2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
67 #include <linux/fault-inject.h>
69 #include <asm/futex.h>
71 #include "locking/rtmutex_common.h"
74 * READ this before attempting to hack on futexes!
76 * Basic futex operation and ordering guarantees
77 * =============================================
79 * The waiter reads the futex value in user space and calls
80 * futex_wait(). This function computes the hash bucket and acquires
81 * the hash bucket lock. After that it reads the futex user space value
82 * again and verifies that the data has not changed. If it has not changed
83 * it enqueues itself into the hash bucket, releases the hash bucket lock
86 * The waker side modifies the user space value of the futex and calls
87 * futex_wake(). This function computes the hash bucket and acquires the
88 * hash bucket lock. Then it looks for waiters on that futex in the hash
89 * bucket and wakes them.
91 * In futex wake up scenarios where no tasks are blocked on a futex, taking
92 * the hb spinlock can be avoided and simply return. In order for this
93 * optimization to work, ordering guarantees must exist so that the waiter
94 * being added to the list is acknowledged when the list is concurrently being
95 * checked by the waker, avoiding scenarios like the following:
99 * sys_futex(WAIT, futex, val);
100 * futex_wait(futex, val);
103 * sys_futex(WAKE, futex);
108 * lock(hash_bucket(futex));
110 * unlock(hash_bucket(futex));
113 * This would cause the waiter on CPU 0 to wait forever because it
114 * missed the transition of the user space value from val to newval
115 * and the waker did not find the waiter in the hash bucket queue.
117 * The correct serialization ensures that a waiter either observes
118 * the changed user space value before blocking or is woken by a
123 * sys_futex(WAIT, futex, val);
124 * futex_wait(futex, val);
127 * smp_mb(); (A) <-- paired with -.
129 * lock(hash_bucket(futex)); |
133 * | sys_futex(WAKE, futex);
134 * | futex_wake(futex);
136 * `--------> smp_mb(); (B)
139 * unlock(hash_bucket(futex));
140 * schedule(); if (waiters)
141 * lock(hash_bucket(futex));
142 * else wake_waiters(futex);
143 * waiters--; (b) unlock(hash_bucket(futex));
145 * Where (A) orders the waiters increment and the futex value read through
146 * atomic operations (see hb_waiters_inc) and where (B) orders the write
147 * to futex and the waiters read -- this is done by the barriers for both
148 * shared and private futexes in get_futex_key_refs().
150 * This yields the following case (where X:=waiters, Y:=futex):
158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
159 * the guarantee that we cannot both miss the futex variable change and the
162 * Note that a new waiter is accounted for in (a) even when it is possible that
163 * the wait call can return error, in which case we backtrack from it in (b).
164 * Refer to the comment in queue_lock().
166 * Similarly, in order to account for waiters being requeued on another
167 * address we always increment the waiters for the destination bucket before
168 * acquiring the lock. It then decrements them again after releasing it -
169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170 * will do the additional required waiter count housekeeping. This is done for
171 * double_lock_hb() and double_unlock_hb(), respectively.
174 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175 int __read_mostly futex_cmpxchg_enabled;
179 * Futex flags used to encode options to functions and preserve them across
183 # define FLAGS_SHARED 0x01
186 * NOMMU does not have per process address space. Let the compiler optimize
189 # define FLAGS_SHARED 0x00
191 #define FLAGS_CLOCKRT 0x02
192 #define FLAGS_HAS_TIMEOUT 0x04
195 * Priority Inheritance state:
197 struct futex_pi_state {
199 * list of 'owned' pi_state instances - these have to be
200 * cleaned up in do_exit() if the task exits prematurely:
202 struct list_head list;
207 struct rt_mutex pi_mutex;
209 struct task_struct *owner;
216 * struct futex_q - The hashed futex queue entry, one per waiting task
217 * @list: priority-sorted list of tasks waiting on this futex
218 * @task: the task waiting on the futex
219 * @lock_ptr: the hash bucket lock
220 * @key: the key the futex is hashed on
221 * @pi_state: optional priority inheritance state
222 * @rt_waiter: rt_waiter storage for use with requeue_pi
223 * @requeue_pi_key: the requeue_pi target futex key
224 * @bitset: bitset for the optional bitmasked wakeup
226 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
227 * we can wake only the relevant ones (hashed queues may be shared).
229 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
230 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
231 * The order of wakeup is always to make the first condition true, then
234 * PI futexes are typically woken before they are removed from the hash list via
235 * the rt_mutex code. See unqueue_me_pi().
238 struct plist_node list;
240 struct task_struct *task;
241 spinlock_t *lock_ptr;
243 struct futex_pi_state *pi_state;
244 struct rt_mutex_waiter *rt_waiter;
245 union futex_key *requeue_pi_key;
249 static const struct futex_q futex_q_init = {
250 /* list gets initialized in queue_me()*/
251 .key = FUTEX_KEY_INIT,
252 .bitset = FUTEX_BITSET_MATCH_ANY
256 * Hash buckets are shared by all the futex_keys that hash to the same
257 * location. Each key may have multiple futex_q structures, one for each task
258 * waiting on a futex.
260 struct futex_hash_bucket {
263 struct plist_head chain;
264 } ____cacheline_aligned_in_smp;
267 * The base of the bucket array and its size are always used together
268 * (after initialization only in hash_futex()), so ensure that they
269 * reside in the same cacheline.
272 struct futex_hash_bucket *queues;
273 unsigned long hashsize;
274 } __futex_data __read_mostly __aligned(2*sizeof(long));
275 #define futex_queues (__futex_data.queues)
276 #define futex_hashsize (__futex_data.hashsize)
280 * Fault injections for futexes.
282 #ifdef CONFIG_FAIL_FUTEX
285 struct fault_attr attr;
289 .attr = FAULT_ATTR_INITIALIZER,
290 .ignore_private = false,
293 static int __init setup_fail_futex(char *str)
295 return setup_fault_attr(&fail_futex.attr, str);
297 __setup("fail_futex=", setup_fail_futex);
299 static bool should_fail_futex(bool fshared)
301 if (fail_futex.ignore_private && !fshared)
304 return should_fail(&fail_futex.attr, 1);
307 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
309 static int __init fail_futex_debugfs(void)
311 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
314 dir = fault_create_debugfs_attr("fail_futex", NULL,
319 if (!debugfs_create_bool("ignore-private", mode, dir,
320 &fail_futex.ignore_private)) {
321 debugfs_remove_recursive(dir);
328 late_initcall(fail_futex_debugfs);
330 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
333 static inline bool should_fail_futex(bool fshared)
337 #endif /* CONFIG_FAIL_FUTEX */
339 static inline void futex_get_mm(union futex_key *key)
341 atomic_inc(&key->private.mm->mm_count);
343 * Ensure futex_get_mm() implies a full barrier such that
344 * get_futex_key() implies a full barrier. This is relied upon
345 * as smp_mb(); (B), see the ordering comment above.
347 smp_mb__after_atomic();
351 * Reflects a new waiter being added to the waitqueue.
353 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
356 atomic_inc(&hb->waiters);
358 * Full barrier (A), see the ordering comment above.
360 smp_mb__after_atomic();
365 * Reflects a waiter being removed from the waitqueue by wakeup
368 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
371 atomic_dec(&hb->waiters);
375 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
378 return atomic_read(&hb->waiters);
385 * hash_futex - Return the hash bucket in the global hash
386 * @key: Pointer to the futex key for which the hash is calculated
388 * We hash on the keys returned from get_futex_key (see below) and return the
389 * corresponding hash bucket in the global hash.
391 static struct futex_hash_bucket *hash_futex(union futex_key *key)
393 u32 hash = jhash2((u32*)&key->both.word,
394 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
396 return &futex_queues[hash & (futex_hashsize - 1)];
401 * match_futex - Check whether two futex keys are equal
402 * @key1: Pointer to key1
403 * @key2: Pointer to key2
405 * Return 1 if two futex_keys are equal, 0 otherwise.
407 static inline int match_futex(union futex_key *key1, union futex_key *key2)
410 && key1->both.word == key2->both.word
411 && key1->both.ptr == key2->both.ptr
412 && key1->both.offset == key2->both.offset);
416 * Take a reference to the resource addressed by a key.
417 * Can be called while holding spinlocks.
420 static void get_futex_key_refs(union futex_key *key)
426 * On MMU less systems futexes are always "private" as there is no per
427 * process address space. We need the smp wmb nevertheless - yes,
428 * arch/blackfin has MMU less SMP ...
430 if (!IS_ENABLED(CONFIG_MMU)) {
431 smp_mb(); /* explicit smp_mb(); (B) */
435 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
437 ihold(key->shared.inode); /* implies smp_mb(); (B) */
439 case FUT_OFF_MMSHARED:
440 futex_get_mm(key); /* implies smp_mb(); (B) */
444 * Private futexes do not hold reference on an inode or
445 * mm, therefore the only purpose of calling get_futex_key_refs
446 * is because we need the barrier for the lockless waiter check.
448 smp_mb(); /* explicit smp_mb(); (B) */
453 * Drop a reference to the resource addressed by a key.
454 * The hash bucket spinlock must not be held. This is
455 * a no-op for private futexes, see comment in the get
458 static void drop_futex_key_refs(union futex_key *key)
460 if (!key->both.ptr) {
461 /* If we're here then we tried to put a key we failed to get */
466 if (!IS_ENABLED(CONFIG_MMU))
469 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
471 iput(key->shared.inode);
473 case FUT_OFF_MMSHARED:
474 mmdrop(key->private.mm);
480 * get_futex_key() - Get parameters which are the keys for a futex
481 * @uaddr: virtual address of the futex
482 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
483 * @key: address where result is stored.
484 * @rw: mapping needs to be read/write (values: VERIFY_READ,
487 * Return: a negative error code or 0
489 * The key words are stored in *key on success.
491 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
492 * offset_within_page). For private mappings, it's (uaddr, current->mm).
493 * We can usually work out the index without swapping in the page.
495 * lock_page() might sleep, the caller should not hold a spinlock.
498 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
500 unsigned long address = (unsigned long)uaddr;
501 struct mm_struct *mm = current->mm;
502 struct page *page, *tail;
503 struct address_space *mapping;
507 * The futex address must be "naturally" aligned.
509 key->both.offset = address % PAGE_SIZE;
510 if (unlikely((address % sizeof(u32)) != 0))
512 address -= key->both.offset;
514 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
517 if (unlikely(should_fail_futex(fshared)))
521 * PROCESS_PRIVATE futexes are fast.
522 * As the mm cannot disappear under us and the 'key' only needs
523 * virtual address, we dont even have to find the underlying vma.
524 * Note : We do have to check 'uaddr' is a valid user address,
525 * but access_ok() should be faster than find_vma()
528 key->private.mm = mm;
529 key->private.address = address;
530 get_futex_key_refs(key); /* implies smp_mb(); (B) */
535 /* Ignore any VERIFY_READ mapping (futex common case) */
536 if (unlikely(should_fail_futex(fshared)))
539 err = get_user_pages_fast(address, 1, 1, &page);
541 * If write access is not required (eg. FUTEX_WAIT), try
542 * and get read-only access.
544 if (err == -EFAULT && rw == VERIFY_READ) {
545 err = get_user_pages_fast(address, 1, 0, &page);
554 * The treatment of mapping from this point on is critical. The page
555 * lock protects many things but in this context the page lock
556 * stabilizes mapping, prevents inode freeing in the shared
557 * file-backed region case and guards against movement to swap cache.
559 * Strictly speaking the page lock is not needed in all cases being
560 * considered here and page lock forces unnecessarily serialization
561 * From this point on, mapping will be re-verified if necessary and
562 * page lock will be acquired only if it is unavoidable
564 * Mapping checks require the head page for any compound page so the
565 * head page and mapping is looked up now. For anonymous pages, it
566 * does not matter if the page splits in the future as the key is
567 * based on the address. For filesystem-backed pages, the tail is
568 * required as the index of the page determines the key. For
569 * base pages, there is no tail page and tail == page.
572 page = compound_head(page);
573 mapping = READ_ONCE(page->mapping);
576 * If page->mapping is NULL, then it cannot be a PageAnon
577 * page; but it might be the ZERO_PAGE or in the gate area or
578 * in a special mapping (all cases which we are happy to fail);
579 * or it may have been a good file page when get_user_pages_fast
580 * found it, but truncated or holepunched or subjected to
581 * invalidate_complete_page2 before we got the page lock (also
582 * cases which we are happy to fail). And we hold a reference,
583 * so refcount care in invalidate_complete_page's remove_mapping
584 * prevents drop_caches from setting mapping to NULL beneath us.
586 * The case we do have to guard against is when memory pressure made
587 * shmem_writepage move it from filecache to swapcache beneath us:
588 * an unlikely race, but we do need to retry for page->mapping.
590 if (unlikely(!mapping)) {
594 * Page lock is required to identify which special case above
595 * applies. If this is really a shmem page then the page lock
596 * will prevent unexpected transitions.
599 shmem_swizzled = PageSwapCache(page) || page->mapping;
610 * Private mappings are handled in a simple way.
612 * If the futex key is stored on an anonymous page, then the associated
613 * object is the mm which is implicitly pinned by the calling process.
615 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
616 * it's a read-only handle, it's expected that futexes attach to
617 * the object not the particular process.
619 if (PageAnon(page)) {
621 * A RO anonymous page will never change and thus doesn't make
622 * sense for futex operations.
624 if (unlikely(should_fail_futex(fshared)) || ro) {
629 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
630 key->private.mm = mm;
631 key->private.address = address;
633 get_futex_key_refs(key); /* implies smp_mb(); (B) */
639 * The associated futex object in this case is the inode and
640 * the page->mapping must be traversed. Ordinarily this should
641 * be stabilised under page lock but it's not strictly
642 * necessary in this case as we just want to pin the inode, not
643 * update the radix tree or anything like that.
645 * The RCU read lock is taken as the inode is finally freed
646 * under RCU. If the mapping still matches expectations then the
647 * mapping->host can be safely accessed as being a valid inode.
651 if (READ_ONCE(page->mapping) != mapping) {
658 inode = READ_ONCE(mapping->host);
667 * Take a reference unless it is about to be freed. Previously
668 * this reference was taken by ihold under the page lock
669 * pinning the inode in place so i_lock was unnecessary. The
670 * only way for this check to fail is if the inode was
671 * truncated in parallel so warn for now if this happens.
673 * We are not calling into get_futex_key_refs() in file-backed
674 * cases, therefore a successful atomic_inc return below will
675 * guarantee that get_futex_key() will still imply smp_mb(); (B).
677 if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
684 /* Should be impossible but lets be paranoid for now */
685 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
693 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
694 key->shared.inode = inode;
695 key->shared.pgoff = basepage_index(tail);
704 static inline void put_futex_key(union futex_key *key)
706 drop_futex_key_refs(key);
710 * fault_in_user_writeable() - Fault in user address and verify RW access
711 * @uaddr: pointer to faulting user space address
713 * Slow path to fixup the fault we just took in the atomic write
716 * We have no generic implementation of a non-destructive write to the
717 * user address. We know that we faulted in the atomic pagefault
718 * disabled section so we can as well avoid the #PF overhead by
719 * calling get_user_pages() right away.
721 static int fault_in_user_writeable(u32 __user *uaddr)
723 struct mm_struct *mm = current->mm;
726 down_read(&mm->mmap_sem);
727 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
728 FAULT_FLAG_WRITE, NULL);
729 up_read(&mm->mmap_sem);
731 return ret < 0 ? ret : 0;
735 * futex_top_waiter() - Return the highest priority waiter on a futex
736 * @hb: the hash bucket the futex_q's reside in
737 * @key: the futex key (to distinguish it from other futex futex_q's)
739 * Must be called with the hb lock held.
741 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
742 union futex_key *key)
744 struct futex_q *this;
746 plist_for_each_entry(this, &hb->chain, list) {
747 if (match_futex(&this->key, key))
753 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
754 u32 uval, u32 newval)
759 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
765 static int get_futex_value_locked(u32 *dest, u32 __user *from)
770 ret = __get_user(*dest, from);
773 return ret ? -EFAULT : 0;
780 static int refill_pi_state_cache(void)
782 struct futex_pi_state *pi_state;
784 if (likely(current->pi_state_cache))
787 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
792 INIT_LIST_HEAD(&pi_state->list);
793 /* pi_mutex gets initialized later */
794 pi_state->owner = NULL;
795 atomic_set(&pi_state->refcount, 1);
796 pi_state->key = FUTEX_KEY_INIT;
798 current->pi_state_cache = pi_state;
803 static struct futex_pi_state * alloc_pi_state(void)
805 struct futex_pi_state *pi_state = current->pi_state_cache;
808 current->pi_state_cache = NULL;
814 * Drops a reference to the pi_state object and frees or caches it
815 * when the last reference is gone.
817 * Must be called with the hb lock held.
819 static void put_pi_state(struct futex_pi_state *pi_state)
824 if (!atomic_dec_and_test(&pi_state->refcount))
828 * If pi_state->owner is NULL, the owner is most probably dying
829 * and has cleaned up the pi_state already
831 if (pi_state->owner) {
832 raw_spin_lock_irq(&pi_state->owner->pi_lock);
833 list_del_init(&pi_state->list);
834 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
836 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
839 if (current->pi_state_cache)
843 * pi_state->list is already empty.
844 * clear pi_state->owner.
845 * refcount is at 0 - put it back to 1.
847 pi_state->owner = NULL;
848 atomic_set(&pi_state->refcount, 1);
849 current->pi_state_cache = pi_state;
854 * Look up the task based on what TID userspace gave us.
857 static struct task_struct * futex_find_get_task(pid_t pid)
859 struct task_struct *p;
862 p = find_task_by_vpid(pid);
872 * This task is holding PI mutexes at exit time => bad.
873 * Kernel cleans up PI-state, but userspace is likely hosed.
874 * (Robust-futex cleanup is separate and might save the day for userspace.)
876 void exit_pi_state_list(struct task_struct *curr)
878 struct list_head *next, *head = &curr->pi_state_list;
879 struct futex_pi_state *pi_state;
880 struct futex_hash_bucket *hb;
881 union futex_key key = FUTEX_KEY_INIT;
883 if (!futex_cmpxchg_enabled)
886 * We are a ZOMBIE and nobody can enqueue itself on
887 * pi_state_list anymore, but we have to be careful
888 * versus waiters unqueueing themselves:
890 raw_spin_lock_irq(&curr->pi_lock);
891 while (!list_empty(head)) {
894 pi_state = list_entry(next, struct futex_pi_state, list);
896 hb = hash_futex(&key);
897 raw_spin_unlock_irq(&curr->pi_lock);
899 spin_lock(&hb->lock);
901 raw_spin_lock_irq(&curr->pi_lock);
903 * We dropped the pi-lock, so re-check whether this
904 * task still owns the PI-state:
906 if (head->next != next) {
907 spin_unlock(&hb->lock);
911 WARN_ON(pi_state->owner != curr);
912 WARN_ON(list_empty(&pi_state->list));
913 list_del_init(&pi_state->list);
914 pi_state->owner = NULL;
915 raw_spin_unlock_irq(&curr->pi_lock);
917 rt_mutex_unlock(&pi_state->pi_mutex);
919 spin_unlock(&hb->lock);
921 raw_spin_lock_irq(&curr->pi_lock);
923 raw_spin_unlock_irq(&curr->pi_lock);
927 * We need to check the following states:
929 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
931 * [1] NULL | --- | --- | 0 | 0/1 | Valid
932 * [2] NULL | --- | --- | >0 | 0/1 | Valid
934 * [3] Found | NULL | -- | Any | 0/1 | Invalid
936 * [4] Found | Found | NULL | 0 | 1 | Valid
937 * [5] Found | Found | NULL | >0 | 1 | Invalid
939 * [6] Found | Found | task | 0 | 1 | Valid
941 * [7] Found | Found | NULL | Any | 0 | Invalid
943 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
944 * [9] Found | Found | task | 0 | 0 | Invalid
945 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
947 * [1] Indicates that the kernel can acquire the futex atomically. We
948 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
950 * [2] Valid, if TID does not belong to a kernel thread. If no matching
951 * thread is found then it indicates that the owner TID has died.
953 * [3] Invalid. The waiter is queued on a non PI futex
955 * [4] Valid state after exit_robust_list(), which sets the user space
956 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
958 * [5] The user space value got manipulated between exit_robust_list()
959 * and exit_pi_state_list()
961 * [6] Valid state after exit_pi_state_list() which sets the new owner in
962 * the pi_state but cannot access the user space value.
964 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
966 * [8] Owner and user space value match
968 * [9] There is no transient state which sets the user space TID to 0
969 * except exit_robust_list(), but this is indicated by the
970 * FUTEX_OWNER_DIED bit. See [4]
972 * [10] There is no transient state which leaves owner and user space
977 * Validate that the existing waiter has a pi_state and sanity check
978 * the pi_state against the user space value. If correct, attach to
981 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
982 struct futex_pi_state **ps)
984 pid_t pid = uval & FUTEX_TID_MASK;
987 * Userspace might have messed up non-PI and PI futexes [3]
989 if (unlikely(!pi_state))
992 WARN_ON(!atomic_read(&pi_state->refcount));
995 * Handle the owner died case:
997 if (uval & FUTEX_OWNER_DIED) {
999 * exit_pi_state_list sets owner to NULL and wakes the
1000 * topmost waiter. The task which acquires the
1001 * pi_state->rt_mutex will fixup owner.
1003 if (!pi_state->owner) {
1005 * No pi state owner, but the user space TID
1006 * is not 0. Inconsistent state. [5]
1011 * Take a ref on the state and return success. [4]
1017 * If TID is 0, then either the dying owner has not
1018 * yet executed exit_pi_state_list() or some waiter
1019 * acquired the rtmutex in the pi state, but did not
1020 * yet fixup the TID in user space.
1022 * Take a ref on the state and return success. [6]
1028 * If the owner died bit is not set, then the pi_state
1029 * must have an owner. [7]
1031 if (!pi_state->owner)
1036 * Bail out if user space manipulated the futex value. If pi
1037 * state exists then the owner TID must be the same as the
1038 * user space TID. [9/10]
1040 if (pid != task_pid_vnr(pi_state->owner))
1043 atomic_inc(&pi_state->refcount);
1049 * Lookup the task for the TID provided from user space and attach to
1050 * it after doing proper sanity checks.
1052 static int attach_to_pi_owner(u32 uval, union futex_key *key,
1053 struct futex_pi_state **ps)
1055 pid_t pid = uval & FUTEX_TID_MASK;
1056 struct futex_pi_state *pi_state;
1057 struct task_struct *p;
1060 * We are the first waiter - try to look up the real owner and attach
1061 * the new pi_state to it, but bail out when TID = 0 [1]
1065 p = futex_find_get_task(pid);
1069 if (unlikely(p->flags & PF_KTHREAD)) {
1075 * We need to look at the task state flags to figure out,
1076 * whether the task is exiting. To protect against the do_exit
1077 * change of the task flags, we do this protected by
1080 raw_spin_lock_irq(&p->pi_lock);
1081 if (unlikely(p->flags & PF_EXITING)) {
1083 * The task is on the way out. When PF_EXITPIDONE is
1084 * set, we know that the task has finished the
1087 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1089 raw_spin_unlock_irq(&p->pi_lock);
1095 * No existing pi state. First waiter. [2]
1097 pi_state = alloc_pi_state();
1100 * Initialize the pi_mutex in locked state and make @p
1103 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1105 /* Store the key for possible exit cleanups: */
1106 pi_state->key = *key;
1108 WARN_ON(!list_empty(&pi_state->list));
1109 list_add(&pi_state->list, &p->pi_state_list);
1110 pi_state->owner = p;
1111 raw_spin_unlock_irq(&p->pi_lock);
1120 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1121 union futex_key *key, struct futex_pi_state **ps)
1123 struct futex_q *match = futex_top_waiter(hb, key);
1126 * If there is a waiter on that futex, validate it and
1127 * attach to the pi_state when the validation succeeds.
1130 return attach_to_pi_state(uval, match->pi_state, ps);
1133 * We are the first waiter - try to look up the owner based on
1134 * @uval and attach to it.
1136 return attach_to_pi_owner(uval, key, ps);
1139 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1141 u32 uninitialized_var(curval);
1143 if (unlikely(should_fail_futex(true)))
1146 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1149 /*If user space value changed, let the caller retry */
1150 return curval != uval ? -EAGAIN : 0;
1154 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1155 * @uaddr: the pi futex user address
1156 * @hb: the pi futex hash bucket
1157 * @key: the futex key associated with uaddr and hb
1158 * @ps: the pi_state pointer where we store the result of the
1160 * @task: the task to perform the atomic lock work for. This will
1161 * be "current" except in the case of requeue pi.
1162 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1165 * 0 - ready to wait;
1166 * 1 - acquired the lock;
1169 * The hb->lock and futex_key refs shall be held by the caller.
1171 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1172 union futex_key *key,
1173 struct futex_pi_state **ps,
1174 struct task_struct *task, int set_waiters)
1176 u32 uval, newval, vpid = task_pid_vnr(task);
1177 struct futex_q *match;
1181 * Read the user space value first so we can validate a few
1182 * things before proceeding further.
1184 if (get_futex_value_locked(&uval, uaddr))
1187 if (unlikely(should_fail_futex(true)))
1193 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1196 if ((unlikely(should_fail_futex(true))))
1200 * Lookup existing state first. If it exists, try to attach to
1203 match = futex_top_waiter(hb, key);
1205 return attach_to_pi_state(uval, match->pi_state, ps);
1208 * No waiter and user TID is 0. We are here because the
1209 * waiters or the owner died bit is set or called from
1210 * requeue_cmp_pi or for whatever reason something took the
1213 if (!(uval & FUTEX_TID_MASK)) {
1215 * We take over the futex. No other waiters and the user space
1216 * TID is 0. We preserve the owner died bit.
1218 newval = uval & FUTEX_OWNER_DIED;
1221 /* The futex requeue_pi code can enforce the waiters bit */
1223 newval |= FUTEX_WAITERS;
1225 ret = lock_pi_update_atomic(uaddr, uval, newval);
1226 /* If the take over worked, return 1 */
1227 return ret < 0 ? ret : 1;
1231 * First waiter. Set the waiters bit before attaching ourself to
1232 * the owner. If owner tries to unlock, it will be forced into
1233 * the kernel and blocked on hb->lock.
1235 newval = uval | FUTEX_WAITERS;
1236 ret = lock_pi_update_atomic(uaddr, uval, newval);
1240 * If the update of the user space value succeeded, we try to
1241 * attach to the owner. If that fails, no harm done, we only
1242 * set the FUTEX_WAITERS bit in the user space variable.
1244 return attach_to_pi_owner(uval, key, ps);
1248 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1249 * @q: The futex_q to unqueue
1251 * The q->lock_ptr must not be NULL and must be held by the caller.
1253 static void __unqueue_futex(struct futex_q *q)
1255 struct futex_hash_bucket *hb;
1257 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1258 || WARN_ON(plist_node_empty(&q->list)))
1261 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1262 plist_del(&q->list, &hb->chain);
1267 * The hash bucket lock must be held when this is called.
1268 * Afterwards, the futex_q must not be accessed. Callers
1269 * must ensure to later call wake_up_q() for the actual
1272 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1274 struct task_struct *p = q->task;
1276 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1280 * Queue the task for later wakeup for after we've released
1281 * the hb->lock. wake_q_add() grabs reference to p.
1283 wake_q_add(wake_q, p);
1286 * The waiting task can free the futex_q as soon as
1287 * q->lock_ptr = NULL is written, without taking any locks. A
1288 * memory barrier is required here to prevent the following
1289 * store to lock_ptr from getting ahead of the plist_del.
1295 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1296 struct futex_hash_bucket *hb)
1298 struct task_struct *new_owner;
1299 struct futex_pi_state *pi_state = this->pi_state;
1300 u32 uninitialized_var(curval), newval;
1309 * If current does not own the pi_state then the futex is
1310 * inconsistent and user space fiddled with the futex value.
1312 if (pi_state->owner != current)
1315 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1316 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1319 * It is possible that the next waiter (the one that brought
1320 * this owner to the kernel) timed out and is no longer
1321 * waiting on the lock.
1324 new_owner = this->task;
1327 * We pass it to the next owner. The WAITERS bit is always
1328 * kept enabled while there is PI state around. We cleanup the
1329 * owner died bit, because we are the owner.
1331 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1333 if (unlikely(should_fail_futex(true)))
1336 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1338 } else if (curval != uval) {
1340 * If a unconditional UNLOCK_PI operation (user space did not
1341 * try the TID->0 transition) raced with a waiter setting the
1342 * FUTEX_WAITERS flag between get_user() and locking the hash
1343 * bucket lock, retry the operation.
1345 if ((FUTEX_TID_MASK & curval) == uval)
1351 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1355 raw_spin_lock(&pi_state->owner->pi_lock);
1356 WARN_ON(list_empty(&pi_state->list));
1357 list_del_init(&pi_state->list);
1358 raw_spin_unlock(&pi_state->owner->pi_lock);
1360 raw_spin_lock(&new_owner->pi_lock);
1361 WARN_ON(!list_empty(&pi_state->list));
1362 list_add(&pi_state->list, &new_owner->pi_state_list);
1363 pi_state->owner = new_owner;
1364 raw_spin_unlock(&new_owner->pi_lock);
1366 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1368 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1371 * First unlock HB so the waiter does not spin on it once he got woken
1372 * up. Second wake up the waiter before the priority is adjusted. If we
1373 * deboost first (and lose our higher priority), then the task might get
1374 * scheduled away before the wake up can take place.
1376 spin_unlock(&hb->lock);
1379 rt_mutex_adjust_prio(current);
1385 * Express the locking dependencies for lockdep:
1388 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1391 spin_lock(&hb1->lock);
1393 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1394 } else { /* hb1 > hb2 */
1395 spin_lock(&hb2->lock);
1396 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1401 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1403 spin_unlock(&hb1->lock);
1405 spin_unlock(&hb2->lock);
1409 * Wake up waiters matching bitset queued on this futex (uaddr).
1412 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1414 struct futex_hash_bucket *hb;
1415 struct futex_q *this, *next;
1416 union futex_key key = FUTEX_KEY_INIT;
1423 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1424 if (unlikely(ret != 0))
1427 hb = hash_futex(&key);
1429 /* Make sure we really have tasks to wakeup */
1430 if (!hb_waiters_pending(hb))
1433 spin_lock(&hb->lock);
1435 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1436 if (match_futex (&this->key, &key)) {
1437 if (this->pi_state || this->rt_waiter) {
1442 /* Check if one of the bits is set in both bitsets */
1443 if (!(this->bitset & bitset))
1446 mark_wake_futex(&wake_q, this);
1447 if (++ret >= nr_wake)
1452 spin_unlock(&hb->lock);
1455 put_futex_key(&key);
1461 * Wake up all waiters hashed on the physical page that is mapped
1462 * to this virtual address:
1465 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1466 int nr_wake, int nr_wake2, int op)
1468 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1469 struct futex_hash_bucket *hb1, *hb2;
1470 struct futex_q *this, *next;
1475 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1476 if (unlikely(ret != 0))
1478 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1479 if (unlikely(ret != 0))
1482 hb1 = hash_futex(&key1);
1483 hb2 = hash_futex(&key2);
1486 double_lock_hb(hb1, hb2);
1487 op_ret = futex_atomic_op_inuser(op, uaddr2);
1488 if (unlikely(op_ret < 0)) {
1490 double_unlock_hb(hb1, hb2);
1494 * we don't get EFAULT from MMU faults if we don't have an MMU,
1495 * but we might get them from range checking
1501 if (unlikely(op_ret != -EFAULT)) {
1506 ret = fault_in_user_writeable(uaddr2);
1510 if (!(flags & FLAGS_SHARED))
1513 put_futex_key(&key2);
1514 put_futex_key(&key1);
1518 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1519 if (match_futex (&this->key, &key1)) {
1520 if (this->pi_state || this->rt_waiter) {
1524 mark_wake_futex(&wake_q, this);
1525 if (++ret >= nr_wake)
1532 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1533 if (match_futex (&this->key, &key2)) {
1534 if (this->pi_state || this->rt_waiter) {
1538 mark_wake_futex(&wake_q, this);
1539 if (++op_ret >= nr_wake2)
1547 double_unlock_hb(hb1, hb2);
1550 put_futex_key(&key2);
1552 put_futex_key(&key1);
1558 * requeue_futex() - Requeue a futex_q from one hb to another
1559 * @q: the futex_q to requeue
1560 * @hb1: the source hash_bucket
1561 * @hb2: the target hash_bucket
1562 * @key2: the new key for the requeued futex_q
1565 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1566 struct futex_hash_bucket *hb2, union futex_key *key2)
1570 * If key1 and key2 hash to the same bucket, no need to
1573 if (likely(&hb1->chain != &hb2->chain)) {
1574 plist_del(&q->list, &hb1->chain);
1575 hb_waiters_dec(hb1);
1576 hb_waiters_inc(hb2);
1577 plist_add(&q->list, &hb2->chain);
1578 q->lock_ptr = &hb2->lock;
1580 get_futex_key_refs(key2);
1585 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1587 * @key: the key of the requeue target futex
1588 * @hb: the hash_bucket of the requeue target futex
1590 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1591 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1592 * to the requeue target futex so the waiter can detect the wakeup on the right
1593 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1594 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1595 * to protect access to the pi_state to fixup the owner later. Must be called
1596 * with both q->lock_ptr and hb->lock held.
1599 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1600 struct futex_hash_bucket *hb)
1602 get_futex_key_refs(key);
1607 WARN_ON(!q->rt_waiter);
1608 q->rt_waiter = NULL;
1610 q->lock_ptr = &hb->lock;
1612 wake_up_state(q->task, TASK_NORMAL);
1616 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1617 * @pifutex: the user address of the to futex
1618 * @hb1: the from futex hash bucket, must be locked by the caller
1619 * @hb2: the to futex hash bucket, must be locked by the caller
1620 * @key1: the from futex key
1621 * @key2: the to futex key
1622 * @ps: address to store the pi_state pointer
1623 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1625 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1626 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1627 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1628 * hb1 and hb2 must be held by the caller.
1631 * 0 - failed to acquire the lock atomically;
1632 * >0 - acquired the lock, return value is vpid of the top_waiter
1635 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1636 struct futex_hash_bucket *hb1,
1637 struct futex_hash_bucket *hb2,
1638 union futex_key *key1, union futex_key *key2,
1639 struct futex_pi_state **ps, int set_waiters)
1641 struct futex_q *top_waiter = NULL;
1645 if (get_futex_value_locked(&curval, pifutex))
1648 if (unlikely(should_fail_futex(true)))
1652 * Find the top_waiter and determine if there are additional waiters.
1653 * If the caller intends to requeue more than 1 waiter to pifutex,
1654 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1655 * as we have means to handle the possible fault. If not, don't set
1656 * the bit unecessarily as it will force the subsequent unlock to enter
1659 top_waiter = futex_top_waiter(hb1, key1);
1661 /* There are no waiters, nothing for us to do. */
1665 /* Ensure we requeue to the expected futex. */
1666 if (!match_futex(top_waiter->requeue_pi_key, key2))
1670 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1671 * the contended case or if set_waiters is 1. The pi_state is returned
1672 * in ps in contended cases.
1674 vpid = task_pid_vnr(top_waiter->task);
1675 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1678 requeue_pi_wake_futex(top_waiter, key2, hb2);
1685 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1686 * @uaddr1: source futex user address
1687 * @flags: futex flags (FLAGS_SHARED, etc.)
1688 * @uaddr2: target futex user address
1689 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1690 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1691 * @cmpval: @uaddr1 expected value (or %NULL)
1692 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1693 * pi futex (pi to pi requeue is not supported)
1695 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1696 * uaddr2 atomically on behalf of the top waiter.
1699 * >=0 - on success, the number of tasks requeued or woken;
1702 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1703 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1704 u32 *cmpval, int requeue_pi)
1706 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1707 int drop_count = 0, task_count = 0, ret;
1708 struct futex_pi_state *pi_state = NULL;
1709 struct futex_hash_bucket *hb1, *hb2;
1710 struct futex_q *this, *next;
1715 * Requeue PI only works on two distinct uaddrs. This
1716 * check is only valid for private futexes. See below.
1718 if (uaddr1 == uaddr2)
1722 * requeue_pi requires a pi_state, try to allocate it now
1723 * without any locks in case it fails.
1725 if (refill_pi_state_cache())
1728 * requeue_pi must wake as many tasks as it can, up to nr_wake
1729 * + nr_requeue, since it acquires the rt_mutex prior to
1730 * returning to userspace, so as to not leave the rt_mutex with
1731 * waiters and no owner. However, second and third wake-ups
1732 * cannot be predicted as they involve race conditions with the
1733 * first wake and a fault while looking up the pi_state. Both
1734 * pthread_cond_signal() and pthread_cond_broadcast() should
1742 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1743 if (unlikely(ret != 0))
1745 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1746 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1747 if (unlikely(ret != 0))
1751 * The check above which compares uaddrs is not sufficient for
1752 * shared futexes. We need to compare the keys:
1754 if (requeue_pi && match_futex(&key1, &key2)) {
1759 hb1 = hash_futex(&key1);
1760 hb2 = hash_futex(&key2);
1763 hb_waiters_inc(hb2);
1764 double_lock_hb(hb1, hb2);
1766 if (likely(cmpval != NULL)) {
1769 ret = get_futex_value_locked(&curval, uaddr1);
1771 if (unlikely(ret)) {
1772 double_unlock_hb(hb1, hb2);
1773 hb_waiters_dec(hb2);
1775 ret = get_user(curval, uaddr1);
1779 if (!(flags & FLAGS_SHARED))
1782 put_futex_key(&key2);
1783 put_futex_key(&key1);
1786 if (curval != *cmpval) {
1792 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1794 * Attempt to acquire uaddr2 and wake the top waiter. If we
1795 * intend to requeue waiters, force setting the FUTEX_WAITERS
1796 * bit. We force this here where we are able to easily handle
1797 * faults rather in the requeue loop below.
1799 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1800 &key2, &pi_state, nr_requeue);
1803 * At this point the top_waiter has either taken uaddr2 or is
1804 * waiting on it. If the former, then the pi_state will not
1805 * exist yet, look it up one more time to ensure we have a
1806 * reference to it. If the lock was taken, ret contains the
1807 * vpid of the top waiter task.
1808 * If the lock was not taken, we have pi_state and an initial
1809 * refcount on it. In case of an error we have nothing.
1816 * If we acquired the lock, then the user space value
1817 * of uaddr2 should be vpid. It cannot be changed by
1818 * the top waiter as it is blocked on hb2 lock if it
1819 * tries to do so. If something fiddled with it behind
1820 * our back the pi state lookup might unearth it. So
1821 * we rather use the known value than rereading and
1822 * handing potential crap to lookup_pi_state.
1824 * If that call succeeds then we have pi_state and an
1825 * initial refcount on it.
1827 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1832 /* We hold a reference on the pi state. */
1835 /* If the above failed, then pi_state is NULL */
1837 double_unlock_hb(hb1, hb2);
1838 hb_waiters_dec(hb2);
1839 put_futex_key(&key2);
1840 put_futex_key(&key1);
1841 ret = fault_in_user_writeable(uaddr2);
1847 * Two reasons for this:
1848 * - Owner is exiting and we just wait for the
1850 * - The user space value changed.
1852 double_unlock_hb(hb1, hb2);
1853 hb_waiters_dec(hb2);
1854 put_futex_key(&key2);
1855 put_futex_key(&key1);
1863 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1864 if (task_count - nr_wake >= nr_requeue)
1867 if (!match_futex(&this->key, &key1))
1871 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1872 * be paired with each other and no other futex ops.
1874 * We should never be requeueing a futex_q with a pi_state,
1875 * which is awaiting a futex_unlock_pi().
1877 if ((requeue_pi && !this->rt_waiter) ||
1878 (!requeue_pi && this->rt_waiter) ||
1885 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1886 * lock, we already woke the top_waiter. If not, it will be
1887 * woken by futex_unlock_pi().
1889 if (++task_count <= nr_wake && !requeue_pi) {
1890 mark_wake_futex(&wake_q, this);
1894 /* Ensure we requeue to the expected futex for requeue_pi. */
1895 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1901 * Requeue nr_requeue waiters and possibly one more in the case
1902 * of requeue_pi if we couldn't acquire the lock atomically.
1906 * Prepare the waiter to take the rt_mutex. Take a
1907 * refcount on the pi_state and store the pointer in
1908 * the futex_q object of the waiter.
1910 atomic_inc(&pi_state->refcount);
1911 this->pi_state = pi_state;
1912 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1917 * We got the lock. We do neither drop the
1918 * refcount on pi_state nor clear
1919 * this->pi_state because the waiter needs the
1920 * pi_state for cleaning up the user space
1921 * value. It will drop the refcount after
1924 requeue_pi_wake_futex(this, &key2, hb2);
1929 * rt_mutex_start_proxy_lock() detected a
1930 * potential deadlock when we tried to queue
1931 * that waiter. Drop the pi_state reference
1932 * which we took above and remove the pointer
1933 * to the state from the waiters futex_q
1936 this->pi_state = NULL;
1937 put_pi_state(pi_state);
1939 * We stop queueing more waiters and let user
1940 * space deal with the mess.
1945 requeue_futex(this, hb1, hb2, &key2);
1950 * We took an extra initial reference to the pi_state either
1951 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
1952 * need to drop it here again.
1954 put_pi_state(pi_state);
1957 double_unlock_hb(hb1, hb2);
1959 hb_waiters_dec(hb2);
1962 * drop_futex_key_refs() must be called outside the spinlocks. During
1963 * the requeue we moved futex_q's from the hash bucket at key1 to the
1964 * one at key2 and updated their key pointer. We no longer need to
1965 * hold the references to key1.
1967 while (--drop_count >= 0)
1968 drop_futex_key_refs(&key1);
1971 put_futex_key(&key2);
1973 put_futex_key(&key1);
1975 return ret ? ret : task_count;
1978 /* The key must be already stored in q->key. */
1979 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1980 __acquires(&hb->lock)
1982 struct futex_hash_bucket *hb;
1984 hb = hash_futex(&q->key);
1987 * Increment the counter before taking the lock so that
1988 * a potential waker won't miss a to-be-slept task that is
1989 * waiting for the spinlock. This is safe as all queue_lock()
1990 * users end up calling queue_me(). Similarly, for housekeeping,
1991 * decrement the counter at queue_unlock() when some error has
1992 * occurred and we don't end up adding the task to the list.
1996 q->lock_ptr = &hb->lock;
1998 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
2003 queue_unlock(struct futex_hash_bucket *hb)
2004 __releases(&hb->lock)
2006 spin_unlock(&hb->lock);
2011 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2012 * @q: The futex_q to enqueue
2013 * @hb: The destination hash bucket
2015 * The hb->lock must be held by the caller, and is released here. A call to
2016 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2017 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2018 * or nothing if the unqueue is done as part of the wake process and the unqueue
2019 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2022 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2023 __releases(&hb->lock)
2028 * The priority used to register this element is
2029 * - either the real thread-priority for the real-time threads
2030 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2031 * - or MAX_RT_PRIO for non-RT threads.
2032 * Thus, all RT-threads are woken first in priority order, and
2033 * the others are woken last, in FIFO order.
2035 prio = min(current->normal_prio, MAX_RT_PRIO);
2037 plist_node_init(&q->list, prio);
2038 plist_add(&q->list, &hb->chain);
2040 spin_unlock(&hb->lock);
2044 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2045 * @q: The futex_q to unqueue
2047 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2048 * be paired with exactly one earlier call to queue_me().
2051 * 1 - if the futex_q was still queued (and we removed unqueued it);
2052 * 0 - if the futex_q was already removed by the waking thread
2054 static int unqueue_me(struct futex_q *q)
2056 spinlock_t *lock_ptr;
2059 /* In the common case we don't take the spinlock, which is nice. */
2062 * q->lock_ptr can change between this read and the following spin_lock.
2063 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2064 * optimizing lock_ptr out of the logic below.
2066 lock_ptr = READ_ONCE(q->lock_ptr);
2067 if (lock_ptr != NULL) {
2068 spin_lock(lock_ptr);
2070 * q->lock_ptr can change between reading it and
2071 * spin_lock(), causing us to take the wrong lock. This
2072 * corrects the race condition.
2074 * Reasoning goes like this: if we have the wrong lock,
2075 * q->lock_ptr must have changed (maybe several times)
2076 * between reading it and the spin_lock(). It can
2077 * change again after the spin_lock() but only if it was
2078 * already changed before the spin_lock(). It cannot,
2079 * however, change back to the original value. Therefore
2080 * we can detect whether we acquired the correct lock.
2082 if (unlikely(lock_ptr != q->lock_ptr)) {
2083 spin_unlock(lock_ptr);
2088 BUG_ON(q->pi_state);
2090 spin_unlock(lock_ptr);
2094 drop_futex_key_refs(&q->key);
2099 * PI futexes can not be requeued and must remove themself from the
2100 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2103 static void unqueue_me_pi(struct futex_q *q)
2104 __releases(q->lock_ptr)
2108 BUG_ON(!q->pi_state);
2109 put_pi_state(q->pi_state);
2112 spin_unlock(q->lock_ptr);
2116 * Fixup the pi_state owner with the new owner.
2118 * Must be called with hash bucket lock held and mm->sem held for non
2121 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2122 struct task_struct *newowner)
2124 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2125 struct futex_pi_state *pi_state = q->pi_state;
2126 struct task_struct *oldowner = pi_state->owner;
2127 u32 uval, uninitialized_var(curval), newval;
2131 if (!pi_state->owner)
2132 newtid |= FUTEX_OWNER_DIED;
2135 * We are here either because we stole the rtmutex from the
2136 * previous highest priority waiter or we are the highest priority
2137 * waiter but failed to get the rtmutex the first time.
2138 * We have to replace the newowner TID in the user space variable.
2139 * This must be atomic as we have to preserve the owner died bit here.
2141 * Note: We write the user space value _before_ changing the pi_state
2142 * because we can fault here. Imagine swapped out pages or a fork
2143 * that marked all the anonymous memory readonly for cow.
2145 * Modifying pi_state _before_ the user space value would
2146 * leave the pi_state in an inconsistent state when we fault
2147 * here, because we need to drop the hash bucket lock to
2148 * handle the fault. This might be observed in the PID check
2149 * in lookup_pi_state.
2152 if (get_futex_value_locked(&uval, uaddr))
2156 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2158 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2166 * We fixed up user space. Now we need to fix the pi_state
2169 if (pi_state->owner != NULL) {
2170 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2171 WARN_ON(list_empty(&pi_state->list));
2172 list_del_init(&pi_state->list);
2173 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2176 pi_state->owner = newowner;
2178 raw_spin_lock_irq(&newowner->pi_lock);
2179 WARN_ON(!list_empty(&pi_state->list));
2180 list_add(&pi_state->list, &newowner->pi_state_list);
2181 raw_spin_unlock_irq(&newowner->pi_lock);
2185 * To handle the page fault we need to drop the hash bucket
2186 * lock here. That gives the other task (either the highest priority
2187 * waiter itself or the task which stole the rtmutex) the
2188 * chance to try the fixup of the pi_state. So once we are
2189 * back from handling the fault we need to check the pi_state
2190 * after reacquiring the hash bucket lock and before trying to
2191 * do another fixup. When the fixup has been done already we
2195 spin_unlock(q->lock_ptr);
2197 ret = fault_in_user_writeable(uaddr);
2199 spin_lock(q->lock_ptr);
2202 * Check if someone else fixed it for us:
2204 if (pi_state->owner != oldowner)
2213 static long futex_wait_restart(struct restart_block *restart);
2216 * fixup_owner() - Post lock pi_state and corner case management
2217 * @uaddr: user address of the futex
2218 * @q: futex_q (contains pi_state and access to the rt_mutex)
2219 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2221 * After attempting to lock an rt_mutex, this function is called to cleanup
2222 * the pi_state owner as well as handle race conditions that may allow us to
2223 * acquire the lock. Must be called with the hb lock held.
2226 * 1 - success, lock taken;
2227 * 0 - success, lock not taken;
2228 * <0 - on error (-EFAULT)
2230 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2232 struct task_struct *owner;
2237 * Got the lock. We might not be the anticipated owner if we
2238 * did a lock-steal - fix up the PI-state in that case:
2240 if (q->pi_state->owner != current)
2241 ret = fixup_pi_state_owner(uaddr, q, current);
2246 * Catch the rare case, where the lock was released when we were on the
2247 * way back before we locked the hash bucket.
2249 if (q->pi_state->owner == current) {
2251 * Try to get the rt_mutex now. This might fail as some other
2252 * task acquired the rt_mutex after we removed ourself from the
2253 * rt_mutex waiters list.
2255 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2261 * pi_state is incorrect, some other task did a lock steal and
2262 * we returned due to timeout or signal without taking the
2263 * rt_mutex. Too late.
2265 raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2266 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2268 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2269 raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2270 ret = fixup_pi_state_owner(uaddr, q, owner);
2275 * Paranoia check. If we did not take the lock, then we should not be
2276 * the owner of the rt_mutex.
2278 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2279 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2280 "pi-state %p\n", ret,
2281 q->pi_state->pi_mutex.owner,
2282 q->pi_state->owner);
2285 return ret ? ret : locked;
2289 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2290 * @hb: the futex hash bucket, must be locked by the caller
2291 * @q: the futex_q to queue up on
2292 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2294 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2295 struct hrtimer_sleeper *timeout)
2298 * The task state is guaranteed to be set before another task can
2299 * wake it. set_current_state() is implemented using smp_store_mb() and
2300 * queue_me() calls spin_unlock() upon completion, both serializing
2301 * access to the hash list and forcing another memory barrier.
2303 set_current_state(TASK_INTERRUPTIBLE);
2308 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2311 * If we have been removed from the hash list, then another task
2312 * has tried to wake us, and we can skip the call to schedule().
2314 if (likely(!plist_node_empty(&q->list))) {
2316 * If the timer has already expired, current will already be
2317 * flagged for rescheduling. Only call schedule if there
2318 * is no timeout, or if it has yet to expire.
2320 if (!timeout || timeout->task)
2321 freezable_schedule();
2323 __set_current_state(TASK_RUNNING);
2327 * futex_wait_setup() - Prepare to wait on a futex
2328 * @uaddr: the futex userspace address
2329 * @val: the expected value
2330 * @flags: futex flags (FLAGS_SHARED, etc.)
2331 * @q: the associated futex_q
2332 * @hb: storage for hash_bucket pointer to be returned to caller
2334 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2335 * compare it with the expected value. Handle atomic faults internally.
2336 * Return with the hb lock held and a q.key reference on success, and unlocked
2337 * with no q.key reference on failure.
2340 * 0 - uaddr contains val and hb has been locked;
2341 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2343 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2344 struct futex_q *q, struct futex_hash_bucket **hb)
2350 * Access the page AFTER the hash-bucket is locked.
2351 * Order is important:
2353 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2354 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2356 * The basic logical guarantee of a futex is that it blocks ONLY
2357 * if cond(var) is known to be true at the time of blocking, for
2358 * any cond. If we locked the hash-bucket after testing *uaddr, that
2359 * would open a race condition where we could block indefinitely with
2360 * cond(var) false, which would violate the guarantee.
2362 * On the other hand, we insert q and release the hash-bucket only
2363 * after testing *uaddr. This guarantees that futex_wait() will NOT
2364 * absorb a wakeup if *uaddr does not match the desired values
2365 * while the syscall executes.
2368 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2369 if (unlikely(ret != 0))
2373 *hb = queue_lock(q);
2375 ret = get_futex_value_locked(&uval, uaddr);
2380 ret = get_user(uval, uaddr);
2384 if (!(flags & FLAGS_SHARED))
2387 put_futex_key(&q->key);
2398 put_futex_key(&q->key);
2402 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2403 ktime_t *abs_time, u32 bitset)
2405 struct hrtimer_sleeper timeout, *to = NULL;
2406 struct restart_block *restart;
2407 struct futex_hash_bucket *hb;
2408 struct futex_q q = futex_q_init;
2418 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2419 CLOCK_REALTIME : CLOCK_MONOTONIC,
2421 hrtimer_init_sleeper(to, current);
2422 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2423 current->timer_slack_ns);
2428 * Prepare to wait on uaddr. On success, holds hb lock and increments
2431 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2435 /* queue_me and wait for wakeup, timeout, or a signal. */
2436 futex_wait_queue_me(hb, &q, to);
2438 /* If we were woken (and unqueued), we succeeded, whatever. */
2440 /* unqueue_me() drops q.key ref */
2441 if (!unqueue_me(&q))
2444 if (to && !to->task)
2448 * We expect signal_pending(current), but we might be the
2449 * victim of a spurious wakeup as well.
2451 if (!signal_pending(current))
2458 restart = ¤t->restart_block;
2459 restart->fn = futex_wait_restart;
2460 restart->futex.uaddr = uaddr;
2461 restart->futex.val = val;
2462 restart->futex.time = abs_time->tv64;
2463 restart->futex.bitset = bitset;
2464 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2466 ret = -ERESTART_RESTARTBLOCK;
2470 hrtimer_cancel(&to->timer);
2471 destroy_hrtimer_on_stack(&to->timer);
2477 static long futex_wait_restart(struct restart_block *restart)
2479 u32 __user *uaddr = restart->futex.uaddr;
2480 ktime_t t, *tp = NULL;
2482 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2483 t.tv64 = restart->futex.time;
2486 restart->fn = do_no_restart_syscall;
2488 return (long)futex_wait(uaddr, restart->futex.flags,
2489 restart->futex.val, tp, restart->futex.bitset);
2494 * Userspace tried a 0 -> TID atomic transition of the futex value
2495 * and failed. The kernel side here does the whole locking operation:
2496 * if there are waiters then it will block as a consequence of relying
2497 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2498 * a 0 value of the futex too.).
2500 * Also serves as futex trylock_pi()'ing, and due semantics.
2502 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2503 ktime_t *time, int trylock)
2505 struct hrtimer_sleeper timeout, *to = NULL;
2506 struct futex_hash_bucket *hb;
2507 struct futex_q q = futex_q_init;
2510 if (refill_pi_state_cache())
2515 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2517 hrtimer_init_sleeper(to, current);
2518 hrtimer_set_expires(&to->timer, *time);
2522 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2523 if (unlikely(ret != 0))
2527 hb = queue_lock(&q);
2529 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2530 if (unlikely(ret)) {
2532 * Atomic work succeeded and we got the lock,
2533 * or failed. Either way, we do _not_ block.
2537 /* We got the lock. */
2539 goto out_unlock_put_key;
2544 * Two reasons for this:
2545 * - Task is exiting and we just wait for the
2547 * - The user space value changed.
2550 put_futex_key(&q.key);
2554 goto out_unlock_put_key;
2559 * Only actually queue now that the atomic ops are done:
2563 WARN_ON(!q.pi_state);
2565 * Block on the PI mutex:
2568 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2570 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2571 /* Fixup the trylock return value: */
2572 ret = ret ? 0 : -EWOULDBLOCK;
2575 spin_lock(q.lock_ptr);
2577 * Fixup the pi_state owner and possibly acquire the lock if we
2580 res = fixup_owner(uaddr, &q, !ret);
2582 * If fixup_owner() returned an error, proprogate that. If it acquired
2583 * the lock, clear our -ETIMEDOUT or -EINTR.
2586 ret = (res < 0) ? res : 0;
2589 * If fixup_owner() faulted and was unable to handle the fault, unlock
2590 * it and return the fault to userspace.
2592 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2593 rt_mutex_unlock(&q.pi_state->pi_mutex);
2595 /* Unqueue and drop the lock */
2604 put_futex_key(&q.key);
2607 destroy_hrtimer_on_stack(&to->timer);
2608 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2613 ret = fault_in_user_writeable(uaddr);
2617 if (!(flags & FLAGS_SHARED))
2620 put_futex_key(&q.key);
2625 * Userspace attempted a TID -> 0 atomic transition, and failed.
2626 * This is the in-kernel slowpath: we look up the PI state (if any),
2627 * and do the rt-mutex unlock.
2629 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2631 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2632 union futex_key key = FUTEX_KEY_INIT;
2633 struct futex_hash_bucket *hb;
2634 struct futex_q *match;
2638 if (get_user(uval, uaddr))
2641 * We release only a lock we actually own:
2643 if ((uval & FUTEX_TID_MASK) != vpid)
2646 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2650 hb = hash_futex(&key);
2651 spin_lock(&hb->lock);
2654 * Check waiters first. We do not trust user space values at
2655 * all and we at least want to know if user space fiddled
2656 * with the futex value instead of blindly unlocking.
2658 match = futex_top_waiter(hb, &key);
2660 ret = wake_futex_pi(uaddr, uval, match, hb);
2662 * In case of success wake_futex_pi dropped the hash
2668 * The atomic access to the futex value generated a
2669 * pagefault, so retry the user-access and the wakeup:
2674 * A unconditional UNLOCK_PI op raced against a waiter
2675 * setting the FUTEX_WAITERS bit. Try again.
2677 if (ret == -EAGAIN) {
2678 spin_unlock(&hb->lock);
2679 put_futex_key(&key);
2683 * wake_futex_pi has detected invalid state. Tell user
2690 * We have no kernel internal state, i.e. no waiters in the
2691 * kernel. Waiters which are about to queue themselves are stuck
2692 * on hb->lock. So we can safely ignore them. We do neither
2693 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2696 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2700 * If uval has changed, let user space handle it.
2702 ret = (curval == uval) ? 0 : -EAGAIN;
2705 spin_unlock(&hb->lock);
2707 put_futex_key(&key);
2711 spin_unlock(&hb->lock);
2712 put_futex_key(&key);
2714 ret = fault_in_user_writeable(uaddr);
2722 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2723 * @hb: the hash_bucket futex_q was original enqueued on
2724 * @q: the futex_q woken while waiting to be requeued
2725 * @key2: the futex_key of the requeue target futex
2726 * @timeout: the timeout associated with the wait (NULL if none)
2728 * Detect if the task was woken on the initial futex as opposed to the requeue
2729 * target futex. If so, determine if it was a timeout or a signal that caused
2730 * the wakeup and return the appropriate error code to the caller. Must be
2731 * called with the hb lock held.
2734 * 0 = no early wakeup detected;
2735 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2738 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2739 struct futex_q *q, union futex_key *key2,
2740 struct hrtimer_sleeper *timeout)
2745 * With the hb lock held, we avoid races while we process the wakeup.
2746 * We only need to hold hb (and not hb2) to ensure atomicity as the
2747 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2748 * It can't be requeued from uaddr2 to something else since we don't
2749 * support a PI aware source futex for requeue.
2751 if (!match_futex(&q->key, key2)) {
2752 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2754 * We were woken prior to requeue by a timeout or a signal.
2755 * Unqueue the futex_q and determine which it was.
2757 plist_del(&q->list, &hb->chain);
2760 /* Handle spurious wakeups gracefully */
2762 if (timeout && !timeout->task)
2764 else if (signal_pending(current))
2765 ret = -ERESTARTNOINTR;
2771 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2772 * @uaddr: the futex we initially wait on (non-pi)
2773 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2774 * the same type, no requeueing from private to shared, etc.
2775 * @val: the expected value of uaddr
2776 * @abs_time: absolute timeout
2777 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2778 * @uaddr2: the pi futex we will take prior to returning to user-space
2780 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2781 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2782 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2783 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2784 * without one, the pi logic would not know which task to boost/deboost, if
2785 * there was a need to.
2787 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2788 * via the following--
2789 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2790 * 2) wakeup on uaddr2 after a requeue
2794 * If 3, cleanup and return -ERESTARTNOINTR.
2796 * If 2, we may then block on trying to take the rt_mutex and return via:
2797 * 5) successful lock
2800 * 8) other lock acquisition failure
2802 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2804 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2810 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2811 u32 val, ktime_t *abs_time, u32 bitset,
2814 struct hrtimer_sleeper timeout, *to = NULL;
2815 struct rt_mutex_waiter rt_waiter;
2816 struct rt_mutex *pi_mutex = NULL;
2817 struct futex_hash_bucket *hb;
2818 union futex_key key2 = FUTEX_KEY_INIT;
2819 struct futex_q q = futex_q_init;
2822 if (uaddr == uaddr2)
2830 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2831 CLOCK_REALTIME : CLOCK_MONOTONIC,
2833 hrtimer_init_sleeper(to, current);
2834 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2835 current->timer_slack_ns);
2839 * The waiter is allocated on our stack, manipulated by the requeue
2840 * code while we sleep on uaddr.
2842 debug_rt_mutex_init_waiter(&rt_waiter);
2843 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2844 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2845 rt_waiter.task = NULL;
2847 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2848 if (unlikely(ret != 0))
2852 q.rt_waiter = &rt_waiter;
2853 q.requeue_pi_key = &key2;
2856 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2859 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2864 * The check above which compares uaddrs is not sufficient for
2865 * shared futexes. We need to compare the keys:
2867 if (match_futex(&q.key, &key2)) {
2873 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2874 futex_wait_queue_me(hb, &q, to);
2876 spin_lock(&hb->lock);
2877 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2878 spin_unlock(&hb->lock);
2883 * In order for us to be here, we know our q.key == key2, and since
2884 * we took the hb->lock above, we also know that futex_requeue() has
2885 * completed and we no longer have to concern ourselves with a wakeup
2886 * race with the atomic proxy lock acquisition by the requeue code. The
2887 * futex_requeue dropped our key1 reference and incremented our key2
2891 /* Check if the requeue code acquired the second futex for us. */
2894 * Got the lock. We might not be the anticipated owner if we
2895 * did a lock-steal - fix up the PI-state in that case.
2897 if (q.pi_state && (q.pi_state->owner != current)) {
2898 spin_lock(q.lock_ptr);
2899 ret = fixup_pi_state_owner(uaddr2, &q, current);
2901 * Drop the reference to the pi state which
2902 * the requeue_pi() code acquired for us.
2904 put_pi_state(q.pi_state);
2905 spin_unlock(q.lock_ptr);
2909 * We have been woken up by futex_unlock_pi(), a timeout, or a
2910 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2913 WARN_ON(!q.pi_state);
2914 pi_mutex = &q.pi_state->pi_mutex;
2915 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2916 debug_rt_mutex_free_waiter(&rt_waiter);
2918 spin_lock(q.lock_ptr);
2920 * Fixup the pi_state owner and possibly acquire the lock if we
2923 res = fixup_owner(uaddr2, &q, !ret);
2925 * If fixup_owner() returned an error, proprogate that. If it
2926 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2929 ret = (res < 0) ? res : 0;
2931 /* Unqueue and drop the lock. */
2936 * If fixup_pi_state_owner() faulted and was unable to handle the
2937 * fault, unlock the rt_mutex and return the fault to userspace.
2939 if (ret == -EFAULT) {
2940 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2941 rt_mutex_unlock(pi_mutex);
2942 } else if (ret == -EINTR) {
2944 * We've already been requeued, but cannot restart by calling
2945 * futex_lock_pi() directly. We could restart this syscall, but
2946 * it would detect that the user space "val" changed and return
2947 * -EWOULDBLOCK. Save the overhead of the restart and return
2948 * -EWOULDBLOCK directly.
2954 put_futex_key(&q.key);
2956 put_futex_key(&key2);
2960 hrtimer_cancel(&to->timer);
2961 destroy_hrtimer_on_stack(&to->timer);
2967 * Support for robust futexes: the kernel cleans up held futexes at
2970 * Implementation: user-space maintains a per-thread list of locks it
2971 * is holding. Upon do_exit(), the kernel carefully walks this list,
2972 * and marks all locks that are owned by this thread with the
2973 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2974 * always manipulated with the lock held, so the list is private and
2975 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2976 * field, to allow the kernel to clean up if the thread dies after
2977 * acquiring the lock, but just before it could have added itself to
2978 * the list. There can only be one such pending lock.
2982 * sys_set_robust_list() - Set the robust-futex list head of a task
2983 * @head: pointer to the list-head
2984 * @len: length of the list-head, as userspace expects
2986 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2989 if (!futex_cmpxchg_enabled)
2992 * The kernel knows only one size for now:
2994 if (unlikely(len != sizeof(*head)))
2997 current->robust_list = head;
3003 * sys_get_robust_list() - Get the robust-futex list head of a task
3004 * @pid: pid of the process [zero for current task]
3005 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3006 * @len_ptr: pointer to a length field, the kernel fills in the header size
3008 SYSCALL_DEFINE3(get_robust_list, int, pid,
3009 struct robust_list_head __user * __user *, head_ptr,
3010 size_t __user *, len_ptr)
3012 struct robust_list_head __user *head;
3014 struct task_struct *p;
3016 if (!futex_cmpxchg_enabled)
3025 p = find_task_by_vpid(pid);
3031 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3034 head = p->robust_list;
3037 if (put_user(sizeof(*head), len_ptr))
3039 return put_user(head, head_ptr);
3048 * Process a futex-list entry, check whether it's owned by the
3049 * dying task, and do notification if so:
3051 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3053 u32 uval, uninitialized_var(nval), mval;
3056 if (get_user(uval, uaddr))
3059 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3061 * Ok, this dying thread is truly holding a futex
3062 * of interest. Set the OWNER_DIED bit atomically
3063 * via cmpxchg, and if the value had FUTEX_WAITERS
3064 * set, wake up a waiter (if any). (We have to do a
3065 * futex_wake() even if OWNER_DIED is already set -
3066 * to handle the rare but possible case of recursive
3067 * thread-death.) The rest of the cleanup is done in
3070 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3072 * We are not holding a lock here, but we want to have
3073 * the pagefault_disable/enable() protection because
3074 * we want to handle the fault gracefully. If the
3075 * access fails we try to fault in the futex with R/W
3076 * verification via get_user_pages. get_user() above
3077 * does not guarantee R/W access. If that fails we
3078 * give up and leave the futex locked.
3080 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3081 if (fault_in_user_writeable(uaddr))
3089 * Wake robust non-PI futexes here. The wakeup of
3090 * PI futexes happens in exit_pi_state():
3092 if (!pi && (uval & FUTEX_WAITERS))
3093 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3099 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3101 static inline int fetch_robust_entry(struct robust_list __user **entry,
3102 struct robust_list __user * __user *head,
3105 unsigned long uentry;
3107 if (get_user(uentry, (unsigned long __user *)head))
3110 *entry = (void __user *)(uentry & ~1UL);
3117 * Walk curr->robust_list (very carefully, it's a userspace list!)
3118 * and mark any locks found there dead, and notify any waiters.
3120 * We silently return on any sign of list-walking problem.
3122 void exit_robust_list(struct task_struct *curr)
3124 struct robust_list_head __user *head = curr->robust_list;
3125 struct robust_list __user *entry, *next_entry, *pending;
3126 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3127 unsigned int uninitialized_var(next_pi);
3128 unsigned long futex_offset;
3131 if (!futex_cmpxchg_enabled)
3135 * Fetch the list head (which was registered earlier, via
3136 * sys_set_robust_list()):
3138 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3141 * Fetch the relative futex offset:
3143 if (get_user(futex_offset, &head->futex_offset))
3146 * Fetch any possibly pending lock-add first, and handle it
3149 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3152 next_entry = NULL; /* avoid warning with gcc */
3153 while (entry != &head->list) {
3155 * Fetch the next entry in the list before calling
3156 * handle_futex_death:
3158 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3160 * A pending lock might already be on the list, so
3161 * don't process it twice:
3163 if (entry != pending)
3164 if (handle_futex_death((void __user *)entry + futex_offset,
3172 * Avoid excessively long or circular lists:
3181 handle_futex_death((void __user *)pending + futex_offset,
3185 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3186 u32 __user *uaddr2, u32 val2, u32 val3)
3188 int cmd = op & FUTEX_CMD_MASK;
3189 unsigned int flags = 0;
3191 if (!(op & FUTEX_PRIVATE_FLAG))
3192 flags |= FLAGS_SHARED;
3194 if (op & FUTEX_CLOCK_REALTIME) {
3195 flags |= FLAGS_CLOCKRT;
3196 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3197 cmd != FUTEX_WAIT_REQUEUE_PI)
3203 case FUTEX_UNLOCK_PI:
3204 case FUTEX_TRYLOCK_PI:
3205 case FUTEX_WAIT_REQUEUE_PI:
3206 case FUTEX_CMP_REQUEUE_PI:
3207 if (!futex_cmpxchg_enabled)
3213 val3 = FUTEX_BITSET_MATCH_ANY;
3214 case FUTEX_WAIT_BITSET:
3215 return futex_wait(uaddr, flags, val, timeout, val3);
3217 val3 = FUTEX_BITSET_MATCH_ANY;
3218 case FUTEX_WAKE_BITSET:
3219 return futex_wake(uaddr, flags, val, val3);
3221 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3222 case FUTEX_CMP_REQUEUE:
3223 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3225 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3227 return futex_lock_pi(uaddr, flags, timeout, 0);
3228 case FUTEX_UNLOCK_PI:
3229 return futex_unlock_pi(uaddr, flags);
3230 case FUTEX_TRYLOCK_PI:
3231 return futex_lock_pi(uaddr, flags, NULL, 1);
3232 case FUTEX_WAIT_REQUEUE_PI:
3233 val3 = FUTEX_BITSET_MATCH_ANY;
3234 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3236 case FUTEX_CMP_REQUEUE_PI:
3237 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3243 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3244 struct timespec __user *, utime, u32 __user *, uaddr2,
3248 ktime_t t, *tp = NULL;
3250 int cmd = op & FUTEX_CMD_MASK;
3252 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3253 cmd == FUTEX_WAIT_BITSET ||
3254 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3255 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3257 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3259 if (!timespec_valid(&ts))
3262 t = timespec_to_ktime(ts);
3263 if (cmd == FUTEX_WAIT)
3264 t = ktime_add_safe(ktime_get(), t);
3268 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3269 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3271 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3272 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3273 val2 = (u32) (unsigned long) utime;
3275 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3278 static void __init futex_detect_cmpxchg(void)
3280 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3284 * This will fail and we want it. Some arch implementations do
3285 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3286 * functionality. We want to know that before we call in any
3287 * of the complex code paths. Also we want to prevent
3288 * registration of robust lists in that case. NULL is
3289 * guaranteed to fault and we get -EFAULT on functional
3290 * implementation, the non-functional ones will return
3293 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3294 futex_cmpxchg_enabled = 1;
3298 static int __init futex_init(void)
3300 unsigned int futex_shift;
3303 #if CONFIG_BASE_SMALL
3304 futex_hashsize = 16;
3306 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3309 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3311 futex_hashsize < 256 ? HASH_SMALL : 0,
3313 futex_hashsize, futex_hashsize);
3314 futex_hashsize = 1UL << futex_shift;
3316 futex_detect_cmpxchg();
3318 for (i = 0; i < futex_hashsize; i++) {
3319 atomic_set(&futex_queues[i].waiters, 0);
3320 plist_head_init(&futex_queues[i].chain);
3321 spin_lock_init(&futex_queues[i].lock);
3326 __initcall(futex_init);