2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
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.
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>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
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.
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.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
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.
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.
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
47 #include <linux/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/memblock.h>
70 #include <linux/fault-inject.h>
72 #include <asm/futex.h>
74 #include "locking/rtmutex_common.h"
77 * READ this before attempting to hack on futexes!
79 * Basic futex operation and ordering guarantees
80 * =============================================
82 * The waiter reads the futex value in user space and calls
83 * futex_wait(). This function computes the hash bucket and acquires
84 * the hash bucket lock. After that it reads the futex user space value
85 * again and verifies that the data has not changed. If it has not changed
86 * it enqueues itself into the hash bucket, releases the hash bucket lock
89 * The waker side modifies the user space value of the futex and calls
90 * futex_wake(). This function computes the hash bucket and acquires the
91 * hash bucket lock. Then it looks for waiters on that futex in the hash
92 * bucket and wakes them.
94 * In futex wake up scenarios where no tasks are blocked on a futex, taking
95 * the hb spinlock can be avoided and simply return. In order for this
96 * optimization to work, ordering guarantees must exist so that the waiter
97 * being added to the list is acknowledged when the list is concurrently being
98 * checked by the waker, avoiding scenarios like the following:
102 * sys_futex(WAIT, futex, val);
103 * futex_wait(futex, val);
106 * sys_futex(WAKE, futex);
111 * lock(hash_bucket(futex));
113 * unlock(hash_bucket(futex));
116 * This would cause the waiter on CPU 0 to wait forever because it
117 * missed the transition of the user space value from val to newval
118 * and the waker did not find the waiter in the hash bucket queue.
120 * The correct serialization ensures that a waiter either observes
121 * the changed user space value before blocking or is woken by a
126 * sys_futex(WAIT, futex, val);
127 * futex_wait(futex, val);
130 * smp_mb(); (A) <-- paired with -.
132 * lock(hash_bucket(futex)); |
136 * | sys_futex(WAKE, futex);
137 * | futex_wake(futex);
139 * `--------> smp_mb(); (B)
142 * unlock(hash_bucket(futex));
143 * schedule(); if (waiters)
144 * lock(hash_bucket(futex));
145 * else wake_waiters(futex);
146 * waiters--; (b) unlock(hash_bucket(futex));
148 * Where (A) orders the waiters increment and the futex value read through
149 * atomic operations (see hb_waiters_inc) and where (B) orders the write
150 * to futex and the waiters read -- this is done by the barriers for both
151 * shared and private futexes in get_futex_key_refs().
153 * This yields the following case (where X:=waiters, Y:=futex):
161 * Which guarantees that x==0 && y==0 is impossible; which translates back into
162 * the guarantee that we cannot both miss the futex variable change and the
165 * Note that a new waiter is accounted for in (a) even when it is possible that
166 * the wait call can return error, in which case we backtrack from it in (b).
167 * Refer to the comment in queue_lock().
169 * Similarly, in order to account for waiters being requeued on another
170 * address we always increment the waiters for the destination bucket before
171 * acquiring the lock. It then decrements them again after releasing it -
172 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173 * will do the additional required waiter count housekeeping. This is done for
174 * double_lock_hb() and double_unlock_hb(), respectively.
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
180 static int __read_mostly futex_cmpxchg_enabled;
184 * Futex flags used to encode options to functions and preserve them across
188 # define FLAGS_SHARED 0x01
191 * NOMMU does not have per process address space. Let the compiler optimize
194 # define FLAGS_SHARED 0x00
196 #define FLAGS_CLOCKRT 0x02
197 #define FLAGS_HAS_TIMEOUT 0x04
200 * Priority Inheritance state:
202 struct futex_pi_state {
204 * list of 'owned' pi_state instances - these have to be
205 * cleaned up in do_exit() if the task exits prematurely:
207 struct list_head list;
212 struct rt_mutex pi_mutex;
214 struct task_struct *owner;
218 } __randomize_layout;
221 * struct futex_q - The hashed futex queue entry, one per waiting task
222 * @list: priority-sorted list of tasks waiting on this futex
223 * @task: the task waiting on the futex
224 * @lock_ptr: the hash bucket lock
225 * @key: the key the futex is hashed on
226 * @pi_state: optional priority inheritance state
227 * @rt_waiter: rt_waiter storage for use with requeue_pi
228 * @requeue_pi_key: the requeue_pi target futex key
229 * @bitset: bitset for the optional bitmasked wakeup
231 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232 * we can wake only the relevant ones (hashed queues may be shared).
234 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236 * The order of wakeup is always to make the first condition true, then
239 * PI futexes are typically woken before they are removed from the hash list via
240 * the rt_mutex code. See unqueue_me_pi().
243 struct plist_node list;
245 struct task_struct *task;
246 spinlock_t *lock_ptr;
248 struct futex_pi_state *pi_state;
249 struct rt_mutex_waiter *rt_waiter;
250 union futex_key *requeue_pi_key;
252 } __randomize_layout;
254 static const struct futex_q futex_q_init = {
255 /* list gets initialized in queue_me()*/
256 .key = FUTEX_KEY_INIT,
257 .bitset = FUTEX_BITSET_MATCH_ANY
261 * Hash buckets are shared by all the futex_keys that hash to the same
262 * location. Each key may have multiple futex_q structures, one for each task
263 * waiting on a futex.
265 struct futex_hash_bucket {
268 struct plist_head chain;
269 } ____cacheline_aligned_in_smp;
272 * The base of the bucket array and its size are always used together
273 * (after initialization only in hash_futex()), so ensure that they
274 * reside in the same cacheline.
277 struct futex_hash_bucket *queues;
278 unsigned long hashsize;
279 } __futex_data __read_mostly __aligned(2*sizeof(long));
280 #define futex_queues (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
285 * Fault injections for futexes.
287 #ifdef CONFIG_FAIL_FUTEX
290 struct fault_attr attr;
294 .attr = FAULT_ATTR_INITIALIZER,
295 .ignore_private = false,
298 static int __init setup_fail_futex(char *str)
300 return setup_fault_attr(&fail_futex.attr, str);
302 __setup("fail_futex=", setup_fail_futex);
304 static bool should_fail_futex(bool fshared)
306 if (fail_futex.ignore_private && !fshared)
309 return should_fail(&fail_futex.attr, 1);
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314 static int __init fail_futex_debugfs(void)
316 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
319 dir = fault_create_debugfs_attr("fail_futex", NULL,
324 if (!debugfs_create_bool("ignore-private", mode, dir,
325 &fail_futex.ignore_private)) {
326 debugfs_remove_recursive(dir);
333 late_initcall(fail_futex_debugfs);
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
338 static inline bool should_fail_futex(bool fshared)
342 #endif /* CONFIG_FAIL_FUTEX */
344 static inline void futex_get_mm(union futex_key *key)
346 mmgrab(key->private.mm);
348 * Ensure futex_get_mm() implies a full barrier such that
349 * get_futex_key() implies a full barrier. This is relied upon
350 * as smp_mb(); (B), see the ordering comment above.
352 smp_mb__after_atomic();
356 * Reflects a new waiter being added to the waitqueue.
358 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
361 atomic_inc(&hb->waiters);
363 * Full barrier (A), see the ordering comment above.
365 smp_mb__after_atomic();
370 * Reflects a waiter being removed from the waitqueue by wakeup
373 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
376 atomic_dec(&hb->waiters);
380 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
383 return atomic_read(&hb->waiters);
390 * hash_futex - Return the hash bucket in the global hash
391 * @key: Pointer to the futex key for which the hash is calculated
393 * We hash on the keys returned from get_futex_key (see below) and return the
394 * corresponding hash bucket in the global hash.
396 static struct futex_hash_bucket *hash_futex(union futex_key *key)
398 u32 hash = jhash2((u32*)&key->both.word,
399 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
401 return &futex_queues[hash & (futex_hashsize - 1)];
406 * match_futex - Check whether two futex keys are equal
407 * @key1: Pointer to key1
408 * @key2: Pointer to key2
410 * Return 1 if two futex_keys are equal, 0 otherwise.
412 static inline int match_futex(union futex_key *key1, union futex_key *key2)
415 && key1->both.word == key2->both.word
416 && key1->both.ptr == key2->both.ptr
417 && key1->both.offset == key2->both.offset);
421 * Take a reference to the resource addressed by a key.
422 * Can be called while holding spinlocks.
425 static void get_futex_key_refs(union futex_key *key)
431 * On MMU less systems futexes are always "private" as there is no per
432 * process address space. We need the smp wmb nevertheless - yes,
433 * arch/blackfin has MMU less SMP ...
435 if (!IS_ENABLED(CONFIG_MMU)) {
436 smp_mb(); /* explicit smp_mb(); (B) */
440 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
442 ihold(key->shared.inode); /* implies smp_mb(); (B) */
444 case FUT_OFF_MMSHARED:
445 futex_get_mm(key); /* implies smp_mb(); (B) */
449 * Private futexes do not hold reference on an inode or
450 * mm, therefore the only purpose of calling get_futex_key_refs
451 * is because we need the barrier for the lockless waiter check.
453 smp_mb(); /* explicit smp_mb(); (B) */
458 * Drop a reference to the resource addressed by a key.
459 * The hash bucket spinlock must not be held. This is
460 * a no-op for private futexes, see comment in the get
463 static void drop_futex_key_refs(union futex_key *key)
465 if (!key->both.ptr) {
466 /* If we're here then we tried to put a key we failed to get */
471 if (!IS_ENABLED(CONFIG_MMU))
474 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
476 iput(key->shared.inode);
478 case FUT_OFF_MMSHARED:
479 mmdrop(key->private.mm);
485 * get_futex_key() - Get parameters which are the keys for a futex
486 * @uaddr: virtual address of the futex
487 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
488 * @key: address where result is stored.
489 * @rw: mapping needs to be read/write (values: VERIFY_READ,
492 * Return: a negative error code or 0
494 * The key words are stored in @key on success.
496 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
497 * offset_within_page). For private mappings, it's (uaddr, current->mm).
498 * We can usually work out the index without swapping in the page.
500 * lock_page() might sleep, the caller should not hold a spinlock.
503 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
505 unsigned long address = (unsigned long)uaddr;
506 struct mm_struct *mm = current->mm;
507 struct page *page, *tail;
508 struct address_space *mapping;
512 * The futex address must be "naturally" aligned.
514 key->both.offset = address % PAGE_SIZE;
515 if (unlikely((address % sizeof(u32)) != 0))
517 address -= key->both.offset;
519 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
522 if (unlikely(should_fail_futex(fshared)))
526 * PROCESS_PRIVATE futexes are fast.
527 * As the mm cannot disappear under us and the 'key' only needs
528 * virtual address, we dont even have to find the underlying vma.
529 * Note : We do have to check 'uaddr' is a valid user address,
530 * but access_ok() should be faster than find_vma()
533 key->private.mm = mm;
534 key->private.address = address;
535 get_futex_key_refs(key); /* implies smp_mb(); (B) */
540 /* Ignore any VERIFY_READ mapping (futex common case) */
541 if (unlikely(should_fail_futex(fshared)))
544 err = get_user_pages_fast(address, 1, 1, &page);
546 * If write access is not required (eg. FUTEX_WAIT), try
547 * and get read-only access.
549 if (err == -EFAULT && rw == VERIFY_READ) {
550 err = get_user_pages_fast(address, 1, 0, &page);
559 * The treatment of mapping from this point on is critical. The page
560 * lock protects many things but in this context the page lock
561 * stabilizes mapping, prevents inode freeing in the shared
562 * file-backed region case and guards against movement to swap cache.
564 * Strictly speaking the page lock is not needed in all cases being
565 * considered here and page lock forces unnecessarily serialization
566 * From this point on, mapping will be re-verified if necessary and
567 * page lock will be acquired only if it is unavoidable
569 * Mapping checks require the head page for any compound page so the
570 * head page and mapping is looked up now. For anonymous pages, it
571 * does not matter if the page splits in the future as the key is
572 * based on the address. For filesystem-backed pages, the tail is
573 * required as the index of the page determines the key. For
574 * base pages, there is no tail page and tail == page.
577 page = compound_head(page);
578 mapping = READ_ONCE(page->mapping);
581 * If page->mapping is NULL, then it cannot be a PageAnon
582 * page; but it might be the ZERO_PAGE or in the gate area or
583 * in a special mapping (all cases which we are happy to fail);
584 * or it may have been a good file page when get_user_pages_fast
585 * found it, but truncated or holepunched or subjected to
586 * invalidate_complete_page2 before we got the page lock (also
587 * cases which we are happy to fail). And we hold a reference,
588 * so refcount care in invalidate_complete_page's remove_mapping
589 * prevents drop_caches from setting mapping to NULL beneath us.
591 * The case we do have to guard against is when memory pressure made
592 * shmem_writepage move it from filecache to swapcache beneath us:
593 * an unlikely race, but we do need to retry for page->mapping.
595 if (unlikely(!mapping)) {
599 * Page lock is required to identify which special case above
600 * applies. If this is really a shmem page then the page lock
601 * will prevent unexpected transitions.
604 shmem_swizzled = PageSwapCache(page) || page->mapping;
615 * Private mappings are handled in a simple way.
617 * If the futex key is stored on an anonymous page, then the associated
618 * object is the mm which is implicitly pinned by the calling process.
620 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
621 * it's a read-only handle, it's expected that futexes attach to
622 * the object not the particular process.
624 if (PageAnon(page)) {
626 * A RO anonymous page will never change and thus doesn't make
627 * sense for futex operations.
629 if (unlikely(should_fail_futex(fshared)) || ro) {
634 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
635 key->private.mm = mm;
636 key->private.address = address;
638 get_futex_key_refs(key); /* implies smp_mb(); (B) */
644 * The associated futex object in this case is the inode and
645 * the page->mapping must be traversed. Ordinarily this should
646 * be stabilised under page lock but it's not strictly
647 * necessary in this case as we just want to pin the inode, not
648 * update the radix tree or anything like that.
650 * The RCU read lock is taken as the inode is finally freed
651 * under RCU. If the mapping still matches expectations then the
652 * mapping->host can be safely accessed as being a valid inode.
656 if (READ_ONCE(page->mapping) != mapping) {
663 inode = READ_ONCE(mapping->host);
672 * Take a reference unless it is about to be freed. Previously
673 * this reference was taken by ihold under the page lock
674 * pinning the inode in place so i_lock was unnecessary. The
675 * only way for this check to fail is if the inode was
676 * truncated in parallel which is almost certainly an
677 * application bug. In such a case, just retry.
679 * We are not calling into get_futex_key_refs() in file-backed
680 * cases, therefore a successful atomic_inc return below will
681 * guarantee that get_futex_key() will still imply smp_mb(); (B).
683 if (!atomic_inc_not_zero(&inode->i_count)) {
690 /* Should be impossible but lets be paranoid for now */
691 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
699 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
700 key->shared.inode = inode;
701 key->shared.pgoff = basepage_index(tail);
710 static inline void put_futex_key(union futex_key *key)
712 drop_futex_key_refs(key);
716 * fault_in_user_writeable() - Fault in user address and verify RW access
717 * @uaddr: pointer to faulting user space address
719 * Slow path to fixup the fault we just took in the atomic write
722 * We have no generic implementation of a non-destructive write to the
723 * user address. We know that we faulted in the atomic pagefault
724 * disabled section so we can as well avoid the #PF overhead by
725 * calling get_user_pages() right away.
727 static int fault_in_user_writeable(u32 __user *uaddr)
729 struct mm_struct *mm = current->mm;
732 down_read(&mm->mmap_sem);
733 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
734 FAULT_FLAG_WRITE, NULL);
735 up_read(&mm->mmap_sem);
737 return ret < 0 ? ret : 0;
741 * futex_top_waiter() - Return the highest priority waiter on a futex
742 * @hb: the hash bucket the futex_q's reside in
743 * @key: the futex key (to distinguish it from other futex futex_q's)
745 * Must be called with the hb lock held.
747 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
748 union futex_key *key)
750 struct futex_q *this;
752 plist_for_each_entry(this, &hb->chain, list) {
753 if (match_futex(&this->key, key))
759 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
760 u32 uval, u32 newval)
765 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
771 static int get_futex_value_locked(u32 *dest, u32 __user *from)
776 ret = __get_user(*dest, from);
779 return ret ? -EFAULT : 0;
786 static int refill_pi_state_cache(void)
788 struct futex_pi_state *pi_state;
790 if (likely(current->pi_state_cache))
793 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
798 INIT_LIST_HEAD(&pi_state->list);
799 /* pi_mutex gets initialized later */
800 pi_state->owner = NULL;
801 atomic_set(&pi_state->refcount, 1);
802 pi_state->key = FUTEX_KEY_INIT;
804 current->pi_state_cache = pi_state;
809 static struct futex_pi_state *alloc_pi_state(void)
811 struct futex_pi_state *pi_state = current->pi_state_cache;
814 current->pi_state_cache = NULL;
819 static void get_pi_state(struct futex_pi_state *pi_state)
821 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
825 * Drops a reference to the pi_state object and frees or caches it
826 * when the last reference is gone.
828 static void put_pi_state(struct futex_pi_state *pi_state)
833 if (!atomic_dec_and_test(&pi_state->refcount))
837 * If pi_state->owner is NULL, the owner is most probably dying
838 * and has cleaned up the pi_state already
840 if (pi_state->owner) {
841 struct task_struct *owner;
843 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
844 owner = pi_state->owner;
846 raw_spin_lock(&owner->pi_lock);
847 list_del_init(&pi_state->list);
848 raw_spin_unlock(&owner->pi_lock);
850 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
851 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
854 if (current->pi_state_cache) {
858 * pi_state->list is already empty.
859 * clear pi_state->owner.
860 * refcount is at 0 - put it back to 1.
862 pi_state->owner = NULL;
863 atomic_set(&pi_state->refcount, 1);
864 current->pi_state_cache = pi_state;
868 #ifdef CONFIG_FUTEX_PI
871 * This task is holding PI mutexes at exit time => bad.
872 * Kernel cleans up PI-state, but userspace is likely hosed.
873 * (Robust-futex cleanup is separate and might save the day for userspace.)
875 void exit_pi_state_list(struct task_struct *curr)
877 struct list_head *next, *head = &curr->pi_state_list;
878 struct futex_pi_state *pi_state;
879 struct futex_hash_bucket *hb;
880 union futex_key key = FUTEX_KEY_INIT;
882 if (!futex_cmpxchg_enabled)
885 * We are a ZOMBIE and nobody can enqueue itself on
886 * pi_state_list anymore, but we have to be careful
887 * versus waiters unqueueing themselves:
889 raw_spin_lock_irq(&curr->pi_lock);
890 while (!list_empty(head)) {
892 pi_state = list_entry(next, struct futex_pi_state, list);
894 hb = hash_futex(&key);
897 * We can race against put_pi_state() removing itself from the
898 * list (a waiter going away). put_pi_state() will first
899 * decrement the reference count and then modify the list, so
900 * its possible to see the list entry but fail this reference
903 * In that case; drop the locks to let put_pi_state() make
904 * progress and retry the loop.
906 if (!atomic_inc_not_zero(&pi_state->refcount)) {
907 raw_spin_unlock_irq(&curr->pi_lock);
909 raw_spin_lock_irq(&curr->pi_lock);
912 raw_spin_unlock_irq(&curr->pi_lock);
914 spin_lock(&hb->lock);
915 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
916 raw_spin_lock(&curr->pi_lock);
918 * We dropped the pi-lock, so re-check whether this
919 * task still owns the PI-state:
921 if (head->next != next) {
922 /* retain curr->pi_lock for the loop invariant */
923 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
924 spin_unlock(&hb->lock);
925 put_pi_state(pi_state);
929 WARN_ON(pi_state->owner != curr);
930 WARN_ON(list_empty(&pi_state->list));
931 list_del_init(&pi_state->list);
932 pi_state->owner = NULL;
934 raw_spin_unlock(&curr->pi_lock);
935 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
936 spin_unlock(&hb->lock);
938 rt_mutex_futex_unlock(&pi_state->pi_mutex);
939 put_pi_state(pi_state);
941 raw_spin_lock_irq(&curr->pi_lock);
943 raw_spin_unlock_irq(&curr->pi_lock);
949 * We need to check the following states:
951 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
953 * [1] NULL | --- | --- | 0 | 0/1 | Valid
954 * [2] NULL | --- | --- | >0 | 0/1 | Valid
956 * [3] Found | NULL | -- | Any | 0/1 | Invalid
958 * [4] Found | Found | NULL | 0 | 1 | Valid
959 * [5] Found | Found | NULL | >0 | 1 | Invalid
961 * [6] Found | Found | task | 0 | 1 | Valid
963 * [7] Found | Found | NULL | Any | 0 | Invalid
965 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
966 * [9] Found | Found | task | 0 | 0 | Invalid
967 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
969 * [1] Indicates that the kernel can acquire the futex atomically. We
970 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
972 * [2] Valid, if TID does not belong to a kernel thread. If no matching
973 * thread is found then it indicates that the owner TID has died.
975 * [3] Invalid. The waiter is queued on a non PI futex
977 * [4] Valid state after exit_robust_list(), which sets the user space
978 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
980 * [5] The user space value got manipulated between exit_robust_list()
981 * and exit_pi_state_list()
983 * [6] Valid state after exit_pi_state_list() which sets the new owner in
984 * the pi_state but cannot access the user space value.
986 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
988 * [8] Owner and user space value match
990 * [9] There is no transient state which sets the user space TID to 0
991 * except exit_robust_list(), but this is indicated by the
992 * FUTEX_OWNER_DIED bit. See [4]
994 * [10] There is no transient state which leaves owner and user space
998 * Serialization and lifetime rules:
1002 * hb -> futex_q, relation
1003 * futex_q -> pi_state, relation
1005 * (cannot be raw because hb can contain arbitrary amount
1008 * pi_mutex->wait_lock:
1012 * (and pi_mutex 'obviously')
1016 * p->pi_state_list -> pi_state->list, relation
1018 * pi_state->refcount:
1026 * pi_mutex->wait_lock
1032 * Validate that the existing waiter has a pi_state and sanity check
1033 * the pi_state against the user space value. If correct, attach to
1036 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1037 struct futex_pi_state *pi_state,
1038 struct futex_pi_state **ps)
1040 pid_t pid = uval & FUTEX_TID_MASK;
1045 * Userspace might have messed up non-PI and PI futexes [3]
1047 if (unlikely(!pi_state))
1051 * We get here with hb->lock held, and having found a
1052 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1053 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1054 * which in turn means that futex_lock_pi() still has a reference on
1057 * The waiter holding a reference on @pi_state also protects against
1058 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1059 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1060 * free pi_state before we can take a reference ourselves.
1062 WARN_ON(!atomic_read(&pi_state->refcount));
1065 * Now that we have a pi_state, we can acquire wait_lock
1066 * and do the state validation.
1068 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1071 * Since {uval, pi_state} is serialized by wait_lock, and our current
1072 * uval was read without holding it, it can have changed. Verify it
1073 * still is what we expect it to be, otherwise retry the entire
1076 if (get_futex_value_locked(&uval2, uaddr))
1083 * Handle the owner died case:
1085 if (uval & FUTEX_OWNER_DIED) {
1087 * exit_pi_state_list sets owner to NULL and wakes the
1088 * topmost waiter. The task which acquires the
1089 * pi_state->rt_mutex will fixup owner.
1091 if (!pi_state->owner) {
1093 * No pi state owner, but the user space TID
1094 * is not 0. Inconsistent state. [5]
1099 * Take a ref on the state and return success. [4]
1105 * If TID is 0, then either the dying owner has not
1106 * yet executed exit_pi_state_list() or some waiter
1107 * acquired the rtmutex in the pi state, but did not
1108 * yet fixup the TID in user space.
1110 * Take a ref on the state and return success. [6]
1116 * If the owner died bit is not set, then the pi_state
1117 * must have an owner. [7]
1119 if (!pi_state->owner)
1124 * Bail out if user space manipulated the futex value. If pi
1125 * state exists then the owner TID must be the same as the
1126 * user space TID. [9/10]
1128 if (pid != task_pid_vnr(pi_state->owner))
1132 get_pi_state(pi_state);
1133 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1150 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1154 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1155 struct task_struct *tsk)
1160 * If PF_EXITPIDONE is not yet set, then try again.
1162 if (tsk && !(tsk->flags & PF_EXITPIDONE))
1166 * Reread the user space value to handle the following situation:
1170 * sys_exit() sys_futex()
1171 * do_exit() futex_lock_pi()
1172 * futex_lock_pi_atomic()
1173 * exit_signals(tsk) No waiters:
1174 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1175 * mm_release(tsk) Set waiter bit
1176 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1177 * Set owner died attach_to_pi_owner() {
1178 * *uaddr = 0xC0000000; tsk = get_task(PID);
1179 * } if (!tsk->flags & PF_EXITING) {
1181 * tsk->flags |= PF_EXITPIDONE; } else {
1182 * if (!(tsk->flags & PF_EXITPIDONE))
1184 * return -ESRCH; <--- FAIL
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1190 * The same logic applies to the case where the exiting task is
1193 if (get_futex_value_locked(&uval2, uaddr))
1196 /* If the user space value has changed, try again. */
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.
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1212 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213 struct futex_pi_state **ps)
1215 pid_t pid = uval & FUTEX_TID_MASK;
1216 struct futex_pi_state *pi_state;
1217 struct task_struct *p;
1220 * We are the first waiter - try to look up the real owner and attach
1221 * the new pi_state to it, but bail out when TID = 0 [1]
1223 * The !pid check is paranoid. None of the call sites should end up
1224 * with pid == 0, but better safe than sorry. Let the caller retry
1228 p = find_get_task_by_vpid(pid);
1230 return handle_exit_race(uaddr, uval, NULL);
1232 if (unlikely(p->flags & PF_KTHREAD)) {
1238 * We need to look at the task state flags to figure out,
1239 * whether the task is exiting. To protect against the do_exit
1240 * change of the task flags, we do this protected by
1243 raw_spin_lock_irq(&p->pi_lock);
1244 if (unlikely(p->flags & PF_EXITING)) {
1246 * The task is on the way out. When PF_EXITPIDONE is
1247 * set, we know that the task has finished the
1250 int ret = handle_exit_race(uaddr, uval, p);
1252 raw_spin_unlock_irq(&p->pi_lock);
1258 * No existing pi state. First waiter. [2]
1260 * This creates pi_state, we have hb->lock held, this means nothing can
1261 * observe this state, wait_lock is irrelevant.
1263 pi_state = alloc_pi_state();
1266 * Initialize the pi_mutex in locked state and make @p
1269 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1271 /* Store the key for possible exit cleanups: */
1272 pi_state->key = *key;
1274 WARN_ON(!list_empty(&pi_state->list));
1275 list_add(&pi_state->list, &p->pi_state_list);
1277 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1278 * because there is no concurrency as the object is not published yet.
1280 pi_state->owner = p;
1281 raw_spin_unlock_irq(&p->pi_lock);
1290 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1291 struct futex_hash_bucket *hb,
1292 union futex_key *key, struct futex_pi_state **ps)
1294 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1297 * If there is a waiter on that futex, validate it and
1298 * attach to the pi_state when the validation succeeds.
1301 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1304 * We are the first waiter - try to look up the owner based on
1305 * @uval and attach to it.
1307 return attach_to_pi_owner(uaddr, uval, key, ps);
1310 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1312 u32 uninitialized_var(curval);
1314 if (unlikely(should_fail_futex(true)))
1317 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1320 /* If user space value changed, let the caller retry */
1321 return curval != uval ? -EAGAIN : 0;
1325 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1326 * @uaddr: the pi futex user address
1327 * @hb: the pi futex hash bucket
1328 * @key: the futex key associated with uaddr and hb
1329 * @ps: the pi_state pointer where we store the result of the
1331 * @task: the task to perform the atomic lock work for. This will
1332 * be "current" except in the case of requeue pi.
1333 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1336 * - 0 - ready to wait;
1337 * - 1 - acquired the lock;
1340 * The hb->lock and futex_key refs shall be held by the caller.
1342 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1343 union futex_key *key,
1344 struct futex_pi_state **ps,
1345 struct task_struct *task, int set_waiters)
1347 u32 uval, newval, vpid = task_pid_vnr(task);
1348 struct futex_q *top_waiter;
1352 * Read the user space value first so we can validate a few
1353 * things before proceeding further.
1355 if (get_futex_value_locked(&uval, uaddr))
1358 if (unlikely(should_fail_futex(true)))
1364 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1367 if ((unlikely(should_fail_futex(true))))
1371 * Lookup existing state first. If it exists, try to attach to
1374 top_waiter = futex_top_waiter(hb, key);
1376 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1379 * No waiter and user TID is 0. We are here because the
1380 * waiters or the owner died bit is set or called from
1381 * requeue_cmp_pi or for whatever reason something took the
1384 if (!(uval & FUTEX_TID_MASK)) {
1386 * We take over the futex. No other waiters and the user space
1387 * TID is 0. We preserve the owner died bit.
1389 newval = uval & FUTEX_OWNER_DIED;
1392 /* The futex requeue_pi code can enforce the waiters bit */
1394 newval |= FUTEX_WAITERS;
1396 ret = lock_pi_update_atomic(uaddr, uval, newval);
1397 /* If the take over worked, return 1 */
1398 return ret < 0 ? ret : 1;
1402 * First waiter. Set the waiters bit before attaching ourself to
1403 * the owner. If owner tries to unlock, it will be forced into
1404 * the kernel and blocked on hb->lock.
1406 newval = uval | FUTEX_WAITERS;
1407 ret = lock_pi_update_atomic(uaddr, uval, newval);
1411 * If the update of the user space value succeeded, we try to
1412 * attach to the owner. If that fails, no harm done, we only
1413 * set the FUTEX_WAITERS bit in the user space variable.
1415 return attach_to_pi_owner(uaddr, newval, key, ps);
1419 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1420 * @q: The futex_q to unqueue
1422 * The q->lock_ptr must not be NULL and must be held by the caller.
1424 static void __unqueue_futex(struct futex_q *q)
1426 struct futex_hash_bucket *hb;
1428 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1430 lockdep_assert_held(q->lock_ptr);
1432 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1433 plist_del(&q->list, &hb->chain);
1438 * The hash bucket lock must be held when this is called.
1439 * Afterwards, the futex_q must not be accessed. Callers
1440 * must ensure to later call wake_up_q() for the actual
1443 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1445 struct task_struct *p = q->task;
1447 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1451 * Queue the task for later wakeup for after we've released
1452 * the hb->lock. wake_q_add() grabs reference to p.
1454 wake_q_add(wake_q, p);
1457 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1458 * is written, without taking any locks. This is possible in the event
1459 * of a spurious wakeup, for example. A memory barrier is required here
1460 * to prevent the following store to lock_ptr from getting ahead of the
1461 * plist_del in __unqueue_futex().
1463 smp_store_release(&q->lock_ptr, NULL);
1467 * Caller must hold a reference on @pi_state.
1469 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1471 u32 uninitialized_var(curval), newval;
1472 struct task_struct *new_owner;
1473 bool postunlock = false;
1474 DEFINE_WAKE_Q(wake_q);
1477 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1478 if (WARN_ON_ONCE(!new_owner)) {
1480 * As per the comment in futex_unlock_pi() this should not happen.
1482 * When this happens, give up our locks and try again, giving
1483 * the futex_lock_pi() instance time to complete, either by
1484 * waiting on the rtmutex or removing itself from the futex
1492 * We pass it to the next owner. The WAITERS bit is always kept
1493 * enabled while there is PI state around. We cleanup the owner
1494 * died bit, because we are the owner.
1496 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1498 if (unlikely(should_fail_futex(true)))
1501 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1504 } else if (curval != uval) {
1506 * If a unconditional UNLOCK_PI operation (user space did not
1507 * try the TID->0 transition) raced with a waiter setting the
1508 * FUTEX_WAITERS flag between get_user() and locking the hash
1509 * bucket lock, retry the operation.
1511 if ((FUTEX_TID_MASK & curval) == uval)
1521 * This is a point of no return; once we modify the uval there is no
1522 * going back and subsequent operations must not fail.
1525 raw_spin_lock(&pi_state->owner->pi_lock);
1526 WARN_ON(list_empty(&pi_state->list));
1527 list_del_init(&pi_state->list);
1528 raw_spin_unlock(&pi_state->owner->pi_lock);
1530 raw_spin_lock(&new_owner->pi_lock);
1531 WARN_ON(!list_empty(&pi_state->list));
1532 list_add(&pi_state->list, &new_owner->pi_state_list);
1533 pi_state->owner = new_owner;
1534 raw_spin_unlock(&new_owner->pi_lock);
1536 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1539 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1542 rt_mutex_postunlock(&wake_q);
1548 * Express the locking dependencies for lockdep:
1551 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1554 spin_lock(&hb1->lock);
1556 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1557 } else { /* hb1 > hb2 */
1558 spin_lock(&hb2->lock);
1559 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1564 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1566 spin_unlock(&hb1->lock);
1568 spin_unlock(&hb2->lock);
1572 * Wake up waiters matching bitset queued on this futex (uaddr).
1575 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1577 struct futex_hash_bucket *hb;
1578 struct futex_q *this, *next;
1579 union futex_key key = FUTEX_KEY_INIT;
1581 DEFINE_WAKE_Q(wake_q);
1586 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1587 if (unlikely(ret != 0))
1590 hb = hash_futex(&key);
1592 /* Make sure we really have tasks to wakeup */
1593 if (!hb_waiters_pending(hb))
1596 spin_lock(&hb->lock);
1598 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1599 if (match_futex (&this->key, &key)) {
1600 if (this->pi_state || this->rt_waiter) {
1605 /* Check if one of the bits is set in both bitsets */
1606 if (!(this->bitset & bitset))
1609 mark_wake_futex(&wake_q, this);
1610 if (++ret >= nr_wake)
1615 spin_unlock(&hb->lock);
1618 put_futex_key(&key);
1623 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1625 unsigned int op = (encoded_op & 0x70000000) >> 28;
1626 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1627 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1628 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1631 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1632 if (oparg < 0 || oparg > 31) {
1633 char comm[sizeof(current->comm)];
1635 * kill this print and return -EINVAL when userspace
1638 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1639 get_task_comm(comm, current), oparg);
1645 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1648 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1653 case FUTEX_OP_CMP_EQ:
1654 return oldval == cmparg;
1655 case FUTEX_OP_CMP_NE:
1656 return oldval != cmparg;
1657 case FUTEX_OP_CMP_LT:
1658 return oldval < cmparg;
1659 case FUTEX_OP_CMP_GE:
1660 return oldval >= cmparg;
1661 case FUTEX_OP_CMP_LE:
1662 return oldval <= cmparg;
1663 case FUTEX_OP_CMP_GT:
1664 return oldval > cmparg;
1671 * Wake up all waiters hashed on the physical page that is mapped
1672 * to this virtual address:
1675 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1676 int nr_wake, int nr_wake2, int op)
1678 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1679 struct futex_hash_bucket *hb1, *hb2;
1680 struct futex_q *this, *next;
1682 DEFINE_WAKE_Q(wake_q);
1685 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1686 if (unlikely(ret != 0))
1688 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1689 if (unlikely(ret != 0))
1692 hb1 = hash_futex(&key1);
1693 hb2 = hash_futex(&key2);
1696 double_lock_hb(hb1, hb2);
1697 op_ret = futex_atomic_op_inuser(op, uaddr2);
1698 if (unlikely(op_ret < 0)) {
1700 double_unlock_hb(hb1, hb2);
1704 * we don't get EFAULT from MMU faults if we don't have an MMU,
1705 * but we might get them from range checking
1711 if (unlikely(op_ret != -EFAULT)) {
1716 ret = fault_in_user_writeable(uaddr2);
1720 if (!(flags & FLAGS_SHARED))
1723 put_futex_key(&key2);
1724 put_futex_key(&key1);
1728 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1729 if (match_futex (&this->key, &key1)) {
1730 if (this->pi_state || this->rt_waiter) {
1734 mark_wake_futex(&wake_q, this);
1735 if (++ret >= nr_wake)
1742 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1743 if (match_futex (&this->key, &key2)) {
1744 if (this->pi_state || this->rt_waiter) {
1748 mark_wake_futex(&wake_q, this);
1749 if (++op_ret >= nr_wake2)
1757 double_unlock_hb(hb1, hb2);
1760 put_futex_key(&key2);
1762 put_futex_key(&key1);
1768 * requeue_futex() - Requeue a futex_q from one hb to another
1769 * @q: the futex_q to requeue
1770 * @hb1: the source hash_bucket
1771 * @hb2: the target hash_bucket
1772 * @key2: the new key for the requeued futex_q
1775 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1776 struct futex_hash_bucket *hb2, union futex_key *key2)
1780 * If key1 and key2 hash to the same bucket, no need to
1783 if (likely(&hb1->chain != &hb2->chain)) {
1784 plist_del(&q->list, &hb1->chain);
1785 hb_waiters_dec(hb1);
1786 hb_waiters_inc(hb2);
1787 plist_add(&q->list, &hb2->chain);
1788 q->lock_ptr = &hb2->lock;
1790 get_futex_key_refs(key2);
1795 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1797 * @key: the key of the requeue target futex
1798 * @hb: the hash_bucket of the requeue target futex
1800 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1801 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1802 * to the requeue target futex so the waiter can detect the wakeup on the right
1803 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1804 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1805 * to protect access to the pi_state to fixup the owner later. Must be called
1806 * with both q->lock_ptr and hb->lock held.
1809 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1810 struct futex_hash_bucket *hb)
1812 get_futex_key_refs(key);
1817 WARN_ON(!q->rt_waiter);
1818 q->rt_waiter = NULL;
1820 q->lock_ptr = &hb->lock;
1822 wake_up_state(q->task, TASK_NORMAL);
1826 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1827 * @pifutex: the user address of the to futex
1828 * @hb1: the from futex hash bucket, must be locked by the caller
1829 * @hb2: the to futex hash bucket, must be locked by the caller
1830 * @key1: the from futex key
1831 * @key2: the to futex key
1832 * @ps: address to store the pi_state pointer
1833 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1835 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1836 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1837 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1838 * hb1 and hb2 must be held by the caller.
1841 * - 0 - failed to acquire the lock atomically;
1842 * - >0 - acquired the lock, return value is vpid of the top_waiter
1845 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1846 struct futex_hash_bucket *hb1,
1847 struct futex_hash_bucket *hb2,
1848 union futex_key *key1, union futex_key *key2,
1849 struct futex_pi_state **ps, int set_waiters)
1851 struct futex_q *top_waiter = NULL;
1855 if (get_futex_value_locked(&curval, pifutex))
1858 if (unlikely(should_fail_futex(true)))
1862 * Find the top_waiter and determine if there are additional waiters.
1863 * If the caller intends to requeue more than 1 waiter to pifutex,
1864 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1865 * as we have means to handle the possible fault. If not, don't set
1866 * the bit unecessarily as it will force the subsequent unlock to enter
1869 top_waiter = futex_top_waiter(hb1, key1);
1871 /* There are no waiters, nothing for us to do. */
1875 /* Ensure we requeue to the expected futex. */
1876 if (!match_futex(top_waiter->requeue_pi_key, key2))
1880 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1881 * the contended case or if set_waiters is 1. The pi_state is returned
1882 * in ps in contended cases.
1884 vpid = task_pid_vnr(top_waiter->task);
1885 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1888 requeue_pi_wake_futex(top_waiter, key2, hb2);
1895 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1896 * @uaddr1: source futex user address
1897 * @flags: futex flags (FLAGS_SHARED, etc.)
1898 * @uaddr2: target futex user address
1899 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1900 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1901 * @cmpval: @uaddr1 expected value (or %NULL)
1902 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1903 * pi futex (pi to pi requeue is not supported)
1905 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1906 * uaddr2 atomically on behalf of the top waiter.
1909 * - >=0 - on success, the number of tasks requeued or woken;
1912 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1913 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1914 u32 *cmpval, int requeue_pi)
1916 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1917 int drop_count = 0, task_count = 0, ret;
1918 struct futex_pi_state *pi_state = NULL;
1919 struct futex_hash_bucket *hb1, *hb2;
1920 struct futex_q *this, *next;
1921 DEFINE_WAKE_Q(wake_q);
1923 if (nr_wake < 0 || nr_requeue < 0)
1927 * When PI not supported: return -ENOSYS if requeue_pi is true,
1928 * consequently the compiler knows requeue_pi is always false past
1929 * this point which will optimize away all the conditional code
1932 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1937 * Requeue PI only works on two distinct uaddrs. This
1938 * check is only valid for private futexes. See below.
1940 if (uaddr1 == uaddr2)
1944 * requeue_pi requires a pi_state, try to allocate it now
1945 * without any locks in case it fails.
1947 if (refill_pi_state_cache())
1950 * requeue_pi must wake as many tasks as it can, up to nr_wake
1951 * + nr_requeue, since it acquires the rt_mutex prior to
1952 * returning to userspace, so as to not leave the rt_mutex with
1953 * waiters and no owner. However, second and third wake-ups
1954 * cannot be predicted as they involve race conditions with the
1955 * first wake and a fault while looking up the pi_state. Both
1956 * pthread_cond_signal() and pthread_cond_broadcast() should
1964 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1965 if (unlikely(ret != 0))
1967 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1968 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1969 if (unlikely(ret != 0))
1973 * The check above which compares uaddrs is not sufficient for
1974 * shared futexes. We need to compare the keys:
1976 if (requeue_pi && match_futex(&key1, &key2)) {
1981 hb1 = hash_futex(&key1);
1982 hb2 = hash_futex(&key2);
1985 hb_waiters_inc(hb2);
1986 double_lock_hb(hb1, hb2);
1988 if (likely(cmpval != NULL)) {
1991 ret = get_futex_value_locked(&curval, uaddr1);
1993 if (unlikely(ret)) {
1994 double_unlock_hb(hb1, hb2);
1995 hb_waiters_dec(hb2);
1997 ret = get_user(curval, uaddr1);
2001 if (!(flags & FLAGS_SHARED))
2004 put_futex_key(&key2);
2005 put_futex_key(&key1);
2008 if (curval != *cmpval) {
2014 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2016 * Attempt to acquire uaddr2 and wake the top waiter. If we
2017 * intend to requeue waiters, force setting the FUTEX_WAITERS
2018 * bit. We force this here where we are able to easily handle
2019 * faults rather in the requeue loop below.
2021 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2022 &key2, &pi_state, nr_requeue);
2025 * At this point the top_waiter has either taken uaddr2 or is
2026 * waiting on it. If the former, then the pi_state will not
2027 * exist yet, look it up one more time to ensure we have a
2028 * reference to it. If the lock was taken, ret contains the
2029 * vpid of the top waiter task.
2030 * If the lock was not taken, we have pi_state and an initial
2031 * refcount on it. In case of an error we have nothing.
2038 * If we acquired the lock, then the user space value
2039 * of uaddr2 should be vpid. It cannot be changed by
2040 * the top waiter as it is blocked on hb2 lock if it
2041 * tries to do so. If something fiddled with it behind
2042 * our back the pi state lookup might unearth it. So
2043 * we rather use the known value than rereading and
2044 * handing potential crap to lookup_pi_state.
2046 * If that call succeeds then we have pi_state and an
2047 * initial refcount on it.
2049 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2054 /* We hold a reference on the pi state. */
2057 /* If the above failed, then pi_state is NULL */
2059 double_unlock_hb(hb1, hb2);
2060 hb_waiters_dec(hb2);
2061 put_futex_key(&key2);
2062 put_futex_key(&key1);
2063 ret = fault_in_user_writeable(uaddr2);
2069 * Two reasons for this:
2070 * - Owner is exiting and we just wait for the
2072 * - The user space value changed.
2074 double_unlock_hb(hb1, hb2);
2075 hb_waiters_dec(hb2);
2076 put_futex_key(&key2);
2077 put_futex_key(&key1);
2085 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2086 if (task_count - nr_wake >= nr_requeue)
2089 if (!match_futex(&this->key, &key1))
2093 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2094 * be paired with each other and no other futex ops.
2096 * We should never be requeueing a futex_q with a pi_state,
2097 * which is awaiting a futex_unlock_pi().
2099 if ((requeue_pi && !this->rt_waiter) ||
2100 (!requeue_pi && this->rt_waiter) ||
2107 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2108 * lock, we already woke the top_waiter. If not, it will be
2109 * woken by futex_unlock_pi().
2111 if (++task_count <= nr_wake && !requeue_pi) {
2112 mark_wake_futex(&wake_q, this);
2116 /* Ensure we requeue to the expected futex for requeue_pi. */
2117 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2123 * Requeue nr_requeue waiters and possibly one more in the case
2124 * of requeue_pi if we couldn't acquire the lock atomically.
2128 * Prepare the waiter to take the rt_mutex. Take a
2129 * refcount on the pi_state and store the pointer in
2130 * the futex_q object of the waiter.
2132 get_pi_state(pi_state);
2133 this->pi_state = pi_state;
2134 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2139 * We got the lock. We do neither drop the
2140 * refcount on pi_state nor clear
2141 * this->pi_state because the waiter needs the
2142 * pi_state for cleaning up the user space
2143 * value. It will drop the refcount after
2146 requeue_pi_wake_futex(this, &key2, hb2);
2151 * rt_mutex_start_proxy_lock() detected a
2152 * potential deadlock when we tried to queue
2153 * that waiter. Drop the pi_state reference
2154 * which we took above and remove the pointer
2155 * to the state from the waiters futex_q
2158 this->pi_state = NULL;
2159 put_pi_state(pi_state);
2161 * We stop queueing more waiters and let user
2162 * space deal with the mess.
2167 requeue_futex(this, hb1, hb2, &key2);
2172 * We took an extra initial reference to the pi_state either
2173 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2174 * need to drop it here again.
2176 put_pi_state(pi_state);
2179 double_unlock_hb(hb1, hb2);
2181 hb_waiters_dec(hb2);
2184 * drop_futex_key_refs() must be called outside the spinlocks. During
2185 * the requeue we moved futex_q's from the hash bucket at key1 to the
2186 * one at key2 and updated their key pointer. We no longer need to
2187 * hold the references to key1.
2189 while (--drop_count >= 0)
2190 drop_futex_key_refs(&key1);
2193 put_futex_key(&key2);
2195 put_futex_key(&key1);
2197 return ret ? ret : task_count;
2200 /* The key must be already stored in q->key. */
2201 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2202 __acquires(&hb->lock)
2204 struct futex_hash_bucket *hb;
2206 hb = hash_futex(&q->key);
2209 * Increment the counter before taking the lock so that
2210 * a potential waker won't miss a to-be-slept task that is
2211 * waiting for the spinlock. This is safe as all queue_lock()
2212 * users end up calling queue_me(). Similarly, for housekeeping,
2213 * decrement the counter at queue_unlock() when some error has
2214 * occurred and we don't end up adding the task to the list.
2218 q->lock_ptr = &hb->lock;
2220 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2225 queue_unlock(struct futex_hash_bucket *hb)
2226 __releases(&hb->lock)
2228 spin_unlock(&hb->lock);
2232 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2237 * The priority used to register this element is
2238 * - either the real thread-priority for the real-time threads
2239 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2240 * - or MAX_RT_PRIO for non-RT threads.
2241 * Thus, all RT-threads are woken first in priority order, and
2242 * the others are woken last, in FIFO order.
2244 prio = min(current->normal_prio, MAX_RT_PRIO);
2246 plist_node_init(&q->list, prio);
2247 plist_add(&q->list, &hb->chain);
2252 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2253 * @q: The futex_q to enqueue
2254 * @hb: The destination hash bucket
2256 * The hb->lock must be held by the caller, and is released here. A call to
2257 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2258 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2259 * or nothing if the unqueue is done as part of the wake process and the unqueue
2260 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2263 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2264 __releases(&hb->lock)
2267 spin_unlock(&hb->lock);
2271 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2272 * @q: The futex_q to unqueue
2274 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2275 * be paired with exactly one earlier call to queue_me().
2278 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2279 * - 0 - if the futex_q was already removed by the waking thread
2281 static int unqueue_me(struct futex_q *q)
2283 spinlock_t *lock_ptr;
2286 /* In the common case we don't take the spinlock, which is nice. */
2289 * q->lock_ptr can change between this read and the following spin_lock.
2290 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2291 * optimizing lock_ptr out of the logic below.
2293 lock_ptr = READ_ONCE(q->lock_ptr);
2294 if (lock_ptr != NULL) {
2295 spin_lock(lock_ptr);
2297 * q->lock_ptr can change between reading it and
2298 * spin_lock(), causing us to take the wrong lock. This
2299 * corrects the race condition.
2301 * Reasoning goes like this: if we have the wrong lock,
2302 * q->lock_ptr must have changed (maybe several times)
2303 * between reading it and the spin_lock(). It can
2304 * change again after the spin_lock() but only if it was
2305 * already changed before the spin_lock(). It cannot,
2306 * however, change back to the original value. Therefore
2307 * we can detect whether we acquired the correct lock.
2309 if (unlikely(lock_ptr != q->lock_ptr)) {
2310 spin_unlock(lock_ptr);
2315 BUG_ON(q->pi_state);
2317 spin_unlock(lock_ptr);
2321 drop_futex_key_refs(&q->key);
2326 * PI futexes can not be requeued and must remove themself from the
2327 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2330 static void unqueue_me_pi(struct futex_q *q)
2331 __releases(q->lock_ptr)
2335 BUG_ON(!q->pi_state);
2336 put_pi_state(q->pi_state);
2339 spin_unlock(q->lock_ptr);
2342 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2343 struct task_struct *argowner)
2345 struct futex_pi_state *pi_state = q->pi_state;
2346 u32 uval, uninitialized_var(curval), newval;
2347 struct task_struct *oldowner, *newowner;
2351 lockdep_assert_held(q->lock_ptr);
2353 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2355 oldowner = pi_state->owner;
2358 * We are here because either:
2360 * - we stole the lock and pi_state->owner needs updating to reflect
2361 * that (@argowner == current),
2365 * - someone stole our lock and we need to fix things to point to the
2366 * new owner (@argowner == NULL).
2368 * Either way, we have to replace the TID in the user space variable.
2369 * This must be atomic as we have to preserve the owner died bit here.
2371 * Note: We write the user space value _before_ changing the pi_state
2372 * because we can fault here. Imagine swapped out pages or a fork
2373 * that marked all the anonymous memory readonly for cow.
2375 * Modifying pi_state _before_ the user space value would leave the
2376 * pi_state in an inconsistent state when we fault here, because we
2377 * need to drop the locks to handle the fault. This might be observed
2378 * in the PID check in lookup_pi_state.
2382 if (oldowner != current) {
2384 * We raced against a concurrent self; things are
2385 * already fixed up. Nothing to do.
2391 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2392 /* We got the lock after all, nothing to fix. */
2398 * Since we just failed the trylock; there must be an owner.
2400 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2403 WARN_ON_ONCE(argowner != current);
2404 if (oldowner == current) {
2406 * We raced against a concurrent self; things are
2407 * already fixed up. Nothing to do.
2412 newowner = argowner;
2415 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2417 if (!pi_state->owner)
2418 newtid |= FUTEX_OWNER_DIED;
2420 if (get_futex_value_locked(&uval, uaddr))
2424 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2426 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2434 * We fixed up user space. Now we need to fix the pi_state
2437 if (pi_state->owner != NULL) {
2438 raw_spin_lock(&pi_state->owner->pi_lock);
2439 WARN_ON(list_empty(&pi_state->list));
2440 list_del_init(&pi_state->list);
2441 raw_spin_unlock(&pi_state->owner->pi_lock);
2444 pi_state->owner = newowner;
2446 raw_spin_lock(&newowner->pi_lock);
2447 WARN_ON(!list_empty(&pi_state->list));
2448 list_add(&pi_state->list, &newowner->pi_state_list);
2449 raw_spin_unlock(&newowner->pi_lock);
2450 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2455 * To handle the page fault we need to drop the locks here. That gives
2456 * the other task (either the highest priority waiter itself or the
2457 * task which stole the rtmutex) the chance to try the fixup of the
2458 * pi_state. So once we are back from handling the fault we need to
2459 * check the pi_state after reacquiring the locks and before trying to
2460 * do another fixup. When the fixup has been done already we simply
2463 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2464 * drop hb->lock since the caller owns the hb -> futex_q relation.
2465 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2468 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2469 spin_unlock(q->lock_ptr);
2471 ret = fault_in_user_writeable(uaddr);
2473 spin_lock(q->lock_ptr);
2474 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2477 * Check if someone else fixed it for us:
2479 if (pi_state->owner != oldowner) {
2490 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2494 static long futex_wait_restart(struct restart_block *restart);
2497 * fixup_owner() - Post lock pi_state and corner case management
2498 * @uaddr: user address of the futex
2499 * @q: futex_q (contains pi_state and access to the rt_mutex)
2500 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2502 * After attempting to lock an rt_mutex, this function is called to cleanup
2503 * the pi_state owner as well as handle race conditions that may allow us to
2504 * acquire the lock. Must be called with the hb lock held.
2507 * - 1 - success, lock taken;
2508 * - 0 - success, lock not taken;
2509 * - <0 - on error (-EFAULT)
2511 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2517 * Got the lock. We might not be the anticipated owner if we
2518 * did a lock-steal - fix up the PI-state in that case:
2520 * Speculative pi_state->owner read (we don't hold wait_lock);
2521 * since we own the lock pi_state->owner == current is the
2522 * stable state, anything else needs more attention.
2524 if (q->pi_state->owner != current)
2525 ret = fixup_pi_state_owner(uaddr, q, current);
2530 * If we didn't get the lock; check if anybody stole it from us. In
2531 * that case, we need to fix up the uval to point to them instead of
2532 * us, otherwise bad things happen. [10]
2534 * Another speculative read; pi_state->owner == current is unstable
2535 * but needs our attention.
2537 if (q->pi_state->owner == current) {
2538 ret = fixup_pi_state_owner(uaddr, q, NULL);
2543 * Paranoia check. If we did not take the lock, then we should not be
2544 * the owner of the rt_mutex.
2546 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2547 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2548 "pi-state %p\n", ret,
2549 q->pi_state->pi_mutex.owner,
2550 q->pi_state->owner);
2554 return ret ? ret : locked;
2558 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2559 * @hb: the futex hash bucket, must be locked by the caller
2560 * @q: the futex_q to queue up on
2561 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2563 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2564 struct hrtimer_sleeper *timeout)
2567 * The task state is guaranteed to be set before another task can
2568 * wake it. set_current_state() is implemented using smp_store_mb() and
2569 * queue_me() calls spin_unlock() upon completion, both serializing
2570 * access to the hash list and forcing another memory barrier.
2572 set_current_state(TASK_INTERRUPTIBLE);
2577 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2580 * If we have been removed from the hash list, then another task
2581 * has tried to wake us, and we can skip the call to schedule().
2583 if (likely(!plist_node_empty(&q->list))) {
2585 * If the timer has already expired, current will already be
2586 * flagged for rescheduling. Only call schedule if there
2587 * is no timeout, or if it has yet to expire.
2589 if (!timeout || timeout->task)
2590 freezable_schedule();
2592 __set_current_state(TASK_RUNNING);
2596 * futex_wait_setup() - Prepare to wait on a futex
2597 * @uaddr: the futex userspace address
2598 * @val: the expected value
2599 * @flags: futex flags (FLAGS_SHARED, etc.)
2600 * @q: the associated futex_q
2601 * @hb: storage for hash_bucket pointer to be returned to caller
2603 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2604 * compare it with the expected value. Handle atomic faults internally.
2605 * Return with the hb lock held and a q.key reference on success, and unlocked
2606 * with no q.key reference on failure.
2609 * - 0 - uaddr contains val and hb has been locked;
2610 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2612 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2613 struct futex_q *q, struct futex_hash_bucket **hb)
2619 * Access the page AFTER the hash-bucket is locked.
2620 * Order is important:
2622 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2623 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2625 * The basic logical guarantee of a futex is that it blocks ONLY
2626 * if cond(var) is known to be true at the time of blocking, for
2627 * any cond. If we locked the hash-bucket after testing *uaddr, that
2628 * would open a race condition where we could block indefinitely with
2629 * cond(var) false, which would violate the guarantee.
2631 * On the other hand, we insert q and release the hash-bucket only
2632 * after testing *uaddr. This guarantees that futex_wait() will NOT
2633 * absorb a wakeup if *uaddr does not match the desired values
2634 * while the syscall executes.
2637 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2638 if (unlikely(ret != 0))
2642 *hb = queue_lock(q);
2644 ret = get_futex_value_locked(&uval, uaddr);
2649 ret = get_user(uval, uaddr);
2653 if (!(flags & FLAGS_SHARED))
2656 put_futex_key(&q->key);
2667 put_futex_key(&q->key);
2671 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2672 ktime_t *abs_time, u32 bitset)
2674 struct hrtimer_sleeper timeout, *to = NULL;
2675 struct restart_block *restart;
2676 struct futex_hash_bucket *hb;
2677 struct futex_q q = futex_q_init;
2687 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2688 CLOCK_REALTIME : CLOCK_MONOTONIC,
2690 hrtimer_init_sleeper(to, current);
2691 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2692 current->timer_slack_ns);
2697 * Prepare to wait on uaddr. On success, holds hb lock and increments
2700 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2704 /* queue_me and wait for wakeup, timeout, or a signal. */
2705 futex_wait_queue_me(hb, &q, to);
2707 /* If we were woken (and unqueued), we succeeded, whatever. */
2709 /* unqueue_me() drops q.key ref */
2710 if (!unqueue_me(&q))
2713 if (to && !to->task)
2717 * We expect signal_pending(current), but we might be the
2718 * victim of a spurious wakeup as well.
2720 if (!signal_pending(current))
2727 restart = ¤t->restart_block;
2728 restart->fn = futex_wait_restart;
2729 restart->futex.uaddr = uaddr;
2730 restart->futex.val = val;
2731 restart->futex.time = *abs_time;
2732 restart->futex.bitset = bitset;
2733 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2735 ret = -ERESTART_RESTARTBLOCK;
2739 hrtimer_cancel(&to->timer);
2740 destroy_hrtimer_on_stack(&to->timer);
2746 static long futex_wait_restart(struct restart_block *restart)
2748 u32 __user *uaddr = restart->futex.uaddr;
2749 ktime_t t, *tp = NULL;
2751 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2752 t = restart->futex.time;
2755 restart->fn = do_no_restart_syscall;
2757 return (long)futex_wait(uaddr, restart->futex.flags,
2758 restart->futex.val, tp, restart->futex.bitset);
2763 * Userspace tried a 0 -> TID atomic transition of the futex value
2764 * and failed. The kernel side here does the whole locking operation:
2765 * if there are waiters then it will block as a consequence of relying
2766 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2767 * a 0 value of the futex too.).
2769 * Also serves as futex trylock_pi()'ing, and due semantics.
2771 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2772 ktime_t *time, int trylock)
2774 struct hrtimer_sleeper timeout, *to = NULL;
2775 struct futex_pi_state *pi_state = NULL;
2776 struct rt_mutex_waiter rt_waiter;
2777 struct futex_hash_bucket *hb;
2778 struct futex_q q = futex_q_init;
2781 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2784 if (refill_pi_state_cache())
2789 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2791 hrtimer_init_sleeper(to, current);
2792 hrtimer_set_expires(&to->timer, *time);
2796 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2797 if (unlikely(ret != 0))
2801 hb = queue_lock(&q);
2803 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2804 if (unlikely(ret)) {
2806 * Atomic work succeeded and we got the lock,
2807 * or failed. Either way, we do _not_ block.
2811 /* We got the lock. */
2813 goto out_unlock_put_key;
2818 * Two reasons for this:
2819 * - Task is exiting and we just wait for the
2821 * - The user space value changed.
2824 put_futex_key(&q.key);
2828 goto out_unlock_put_key;
2832 WARN_ON(!q.pi_state);
2835 * Only actually queue now that the atomic ops are done:
2840 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2841 /* Fixup the trylock return value: */
2842 ret = ret ? 0 : -EWOULDBLOCK;
2846 rt_mutex_init_waiter(&rt_waiter);
2849 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2850 * hold it while doing rt_mutex_start_proxy(), because then it will
2851 * include hb->lock in the blocking chain, even through we'll not in
2852 * fact hold it while blocking. This will lead it to report -EDEADLK
2853 * and BUG when futex_unlock_pi() interleaves with this.
2855 * Therefore acquire wait_lock while holding hb->lock, but drop the
2856 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2857 * serializes against futex_unlock_pi() as that does the exact same
2858 * lock handoff sequence.
2860 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2861 spin_unlock(q.lock_ptr);
2862 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2863 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2869 spin_lock(q.lock_ptr);
2875 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2877 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2879 spin_lock(q.lock_ptr);
2881 * If we failed to acquire the lock (signal/timeout), we must
2882 * first acquire the hb->lock before removing the lock from the
2883 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2884 * wait lists consistent.
2886 * In particular; it is important that futex_unlock_pi() can not
2887 * observe this inconsistency.
2889 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2894 * Fixup the pi_state owner and possibly acquire the lock if we
2897 res = fixup_owner(uaddr, &q, !ret);
2899 * If fixup_owner() returned an error, proprogate that. If it acquired
2900 * the lock, clear our -ETIMEDOUT or -EINTR.
2903 ret = (res < 0) ? res : 0;
2906 * If fixup_owner() faulted and was unable to handle the fault, unlock
2907 * it and return the fault to userspace.
2909 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2910 pi_state = q.pi_state;
2911 get_pi_state(pi_state);
2914 /* Unqueue and drop the lock */
2918 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2919 put_pi_state(pi_state);
2928 put_futex_key(&q.key);
2931 hrtimer_cancel(&to->timer);
2932 destroy_hrtimer_on_stack(&to->timer);
2934 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2939 ret = fault_in_user_writeable(uaddr);
2943 if (!(flags & FLAGS_SHARED))
2946 put_futex_key(&q.key);
2951 * Userspace attempted a TID -> 0 atomic transition, and failed.
2952 * This is the in-kernel slowpath: we look up the PI state (if any),
2953 * and do the rt-mutex unlock.
2955 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2957 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2958 union futex_key key = FUTEX_KEY_INIT;
2959 struct futex_hash_bucket *hb;
2960 struct futex_q *top_waiter;
2963 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2967 if (get_user(uval, uaddr))
2970 * We release only a lock we actually own:
2972 if ((uval & FUTEX_TID_MASK) != vpid)
2975 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2979 hb = hash_futex(&key);
2980 spin_lock(&hb->lock);
2983 * Check waiters first. We do not trust user space values at
2984 * all and we at least want to know if user space fiddled
2985 * with the futex value instead of blindly unlocking.
2987 top_waiter = futex_top_waiter(hb, &key);
2989 struct futex_pi_state *pi_state = top_waiter->pi_state;
2996 * If current does not own the pi_state then the futex is
2997 * inconsistent and user space fiddled with the futex value.
2999 if (pi_state->owner != current)
3002 get_pi_state(pi_state);
3004 * By taking wait_lock while still holding hb->lock, we ensure
3005 * there is no point where we hold neither; and therefore
3006 * wake_futex_pi() must observe a state consistent with what we
3009 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3010 spin_unlock(&hb->lock);
3012 /* drops pi_state->pi_mutex.wait_lock */
3013 ret = wake_futex_pi(uaddr, uval, pi_state);
3015 put_pi_state(pi_state);
3018 * Success, we're done! No tricky corner cases.
3023 * The atomic access to the futex value generated a
3024 * pagefault, so retry the user-access and the wakeup:
3029 * A unconditional UNLOCK_PI op raced against a waiter
3030 * setting the FUTEX_WAITERS bit. Try again.
3032 if (ret == -EAGAIN) {
3033 put_futex_key(&key);
3037 * wake_futex_pi has detected invalid state. Tell user
3044 * We have no kernel internal state, i.e. no waiters in the
3045 * kernel. Waiters which are about to queue themselves are stuck
3046 * on hb->lock. So we can safely ignore them. We do neither
3047 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3050 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
3051 spin_unlock(&hb->lock);
3056 * If uval has changed, let user space handle it.
3058 ret = (curval == uval) ? 0 : -EAGAIN;
3061 spin_unlock(&hb->lock);
3063 put_futex_key(&key);
3067 put_futex_key(&key);
3069 ret = fault_in_user_writeable(uaddr);
3077 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3078 * @hb: the hash_bucket futex_q was original enqueued on
3079 * @q: the futex_q woken while waiting to be requeued
3080 * @key2: the futex_key of the requeue target futex
3081 * @timeout: the timeout associated with the wait (NULL if none)
3083 * Detect if the task was woken on the initial futex as opposed to the requeue
3084 * target futex. If so, determine if it was a timeout or a signal that caused
3085 * the wakeup and return the appropriate error code to the caller. Must be
3086 * called with the hb lock held.
3089 * - 0 = no early wakeup detected;
3090 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3093 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3094 struct futex_q *q, union futex_key *key2,
3095 struct hrtimer_sleeper *timeout)
3100 * With the hb lock held, we avoid races while we process the wakeup.
3101 * We only need to hold hb (and not hb2) to ensure atomicity as the
3102 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3103 * It can't be requeued from uaddr2 to something else since we don't
3104 * support a PI aware source futex for requeue.
3106 if (!match_futex(&q->key, key2)) {
3107 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3109 * We were woken prior to requeue by a timeout or a signal.
3110 * Unqueue the futex_q and determine which it was.
3112 plist_del(&q->list, &hb->chain);
3115 /* Handle spurious wakeups gracefully */
3117 if (timeout && !timeout->task)
3119 else if (signal_pending(current))
3120 ret = -ERESTARTNOINTR;
3126 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3127 * @uaddr: the futex we initially wait on (non-pi)
3128 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3129 * the same type, no requeueing from private to shared, etc.
3130 * @val: the expected value of uaddr
3131 * @abs_time: absolute timeout
3132 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3133 * @uaddr2: the pi futex we will take prior to returning to user-space
3135 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3136 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3137 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3138 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3139 * without one, the pi logic would not know which task to boost/deboost, if
3140 * there was a need to.
3142 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3143 * via the following--
3144 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3145 * 2) wakeup on uaddr2 after a requeue
3149 * If 3, cleanup and return -ERESTARTNOINTR.
3151 * If 2, we may then block on trying to take the rt_mutex and return via:
3152 * 5) successful lock
3155 * 8) other lock acquisition failure
3157 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3159 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3165 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3166 u32 val, ktime_t *abs_time, u32 bitset,
3169 struct hrtimer_sleeper timeout, *to = NULL;
3170 struct futex_pi_state *pi_state = NULL;
3171 struct rt_mutex_waiter rt_waiter;
3172 struct futex_hash_bucket *hb;
3173 union futex_key key2 = FUTEX_KEY_INIT;
3174 struct futex_q q = futex_q_init;
3177 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3180 if (uaddr == uaddr2)
3188 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3189 CLOCK_REALTIME : CLOCK_MONOTONIC,
3191 hrtimer_init_sleeper(to, current);
3192 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3193 current->timer_slack_ns);
3197 * The waiter is allocated on our stack, manipulated by the requeue
3198 * code while we sleep on uaddr.
3200 rt_mutex_init_waiter(&rt_waiter);
3202 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3203 if (unlikely(ret != 0))
3207 q.rt_waiter = &rt_waiter;
3208 q.requeue_pi_key = &key2;
3211 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3214 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3219 * The check above which compares uaddrs is not sufficient for
3220 * shared futexes. We need to compare the keys:
3222 if (match_futex(&q.key, &key2)) {
3228 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3229 futex_wait_queue_me(hb, &q, to);
3231 spin_lock(&hb->lock);
3232 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3233 spin_unlock(&hb->lock);
3238 * In order for us to be here, we know our q.key == key2, and since
3239 * we took the hb->lock above, we also know that futex_requeue() has
3240 * completed and we no longer have to concern ourselves with a wakeup
3241 * race with the atomic proxy lock acquisition by the requeue code. The
3242 * futex_requeue dropped our key1 reference and incremented our key2
3246 /* Check if the requeue code acquired the second futex for us. */
3249 * Got the lock. We might not be the anticipated owner if we
3250 * did a lock-steal - fix up the PI-state in that case.
3252 if (q.pi_state && (q.pi_state->owner != current)) {
3253 spin_lock(q.lock_ptr);
3254 ret = fixup_pi_state_owner(uaddr2, &q, current);
3255 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3256 pi_state = q.pi_state;
3257 get_pi_state(pi_state);
3260 * Drop the reference to the pi state which
3261 * the requeue_pi() code acquired for us.
3263 put_pi_state(q.pi_state);
3264 spin_unlock(q.lock_ptr);
3267 struct rt_mutex *pi_mutex;
3270 * We have been woken up by futex_unlock_pi(), a timeout, or a
3271 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3274 WARN_ON(!q.pi_state);
3275 pi_mutex = &q.pi_state->pi_mutex;
3276 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3278 spin_lock(q.lock_ptr);
3279 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3282 debug_rt_mutex_free_waiter(&rt_waiter);
3284 * Fixup the pi_state owner and possibly acquire the lock if we
3287 res = fixup_owner(uaddr2, &q, !ret);
3289 * If fixup_owner() returned an error, proprogate that. If it
3290 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3293 ret = (res < 0) ? res : 0;
3296 * If fixup_pi_state_owner() faulted and was unable to handle
3297 * the fault, unlock the rt_mutex and return the fault to
3300 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3301 pi_state = q.pi_state;
3302 get_pi_state(pi_state);
3305 /* Unqueue and drop the lock. */
3310 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3311 put_pi_state(pi_state);
3314 if (ret == -EINTR) {
3316 * We've already been requeued, but cannot restart by calling
3317 * futex_lock_pi() directly. We could restart this syscall, but
3318 * it would detect that the user space "val" changed and return
3319 * -EWOULDBLOCK. Save the overhead of the restart and return
3320 * -EWOULDBLOCK directly.
3326 put_futex_key(&q.key);
3328 put_futex_key(&key2);
3332 hrtimer_cancel(&to->timer);
3333 destroy_hrtimer_on_stack(&to->timer);
3339 * Support for robust futexes: the kernel cleans up held futexes at
3342 * Implementation: user-space maintains a per-thread list of locks it
3343 * is holding. Upon do_exit(), the kernel carefully walks this list,
3344 * and marks all locks that are owned by this thread with the
3345 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3346 * always manipulated with the lock held, so the list is private and
3347 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3348 * field, to allow the kernel to clean up if the thread dies after
3349 * acquiring the lock, but just before it could have added itself to
3350 * the list. There can only be one such pending lock.
3354 * sys_set_robust_list() - Set the robust-futex list head of a task
3355 * @head: pointer to the list-head
3356 * @len: length of the list-head, as userspace expects
3358 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3361 if (!futex_cmpxchg_enabled)
3364 * The kernel knows only one size for now:
3366 if (unlikely(len != sizeof(*head)))
3369 current->robust_list = head;
3375 * sys_get_robust_list() - Get the robust-futex list head of a task
3376 * @pid: pid of the process [zero for current task]
3377 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3378 * @len_ptr: pointer to a length field, the kernel fills in the header size
3380 SYSCALL_DEFINE3(get_robust_list, int, pid,
3381 struct robust_list_head __user * __user *, head_ptr,
3382 size_t __user *, len_ptr)
3384 struct robust_list_head __user *head;
3386 struct task_struct *p;
3388 if (!futex_cmpxchg_enabled)
3397 p = find_task_by_vpid(pid);
3403 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3406 head = p->robust_list;
3409 if (put_user(sizeof(*head), len_ptr))
3411 return put_user(head, head_ptr);
3420 * Process a futex-list entry, check whether it's owned by the
3421 * dying task, and do notification if so:
3423 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3425 u32 uval, uninitialized_var(nval), mval;
3428 if (get_user(uval, uaddr))
3431 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3433 * Ok, this dying thread is truly holding a futex
3434 * of interest. Set the OWNER_DIED bit atomically
3435 * via cmpxchg, and if the value had FUTEX_WAITERS
3436 * set, wake up a waiter (if any). (We have to do a
3437 * futex_wake() even if OWNER_DIED is already set -
3438 * to handle the rare but possible case of recursive
3439 * thread-death.) The rest of the cleanup is done in
3442 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3444 * We are not holding a lock here, but we want to have
3445 * the pagefault_disable/enable() protection because
3446 * we want to handle the fault gracefully. If the
3447 * access fails we try to fault in the futex with R/W
3448 * verification via get_user_pages. get_user() above
3449 * does not guarantee R/W access. If that fails we
3450 * give up and leave the futex locked.
3452 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3453 if (fault_in_user_writeable(uaddr))
3461 * Wake robust non-PI futexes here. The wakeup of
3462 * PI futexes happens in exit_pi_state():
3464 if (!pi && (uval & FUTEX_WAITERS))
3465 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3471 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3473 static inline int fetch_robust_entry(struct robust_list __user **entry,
3474 struct robust_list __user * __user *head,
3477 unsigned long uentry;
3479 if (get_user(uentry, (unsigned long __user *)head))
3482 *entry = (void __user *)(uentry & ~1UL);
3489 * Walk curr->robust_list (very carefully, it's a userspace list!)
3490 * and mark any locks found there dead, and notify any waiters.
3492 * We silently return on any sign of list-walking problem.
3494 void exit_robust_list(struct task_struct *curr)
3496 struct robust_list_head __user *head = curr->robust_list;
3497 struct robust_list __user *entry, *next_entry, *pending;
3498 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3499 unsigned int uninitialized_var(next_pi);
3500 unsigned long futex_offset;
3503 if (!futex_cmpxchg_enabled)
3507 * Fetch the list head (which was registered earlier, via
3508 * sys_set_robust_list()):
3510 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3513 * Fetch the relative futex offset:
3515 if (get_user(futex_offset, &head->futex_offset))
3518 * Fetch any possibly pending lock-add first, and handle it
3521 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3524 next_entry = NULL; /* avoid warning with gcc */
3525 while (entry != &head->list) {
3527 * Fetch the next entry in the list before calling
3528 * handle_futex_death:
3530 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3532 * A pending lock might already be on the list, so
3533 * don't process it twice:
3535 if (entry != pending)
3536 if (handle_futex_death((void __user *)entry + futex_offset,
3544 * Avoid excessively long or circular lists:
3553 handle_futex_death((void __user *)pending + futex_offset,
3557 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3558 u32 __user *uaddr2, u32 val2, u32 val3)
3560 int cmd = op & FUTEX_CMD_MASK;
3561 unsigned int flags = 0;
3563 if (!(op & FUTEX_PRIVATE_FLAG))
3564 flags |= FLAGS_SHARED;
3566 if (op & FUTEX_CLOCK_REALTIME) {
3567 flags |= FLAGS_CLOCKRT;
3568 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3569 cmd != FUTEX_WAIT_REQUEUE_PI)
3575 case FUTEX_UNLOCK_PI:
3576 case FUTEX_TRYLOCK_PI:
3577 case FUTEX_WAIT_REQUEUE_PI:
3578 case FUTEX_CMP_REQUEUE_PI:
3579 if (!futex_cmpxchg_enabled)
3585 val3 = FUTEX_BITSET_MATCH_ANY;
3587 case FUTEX_WAIT_BITSET:
3588 return futex_wait(uaddr, flags, val, timeout, val3);
3590 val3 = FUTEX_BITSET_MATCH_ANY;
3592 case FUTEX_WAKE_BITSET:
3593 return futex_wake(uaddr, flags, val, val3);
3595 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3596 case FUTEX_CMP_REQUEUE:
3597 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3599 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3601 return futex_lock_pi(uaddr, flags, timeout, 0);
3602 case FUTEX_UNLOCK_PI:
3603 return futex_unlock_pi(uaddr, flags);
3604 case FUTEX_TRYLOCK_PI:
3605 return futex_lock_pi(uaddr, flags, NULL, 1);
3606 case FUTEX_WAIT_REQUEUE_PI:
3607 val3 = FUTEX_BITSET_MATCH_ANY;
3608 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3610 case FUTEX_CMP_REQUEUE_PI:
3611 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3617 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3618 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3621 struct timespec64 ts;
3622 ktime_t t, *tp = NULL;
3624 int cmd = op & FUTEX_CMD_MASK;
3626 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3627 cmd == FUTEX_WAIT_BITSET ||
3628 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3629 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3631 if (get_timespec64(&ts, utime))
3633 if (!timespec64_valid(&ts))
3636 t = timespec64_to_ktime(ts);
3637 if (cmd == FUTEX_WAIT)
3638 t = ktime_add_safe(ktime_get(), t);
3642 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3643 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3645 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3646 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3647 val2 = (u32) (unsigned long) utime;
3649 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3652 #ifdef CONFIG_COMPAT
3654 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3657 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3658 compat_uptr_t __user *head, unsigned int *pi)
3660 if (get_user(*uentry, head))
3663 *entry = compat_ptr((*uentry) & ~1);
3664 *pi = (unsigned int)(*uentry) & 1;
3669 static void __user *futex_uaddr(struct robust_list __user *entry,
3670 compat_long_t futex_offset)
3672 compat_uptr_t base = ptr_to_compat(entry);
3673 void __user *uaddr = compat_ptr(base + futex_offset);
3679 * Walk curr->robust_list (very carefully, it's a userspace list!)
3680 * and mark any locks found there dead, and notify any waiters.
3682 * We silently return on any sign of list-walking problem.
3684 void compat_exit_robust_list(struct task_struct *curr)
3686 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3687 struct robust_list __user *entry, *next_entry, *pending;
3688 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3689 unsigned int uninitialized_var(next_pi);
3690 compat_uptr_t uentry, next_uentry, upending;
3691 compat_long_t futex_offset;
3694 if (!futex_cmpxchg_enabled)
3698 * Fetch the list head (which was registered earlier, via
3699 * sys_set_robust_list()):
3701 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3704 * Fetch the relative futex offset:
3706 if (get_user(futex_offset, &head->futex_offset))
3709 * Fetch any possibly pending lock-add first, and handle it
3712 if (compat_fetch_robust_entry(&upending, &pending,
3713 &head->list_op_pending, &pip))
3716 next_entry = NULL; /* avoid warning with gcc */
3717 while (entry != (struct robust_list __user *) &head->list) {
3719 * Fetch the next entry in the list before calling
3720 * handle_futex_death:
3722 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3723 (compat_uptr_t __user *)&entry->next, &next_pi);
3725 * A pending lock might already be on the list, so
3726 * dont process it twice:
3728 if (entry != pending) {
3729 void __user *uaddr = futex_uaddr(entry, futex_offset);
3731 if (handle_futex_death(uaddr, curr, pi))
3736 uentry = next_uentry;
3740 * Avoid excessively long or circular lists:
3748 void __user *uaddr = futex_uaddr(pending, futex_offset);
3750 handle_futex_death(uaddr, curr, pip);
3754 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3755 struct compat_robust_list_head __user *, head,
3758 if (!futex_cmpxchg_enabled)
3761 if (unlikely(len != sizeof(*head)))
3764 current->compat_robust_list = head;
3769 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3770 compat_uptr_t __user *, head_ptr,
3771 compat_size_t __user *, len_ptr)
3773 struct compat_robust_list_head __user *head;
3775 struct task_struct *p;
3777 if (!futex_cmpxchg_enabled)
3786 p = find_task_by_vpid(pid);
3792 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3795 head = p->compat_robust_list;
3798 if (put_user(sizeof(*head), len_ptr))
3800 return put_user(ptr_to_compat(head), head_ptr);
3807 #endif /* CONFIG_COMPAT */
3809 #ifdef CONFIG_COMPAT_32BIT_TIME
3810 COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3811 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3814 struct timespec64 ts;
3815 ktime_t t, *tp = NULL;
3817 int cmd = op & FUTEX_CMD_MASK;
3819 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3820 cmd == FUTEX_WAIT_BITSET ||
3821 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3822 if (get_old_timespec32(&ts, utime))
3824 if (!timespec64_valid(&ts))
3827 t = timespec64_to_ktime(ts);
3828 if (cmd == FUTEX_WAIT)
3829 t = ktime_add_safe(ktime_get(), t);
3832 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3833 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3834 val2 = (int) (unsigned long) utime;
3836 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3838 #endif /* CONFIG_COMPAT_32BIT_TIME */
3840 static void __init futex_detect_cmpxchg(void)
3842 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3846 * This will fail and we want it. Some arch implementations do
3847 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3848 * functionality. We want to know that before we call in any
3849 * of the complex code paths. Also we want to prevent
3850 * registration of robust lists in that case. NULL is
3851 * guaranteed to fault and we get -EFAULT on functional
3852 * implementation, the non-functional ones will return
3855 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3856 futex_cmpxchg_enabled = 1;
3860 static int __init futex_init(void)
3862 unsigned int futex_shift;
3865 #if CONFIG_BASE_SMALL
3866 futex_hashsize = 16;
3868 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3871 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3873 futex_hashsize < 256 ? HASH_SMALL : 0,
3875 futex_hashsize, futex_hashsize);
3876 futex_hashsize = 1UL << futex_shift;
3878 futex_detect_cmpxchg();
3880 for (i = 0; i < futex_hashsize; i++) {
3881 atomic_set(&futex_queues[i].waiters, 0);
3882 plist_head_init(&futex_queues[i].chain);
3883 spin_lock_init(&futex_queues[i].lock);
3888 core_initcall(futex_init);