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 * Drops a reference to the pi_state object and frees or caches it
729 * when the last reference is gone.
731 * Must be called with the hb lock held.
733 static void put_pi_state(struct futex_pi_state *pi_state)
738 if (!atomic_dec_and_test(&pi_state->refcount))
742 * If pi_state->owner is NULL, the owner is most probably dying
743 * and has cleaned up the pi_state already
745 if (pi_state->owner) {
746 raw_spin_lock_irq(&pi_state->owner->pi_lock);
747 list_del_init(&pi_state->list);
748 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
750 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
753 if (current->pi_state_cache)
757 * pi_state->list is already empty.
758 * clear pi_state->owner.
759 * refcount is at 0 - put it back to 1.
761 pi_state->owner = NULL;
762 atomic_set(&pi_state->refcount, 1);
763 current->pi_state_cache = pi_state;
768 * Look up the task based on what TID userspace gave us.
771 static struct task_struct * futex_find_get_task(pid_t pid)
773 struct task_struct *p;
776 p = find_task_by_vpid(pid);
786 * This task is holding PI mutexes at exit time => bad.
787 * Kernel cleans up PI-state, but userspace is likely hosed.
788 * (Robust-futex cleanup is separate and might save the day for userspace.)
790 void exit_pi_state_list(struct task_struct *curr)
792 struct list_head *next, *head = &curr->pi_state_list;
793 struct futex_pi_state *pi_state;
794 struct futex_hash_bucket *hb;
795 union futex_key key = FUTEX_KEY_INIT;
797 if (!futex_cmpxchg_enabled)
800 * We are a ZOMBIE and nobody can enqueue itself on
801 * pi_state_list anymore, but we have to be careful
802 * versus waiters unqueueing themselves:
804 raw_spin_lock_irq(&curr->pi_lock);
805 while (!list_empty(head)) {
808 pi_state = list_entry(next, struct futex_pi_state, list);
810 hb = hash_futex(&key);
811 raw_spin_unlock_irq(&curr->pi_lock);
813 spin_lock(&hb->lock);
815 raw_spin_lock_irq(&curr->pi_lock);
817 * We dropped the pi-lock, so re-check whether this
818 * task still owns the PI-state:
820 if (head->next != next) {
821 spin_unlock(&hb->lock);
825 WARN_ON(pi_state->owner != curr);
826 WARN_ON(list_empty(&pi_state->list));
827 list_del_init(&pi_state->list);
828 pi_state->owner = NULL;
829 raw_spin_unlock_irq(&curr->pi_lock);
831 rt_mutex_unlock(&pi_state->pi_mutex);
833 spin_unlock(&hb->lock);
835 raw_spin_lock_irq(&curr->pi_lock);
837 raw_spin_unlock_irq(&curr->pi_lock);
841 * We need to check the following states:
843 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
845 * [1] NULL | --- | --- | 0 | 0/1 | Valid
846 * [2] NULL | --- | --- | >0 | 0/1 | Valid
848 * [3] Found | NULL | -- | Any | 0/1 | Invalid
850 * [4] Found | Found | NULL | 0 | 1 | Valid
851 * [5] Found | Found | NULL | >0 | 1 | Invalid
853 * [6] Found | Found | task | 0 | 1 | Valid
855 * [7] Found | Found | NULL | Any | 0 | Invalid
857 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
858 * [9] Found | Found | task | 0 | 0 | Invalid
859 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
861 * [1] Indicates that the kernel can acquire the futex atomically. We
862 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
864 * [2] Valid, if TID does not belong to a kernel thread. If no matching
865 * thread is found then it indicates that the owner TID has died.
867 * [3] Invalid. The waiter is queued on a non PI futex
869 * [4] Valid state after exit_robust_list(), which sets the user space
870 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
872 * [5] The user space value got manipulated between exit_robust_list()
873 * and exit_pi_state_list()
875 * [6] Valid state after exit_pi_state_list() which sets the new owner in
876 * the pi_state but cannot access the user space value.
878 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
880 * [8] Owner and user space value match
882 * [9] There is no transient state which sets the user space TID to 0
883 * except exit_robust_list(), but this is indicated by the
884 * FUTEX_OWNER_DIED bit. See [4]
886 * [10] There is no transient state which leaves owner and user space
891 * Validate that the existing waiter has a pi_state and sanity check
892 * the pi_state against the user space value. If correct, attach to
895 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
896 struct futex_pi_state **ps)
898 pid_t pid = uval & FUTEX_TID_MASK;
901 * Userspace might have messed up non-PI and PI futexes [3]
903 if (unlikely(!pi_state))
906 WARN_ON(!atomic_read(&pi_state->refcount));
909 * Handle the owner died case:
911 if (uval & FUTEX_OWNER_DIED) {
913 * exit_pi_state_list sets owner to NULL and wakes the
914 * topmost waiter. The task which acquires the
915 * pi_state->rt_mutex will fixup owner.
917 if (!pi_state->owner) {
919 * No pi state owner, but the user space TID
920 * is not 0. Inconsistent state. [5]
925 * Take a ref on the state and return success. [4]
931 * If TID is 0, then either the dying owner has not
932 * yet executed exit_pi_state_list() or some waiter
933 * acquired the rtmutex in the pi state, but did not
934 * yet fixup the TID in user space.
936 * Take a ref on the state and return success. [6]
942 * If the owner died bit is not set, then the pi_state
943 * must have an owner. [7]
945 if (!pi_state->owner)
950 * Bail out if user space manipulated the futex value. If pi
951 * state exists then the owner TID must be the same as the
952 * user space TID. [9/10]
954 if (pid != task_pid_vnr(pi_state->owner))
957 atomic_inc(&pi_state->refcount);
963 * Lookup the task for the TID provided from user space and attach to
964 * it after doing proper sanity checks.
966 static int attach_to_pi_owner(u32 uval, union futex_key *key,
967 struct futex_pi_state **ps)
969 pid_t pid = uval & FUTEX_TID_MASK;
970 struct futex_pi_state *pi_state;
971 struct task_struct *p;
974 * We are the first waiter - try to look up the real owner and attach
975 * the new pi_state to it, but bail out when TID = 0 [1]
979 p = futex_find_get_task(pid);
983 if (unlikely(p->flags & PF_KTHREAD)) {
989 * We need to look at the task state flags to figure out,
990 * whether the task is exiting. To protect against the do_exit
991 * change of the task flags, we do this protected by
994 raw_spin_lock_irq(&p->pi_lock);
995 if (unlikely(p->flags & PF_EXITING)) {
997 * The task is on the way out. When PF_EXITPIDONE is
998 * set, we know that the task has finished the
1001 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1003 raw_spin_unlock_irq(&p->pi_lock);
1009 * No existing pi state. First waiter. [2]
1011 pi_state = alloc_pi_state();
1014 * Initialize the pi_mutex in locked state and make @p
1017 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1019 /* Store the key for possible exit cleanups: */
1020 pi_state->key = *key;
1022 WARN_ON(!list_empty(&pi_state->list));
1023 list_add(&pi_state->list, &p->pi_state_list);
1024 pi_state->owner = p;
1025 raw_spin_unlock_irq(&p->pi_lock);
1034 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1035 union futex_key *key, struct futex_pi_state **ps)
1037 struct futex_q *match = futex_top_waiter(hb, key);
1040 * If there is a waiter on that futex, validate it and
1041 * attach to the pi_state when the validation succeeds.
1044 return attach_to_pi_state(uval, match->pi_state, ps);
1047 * We are the first waiter - try to look up the owner based on
1048 * @uval and attach to it.
1050 return attach_to_pi_owner(uval, key, ps);
1053 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1055 u32 uninitialized_var(curval);
1057 if (unlikely(should_fail_futex(true)))
1060 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1063 /*If user space value changed, let the caller retry */
1064 return curval != uval ? -EAGAIN : 0;
1068 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1069 * @uaddr: the pi futex user address
1070 * @hb: the pi futex hash bucket
1071 * @key: the futex key associated with uaddr and hb
1072 * @ps: the pi_state pointer where we store the result of the
1074 * @task: the task to perform the atomic lock work for. This will
1075 * be "current" except in the case of requeue pi.
1076 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1079 * 0 - ready to wait;
1080 * 1 - acquired the lock;
1083 * The hb->lock and futex_key refs shall be held by the caller.
1085 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1086 union futex_key *key,
1087 struct futex_pi_state **ps,
1088 struct task_struct *task, int set_waiters)
1090 u32 uval, newval, vpid = task_pid_vnr(task);
1091 struct futex_q *match;
1095 * Read the user space value first so we can validate a few
1096 * things before proceeding further.
1098 if (get_futex_value_locked(&uval, uaddr))
1101 if (unlikely(should_fail_futex(true)))
1107 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1110 if ((unlikely(should_fail_futex(true))))
1114 * Lookup existing state first. If it exists, try to attach to
1117 match = futex_top_waiter(hb, key);
1119 return attach_to_pi_state(uval, match->pi_state, ps);
1122 * No waiter and user TID is 0. We are here because the
1123 * waiters or the owner died bit is set or called from
1124 * requeue_cmp_pi or for whatever reason something took the
1127 if (!(uval & FUTEX_TID_MASK)) {
1129 * We take over the futex. No other waiters and the user space
1130 * TID is 0. We preserve the owner died bit.
1132 newval = uval & FUTEX_OWNER_DIED;
1135 /* The futex requeue_pi code can enforce the waiters bit */
1137 newval |= FUTEX_WAITERS;
1139 ret = lock_pi_update_atomic(uaddr, uval, newval);
1140 /* If the take over worked, return 1 */
1141 return ret < 0 ? ret : 1;
1145 * First waiter. Set the waiters bit before attaching ourself to
1146 * the owner. If owner tries to unlock, it will be forced into
1147 * the kernel and blocked on hb->lock.
1149 newval = uval | FUTEX_WAITERS;
1150 ret = lock_pi_update_atomic(uaddr, uval, newval);
1154 * If the update of the user space value succeeded, we try to
1155 * attach to the owner. If that fails, no harm done, we only
1156 * set the FUTEX_WAITERS bit in the user space variable.
1158 return attach_to_pi_owner(uval, key, ps);
1162 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1163 * @q: The futex_q to unqueue
1165 * The q->lock_ptr must not be NULL and must be held by the caller.
1167 static void __unqueue_futex(struct futex_q *q)
1169 struct futex_hash_bucket *hb;
1171 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1172 || WARN_ON(plist_node_empty(&q->list)))
1175 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1176 plist_del(&q->list, &hb->chain);
1181 * The hash bucket lock must be held when this is called.
1182 * Afterwards, the futex_q must not be accessed. Callers
1183 * must ensure to later call wake_up_q() for the actual
1186 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1188 struct task_struct *p = q->task;
1190 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1194 * Queue the task for later wakeup for after we've released
1195 * the hb->lock. wake_q_add() grabs reference to p.
1197 wake_q_add(wake_q, p);
1200 * The waiting task can free the futex_q as soon as
1201 * q->lock_ptr = NULL is written, without taking any locks. A
1202 * memory barrier is required here to prevent the following
1203 * store to lock_ptr from getting ahead of the plist_del.
1209 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1210 struct futex_hash_bucket *hb)
1212 struct task_struct *new_owner;
1213 struct futex_pi_state *pi_state = this->pi_state;
1214 u32 uninitialized_var(curval), newval;
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(&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)
1255 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1259 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1260 WARN_ON(list_empty(&pi_state->list));
1261 list_del_init(&pi_state->list);
1262 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1264 raw_spin_lock_irq(&new_owner->pi_lock);
1265 WARN_ON(!list_empty(&pi_state->list));
1266 list_add(&pi_state->list, &new_owner->pi_state_list);
1267 pi_state->owner = new_owner;
1268 raw_spin_unlock_irq(&new_owner->pi_lock);
1270 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1272 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1275 * First unlock HB so the waiter does not spin on it once he got woken
1276 * up. Second wake up the waiter before the priority is adjusted. If we
1277 * deboost first (and lose our higher priority), then the task might get
1278 * scheduled away before the wake up can take place.
1280 spin_unlock(&hb->lock);
1283 rt_mutex_adjust_prio(current);
1289 * Express the locking dependencies for lockdep:
1292 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1295 spin_lock(&hb1->lock);
1297 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1298 } else { /* hb1 > hb2 */
1299 spin_lock(&hb2->lock);
1300 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1305 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1307 spin_unlock(&hb1->lock);
1309 spin_unlock(&hb2->lock);
1313 * Wake up waiters matching bitset queued on this futex (uaddr).
1316 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1318 struct futex_hash_bucket *hb;
1319 struct futex_q *this, *next;
1320 union futex_key key = FUTEX_KEY_INIT;
1327 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1328 if (unlikely(ret != 0))
1331 hb = hash_futex(&key);
1333 /* Make sure we really have tasks to wakeup */
1334 if (!hb_waiters_pending(hb))
1337 spin_lock(&hb->lock);
1339 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1340 if (match_futex (&this->key, &key)) {
1341 if (this->pi_state || this->rt_waiter) {
1346 /* Check if one of the bits is set in both bitsets */
1347 if (!(this->bitset & bitset))
1350 mark_wake_futex(&wake_q, this);
1351 if (++ret >= nr_wake)
1356 spin_unlock(&hb->lock);
1359 put_futex_key(&key);
1365 * Wake up all waiters hashed on the physical page that is mapped
1366 * to this virtual address:
1369 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1370 int nr_wake, int nr_wake2, int op)
1372 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1373 struct futex_hash_bucket *hb1, *hb2;
1374 struct futex_q *this, *next;
1379 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1380 if (unlikely(ret != 0))
1382 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1383 if (unlikely(ret != 0))
1386 hb1 = hash_futex(&key1);
1387 hb2 = hash_futex(&key2);
1390 double_lock_hb(hb1, hb2);
1391 op_ret = futex_atomic_op_inuser(op, uaddr2);
1392 if (unlikely(op_ret < 0)) {
1394 double_unlock_hb(hb1, hb2);
1398 * we don't get EFAULT from MMU faults if we don't have an MMU,
1399 * but we might get them from range checking
1405 if (unlikely(op_ret != -EFAULT)) {
1410 ret = fault_in_user_writeable(uaddr2);
1414 if (!(flags & FLAGS_SHARED))
1417 put_futex_key(&key2);
1418 put_futex_key(&key1);
1422 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1423 if (match_futex (&this->key, &key1)) {
1424 if (this->pi_state || this->rt_waiter) {
1428 mark_wake_futex(&wake_q, this);
1429 if (++ret >= nr_wake)
1436 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1437 if (match_futex (&this->key, &key2)) {
1438 if (this->pi_state || this->rt_waiter) {
1442 mark_wake_futex(&wake_q, this);
1443 if (++op_ret >= nr_wake2)
1451 double_unlock_hb(hb1, hb2);
1454 put_futex_key(&key2);
1456 put_futex_key(&key1);
1462 * requeue_futex() - Requeue a futex_q from one hb to another
1463 * @q: the futex_q to requeue
1464 * @hb1: the source hash_bucket
1465 * @hb2: the target hash_bucket
1466 * @key2: the new key for the requeued futex_q
1469 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1470 struct futex_hash_bucket *hb2, union futex_key *key2)
1474 * If key1 and key2 hash to the same bucket, no need to
1477 if (likely(&hb1->chain != &hb2->chain)) {
1478 plist_del(&q->list, &hb1->chain);
1479 hb_waiters_dec(hb1);
1480 plist_add(&q->list, &hb2->chain);
1481 hb_waiters_inc(hb2);
1482 q->lock_ptr = &hb2->lock;
1484 get_futex_key_refs(key2);
1489 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1491 * @key: the key of the requeue target futex
1492 * @hb: the hash_bucket of the requeue target futex
1494 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1495 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1496 * to the requeue target futex so the waiter can detect the wakeup on the right
1497 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1498 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1499 * to protect access to the pi_state to fixup the owner later. Must be called
1500 * with both q->lock_ptr and hb->lock held.
1503 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1504 struct futex_hash_bucket *hb)
1506 get_futex_key_refs(key);
1511 WARN_ON(!q->rt_waiter);
1512 q->rt_waiter = NULL;
1514 q->lock_ptr = &hb->lock;
1516 wake_up_state(q->task, TASK_NORMAL);
1520 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1521 * @pifutex: the user address of the to futex
1522 * @hb1: the from futex hash bucket, must be locked by the caller
1523 * @hb2: the to futex hash bucket, must be locked by the caller
1524 * @key1: the from futex key
1525 * @key2: the to futex key
1526 * @ps: address to store the pi_state pointer
1527 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1529 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1530 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1531 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1532 * hb1 and hb2 must be held by the caller.
1535 * 0 - failed to acquire the lock atomically;
1536 * >0 - acquired the lock, return value is vpid of the top_waiter
1539 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1540 struct futex_hash_bucket *hb1,
1541 struct futex_hash_bucket *hb2,
1542 union futex_key *key1, union futex_key *key2,
1543 struct futex_pi_state **ps, int set_waiters)
1545 struct futex_q *top_waiter = NULL;
1549 if (get_futex_value_locked(&curval, pifutex))
1552 if (unlikely(should_fail_futex(true)))
1556 * Find the top_waiter and determine if there are additional waiters.
1557 * If the caller intends to requeue more than 1 waiter to pifutex,
1558 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1559 * as we have means to handle the possible fault. If not, don't set
1560 * the bit unecessarily as it will force the subsequent unlock to enter
1563 top_waiter = futex_top_waiter(hb1, key1);
1565 /* There are no waiters, nothing for us to do. */
1569 /* Ensure we requeue to the expected futex. */
1570 if (!match_futex(top_waiter->requeue_pi_key, key2))
1574 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1575 * the contended case or if set_waiters is 1. The pi_state is returned
1576 * in ps in contended cases.
1578 vpid = task_pid_vnr(top_waiter->task);
1579 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1582 requeue_pi_wake_futex(top_waiter, key2, hb2);
1589 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1590 * @uaddr1: source futex user address
1591 * @flags: futex flags (FLAGS_SHARED, etc.)
1592 * @uaddr2: target futex user address
1593 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1594 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1595 * @cmpval: @uaddr1 expected value (or %NULL)
1596 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1597 * pi futex (pi to pi requeue is not supported)
1599 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1600 * uaddr2 atomically on behalf of the top waiter.
1603 * >=0 - on success, the number of tasks requeued or woken;
1606 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1607 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1608 u32 *cmpval, int requeue_pi)
1610 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1611 int drop_count = 0, task_count = 0, ret;
1612 struct futex_pi_state *pi_state = NULL;
1613 struct futex_hash_bucket *hb1, *hb2;
1614 struct futex_q *this, *next;
1619 * Requeue PI only works on two distinct uaddrs. This
1620 * check is only valid for private futexes. See below.
1622 if (uaddr1 == uaddr2)
1626 * requeue_pi requires a pi_state, try to allocate it now
1627 * without any locks in case it fails.
1629 if (refill_pi_state_cache())
1632 * requeue_pi must wake as many tasks as it can, up to nr_wake
1633 * + nr_requeue, since it acquires the rt_mutex prior to
1634 * returning to userspace, so as to not leave the rt_mutex with
1635 * waiters and no owner. However, second and third wake-ups
1636 * cannot be predicted as they involve race conditions with the
1637 * first wake and a fault while looking up the pi_state. Both
1638 * pthread_cond_signal() and pthread_cond_broadcast() should
1646 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1647 if (unlikely(ret != 0))
1649 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1650 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1651 if (unlikely(ret != 0))
1655 * The check above which compares uaddrs is not sufficient for
1656 * shared futexes. We need to compare the keys:
1658 if (requeue_pi && match_futex(&key1, &key2)) {
1663 hb1 = hash_futex(&key1);
1664 hb2 = hash_futex(&key2);
1667 hb_waiters_inc(hb2);
1668 double_lock_hb(hb1, hb2);
1670 if (likely(cmpval != NULL)) {
1673 ret = get_futex_value_locked(&curval, uaddr1);
1675 if (unlikely(ret)) {
1676 double_unlock_hb(hb1, hb2);
1677 hb_waiters_dec(hb2);
1679 ret = get_user(curval, uaddr1);
1683 if (!(flags & FLAGS_SHARED))
1686 put_futex_key(&key2);
1687 put_futex_key(&key1);
1690 if (curval != *cmpval) {
1696 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1698 * Attempt to acquire uaddr2 and wake the top waiter. If we
1699 * intend to requeue waiters, force setting the FUTEX_WAITERS
1700 * bit. We force this here where we are able to easily handle
1701 * faults rather in the requeue loop below.
1703 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1704 &key2, &pi_state, nr_requeue);
1707 * At this point the top_waiter has either taken uaddr2 or is
1708 * waiting on it. If the former, then the pi_state will not
1709 * exist yet, look it up one more time to ensure we have a
1710 * reference to it. If the lock was taken, ret contains the
1711 * vpid of the top waiter task.
1718 * If we acquired the lock, then the user
1719 * space value of uaddr2 should be vpid. It
1720 * cannot be changed by the top waiter as it
1721 * is blocked on hb2 lock if it tries to do
1722 * so. If something fiddled with it behind our
1723 * back the pi state lookup might unearth
1724 * it. So we rather use the known value than
1725 * rereading and handing potential crap to
1728 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1735 put_pi_state(pi_state);
1737 double_unlock_hb(hb1, hb2);
1738 hb_waiters_dec(hb2);
1739 put_futex_key(&key2);
1740 put_futex_key(&key1);
1741 ret = fault_in_user_writeable(uaddr2);
1747 * Two reasons for this:
1748 * - Owner is exiting and we just wait for the
1750 * - The user space value changed.
1752 put_pi_state(pi_state);
1754 double_unlock_hb(hb1, hb2);
1755 hb_waiters_dec(hb2);
1756 put_futex_key(&key2);
1757 put_futex_key(&key1);
1765 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1766 if (task_count - nr_wake >= nr_requeue)
1769 if (!match_futex(&this->key, &key1))
1773 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1774 * be paired with each other and no other futex ops.
1776 * We should never be requeueing a futex_q with a pi_state,
1777 * which is awaiting a futex_unlock_pi().
1779 if ((requeue_pi && !this->rt_waiter) ||
1780 (!requeue_pi && this->rt_waiter) ||
1787 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1788 * lock, we already woke the top_waiter. If not, it will be
1789 * woken by futex_unlock_pi().
1791 if (++task_count <= nr_wake && !requeue_pi) {
1792 mark_wake_futex(&wake_q, this);
1796 /* Ensure we requeue to the expected futex for requeue_pi. */
1797 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1803 * Requeue nr_requeue waiters and possibly one more in the case
1804 * of requeue_pi if we couldn't acquire the lock atomically.
1807 /* Prepare the waiter to take the rt_mutex. */
1808 atomic_inc(&pi_state->refcount);
1809 this->pi_state = pi_state;
1810 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1814 /* We got the lock. */
1815 requeue_pi_wake_futex(this, &key2, hb2);
1820 this->pi_state = NULL;
1821 put_pi_state(pi_state);
1825 requeue_futex(this, hb1, hb2, &key2);
1830 put_pi_state(pi_state);
1831 double_unlock_hb(hb1, hb2);
1833 hb_waiters_dec(hb2);
1836 * drop_futex_key_refs() must be called outside the spinlocks. During
1837 * the requeue we moved futex_q's from the hash bucket at key1 to the
1838 * one at key2 and updated their key pointer. We no longer need to
1839 * hold the references to key1.
1841 while (--drop_count >= 0)
1842 drop_futex_key_refs(&key1);
1845 put_futex_key(&key2);
1847 put_futex_key(&key1);
1849 return ret ? ret : task_count;
1852 /* The key must be already stored in q->key. */
1853 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1854 __acquires(&hb->lock)
1856 struct futex_hash_bucket *hb;
1858 hb = hash_futex(&q->key);
1861 * Increment the counter before taking the lock so that
1862 * a potential waker won't miss a to-be-slept task that is
1863 * waiting for the spinlock. This is safe as all queue_lock()
1864 * users end up calling queue_me(). Similarly, for housekeeping,
1865 * decrement the counter at queue_unlock() when some error has
1866 * occurred and we don't end up adding the task to the list.
1870 q->lock_ptr = &hb->lock;
1872 spin_lock(&hb->lock); /* implies MB (A) */
1877 queue_unlock(struct futex_hash_bucket *hb)
1878 __releases(&hb->lock)
1880 spin_unlock(&hb->lock);
1885 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1886 * @q: The futex_q to enqueue
1887 * @hb: The destination hash bucket
1889 * The hb->lock must be held by the caller, and is released here. A call to
1890 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1891 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1892 * or nothing if the unqueue is done as part of the wake process and the unqueue
1893 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1896 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1897 __releases(&hb->lock)
1902 * The priority used to register this element is
1903 * - either the real thread-priority for the real-time threads
1904 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1905 * - or MAX_RT_PRIO for non-RT threads.
1906 * Thus, all RT-threads are woken first in priority order, and
1907 * the others are woken last, in FIFO order.
1909 prio = min(current->normal_prio, MAX_RT_PRIO);
1911 plist_node_init(&q->list, prio);
1912 plist_add(&q->list, &hb->chain);
1914 spin_unlock(&hb->lock);
1918 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1919 * @q: The futex_q to unqueue
1921 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1922 * be paired with exactly one earlier call to queue_me().
1925 * 1 - if the futex_q was still queued (and we removed unqueued it);
1926 * 0 - if the futex_q was already removed by the waking thread
1928 static int unqueue_me(struct futex_q *q)
1930 spinlock_t *lock_ptr;
1933 /* In the common case we don't take the spinlock, which is nice. */
1935 lock_ptr = q->lock_ptr;
1937 if (lock_ptr != NULL) {
1938 spin_lock(lock_ptr);
1940 * q->lock_ptr can change between reading it and
1941 * spin_lock(), causing us to take the wrong lock. This
1942 * corrects the race condition.
1944 * Reasoning goes like this: if we have the wrong lock,
1945 * q->lock_ptr must have changed (maybe several times)
1946 * between reading it and the spin_lock(). It can
1947 * change again after the spin_lock() but only if it was
1948 * already changed before the spin_lock(). It cannot,
1949 * however, change back to the original value. Therefore
1950 * we can detect whether we acquired the correct lock.
1952 if (unlikely(lock_ptr != q->lock_ptr)) {
1953 spin_unlock(lock_ptr);
1958 BUG_ON(q->pi_state);
1960 spin_unlock(lock_ptr);
1964 drop_futex_key_refs(&q->key);
1969 * PI futexes can not be requeued and must remove themself from the
1970 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1973 static void unqueue_me_pi(struct futex_q *q)
1974 __releases(q->lock_ptr)
1978 BUG_ON(!q->pi_state);
1979 put_pi_state(q->pi_state);
1982 spin_unlock(q->lock_ptr);
1986 * Fixup the pi_state owner with the new owner.
1988 * Must be called with hash bucket lock held and mm->sem held for non
1991 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1992 struct task_struct *newowner)
1994 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1995 struct futex_pi_state *pi_state = q->pi_state;
1996 struct task_struct *oldowner = pi_state->owner;
1997 u32 uval, uninitialized_var(curval), newval;
2001 if (!pi_state->owner)
2002 newtid |= FUTEX_OWNER_DIED;
2005 * We are here either because we stole the rtmutex from the
2006 * previous highest priority waiter or we are the highest priority
2007 * waiter but failed to get the rtmutex the first time.
2008 * We have to replace the newowner TID in the user space variable.
2009 * This must be atomic as we have to preserve the owner died bit here.
2011 * Note: We write the user space value _before_ changing the pi_state
2012 * because we can fault here. Imagine swapped out pages or a fork
2013 * that marked all the anonymous memory readonly for cow.
2015 * Modifying pi_state _before_ the user space value would
2016 * leave the pi_state in an inconsistent state when we fault
2017 * here, because we need to drop the hash bucket lock to
2018 * handle the fault. This might be observed in the PID check
2019 * in lookup_pi_state.
2022 if (get_futex_value_locked(&uval, uaddr))
2026 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2028 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2036 * We fixed up user space. Now we need to fix the pi_state
2039 if (pi_state->owner != NULL) {
2040 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2041 WARN_ON(list_empty(&pi_state->list));
2042 list_del_init(&pi_state->list);
2043 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2046 pi_state->owner = newowner;
2048 raw_spin_lock_irq(&newowner->pi_lock);
2049 WARN_ON(!list_empty(&pi_state->list));
2050 list_add(&pi_state->list, &newowner->pi_state_list);
2051 raw_spin_unlock_irq(&newowner->pi_lock);
2055 * To handle the page fault we need to drop the hash bucket
2056 * lock here. That gives the other task (either the highest priority
2057 * waiter itself or the task which stole the rtmutex) the
2058 * chance to try the fixup of the pi_state. So once we are
2059 * back from handling the fault we need to check the pi_state
2060 * after reacquiring the hash bucket lock and before trying to
2061 * do another fixup. When the fixup has been done already we
2065 spin_unlock(q->lock_ptr);
2067 ret = fault_in_user_writeable(uaddr);
2069 spin_lock(q->lock_ptr);
2072 * Check if someone else fixed it for us:
2074 if (pi_state->owner != oldowner)
2083 static long futex_wait_restart(struct restart_block *restart);
2086 * fixup_owner() - Post lock pi_state and corner case management
2087 * @uaddr: user address of the futex
2088 * @q: futex_q (contains pi_state and access to the rt_mutex)
2089 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2091 * After attempting to lock an rt_mutex, this function is called to cleanup
2092 * the pi_state owner as well as handle race conditions that may allow us to
2093 * acquire the lock. Must be called with the hb lock held.
2096 * 1 - success, lock taken;
2097 * 0 - success, lock not taken;
2098 * <0 - on error (-EFAULT)
2100 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2102 struct task_struct *owner;
2107 * Got the lock. We might not be the anticipated owner if we
2108 * did a lock-steal - fix up the PI-state in that case:
2110 if (q->pi_state->owner != current)
2111 ret = fixup_pi_state_owner(uaddr, q, current);
2116 * Catch the rare case, where the lock was released when we were on the
2117 * way back before we locked the hash bucket.
2119 if (q->pi_state->owner == current) {
2121 * Try to get the rt_mutex now. This might fail as some other
2122 * task acquired the rt_mutex after we removed ourself from the
2123 * rt_mutex waiters list.
2125 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2131 * pi_state is incorrect, some other task did a lock steal and
2132 * we returned due to timeout or signal without taking the
2133 * rt_mutex. Too late.
2135 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2136 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2138 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2139 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2140 ret = fixup_pi_state_owner(uaddr, q, owner);
2145 * Paranoia check. If we did not take the lock, then we should not be
2146 * the owner of the rt_mutex.
2148 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2149 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2150 "pi-state %p\n", ret,
2151 q->pi_state->pi_mutex.owner,
2152 q->pi_state->owner);
2155 return ret ? ret : locked;
2159 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2160 * @hb: the futex hash bucket, must be locked by the caller
2161 * @q: the futex_q to queue up on
2162 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2164 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2165 struct hrtimer_sleeper *timeout)
2168 * The task state is guaranteed to be set before another task can
2169 * wake it. set_current_state() is implemented using smp_store_mb() and
2170 * queue_me() calls spin_unlock() upon completion, both serializing
2171 * access to the hash list and forcing another memory barrier.
2173 set_current_state(TASK_INTERRUPTIBLE);
2178 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2181 * If we have been removed from the hash list, then another task
2182 * has tried to wake us, and we can skip the call to schedule().
2184 if (likely(!plist_node_empty(&q->list))) {
2186 * If the timer has already expired, current will already be
2187 * flagged for rescheduling. Only call schedule if there
2188 * is no timeout, or if it has yet to expire.
2190 if (!timeout || timeout->task)
2191 freezable_schedule();
2193 __set_current_state(TASK_RUNNING);
2197 * futex_wait_setup() - Prepare to wait on a futex
2198 * @uaddr: the futex userspace address
2199 * @val: the expected value
2200 * @flags: futex flags (FLAGS_SHARED, etc.)
2201 * @q: the associated futex_q
2202 * @hb: storage for hash_bucket pointer to be returned to caller
2204 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2205 * compare it with the expected value. Handle atomic faults internally.
2206 * Return with the hb lock held and a q.key reference on success, and unlocked
2207 * with no q.key reference on failure.
2210 * 0 - uaddr contains val and hb has been locked;
2211 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2213 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2214 struct futex_q *q, struct futex_hash_bucket **hb)
2220 * Access the page AFTER the hash-bucket is locked.
2221 * Order is important:
2223 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2224 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2226 * The basic logical guarantee of a futex is that it blocks ONLY
2227 * if cond(var) is known to be true at the time of blocking, for
2228 * any cond. If we locked the hash-bucket after testing *uaddr, that
2229 * would open a race condition where we could block indefinitely with
2230 * cond(var) false, which would violate the guarantee.
2232 * On the other hand, we insert q and release the hash-bucket only
2233 * after testing *uaddr. This guarantees that futex_wait() will NOT
2234 * absorb a wakeup if *uaddr does not match the desired values
2235 * while the syscall executes.
2238 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2239 if (unlikely(ret != 0))
2243 *hb = queue_lock(q);
2245 ret = get_futex_value_locked(&uval, uaddr);
2250 ret = get_user(uval, uaddr);
2254 if (!(flags & FLAGS_SHARED))
2257 put_futex_key(&q->key);
2268 put_futex_key(&q->key);
2272 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2273 ktime_t *abs_time, u32 bitset)
2275 struct hrtimer_sleeper timeout, *to = NULL;
2276 struct restart_block *restart;
2277 struct futex_hash_bucket *hb;
2278 struct futex_q q = futex_q_init;
2288 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2289 CLOCK_REALTIME : CLOCK_MONOTONIC,
2291 hrtimer_init_sleeper(to, current);
2292 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2293 current->timer_slack_ns);
2298 * Prepare to wait on uaddr. On success, holds hb lock and increments
2301 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2305 /* queue_me and wait for wakeup, timeout, or a signal. */
2306 futex_wait_queue_me(hb, &q, to);
2308 /* If we were woken (and unqueued), we succeeded, whatever. */
2310 /* unqueue_me() drops q.key ref */
2311 if (!unqueue_me(&q))
2314 if (to && !to->task)
2318 * We expect signal_pending(current), but we might be the
2319 * victim of a spurious wakeup as well.
2321 if (!signal_pending(current))
2328 restart = ¤t->restart_block;
2329 restart->fn = futex_wait_restart;
2330 restart->futex.uaddr = uaddr;
2331 restart->futex.val = val;
2332 restart->futex.time = abs_time->tv64;
2333 restart->futex.bitset = bitset;
2334 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2336 ret = -ERESTART_RESTARTBLOCK;
2340 hrtimer_cancel(&to->timer);
2341 destroy_hrtimer_on_stack(&to->timer);
2347 static long futex_wait_restart(struct restart_block *restart)
2349 u32 __user *uaddr = restart->futex.uaddr;
2350 ktime_t t, *tp = NULL;
2352 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2353 t.tv64 = restart->futex.time;
2356 restart->fn = do_no_restart_syscall;
2358 return (long)futex_wait(uaddr, restart->futex.flags,
2359 restart->futex.val, tp, restart->futex.bitset);
2364 * Userspace tried a 0 -> TID atomic transition of the futex value
2365 * and failed. The kernel side here does the whole locking operation:
2366 * if there are waiters then it will block as a consequence of relying
2367 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2368 * a 0 value of the futex too.).
2370 * Also serves as futex trylock_pi()'ing, and due semantics.
2372 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2373 ktime_t *time, int trylock)
2375 struct hrtimer_sleeper timeout, *to = NULL;
2376 struct futex_hash_bucket *hb;
2377 struct futex_q q = futex_q_init;
2380 if (refill_pi_state_cache())
2385 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2387 hrtimer_init_sleeper(to, current);
2388 hrtimer_set_expires(&to->timer, *time);
2392 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2393 if (unlikely(ret != 0))
2397 hb = queue_lock(&q);
2399 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2400 if (unlikely(ret)) {
2402 * Atomic work succeeded and we got the lock,
2403 * or failed. Either way, we do _not_ block.
2407 /* We got the lock. */
2409 goto out_unlock_put_key;
2414 * Two reasons for this:
2415 * - Task is exiting and we just wait for the
2417 * - The user space value changed.
2420 put_futex_key(&q.key);
2424 goto out_unlock_put_key;
2429 * Only actually queue now that the atomic ops are done:
2433 WARN_ON(!q.pi_state);
2435 * Block on the PI mutex:
2438 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2440 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2441 /* Fixup the trylock return value: */
2442 ret = ret ? 0 : -EWOULDBLOCK;
2445 spin_lock(q.lock_ptr);
2447 * Fixup the pi_state owner and possibly acquire the lock if we
2450 res = fixup_owner(uaddr, &q, !ret);
2452 * If fixup_owner() returned an error, proprogate that. If it acquired
2453 * the lock, clear our -ETIMEDOUT or -EINTR.
2456 ret = (res < 0) ? res : 0;
2459 * If fixup_owner() faulted and was unable to handle the fault, unlock
2460 * it and return the fault to userspace.
2462 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2463 rt_mutex_unlock(&q.pi_state->pi_mutex);
2465 /* Unqueue and drop the lock */
2474 put_futex_key(&q.key);
2477 destroy_hrtimer_on_stack(&to->timer);
2478 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2483 ret = fault_in_user_writeable(uaddr);
2487 if (!(flags & FLAGS_SHARED))
2490 put_futex_key(&q.key);
2495 * Userspace attempted a TID -> 0 atomic transition, and failed.
2496 * This is the in-kernel slowpath: we look up the PI state (if any),
2497 * and do the rt-mutex unlock.
2499 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2501 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2502 union futex_key key = FUTEX_KEY_INIT;
2503 struct futex_hash_bucket *hb;
2504 struct futex_q *match;
2508 if (get_user(uval, uaddr))
2511 * We release only a lock we actually own:
2513 if ((uval & FUTEX_TID_MASK) != vpid)
2516 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2520 hb = hash_futex(&key);
2521 spin_lock(&hb->lock);
2524 * Check waiters first. We do not trust user space values at
2525 * all and we at least want to know if user space fiddled
2526 * with the futex value instead of blindly unlocking.
2528 match = futex_top_waiter(hb, &key);
2530 ret = wake_futex_pi(uaddr, uval, match, hb);
2532 * In case of success wake_futex_pi dropped the hash
2538 * The atomic access to the futex value generated a
2539 * pagefault, so retry the user-access and the wakeup:
2544 * wake_futex_pi has detected invalid state. Tell user
2551 * We have no kernel internal state, i.e. no waiters in the
2552 * kernel. Waiters which are about to queue themselves are stuck
2553 * on hb->lock. So we can safely ignore them. We do neither
2554 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2557 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2561 * If uval has changed, let user space handle it.
2563 ret = (curval == uval) ? 0 : -EAGAIN;
2566 spin_unlock(&hb->lock);
2568 put_futex_key(&key);
2572 spin_unlock(&hb->lock);
2573 put_futex_key(&key);
2575 ret = fault_in_user_writeable(uaddr);
2583 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2584 * @hb: the hash_bucket futex_q was original enqueued on
2585 * @q: the futex_q woken while waiting to be requeued
2586 * @key2: the futex_key of the requeue target futex
2587 * @timeout: the timeout associated with the wait (NULL if none)
2589 * Detect if the task was woken on the initial futex as opposed to the requeue
2590 * target futex. If so, determine if it was a timeout or a signal that caused
2591 * the wakeup and return the appropriate error code to the caller. Must be
2592 * called with the hb lock held.
2595 * 0 = no early wakeup detected;
2596 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2599 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2600 struct futex_q *q, union futex_key *key2,
2601 struct hrtimer_sleeper *timeout)
2606 * With the hb lock held, we avoid races while we process the wakeup.
2607 * We only need to hold hb (and not hb2) to ensure atomicity as the
2608 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2609 * It can't be requeued from uaddr2 to something else since we don't
2610 * support a PI aware source futex for requeue.
2612 if (!match_futex(&q->key, key2)) {
2613 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2615 * We were woken prior to requeue by a timeout or a signal.
2616 * Unqueue the futex_q and determine which it was.
2618 plist_del(&q->list, &hb->chain);
2621 /* Handle spurious wakeups gracefully */
2623 if (timeout && !timeout->task)
2625 else if (signal_pending(current))
2626 ret = -ERESTARTNOINTR;
2632 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2633 * @uaddr: the futex we initially wait on (non-pi)
2634 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2635 * the same type, no requeueing from private to shared, etc.
2636 * @val: the expected value of uaddr
2637 * @abs_time: absolute timeout
2638 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2639 * @uaddr2: the pi futex we will take prior to returning to user-space
2641 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2642 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2643 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2644 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2645 * without one, the pi logic would not know which task to boost/deboost, if
2646 * there was a need to.
2648 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2649 * via the following--
2650 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2651 * 2) wakeup on uaddr2 after a requeue
2655 * If 3, cleanup and return -ERESTARTNOINTR.
2657 * If 2, we may then block on trying to take the rt_mutex and return via:
2658 * 5) successful lock
2661 * 8) other lock acquisition failure
2663 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2665 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2671 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2672 u32 val, ktime_t *abs_time, u32 bitset,
2675 struct hrtimer_sleeper timeout, *to = NULL;
2676 struct rt_mutex_waiter rt_waiter;
2677 struct rt_mutex *pi_mutex = NULL;
2678 struct futex_hash_bucket *hb;
2679 union futex_key key2 = FUTEX_KEY_INIT;
2680 struct futex_q q = futex_q_init;
2683 if (uaddr == uaddr2)
2691 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2692 CLOCK_REALTIME : CLOCK_MONOTONIC,
2694 hrtimer_init_sleeper(to, current);
2695 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2696 current->timer_slack_ns);
2700 * The waiter is allocated on our stack, manipulated by the requeue
2701 * code while we sleep on uaddr.
2703 debug_rt_mutex_init_waiter(&rt_waiter);
2704 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2705 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2706 rt_waiter.task = NULL;
2708 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2709 if (unlikely(ret != 0))
2713 q.rt_waiter = &rt_waiter;
2714 q.requeue_pi_key = &key2;
2717 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2720 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2725 * The check above which compares uaddrs is not sufficient for
2726 * shared futexes. We need to compare the keys:
2728 if (match_futex(&q.key, &key2)) {
2734 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2735 futex_wait_queue_me(hb, &q, to);
2737 spin_lock(&hb->lock);
2738 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2739 spin_unlock(&hb->lock);
2744 * In order for us to be here, we know our q.key == key2, and since
2745 * we took the hb->lock above, we also know that futex_requeue() has
2746 * completed and we no longer have to concern ourselves with a wakeup
2747 * race with the atomic proxy lock acquisition by the requeue code. The
2748 * futex_requeue dropped our key1 reference and incremented our key2
2752 /* Check if the requeue code acquired the second futex for us. */
2755 * Got the lock. We might not be the anticipated owner if we
2756 * did a lock-steal - fix up the PI-state in that case.
2758 if (q.pi_state && (q.pi_state->owner != current)) {
2759 spin_lock(q.lock_ptr);
2760 ret = fixup_pi_state_owner(uaddr2, &q, current);
2762 * Drop the reference to the pi state which
2763 * the requeue_pi() code acquired for us.
2765 put_pi_state(q.pi_state);
2766 spin_unlock(q.lock_ptr);
2770 * We have been woken up by futex_unlock_pi(), a timeout, or a
2771 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2774 WARN_ON(!q.pi_state);
2775 pi_mutex = &q.pi_state->pi_mutex;
2776 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2777 debug_rt_mutex_free_waiter(&rt_waiter);
2779 spin_lock(q.lock_ptr);
2781 * Fixup the pi_state owner and possibly acquire the lock if we
2784 res = fixup_owner(uaddr2, &q, !ret);
2786 * If fixup_owner() returned an error, proprogate that. If it
2787 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2790 ret = (res < 0) ? res : 0;
2792 /* Unqueue and drop the lock. */
2797 * If fixup_pi_state_owner() faulted and was unable to handle the
2798 * fault, unlock the rt_mutex and return the fault to userspace.
2800 if (ret == -EFAULT) {
2801 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2802 rt_mutex_unlock(pi_mutex);
2803 } else if (ret == -EINTR) {
2805 * We've already been requeued, but cannot restart by calling
2806 * futex_lock_pi() directly. We could restart this syscall, but
2807 * it would detect that the user space "val" changed and return
2808 * -EWOULDBLOCK. Save the overhead of the restart and return
2809 * -EWOULDBLOCK directly.
2815 put_futex_key(&q.key);
2817 put_futex_key(&key2);
2821 hrtimer_cancel(&to->timer);
2822 destroy_hrtimer_on_stack(&to->timer);
2828 * Support for robust futexes: the kernel cleans up held futexes at
2831 * Implementation: user-space maintains a per-thread list of locks it
2832 * is holding. Upon do_exit(), the kernel carefully walks this list,
2833 * and marks all locks that are owned by this thread with the
2834 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2835 * always manipulated with the lock held, so the list is private and
2836 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2837 * field, to allow the kernel to clean up if the thread dies after
2838 * acquiring the lock, but just before it could have added itself to
2839 * the list. There can only be one such pending lock.
2843 * sys_set_robust_list() - Set the robust-futex list head of a task
2844 * @head: pointer to the list-head
2845 * @len: length of the list-head, as userspace expects
2847 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2850 if (!futex_cmpxchg_enabled)
2853 * The kernel knows only one size for now:
2855 if (unlikely(len != sizeof(*head)))
2858 current->robust_list = head;
2864 * sys_get_robust_list() - Get the robust-futex list head of a task
2865 * @pid: pid of the process [zero for current task]
2866 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2867 * @len_ptr: pointer to a length field, the kernel fills in the header size
2869 SYSCALL_DEFINE3(get_robust_list, int, pid,
2870 struct robust_list_head __user * __user *, head_ptr,
2871 size_t __user *, len_ptr)
2873 struct robust_list_head __user *head;
2875 struct task_struct *p;
2877 if (!futex_cmpxchg_enabled)
2886 p = find_task_by_vpid(pid);
2892 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2895 head = p->robust_list;
2898 if (put_user(sizeof(*head), len_ptr))
2900 return put_user(head, head_ptr);
2909 * Process a futex-list entry, check whether it's owned by the
2910 * dying task, and do notification if so:
2912 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2914 u32 uval, uninitialized_var(nval), mval;
2917 if (get_user(uval, uaddr))
2920 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2922 * Ok, this dying thread is truly holding a futex
2923 * of interest. Set the OWNER_DIED bit atomically
2924 * via cmpxchg, and if the value had FUTEX_WAITERS
2925 * set, wake up a waiter (if any). (We have to do a
2926 * futex_wake() even if OWNER_DIED is already set -
2927 * to handle the rare but possible case of recursive
2928 * thread-death.) The rest of the cleanup is done in
2931 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2933 * We are not holding a lock here, but we want to have
2934 * the pagefault_disable/enable() protection because
2935 * we want to handle the fault gracefully. If the
2936 * access fails we try to fault in the futex with R/W
2937 * verification via get_user_pages. get_user() above
2938 * does not guarantee R/W access. If that fails we
2939 * give up and leave the futex locked.
2941 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2942 if (fault_in_user_writeable(uaddr))
2950 * Wake robust non-PI futexes here. The wakeup of
2951 * PI futexes happens in exit_pi_state():
2953 if (!pi && (uval & FUTEX_WAITERS))
2954 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2960 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2962 static inline int fetch_robust_entry(struct robust_list __user **entry,
2963 struct robust_list __user * __user *head,
2966 unsigned long uentry;
2968 if (get_user(uentry, (unsigned long __user *)head))
2971 *entry = (void __user *)(uentry & ~1UL);
2978 * Walk curr->robust_list (very carefully, it's a userspace list!)
2979 * and mark any locks found there dead, and notify any waiters.
2981 * We silently return on any sign of list-walking problem.
2983 void exit_robust_list(struct task_struct *curr)
2985 struct robust_list_head __user *head = curr->robust_list;
2986 struct robust_list __user *entry, *next_entry, *pending;
2987 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2988 unsigned int uninitialized_var(next_pi);
2989 unsigned long futex_offset;
2992 if (!futex_cmpxchg_enabled)
2996 * Fetch the list head (which was registered earlier, via
2997 * sys_set_robust_list()):
2999 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3002 * Fetch the relative futex offset:
3004 if (get_user(futex_offset, &head->futex_offset))
3007 * Fetch any possibly pending lock-add first, and handle it
3010 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3013 next_entry = NULL; /* avoid warning with gcc */
3014 while (entry != &head->list) {
3016 * Fetch the next entry in the list before calling
3017 * handle_futex_death:
3019 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3021 * A pending lock might already be on the list, so
3022 * don't process it twice:
3024 if (entry != pending)
3025 if (handle_futex_death((void __user *)entry + futex_offset,
3033 * Avoid excessively long or circular lists:
3042 handle_futex_death((void __user *)pending + futex_offset,
3046 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3047 u32 __user *uaddr2, u32 val2, u32 val3)
3049 int cmd = op & FUTEX_CMD_MASK;
3050 unsigned int flags = 0;
3052 if (!(op & FUTEX_PRIVATE_FLAG))
3053 flags |= FLAGS_SHARED;
3055 if (op & FUTEX_CLOCK_REALTIME) {
3056 flags |= FLAGS_CLOCKRT;
3057 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3063 case FUTEX_UNLOCK_PI:
3064 case FUTEX_TRYLOCK_PI:
3065 case FUTEX_WAIT_REQUEUE_PI:
3066 case FUTEX_CMP_REQUEUE_PI:
3067 if (!futex_cmpxchg_enabled)
3073 val3 = FUTEX_BITSET_MATCH_ANY;
3074 case FUTEX_WAIT_BITSET:
3075 return futex_wait(uaddr, flags, val, timeout, val3);
3077 val3 = FUTEX_BITSET_MATCH_ANY;
3078 case FUTEX_WAKE_BITSET:
3079 return futex_wake(uaddr, flags, val, val3);
3081 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3082 case FUTEX_CMP_REQUEUE:
3083 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3085 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3087 return futex_lock_pi(uaddr, flags, timeout, 0);
3088 case FUTEX_UNLOCK_PI:
3089 return futex_unlock_pi(uaddr, flags);
3090 case FUTEX_TRYLOCK_PI:
3091 return futex_lock_pi(uaddr, flags, NULL, 1);
3092 case FUTEX_WAIT_REQUEUE_PI:
3093 val3 = FUTEX_BITSET_MATCH_ANY;
3094 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3096 case FUTEX_CMP_REQUEUE_PI:
3097 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3103 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3104 struct timespec __user *, utime, u32 __user *, uaddr2,
3108 ktime_t t, *tp = NULL;
3110 int cmd = op & FUTEX_CMD_MASK;
3112 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3113 cmd == FUTEX_WAIT_BITSET ||
3114 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3115 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3117 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3119 if (!timespec_valid(&ts))
3122 t = timespec_to_ktime(ts);
3123 if (cmd == FUTEX_WAIT)
3124 t = ktime_add_safe(ktime_get(), t);
3128 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3129 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3131 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3132 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3133 val2 = (u32) (unsigned long) utime;
3135 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3138 static void __init futex_detect_cmpxchg(void)
3140 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3144 * This will fail and we want it. Some arch implementations do
3145 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3146 * functionality. We want to know that before we call in any
3147 * of the complex code paths. Also we want to prevent
3148 * registration of robust lists in that case. NULL is
3149 * guaranteed to fault and we get -EFAULT on functional
3150 * implementation, the non-functional ones will return
3153 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3154 futex_cmpxchg_enabled = 1;
3158 static int __init futex_init(void)
3160 unsigned int futex_shift;
3163 #if CONFIG_BASE_SMALL
3164 futex_hashsize = 16;
3166 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3169 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3171 futex_hashsize < 256 ? HASH_SMALL : 0,
3173 futex_hashsize, futex_hashsize);
3174 futex_hashsize = 1UL << futex_shift;
3176 futex_detect_cmpxchg();
3178 for (i = 0; i < futex_hashsize; i++) {
3179 atomic_set(&futex_queues[i].waiters, 0);
3180 plist_head_init(&futex_queues[i].chain);
3181 spin_lock_init(&futex_queues[i].lock);
3186 __initcall(futex_init);