
868 * progress and retry the loop.
869 */
870 if (!refcount_inc_not_zero(&pi_state->refcount)) {
871 raw_spin_unlock_irq(&curr->pi_lock);
872 cpu_relax();
873 raw_spin_lock_irq(&curr->pi_lock);
874 continue;
875 }
876 raw_spin_unlock_irq(&curr->pi_lock);
877
878 spin_lock(&hb->lock);
879 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
880 raw_spin_lock(&curr->pi_lock);
881 /*
882 * We dropped the pi-lock, so re-check whether this
883 * task still owns the PI-state:
884 */
885 if (head->next != next) {
886 /* retain curr->pi_lock for the loop invariant */
887 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
888 spin_unlock(&hb->lock);
889 put_pi_state(pi_state);
890 continue;
891 }
892
893 WARN_ON(pi_state->owner != curr);
894 WARN_ON(list_empty(&pi_state->list));
895 list_del_init(&pi_state->list);
896 pi_state->owner = NULL;
897
898 raw_spin_unlock(&curr->pi_lock);
899 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
900 spin_unlock(&hb->lock);
901
902 rt_mutex_futex_unlock(&pi_state->pi_mutex);
903 put_pi_state(pi_state);
904
905 raw_spin_lock_irq(&curr->pi_lock);
906 }
907 raw_spin_unlock_irq(&curr->pi_lock);
908}
909#else
910static inline void exit_pi_state_list(struct task_struct *curr) { }
911#endif
912
913/*
914 * We need to check the following states:
915 *
916 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
917 *
918 * [1] NULL | --- | --- | 0 | 0/1 | Valid
919 * [2] NULL | --- | --- | >0 | 0/1 | Valid
920 *
921 * [3] Found | NULL | -- | Any | 0/1 | Invalid
922 *
923 * [4] Found | Found | NULL | 0 | 1 | Valid
924 * [5] Found | Found | NULL | >0 | 1 | Invalid
925 *
926 * [6] Found | Found | task | 0 | 1 | Valid
927 *
928 * [7] Found | Found | NULL | Any | 0 | Invalid
929 *
930 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
931 * [9] Found | Found | task | 0 | 0 | Invalid
932 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
933 *
934 * [1] Indicates that the kernel can acquire the futex atomically. We
935 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
936 *
937 * [2] Valid, if TID does not belong to a kernel thread. If no matching
938 * thread is found then it indicates that the owner TID has died.
939 *
940 * [3] Invalid. The waiter is queued on a non PI futex
941 *
942 * [4] Valid state after exit_robust_list(), which sets the user space
943 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
944 *
945 * [5] The user space value got manipulated between exit_robust_list()
946 * and exit_pi_state_list()
947 *
948 * [6] Valid state after exit_pi_state_list() which sets the new owner in
949 * the pi_state but cannot access the user space value.
950 *
951 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
952 *
953 * [8] Owner and user space value match
954 *
955 * [9] There is no transient state which sets the user space TID to 0
956 * except exit_robust_list(), but this is indicated by the
957 * FUTEX_OWNER_DIED bit. See [4]
958 *
959 * [10] There is no transient state which leaves owner and user space
960 * TID out of sync. Except one error case where the kernel is denied
961 * write access to the user address, see fixup_pi_state_owner().
962 *
963 *
964 * Serialization and lifetime rules:
965 *
966 * hb->lock:
967 *
968 * hb -> futex_q, relation
969 * futex_q -> pi_state, relation
970 *
971 * (cannot be raw because hb can contain arbitrary amount
972 * of futex_q's)
973 *
974 * pi_mutex->wait_lock:
975 *
976 * {uval, pi_state}
977 *
978 * (and pi_mutex 'obviously')
979 *
980 * p->pi_lock:
981 *
982 * p->pi_state_list -> pi_state->list, relation
983 * pi_mutex->owner -> pi_state->owner, relation
984 *
985 * pi_state->refcount:
986 *
987 * pi_state lifetime
988 *
989 *
990 * Lock order:
991 *
992 * hb->lock
993 * pi_mutex->wait_lock
994 * p->pi_lock
995 *
996 */
997
998/*
999 * Validate that the existing waiter has a pi_state and sanity check
1000 * the pi_state against the user space value. If correct, attach to
1001 * it.
1002 */
1003static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1004 struct futex_pi_state *pi_state,
1005 struct futex_pi_state **ps)
1006{
1007 pid_t pid = uval & FUTEX_TID_MASK;
1008 u32 uval2;
1009 int ret;
1010
1011 /*
1012 * Userspace might have messed up non-PI and PI futexes [3]
1013 */
1014 if (unlikely(!pi_state))
1015 return -EINVAL;
1016
1017 /*
1018 * We get here with hb->lock held, and having found a
1019 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1020 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1021 * which in turn means that futex_lock_pi() still has a reference on
1022 * our pi_state.
1023 *
1024 * The waiter holding a reference on @pi_state also protects against
1025 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1026 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1027 * free pi_state before we can take a reference ourselves.
1028 */
1029 WARN_ON(!refcount_read(&pi_state->refcount));
1030
1031 /*
1032 * Now that we have a pi_state, we can acquire wait_lock
1033 * and do the state validation.
1034 */
1035 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1036
1037 /*
1038 * Since {uval, pi_state} is serialized by wait_lock, and our current
1039 * uval was read without holding it, it can have changed. Verify it
1040 * still is what we expect it to be, otherwise retry the entire
1041 * operation.
1042 */
1043 if (get_futex_value_locked(&uval2, uaddr))
1044 goto out_efault;
1045
1046 if (uval != uval2)
1047 goto out_eagain;
1048
1049 /*
1050 * Handle the owner died case:
1051 */
1052 if (uval & FUTEX_OWNER_DIED) {
1053 /*
1054 * exit_pi_state_list sets owner to NULL and wakes the
1055 * topmost waiter. The task which acquires the
1056 * pi_state->rt_mutex will fixup owner.
1057 */
1058 if (!pi_state->owner) {
1059 /*
1060 * No pi state owner, but the user space TID
1061 * is not 0. Inconsistent state. [5]
1062 */
1063 if (pid)
1064 goto out_einval;
1065 /*
1066 * Take a ref on the state and return success. [4]
1067 */
1068 goto out_attach;
1069 }
1070
1071 /*
1072 * If TID is 0, then either the dying owner has not
1073 * yet executed exit_pi_state_list() or some waiter
1074 * acquired the rtmutex in the pi state, but did not
1075 * yet fixup the TID in user space.
1076 *
1077 * Take a ref on the state and return success. [6]
1078 */
1079 if (!pid)
1080 goto out_attach;
1081 } else {
1082 /*
1083 * If the owner died bit is not set, then the pi_state
1084 * must have an owner. [7]
1085 */
1086 if (!pi_state->owner)
1087 goto out_einval;
1088 }
1089
1090 /*
1091 * Bail out if user space manipulated the futex value. If pi
1092 * state exists then the owner TID must be the same as the
1093 * user space TID. [9/10]
1094 */
1095 if (pid != task_pid_vnr(pi_state->owner))
1096 goto out_einval;
1097
1098out_attach:
1099 get_pi_state(pi_state);
1100 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1101 *ps = pi_state;
1102 return 0;
1103
1104out_einval:
1105 ret = -EINVAL;
1106 goto out_error;
1107
1108out_eagain:
1109 ret = -EAGAIN;
1110 goto out_error;
1111
1112out_efault:
1113 ret = -EFAULT;
1114 goto out_error;
1115
1116out_error:
1117 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1118 return ret;
1119}
1120
1121/**
1122 * wait_for_owner_exiting - Block until the owner has exited
1123 * @ret: owner's current futex lock status
1124 * @exiting: Pointer to the exiting task
1125 *
1126 * Caller must hold a refcount on @exiting.
1127 */
1128static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1129{
1130 if (ret != -EBUSY) {
1131 WARN_ON_ONCE(exiting);
1132 return;
1133 }
1134
1135 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1136 return;
1137
1138 mutex_lock(&exiting->futex_exit_mutex);
1139 /*
1140 * No point in doing state checking here. If the waiter got here
1141 * while the task was in exec()->exec_futex_release() then it can
1142 * have any FUTEX_STATE_* value when the waiter has acquired the
1143 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1144 * already. Highly unlikely and not a problem. Just one more round
1145 * through the futex maze.
1146 */
1147 mutex_unlock(&exiting->futex_exit_mutex);
1148
1149 put_task_struct(exiting);
1150}
1151
1152static int handle_exit_race(u32 __user *uaddr, u32 uval,
1153 struct task_struct *tsk)
1154{
1155 u32 uval2;
1156
1157 /*
1158 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1159 * caller that the alleged owner is busy.
1160 */
1161 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1162 return -EBUSY;
1163
1164 /*
1165 * Reread the user space value to handle the following situation:
1166 *
1167 * CPU0 CPU1
1168 *
1169 * sys_exit() sys_futex()
1170 * do_exit() futex_lock_pi()
1171 * futex_lock_pi_atomic()
1172 * exit_signals(tsk) No waiters:
1173 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1174 * mm_release(tsk) Set waiter bit
1175 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1176 * Set owner died attach_to_pi_owner() {
1177 * *uaddr = 0xC0000000; tsk = get_task(PID);
1178 * } if (!tsk->flags & PF_EXITING) {
1179 * ... attach();
1180 * tsk->futex_state = } else {
1181 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1182 * FUTEX_STATE_DEAD)
1183 * return -EAGAIN;
1184 * return -ESRCH; <--- FAIL
1185 * }
1186 *
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1189 *
1190 * The same logic applies to the case where the exiting task is
1191 * already gone.
1192 */
1193 if (get_futex_value_locked(&uval2, uaddr))
1194 return -EFAULT;
1195
1196 /* If the user space value has changed, try again. */
1197 if (uval2 != uval)
1198 return -EAGAIN;
1199
1200 /*
1201 * The exiting task did not have a robust list, the robust list was
1202 * corrupted or the user space value in *uaddr is simply bogus.
1203 * Give up and tell user space.
1204 */
1205 return -ESRCH;
1206}
1207
1208/*
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1211 */
1212static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213 struct futex_pi_state **ps,
1214 struct task_struct **exiting)
1215{
1216 pid_t pid = uval & FUTEX_TID_MASK;
1217 struct futex_pi_state *pi_state;
1218 struct task_struct *p;
1219
1220 /*
1221 * We are the first waiter - try to look up the real owner and attach
1222 * the new pi_state to it, but bail out when TID = 0 [1]
1223 *
1224 * The !pid check is paranoid. None of the call sites should end up
1225 * with pid == 0, but better safe than sorry. Let the caller retry
1226 */
1227 if (!pid)
1228 return -EAGAIN;
1229 p = find_get_task_by_vpid(pid);
1230 if (!p)
1231 return handle_exit_race(uaddr, uval, NULL);
1232
1233 if (unlikely(p->flags & PF_KTHREAD)) {
1234 put_task_struct(p);
1235 return -EPERM;
1236 }
1237
1238 /*
1239 * We need to look at the task state to figure out, whether the
1240 * task is exiting. To protect against the change of the task state
1241 * in futex_exit_release(), we do this protected by p->pi_lock:
1242 */
1243 raw_spin_lock_irq(&p->pi_lock);
1244 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1245 /*
1246 * The task is on the way out. When the futex state is
1247 * FUTEX_STATE_DEAD, we know that the task has finished
1248 * the cleanup:
1249 */
1250 int ret = handle_exit_race(uaddr, uval, p);
1251
1252 raw_spin_unlock_irq(&p->pi_lock);
1253 /*
1254 * If the owner task is between FUTEX_STATE_EXITING and
1255 * FUTEX_STATE_DEAD then store the task pointer and keep
1256 * the reference on the task struct. The calling code will
1257 * drop all locks, wait for the task to reach
1258 * FUTEX_STATE_DEAD and then drop the refcount. This is
1259 * required to prevent a live lock when the current task
1260 * preempted the exiting task between the two states.
1261 */
1262 if (ret == -EBUSY)
1263 *exiting = p;
1264 else
1265 put_task_struct(p);
1266 return ret;
1267 }
1268
1269 /*
1270 * No existing pi state. First waiter. [2]
1271 *
1272 * This creates pi_state, we have hb->lock held, this means nothing can
1273 * observe this state, wait_lock is irrelevant.
1274 */
1275 pi_state = alloc_pi_state();
1276
1277 /*
1278 * Initialize the pi_mutex in locked state and make @p
1279 * the owner of it:
1280 */
1281 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1282
1283 /* Store the key for possible exit cleanups: */
1284 pi_state->key = *key;
1285
1286 WARN_ON(!list_empty(&pi_state->list));
1287 list_add(&pi_state->list, &p->pi_state_list);
1288 /*
1289 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1290 * because there is no concurrency as the object is not published yet.
1291 */
1292 pi_state->owner = p;
1293 raw_spin_unlock_irq(&p->pi_lock);
1294
1295 put_task_struct(p);
1296
1297 *ps = pi_state;
1298
1299 return 0;
1300}
1301
1302static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1303 struct futex_hash_bucket *hb,
1304 union futex_key *key, struct futex_pi_state **ps,
1305 struct task_struct **exiting)
1306{
1307 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1308
1309 /*
1310 * If there is a waiter on that futex, validate it and
1311 * attach to the pi_state when the validation succeeds.
1312 */
1313 if (top_waiter)
1314 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1315
1316 /*
1317 * We are the first waiter - try to look up the owner based on
1318 * @uval and attach to it.
1319 */
1320 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1321}
1322
1323static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1324{
1325 int err;
1326 u32 curval;
1327
1328 if (unlikely(should_fail_futex(true)))
1329 return -EFAULT;
1330
1331 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1332 if (unlikely(err))
1333 return err;
1334
1335 /* If user space value changed, let the caller retry */
1336 return curval != uval ? -EAGAIN : 0;
1337}
1338
1339/**
1340 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1341 * @uaddr: the pi futex user address
1342 * @hb: the pi futex hash bucket
1343 * @key: the futex key associated with uaddr and hb
1344 * @ps: the pi_state pointer where we store the result of the
1345 * lookup
1346 * @task: the task to perform the atomic lock work for. This will
1347 * be "current" except in the case of requeue pi.
1348 * @exiting: Pointer to store the task pointer of the owner task
1349 * which is in the middle of exiting
1350 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1351 *
1352 * Return:
1353 * - 0 - ready to wait;
1354 * - 1 - acquired the lock;
1355 * - <0 - error
1356 *
1357 * The hb->lock and futex_key refs shall be held by the caller.
1358 *
1359 * @exiting is only set when the return value is -EBUSY. If so, this holds
1360 * a refcount on the exiting task on return and the caller needs to drop it
1361 * after waiting for the exit to complete.
1362 */
1363static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1364 union futex_key *key,
1365 struct futex_pi_state **ps,
1366 struct task_struct *task,
1367 struct task_struct **exiting,
1368 int set_waiters)
1369{
1370 u32 uval, newval, vpid = task_pid_vnr(task);
1371 struct futex_q *top_waiter;
1372 int ret;
1373
1374 /*
1375 * Read the user space value first so we can validate a few
1376 * things before proceeding further.
1377 */
1378 if (get_futex_value_locked(&uval, uaddr))
1379 return -EFAULT;
1380
1381 if (unlikely(should_fail_futex(true)))
1382 return -EFAULT;
1383
1384 /*
1385 * Detect deadlocks.
1386 */
1387 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1388 return -EDEADLK;
1389
1390 if ((unlikely(should_fail_futex(true))))
1391 return -EDEADLK;
1392
1393 /*
1394 * Lookup existing state first. If it exists, try to attach to
1395 * its pi_state.
1396 */
1397 top_waiter = futex_top_waiter(hb, key);
1398 if (top_waiter)
1399 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1400
1401 /*
1402 * No waiter and user TID is 0. We are here because the
1403 * waiters or the owner died bit is set or called from
1404 * requeue_cmp_pi or for whatever reason something took the
1405 * syscall.
1406 */
1407 if (!(uval & FUTEX_TID_MASK)) {
1408 /*
1409 * We take over the futex. No other waiters and the user space
1410 * TID is 0. We preserve the owner died bit.
1411 */
1412 newval = uval & FUTEX_OWNER_DIED;
1413 newval |= vpid;
1414
1415 /* The futex requeue_pi code can enforce the waiters bit */
1416 if (set_waiters)
1417 newval |= FUTEX_WAITERS;
1418
1419 ret = lock_pi_update_atomic(uaddr, uval, newval);
1420 /* If the take over worked, return 1 */
1421 return ret < 0 ? ret : 1;
1422 }
1423
1424 /*
1425 * First waiter. Set the waiters bit before attaching ourself to
1426 * the owner. If owner tries to unlock, it will be forced into
1427 * the kernel and blocked on hb->lock.
1428 */
1429 newval = uval | FUTEX_WAITERS;
1430 ret = lock_pi_update_atomic(uaddr, uval, newval);
1431 if (ret)
1432 return ret;
1433 /*
1434 * If the update of the user space value succeeded, we try to
1435 * attach to the owner. If that fails, no harm done, we only
1436 * set the FUTEX_WAITERS bit in the user space variable.
1437 */
1438 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1439}
1440
1441/**
1442 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1443 * @q: The futex_q to unqueue
1444 *
1445 * The q->lock_ptr must not be NULL and must be held by the caller.
1446 */
1447static void __unqueue_futex(struct futex_q *q)
1448{
1449 struct futex_hash_bucket *hb;
1450
1451 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1452 return;
1453 lockdep_assert_held(q->lock_ptr);
1454
1455 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1456 plist_del(&q->list, &hb->chain);
1457 hb_waiters_dec(hb);
1458}
1459
1460/*
1461 * The hash bucket lock must be held when this is called.
1462 * Afterwards, the futex_q must not be accessed. Callers
1463 * must ensure to later call wake_up_q() for the actual
1464 * wakeups to occur.
1465 */
1466static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1467{
1468 struct task_struct *p = q->task;
1469
1470 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1471 return;
1472
1473 get_task_struct(p);
1474 __unqueue_futex(q);
1475 /*
1476 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1477 * is written, without taking any locks. This is possible in the event
1478 * of a spurious wakeup, for example. A memory barrier is required here
1479 * to prevent the following store to lock_ptr from getting ahead of the
1480 * plist_del in __unqueue_futex().
1481 */
1482 smp_store_release(&q->lock_ptr, NULL);
1483
1484 /*
1485 * Queue the task for later wakeup for after we've released
1486 * the hb->lock.
1487 */
1488 wake_q_add_safe(wake_q, p);
1489}
1490
1491/*
1492 * Caller must hold a reference on @pi_state.
1493 */
1494static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1495{
1496 u32 curval, newval;
1497 struct rt_mutex_waiter *top_waiter;
1498 struct task_struct *new_owner;
1499 bool postunlock = false;
1500 DEFINE_WAKE_Q(wake_q);
1501 int ret = 0;
1502
1503 top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1504 if (WARN_ON_ONCE(!top_waiter)) {
1505 /*
1506 * As per the comment in futex_unlock_pi() this should not happen.
1507 *
1508 * When this happens, give up our locks and try again, giving
1509 * the futex_lock_pi() instance time to complete, either by
1510 * waiting on the rtmutex or removing itself from the futex
1511 * queue.
1512 */
1513 ret = -EAGAIN;
1514 goto out_unlock;
1515 }
1516
1517 new_owner = top_waiter->task;
1518
1519 /*
1520 * We pass it to the next owner. The WAITERS bit is always kept
1521 * enabled while there is PI state around. We cleanup the owner
1522 * died bit, because we are the owner.
1523 */
1524 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1525
1526 if (unlikely(should_fail_futex(true))) {
1527 ret = -EFAULT;
1528 goto out_unlock;
1529 }
1530
1531 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1532 if (!ret && (curval != uval)) {
1533 /*
1534 * If a unconditional UNLOCK_PI operation (user space did not
1535 * try the TID->0 transition) raced with a waiter setting the
1536 * FUTEX_WAITERS flag between get_user() and locking the hash
1537 * bucket lock, retry the operation.
1538 */
1539 if ((FUTEX_TID_MASK & curval) == uval)
1540 ret = -EAGAIN;
1541 else
1542 ret = -EINVAL;
1543 }
1544
1545 if (!ret) {
1546 /*
1547 * This is a point of no return; once we modified the uval
1548 * there is no going back and subsequent operations must
1549 * not fail.
1550 */
1551 pi_state_update_owner(pi_state, new_owner);
1552 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1553 }
1554
1555out_unlock:
1556 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1557
1558 if (postunlock)
1559 rt_mutex_postunlock(&wake_q);
1560
1561 return ret;
1562}
1563
1564/*
1565 * Express the locking dependencies for lockdep:
1566 */
1567static inline void
1568double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1569{
1570 if (hb1 <= hb2) {
1571 spin_lock(&hb1->lock);
1572 if (hb1 < hb2)
1573 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1574 } else { /* hb1 > hb2 */
1575 spin_lock(&hb2->lock);
1576 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1577 }
1578}
1579
1580static inline void
1581double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1582{
1583 spin_unlock(&hb1->lock);
1584 if (hb1 != hb2)
1585 spin_unlock(&hb2->lock);
1586}
1587
1588/*
1589 * Wake up waiters matching bitset queued on this futex (uaddr).
1590 */
1591static int
1592futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1593{
1594 struct futex_hash_bucket *hb;
1595 struct futex_q *this, *next;
1596 union futex_key key = FUTEX_KEY_INIT;
1597 int ret;
1598 DEFINE_WAKE_Q(wake_q);
1599
1600 if (!bitset)
1601 return -EINVAL;
1602
1603 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1604 if (unlikely(ret != 0))
1605 return ret;
1606
1607 hb = hash_futex(&key);
1608
1609 /* Make sure we really have tasks to wakeup */
1610 if (!hb_waiters_pending(hb))
1611 return ret;
1612
1613 spin_lock(&hb->lock);
1614
1615 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1616 if (match_futex (&this->key, &key)) {
1617 if (this->pi_state || this->rt_waiter) {
1618 ret = -EINVAL;
1619 break;
1620 }
1621
1622 /* Check if one of the bits is set in both bitsets */
1623 if (!(this->bitset & bitset))
1624 continue;
1625
1626 mark_wake_futex(&wake_q, this);
1627 if (++ret >= nr_wake)
1628 break;
1629 }
1630 }
1631
1632 spin_unlock(&hb->lock);
1633 wake_up_q(&wake_q);
1634 return ret;
1635}
1636
1637static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1638{
1639 unsigned int op = (encoded_op & 0x70000000) >> 28;
1640 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1641 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1642 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1643 int oldval, ret;
1644
1645 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1646 if (oparg < 0 || oparg > 31) {
1647 char comm[sizeof(current->comm)];
1648 /*
1649 * kill this print and return -EINVAL when userspace
1650 * is sane again
1651 */
1652 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1653 get_task_comm(comm, current), oparg);
1654 oparg &= 31;
1655 }
1656 oparg = 1 << oparg;
1657 }
1658
1659 pagefault_disable();
1660 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1661 pagefault_enable();
1662 if (ret)
1663 return ret;
1664
1665 switch (cmp) {
1666 case FUTEX_OP_CMP_EQ:
1667 return oldval == cmparg;
1668 case FUTEX_OP_CMP_NE:
1669 return oldval != cmparg;
1670 case FUTEX_OP_CMP_LT:
1671 return oldval < cmparg;
1672 case FUTEX_OP_CMP_GE:
1673 return oldval >= cmparg;
1674 case FUTEX_OP_CMP_LE:
1675 return oldval <= cmparg;
1676 case FUTEX_OP_CMP_GT:
1677 return oldval > cmparg;
1678 default:
1679 return -ENOSYS;
1680 }
1681}
1682
1683/*
1684 * Wake up all waiters hashed on the physical page that is mapped
1685 * to this virtual address:
1686 */
1687static int
1688futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1689 int nr_wake, int nr_wake2, int op)
1690{
1691 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1692 struct futex_hash_bucket *hb1, *hb2;
1693 struct futex_q *this, *next;
1694 int ret, op_ret;
1695 DEFINE_WAKE_Q(wake_q);
1696
1697retry:
1698 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1699 if (unlikely(ret != 0))
1700 return ret;
1701 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1702 if (unlikely(ret != 0))
1703 return ret;
1704
1705 hb1 = hash_futex(&key1);
1706 hb2 = hash_futex(&key2);
1707
1708retry_private:
1709 double_lock_hb(hb1, hb2);
1710 op_ret = futex_atomic_op_inuser(op, uaddr2);
1711 if (unlikely(op_ret < 0)) {
1712 double_unlock_hb(hb1, hb2);
1713
1714 if (!IS_ENABLED(CONFIG_MMU) ||
1715 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1716 /*
1717 * we don't get EFAULT from MMU faults if we don't have
1718 * an MMU, but we might get them from range checking
1719 */
1720 ret = op_ret;
1721 return ret;
1722 }
1723
1724 if (op_ret == -EFAULT) {
1725 ret = fault_in_user_writeable(uaddr2);
1726 if (ret)
1727 return ret;
1728 }
1729
1730 cond_resched();
1731 if (!(flags & FLAGS_SHARED))
1732 goto retry_private;
1733 goto retry;
1734 }
1735
1736 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1737 if (match_futex (&this->key, &key1)) {
1738 if (this->pi_state || this->rt_waiter) {
1739 ret = -EINVAL;
1740 goto out_unlock;
1741 }
1742 mark_wake_futex(&wake_q, this);
1743 if (++ret >= nr_wake)
1744 break;
1745 }
1746 }
1747
1748 if (op_ret > 0) {
1749 op_ret = 0;
1750 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1751 if (match_futex (&this->key, &key2)) {
1752 if (this->pi_state || this->rt_waiter) {
1753 ret = -EINVAL;
1754 goto out_unlock;
1755 }
1756 mark_wake_futex(&wake_q, this);
1757 if (++op_ret >= nr_wake2)
1758 break;
1759 }
1760 }
1761 ret += op_ret;
1762 }
1763
1764out_unlock:
1765 double_unlock_hb(hb1, hb2);
1766 wake_up_q(&wake_q);
1767 return ret;
1768}
1769
1770/**
1771 * requeue_futex() - Requeue a futex_q from one hb to another
1772 * @q: the futex_q to requeue
1773 * @hb1: the source hash_bucket
1774 * @hb2: the target hash_bucket
1775 * @key2: the new key for the requeued futex_q
1776 */
1777static inline
1778void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1779 struct futex_hash_bucket *hb2, union futex_key *key2)
1780{
1781
1782 /*
1783 * If key1 and key2 hash to the same bucket, no need to
1784 * requeue.
1785 */
1786 if (likely(&hb1->chain != &hb2->chain)) {
1787 plist_del(&q->list, &hb1->chain);
1788 hb_waiters_dec(hb1);
1789 hb_waiters_inc(hb2);
1790 plist_add(&q->list, &hb2->chain);
1791 q->lock_ptr = &hb2->lock;
1792 }
1793 q->key = *key2;
1794}
1795
1796/**
1797 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1798 * @q: the futex_q
1799 * @key: the key of the requeue target futex
1800 * @hb: the hash_bucket of the requeue target futex
1801 *
1802 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1803 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1804 * to the requeue target futex so the waiter can detect the wakeup on the right
1805 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1806 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1807 * to protect access to the pi_state to fixup the owner later. Must be called
1808 * with both q->lock_ptr and hb->lock held.
1809 */
1810static inline
1811void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1812 struct futex_hash_bucket *hb)
1813{
1814 q->key = *key;
1815
1816 __unqueue_futex(q);
1817
1818 WARN_ON(!q->rt_waiter);
1819 q->rt_waiter = NULL;
1820
1821 q->lock_ptr = &hb->lock;
1822
1823 wake_up_state(q->task, TASK_NORMAL);
1824}
1825
1826/**
1827 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1828 * @pifutex: the user address of the to futex
1829 * @hb1: the from futex hash bucket, must be locked by the caller
1830 * @hb2: the to futex hash bucket, must be locked by the caller
1831 * @key1: the from futex key
1832 * @key2: the to futex key
1833 * @ps: address to store the pi_state pointer
1834 * @exiting: Pointer to store the task pointer of the owner task
1835 * which is in the middle of exiting
1836 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1837 *
1838 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1839 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1840 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1841 * hb1 and hb2 must be held by the caller.
1842 *
1843 * @exiting is only set when the return value is -EBUSY. If so, this holds
1844 * a refcount on the exiting task on return and the caller needs to drop it
1845 * after waiting for the exit to complete.
1846 *
1847 * Return:
1848 * - 0 - failed to acquire the lock atomically;
1849 * - >0 - acquired the lock, return value is vpid of the top_waiter
1850 * - <0 - error
1851 */
1852static int
1853futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1854 struct futex_hash_bucket *hb2, union futex_key *key1,
1855 union futex_key *key2, struct futex_pi_state **ps,
1856 struct task_struct **exiting, int set_waiters)
1857{
1858 struct futex_q *top_waiter = NULL;
1859 u32 curval;
1860 int ret, vpid;
1861
1862 if (get_futex_value_locked(&curval, pifutex))
1863 return -EFAULT;
1864
1865 if (unlikely(should_fail_futex(true)))
1866 return -EFAULT;
1867
1868 /*
1869 * Find the top_waiter and determine if there are additional waiters.
1870 * If the caller intends to requeue more than 1 waiter to pifutex,
1871 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1872 * as we have means to handle the possible fault. If not, don't set
1873 * the bit unnecessarily as it will force the subsequent unlock to enter
1874 * the kernel.
1875 */
1876 top_waiter = futex_top_waiter(hb1, key1);
1877
1878 /* There are no waiters, nothing for us to do. */
1879 if (!top_waiter)
1880 return 0;
1881
1882 /* Ensure we requeue to the expected futex. */
1883 if (!match_futex(top_waiter->requeue_pi_key, key2))
1884 return -EINVAL;
1885
1886 /*
1887 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1888 * the contended case or if set_waiters is 1. The pi_state is returned
1889 * in ps in contended cases.
1890 */
1891 vpid = task_pid_vnr(top_waiter->task);
1892 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1893 exiting, set_waiters);
1894 if (ret == 1) {
1895 requeue_pi_wake_futex(top_waiter, key2, hb2);
1896 return vpid;
1897 }
1898 return ret;
1899}
1900
1901/**
1902 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1903 * @uaddr1: source futex user address
1904 * @flags: futex flags (FLAGS_SHARED, etc.)
1905 * @uaddr2: target futex user address
1906 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1907 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1908 * @cmpval: @uaddr1 expected value (or %NULL)
1909 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1910 * pi futex (pi to pi requeue is not supported)
1911 *
1912 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1913 * uaddr2 atomically on behalf of the top waiter.
1914 *
1915 * Return:
1916 * - >=0 - on success, the number of tasks requeued or woken;
1917 * - <0 - on error
1918 */
1919static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1920 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1921 u32 *cmpval, int requeue_pi)
1922{
1923 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1924 int task_count = 0, ret;
1925 struct futex_pi_state *pi_state = NULL;
1926 struct futex_hash_bucket *hb1, *hb2;
1927 struct futex_q *this, *next;
1928 DEFINE_WAKE_Q(wake_q);
1929
1930 if (nr_wake < 0 || nr_requeue < 0)
1931 return -EINVAL;
1932
1933 /*
1934 * When PI not supported: return -ENOSYS if requeue_pi is true,
1935 * consequently the compiler knows requeue_pi is always false past
1936 * this point which will optimize away all the conditional code
1937 * further down.
1938 */
1939 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1940 return -ENOSYS;
1941
1942 if (requeue_pi) {
1943 /*
1944 * Requeue PI only works on two distinct uaddrs. This
1945 * check is only valid for private futexes. See below.
1946 */
1947 if (uaddr1 == uaddr2)
1948 return -EINVAL;
1949
1950 /*
1951 * requeue_pi requires a pi_state, try to allocate it now
1952 * without any locks in case it fails.
1953 */
1954 if (refill_pi_state_cache())
1955 return -ENOMEM;
1956 /*
1957 * requeue_pi must wake as many tasks as it can, up to nr_wake
1958 * + nr_requeue, since it acquires the rt_mutex prior to
1959 * returning to userspace, so as to not leave the rt_mutex with
1960 * waiters and no owner. However, second and third wake-ups
1961 * cannot be predicted as they involve race conditions with the
1962 * first wake and a fault while looking up the pi_state. Both
1963 * pthread_cond_signal() and pthread_cond_broadcast() should
1964 * use nr_wake=1.
1965 */
1966 if (nr_wake != 1)
1967 return -EINVAL;
1968 }
1969
1970retry:
1971 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1972 if (unlikely(ret != 0))
1973 return ret;
1974 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1975 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1976 if (unlikely(ret != 0))
1977 return ret;
1978
1979 /*
1980 * The check above which compares uaddrs is not sufficient for
1981 * shared futexes. We need to compare the keys:
1982 */
1983 if (requeue_pi && match_futex(&key1, &key2))
1984 return -EINVAL;
1985
1986 hb1 = hash_futex(&key1);
1987 hb2 = hash_futex(&key2);
1988
1989retry_private:
1990 hb_waiters_inc(hb2);
1991 double_lock_hb(hb1, hb2);
1992
1993 if (likely(cmpval != NULL)) {
1994 u32 curval;
1995
1996 ret = get_futex_value_locked(&curval, uaddr1);
1997
1998 if (unlikely(ret)) {
1999 double_unlock_hb(hb1, hb2);
2000 hb_waiters_dec(hb2);
2001
2002 ret = get_user(curval, uaddr1);
2003 if (ret)
2004 return ret;
2005
2006 if (!(flags & FLAGS_SHARED))
2007 goto retry_private;
2008
2009 goto retry;
2010 }
2011 if (curval != *cmpval) {
2012 ret = -EAGAIN;
2013 goto out_unlock;
2014 }
2015 }
2016
2017 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2018 struct task_struct *exiting = NULL;
2019
2020 /*
2021 * Attempt to acquire uaddr2 and wake the top waiter. If we
2022 * intend to requeue waiters, force setting the FUTEX_WAITERS
2023 * bit. We force this here where we are able to easily handle
2024 * faults rather in the requeue loop below.
2025 */
2026 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2027 &key2, &pi_state,
2028 &exiting, nr_requeue);
2029
2030 /*
2031 * At this point the top_waiter has either taken uaddr2 or is
2032 * waiting on it. If the former, then the pi_state will not
2033 * exist yet, look it up one more time to ensure we have a
2034 * reference to it. If the lock was taken, ret contains the
2035 * vpid of the top waiter task.
2036 * If the lock was not taken, we have pi_state and an initial
2037 * refcount on it. In case of an error we have nothing.
2038 */
2039 if (ret > 0) {
2040 WARN_ON(pi_state);
2041 task_count++;
2042 /*
2043 * If we acquired the lock, then the user space value
2044 * of uaddr2 should be vpid. It cannot be changed by
2045 * the top waiter as it is blocked on hb2 lock if it
2046 * tries to do so. If something fiddled with it behind
2047 * our back the pi state lookup might unearth it. So
2048 * we rather use the known value than rereading and
2049 * handing potential crap to lookup_pi_state.
2050 *
2051 * If that call succeeds then we have pi_state and an
2052 * initial refcount on it.
2053 */
2054 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2055 &pi_state, &exiting);
2056 }
2057
2058 switch (ret) {
2059 case 0:
2060 /* We hold a reference on the pi state. */
2061 break;
2062
2063 /* If the above failed, then pi_state is NULL */
2064 case -EFAULT:
2065 double_unlock_hb(hb1, hb2);
2066 hb_waiters_dec(hb2);
2067 ret = fault_in_user_writeable(uaddr2);
2068 if (!ret)
2069 goto retry;
2070 return ret;
2071 case -EBUSY:
2072 case -EAGAIN:
2073 /*
2074 * Two reasons for this:
2075 * - EBUSY: Owner is exiting and we just wait for the
2076 * exit to complete.
2077 * - EAGAIN: The user space value changed.
2078 */
2079 double_unlock_hb(hb1, hb2);
2080 hb_waiters_dec(hb2);
2081 /*
2082 * Handle the case where the owner is in the middle of
2083 * exiting. Wait for the exit to complete otherwise
2084 * this task might loop forever, aka. live lock.
2085 */
2086 wait_for_owner_exiting(ret, exiting);
2087 cond_resched();
2088 goto retry;
2089 default:
2090 goto out_unlock;
2091 }
2092 }
2093
2094 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2095 if (task_count - nr_wake >= nr_requeue)
2096 break;
2097
2098 if (!match_futex(&this->key, &key1))
2099 continue;
2100
2101 /*
2102 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2103 * be paired with each other and no other futex ops.
2104 *
2105 * We should never be requeueing a futex_q with a pi_state,
2106 * which is awaiting a futex_unlock_pi().
2107 */
2108 if ((requeue_pi && !this->rt_waiter) ||
2109 (!requeue_pi && this->rt_waiter) ||
2110 this->pi_state) {
2111 ret = -EINVAL;
2112 break;
2113 }
2114
2115 /*
2116 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2117 * lock, we already woke the top_waiter. If not, it will be
2118 * woken by futex_unlock_pi().
2119 */
2120 if (++task_count <= nr_wake && !requeue_pi) {
2121 mark_wake_futex(&wake_q, this);
2122 continue;
2123 }
2124
2125 /* Ensure we requeue to the expected futex for requeue_pi. */
2126 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2127 ret = -EINVAL;
2128 break;
2129 }
2130
2131 /*
2132 * Requeue nr_requeue waiters and possibly one more in the case
2133 * of requeue_pi if we couldn't acquire the lock atomically.
2134 */
2135 if (requeue_pi) {
2136 /*
2137 * Prepare the waiter to take the rt_mutex. Take a
2138 * refcount on the pi_state and store the pointer in
2139 * the futex_q object of the waiter.
2140 */
2141 get_pi_state(pi_state);
2142 this->pi_state = pi_state;
2143 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2144 this->rt_waiter,
2145 this->task);
2146 if (ret == 1) {
2147 /*
2148 * We got the lock. We do neither drop the
2149 * refcount on pi_state nor clear
2150 * this->pi_state because the waiter needs the
2151 * pi_state for cleaning up the user space
2152 * value. It will drop the refcount after
2153 * doing so.
2154 */
2155 requeue_pi_wake_futex(this, &key2, hb2);
2156 continue;
2157 } else if (ret) {
2158 /*
2159 * rt_mutex_start_proxy_lock() detected a
2160 * potential deadlock when we tried to queue
2161 * that waiter. Drop the pi_state reference
2162 * which we took above and remove the pointer
2163 * to the state from the waiters futex_q
2164 * object.
2165 */
2166 this->pi_state = NULL;
2167 put_pi_state(pi_state);
2168 /*
2169 * We stop queueing more waiters and let user
2170 * space deal with the mess.
2171 */
2172 break;
2173 }
2174 }
2175 requeue_futex(this, hb1, hb2, &key2);
2176 }
2177
2178 /*
2179 * We took an extra initial reference to the pi_state either
2180 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2181 * need to drop it here again.
2182 */
2183 put_pi_state(pi_state);
2184
2185out_unlock:
2186 double_unlock_hb(hb1, hb2);
2187 wake_up_q(&wake_q);
2188 hb_waiters_dec(hb2);
2189 return ret ? ret : task_count;
2190}
2191
2192/* The key must be already stored in q->key. */
2193static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2194 __acquires(&hb->lock)
2195{
2196 struct futex_hash_bucket *hb;
2197
2198 hb = hash_futex(&q->key);
2199
2200 /*
2201 * Increment the counter before taking the lock so that
2202 * a potential waker won't miss a to-be-slept task that is
2203 * waiting for the spinlock. This is safe as all queue_lock()
2204 * users end up calling queue_me(). Similarly, for housekeeping,
2205 * decrement the counter at queue_unlock() when some error has
2206 * occurred and we don't end up adding the task to the list.
2207 */
2208 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2209
2210 q->lock_ptr = &hb->lock;
2211
2212 spin_lock(&hb->lock);
2213 return hb;
2214}
2215
2216static inline void
2217queue_unlock(struct futex_hash_bucket *hb)
2218 __releases(&hb->lock)
2219{
2220 spin_unlock(&hb->lock);
2221 hb_waiters_dec(hb);
2222}
2223
2224static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2225{
2226 int prio;
2227
2228 /*
2229 * The priority used to register this element is
2230 * - either the real thread-priority for the real-time threads
2231 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2232 * - or MAX_RT_PRIO for non-RT threads.
2233 * Thus, all RT-threads are woken first in priority order, and
2234 * the others are woken last, in FIFO order.
2235 */
2236 prio = min(current->normal_prio, MAX_RT_PRIO);
2237
2238 plist_node_init(&q->list, prio);
2239 plist_add(&q->list, &hb->chain);
2240 q->task = current;
2241}
2242
2243/**
2244 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2245 * @q: The futex_q to enqueue
2246 * @hb: The destination hash bucket
2247 *
2248 * The hb->lock must be held by the caller, and is released here. A call to
2249 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2250 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2251 * or nothing if the unqueue is done as part of the wake process and the unqueue
2252 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2253 * an example).
2254 */
2255static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2256 __releases(&hb->lock)
2257{
2258 __queue_me(q, hb);
2259 spin_unlock(&hb->lock);
2260}
2261
2262/**
2263 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2264 * @q: The futex_q to unqueue
2265 *
2266 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2267 * be paired with exactly one earlier call to queue_me().
2268 *
2269 * Return:
2270 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2271 * - 0 - if the futex_q was already removed by the waking thread
2272 */
2273static int unqueue_me(struct futex_q *q)
2274{
2275 spinlock_t *lock_ptr;
2276 int ret = 0;
2277
2278 /* In the common case we don't take the spinlock, which is nice. */
2279retry:
2280 /*
2281 * q->lock_ptr can change between this read and the following spin_lock.
2282 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2283 * optimizing lock_ptr out of the logic below.
2284 */
2285 lock_ptr = READ_ONCE(q->lock_ptr);
2286 if (lock_ptr != NULL) {
2287 spin_lock(lock_ptr);
2288 /*
2289 * q->lock_ptr can change between reading it and
2290 * spin_lock(), causing us to take the wrong lock. This
2291 * corrects the race condition.
2292 *
2293 * Reasoning goes like this: if we have the wrong lock,
2294 * q->lock_ptr must have changed (maybe several times)
2295 * between reading it and the spin_lock(). It can
2296 * change again after the spin_lock() but only if it was
2297 * already changed before the spin_lock(). It cannot,
2298 * however, change back to the original value. Therefore
2299 * we can detect whether we acquired the correct lock.
2300 */
2301 if (unlikely(lock_ptr != q->lock_ptr)) {
2302 spin_unlock(lock_ptr);
2303 goto retry;
2304 }
2305 __unqueue_futex(q);
2306
2307 BUG_ON(q->pi_state);
2308
2309 spin_unlock(lock_ptr);
2310 ret = 1;
2311 }
2312
2313 return ret;
2314}
2315
2316/*
2317 * PI futexes can not be requeued and must remove themselves from the
2318 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
2319 */
2320static void unqueue_me_pi(struct futex_q *q)
2321{
2322 __unqueue_futex(q);
2323
2324 BUG_ON(!q->pi_state);
2325 put_pi_state(q->pi_state);
2326 q->pi_state = NULL;
2327}
2328
2329static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2330 struct task_struct *argowner)
2331{
2332 struct futex_pi_state *pi_state = q->pi_state;
2333 struct task_struct *oldowner, *newowner;
2334 u32 uval, curval, newval, newtid;
2335 int err = 0;
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 return 0;
2370 }
2371
2372 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2373 /* We got the lock. pi_state is correct. Tell caller. */
2374 return 1;
2375 }
2376
2377 /*
2378 * The trylock just failed, so either there is an owner or
2379 * there is a higher priority waiter than this one.
2380 */
2381 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2382 /*
2383 * If the higher priority waiter has not yet taken over the
2384 * rtmutex then newowner is NULL. We can't return here with
2385 * that state because it's inconsistent vs. the user space
2386 * state. So drop the locks and try again. It's a valid
2387 * situation and not any different from the other retry
2388 * conditions.
2389 */
2390 if (unlikely(!newowner)) {
2391 err = -EAGAIN;
2392 goto handle_err;
2393 }
2394 } else {
2395 WARN_ON_ONCE(argowner != current);
2396 if (oldowner == current) {
2397 /*
2398 * We raced against a concurrent self; things are
2399 * already fixed up. Nothing to do.
2400 */
2401 return 1;
2402 }
2403 newowner = argowner;
2404 }
2405
2406 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2407 /* Owner died? */
2408 if (!pi_state->owner)
2409 newtid |= FUTEX_OWNER_DIED;
2410
2411 err = get_futex_value_locked(&uval, uaddr);
2412 if (err)
2413 goto handle_err;
2414
2415 for (;;) {
2416 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2417
2418 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2419 if (err)
2420 goto handle_err;
2421
2422 if (curval == uval)
2423 break;
2424 uval = curval;
2425 }
2426
2427 /*
2428 * We fixed up user space. Now we need to fix the pi_state
2429 * itself.
2430 */
2431 pi_state_update_owner(pi_state, newowner);
2432
2433 return argowner == current;
2434
2435 /*
2436 * In order to reschedule or handle a page fault, we need to drop the
2437 * locks here. In the case of a fault, this gives the other task
2438 * (either the highest priority waiter itself or the task which stole
2439 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2440 * are back from handling the fault we need to check the pi_state after
2441 * reacquiring the locks and before trying to do another fixup. When
2442 * the fixup has been done already we simply return.
2443 *
2444 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2445 * drop hb->lock since the caller owns the hb -> futex_q relation.
2446 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2447 */
2448handle_err:
2449 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2450 spin_unlock(q->lock_ptr);
2451
2452 switch (err) {
2453 case -EFAULT:
2454 err = fault_in_user_writeable(uaddr);
2455 break;
2456
2457 case -EAGAIN:
2458 cond_resched();
2459 err = 0;
2460 break;
2461
2462 default:
2463 WARN_ON_ONCE(1);
2464 break;
2465 }
2466
2467 spin_lock(q->lock_ptr);
2468 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2469
2470 /*
2471 * Check if someone else fixed it for us:
2472 */
2473 if (pi_state->owner != oldowner)
2474 return argowner == current;
2475
2476 /* Retry if err was -EAGAIN or the fault in succeeded */
2477 if (!err)
2478 goto retry;
2479
2480 /*
2481 * fault_in_user_writeable() failed so user state is immutable. At
2482 * best we can make the kernel state consistent but user state will
2483 * be most likely hosed and any subsequent unlock operation will be
2484 * rejected due to PI futex rule [10].
2485 *
2486 * Ensure that the rtmutex owner is also the pi_state owner despite
2487 * the user space value claiming something different. There is no
2488 * point in unlocking the rtmutex if current is the owner as it
2489 * would need to wait until the next waiter has taken the rtmutex
2490 * to guarantee consistent state. Keep it simple. Userspace asked
2491 * for this wreckaged state.
2492 *
2493 * The rtmutex has an owner - either current or some other
2494 * task. See the EAGAIN loop above.
2495 */
2496 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2497
2498 return err;
2499}
2500
2501static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2502 struct task_struct *argowner)
2503{
2504 struct futex_pi_state *pi_state = q->pi_state;
2505 int ret;
2506
2507 lockdep_assert_held(q->lock_ptr);
2508
2509 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2510 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2511 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2512 return ret;
2513}
2514
2515static long futex_wait_restart(struct restart_block *restart);
2516
2517/**
2518 * fixup_owner() - Post lock pi_state and corner case management
2519 * @uaddr: user address of the futex
2520 * @q: futex_q (contains pi_state and access to the rt_mutex)
2521 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2522 *
2523 * After attempting to lock an rt_mutex, this function is called to cleanup
2524 * the pi_state owner as well as handle race conditions that may allow us to
2525 * acquire the lock. Must be called with the hb lock held.
2526 *
2527 * Return:
2528 * - 1 - success, lock taken;
2529 * - 0 - success, lock not taken;
2530 * - <0 - on error (-EFAULT)
2531 */
2532static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2533{
2534 if (locked) {
2535 /*
2536 * Got the lock. We might not be the anticipated owner if we
2537 * did a lock-steal - fix up the PI-state in that case:
2538 *
2539 * Speculative pi_state->owner read (we don't hold wait_lock);
2540 * since we own the lock pi_state->owner == current is the
2541 * stable state, anything else needs more attention.
2542 */
2543 if (q->pi_state->owner != current)
2544 return fixup_pi_state_owner(uaddr, q, current);
2545 return 1;
2546 }
2547
2548 /*
2549 * If we didn't get the lock; check if anybody stole it from us. In
2550 * that case, we need to fix up the uval to point to them instead of
2551 * us, otherwise bad things happen. [10]
2552 *
2553 * Another speculative read; pi_state->owner == current is unstable
2554 * but needs our attention.
2555 */
2556 if (q->pi_state->owner == current)
2557 return fixup_pi_state_owner(uaddr, q, NULL);
2558
2559 /*
2560 * Paranoia check. If we did not take the lock, then we should not be
2561 * the owner of the rt_mutex. Warn and establish consistent state.
2562 */
2563 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2564 return fixup_pi_state_owner(uaddr, q, current);
2565
2566 return 0;
2567}
2568
2569/**
2570 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571 * @hb: the futex hash bucket, must be locked by the caller
2572 * @q: the futex_q to queue up on
2573 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2574 */
2575static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2576 struct hrtimer_sleeper *timeout)
2577{
2578 /*
2579 * The task state is guaranteed to be set before another task can
2580 * wake it. set_current_state() is implemented using smp_store_mb() and
2581 * queue_me() calls spin_unlock() upon completion, both serializing
2582 * access to the hash list and forcing another memory barrier.
2583 */
2584 set_current_state(TASK_INTERRUPTIBLE);
2585 queue_me(q, hb);
2586
2587 /* Arm the timer */
2588 if (timeout)
2589 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2590
2591 /*
2592 * If we have been removed from the hash list, then another task
2593 * has tried to wake us, and we can skip the call to schedule().
2594 */
2595 if (likely(!plist_node_empty(&q->list))) {
2596 /*
2597 * If the timer has already expired, current will already be
2598 * flagged for rescheduling. Only call schedule if there
2599 * is no timeout, or if it has yet to expire.
2600 */
2601 if (!timeout || timeout->task)
2602 freezable_schedule();
2603 }
2604 __set_current_state(TASK_RUNNING);
2605}
2606
2607/**
2608 * futex_wait_setup() - Prepare to wait on a futex
2609 * @uaddr: the futex userspace address
2610 * @val: the expected value
2611 * @flags: futex flags (FLAGS_SHARED, etc.)
2612 * @q: the associated futex_q
2613 * @hb: storage for hash_bucket pointer to be returned to caller
2614 *
2615 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2616 * compare it with the expected value. Handle atomic faults internally.
2617 * Return with the hb lock held and a q.key reference on success, and unlocked
2618 * with no q.key reference on failure.
2619 *
2620 * Return:
2621 * - 0 - uaddr contains val and hb has been locked;
2622 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2623 */
2624static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2625 struct futex_q *q, struct futex_hash_bucket **hb)
2626{
2627 u32 uval;
2628 int ret;
2629
2630 /*
2631 * Access the page AFTER the hash-bucket is locked.
2632 * Order is important:
2633 *
2634 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2635 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2636 *
2637 * The basic logical guarantee of a futex is that it blocks ONLY
2638 * if cond(var) is known to be true at the time of blocking, for
2639 * any cond. If we locked the hash-bucket after testing *uaddr, that
2640 * would open a race condition where we could block indefinitely with
2641 * cond(var) false, which would violate the guarantee.
2642 *
2643 * On the other hand, we insert q and release the hash-bucket only
2644 * after testing *uaddr. This guarantees that futex_wait() will NOT
2645 * absorb a wakeup if *uaddr does not match the desired values
2646 * while the syscall executes.
2647 */
2648retry:
2649 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2650 if (unlikely(ret != 0))
2651 return ret;
2652
2653retry_private:
2654 *hb = queue_lock(q);
2655
2656 ret = get_futex_value_locked(&uval, uaddr);
2657
2658 if (ret) {
2659 queue_unlock(*hb);
2660
2661 ret = get_user(uval, uaddr);
2662 if (ret)
2663 return ret;
2664
2665 if (!(flags & FLAGS_SHARED))
2666 goto retry_private;
2667
2668 goto retry;
2669 }
2670
2671 if (uval != val) {
2672 queue_unlock(*hb);
2673 ret = -EWOULDBLOCK;
2674 }
2675
2676 return ret;
2677}
2678
2679static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2680 ktime_t *abs_time, u32 bitset)
2681{
2682 struct hrtimer_sleeper timeout, *to;
2683 struct restart_block *restart;
2684 struct futex_hash_bucket *hb;
2685 struct futex_q q = futex_q_init;
2686 int ret;
2687
2688 if (!bitset)
2689 return -EINVAL;
2690 q.bitset = bitset;
2691
2692 to = futex_setup_timer(abs_time, &timeout, flags,
2693 current->timer_slack_ns);
2694retry:
2695 /*
2696 * Prepare to wait on uaddr. On success, holds hb lock and increments
2697 * q.key refs.
2698 */
2699 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2700 if (ret)
2701 goto out;
2702
2703 /* queue_me and wait for wakeup, timeout, or a signal. */
2704 futex_wait_queue_me(hb, &q, to);
2705
2706 /* If we were woken (and unqueued), we succeeded, whatever. */
2707 ret = 0;
2708 /* unqueue_me() drops q.key ref */
2709 if (!unqueue_me(&q))
2710 goto out;
2711 ret = -ETIMEDOUT;
2712 if (to && !to->task)
2713 goto out;
2714
2715 /*
2716 * We expect signal_pending(current), but we might be the
2717 * victim of a spurious wakeup as well.
2718 */
2719 if (!signal_pending(current))
2720 goto retry;
2721
2722 ret = -ERESTARTSYS;
2723 if (!abs_time)
2724 goto out;
2725
2726 restart = ¤t->restart_block;
2727 restart->futex.uaddr = uaddr;
2728 restart->futex.val = val;
2729 restart->futex.time = *abs_time;
2730 restart->futex.bitset = bitset;
2731 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2732
2733 ret = set_restart_fn(restart, futex_wait_restart);
2734
2735out:
2736 if (to) {
2737 hrtimer_cancel(&to->timer);
2738 destroy_hrtimer_on_stack(&to->timer);
2739 }
2740 return ret;
2741}
2742
2743
2744static long futex_wait_restart(struct restart_block *restart)
2745{
2746 u32 __user *uaddr = restart->futex.uaddr;
2747 ktime_t t, *tp = NULL;
2748
2749 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2750 t = restart->futex.time;
2751 tp = &t;
2752 }
2753 restart->fn = do_no_restart_syscall;
2754
2755 return (long)futex_wait(uaddr, restart->futex.flags,
2756 restart->futex.val, tp, restart->futex.bitset);
2757}
2758
2759
2760/*
2761 * Userspace tried a 0 -> TID atomic transition of the futex value
2762 * and failed. The kernel side here does the whole locking operation:
2763 * if there are waiters then it will block as a consequence of relying
2764 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2765 * a 0 value of the futex too.).
2766 *
2767 * Also serves as futex trylock_pi()'ing, and due semantics.
2768 */
2769static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2770 ktime_t *time, int trylock)
2771{
2772 struct hrtimer_sleeper timeout, *to;
2773 struct task_struct *exiting = NULL;
2774 struct rt_mutex_waiter rt_waiter;
2775 struct futex_hash_bucket *hb;
2776 struct futex_q q = futex_q_init;
2777 int res, ret;
2778
2779 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2780 return -ENOSYS;
2781
2782 if (refill_pi_state_cache())
2783 return -ENOMEM;
2784
2785 to = futex_setup_timer(time, &timeout, flags, 0);
2786
2787retry:
2788 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2789 if (unlikely(ret != 0))
2790 goto out;
2791
2792retry_private:
2793 hb = queue_lock(&q);
2794
2795 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2796 &exiting, 0);
2797 if (unlikely(ret)) {
2798 /*
2799 * Atomic work succeeded and we got the lock,
2800 * or failed. Either way, we do _not_ block.
2801 */
2802 switch (ret) {
2803 case 1:
2804 /* We got the lock. */
2805 ret = 0;
2806 goto out_unlock_put_key;
2807 case -EFAULT:
2808 goto uaddr_faulted;
2809 case -EBUSY:
2810 case -EAGAIN:
2811 /*
2812 * Two reasons for this:
2813 * - EBUSY: Task is exiting and we just wait for the
2814 * exit to complete.
2815 * - EAGAIN: The user space value changed.
2816 */
2817 queue_unlock(hb);
2818 /*
2819 * Handle the case where the owner is in the middle of
2820 * exiting. Wait for the exit to complete otherwise
2821 * this task might loop forever, aka. live lock.
2822 */
2823 wait_for_owner_exiting(ret, exiting);
2824 cond_resched();
2825 goto retry;
2826 default:
2827 goto out_unlock_put_key;
2828 }
2829 }
2830
2831 WARN_ON(!q.pi_state);
2832
2833 /*
2834 * Only actually queue now that the atomic ops are done:
2835 */
2836 __queue_me(&q, hb);
2837
2838 if (trylock) {
2839 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2840 /* Fixup the trylock return value: */
2841 ret = ret ? 0 : -EWOULDBLOCK;
2842 goto no_block;
2843 }
2844
2845 rt_mutex_init_waiter(&rt_waiter);
2846
2847 /*
2848 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2849 * hold it while doing rt_mutex_start_proxy(), because then it will
2850 * include hb->lock in the blocking chain, even through we'll not in
2851 * fact hold it while blocking. This will lead it to report -EDEADLK
2852 * and BUG when futex_unlock_pi() interleaves with this.
2853 *
2854 * Therefore acquire wait_lock while holding hb->lock, but drop the
2855 * latter before calling __rt_mutex_start_proxy_lock(). This
2856 * interleaves with futex_unlock_pi() -- which does a similar lock
2857 * handoff -- such that the latter can observe the futex_q::pi_state
2858 * before __rt_mutex_start_proxy_lock() is done.
2859 */
2860 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2861 spin_unlock(q.lock_ptr);
2862 /*
2863 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2864 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2865 * it sees the futex_q::pi_state.
2866 */
2867 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2868 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2869
2870 if (ret) {
2871 if (ret == 1)
2872 ret = 0;
2873 goto cleanup;
2874 }
2875
2876 if (unlikely(to))
2877 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2878
2879 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2880
2881cleanup:
2882 spin_lock(q.lock_ptr);
2883 /*
2884 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2885 * first acquire the hb->lock before removing the lock from the
2886 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2887 * lists consistent.
2888 *
2889 * In particular; it is important that futex_unlock_pi() can not
2890 * observe this inconsistency.
2891 */
2892 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2893 ret = 0;
2894
2895no_block:
2896 /*
2897 * Fixup the pi_state owner and possibly acquire the lock if we
2898 * haven't already.
2899 */
2900 res = fixup_owner(uaddr, &q, !ret);
2901 /*
2902 * If fixup_owner() returned an error, propagate that. If it acquired
2903 * the lock, clear our -ETIMEDOUT or -EINTR.
2904 */
2905 if (res)
2906 ret = (res < 0) ? res : 0;
2907
2908 unqueue_me_pi(&q);
2909 spin_unlock(q.lock_ptr);
2910 goto out;
2911
2912out_unlock_put_key:
2913 queue_unlock(hb);
2914
2915out:
2916 if (to) {
2917 hrtimer_cancel(&to->timer);
2918 destroy_hrtimer_on_stack(&to->timer);
2919 }
2920 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2921
2922uaddr_faulted:
2923 queue_unlock(hb);
2924
2925 ret = fault_in_user_writeable(uaddr);
2926 if (ret)
2927 goto out;
2928
2929 if (!(flags & FLAGS_SHARED))
2930 goto retry_private;
2931
2932 goto retry;
2933}
2934
2935/*
2936 * Userspace attempted a TID -> 0 atomic transition, and failed.
2937 * This is the in-kernel slowpath: we look up the PI state (if any),
2938 * and do the rt-mutex unlock.
2939 */
2940static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2941{
2942 u32 curval, uval, vpid = task_pid_vnr(current);
2943 union futex_key key = FUTEX_KEY_INIT;
2944 struct futex_hash_bucket *hb;
2945 struct futex_q *top_waiter;
2946 int ret;
2947
2948 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2949 return -ENOSYS;
2950
2951retry:
2952 if (get_user(uval, uaddr))
2953 return -EFAULT;
2954 /*
2955 * We release only a lock we actually own:
2956 */
2957 if ((uval & FUTEX_TID_MASK) != vpid)
2958 return -EPERM;
2959
2960 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2961 if (ret)
2962 return ret;
2963
2964 hb = hash_futex(&key);
2965 spin_lock(&hb->lock);
2966
2967 /*
2968 * Check waiters first. We do not trust user space values at
2969 * all and we at least want to know if user space fiddled
2970 * with the futex value instead of blindly unlocking.
2971 */
2972 top_waiter = futex_top_waiter(hb, &key);
2973 if (top_waiter) {
2974 struct futex_pi_state *pi_state = top_waiter->pi_state;
2975
2976 ret = -EINVAL;
2977 if (!pi_state)
2978 goto out_unlock;
2979
2980 /*
2981 * If current does not own the pi_state then the futex is
2982 * inconsistent and user space fiddled with the futex value.
2983 */
2984 if (pi_state->owner != current)
2985 goto out_unlock;
2986
2987 get_pi_state(pi_state);
2988 /*
2989 * By taking wait_lock while still holding hb->lock, we ensure
2990 * there is no point where we hold neither; and therefore
2991 * wake_futex_pi() must observe a state consistent with what we
2992 * observed.
2993 *
2994 * In particular; this forces __rt_mutex_start_proxy() to
2995 * complete such that we're guaranteed to observe the
2996 * rt_waiter. Also see the WARN in wake_futex_pi().
2997 */
2998 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2999 spin_unlock(&hb->lock);
3000
3001 /* drops pi_state->pi_mutex.wait_lock */
3002 ret = wake_futex_pi(uaddr, uval, pi_state);
3003
3004 put_pi_state(pi_state);
3005
3006 /*
3007 * Success, we're done! No tricky corner cases.
3008 */
3009 if (!ret)
3010 return ret;
3011 /*
3012 * The atomic access to the futex value generated a
3013 * pagefault, so retry the user-access and the wakeup:
3014 */
3015 if (ret == -EFAULT)
3016 goto pi_faulted;
3017 /*
3018 * A unconditional UNLOCK_PI op raced against a waiter
3019 * setting the FUTEX_WAITERS bit. Try again.
3020 */
3021 if (ret == -EAGAIN)
3022 goto pi_retry;
3023 /*
3024 * wake_futex_pi has detected invalid state. Tell user
3025 * space.
3026 */
3027 return ret;
3028 }
3029
3030 /*
3031 * We have no kernel internal state, i.e. no waiters in the
3032 * kernel. Waiters which are about to queue themselves are stuck
3033 * on hb->lock. So we can safely ignore them. We do neither
3034 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3035 * owner.
3036 */
3037 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3038 spin_unlock(&hb->lock);
3039 switch (ret) {
3040 case -EFAULT:
3041 goto pi_faulted;
3042
3043 case -EAGAIN:
3044 goto pi_retry;
3045
3046 default:
3047 WARN_ON_ONCE(1);
3048 return ret;
3049 }
3050 }
3051
3052 /*
3053 * If uval has changed, let user space handle it.
3054 */
3055 ret = (curval == uval) ? 0 : -EAGAIN;
3056
3057out_unlock:
3058 spin_unlock(&hb->lock);
3059 return ret;
3060
3061pi_retry:
3062 cond_resched();
3063 goto retry;
3064
3065pi_faulted:
3066
3067 ret = fault_in_user_writeable(uaddr);
3068 if (!ret)
3069 goto retry;
3070
3071 return ret;
3072}
3073
3074/**
3075 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3076 * @hb: the hash_bucket futex_q was original enqueued on
3077 * @q: the futex_q woken while waiting to be requeued
3078 * @key2: the futex_key of the requeue target futex
3079 * @timeout: the timeout associated with the wait (NULL if none)
3080 *
3081 * Detect if the task was woken on the initial futex as opposed to the requeue
3082 * target futex. If so, determine if it was a timeout or a signal that caused
3083 * the wakeup and return the appropriate error code to the caller. Must be
3084 * called with the hb lock held.
3085 *
3086 * Return:
3087 * - 0 = no early wakeup detected;
3088 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3089 */
3090static inline
3091int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3092 struct futex_q *q, union futex_key *key2,
3093 struct hrtimer_sleeper *timeout)
3094{
3095 int ret = 0;
3096
3097 /*
3098 * With the hb lock held, we avoid races while we process the wakeup.
3099 * We only need to hold hb (and not hb2) to ensure atomicity as the
3100 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3101 * It can't be requeued from uaddr2 to something else since we don't
3102 * support a PI aware source futex for requeue.
3103 */
3104 if (!match_futex(&q->key, key2)) {
3105 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3106 /*
3107 * We were woken prior to requeue by a timeout or a signal.
3108 * Unqueue the futex_q and determine which it was.
3109 */
3110 plist_del(&q->list, &hb->chain);
3111 hb_waiters_dec(hb);
3112
3113 /* Handle spurious wakeups gracefully */
3114 ret = -EWOULDBLOCK;
3115 if (timeout && !timeout->task)
3116 ret = -ETIMEDOUT;
3117 else if (signal_pending(current))
3118 ret = -ERESTARTNOINTR;
3119 }
3120 return ret;
3121}
3122
3123/**
3124 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3125 * @uaddr: the futex we initially wait on (non-pi)
3126 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3127 * the same type, no requeueing from private to shared, etc.
3128 * @val: the expected value of uaddr
3129 * @abs_time: absolute timeout
3130 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3131 * @uaddr2: the pi futex we will take prior to returning to user-space
3132 *
3133 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3134 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3135 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3136 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3137 * without one, the pi logic would not know which task to boost/deboost, if
3138 * there was a need to.
3139 *
3140 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3141 * via the following--
3142 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3143 * 2) wakeup on uaddr2 after a requeue
3144 * 3) signal
3145 * 4) timeout
3146 *
3147 * If 3, cleanup and return -ERESTARTNOINTR.
3148 *
3149 * If 2, we may then block on trying to take the rt_mutex and return via:
3150 * 5) successful lock
3151 * 6) signal
3152 * 7) timeout
3153 * 8) other lock acquisition failure
3154 *
3155 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3156 *
3157 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3158 *
3159 * Return:
3160 * - 0 - On success;
3161 * - <0 - On error
3162 */
3163static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3164 u32 val, ktime_t *abs_time, u32 bitset,
3165 u32 __user *uaddr2)
3166{
3167 struct hrtimer_sleeper timeout, *to;
3168 struct rt_mutex_waiter rt_waiter;
3169 struct futex_hash_bucket *hb;
3170 union futex_key key2 = FUTEX_KEY_INIT;
3171 struct futex_q q = futex_q_init;
3172 int res, ret;
3173
3174 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3175 return -ENOSYS;
3176
3177 if (uaddr == uaddr2)
3178 return -EINVAL;
3179
3180 if (!bitset)
3181 return -EINVAL;
3182
3183 to = futex_setup_timer(abs_time, &timeout, flags,
3184 current->timer_slack_ns);
3185
3186 /*
3187 * The waiter is allocated on our stack, manipulated by the requeue
3188 * code while we sleep on uaddr.
3189 */
3190 rt_mutex_init_waiter(&rt_waiter);
3191
3192 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3193 if (unlikely(ret != 0))
3194 goto out;
3195
3196 q.bitset = bitset;
3197 q.rt_waiter = &rt_waiter;
3198 q.requeue_pi_key = &key2;
3199
3200 /*
3201 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3202 * count.
3203 */
3204 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3205 if (ret)
3206 goto out;
3207
3208 /*
3209 * The check above which compares uaddrs is not sufficient for
3210 * shared futexes. We need to compare the keys:
3211 */
3212 if (match_futex(&q.key, &key2)) {
3213 queue_unlock(hb);
3214 ret = -EINVAL;
3215 goto out;
3216 }
3217
3218 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3219 futex_wait_queue_me(hb, &q, to);
3220
3221 spin_lock(&hb->lock);
3222 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3223 spin_unlock(&hb->lock);
3224 if (ret)
3225 goto out;
3226
3227 /*
3228 * In order for us to be here, we know our q.key == key2, and since
3229 * we took the hb->lock above, we also know that futex_requeue() has
3230 * completed and we no longer have to concern ourselves with a wakeup
3231 * race with the atomic proxy lock acquisition by the requeue code. The
3232 * futex_requeue dropped our key1 reference and incremented our key2
3233 * reference count.
3234 */
3235
3236 /*
3237 * Check if the requeue code acquired the second futex for us and do
3238 * any pertinent fixup.
3239 */
3240 if (!q.rt_waiter) {
3241 if (q.pi_state && (q.pi_state->owner != current)) {
3242 spin_lock(q.lock_ptr);
3243 ret = fixup_owner(uaddr2, &q, true);
3244 /*
3245 * Drop the reference to the pi state which
3246 * the requeue_pi() code acquired for us.
3247 */
3248 put_pi_state(q.pi_state);
3249 spin_unlock(q.lock_ptr);
3250 /*
3251 * Adjust the return value. It's either -EFAULT or
3252 * success (1) but the caller expects 0 for success.
3253 */
3254 ret = ret < 0 ? ret : 0;
3255 }
3256 } else {
3257 struct rt_mutex *pi_mutex;
3258
3259 /*
3260 * We have been woken up by futex_unlock_pi(), a timeout, or a
3261 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3262 * the pi_state.
3263 */
3264 WARN_ON(!q.pi_state);
3265 pi_mutex = &q.pi_state->pi_mutex;
3266 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3267
3268 spin_lock(q.lock_ptr);
3269 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3270 ret = 0;
3271
3272 debug_rt_mutex_free_waiter(&rt_waiter);
3273 /*
3274 * Fixup the pi_state owner and possibly acquire the lock if we
3275 * haven't already.
3276 */
3277 res = fixup_owner(uaddr2, &q, !ret);
3278 /*
3279 * If fixup_owner() returned an error, propagate that. If it
3280 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3281 */
3282 if (res)
3283 ret = (res < 0) ? res : 0;
3284
3285 unqueue_me_pi(&q);
3286 spin_unlock(q.lock_ptr);
3287 }
3288
3289 if (ret == -EINTR) {
3290 /*
3291 * We've already been requeued, but cannot restart by calling
3292 * futex_lock_pi() directly. We could restart this syscall, but
3293 * it would detect that the user space "val" changed and return
3294 * -EWOULDBLOCK. Save the overhead of the restart and return
3295 * -EWOULDBLOCK directly.
3296 */
3297 ret = -EWOULDBLOCK;
3298 }
3299
3300out:
3301 if (to) {
3302 hrtimer_cancel(&to->timer);
3303 destroy_hrtimer_on_stack(&to->timer);
3304 }
3305 return ret;
3306}
3307
3308/*
3309 * Support for robust futexes: the kernel cleans up held futexes at
3310 * thread exit time.
3311 *
3312 * Implementation: user-space maintains a per-thread list of locks it
3313 * is holding. Upon do_exit(), the kernel carefully walks this list,
3314 * and marks all locks that are owned by this thread with the
3315 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3316 * always manipulated with the lock held, so the list is private and
3317 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3318 * field, to allow the kernel to clean up if the thread dies after
3319 * acquiring the lock, but just before it could have added itself to
3320 * the list. There can only be one such pending lock.
3321 */
3322
3323/**
3324 * sys_set_robust_list() - Set the robust-futex list head of a task
3325 * @head: pointer to the list-head
3326 * @len: length of the list-head, as userspace expects
3327 */
3328SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3329 size_t, len)
3330{
3331 if (!futex_cmpxchg_enabled)
3332 return -ENOSYS;
3333 /*
3334 * The kernel knows only one size for now:
3335 */
3336 if (unlikely(len != sizeof(*head)))
3337 return -EINVAL;
3338
3339 current->robust_list = head;
3340
3341 return 0;
3342}
3343
3344/**
3345 * sys_get_robust_list() - Get the robust-futex list head of a task
3346 * @pid: pid of the process [zero for current task]
3347 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3348 * @len_ptr: pointer to a length field, the kernel fills in the header size
3349 */
3350SYSCALL_DEFINE3(get_robust_list, int, pid,
3351 struct robust_list_head __user * __user *, head_ptr,
3352 size_t __user *, len_ptr)
3353{
3354 struct robust_list_head __user *head;
3355 unsigned long ret;
3356 struct task_struct *p;
3357
3358 if (!futex_cmpxchg_enabled)
3359 return -ENOSYS;
3360
3361 rcu_read_lock();
3362
3363 ret = -ESRCH;
3364 if (!pid)
3365 p = current;
3366 else {
3367 p = find_task_by_vpid(pid);
3368 if (!p)
3369 goto err_unlock;
3370 }
3371
3372 ret = -EPERM;
3373 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3374 goto err_unlock;
3375
3376 head = p->robust_list;
3377 rcu_read_unlock();
3378
3379 if (put_user(sizeof(*head), len_ptr))
3380 return -EFAULT;
3381 return put_user(head, head_ptr);
3382
3383err_unlock:
3384 rcu_read_unlock();
3385
3386 return ret;
3387}
3388
3389/* Constants for the pending_op argument of handle_futex_death */
3390#define HANDLE_DEATH_PENDING true
3391#define HANDLE_DEATH_LIST false
3392
3393/*
3394 * Process a futex-list entry, check whether it's owned by the
3395 * dying task, and do notification if so:
3396 */
3397static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3398 bool pi, bool pending_op)
3399{
3400 u32 uval, nval, mval;
3401 int err;
3402
3403 /* Futex address must be 32bit aligned */
3404 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3405 return -1;
3406
3407retry:
3408 if (get_user(uval, uaddr))
3409 return -1;
3410
3411 /*
3412 * Special case for regular (non PI) futexes. The unlock path in
3413 * user space has two race scenarios:
3414 *
3415 * 1. The unlock path releases the user space futex value and
3416 * before it can execute the futex() syscall to wake up
3417 * waiters it is killed.
3418 *
3419 * 2. A woken up waiter is killed before it can acquire the
3420 * futex in user space.
3421 *
3422 * In both cases the TID validation below prevents a wakeup of
3423 * potential waiters which can cause these waiters to block
3424 * forever.
3425 *
3426 * In both cases the following conditions are met:
3427 *
3428 * 1) task->robust_list->list_op_pending != NULL
3429 * @pending_op == true
3430 * 2) User space futex value == 0
3431 * 3) Regular futex: @pi == false
3432 *
3433 * If these conditions are met, it is safe to attempt waking up a
3434 * potential waiter without touching the user space futex value and
3435 * trying to set the OWNER_DIED bit. The user space futex value is
3436 * uncontended and the rest of the user space mutex state is
3437 * consistent, so a woken waiter will just take over the
3438 * uncontended futex. Setting the OWNER_DIED bit would create
3439 * inconsistent state and malfunction of the user space owner died
3440 * handling.
3441 */
3442 if (pending_op && !pi && !uval) {
3443 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3444 return 0;
3445 }
3446
3447 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3448 return 0;
3449
3450 /*
3451 * Ok, this dying thread is truly holding a futex
3452 * of interest. Set the OWNER_DIED bit atomically
3453 * via cmpxchg, and if the value had FUTEX_WAITERS
3454 * set, wake up a waiter (if any). (We have to do a
3455 * futex_wake() even if OWNER_DIED is already set -
3456 * to handle the rare but possible case of recursive
3457 * thread-death.) The rest of the cleanup is done in
3458 * userspace.
3459 */
3460 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3461
3462 /*
3463 * We are not holding a lock here, but we want to have
3464 * the pagefault_disable/enable() protection because
3465 * we want to handle the fault gracefully. If the
3466 * access fails we try to fault in the futex with R/W
3467 * verification via get_user_pages. get_user() above
3468 * does not guarantee R/W access. If that fails we
3469 * give up and leave the futex locked.
3470 */
3471 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3472 switch (err) {
3473 case -EFAULT:
3474 if (fault_in_user_writeable(uaddr))
3475 return -1;
3476 goto retry;
3477
3478 case -EAGAIN:
3479 cond_resched();
3480 goto retry;
3481
3482 default:
3483 WARN_ON_ONCE(1);
3484 return err;
3485 }
3486 }
3487
3488 if (nval != uval)
3489 goto retry;
3490
3491 /*
3492 * Wake robust non-PI futexes here. The wakeup of
3493 * PI futexes happens in exit_pi_state():
3494 */
3495 if (!pi && (uval & FUTEX_WAITERS))
3496 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3497
3498 return 0;
3499}
3500
3501/*
3502 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3503 */
3504static inline int fetch_robust_entry(struct robust_list __user **entry,
3505 struct robust_list __user * __user *head,
3506 unsigned int *pi)
3507{
3508 unsigned long uentry;
3509
3510 if (get_user(uentry, (unsigned long __user *)head))
3511 return -EFAULT;
3512
3513 *entry = (void __user *)(uentry & ~1UL);
3514 *pi = uentry & 1;
3515
3516 return 0;
3517}
3518
3519/*
3520 * Walk curr->robust_list (very carefully, it's a userspace list!)
3521 * and mark any locks found there dead, and notify any waiters.
3522 *
3523 * We silently return on any sign of list-walking problem.
3524 */
3525static void exit_robust_list(struct task_struct *curr)
3526{
3527 struct robust_list_head __user *head = curr->robust_list;
3528 struct robust_list __user *entry, *next_entry, *pending;
3529 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3530 unsigned int next_pi;
3531 unsigned long futex_offset;
3532 int rc;
3533
3534 if (!futex_cmpxchg_enabled)
3535 return;
3536
3537 /*
3538 * Fetch the list head (which was registered earlier, via
3539 * sys_set_robust_list()):
3540 */
3541 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3542 return;
3543 /*
3544 * Fetch the relative futex offset:
3545 */
3546 if (get_user(futex_offset, &head->futex_offset))
3547 return;
3548 /*
3549 * Fetch any possibly pending lock-add first, and handle it
3550 * if it exists:
3551 */
3552 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3553 return;
3554
3555 next_entry = NULL; /* avoid warning with gcc */
3556 while (entry != &head->list) {
3557 /*
3558 * Fetch the next entry in the list before calling
3559 * handle_futex_death:
3560 */
3561 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3562 /*
3563 * A pending lock might already be on the list, so
3564 * don't process it twice:
3565 */
3566 if (entry != pending) {
3567 if (handle_futex_death((void __user *)entry + futex_offset,
3568 curr, pi, HANDLE_DEATH_LIST))
3569 return;
3570 }
3571 if (rc)
3572 return;
3573 entry = next_entry;
3574 pi = next_pi;
3575 /*
3576 * Avoid excessively long or circular lists:
3577 */
3578 if (!--limit)
3579 break;
3580
3581 cond_resched();
3582 }
3583
3584 if (pending) {
3585 handle_futex_death((void __user *)pending + futex_offset,
3586 curr, pip, HANDLE_DEATH_PENDING);
3587 }
3588}
3589
3590static void futex_cleanup(struct task_struct *tsk)
3591{
3592 if (unlikely(tsk->robust_list)) {
3593 exit_robust_list(tsk);
3594 tsk->robust_list = NULL;
3595 }
3596
3597#ifdef CONFIG_COMPAT
3598 if (unlikely(tsk->compat_robust_list)) {
3599 compat_exit_robust_list(tsk);
3600 tsk->compat_robust_list = NULL;
3601 }
3602#endif
3603
3604 if (unlikely(!list_empty(&tsk->pi_state_list)))
3605 exit_pi_state_list(tsk);
3606}
3607
3608/**
3609 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3610 * @tsk: task to set the state on
3611 *
3612 * Set the futex exit state of the task lockless. The futex waiter code
3613 * observes that state when a task is exiting and loops until the task has
3614 * actually finished the futex cleanup. The worst case for this is that the
3615 * waiter runs through the wait loop until the state becomes visible.
3616 *
3617 * This is called from the recursive fault handling path in do_exit().
3618 *
3619 * This is best effort. Either the futex exit code has run already or
3620 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3621 * take it over. If not, the problem is pushed back to user space. If the
3622 * futex exit code did not run yet, then an already queued waiter might
3623 * block forever, but there is nothing which can be done about that.
3624 */
3625void futex_exit_recursive(struct task_struct *tsk)
3626{
3627 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3628 if (tsk->futex_state == FUTEX_STATE_EXITING)
3629 mutex_unlock(&tsk->futex_exit_mutex);
3630 tsk->futex_state = FUTEX_STATE_DEAD;
3631}
3632
3633static void futex_cleanup_begin(struct task_struct *tsk)
3634{
3635 /*
3636 * Prevent various race issues against a concurrent incoming waiter
3637 * including live locks by forcing the waiter to block on
3638 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3639 * attach_to_pi_owner().
3640 */
3641 mutex_lock(&tsk->futex_exit_mutex);
3642
3643 /*
3644 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3645 *
3646 * This ensures that all subsequent checks of tsk->futex_state in
3647 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3648 * tsk->pi_lock held.
3649 *
3650 * It guarantees also that a pi_state which was queued right before
3651 * the state change under tsk->pi_lock by a concurrent waiter must
3652 * be observed in exit_pi_state_list().
3653 */
3654 raw_spin_lock_irq(&tsk->pi_lock);
3655 tsk->futex_state = FUTEX_STATE_EXITING;
3656 raw_spin_unlock_irq(&tsk->pi_lock);
3657}
3658
3659static void futex_cleanup_end(struct task_struct *tsk, int state)
3660{
3661 /*
3662 * Lockless store. The only side effect is that an observer might
3663 * take another loop until it becomes visible.
3664 */
3665 tsk->futex_state = state;
3666 /*
3667 * Drop the exit protection. This unblocks waiters which observed
3668 * FUTEX_STATE_EXITING to reevaluate the state.
3669 */
3670 mutex_unlock(&tsk->futex_exit_mutex);
3671}
3672
3673void futex_exec_release(struct task_struct *tsk)
3674{
3675 /*
3676 * The state handling is done for consistency, but in the case of
3677 * exec() there is no way to prevent further damage as the PID stays
3678 * the same. But for the unlikely and arguably buggy case that a
3679 * futex is held on exec(), this provides at least as much state
3680 * consistency protection which is possible.
3681 */
3682 futex_cleanup_begin(tsk);
3683 futex_cleanup(tsk);
3684 /*
3685 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3686 * exec a new binary.
3687 */
3688 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3689}
3690
3691void futex_exit_release(struct task_struct *tsk)
3692{
3693 futex_cleanup_begin(tsk);
3694 futex_cleanup(tsk);
3695 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3696}
3697
3698long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3699 u32 __user *uaddr2, u32 val2, u32 val3)
3700{
3701 int cmd = op & FUTEX_CMD_MASK;
3702 unsigned int flags = 0;
3703
3704 if (!(op & FUTEX_PRIVATE_FLAG))
3705 flags |= FLAGS_SHARED;
3706
3707 if (op & FUTEX_CLOCK_REALTIME) {
3708 flags |= FLAGS_CLOCKRT;
3709 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI &&
3710 cmd != FUTEX_LOCK_PI2)
3711 return -ENOSYS;
3712 }
3713
3714 switch (cmd) {
3715 case FUTEX_LOCK_PI:
3716 case FUTEX_LOCK_PI2:
3717 case FUTEX_UNLOCK_PI:
3718 case FUTEX_TRYLOCK_PI:
3719 case FUTEX_WAIT_REQUEUE_PI:
3720 case FUTEX_CMP_REQUEUE_PI:
3721 if (!futex_cmpxchg_enabled)
3722 return -ENOSYS;
3723 }
3724
3725 switch (cmd) {
3726 case FUTEX_WAIT:
3727 val3 = FUTEX_BITSET_MATCH_ANY;
3728 fallthrough;
3729 case FUTEX_WAIT_BITSET:
3730 return futex_wait(uaddr, flags, val, timeout, val3);
3731 case FUTEX_WAKE:
3732 val3 = FUTEX_BITSET_MATCH_ANY;
3733 fallthrough;
3734 case FUTEX_WAKE_BITSET:
3735 return futex_wake(uaddr, flags, val, val3);
3736 case FUTEX_REQUEUE:
3737 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3738 case FUTEX_CMP_REQUEUE:
3739 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3740 case FUTEX_WAKE_OP:
3741 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3742 case FUTEX_LOCK_PI:
3743 flags |= FLAGS_CLOCKRT;
3744 fallthrough;
3745 case FUTEX_LOCK_PI2:
3746 return futex_lock_pi(uaddr, flags, timeout, 0);
3747 case FUTEX_UNLOCK_PI:
3748 return futex_unlock_pi(uaddr, flags);
3749 case FUTEX_TRYLOCK_PI:
3750 return futex_lock_pi(uaddr, flags, NULL, 1);
3751 case FUTEX_WAIT_REQUEUE_PI:
3752 val3 = FUTEX_BITSET_MATCH_ANY;
3753 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3754 uaddr2);
3755 case FUTEX_CMP_REQUEUE_PI:
3756 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3757 }
3758 return -ENOSYS;
3759}
3760
3761static __always_inline bool futex_cmd_has_timeout(u32 cmd)
3762{
3763 switch (cmd) {
3764 case FUTEX_WAIT:
3765 case FUTEX_LOCK_PI:
3766 case FUTEX_LOCK_PI2:
3767 case FUTEX_WAIT_BITSET:
3768 case FUTEX_WAIT_REQUEUE_PI:
3769 return true;
3770 }
3771 return false;
3772}
3773
3774static __always_inline int
3775futex_init_timeout(u32 cmd, u32 op, struct timespec64 *ts, ktime_t *t)
3776{
3777 if (!timespec64_valid(ts))
3778 return -EINVAL;
3779
3780 *t = timespec64_to_ktime(*ts);
3781 if (cmd == FUTEX_WAIT)
3782 *t = ktime_add_safe(ktime_get(), *t);
3783 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3784 *t = timens_ktime_to_host(CLOCK_MONOTONIC, *t);
3785 return 0;
3786}
3787
3788SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3789 const struct __kernel_timespec __user *, utime,
3790 u32 __user *, uaddr2, u32, val3)
3791{
3792 int ret, cmd = op & FUTEX_CMD_MASK;
3793 ktime_t t, *tp = NULL;
3794 struct timespec64 ts;
3795
3796 if (utime && futex_cmd_has_timeout(cmd)) {
3797 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3798 return -EFAULT;
3799 if (get_timespec64(&ts, utime))
3800 return -EFAULT;
3801 ret = futex_init_timeout(cmd, op, &ts, &t);
3802 if (ret)
3803 return ret;
3804 tp = &t;
3805 }
3806
3807 return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3808}
3809
3810#ifdef CONFIG_COMPAT
3811/*
3812 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3813 */
3814static inline int
3815compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3816 compat_uptr_t __user *head, unsigned int *pi)
3817{
3818 if (get_user(*uentry, head))
3819 return -EFAULT;
3820
3821 *entry = compat_ptr((*uentry) & ~1);
3822 *pi = (unsigned int)(*uentry) & 1;
3823
3824 return 0;
3825}
3826
3827static void __user *futex_uaddr(struct robust_list __user *entry,
3828 compat_long_t futex_offset)
3829{
3830 compat_uptr_t base = ptr_to_compat(entry);
3831 void __user *uaddr = compat_ptr(base + futex_offset);
3832
3833 return uaddr;
3834}
3835
3836/*
3837 * Walk curr->robust_list (very carefully, it's a userspace list!)
3838 * and mark any locks found there dead, and notify any waiters.
3839 *
3840 * We silently return on any sign of list-walking problem.
3841 */
3842static void compat_exit_robust_list(struct task_struct *curr)
3843{
3844 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3845 struct robust_list __user *entry, *next_entry, *pending;
3846 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3847 unsigned int next_pi;
3848 compat_uptr_t uentry, next_uentry, upending;
3849 compat_long_t futex_offset;
3850 int rc;
3851
3852 if (!futex_cmpxchg_enabled)
3853 return;
3854
3855 /*
3856 * Fetch the list head (which was registered earlier, via
3857 * sys_set_robust_list()):
3858 */
3859 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3860 return;
3861 /*
3862 * Fetch the relative futex offset:
3863 */
3864 if (get_user(futex_offset, &head->futex_offset))
3865 return;
3866 /*
3867 * Fetch any possibly pending lock-add first, and handle it
3868 * if it exists:
3869 */
3870 if (compat_fetch_robust_entry(&upending, &pending,
3871 &head->list_op_pending, &pip))
3872 return;
3873
3874 next_entry = NULL; /* avoid warning with gcc */
3875 while (entry != (struct robust_list __user *) &head->list) {
3876 /*
3877 * Fetch the next entry in the list before calling
3878 * handle_futex_death:
3879 */
3880 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3881 (compat_uptr_t __user *)&entry->next, &next_pi);
3882 /*
3883 * A pending lock might already be on the list, so
3884 * dont process it twice:
3885 */
3886 if (entry != pending) {
3887 void __user *uaddr = futex_uaddr(entry, futex_offset);
3888
3889 if (handle_futex_death(uaddr, curr, pi,
3890 HANDLE_DEATH_LIST))
3891 return;
3892 }
3893 if (rc)
3894 return;
3895 uentry = next_uentry;
3896 entry = next_entry;
3897 pi = next_pi;
3898 /*
3899 * Avoid excessively long or circular lists:
3900 */
3901 if (!--limit)
3902 break;
3903
3904 cond_resched();
3905 }
3906 if (pending) {
3907 void __user *uaddr = futex_uaddr(pending, futex_offset);
3908
3909 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3910 }
3911}
3912
3913COMPAT_SYSCALL_DEFINE2(set_robust_list,
3914 struct compat_robust_list_head __user *, head,
3915 compat_size_t, len)
3916{
3917 if (!futex_cmpxchg_enabled)
3918 return -ENOSYS;
3919
3920 if (unlikely(len != sizeof(*head)))
3921 return -EINVAL;
3922
3923 current->compat_robust_list = head;
3924
3925 return 0;
3926}
3927
3928COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3929 compat_uptr_t __user *, head_ptr,
3930 compat_size_t __user *, len_ptr)
3931{
3932 struct compat_robust_list_head __user *head;
3933 unsigned long ret;
3934 struct task_struct *p;
3935
3936 if (!futex_cmpxchg_enabled)
3937 return -ENOSYS;
3938
3939 rcu_read_lock();
3940
3941 ret = -ESRCH;
3942 if (!pid)
3943 p = current;
3944 else {
3945 p = find_task_by_vpid(pid);
3946 if (!p)
3947 goto err_unlock;
3948 }
3949
3950 ret = -EPERM;
3951 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3952 goto err_unlock;
3953
3954 head = p->compat_robust_list;
3955 rcu_read_unlock();
3956
3957 if (put_user(sizeof(*head), len_ptr))
3958 return -EFAULT;
3959 return put_user(ptr_to_compat(head), head_ptr);
3960
3961err_unlock:
3962 rcu_read_unlock();
3963
3964 return ret;
3965}
3966#endif /* CONFIG_COMPAT */
3967
3968#ifdef CONFIG_COMPAT_32BIT_TIME
3969SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3970 const struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3971 u32, val3)
3972{
3973 int ret, cmd = op & FUTEX_CMD_MASK;
3974 ktime_t t, *tp = NULL;
3975 struct timespec64 ts;
3976
3977 if (utime && futex_cmd_has_timeout(cmd)) {
3978 if (get_old_timespec32(&ts, utime))
3979 return -EFAULT;
3980 ret = futex_init_timeout(cmd, op, &ts, &t);
3981 if (ret)
3982 return ret;
3983 tp = &t;
3984 }
3985
3986 return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3987}
3988#endif /* CONFIG_COMPAT_32BIT_TIME */
3989
3990static void __init futex_detect_cmpxchg(void)
3991{
3992#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3993 u32 curval;
3994
3995 /*
3996 * This will fail and we want it. Some arch implementations do
3997 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3998 * functionality. We want to know that before we call in any
3999 * of the complex code paths. Also we want to prevent
4000 * registration of robust lists in that case. NULL is
4001 * guaranteed to fault and we get -EFAULT on functional
4002 * implementation, the non-functional ones will return
4003 * -ENOSYS.
4004 */
4005 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4006 futex_cmpxchg_enabled = 1;
4007#endif
4008}
4009
4010static int __init futex_init(void)
4011{
4012 unsigned int futex_shift;
4013 unsigned long i;
4014
4015#if CONFIG_BASE_SMALL
4016 futex_hashsize = 16;
4017#else
4018 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4019#endif
4020
4021 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4022 futex_hashsize, 0,
4023 futex_hashsize < 256 ? HASH_SMALL : 0,
4024 &futex_shift, NULL,
4025 futex_hashsize, futex_hashsize);
4026 futex_hashsize = 1UL << futex_shift;
4027
4028 futex_detect_cmpxchg();
4029
4030 for (i = 0; i < futex_hashsize; i++) {
4031 atomic_set(&futex_queues[i].waiters, 0);
4032 plist_head_init(&futex_queues[i].chain);
4033 spin_lock_init(&futex_queues[i].lock);
4034 }
4035
4036 return 0;
4037}
4038core_initcall(futex_init);