2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/wake_q.h>
65 #include <linux/sched/mm.h>
66 #include <linux/hugetlb.h>
67 #include <linux/freezer.h>
68 #include <linux/bootmem.h>
69 #include <linux/fault-inject.h>
71 #include <asm/futex.h>
73 #include "locking/rtmutex_common.h"
76 * READ this before attempting to hack on futexes!
78 * Basic futex operation and ordering guarantees
79 * =============================================
81 * The waiter reads the futex value in user space and calls
82 * futex_wait(). This function computes the hash bucket and acquires
83 * the hash bucket lock. After that it reads the futex user space value
84 * again and verifies that the data has not changed. If it has not changed
85 * it enqueues itself into the hash bucket, releases the hash bucket lock
88 * The waker side modifies the user space value of the futex and calls
89 * futex_wake(). This function computes the hash bucket and acquires the
90 * hash bucket lock. Then it looks for waiters on that futex in the hash
91 * bucket and wakes them.
93 * In futex wake up scenarios where no tasks are blocked on a futex, taking
94 * the hb spinlock can be avoided and simply return. In order for this
95 * optimization to work, ordering guarantees must exist so that the waiter
96 * being added to the list is acknowledged when the list is concurrently being
97 * checked by the waker, avoiding scenarios like the following:
101 * sys_futex(WAIT, futex, val);
102 * futex_wait(futex, val);
105 * sys_futex(WAKE, futex);
110 * lock(hash_bucket(futex));
112 * unlock(hash_bucket(futex));
115 * This would cause the waiter on CPU 0 to wait forever because it
116 * missed the transition of the user space value from val to newval
117 * and the waker did not find the waiter in the hash bucket queue.
119 * The correct serialization ensures that a waiter either observes
120 * the changed user space value before blocking or is woken by a
125 * sys_futex(WAIT, futex, val);
126 * futex_wait(futex, val);
129 * smp_mb(); (A) <-- paired with -.
131 * lock(hash_bucket(futex)); |
135 * | sys_futex(WAKE, futex);
136 * | futex_wake(futex);
138 * `--------> smp_mb(); (B)
141 * unlock(hash_bucket(futex));
142 * schedule(); if (waiters)
143 * lock(hash_bucket(futex));
144 * else wake_waiters(futex);
145 * waiters--; (b) unlock(hash_bucket(futex));
147 * Where (A) orders the waiters increment and the futex value read through
148 * atomic operations (see hb_waiters_inc) and where (B) orders the write
149 * to futex and the waiters read -- this is done by the barriers for both
150 * shared and private futexes in get_futex_key_refs().
152 * This yields the following case (where X:=waiters, Y:=futex):
160 * Which guarantees that x==0 && y==0 is impossible; which translates back into
161 * the guarantee that we cannot both miss the futex variable change and the
164 * Note that a new waiter is accounted for in (a) even when it is possible that
165 * the wait call can return error, in which case we backtrack from it in (b).
166 * Refer to the comment in queue_lock().
168 * Similarly, in order to account for waiters being requeued on another
169 * address we always increment the waiters for the destination bucket before
170 * acquiring the lock. It then decrements them again after releasing it -
171 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172 * will do the additional required waiter count housekeeping. This is done for
173 * double_lock_hb() and double_unlock_hb(), respectively.
176 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
177 int __read_mostly futex_cmpxchg_enabled;
181 * Futex flags used to encode options to functions and preserve them across
185 # define FLAGS_SHARED 0x01
188 * NOMMU does not have per process address space. Let the compiler optimize
191 # define FLAGS_SHARED 0x00
193 #define FLAGS_CLOCKRT 0x02
194 #define FLAGS_HAS_TIMEOUT 0x04
197 * Priority Inheritance state:
199 struct futex_pi_state {
201 * list of 'owned' pi_state instances - these have to be
202 * cleaned up in do_exit() if the task exits prematurely:
204 struct list_head list;
209 struct rt_mutex pi_mutex;
211 struct task_struct *owner;
215 } __randomize_layout;
218 * struct futex_q - The hashed futex queue entry, one per waiting task
219 * @list: priority-sorted list of tasks waiting on this futex
220 * @task: the task waiting on the futex
221 * @lock_ptr: the hash bucket lock
222 * @key: the key the futex is hashed on
223 * @pi_state: optional priority inheritance state
224 * @rt_waiter: rt_waiter storage for use with requeue_pi
225 * @requeue_pi_key: the requeue_pi target futex key
226 * @bitset: bitset for the optional bitmasked wakeup
228 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
229 * we can wake only the relevant ones (hashed queues may be shared).
231 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
232 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
233 * The order of wakeup is always to make the first condition true, then
236 * PI futexes are typically woken before they are removed from the hash list via
237 * the rt_mutex code. See unqueue_me_pi().
240 struct plist_node list;
242 struct task_struct *task;
243 spinlock_t *lock_ptr;
245 struct futex_pi_state *pi_state;
246 struct rt_mutex_waiter *rt_waiter;
247 union futex_key *requeue_pi_key;
249 } __randomize_layout;
251 static const struct futex_q futex_q_init = {
252 /* list gets initialized in queue_me()*/
253 .key = FUTEX_KEY_INIT,
254 .bitset = FUTEX_BITSET_MATCH_ANY
258 * Hash buckets are shared by all the futex_keys that hash to the same
259 * location. Each key may have multiple futex_q structures, one for each task
260 * waiting on a futex.
262 struct futex_hash_bucket {
265 struct plist_head chain;
266 } ____cacheline_aligned_in_smp;
269 * The base of the bucket array and its size are always used together
270 * (after initialization only in hash_futex()), so ensure that they
271 * reside in the same cacheline.
274 struct futex_hash_bucket *queues;
275 unsigned long hashsize;
276 } __futex_data __read_mostly __aligned(2*sizeof(long));
277 #define futex_queues (__futex_data.queues)
278 #define futex_hashsize (__futex_data.hashsize)
282 * Fault injections for futexes.
284 #ifdef CONFIG_FAIL_FUTEX
287 struct fault_attr attr;
291 .attr = FAULT_ATTR_INITIALIZER,
292 .ignore_private = false,
295 static int __init setup_fail_futex(char *str)
297 return setup_fault_attr(&fail_futex.attr, str);
299 __setup("fail_futex=", setup_fail_futex);
301 static bool should_fail_futex(bool fshared)
303 if (fail_futex.ignore_private && !fshared)
306 return should_fail(&fail_futex.attr, 1);
309 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
311 static int __init fail_futex_debugfs(void)
313 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
316 dir = fault_create_debugfs_attr("fail_futex", NULL,
321 if (!debugfs_create_bool("ignore-private", mode, dir,
322 &fail_futex.ignore_private)) {
323 debugfs_remove_recursive(dir);
330 late_initcall(fail_futex_debugfs);
332 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
335 static inline bool should_fail_futex(bool fshared)
339 #endif /* CONFIG_FAIL_FUTEX */
341 static inline void futex_get_mm(union futex_key *key)
343 mmgrab(key->private.mm);
345 * Ensure futex_get_mm() implies a full barrier such that
346 * get_futex_key() implies a full barrier. This is relied upon
347 * as smp_mb(); (B), see the ordering comment above.
349 smp_mb__after_atomic();
353 * Reflects a new waiter being added to the waitqueue.
355 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
358 atomic_inc(&hb->waiters);
360 * Full barrier (A), see the ordering comment above.
362 smp_mb__after_atomic();
367 * Reflects a waiter being removed from the waitqueue by wakeup
370 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
373 atomic_dec(&hb->waiters);
377 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
380 return atomic_read(&hb->waiters);
387 * hash_futex - Return the hash bucket in the global hash
388 * @key: Pointer to the futex key for which the hash is calculated
390 * We hash on the keys returned from get_futex_key (see below) and return the
391 * corresponding hash bucket in the global hash.
393 static struct futex_hash_bucket *hash_futex(union futex_key *key)
395 u32 hash = jhash2((u32*)&key->both.word,
396 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
398 return &futex_queues[hash & (futex_hashsize - 1)];
403 * match_futex - Check whether two futex keys are equal
404 * @key1: Pointer to key1
405 * @key2: Pointer to key2
407 * Return 1 if two futex_keys are equal, 0 otherwise.
409 static inline int match_futex(union futex_key *key1, union futex_key *key2)
412 && key1->both.word == key2->both.word
413 && key1->both.ptr == key2->both.ptr
414 && key1->both.offset == key2->both.offset);
418 * Take a reference to the resource addressed by a key.
419 * Can be called while holding spinlocks.
422 static void get_futex_key_refs(union futex_key *key)
428 * On MMU less systems futexes are always "private" as there is no per
429 * process address space. We need the smp wmb nevertheless - yes,
430 * arch/blackfin has MMU less SMP ...
432 if (!IS_ENABLED(CONFIG_MMU)) {
433 smp_mb(); /* explicit smp_mb(); (B) */
437 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
439 ihold(key->shared.inode); /* implies smp_mb(); (B) */
441 case FUT_OFF_MMSHARED:
442 futex_get_mm(key); /* implies smp_mb(); (B) */
446 * Private futexes do not hold reference on an inode or
447 * mm, therefore the only purpose of calling get_futex_key_refs
448 * is because we need the barrier for the lockless waiter check.
450 smp_mb(); /* explicit smp_mb(); (B) */
455 * Drop a reference to the resource addressed by a key.
456 * The hash bucket spinlock must not be held. This is
457 * a no-op for private futexes, see comment in the get
460 static void drop_futex_key_refs(union futex_key *key)
462 if (!key->both.ptr) {
463 /* If we're here then we tried to put a key we failed to get */
468 if (!IS_ENABLED(CONFIG_MMU))
471 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
473 iput(key->shared.inode);
475 case FUT_OFF_MMSHARED:
476 mmdrop(key->private.mm);
482 * get_futex_key() - Get parameters which are the keys for a futex
483 * @uaddr: virtual address of the futex
484 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485 * @key: address where result is stored.
486 * @rw: mapping needs to be read/write (values: VERIFY_READ,
489 * Return: a negative error code or 0
491 * The key words are stored in @key on success.
493 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
494 * offset_within_page). For private mappings, it's (uaddr, current->mm).
495 * We can usually work out the index without swapping in the page.
497 * lock_page() might sleep, the caller should not hold a spinlock.
500 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
502 unsigned long address = (unsigned long)uaddr;
503 struct mm_struct *mm = current->mm;
504 struct page *page, *tail;
505 struct address_space *mapping;
509 * The futex address must be "naturally" aligned.
511 key->both.offset = address % PAGE_SIZE;
512 if (unlikely((address % sizeof(u32)) != 0))
514 address -= key->both.offset;
516 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
519 if (unlikely(should_fail_futex(fshared)))
523 * PROCESS_PRIVATE futexes are fast.
524 * As the mm cannot disappear under us and the 'key' only needs
525 * virtual address, we dont even have to find the underlying vma.
526 * Note : We do have to check 'uaddr' is a valid user address,
527 * but access_ok() should be faster than find_vma()
530 key->private.mm = mm;
531 key->private.address = address;
532 get_futex_key_refs(key); /* implies smp_mb(); (B) */
537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared)))
541 err = get_user_pages_fast(address, 1, 1, &page);
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
546 if (err == -EFAULT && rw == VERIFY_READ) {
547 err = get_user_pages_fast(address, 1, 0, &page);
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
574 page = compound_head(page);
575 mapping = READ_ONCE(page->mapping);
578 * If page->mapping is NULL, then it cannot be a PageAnon
579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
590 * an unlikely race, but we do need to retry for page->mapping.
592 if (unlikely(!mapping)) {
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
601 shmem_swizzled = PageSwapCache(page) || page->mapping;
612 * Private mappings are handled in a simple way.
614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
619 * the object not the particular process.
621 if (PageAnon(page)) {
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
626 if (unlikely(should_fail_futex(fshared)) || ro) {
631 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
632 key->private.mm = mm;
633 key->private.address = address;
635 get_futex_key_refs(key); /* implies smp_mb(); (B) */
641 * The associated futex object in this case is the inode and
642 * the page->mapping must be traversed. Ordinarily this should
643 * be stabilised under page lock but it's not strictly
644 * necessary in this case as we just want to pin the inode, not
645 * update the radix tree or anything like that.
647 * The RCU read lock is taken as the inode is finally freed
648 * under RCU. If the mapping still matches expectations then the
649 * mapping->host can be safely accessed as being a valid inode.
653 if (READ_ONCE(page->mapping) != mapping) {
660 inode = READ_ONCE(mapping->host);
669 * Take a reference unless it is about to be freed. Previously
670 * this reference was taken by ihold under the page lock
671 * pinning the inode in place so i_lock was unnecessary. The
672 * only way for this check to fail is if the inode was
673 * truncated in parallel which is almost certainly an
674 * application bug. In such a case, just retry.
676 * We are not calling into get_futex_key_refs() in file-backed
677 * cases, therefore a successful atomic_inc return below will
678 * guarantee that get_futex_key() will still imply smp_mb(); (B).
680 if (!atomic_inc_not_zero(&inode->i_count)) {
687 /* Should be impossible but lets be paranoid for now */
688 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
696 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
697 key->shared.inode = inode;
698 key->shared.pgoff = basepage_index(tail);
707 static inline void put_futex_key(union futex_key *key)
709 drop_futex_key_refs(key);
713 * fault_in_user_writeable() - Fault in user address and verify RW access
714 * @uaddr: pointer to faulting user space address
716 * Slow path to fixup the fault we just took in the atomic write
719 * We have no generic implementation of a non-destructive write to the
720 * user address. We know that we faulted in the atomic pagefault
721 * disabled section so we can as well avoid the #PF overhead by
722 * calling get_user_pages() right away.
724 static int fault_in_user_writeable(u32 __user *uaddr)
726 struct mm_struct *mm = current->mm;
729 down_read(&mm->mmap_sem);
730 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
731 FAULT_FLAG_WRITE, NULL);
732 up_read(&mm->mmap_sem);
734 return ret < 0 ? ret : 0;
738 * futex_top_waiter() - Return the highest priority waiter on a futex
739 * @hb: the hash bucket the futex_q's reside in
740 * @key: the futex key (to distinguish it from other futex futex_q's)
742 * Must be called with the hb lock held.
744 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
745 union futex_key *key)
747 struct futex_q *this;
749 plist_for_each_entry(this, &hb->chain, list) {
750 if (match_futex(&this->key, key))
756 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
757 u32 uval, u32 newval)
762 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
768 static int get_futex_value_locked(u32 *dest, u32 __user *from)
773 ret = __get_user(*dest, from);
776 return ret ? -EFAULT : 0;
783 static int refill_pi_state_cache(void)
785 struct futex_pi_state *pi_state;
787 if (likely(current->pi_state_cache))
790 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
795 INIT_LIST_HEAD(&pi_state->list);
796 /* pi_mutex gets initialized later */
797 pi_state->owner = NULL;
798 atomic_set(&pi_state->refcount, 1);
799 pi_state->key = FUTEX_KEY_INIT;
801 current->pi_state_cache = pi_state;
806 static struct futex_pi_state *alloc_pi_state(void)
808 struct futex_pi_state *pi_state = current->pi_state_cache;
811 current->pi_state_cache = NULL;
816 static void get_pi_state(struct futex_pi_state *pi_state)
818 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
822 * Drops a reference to the pi_state object and frees or caches it
823 * when the last reference is gone.
825 static void put_pi_state(struct futex_pi_state *pi_state)
830 if (!atomic_dec_and_test(&pi_state->refcount))
834 * If pi_state->owner is NULL, the owner is most probably dying
835 * and has cleaned up the pi_state already
837 if (pi_state->owner) {
838 struct task_struct *owner;
840 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
841 owner = pi_state->owner;
843 raw_spin_lock(&owner->pi_lock);
844 list_del_init(&pi_state->list);
845 raw_spin_unlock(&owner->pi_lock);
847 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
848 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
851 if (current->pi_state_cache) {
855 * pi_state->list is already empty.
856 * clear pi_state->owner.
857 * refcount is at 0 - put it back to 1.
859 pi_state->owner = NULL;
860 atomic_set(&pi_state->refcount, 1);
861 current->pi_state_cache = pi_state;
865 #ifdef CONFIG_FUTEX_PI
868 * This task is holding PI mutexes at exit time => bad.
869 * Kernel cleans up PI-state, but userspace is likely hosed.
870 * (Robust-futex cleanup is separate and might save the day for userspace.)
872 void exit_pi_state_list(struct task_struct *curr)
874 struct list_head *next, *head = &curr->pi_state_list;
875 struct futex_pi_state *pi_state;
876 struct futex_hash_bucket *hb;
877 union futex_key key = FUTEX_KEY_INIT;
879 if (!futex_cmpxchg_enabled)
882 * We are a ZOMBIE and nobody can enqueue itself on
883 * pi_state_list anymore, but we have to be careful
884 * versus waiters unqueueing themselves:
886 raw_spin_lock_irq(&curr->pi_lock);
887 while (!list_empty(head)) {
889 pi_state = list_entry(next, struct futex_pi_state, list);
891 hb = hash_futex(&key);
894 * We can race against put_pi_state() removing itself from the
895 * list (a waiter going away). put_pi_state() will first
896 * decrement the reference count and then modify the list, so
897 * its possible to see the list entry but fail this reference
900 * In that case; drop the locks to let put_pi_state() make
901 * progress and retry the loop.
903 if (!atomic_inc_not_zero(&pi_state->refcount)) {
904 raw_spin_unlock_irq(&curr->pi_lock);
906 raw_spin_lock_irq(&curr->pi_lock);
909 raw_spin_unlock_irq(&curr->pi_lock);
911 spin_lock(&hb->lock);
912 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
913 raw_spin_lock(&curr->pi_lock);
915 * We dropped the pi-lock, so re-check whether this
916 * task still owns the PI-state:
918 if (head->next != next) {
919 /* retain curr->pi_lock for the loop invariant */
920 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
921 spin_unlock(&hb->lock);
922 put_pi_state(pi_state);
926 WARN_ON(pi_state->owner != curr);
927 WARN_ON(list_empty(&pi_state->list));
928 list_del_init(&pi_state->list);
929 pi_state->owner = NULL;
931 raw_spin_unlock(&curr->pi_lock);
932 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
933 spin_unlock(&hb->lock);
935 rt_mutex_futex_unlock(&pi_state->pi_mutex);
936 put_pi_state(pi_state);
938 raw_spin_lock_irq(&curr->pi_lock);
940 raw_spin_unlock_irq(&curr->pi_lock);
946 * We need to check the following states:
948 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
950 * [1] NULL | --- | --- | 0 | 0/1 | Valid
951 * [2] NULL | --- | --- | >0 | 0/1 | Valid
953 * [3] Found | NULL | -- | Any | 0/1 | Invalid
955 * [4] Found | Found | NULL | 0 | 1 | Valid
956 * [5] Found | Found | NULL | >0 | 1 | Invalid
958 * [6] Found | Found | task | 0 | 1 | Valid
960 * [7] Found | Found | NULL | Any | 0 | Invalid
962 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
963 * [9] Found | Found | task | 0 | 0 | Invalid
964 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
966 * [1] Indicates that the kernel can acquire the futex atomically. We
967 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
969 * [2] Valid, if TID does not belong to a kernel thread. If no matching
970 * thread is found then it indicates that the owner TID has died.
972 * [3] Invalid. The waiter is queued on a non PI futex
974 * [4] Valid state after exit_robust_list(), which sets the user space
975 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
977 * [5] The user space value got manipulated between exit_robust_list()
978 * and exit_pi_state_list()
980 * [6] Valid state after exit_pi_state_list() which sets the new owner in
981 * the pi_state but cannot access the user space value.
983 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
985 * [8] Owner and user space value match
987 * [9] There is no transient state which sets the user space TID to 0
988 * except exit_robust_list(), but this is indicated by the
989 * FUTEX_OWNER_DIED bit. See [4]
991 * [10] There is no transient state which leaves owner and user space
995 * Serialization and lifetime rules:
999 * hb -> futex_q, relation
1000 * futex_q -> pi_state, relation
1002 * (cannot be raw because hb can contain arbitrary amount
1005 * pi_mutex->wait_lock:
1009 * (and pi_mutex 'obviously')
1013 * p->pi_state_list -> pi_state->list, relation
1015 * pi_state->refcount:
1023 * pi_mutex->wait_lock
1029 * Validate that the existing waiter has a pi_state and sanity check
1030 * the pi_state against the user space value. If correct, attach to
1033 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1034 struct futex_pi_state *pi_state,
1035 struct futex_pi_state **ps)
1037 pid_t pid = uval & FUTEX_TID_MASK;
1042 * Userspace might have messed up non-PI and PI futexes [3]
1044 if (unlikely(!pi_state))
1048 * We get here with hb->lock held, and having found a
1049 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1050 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1051 * which in turn means that futex_lock_pi() still has a reference on
1054 * The waiter holding a reference on @pi_state also protects against
1055 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1056 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1057 * free pi_state before we can take a reference ourselves.
1059 WARN_ON(!atomic_read(&pi_state->refcount));
1062 * Now that we have a pi_state, we can acquire wait_lock
1063 * and do the state validation.
1065 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1068 * Since {uval, pi_state} is serialized by wait_lock, and our current
1069 * uval was read without holding it, it can have changed. Verify it
1070 * still is what we expect it to be, otherwise retry the entire
1073 if (get_futex_value_locked(&uval2, uaddr))
1080 * Handle the owner died case:
1082 if (uval & FUTEX_OWNER_DIED) {
1084 * exit_pi_state_list sets owner to NULL and wakes the
1085 * topmost waiter. The task which acquires the
1086 * pi_state->rt_mutex will fixup owner.
1088 if (!pi_state->owner) {
1090 * No pi state owner, but the user space TID
1091 * is not 0. Inconsistent state. [5]
1096 * Take a ref on the state and return success. [4]
1102 * If TID is 0, then either the dying owner has not
1103 * yet executed exit_pi_state_list() or some waiter
1104 * acquired the rtmutex in the pi state, but did not
1105 * yet fixup the TID in user space.
1107 * Take a ref on the state and return success. [6]
1113 * If the owner died bit is not set, then the pi_state
1114 * must have an owner. [7]
1116 if (!pi_state->owner)
1121 * Bail out if user space manipulated the futex value. If pi
1122 * state exists then the owner TID must be the same as the
1123 * user space TID. [9/10]
1125 if (pid != task_pid_vnr(pi_state->owner))
1129 get_pi_state(pi_state);
1130 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1147 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1151 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1152 struct task_struct *tsk)
1157 * If PF_EXITPIDONE is not yet set, then try again.
1159 if (tsk && !(tsk->flags & PF_EXITPIDONE))
1163 * Reread the user space value to handle the following situation:
1167 * sys_exit() sys_futex()
1168 * do_exit() futex_lock_pi()
1169 * futex_lock_pi_atomic()
1170 * exit_signals(tsk) No waiters:
1171 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1172 * mm_release(tsk) Set waiter bit
1173 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1174 * Set owner died attach_to_pi_owner() {
1175 * *uaddr = 0xC0000000; tsk = get_task(PID);
1176 * } if (!tsk->flags & PF_EXITING) {
1178 * tsk->flags |= PF_EXITPIDONE; } else {
1179 * if (!(tsk->flags & PF_EXITPIDONE))
1181 * return -ESRCH; <--- FAIL
1184 * Returning ESRCH unconditionally is wrong here because the
1185 * user space value has been changed by the exiting task.
1187 * The same logic applies to the case where the exiting task is
1190 if (get_futex_value_locked(&uval2, uaddr))
1193 /* If the user space value has changed, try again. */
1198 * The exiting task did not have a robust list, the robust list was
1199 * corrupted or the user space value in *uaddr is simply bogus.
1200 * Give up and tell user space.
1206 * Lookup the task for the TID provided from user space and attach to
1207 * it after doing proper sanity checks.
1209 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1210 struct futex_pi_state **ps)
1212 pid_t pid = uval & FUTEX_TID_MASK;
1213 struct futex_pi_state *pi_state;
1214 struct task_struct *p;
1217 * We are the first waiter - try to look up the real owner and attach
1218 * the new pi_state to it, but bail out when TID = 0 [1]
1220 * The !pid check is paranoid. None of the call sites should end up
1221 * with pid == 0, but better safe than sorry. Let the caller retry
1225 p = find_get_task_by_vpid(pid);
1227 return handle_exit_race(uaddr, uval, NULL);
1229 if (unlikely(p->flags & PF_KTHREAD)) {
1235 * We need to look at the task state flags to figure out,
1236 * whether the task is exiting. To protect against the do_exit
1237 * change of the task flags, we do this protected by
1240 raw_spin_lock_irq(&p->pi_lock);
1241 if (unlikely(p->flags & PF_EXITING)) {
1243 * The task is on the way out. When PF_EXITPIDONE is
1244 * set, we know that the task has finished the
1247 int ret = handle_exit_race(uaddr, uval, p);
1249 raw_spin_unlock_irq(&p->pi_lock);
1255 * No existing pi state. First waiter. [2]
1257 * This creates pi_state, we have hb->lock held, this means nothing can
1258 * observe this state, wait_lock is irrelevant.
1260 pi_state = alloc_pi_state();
1263 * Initialize the pi_mutex in locked state and make @p
1266 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1268 /* Store the key for possible exit cleanups: */
1269 pi_state->key = *key;
1271 WARN_ON(!list_empty(&pi_state->list));
1272 list_add(&pi_state->list, &p->pi_state_list);
1274 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1275 * because there is no concurrency as the object is not published yet.
1277 pi_state->owner = p;
1278 raw_spin_unlock_irq(&p->pi_lock);
1287 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1288 struct futex_hash_bucket *hb,
1289 union futex_key *key, struct futex_pi_state **ps)
1291 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1294 * If there is a waiter on that futex, validate it and
1295 * attach to the pi_state when the validation succeeds.
1298 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1301 * We are the first waiter - try to look up the owner based on
1302 * @uval and attach to it.
1304 return attach_to_pi_owner(uaddr, uval, key, ps);
1307 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1310 u32 uninitialized_var(curval);
1312 if (unlikely(should_fail_futex(true)))
1315 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1319 /* If user space value changed, let the caller retry */
1320 return curval != uval ? -EAGAIN : 0;
1324 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1325 * @uaddr: the pi futex user address
1326 * @hb: the pi futex hash bucket
1327 * @key: the futex key associated with uaddr and hb
1328 * @ps: the pi_state pointer where we store the result of the
1330 * @task: the task to perform the atomic lock work for. This will
1331 * be "current" except in the case of requeue pi.
1332 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1335 * - 0 - ready to wait;
1336 * - 1 - acquired the lock;
1339 * The hb->lock and futex_key refs shall be held by the caller.
1341 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1342 union futex_key *key,
1343 struct futex_pi_state **ps,
1344 struct task_struct *task, int set_waiters)
1346 u32 uval, newval, vpid = task_pid_vnr(task);
1347 struct futex_q *top_waiter;
1351 * Read the user space value first so we can validate a few
1352 * things before proceeding further.
1354 if (get_futex_value_locked(&uval, uaddr))
1357 if (unlikely(should_fail_futex(true)))
1363 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1366 if ((unlikely(should_fail_futex(true))))
1370 * Lookup existing state first. If it exists, try to attach to
1373 top_waiter = futex_top_waiter(hb, key);
1375 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1378 * No waiter and user TID is 0. We are here because the
1379 * waiters or the owner died bit is set or called from
1380 * requeue_cmp_pi or for whatever reason something took the
1383 if (!(uval & FUTEX_TID_MASK)) {
1385 * We take over the futex. No other waiters and the user space
1386 * TID is 0. We preserve the owner died bit.
1388 newval = uval & FUTEX_OWNER_DIED;
1391 /* The futex requeue_pi code can enforce the waiters bit */
1393 newval |= FUTEX_WAITERS;
1395 ret = lock_pi_update_atomic(uaddr, uval, newval);
1396 /* If the take over worked, return 1 */
1397 return ret < 0 ? ret : 1;
1401 * First waiter. Set the waiters bit before attaching ourself to
1402 * the owner. If owner tries to unlock, it will be forced into
1403 * the kernel and blocked on hb->lock.
1405 newval = uval | FUTEX_WAITERS;
1406 ret = lock_pi_update_atomic(uaddr, uval, newval);
1410 * If the update of the user space value succeeded, we try to
1411 * attach to the owner. If that fails, no harm done, we only
1412 * set the FUTEX_WAITERS bit in the user space variable.
1414 return attach_to_pi_owner(uaddr, newval, key, ps);
1418 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1419 * @q: The futex_q to unqueue
1421 * The q->lock_ptr must not be NULL and must be held by the caller.
1423 static void __unqueue_futex(struct futex_q *q)
1425 struct futex_hash_bucket *hb;
1427 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1428 || WARN_ON(plist_node_empty(&q->list)))
1431 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1432 plist_del(&q->list, &hb->chain);
1437 * The hash bucket lock must be held when this is called.
1438 * Afterwards, the futex_q must not be accessed. Callers
1439 * must ensure to later call wake_up_q() for the actual
1442 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1444 struct task_struct *p = q->task;
1446 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1452 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1453 * is written, without taking any locks. This is possible in the event
1454 * of a spurious wakeup, for example. A memory barrier is required here
1455 * to prevent the following store to lock_ptr from getting ahead of the
1456 * plist_del in __unqueue_futex().
1458 smp_store_release(&q->lock_ptr, NULL);
1461 * Queue the task for later wakeup for after we've released
1462 * the hb->lock. wake_q_add() grabs reference to p.
1464 wake_q_add(wake_q, p);
1469 * Caller must hold a reference on @pi_state.
1471 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1473 u32 uninitialized_var(curval), newval;
1474 struct task_struct *new_owner;
1475 bool postunlock = false;
1476 DEFINE_WAKE_Q(wake_q);
1479 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1480 if (WARN_ON_ONCE(!new_owner)) {
1482 * As per the comment in futex_unlock_pi() this should not happen.
1484 * When this happens, give up our locks and try again, giving
1485 * the futex_lock_pi() instance time to complete, either by
1486 * waiting on the rtmutex or removing itself from the futex
1494 * We pass it to the next owner. The WAITERS bit is always kept
1495 * enabled while there is PI state around. We cleanup the owner
1496 * died bit, because we are the owner.
1498 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1500 if (unlikely(should_fail_futex(true)))
1503 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1504 if (!ret && (curval != uval)) {
1506 * If a unconditional UNLOCK_PI operation (user space did not
1507 * try the TID->0 transition) raced with a waiter setting the
1508 * FUTEX_WAITERS flag between get_user() and locking the hash
1509 * bucket lock, retry the operation.
1511 if ((FUTEX_TID_MASK & curval) == uval)
1521 * This is a point of no return; once we modify the uval there is no
1522 * going back and subsequent operations must not fail.
1525 raw_spin_lock(&pi_state->owner->pi_lock);
1526 WARN_ON(list_empty(&pi_state->list));
1527 list_del_init(&pi_state->list);
1528 raw_spin_unlock(&pi_state->owner->pi_lock);
1530 raw_spin_lock(&new_owner->pi_lock);
1531 WARN_ON(!list_empty(&pi_state->list));
1532 list_add(&pi_state->list, &new_owner->pi_state_list);
1533 pi_state->owner = new_owner;
1534 raw_spin_unlock(&new_owner->pi_lock);
1536 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1539 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1542 rt_mutex_postunlock(&wake_q);
1548 * Express the locking dependencies for lockdep:
1551 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1554 spin_lock(&hb1->lock);
1556 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1557 } else { /* hb1 > hb2 */
1558 spin_lock(&hb2->lock);
1559 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1564 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1566 spin_unlock(&hb1->lock);
1568 spin_unlock(&hb2->lock);
1572 * Wake up waiters matching bitset queued on this futex (uaddr).
1575 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1577 struct futex_hash_bucket *hb;
1578 struct futex_q *this, *next;
1579 union futex_key key = FUTEX_KEY_INIT;
1581 DEFINE_WAKE_Q(wake_q);
1586 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1587 if (unlikely(ret != 0))
1590 hb = hash_futex(&key);
1592 /* Make sure we really have tasks to wakeup */
1593 if (!hb_waiters_pending(hb))
1596 spin_lock(&hb->lock);
1598 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1599 if (match_futex (&this->key, &key)) {
1600 if (this->pi_state || this->rt_waiter) {
1605 /* Check if one of the bits is set in both bitsets */
1606 if (!(this->bitset & bitset))
1609 mark_wake_futex(&wake_q, this);
1610 if (++ret >= nr_wake)
1615 spin_unlock(&hb->lock);
1618 put_futex_key(&key);
1623 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1625 unsigned int op = (encoded_op & 0x70000000) >> 28;
1626 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1627 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1628 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1631 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1632 if (oparg < 0 || oparg > 31) {
1633 char comm[sizeof(current->comm)];
1635 * kill this print and return -EINVAL when userspace
1638 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1639 get_task_comm(comm, current), oparg);
1645 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1648 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1653 case FUTEX_OP_CMP_EQ:
1654 return oldval == cmparg;
1655 case FUTEX_OP_CMP_NE:
1656 return oldval != cmparg;
1657 case FUTEX_OP_CMP_LT:
1658 return oldval < cmparg;
1659 case FUTEX_OP_CMP_GE:
1660 return oldval >= cmparg;
1661 case FUTEX_OP_CMP_LE:
1662 return oldval <= cmparg;
1663 case FUTEX_OP_CMP_GT:
1664 return oldval > cmparg;
1671 * Wake up all waiters hashed on the physical page that is mapped
1672 * to this virtual address:
1675 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1676 int nr_wake, int nr_wake2, int op)
1678 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1679 struct futex_hash_bucket *hb1, *hb2;
1680 struct futex_q *this, *next;
1682 DEFINE_WAKE_Q(wake_q);
1685 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1686 if (unlikely(ret != 0))
1688 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1689 if (unlikely(ret != 0))
1692 hb1 = hash_futex(&key1);
1693 hb2 = hash_futex(&key2);
1696 double_lock_hb(hb1, hb2);
1697 op_ret = futex_atomic_op_inuser(op, uaddr2);
1698 if (unlikely(op_ret < 0)) {
1699 double_unlock_hb(hb1, hb2);
1701 if (!IS_ENABLED(CONFIG_MMU) ||
1702 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1704 * we don't get EFAULT from MMU faults if we don't have
1705 * an MMU, but we might get them from range checking
1711 if (op_ret == -EFAULT) {
1712 ret = fault_in_user_writeable(uaddr2);
1717 if (!(flags & FLAGS_SHARED)) {
1722 put_futex_key(&key2);
1723 put_futex_key(&key1);
1728 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1729 if (match_futex (&this->key, &key1)) {
1730 if (this->pi_state || this->rt_waiter) {
1734 mark_wake_futex(&wake_q, this);
1735 if (++ret >= nr_wake)
1742 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1743 if (match_futex (&this->key, &key2)) {
1744 if (this->pi_state || this->rt_waiter) {
1748 mark_wake_futex(&wake_q, this);
1749 if (++op_ret >= nr_wake2)
1757 double_unlock_hb(hb1, hb2);
1760 put_futex_key(&key2);
1762 put_futex_key(&key1);
1768 * requeue_futex() - Requeue a futex_q from one hb to another
1769 * @q: the futex_q to requeue
1770 * @hb1: the source hash_bucket
1771 * @hb2: the target hash_bucket
1772 * @key2: the new key for the requeued futex_q
1775 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1776 struct futex_hash_bucket *hb2, union futex_key *key2)
1780 * If key1 and key2 hash to the same bucket, no need to
1783 if (likely(&hb1->chain != &hb2->chain)) {
1784 plist_del(&q->list, &hb1->chain);
1785 hb_waiters_dec(hb1);
1786 hb_waiters_inc(hb2);
1787 plist_add(&q->list, &hb2->chain);
1788 q->lock_ptr = &hb2->lock;
1790 get_futex_key_refs(key2);
1795 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1797 * @key: the key of the requeue target futex
1798 * @hb: the hash_bucket of the requeue target futex
1800 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1801 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1802 * to the requeue target futex so the waiter can detect the wakeup on the right
1803 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1804 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1805 * to protect access to the pi_state to fixup the owner later. Must be called
1806 * with both q->lock_ptr and hb->lock held.
1809 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1810 struct futex_hash_bucket *hb)
1812 get_futex_key_refs(key);
1817 WARN_ON(!q->rt_waiter);
1818 q->rt_waiter = NULL;
1820 q->lock_ptr = &hb->lock;
1822 wake_up_state(q->task, TASK_NORMAL);
1826 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1827 * @pifutex: the user address of the to futex
1828 * @hb1: the from futex hash bucket, must be locked by the caller
1829 * @hb2: the to futex hash bucket, must be locked by the caller
1830 * @key1: the from futex key
1831 * @key2: the to futex key
1832 * @ps: address to store the pi_state pointer
1833 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1835 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1836 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1837 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1838 * hb1 and hb2 must be held by the caller.
1841 * - 0 - failed to acquire the lock atomically;
1842 * - >0 - acquired the lock, return value is vpid of the top_waiter
1845 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1846 struct futex_hash_bucket *hb1,
1847 struct futex_hash_bucket *hb2,
1848 union futex_key *key1, union futex_key *key2,
1849 struct futex_pi_state **ps, int set_waiters)
1851 struct futex_q *top_waiter = NULL;
1855 if (get_futex_value_locked(&curval, pifutex))
1858 if (unlikely(should_fail_futex(true)))
1862 * Find the top_waiter and determine if there are additional waiters.
1863 * If the caller intends to requeue more than 1 waiter to pifutex,
1864 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1865 * as we have means to handle the possible fault. If not, don't set
1866 * the bit unecessarily as it will force the subsequent unlock to enter
1869 top_waiter = futex_top_waiter(hb1, key1);
1871 /* There are no waiters, nothing for us to do. */
1875 /* Ensure we requeue to the expected futex. */
1876 if (!match_futex(top_waiter->requeue_pi_key, key2))
1880 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1881 * the contended case or if set_waiters is 1. The pi_state is returned
1882 * in ps in contended cases.
1884 vpid = task_pid_vnr(top_waiter->task);
1885 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1888 requeue_pi_wake_futex(top_waiter, key2, hb2);
1895 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1896 * @uaddr1: source futex user address
1897 * @flags: futex flags (FLAGS_SHARED, etc.)
1898 * @uaddr2: target futex user address
1899 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1900 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1901 * @cmpval: @uaddr1 expected value (or %NULL)
1902 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1903 * pi futex (pi to pi requeue is not supported)
1905 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1906 * uaddr2 atomically on behalf of the top waiter.
1909 * - >=0 - on success, the number of tasks requeued or woken;
1912 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1913 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1914 u32 *cmpval, int requeue_pi)
1916 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1917 int drop_count = 0, task_count = 0, ret;
1918 struct futex_pi_state *pi_state = NULL;
1919 struct futex_hash_bucket *hb1, *hb2;
1920 struct futex_q *this, *next;
1921 DEFINE_WAKE_Q(wake_q);
1923 if (nr_wake < 0 || nr_requeue < 0)
1927 * When PI not supported: return -ENOSYS if requeue_pi is true,
1928 * consequently the compiler knows requeue_pi is always false past
1929 * this point which will optimize away all the conditional code
1932 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1937 * Requeue PI only works on two distinct uaddrs. This
1938 * check is only valid for private futexes. See below.
1940 if (uaddr1 == uaddr2)
1944 * requeue_pi requires a pi_state, try to allocate it now
1945 * without any locks in case it fails.
1947 if (refill_pi_state_cache())
1950 * requeue_pi must wake as many tasks as it can, up to nr_wake
1951 * + nr_requeue, since it acquires the rt_mutex prior to
1952 * returning to userspace, so as to not leave the rt_mutex with
1953 * waiters and no owner. However, second and third wake-ups
1954 * cannot be predicted as they involve race conditions with the
1955 * first wake and a fault while looking up the pi_state. Both
1956 * pthread_cond_signal() and pthread_cond_broadcast() should
1964 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1965 if (unlikely(ret != 0))
1967 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1968 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1969 if (unlikely(ret != 0))
1973 * The check above which compares uaddrs is not sufficient for
1974 * shared futexes. We need to compare the keys:
1976 if (requeue_pi && match_futex(&key1, &key2)) {
1981 hb1 = hash_futex(&key1);
1982 hb2 = hash_futex(&key2);
1985 hb_waiters_inc(hb2);
1986 double_lock_hb(hb1, hb2);
1988 if (likely(cmpval != NULL)) {
1991 ret = get_futex_value_locked(&curval, uaddr1);
1993 if (unlikely(ret)) {
1994 double_unlock_hb(hb1, hb2);
1995 hb_waiters_dec(hb2);
1997 ret = get_user(curval, uaddr1);
2001 if (!(flags & FLAGS_SHARED))
2004 put_futex_key(&key2);
2005 put_futex_key(&key1);
2008 if (curval != *cmpval) {
2014 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2016 * Attempt to acquire uaddr2 and wake the top waiter. If we
2017 * intend to requeue waiters, force setting the FUTEX_WAITERS
2018 * bit. We force this here where we are able to easily handle
2019 * faults rather in the requeue loop below.
2021 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2022 &key2, &pi_state, nr_requeue);
2025 * At this point the top_waiter has either taken uaddr2 or is
2026 * waiting on it. If the former, then the pi_state will not
2027 * exist yet, look it up one more time to ensure we have a
2028 * reference to it. If the lock was taken, ret contains the
2029 * vpid of the top waiter task.
2030 * If the lock was not taken, we have pi_state and an initial
2031 * refcount on it. In case of an error we have nothing.
2038 * If we acquired the lock, then the user space value
2039 * of uaddr2 should be vpid. It cannot be changed by
2040 * the top waiter as it is blocked on hb2 lock if it
2041 * tries to do so. If something fiddled with it behind
2042 * our back the pi state lookup might unearth it. So
2043 * we rather use the known value than rereading and
2044 * handing potential crap to lookup_pi_state.
2046 * If that call succeeds then we have pi_state and an
2047 * initial refcount on it.
2049 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2054 /* We hold a reference on the pi state. */
2057 /* If the above failed, then pi_state is NULL */
2059 double_unlock_hb(hb1, hb2);
2060 hb_waiters_dec(hb2);
2061 put_futex_key(&key2);
2062 put_futex_key(&key1);
2063 ret = fault_in_user_writeable(uaddr2);
2069 * Two reasons for this:
2070 * - Owner is exiting and we just wait for the
2072 * - The user space value changed.
2074 double_unlock_hb(hb1, hb2);
2075 hb_waiters_dec(hb2);
2076 put_futex_key(&key2);
2077 put_futex_key(&key1);
2085 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2086 if (task_count - nr_wake >= nr_requeue)
2089 if (!match_futex(&this->key, &key1))
2093 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2094 * be paired with each other and no other futex ops.
2096 * We should never be requeueing a futex_q with a pi_state,
2097 * which is awaiting a futex_unlock_pi().
2099 if ((requeue_pi && !this->rt_waiter) ||
2100 (!requeue_pi && this->rt_waiter) ||
2107 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2108 * lock, we already woke the top_waiter. If not, it will be
2109 * woken by futex_unlock_pi().
2111 if (++task_count <= nr_wake && !requeue_pi) {
2112 mark_wake_futex(&wake_q, this);
2116 /* Ensure we requeue to the expected futex for requeue_pi. */
2117 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2123 * Requeue nr_requeue waiters and possibly one more in the case
2124 * of requeue_pi if we couldn't acquire the lock atomically.
2128 * Prepare the waiter to take the rt_mutex. Take a
2129 * refcount on the pi_state and store the pointer in
2130 * the futex_q object of the waiter.
2132 get_pi_state(pi_state);
2133 this->pi_state = pi_state;
2134 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2139 * We got the lock. We do neither drop the
2140 * refcount on pi_state nor clear
2141 * this->pi_state because the waiter needs the
2142 * pi_state for cleaning up the user space
2143 * value. It will drop the refcount after
2146 requeue_pi_wake_futex(this, &key2, hb2);
2151 * rt_mutex_start_proxy_lock() detected a
2152 * potential deadlock when we tried to queue
2153 * that waiter. Drop the pi_state reference
2154 * which we took above and remove the pointer
2155 * to the state from the waiters futex_q
2158 this->pi_state = NULL;
2159 put_pi_state(pi_state);
2161 * We stop queueing more waiters and let user
2162 * space deal with the mess.
2167 requeue_futex(this, hb1, hb2, &key2);
2172 * We took an extra initial reference to the pi_state either
2173 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2174 * need to drop it here again.
2176 put_pi_state(pi_state);
2179 double_unlock_hb(hb1, hb2);
2181 hb_waiters_dec(hb2);
2184 * drop_futex_key_refs() must be called outside the spinlocks. During
2185 * the requeue we moved futex_q's from the hash bucket at key1 to the
2186 * one at key2 and updated their key pointer. We no longer need to
2187 * hold the references to key1.
2189 while (--drop_count >= 0)
2190 drop_futex_key_refs(&key1);
2193 put_futex_key(&key2);
2195 put_futex_key(&key1);
2197 return ret ? ret : task_count;
2200 /* The key must be already stored in q->key. */
2201 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2202 __acquires(&hb->lock)
2204 struct futex_hash_bucket *hb;
2206 hb = hash_futex(&q->key);
2209 * Increment the counter before taking the lock so that
2210 * a potential waker won't miss a to-be-slept task that is
2211 * waiting for the spinlock. This is safe as all queue_lock()
2212 * users end up calling queue_me(). Similarly, for housekeeping,
2213 * decrement the counter at queue_unlock() when some error has
2214 * occurred and we don't end up adding the task to the list.
2218 q->lock_ptr = &hb->lock;
2220 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2225 queue_unlock(struct futex_hash_bucket *hb)
2226 __releases(&hb->lock)
2228 spin_unlock(&hb->lock);
2232 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2237 * The priority used to register this element is
2238 * - either the real thread-priority for the real-time threads
2239 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2240 * - or MAX_RT_PRIO for non-RT threads.
2241 * Thus, all RT-threads are woken first in priority order, and
2242 * the others are woken last, in FIFO order.
2244 prio = min(current->normal_prio, MAX_RT_PRIO);
2246 plist_node_init(&q->list, prio);
2247 plist_add(&q->list, &hb->chain);
2252 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2253 * @q: The futex_q to enqueue
2254 * @hb: The destination hash bucket
2256 * The hb->lock must be held by the caller, and is released here. A call to
2257 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2258 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2259 * or nothing if the unqueue is done as part of the wake process and the unqueue
2260 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2263 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2264 __releases(&hb->lock)
2267 spin_unlock(&hb->lock);
2271 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2272 * @q: The futex_q to unqueue
2274 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2275 * be paired with exactly one earlier call to queue_me().
2278 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2279 * - 0 - if the futex_q was already removed by the waking thread
2281 static int unqueue_me(struct futex_q *q)
2283 spinlock_t *lock_ptr;
2286 /* In the common case we don't take the spinlock, which is nice. */
2289 * q->lock_ptr can change between this read and the following spin_lock.
2290 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2291 * optimizing lock_ptr out of the logic below.
2293 lock_ptr = READ_ONCE(q->lock_ptr);
2294 if (lock_ptr != NULL) {
2295 spin_lock(lock_ptr);
2297 * q->lock_ptr can change between reading it and
2298 * spin_lock(), causing us to take the wrong lock. This
2299 * corrects the race condition.
2301 * Reasoning goes like this: if we have the wrong lock,
2302 * q->lock_ptr must have changed (maybe several times)
2303 * between reading it and the spin_lock(). It can
2304 * change again after the spin_lock() but only if it was
2305 * already changed before the spin_lock(). It cannot,
2306 * however, change back to the original value. Therefore
2307 * we can detect whether we acquired the correct lock.
2309 if (unlikely(lock_ptr != q->lock_ptr)) {
2310 spin_unlock(lock_ptr);
2315 BUG_ON(q->pi_state);
2317 spin_unlock(lock_ptr);
2321 drop_futex_key_refs(&q->key);
2326 * PI futexes can not be requeued and must remove themself from the
2327 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2330 static void unqueue_me_pi(struct futex_q *q)
2331 __releases(q->lock_ptr)
2335 BUG_ON(!q->pi_state);
2336 put_pi_state(q->pi_state);
2339 spin_unlock(q->lock_ptr);
2342 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2343 struct task_struct *argowner)
2345 struct futex_pi_state *pi_state = q->pi_state;
2346 u32 uval, uninitialized_var(curval), newval;
2347 struct task_struct *oldowner, *newowner;
2351 lockdep_assert_held(q->lock_ptr);
2353 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2355 oldowner = pi_state->owner;
2358 * We are here because either:
2360 * - we stole the lock and pi_state->owner needs updating to reflect
2361 * that (@argowner == current),
2365 * - someone stole our lock and we need to fix things to point to the
2366 * new owner (@argowner == NULL).
2368 * Either way, we have to replace the TID in the user space variable.
2369 * This must be atomic as we have to preserve the owner died bit here.
2371 * Note: We write the user space value _before_ changing the pi_state
2372 * because we can fault here. Imagine swapped out pages or a fork
2373 * that marked all the anonymous memory readonly for cow.
2375 * Modifying pi_state _before_ the user space value would leave the
2376 * pi_state in an inconsistent state when we fault here, because we
2377 * need to drop the locks to handle the fault. This might be observed
2378 * in the PID check in lookup_pi_state.
2382 if (oldowner != current) {
2384 * We raced against a concurrent self; things are
2385 * already fixed up. Nothing to do.
2391 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2392 /* We got the lock after all, nothing to fix. */
2398 * Since we just failed the trylock; there must be an owner.
2400 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2403 WARN_ON_ONCE(argowner != current);
2404 if (oldowner == current) {
2406 * We raced against a concurrent self; things are
2407 * already fixed up. Nothing to do.
2412 newowner = argowner;
2415 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2417 if (!pi_state->owner)
2418 newtid |= FUTEX_OWNER_DIED;
2420 err = get_futex_value_locked(&uval, uaddr);
2425 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2427 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2437 * We fixed up user space. Now we need to fix the pi_state
2440 if (pi_state->owner != NULL) {
2441 raw_spin_lock(&pi_state->owner->pi_lock);
2442 WARN_ON(list_empty(&pi_state->list));
2443 list_del_init(&pi_state->list);
2444 raw_spin_unlock(&pi_state->owner->pi_lock);
2447 pi_state->owner = newowner;
2449 raw_spin_lock(&newowner->pi_lock);
2450 WARN_ON(!list_empty(&pi_state->list));
2451 list_add(&pi_state->list, &newowner->pi_state_list);
2452 raw_spin_unlock(&newowner->pi_lock);
2453 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2458 * In order to reschedule or handle a page fault, we need to drop the
2459 * locks here. In the case of a fault, this gives the other task
2460 * (either the highest priority waiter itself or the task which stole
2461 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2462 * are back from handling the fault we need to check the pi_state after
2463 * reacquiring the locks and before trying to do another fixup. When
2464 * the fixup has been done already we simply return.
2466 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2467 * drop hb->lock since the caller owns the hb -> futex_q relation.
2468 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2471 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2472 spin_unlock(q->lock_ptr);
2476 ret = fault_in_user_writeable(uaddr);
2490 spin_lock(q->lock_ptr);
2491 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2494 * Check if someone else fixed it for us:
2496 if (pi_state->owner != oldowner) {
2507 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2511 static long futex_wait_restart(struct restart_block *restart);
2514 * fixup_owner() - Post lock pi_state and corner case management
2515 * @uaddr: user address of the futex
2516 * @q: futex_q (contains pi_state and access to the rt_mutex)
2517 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2519 * After attempting to lock an rt_mutex, this function is called to cleanup
2520 * the pi_state owner as well as handle race conditions that may allow us to
2521 * acquire the lock. Must be called with the hb lock held.
2524 * - 1 - success, lock taken;
2525 * - 0 - success, lock not taken;
2526 * - <0 - on error (-EFAULT)
2528 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2534 * Got the lock. We might not be the anticipated owner if we
2535 * did a lock-steal - fix up the PI-state in that case:
2537 * Speculative pi_state->owner read (we don't hold wait_lock);
2538 * since we own the lock pi_state->owner == current is the
2539 * stable state, anything else needs more attention.
2541 if (q->pi_state->owner != current)
2542 ret = fixup_pi_state_owner(uaddr, q, current);
2547 * If we didn't get the lock; check if anybody stole it from us. In
2548 * that case, we need to fix up the uval to point to them instead of
2549 * us, otherwise bad things happen. [10]
2551 * Another speculative read; pi_state->owner == current is unstable
2552 * but needs our attention.
2554 if (q->pi_state->owner == current) {
2555 ret = fixup_pi_state_owner(uaddr, q, NULL);
2560 * Paranoia check. If we did not take the lock, then we should not be
2561 * the owner of the rt_mutex.
2563 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2564 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2565 "pi-state %p\n", ret,
2566 q->pi_state->pi_mutex.owner,
2567 q->pi_state->owner);
2571 return ret ? ret : locked;
2575 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2576 * @hb: the futex hash bucket, must be locked by the caller
2577 * @q: the futex_q to queue up on
2578 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2580 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2581 struct hrtimer_sleeper *timeout)
2584 * The task state is guaranteed to be set before another task can
2585 * wake it. set_current_state() is implemented using smp_store_mb() and
2586 * queue_me() calls spin_unlock() upon completion, both serializing
2587 * access to the hash list and forcing another memory barrier.
2589 set_current_state(TASK_INTERRUPTIBLE);
2594 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2597 * If we have been removed from the hash list, then another task
2598 * has tried to wake us, and we can skip the call to schedule().
2600 if (likely(!plist_node_empty(&q->list))) {
2602 * If the timer has already expired, current will already be
2603 * flagged for rescheduling. Only call schedule if there
2604 * is no timeout, or if it has yet to expire.
2606 if (!timeout || timeout->task)
2607 freezable_schedule();
2609 __set_current_state(TASK_RUNNING);
2613 * futex_wait_setup() - Prepare to wait on a futex
2614 * @uaddr: the futex userspace address
2615 * @val: the expected value
2616 * @flags: futex flags (FLAGS_SHARED, etc.)
2617 * @q: the associated futex_q
2618 * @hb: storage for hash_bucket pointer to be returned to caller
2620 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2621 * compare it with the expected value. Handle atomic faults internally.
2622 * Return with the hb lock held and a q.key reference on success, and unlocked
2623 * with no q.key reference on failure.
2626 * - 0 - uaddr contains val and hb has been locked;
2627 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2629 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2630 struct futex_q *q, struct futex_hash_bucket **hb)
2636 * Access the page AFTER the hash-bucket is locked.
2637 * Order is important:
2639 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2640 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2642 * The basic logical guarantee of a futex is that it blocks ONLY
2643 * if cond(var) is known to be true at the time of blocking, for
2644 * any cond. If we locked the hash-bucket after testing *uaddr, that
2645 * would open a race condition where we could block indefinitely with
2646 * cond(var) false, which would violate the guarantee.
2648 * On the other hand, we insert q and release the hash-bucket only
2649 * after testing *uaddr. This guarantees that futex_wait() will NOT
2650 * absorb a wakeup if *uaddr does not match the desired values
2651 * while the syscall executes.
2654 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2655 if (unlikely(ret != 0))
2659 *hb = queue_lock(q);
2661 ret = get_futex_value_locked(&uval, uaddr);
2666 ret = get_user(uval, uaddr);
2670 if (!(flags & FLAGS_SHARED))
2673 put_futex_key(&q->key);
2684 put_futex_key(&q->key);
2688 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2689 ktime_t *abs_time, u32 bitset)
2691 struct hrtimer_sleeper timeout, *to = NULL;
2692 struct restart_block *restart;
2693 struct futex_hash_bucket *hb;
2694 struct futex_q q = futex_q_init;
2704 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2705 CLOCK_REALTIME : CLOCK_MONOTONIC,
2707 hrtimer_init_sleeper(to, current);
2708 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2709 current->timer_slack_ns);
2714 * Prepare to wait on uaddr. On success, holds hb lock and increments
2717 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2721 /* queue_me and wait for wakeup, timeout, or a signal. */
2722 futex_wait_queue_me(hb, &q, to);
2724 /* If we were woken (and unqueued), we succeeded, whatever. */
2726 /* unqueue_me() drops q.key ref */
2727 if (!unqueue_me(&q))
2730 if (to && !to->task)
2734 * We expect signal_pending(current), but we might be the
2735 * victim of a spurious wakeup as well.
2737 if (!signal_pending(current))
2744 restart = ¤t->restart_block;
2745 restart->fn = futex_wait_restart;
2746 restart->futex.uaddr = uaddr;
2747 restart->futex.val = val;
2748 restart->futex.time = *abs_time;
2749 restart->futex.bitset = bitset;
2750 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2752 ret = -ERESTART_RESTARTBLOCK;
2756 hrtimer_cancel(&to->timer);
2757 destroy_hrtimer_on_stack(&to->timer);
2763 static long futex_wait_restart(struct restart_block *restart)
2765 u32 __user *uaddr = restart->futex.uaddr;
2766 ktime_t t, *tp = NULL;
2768 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2769 t = restart->futex.time;
2772 restart->fn = do_no_restart_syscall;
2774 return (long)futex_wait(uaddr, restart->futex.flags,
2775 restart->futex.val, tp, restart->futex.bitset);
2780 * Userspace tried a 0 -> TID atomic transition of the futex value
2781 * and failed. The kernel side here does the whole locking operation:
2782 * if there are waiters then it will block as a consequence of relying
2783 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2784 * a 0 value of the futex too.).
2786 * Also serves as futex trylock_pi()'ing, and due semantics.
2788 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2789 ktime_t *time, int trylock)
2791 struct hrtimer_sleeper timeout, *to = NULL;
2792 struct futex_pi_state *pi_state = NULL;
2793 struct rt_mutex_waiter rt_waiter;
2794 struct futex_hash_bucket *hb;
2795 struct futex_q q = futex_q_init;
2798 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2801 if (refill_pi_state_cache())
2806 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2808 hrtimer_init_sleeper(to, current);
2809 hrtimer_set_expires(&to->timer, *time);
2813 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2814 if (unlikely(ret != 0))
2818 hb = queue_lock(&q);
2820 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2821 if (unlikely(ret)) {
2823 * Atomic work succeeded and we got the lock,
2824 * or failed. Either way, we do _not_ block.
2828 /* We got the lock. */
2830 goto out_unlock_put_key;
2835 * Two reasons for this:
2836 * - Task is exiting and we just wait for the
2838 * - The user space value changed.
2841 put_futex_key(&q.key);
2845 goto out_unlock_put_key;
2849 WARN_ON(!q.pi_state);
2852 * Only actually queue now that the atomic ops are done:
2857 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2858 /* Fixup the trylock return value: */
2859 ret = ret ? 0 : -EWOULDBLOCK;
2863 rt_mutex_init_waiter(&rt_waiter);
2866 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2867 * hold it while doing rt_mutex_start_proxy(), because then it will
2868 * include hb->lock in the blocking chain, even through we'll not in
2869 * fact hold it while blocking. This will lead it to report -EDEADLK
2870 * and BUG when futex_unlock_pi() interleaves with this.
2872 * Therefore acquire wait_lock while holding hb->lock, but drop the
2873 * latter before calling __rt_mutex_start_proxy_lock(). This
2874 * interleaves with futex_unlock_pi() -- which does a similar lock
2875 * handoff -- such that the latter can observe the futex_q::pi_state
2876 * before __rt_mutex_start_proxy_lock() is done.
2878 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2879 spin_unlock(q.lock_ptr);
2881 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2882 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2883 * it sees the futex_q::pi_state.
2885 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2886 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2895 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2897 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2900 spin_lock(q.lock_ptr);
2902 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2903 * first acquire the hb->lock before removing the lock from the
2904 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2907 * In particular; it is important that futex_unlock_pi() can not
2908 * observe this inconsistency.
2910 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2915 * Fixup the pi_state owner and possibly acquire the lock if we
2918 res = fixup_owner(uaddr, &q, !ret);
2920 * If fixup_owner() returned an error, proprogate that. If it acquired
2921 * the lock, clear our -ETIMEDOUT or -EINTR.
2924 ret = (res < 0) ? res : 0;
2927 * If fixup_owner() faulted and was unable to handle the fault, unlock
2928 * it and return the fault to userspace.
2930 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2931 pi_state = q.pi_state;
2932 get_pi_state(pi_state);
2935 /* Unqueue and drop the lock */
2939 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2940 put_pi_state(pi_state);
2949 put_futex_key(&q.key);
2952 hrtimer_cancel(&to->timer);
2953 destroy_hrtimer_on_stack(&to->timer);
2955 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2960 ret = fault_in_user_writeable(uaddr);
2964 if (!(flags & FLAGS_SHARED))
2967 put_futex_key(&q.key);
2972 * Userspace attempted a TID -> 0 atomic transition, and failed.
2973 * This is the in-kernel slowpath: we look up the PI state (if any),
2974 * and do the rt-mutex unlock.
2976 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2978 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2979 union futex_key key = FUTEX_KEY_INIT;
2980 struct futex_hash_bucket *hb;
2981 struct futex_q *top_waiter;
2984 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2988 if (get_user(uval, uaddr))
2991 * We release only a lock we actually own:
2993 if ((uval & FUTEX_TID_MASK) != vpid)
2996 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
3000 hb = hash_futex(&key);
3001 spin_lock(&hb->lock);
3004 * Check waiters first. We do not trust user space values at
3005 * all and we at least want to know if user space fiddled
3006 * with the futex value instead of blindly unlocking.
3008 top_waiter = futex_top_waiter(hb, &key);
3010 struct futex_pi_state *pi_state = top_waiter->pi_state;
3017 * If current does not own the pi_state then the futex is
3018 * inconsistent and user space fiddled with the futex value.
3020 if (pi_state->owner != current)
3023 get_pi_state(pi_state);
3025 * By taking wait_lock while still holding hb->lock, we ensure
3026 * there is no point where we hold neither; and therefore
3027 * wake_futex_pi() must observe a state consistent with what we
3030 * In particular; this forces __rt_mutex_start_proxy() to
3031 * complete such that we're guaranteed to observe the
3032 * rt_waiter. Also see the WARN in wake_futex_pi().
3034 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3035 spin_unlock(&hb->lock);
3037 /* drops pi_state->pi_mutex.wait_lock */
3038 ret = wake_futex_pi(uaddr, uval, pi_state);
3040 put_pi_state(pi_state);
3043 * Success, we're done! No tricky corner cases.
3048 * The atomic access to the futex value generated a
3049 * pagefault, so retry the user-access and the wakeup:
3054 * A unconditional UNLOCK_PI op raced against a waiter
3055 * setting the FUTEX_WAITERS bit. Try again.
3060 * wake_futex_pi has detected invalid state. Tell user
3067 * We have no kernel internal state, i.e. no waiters in the
3068 * kernel. Waiters which are about to queue themselves are stuck
3069 * on hb->lock. So we can safely ignore them. We do neither
3070 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3073 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3074 spin_unlock(&hb->lock);
3089 * If uval has changed, let user space handle it.
3091 ret = (curval == uval) ? 0 : -EAGAIN;
3094 spin_unlock(&hb->lock);
3096 put_futex_key(&key);
3100 put_futex_key(&key);
3105 put_futex_key(&key);
3107 ret = fault_in_user_writeable(uaddr);
3115 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3116 * @hb: the hash_bucket futex_q was original enqueued on
3117 * @q: the futex_q woken while waiting to be requeued
3118 * @key2: the futex_key of the requeue target futex
3119 * @timeout: the timeout associated with the wait (NULL if none)
3121 * Detect if the task was woken on the initial futex as opposed to the requeue
3122 * target futex. If so, determine if it was a timeout or a signal that caused
3123 * the wakeup and return the appropriate error code to the caller. Must be
3124 * called with the hb lock held.
3127 * - 0 = no early wakeup detected;
3128 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3131 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3132 struct futex_q *q, union futex_key *key2,
3133 struct hrtimer_sleeper *timeout)
3138 * With the hb lock held, we avoid races while we process the wakeup.
3139 * We only need to hold hb (and not hb2) to ensure atomicity as the
3140 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3141 * It can't be requeued from uaddr2 to something else since we don't
3142 * support a PI aware source futex for requeue.
3144 if (!match_futex(&q->key, key2)) {
3145 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3147 * We were woken prior to requeue by a timeout or a signal.
3148 * Unqueue the futex_q and determine which it was.
3150 plist_del(&q->list, &hb->chain);
3153 /* Handle spurious wakeups gracefully */
3155 if (timeout && !timeout->task)
3157 else if (signal_pending(current))
3158 ret = -ERESTARTNOINTR;
3164 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3165 * @uaddr: the futex we initially wait on (non-pi)
3166 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3167 * the same type, no requeueing from private to shared, etc.
3168 * @val: the expected value of uaddr
3169 * @abs_time: absolute timeout
3170 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3171 * @uaddr2: the pi futex we will take prior to returning to user-space
3173 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3174 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3175 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3176 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3177 * without one, the pi logic would not know which task to boost/deboost, if
3178 * there was a need to.
3180 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3181 * via the following--
3182 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3183 * 2) wakeup on uaddr2 after a requeue
3187 * If 3, cleanup and return -ERESTARTNOINTR.
3189 * If 2, we may then block on trying to take the rt_mutex and return via:
3190 * 5) successful lock
3193 * 8) other lock acquisition failure
3195 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3197 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3203 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3204 u32 val, ktime_t *abs_time, u32 bitset,
3207 struct hrtimer_sleeper timeout, *to = NULL;
3208 struct futex_pi_state *pi_state = NULL;
3209 struct rt_mutex_waiter rt_waiter;
3210 struct futex_hash_bucket *hb;
3211 union futex_key key2 = FUTEX_KEY_INIT;
3212 struct futex_q q = futex_q_init;
3215 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3218 if (uaddr == uaddr2)
3226 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3227 CLOCK_REALTIME : CLOCK_MONOTONIC,
3229 hrtimer_init_sleeper(to, current);
3230 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3231 current->timer_slack_ns);
3235 * The waiter is allocated on our stack, manipulated by the requeue
3236 * code while we sleep on uaddr.
3238 rt_mutex_init_waiter(&rt_waiter);
3240 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3241 if (unlikely(ret != 0))
3245 q.rt_waiter = &rt_waiter;
3246 q.requeue_pi_key = &key2;
3249 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3252 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3257 * The check above which compares uaddrs is not sufficient for
3258 * shared futexes. We need to compare the keys:
3260 if (match_futex(&q.key, &key2)) {
3266 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3267 futex_wait_queue_me(hb, &q, to);
3269 spin_lock(&hb->lock);
3270 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3271 spin_unlock(&hb->lock);
3276 * In order for us to be here, we know our q.key == key2, and since
3277 * we took the hb->lock above, we also know that futex_requeue() has
3278 * completed and we no longer have to concern ourselves with a wakeup
3279 * race with the atomic proxy lock acquisition by the requeue code. The
3280 * futex_requeue dropped our key1 reference and incremented our key2
3284 /* Check if the requeue code acquired the second futex for us. */
3287 * Got the lock. We might not be the anticipated owner if we
3288 * did a lock-steal - fix up the PI-state in that case.
3290 if (q.pi_state && (q.pi_state->owner != current)) {
3291 spin_lock(q.lock_ptr);
3292 ret = fixup_pi_state_owner(uaddr2, &q, current);
3293 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3294 pi_state = q.pi_state;
3295 get_pi_state(pi_state);
3298 * Drop the reference to the pi state which
3299 * the requeue_pi() code acquired for us.
3301 put_pi_state(q.pi_state);
3302 spin_unlock(q.lock_ptr);
3305 struct rt_mutex *pi_mutex;
3308 * We have been woken up by futex_unlock_pi(), a timeout, or a
3309 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3312 WARN_ON(!q.pi_state);
3313 pi_mutex = &q.pi_state->pi_mutex;
3314 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3316 spin_lock(q.lock_ptr);
3317 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3320 debug_rt_mutex_free_waiter(&rt_waiter);
3322 * Fixup the pi_state owner and possibly acquire the lock if we
3325 res = fixup_owner(uaddr2, &q, !ret);
3327 * If fixup_owner() returned an error, proprogate that. If it
3328 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3331 ret = (res < 0) ? res : 0;
3334 * If fixup_pi_state_owner() faulted and was unable to handle
3335 * the fault, unlock the rt_mutex and return the fault to
3338 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3339 pi_state = q.pi_state;
3340 get_pi_state(pi_state);
3343 /* Unqueue and drop the lock. */
3348 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3349 put_pi_state(pi_state);
3352 if (ret == -EINTR) {
3354 * We've already been requeued, but cannot restart by calling
3355 * futex_lock_pi() directly. We could restart this syscall, but
3356 * it would detect that the user space "val" changed and return
3357 * -EWOULDBLOCK. Save the overhead of the restart and return
3358 * -EWOULDBLOCK directly.
3364 put_futex_key(&q.key);
3366 put_futex_key(&key2);
3370 hrtimer_cancel(&to->timer);
3371 destroy_hrtimer_on_stack(&to->timer);
3377 * Support for robust futexes: the kernel cleans up held futexes at
3380 * Implementation: user-space maintains a per-thread list of locks it
3381 * is holding. Upon do_exit(), the kernel carefully walks this list,
3382 * and marks all locks that are owned by this thread with the
3383 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3384 * always manipulated with the lock held, so the list is private and
3385 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3386 * field, to allow the kernel to clean up if the thread dies after
3387 * acquiring the lock, but just before it could have added itself to
3388 * the list. There can only be one such pending lock.
3392 * sys_set_robust_list() - Set the robust-futex list head of a task
3393 * @head: pointer to the list-head
3394 * @len: length of the list-head, as userspace expects
3396 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3399 if (!futex_cmpxchg_enabled)
3402 * The kernel knows only one size for now:
3404 if (unlikely(len != sizeof(*head)))
3407 current->robust_list = head;
3413 * sys_get_robust_list() - Get the robust-futex list head of a task
3414 * @pid: pid of the process [zero for current task]
3415 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3416 * @len_ptr: pointer to a length field, the kernel fills in the header size
3418 SYSCALL_DEFINE3(get_robust_list, int, pid,
3419 struct robust_list_head __user * __user *, head_ptr,
3420 size_t __user *, len_ptr)
3422 struct robust_list_head __user *head;
3424 struct task_struct *p;
3426 if (!futex_cmpxchg_enabled)
3435 p = find_task_by_vpid(pid);
3441 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3444 head = p->robust_list;
3447 if (put_user(sizeof(*head), len_ptr))
3449 return put_user(head, head_ptr);
3458 * Process a futex-list entry, check whether it's owned by the
3459 * dying task, and do notification if so:
3461 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3463 u32 uval, uninitialized_var(nval), mval;
3466 /* Futex address must be 32bit aligned */
3467 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3471 if (get_user(uval, uaddr))
3474 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3478 * Ok, this dying thread is truly holding a futex
3479 * of interest. Set the OWNER_DIED bit atomically
3480 * via cmpxchg, and if the value had FUTEX_WAITERS
3481 * set, wake up a waiter (if any). (We have to do a
3482 * futex_wake() even if OWNER_DIED is already set -
3483 * to handle the rare but possible case of recursive
3484 * thread-death.) The rest of the cleanup is done in
3487 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3490 * We are not holding a lock here, but we want to have
3491 * the pagefault_disable/enable() protection because
3492 * we want to handle the fault gracefully. If the
3493 * access fails we try to fault in the futex with R/W
3494 * verification via get_user_pages. get_user() above
3495 * does not guarantee R/W access. If that fails we
3496 * give up and leave the futex locked.
3498 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3501 if (fault_in_user_writeable(uaddr))
3519 * Wake robust non-PI futexes here. The wakeup of
3520 * PI futexes happens in exit_pi_state():
3522 if (!pi && (uval & FUTEX_WAITERS))
3523 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3529 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3531 static inline int fetch_robust_entry(struct robust_list __user **entry,
3532 struct robust_list __user * __user *head,
3535 unsigned long uentry;
3537 if (get_user(uentry, (unsigned long __user *)head))
3540 *entry = (void __user *)(uentry & ~1UL);
3547 * Walk curr->robust_list (very carefully, it's a userspace list!)
3548 * and mark any locks found there dead, and notify any waiters.
3550 * We silently return on any sign of list-walking problem.
3552 void exit_robust_list(struct task_struct *curr)
3554 struct robust_list_head __user *head = curr->robust_list;
3555 struct robust_list __user *entry, *next_entry, *pending;
3556 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3557 unsigned int uninitialized_var(next_pi);
3558 unsigned long futex_offset;
3561 if (!futex_cmpxchg_enabled)
3565 * Fetch the list head (which was registered earlier, via
3566 * sys_set_robust_list()):
3568 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3571 * Fetch the relative futex offset:
3573 if (get_user(futex_offset, &head->futex_offset))
3576 * Fetch any possibly pending lock-add first, and handle it
3579 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3582 next_entry = NULL; /* avoid warning with gcc */
3583 while (entry != &head->list) {
3585 * Fetch the next entry in the list before calling
3586 * handle_futex_death:
3588 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3590 * A pending lock might already be on the list, so
3591 * don't process it twice:
3593 if (entry != pending)
3594 if (handle_futex_death((void __user *)entry + futex_offset,
3602 * Avoid excessively long or circular lists:
3611 handle_futex_death((void __user *)pending + futex_offset,
3615 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3616 u32 __user *uaddr2, u32 val2, u32 val3)
3618 int cmd = op & FUTEX_CMD_MASK;
3619 unsigned int flags = 0;
3621 if (!(op & FUTEX_PRIVATE_FLAG))
3622 flags |= FLAGS_SHARED;
3624 if (op & FUTEX_CLOCK_REALTIME) {
3625 flags |= FLAGS_CLOCKRT;
3626 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3627 cmd != FUTEX_WAIT_REQUEUE_PI)
3633 case FUTEX_UNLOCK_PI:
3634 case FUTEX_TRYLOCK_PI:
3635 case FUTEX_WAIT_REQUEUE_PI:
3636 case FUTEX_CMP_REQUEUE_PI:
3637 if (!futex_cmpxchg_enabled)
3643 val3 = FUTEX_BITSET_MATCH_ANY;
3645 case FUTEX_WAIT_BITSET:
3646 return futex_wait(uaddr, flags, val, timeout, val3);
3648 val3 = FUTEX_BITSET_MATCH_ANY;
3650 case FUTEX_WAKE_BITSET:
3651 return futex_wake(uaddr, flags, val, val3);
3653 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3654 case FUTEX_CMP_REQUEUE:
3655 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3657 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3659 return futex_lock_pi(uaddr, flags, timeout, 0);
3660 case FUTEX_UNLOCK_PI:
3661 return futex_unlock_pi(uaddr, flags);
3662 case FUTEX_TRYLOCK_PI:
3663 return futex_lock_pi(uaddr, flags, NULL, 1);
3664 case FUTEX_WAIT_REQUEUE_PI:
3665 val3 = FUTEX_BITSET_MATCH_ANY;
3666 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3668 case FUTEX_CMP_REQUEUE_PI:
3669 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3675 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3676 struct timespec __user *, utime, u32 __user *, uaddr2,
3680 ktime_t t, *tp = NULL;
3682 int cmd = op & FUTEX_CMD_MASK;
3684 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3685 cmd == FUTEX_WAIT_BITSET ||
3686 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3687 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3689 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3691 if (!timespec_valid(&ts))
3694 t = timespec_to_ktime(ts);
3695 if (cmd == FUTEX_WAIT)
3696 t = ktime_add_safe(ktime_get(), t);
3700 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3701 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3703 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3704 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3705 val2 = (u32) (unsigned long) utime;
3707 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3710 static void __init futex_detect_cmpxchg(void)
3712 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3716 * This will fail and we want it. Some arch implementations do
3717 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3718 * functionality. We want to know that before we call in any
3719 * of the complex code paths. Also we want to prevent
3720 * registration of robust lists in that case. NULL is
3721 * guaranteed to fault and we get -EFAULT on functional
3722 * implementation, the non-functional ones will return
3725 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3726 futex_cmpxchg_enabled = 1;
3730 static int __init futex_init(void)
3732 unsigned int futex_shift;
3735 #if CONFIG_BASE_SMALL
3736 futex_hashsize = 16;
3738 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3741 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3743 futex_hashsize < 256 ? HASH_SMALL : 0,
3745 futex_hashsize, futex_hashsize);
3746 futex_hashsize = 1UL << futex_shift;
3748 futex_detect_cmpxchg();
3750 for (i = 0; i < futex_hashsize; i++) {
3751 atomic_set(&futex_queues[i].waiters, 0);
3752 plist_head_init(&futex_queues[i].chain);
3753 spin_lock_init(&futex_queues[i].lock);
3758 core_initcall(futex_init);