1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/syscalls.h>
38 #include <linux/freezer.h>
39 #include <linux/memblock.h>
40 #include <linux/fault-inject.h>
41 #include <linux/time_namespace.h>
43 #include <asm/futex.h>
45 #include "locking/rtmutex_common.h"
48 * READ this before attempting to hack on futexes!
50 * Basic futex operation and ordering guarantees
51 * =============================================
53 * The waiter reads the futex value in user space and calls
54 * futex_wait(). This function computes the hash bucket and acquires
55 * the hash bucket lock. After that it reads the futex user space value
56 * again and verifies that the data has not changed. If it has not changed
57 * it enqueues itself into the hash bucket, releases the hash bucket lock
60 * The waker side modifies the user space value of the futex and calls
61 * futex_wake(). This function computes the hash bucket and acquires the
62 * hash bucket lock. Then it looks for waiters on that futex in the hash
63 * bucket and wakes them.
65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
66 * the hb spinlock can be avoided and simply return. In order for this
67 * optimization to work, ordering guarantees must exist so that the waiter
68 * being added to the list is acknowledged when the list is concurrently being
69 * checked by the waker, avoiding scenarios like the following:
73 * sys_futex(WAIT, futex, val);
74 * futex_wait(futex, val);
77 * sys_futex(WAKE, futex);
82 * lock(hash_bucket(futex));
84 * unlock(hash_bucket(futex));
87 * This would cause the waiter on CPU 0 to wait forever because it
88 * missed the transition of the user space value from val to newval
89 * and the waker did not find the waiter in the hash bucket queue.
91 * The correct serialization ensures that a waiter either observes
92 * the changed user space value before blocking or is woken by a
97 * sys_futex(WAIT, futex, val);
98 * futex_wait(futex, val);
101 * smp_mb(); (A) <-- paired with -.
103 * lock(hash_bucket(futex)); |
107 * | sys_futex(WAKE, futex);
108 * | futex_wake(futex);
110 * `--------> smp_mb(); (B)
113 * unlock(hash_bucket(futex));
114 * schedule(); if (waiters)
115 * lock(hash_bucket(futex));
116 * else wake_waiters(futex);
117 * waiters--; (b) unlock(hash_bucket(futex));
119 * Where (A) orders the waiters increment and the futex value read through
120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
121 * to futex and the waiters read (see hb_waiters_pending()).
123 * This yields the following case (where X:=waiters, Y:=futex):
131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
132 * the guarantee that we cannot both miss the futex variable change and the
135 * Note that a new waiter is accounted for in (a) even when it is possible that
136 * the wait call can return error, in which case we backtrack from it in (b).
137 * Refer to the comment in queue_lock().
139 * Similarly, in order to account for waiters being requeued on another
140 * address we always increment the waiters for the destination bucket before
141 * acquiring the lock. It then decrements them again after releasing it -
142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
143 * will do the additional required waiter count housekeeping. This is done for
144 * double_lock_hb() and double_unlock_hb(), respectively.
147 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
148 #define futex_cmpxchg_enabled 1
150 static int __read_mostly futex_cmpxchg_enabled;
154 * Futex flags used to encode options to functions and preserve them across
158 # define FLAGS_SHARED 0x01
161 * NOMMU does not have per process address space. Let the compiler optimize
164 # define FLAGS_SHARED 0x00
166 #define FLAGS_CLOCKRT 0x02
167 #define FLAGS_HAS_TIMEOUT 0x04
170 * Priority Inheritance state:
172 struct futex_pi_state {
174 * list of 'owned' pi_state instances - these have to be
175 * cleaned up in do_exit() if the task exits prematurely:
177 struct list_head list;
182 struct rt_mutex pi_mutex;
184 struct task_struct *owner;
188 } __randomize_layout;
191 * struct futex_q - The hashed futex queue entry, one per waiting task
192 * @list: priority-sorted list of tasks waiting on this futex
193 * @task: the task waiting on the futex
194 * @lock_ptr: the hash bucket lock
195 * @key: the key the futex is hashed on
196 * @pi_state: optional priority inheritance state
197 * @rt_waiter: rt_waiter storage for use with requeue_pi
198 * @requeue_pi_key: the requeue_pi target futex key
199 * @bitset: bitset for the optional bitmasked wakeup
201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
202 * we can wake only the relevant ones (hashed queues may be shared).
204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
206 * The order of wakeup is always to make the first condition true, then
209 * PI futexes are typically woken before they are removed from the hash list via
210 * the rt_mutex code. See unqueue_me_pi().
213 struct plist_node list;
215 struct task_struct *task;
216 spinlock_t *lock_ptr;
218 struct futex_pi_state *pi_state;
219 struct rt_mutex_waiter *rt_waiter;
220 union futex_key *requeue_pi_key;
222 } __randomize_layout;
224 static const struct futex_q futex_q_init = {
225 /* list gets initialized in queue_me()*/
226 .key = FUTEX_KEY_INIT,
227 .bitset = FUTEX_BITSET_MATCH_ANY
231 * Hash buckets are shared by all the futex_keys that hash to the same
232 * location. Each key may have multiple futex_q structures, one for each task
233 * waiting on a futex.
235 struct futex_hash_bucket {
238 struct plist_head chain;
239 } ____cacheline_aligned_in_smp;
242 * The base of the bucket array and its size are always used together
243 * (after initialization only in hash_futex()), so ensure that they
244 * reside in the same cacheline.
247 struct futex_hash_bucket *queues;
248 unsigned long hashsize;
249 } __futex_data __read_mostly __aligned(2*sizeof(long));
250 #define futex_queues (__futex_data.queues)
251 #define futex_hashsize (__futex_data.hashsize)
255 * Fault injections for futexes.
257 #ifdef CONFIG_FAIL_FUTEX
260 struct fault_attr attr;
264 .attr = FAULT_ATTR_INITIALIZER,
265 .ignore_private = false,
268 static int __init setup_fail_futex(char *str)
270 return setup_fault_attr(&fail_futex.attr, str);
272 __setup("fail_futex=", setup_fail_futex);
274 static bool should_fail_futex(bool fshared)
276 if (fail_futex.ignore_private && !fshared)
279 return should_fail(&fail_futex.attr, 1);
282 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
284 static int __init fail_futex_debugfs(void)
286 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
289 dir = fault_create_debugfs_attr("fail_futex", NULL,
294 debugfs_create_bool("ignore-private", mode, dir,
295 &fail_futex.ignore_private);
299 late_initcall(fail_futex_debugfs);
301 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
304 static inline bool should_fail_futex(bool fshared)
308 #endif /* CONFIG_FAIL_FUTEX */
311 static void compat_exit_robust_list(struct task_struct *curr);
313 static inline void compat_exit_robust_list(struct task_struct *curr) { }
317 * Reflects a new waiter being added to the waitqueue.
319 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
322 atomic_inc(&hb->waiters);
324 * Full barrier (A), see the ordering comment above.
326 smp_mb__after_atomic();
331 * Reflects a waiter being removed from the waitqueue by wakeup
334 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
337 atomic_dec(&hb->waiters);
341 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
345 * Full barrier (B), see the ordering comment above.
348 return atomic_read(&hb->waiters);
355 * hash_futex - Return the hash bucket in the global hash
356 * @key: Pointer to the futex key for which the hash is calculated
358 * We hash on the keys returned from get_futex_key (see below) and return the
359 * corresponding hash bucket in the global hash.
361 static struct futex_hash_bucket *hash_futex(union futex_key *key)
363 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
366 return &futex_queues[hash & (futex_hashsize - 1)];
371 * match_futex - Check whether two futex keys are equal
372 * @key1: Pointer to key1
373 * @key2: Pointer to key2
375 * Return 1 if two futex_keys are equal, 0 otherwise.
377 static inline int match_futex(union futex_key *key1, union futex_key *key2)
380 && key1->both.word == key2->both.word
381 && key1->both.ptr == key2->both.ptr
382 && key1->both.offset == key2->both.offset);
391 * futex_setup_timer - set up the sleeping hrtimer.
392 * @time: ptr to the given timeout value
393 * @timeout: the hrtimer_sleeper structure to be set up
394 * @flags: futex flags
395 * @range_ns: optional range in ns
397 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
400 static inline struct hrtimer_sleeper *
401 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
402 int flags, u64 range_ns)
407 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
408 CLOCK_REALTIME : CLOCK_MONOTONIC,
411 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
412 * effectively the same as calling hrtimer_set_expires().
414 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
420 * Generate a machine wide unique identifier for this inode.
422 * This relies on u64 not wrapping in the life-time of the machine; which with
423 * 1ns resolution means almost 585 years.
425 * This further relies on the fact that a well formed program will not unmap
426 * the file while it has a (shared) futex waiting on it. This mapping will have
427 * a file reference which pins the mount and inode.
429 * If for some reason an inode gets evicted and read back in again, it will get
430 * a new sequence number and will _NOT_ match, even though it is the exact same
433 * It is important that match_futex() will never have a false-positive, esp.
434 * for PI futexes that can mess up the state. The above argues that false-negatives
435 * are only possible for malformed programs.
437 static u64 get_inode_sequence_number(struct inode *inode)
439 static atomic64_t i_seq;
442 /* Does the inode already have a sequence number? */
443 old = atomic64_read(&inode->i_sequence);
448 u64 new = atomic64_add_return(1, &i_seq);
449 if (WARN_ON_ONCE(!new))
452 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
460 * get_futex_key() - Get parameters which are the keys for a futex
461 * @uaddr: virtual address of the futex
462 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
463 * @key: address where result is stored.
464 * @rw: mapping needs to be read/write (values: FUTEX_READ,
467 * Return: a negative error code or 0
469 * The key words are stored in @key on success.
471 * For shared mappings (when @fshared), the key is:
473 * ( inode->i_sequence, page->index, offset_within_page )
475 * [ also see get_inode_sequence_number() ]
477 * For private mappings (or when !@fshared), the key is:
479 * ( current->mm, address, 0 )
481 * This allows (cross process, where applicable) identification of the futex
482 * without keeping the page pinned for the duration of the FUTEX_WAIT.
484 * lock_page() might sleep, the caller should not hold a spinlock.
486 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
487 enum futex_access rw)
489 unsigned long address = (unsigned long)uaddr;
490 struct mm_struct *mm = current->mm;
491 struct page *page, *tail;
492 struct address_space *mapping;
496 * The futex address must be "naturally" aligned.
498 key->both.offset = address % PAGE_SIZE;
499 if (unlikely((address % sizeof(u32)) != 0))
501 address -= key->both.offset;
503 if (unlikely(!access_ok(uaddr, sizeof(u32))))
506 if (unlikely(should_fail_futex(fshared)))
510 * PROCESS_PRIVATE futexes are fast.
511 * As the mm cannot disappear under us and the 'key' only needs
512 * virtual address, we dont even have to find the underlying vma.
513 * Note : We do have to check 'uaddr' is a valid user address,
514 * but access_ok() should be faster than find_vma()
517 key->private.mm = mm;
518 key->private.address = address;
523 /* Ignore any VERIFY_READ mapping (futex common case) */
524 if (unlikely(should_fail_futex(true)))
527 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
529 * If write access is not required (eg. FUTEX_WAIT), try
530 * and get read-only access.
532 if (err == -EFAULT && rw == FUTEX_READ) {
533 err = get_user_pages_fast(address, 1, 0, &page);
542 * The treatment of mapping from this point on is critical. The page
543 * lock protects many things but in this context the page lock
544 * stabilizes mapping, prevents inode freeing in the shared
545 * file-backed region case and guards against movement to swap cache.
547 * Strictly speaking the page lock is not needed in all cases being
548 * considered here and page lock forces unnecessarily serialization
549 * From this point on, mapping will be re-verified if necessary and
550 * page lock will be acquired only if it is unavoidable
552 * Mapping checks require the head page for any compound page so the
553 * head page and mapping is looked up now. For anonymous pages, it
554 * does not matter if the page splits in the future as the key is
555 * based on the address. For filesystem-backed pages, the tail is
556 * required as the index of the page determines the key. For
557 * base pages, there is no tail page and tail == page.
560 page = compound_head(page);
561 mapping = READ_ONCE(page->mapping);
564 * If page->mapping is NULL, then it cannot be a PageAnon
565 * page; but it might be the ZERO_PAGE or in the gate area or
566 * in a special mapping (all cases which we are happy to fail);
567 * or it may have been a good file page when get_user_pages_fast
568 * found it, but truncated or holepunched or subjected to
569 * invalidate_complete_page2 before we got the page lock (also
570 * cases which we are happy to fail). And we hold a reference,
571 * so refcount care in invalidate_complete_page's remove_mapping
572 * prevents drop_caches from setting mapping to NULL beneath us.
574 * The case we do have to guard against is when memory pressure made
575 * shmem_writepage move it from filecache to swapcache beneath us:
576 * an unlikely race, but we do need to retry for page->mapping.
578 if (unlikely(!mapping)) {
582 * Page lock is required to identify which special case above
583 * applies. If this is really a shmem page then the page lock
584 * will prevent unexpected transitions.
587 shmem_swizzled = PageSwapCache(page) || page->mapping;
598 * Private mappings are handled in a simple way.
600 * If the futex key is stored on an anonymous page, then the associated
601 * object is the mm which is implicitly pinned by the calling process.
603 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
604 * it's a read-only handle, it's expected that futexes attach to
605 * the object not the particular process.
607 if (PageAnon(page)) {
609 * A RO anonymous page will never change and thus doesn't make
610 * sense for futex operations.
612 if (unlikely(should_fail_futex(true)) || ro) {
617 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
618 key->private.mm = mm;
619 key->private.address = address;
625 * The associated futex object in this case is the inode and
626 * the page->mapping must be traversed. Ordinarily this should
627 * be stabilised under page lock but it's not strictly
628 * necessary in this case as we just want to pin the inode, not
629 * update the radix tree or anything like that.
631 * The RCU read lock is taken as the inode is finally freed
632 * under RCU. If the mapping still matches expectations then the
633 * mapping->host can be safely accessed as being a valid inode.
637 if (READ_ONCE(page->mapping) != mapping) {
644 inode = READ_ONCE(mapping->host);
652 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
653 key->shared.i_seq = get_inode_sequence_number(inode);
654 key->shared.pgoff = page_to_pgoff(tail);
664 * fault_in_user_writeable() - Fault in user address and verify RW access
665 * @uaddr: pointer to faulting user space address
667 * Slow path to fixup the fault we just took in the atomic write
670 * We have no generic implementation of a non-destructive write to the
671 * user address. We know that we faulted in the atomic pagefault
672 * disabled section so we can as well avoid the #PF overhead by
673 * calling get_user_pages() right away.
675 static int fault_in_user_writeable(u32 __user *uaddr)
677 struct mm_struct *mm = current->mm;
681 ret = fixup_user_fault(mm, (unsigned long)uaddr,
682 FAULT_FLAG_WRITE, NULL);
683 mmap_read_unlock(mm);
685 return ret < 0 ? ret : 0;
689 * futex_top_waiter() - Return the highest priority waiter on a futex
690 * @hb: the hash bucket the futex_q's reside in
691 * @key: the futex key (to distinguish it from other futex futex_q's)
693 * Must be called with the hb lock held.
695 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
696 union futex_key *key)
698 struct futex_q *this;
700 plist_for_each_entry(this, &hb->chain, list) {
701 if (match_futex(&this->key, key))
707 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
708 u32 uval, u32 newval)
713 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
719 static int get_futex_value_locked(u32 *dest, u32 __user *from)
724 ret = __get_user(*dest, from);
727 return ret ? -EFAULT : 0;
734 static int refill_pi_state_cache(void)
736 struct futex_pi_state *pi_state;
738 if (likely(current->pi_state_cache))
741 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
746 INIT_LIST_HEAD(&pi_state->list);
747 /* pi_mutex gets initialized later */
748 pi_state->owner = NULL;
749 refcount_set(&pi_state->refcount, 1);
750 pi_state->key = FUTEX_KEY_INIT;
752 current->pi_state_cache = pi_state;
757 static struct futex_pi_state *alloc_pi_state(void)
759 struct futex_pi_state *pi_state = current->pi_state_cache;
762 current->pi_state_cache = NULL;
767 static void pi_state_update_owner(struct futex_pi_state *pi_state,
768 struct task_struct *new_owner)
770 struct task_struct *old_owner = pi_state->owner;
772 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
775 raw_spin_lock(&old_owner->pi_lock);
776 WARN_ON(list_empty(&pi_state->list));
777 list_del_init(&pi_state->list);
778 raw_spin_unlock(&old_owner->pi_lock);
782 raw_spin_lock(&new_owner->pi_lock);
783 WARN_ON(!list_empty(&pi_state->list));
784 list_add(&pi_state->list, &new_owner->pi_state_list);
785 pi_state->owner = new_owner;
786 raw_spin_unlock(&new_owner->pi_lock);
790 static void get_pi_state(struct futex_pi_state *pi_state)
792 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
796 * Drops a reference to the pi_state object and frees or caches it
797 * when the last reference is gone.
799 static void put_pi_state(struct futex_pi_state *pi_state)
804 if (!refcount_dec_and_test(&pi_state->refcount))
808 * If pi_state->owner is NULL, the owner is most probably dying
809 * and has cleaned up the pi_state already
811 if (pi_state->owner) {
814 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
815 pi_state_update_owner(pi_state, NULL);
816 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
817 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
820 if (current->pi_state_cache) {
824 * pi_state->list is already empty.
825 * clear pi_state->owner.
826 * refcount is at 0 - put it back to 1.
828 pi_state->owner = NULL;
829 refcount_set(&pi_state->refcount, 1);
830 current->pi_state_cache = pi_state;
834 #ifdef CONFIG_FUTEX_PI
837 * This task is holding PI mutexes at exit time => bad.
838 * Kernel cleans up PI-state, but userspace is likely hosed.
839 * (Robust-futex cleanup is separate and might save the day for userspace.)
841 static void exit_pi_state_list(struct task_struct *curr)
843 struct list_head *next, *head = &curr->pi_state_list;
844 struct futex_pi_state *pi_state;
845 struct futex_hash_bucket *hb;
846 union futex_key key = FUTEX_KEY_INIT;
848 if (!futex_cmpxchg_enabled)
851 * We are a ZOMBIE and nobody can enqueue itself on
852 * pi_state_list anymore, but we have to be careful
853 * versus waiters unqueueing themselves:
855 raw_spin_lock_irq(&curr->pi_lock);
856 while (!list_empty(head)) {
858 pi_state = list_entry(next, struct futex_pi_state, list);
860 hb = hash_futex(&key);
863 * We can race against put_pi_state() removing itself from the
864 * list (a waiter going away). put_pi_state() will first
865 * decrement the reference count and then modify the list, so
866 * its possible to see the list entry but fail this reference
869 * In that case; drop the locks to let put_pi_state() make
870 * progress and retry the loop.
872 if (!refcount_inc_not_zero(&pi_state->refcount)) {
873 raw_spin_unlock_irq(&curr->pi_lock);
875 raw_spin_lock_irq(&curr->pi_lock);
878 raw_spin_unlock_irq(&curr->pi_lock);
880 spin_lock(&hb->lock);
881 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
882 raw_spin_lock(&curr->pi_lock);
884 * We dropped the pi-lock, so re-check whether this
885 * task still owns the PI-state:
887 if (head->next != next) {
888 /* retain curr->pi_lock for the loop invariant */
889 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
890 spin_unlock(&hb->lock);
891 put_pi_state(pi_state);
895 WARN_ON(pi_state->owner != curr);
896 WARN_ON(list_empty(&pi_state->list));
897 list_del_init(&pi_state->list);
898 pi_state->owner = NULL;
900 raw_spin_unlock(&curr->pi_lock);
901 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
902 spin_unlock(&hb->lock);
904 rt_mutex_futex_unlock(&pi_state->pi_mutex);
905 put_pi_state(pi_state);
907 raw_spin_lock_irq(&curr->pi_lock);
909 raw_spin_unlock_irq(&curr->pi_lock);
912 static inline void exit_pi_state_list(struct task_struct *curr) { }
916 * We need to check the following states:
918 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
920 * [1] NULL | --- | --- | 0 | 0/1 | Valid
921 * [2] NULL | --- | --- | >0 | 0/1 | Valid
923 * [3] Found | NULL | -- | Any | 0/1 | Invalid
925 * [4] Found | Found | NULL | 0 | 1 | Valid
926 * [5] Found | Found | NULL | >0 | 1 | Invalid
928 * [6] Found | Found | task | 0 | 1 | Valid
930 * [7] Found | Found | NULL | Any | 0 | Invalid
932 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
933 * [9] Found | Found | task | 0 | 0 | Invalid
934 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
936 * [1] Indicates that the kernel can acquire the futex atomically. We
937 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
939 * [2] Valid, if TID does not belong to a kernel thread. If no matching
940 * thread is found then it indicates that the owner TID has died.
942 * [3] Invalid. The waiter is queued on a non PI futex
944 * [4] Valid state after exit_robust_list(), which sets the user space
945 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
947 * [5] The user space value got manipulated between exit_robust_list()
948 * and exit_pi_state_list()
950 * [6] Valid state after exit_pi_state_list() which sets the new owner in
951 * the pi_state but cannot access the user space value.
953 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
955 * [8] Owner and user space value match
957 * [9] There is no transient state which sets the user space TID to 0
958 * except exit_robust_list(), but this is indicated by the
959 * FUTEX_OWNER_DIED bit. See [4]
961 * [10] There is no transient state which leaves owner and user space
962 * TID out of sync. Except one error case where the kernel is denied
963 * write access to the user address, see fixup_pi_state_owner().
966 * Serialization and lifetime rules:
970 * hb -> futex_q, relation
971 * futex_q -> pi_state, relation
973 * (cannot be raw because hb can contain arbitrary amount
976 * pi_mutex->wait_lock:
980 * (and pi_mutex 'obviously')
984 * p->pi_state_list -> pi_state->list, relation
986 * pi_state->refcount:
994 * pi_mutex->wait_lock
1000 * Validate that the existing waiter has a pi_state and sanity check
1001 * the pi_state against the user space value. If correct, attach to
1004 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1005 struct futex_pi_state *pi_state,
1006 struct futex_pi_state **ps)
1008 pid_t pid = uval & FUTEX_TID_MASK;
1013 * Userspace might have messed up non-PI and PI futexes [3]
1015 if (unlikely(!pi_state))
1019 * We get here with hb->lock held, and having found a
1020 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1021 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1022 * which in turn means that futex_lock_pi() still has a reference on
1025 * The waiter holding a reference on @pi_state also protects against
1026 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1027 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1028 * free pi_state before we can take a reference ourselves.
1030 WARN_ON(!refcount_read(&pi_state->refcount));
1033 * Now that we have a pi_state, we can acquire wait_lock
1034 * and do the state validation.
1036 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1039 * Since {uval, pi_state} is serialized by wait_lock, and our current
1040 * uval was read without holding it, it can have changed. Verify it
1041 * still is what we expect it to be, otherwise retry the entire
1044 if (get_futex_value_locked(&uval2, uaddr))
1051 * Handle the owner died case:
1053 if (uval & FUTEX_OWNER_DIED) {
1055 * exit_pi_state_list sets owner to NULL and wakes the
1056 * topmost waiter. The task which acquires the
1057 * pi_state->rt_mutex will fixup owner.
1059 if (!pi_state->owner) {
1061 * No pi state owner, but the user space TID
1062 * is not 0. Inconsistent state. [5]
1067 * Take a ref on the state and return success. [4]
1073 * If TID is 0, then either the dying owner has not
1074 * yet executed exit_pi_state_list() or some waiter
1075 * acquired the rtmutex in the pi state, but did not
1076 * yet fixup the TID in user space.
1078 * Take a ref on the state and return success. [6]
1084 * If the owner died bit is not set, then the pi_state
1085 * must have an owner. [7]
1087 if (!pi_state->owner)
1092 * Bail out if user space manipulated the futex value. If pi
1093 * state exists then the owner TID must be the same as the
1094 * user space TID. [9/10]
1096 if (pid != task_pid_vnr(pi_state->owner))
1100 get_pi_state(pi_state);
1101 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1118 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1123 * wait_for_owner_exiting - Block until the owner has exited
1124 * @ret: owner's current futex lock status
1125 * @exiting: Pointer to the exiting task
1127 * Caller must hold a refcount on @exiting.
1129 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1131 if (ret != -EBUSY) {
1132 WARN_ON_ONCE(exiting);
1136 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1139 mutex_lock(&exiting->futex_exit_mutex);
1141 * No point in doing state checking here. If the waiter got here
1142 * while the task was in exec()->exec_futex_release() then it can
1143 * have any FUTEX_STATE_* value when the waiter has acquired the
1144 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1145 * already. Highly unlikely and not a problem. Just one more round
1146 * through the futex maze.
1148 mutex_unlock(&exiting->futex_exit_mutex);
1150 put_task_struct(exiting);
1153 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1154 struct task_struct *tsk)
1159 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1160 * caller that the alleged owner is busy.
1162 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1166 * Reread the user space value to handle the following situation:
1170 * sys_exit() sys_futex()
1171 * do_exit() futex_lock_pi()
1172 * futex_lock_pi_atomic()
1173 * exit_signals(tsk) No waiters:
1174 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1175 * mm_release(tsk) Set waiter bit
1176 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1177 * Set owner died attach_to_pi_owner() {
1178 * *uaddr = 0xC0000000; tsk = get_task(PID);
1179 * } if (!tsk->flags & PF_EXITING) {
1181 * tsk->futex_state = } else {
1182 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1185 * return -ESRCH; <--- FAIL
1188 * Returning ESRCH unconditionally is wrong here because the
1189 * user space value has been changed by the exiting task.
1191 * The same logic applies to the case where the exiting task is
1194 if (get_futex_value_locked(&uval2, uaddr))
1197 /* If the user space value has changed, try again. */
1202 * The exiting task did not have a robust list, the robust list was
1203 * corrupted or the user space value in *uaddr is simply bogus.
1204 * Give up and tell user space.
1210 * Lookup the task for the TID provided from user space and attach to
1211 * it after doing proper sanity checks.
1213 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1214 struct futex_pi_state **ps,
1215 struct task_struct **exiting)
1217 pid_t pid = uval & FUTEX_TID_MASK;
1218 struct futex_pi_state *pi_state;
1219 struct task_struct *p;
1222 * We are the first waiter - try to look up the real owner and attach
1223 * the new pi_state to it, but bail out when TID = 0 [1]
1225 * The !pid check is paranoid. None of the call sites should end up
1226 * with pid == 0, but better safe than sorry. Let the caller retry
1230 p = find_get_task_by_vpid(pid);
1232 return handle_exit_race(uaddr, uval, NULL);
1234 if (unlikely(p->flags & PF_KTHREAD)) {
1240 * We need to look at the task state to figure out, whether the
1241 * task is exiting. To protect against the change of the task state
1242 * in futex_exit_release(), we do this protected by p->pi_lock:
1244 raw_spin_lock_irq(&p->pi_lock);
1245 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1247 * The task is on the way out. When the futex state is
1248 * FUTEX_STATE_DEAD, we know that the task has finished
1251 int ret = handle_exit_race(uaddr, uval, p);
1253 raw_spin_unlock_irq(&p->pi_lock);
1255 * If the owner task is between FUTEX_STATE_EXITING and
1256 * FUTEX_STATE_DEAD then store the task pointer and keep
1257 * the reference on the task struct. The calling code will
1258 * drop all locks, wait for the task to reach
1259 * FUTEX_STATE_DEAD and then drop the refcount. This is
1260 * required to prevent a live lock when the current task
1261 * preempted the exiting task between the two states.
1271 * No existing pi state. First waiter. [2]
1273 * This creates pi_state, we have hb->lock held, this means nothing can
1274 * observe this state, wait_lock is irrelevant.
1276 pi_state = alloc_pi_state();
1279 * Initialize the pi_mutex in locked state and make @p
1282 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1284 /* Store the key for possible exit cleanups: */
1285 pi_state->key = *key;
1287 WARN_ON(!list_empty(&pi_state->list));
1288 list_add(&pi_state->list, &p->pi_state_list);
1290 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1291 * because there is no concurrency as the object is not published yet.
1293 pi_state->owner = p;
1294 raw_spin_unlock_irq(&p->pi_lock);
1303 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1304 struct futex_hash_bucket *hb,
1305 union futex_key *key, struct futex_pi_state **ps,
1306 struct task_struct **exiting)
1308 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1311 * If there is a waiter on that futex, validate it and
1312 * attach to the pi_state when the validation succeeds.
1315 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1318 * We are the first waiter - try to look up the owner based on
1319 * @uval and attach to it.
1321 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1324 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1329 if (unlikely(should_fail_futex(true)))
1332 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1336 /* If user space value changed, let the caller retry */
1337 return curval != uval ? -EAGAIN : 0;
1341 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1342 * @uaddr: the pi futex user address
1343 * @hb: the pi futex hash bucket
1344 * @key: the futex key associated with uaddr and hb
1345 * @ps: the pi_state pointer where we store the result of the
1347 * @task: the task to perform the atomic lock work for. This will
1348 * be "current" except in the case of requeue pi.
1349 * @exiting: Pointer to store the task pointer of the owner task
1350 * which is in the middle of exiting
1351 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1354 * - 0 - ready to wait;
1355 * - 1 - acquired the lock;
1358 * The hb->lock and futex_key refs shall be held by the caller.
1360 * @exiting is only set when the return value is -EBUSY. If so, this holds
1361 * a refcount on the exiting task on return and the caller needs to drop it
1362 * after waiting for the exit to complete.
1364 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1365 union futex_key *key,
1366 struct futex_pi_state **ps,
1367 struct task_struct *task,
1368 struct task_struct **exiting,
1371 u32 uval, newval, vpid = task_pid_vnr(task);
1372 struct futex_q *top_waiter;
1376 * Read the user space value first so we can validate a few
1377 * things before proceeding further.
1379 if (get_futex_value_locked(&uval, uaddr))
1382 if (unlikely(should_fail_futex(true)))
1388 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1391 if ((unlikely(should_fail_futex(true))))
1395 * Lookup existing state first. If it exists, try to attach to
1398 top_waiter = futex_top_waiter(hb, key);
1400 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1403 * No waiter and user TID is 0. We are here because the
1404 * waiters or the owner died bit is set or called from
1405 * requeue_cmp_pi or for whatever reason something took the
1408 if (!(uval & FUTEX_TID_MASK)) {
1410 * We take over the futex. No other waiters and the user space
1411 * TID is 0. We preserve the owner died bit.
1413 newval = uval & FUTEX_OWNER_DIED;
1416 /* The futex requeue_pi code can enforce the waiters bit */
1418 newval |= FUTEX_WAITERS;
1420 ret = lock_pi_update_atomic(uaddr, uval, newval);
1421 /* If the take over worked, return 1 */
1422 return ret < 0 ? ret : 1;
1426 * First waiter. Set the waiters bit before attaching ourself to
1427 * the owner. If owner tries to unlock, it will be forced into
1428 * the kernel and blocked on hb->lock.
1430 newval = uval | FUTEX_WAITERS;
1431 ret = lock_pi_update_atomic(uaddr, uval, newval);
1435 * If the update of the user space value succeeded, we try to
1436 * attach to the owner. If that fails, no harm done, we only
1437 * set the FUTEX_WAITERS bit in the user space variable.
1439 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1443 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1444 * @q: The futex_q to unqueue
1446 * The q->lock_ptr must not be NULL and must be held by the caller.
1448 static void __unqueue_futex(struct futex_q *q)
1450 struct futex_hash_bucket *hb;
1452 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1454 lockdep_assert_held(q->lock_ptr);
1456 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1457 plist_del(&q->list, &hb->chain);
1462 * The hash bucket lock must be held when this is called.
1463 * Afterwards, the futex_q must not be accessed. Callers
1464 * must ensure to later call wake_up_q() for the actual
1467 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1469 struct task_struct *p = q->task;
1471 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1477 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1478 * is written, without taking any locks. This is possible in the event
1479 * of a spurious wakeup, for example. A memory barrier is required here
1480 * to prevent the following store to lock_ptr from getting ahead of the
1481 * plist_del in __unqueue_futex().
1483 smp_store_release(&q->lock_ptr, NULL);
1486 * Queue the task for later wakeup for after we've released
1489 wake_q_add_safe(wake_q, p);
1493 * Caller must hold a reference on @pi_state.
1495 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1498 struct task_struct *new_owner;
1499 bool postunlock = false;
1500 DEFINE_WAKE_Q(wake_q);
1501 DEFINE_WAKE_Q(wake_sleeper_q);
1504 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1505 if (WARN_ON_ONCE(!new_owner)) {
1507 * As per the comment in futex_unlock_pi() this should not happen.
1509 * When this happens, give up our locks and try again, giving
1510 * the futex_lock_pi() instance time to complete, either by
1511 * waiting on the rtmutex or removing itself from the futex
1519 * We pass it to the next owner. The WAITERS bit is always kept
1520 * enabled while there is PI state around. We cleanup the owner
1521 * died bit, because we are the owner.
1523 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1525 if (unlikely(should_fail_futex(true))) {
1530 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1531 if (!ret && (curval != uval)) {
1533 * If a unconditional UNLOCK_PI operation (user space did not
1534 * try the TID->0 transition) raced with a waiter setting the
1535 * FUTEX_WAITERS flag between get_user() and locking the hash
1536 * bucket lock, retry the operation.
1538 if ((FUTEX_TID_MASK & curval) == uval)
1546 * This is a point of no return; once we modified the uval
1547 * there is no going back and subsequent operations must
1550 pi_state_update_owner(pi_state, new_owner);
1551 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q,
1556 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1559 rt_mutex_postunlock(&wake_q, &wake_sleeper_q);
1565 * Express the locking dependencies for lockdep:
1568 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1571 spin_lock(&hb1->lock);
1573 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1574 } else { /* hb1 > hb2 */
1575 spin_lock(&hb2->lock);
1576 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1581 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1583 spin_unlock(&hb1->lock);
1585 spin_unlock(&hb2->lock);
1589 * Wake up waiters matching bitset queued on this futex (uaddr).
1592 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1594 struct futex_hash_bucket *hb;
1595 struct futex_q *this, *next;
1596 union futex_key key = FUTEX_KEY_INIT;
1598 DEFINE_WAKE_Q(wake_q);
1603 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1604 if (unlikely(ret != 0))
1607 hb = hash_futex(&key);
1609 /* Make sure we really have tasks to wakeup */
1610 if (!hb_waiters_pending(hb))
1613 spin_lock(&hb->lock);
1615 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1616 if (match_futex (&this->key, &key)) {
1617 if (this->pi_state || this->rt_waiter) {
1622 /* Check if one of the bits is set in both bitsets */
1623 if (!(this->bitset & bitset))
1626 mark_wake_futex(&wake_q, this);
1627 if (++ret >= nr_wake)
1632 spin_unlock(&hb->lock);
1637 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1639 unsigned int op = (encoded_op & 0x70000000) >> 28;
1640 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1641 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1642 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1645 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1646 if (oparg < 0 || oparg > 31) {
1647 char comm[sizeof(current->comm)];
1649 * kill this print and return -EINVAL when userspace
1652 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1653 get_task_comm(comm, current), oparg);
1659 pagefault_disable();
1660 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1666 case FUTEX_OP_CMP_EQ:
1667 return oldval == cmparg;
1668 case FUTEX_OP_CMP_NE:
1669 return oldval != cmparg;
1670 case FUTEX_OP_CMP_LT:
1671 return oldval < cmparg;
1672 case FUTEX_OP_CMP_GE:
1673 return oldval >= cmparg;
1674 case FUTEX_OP_CMP_LE:
1675 return oldval <= cmparg;
1676 case FUTEX_OP_CMP_GT:
1677 return oldval > cmparg;
1684 * Wake up all waiters hashed on the physical page that is mapped
1685 * to this virtual address:
1688 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1689 int nr_wake, int nr_wake2, int op)
1691 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1692 struct futex_hash_bucket *hb1, *hb2;
1693 struct futex_q *this, *next;
1695 DEFINE_WAKE_Q(wake_q);
1698 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1699 if (unlikely(ret != 0))
1701 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1702 if (unlikely(ret != 0))
1705 hb1 = hash_futex(&key1);
1706 hb2 = hash_futex(&key2);
1709 double_lock_hb(hb1, hb2);
1710 op_ret = futex_atomic_op_inuser(op, uaddr2);
1711 if (unlikely(op_ret < 0)) {
1712 double_unlock_hb(hb1, hb2);
1714 if (!IS_ENABLED(CONFIG_MMU) ||
1715 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1717 * we don't get EFAULT from MMU faults if we don't have
1718 * an MMU, but we might get them from range checking
1724 if (op_ret == -EFAULT) {
1725 ret = fault_in_user_writeable(uaddr2);
1730 if (!(flags & FLAGS_SHARED)) {
1739 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1740 if (match_futex (&this->key, &key1)) {
1741 if (this->pi_state || this->rt_waiter) {
1745 mark_wake_futex(&wake_q, this);
1746 if (++ret >= nr_wake)
1753 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1754 if (match_futex (&this->key, &key2)) {
1755 if (this->pi_state || this->rt_waiter) {
1759 mark_wake_futex(&wake_q, this);
1760 if (++op_ret >= nr_wake2)
1768 double_unlock_hb(hb1, hb2);
1774 * requeue_futex() - Requeue a futex_q from one hb to another
1775 * @q: the futex_q to requeue
1776 * @hb1: the source hash_bucket
1777 * @hb2: the target hash_bucket
1778 * @key2: the new key for the requeued futex_q
1781 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1782 struct futex_hash_bucket *hb2, union futex_key *key2)
1786 * If key1 and key2 hash to the same bucket, no need to
1789 if (likely(&hb1->chain != &hb2->chain)) {
1790 plist_del(&q->list, &hb1->chain);
1791 hb_waiters_dec(hb1);
1792 hb_waiters_inc(hb2);
1793 plist_add(&q->list, &hb2->chain);
1794 q->lock_ptr = &hb2->lock;
1800 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1802 * @key: the key of the requeue target futex
1803 * @hb: the hash_bucket of the requeue target futex
1805 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1806 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1807 * to the requeue target futex so the waiter can detect the wakeup on the right
1808 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1809 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1810 * to protect access to the pi_state to fixup the owner later. Must be called
1811 * with both q->lock_ptr and hb->lock held.
1814 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1815 struct futex_hash_bucket *hb)
1821 WARN_ON(!q->rt_waiter);
1822 q->rt_waiter = NULL;
1824 q->lock_ptr = &hb->lock;
1826 wake_up_state(q->task, TASK_NORMAL);
1830 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1831 * @pifutex: the user address of the to futex
1832 * @hb1: the from futex hash bucket, must be locked by the caller
1833 * @hb2: the to futex hash bucket, must be locked by the caller
1834 * @key1: the from futex key
1835 * @key2: the to futex key
1836 * @ps: address to store the pi_state pointer
1837 * @exiting: Pointer to store the task pointer of the owner task
1838 * which is in the middle of exiting
1839 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1841 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1842 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1843 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1844 * hb1 and hb2 must be held by the caller.
1846 * @exiting is only set when the return value is -EBUSY. If so, this holds
1847 * a refcount on the exiting task on return and the caller needs to drop it
1848 * after waiting for the exit to complete.
1851 * - 0 - failed to acquire the lock atomically;
1852 * - >0 - acquired the lock, return value is vpid of the top_waiter
1856 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1857 struct futex_hash_bucket *hb2, union futex_key *key1,
1858 union futex_key *key2, struct futex_pi_state **ps,
1859 struct task_struct **exiting, int set_waiters)
1861 struct futex_q *top_waiter = NULL;
1865 if (get_futex_value_locked(&curval, pifutex))
1868 if (unlikely(should_fail_futex(true)))
1872 * Find the top_waiter and determine if there are additional waiters.
1873 * If the caller intends to requeue more than 1 waiter to pifutex,
1874 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1875 * as we have means to handle the possible fault. If not, don't set
1876 * the bit unecessarily as it will force the subsequent unlock to enter
1879 top_waiter = futex_top_waiter(hb1, key1);
1881 /* There are no waiters, nothing for us to do. */
1885 /* Ensure we requeue to the expected futex. */
1886 if (!match_futex(top_waiter->requeue_pi_key, key2))
1890 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1891 * the contended case or if set_waiters is 1. The pi_state is returned
1892 * in ps in contended cases.
1894 vpid = task_pid_vnr(top_waiter->task);
1895 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1896 exiting, set_waiters);
1898 requeue_pi_wake_futex(top_waiter, key2, hb2);
1905 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1906 * @uaddr1: source futex user address
1907 * @flags: futex flags (FLAGS_SHARED, etc.)
1908 * @uaddr2: target futex user address
1909 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1910 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1911 * @cmpval: @uaddr1 expected value (or %NULL)
1912 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1913 * pi futex (pi to pi requeue is not supported)
1915 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1916 * uaddr2 atomically on behalf of the top waiter.
1919 * - >=0 - on success, the number of tasks requeued or woken;
1922 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1923 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1924 u32 *cmpval, int requeue_pi)
1926 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1927 int task_count = 0, ret;
1928 struct futex_pi_state *pi_state = NULL;
1929 struct futex_hash_bucket *hb1, *hb2;
1930 struct futex_q *this, *next;
1931 DEFINE_WAKE_Q(wake_q);
1933 if (nr_wake < 0 || nr_requeue < 0)
1937 * When PI not supported: return -ENOSYS if requeue_pi is true,
1938 * consequently the compiler knows requeue_pi is always false past
1939 * this point which will optimize away all the conditional code
1942 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1947 * Requeue PI only works on two distinct uaddrs. This
1948 * check is only valid for private futexes. See below.
1950 if (uaddr1 == uaddr2)
1954 * requeue_pi requires a pi_state, try to allocate it now
1955 * without any locks in case it fails.
1957 if (refill_pi_state_cache())
1960 * requeue_pi must wake as many tasks as it can, up to nr_wake
1961 * + nr_requeue, since it acquires the rt_mutex prior to
1962 * returning to userspace, so as to not leave the rt_mutex with
1963 * waiters and no owner. However, second and third wake-ups
1964 * cannot be predicted as they involve race conditions with the
1965 * first wake and a fault while looking up the pi_state. Both
1966 * pthread_cond_signal() and pthread_cond_broadcast() should
1974 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1975 if (unlikely(ret != 0))
1977 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1978 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1979 if (unlikely(ret != 0))
1983 * The check above which compares uaddrs is not sufficient for
1984 * shared futexes. We need to compare the keys:
1986 if (requeue_pi && match_futex(&key1, &key2))
1989 hb1 = hash_futex(&key1);
1990 hb2 = hash_futex(&key2);
1993 hb_waiters_inc(hb2);
1994 double_lock_hb(hb1, hb2);
1996 if (likely(cmpval != NULL)) {
1999 ret = get_futex_value_locked(&curval, uaddr1);
2001 if (unlikely(ret)) {
2002 double_unlock_hb(hb1, hb2);
2003 hb_waiters_dec(hb2);
2005 ret = get_user(curval, uaddr1);
2009 if (!(flags & FLAGS_SHARED))
2014 if (curval != *cmpval) {
2020 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2021 struct task_struct *exiting = NULL;
2024 * Attempt to acquire uaddr2 and wake the top waiter. If we
2025 * intend to requeue waiters, force setting the FUTEX_WAITERS
2026 * bit. We force this here where we are able to easily handle
2027 * faults rather in the requeue loop below.
2029 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2031 &exiting, nr_requeue);
2034 * At this point the top_waiter has either taken uaddr2 or is
2035 * waiting on it. If the former, then the pi_state will not
2036 * exist yet, look it up one more time to ensure we have a
2037 * reference to it. If the lock was taken, ret contains the
2038 * vpid of the top waiter task.
2039 * If the lock was not taken, we have pi_state and an initial
2040 * refcount on it. In case of an error we have nothing.
2046 * If we acquired the lock, then the user space value
2047 * of uaddr2 should be vpid. It cannot be changed by
2048 * the top waiter as it is blocked on hb2 lock if it
2049 * tries to do so. If something fiddled with it behind
2050 * our back the pi state lookup might unearth it. So
2051 * we rather use the known value than rereading and
2052 * handing potential crap to lookup_pi_state.
2054 * If that call succeeds then we have pi_state and an
2055 * initial refcount on it.
2057 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2058 &pi_state, &exiting);
2063 /* We hold a reference on the pi state. */
2066 /* If the above failed, then pi_state is NULL */
2068 double_unlock_hb(hb1, hb2);
2069 hb_waiters_dec(hb2);
2070 ret = fault_in_user_writeable(uaddr2);
2077 * Two reasons for this:
2078 * - EBUSY: Owner is exiting and we just wait for the
2080 * - EAGAIN: The user space value changed.
2082 double_unlock_hb(hb1, hb2);
2083 hb_waiters_dec(hb2);
2085 * Handle the case where the owner is in the middle of
2086 * exiting. Wait for the exit to complete otherwise
2087 * this task might loop forever, aka. live lock.
2089 wait_for_owner_exiting(ret, exiting);
2097 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2098 if (task_count - nr_wake >= nr_requeue)
2101 if (!match_futex(&this->key, &key1))
2105 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2106 * be paired with each other and no other futex ops.
2108 * We should never be requeueing a futex_q with a pi_state,
2109 * which is awaiting a futex_unlock_pi().
2111 if ((requeue_pi && !this->rt_waiter) ||
2112 (!requeue_pi && this->rt_waiter) ||
2119 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2120 * lock, we already woke the top_waiter. If not, it will be
2121 * woken by futex_unlock_pi().
2123 if (++task_count <= nr_wake && !requeue_pi) {
2124 mark_wake_futex(&wake_q, this);
2128 /* Ensure we requeue to the expected futex for requeue_pi. */
2129 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2135 * Requeue nr_requeue waiters and possibly one more in the case
2136 * of requeue_pi if we couldn't acquire the lock atomically.
2140 * Prepare the waiter to take the rt_mutex. Take a
2141 * refcount on the pi_state and store the pointer in
2142 * the futex_q object of the waiter.
2144 get_pi_state(pi_state);
2145 this->pi_state = pi_state;
2146 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2151 * We got the lock. We do neither drop the
2152 * refcount on pi_state nor clear
2153 * this->pi_state because the waiter needs the
2154 * pi_state for cleaning up the user space
2155 * value. It will drop the refcount after
2158 requeue_pi_wake_futex(this, &key2, hb2);
2160 } else if (ret == -EAGAIN) {
2162 * Waiter was woken by timeout or
2163 * signal and has set pi_blocked_on to
2164 * PI_WAKEUP_INPROGRESS before we
2165 * tried to enqueue it on the rtmutex.
2167 this->pi_state = NULL;
2168 put_pi_state(pi_state);
2172 * rt_mutex_start_proxy_lock() detected a
2173 * potential deadlock when we tried to queue
2174 * that waiter. Drop the pi_state reference
2175 * which we took above and remove the pointer
2176 * to the state from the waiters futex_q
2179 this->pi_state = NULL;
2180 put_pi_state(pi_state);
2182 * We stop queueing more waiters and let user
2183 * space deal with the mess.
2188 requeue_futex(this, hb1, hb2, &key2);
2192 * We took an extra initial reference to the pi_state either
2193 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2194 * need to drop it here again.
2196 put_pi_state(pi_state);
2199 double_unlock_hb(hb1, hb2);
2201 hb_waiters_dec(hb2);
2202 return ret ? ret : task_count;
2205 /* The key must be already stored in q->key. */
2206 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2207 __acquires(&hb->lock)
2209 struct futex_hash_bucket *hb;
2211 hb = hash_futex(&q->key);
2214 * Increment the counter before taking the lock so that
2215 * a potential waker won't miss a to-be-slept task that is
2216 * waiting for the spinlock. This is safe as all queue_lock()
2217 * users end up calling queue_me(). Similarly, for housekeeping,
2218 * decrement the counter at queue_unlock() when some error has
2219 * occurred and we don't end up adding the task to the list.
2221 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2223 q->lock_ptr = &hb->lock;
2225 spin_lock(&hb->lock);
2230 queue_unlock(struct futex_hash_bucket *hb)
2231 __releases(&hb->lock)
2233 spin_unlock(&hb->lock);
2237 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2242 * The priority used to register this element is
2243 * - either the real thread-priority for the real-time threads
2244 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2245 * - or MAX_RT_PRIO for non-RT threads.
2246 * Thus, all RT-threads are woken first in priority order, and
2247 * the others are woken last, in FIFO order.
2249 prio = min(current->normal_prio, MAX_RT_PRIO);
2251 plist_node_init(&q->list, prio);
2252 plist_add(&q->list, &hb->chain);
2257 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2258 * @q: The futex_q to enqueue
2259 * @hb: The destination hash bucket
2261 * The hb->lock must be held by the caller, and is released here. A call to
2262 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2263 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2264 * or nothing if the unqueue is done as part of the wake process and the unqueue
2265 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2268 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2269 __releases(&hb->lock)
2272 spin_unlock(&hb->lock);
2276 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2277 * @q: The futex_q to unqueue
2279 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2280 * be paired with exactly one earlier call to queue_me().
2283 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2284 * - 0 - if the futex_q was already removed by the waking thread
2286 static int unqueue_me(struct futex_q *q)
2288 spinlock_t *lock_ptr;
2291 /* In the common case we don't take the spinlock, which is nice. */
2294 * q->lock_ptr can change between this read and the following spin_lock.
2295 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2296 * optimizing lock_ptr out of the logic below.
2298 lock_ptr = READ_ONCE(q->lock_ptr);
2299 if (lock_ptr != NULL) {
2300 spin_lock(lock_ptr);
2302 * q->lock_ptr can change between reading it and
2303 * spin_lock(), causing us to take the wrong lock. This
2304 * corrects the race condition.
2306 * Reasoning goes like this: if we have the wrong lock,
2307 * q->lock_ptr must have changed (maybe several times)
2308 * between reading it and the spin_lock(). It can
2309 * change again after the spin_lock() but only if it was
2310 * already changed before the spin_lock(). It cannot,
2311 * however, change back to the original value. Therefore
2312 * we can detect whether we acquired the correct lock.
2314 if (unlikely(lock_ptr != q->lock_ptr)) {
2315 spin_unlock(lock_ptr);
2320 BUG_ON(q->pi_state);
2322 spin_unlock(lock_ptr);
2330 * PI futexes can not be requeued and must remove themself from the
2331 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2334 static void unqueue_me_pi(struct futex_q *q)
2335 __releases(q->lock_ptr)
2339 BUG_ON(!q->pi_state);
2340 put_pi_state(q->pi_state);
2343 spin_unlock(q->lock_ptr);
2346 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2347 struct task_struct *argowner)
2349 struct futex_pi_state *pi_state = q->pi_state;
2350 struct task_struct *oldowner, *newowner;
2351 u32 uval, curval, newval, newtid;
2354 oldowner = pi_state->owner;
2357 * We are here because either:
2359 * - we stole the lock and pi_state->owner needs updating to reflect
2360 * that (@argowner == current),
2364 * - someone stole our lock and we need to fix things to point to the
2365 * new owner (@argowner == NULL).
2367 * Either way, we have to replace the TID in the user space variable.
2368 * This must be atomic as we have to preserve the owner died bit here.
2370 * Note: We write the user space value _before_ changing the pi_state
2371 * because we can fault here. Imagine swapped out pages or a fork
2372 * that marked all the anonymous memory readonly for cow.
2374 * Modifying pi_state _before_ the user space value would leave the
2375 * pi_state in an inconsistent state when we fault here, because we
2376 * need to drop the locks to handle the fault. This might be observed
2377 * in the PID check in lookup_pi_state.
2381 if (oldowner != current) {
2383 * We raced against a concurrent self; things are
2384 * already fixed up. Nothing to do.
2389 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2390 /* We got the lock. pi_state is correct. Tell caller. */
2395 * The trylock just failed, so either there is an owner or
2396 * there is a higher priority waiter than this one.
2398 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2400 * If the higher priority waiter has not yet taken over the
2401 * rtmutex then newowner is NULL. We can't return here with
2402 * that state because it's inconsistent vs. the user space
2403 * state. So drop the locks and try again. It's a valid
2404 * situation and not any different from the other retry
2407 if (unlikely(!newowner)) {
2412 WARN_ON_ONCE(argowner != current);
2413 if (oldowner == current) {
2415 * We raced against a concurrent self; things are
2416 * already fixed up. Nothing to do.
2420 newowner = argowner;
2423 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2425 if (!pi_state->owner)
2426 newtid |= FUTEX_OWNER_DIED;
2428 err = get_futex_value_locked(&uval, uaddr);
2433 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2435 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2445 * We fixed up user space. Now we need to fix the pi_state
2448 pi_state_update_owner(pi_state, newowner);
2450 return argowner == current;
2453 * In order to reschedule or handle a page fault, we need to drop the
2454 * locks here. In the case of a fault, this gives the other task
2455 * (either the highest priority waiter itself or the task which stole
2456 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2457 * are back from handling the fault we need to check the pi_state after
2458 * reacquiring the locks and before trying to do another fixup. When
2459 * the fixup has been done already we simply return.
2461 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2462 * drop hb->lock since the caller owns the hb -> futex_q relation.
2463 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2466 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2467 spin_unlock(q->lock_ptr);
2471 err = fault_in_user_writeable(uaddr);
2484 spin_lock(q->lock_ptr);
2485 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2488 * Check if someone else fixed it for us:
2490 if (pi_state->owner != oldowner)
2491 return argowner == current;
2493 /* Retry if err was -EAGAIN or the fault in succeeded */
2498 * fault_in_user_writeable() failed so user state is immutable. At
2499 * best we can make the kernel state consistent but user state will
2500 * be most likely hosed and any subsequent unlock operation will be
2501 * rejected due to PI futex rule [10].
2503 * Ensure that the rtmutex owner is also the pi_state owner despite
2504 * the user space value claiming something different. There is no
2505 * point in unlocking the rtmutex if current is the owner as it
2506 * would need to wait until the next waiter has taken the rtmutex
2507 * to guarantee consistent state. Keep it simple. Userspace asked
2508 * for this wreckaged state.
2510 * The rtmutex has an owner - either current or some other
2511 * task. See the EAGAIN loop above.
2513 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2518 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2519 struct task_struct *argowner)
2521 struct futex_pi_state *pi_state = q->pi_state;
2524 lockdep_assert_held(q->lock_ptr);
2526 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2527 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2528 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2532 static long futex_wait_restart(struct restart_block *restart);
2535 * fixup_owner() - Post lock pi_state and corner case management
2536 * @uaddr: user address of the futex
2537 * @q: futex_q (contains pi_state and access to the rt_mutex)
2538 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2540 * After attempting to lock an rt_mutex, this function is called to cleanup
2541 * the pi_state owner as well as handle race conditions that may allow us to
2542 * acquire the lock. Must be called with the hb lock held.
2545 * - 1 - success, lock taken;
2546 * - 0 - success, lock not taken;
2547 * - <0 - on error (-EFAULT)
2549 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2553 * Got the lock. We might not be the anticipated owner if we
2554 * did a lock-steal - fix up the PI-state in that case:
2556 * Speculative pi_state->owner read (we don't hold wait_lock);
2557 * since we own the lock pi_state->owner == current is the
2558 * stable state, anything else needs more attention.
2560 if (q->pi_state->owner != current)
2561 return fixup_pi_state_owner(uaddr, q, current);
2566 * If we didn't get the lock; check if anybody stole it from us. In
2567 * that case, we need to fix up the uval to point to them instead of
2568 * us, otherwise bad things happen. [10]
2570 * Another speculative read; pi_state->owner == current is unstable
2571 * but needs our attention.
2573 if (q->pi_state->owner == current)
2574 return fixup_pi_state_owner(uaddr, q, NULL);
2577 * Paranoia check. If we did not take the lock, then we should not be
2578 * the owner of the rt_mutex. Warn and establish consistent state.
2580 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2581 return fixup_pi_state_owner(uaddr, q, current);
2587 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2588 * @hb: the futex hash bucket, must be locked by the caller
2589 * @q: the futex_q to queue up on
2590 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2592 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2593 struct hrtimer_sleeper *timeout)
2596 * The task state is guaranteed to be set before another task can
2597 * wake it. set_current_state() is implemented using smp_store_mb() and
2598 * queue_me() calls spin_unlock() upon completion, both serializing
2599 * access to the hash list and forcing another memory barrier.
2601 set_current_state(TASK_INTERRUPTIBLE);
2606 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2609 * If we have been removed from the hash list, then another task
2610 * has tried to wake us, and we can skip the call to schedule().
2612 if (likely(!plist_node_empty(&q->list))) {
2614 * If the timer has already expired, current will already be
2615 * flagged for rescheduling. Only call schedule if there
2616 * is no timeout, or if it has yet to expire.
2618 if (!timeout || timeout->task)
2619 freezable_schedule();
2621 __set_current_state(TASK_RUNNING);
2625 * futex_wait_setup() - Prepare to wait on a futex
2626 * @uaddr: the futex userspace address
2627 * @val: the expected value
2628 * @flags: futex flags (FLAGS_SHARED, etc.)
2629 * @q: the associated futex_q
2630 * @hb: storage for hash_bucket pointer to be returned to caller
2632 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2633 * compare it with the expected value. Handle atomic faults internally.
2634 * Return with the hb lock held and a q.key reference on success, and unlocked
2635 * with no q.key reference on failure.
2638 * - 0 - uaddr contains val and hb has been locked;
2639 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2641 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2642 struct futex_q *q, struct futex_hash_bucket **hb)
2648 * Access the page AFTER the hash-bucket is locked.
2649 * Order is important:
2651 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2652 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2654 * The basic logical guarantee of a futex is that it blocks ONLY
2655 * if cond(var) is known to be true at the time of blocking, for
2656 * any cond. If we locked the hash-bucket after testing *uaddr, that
2657 * would open a race condition where we could block indefinitely with
2658 * cond(var) false, which would violate the guarantee.
2660 * On the other hand, we insert q and release the hash-bucket only
2661 * after testing *uaddr. This guarantees that futex_wait() will NOT
2662 * absorb a wakeup if *uaddr does not match the desired values
2663 * while the syscall executes.
2666 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2667 if (unlikely(ret != 0))
2671 *hb = queue_lock(q);
2673 ret = get_futex_value_locked(&uval, uaddr);
2678 ret = get_user(uval, uaddr);
2682 if (!(flags & FLAGS_SHARED))
2696 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2697 ktime_t *abs_time, u32 bitset)
2699 struct hrtimer_sleeper timeout, *to;
2700 struct restart_block *restart;
2701 struct futex_hash_bucket *hb;
2702 struct futex_q q = futex_q_init;
2709 to = futex_setup_timer(abs_time, &timeout, flags,
2710 current->timer_slack_ns);
2713 * Prepare to wait on uaddr. On success, holds hb lock and increments
2716 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2720 /* queue_me and wait for wakeup, timeout, or a signal. */
2721 futex_wait_queue_me(hb, &q, to);
2723 /* If we were woken (and unqueued), we succeeded, whatever. */
2725 /* unqueue_me() drops q.key ref */
2726 if (!unqueue_me(&q))
2729 if (to && !to->task)
2733 * We expect signal_pending(current), but we might be the
2734 * victim of a spurious wakeup as well.
2736 if (!signal_pending(current))
2743 restart = ¤t->restart_block;
2744 restart->futex.uaddr = uaddr;
2745 restart->futex.val = val;
2746 restart->futex.time = *abs_time;
2747 restart->futex.bitset = bitset;
2748 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2750 ret = set_restart_fn(restart, futex_wait_restart);
2754 hrtimer_cancel(&to->timer);
2755 destroy_hrtimer_on_stack(&to->timer);
2761 static long futex_wait_restart(struct restart_block *restart)
2763 u32 __user *uaddr = restart->futex.uaddr;
2764 ktime_t t, *tp = NULL;
2766 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2767 t = restart->futex.time;
2770 restart->fn = do_no_restart_syscall;
2772 return (long)futex_wait(uaddr, restart->futex.flags,
2773 restart->futex.val, tp, restart->futex.bitset);
2778 * Userspace tried a 0 -> TID atomic transition of the futex value
2779 * and failed. The kernel side here does the whole locking operation:
2780 * if there are waiters then it will block as a consequence of relying
2781 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2782 * a 0 value of the futex too.).
2784 * Also serves as futex trylock_pi()'ing, and due semantics.
2786 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2787 ktime_t *time, int trylock)
2789 struct hrtimer_sleeper timeout, *to;
2790 struct task_struct *exiting = NULL;
2791 struct rt_mutex_waiter rt_waiter;
2792 struct futex_hash_bucket *hb;
2793 struct futex_q q = futex_q_init;
2796 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2799 if (refill_pi_state_cache())
2802 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2805 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2806 if (unlikely(ret != 0))
2810 hb = queue_lock(&q);
2812 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2814 if (unlikely(ret)) {
2816 * Atomic work succeeded and we got the lock,
2817 * or failed. Either way, we do _not_ block.
2821 /* We got the lock. */
2823 goto out_unlock_put_key;
2829 * Two reasons for this:
2830 * - EBUSY: Task is exiting and we just wait for the
2832 * - EAGAIN: The user space value changed.
2836 * Handle the case where the owner is in the middle of
2837 * exiting. Wait for the exit to complete otherwise
2838 * this task might loop forever, aka. live lock.
2840 wait_for_owner_exiting(ret, exiting);
2844 goto out_unlock_put_key;
2848 WARN_ON(!q.pi_state);
2851 * Only actually queue now that the atomic ops are done:
2856 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2857 /* Fixup the trylock return value: */
2858 ret = ret ? 0 : -EWOULDBLOCK;
2862 rt_mutex_init_waiter(&rt_waiter, false);
2865 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2866 * hold it while doing rt_mutex_start_proxy(), because then it will
2867 * include hb->lock in the blocking chain, even through we'll not in
2868 * fact hold it while blocking. This will lead it to report -EDEADLK
2869 * and BUG when futex_unlock_pi() interleaves with this.
2871 * Therefore acquire wait_lock while holding hb->lock, but drop the
2872 * latter before calling __rt_mutex_start_proxy_lock(). This
2873 * interleaves with futex_unlock_pi() -- which does a similar lock
2874 * handoff -- such that the latter can observe the futex_q::pi_state
2875 * before __rt_mutex_start_proxy_lock() is done.
2877 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2878 spin_unlock(q.lock_ptr);
2880 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2881 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2882 * it sees the futex_q::pi_state.
2884 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2885 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2894 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2896 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2899 spin_lock(q.lock_ptr);
2901 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2902 * first acquire the hb->lock before removing the lock from the
2903 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2906 * In particular; it is important that futex_unlock_pi() can not
2907 * observe this inconsistency.
2909 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2914 * Fixup the pi_state owner and possibly acquire the lock if we
2917 res = fixup_owner(uaddr, &q, !ret);
2919 * If fixup_owner() returned an error, proprogate that. If it acquired
2920 * the lock, clear our -ETIMEDOUT or -EINTR.
2923 ret = (res < 0) ? res : 0;
2925 /* Unqueue and drop the lock */
2934 hrtimer_cancel(&to->timer);
2935 destroy_hrtimer_on_stack(&to->timer);
2937 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2942 ret = fault_in_user_writeable(uaddr);
2946 if (!(flags & FLAGS_SHARED))
2953 * Userspace attempted a TID -> 0 atomic transition, and failed.
2954 * This is the in-kernel slowpath: we look up the PI state (if any),
2955 * and do the rt-mutex unlock.
2957 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2959 u32 curval, uval, vpid = task_pid_vnr(current);
2960 union futex_key key = FUTEX_KEY_INIT;
2961 struct futex_hash_bucket *hb;
2962 struct futex_q *top_waiter;
2965 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2969 if (get_user(uval, uaddr))
2972 * We release only a lock we actually own:
2974 if ((uval & FUTEX_TID_MASK) != vpid)
2977 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2981 hb = hash_futex(&key);
2982 spin_lock(&hb->lock);
2985 * Check waiters first. We do not trust user space values at
2986 * all and we at least want to know if user space fiddled
2987 * with the futex value instead of blindly unlocking.
2989 top_waiter = futex_top_waiter(hb, &key);
2991 struct futex_pi_state *pi_state = top_waiter->pi_state;
2998 * If current does not own the pi_state then the futex is
2999 * inconsistent and user space fiddled with the futex value.
3001 if (pi_state->owner != current)
3004 get_pi_state(pi_state);
3006 * By taking wait_lock while still holding hb->lock, we ensure
3007 * there is no point where we hold neither; and therefore
3008 * wake_futex_pi() must observe a state consistent with what we
3011 * In particular; this forces __rt_mutex_start_proxy() to
3012 * complete such that we're guaranteed to observe the
3013 * rt_waiter. Also see the WARN in wake_futex_pi().
3015 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3016 spin_unlock(&hb->lock);
3018 /* drops pi_state->pi_mutex.wait_lock */
3019 ret = wake_futex_pi(uaddr, uval, pi_state);
3021 put_pi_state(pi_state);
3024 * Success, we're done! No tricky corner cases.
3029 * The atomic access to the futex value generated a
3030 * pagefault, so retry the user-access and the wakeup:
3035 * A unconditional UNLOCK_PI op raced against a waiter
3036 * setting the FUTEX_WAITERS bit. Try again.
3041 * wake_futex_pi has detected invalid state. Tell user
3048 * We have no kernel internal state, i.e. no waiters in the
3049 * kernel. Waiters which are about to queue themselves are stuck
3050 * on hb->lock. So we can safely ignore them. We do neither
3051 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3054 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3055 spin_unlock(&hb->lock);
3070 * If uval has changed, let user space handle it.
3072 ret = (curval == uval) ? 0 : -EAGAIN;
3075 spin_unlock(&hb->lock);
3085 ret = fault_in_user_writeable(uaddr);
3093 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3094 * @hb: the hash_bucket futex_q was original enqueued on
3095 * @q: the futex_q woken while waiting to be requeued
3096 * @key2: the futex_key of the requeue target futex
3097 * @timeout: the timeout associated with the wait (NULL if none)
3099 * Detect if the task was woken on the initial futex as opposed to the requeue
3100 * target futex. If so, determine if it was a timeout or a signal that caused
3101 * the wakeup and return the appropriate error code to the caller. Must be
3102 * called with the hb lock held.
3105 * - 0 = no early wakeup detected;
3106 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3109 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3110 struct futex_q *q, union futex_key *key2,
3111 struct hrtimer_sleeper *timeout)
3116 * With the hb lock held, we avoid races while we process the wakeup.
3117 * We only need to hold hb (and not hb2) to ensure atomicity as the
3118 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3119 * It can't be requeued from uaddr2 to something else since we don't
3120 * support a PI aware source futex for requeue.
3122 if (!match_futex(&q->key, key2)) {
3123 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3125 * We were woken prior to requeue by a timeout or a signal.
3126 * Unqueue the futex_q and determine which it was.
3128 plist_del(&q->list, &hb->chain);
3131 /* Handle spurious wakeups gracefully */
3133 if (timeout && !timeout->task)
3135 else if (signal_pending(current))
3136 ret = -ERESTARTNOINTR;
3142 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3143 * @uaddr: the futex we initially wait on (non-pi)
3144 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3145 * the same type, no requeueing from private to shared, etc.
3146 * @val: the expected value of uaddr
3147 * @abs_time: absolute timeout
3148 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3149 * @uaddr2: the pi futex we will take prior to returning to user-space
3151 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3152 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3153 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3154 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3155 * without one, the pi logic would not know which task to boost/deboost, if
3156 * there was a need to.
3158 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3159 * via the following--
3160 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3161 * 2) wakeup on uaddr2 after a requeue
3165 * If 3, cleanup and return -ERESTARTNOINTR.
3167 * If 2, we may then block on trying to take the rt_mutex and return via:
3168 * 5) successful lock
3171 * 8) other lock acquisition failure
3173 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3175 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3181 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3182 u32 val, ktime_t *abs_time, u32 bitset,
3185 struct hrtimer_sleeper timeout, *to;
3186 struct rt_mutex_waiter rt_waiter;
3187 struct futex_hash_bucket *hb, *hb2;
3188 union futex_key key2 = FUTEX_KEY_INIT;
3189 struct futex_q q = futex_q_init;
3192 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3195 if (uaddr == uaddr2)
3201 to = futex_setup_timer(abs_time, &timeout, flags,
3202 current->timer_slack_ns);
3205 * The waiter is allocated on our stack, manipulated by the requeue
3206 * code while we sleep on uaddr.
3208 rt_mutex_init_waiter(&rt_waiter, false);
3210 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3211 if (unlikely(ret != 0))
3215 q.rt_waiter = &rt_waiter;
3216 q.requeue_pi_key = &key2;
3219 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3222 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3227 * The check above which compares uaddrs is not sufficient for
3228 * shared futexes. We need to compare the keys:
3230 if (match_futex(&q.key, &key2)) {
3236 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3237 futex_wait_queue_me(hb, &q, to);
3240 * On RT we must avoid races with requeue and trying to block
3241 * on two mutexes (hb->lock and uaddr2's rtmutex) by
3242 * serializing access to pi_blocked_on with pi_lock.
3244 raw_spin_lock_irq(¤t->pi_lock);
3245 if (current->pi_blocked_on) {
3247 * We have been requeued or are in the process of
3250 raw_spin_unlock_irq(¤t->pi_lock);
3253 * Setting pi_blocked_on to PI_WAKEUP_INPROGRESS
3254 * prevents a concurrent requeue from moving us to the
3255 * uaddr2 rtmutex. After that we can safely acquire
3256 * (and possibly block on) hb->lock.
3258 current->pi_blocked_on = PI_WAKEUP_INPROGRESS;
3259 raw_spin_unlock_irq(¤t->pi_lock);
3261 spin_lock(&hb->lock);
3264 * Clean up pi_blocked_on. We might leak it otherwise
3265 * when we succeeded with the hb->lock in the fast
3268 raw_spin_lock_irq(¤t->pi_lock);
3269 current->pi_blocked_on = NULL;
3270 raw_spin_unlock_irq(¤t->pi_lock);
3272 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3273 spin_unlock(&hb->lock);
3279 * In order to be here, we have either been requeued, are in
3280 * the process of being requeued, or requeue successfully
3281 * acquired uaddr2 on our behalf. If pi_blocked_on was
3282 * non-null above, we may be racing with a requeue. Do not
3283 * rely on q->lock_ptr to be hb2->lock until after blocking on
3284 * hb->lock or hb2->lock. The futex_requeue dropped our key1
3285 * reference and incremented our key2 reference count.
3287 hb2 = hash_futex(&key2);
3289 /* Check if the requeue code acquired the second futex for us. */
3292 * Got the lock. We might not be the anticipated owner if we
3293 * did a lock-steal - fix up the PI-state in that case.
3295 if (q.pi_state && (q.pi_state->owner != current)) {
3296 spin_lock(&hb2->lock);
3297 BUG_ON(&hb2->lock != q.lock_ptr);
3298 ret = fixup_pi_state_owner(uaddr2, &q, current);
3300 * Drop the reference to the pi state which
3301 * the requeue_pi() code acquired for us.
3303 put_pi_state(q.pi_state);
3304 spin_unlock(&hb2->lock);
3306 * Adjust the return value. It's either -EFAULT or
3307 * success (1) but the caller expects 0 for success.
3309 ret = ret < 0 ? ret : 0;
3312 struct rt_mutex *pi_mutex;
3315 * We have been woken up by futex_unlock_pi(), a timeout, or a
3316 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3319 WARN_ON(!q.pi_state);
3320 pi_mutex = &q.pi_state->pi_mutex;
3321 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3323 spin_lock(&hb2->lock);
3324 BUG_ON(&hb2->lock != q.lock_ptr);
3325 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3328 debug_rt_mutex_free_waiter(&rt_waiter);
3330 * Fixup the pi_state owner and possibly acquire the lock if we
3333 res = fixup_owner(uaddr2, &q, !ret);
3335 * If fixup_owner() returned an error, proprogate that. If it
3336 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3339 ret = (res < 0) ? res : 0;
3341 /* Unqueue and drop the lock. */
3345 if (ret == -EINTR) {
3347 * We've already been requeued, but cannot restart by calling
3348 * futex_lock_pi() directly. We could restart this syscall, but
3349 * it would detect that the user space "val" changed and return
3350 * -EWOULDBLOCK. Save the overhead of the restart and return
3351 * -EWOULDBLOCK directly.
3358 hrtimer_cancel(&to->timer);
3359 destroy_hrtimer_on_stack(&to->timer);
3365 * Support for robust futexes: the kernel cleans up held futexes at
3368 * Implementation: user-space maintains a per-thread list of locks it
3369 * is holding. Upon do_exit(), the kernel carefully walks this list,
3370 * and marks all locks that are owned by this thread with the
3371 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3372 * always manipulated with the lock held, so the list is private and
3373 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3374 * field, to allow the kernel to clean up if the thread dies after
3375 * acquiring the lock, but just before it could have added itself to
3376 * the list. There can only be one such pending lock.
3380 * sys_set_robust_list() - Set the robust-futex list head of a task
3381 * @head: pointer to the list-head
3382 * @len: length of the list-head, as userspace expects
3384 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3387 if (!futex_cmpxchg_enabled)
3390 * The kernel knows only one size for now:
3392 if (unlikely(len != sizeof(*head)))
3395 current->robust_list = head;
3401 * sys_get_robust_list() - Get the robust-futex list head of a task
3402 * @pid: pid of the process [zero for current task]
3403 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3404 * @len_ptr: pointer to a length field, the kernel fills in the header size
3406 SYSCALL_DEFINE3(get_robust_list, int, pid,
3407 struct robust_list_head __user * __user *, head_ptr,
3408 size_t __user *, len_ptr)
3410 struct robust_list_head __user *head;
3412 struct task_struct *p;
3414 if (!futex_cmpxchg_enabled)
3423 p = find_task_by_vpid(pid);
3429 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3432 head = p->robust_list;
3435 if (put_user(sizeof(*head), len_ptr))
3437 return put_user(head, head_ptr);
3445 /* Constants for the pending_op argument of handle_futex_death */
3446 #define HANDLE_DEATH_PENDING true
3447 #define HANDLE_DEATH_LIST false
3450 * Process a futex-list entry, check whether it's owned by the
3451 * dying task, and do notification if so:
3453 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3454 bool pi, bool pending_op)
3456 u32 uval, nval, mval;
3459 /* Futex address must be 32bit aligned */
3460 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3464 if (get_user(uval, uaddr))
3468 * Special case for regular (non PI) futexes. The unlock path in
3469 * user space has two race scenarios:
3471 * 1. The unlock path releases the user space futex value and
3472 * before it can execute the futex() syscall to wake up
3473 * waiters it is killed.
3475 * 2. A woken up waiter is killed before it can acquire the
3476 * futex in user space.
3478 * In both cases the TID validation below prevents a wakeup of
3479 * potential waiters which can cause these waiters to block
3482 * In both cases the following conditions are met:
3484 * 1) task->robust_list->list_op_pending != NULL
3485 * @pending_op == true
3486 * 2) User space futex value == 0
3487 * 3) Regular futex: @pi == false
3489 * If these conditions are met, it is safe to attempt waking up a
3490 * potential waiter without touching the user space futex value and
3491 * trying to set the OWNER_DIED bit. The user space futex value is
3492 * uncontended and the rest of the user space mutex state is
3493 * consistent, so a woken waiter will just take over the
3494 * uncontended futex. Setting the OWNER_DIED bit would create
3495 * inconsistent state and malfunction of the user space owner died
3498 if (pending_op && !pi && !uval) {
3499 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3503 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3507 * Ok, this dying thread is truly holding a futex
3508 * of interest. Set the OWNER_DIED bit atomically
3509 * via cmpxchg, and if the value had FUTEX_WAITERS
3510 * set, wake up a waiter (if any). (We have to do a
3511 * futex_wake() even if OWNER_DIED is already set -
3512 * to handle the rare but possible case of recursive
3513 * thread-death.) The rest of the cleanup is done in
3516 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3519 * We are not holding a lock here, but we want to have
3520 * the pagefault_disable/enable() protection because
3521 * we want to handle the fault gracefully. If the
3522 * access fails we try to fault in the futex with R/W
3523 * verification via get_user_pages. get_user() above
3524 * does not guarantee R/W access. If that fails we
3525 * give up and leave the futex locked.
3527 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3530 if (fault_in_user_writeable(uaddr))
3548 * Wake robust non-PI futexes here. The wakeup of
3549 * PI futexes happens in exit_pi_state():
3551 if (!pi && (uval & FUTEX_WAITERS))
3552 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3558 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3560 static inline int fetch_robust_entry(struct robust_list __user **entry,
3561 struct robust_list __user * __user *head,
3564 unsigned long uentry;
3566 if (get_user(uentry, (unsigned long __user *)head))
3569 *entry = (void __user *)(uentry & ~1UL);
3576 * Walk curr->robust_list (very carefully, it's a userspace list!)
3577 * and mark any locks found there dead, and notify any waiters.
3579 * We silently return on any sign of list-walking problem.
3581 static void exit_robust_list(struct task_struct *curr)
3583 struct robust_list_head __user *head = curr->robust_list;
3584 struct robust_list __user *entry, *next_entry, *pending;
3585 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3586 unsigned int next_pi;
3587 unsigned long futex_offset;
3590 if (!futex_cmpxchg_enabled)
3594 * Fetch the list head (which was registered earlier, via
3595 * sys_set_robust_list()):
3597 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3600 * Fetch the relative futex offset:
3602 if (get_user(futex_offset, &head->futex_offset))
3605 * Fetch any possibly pending lock-add first, and handle it
3608 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3611 next_entry = NULL; /* avoid warning with gcc */
3612 while (entry != &head->list) {
3614 * Fetch the next entry in the list before calling
3615 * handle_futex_death:
3617 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3619 * A pending lock might already be on the list, so
3620 * don't process it twice:
3622 if (entry != pending) {
3623 if (handle_futex_death((void __user *)entry + futex_offset,
3624 curr, pi, HANDLE_DEATH_LIST))
3632 * Avoid excessively long or circular lists:
3641 handle_futex_death((void __user *)pending + futex_offset,
3642 curr, pip, HANDLE_DEATH_PENDING);
3646 static void futex_cleanup(struct task_struct *tsk)
3648 if (unlikely(tsk->robust_list)) {
3649 exit_robust_list(tsk);
3650 tsk->robust_list = NULL;
3653 #ifdef CONFIG_COMPAT
3654 if (unlikely(tsk->compat_robust_list)) {
3655 compat_exit_robust_list(tsk);
3656 tsk->compat_robust_list = NULL;
3660 if (unlikely(!list_empty(&tsk->pi_state_list)))
3661 exit_pi_state_list(tsk);
3665 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3666 * @tsk: task to set the state on
3668 * Set the futex exit state of the task lockless. The futex waiter code
3669 * observes that state when a task is exiting and loops until the task has
3670 * actually finished the futex cleanup. The worst case for this is that the
3671 * waiter runs through the wait loop until the state becomes visible.
3673 * This is called from the recursive fault handling path in do_exit().
3675 * This is best effort. Either the futex exit code has run already or
3676 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3677 * take it over. If not, the problem is pushed back to user space. If the
3678 * futex exit code did not run yet, then an already queued waiter might
3679 * block forever, but there is nothing which can be done about that.
3681 void futex_exit_recursive(struct task_struct *tsk)
3683 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3684 if (tsk->futex_state == FUTEX_STATE_EXITING)
3685 mutex_unlock(&tsk->futex_exit_mutex);
3686 tsk->futex_state = FUTEX_STATE_DEAD;
3689 static void futex_cleanup_begin(struct task_struct *tsk)
3692 * Prevent various race issues against a concurrent incoming waiter
3693 * including live locks by forcing the waiter to block on
3694 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3695 * attach_to_pi_owner().
3697 mutex_lock(&tsk->futex_exit_mutex);
3700 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3702 * This ensures that all subsequent checks of tsk->futex_state in
3703 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3704 * tsk->pi_lock held.
3706 * It guarantees also that a pi_state which was queued right before
3707 * the state change under tsk->pi_lock by a concurrent waiter must
3708 * be observed in exit_pi_state_list().
3710 raw_spin_lock_irq(&tsk->pi_lock);
3711 tsk->futex_state = FUTEX_STATE_EXITING;
3712 raw_spin_unlock_irq(&tsk->pi_lock);
3715 static void futex_cleanup_end(struct task_struct *tsk, int state)
3718 * Lockless store. The only side effect is that an observer might
3719 * take another loop until it becomes visible.
3721 tsk->futex_state = state;
3723 * Drop the exit protection. This unblocks waiters which observed
3724 * FUTEX_STATE_EXITING to reevaluate the state.
3726 mutex_unlock(&tsk->futex_exit_mutex);
3729 void futex_exec_release(struct task_struct *tsk)
3732 * The state handling is done for consistency, but in the case of
3733 * exec() there is no way to prevent futher damage as the PID stays
3734 * the same. But for the unlikely and arguably buggy case that a
3735 * futex is held on exec(), this provides at least as much state
3736 * consistency protection which is possible.
3738 futex_cleanup_begin(tsk);
3741 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3742 * exec a new binary.
3744 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3747 void futex_exit_release(struct task_struct *tsk)
3749 futex_cleanup_begin(tsk);
3751 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3754 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3755 u32 __user *uaddr2, u32 val2, u32 val3)
3757 int cmd = op & FUTEX_CMD_MASK;
3758 unsigned int flags = 0;
3760 if (!(op & FUTEX_PRIVATE_FLAG))
3761 flags |= FLAGS_SHARED;
3763 if (op & FUTEX_CLOCK_REALTIME) {
3764 flags |= FLAGS_CLOCKRT;
3765 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3771 case FUTEX_UNLOCK_PI:
3772 case FUTEX_TRYLOCK_PI:
3773 case FUTEX_WAIT_REQUEUE_PI:
3774 case FUTEX_CMP_REQUEUE_PI:
3775 if (!futex_cmpxchg_enabled)
3781 val3 = FUTEX_BITSET_MATCH_ANY;
3783 case FUTEX_WAIT_BITSET:
3784 return futex_wait(uaddr, flags, val, timeout, val3);
3786 val3 = FUTEX_BITSET_MATCH_ANY;
3788 case FUTEX_WAKE_BITSET:
3789 return futex_wake(uaddr, flags, val, val3);
3791 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3792 case FUTEX_CMP_REQUEUE:
3793 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3795 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3797 return futex_lock_pi(uaddr, flags, timeout, 0);
3798 case FUTEX_UNLOCK_PI:
3799 return futex_unlock_pi(uaddr, flags);
3800 case FUTEX_TRYLOCK_PI:
3801 return futex_lock_pi(uaddr, flags, NULL, 1);
3802 case FUTEX_WAIT_REQUEUE_PI:
3803 val3 = FUTEX_BITSET_MATCH_ANY;
3804 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3806 case FUTEX_CMP_REQUEUE_PI:
3807 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3813 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3814 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3817 struct timespec64 ts;
3818 ktime_t t, *tp = NULL;
3820 int cmd = op & FUTEX_CMD_MASK;
3822 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3823 cmd == FUTEX_WAIT_BITSET ||
3824 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3825 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3827 if (get_timespec64(&ts, utime))
3829 if (!timespec64_valid(&ts))
3832 t = timespec64_to_ktime(ts);
3833 if (cmd == FUTEX_WAIT)
3834 t = ktime_add_safe(ktime_get(), t);
3835 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3836 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3840 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3841 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3843 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3844 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3845 val2 = (u32) (unsigned long) utime;
3847 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3850 #ifdef CONFIG_COMPAT
3852 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3855 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3856 compat_uptr_t __user *head, unsigned int *pi)
3858 if (get_user(*uentry, head))
3861 *entry = compat_ptr((*uentry) & ~1);
3862 *pi = (unsigned int)(*uentry) & 1;
3867 static void __user *futex_uaddr(struct robust_list __user *entry,
3868 compat_long_t futex_offset)
3870 compat_uptr_t base = ptr_to_compat(entry);
3871 void __user *uaddr = compat_ptr(base + futex_offset);
3877 * Walk curr->robust_list (very carefully, it's a userspace list!)
3878 * and mark any locks found there dead, and notify any waiters.
3880 * We silently return on any sign of list-walking problem.
3882 static void compat_exit_robust_list(struct task_struct *curr)
3884 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3885 struct robust_list __user *entry, *next_entry, *pending;
3886 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3887 unsigned int next_pi;
3888 compat_uptr_t uentry, next_uentry, upending;
3889 compat_long_t futex_offset;
3892 if (!futex_cmpxchg_enabled)
3896 * Fetch the list head (which was registered earlier, via
3897 * sys_set_robust_list()):
3899 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3902 * Fetch the relative futex offset:
3904 if (get_user(futex_offset, &head->futex_offset))
3907 * Fetch any possibly pending lock-add first, and handle it
3910 if (compat_fetch_robust_entry(&upending, &pending,
3911 &head->list_op_pending, &pip))
3914 next_entry = NULL; /* avoid warning with gcc */
3915 while (entry != (struct robust_list __user *) &head->list) {
3917 * Fetch the next entry in the list before calling
3918 * handle_futex_death:
3920 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3921 (compat_uptr_t __user *)&entry->next, &next_pi);
3923 * A pending lock might already be on the list, so
3924 * dont process it twice:
3926 if (entry != pending) {
3927 void __user *uaddr = futex_uaddr(entry, futex_offset);
3929 if (handle_futex_death(uaddr, curr, pi,
3935 uentry = next_uentry;
3939 * Avoid excessively long or circular lists:
3947 void __user *uaddr = futex_uaddr(pending, futex_offset);
3949 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3953 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3954 struct compat_robust_list_head __user *, head,
3957 if (!futex_cmpxchg_enabled)
3960 if (unlikely(len != sizeof(*head)))
3963 current->compat_robust_list = head;
3968 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3969 compat_uptr_t __user *, head_ptr,
3970 compat_size_t __user *, len_ptr)
3972 struct compat_robust_list_head __user *head;
3974 struct task_struct *p;
3976 if (!futex_cmpxchg_enabled)
3985 p = find_task_by_vpid(pid);
3991 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3994 head = p->compat_robust_list;
3997 if (put_user(sizeof(*head), len_ptr))
3999 return put_user(ptr_to_compat(head), head_ptr);
4006 #endif /* CONFIG_COMPAT */
4008 #ifdef CONFIG_COMPAT_32BIT_TIME
4009 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
4010 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
4013 struct timespec64 ts;
4014 ktime_t t, *tp = NULL;
4016 int cmd = op & FUTEX_CMD_MASK;
4018 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
4019 cmd == FUTEX_WAIT_BITSET ||
4020 cmd == FUTEX_WAIT_REQUEUE_PI)) {
4021 if (get_old_timespec32(&ts, utime))
4023 if (!timespec64_valid(&ts))
4026 t = timespec64_to_ktime(ts);
4027 if (cmd == FUTEX_WAIT)
4028 t = ktime_add_safe(ktime_get(), t);
4029 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
4030 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
4033 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4034 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4035 val2 = (int) (unsigned long) utime;
4037 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4039 #endif /* CONFIG_COMPAT_32BIT_TIME */
4041 static void __init futex_detect_cmpxchg(void)
4043 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4047 * This will fail and we want it. Some arch implementations do
4048 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4049 * functionality. We want to know that before we call in any
4050 * of the complex code paths. Also we want to prevent
4051 * registration of robust lists in that case. NULL is
4052 * guaranteed to fault and we get -EFAULT on functional
4053 * implementation, the non-functional ones will return
4056 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4057 futex_cmpxchg_enabled = 1;
4061 static int __init futex_init(void)
4063 unsigned int futex_shift;
4066 #if CONFIG_BASE_SMALL
4067 futex_hashsize = 16;
4069 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4072 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4074 futex_hashsize < 256 ? HASH_SMALL : 0,
4076 futex_hashsize, futex_hashsize);
4077 futex_hashsize = 1UL << futex_shift;
4079 futex_detect_cmpxchg();
4081 for (i = 0; i < futex_hashsize; i++) {
4082 atomic_set(&futex_queues[i].waiters, 0);
4083 plist_head_init(&futex_queues[i].chain);
4084 spin_lock_init(&futex_queues[i].lock);
4089 core_initcall(futex_init);