2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/compat.h>
48 #include <linux/slab.h>
49 #include <linux/poll.h>
51 #include <linux/file.h>
52 #include <linux/jhash.h>
53 #include <linux/init.h>
54 #include <linux/futex.h>
55 #include <linux/mount.h>
56 #include <linux/pagemap.h>
57 #include <linux/syscalls.h>
58 #include <linux/signal.h>
59 #include <linux/export.h>
60 #include <linux/magic.h>
61 #include <linux/pid.h>
62 #include <linux/nsproxy.h>
63 #include <linux/ptrace.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/wake_q.h>
66 #include <linux/sched/mm.h>
67 #include <linux/hugetlb.h>
68 #include <linux/freezer.h>
69 #include <linux/bootmem.h>
70 #include <linux/fault-inject.h>
72 #include <asm/futex.h>
74 #include "locking/rtmutex_common.h"
77 * READ this before attempting to hack on futexes!
79 * Basic futex operation and ordering guarantees
80 * =============================================
82 * The waiter reads the futex value in user space and calls
83 * futex_wait(). This function computes the hash bucket and acquires
84 * the hash bucket lock. After that it reads the futex user space value
85 * again and verifies that the data has not changed. If it has not changed
86 * it enqueues itself into the hash bucket, releases the hash bucket lock
89 * The waker side modifies the user space value of the futex and calls
90 * futex_wake(). This function computes the hash bucket and acquires the
91 * hash bucket lock. Then it looks for waiters on that futex in the hash
92 * bucket and wakes them.
94 * In futex wake up scenarios where no tasks are blocked on a futex, taking
95 * the hb spinlock can be avoided and simply return. In order for this
96 * optimization to work, ordering guarantees must exist so that the waiter
97 * being added to the list is acknowledged when the list is concurrently being
98 * checked by the waker, avoiding scenarios like the following:
102 * sys_futex(WAIT, futex, val);
103 * futex_wait(futex, val);
106 * sys_futex(WAKE, futex);
111 * lock(hash_bucket(futex));
113 * unlock(hash_bucket(futex));
116 * This would cause the waiter on CPU 0 to wait forever because it
117 * missed the transition of the user space value from val to newval
118 * and the waker did not find the waiter in the hash bucket queue.
120 * The correct serialization ensures that a waiter either observes
121 * the changed user space value before blocking or is woken by a
126 * sys_futex(WAIT, futex, val);
127 * futex_wait(futex, val);
130 * smp_mb(); (A) <-- paired with -.
132 * lock(hash_bucket(futex)); |
136 * | sys_futex(WAKE, futex);
137 * | futex_wake(futex);
139 * `--------> smp_mb(); (B)
142 * unlock(hash_bucket(futex));
143 * schedule(); if (waiters)
144 * lock(hash_bucket(futex));
145 * else wake_waiters(futex);
146 * waiters--; (b) unlock(hash_bucket(futex));
148 * Where (A) orders the waiters increment and the futex value read through
149 * atomic operations (see hb_waiters_inc) and where (B) orders the write
150 * to futex and the waiters read -- this is done by the barriers for both
151 * shared and private futexes in get_futex_key_refs().
153 * This yields the following case (where X:=waiters, Y:=futex):
161 * Which guarantees that x==0 && y==0 is impossible; which translates back into
162 * the guarantee that we cannot both miss the futex variable change and the
165 * Note that a new waiter is accounted for in (a) even when it is possible that
166 * the wait call can return error, in which case we backtrack from it in (b).
167 * Refer to the comment in queue_lock().
169 * Similarly, in order to account for waiters being requeued on another
170 * address we always increment the waiters for the destination bucket before
171 * acquiring the lock. It then decrements them again after releasing it -
172 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
173 * will do the additional required waiter count housekeeping. This is done for
174 * double_lock_hb() and double_unlock_hb(), respectively.
177 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
178 #define futex_cmpxchg_enabled 1
180 static int __read_mostly futex_cmpxchg_enabled;
184 * Futex flags used to encode options to functions and preserve them across
188 # define FLAGS_SHARED 0x01
191 * NOMMU does not have per process address space. Let the compiler optimize
194 # define FLAGS_SHARED 0x00
196 #define FLAGS_CLOCKRT 0x02
197 #define FLAGS_HAS_TIMEOUT 0x04
200 * Priority Inheritance state:
202 struct futex_pi_state {
204 * list of 'owned' pi_state instances - these have to be
205 * cleaned up in do_exit() if the task exits prematurely:
207 struct list_head list;
212 struct rt_mutex pi_mutex;
214 struct task_struct *owner;
218 } __randomize_layout;
221 * struct futex_q - The hashed futex queue entry, one per waiting task
222 * @list: priority-sorted list of tasks waiting on this futex
223 * @task: the task waiting on the futex
224 * @lock_ptr: the hash bucket lock
225 * @key: the key the futex is hashed on
226 * @pi_state: optional priority inheritance state
227 * @rt_waiter: rt_waiter storage for use with requeue_pi
228 * @requeue_pi_key: the requeue_pi target futex key
229 * @bitset: bitset for the optional bitmasked wakeup
231 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
232 * we can wake only the relevant ones (hashed queues may be shared).
234 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
235 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
236 * The order of wakeup is always to make the first condition true, then
239 * PI futexes are typically woken before they are removed from the hash list via
240 * the rt_mutex code. See unqueue_me_pi().
243 struct plist_node list;
245 struct task_struct *task;
246 spinlock_t *lock_ptr;
248 struct futex_pi_state *pi_state;
249 struct rt_mutex_waiter *rt_waiter;
250 union futex_key *requeue_pi_key;
252 } __randomize_layout;
254 static const struct futex_q futex_q_init = {
255 /* list gets initialized in queue_me()*/
256 .key = FUTEX_KEY_INIT,
257 .bitset = FUTEX_BITSET_MATCH_ANY
261 * Hash buckets are shared by all the futex_keys that hash to the same
262 * location. Each key may have multiple futex_q structures, one for each task
263 * waiting on a futex.
265 struct futex_hash_bucket {
268 struct plist_head chain;
269 } ____cacheline_aligned_in_smp;
272 * The base of the bucket array and its size are always used together
273 * (after initialization only in hash_futex()), so ensure that they
274 * reside in the same cacheline.
277 struct futex_hash_bucket *queues;
278 unsigned long hashsize;
279 } __futex_data __read_mostly __aligned(2*sizeof(long));
280 #define futex_queues (__futex_data.queues)
281 #define futex_hashsize (__futex_data.hashsize)
285 * Fault injections for futexes.
287 #ifdef CONFIG_FAIL_FUTEX
290 struct fault_attr attr;
294 .attr = FAULT_ATTR_INITIALIZER,
295 .ignore_private = false,
298 static int __init setup_fail_futex(char *str)
300 return setup_fault_attr(&fail_futex.attr, str);
302 __setup("fail_futex=", setup_fail_futex);
304 static bool should_fail_futex(bool fshared)
306 if (fail_futex.ignore_private && !fshared)
309 return should_fail(&fail_futex.attr, 1);
312 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
314 static int __init fail_futex_debugfs(void)
316 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
319 dir = fault_create_debugfs_attr("fail_futex", NULL,
324 if (!debugfs_create_bool("ignore-private", mode, dir,
325 &fail_futex.ignore_private)) {
326 debugfs_remove_recursive(dir);
333 late_initcall(fail_futex_debugfs);
335 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
338 static inline bool should_fail_futex(bool fshared)
342 #endif /* CONFIG_FAIL_FUTEX */
344 static inline void futex_get_mm(union futex_key *key)
346 mmgrab(key->private.mm);
348 * Ensure futex_get_mm() implies a full barrier such that
349 * get_futex_key() implies a full barrier. This is relied upon
350 * as smp_mb(); (B), see the ordering comment above.
352 smp_mb__after_atomic();
356 * Reflects a new waiter being added to the waitqueue.
358 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
361 atomic_inc(&hb->waiters);
363 * Full barrier (A), see the ordering comment above.
365 smp_mb__after_atomic();
370 * Reflects a waiter being removed from the waitqueue by wakeup
373 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
376 atomic_dec(&hb->waiters);
380 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
383 return atomic_read(&hb->waiters);
390 * hash_futex - Return the hash bucket in the global hash
391 * @key: Pointer to the futex key for which the hash is calculated
393 * We hash on the keys returned from get_futex_key (see below) and return the
394 * corresponding hash bucket in the global hash.
396 static struct futex_hash_bucket *hash_futex(union futex_key *key)
398 u32 hash = jhash2((u32*)&key->both.word,
399 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
401 return &futex_queues[hash & (futex_hashsize - 1)];
406 * match_futex - Check whether two futex keys are equal
407 * @key1: Pointer to key1
408 * @key2: Pointer to key2
410 * Return 1 if two futex_keys are equal, 0 otherwise.
412 static inline int match_futex(union futex_key *key1, union futex_key *key2)
415 && key1->both.word == key2->both.word
416 && key1->both.ptr == key2->both.ptr
417 && key1->both.offset == key2->both.offset);
421 * Take a reference to the resource addressed by a key.
422 * Can be called while holding spinlocks.
425 static void get_futex_key_refs(union futex_key *key)
431 * On MMU less systems futexes are always "private" as there is no per
432 * process address space. We need the smp wmb nevertheless - yes,
433 * arch/blackfin has MMU less SMP ...
435 if (!IS_ENABLED(CONFIG_MMU)) {
436 smp_mb(); /* explicit smp_mb(); (B) */
440 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
442 ihold(key->shared.inode); /* implies smp_mb(); (B) */
444 case FUT_OFF_MMSHARED:
445 futex_get_mm(key); /* implies smp_mb(); (B) */
449 * Private futexes do not hold reference on an inode or
450 * mm, therefore the only purpose of calling get_futex_key_refs
451 * is because we need the barrier for the lockless waiter check.
453 smp_mb(); /* explicit smp_mb(); (B) */
458 * Drop a reference to the resource addressed by a key.
459 * The hash bucket spinlock must not be held. This is
460 * a no-op for private futexes, see comment in the get
463 static void drop_futex_key_refs(union futex_key *key)
465 if (!key->both.ptr) {
466 /* If we're here then we tried to put a key we failed to get */
471 if (!IS_ENABLED(CONFIG_MMU))
474 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
476 iput(key->shared.inode);
478 case FUT_OFF_MMSHARED:
479 mmdrop(key->private.mm);
485 * get_futex_key() - Get parameters which are the keys for a futex
486 * @uaddr: virtual address of the futex
487 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
488 * @key: address where result is stored.
489 * @rw: mapping needs to be read/write (values: VERIFY_READ,
492 * Return: a negative error code or 0
494 * The key words are stored in @key on success.
496 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
497 * offset_within_page). For private mappings, it's (uaddr, current->mm).
498 * We can usually work out the index without swapping in the page.
500 * lock_page() might sleep, the caller should not hold a spinlock.
503 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
505 unsigned long address = (unsigned long)uaddr;
506 struct mm_struct *mm = current->mm;
507 struct page *page, *tail;
508 struct address_space *mapping;
512 * The futex address must be "naturally" aligned.
514 key->both.offset = address % PAGE_SIZE;
515 if (unlikely((address % sizeof(u32)) != 0))
517 address -= key->both.offset;
519 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
522 if (unlikely(should_fail_futex(fshared)))
526 * PROCESS_PRIVATE futexes are fast.
527 * As the mm cannot disappear under us and the 'key' only needs
528 * virtual address, we dont even have to find the underlying vma.
529 * Note : We do have to check 'uaddr' is a valid user address,
530 * but access_ok() should be faster than find_vma()
533 key->private.mm = mm;
534 key->private.address = address;
535 get_futex_key_refs(key); /* implies smp_mb(); (B) */
540 /* Ignore any VERIFY_READ mapping (futex common case) */
541 if (unlikely(should_fail_futex(fshared)))
544 err = get_user_pages_fast(address, 1, 1, &page);
546 * If write access is not required (eg. FUTEX_WAIT), try
547 * and get read-only access.
549 if (err == -EFAULT && rw == VERIFY_READ) {
550 err = get_user_pages_fast(address, 1, 0, &page);
559 * The treatment of mapping from this point on is critical. The page
560 * lock protects many things but in this context the page lock
561 * stabilizes mapping, prevents inode freeing in the shared
562 * file-backed region case and guards against movement to swap cache.
564 * Strictly speaking the page lock is not needed in all cases being
565 * considered here and page lock forces unnecessarily serialization
566 * From this point on, mapping will be re-verified if necessary and
567 * page lock will be acquired only if it is unavoidable
569 * Mapping checks require the head page for any compound page so the
570 * head page and mapping is looked up now. For anonymous pages, it
571 * does not matter if the page splits in the future as the key is
572 * based on the address. For filesystem-backed pages, the tail is
573 * required as the index of the page determines the key. For
574 * base pages, there is no tail page and tail == page.
577 page = compound_head(page);
578 mapping = READ_ONCE(page->mapping);
581 * If page->mapping is NULL, then it cannot be a PageAnon
582 * page; but it might be the ZERO_PAGE or in the gate area or
583 * in a special mapping (all cases which we are happy to fail);
584 * or it may have been a good file page when get_user_pages_fast
585 * found it, but truncated or holepunched or subjected to
586 * invalidate_complete_page2 before we got the page lock (also
587 * cases which we are happy to fail). And we hold a reference,
588 * so refcount care in invalidate_complete_page's remove_mapping
589 * prevents drop_caches from setting mapping to NULL beneath us.
591 * The case we do have to guard against is when memory pressure made
592 * shmem_writepage move it from filecache to swapcache beneath us:
593 * an unlikely race, but we do need to retry for page->mapping.
595 if (unlikely(!mapping)) {
599 * Page lock is required to identify which special case above
600 * applies. If this is really a shmem page then the page lock
601 * will prevent unexpected transitions.
604 shmem_swizzled = PageSwapCache(page) || page->mapping;
615 * Private mappings are handled in a simple way.
617 * If the futex key is stored on an anonymous page, then the associated
618 * object is the mm which is implicitly pinned by the calling process.
620 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
621 * it's a read-only handle, it's expected that futexes attach to
622 * the object not the particular process.
624 if (PageAnon(page)) {
626 * A RO anonymous page will never change and thus doesn't make
627 * sense for futex operations.
629 if (unlikely(should_fail_futex(fshared)) || ro) {
634 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
635 key->private.mm = mm;
636 key->private.address = address;
638 get_futex_key_refs(key); /* implies smp_mb(); (B) */
644 * The associated futex object in this case is the inode and
645 * the page->mapping must be traversed. Ordinarily this should
646 * be stabilised under page lock but it's not strictly
647 * necessary in this case as we just want to pin the inode, not
648 * update the radix tree or anything like that.
650 * The RCU read lock is taken as the inode is finally freed
651 * under RCU. If the mapping still matches expectations then the
652 * mapping->host can be safely accessed as being a valid inode.
656 if (READ_ONCE(page->mapping) != mapping) {
663 inode = READ_ONCE(mapping->host);
672 * Take a reference unless it is about to be freed. Previously
673 * this reference was taken by ihold under the page lock
674 * pinning the inode in place so i_lock was unnecessary. The
675 * only way for this check to fail is if the inode was
676 * truncated in parallel which is almost certainly an
677 * application bug. In such a case, just retry.
679 * We are not calling into get_futex_key_refs() in file-backed
680 * cases, therefore a successful atomic_inc return below will
681 * guarantee that get_futex_key() will still imply smp_mb(); (B).
683 if (!atomic_inc_not_zero(&inode->i_count)) {
690 /* Should be impossible but lets be paranoid for now */
691 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
699 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
700 key->shared.inode = inode;
701 key->shared.pgoff = basepage_index(tail);
710 static inline void put_futex_key(union futex_key *key)
712 drop_futex_key_refs(key);
716 * fault_in_user_writeable() - Fault in user address and verify RW access
717 * @uaddr: pointer to faulting user space address
719 * Slow path to fixup the fault we just took in the atomic write
722 * We have no generic implementation of a non-destructive write to the
723 * user address. We know that we faulted in the atomic pagefault
724 * disabled section so we can as well avoid the #PF overhead by
725 * calling get_user_pages() right away.
727 static int fault_in_user_writeable(u32 __user *uaddr)
729 struct mm_struct *mm = current->mm;
732 down_read(&mm->mmap_sem);
733 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
734 FAULT_FLAG_WRITE, NULL);
735 up_read(&mm->mmap_sem);
737 return ret < 0 ? ret : 0;
741 * futex_top_waiter() - Return the highest priority waiter on a futex
742 * @hb: the hash bucket the futex_q's reside in
743 * @key: the futex key (to distinguish it from other futex futex_q's)
745 * Must be called with the hb lock held.
747 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
748 union futex_key *key)
750 struct futex_q *this;
752 plist_for_each_entry(this, &hb->chain, list) {
753 if (match_futex(&this->key, key))
759 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
760 u32 uval, u32 newval)
765 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
771 static int get_futex_value_locked(u32 *dest, u32 __user *from)
776 ret = __get_user(*dest, from);
779 return ret ? -EFAULT : 0;
786 static int refill_pi_state_cache(void)
788 struct futex_pi_state *pi_state;
790 if (likely(current->pi_state_cache))
793 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
798 INIT_LIST_HEAD(&pi_state->list);
799 /* pi_mutex gets initialized later */
800 pi_state->owner = NULL;
801 atomic_set(&pi_state->refcount, 1);
802 pi_state->key = FUTEX_KEY_INIT;
804 current->pi_state_cache = pi_state;
809 static struct futex_pi_state *alloc_pi_state(void)
811 struct futex_pi_state *pi_state = current->pi_state_cache;
814 current->pi_state_cache = NULL;
819 static void get_pi_state(struct futex_pi_state *pi_state)
821 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
825 * Drops a reference to the pi_state object and frees or caches it
826 * when the last reference is gone.
828 static void put_pi_state(struct futex_pi_state *pi_state)
833 if (!atomic_dec_and_test(&pi_state->refcount))
837 * If pi_state->owner is NULL, the owner is most probably dying
838 * and has cleaned up the pi_state already
840 if (pi_state->owner) {
841 struct task_struct *owner;
843 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
844 owner = pi_state->owner;
846 raw_spin_lock(&owner->pi_lock);
847 list_del_init(&pi_state->list);
848 raw_spin_unlock(&owner->pi_lock);
850 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
851 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
854 if (current->pi_state_cache) {
858 * pi_state->list is already empty.
859 * clear pi_state->owner.
860 * refcount is at 0 - put it back to 1.
862 pi_state->owner = NULL;
863 atomic_set(&pi_state->refcount, 1);
864 current->pi_state_cache = pi_state;
868 #ifdef CONFIG_FUTEX_PI
871 * This task is holding PI mutexes at exit time => bad.
872 * Kernel cleans up PI-state, but userspace is likely hosed.
873 * (Robust-futex cleanup is separate and might save the day for userspace.)
875 void exit_pi_state_list(struct task_struct *curr)
877 struct list_head *next, *head = &curr->pi_state_list;
878 struct futex_pi_state *pi_state;
879 struct futex_hash_bucket *hb;
880 union futex_key key = FUTEX_KEY_INIT;
882 if (!futex_cmpxchg_enabled)
885 * We are a ZOMBIE and nobody can enqueue itself on
886 * pi_state_list anymore, but we have to be careful
887 * versus waiters unqueueing themselves:
889 raw_spin_lock_irq(&curr->pi_lock);
890 while (!list_empty(head)) {
892 pi_state = list_entry(next, struct futex_pi_state, list);
894 hb = hash_futex(&key);
897 * We can race against put_pi_state() removing itself from the
898 * list (a waiter going away). put_pi_state() will first
899 * decrement the reference count and then modify the list, so
900 * its possible to see the list entry but fail this reference
903 * In that case; drop the locks to let put_pi_state() make
904 * progress and retry the loop.
906 if (!atomic_inc_not_zero(&pi_state->refcount)) {
907 raw_spin_unlock_irq(&curr->pi_lock);
909 raw_spin_lock_irq(&curr->pi_lock);
912 raw_spin_unlock_irq(&curr->pi_lock);
914 spin_lock(&hb->lock);
915 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
916 raw_spin_lock(&curr->pi_lock);
918 * We dropped the pi-lock, so re-check whether this
919 * task still owns the PI-state:
921 if (head->next != next) {
922 /* retain curr->pi_lock for the loop invariant */
923 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
924 spin_unlock(&hb->lock);
925 put_pi_state(pi_state);
929 WARN_ON(pi_state->owner != curr);
930 WARN_ON(list_empty(&pi_state->list));
931 list_del_init(&pi_state->list);
932 pi_state->owner = NULL;
934 raw_spin_unlock(&curr->pi_lock);
935 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
936 spin_unlock(&hb->lock);
938 rt_mutex_futex_unlock(&pi_state->pi_mutex);
939 put_pi_state(pi_state);
941 raw_spin_lock_irq(&curr->pi_lock);
943 raw_spin_unlock_irq(&curr->pi_lock);
949 * We need to check the following states:
951 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
953 * [1] NULL | --- | --- | 0 | 0/1 | Valid
954 * [2] NULL | --- | --- | >0 | 0/1 | Valid
956 * [3] Found | NULL | -- | Any | 0/1 | Invalid
958 * [4] Found | Found | NULL | 0 | 1 | Valid
959 * [5] Found | Found | NULL | >0 | 1 | Invalid
961 * [6] Found | Found | task | 0 | 1 | Valid
963 * [7] Found | Found | NULL | Any | 0 | Invalid
965 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
966 * [9] Found | Found | task | 0 | 0 | Invalid
967 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
969 * [1] Indicates that the kernel can acquire the futex atomically. We
970 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
972 * [2] Valid, if TID does not belong to a kernel thread. If no matching
973 * thread is found then it indicates that the owner TID has died.
975 * [3] Invalid. The waiter is queued on a non PI futex
977 * [4] Valid state after exit_robust_list(), which sets the user space
978 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
980 * [5] The user space value got manipulated between exit_robust_list()
981 * and exit_pi_state_list()
983 * [6] Valid state after exit_pi_state_list() which sets the new owner in
984 * the pi_state but cannot access the user space value.
986 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
988 * [8] Owner and user space value match
990 * [9] There is no transient state which sets the user space TID to 0
991 * except exit_robust_list(), but this is indicated by the
992 * FUTEX_OWNER_DIED bit. See [4]
994 * [10] There is no transient state which leaves owner and user space
998 * Serialization and lifetime rules:
1002 * hb -> futex_q, relation
1003 * futex_q -> pi_state, relation
1005 * (cannot be raw because hb can contain arbitrary amount
1008 * pi_mutex->wait_lock:
1012 * (and pi_mutex 'obviously')
1016 * p->pi_state_list -> pi_state->list, relation
1018 * pi_state->refcount:
1026 * pi_mutex->wait_lock
1032 * Validate that the existing waiter has a pi_state and sanity check
1033 * the pi_state against the user space value. If correct, attach to
1036 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1037 struct futex_pi_state *pi_state,
1038 struct futex_pi_state **ps)
1040 pid_t pid = uval & FUTEX_TID_MASK;
1045 * Userspace might have messed up non-PI and PI futexes [3]
1047 if (unlikely(!pi_state))
1051 * We get here with hb->lock held, and having found a
1052 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1053 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1054 * which in turn means that futex_lock_pi() still has a reference on
1057 * The waiter holding a reference on @pi_state also protects against
1058 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1059 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1060 * free pi_state before we can take a reference ourselves.
1062 WARN_ON(!atomic_read(&pi_state->refcount));
1065 * Now that we have a pi_state, we can acquire wait_lock
1066 * and do the state validation.
1068 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1071 * Since {uval, pi_state} is serialized by wait_lock, and our current
1072 * uval was read without holding it, it can have changed. Verify it
1073 * still is what we expect it to be, otherwise retry the entire
1076 if (get_futex_value_locked(&uval2, uaddr))
1083 * Handle the owner died case:
1085 if (uval & FUTEX_OWNER_DIED) {
1087 * exit_pi_state_list sets owner to NULL and wakes the
1088 * topmost waiter. The task which acquires the
1089 * pi_state->rt_mutex will fixup owner.
1091 if (!pi_state->owner) {
1093 * No pi state owner, but the user space TID
1094 * is not 0. Inconsistent state. [5]
1099 * Take a ref on the state and return success. [4]
1105 * If TID is 0, then either the dying owner has not
1106 * yet executed exit_pi_state_list() or some waiter
1107 * acquired the rtmutex in the pi state, but did not
1108 * yet fixup the TID in user space.
1110 * Take a ref on the state and return success. [6]
1116 * If the owner died bit is not set, then the pi_state
1117 * must have an owner. [7]
1119 if (!pi_state->owner)
1124 * Bail out if user space manipulated the futex value. If pi
1125 * state exists then the owner TID must be the same as the
1126 * user space TID. [9/10]
1128 if (pid != task_pid_vnr(pi_state->owner))
1132 get_pi_state(pi_state);
1133 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1150 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1154 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1155 struct task_struct *tsk)
1160 * If PF_EXITPIDONE is not yet set, then try again.
1162 if (tsk && !(tsk->flags & PF_EXITPIDONE))
1166 * Reread the user space value to handle the following situation:
1170 * sys_exit() sys_futex()
1171 * do_exit() futex_lock_pi()
1172 * futex_lock_pi_atomic()
1173 * exit_signals(tsk) No waiters:
1174 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1175 * mm_release(tsk) Set waiter bit
1176 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1177 * Set owner died attach_to_pi_owner() {
1178 * *uaddr = 0xC0000000; tsk = get_task(PID);
1179 * } if (!tsk->flags & PF_EXITING) {
1181 * tsk->flags |= PF_EXITPIDONE; } else {
1182 * if (!(tsk->flags & PF_EXITPIDONE))
1184 * return -ESRCH; <--- FAIL
1187 * Returning ESRCH unconditionally is wrong here because the
1188 * user space value has been changed by the exiting task.
1190 * The same logic applies to the case where the exiting task is
1193 if (get_futex_value_locked(&uval2, uaddr))
1196 /* If the user space value has changed, try again. */
1201 * The exiting task did not have a robust list, the robust list was
1202 * corrupted or the user space value in *uaddr is simply bogus.
1203 * Give up and tell user space.
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1212 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213 struct futex_pi_state **ps)
1215 pid_t pid = uval & FUTEX_TID_MASK;
1216 struct futex_pi_state *pi_state;
1217 struct task_struct *p;
1220 * We are the first waiter - try to look up the real owner and attach
1221 * the new pi_state to it, but bail out when TID = 0 [1]
1223 * The !pid check is paranoid. None of the call sites should end up
1224 * with pid == 0, but better safe than sorry. Let the caller retry
1228 p = find_get_task_by_vpid(pid);
1230 return handle_exit_race(uaddr, uval, NULL);
1232 if (unlikely(p->flags & PF_KTHREAD)) {
1238 * We need to look at the task state flags to figure out,
1239 * whether the task is exiting. To protect against the do_exit
1240 * change of the task flags, we do this protected by
1243 raw_spin_lock_irq(&p->pi_lock);
1244 if (unlikely(p->flags & PF_EXITING)) {
1246 * The task is on the way out. When PF_EXITPIDONE is
1247 * set, we know that the task has finished the
1250 int ret = handle_exit_race(uaddr, uval, p);
1252 raw_spin_unlock_irq(&p->pi_lock);
1258 * No existing pi state. First waiter. [2]
1260 * This creates pi_state, we have hb->lock held, this means nothing can
1261 * observe this state, wait_lock is irrelevant.
1263 pi_state = alloc_pi_state();
1266 * Initialize the pi_mutex in locked state and make @p
1269 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1271 /* Store the key for possible exit cleanups: */
1272 pi_state->key = *key;
1274 WARN_ON(!list_empty(&pi_state->list));
1275 list_add(&pi_state->list, &p->pi_state_list);
1277 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1278 * because there is no concurrency as the object is not published yet.
1280 pi_state->owner = p;
1281 raw_spin_unlock_irq(&p->pi_lock);
1290 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1291 struct futex_hash_bucket *hb,
1292 union futex_key *key, struct futex_pi_state **ps)
1294 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1297 * If there is a waiter on that futex, validate it and
1298 * attach to the pi_state when the validation succeeds.
1301 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1304 * We are the first waiter - try to look up the owner based on
1305 * @uval and attach to it.
1307 return attach_to_pi_owner(uaddr, uval, key, ps);
1310 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1313 u32 uninitialized_var(curval);
1315 if (unlikely(should_fail_futex(true)))
1318 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1322 /* If user space value changed, let the caller retry */
1323 return curval != uval ? -EAGAIN : 0;
1327 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1328 * @uaddr: the pi futex user address
1329 * @hb: the pi futex hash bucket
1330 * @key: the futex key associated with uaddr and hb
1331 * @ps: the pi_state pointer where we store the result of the
1333 * @task: the task to perform the atomic lock work for. This will
1334 * be "current" except in the case of requeue pi.
1335 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1338 * - 0 - ready to wait;
1339 * - 1 - acquired the lock;
1342 * The hb->lock and futex_key refs shall be held by the caller.
1344 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1345 union futex_key *key,
1346 struct futex_pi_state **ps,
1347 struct task_struct *task, int set_waiters)
1349 u32 uval, newval, vpid = task_pid_vnr(task);
1350 struct futex_q *top_waiter;
1354 * Read the user space value first so we can validate a few
1355 * things before proceeding further.
1357 if (get_futex_value_locked(&uval, uaddr))
1360 if (unlikely(should_fail_futex(true)))
1366 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1369 if ((unlikely(should_fail_futex(true))))
1373 * Lookup existing state first. If it exists, try to attach to
1376 top_waiter = futex_top_waiter(hb, key);
1378 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1381 * No waiter and user TID is 0. We are here because the
1382 * waiters or the owner died bit is set or called from
1383 * requeue_cmp_pi or for whatever reason something took the
1386 if (!(uval & FUTEX_TID_MASK)) {
1388 * We take over the futex. No other waiters and the user space
1389 * TID is 0. We preserve the owner died bit.
1391 newval = uval & FUTEX_OWNER_DIED;
1394 /* The futex requeue_pi code can enforce the waiters bit */
1396 newval |= FUTEX_WAITERS;
1398 ret = lock_pi_update_atomic(uaddr, uval, newval);
1399 /* If the take over worked, return 1 */
1400 return ret < 0 ? ret : 1;
1404 * First waiter. Set the waiters bit before attaching ourself to
1405 * the owner. If owner tries to unlock, it will be forced into
1406 * the kernel and blocked on hb->lock.
1408 newval = uval | FUTEX_WAITERS;
1409 ret = lock_pi_update_atomic(uaddr, uval, newval);
1413 * If the update of the user space value succeeded, we try to
1414 * attach to the owner. If that fails, no harm done, we only
1415 * set the FUTEX_WAITERS bit in the user space variable.
1417 return attach_to_pi_owner(uaddr, newval, key, ps);
1421 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1422 * @q: The futex_q to unqueue
1424 * The q->lock_ptr must not be NULL and must be held by the caller.
1426 static void __unqueue_futex(struct futex_q *q)
1428 struct futex_hash_bucket *hb;
1430 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1431 || WARN_ON(plist_node_empty(&q->list)))
1434 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1435 plist_del(&q->list, &hb->chain);
1440 * The hash bucket lock must be held when this is called.
1441 * Afterwards, the futex_q must not be accessed. Callers
1442 * must ensure to later call wake_up_q() for the actual
1445 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1447 struct task_struct *p = q->task;
1449 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1455 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1456 * is written, without taking any locks. This is possible in the event
1457 * of a spurious wakeup, for example. A memory barrier is required here
1458 * to prevent the following store to lock_ptr from getting ahead of the
1459 * plist_del in __unqueue_futex().
1461 smp_store_release(&q->lock_ptr, NULL);
1464 * Queue the task for later wakeup for after we've released
1465 * the hb->lock. wake_q_add() grabs reference to p.
1467 wake_q_add(wake_q, p);
1472 * Caller must hold a reference on @pi_state.
1474 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1476 u32 uninitialized_var(curval), newval;
1477 struct task_struct *new_owner;
1478 bool postunlock = false;
1479 DEFINE_WAKE_Q(wake_q);
1482 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1483 if (WARN_ON_ONCE(!new_owner)) {
1485 * As per the comment in futex_unlock_pi() this should not happen.
1487 * When this happens, give up our locks and try again, giving
1488 * the futex_lock_pi() instance time to complete, either by
1489 * waiting on the rtmutex or removing itself from the futex
1497 * We pass it to the next owner. The WAITERS bit is always kept
1498 * enabled while there is PI state around. We cleanup the owner
1499 * died bit, because we are the owner.
1501 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1503 if (unlikely(should_fail_futex(true)))
1506 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1507 if (!ret && (curval != uval)) {
1509 * If a unconditional UNLOCK_PI operation (user space did not
1510 * try the TID->0 transition) raced with a waiter setting the
1511 * FUTEX_WAITERS flag between get_user() and locking the hash
1512 * bucket lock, retry the operation.
1514 if ((FUTEX_TID_MASK & curval) == uval)
1524 * This is a point of no return; once we modify the uval there is no
1525 * going back and subsequent operations must not fail.
1528 raw_spin_lock(&pi_state->owner->pi_lock);
1529 WARN_ON(list_empty(&pi_state->list));
1530 list_del_init(&pi_state->list);
1531 raw_spin_unlock(&pi_state->owner->pi_lock);
1533 raw_spin_lock(&new_owner->pi_lock);
1534 WARN_ON(!list_empty(&pi_state->list));
1535 list_add(&pi_state->list, &new_owner->pi_state_list);
1536 pi_state->owner = new_owner;
1537 raw_spin_unlock(&new_owner->pi_lock);
1539 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1542 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1545 rt_mutex_postunlock(&wake_q);
1551 * Express the locking dependencies for lockdep:
1554 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1557 spin_lock(&hb1->lock);
1559 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1560 } else { /* hb1 > hb2 */
1561 spin_lock(&hb2->lock);
1562 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1567 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1569 spin_unlock(&hb1->lock);
1571 spin_unlock(&hb2->lock);
1575 * Wake up waiters matching bitset queued on this futex (uaddr).
1578 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1580 struct futex_hash_bucket *hb;
1581 struct futex_q *this, *next;
1582 union futex_key key = FUTEX_KEY_INIT;
1584 DEFINE_WAKE_Q(wake_q);
1589 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1590 if (unlikely(ret != 0))
1593 hb = hash_futex(&key);
1595 /* Make sure we really have tasks to wakeup */
1596 if (!hb_waiters_pending(hb))
1599 spin_lock(&hb->lock);
1601 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1602 if (match_futex (&this->key, &key)) {
1603 if (this->pi_state || this->rt_waiter) {
1608 /* Check if one of the bits is set in both bitsets */
1609 if (!(this->bitset & bitset))
1612 mark_wake_futex(&wake_q, this);
1613 if (++ret >= nr_wake)
1618 spin_unlock(&hb->lock);
1621 put_futex_key(&key);
1626 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1628 unsigned int op = (encoded_op & 0x70000000) >> 28;
1629 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1630 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1631 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1634 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1635 if (oparg < 0 || oparg > 31) {
1636 char comm[sizeof(current->comm)];
1638 * kill this print and return -EINVAL when userspace
1641 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1642 get_task_comm(comm, current), oparg);
1648 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1651 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1656 case FUTEX_OP_CMP_EQ:
1657 return oldval == cmparg;
1658 case FUTEX_OP_CMP_NE:
1659 return oldval != cmparg;
1660 case FUTEX_OP_CMP_LT:
1661 return oldval < cmparg;
1662 case FUTEX_OP_CMP_GE:
1663 return oldval >= cmparg;
1664 case FUTEX_OP_CMP_LE:
1665 return oldval <= cmparg;
1666 case FUTEX_OP_CMP_GT:
1667 return oldval > cmparg;
1674 * Wake up all waiters hashed on the physical page that is mapped
1675 * to this virtual address:
1678 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1679 int nr_wake, int nr_wake2, int op)
1681 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1682 struct futex_hash_bucket *hb1, *hb2;
1683 struct futex_q *this, *next;
1685 DEFINE_WAKE_Q(wake_q);
1688 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1689 if (unlikely(ret != 0))
1691 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1692 if (unlikely(ret != 0))
1695 hb1 = hash_futex(&key1);
1696 hb2 = hash_futex(&key2);
1699 double_lock_hb(hb1, hb2);
1700 op_ret = futex_atomic_op_inuser(op, uaddr2);
1701 if (unlikely(op_ret < 0)) {
1702 double_unlock_hb(hb1, hb2);
1704 if (!IS_ENABLED(CONFIG_MMU) ||
1705 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1707 * we don't get EFAULT from MMU faults if we don't have
1708 * an MMU, but we might get them from range checking
1714 if (op_ret == -EFAULT) {
1715 ret = fault_in_user_writeable(uaddr2);
1720 if (!(flags & FLAGS_SHARED)) {
1725 put_futex_key(&key2);
1726 put_futex_key(&key1);
1731 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1732 if (match_futex (&this->key, &key1)) {
1733 if (this->pi_state || this->rt_waiter) {
1737 mark_wake_futex(&wake_q, this);
1738 if (++ret >= nr_wake)
1745 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1746 if (match_futex (&this->key, &key2)) {
1747 if (this->pi_state || this->rt_waiter) {
1751 mark_wake_futex(&wake_q, this);
1752 if (++op_ret >= nr_wake2)
1760 double_unlock_hb(hb1, hb2);
1763 put_futex_key(&key2);
1765 put_futex_key(&key1);
1771 * requeue_futex() - Requeue a futex_q from one hb to another
1772 * @q: the futex_q to requeue
1773 * @hb1: the source hash_bucket
1774 * @hb2: the target hash_bucket
1775 * @key2: the new key for the requeued futex_q
1778 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1779 struct futex_hash_bucket *hb2, union futex_key *key2)
1783 * If key1 and key2 hash to the same bucket, no need to
1786 if (likely(&hb1->chain != &hb2->chain)) {
1787 plist_del(&q->list, &hb1->chain);
1788 hb_waiters_dec(hb1);
1789 hb_waiters_inc(hb2);
1790 plist_add(&q->list, &hb2->chain);
1791 q->lock_ptr = &hb2->lock;
1793 get_futex_key_refs(key2);
1798 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1800 * @key: the key of the requeue target futex
1801 * @hb: the hash_bucket of the requeue target futex
1803 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1804 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1805 * to the requeue target futex so the waiter can detect the wakeup on the right
1806 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1807 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1808 * to protect access to the pi_state to fixup the owner later. Must be called
1809 * with both q->lock_ptr and hb->lock held.
1812 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1813 struct futex_hash_bucket *hb)
1815 get_futex_key_refs(key);
1820 WARN_ON(!q->rt_waiter);
1821 q->rt_waiter = NULL;
1823 q->lock_ptr = &hb->lock;
1825 wake_up_state(q->task, TASK_NORMAL);
1829 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1830 * @pifutex: the user address of the to futex
1831 * @hb1: the from futex hash bucket, must be locked by the caller
1832 * @hb2: the to futex hash bucket, must be locked by the caller
1833 * @key1: the from futex key
1834 * @key2: the to futex key
1835 * @ps: address to store the pi_state pointer
1836 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1838 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1839 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1840 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1841 * hb1 and hb2 must be held by the caller.
1844 * - 0 - failed to acquire the lock atomically;
1845 * - >0 - acquired the lock, return value is vpid of the top_waiter
1848 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1849 struct futex_hash_bucket *hb1,
1850 struct futex_hash_bucket *hb2,
1851 union futex_key *key1, union futex_key *key2,
1852 struct futex_pi_state **ps, int set_waiters)
1854 struct futex_q *top_waiter = NULL;
1858 if (get_futex_value_locked(&curval, pifutex))
1861 if (unlikely(should_fail_futex(true)))
1865 * Find the top_waiter and determine if there are additional waiters.
1866 * If the caller intends to requeue more than 1 waiter to pifutex,
1867 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1868 * as we have means to handle the possible fault. If not, don't set
1869 * the bit unecessarily as it will force the subsequent unlock to enter
1872 top_waiter = futex_top_waiter(hb1, key1);
1874 /* There are no waiters, nothing for us to do. */
1878 /* Ensure we requeue to the expected futex. */
1879 if (!match_futex(top_waiter->requeue_pi_key, key2))
1883 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1884 * the contended case or if set_waiters is 1. The pi_state is returned
1885 * in ps in contended cases.
1887 vpid = task_pid_vnr(top_waiter->task);
1888 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1891 requeue_pi_wake_futex(top_waiter, key2, hb2);
1898 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1899 * @uaddr1: source futex user address
1900 * @flags: futex flags (FLAGS_SHARED, etc.)
1901 * @uaddr2: target futex user address
1902 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1903 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1904 * @cmpval: @uaddr1 expected value (or %NULL)
1905 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1906 * pi futex (pi to pi requeue is not supported)
1908 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1909 * uaddr2 atomically on behalf of the top waiter.
1912 * - >=0 - on success, the number of tasks requeued or woken;
1915 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1916 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1917 u32 *cmpval, int requeue_pi)
1919 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1920 int drop_count = 0, task_count = 0, ret;
1921 struct futex_pi_state *pi_state = NULL;
1922 struct futex_hash_bucket *hb1, *hb2;
1923 struct futex_q *this, *next;
1924 DEFINE_WAKE_Q(wake_q);
1926 if (nr_wake < 0 || nr_requeue < 0)
1930 * When PI not supported: return -ENOSYS if requeue_pi is true,
1931 * consequently the compiler knows requeue_pi is always false past
1932 * this point which will optimize away all the conditional code
1935 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1940 * Requeue PI only works on two distinct uaddrs. This
1941 * check is only valid for private futexes. See below.
1943 if (uaddr1 == uaddr2)
1947 * requeue_pi requires a pi_state, try to allocate it now
1948 * without any locks in case it fails.
1950 if (refill_pi_state_cache())
1953 * requeue_pi must wake as many tasks as it can, up to nr_wake
1954 * + nr_requeue, since it acquires the rt_mutex prior to
1955 * returning to userspace, so as to not leave the rt_mutex with
1956 * waiters and no owner. However, second and third wake-ups
1957 * cannot be predicted as they involve race conditions with the
1958 * first wake and a fault while looking up the pi_state. Both
1959 * pthread_cond_signal() and pthread_cond_broadcast() should
1967 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1968 if (unlikely(ret != 0))
1970 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1971 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1972 if (unlikely(ret != 0))
1976 * The check above which compares uaddrs is not sufficient for
1977 * shared futexes. We need to compare the keys:
1979 if (requeue_pi && match_futex(&key1, &key2)) {
1984 hb1 = hash_futex(&key1);
1985 hb2 = hash_futex(&key2);
1988 hb_waiters_inc(hb2);
1989 double_lock_hb(hb1, hb2);
1991 if (likely(cmpval != NULL)) {
1994 ret = get_futex_value_locked(&curval, uaddr1);
1996 if (unlikely(ret)) {
1997 double_unlock_hb(hb1, hb2);
1998 hb_waiters_dec(hb2);
2000 ret = get_user(curval, uaddr1);
2004 if (!(flags & FLAGS_SHARED))
2007 put_futex_key(&key2);
2008 put_futex_key(&key1);
2011 if (curval != *cmpval) {
2017 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2019 * Attempt to acquire uaddr2 and wake the top waiter. If we
2020 * intend to requeue waiters, force setting the FUTEX_WAITERS
2021 * bit. We force this here where we are able to easily handle
2022 * faults rather in the requeue loop below.
2024 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2025 &key2, &pi_state, nr_requeue);
2028 * At this point the top_waiter has either taken uaddr2 or is
2029 * waiting on it. If the former, then the pi_state will not
2030 * exist yet, look it up one more time to ensure we have a
2031 * reference to it. If the lock was taken, ret contains the
2032 * vpid of the top waiter task.
2033 * If the lock was not taken, we have pi_state and an initial
2034 * refcount on it. In case of an error we have nothing.
2041 * If we acquired the lock, then the user space value
2042 * of uaddr2 should be vpid. It cannot be changed by
2043 * the top waiter as it is blocked on hb2 lock if it
2044 * tries to do so. If something fiddled with it behind
2045 * our back the pi state lookup might unearth it. So
2046 * we rather use the known value than rereading and
2047 * handing potential crap to lookup_pi_state.
2049 * If that call succeeds then we have pi_state and an
2050 * initial refcount on it.
2052 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
2057 /* We hold a reference on the pi state. */
2060 /* If the above failed, then pi_state is NULL */
2062 double_unlock_hb(hb1, hb2);
2063 hb_waiters_dec(hb2);
2064 put_futex_key(&key2);
2065 put_futex_key(&key1);
2066 ret = fault_in_user_writeable(uaddr2);
2072 * Two reasons for this:
2073 * - Owner is exiting and we just wait for the
2075 * - The user space value changed.
2077 double_unlock_hb(hb1, hb2);
2078 hb_waiters_dec(hb2);
2079 put_futex_key(&key2);
2080 put_futex_key(&key1);
2088 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2089 if (task_count - nr_wake >= nr_requeue)
2092 if (!match_futex(&this->key, &key1))
2096 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2097 * be paired with each other and no other futex ops.
2099 * We should never be requeueing a futex_q with a pi_state,
2100 * which is awaiting a futex_unlock_pi().
2102 if ((requeue_pi && !this->rt_waiter) ||
2103 (!requeue_pi && this->rt_waiter) ||
2110 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2111 * lock, we already woke the top_waiter. If not, it will be
2112 * woken by futex_unlock_pi().
2114 if (++task_count <= nr_wake && !requeue_pi) {
2115 mark_wake_futex(&wake_q, this);
2119 /* Ensure we requeue to the expected futex for requeue_pi. */
2120 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2126 * Requeue nr_requeue waiters and possibly one more in the case
2127 * of requeue_pi if we couldn't acquire the lock atomically.
2131 * Prepare the waiter to take the rt_mutex. Take a
2132 * refcount on the pi_state and store the pointer in
2133 * the futex_q object of the waiter.
2135 get_pi_state(pi_state);
2136 this->pi_state = pi_state;
2137 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2142 * We got the lock. We do neither drop the
2143 * refcount on pi_state nor clear
2144 * this->pi_state because the waiter needs the
2145 * pi_state for cleaning up the user space
2146 * value. It will drop the refcount after
2149 requeue_pi_wake_futex(this, &key2, hb2);
2154 * rt_mutex_start_proxy_lock() detected a
2155 * potential deadlock when we tried to queue
2156 * that waiter. Drop the pi_state reference
2157 * which we took above and remove the pointer
2158 * to the state from the waiters futex_q
2161 this->pi_state = NULL;
2162 put_pi_state(pi_state);
2164 * We stop queueing more waiters and let user
2165 * space deal with the mess.
2170 requeue_futex(this, hb1, hb2, &key2);
2175 * We took an extra initial reference to the pi_state either
2176 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2177 * need to drop it here again.
2179 put_pi_state(pi_state);
2182 double_unlock_hb(hb1, hb2);
2184 hb_waiters_dec(hb2);
2187 * drop_futex_key_refs() must be called outside the spinlocks. During
2188 * the requeue we moved futex_q's from the hash bucket at key1 to the
2189 * one at key2 and updated their key pointer. We no longer need to
2190 * hold the references to key1.
2192 while (--drop_count >= 0)
2193 drop_futex_key_refs(&key1);
2196 put_futex_key(&key2);
2198 put_futex_key(&key1);
2200 return ret ? ret : task_count;
2203 /* The key must be already stored in q->key. */
2204 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2205 __acquires(&hb->lock)
2207 struct futex_hash_bucket *hb;
2209 hb = hash_futex(&q->key);
2212 * Increment the counter before taking the lock so that
2213 * a potential waker won't miss a to-be-slept task that is
2214 * waiting for the spinlock. This is safe as all queue_lock()
2215 * users end up calling queue_me(). Similarly, for housekeeping,
2216 * decrement the counter at queue_unlock() when some error has
2217 * occurred and we don't end up adding the task to the list.
2221 q->lock_ptr = &hb->lock;
2223 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2228 queue_unlock(struct futex_hash_bucket *hb)
2229 __releases(&hb->lock)
2231 spin_unlock(&hb->lock);
2235 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2240 * The priority used to register this element is
2241 * - either the real thread-priority for the real-time threads
2242 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2243 * - or MAX_RT_PRIO for non-RT threads.
2244 * Thus, all RT-threads are woken first in priority order, and
2245 * the others are woken last, in FIFO order.
2247 prio = min(current->normal_prio, MAX_RT_PRIO);
2249 plist_node_init(&q->list, prio);
2250 plist_add(&q->list, &hb->chain);
2255 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2256 * @q: The futex_q to enqueue
2257 * @hb: The destination hash bucket
2259 * The hb->lock must be held by the caller, and is released here. A call to
2260 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2261 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2262 * or nothing if the unqueue is done as part of the wake process and the unqueue
2263 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2266 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2267 __releases(&hb->lock)
2270 spin_unlock(&hb->lock);
2274 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2275 * @q: The futex_q to unqueue
2277 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2278 * be paired with exactly one earlier call to queue_me().
2281 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2282 * - 0 - if the futex_q was already removed by the waking thread
2284 static int unqueue_me(struct futex_q *q)
2286 spinlock_t *lock_ptr;
2289 /* In the common case we don't take the spinlock, which is nice. */
2292 * q->lock_ptr can change between this read and the following spin_lock.
2293 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2294 * optimizing lock_ptr out of the logic below.
2296 lock_ptr = READ_ONCE(q->lock_ptr);
2297 if (lock_ptr != NULL) {
2298 spin_lock(lock_ptr);
2300 * q->lock_ptr can change between reading it and
2301 * spin_lock(), causing us to take the wrong lock. This
2302 * corrects the race condition.
2304 * Reasoning goes like this: if we have the wrong lock,
2305 * q->lock_ptr must have changed (maybe several times)
2306 * between reading it and the spin_lock(). It can
2307 * change again after the spin_lock() but only if it was
2308 * already changed before the spin_lock(). It cannot,
2309 * however, change back to the original value. Therefore
2310 * we can detect whether we acquired the correct lock.
2312 if (unlikely(lock_ptr != q->lock_ptr)) {
2313 spin_unlock(lock_ptr);
2318 BUG_ON(q->pi_state);
2320 spin_unlock(lock_ptr);
2324 drop_futex_key_refs(&q->key);
2329 * PI futexes can not be requeued and must remove themself from the
2330 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2333 static void unqueue_me_pi(struct futex_q *q)
2334 __releases(q->lock_ptr)
2338 BUG_ON(!q->pi_state);
2339 put_pi_state(q->pi_state);
2342 spin_unlock(q->lock_ptr);
2345 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2346 struct task_struct *argowner)
2348 struct futex_pi_state *pi_state = q->pi_state;
2349 u32 uval, uninitialized_var(curval), newval;
2350 struct task_struct *oldowner, *newowner;
2354 lockdep_assert_held(q->lock_ptr);
2356 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2358 oldowner = pi_state->owner;
2361 * We are here because either:
2363 * - we stole the lock and pi_state->owner needs updating to reflect
2364 * that (@argowner == current),
2368 * - someone stole our lock and we need to fix things to point to the
2369 * new owner (@argowner == NULL).
2371 * Either way, we have to replace the TID in the user space variable.
2372 * This must be atomic as we have to preserve the owner died bit here.
2374 * Note: We write the user space value _before_ changing the pi_state
2375 * because we can fault here. Imagine swapped out pages or a fork
2376 * that marked all the anonymous memory readonly for cow.
2378 * Modifying pi_state _before_ the user space value would leave the
2379 * pi_state in an inconsistent state when we fault here, because we
2380 * need to drop the locks to handle the fault. This might be observed
2381 * in the PID check in lookup_pi_state.
2385 if (oldowner != current) {
2387 * We raced against a concurrent self; things are
2388 * already fixed up. Nothing to do.
2394 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2395 /* We got the lock after all, nothing to fix. */
2401 * Since we just failed the trylock; there must be an owner.
2403 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2406 WARN_ON_ONCE(argowner != current);
2407 if (oldowner == current) {
2409 * We raced against a concurrent self; things are
2410 * already fixed up. Nothing to do.
2415 newowner = argowner;
2418 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2420 if (!pi_state->owner)
2421 newtid |= FUTEX_OWNER_DIED;
2423 err = get_futex_value_locked(&uval, uaddr);
2428 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2430 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2440 * We fixed up user space. Now we need to fix the pi_state
2443 if (pi_state->owner != NULL) {
2444 raw_spin_lock(&pi_state->owner->pi_lock);
2445 WARN_ON(list_empty(&pi_state->list));
2446 list_del_init(&pi_state->list);
2447 raw_spin_unlock(&pi_state->owner->pi_lock);
2450 pi_state->owner = newowner;
2452 raw_spin_lock(&newowner->pi_lock);
2453 WARN_ON(!list_empty(&pi_state->list));
2454 list_add(&pi_state->list, &newowner->pi_state_list);
2455 raw_spin_unlock(&newowner->pi_lock);
2456 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2461 * In order to reschedule or handle a page fault, we need to drop the
2462 * locks here. In the case of a fault, this gives the other task
2463 * (either the highest priority waiter itself or the task which stole
2464 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2465 * are back from handling the fault we need to check the pi_state after
2466 * reacquiring the locks and before trying to do another fixup. When
2467 * the fixup has been done already we simply return.
2469 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2470 * drop hb->lock since the caller owns the hb -> futex_q relation.
2471 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2474 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2475 spin_unlock(q->lock_ptr);
2479 ret = fault_in_user_writeable(uaddr);
2493 spin_lock(q->lock_ptr);
2494 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2497 * Check if someone else fixed it for us:
2499 if (pi_state->owner != oldowner) {
2510 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2514 static long futex_wait_restart(struct restart_block *restart);
2517 * fixup_owner() - Post lock pi_state and corner case management
2518 * @uaddr: user address of the futex
2519 * @q: futex_q (contains pi_state and access to the rt_mutex)
2520 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2522 * After attempting to lock an rt_mutex, this function is called to cleanup
2523 * the pi_state owner as well as handle race conditions that may allow us to
2524 * acquire the lock. Must be called with the hb lock held.
2527 * - 1 - success, lock taken;
2528 * - 0 - success, lock not taken;
2529 * - <0 - on error (-EFAULT)
2531 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2537 * Got the lock. We might not be the anticipated owner if we
2538 * did a lock-steal - fix up the PI-state in that case:
2540 * Speculative pi_state->owner read (we don't hold wait_lock);
2541 * since we own the lock pi_state->owner == current is the
2542 * stable state, anything else needs more attention.
2544 if (q->pi_state->owner != current)
2545 ret = fixup_pi_state_owner(uaddr, q, current);
2550 * If we didn't get the lock; check if anybody stole it from us. In
2551 * that case, we need to fix up the uval to point to them instead of
2552 * us, otherwise bad things happen. [10]
2554 * Another speculative read; pi_state->owner == current is unstable
2555 * but needs our attention.
2557 if (q->pi_state->owner == current) {
2558 ret = fixup_pi_state_owner(uaddr, q, NULL);
2563 * Paranoia check. If we did not take the lock, then we should not be
2564 * the owner of the rt_mutex.
2566 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
2567 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2568 "pi-state %p\n", ret,
2569 q->pi_state->pi_mutex.owner,
2570 q->pi_state->owner);
2574 return ret ? ret : locked;
2578 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2579 * @hb: the futex hash bucket, must be locked by the caller
2580 * @q: the futex_q to queue up on
2581 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2583 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2584 struct hrtimer_sleeper *timeout)
2587 * The task state is guaranteed to be set before another task can
2588 * wake it. set_current_state() is implemented using smp_store_mb() and
2589 * queue_me() calls spin_unlock() upon completion, both serializing
2590 * access to the hash list and forcing another memory barrier.
2592 set_current_state(TASK_INTERRUPTIBLE);
2597 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2600 * If we have been removed from the hash list, then another task
2601 * has tried to wake us, and we can skip the call to schedule().
2603 if (likely(!plist_node_empty(&q->list))) {
2605 * If the timer has already expired, current will already be
2606 * flagged for rescheduling. Only call schedule if there
2607 * is no timeout, or if it has yet to expire.
2609 if (!timeout || timeout->task)
2610 freezable_schedule();
2612 __set_current_state(TASK_RUNNING);
2616 * futex_wait_setup() - Prepare to wait on a futex
2617 * @uaddr: the futex userspace address
2618 * @val: the expected value
2619 * @flags: futex flags (FLAGS_SHARED, etc.)
2620 * @q: the associated futex_q
2621 * @hb: storage for hash_bucket pointer to be returned to caller
2623 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2624 * compare it with the expected value. Handle atomic faults internally.
2625 * Return with the hb lock held and a q.key reference on success, and unlocked
2626 * with no q.key reference on failure.
2629 * - 0 - uaddr contains val and hb has been locked;
2630 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2632 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2633 struct futex_q *q, struct futex_hash_bucket **hb)
2639 * Access the page AFTER the hash-bucket is locked.
2640 * Order is important:
2642 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2643 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2645 * The basic logical guarantee of a futex is that it blocks ONLY
2646 * if cond(var) is known to be true at the time of blocking, for
2647 * any cond. If we locked the hash-bucket after testing *uaddr, that
2648 * would open a race condition where we could block indefinitely with
2649 * cond(var) false, which would violate the guarantee.
2651 * On the other hand, we insert q and release the hash-bucket only
2652 * after testing *uaddr. This guarantees that futex_wait() will NOT
2653 * absorb a wakeup if *uaddr does not match the desired values
2654 * while the syscall executes.
2657 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2658 if (unlikely(ret != 0))
2662 *hb = queue_lock(q);
2664 ret = get_futex_value_locked(&uval, uaddr);
2669 ret = get_user(uval, uaddr);
2673 if (!(flags & FLAGS_SHARED))
2676 put_futex_key(&q->key);
2687 put_futex_key(&q->key);
2691 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2692 ktime_t *abs_time, u32 bitset)
2694 struct hrtimer_sleeper timeout, *to = NULL;
2695 struct restart_block *restart;
2696 struct futex_hash_bucket *hb;
2697 struct futex_q q = futex_q_init;
2707 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2708 CLOCK_REALTIME : CLOCK_MONOTONIC,
2710 hrtimer_init_sleeper(to, current);
2711 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2712 current->timer_slack_ns);
2717 * Prepare to wait on uaddr. On success, holds hb lock and increments
2720 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2724 /* queue_me and wait for wakeup, timeout, or a signal. */
2725 futex_wait_queue_me(hb, &q, to);
2727 /* If we were woken (and unqueued), we succeeded, whatever. */
2729 /* unqueue_me() drops q.key ref */
2730 if (!unqueue_me(&q))
2733 if (to && !to->task)
2737 * We expect signal_pending(current), but we might be the
2738 * victim of a spurious wakeup as well.
2740 if (!signal_pending(current))
2747 restart = ¤t->restart_block;
2748 restart->fn = futex_wait_restart;
2749 restart->futex.uaddr = uaddr;
2750 restart->futex.val = val;
2751 restart->futex.time = *abs_time;
2752 restart->futex.bitset = bitset;
2753 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2755 ret = -ERESTART_RESTARTBLOCK;
2759 hrtimer_cancel(&to->timer);
2760 destroy_hrtimer_on_stack(&to->timer);
2766 static long futex_wait_restart(struct restart_block *restart)
2768 u32 __user *uaddr = restart->futex.uaddr;
2769 ktime_t t, *tp = NULL;
2771 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2772 t = restart->futex.time;
2775 restart->fn = do_no_restart_syscall;
2777 return (long)futex_wait(uaddr, restart->futex.flags,
2778 restart->futex.val, tp, restart->futex.bitset);
2783 * Userspace tried a 0 -> TID atomic transition of the futex value
2784 * and failed. The kernel side here does the whole locking operation:
2785 * if there are waiters then it will block as a consequence of relying
2786 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2787 * a 0 value of the futex too.).
2789 * Also serves as futex trylock_pi()'ing, and due semantics.
2791 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2792 ktime_t *time, int trylock)
2794 struct hrtimer_sleeper timeout, *to = NULL;
2795 struct futex_pi_state *pi_state = NULL;
2796 struct rt_mutex_waiter rt_waiter;
2797 struct futex_hash_bucket *hb;
2798 struct futex_q q = futex_q_init;
2801 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2804 if (refill_pi_state_cache())
2809 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2811 hrtimer_init_sleeper(to, current);
2812 hrtimer_set_expires(&to->timer, *time);
2816 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2817 if (unlikely(ret != 0))
2821 hb = queue_lock(&q);
2823 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2824 if (unlikely(ret)) {
2826 * Atomic work succeeded and we got the lock,
2827 * or failed. Either way, we do _not_ block.
2831 /* We got the lock. */
2833 goto out_unlock_put_key;
2838 * Two reasons for this:
2839 * - Task is exiting and we just wait for the
2841 * - The user space value changed.
2844 put_futex_key(&q.key);
2848 goto out_unlock_put_key;
2852 WARN_ON(!q.pi_state);
2855 * Only actually queue now that the atomic ops are done:
2860 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2861 /* Fixup the trylock return value: */
2862 ret = ret ? 0 : -EWOULDBLOCK;
2866 rt_mutex_init_waiter(&rt_waiter);
2869 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2870 * hold it while doing rt_mutex_start_proxy(), because then it will
2871 * include hb->lock in the blocking chain, even through we'll not in
2872 * fact hold it while blocking. This will lead it to report -EDEADLK
2873 * and BUG when futex_unlock_pi() interleaves with this.
2875 * Therefore acquire wait_lock while holding hb->lock, but drop the
2876 * latter before calling __rt_mutex_start_proxy_lock(). This
2877 * interleaves with futex_unlock_pi() -- which does a similar lock
2878 * handoff -- such that the latter can observe the futex_q::pi_state
2879 * before __rt_mutex_start_proxy_lock() is done.
2881 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2882 spin_unlock(q.lock_ptr);
2884 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2885 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2886 * it sees the futex_q::pi_state.
2888 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2889 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2898 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2900 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2903 spin_lock(q.lock_ptr);
2905 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2906 * first acquire the hb->lock before removing the lock from the
2907 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2910 * In particular; it is important that futex_unlock_pi() can not
2911 * observe this inconsistency.
2913 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2918 * Fixup the pi_state owner and possibly acquire the lock if we
2921 res = fixup_owner(uaddr, &q, !ret);
2923 * If fixup_owner() returned an error, proprogate that. If it acquired
2924 * the lock, clear our -ETIMEDOUT or -EINTR.
2927 ret = (res < 0) ? res : 0;
2930 * If fixup_owner() faulted and was unable to handle the fault, unlock
2931 * it and return the fault to userspace.
2933 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2934 pi_state = q.pi_state;
2935 get_pi_state(pi_state);
2938 /* Unqueue and drop the lock */
2942 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2943 put_pi_state(pi_state);
2952 put_futex_key(&q.key);
2955 hrtimer_cancel(&to->timer);
2956 destroy_hrtimer_on_stack(&to->timer);
2958 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2963 ret = fault_in_user_writeable(uaddr);
2967 if (!(flags & FLAGS_SHARED))
2970 put_futex_key(&q.key);
2975 * Userspace attempted a TID -> 0 atomic transition, and failed.
2976 * This is the in-kernel slowpath: we look up the PI state (if any),
2977 * and do the rt-mutex unlock.
2979 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2981 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2982 union futex_key key = FUTEX_KEY_INIT;
2983 struct futex_hash_bucket *hb;
2984 struct futex_q *top_waiter;
2987 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2991 if (get_user(uval, uaddr))
2994 * We release only a lock we actually own:
2996 if ((uval & FUTEX_TID_MASK) != vpid)
2999 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
3003 hb = hash_futex(&key);
3004 spin_lock(&hb->lock);
3007 * Check waiters first. We do not trust user space values at
3008 * all and we at least want to know if user space fiddled
3009 * with the futex value instead of blindly unlocking.
3011 top_waiter = futex_top_waiter(hb, &key);
3013 struct futex_pi_state *pi_state = top_waiter->pi_state;
3020 * If current does not own the pi_state then the futex is
3021 * inconsistent and user space fiddled with the futex value.
3023 if (pi_state->owner != current)
3026 get_pi_state(pi_state);
3028 * By taking wait_lock while still holding hb->lock, we ensure
3029 * there is no point where we hold neither; and therefore
3030 * wake_futex_pi() must observe a state consistent with what we
3033 * In particular; this forces __rt_mutex_start_proxy() to
3034 * complete such that we're guaranteed to observe the
3035 * rt_waiter. Also see the WARN in wake_futex_pi().
3037 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3038 spin_unlock(&hb->lock);
3040 /* drops pi_state->pi_mutex.wait_lock */
3041 ret = wake_futex_pi(uaddr, uval, pi_state);
3043 put_pi_state(pi_state);
3046 * Success, we're done! No tricky corner cases.
3051 * The atomic access to the futex value generated a
3052 * pagefault, so retry the user-access and the wakeup:
3057 * A unconditional UNLOCK_PI op raced against a waiter
3058 * setting the FUTEX_WAITERS bit. Try again.
3063 * wake_futex_pi has detected invalid state. Tell user
3070 * We have no kernel internal state, i.e. no waiters in the
3071 * kernel. Waiters which are about to queue themselves are stuck
3072 * on hb->lock. So we can safely ignore them. We do neither
3073 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3076 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3077 spin_unlock(&hb->lock);
3092 * If uval has changed, let user space handle it.
3094 ret = (curval == uval) ? 0 : -EAGAIN;
3097 spin_unlock(&hb->lock);
3099 put_futex_key(&key);
3103 put_futex_key(&key);
3108 put_futex_key(&key);
3110 ret = fault_in_user_writeable(uaddr);
3118 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3119 * @hb: the hash_bucket futex_q was original enqueued on
3120 * @q: the futex_q woken while waiting to be requeued
3121 * @key2: the futex_key of the requeue target futex
3122 * @timeout: the timeout associated with the wait (NULL if none)
3124 * Detect if the task was woken on the initial futex as opposed to the requeue
3125 * target futex. If so, determine if it was a timeout or a signal that caused
3126 * the wakeup and return the appropriate error code to the caller. Must be
3127 * called with the hb lock held.
3130 * - 0 = no early wakeup detected;
3131 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3134 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3135 struct futex_q *q, union futex_key *key2,
3136 struct hrtimer_sleeper *timeout)
3141 * With the hb lock held, we avoid races while we process the wakeup.
3142 * We only need to hold hb (and not hb2) to ensure atomicity as the
3143 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3144 * It can't be requeued from uaddr2 to something else since we don't
3145 * support a PI aware source futex for requeue.
3147 if (!match_futex(&q->key, key2)) {
3148 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3150 * We were woken prior to requeue by a timeout or a signal.
3151 * Unqueue the futex_q and determine which it was.
3153 plist_del(&q->list, &hb->chain);
3156 /* Handle spurious wakeups gracefully */
3158 if (timeout && !timeout->task)
3160 else if (signal_pending(current))
3161 ret = -ERESTARTNOINTR;
3167 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3168 * @uaddr: the futex we initially wait on (non-pi)
3169 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3170 * the same type, no requeueing from private to shared, etc.
3171 * @val: the expected value of uaddr
3172 * @abs_time: absolute timeout
3173 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3174 * @uaddr2: the pi futex we will take prior to returning to user-space
3176 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3177 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3178 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3179 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3180 * without one, the pi logic would not know which task to boost/deboost, if
3181 * there was a need to.
3183 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3184 * via the following--
3185 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3186 * 2) wakeup on uaddr2 after a requeue
3190 * If 3, cleanup and return -ERESTARTNOINTR.
3192 * If 2, we may then block on trying to take the rt_mutex and return via:
3193 * 5) successful lock
3196 * 8) other lock acquisition failure
3198 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3200 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3206 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3207 u32 val, ktime_t *abs_time, u32 bitset,
3210 struct hrtimer_sleeper timeout, *to = NULL;
3211 struct futex_pi_state *pi_state = NULL;
3212 struct rt_mutex_waiter rt_waiter;
3213 struct futex_hash_bucket *hb;
3214 union futex_key key2 = FUTEX_KEY_INIT;
3215 struct futex_q q = futex_q_init;
3218 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3221 if (uaddr == uaddr2)
3229 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3230 CLOCK_REALTIME : CLOCK_MONOTONIC,
3232 hrtimer_init_sleeper(to, current);
3233 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3234 current->timer_slack_ns);
3238 * The waiter is allocated on our stack, manipulated by the requeue
3239 * code while we sleep on uaddr.
3241 rt_mutex_init_waiter(&rt_waiter);
3243 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
3244 if (unlikely(ret != 0))
3248 q.rt_waiter = &rt_waiter;
3249 q.requeue_pi_key = &key2;
3252 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3255 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3260 * The check above which compares uaddrs is not sufficient for
3261 * shared futexes. We need to compare the keys:
3263 if (match_futex(&q.key, &key2)) {
3269 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3270 futex_wait_queue_me(hb, &q, to);
3272 spin_lock(&hb->lock);
3273 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3274 spin_unlock(&hb->lock);
3279 * In order for us to be here, we know our q.key == key2, and since
3280 * we took the hb->lock above, we also know that futex_requeue() has
3281 * completed and we no longer have to concern ourselves with a wakeup
3282 * race with the atomic proxy lock acquisition by the requeue code. The
3283 * futex_requeue dropped our key1 reference and incremented our key2
3287 /* Check if the requeue code acquired the second futex for us. */
3290 * Got the lock. We might not be the anticipated owner if we
3291 * did a lock-steal - fix up the PI-state in that case.
3293 if (q.pi_state && (q.pi_state->owner != current)) {
3294 spin_lock(q.lock_ptr);
3295 ret = fixup_pi_state_owner(uaddr2, &q, current);
3296 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3297 pi_state = q.pi_state;
3298 get_pi_state(pi_state);
3301 * Drop the reference to the pi state which
3302 * the requeue_pi() code acquired for us.
3304 put_pi_state(q.pi_state);
3305 spin_unlock(q.lock_ptr);
3308 struct rt_mutex *pi_mutex;
3311 * We have been woken up by futex_unlock_pi(), a timeout, or a
3312 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3315 WARN_ON(!q.pi_state);
3316 pi_mutex = &q.pi_state->pi_mutex;
3317 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3319 spin_lock(q.lock_ptr);
3320 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3323 debug_rt_mutex_free_waiter(&rt_waiter);
3325 * Fixup the pi_state owner and possibly acquire the lock if we
3328 res = fixup_owner(uaddr2, &q, !ret);
3330 * If fixup_owner() returned an error, proprogate that. If it
3331 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3334 ret = (res < 0) ? res : 0;
3337 * If fixup_pi_state_owner() faulted and was unable to handle
3338 * the fault, unlock the rt_mutex and return the fault to
3341 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3342 pi_state = q.pi_state;
3343 get_pi_state(pi_state);
3346 /* Unqueue and drop the lock. */
3351 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3352 put_pi_state(pi_state);
3355 if (ret == -EINTR) {
3357 * We've already been requeued, but cannot restart by calling
3358 * futex_lock_pi() directly. We could restart this syscall, but
3359 * it would detect that the user space "val" changed and return
3360 * -EWOULDBLOCK. Save the overhead of the restart and return
3361 * -EWOULDBLOCK directly.
3367 put_futex_key(&q.key);
3369 put_futex_key(&key2);
3373 hrtimer_cancel(&to->timer);
3374 destroy_hrtimer_on_stack(&to->timer);
3380 * Support for robust futexes: the kernel cleans up held futexes at
3383 * Implementation: user-space maintains a per-thread list of locks it
3384 * is holding. Upon do_exit(), the kernel carefully walks this list,
3385 * and marks all locks that are owned by this thread with the
3386 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3387 * always manipulated with the lock held, so the list is private and
3388 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3389 * field, to allow the kernel to clean up if the thread dies after
3390 * acquiring the lock, but just before it could have added itself to
3391 * the list. There can only be one such pending lock.
3395 * sys_set_robust_list() - Set the robust-futex list head of a task
3396 * @head: pointer to the list-head
3397 * @len: length of the list-head, as userspace expects
3399 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3402 if (!futex_cmpxchg_enabled)
3405 * The kernel knows only one size for now:
3407 if (unlikely(len != sizeof(*head)))
3410 current->robust_list = head;
3416 * sys_get_robust_list() - Get the robust-futex list head of a task
3417 * @pid: pid of the process [zero for current task]
3418 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3419 * @len_ptr: pointer to a length field, the kernel fills in the header size
3421 SYSCALL_DEFINE3(get_robust_list, int, pid,
3422 struct robust_list_head __user * __user *, head_ptr,
3423 size_t __user *, len_ptr)
3425 struct robust_list_head __user *head;
3427 struct task_struct *p;
3429 if (!futex_cmpxchg_enabled)
3438 p = find_task_by_vpid(pid);
3444 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3447 head = p->robust_list;
3450 if (put_user(sizeof(*head), len_ptr))
3452 return put_user(head, head_ptr);
3460 /* Constants for the pending_op argument of handle_futex_death */
3461 #define HANDLE_DEATH_PENDING true
3462 #define HANDLE_DEATH_LIST false
3465 * Process a futex-list entry, check whether it's owned by the
3466 * dying task, and do notification if so:
3468 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3469 bool pi, bool pending_op)
3471 u32 uval, uninitialized_var(nval), mval;
3474 /* Futex address must be 32bit aligned */
3475 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3479 if (get_user(uval, uaddr))
3483 * Special case for regular (non PI) futexes. The unlock path in
3484 * user space has two race scenarios:
3486 * 1. The unlock path releases the user space futex value and
3487 * before it can execute the futex() syscall to wake up
3488 * waiters it is killed.
3490 * 2. A woken up waiter is killed before it can acquire the
3491 * futex in user space.
3493 * In both cases the TID validation below prevents a wakeup of
3494 * potential waiters which can cause these waiters to block
3497 * In both cases the following conditions are met:
3499 * 1) task->robust_list->list_op_pending != NULL
3500 * @pending_op == true
3501 * 2) User space futex value == 0
3502 * 3) Regular futex: @pi == false
3504 * If these conditions are met, it is safe to attempt waking up a
3505 * potential waiter without touching the user space futex value and
3506 * trying to set the OWNER_DIED bit. The user space futex value is
3507 * uncontended and the rest of the user space mutex state is
3508 * consistent, so a woken waiter will just take over the
3509 * uncontended futex. Setting the OWNER_DIED bit would create
3510 * inconsistent state and malfunction of the user space owner died
3513 if (pending_op && !pi && !uval) {
3514 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3518 if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3522 * Ok, this dying thread is truly holding a futex
3523 * of interest. Set the OWNER_DIED bit atomically
3524 * via cmpxchg, and if the value had FUTEX_WAITERS
3525 * set, wake up a waiter (if any). (We have to do a
3526 * futex_wake() even if OWNER_DIED is already set -
3527 * to handle the rare but possible case of recursive
3528 * thread-death.) The rest of the cleanup is done in
3531 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3534 * We are not holding a lock here, but we want to have
3535 * the pagefault_disable/enable() protection because
3536 * we want to handle the fault gracefully. If the
3537 * access fails we try to fault in the futex with R/W
3538 * verification via get_user_pages. get_user() above
3539 * does not guarantee R/W access. If that fails we
3540 * give up and leave the futex locked.
3542 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3545 if (fault_in_user_writeable(uaddr))
3563 * Wake robust non-PI futexes here. The wakeup of
3564 * PI futexes happens in exit_pi_state():
3566 if (!pi && (uval & FUTEX_WAITERS))
3567 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3573 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3575 static inline int fetch_robust_entry(struct robust_list __user **entry,
3576 struct robust_list __user * __user *head,
3579 unsigned long uentry;
3581 if (get_user(uentry, (unsigned long __user *)head))
3584 *entry = (void __user *)(uentry & ~1UL);
3591 * Walk curr->robust_list (very carefully, it's a userspace list!)
3592 * and mark any locks found there dead, and notify any waiters.
3594 * We silently return on any sign of list-walking problem.
3596 void exit_robust_list(struct task_struct *curr)
3598 struct robust_list_head __user *head = curr->robust_list;
3599 struct robust_list __user *entry, *next_entry, *pending;
3600 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3601 unsigned int uninitialized_var(next_pi);
3602 unsigned long futex_offset;
3605 if (!futex_cmpxchg_enabled)
3609 * Fetch the list head (which was registered earlier, via
3610 * sys_set_robust_list()):
3612 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3615 * Fetch the relative futex offset:
3617 if (get_user(futex_offset, &head->futex_offset))
3620 * Fetch any possibly pending lock-add first, and handle it
3623 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3626 next_entry = NULL; /* avoid warning with gcc */
3627 while (entry != &head->list) {
3629 * Fetch the next entry in the list before calling
3630 * handle_futex_death:
3632 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3634 * A pending lock might already be on the list, so
3635 * don't process it twice:
3637 if (entry != pending) {
3638 if (handle_futex_death((void __user *)entry + futex_offset,
3639 curr, pi, HANDLE_DEATH_LIST))
3647 * Avoid excessively long or circular lists:
3656 handle_futex_death((void __user *)pending + futex_offset,
3657 curr, pip, HANDLE_DEATH_PENDING);
3661 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3662 u32 __user *uaddr2, u32 val2, u32 val3)
3664 int cmd = op & FUTEX_CMD_MASK;
3665 unsigned int flags = 0;
3667 if (!(op & FUTEX_PRIVATE_FLAG))
3668 flags |= FLAGS_SHARED;
3670 if (op & FUTEX_CLOCK_REALTIME) {
3671 flags |= FLAGS_CLOCKRT;
3672 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3673 cmd != FUTEX_WAIT_REQUEUE_PI)
3679 case FUTEX_UNLOCK_PI:
3680 case FUTEX_TRYLOCK_PI:
3681 case FUTEX_WAIT_REQUEUE_PI:
3682 case FUTEX_CMP_REQUEUE_PI:
3683 if (!futex_cmpxchg_enabled)
3689 val3 = FUTEX_BITSET_MATCH_ANY;
3691 case FUTEX_WAIT_BITSET:
3692 return futex_wait(uaddr, flags, val, timeout, val3);
3694 val3 = FUTEX_BITSET_MATCH_ANY;
3696 case FUTEX_WAKE_BITSET:
3697 return futex_wake(uaddr, flags, val, val3);
3699 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3700 case FUTEX_CMP_REQUEUE:
3701 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3703 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3705 return futex_lock_pi(uaddr, flags, timeout, 0);
3706 case FUTEX_UNLOCK_PI:
3707 return futex_unlock_pi(uaddr, flags);
3708 case FUTEX_TRYLOCK_PI:
3709 return futex_lock_pi(uaddr, flags, NULL, 1);
3710 case FUTEX_WAIT_REQUEUE_PI:
3711 val3 = FUTEX_BITSET_MATCH_ANY;
3712 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3714 case FUTEX_CMP_REQUEUE_PI:
3715 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3721 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3722 struct timespec __user *, utime, u32 __user *, uaddr2,
3726 ktime_t t, *tp = NULL;
3728 int cmd = op & FUTEX_CMD_MASK;
3730 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3731 cmd == FUTEX_WAIT_BITSET ||
3732 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3733 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3735 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3737 if (!timespec_valid(&ts))
3740 t = timespec_to_ktime(ts);
3741 if (cmd == FUTEX_WAIT)
3742 t = ktime_add_safe(ktime_get(), t);
3746 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3747 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3749 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3750 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3751 val2 = (u32) (unsigned long) utime;
3753 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3756 #ifdef CONFIG_COMPAT
3758 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3761 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3762 compat_uptr_t __user *head, unsigned int *pi)
3764 if (get_user(*uentry, head))
3767 *entry = compat_ptr((*uentry) & ~1);
3768 *pi = (unsigned int)(*uentry) & 1;
3773 static void __user *futex_uaddr(struct robust_list __user *entry,
3774 compat_long_t futex_offset)
3776 compat_uptr_t base = ptr_to_compat(entry);
3777 void __user *uaddr = compat_ptr(base + futex_offset);
3783 * Walk curr->robust_list (very carefully, it's a userspace list!)
3784 * and mark any locks found there dead, and notify any waiters.
3786 * We silently return on any sign of list-walking problem.
3788 void compat_exit_robust_list(struct task_struct *curr)
3790 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3791 struct robust_list __user *entry, *next_entry, *pending;
3792 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3793 unsigned int uninitialized_var(next_pi);
3794 compat_uptr_t uentry, next_uentry, upending;
3795 compat_long_t futex_offset;
3798 if (!futex_cmpxchg_enabled)
3802 * Fetch the list head (which was registered earlier, via
3803 * sys_set_robust_list()):
3805 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3808 * Fetch the relative futex offset:
3810 if (get_user(futex_offset, &head->futex_offset))
3813 * Fetch any possibly pending lock-add first, and handle it
3816 if (compat_fetch_robust_entry(&upending, &pending,
3817 &head->list_op_pending, &pip))
3820 next_entry = NULL; /* avoid warning with gcc */
3821 while (entry != (struct robust_list __user *) &head->list) {
3823 * Fetch the next entry in the list before calling
3824 * handle_futex_death:
3826 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3827 (compat_uptr_t __user *)&entry->next, &next_pi);
3829 * A pending lock might already be on the list, so
3830 * dont process it twice:
3832 if (entry != pending) {
3833 void __user *uaddr = futex_uaddr(entry, futex_offset);
3835 if (handle_futex_death(uaddr, curr, pi,
3841 uentry = next_uentry;
3845 * Avoid excessively long or circular lists:
3853 void __user *uaddr = futex_uaddr(pending, futex_offset);
3855 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3859 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3860 struct compat_robust_list_head __user *, head,
3863 if (!futex_cmpxchg_enabled)
3866 if (unlikely(len != sizeof(*head)))
3869 current->compat_robust_list = head;
3874 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3875 compat_uptr_t __user *, head_ptr,
3876 compat_size_t __user *, len_ptr)
3878 struct compat_robust_list_head __user *head;
3880 struct task_struct *p;
3882 if (!futex_cmpxchg_enabled)
3891 p = find_task_by_vpid(pid);
3897 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3900 head = p->compat_robust_list;
3903 if (put_user(sizeof(*head), len_ptr))
3905 return put_user(ptr_to_compat(head), head_ptr);
3913 COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3914 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3918 ktime_t t, *tp = NULL;
3920 int cmd = op & FUTEX_CMD_MASK;
3922 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3923 cmd == FUTEX_WAIT_BITSET ||
3924 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3925 if (compat_get_timespec(&ts, utime))
3927 if (!timespec_valid(&ts))
3930 t = timespec_to_ktime(ts);
3931 if (cmd == FUTEX_WAIT)
3932 t = ktime_add_safe(ktime_get(), t);
3935 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3936 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3937 val2 = (int) (unsigned long) utime;
3939 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3941 #endif /* CONFIG_COMPAT */
3943 static void __init futex_detect_cmpxchg(void)
3945 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3949 * This will fail and we want it. Some arch implementations do
3950 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3951 * functionality. We want to know that before we call in any
3952 * of the complex code paths. Also we want to prevent
3953 * registration of robust lists in that case. NULL is
3954 * guaranteed to fault and we get -EFAULT on functional
3955 * implementation, the non-functional ones will return
3958 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3959 futex_cmpxchg_enabled = 1;
3963 static int __init futex_init(void)
3965 unsigned int futex_shift;
3968 #if CONFIG_BASE_SMALL
3969 futex_hashsize = 16;
3971 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3974 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3976 futex_hashsize < 256 ? HASH_SMALL : 0,
3978 futex_hashsize, futex_hashsize);
3979 futex_hashsize = 1UL << futex_shift;
3981 futex_detect_cmpxchg();
3983 for (i = 0; i < futex_hashsize; i++) {
3984 atomic_set(&futex_queues[i].waiters, 0);
3985 plist_head_init(&futex_queues[i].chain);
3986 spin_lock_init(&futex_queues[i].lock);
3991 core_initcall(futex_init);