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/bootmem.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 */
345 static void compat_exit_robust_list(struct task_struct *curr);
347 static inline void compat_exit_robust_list(struct task_struct *curr) { }
350 static inline void futex_get_mm(union futex_key *key)
352 mmgrab(key->private.mm);
354 * Ensure futex_get_mm() implies a full barrier such that
355 * get_futex_key() implies a full barrier. This is relied upon
356 * as smp_mb(); (B), see the ordering comment above.
358 smp_mb__after_atomic();
362 * Reflects a new waiter being added to the waitqueue.
364 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
367 atomic_inc(&hb->waiters);
369 * Full barrier (A), see the ordering comment above.
371 smp_mb__after_atomic();
376 * Reflects a waiter being removed from the waitqueue by wakeup
379 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
382 atomic_dec(&hb->waiters);
386 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
389 return atomic_read(&hb->waiters);
396 * hash_futex - Return the hash bucket in the global hash
397 * @key: Pointer to the futex key for which the hash is calculated
399 * We hash on the keys returned from get_futex_key (see below) and return the
400 * corresponding hash bucket in the global hash.
402 static struct futex_hash_bucket *hash_futex(union futex_key *key)
404 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
407 return &futex_queues[hash & (futex_hashsize - 1)];
412 * match_futex - Check whether two futex keys are equal
413 * @key1: Pointer to key1
414 * @key2: Pointer to key2
416 * Return 1 if two futex_keys are equal, 0 otherwise.
418 static inline int match_futex(union futex_key *key1, union futex_key *key2)
421 && key1->both.word == key2->both.word
422 && key1->both.ptr == key2->both.ptr
423 && key1->both.offset == key2->both.offset);
427 * Take a reference to the resource addressed by a key.
428 * Can be called while holding spinlocks.
431 static void get_futex_key_refs(union futex_key *key)
437 * On MMU less systems futexes are always "private" as there is no per
438 * process address space. We need the smp wmb nevertheless - yes,
439 * arch/blackfin has MMU less SMP ...
441 if (!IS_ENABLED(CONFIG_MMU)) {
442 smp_mb(); /* explicit smp_mb(); (B) */
446 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
448 smp_mb(); /* explicit smp_mb(); (B) */
450 case FUT_OFF_MMSHARED:
451 futex_get_mm(key); /* implies smp_mb(); (B) */
455 * Private futexes do not hold reference on an inode or
456 * mm, therefore the only purpose of calling get_futex_key_refs
457 * is because we need the barrier for the lockless waiter check.
459 smp_mb(); /* explicit smp_mb(); (B) */
464 * Drop a reference to the resource addressed by a key.
465 * The hash bucket spinlock must not be held. This is
466 * a no-op for private futexes, see comment in the get
469 static void drop_futex_key_refs(union futex_key *key)
471 if (!key->both.ptr) {
472 /* If we're here then we tried to put a key we failed to get */
477 if (!IS_ENABLED(CONFIG_MMU))
480 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
483 case FUT_OFF_MMSHARED:
484 mmdrop(key->private.mm);
490 * Generate a machine wide unique identifier for this inode.
492 * This relies on u64 not wrapping in the life-time of the machine; which with
493 * 1ns resolution means almost 585 years.
495 * This further relies on the fact that a well formed program will not unmap
496 * the file while it has a (shared) futex waiting on it. This mapping will have
497 * a file reference which pins the mount and inode.
499 * If for some reason an inode gets evicted and read back in again, it will get
500 * a new sequence number and will _NOT_ match, even though it is the exact same
503 * It is important that match_futex() will never have a false-positive, esp.
504 * for PI futexes that can mess up the state. The above argues that false-negatives
505 * are only possible for malformed programs.
507 static u64 get_inode_sequence_number(struct inode *inode)
509 static atomic64_t i_seq;
512 /* Does the inode already have a sequence number? */
513 old = atomic64_read(&inode->i_sequence);
518 u64 new = atomic64_add_return(1, &i_seq);
519 if (WARN_ON_ONCE(!new))
522 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
530 * get_futex_key() - Get parameters which are the keys for a futex
531 * @uaddr: virtual address of the futex
532 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
533 * @key: address where result is stored.
534 * @rw: mapping needs to be read/write (values: VERIFY_READ,
537 * Return: a negative error code or 0
539 * The key words are stored in @key on success.
541 * For shared mappings (when @fshared), the key is:
542 * ( inode->i_sequence, page->index, offset_within_page )
543 * [ also see get_inode_sequence_number() ]
545 * For private mappings (or when !@fshared), the key is:
546 * ( current->mm, address, 0 )
548 * This allows (cross process, where applicable) identification of the futex
549 * without keeping the page pinned for the duration of the FUTEX_WAIT.
551 * lock_page() might sleep, the caller should not hold a spinlock.
554 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
556 unsigned long address = (unsigned long)uaddr;
557 struct mm_struct *mm = current->mm;
558 struct page *page, *tail;
559 struct address_space *mapping;
563 * The futex address must be "naturally" aligned.
565 key->both.offset = address % PAGE_SIZE;
566 if (unlikely((address % sizeof(u32)) != 0))
568 address -= key->both.offset;
570 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
573 if (unlikely(should_fail_futex(fshared)))
577 * PROCESS_PRIVATE futexes are fast.
578 * As the mm cannot disappear under us and the 'key' only needs
579 * virtual address, we dont even have to find the underlying vma.
580 * Note : We do have to check 'uaddr' is a valid user address,
581 * but access_ok() should be faster than find_vma()
584 key->private.mm = mm;
585 key->private.address = address;
586 get_futex_key_refs(key); /* implies smp_mb(); (B) */
591 /* Ignore any VERIFY_READ mapping (futex common case) */
592 if (unlikely(should_fail_futex(fshared)))
595 err = get_user_pages_fast(address, 1, 1, &page);
597 * If write access is not required (eg. FUTEX_WAIT), try
598 * and get read-only access.
600 if (err == -EFAULT && rw == VERIFY_READ) {
601 err = get_user_pages_fast(address, 1, 0, &page);
610 * The treatment of mapping from this point on is critical. The page
611 * lock protects many things but in this context the page lock
612 * stabilizes mapping, prevents inode freeing in the shared
613 * file-backed region case and guards against movement to swap cache.
615 * Strictly speaking the page lock is not needed in all cases being
616 * considered here and page lock forces unnecessarily serialization
617 * From this point on, mapping will be re-verified if necessary and
618 * page lock will be acquired only if it is unavoidable
620 * Mapping checks require the head page for any compound page so the
621 * head page and mapping is looked up now. For anonymous pages, it
622 * does not matter if the page splits in the future as the key is
623 * based on the address. For filesystem-backed pages, the tail is
624 * required as the index of the page determines the key. For
625 * base pages, there is no tail page and tail == page.
628 page = compound_head(page);
629 mapping = READ_ONCE(page->mapping);
632 * If page->mapping is NULL, then it cannot be a PageAnon
633 * page; but it might be the ZERO_PAGE or in the gate area or
634 * in a special mapping (all cases which we are happy to fail);
635 * or it may have been a good file page when get_user_pages_fast
636 * found it, but truncated or holepunched or subjected to
637 * invalidate_complete_page2 before we got the page lock (also
638 * cases which we are happy to fail). And we hold a reference,
639 * so refcount care in invalidate_complete_page's remove_mapping
640 * prevents drop_caches from setting mapping to NULL beneath us.
642 * The case we do have to guard against is when memory pressure made
643 * shmem_writepage move it from filecache to swapcache beneath us:
644 * an unlikely race, but we do need to retry for page->mapping.
646 if (unlikely(!mapping)) {
650 * Page lock is required to identify which special case above
651 * applies. If this is really a shmem page then the page lock
652 * will prevent unexpected transitions.
655 shmem_swizzled = PageSwapCache(page) || page->mapping;
666 * Private mappings are handled in a simple way.
668 * If the futex key is stored on an anonymous page, then the associated
669 * object is the mm which is implicitly pinned by the calling process.
671 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
672 * it's a read-only handle, it's expected that futexes attach to
673 * the object not the particular process.
675 if (PageAnon(page)) {
677 * A RO anonymous page will never change and thus doesn't make
678 * sense for futex operations.
680 if (unlikely(should_fail_futex(fshared)) || ro) {
685 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
686 key->private.mm = mm;
687 key->private.address = address;
693 * The associated futex object in this case is the inode and
694 * the page->mapping must be traversed. Ordinarily this should
695 * be stabilised under page lock but it's not strictly
696 * necessary in this case as we just want to pin the inode, not
697 * update the radix tree or anything like that.
699 * The RCU read lock is taken as the inode is finally freed
700 * under RCU. If the mapping still matches expectations then the
701 * mapping->host can be safely accessed as being a valid inode.
705 if (READ_ONCE(page->mapping) != mapping) {
712 inode = READ_ONCE(mapping->host);
720 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
721 key->shared.i_seq = get_inode_sequence_number(inode);
722 key->shared.pgoff = page_to_pgoff(tail);
726 get_futex_key_refs(key); /* implies smp_mb(); (B) */
733 static inline void put_futex_key(union futex_key *key)
735 drop_futex_key_refs(key);
739 * fault_in_user_writeable() - Fault in user address and verify RW access
740 * @uaddr: pointer to faulting user space address
742 * Slow path to fixup the fault we just took in the atomic write
745 * We have no generic implementation of a non-destructive write to the
746 * user address. We know that we faulted in the atomic pagefault
747 * disabled section so we can as well avoid the #PF overhead by
748 * calling get_user_pages() right away.
750 static int fault_in_user_writeable(u32 __user *uaddr)
752 struct mm_struct *mm = current->mm;
755 down_read(&mm->mmap_sem);
756 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
757 FAULT_FLAG_WRITE, NULL);
758 up_read(&mm->mmap_sem);
760 return ret < 0 ? ret : 0;
764 * futex_top_waiter() - Return the highest priority waiter on a futex
765 * @hb: the hash bucket the futex_q's reside in
766 * @key: the futex key (to distinguish it from other futex futex_q's)
768 * Must be called with the hb lock held.
770 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
771 union futex_key *key)
773 struct futex_q *this;
775 plist_for_each_entry(this, &hb->chain, list) {
776 if (match_futex(&this->key, key))
782 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
783 u32 uval, u32 newval)
788 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
794 static int get_futex_value_locked(u32 *dest, u32 __user *from)
799 ret = __get_user(*dest, from);
802 return ret ? -EFAULT : 0;
809 static int refill_pi_state_cache(void)
811 struct futex_pi_state *pi_state;
813 if (likely(current->pi_state_cache))
816 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
821 INIT_LIST_HEAD(&pi_state->list);
822 /* pi_mutex gets initialized later */
823 pi_state->owner = NULL;
824 atomic_set(&pi_state->refcount, 1);
825 pi_state->key = FUTEX_KEY_INIT;
827 current->pi_state_cache = pi_state;
832 static struct futex_pi_state *alloc_pi_state(void)
834 struct futex_pi_state *pi_state = current->pi_state_cache;
837 current->pi_state_cache = NULL;
842 static void pi_state_update_owner(struct futex_pi_state *pi_state,
843 struct task_struct *new_owner)
845 struct task_struct *old_owner = pi_state->owner;
847 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
850 raw_spin_lock(&old_owner->pi_lock);
851 WARN_ON(list_empty(&pi_state->list));
852 list_del_init(&pi_state->list);
853 raw_spin_unlock(&old_owner->pi_lock);
857 raw_spin_lock(&new_owner->pi_lock);
858 WARN_ON(!list_empty(&pi_state->list));
859 list_add(&pi_state->list, &new_owner->pi_state_list);
860 pi_state->owner = new_owner;
861 raw_spin_unlock(&new_owner->pi_lock);
865 static void get_pi_state(struct futex_pi_state *pi_state)
867 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
871 * Drops a reference to the pi_state object and frees or caches it
872 * when the last reference is gone.
874 static void put_pi_state(struct futex_pi_state *pi_state)
879 if (!atomic_dec_and_test(&pi_state->refcount))
883 * If pi_state->owner is NULL, the owner is most probably dying
884 * and has cleaned up the pi_state already
886 if (pi_state->owner) {
889 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
890 pi_state_update_owner(pi_state, NULL);
891 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
892 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
895 if (current->pi_state_cache) {
899 * pi_state->list is already empty.
900 * clear pi_state->owner.
901 * refcount is at 0 - put it back to 1.
903 pi_state->owner = NULL;
904 atomic_set(&pi_state->refcount, 1);
905 current->pi_state_cache = pi_state;
910 * Look up the task based on what TID userspace gave us.
913 static struct task_struct *futex_find_get_task(pid_t pid)
915 struct task_struct *p;
918 p = find_task_by_vpid(pid);
927 #ifdef CONFIG_FUTEX_PI
930 * This task is holding PI mutexes at exit time => bad.
931 * Kernel cleans up PI-state, but userspace is likely hosed.
932 * (Robust-futex cleanup is separate and might save the day for userspace.)
934 static void exit_pi_state_list(struct task_struct *curr)
936 struct list_head *next, *head = &curr->pi_state_list;
937 struct futex_pi_state *pi_state;
938 struct futex_hash_bucket *hb;
939 union futex_key key = FUTEX_KEY_INIT;
941 if (!futex_cmpxchg_enabled)
944 * We are a ZOMBIE and nobody can enqueue itself on
945 * pi_state_list anymore, but we have to be careful
946 * versus waiters unqueueing themselves:
948 raw_spin_lock_irq(&curr->pi_lock);
949 while (!list_empty(head)) {
951 pi_state = list_entry(next, struct futex_pi_state, list);
953 hb = hash_futex(&key);
956 * We can race against put_pi_state() removing itself from the
957 * list (a waiter going away). put_pi_state() will first
958 * decrement the reference count and then modify the list, so
959 * its possible to see the list entry but fail this reference
962 * In that case; drop the locks to let put_pi_state() make
963 * progress and retry the loop.
965 if (!atomic_inc_not_zero(&pi_state->refcount)) {
966 raw_spin_unlock_irq(&curr->pi_lock);
968 raw_spin_lock_irq(&curr->pi_lock);
971 raw_spin_unlock_irq(&curr->pi_lock);
973 spin_lock(&hb->lock);
974 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
975 raw_spin_lock(&curr->pi_lock);
977 * We dropped the pi-lock, so re-check whether this
978 * task still owns the PI-state:
980 if (head->next != next) {
981 /* retain curr->pi_lock for the loop invariant */
982 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
983 spin_unlock(&hb->lock);
984 put_pi_state(pi_state);
988 WARN_ON(pi_state->owner != curr);
989 WARN_ON(list_empty(&pi_state->list));
990 list_del_init(&pi_state->list);
991 pi_state->owner = NULL;
993 raw_spin_unlock(&curr->pi_lock);
994 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
995 spin_unlock(&hb->lock);
997 rt_mutex_futex_unlock(&pi_state->pi_mutex);
998 put_pi_state(pi_state);
1000 raw_spin_lock_irq(&curr->pi_lock);
1002 raw_spin_unlock_irq(&curr->pi_lock);
1005 static inline void exit_pi_state_list(struct task_struct *curr) { }
1009 * We need to check the following states:
1011 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
1013 * [1] NULL | --- | --- | 0 | 0/1 | Valid
1014 * [2] NULL | --- | --- | >0 | 0/1 | Valid
1016 * [3] Found | NULL | -- | Any | 0/1 | Invalid
1018 * [4] Found | Found | NULL | 0 | 1 | Valid
1019 * [5] Found | Found | NULL | >0 | 1 | Invalid
1021 * [6] Found | Found | task | 0 | 1 | Valid
1023 * [7] Found | Found | NULL | Any | 0 | Invalid
1025 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
1026 * [9] Found | Found | task | 0 | 0 | Invalid
1027 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
1029 * [1] Indicates that the kernel can acquire the futex atomically. We
1030 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
1032 * [2] Valid, if TID does not belong to a kernel thread. If no matching
1033 * thread is found then it indicates that the owner TID has died.
1035 * [3] Invalid. The waiter is queued on a non PI futex
1037 * [4] Valid state after exit_robust_list(), which sets the user space
1038 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
1040 * [5] The user space value got manipulated between exit_robust_list()
1041 * and exit_pi_state_list()
1043 * [6] Valid state after exit_pi_state_list() which sets the new owner in
1044 * the pi_state but cannot access the user space value.
1046 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1048 * [8] Owner and user space value match
1050 * [9] There is no transient state which sets the user space TID to 0
1051 * except exit_robust_list(), but this is indicated by the
1052 * FUTEX_OWNER_DIED bit. See [4]
1054 * [10] There is no transient state which leaves owner and user space
1055 * TID out of sync. Except one error case where the kernel is denied
1056 * write access to the user address, see fixup_pi_state_owner().
1059 * Serialization and lifetime rules:
1063 * hb -> futex_q, relation
1064 * futex_q -> pi_state, relation
1066 * (cannot be raw because hb can contain arbitrary amount
1069 * pi_mutex->wait_lock:
1073 * (and pi_mutex 'obviously')
1077 * p->pi_state_list -> pi_state->list, relation
1079 * pi_state->refcount:
1087 * pi_mutex->wait_lock
1093 * Validate that the existing waiter has a pi_state and sanity check
1094 * the pi_state against the user space value. If correct, attach to
1097 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1098 struct futex_pi_state *pi_state,
1099 struct futex_pi_state **ps)
1101 pid_t pid = uval & FUTEX_TID_MASK;
1106 * Userspace might have messed up non-PI and PI futexes [3]
1108 if (unlikely(!pi_state))
1112 * We get here with hb->lock held, and having found a
1113 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1114 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1115 * which in turn means that futex_lock_pi() still has a reference on
1118 * The waiter holding a reference on @pi_state also protects against
1119 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1120 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1121 * free pi_state before we can take a reference ourselves.
1123 WARN_ON(!atomic_read(&pi_state->refcount));
1126 * Now that we have a pi_state, we can acquire wait_lock
1127 * and do the state validation.
1129 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1132 * Since {uval, pi_state} is serialized by wait_lock, and our current
1133 * uval was read without holding it, it can have changed. Verify it
1134 * still is what we expect it to be, otherwise retry the entire
1137 if (get_futex_value_locked(&uval2, uaddr))
1144 * Handle the owner died case:
1146 if (uval & FUTEX_OWNER_DIED) {
1148 * exit_pi_state_list sets owner to NULL and wakes the
1149 * topmost waiter. The task which acquires the
1150 * pi_state->rt_mutex will fixup owner.
1152 if (!pi_state->owner) {
1154 * No pi state owner, but the user space TID
1155 * is not 0. Inconsistent state. [5]
1160 * Take a ref on the state and return success. [4]
1166 * If TID is 0, then either the dying owner has not
1167 * yet executed exit_pi_state_list() or some waiter
1168 * acquired the rtmutex in the pi state, but did not
1169 * yet fixup the TID in user space.
1171 * Take a ref on the state and return success. [6]
1177 * If the owner died bit is not set, then the pi_state
1178 * must have an owner. [7]
1180 if (!pi_state->owner)
1185 * Bail out if user space manipulated the futex value. If pi
1186 * state exists then the owner TID must be the same as the
1187 * user space TID. [9/10]
1189 if (pid != task_pid_vnr(pi_state->owner))
1193 get_pi_state(pi_state);
1194 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1211 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1216 * wait_for_owner_exiting - Block until the owner has exited
1217 * @exiting: Pointer to the exiting task
1219 * Caller must hold a refcount on @exiting.
1221 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1223 if (ret != -EBUSY) {
1224 WARN_ON_ONCE(exiting);
1228 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1231 mutex_lock(&exiting->futex_exit_mutex);
1233 * No point in doing state checking here. If the waiter got here
1234 * while the task was in exec()->exec_futex_release() then it can
1235 * have any FUTEX_STATE_* value when the waiter has acquired the
1236 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1237 * already. Highly unlikely and not a problem. Just one more round
1238 * through the futex maze.
1240 mutex_unlock(&exiting->futex_exit_mutex);
1242 put_task_struct(exiting);
1245 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1246 struct task_struct *tsk)
1251 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1252 * caller that the alleged owner is busy.
1254 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1258 * Reread the user space value to handle the following situation:
1262 * sys_exit() sys_futex()
1263 * do_exit() futex_lock_pi()
1264 * futex_lock_pi_atomic()
1265 * exit_signals(tsk) No waiters:
1266 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1267 * mm_release(tsk) Set waiter bit
1268 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1269 * Set owner died attach_to_pi_owner() {
1270 * *uaddr = 0xC0000000; tsk = get_task(PID);
1271 * } if (!tsk->flags & PF_EXITING) {
1273 * tsk->futex_state = } else {
1274 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1277 * return -ESRCH; <--- FAIL
1280 * Returning ESRCH unconditionally is wrong here because the
1281 * user space value has been changed by the exiting task.
1283 * The same logic applies to the case where the exiting task is
1286 if (get_futex_value_locked(&uval2, uaddr))
1289 /* If the user space value has changed, try again. */
1294 * The exiting task did not have a robust list, the robust list was
1295 * corrupted or the user space value in *uaddr is simply bogus.
1296 * Give up and tell user space.
1302 * Lookup the task for the TID provided from user space and attach to
1303 * it after doing proper sanity checks.
1305 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1306 struct futex_pi_state **ps,
1307 struct task_struct **exiting)
1309 pid_t pid = uval & FUTEX_TID_MASK;
1310 struct futex_pi_state *pi_state;
1311 struct task_struct *p;
1314 * We are the first waiter - try to look up the real owner and attach
1315 * the new pi_state to it, but bail out when TID = 0 [1]
1317 * The !pid check is paranoid. None of the call sites should end up
1318 * with pid == 0, but better safe than sorry. Let the caller retry
1322 p = futex_find_get_task(pid);
1324 return handle_exit_race(uaddr, uval, NULL);
1326 if (unlikely(p->flags & PF_KTHREAD)) {
1332 * We need to look at the task state to figure out, whether the
1333 * task is exiting. To protect against the change of the task state
1334 * in futex_exit_release(), we do this protected by p->pi_lock:
1336 raw_spin_lock_irq(&p->pi_lock);
1337 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1339 * The task is on the way out. When the futex state is
1340 * FUTEX_STATE_DEAD, we know that the task has finished
1343 int ret = handle_exit_race(uaddr, uval, p);
1345 raw_spin_unlock_irq(&p->pi_lock);
1347 * If the owner task is between FUTEX_STATE_EXITING and
1348 * FUTEX_STATE_DEAD then store the task pointer and keep
1349 * the reference on the task struct. The calling code will
1350 * drop all locks, wait for the task to reach
1351 * FUTEX_STATE_DEAD and then drop the refcount. This is
1352 * required to prevent a live lock when the current task
1353 * preempted the exiting task between the two states.
1363 * No existing pi state. First waiter. [2]
1365 * This creates pi_state, we have hb->lock held, this means nothing can
1366 * observe this state, wait_lock is irrelevant.
1368 pi_state = alloc_pi_state();
1371 * Initialize the pi_mutex in locked state and make @p
1374 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1376 /* Store the key for possible exit cleanups: */
1377 pi_state->key = *key;
1379 WARN_ON(!list_empty(&pi_state->list));
1380 list_add(&pi_state->list, &p->pi_state_list);
1382 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1383 * because there is no concurrency as the object is not published yet.
1385 pi_state->owner = p;
1386 raw_spin_unlock_irq(&p->pi_lock);
1395 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1396 struct futex_hash_bucket *hb,
1397 union futex_key *key, struct futex_pi_state **ps,
1398 struct task_struct **exiting)
1400 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1403 * If there is a waiter on that futex, validate it and
1404 * attach to the pi_state when the validation succeeds.
1407 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1410 * We are the first waiter - try to look up the owner based on
1411 * @uval and attach to it.
1413 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1416 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1419 u32 uninitialized_var(curval);
1421 if (unlikely(should_fail_futex(true)))
1424 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1428 /* If user space value changed, let the caller retry */
1429 return curval != uval ? -EAGAIN : 0;
1433 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1434 * @uaddr: the pi futex user address
1435 * @hb: the pi futex hash bucket
1436 * @key: the futex key associated with uaddr and hb
1437 * @ps: the pi_state pointer where we store the result of the
1439 * @task: the task to perform the atomic lock work for. This will
1440 * be "current" except in the case of requeue pi.
1441 * @exiting: Pointer to store the task pointer of the owner task
1442 * which is in the middle of exiting
1443 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1446 * - 0 - ready to wait;
1447 * - 1 - acquired the lock;
1450 * The hb->lock and futex_key refs shall be held by the caller.
1452 * @exiting is only set when the return value is -EBUSY. If so, this holds
1453 * a refcount on the exiting task on return and the caller needs to drop it
1454 * after waiting for the exit to complete.
1456 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1457 union futex_key *key,
1458 struct futex_pi_state **ps,
1459 struct task_struct *task,
1460 struct task_struct **exiting,
1463 u32 uval, newval, vpid = task_pid_vnr(task);
1464 struct futex_q *top_waiter;
1468 * Read the user space value first so we can validate a few
1469 * things before proceeding further.
1471 if (get_futex_value_locked(&uval, uaddr))
1474 if (unlikely(should_fail_futex(true)))
1480 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1483 if ((unlikely(should_fail_futex(true))))
1487 * Lookup existing state first. If it exists, try to attach to
1490 top_waiter = futex_top_waiter(hb, key);
1492 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1495 * No waiter and user TID is 0. We are here because the
1496 * waiters or the owner died bit is set or called from
1497 * requeue_cmp_pi or for whatever reason something took the
1500 if (!(uval & FUTEX_TID_MASK)) {
1502 * We take over the futex. No other waiters and the user space
1503 * TID is 0. We preserve the owner died bit.
1505 newval = uval & FUTEX_OWNER_DIED;
1508 /* The futex requeue_pi code can enforce the waiters bit */
1510 newval |= FUTEX_WAITERS;
1512 ret = lock_pi_update_atomic(uaddr, uval, newval);
1513 /* If the take over worked, return 1 */
1514 return ret < 0 ? ret : 1;
1518 * First waiter. Set the waiters bit before attaching ourself to
1519 * the owner. If owner tries to unlock, it will be forced into
1520 * the kernel and blocked on hb->lock.
1522 newval = uval | FUTEX_WAITERS;
1523 ret = lock_pi_update_atomic(uaddr, uval, newval);
1527 * If the update of the user space value succeeded, we try to
1528 * attach to the owner. If that fails, no harm done, we only
1529 * set the FUTEX_WAITERS bit in the user space variable.
1531 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1535 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1536 * @q: The futex_q to unqueue
1538 * The q->lock_ptr must not be NULL and must be held by the caller.
1540 static void __unqueue_futex(struct futex_q *q)
1542 struct futex_hash_bucket *hb;
1544 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1545 || WARN_ON(plist_node_empty(&q->list)))
1548 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1549 plist_del(&q->list, &hb->chain);
1554 * The hash bucket lock must be held when this is called.
1555 * Afterwards, the futex_q must not be accessed. Callers
1556 * must ensure to later call wake_up_q() for the actual
1559 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1561 struct task_struct *p = q->task;
1563 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1569 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1570 * is written, without taking any locks. This is possible in the event
1571 * of a spurious wakeup, for example. A memory barrier is required here
1572 * to prevent the following store to lock_ptr from getting ahead of the
1573 * plist_del in __unqueue_futex().
1575 smp_store_release(&q->lock_ptr, NULL);
1578 * Queue the task for later wakeup for after we've released
1579 * the hb->lock. wake_q_add() grabs reference to p.
1581 wake_q_add(wake_q, p);
1586 * Caller must hold a reference on @pi_state.
1588 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1590 u32 uninitialized_var(curval), newval;
1591 struct task_struct *new_owner;
1592 bool postunlock = false;
1593 DEFINE_WAKE_Q(wake_q);
1596 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1597 if (WARN_ON_ONCE(!new_owner)) {
1599 * As per the comment in futex_unlock_pi() this should not happen.
1601 * When this happens, give up our locks and try again, giving
1602 * the futex_lock_pi() instance time to complete, either by
1603 * waiting on the rtmutex or removing itself from the futex
1611 * We pass it to the next owner. The WAITERS bit is always kept
1612 * enabled while there is PI state around. We cleanup the owner
1613 * died bit, because we are the owner.
1615 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1617 if (unlikely(should_fail_futex(true))) {
1622 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1623 if (!ret && (curval != uval)) {
1625 * If a unconditional UNLOCK_PI operation (user space did not
1626 * try the TID->0 transition) raced with a waiter setting the
1627 * FUTEX_WAITERS flag between get_user() and locking the hash
1628 * bucket lock, retry the operation.
1630 if ((FUTEX_TID_MASK & curval) == uval)
1638 * This is a point of no return; once we modified the uval
1639 * there is no going back and subsequent operations must
1642 pi_state_update_owner(pi_state, new_owner);
1643 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1647 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1650 rt_mutex_postunlock(&wake_q);
1656 * Express the locking dependencies for lockdep:
1659 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1662 spin_lock(&hb1->lock);
1664 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1665 } else { /* hb1 > hb2 */
1666 spin_lock(&hb2->lock);
1667 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1672 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1674 spin_unlock(&hb1->lock);
1676 spin_unlock(&hb2->lock);
1680 * Wake up waiters matching bitset queued on this futex (uaddr).
1683 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1685 struct futex_hash_bucket *hb;
1686 struct futex_q *this, *next;
1687 union futex_key key = FUTEX_KEY_INIT;
1689 DEFINE_WAKE_Q(wake_q);
1694 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1695 if (unlikely(ret != 0))
1698 hb = hash_futex(&key);
1700 /* Make sure we really have tasks to wakeup */
1701 if (!hb_waiters_pending(hb))
1704 spin_lock(&hb->lock);
1706 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1707 if (match_futex (&this->key, &key)) {
1708 if (this->pi_state || this->rt_waiter) {
1713 /* Check if one of the bits is set in both bitsets */
1714 if (!(this->bitset & bitset))
1717 mark_wake_futex(&wake_q, this);
1718 if (++ret >= nr_wake)
1723 spin_unlock(&hb->lock);
1726 put_futex_key(&key);
1731 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1733 unsigned int op = (encoded_op & 0x70000000) >> 28;
1734 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1735 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1736 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1739 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1740 if (oparg < 0 || oparg > 31) {
1741 char comm[sizeof(current->comm)];
1743 * kill this print and return -EINVAL when userspace
1746 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1747 get_task_comm(comm, current), oparg);
1753 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1756 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1761 case FUTEX_OP_CMP_EQ:
1762 return oldval == cmparg;
1763 case FUTEX_OP_CMP_NE:
1764 return oldval != cmparg;
1765 case FUTEX_OP_CMP_LT:
1766 return oldval < cmparg;
1767 case FUTEX_OP_CMP_GE:
1768 return oldval >= cmparg;
1769 case FUTEX_OP_CMP_LE:
1770 return oldval <= cmparg;
1771 case FUTEX_OP_CMP_GT:
1772 return oldval > cmparg;
1779 * Wake up all waiters hashed on the physical page that is mapped
1780 * to this virtual address:
1783 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1784 int nr_wake, int nr_wake2, int op)
1786 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1787 struct futex_hash_bucket *hb1, *hb2;
1788 struct futex_q *this, *next;
1790 DEFINE_WAKE_Q(wake_q);
1793 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1794 if (unlikely(ret != 0))
1796 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1797 if (unlikely(ret != 0))
1800 hb1 = hash_futex(&key1);
1801 hb2 = hash_futex(&key2);
1804 double_lock_hb(hb1, hb2);
1805 op_ret = futex_atomic_op_inuser(op, uaddr2);
1806 if (unlikely(op_ret < 0)) {
1807 double_unlock_hb(hb1, hb2);
1809 if (!IS_ENABLED(CONFIG_MMU) ||
1810 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1812 * we don't get EFAULT from MMU faults if we don't have
1813 * an MMU, but we might get them from range checking
1819 if (op_ret == -EFAULT) {
1820 ret = fault_in_user_writeable(uaddr2);
1825 if (!(flags & FLAGS_SHARED)) {
1830 put_futex_key(&key2);
1831 put_futex_key(&key1);
1836 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1837 if (match_futex (&this->key, &key1)) {
1838 if (this->pi_state || this->rt_waiter) {
1842 mark_wake_futex(&wake_q, this);
1843 if (++ret >= nr_wake)
1850 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1851 if (match_futex (&this->key, &key2)) {
1852 if (this->pi_state || this->rt_waiter) {
1856 mark_wake_futex(&wake_q, this);
1857 if (++op_ret >= nr_wake2)
1865 double_unlock_hb(hb1, hb2);
1868 put_futex_key(&key2);
1870 put_futex_key(&key1);
1876 * requeue_futex() - Requeue a futex_q from one hb to another
1877 * @q: the futex_q to requeue
1878 * @hb1: the source hash_bucket
1879 * @hb2: the target hash_bucket
1880 * @key2: the new key for the requeued futex_q
1883 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1884 struct futex_hash_bucket *hb2, union futex_key *key2)
1888 * If key1 and key2 hash to the same bucket, no need to
1891 if (likely(&hb1->chain != &hb2->chain)) {
1892 plist_del(&q->list, &hb1->chain);
1893 hb_waiters_dec(hb1);
1894 hb_waiters_inc(hb2);
1895 plist_add(&q->list, &hb2->chain);
1896 q->lock_ptr = &hb2->lock;
1898 get_futex_key_refs(key2);
1903 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1905 * @key: the key of the requeue target futex
1906 * @hb: the hash_bucket of the requeue target futex
1908 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1909 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1910 * to the requeue target futex so the waiter can detect the wakeup on the right
1911 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1912 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1913 * to protect access to the pi_state to fixup the owner later. Must be called
1914 * with both q->lock_ptr and hb->lock held.
1917 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1918 struct futex_hash_bucket *hb)
1920 get_futex_key_refs(key);
1925 WARN_ON(!q->rt_waiter);
1926 q->rt_waiter = NULL;
1928 q->lock_ptr = &hb->lock;
1930 wake_up_state(q->task, TASK_NORMAL);
1934 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1935 * @pifutex: the user address of the to futex
1936 * @hb1: the from futex hash bucket, must be locked by the caller
1937 * @hb2: the to futex hash bucket, must be locked by the caller
1938 * @key1: the from futex key
1939 * @key2: the to futex key
1940 * @ps: address to store the pi_state pointer
1941 * @exiting: Pointer to store the task pointer of the owner task
1942 * which is in the middle of exiting
1943 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1945 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1946 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1947 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1948 * hb1 and hb2 must be held by the caller.
1950 * @exiting is only set when the return value is -EBUSY. If so, this holds
1951 * a refcount on the exiting task on return and the caller needs to drop it
1952 * after waiting for the exit to complete.
1955 * - 0 - failed to acquire the lock atomically;
1956 * - >0 - acquired the lock, return value is vpid of the top_waiter
1960 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1961 struct futex_hash_bucket *hb2, union futex_key *key1,
1962 union futex_key *key2, struct futex_pi_state **ps,
1963 struct task_struct **exiting, int set_waiters)
1965 struct futex_q *top_waiter = NULL;
1969 if (get_futex_value_locked(&curval, pifutex))
1972 if (unlikely(should_fail_futex(true)))
1976 * Find the top_waiter and determine if there are additional waiters.
1977 * If the caller intends to requeue more than 1 waiter to pifutex,
1978 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1979 * as we have means to handle the possible fault. If not, don't set
1980 * the bit unecessarily as it will force the subsequent unlock to enter
1983 top_waiter = futex_top_waiter(hb1, key1);
1985 /* There are no waiters, nothing for us to do. */
1989 /* Ensure we requeue to the expected futex. */
1990 if (!match_futex(top_waiter->requeue_pi_key, key2))
1994 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1995 * the contended case or if set_waiters is 1. The pi_state is returned
1996 * in ps in contended cases.
1998 vpid = task_pid_vnr(top_waiter->task);
1999 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
2000 exiting, set_waiters);
2002 requeue_pi_wake_futex(top_waiter, key2, hb2);
2009 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
2010 * @uaddr1: source futex user address
2011 * @flags: futex flags (FLAGS_SHARED, etc.)
2012 * @uaddr2: target futex user address
2013 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
2014 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
2015 * @cmpval: @uaddr1 expected value (or %NULL)
2016 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
2017 * pi futex (pi to pi requeue is not supported)
2019 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
2020 * uaddr2 atomically on behalf of the top waiter.
2023 * - >=0 - on success, the number of tasks requeued or woken;
2026 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
2027 u32 __user *uaddr2, int nr_wake, int nr_requeue,
2028 u32 *cmpval, int requeue_pi)
2030 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
2031 int drop_count = 0, task_count = 0, ret;
2032 struct futex_pi_state *pi_state = NULL;
2033 struct futex_hash_bucket *hb1, *hb2;
2034 struct futex_q *this, *next;
2035 DEFINE_WAKE_Q(wake_q);
2037 if (nr_wake < 0 || nr_requeue < 0)
2041 * When PI not supported: return -ENOSYS if requeue_pi is true,
2042 * consequently the compiler knows requeue_pi is always false past
2043 * this point which will optimize away all the conditional code
2046 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
2051 * Requeue PI only works on two distinct uaddrs. This
2052 * check is only valid for private futexes. See below.
2054 if (uaddr1 == uaddr2)
2058 * requeue_pi requires a pi_state, try to allocate it now
2059 * without any locks in case it fails.
2061 if (refill_pi_state_cache())
2064 * requeue_pi must wake as many tasks as it can, up to nr_wake
2065 * + nr_requeue, since it acquires the rt_mutex prior to
2066 * returning to userspace, so as to not leave the rt_mutex with
2067 * waiters and no owner. However, second and third wake-ups
2068 * cannot be predicted as they involve race conditions with the
2069 * first wake and a fault while looking up the pi_state. Both
2070 * pthread_cond_signal() and pthread_cond_broadcast() should
2078 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
2079 if (unlikely(ret != 0))
2081 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
2082 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
2083 if (unlikely(ret != 0))
2087 * The check above which compares uaddrs is not sufficient for
2088 * shared futexes. We need to compare the keys:
2090 if (requeue_pi && match_futex(&key1, &key2)) {
2095 hb1 = hash_futex(&key1);
2096 hb2 = hash_futex(&key2);
2099 hb_waiters_inc(hb2);
2100 double_lock_hb(hb1, hb2);
2102 if (likely(cmpval != NULL)) {
2105 ret = get_futex_value_locked(&curval, uaddr1);
2107 if (unlikely(ret)) {
2108 double_unlock_hb(hb1, hb2);
2109 hb_waiters_dec(hb2);
2111 ret = get_user(curval, uaddr1);
2115 if (!(flags & FLAGS_SHARED))
2118 put_futex_key(&key2);
2119 put_futex_key(&key1);
2122 if (curval != *cmpval) {
2128 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2129 struct task_struct *exiting = NULL;
2132 * Attempt to acquire uaddr2 and wake the top waiter. If we
2133 * intend to requeue waiters, force setting the FUTEX_WAITERS
2134 * bit. We force this here where we are able to easily handle
2135 * faults rather in the requeue loop below.
2137 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2139 &exiting, nr_requeue);
2142 * At this point the top_waiter has either taken uaddr2 or is
2143 * waiting on it. If the former, then the pi_state will not
2144 * exist yet, look it up one more time to ensure we have a
2145 * reference to it. If the lock was taken, ret contains the
2146 * vpid of the top waiter task.
2147 * If the lock was not taken, we have pi_state and an initial
2148 * refcount on it. In case of an error we have nothing.
2155 * If we acquired the lock, then the user space value
2156 * of uaddr2 should be vpid. It cannot be changed by
2157 * the top waiter as it is blocked on hb2 lock if it
2158 * tries to do so. If something fiddled with it behind
2159 * our back the pi state lookup might unearth it. So
2160 * we rather use the known value than rereading and
2161 * handing potential crap to lookup_pi_state.
2163 * If that call succeeds then we have pi_state and an
2164 * initial refcount on it.
2166 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2167 &pi_state, &exiting);
2172 /* We hold a reference on the pi state. */
2175 /* If the above failed, then pi_state is NULL */
2177 double_unlock_hb(hb1, hb2);
2178 hb_waiters_dec(hb2);
2179 put_futex_key(&key2);
2180 put_futex_key(&key1);
2181 ret = fault_in_user_writeable(uaddr2);
2188 * Two reasons for this:
2189 * - EBUSY: Owner is exiting and we just wait for the
2191 * - EAGAIN: The user space value changed.
2193 double_unlock_hb(hb1, hb2);
2194 hb_waiters_dec(hb2);
2195 put_futex_key(&key2);
2196 put_futex_key(&key1);
2198 * Handle the case where the owner is in the middle of
2199 * exiting. Wait for the exit to complete otherwise
2200 * this task might loop forever, aka. live lock.
2202 wait_for_owner_exiting(ret, exiting);
2210 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2211 if (task_count - nr_wake >= nr_requeue)
2214 if (!match_futex(&this->key, &key1))
2218 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2219 * be paired with each other and no other futex ops.
2221 * We should never be requeueing a futex_q with a pi_state,
2222 * which is awaiting a futex_unlock_pi().
2224 if ((requeue_pi && !this->rt_waiter) ||
2225 (!requeue_pi && this->rt_waiter) ||
2232 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2233 * lock, we already woke the top_waiter. If not, it will be
2234 * woken by futex_unlock_pi().
2236 if (++task_count <= nr_wake && !requeue_pi) {
2237 mark_wake_futex(&wake_q, this);
2241 /* Ensure we requeue to the expected futex for requeue_pi. */
2242 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2248 * Requeue nr_requeue waiters and possibly one more in the case
2249 * of requeue_pi if we couldn't acquire the lock atomically.
2253 * Prepare the waiter to take the rt_mutex. Take a
2254 * refcount on the pi_state and store the pointer in
2255 * the futex_q object of the waiter.
2257 get_pi_state(pi_state);
2258 this->pi_state = pi_state;
2259 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2264 * We got the lock. We do neither drop the
2265 * refcount on pi_state nor clear
2266 * this->pi_state because the waiter needs the
2267 * pi_state for cleaning up the user space
2268 * value. It will drop the refcount after
2271 requeue_pi_wake_futex(this, &key2, hb2);
2276 * rt_mutex_start_proxy_lock() detected a
2277 * potential deadlock when we tried to queue
2278 * that waiter. Drop the pi_state reference
2279 * which we took above and remove the pointer
2280 * to the state from the waiters futex_q
2283 this->pi_state = NULL;
2284 put_pi_state(pi_state);
2286 * We stop queueing more waiters and let user
2287 * space deal with the mess.
2292 requeue_futex(this, hb1, hb2, &key2);
2297 * We took an extra initial reference to the pi_state either
2298 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2299 * need to drop it here again.
2301 put_pi_state(pi_state);
2304 double_unlock_hb(hb1, hb2);
2306 hb_waiters_dec(hb2);
2309 * drop_futex_key_refs() must be called outside the spinlocks. During
2310 * the requeue we moved futex_q's from the hash bucket at key1 to the
2311 * one at key2 and updated their key pointer. We no longer need to
2312 * hold the references to key1.
2314 while (--drop_count >= 0)
2315 drop_futex_key_refs(&key1);
2318 put_futex_key(&key2);
2320 put_futex_key(&key1);
2322 return ret ? ret : task_count;
2325 /* The key must be already stored in q->key. */
2326 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2327 __acquires(&hb->lock)
2329 struct futex_hash_bucket *hb;
2331 hb = hash_futex(&q->key);
2334 * Increment the counter before taking the lock so that
2335 * a potential waker won't miss a to-be-slept task that is
2336 * waiting for the spinlock. This is safe as all queue_lock()
2337 * users end up calling queue_me(). Similarly, for housekeeping,
2338 * decrement the counter at queue_unlock() when some error has
2339 * occurred and we don't end up adding the task to the list.
2343 q->lock_ptr = &hb->lock;
2345 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2350 queue_unlock(struct futex_hash_bucket *hb)
2351 __releases(&hb->lock)
2353 spin_unlock(&hb->lock);
2357 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2362 * The priority used to register this element is
2363 * - either the real thread-priority for the real-time threads
2364 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2365 * - or MAX_RT_PRIO for non-RT threads.
2366 * Thus, all RT-threads are woken first in priority order, and
2367 * the others are woken last, in FIFO order.
2369 prio = min(current->normal_prio, MAX_RT_PRIO);
2371 plist_node_init(&q->list, prio);
2372 plist_add(&q->list, &hb->chain);
2377 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2378 * @q: The futex_q to enqueue
2379 * @hb: The destination hash bucket
2381 * The hb->lock must be held by the caller, and is released here. A call to
2382 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2383 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2384 * or nothing if the unqueue is done as part of the wake process and the unqueue
2385 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2388 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2389 __releases(&hb->lock)
2392 spin_unlock(&hb->lock);
2396 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2397 * @q: The futex_q to unqueue
2399 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2400 * be paired with exactly one earlier call to queue_me().
2403 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2404 * - 0 - if the futex_q was already removed by the waking thread
2406 static int unqueue_me(struct futex_q *q)
2408 spinlock_t *lock_ptr;
2411 /* In the common case we don't take the spinlock, which is nice. */
2414 * q->lock_ptr can change between this read and the following spin_lock.
2415 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2416 * optimizing lock_ptr out of the logic below.
2418 lock_ptr = READ_ONCE(q->lock_ptr);
2419 if (lock_ptr != NULL) {
2420 spin_lock(lock_ptr);
2422 * q->lock_ptr can change between reading it and
2423 * spin_lock(), causing us to take the wrong lock. This
2424 * corrects the race condition.
2426 * Reasoning goes like this: if we have the wrong lock,
2427 * q->lock_ptr must have changed (maybe several times)
2428 * between reading it and the spin_lock(). It can
2429 * change again after the spin_lock() but only if it was
2430 * already changed before the spin_lock(). It cannot,
2431 * however, change back to the original value. Therefore
2432 * we can detect whether we acquired the correct lock.
2434 if (unlikely(lock_ptr != q->lock_ptr)) {
2435 spin_unlock(lock_ptr);
2440 BUG_ON(q->pi_state);
2442 spin_unlock(lock_ptr);
2446 drop_futex_key_refs(&q->key);
2451 * PI futexes can not be requeued and must remove themself from the
2452 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2455 static void unqueue_me_pi(struct futex_q *q)
2456 __releases(q->lock_ptr)
2460 BUG_ON(!q->pi_state);
2461 put_pi_state(q->pi_state);
2464 spin_unlock(q->lock_ptr);
2467 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2468 struct task_struct *argowner)
2470 u32 uval, uninitialized_var(curval), newval, newtid;
2471 struct futex_pi_state *pi_state = q->pi_state;
2472 struct task_struct *oldowner, *newowner;
2475 oldowner = pi_state->owner;
2478 * We are here because either:
2480 * - we stole the lock and pi_state->owner needs updating to reflect
2481 * that (@argowner == current),
2485 * - someone stole our lock and we need to fix things to point to the
2486 * new owner (@argowner == NULL).
2488 * Either way, we have to replace the TID in the user space variable.
2489 * This must be atomic as we have to preserve the owner died bit here.
2491 * Note: We write the user space value _before_ changing the pi_state
2492 * because we can fault here. Imagine swapped out pages or a fork
2493 * that marked all the anonymous memory readonly for cow.
2495 * Modifying pi_state _before_ the user space value would leave the
2496 * pi_state in an inconsistent state when we fault here, because we
2497 * need to drop the locks to handle the fault. This might be observed
2498 * in the PID check in lookup_pi_state.
2502 if (oldowner != current) {
2504 * We raced against a concurrent self; things are
2505 * already fixed up. Nothing to do.
2510 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2511 /* We got the lock. pi_state is correct. Tell caller. */
2516 * The trylock just failed, so either there is an owner or
2517 * there is a higher priority waiter than this one.
2519 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2521 * If the higher priority waiter has not yet taken over the
2522 * rtmutex then newowner is NULL. We can't return here with
2523 * that state because it's inconsistent vs. the user space
2524 * state. So drop the locks and try again. It's a valid
2525 * situation and not any different from the other retry
2528 if (unlikely(!newowner)) {
2533 WARN_ON_ONCE(argowner != current);
2534 if (oldowner == current) {
2536 * We raced against a concurrent self; things are
2537 * already fixed up. Nothing to do.
2541 newowner = argowner;
2544 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2546 if (!pi_state->owner)
2547 newtid |= FUTEX_OWNER_DIED;
2549 err = get_futex_value_locked(&uval, uaddr);
2554 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2556 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2566 * We fixed up user space. Now we need to fix the pi_state
2569 pi_state_update_owner(pi_state, newowner);
2571 return argowner == current;
2574 * In order to reschedule or handle a page fault, we need to drop the
2575 * locks here. In the case of a fault, this gives the other task
2576 * (either the highest priority waiter itself or the task which stole
2577 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2578 * are back from handling the fault we need to check the pi_state after
2579 * reacquiring the locks and before trying to do another fixup. When
2580 * the fixup has been done already we simply return.
2582 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2583 * drop hb->lock since the caller owns the hb -> futex_q relation.
2584 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2587 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2588 spin_unlock(q->lock_ptr);
2592 err = fault_in_user_writeable(uaddr);
2605 spin_lock(q->lock_ptr);
2606 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2609 * Check if someone else fixed it for us:
2611 if (pi_state->owner != oldowner)
2612 return argowner == current;
2614 /* Retry if err was -EAGAIN or the fault in succeeded */
2619 * fault_in_user_writeable() failed so user state is immutable. At
2620 * best we can make the kernel state consistent but user state will
2621 * be most likely hosed and any subsequent unlock operation will be
2622 * rejected due to PI futex rule [10].
2624 * Ensure that the rtmutex owner is also the pi_state owner despite
2625 * the user space value claiming something different. There is no
2626 * point in unlocking the rtmutex if current is the owner as it
2627 * would need to wait until the next waiter has taken the rtmutex
2628 * to guarantee consistent state. Keep it simple. Userspace asked
2629 * for this wreckaged state.
2631 * The rtmutex has an owner - either current or some other
2632 * task. See the EAGAIN loop above.
2634 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2639 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2640 struct task_struct *argowner)
2642 struct futex_pi_state *pi_state = q->pi_state;
2645 lockdep_assert_held(q->lock_ptr);
2647 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2648 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2649 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2653 static long futex_wait_restart(struct restart_block *restart);
2656 * fixup_owner() - Post lock pi_state and corner case management
2657 * @uaddr: user address of the futex
2658 * @q: futex_q (contains pi_state and access to the rt_mutex)
2659 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2661 * After attempting to lock an rt_mutex, this function is called to cleanup
2662 * the pi_state owner as well as handle race conditions that may allow us to
2663 * acquire the lock. Must be called with the hb lock held.
2666 * - 1 - success, lock taken;
2667 * - 0 - success, lock not taken;
2668 * - <0 - on error (-EFAULT)
2670 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2674 * Got the lock. We might not be the anticipated owner if we
2675 * did a lock-steal - fix up the PI-state in that case:
2677 * Speculative pi_state->owner read (we don't hold wait_lock);
2678 * since we own the lock pi_state->owner == current is the
2679 * stable state, anything else needs more attention.
2681 if (q->pi_state->owner != current)
2682 return fixup_pi_state_owner(uaddr, q, current);
2687 * If we didn't get the lock; check if anybody stole it from us. In
2688 * that case, we need to fix up the uval to point to them instead of
2689 * us, otherwise bad things happen. [10]
2691 * Another speculative read; pi_state->owner == current is unstable
2692 * but needs our attention.
2694 if (q->pi_state->owner == current)
2695 return fixup_pi_state_owner(uaddr, q, NULL);
2698 * Paranoia check. If we did not take the lock, then we should not be
2699 * the owner of the rt_mutex. Warn and establish consistent state.
2701 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2702 return fixup_pi_state_owner(uaddr, q, current);
2708 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2709 * @hb: the futex hash bucket, must be locked by the caller
2710 * @q: the futex_q to queue up on
2711 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2713 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2714 struct hrtimer_sleeper *timeout)
2717 * The task state is guaranteed to be set before another task can
2718 * wake it. set_current_state() is implemented using smp_store_mb() and
2719 * queue_me() calls spin_unlock() upon completion, both serializing
2720 * access to the hash list and forcing another memory barrier.
2722 set_current_state(TASK_INTERRUPTIBLE);
2727 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2730 * If we have been removed from the hash list, then another task
2731 * has tried to wake us, and we can skip the call to schedule().
2733 if (likely(!plist_node_empty(&q->list))) {
2735 * If the timer has already expired, current will already be
2736 * flagged for rescheduling. Only call schedule if there
2737 * is no timeout, or if it has yet to expire.
2739 if (!timeout || timeout->task)
2740 freezable_schedule();
2742 __set_current_state(TASK_RUNNING);
2746 * futex_wait_setup() - Prepare to wait on a futex
2747 * @uaddr: the futex userspace address
2748 * @val: the expected value
2749 * @flags: futex flags (FLAGS_SHARED, etc.)
2750 * @q: the associated futex_q
2751 * @hb: storage for hash_bucket pointer to be returned to caller
2753 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2754 * compare it with the expected value. Handle atomic faults internally.
2755 * Return with the hb lock held and a q.key reference on success, and unlocked
2756 * with no q.key reference on failure.
2759 * - 0 - uaddr contains val and hb has been locked;
2760 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2762 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2763 struct futex_q *q, struct futex_hash_bucket **hb)
2769 * Access the page AFTER the hash-bucket is locked.
2770 * Order is important:
2772 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2773 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2775 * The basic logical guarantee of a futex is that it blocks ONLY
2776 * if cond(var) is known to be true at the time of blocking, for
2777 * any cond. If we locked the hash-bucket after testing *uaddr, that
2778 * would open a race condition where we could block indefinitely with
2779 * cond(var) false, which would violate the guarantee.
2781 * On the other hand, we insert q and release the hash-bucket only
2782 * after testing *uaddr. This guarantees that futex_wait() will NOT
2783 * absorb a wakeup if *uaddr does not match the desired values
2784 * while the syscall executes.
2787 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2788 if (unlikely(ret != 0))
2792 *hb = queue_lock(q);
2794 ret = get_futex_value_locked(&uval, uaddr);
2799 ret = get_user(uval, uaddr);
2803 if (!(flags & FLAGS_SHARED))
2806 put_futex_key(&q->key);
2817 put_futex_key(&q->key);
2821 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2822 ktime_t *abs_time, u32 bitset)
2824 struct hrtimer_sleeper timeout, *to = NULL;
2825 struct restart_block *restart;
2826 struct futex_hash_bucket *hb;
2827 struct futex_q q = futex_q_init;
2837 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2838 CLOCK_REALTIME : CLOCK_MONOTONIC,
2840 hrtimer_init_sleeper(to, current);
2841 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2842 current->timer_slack_ns);
2847 * Prepare to wait on uaddr. On success, holds hb lock and increments
2850 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2854 /* queue_me and wait for wakeup, timeout, or a signal. */
2855 futex_wait_queue_me(hb, &q, to);
2857 /* If we were woken (and unqueued), we succeeded, whatever. */
2859 /* unqueue_me() drops q.key ref */
2860 if (!unqueue_me(&q))
2863 if (to && !to->task)
2867 * We expect signal_pending(current), but we might be the
2868 * victim of a spurious wakeup as well.
2870 if (!signal_pending(current))
2877 restart = ¤t->restart_block;
2878 restart->futex.uaddr = uaddr;
2879 restart->futex.val = val;
2880 restart->futex.time = *abs_time;
2881 restart->futex.bitset = bitset;
2882 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2884 ret = set_restart_fn(restart, futex_wait_restart);
2888 hrtimer_cancel(&to->timer);
2889 destroy_hrtimer_on_stack(&to->timer);
2895 static long futex_wait_restart(struct restart_block *restart)
2897 u32 __user *uaddr = restart->futex.uaddr;
2898 ktime_t t, *tp = NULL;
2900 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2901 t = restart->futex.time;
2904 restart->fn = do_no_restart_syscall;
2906 return (long)futex_wait(uaddr, restart->futex.flags,
2907 restart->futex.val, tp, restart->futex.bitset);
2912 * Userspace tried a 0 -> TID atomic transition of the futex value
2913 * and failed. The kernel side here does the whole locking operation:
2914 * if there are waiters then it will block as a consequence of relying
2915 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2916 * a 0 value of the futex too.).
2918 * Also serves as futex trylock_pi()'ing, and due semantics.
2920 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2921 ktime_t *time, int trylock)
2923 struct hrtimer_sleeper timeout, *to = NULL;
2924 struct task_struct *exiting = NULL;
2925 struct rt_mutex_waiter rt_waiter;
2926 struct futex_hash_bucket *hb;
2927 struct futex_q q = futex_q_init;
2930 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2933 if (refill_pi_state_cache())
2938 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2940 hrtimer_init_sleeper(to, current);
2941 hrtimer_set_expires(&to->timer, *time);
2945 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2946 if (unlikely(ret != 0))
2950 hb = queue_lock(&q);
2952 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2954 if (unlikely(ret)) {
2956 * Atomic work succeeded and we got the lock,
2957 * or failed. Either way, we do _not_ block.
2961 /* We got the lock. */
2963 goto out_unlock_put_key;
2969 * Two reasons for this:
2970 * - EBUSY: Task is exiting and we just wait for the
2972 * - EAGAIN: The user space value changed.
2975 put_futex_key(&q.key);
2977 * Handle the case where the owner is in the middle of
2978 * exiting. Wait for the exit to complete otherwise
2979 * this task might loop forever, aka. live lock.
2981 wait_for_owner_exiting(ret, exiting);
2985 goto out_unlock_put_key;
2989 WARN_ON(!q.pi_state);
2992 * Only actually queue now that the atomic ops are done:
2997 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2998 /* Fixup the trylock return value: */
2999 ret = ret ? 0 : -EWOULDBLOCK;
3003 rt_mutex_init_waiter(&rt_waiter);
3006 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
3007 * hold it while doing rt_mutex_start_proxy(), because then it will
3008 * include hb->lock in the blocking chain, even through we'll not in
3009 * fact hold it while blocking. This will lead it to report -EDEADLK
3010 * and BUG when futex_unlock_pi() interleaves with this.
3012 * Therefore acquire wait_lock while holding hb->lock, but drop the
3013 * latter before calling __rt_mutex_start_proxy_lock(). This
3014 * interleaves with futex_unlock_pi() -- which does a similar lock
3015 * handoff -- such that the latter can observe the futex_q::pi_state
3016 * before __rt_mutex_start_proxy_lock() is done.
3018 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
3019 spin_unlock(q.lock_ptr);
3021 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
3022 * such that futex_unlock_pi() is guaranteed to observe the waiter when
3023 * it sees the futex_q::pi_state.
3025 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
3026 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
3035 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
3037 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
3040 spin_lock(q.lock_ptr);
3042 * If we failed to acquire the lock (deadlock/signal/timeout), we must
3043 * first acquire the hb->lock before removing the lock from the
3044 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
3047 * In particular; it is important that futex_unlock_pi() can not
3048 * observe this inconsistency.
3050 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
3055 * Fixup the pi_state owner and possibly acquire the lock if we
3058 res = fixup_owner(uaddr, &q, !ret);
3060 * If fixup_owner() returned an error, proprogate that. If it acquired
3061 * the lock, clear our -ETIMEDOUT or -EINTR.
3064 ret = (res < 0) ? res : 0;
3066 /* Unqueue and drop the lock */
3075 put_futex_key(&q.key);
3078 hrtimer_cancel(&to->timer);
3079 destroy_hrtimer_on_stack(&to->timer);
3081 return ret != -EINTR ? ret : -ERESTARTNOINTR;
3086 ret = fault_in_user_writeable(uaddr);
3090 if (!(flags & FLAGS_SHARED))
3093 put_futex_key(&q.key);
3098 * Userspace attempted a TID -> 0 atomic transition, and failed.
3099 * This is the in-kernel slowpath: we look up the PI state (if any),
3100 * and do the rt-mutex unlock.
3102 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
3104 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
3105 union futex_key key = FUTEX_KEY_INIT;
3106 struct futex_hash_bucket *hb;
3107 struct futex_q *top_waiter;
3110 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3114 if (get_user(uval, uaddr))
3117 * We release only a lock we actually own:
3119 if ((uval & FUTEX_TID_MASK) != vpid)
3122 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
3126 hb = hash_futex(&key);
3127 spin_lock(&hb->lock);
3130 * Check waiters first. We do not trust user space values at
3131 * all and we at least want to know if user space fiddled
3132 * with the futex value instead of blindly unlocking.
3134 top_waiter = futex_top_waiter(hb, &key);
3136 struct futex_pi_state *pi_state = top_waiter->pi_state;
3143 * If current does not own the pi_state then the futex is
3144 * inconsistent and user space fiddled with the futex value.
3146 if (pi_state->owner != current)
3149 get_pi_state(pi_state);
3151 * By taking wait_lock while still holding hb->lock, we ensure
3152 * there is no point where we hold neither; and therefore
3153 * wake_futex_pi() must observe a state consistent with what we
3156 * In particular; this forces __rt_mutex_start_proxy() to
3157 * complete such that we're guaranteed to observe the
3158 * rt_waiter. Also see the WARN in wake_futex_pi().
3160 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3161 spin_unlock(&hb->lock);
3163 /* drops pi_state->pi_mutex.wait_lock */
3164 ret = wake_futex_pi(uaddr, uval, pi_state);
3166 put_pi_state(pi_state);
3169 * Success, we're done! No tricky corner cases.
3174 * The atomic access to the futex value generated a
3175 * pagefault, so retry the user-access and the wakeup:
3180 * A unconditional UNLOCK_PI op raced against a waiter
3181 * setting the FUTEX_WAITERS bit. Try again.
3186 * wake_futex_pi has detected invalid state. Tell user
3193 * We have no kernel internal state, i.e. no waiters in the
3194 * kernel. Waiters which are about to queue themselves are stuck
3195 * on hb->lock. So we can safely ignore them. We do neither
3196 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3199 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3200 spin_unlock(&hb->lock);
3215 * If uval has changed, let user space handle it.
3217 ret = (curval == uval) ? 0 : -EAGAIN;
3220 spin_unlock(&hb->lock);
3222 put_futex_key(&key);
3226 put_futex_key(&key);
3231 put_futex_key(&key);
3233 ret = fault_in_user_writeable(uaddr);
3241 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3242 * @hb: the hash_bucket futex_q was original enqueued on
3243 * @q: the futex_q woken while waiting to be requeued
3244 * @key2: the futex_key of the requeue target futex
3245 * @timeout: the timeout associated with the wait (NULL if none)
3247 * Detect if the task was woken on the initial futex as opposed to the requeue
3248 * target futex. If so, determine if it was a timeout or a signal that caused
3249 * the wakeup and return the appropriate error code to the caller. Must be
3250 * called with the hb lock held.
3253 * - 0 = no early wakeup detected;
3254 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3257 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3258 struct futex_q *q, union futex_key *key2,
3259 struct hrtimer_sleeper *timeout)
3264 * With the hb lock held, we avoid races while we process the wakeup.
3265 * We only need to hold hb (and not hb2) to ensure atomicity as the
3266 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3267 * It can't be requeued from uaddr2 to something else since we don't
3268 * support a PI aware source futex for requeue.
3270 if (!match_futex(&q->key, key2)) {
3271 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3273 * We were woken prior to requeue by a timeout or a signal.
3274 * Unqueue the futex_q and determine which it was.
3276 plist_del(&q->list, &hb->chain);
3279 /* Handle spurious wakeups gracefully */
3281 if (timeout && !timeout->task)
3283 else if (signal_pending(current))
3284 ret = -ERESTARTNOINTR;
3290 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3291 * @uaddr: the futex we initially wait on (non-pi)
3292 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3293 * the same type, no requeueing from private to shared, etc.
3294 * @val: the expected value of uaddr
3295 * @abs_time: absolute timeout
3296 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3297 * @uaddr2: the pi futex we will take prior to returning to user-space
3299 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3300 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3301 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3302 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3303 * without one, the pi logic would not know which task to boost/deboost, if
3304 * there was a need to.
3306 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3307 * via the following--
3308 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3309 * 2) wakeup on uaddr2 after a requeue
3313 * If 3, cleanup and return -ERESTARTNOINTR.
3315 * If 2, we may then block on trying to take the rt_mutex and return via:
3316 * 5) successful lock
3319 * 8) other lock acquisition failure
3321 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3323 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3329 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3330 u32 val, ktime_t *abs_time, u32 bitset,
3333 struct hrtimer_sleeper timeout, *to = NULL;
3334 struct rt_mutex_waiter rt_waiter;
3335 struct futex_hash_bucket *hb;
3336 union futex_key key2 = FUTEX_KEY_INIT;
3337 struct futex_q q = futex_q_init;
3340 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3343 if (uaddr == uaddr2)
3351 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3352 CLOCK_REALTIME : CLOCK_MONOTONIC,
3354 hrtimer_init_sleeper(to, current);
3355 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3356 current->timer_slack_ns);
3360 * The waiter is allocated on our stack, manipulated by the requeue
3361 * code while we sleep on uaddr.
3363 rt_mutex_init_waiter(&rt_waiter);
3365 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3366 if (unlikely(ret != 0))
3370 q.rt_waiter = &rt_waiter;
3371 q.requeue_pi_key = &key2;
3374 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3377 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3382 * The check above which compares uaddrs is not sufficient for
3383 * shared futexes. We need to compare the keys:
3385 if (match_futex(&q.key, &key2)) {
3391 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3392 futex_wait_queue_me(hb, &q, to);
3394 spin_lock(&hb->lock);
3395 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3396 spin_unlock(&hb->lock);
3401 * In order for us to be here, we know our q.key == key2, and since
3402 * we took the hb->lock above, we also know that futex_requeue() has
3403 * completed and we no longer have to concern ourselves with a wakeup
3404 * race with the atomic proxy lock acquisition by the requeue code. The
3405 * futex_requeue dropped our key1 reference and incremented our key2
3409 /* Check if the requeue code acquired the second futex for us. */
3412 * Got the lock. We might not be the anticipated owner if we
3413 * did a lock-steal - fix up the PI-state in that case.
3415 if (q.pi_state && (q.pi_state->owner != current)) {
3416 spin_lock(q.lock_ptr);
3417 ret = fixup_pi_state_owner(uaddr2, &q, current);
3419 * Drop the reference to the pi state which
3420 * the requeue_pi() code acquired for us.
3422 put_pi_state(q.pi_state);
3423 spin_unlock(q.lock_ptr);
3425 * Adjust the return value. It's either -EFAULT or
3426 * success (1) but the caller expects 0 for success.
3428 ret = ret < 0 ? ret : 0;
3431 struct rt_mutex *pi_mutex;
3434 * We have been woken up by futex_unlock_pi(), a timeout, or a
3435 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3438 WARN_ON(!q.pi_state);
3439 pi_mutex = &q.pi_state->pi_mutex;
3440 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3442 spin_lock(q.lock_ptr);
3443 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3446 debug_rt_mutex_free_waiter(&rt_waiter);
3448 * Fixup the pi_state owner and possibly acquire the lock if we
3451 res = fixup_owner(uaddr2, &q, !ret);
3453 * If fixup_owner() returned an error, proprogate that. If it
3454 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3457 ret = (res < 0) ? res : 0;
3459 /* Unqueue and drop the lock. */
3463 if (ret == -EINTR) {
3465 * We've already been requeued, but cannot restart by calling
3466 * futex_lock_pi() directly. We could restart this syscall, but
3467 * it would detect that the user space "val" changed and return
3468 * -EWOULDBLOCK. Save the overhead of the restart and return
3469 * -EWOULDBLOCK directly.
3475 put_futex_key(&q.key);
3477 put_futex_key(&key2);
3481 hrtimer_cancel(&to->timer);
3482 destroy_hrtimer_on_stack(&to->timer);
3488 * Support for robust futexes: the kernel cleans up held futexes at
3491 * Implementation: user-space maintains a per-thread list of locks it
3492 * is holding. Upon do_exit(), the kernel carefully walks this list,
3493 * and marks all locks that are owned by this thread with the
3494 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3495 * always manipulated with the lock held, so the list is private and
3496 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3497 * field, to allow the kernel to clean up if the thread dies after
3498 * acquiring the lock, but just before it could have added itself to
3499 * the list. There can only be one such pending lock.
3503 * sys_set_robust_list() - Set the robust-futex list head of a task
3504 * @head: pointer to the list-head
3505 * @len: length of the list-head, as userspace expects
3507 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3510 if (!futex_cmpxchg_enabled)
3513 * The kernel knows only one size for now:
3515 if (unlikely(len != sizeof(*head)))
3518 current->robust_list = head;
3524 * sys_get_robust_list() - Get the robust-futex list head of a task
3525 * @pid: pid of the process [zero for current task]
3526 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3527 * @len_ptr: pointer to a length field, the kernel fills in the header size
3529 SYSCALL_DEFINE3(get_robust_list, int, pid,
3530 struct robust_list_head __user * __user *, head_ptr,
3531 size_t __user *, len_ptr)
3533 struct robust_list_head __user *head;
3535 struct task_struct *p;
3537 if (!futex_cmpxchg_enabled)
3546 p = find_task_by_vpid(pid);
3552 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3555 head = p->robust_list;
3558 if (put_user(sizeof(*head), len_ptr))
3560 return put_user(head, head_ptr);
3568 /* Constants for the pending_op argument of handle_futex_death */
3569 #define HANDLE_DEATH_PENDING true
3570 #define HANDLE_DEATH_LIST false
3573 * Process a futex-list entry, check whether it's owned by the
3574 * dying task, and do notification if so:
3576 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3577 bool pi, bool pending_op)
3579 u32 uval, uninitialized_var(nval), mval;
3582 /* Futex address must be 32bit aligned */
3583 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3587 if (get_user(uval, uaddr))
3591 * Special case for regular (non PI) futexes. The unlock path in
3592 * user space has two race scenarios:
3594 * 1. The unlock path releases the user space futex value and
3595 * before it can execute the futex() syscall to wake up
3596 * waiters it is killed.
3598 * 2. A woken up waiter is killed before it can acquire the
3599 * futex in user space.
3601 * In both cases the TID validation below prevents a wakeup of
3602 * potential waiters which can cause these waiters to block
3605 * In both cases the following conditions are met:
3607 * 1) task->robust_list->list_op_pending != NULL
3608 * @pending_op == true
3609 * 2) User space futex value == 0
3610 * 3) Regular futex: @pi == false
3612 * If these conditions are met, it is safe to attempt waking up a
3613 * potential waiter without touching the user space futex value and
3614 * trying to set the OWNER_DIED bit. The user space futex value is
3615 * uncontended and the rest of the user space mutex state is
3616 * consistent, so a woken waiter will just take over the
3617 * uncontended futex. Setting the OWNER_DIED bit would create
3618 * inconsistent state and malfunction of the user space owner died
3621 if (pending_op && !pi && !uval) {
3622 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3626 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3630 * Ok, this dying thread is truly holding a futex
3631 * of interest. Set the OWNER_DIED bit atomically
3632 * via cmpxchg, and if the value had FUTEX_WAITERS
3633 * set, wake up a waiter (if any). (We have to do a
3634 * futex_wake() even if OWNER_DIED is already set -
3635 * to handle the rare but possible case of recursive
3636 * thread-death.) The rest of the cleanup is done in
3639 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3642 * We are not holding a lock here, but we want to have
3643 * the pagefault_disable/enable() protection because
3644 * we want to handle the fault gracefully. If the
3645 * access fails we try to fault in the futex with R/W
3646 * verification via get_user_pages. get_user() above
3647 * does not guarantee R/W access. If that fails we
3648 * give up and leave the futex locked.
3650 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3653 if (fault_in_user_writeable(uaddr))
3671 * Wake robust non-PI futexes here. The wakeup of
3672 * PI futexes happens in exit_pi_state():
3674 if (!pi && (uval & FUTEX_WAITERS))
3675 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3681 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3683 static inline int fetch_robust_entry(struct robust_list __user **entry,
3684 struct robust_list __user * __user *head,
3687 unsigned long uentry;
3689 if (get_user(uentry, (unsigned long __user *)head))
3692 *entry = (void __user *)(uentry & ~1UL);
3699 * Walk curr->robust_list (very carefully, it's a userspace list!)
3700 * and mark any locks found there dead, and notify any waiters.
3702 * We silently return on any sign of list-walking problem.
3704 static void exit_robust_list(struct task_struct *curr)
3706 struct robust_list_head __user *head = curr->robust_list;
3707 struct robust_list __user *entry, *next_entry, *pending;
3708 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3709 unsigned int uninitialized_var(next_pi);
3710 unsigned long futex_offset;
3713 if (!futex_cmpxchg_enabled)
3717 * Fetch the list head (which was registered earlier, via
3718 * sys_set_robust_list()):
3720 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3723 * Fetch the relative futex offset:
3725 if (get_user(futex_offset, &head->futex_offset))
3728 * Fetch any possibly pending lock-add first, and handle it
3731 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3734 next_entry = NULL; /* avoid warning with gcc */
3735 while (entry != &head->list) {
3737 * Fetch the next entry in the list before calling
3738 * handle_futex_death:
3740 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3742 * A pending lock might already be on the list, so
3743 * don't process it twice:
3745 if (entry != pending) {
3746 if (handle_futex_death((void __user *)entry + futex_offset,
3747 curr, pi, HANDLE_DEATH_LIST))
3755 * Avoid excessively long or circular lists:
3764 handle_futex_death((void __user *)pending + futex_offset,
3765 curr, pip, HANDLE_DEATH_PENDING);
3769 static void futex_cleanup(struct task_struct *tsk)
3771 if (unlikely(tsk->robust_list)) {
3772 exit_robust_list(tsk);
3773 tsk->robust_list = NULL;
3776 #ifdef CONFIG_COMPAT
3777 if (unlikely(tsk->compat_robust_list)) {
3778 compat_exit_robust_list(tsk);
3779 tsk->compat_robust_list = NULL;
3783 if (unlikely(!list_empty(&tsk->pi_state_list)))
3784 exit_pi_state_list(tsk);
3788 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3789 * @tsk: task to set the state on
3791 * Set the futex exit state of the task lockless. The futex waiter code
3792 * observes that state when a task is exiting and loops until the task has
3793 * actually finished the futex cleanup. The worst case for this is that the
3794 * waiter runs through the wait loop until the state becomes visible.
3796 * This is called from the recursive fault handling path in do_exit().
3798 * This is best effort. Either the futex exit code has run already or
3799 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3800 * take it over. If not, the problem is pushed back to user space. If the
3801 * futex exit code did not run yet, then an already queued waiter might
3802 * block forever, but there is nothing which can be done about that.
3804 void futex_exit_recursive(struct task_struct *tsk)
3806 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3807 if (tsk->futex_state == FUTEX_STATE_EXITING)
3808 mutex_unlock(&tsk->futex_exit_mutex);
3809 tsk->futex_state = FUTEX_STATE_DEAD;
3812 static void futex_cleanup_begin(struct task_struct *tsk)
3815 * Prevent various race issues against a concurrent incoming waiter
3816 * including live locks by forcing the waiter to block on
3817 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3818 * attach_to_pi_owner().
3820 mutex_lock(&tsk->futex_exit_mutex);
3823 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3825 * This ensures that all subsequent checks of tsk->futex_state in
3826 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3827 * tsk->pi_lock held.
3829 * It guarantees also that a pi_state which was queued right before
3830 * the state change under tsk->pi_lock by a concurrent waiter must
3831 * be observed in exit_pi_state_list().
3833 raw_spin_lock_irq(&tsk->pi_lock);
3834 tsk->futex_state = FUTEX_STATE_EXITING;
3835 raw_spin_unlock_irq(&tsk->pi_lock);
3838 static void futex_cleanup_end(struct task_struct *tsk, int state)
3841 * Lockless store. The only side effect is that an observer might
3842 * take another loop until it becomes visible.
3844 tsk->futex_state = state;
3846 * Drop the exit protection. This unblocks waiters which observed
3847 * FUTEX_STATE_EXITING to reevaluate the state.
3849 mutex_unlock(&tsk->futex_exit_mutex);
3852 void futex_exec_release(struct task_struct *tsk)
3855 * The state handling is done for consistency, but in the case of
3856 * exec() there is no way to prevent futher damage as the PID stays
3857 * the same. But for the unlikely and arguably buggy case that a
3858 * futex is held on exec(), this provides at least as much state
3859 * consistency protection which is possible.
3861 futex_cleanup_begin(tsk);
3864 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3865 * exec a new binary.
3867 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3870 void futex_exit_release(struct task_struct *tsk)
3872 futex_cleanup_begin(tsk);
3874 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3877 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3878 u32 __user *uaddr2, u32 val2, u32 val3)
3880 int cmd = op & FUTEX_CMD_MASK;
3881 unsigned int flags = 0;
3883 if (!(op & FUTEX_PRIVATE_FLAG))
3884 flags |= FLAGS_SHARED;
3886 if (op & FUTEX_CLOCK_REALTIME) {
3887 flags |= FLAGS_CLOCKRT;
3888 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3894 case FUTEX_UNLOCK_PI:
3895 case FUTEX_TRYLOCK_PI:
3896 case FUTEX_WAIT_REQUEUE_PI:
3897 case FUTEX_CMP_REQUEUE_PI:
3898 if (!futex_cmpxchg_enabled)
3904 val3 = FUTEX_BITSET_MATCH_ANY;
3905 case FUTEX_WAIT_BITSET:
3906 return futex_wait(uaddr, flags, val, timeout, val3);
3908 val3 = FUTEX_BITSET_MATCH_ANY;
3909 case FUTEX_WAKE_BITSET:
3910 return futex_wake(uaddr, flags, val, val3);
3912 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3913 case FUTEX_CMP_REQUEUE:
3914 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3916 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3918 return futex_lock_pi(uaddr, flags, timeout, 0);
3919 case FUTEX_UNLOCK_PI:
3920 return futex_unlock_pi(uaddr, flags);
3921 case FUTEX_TRYLOCK_PI:
3922 return futex_lock_pi(uaddr, flags, NULL, 1);
3923 case FUTEX_WAIT_REQUEUE_PI:
3924 val3 = FUTEX_BITSET_MATCH_ANY;
3925 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3927 case FUTEX_CMP_REQUEUE_PI:
3928 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3934 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3935 struct timespec __user *, utime, u32 __user *, uaddr2,
3939 ktime_t t, *tp = NULL;
3941 int cmd = op & FUTEX_CMD_MASK;
3943 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3944 cmd == FUTEX_WAIT_BITSET ||
3945 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3946 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3948 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3950 if (!timespec_valid(&ts))
3953 t = timespec_to_ktime(ts);
3954 if (cmd == FUTEX_WAIT)
3955 t = ktime_add_safe(ktime_get(), t);
3959 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3960 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3962 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3963 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3964 val2 = (u32) (unsigned long) utime;
3966 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3969 #ifdef CONFIG_COMPAT
3971 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3974 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3975 compat_uptr_t __user *head, unsigned int *pi)
3977 if (get_user(*uentry, head))
3980 *entry = compat_ptr((*uentry) & ~1);
3981 *pi = (unsigned int)(*uentry) & 1;
3986 static void __user *futex_uaddr(struct robust_list __user *entry,
3987 compat_long_t futex_offset)
3989 compat_uptr_t base = ptr_to_compat(entry);
3990 void __user *uaddr = compat_ptr(base + futex_offset);
3996 * Walk curr->robust_list (very carefully, it's a userspace list!)
3997 * and mark any locks found there dead, and notify any waiters.
3999 * We silently return on any sign of list-walking problem.
4001 static void compat_exit_robust_list(struct task_struct *curr)
4003 struct compat_robust_list_head __user *head = curr->compat_robust_list;
4004 struct robust_list __user *entry, *next_entry, *pending;
4005 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
4006 unsigned int uninitialized_var(next_pi);
4007 compat_uptr_t uentry, next_uentry, upending;
4008 compat_long_t futex_offset;
4011 if (!futex_cmpxchg_enabled)
4015 * Fetch the list head (which was registered earlier, via
4016 * sys_set_robust_list()):
4018 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
4021 * Fetch the relative futex offset:
4023 if (get_user(futex_offset, &head->futex_offset))
4026 * Fetch any possibly pending lock-add first, and handle it
4029 if (compat_fetch_robust_entry(&upending, &pending,
4030 &head->list_op_pending, &pip))
4033 next_entry = NULL; /* avoid warning with gcc */
4034 while (entry != (struct robust_list __user *) &head->list) {
4036 * Fetch the next entry in the list before calling
4037 * handle_futex_death:
4039 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
4040 (compat_uptr_t __user *)&entry->next, &next_pi);
4042 * A pending lock might already be on the list, so
4043 * dont process it twice:
4045 if (entry != pending) {
4046 void __user *uaddr = futex_uaddr(entry, futex_offset);
4048 if (handle_futex_death(uaddr, curr, pi,
4054 uentry = next_uentry;
4058 * Avoid excessively long or circular lists:
4066 void __user *uaddr = futex_uaddr(pending, futex_offset);
4068 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
4072 COMPAT_SYSCALL_DEFINE2(set_robust_list,
4073 struct compat_robust_list_head __user *, head,
4076 if (!futex_cmpxchg_enabled)
4079 if (unlikely(len != sizeof(*head)))
4082 current->compat_robust_list = head;
4087 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
4088 compat_uptr_t __user *, head_ptr,
4089 compat_size_t __user *, len_ptr)
4091 struct compat_robust_list_head __user *head;
4093 struct task_struct *p;
4095 if (!futex_cmpxchg_enabled)
4104 p = find_task_by_vpid(pid);
4110 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
4113 head = p->compat_robust_list;
4116 if (put_user(sizeof(*head), len_ptr))
4118 return put_user(ptr_to_compat(head), head_ptr);
4126 COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
4127 struct compat_timespec __user *, utime, u32 __user *, uaddr2,
4131 ktime_t t, *tp = NULL;
4133 int cmd = op & FUTEX_CMD_MASK;
4135 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4136 cmd == FUTEX_WAIT_BITSET ||
4137 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4138 if (compat_get_timespec(&ts, utime))
4140 if (!timespec_valid(&ts))
4143 t = timespec_to_ktime(ts);
4144 if (cmd == FUTEX_WAIT)
4145 t = ktime_add_safe(ktime_get(), t);
4148 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4149 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4150 val2 = (int) (unsigned long) utime;
4152 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4154 #endif /* CONFIG_COMPAT */
4156 static void __init futex_detect_cmpxchg(void)
4158 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4162 * This will fail and we want it. Some arch implementations do
4163 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4164 * functionality. We want to know that before we call in any
4165 * of the complex code paths. Also we want to prevent
4166 * registration of robust lists in that case. NULL is
4167 * guaranteed to fault and we get -EFAULT on functional
4168 * implementation, the non-functional ones will return
4171 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4172 futex_cmpxchg_enabled = 1;
4176 static int __init futex_init(void)
4178 unsigned int futex_shift;
4181 #if CONFIG_BASE_SMALL
4182 futex_hashsize = 16;
4184 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4187 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4189 futex_hashsize < 256 ? HASH_SMALL : 0,
4191 futex_hashsize, futex_hashsize);
4192 futex_hashsize = 1UL << futex_shift;
4194 futex_detect_cmpxchg();
4196 for (i = 0; i < futex_hashsize; i++) {
4197 atomic_set(&futex_queues[i].waiters, 0);
4198 plist_head_init(&futex_queues[i].chain);
4199 spin_lock_init(&futex_queues[i].lock);
4204 core_initcall(futex_init);