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 * mb(); (A) <-- paired with -.
129 * lock(hash_bucket(futex)); |
133 * | sys_futex(WAKE, futex);
134 * | futex_wake(futex);
136 * `-------> 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
182 #define FLAGS_SHARED 0x01
183 #define FLAGS_CLOCKRT 0x02
184 #define FLAGS_HAS_TIMEOUT 0x04
187 * Priority Inheritance state:
189 struct futex_pi_state {
191 * list of 'owned' pi_state instances - these have to be
192 * cleaned up in do_exit() if the task exits prematurely:
194 struct list_head list;
199 struct rt_mutex pi_mutex;
201 struct task_struct *owner;
208 * struct futex_q - The hashed futex queue entry, one per waiting task
209 * @list: priority-sorted list of tasks waiting on this futex
210 * @task: the task waiting on the futex
211 * @lock_ptr: the hash bucket lock
212 * @key: the key the futex is hashed on
213 * @pi_state: optional priority inheritance state
214 * @rt_waiter: rt_waiter storage for use with requeue_pi
215 * @requeue_pi_key: the requeue_pi target futex key
216 * @bitset: bitset for the optional bitmasked wakeup
218 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
219 * we can wake only the relevant ones (hashed queues may be shared).
221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223 * The order of wakeup is always to make the first condition true, then
226 * PI futexes are typically woken before they are removed from the hash list via
227 * the rt_mutex code. See unqueue_me_pi().
230 struct plist_node list;
232 struct task_struct *task;
233 spinlock_t *lock_ptr;
235 struct futex_pi_state *pi_state;
236 struct rt_mutex_waiter *rt_waiter;
237 union futex_key *requeue_pi_key;
241 static const struct futex_q futex_q_init = {
242 /* list gets initialized in queue_me()*/
243 .key = FUTEX_KEY_INIT,
244 .bitset = FUTEX_BITSET_MATCH_ANY
248 * Hash buckets are shared by all the futex_keys that hash to the same
249 * location. Each key may have multiple futex_q structures, one for each task
250 * waiting on a futex.
252 struct futex_hash_bucket {
255 struct plist_head chain;
256 } ____cacheline_aligned_in_smp;
259 * The base of the bucket array and its size are always used together
260 * (after initialization only in hash_futex()), so ensure that they
261 * reside in the same cacheline.
264 struct futex_hash_bucket *queues;
265 unsigned long hashsize;
266 } __futex_data __read_mostly __aligned(2*sizeof(long));
267 #define futex_queues (__futex_data.queues)
268 #define futex_hashsize (__futex_data.hashsize)
272 * Fault injections for futexes.
274 #ifdef CONFIG_FAIL_FUTEX
277 struct fault_attr attr;
281 .attr = FAULT_ATTR_INITIALIZER,
282 .ignore_private = false,
285 static int __init setup_fail_futex(char *str)
287 return setup_fault_attr(&fail_futex.attr, str);
289 __setup("fail_futex=", setup_fail_futex);
291 static bool should_fail_futex(bool fshared)
293 if (fail_futex.ignore_private && !fshared)
296 return should_fail(&fail_futex.attr, 1);
299 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
301 static int __init fail_futex_debugfs(void)
303 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
306 dir = fault_create_debugfs_attr("fail_futex", NULL,
311 if (!debugfs_create_bool("ignore-private", mode, dir,
312 &fail_futex.ignore_private)) {
313 debugfs_remove_recursive(dir);
320 late_initcall(fail_futex_debugfs);
322 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
325 static inline bool should_fail_futex(bool fshared)
329 #endif /* CONFIG_FAIL_FUTEX */
331 static inline void futex_get_mm(union futex_key *key)
333 atomic_inc(&key->private.mm->mm_count);
335 * Ensure futex_get_mm() implies a full barrier such that
336 * get_futex_key() implies a full barrier. This is relied upon
337 * as full barrier (B), see the ordering comment above.
339 smp_mb__after_atomic();
343 * Reflects a new waiter being added to the waitqueue.
345 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
348 atomic_inc(&hb->waiters);
350 * Full barrier (A), see the ordering comment above.
352 smp_mb__after_atomic();
357 * Reflects a waiter being removed from the waitqueue by wakeup
360 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
363 atomic_dec(&hb->waiters);
367 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
370 return atomic_read(&hb->waiters);
377 * We hash on the keys returned from get_futex_key (see below).
379 static struct futex_hash_bucket *hash_futex(union futex_key *key)
381 u32 hash = jhash2((u32*)&key->both.word,
382 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
384 return &futex_queues[hash & (futex_hashsize - 1)];
388 * Return 1 if two futex_keys are equal, 0 otherwise.
390 static inline int match_futex(union futex_key *key1, union futex_key *key2)
393 && key1->both.word == key2->both.word
394 && key1->both.ptr == key2->both.ptr
395 && key1->both.offset == key2->both.offset);
399 * Take a reference to the resource addressed by a key.
400 * Can be called while holding spinlocks.
403 static void get_futex_key_refs(union futex_key *key)
408 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
410 ihold(key->shared.inode); /* implies MB (B) */
412 case FUT_OFF_MMSHARED:
413 futex_get_mm(key); /* implies MB (B) */
417 * Private futexes do not hold reference on an inode or
418 * mm, therefore the only purpose of calling get_futex_key_refs
419 * is because we need the barrier for the lockless waiter check.
421 smp_mb(); /* explicit MB (B) */
426 * Drop a reference to the resource addressed by a key.
427 * The hash bucket spinlock must not be held. This is
428 * a no-op for private futexes, see comment in the get
431 static void drop_futex_key_refs(union futex_key *key)
433 if (!key->both.ptr) {
434 /* If we're here then we tried to put a key we failed to get */
439 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
441 iput(key->shared.inode);
443 case FUT_OFF_MMSHARED:
444 mmdrop(key->private.mm);
450 * get_futex_key() - Get parameters which are the keys for a futex
451 * @uaddr: virtual address of the futex
452 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
453 * @key: address where result is stored.
454 * @rw: mapping needs to be read/write (values: VERIFY_READ,
457 * Return: a negative error code or 0
459 * The key words are stored in *key on success.
461 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
462 * offset_within_page). For private mappings, it's (uaddr, current->mm).
463 * We can usually work out the index without swapping in the page.
465 * lock_page() might sleep, the caller should not hold a spinlock.
468 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
470 unsigned long address = (unsigned long)uaddr;
471 struct mm_struct *mm = current->mm;
472 struct page *page, *page_head;
476 * The futex address must be "naturally" aligned.
478 key->both.offset = address % PAGE_SIZE;
479 if (unlikely((address % sizeof(u32)) != 0))
481 address -= key->both.offset;
483 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
486 if (unlikely(should_fail_futex(fshared)))
490 * PROCESS_PRIVATE futexes are fast.
491 * As the mm cannot disappear under us and the 'key' only needs
492 * virtual address, we dont even have to find the underlying vma.
493 * Note : We do have to check 'uaddr' is a valid user address,
494 * but access_ok() should be faster than find_vma()
497 key->private.mm = mm;
498 key->private.address = address;
499 get_futex_key_refs(key); /* implies MB (B) */
504 /* Ignore any VERIFY_READ mapping (futex common case) */
505 if (unlikely(should_fail_futex(fshared)))
508 err = get_user_pages_fast(address, 1, 1, &page);
510 * If write access is not required (eg. FUTEX_WAIT), try
511 * and get read-only access.
513 if (err == -EFAULT && rw == VERIFY_READ) {
514 err = get_user_pages_fast(address, 1, 0, &page);
522 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (unlikely(PageTail(page))) {
526 /* serialize against __split_huge_page_splitting() */
528 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
529 page_head = compound_head(page);
531 * page_head is valid pointer but we must pin
532 * it before taking the PG_lock and/or
533 * PG_compound_lock. The moment we re-enable
534 * irqs __split_huge_page_splitting() can
535 * return and the head page can be freed from
536 * under us. We can't take the PG_lock and/or
537 * PG_compound_lock on a page that could be
538 * freed from under us.
540 if (page != page_head) {
551 page_head = compound_head(page);
552 if (page != page_head) {
558 lock_page(page_head);
561 * If page_head->mapping is NULL, then it cannot be a PageAnon
562 * page; but it might be the ZERO_PAGE or in the gate area or
563 * in a special mapping (all cases which we are happy to fail);
564 * or it may have been a good file page when get_user_pages_fast
565 * found it, but truncated or holepunched or subjected to
566 * invalidate_complete_page2 before we got the page lock (also
567 * cases which we are happy to fail). And we hold a reference,
568 * so refcount care in invalidate_complete_page's remove_mapping
569 * prevents drop_caches from setting mapping to NULL beneath us.
571 * The case we do have to guard against is when memory pressure made
572 * shmem_writepage move it from filecache to swapcache beneath us:
573 * an unlikely race, but we do need to retry for page_head->mapping.
575 if (!page_head->mapping) {
576 int shmem_swizzled = PageSwapCache(page_head);
577 unlock_page(page_head);
585 * Private mappings are handled in a simple way.
587 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
588 * it's a read-only handle, it's expected that futexes attach to
589 * the object not the particular process.
591 if (PageAnon(page_head)) {
593 * A RO anonymous page will never change and thus doesn't make
594 * sense for futex operations.
596 if (unlikely(should_fail_futex(fshared)) || ro) {
601 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
602 key->private.mm = mm;
603 key->private.address = address;
605 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
606 key->shared.inode = page_head->mapping->host;
607 key->shared.pgoff = basepage_index(page);
610 get_futex_key_refs(key); /* implies MB (B) */
613 unlock_page(page_head);
618 static inline void put_futex_key(union futex_key *key)
620 drop_futex_key_refs(key);
624 * fault_in_user_writeable() - Fault in user address and verify RW access
625 * @uaddr: pointer to faulting user space address
627 * Slow path to fixup the fault we just took in the atomic write
630 * We have no generic implementation of a non-destructive write to the
631 * user address. We know that we faulted in the atomic pagefault
632 * disabled section so we can as well avoid the #PF overhead by
633 * calling get_user_pages() right away.
635 static int fault_in_user_writeable(u32 __user *uaddr)
637 struct mm_struct *mm = current->mm;
640 down_read(&mm->mmap_sem);
641 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
643 up_read(&mm->mmap_sem);
645 return ret < 0 ? ret : 0;
649 * futex_top_waiter() - Return the highest priority waiter on a futex
650 * @hb: the hash bucket the futex_q's reside in
651 * @key: the futex key (to distinguish it from other futex futex_q's)
653 * Must be called with the hb lock held.
655 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
656 union futex_key *key)
658 struct futex_q *this;
660 plist_for_each_entry(this, &hb->chain, list) {
661 if (match_futex(&this->key, key))
667 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
668 u32 uval, u32 newval)
673 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
679 static int get_futex_value_locked(u32 *dest, u32 __user *from)
684 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
687 return ret ? -EFAULT : 0;
694 static int refill_pi_state_cache(void)
696 struct futex_pi_state *pi_state;
698 if (likely(current->pi_state_cache))
701 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
706 INIT_LIST_HEAD(&pi_state->list);
707 /* pi_mutex gets initialized later */
708 pi_state->owner = NULL;
709 atomic_set(&pi_state->refcount, 1);
710 pi_state->key = FUTEX_KEY_INIT;
712 current->pi_state_cache = pi_state;
717 static struct futex_pi_state * alloc_pi_state(void)
719 struct futex_pi_state *pi_state = current->pi_state_cache;
722 current->pi_state_cache = NULL;
728 * Must be called with the hb lock held.
730 static void free_pi_state(struct futex_pi_state *pi_state)
735 if (!atomic_dec_and_test(&pi_state->refcount))
739 * If pi_state->owner is NULL, the owner is most probably dying
740 * and has cleaned up the pi_state already
742 if (pi_state->owner) {
743 raw_spin_lock_irq(&pi_state->owner->pi_lock);
744 list_del_init(&pi_state->list);
745 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
747 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
750 if (current->pi_state_cache)
754 * pi_state->list is already empty.
755 * clear pi_state->owner.
756 * refcount is at 0 - put it back to 1.
758 pi_state->owner = NULL;
759 atomic_set(&pi_state->refcount, 1);
760 current->pi_state_cache = pi_state;
765 * Look up the task based on what TID userspace gave us.
768 static struct task_struct * futex_find_get_task(pid_t pid)
770 struct task_struct *p;
773 p = find_task_by_vpid(pid);
783 * This task is holding PI mutexes at exit time => bad.
784 * Kernel cleans up PI-state, but userspace is likely hosed.
785 * (Robust-futex cleanup is separate and might save the day for userspace.)
787 void exit_pi_state_list(struct task_struct *curr)
789 struct list_head *next, *head = &curr->pi_state_list;
790 struct futex_pi_state *pi_state;
791 struct futex_hash_bucket *hb;
792 union futex_key key = FUTEX_KEY_INIT;
794 if (!futex_cmpxchg_enabled)
797 * We are a ZOMBIE and nobody can enqueue itself on
798 * pi_state_list anymore, but we have to be careful
799 * versus waiters unqueueing themselves:
801 raw_spin_lock_irq(&curr->pi_lock);
802 while (!list_empty(head)) {
805 pi_state = list_entry(next, struct futex_pi_state, list);
807 hb = hash_futex(&key);
808 raw_spin_unlock_irq(&curr->pi_lock);
810 spin_lock(&hb->lock);
812 raw_spin_lock_irq(&curr->pi_lock);
814 * We dropped the pi-lock, so re-check whether this
815 * task still owns the PI-state:
817 if (head->next != next) {
818 raw_spin_unlock_irq(&curr->pi_lock);
819 spin_unlock(&hb->lock);
820 raw_spin_lock_irq(&curr->pi_lock);
824 WARN_ON(pi_state->owner != curr);
825 WARN_ON(list_empty(&pi_state->list));
826 list_del_init(&pi_state->list);
827 pi_state->owner = NULL;
828 raw_spin_unlock_irq(&curr->pi_lock);
830 rt_mutex_unlock(&pi_state->pi_mutex);
832 spin_unlock(&hb->lock);
834 raw_spin_lock_irq(&curr->pi_lock);
836 raw_spin_unlock_irq(&curr->pi_lock);
840 * We need to check the following states:
842 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
844 * [1] NULL | --- | --- | 0 | 0/1 | Valid
845 * [2] NULL | --- | --- | >0 | 0/1 | Valid
847 * [3] Found | NULL | -- | Any | 0/1 | Invalid
849 * [4] Found | Found | NULL | 0 | 1 | Valid
850 * [5] Found | Found | NULL | >0 | 1 | Invalid
852 * [6] Found | Found | task | 0 | 1 | Valid
854 * [7] Found | Found | NULL | Any | 0 | Invalid
856 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
857 * [9] Found | Found | task | 0 | 0 | Invalid
858 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
860 * [1] Indicates that the kernel can acquire the futex atomically. We
861 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
863 * [2] Valid, if TID does not belong to a kernel thread. If no matching
864 * thread is found then it indicates that the owner TID has died.
866 * [3] Invalid. The waiter is queued on a non PI futex
868 * [4] Valid state after exit_robust_list(), which sets the user space
869 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
871 * [5] The user space value got manipulated between exit_robust_list()
872 * and exit_pi_state_list()
874 * [6] Valid state after exit_pi_state_list() which sets the new owner in
875 * the pi_state but cannot access the user space value.
877 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
879 * [8] Owner and user space value match
881 * [9] There is no transient state which sets the user space TID to 0
882 * except exit_robust_list(), but this is indicated by the
883 * FUTEX_OWNER_DIED bit. See [4]
885 * [10] There is no transient state which leaves owner and user space
890 * Validate that the existing waiter has a pi_state and sanity check
891 * the pi_state against the user space value. If correct, attach to
894 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
895 struct futex_pi_state **ps)
897 pid_t pid = uval & FUTEX_TID_MASK;
900 * Userspace might have messed up non-PI and PI futexes [3]
902 if (unlikely(!pi_state))
905 WARN_ON(!atomic_read(&pi_state->refcount));
908 * Handle the owner died case:
910 if (uval & FUTEX_OWNER_DIED) {
912 * exit_pi_state_list sets owner to NULL and wakes the
913 * topmost waiter. The task which acquires the
914 * pi_state->rt_mutex will fixup owner.
916 if (!pi_state->owner) {
918 * No pi state owner, but the user space TID
919 * is not 0. Inconsistent state. [5]
924 * Take a ref on the state and return success. [4]
930 * If TID is 0, then either the dying owner has not
931 * yet executed exit_pi_state_list() or some waiter
932 * acquired the rtmutex in the pi state, but did not
933 * yet fixup the TID in user space.
935 * Take a ref on the state and return success. [6]
941 * If the owner died bit is not set, then the pi_state
942 * must have an owner. [7]
944 if (!pi_state->owner)
949 * Bail out if user space manipulated the futex value. If pi
950 * state exists then the owner TID must be the same as the
951 * user space TID. [9/10]
953 if (pid != task_pid_vnr(pi_state->owner))
956 atomic_inc(&pi_state->refcount);
962 * Lookup the task for the TID provided from user space and attach to
963 * it after doing proper sanity checks.
965 static int attach_to_pi_owner(u32 uval, union futex_key *key,
966 struct futex_pi_state **ps)
968 pid_t pid = uval & FUTEX_TID_MASK;
969 struct futex_pi_state *pi_state;
970 struct task_struct *p;
973 * We are the first waiter - try to look up the real owner and attach
974 * the new pi_state to it, but bail out when TID = 0 [1]
978 p = futex_find_get_task(pid);
982 if (unlikely(p->flags & PF_KTHREAD)) {
988 * We need to look at the task state flags to figure out,
989 * whether the task is exiting. To protect against the do_exit
990 * change of the task flags, we do this protected by
993 raw_spin_lock_irq(&p->pi_lock);
994 if (unlikely(p->flags & PF_EXITING)) {
996 * The task is on the way out. When PF_EXITPIDONE is
997 * set, we know that the task has finished the
1000 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1002 raw_spin_unlock_irq(&p->pi_lock);
1008 * No existing pi state. First waiter. [2]
1010 pi_state = alloc_pi_state();
1013 * Initialize the pi_mutex in locked state and make @p
1016 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1018 /* Store the key for possible exit cleanups: */
1019 pi_state->key = *key;
1021 WARN_ON(!list_empty(&pi_state->list));
1022 list_add(&pi_state->list, &p->pi_state_list);
1023 pi_state->owner = p;
1024 raw_spin_unlock_irq(&p->pi_lock);
1033 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1034 union futex_key *key, struct futex_pi_state **ps)
1036 struct futex_q *match = futex_top_waiter(hb, key);
1039 * If there is a waiter on that futex, validate it and
1040 * attach to the pi_state when the validation succeeds.
1043 return attach_to_pi_state(uval, match->pi_state, ps);
1046 * We are the first waiter - try to look up the owner based on
1047 * @uval and attach to it.
1049 return attach_to_pi_owner(uval, key, ps);
1052 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1054 u32 uninitialized_var(curval);
1056 if (unlikely(should_fail_futex(true)))
1059 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1062 /*If user space value changed, let the caller retry */
1063 return curval != uval ? -EAGAIN : 0;
1067 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1068 * @uaddr: the pi futex user address
1069 * @hb: the pi futex hash bucket
1070 * @key: the futex key associated with uaddr and hb
1071 * @ps: the pi_state pointer where we store the result of the
1073 * @task: the task to perform the atomic lock work for. This will
1074 * be "current" except in the case of requeue pi.
1075 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1078 * 0 - ready to wait;
1079 * 1 - acquired the lock;
1082 * The hb->lock and futex_key refs shall be held by the caller.
1084 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1085 union futex_key *key,
1086 struct futex_pi_state **ps,
1087 struct task_struct *task, int set_waiters)
1089 u32 uval, newval, vpid = task_pid_vnr(task);
1090 struct futex_q *match;
1094 * Read the user space value first so we can validate a few
1095 * things before proceeding further.
1097 if (get_futex_value_locked(&uval, uaddr))
1100 if (unlikely(should_fail_futex(true)))
1106 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1109 if ((unlikely(should_fail_futex(true))))
1113 * Lookup existing state first. If it exists, try to attach to
1116 match = futex_top_waiter(hb, key);
1118 return attach_to_pi_state(uval, match->pi_state, ps);
1121 * No waiter and user TID is 0. We are here because the
1122 * waiters or the owner died bit is set or called from
1123 * requeue_cmp_pi or for whatever reason something took the
1126 if (!(uval & FUTEX_TID_MASK)) {
1128 * We take over the futex. No other waiters and the user space
1129 * TID is 0. We preserve the owner died bit.
1131 newval = uval & FUTEX_OWNER_DIED;
1134 /* The futex requeue_pi code can enforce the waiters bit */
1136 newval |= FUTEX_WAITERS;
1138 ret = lock_pi_update_atomic(uaddr, uval, newval);
1139 /* If the take over worked, return 1 */
1140 return ret < 0 ? ret : 1;
1144 * First waiter. Set the waiters bit before attaching ourself to
1145 * the owner. If owner tries to unlock, it will be forced into
1146 * the kernel and blocked on hb->lock.
1148 newval = uval | FUTEX_WAITERS;
1149 ret = lock_pi_update_atomic(uaddr, uval, newval);
1153 * If the update of the user space value succeeded, we try to
1154 * attach to the owner. If that fails, no harm done, we only
1155 * set the FUTEX_WAITERS bit in the user space variable.
1157 return attach_to_pi_owner(uval, key, ps);
1161 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1162 * @q: The futex_q to unqueue
1164 * The q->lock_ptr must not be NULL and must be held by the caller.
1166 static void __unqueue_futex(struct futex_q *q)
1168 struct futex_hash_bucket *hb;
1170 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1171 || WARN_ON(plist_node_empty(&q->list)))
1174 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1175 plist_del(&q->list, &hb->chain);
1180 * The hash bucket lock must be held when this is called.
1181 * Afterwards, the futex_q must not be accessed. Callers
1182 * must ensure to later call wake_up_q() for the actual
1185 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1187 struct task_struct *p = q->task;
1189 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1193 * Queue the task for later wakeup for after we've released
1194 * the hb->lock. wake_q_add() grabs reference to p.
1196 wake_q_add(wake_q, p);
1199 * The waiting task can free the futex_q as soon as
1200 * q->lock_ptr = NULL is written, without taking any locks. A
1201 * memory barrier is required here to prevent the following
1202 * store to lock_ptr from getting ahead of the plist_del.
1208 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1209 struct futex_hash_bucket *hb)
1211 struct task_struct *new_owner;
1212 struct futex_pi_state *pi_state = this->pi_state;
1213 u32 uninitialized_var(curval), newval;
1215 WAKE_Q(wake_sleeper_q);
1223 * If current does not own the pi_state then the futex is
1224 * inconsistent and user space fiddled with the futex value.
1226 if (pi_state->owner != current)
1229 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1230 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1233 * It is possible that the next waiter (the one that brought
1234 * this owner to the kernel) timed out and is no longer
1235 * waiting on the lock.
1238 new_owner = this->task;
1241 * We pass it to the next owner. The WAITERS bit is always
1242 * kept enabled while there is PI state around. We cleanup the
1243 * owner died bit, because we are the owner.
1245 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1247 if (unlikely(should_fail_futex(true)))
1250 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1252 } else if (curval != uval) {
1254 * If a unconditional UNLOCK_PI operation (user space did not
1255 * try the TID->0 transition) raced with a waiter setting the
1256 * FUTEX_WAITERS flag between get_user() and locking the hash
1257 * bucket lock, retry the operation.
1259 if ((FUTEX_TID_MASK & curval) == uval)
1265 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1269 raw_spin_lock(&pi_state->owner->pi_lock);
1270 WARN_ON(list_empty(&pi_state->list));
1271 list_del_init(&pi_state->list);
1272 raw_spin_unlock(&pi_state->owner->pi_lock);
1274 raw_spin_lock(&new_owner->pi_lock);
1275 WARN_ON(!list_empty(&pi_state->list));
1276 list_add(&pi_state->list, &new_owner->pi_state_list);
1277 pi_state->owner = new_owner;
1278 raw_spin_unlock(&new_owner->pi_lock);
1280 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1282 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q,
1286 * First unlock HB so the waiter does not spin on it once he got woken
1287 * up. Second wake up the waiter before the priority is adjusted. If we
1288 * deboost first (and lose our higher priority), then the task might get
1289 * scheduled away before the wake up can take place.
1291 deboost |= spin_unlock_no_deboost(&hb->lock);
1293 wake_up_q_sleeper(&wake_sleeper_q);
1295 rt_mutex_adjust_prio(current);
1301 * Express the locking dependencies for lockdep:
1304 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1307 spin_lock(&hb1->lock);
1309 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1310 } else { /* hb1 > hb2 */
1311 spin_lock(&hb2->lock);
1312 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1317 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1319 spin_unlock(&hb1->lock);
1321 spin_unlock(&hb2->lock);
1325 * Wake up waiters matching bitset queued on this futex (uaddr).
1328 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1330 struct futex_hash_bucket *hb;
1331 struct futex_q *this, *next;
1332 union futex_key key = FUTEX_KEY_INIT;
1339 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1340 if (unlikely(ret != 0))
1343 hb = hash_futex(&key);
1345 /* Make sure we really have tasks to wakeup */
1346 if (!hb_waiters_pending(hb))
1349 spin_lock(&hb->lock);
1351 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1352 if (match_futex (&this->key, &key)) {
1353 if (this->pi_state || this->rt_waiter) {
1358 /* Check if one of the bits is set in both bitsets */
1359 if (!(this->bitset & bitset))
1362 mark_wake_futex(&wake_q, this);
1363 if (++ret >= nr_wake)
1368 spin_unlock(&hb->lock);
1371 put_futex_key(&key);
1377 * Wake up all waiters hashed on the physical page that is mapped
1378 * to this virtual address:
1381 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1382 int nr_wake, int nr_wake2, int op)
1384 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1385 struct futex_hash_bucket *hb1, *hb2;
1386 struct futex_q *this, *next;
1391 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1392 if (unlikely(ret != 0))
1394 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1395 if (unlikely(ret != 0))
1398 hb1 = hash_futex(&key1);
1399 hb2 = hash_futex(&key2);
1402 double_lock_hb(hb1, hb2);
1403 op_ret = futex_atomic_op_inuser(op, uaddr2);
1404 if (unlikely(op_ret < 0)) {
1406 double_unlock_hb(hb1, hb2);
1410 * we don't get EFAULT from MMU faults if we don't have an MMU,
1411 * but we might get them from range checking
1417 if (unlikely(op_ret != -EFAULT)) {
1422 ret = fault_in_user_writeable(uaddr2);
1426 if (!(flags & FLAGS_SHARED))
1429 put_futex_key(&key2);
1430 put_futex_key(&key1);
1434 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1435 if (match_futex (&this->key, &key1)) {
1436 if (this->pi_state || this->rt_waiter) {
1440 mark_wake_futex(&wake_q, this);
1441 if (++ret >= nr_wake)
1448 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1449 if (match_futex (&this->key, &key2)) {
1450 if (this->pi_state || this->rt_waiter) {
1454 mark_wake_futex(&wake_q, this);
1455 if (++op_ret >= nr_wake2)
1463 double_unlock_hb(hb1, hb2);
1466 put_futex_key(&key2);
1468 put_futex_key(&key1);
1474 * requeue_futex() - Requeue a futex_q from one hb to another
1475 * @q: the futex_q to requeue
1476 * @hb1: the source hash_bucket
1477 * @hb2: the target hash_bucket
1478 * @key2: the new key for the requeued futex_q
1481 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1482 struct futex_hash_bucket *hb2, union futex_key *key2)
1486 * If key1 and key2 hash to the same bucket, no need to
1489 if (likely(&hb1->chain != &hb2->chain)) {
1490 plist_del(&q->list, &hb1->chain);
1491 hb_waiters_dec(hb1);
1492 hb_waiters_inc(hb2);
1493 plist_add(&q->list, &hb2->chain);
1494 q->lock_ptr = &hb2->lock;
1496 get_futex_key_refs(key2);
1501 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1503 * @key: the key of the requeue target futex
1504 * @hb: the hash_bucket of the requeue target futex
1506 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1507 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1508 * to the requeue target futex so the waiter can detect the wakeup on the right
1509 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1510 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1511 * to protect access to the pi_state to fixup the owner later. Must be called
1512 * with both q->lock_ptr and hb->lock held.
1515 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1516 struct futex_hash_bucket *hb)
1518 get_futex_key_refs(key);
1523 WARN_ON(!q->rt_waiter);
1524 q->rt_waiter = NULL;
1526 q->lock_ptr = &hb->lock;
1528 wake_up_state(q->task, TASK_NORMAL);
1532 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1533 * @pifutex: the user address of the to futex
1534 * @hb1: the from futex hash bucket, must be locked by the caller
1535 * @hb2: the to futex hash bucket, must be locked by the caller
1536 * @key1: the from futex key
1537 * @key2: the to futex key
1538 * @ps: address to store the pi_state pointer
1539 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1541 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1542 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1543 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1544 * hb1 and hb2 must be held by the caller.
1547 * 0 - failed to acquire the lock atomically;
1548 * >0 - acquired the lock, return value is vpid of the top_waiter
1551 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1552 struct futex_hash_bucket *hb1,
1553 struct futex_hash_bucket *hb2,
1554 union futex_key *key1, union futex_key *key2,
1555 struct futex_pi_state **ps, int set_waiters)
1557 struct futex_q *top_waiter = NULL;
1561 if (get_futex_value_locked(&curval, pifutex))
1564 if (unlikely(should_fail_futex(true)))
1568 * Find the top_waiter and determine if there are additional waiters.
1569 * If the caller intends to requeue more than 1 waiter to pifutex,
1570 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1571 * as we have means to handle the possible fault. If not, don't set
1572 * the bit unecessarily as it will force the subsequent unlock to enter
1575 top_waiter = futex_top_waiter(hb1, key1);
1577 /* There are no waiters, nothing for us to do. */
1581 /* Ensure we requeue to the expected futex. */
1582 if (!match_futex(top_waiter->requeue_pi_key, key2))
1586 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1587 * the contended case or if set_waiters is 1. The pi_state is returned
1588 * in ps in contended cases.
1590 vpid = task_pid_vnr(top_waiter->task);
1591 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1594 requeue_pi_wake_futex(top_waiter, key2, hb2);
1601 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1602 * @uaddr1: source futex user address
1603 * @flags: futex flags (FLAGS_SHARED, etc.)
1604 * @uaddr2: target futex user address
1605 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1606 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1607 * @cmpval: @uaddr1 expected value (or %NULL)
1608 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1609 * pi futex (pi to pi requeue is not supported)
1611 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1612 * uaddr2 atomically on behalf of the top waiter.
1615 * >=0 - on success, the number of tasks requeued or woken;
1618 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1619 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1620 u32 *cmpval, int requeue_pi)
1622 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1623 int drop_count = 0, task_count = 0, ret;
1624 struct futex_pi_state *pi_state = NULL;
1625 struct futex_hash_bucket *hb1, *hb2;
1626 struct futex_q *this, *next;
1631 * Requeue PI only works on two distinct uaddrs. This
1632 * check is only valid for private futexes. See below.
1634 if (uaddr1 == uaddr2)
1638 * requeue_pi requires a pi_state, try to allocate it now
1639 * without any locks in case it fails.
1641 if (refill_pi_state_cache())
1644 * requeue_pi must wake as many tasks as it can, up to nr_wake
1645 * + nr_requeue, since it acquires the rt_mutex prior to
1646 * returning to userspace, so as to not leave the rt_mutex with
1647 * waiters and no owner. However, second and third wake-ups
1648 * cannot be predicted as they involve race conditions with the
1649 * first wake and a fault while looking up the pi_state. Both
1650 * pthread_cond_signal() and pthread_cond_broadcast() should
1658 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1659 if (unlikely(ret != 0))
1661 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1662 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1663 if (unlikely(ret != 0))
1667 * The check above which compares uaddrs is not sufficient for
1668 * shared futexes. We need to compare the keys:
1670 if (requeue_pi && match_futex(&key1, &key2)) {
1675 hb1 = hash_futex(&key1);
1676 hb2 = hash_futex(&key2);
1679 hb_waiters_inc(hb2);
1680 double_lock_hb(hb1, hb2);
1682 if (likely(cmpval != NULL)) {
1685 ret = get_futex_value_locked(&curval, uaddr1);
1687 if (unlikely(ret)) {
1688 double_unlock_hb(hb1, hb2);
1689 hb_waiters_dec(hb2);
1691 ret = get_user(curval, uaddr1);
1695 if (!(flags & FLAGS_SHARED))
1698 put_futex_key(&key2);
1699 put_futex_key(&key1);
1702 if (curval != *cmpval) {
1708 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1710 * Attempt to acquire uaddr2 and wake the top waiter. If we
1711 * intend to requeue waiters, force setting the FUTEX_WAITERS
1712 * bit. We force this here where we are able to easily handle
1713 * faults rather in the requeue loop below.
1715 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1716 &key2, &pi_state, nr_requeue);
1719 * At this point the top_waiter has either taken uaddr2 or is
1720 * waiting on it. If the former, then the pi_state will not
1721 * exist yet, look it up one more time to ensure we have a
1722 * reference to it. If the lock was taken, ret contains the
1723 * vpid of the top waiter task.
1730 * If we acquired the lock, then the user
1731 * space value of uaddr2 should be vpid. It
1732 * cannot be changed by the top waiter as it
1733 * is blocked on hb2 lock if it tries to do
1734 * so. If something fiddled with it behind our
1735 * back the pi state lookup might unearth
1736 * it. So we rather use the known value than
1737 * rereading and handing potential crap to
1740 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1747 free_pi_state(pi_state);
1749 double_unlock_hb(hb1, hb2);
1750 hb_waiters_dec(hb2);
1751 put_futex_key(&key2);
1752 put_futex_key(&key1);
1753 ret = fault_in_user_writeable(uaddr2);
1759 * Two reasons for this:
1760 * - Owner is exiting and we just wait for the
1762 * - The user space value changed.
1764 free_pi_state(pi_state);
1766 double_unlock_hb(hb1, hb2);
1767 hb_waiters_dec(hb2);
1768 put_futex_key(&key2);
1769 put_futex_key(&key1);
1777 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1778 if (task_count - nr_wake >= nr_requeue)
1781 if (!match_futex(&this->key, &key1))
1785 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1786 * be paired with each other and no other futex ops.
1788 * We should never be requeueing a futex_q with a pi_state,
1789 * which is awaiting a futex_unlock_pi().
1791 if ((requeue_pi && !this->rt_waiter) ||
1792 (!requeue_pi && this->rt_waiter) ||
1799 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1800 * lock, we already woke the top_waiter. If not, it will be
1801 * woken by futex_unlock_pi().
1803 if (++task_count <= nr_wake && !requeue_pi) {
1804 mark_wake_futex(&wake_q, this);
1808 /* Ensure we requeue to the expected futex for requeue_pi. */
1809 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1815 * Requeue nr_requeue waiters and possibly one more in the case
1816 * of requeue_pi if we couldn't acquire the lock atomically.
1819 /* Prepare the waiter to take the rt_mutex. */
1820 atomic_inc(&pi_state->refcount);
1821 this->pi_state = pi_state;
1822 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1826 /* We got the lock. */
1827 requeue_pi_wake_futex(this, &key2, hb2);
1830 } else if (ret == -EAGAIN) {
1832 * Waiter was woken by timeout or
1833 * signal and has set pi_blocked_on to
1834 * PI_WAKEUP_INPROGRESS before we
1835 * tried to enqueue it on the rtmutex.
1837 this->pi_state = NULL;
1838 free_pi_state(pi_state);
1842 this->pi_state = NULL;
1843 free_pi_state(pi_state);
1847 requeue_futex(this, hb1, hb2, &key2);
1852 free_pi_state(pi_state);
1853 double_unlock_hb(hb1, hb2);
1855 hb_waiters_dec(hb2);
1858 * drop_futex_key_refs() must be called outside the spinlocks. During
1859 * the requeue we moved futex_q's from the hash bucket at key1 to the
1860 * one at key2 and updated their key pointer. We no longer need to
1861 * hold the references to key1.
1863 while (--drop_count >= 0)
1864 drop_futex_key_refs(&key1);
1867 put_futex_key(&key2);
1869 put_futex_key(&key1);
1871 return ret ? ret : task_count;
1874 /* The key must be already stored in q->key. */
1875 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1876 __acquires(&hb->lock)
1878 struct futex_hash_bucket *hb;
1880 hb = hash_futex(&q->key);
1883 * Increment the counter before taking the lock so that
1884 * a potential waker won't miss a to-be-slept task that is
1885 * waiting for the spinlock. This is safe as all queue_lock()
1886 * users end up calling queue_me(). Similarly, for housekeeping,
1887 * decrement the counter at queue_unlock() when some error has
1888 * occurred and we don't end up adding the task to the list.
1892 q->lock_ptr = &hb->lock;
1894 spin_lock(&hb->lock); /* implies MB (A) */
1899 queue_unlock(struct futex_hash_bucket *hb)
1900 __releases(&hb->lock)
1902 spin_unlock(&hb->lock);
1907 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1908 * @q: The futex_q to enqueue
1909 * @hb: The destination hash bucket
1911 * The hb->lock must be held by the caller, and is released here. A call to
1912 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1913 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1914 * or nothing if the unqueue is done as part of the wake process and the unqueue
1915 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1918 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1919 __releases(&hb->lock)
1924 * The priority used to register this element is
1925 * - either the real thread-priority for the real-time threads
1926 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1927 * - or MAX_RT_PRIO for non-RT threads.
1928 * Thus, all RT-threads are woken first in priority order, and
1929 * the others are woken last, in FIFO order.
1931 prio = min(current->normal_prio, MAX_RT_PRIO);
1933 plist_node_init(&q->list, prio);
1934 plist_add(&q->list, &hb->chain);
1936 spin_unlock(&hb->lock);
1940 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1941 * @q: The futex_q to unqueue
1943 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1944 * be paired with exactly one earlier call to queue_me().
1947 * 1 - if the futex_q was still queued (and we removed unqueued it);
1948 * 0 - if the futex_q was already removed by the waking thread
1950 static int unqueue_me(struct futex_q *q)
1952 spinlock_t *lock_ptr;
1955 /* In the common case we don't take the spinlock, which is nice. */
1957 lock_ptr = q->lock_ptr;
1959 if (lock_ptr != NULL) {
1960 spin_lock(lock_ptr);
1962 * q->lock_ptr can change between reading it and
1963 * spin_lock(), causing us to take the wrong lock. This
1964 * corrects the race condition.
1966 * Reasoning goes like this: if we have the wrong lock,
1967 * q->lock_ptr must have changed (maybe several times)
1968 * between reading it and the spin_lock(). It can
1969 * change again after the spin_lock() but only if it was
1970 * already changed before the spin_lock(). It cannot,
1971 * however, change back to the original value. Therefore
1972 * we can detect whether we acquired the correct lock.
1974 if (unlikely(lock_ptr != q->lock_ptr)) {
1975 spin_unlock(lock_ptr);
1980 BUG_ON(q->pi_state);
1982 spin_unlock(lock_ptr);
1986 drop_futex_key_refs(&q->key);
1991 * PI futexes can not be requeued and must remove themself from the
1992 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1995 static void unqueue_me_pi(struct futex_q *q)
1996 __releases(q->lock_ptr)
2000 BUG_ON(!q->pi_state);
2001 free_pi_state(q->pi_state);
2004 spin_unlock(q->lock_ptr);
2008 * Fixup the pi_state owner with the new owner.
2010 * Must be called with hash bucket lock held and mm->sem held for non
2013 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2014 struct task_struct *newowner)
2016 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2017 struct futex_pi_state *pi_state = q->pi_state;
2018 struct task_struct *oldowner = pi_state->owner;
2019 u32 uval, uninitialized_var(curval), newval;
2023 if (!pi_state->owner)
2024 newtid |= FUTEX_OWNER_DIED;
2027 * We are here either because we stole the rtmutex from the
2028 * previous highest priority waiter or we are the highest priority
2029 * waiter but failed to get the rtmutex the first time.
2030 * We have to replace the newowner TID in the user space variable.
2031 * This must be atomic as we have to preserve the owner died bit here.
2033 * Note: We write the user space value _before_ changing the pi_state
2034 * because we can fault here. Imagine swapped out pages or a fork
2035 * that marked all the anonymous memory readonly for cow.
2037 * Modifying pi_state _before_ the user space value would
2038 * leave the pi_state in an inconsistent state when we fault
2039 * here, because we need to drop the hash bucket lock to
2040 * handle the fault. This might be observed in the PID check
2041 * in lookup_pi_state.
2044 if (get_futex_value_locked(&uval, uaddr))
2048 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2050 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2058 * We fixed up user space. Now we need to fix the pi_state
2061 if (pi_state->owner != NULL) {
2062 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2063 WARN_ON(list_empty(&pi_state->list));
2064 list_del_init(&pi_state->list);
2065 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2068 pi_state->owner = newowner;
2070 raw_spin_lock_irq(&newowner->pi_lock);
2071 WARN_ON(!list_empty(&pi_state->list));
2072 list_add(&pi_state->list, &newowner->pi_state_list);
2073 raw_spin_unlock_irq(&newowner->pi_lock);
2077 * To handle the page fault we need to drop the hash bucket
2078 * lock here. That gives the other task (either the highest priority
2079 * waiter itself or the task which stole the rtmutex) the
2080 * chance to try the fixup of the pi_state. So once we are
2081 * back from handling the fault we need to check the pi_state
2082 * after reacquiring the hash bucket lock and before trying to
2083 * do another fixup. When the fixup has been done already we
2087 spin_unlock(q->lock_ptr);
2089 ret = fault_in_user_writeable(uaddr);
2091 spin_lock(q->lock_ptr);
2094 * Check if someone else fixed it for us:
2096 if (pi_state->owner != oldowner)
2105 static long futex_wait_restart(struct restart_block *restart);
2108 * fixup_owner() - Post lock pi_state and corner case management
2109 * @uaddr: user address of the futex
2110 * @q: futex_q (contains pi_state and access to the rt_mutex)
2111 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2113 * After attempting to lock an rt_mutex, this function is called to cleanup
2114 * the pi_state owner as well as handle race conditions that may allow us to
2115 * acquire the lock. Must be called with the hb lock held.
2118 * 1 - success, lock taken;
2119 * 0 - success, lock not taken;
2120 * <0 - on error (-EFAULT)
2122 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2124 struct task_struct *owner;
2129 * Got the lock. We might not be the anticipated owner if we
2130 * did a lock-steal - fix up the PI-state in that case:
2132 if (q->pi_state->owner != current)
2133 ret = fixup_pi_state_owner(uaddr, q, current);
2138 * Catch the rare case, where the lock was released when we were on the
2139 * way back before we locked the hash bucket.
2141 if (q->pi_state->owner == current) {
2143 * Try to get the rt_mutex now. This might fail as some other
2144 * task acquired the rt_mutex after we removed ourself from the
2145 * rt_mutex waiters list.
2147 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2153 * pi_state is incorrect, some other task did a lock steal and
2154 * we returned due to timeout or signal without taking the
2155 * rt_mutex. Too late.
2157 raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2158 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2160 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2161 raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2162 ret = fixup_pi_state_owner(uaddr, q, owner);
2167 * Paranoia check. If we did not take the lock, then we should not be
2168 * the owner of the rt_mutex.
2170 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2171 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2172 "pi-state %p\n", ret,
2173 q->pi_state->pi_mutex.owner,
2174 q->pi_state->owner);
2177 return ret ? ret : locked;
2181 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2182 * @hb: the futex hash bucket, must be locked by the caller
2183 * @q: the futex_q to queue up on
2184 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2186 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2187 struct hrtimer_sleeper *timeout)
2190 * The task state is guaranteed to be set before another task can
2191 * wake it. set_current_state() is implemented using smp_store_mb() and
2192 * queue_me() calls spin_unlock() upon completion, both serializing
2193 * access to the hash list and forcing another memory barrier.
2195 set_current_state(TASK_INTERRUPTIBLE);
2200 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2203 * If we have been removed from the hash list, then another task
2204 * has tried to wake us, and we can skip the call to schedule().
2206 if (likely(!plist_node_empty(&q->list))) {
2208 * If the timer has already expired, current will already be
2209 * flagged for rescheduling. Only call schedule if there
2210 * is no timeout, or if it has yet to expire.
2212 if (!timeout || timeout->task)
2213 freezable_schedule();
2215 __set_current_state(TASK_RUNNING);
2219 * futex_wait_setup() - Prepare to wait on a futex
2220 * @uaddr: the futex userspace address
2221 * @val: the expected value
2222 * @flags: futex flags (FLAGS_SHARED, etc.)
2223 * @q: the associated futex_q
2224 * @hb: storage for hash_bucket pointer to be returned to caller
2226 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2227 * compare it with the expected value. Handle atomic faults internally.
2228 * Return with the hb lock held and a q.key reference on success, and unlocked
2229 * with no q.key reference on failure.
2232 * 0 - uaddr contains val and hb has been locked;
2233 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2235 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2236 struct futex_q *q, struct futex_hash_bucket **hb)
2242 * Access the page AFTER the hash-bucket is locked.
2243 * Order is important:
2245 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2246 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2248 * The basic logical guarantee of a futex is that it blocks ONLY
2249 * if cond(var) is known to be true at the time of blocking, for
2250 * any cond. If we locked the hash-bucket after testing *uaddr, that
2251 * would open a race condition where we could block indefinitely with
2252 * cond(var) false, which would violate the guarantee.
2254 * On the other hand, we insert q and release the hash-bucket only
2255 * after testing *uaddr. This guarantees that futex_wait() will NOT
2256 * absorb a wakeup if *uaddr does not match the desired values
2257 * while the syscall executes.
2260 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2261 if (unlikely(ret != 0))
2265 *hb = queue_lock(q);
2267 ret = get_futex_value_locked(&uval, uaddr);
2272 ret = get_user(uval, uaddr);
2276 if (!(flags & FLAGS_SHARED))
2279 put_futex_key(&q->key);
2290 put_futex_key(&q->key);
2294 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2295 ktime_t *abs_time, u32 bitset)
2297 struct hrtimer_sleeper timeout, *to = NULL;
2298 struct restart_block *restart;
2299 struct futex_hash_bucket *hb;
2300 struct futex_q q = futex_q_init;
2310 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2311 CLOCK_REALTIME : CLOCK_MONOTONIC,
2313 hrtimer_init_sleeper(to, current);
2314 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2315 current->timer_slack_ns);
2320 * Prepare to wait on uaddr. On success, holds hb lock and increments
2323 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2327 /* queue_me and wait for wakeup, timeout, or a signal. */
2328 futex_wait_queue_me(hb, &q, to);
2330 /* If we were woken (and unqueued), we succeeded, whatever. */
2332 /* unqueue_me() drops q.key ref */
2333 if (!unqueue_me(&q))
2336 if (to && !to->task)
2340 * We expect signal_pending(current), but we might be the
2341 * victim of a spurious wakeup as well.
2343 if (!signal_pending(current))
2350 restart = ¤t->restart_block;
2351 restart->fn = futex_wait_restart;
2352 restart->futex.uaddr = uaddr;
2353 restart->futex.val = val;
2354 restart->futex.time = abs_time->tv64;
2355 restart->futex.bitset = bitset;
2356 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2358 ret = -ERESTART_RESTARTBLOCK;
2362 hrtimer_cancel(&to->timer);
2363 destroy_hrtimer_on_stack(&to->timer);
2369 static long futex_wait_restart(struct restart_block *restart)
2371 u32 __user *uaddr = restart->futex.uaddr;
2372 ktime_t t, *tp = NULL;
2374 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2375 t.tv64 = restart->futex.time;
2378 restart->fn = do_no_restart_syscall;
2380 return (long)futex_wait(uaddr, restart->futex.flags,
2381 restart->futex.val, tp, restart->futex.bitset);
2386 * Userspace tried a 0 -> TID atomic transition of the futex value
2387 * and failed. The kernel side here does the whole locking operation:
2388 * if there are waiters then it will block as a consequence of relying
2389 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2390 * a 0 value of the futex too.).
2392 * Also serves as futex trylock_pi()'ing, and due semantics.
2394 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2395 ktime_t *time, int trylock)
2397 struct hrtimer_sleeper timeout, *to = NULL;
2398 struct futex_hash_bucket *hb;
2399 struct futex_q q = futex_q_init;
2402 if (refill_pi_state_cache())
2407 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2409 hrtimer_init_sleeper(to, current);
2410 hrtimer_set_expires(&to->timer, *time);
2414 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2415 if (unlikely(ret != 0))
2419 hb = queue_lock(&q);
2421 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2422 if (unlikely(ret)) {
2424 * Atomic work succeeded and we got the lock,
2425 * or failed. Either way, we do _not_ block.
2429 /* We got the lock. */
2431 goto out_unlock_put_key;
2436 * Two reasons for this:
2437 * - Task is exiting and we just wait for the
2439 * - The user space value changed.
2442 put_futex_key(&q.key);
2446 goto out_unlock_put_key;
2451 * Only actually queue now that the atomic ops are done:
2455 WARN_ON(!q.pi_state);
2457 * Block on the PI mutex:
2460 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2462 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2463 /* Fixup the trylock return value: */
2464 ret = ret ? 0 : -EWOULDBLOCK;
2467 spin_lock(q.lock_ptr);
2469 * Fixup the pi_state owner and possibly acquire the lock if we
2472 res = fixup_owner(uaddr, &q, !ret);
2474 * If fixup_owner() returned an error, proprogate that. If it acquired
2475 * the lock, clear our -ETIMEDOUT or -EINTR.
2478 ret = (res < 0) ? res : 0;
2481 * If fixup_owner() faulted and was unable to handle the fault, unlock
2482 * it and return the fault to userspace.
2484 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2485 rt_mutex_unlock(&q.pi_state->pi_mutex);
2487 /* Unqueue and drop the lock */
2496 put_futex_key(&q.key);
2499 destroy_hrtimer_on_stack(&to->timer);
2500 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2505 ret = fault_in_user_writeable(uaddr);
2509 if (!(flags & FLAGS_SHARED))
2512 put_futex_key(&q.key);
2517 * Userspace attempted a TID -> 0 atomic transition, and failed.
2518 * This is the in-kernel slowpath: we look up the PI state (if any),
2519 * and do the rt-mutex unlock.
2521 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2523 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2524 union futex_key key = FUTEX_KEY_INIT;
2525 struct futex_hash_bucket *hb;
2526 struct futex_q *match;
2530 if (get_user(uval, uaddr))
2533 * We release only a lock we actually own:
2535 if ((uval & FUTEX_TID_MASK) != vpid)
2538 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2542 hb = hash_futex(&key);
2543 spin_lock(&hb->lock);
2546 * Check waiters first. We do not trust user space values at
2547 * all and we at least want to know if user space fiddled
2548 * with the futex value instead of blindly unlocking.
2550 match = futex_top_waiter(hb, &key);
2552 ret = wake_futex_pi(uaddr, uval, match, hb);
2554 * In case of success wake_futex_pi dropped the hash
2560 * The atomic access to the futex value generated a
2561 * pagefault, so retry the user-access and the wakeup:
2566 * A unconditional UNLOCK_PI op raced against a waiter
2567 * setting the FUTEX_WAITERS bit. Try again.
2569 if (ret == -EAGAIN) {
2570 spin_unlock(&hb->lock);
2571 put_futex_key(&key);
2575 * wake_futex_pi has detected invalid state. Tell user
2582 * We have no kernel internal state, i.e. no waiters in the
2583 * kernel. Waiters which are about to queue themselves are stuck
2584 * on hb->lock. So we can safely ignore them. We do neither
2585 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2588 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2592 * If uval has changed, let user space handle it.
2594 ret = (curval == uval) ? 0 : -EAGAIN;
2597 spin_unlock(&hb->lock);
2599 put_futex_key(&key);
2603 spin_unlock(&hb->lock);
2604 put_futex_key(&key);
2606 ret = fault_in_user_writeable(uaddr);
2614 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2615 * @hb: the hash_bucket futex_q was original enqueued on
2616 * @q: the futex_q woken while waiting to be requeued
2617 * @key2: the futex_key of the requeue target futex
2618 * @timeout: the timeout associated with the wait (NULL if none)
2620 * Detect if the task was woken on the initial futex as opposed to the requeue
2621 * target futex. If so, determine if it was a timeout or a signal that caused
2622 * the wakeup and return the appropriate error code to the caller. Must be
2623 * called with the hb lock held.
2626 * 0 = no early wakeup detected;
2627 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2630 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2631 struct futex_q *q, union futex_key *key2,
2632 struct hrtimer_sleeper *timeout)
2637 * With the hb lock held, we avoid races while we process the wakeup.
2638 * We only need to hold hb (and not hb2) to ensure atomicity as the
2639 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2640 * It can't be requeued from uaddr2 to something else since we don't
2641 * support a PI aware source futex for requeue.
2643 if (!match_futex(&q->key, key2)) {
2644 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2646 * We were woken prior to requeue by a timeout or a signal.
2647 * Unqueue the futex_q and determine which it was.
2649 plist_del(&q->list, &hb->chain);
2652 /* Handle spurious wakeups gracefully */
2654 if (timeout && !timeout->task)
2656 else if (signal_pending(current))
2657 ret = -ERESTARTNOINTR;
2663 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2664 * @uaddr: the futex we initially wait on (non-pi)
2665 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2666 * the same type, no requeueing from private to shared, etc.
2667 * @val: the expected value of uaddr
2668 * @abs_time: absolute timeout
2669 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2670 * @uaddr2: the pi futex we will take prior to returning to user-space
2672 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2673 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2674 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2675 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2676 * without one, the pi logic would not know which task to boost/deboost, if
2677 * there was a need to.
2679 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2680 * via the following--
2681 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2682 * 2) wakeup on uaddr2 after a requeue
2686 * If 3, cleanup and return -ERESTARTNOINTR.
2688 * If 2, we may then block on trying to take the rt_mutex and return via:
2689 * 5) successful lock
2692 * 8) other lock acquisition failure
2694 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2696 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2702 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2703 u32 val, ktime_t *abs_time, u32 bitset,
2706 struct hrtimer_sleeper timeout, *to = NULL;
2707 struct rt_mutex_waiter rt_waiter;
2708 struct rt_mutex *pi_mutex = NULL;
2709 struct futex_hash_bucket *hb, *hb2;
2710 union futex_key key2 = FUTEX_KEY_INIT;
2711 struct futex_q q = futex_q_init;
2714 if (uaddr == uaddr2)
2722 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2723 CLOCK_REALTIME : CLOCK_MONOTONIC,
2725 hrtimer_init_sleeper(to, current);
2726 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2727 current->timer_slack_ns);
2731 * The waiter is allocated on our stack, manipulated by the requeue
2732 * code while we sleep on uaddr.
2734 rt_mutex_init_waiter(&rt_waiter, false);
2736 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2737 if (unlikely(ret != 0))
2741 q.rt_waiter = &rt_waiter;
2742 q.requeue_pi_key = &key2;
2745 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2748 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2753 * The check above which compares uaddrs is not sufficient for
2754 * shared futexes. We need to compare the keys:
2756 if (match_futex(&q.key, &key2)) {
2762 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2763 futex_wait_queue_me(hb, &q, to);
2766 * On RT we must avoid races with requeue and trying to block
2767 * on two mutexes (hb->lock and uaddr2's rtmutex) by
2768 * serializing access to pi_blocked_on with pi_lock.
2770 raw_spin_lock_irq(¤t->pi_lock);
2771 if (current->pi_blocked_on) {
2773 * We have been requeued or are in the process of
2776 raw_spin_unlock_irq(¤t->pi_lock);
2779 * Setting pi_blocked_on to PI_WAKEUP_INPROGRESS
2780 * prevents a concurrent requeue from moving us to the
2781 * uaddr2 rtmutex. After that we can safely acquire
2782 * (and possibly block on) hb->lock.
2784 current->pi_blocked_on = PI_WAKEUP_INPROGRESS;
2785 raw_spin_unlock_irq(¤t->pi_lock);
2787 spin_lock(&hb->lock);
2790 * Clean up pi_blocked_on. We might leak it otherwise
2791 * when we succeeded with the hb->lock in the fast
2794 raw_spin_lock_irq(¤t->pi_lock);
2795 current->pi_blocked_on = NULL;
2796 raw_spin_unlock_irq(¤t->pi_lock);
2798 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2799 spin_unlock(&hb->lock);
2805 * In order to be here, we have either been requeued, are in
2806 * the process of being requeued, or requeue successfully
2807 * acquired uaddr2 on our behalf. If pi_blocked_on was
2808 * non-null above, we may be racing with a requeue. Do not
2809 * rely on q->lock_ptr to be hb2->lock until after blocking on
2810 * hb->lock or hb2->lock. The futex_requeue dropped our key1
2811 * reference and incremented our key2 reference count.
2813 hb2 = hash_futex(&key2);
2815 /* Check if the requeue code acquired the second futex for us. */
2818 * Got the lock. We might not be the anticipated owner if we
2819 * did a lock-steal - fix up the PI-state in that case.
2821 if (q.pi_state && (q.pi_state->owner != current)) {
2822 spin_lock(&hb2->lock);
2823 BUG_ON(&hb2->lock != q.lock_ptr);
2824 ret = fixup_pi_state_owner(uaddr2, &q, current);
2826 * Drop the reference to the pi state which
2827 * the requeue_pi() code acquired for us.
2829 free_pi_state(q.pi_state);
2830 spin_unlock(&hb2->lock);
2834 * We have been woken up by futex_unlock_pi(), a timeout, or a
2835 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2838 WARN_ON(!q.pi_state);
2839 pi_mutex = &q.pi_state->pi_mutex;
2840 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2841 debug_rt_mutex_free_waiter(&rt_waiter);
2843 spin_lock(&hb2->lock);
2844 BUG_ON(&hb2->lock != q.lock_ptr);
2846 * Fixup the pi_state owner and possibly acquire the lock if we
2849 res = fixup_owner(uaddr2, &q, !ret);
2851 * If fixup_owner() returned an error, proprogate that. If it
2852 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2855 ret = (res < 0) ? res : 0;
2857 /* Unqueue and drop the lock. */
2862 * If fixup_pi_state_owner() faulted and was unable to handle the
2863 * fault, unlock the rt_mutex and return the fault to userspace.
2865 if (ret == -EFAULT) {
2866 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2867 rt_mutex_unlock(pi_mutex);
2868 } else if (ret == -EINTR) {
2870 * We've already been requeued, but cannot restart by calling
2871 * futex_lock_pi() directly. We could restart this syscall, but
2872 * it would detect that the user space "val" changed and return
2873 * -EWOULDBLOCK. Save the overhead of the restart and return
2874 * -EWOULDBLOCK directly.
2880 put_futex_key(&q.key);
2882 put_futex_key(&key2);
2886 hrtimer_cancel(&to->timer);
2887 destroy_hrtimer_on_stack(&to->timer);
2893 * Support for robust futexes: the kernel cleans up held futexes at
2896 * Implementation: user-space maintains a per-thread list of locks it
2897 * is holding. Upon do_exit(), the kernel carefully walks this list,
2898 * and marks all locks that are owned by this thread with the
2899 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2900 * always manipulated with the lock held, so the list is private and
2901 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2902 * field, to allow the kernel to clean up if the thread dies after
2903 * acquiring the lock, but just before it could have added itself to
2904 * the list. There can only be one such pending lock.
2908 * sys_set_robust_list() - Set the robust-futex list head of a task
2909 * @head: pointer to the list-head
2910 * @len: length of the list-head, as userspace expects
2912 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2915 if (!futex_cmpxchg_enabled)
2918 * The kernel knows only one size for now:
2920 if (unlikely(len != sizeof(*head)))
2923 current->robust_list = head;
2929 * sys_get_robust_list() - Get the robust-futex list head of a task
2930 * @pid: pid of the process [zero for current task]
2931 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2932 * @len_ptr: pointer to a length field, the kernel fills in the header size
2934 SYSCALL_DEFINE3(get_robust_list, int, pid,
2935 struct robust_list_head __user * __user *, head_ptr,
2936 size_t __user *, len_ptr)
2938 struct robust_list_head __user *head;
2940 struct task_struct *p;
2942 if (!futex_cmpxchg_enabled)
2951 p = find_task_by_vpid(pid);
2957 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
2960 head = p->robust_list;
2963 if (put_user(sizeof(*head), len_ptr))
2965 return put_user(head, head_ptr);
2974 * Process a futex-list entry, check whether it's owned by the
2975 * dying task, and do notification if so:
2977 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2979 u32 uval, uninitialized_var(nval), mval;
2982 if (get_user(uval, uaddr))
2985 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2987 * Ok, this dying thread is truly holding a futex
2988 * of interest. Set the OWNER_DIED bit atomically
2989 * via cmpxchg, and if the value had FUTEX_WAITERS
2990 * set, wake up a waiter (if any). (We have to do a
2991 * futex_wake() even if OWNER_DIED is already set -
2992 * to handle the rare but possible case of recursive
2993 * thread-death.) The rest of the cleanup is done in
2996 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2998 * We are not holding a lock here, but we want to have
2999 * the pagefault_disable/enable() protection because
3000 * we want to handle the fault gracefully. If the
3001 * access fails we try to fault in the futex with R/W
3002 * verification via get_user_pages. get_user() above
3003 * does not guarantee R/W access. If that fails we
3004 * give up and leave the futex locked.
3006 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3007 if (fault_in_user_writeable(uaddr))
3015 * Wake robust non-PI futexes here. The wakeup of
3016 * PI futexes happens in exit_pi_state():
3018 if (!pi && (uval & FUTEX_WAITERS))
3019 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3025 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3027 static inline int fetch_robust_entry(struct robust_list __user **entry,
3028 struct robust_list __user * __user *head,
3031 unsigned long uentry;
3033 if (get_user(uentry, (unsigned long __user *)head))
3036 *entry = (void __user *)(uentry & ~1UL);
3043 * Walk curr->robust_list (very carefully, it's a userspace list!)
3044 * and mark any locks found there dead, and notify any waiters.
3046 * We silently return on any sign of list-walking problem.
3048 void exit_robust_list(struct task_struct *curr)
3050 struct robust_list_head __user *head = curr->robust_list;
3051 struct robust_list __user *entry, *next_entry, *pending;
3052 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3053 unsigned int uninitialized_var(next_pi);
3054 unsigned long futex_offset;
3057 if (!futex_cmpxchg_enabled)
3061 * Fetch the list head (which was registered earlier, via
3062 * sys_set_robust_list()):
3064 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3067 * Fetch the relative futex offset:
3069 if (get_user(futex_offset, &head->futex_offset))
3072 * Fetch any possibly pending lock-add first, and handle it
3075 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3078 next_entry = NULL; /* avoid warning with gcc */
3079 while (entry != &head->list) {
3081 * Fetch the next entry in the list before calling
3082 * handle_futex_death:
3084 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3086 * A pending lock might already be on the list, so
3087 * don't process it twice:
3089 if (entry != pending)
3090 if (handle_futex_death((void __user *)entry + futex_offset,
3098 * Avoid excessively long or circular lists:
3107 handle_futex_death((void __user *)pending + futex_offset,
3111 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3112 u32 __user *uaddr2, u32 val2, u32 val3)
3114 int cmd = op & FUTEX_CMD_MASK;
3115 unsigned int flags = 0;
3117 if (!(op & FUTEX_PRIVATE_FLAG))
3118 flags |= FLAGS_SHARED;
3120 if (op & FUTEX_CLOCK_REALTIME) {
3121 flags |= FLAGS_CLOCKRT;
3122 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3128 case FUTEX_UNLOCK_PI:
3129 case FUTEX_TRYLOCK_PI:
3130 case FUTEX_WAIT_REQUEUE_PI:
3131 case FUTEX_CMP_REQUEUE_PI:
3132 if (!futex_cmpxchg_enabled)
3138 val3 = FUTEX_BITSET_MATCH_ANY;
3139 case FUTEX_WAIT_BITSET:
3140 return futex_wait(uaddr, flags, val, timeout, val3);
3142 val3 = FUTEX_BITSET_MATCH_ANY;
3143 case FUTEX_WAKE_BITSET:
3144 return futex_wake(uaddr, flags, val, val3);
3146 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3147 case FUTEX_CMP_REQUEUE:
3148 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3150 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3152 return futex_lock_pi(uaddr, flags, timeout, 0);
3153 case FUTEX_UNLOCK_PI:
3154 return futex_unlock_pi(uaddr, flags);
3155 case FUTEX_TRYLOCK_PI:
3156 return futex_lock_pi(uaddr, flags, NULL, 1);
3157 case FUTEX_WAIT_REQUEUE_PI:
3158 val3 = FUTEX_BITSET_MATCH_ANY;
3159 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3161 case FUTEX_CMP_REQUEUE_PI:
3162 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3168 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3169 struct timespec __user *, utime, u32 __user *, uaddr2,
3173 ktime_t t, *tp = NULL;
3175 int cmd = op & FUTEX_CMD_MASK;
3177 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3178 cmd == FUTEX_WAIT_BITSET ||
3179 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3180 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3182 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3184 if (!timespec_valid(&ts))
3187 t = timespec_to_ktime(ts);
3188 if (cmd == FUTEX_WAIT)
3189 t = ktime_add_safe(ktime_get(), t);
3193 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3194 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3196 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3197 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3198 val2 = (u32) (unsigned long) utime;
3200 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3203 static void __init futex_detect_cmpxchg(void)
3205 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3209 * This will fail and we want it. Some arch implementations do
3210 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3211 * functionality. We want to know that before we call in any
3212 * of the complex code paths. Also we want to prevent
3213 * registration of robust lists in that case. NULL is
3214 * guaranteed to fault and we get -EFAULT on functional
3215 * implementation, the non-functional ones will return
3218 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3219 futex_cmpxchg_enabled = 1;
3223 static int __init futex_init(void)
3225 unsigned int futex_shift;
3228 #if CONFIG_BASE_SMALL
3229 futex_hashsize = 16;
3231 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3234 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3236 futex_hashsize < 256 ? HASH_SMALL : 0,
3238 futex_hashsize, futex_hashsize);
3239 futex_hashsize = 1UL << futex_shift;
3241 futex_detect_cmpxchg();
3243 for (i = 0; i < futex_hashsize; i++) {
3244 atomic_set(&futex_queues[i].waiters, 0);
3245 plist_head_init(&futex_queues[i].chain);
3246 spin_lock_init(&futex_queues[i].lock);
3251 __initcall(futex_init);