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;
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);
524 * If page->mapping is NULL, then it cannot be a PageAnon
525 * page; but it might be the ZERO_PAGE or in the gate area or
526 * in a special mapping (all cases which we are happy to fail);
527 * or it may have been a good file page when get_user_pages_fast
528 * found it, but truncated or holepunched or subjected to
529 * invalidate_complete_page2 before we got the page lock (also
530 * cases which we are happy to fail). And we hold a reference,
531 * so refcount care in invalidate_complete_page's remove_mapping
532 * prevents drop_caches from setting mapping to NULL beneath us.
534 * The case we do have to guard against is when memory pressure made
535 * shmem_writepage move it from filecache to swapcache beneath us:
536 * an unlikely race, but we do need to retry for page->mapping.
538 if (!page->mapping) {
539 int shmem_swizzled = PageSwapCache(page);
548 * Private mappings are handled in a simple way.
550 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
551 * it's a read-only handle, it's expected that futexes attach to
552 * the object not the particular process.
554 if (PageAnon(page)) {
556 * A RO anonymous page will never change and thus doesn't make
557 * sense for futex operations.
559 if (unlikely(should_fail_futex(fshared)) || ro) {
564 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
565 key->private.mm = mm;
566 key->private.address = address;
568 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
569 key->shared.inode = page->mapping->host;
570 key->shared.pgoff = basepage_index(page);
573 get_futex_key_refs(key); /* implies MB (B) */
581 static inline void put_futex_key(union futex_key *key)
583 drop_futex_key_refs(key);
587 * fault_in_user_writeable() - Fault in user address and verify RW access
588 * @uaddr: pointer to faulting user space address
590 * Slow path to fixup the fault we just took in the atomic write
593 * We have no generic implementation of a non-destructive write to the
594 * user address. We know that we faulted in the atomic pagefault
595 * disabled section so we can as well avoid the #PF overhead by
596 * calling get_user_pages() right away.
598 static int fault_in_user_writeable(u32 __user *uaddr)
600 struct mm_struct *mm = current->mm;
603 down_read(&mm->mmap_sem);
604 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
606 up_read(&mm->mmap_sem);
608 return ret < 0 ? ret : 0;
612 * futex_top_waiter() - Return the highest priority waiter on a futex
613 * @hb: the hash bucket the futex_q's reside in
614 * @key: the futex key (to distinguish it from other futex futex_q's)
616 * Must be called with the hb lock held.
618 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
619 union futex_key *key)
621 struct futex_q *this;
623 plist_for_each_entry(this, &hb->chain, list) {
624 if (match_futex(&this->key, key))
630 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
631 u32 uval, u32 newval)
636 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
642 static int get_futex_value_locked(u32 *dest, u32 __user *from)
647 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
650 return ret ? -EFAULT : 0;
657 static int refill_pi_state_cache(void)
659 struct futex_pi_state *pi_state;
661 if (likely(current->pi_state_cache))
664 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
669 INIT_LIST_HEAD(&pi_state->list);
670 /* pi_mutex gets initialized later */
671 pi_state->owner = NULL;
672 atomic_set(&pi_state->refcount, 1);
673 pi_state->key = FUTEX_KEY_INIT;
675 current->pi_state_cache = pi_state;
680 static struct futex_pi_state * alloc_pi_state(void)
682 struct futex_pi_state *pi_state = current->pi_state_cache;
685 current->pi_state_cache = NULL;
691 * Must be called with the hb lock held.
693 static void free_pi_state(struct futex_pi_state *pi_state)
698 if (!atomic_dec_and_test(&pi_state->refcount))
702 * If pi_state->owner is NULL, the owner is most probably dying
703 * and has cleaned up the pi_state already
705 if (pi_state->owner) {
706 raw_spin_lock_irq(&pi_state->owner->pi_lock);
707 list_del_init(&pi_state->list);
708 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
710 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
713 if (current->pi_state_cache)
717 * pi_state->list is already empty.
718 * clear pi_state->owner.
719 * refcount is at 0 - put it back to 1.
721 pi_state->owner = NULL;
722 atomic_set(&pi_state->refcount, 1);
723 current->pi_state_cache = pi_state;
728 * Look up the task based on what TID userspace gave us.
731 static struct task_struct * futex_find_get_task(pid_t pid)
733 struct task_struct *p;
736 p = find_task_by_vpid(pid);
746 * This task is holding PI mutexes at exit time => bad.
747 * Kernel cleans up PI-state, but userspace is likely hosed.
748 * (Robust-futex cleanup is separate and might save the day for userspace.)
750 void exit_pi_state_list(struct task_struct *curr)
752 struct list_head *next, *head = &curr->pi_state_list;
753 struct futex_pi_state *pi_state;
754 struct futex_hash_bucket *hb;
755 union futex_key key = FUTEX_KEY_INIT;
757 if (!futex_cmpxchg_enabled)
760 * We are a ZOMBIE and nobody can enqueue itself on
761 * pi_state_list anymore, but we have to be careful
762 * versus waiters unqueueing themselves:
764 raw_spin_lock_irq(&curr->pi_lock);
765 while (!list_empty(head)) {
768 pi_state = list_entry(next, struct futex_pi_state, list);
770 hb = hash_futex(&key);
771 raw_spin_unlock_irq(&curr->pi_lock);
773 spin_lock(&hb->lock);
775 raw_spin_lock_irq(&curr->pi_lock);
777 * We dropped the pi-lock, so re-check whether this
778 * task still owns the PI-state:
780 if (head->next != next) {
781 spin_unlock(&hb->lock);
785 WARN_ON(pi_state->owner != curr);
786 WARN_ON(list_empty(&pi_state->list));
787 list_del_init(&pi_state->list);
788 pi_state->owner = NULL;
789 raw_spin_unlock_irq(&curr->pi_lock);
791 rt_mutex_unlock(&pi_state->pi_mutex);
793 spin_unlock(&hb->lock);
795 raw_spin_lock_irq(&curr->pi_lock);
797 raw_spin_unlock_irq(&curr->pi_lock);
801 * We need to check the following states:
803 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
805 * [1] NULL | --- | --- | 0 | 0/1 | Valid
806 * [2] NULL | --- | --- | >0 | 0/1 | Valid
808 * [3] Found | NULL | -- | Any | 0/1 | Invalid
810 * [4] Found | Found | NULL | 0 | 1 | Valid
811 * [5] Found | Found | NULL | >0 | 1 | Invalid
813 * [6] Found | Found | task | 0 | 1 | Valid
815 * [7] Found | Found | NULL | Any | 0 | Invalid
817 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
818 * [9] Found | Found | task | 0 | 0 | Invalid
819 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
821 * [1] Indicates that the kernel can acquire the futex atomically. We
822 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
824 * [2] Valid, if TID does not belong to a kernel thread. If no matching
825 * thread is found then it indicates that the owner TID has died.
827 * [3] Invalid. The waiter is queued on a non PI futex
829 * [4] Valid state after exit_robust_list(), which sets the user space
830 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
832 * [5] The user space value got manipulated between exit_robust_list()
833 * and exit_pi_state_list()
835 * [6] Valid state after exit_pi_state_list() which sets the new owner in
836 * the pi_state but cannot access the user space value.
838 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
840 * [8] Owner and user space value match
842 * [9] There is no transient state which sets the user space TID to 0
843 * except exit_robust_list(), but this is indicated by the
844 * FUTEX_OWNER_DIED bit. See [4]
846 * [10] There is no transient state which leaves owner and user space
851 * Validate that the existing waiter has a pi_state and sanity check
852 * the pi_state against the user space value. If correct, attach to
855 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
856 struct futex_pi_state **ps)
858 pid_t pid = uval & FUTEX_TID_MASK;
861 * Userspace might have messed up non-PI and PI futexes [3]
863 if (unlikely(!pi_state))
866 WARN_ON(!atomic_read(&pi_state->refcount));
869 * Handle the owner died case:
871 if (uval & FUTEX_OWNER_DIED) {
873 * exit_pi_state_list sets owner to NULL and wakes the
874 * topmost waiter. The task which acquires the
875 * pi_state->rt_mutex will fixup owner.
877 if (!pi_state->owner) {
879 * No pi state owner, but the user space TID
880 * is not 0. Inconsistent state. [5]
885 * Take a ref on the state and return success. [4]
891 * If TID is 0, then either the dying owner has not
892 * yet executed exit_pi_state_list() or some waiter
893 * acquired the rtmutex in the pi state, but did not
894 * yet fixup the TID in user space.
896 * Take a ref on the state and return success. [6]
902 * If the owner died bit is not set, then the pi_state
903 * must have an owner. [7]
905 if (!pi_state->owner)
910 * Bail out if user space manipulated the futex value. If pi
911 * state exists then the owner TID must be the same as the
912 * user space TID. [9/10]
914 if (pid != task_pid_vnr(pi_state->owner))
917 atomic_inc(&pi_state->refcount);
923 * Lookup the task for the TID provided from user space and attach to
924 * it after doing proper sanity checks.
926 static int attach_to_pi_owner(u32 uval, union futex_key *key,
927 struct futex_pi_state **ps)
929 pid_t pid = uval & FUTEX_TID_MASK;
930 struct futex_pi_state *pi_state;
931 struct task_struct *p;
934 * We are the first waiter - try to look up the real owner and attach
935 * the new pi_state to it, but bail out when TID = 0 [1]
939 p = futex_find_get_task(pid);
943 if (unlikely(p->flags & PF_KTHREAD)) {
949 * We need to look at the task state flags to figure out,
950 * whether the task is exiting. To protect against the do_exit
951 * change of the task flags, we do this protected by
954 raw_spin_lock_irq(&p->pi_lock);
955 if (unlikely(p->flags & PF_EXITING)) {
957 * The task is on the way out. When PF_EXITPIDONE is
958 * set, we know that the task has finished the
961 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
963 raw_spin_unlock_irq(&p->pi_lock);
969 * No existing pi state. First waiter. [2]
971 pi_state = alloc_pi_state();
974 * Initialize the pi_mutex in locked state and make @p
977 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
979 /* Store the key for possible exit cleanups: */
980 pi_state->key = *key;
982 WARN_ON(!list_empty(&pi_state->list));
983 list_add(&pi_state->list, &p->pi_state_list);
985 raw_spin_unlock_irq(&p->pi_lock);
994 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
995 union futex_key *key, struct futex_pi_state **ps)
997 struct futex_q *match = futex_top_waiter(hb, key);
1000 * If there is a waiter on that futex, validate it and
1001 * attach to the pi_state when the validation succeeds.
1004 return attach_to_pi_state(uval, match->pi_state, ps);
1007 * We are the first waiter - try to look up the owner based on
1008 * @uval and attach to it.
1010 return attach_to_pi_owner(uval, key, ps);
1013 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1015 u32 uninitialized_var(curval);
1017 if (unlikely(should_fail_futex(true)))
1020 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1023 /*If user space value changed, let the caller retry */
1024 return curval != uval ? -EAGAIN : 0;
1028 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1029 * @uaddr: the pi futex user address
1030 * @hb: the pi futex hash bucket
1031 * @key: the futex key associated with uaddr and hb
1032 * @ps: the pi_state pointer where we store the result of the
1034 * @task: the task to perform the atomic lock work for. This will
1035 * be "current" except in the case of requeue pi.
1036 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1039 * 0 - ready to wait;
1040 * 1 - acquired the lock;
1043 * The hb->lock and futex_key refs shall be held by the caller.
1045 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1046 union futex_key *key,
1047 struct futex_pi_state **ps,
1048 struct task_struct *task, int set_waiters)
1050 u32 uval, newval, vpid = task_pid_vnr(task);
1051 struct futex_q *match;
1055 * Read the user space value first so we can validate a few
1056 * things before proceeding further.
1058 if (get_futex_value_locked(&uval, uaddr))
1061 if (unlikely(should_fail_futex(true)))
1067 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1070 if ((unlikely(should_fail_futex(true))))
1074 * Lookup existing state first. If it exists, try to attach to
1077 match = futex_top_waiter(hb, key);
1079 return attach_to_pi_state(uval, match->pi_state, ps);
1082 * No waiter and user TID is 0. We are here because the
1083 * waiters or the owner died bit is set or called from
1084 * requeue_cmp_pi or for whatever reason something took the
1087 if (!(uval & FUTEX_TID_MASK)) {
1089 * We take over the futex. No other waiters and the user space
1090 * TID is 0. We preserve the owner died bit.
1092 newval = uval & FUTEX_OWNER_DIED;
1095 /* The futex requeue_pi code can enforce the waiters bit */
1097 newval |= FUTEX_WAITERS;
1099 ret = lock_pi_update_atomic(uaddr, uval, newval);
1100 /* If the take over worked, return 1 */
1101 return ret < 0 ? ret : 1;
1105 * First waiter. Set the waiters bit before attaching ourself to
1106 * the owner. If owner tries to unlock, it will be forced into
1107 * the kernel and blocked on hb->lock.
1109 newval = uval | FUTEX_WAITERS;
1110 ret = lock_pi_update_atomic(uaddr, uval, newval);
1114 * If the update of the user space value succeeded, we try to
1115 * attach to the owner. If that fails, no harm done, we only
1116 * set the FUTEX_WAITERS bit in the user space variable.
1118 return attach_to_pi_owner(uval, key, ps);
1122 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1123 * @q: The futex_q to unqueue
1125 * The q->lock_ptr must not be NULL and must be held by the caller.
1127 static void __unqueue_futex(struct futex_q *q)
1129 struct futex_hash_bucket *hb;
1131 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1132 || WARN_ON(plist_node_empty(&q->list)))
1135 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1136 plist_del(&q->list, &hb->chain);
1141 * The hash bucket lock must be held when this is called.
1142 * Afterwards, the futex_q must not be accessed. Callers
1143 * must ensure to later call wake_up_q() for the actual
1146 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1148 struct task_struct *p = q->task;
1150 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1154 * Queue the task for later wakeup for after we've released
1155 * the hb->lock. wake_q_add() grabs reference to p.
1157 wake_q_add(wake_q, p);
1160 * The waiting task can free the futex_q as soon as
1161 * q->lock_ptr = NULL is written, without taking any locks. A
1162 * memory barrier is required here to prevent the following
1163 * store to lock_ptr from getting ahead of the plist_del.
1169 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1170 struct futex_hash_bucket *hb)
1172 struct task_struct *new_owner;
1173 struct futex_pi_state *pi_state = this->pi_state;
1174 u32 uninitialized_var(curval), newval;
1183 * If current does not own the pi_state then the futex is
1184 * inconsistent and user space fiddled with the futex value.
1186 if (pi_state->owner != current)
1189 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1190 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1193 * It is possible that the next waiter (the one that brought
1194 * this owner to the kernel) timed out and is no longer
1195 * waiting on the lock.
1198 new_owner = this->task;
1201 * We pass it to the next owner. The WAITERS bit is always
1202 * kept enabled while there is PI state around. We cleanup the
1203 * owner died bit, because we are the owner.
1205 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1207 if (unlikely(should_fail_futex(true)))
1210 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1212 else if (curval != uval)
1215 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1219 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1220 WARN_ON(list_empty(&pi_state->list));
1221 list_del_init(&pi_state->list);
1222 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1224 raw_spin_lock_irq(&new_owner->pi_lock);
1225 WARN_ON(!list_empty(&pi_state->list));
1226 list_add(&pi_state->list, &new_owner->pi_state_list);
1227 pi_state->owner = new_owner;
1228 raw_spin_unlock_irq(&new_owner->pi_lock);
1230 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1232 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1235 * First unlock HB so the waiter does not spin on it once he got woken
1236 * up. Second wake up the waiter before the priority is adjusted. If we
1237 * deboost first (and lose our higher priority), then the task might get
1238 * scheduled away before the wake up can take place.
1240 spin_unlock(&hb->lock);
1243 rt_mutex_adjust_prio(current);
1249 * Express the locking dependencies for lockdep:
1252 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1255 spin_lock(&hb1->lock);
1257 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1258 } else { /* hb1 > hb2 */
1259 spin_lock(&hb2->lock);
1260 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1265 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1267 spin_unlock(&hb1->lock);
1269 spin_unlock(&hb2->lock);
1273 * Wake up waiters matching bitset queued on this futex (uaddr).
1276 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1278 struct futex_hash_bucket *hb;
1279 struct futex_q *this, *next;
1280 union futex_key key = FUTEX_KEY_INIT;
1287 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1288 if (unlikely(ret != 0))
1291 hb = hash_futex(&key);
1293 /* Make sure we really have tasks to wakeup */
1294 if (!hb_waiters_pending(hb))
1297 spin_lock(&hb->lock);
1299 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1300 if (match_futex (&this->key, &key)) {
1301 if (this->pi_state || this->rt_waiter) {
1306 /* Check if one of the bits is set in both bitsets */
1307 if (!(this->bitset & bitset))
1310 mark_wake_futex(&wake_q, this);
1311 if (++ret >= nr_wake)
1316 spin_unlock(&hb->lock);
1319 put_futex_key(&key);
1325 * Wake up all waiters hashed on the physical page that is mapped
1326 * to this virtual address:
1329 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1330 int nr_wake, int nr_wake2, int op)
1332 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1333 struct futex_hash_bucket *hb1, *hb2;
1334 struct futex_q *this, *next;
1339 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1340 if (unlikely(ret != 0))
1342 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1343 if (unlikely(ret != 0))
1346 hb1 = hash_futex(&key1);
1347 hb2 = hash_futex(&key2);
1350 double_lock_hb(hb1, hb2);
1351 op_ret = futex_atomic_op_inuser(op, uaddr2);
1352 if (unlikely(op_ret < 0)) {
1354 double_unlock_hb(hb1, hb2);
1358 * we don't get EFAULT from MMU faults if we don't have an MMU,
1359 * but we might get them from range checking
1365 if (unlikely(op_ret != -EFAULT)) {
1370 ret = fault_in_user_writeable(uaddr2);
1374 if (!(flags & FLAGS_SHARED))
1377 put_futex_key(&key2);
1378 put_futex_key(&key1);
1382 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1383 if (match_futex (&this->key, &key1)) {
1384 if (this->pi_state || this->rt_waiter) {
1388 mark_wake_futex(&wake_q, this);
1389 if (++ret >= nr_wake)
1396 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1397 if (match_futex (&this->key, &key2)) {
1398 if (this->pi_state || this->rt_waiter) {
1402 mark_wake_futex(&wake_q, this);
1403 if (++op_ret >= nr_wake2)
1411 double_unlock_hb(hb1, hb2);
1414 put_futex_key(&key2);
1416 put_futex_key(&key1);
1422 * requeue_futex() - Requeue a futex_q from one hb to another
1423 * @q: the futex_q to requeue
1424 * @hb1: the source hash_bucket
1425 * @hb2: the target hash_bucket
1426 * @key2: the new key for the requeued futex_q
1429 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1430 struct futex_hash_bucket *hb2, union futex_key *key2)
1434 * If key1 and key2 hash to the same bucket, no need to
1437 if (likely(&hb1->chain != &hb2->chain)) {
1438 plist_del(&q->list, &hb1->chain);
1439 hb_waiters_dec(hb1);
1440 plist_add(&q->list, &hb2->chain);
1441 hb_waiters_inc(hb2);
1442 q->lock_ptr = &hb2->lock;
1444 get_futex_key_refs(key2);
1449 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1451 * @key: the key of the requeue target futex
1452 * @hb: the hash_bucket of the requeue target futex
1454 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1455 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1456 * to the requeue target futex so the waiter can detect the wakeup on the right
1457 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1458 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1459 * to protect access to the pi_state to fixup the owner later. Must be called
1460 * with both q->lock_ptr and hb->lock held.
1463 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1464 struct futex_hash_bucket *hb)
1466 get_futex_key_refs(key);
1471 WARN_ON(!q->rt_waiter);
1472 q->rt_waiter = NULL;
1474 q->lock_ptr = &hb->lock;
1476 wake_up_state(q->task, TASK_NORMAL);
1480 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1481 * @pifutex: the user address of the to futex
1482 * @hb1: the from futex hash bucket, must be locked by the caller
1483 * @hb2: the to futex hash bucket, must be locked by the caller
1484 * @key1: the from futex key
1485 * @key2: the to futex key
1486 * @ps: address to store the pi_state pointer
1487 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1489 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1490 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1491 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1492 * hb1 and hb2 must be held by the caller.
1495 * 0 - failed to acquire the lock atomically;
1496 * >0 - acquired the lock, return value is vpid of the top_waiter
1499 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1500 struct futex_hash_bucket *hb1,
1501 struct futex_hash_bucket *hb2,
1502 union futex_key *key1, union futex_key *key2,
1503 struct futex_pi_state **ps, int set_waiters)
1505 struct futex_q *top_waiter = NULL;
1509 if (get_futex_value_locked(&curval, pifutex))
1512 if (unlikely(should_fail_futex(true)))
1516 * Find the top_waiter and determine if there are additional waiters.
1517 * If the caller intends to requeue more than 1 waiter to pifutex,
1518 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1519 * as we have means to handle the possible fault. If not, don't set
1520 * the bit unecessarily as it will force the subsequent unlock to enter
1523 top_waiter = futex_top_waiter(hb1, key1);
1525 /* There are no waiters, nothing for us to do. */
1529 /* Ensure we requeue to the expected futex. */
1530 if (!match_futex(top_waiter->requeue_pi_key, key2))
1534 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1535 * the contended case or if set_waiters is 1. The pi_state is returned
1536 * in ps in contended cases.
1538 vpid = task_pid_vnr(top_waiter->task);
1539 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1542 requeue_pi_wake_futex(top_waiter, key2, hb2);
1549 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1550 * @uaddr1: source futex user address
1551 * @flags: futex flags (FLAGS_SHARED, etc.)
1552 * @uaddr2: target futex user address
1553 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1554 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1555 * @cmpval: @uaddr1 expected value (or %NULL)
1556 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1557 * pi futex (pi to pi requeue is not supported)
1559 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1560 * uaddr2 atomically on behalf of the top waiter.
1563 * >=0 - on success, the number of tasks requeued or woken;
1566 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1567 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1568 u32 *cmpval, int requeue_pi)
1570 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1571 int drop_count = 0, task_count = 0, ret;
1572 struct futex_pi_state *pi_state = NULL;
1573 struct futex_hash_bucket *hb1, *hb2;
1574 struct futex_q *this, *next;
1579 * Requeue PI only works on two distinct uaddrs. This
1580 * check is only valid for private futexes. See below.
1582 if (uaddr1 == uaddr2)
1586 * requeue_pi requires a pi_state, try to allocate it now
1587 * without any locks in case it fails.
1589 if (refill_pi_state_cache())
1592 * requeue_pi must wake as many tasks as it can, up to nr_wake
1593 * + nr_requeue, since it acquires the rt_mutex prior to
1594 * returning to userspace, so as to not leave the rt_mutex with
1595 * waiters and no owner. However, second and third wake-ups
1596 * cannot be predicted as they involve race conditions with the
1597 * first wake and a fault while looking up the pi_state. Both
1598 * pthread_cond_signal() and pthread_cond_broadcast() should
1606 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1607 if (unlikely(ret != 0))
1609 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1610 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1611 if (unlikely(ret != 0))
1615 * The check above which compares uaddrs is not sufficient for
1616 * shared futexes. We need to compare the keys:
1618 if (requeue_pi && match_futex(&key1, &key2)) {
1623 hb1 = hash_futex(&key1);
1624 hb2 = hash_futex(&key2);
1627 hb_waiters_inc(hb2);
1628 double_lock_hb(hb1, hb2);
1630 if (likely(cmpval != NULL)) {
1633 ret = get_futex_value_locked(&curval, uaddr1);
1635 if (unlikely(ret)) {
1636 double_unlock_hb(hb1, hb2);
1637 hb_waiters_dec(hb2);
1639 ret = get_user(curval, uaddr1);
1643 if (!(flags & FLAGS_SHARED))
1646 put_futex_key(&key2);
1647 put_futex_key(&key1);
1650 if (curval != *cmpval) {
1656 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1658 * Attempt to acquire uaddr2 and wake the top waiter. If we
1659 * intend to requeue waiters, force setting the FUTEX_WAITERS
1660 * bit. We force this here where we are able to easily handle
1661 * faults rather in the requeue loop below.
1663 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1664 &key2, &pi_state, nr_requeue);
1667 * At this point the top_waiter has either taken uaddr2 or is
1668 * waiting on it. If the former, then the pi_state will not
1669 * exist yet, look it up one more time to ensure we have a
1670 * reference to it. If the lock was taken, ret contains the
1671 * vpid of the top waiter task.
1678 * If we acquired the lock, then the user
1679 * space value of uaddr2 should be vpid. It
1680 * cannot be changed by the top waiter as it
1681 * is blocked on hb2 lock if it tries to do
1682 * so. If something fiddled with it behind our
1683 * back the pi state lookup might unearth
1684 * it. So we rather use the known value than
1685 * rereading and handing potential crap to
1688 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1695 free_pi_state(pi_state);
1697 double_unlock_hb(hb1, hb2);
1698 hb_waiters_dec(hb2);
1699 put_futex_key(&key2);
1700 put_futex_key(&key1);
1701 ret = fault_in_user_writeable(uaddr2);
1707 * Two reasons for this:
1708 * - Owner is exiting and we just wait for the
1710 * - The user space value changed.
1712 free_pi_state(pi_state);
1714 double_unlock_hb(hb1, hb2);
1715 hb_waiters_dec(hb2);
1716 put_futex_key(&key2);
1717 put_futex_key(&key1);
1725 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1726 if (task_count - nr_wake >= nr_requeue)
1729 if (!match_futex(&this->key, &key1))
1733 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1734 * be paired with each other and no other futex ops.
1736 * We should never be requeueing a futex_q with a pi_state,
1737 * which is awaiting a futex_unlock_pi().
1739 if ((requeue_pi && !this->rt_waiter) ||
1740 (!requeue_pi && this->rt_waiter) ||
1747 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1748 * lock, we already woke the top_waiter. If not, it will be
1749 * woken by futex_unlock_pi().
1751 if (++task_count <= nr_wake && !requeue_pi) {
1752 mark_wake_futex(&wake_q, this);
1756 /* Ensure we requeue to the expected futex for requeue_pi. */
1757 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1763 * Requeue nr_requeue waiters and possibly one more in the case
1764 * of requeue_pi if we couldn't acquire the lock atomically.
1767 /* Prepare the waiter to take the rt_mutex. */
1768 atomic_inc(&pi_state->refcount);
1769 this->pi_state = pi_state;
1770 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1774 /* We got the lock. */
1775 requeue_pi_wake_futex(this, &key2, hb2);
1780 this->pi_state = NULL;
1781 free_pi_state(pi_state);
1785 requeue_futex(this, hb1, hb2, &key2);
1790 free_pi_state(pi_state);
1791 double_unlock_hb(hb1, hb2);
1793 hb_waiters_dec(hb2);
1796 * drop_futex_key_refs() must be called outside the spinlocks. During
1797 * the requeue we moved futex_q's from the hash bucket at key1 to the
1798 * one at key2 and updated their key pointer. We no longer need to
1799 * hold the references to key1.
1801 while (--drop_count >= 0)
1802 drop_futex_key_refs(&key1);
1805 put_futex_key(&key2);
1807 put_futex_key(&key1);
1809 return ret ? ret : task_count;
1812 /* The key must be already stored in q->key. */
1813 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1814 __acquires(&hb->lock)
1816 struct futex_hash_bucket *hb;
1818 hb = hash_futex(&q->key);
1821 * Increment the counter before taking the lock so that
1822 * a potential waker won't miss a to-be-slept task that is
1823 * waiting for the spinlock. This is safe as all queue_lock()
1824 * users end up calling queue_me(). Similarly, for housekeeping,
1825 * decrement the counter at queue_unlock() when some error has
1826 * occurred and we don't end up adding the task to the list.
1830 q->lock_ptr = &hb->lock;
1832 spin_lock(&hb->lock); /* implies MB (A) */
1837 queue_unlock(struct futex_hash_bucket *hb)
1838 __releases(&hb->lock)
1840 spin_unlock(&hb->lock);
1845 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1846 * @q: The futex_q to enqueue
1847 * @hb: The destination hash bucket
1849 * The hb->lock must be held by the caller, and is released here. A call to
1850 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1851 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1852 * or nothing if the unqueue is done as part of the wake process and the unqueue
1853 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1856 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1857 __releases(&hb->lock)
1862 * The priority used to register this element is
1863 * - either the real thread-priority for the real-time threads
1864 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1865 * - or MAX_RT_PRIO for non-RT threads.
1866 * Thus, all RT-threads are woken first in priority order, and
1867 * the others are woken last, in FIFO order.
1869 prio = min(current->normal_prio, MAX_RT_PRIO);
1871 plist_node_init(&q->list, prio);
1872 plist_add(&q->list, &hb->chain);
1874 spin_unlock(&hb->lock);
1878 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1879 * @q: The futex_q to unqueue
1881 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1882 * be paired with exactly one earlier call to queue_me().
1885 * 1 - if the futex_q was still queued (and we removed unqueued it);
1886 * 0 - if the futex_q was already removed by the waking thread
1888 static int unqueue_me(struct futex_q *q)
1890 spinlock_t *lock_ptr;
1893 /* In the common case we don't take the spinlock, which is nice. */
1895 lock_ptr = q->lock_ptr;
1897 if (lock_ptr != NULL) {
1898 spin_lock(lock_ptr);
1900 * q->lock_ptr can change between reading it and
1901 * spin_lock(), causing us to take the wrong lock. This
1902 * corrects the race condition.
1904 * Reasoning goes like this: if we have the wrong lock,
1905 * q->lock_ptr must have changed (maybe several times)
1906 * between reading it and the spin_lock(). It can
1907 * change again after the spin_lock() but only if it was
1908 * already changed before the spin_lock(). It cannot,
1909 * however, change back to the original value. Therefore
1910 * we can detect whether we acquired the correct lock.
1912 if (unlikely(lock_ptr != q->lock_ptr)) {
1913 spin_unlock(lock_ptr);
1918 BUG_ON(q->pi_state);
1920 spin_unlock(lock_ptr);
1924 drop_futex_key_refs(&q->key);
1929 * PI futexes can not be requeued and must remove themself from the
1930 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1933 static void unqueue_me_pi(struct futex_q *q)
1934 __releases(q->lock_ptr)
1938 BUG_ON(!q->pi_state);
1939 free_pi_state(q->pi_state);
1942 spin_unlock(q->lock_ptr);
1946 * Fixup the pi_state owner with the new owner.
1948 * Must be called with hash bucket lock held and mm->sem held for non
1951 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1952 struct task_struct *newowner)
1954 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1955 struct futex_pi_state *pi_state = q->pi_state;
1956 struct task_struct *oldowner = pi_state->owner;
1957 u32 uval, uninitialized_var(curval), newval;
1961 if (!pi_state->owner)
1962 newtid |= FUTEX_OWNER_DIED;
1965 * We are here either because we stole the rtmutex from the
1966 * previous highest priority waiter or we are the highest priority
1967 * waiter but failed to get the rtmutex the first time.
1968 * We have to replace the newowner TID in the user space variable.
1969 * This must be atomic as we have to preserve the owner died bit here.
1971 * Note: We write the user space value _before_ changing the pi_state
1972 * because we can fault here. Imagine swapped out pages or a fork
1973 * that marked all the anonymous memory readonly for cow.
1975 * Modifying pi_state _before_ the user space value would
1976 * leave the pi_state in an inconsistent state when we fault
1977 * here, because we need to drop the hash bucket lock to
1978 * handle the fault. This might be observed in the PID check
1979 * in lookup_pi_state.
1982 if (get_futex_value_locked(&uval, uaddr))
1986 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1988 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1996 * We fixed up user space. Now we need to fix the pi_state
1999 if (pi_state->owner != NULL) {
2000 raw_spin_lock_irq(&pi_state->owner->pi_lock);
2001 WARN_ON(list_empty(&pi_state->list));
2002 list_del_init(&pi_state->list);
2003 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2006 pi_state->owner = newowner;
2008 raw_spin_lock_irq(&newowner->pi_lock);
2009 WARN_ON(!list_empty(&pi_state->list));
2010 list_add(&pi_state->list, &newowner->pi_state_list);
2011 raw_spin_unlock_irq(&newowner->pi_lock);
2015 * To handle the page fault we need to drop the hash bucket
2016 * lock here. That gives the other task (either the highest priority
2017 * waiter itself or the task which stole the rtmutex) the
2018 * chance to try the fixup of the pi_state. So once we are
2019 * back from handling the fault we need to check the pi_state
2020 * after reacquiring the hash bucket lock and before trying to
2021 * do another fixup. When the fixup has been done already we
2025 spin_unlock(q->lock_ptr);
2027 ret = fault_in_user_writeable(uaddr);
2029 spin_lock(q->lock_ptr);
2032 * Check if someone else fixed it for us:
2034 if (pi_state->owner != oldowner)
2043 static long futex_wait_restart(struct restart_block *restart);
2046 * fixup_owner() - Post lock pi_state and corner case management
2047 * @uaddr: user address of the futex
2048 * @q: futex_q (contains pi_state and access to the rt_mutex)
2049 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2051 * After attempting to lock an rt_mutex, this function is called to cleanup
2052 * the pi_state owner as well as handle race conditions that may allow us to
2053 * acquire the lock. Must be called with the hb lock held.
2056 * 1 - success, lock taken;
2057 * 0 - success, lock not taken;
2058 * <0 - on error (-EFAULT)
2060 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2062 struct task_struct *owner;
2067 * Got the lock. We might not be the anticipated owner if we
2068 * did a lock-steal - fix up the PI-state in that case:
2070 if (q->pi_state->owner != current)
2071 ret = fixup_pi_state_owner(uaddr, q, current);
2076 * Catch the rare case, where the lock was released when we were on the
2077 * way back before we locked the hash bucket.
2079 if (q->pi_state->owner == current) {
2081 * Try to get the rt_mutex now. This might fail as some other
2082 * task acquired the rt_mutex after we removed ourself from the
2083 * rt_mutex waiters list.
2085 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2091 * pi_state is incorrect, some other task did a lock steal and
2092 * we returned due to timeout or signal without taking the
2093 * rt_mutex. Too late.
2095 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2096 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2098 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2099 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2100 ret = fixup_pi_state_owner(uaddr, q, owner);
2105 * Paranoia check. If we did not take the lock, then we should not be
2106 * the owner of the rt_mutex.
2108 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2109 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2110 "pi-state %p\n", ret,
2111 q->pi_state->pi_mutex.owner,
2112 q->pi_state->owner);
2115 return ret ? ret : locked;
2119 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2120 * @hb: the futex hash bucket, must be locked by the caller
2121 * @q: the futex_q to queue up on
2122 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2124 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2125 struct hrtimer_sleeper *timeout)
2128 * The task state is guaranteed to be set before another task can
2129 * wake it. set_current_state() is implemented using smp_store_mb() and
2130 * queue_me() calls spin_unlock() upon completion, both serializing
2131 * access to the hash list and forcing another memory barrier.
2133 set_current_state(TASK_INTERRUPTIBLE);
2138 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2141 * If we have been removed from the hash list, then another task
2142 * has tried to wake us, and we can skip the call to schedule().
2144 if (likely(!plist_node_empty(&q->list))) {
2146 * If the timer has already expired, current will already be
2147 * flagged for rescheduling. Only call schedule if there
2148 * is no timeout, or if it has yet to expire.
2150 if (!timeout || timeout->task)
2151 freezable_schedule();
2153 __set_current_state(TASK_RUNNING);
2157 * futex_wait_setup() - Prepare to wait on a futex
2158 * @uaddr: the futex userspace address
2159 * @val: the expected value
2160 * @flags: futex flags (FLAGS_SHARED, etc.)
2161 * @q: the associated futex_q
2162 * @hb: storage for hash_bucket pointer to be returned to caller
2164 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2165 * compare it with the expected value. Handle atomic faults internally.
2166 * Return with the hb lock held and a q.key reference on success, and unlocked
2167 * with no q.key reference on failure.
2170 * 0 - uaddr contains val and hb has been locked;
2171 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2173 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2174 struct futex_q *q, struct futex_hash_bucket **hb)
2180 * Access the page AFTER the hash-bucket is locked.
2181 * Order is important:
2183 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2184 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2186 * The basic logical guarantee of a futex is that it blocks ONLY
2187 * if cond(var) is known to be true at the time of blocking, for
2188 * any cond. If we locked the hash-bucket after testing *uaddr, that
2189 * would open a race condition where we could block indefinitely with
2190 * cond(var) false, which would violate the guarantee.
2192 * On the other hand, we insert q and release the hash-bucket only
2193 * after testing *uaddr. This guarantees that futex_wait() will NOT
2194 * absorb a wakeup if *uaddr does not match the desired values
2195 * while the syscall executes.
2198 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2199 if (unlikely(ret != 0))
2203 *hb = queue_lock(q);
2205 ret = get_futex_value_locked(&uval, uaddr);
2210 ret = get_user(uval, uaddr);
2214 if (!(flags & FLAGS_SHARED))
2217 put_futex_key(&q->key);
2228 put_futex_key(&q->key);
2232 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2233 ktime_t *abs_time, u32 bitset)
2235 struct hrtimer_sleeper timeout, *to = NULL;
2236 struct restart_block *restart;
2237 struct futex_hash_bucket *hb;
2238 struct futex_q q = futex_q_init;
2248 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2249 CLOCK_REALTIME : CLOCK_MONOTONIC,
2251 hrtimer_init_sleeper(to, current);
2252 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2253 current->timer_slack_ns);
2258 * Prepare to wait on uaddr. On success, holds hb lock and increments
2261 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2265 /* queue_me and wait for wakeup, timeout, or a signal. */
2266 futex_wait_queue_me(hb, &q, to);
2268 /* If we were woken (and unqueued), we succeeded, whatever. */
2270 /* unqueue_me() drops q.key ref */
2271 if (!unqueue_me(&q))
2274 if (to && !to->task)
2278 * We expect signal_pending(current), but we might be the
2279 * victim of a spurious wakeup as well.
2281 if (!signal_pending(current))
2288 restart = ¤t->restart_block;
2289 restart->fn = futex_wait_restart;
2290 restart->futex.uaddr = uaddr;
2291 restart->futex.val = val;
2292 restart->futex.time = abs_time->tv64;
2293 restart->futex.bitset = bitset;
2294 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2296 ret = -ERESTART_RESTARTBLOCK;
2300 hrtimer_cancel(&to->timer);
2301 destroy_hrtimer_on_stack(&to->timer);
2307 static long futex_wait_restart(struct restart_block *restart)
2309 u32 __user *uaddr = restart->futex.uaddr;
2310 ktime_t t, *tp = NULL;
2312 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2313 t.tv64 = restart->futex.time;
2316 restart->fn = do_no_restart_syscall;
2318 return (long)futex_wait(uaddr, restart->futex.flags,
2319 restart->futex.val, tp, restart->futex.bitset);
2324 * Userspace tried a 0 -> TID atomic transition of the futex value
2325 * and failed. The kernel side here does the whole locking operation:
2326 * if there are waiters then it will block as a consequence of relying
2327 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2328 * a 0 value of the futex too.).
2330 * Also serves as futex trylock_pi()'ing, and due semantics.
2332 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2333 ktime_t *time, int trylock)
2335 struct hrtimer_sleeper timeout, *to = NULL;
2336 struct futex_hash_bucket *hb;
2337 struct futex_q q = futex_q_init;
2340 if (refill_pi_state_cache())
2345 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2347 hrtimer_init_sleeper(to, current);
2348 hrtimer_set_expires(&to->timer, *time);
2352 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2353 if (unlikely(ret != 0))
2357 hb = queue_lock(&q);
2359 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2360 if (unlikely(ret)) {
2362 * Atomic work succeeded and we got the lock,
2363 * or failed. Either way, we do _not_ block.
2367 /* We got the lock. */
2369 goto out_unlock_put_key;
2374 * Two reasons for this:
2375 * - Task is exiting and we just wait for the
2377 * - The user space value changed.
2380 put_futex_key(&q.key);
2384 goto out_unlock_put_key;
2389 * Only actually queue now that the atomic ops are done:
2393 WARN_ON(!q.pi_state);
2395 * Block on the PI mutex:
2398 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2400 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2401 /* Fixup the trylock return value: */
2402 ret = ret ? 0 : -EWOULDBLOCK;
2405 spin_lock(q.lock_ptr);
2407 * Fixup the pi_state owner and possibly acquire the lock if we
2410 res = fixup_owner(uaddr, &q, !ret);
2412 * If fixup_owner() returned an error, proprogate that. If it acquired
2413 * the lock, clear our -ETIMEDOUT or -EINTR.
2416 ret = (res < 0) ? res : 0;
2419 * If fixup_owner() faulted and was unable to handle the fault, unlock
2420 * it and return the fault to userspace.
2422 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2423 rt_mutex_unlock(&q.pi_state->pi_mutex);
2425 /* Unqueue and drop the lock */
2434 put_futex_key(&q.key);
2437 destroy_hrtimer_on_stack(&to->timer);
2438 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2443 ret = fault_in_user_writeable(uaddr);
2447 if (!(flags & FLAGS_SHARED))
2450 put_futex_key(&q.key);
2455 * Userspace attempted a TID -> 0 atomic transition, and failed.
2456 * This is the in-kernel slowpath: we look up the PI state (if any),
2457 * and do the rt-mutex unlock.
2459 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2461 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2462 union futex_key key = FUTEX_KEY_INIT;
2463 struct futex_hash_bucket *hb;
2464 struct futex_q *match;
2468 if (get_user(uval, uaddr))
2471 * We release only a lock we actually own:
2473 if ((uval & FUTEX_TID_MASK) != vpid)
2476 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2480 hb = hash_futex(&key);
2481 spin_lock(&hb->lock);
2484 * Check waiters first. We do not trust user space values at
2485 * all and we at least want to know if user space fiddled
2486 * with the futex value instead of blindly unlocking.
2488 match = futex_top_waiter(hb, &key);
2490 ret = wake_futex_pi(uaddr, uval, match, hb);
2492 * In case of success wake_futex_pi dropped the hash
2498 * The atomic access to the futex value generated a
2499 * pagefault, so retry the user-access and the wakeup:
2504 * wake_futex_pi has detected invalid state. Tell user
2511 * We have no kernel internal state, i.e. no waiters in the
2512 * kernel. Waiters which are about to queue themselves are stuck
2513 * on hb->lock. So we can safely ignore them. We do neither
2514 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2517 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2521 * If uval has changed, let user space handle it.
2523 ret = (curval == uval) ? 0 : -EAGAIN;
2526 spin_unlock(&hb->lock);
2528 put_futex_key(&key);
2532 spin_unlock(&hb->lock);
2533 put_futex_key(&key);
2535 ret = fault_in_user_writeable(uaddr);
2543 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2544 * @hb: the hash_bucket futex_q was original enqueued on
2545 * @q: the futex_q woken while waiting to be requeued
2546 * @key2: the futex_key of the requeue target futex
2547 * @timeout: the timeout associated with the wait (NULL if none)
2549 * Detect if the task was woken on the initial futex as opposed to the requeue
2550 * target futex. If so, determine if it was a timeout or a signal that caused
2551 * the wakeup and return the appropriate error code to the caller. Must be
2552 * called with the hb lock held.
2555 * 0 = no early wakeup detected;
2556 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2559 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2560 struct futex_q *q, union futex_key *key2,
2561 struct hrtimer_sleeper *timeout)
2566 * With the hb lock held, we avoid races while we process the wakeup.
2567 * We only need to hold hb (and not hb2) to ensure atomicity as the
2568 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2569 * It can't be requeued from uaddr2 to something else since we don't
2570 * support a PI aware source futex for requeue.
2572 if (!match_futex(&q->key, key2)) {
2573 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2575 * We were woken prior to requeue by a timeout or a signal.
2576 * Unqueue the futex_q and determine which it was.
2578 plist_del(&q->list, &hb->chain);
2581 /* Handle spurious wakeups gracefully */
2583 if (timeout && !timeout->task)
2585 else if (signal_pending(current))
2586 ret = -ERESTARTNOINTR;
2592 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2593 * @uaddr: the futex we initially wait on (non-pi)
2594 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2595 * the same type, no requeueing from private to shared, etc.
2596 * @val: the expected value of uaddr
2597 * @abs_time: absolute timeout
2598 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2599 * @uaddr2: the pi futex we will take prior to returning to user-space
2601 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2602 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2603 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2604 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2605 * without one, the pi logic would not know which task to boost/deboost, if
2606 * there was a need to.
2608 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2609 * via the following--
2610 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2611 * 2) wakeup on uaddr2 after a requeue
2615 * If 3, cleanup and return -ERESTARTNOINTR.
2617 * If 2, we may then block on trying to take the rt_mutex and return via:
2618 * 5) successful lock
2621 * 8) other lock acquisition failure
2623 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2625 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2631 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2632 u32 val, ktime_t *abs_time, u32 bitset,
2635 struct hrtimer_sleeper timeout, *to = NULL;
2636 struct rt_mutex_waiter rt_waiter;
2637 struct rt_mutex *pi_mutex = NULL;
2638 struct futex_hash_bucket *hb;
2639 union futex_key key2 = FUTEX_KEY_INIT;
2640 struct futex_q q = futex_q_init;
2643 if (uaddr == uaddr2)
2651 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2652 CLOCK_REALTIME : CLOCK_MONOTONIC,
2654 hrtimer_init_sleeper(to, current);
2655 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2656 current->timer_slack_ns);
2660 * The waiter is allocated on our stack, manipulated by the requeue
2661 * code while we sleep on uaddr.
2663 debug_rt_mutex_init_waiter(&rt_waiter);
2664 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2665 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2666 rt_waiter.task = NULL;
2668 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2669 if (unlikely(ret != 0))
2673 q.rt_waiter = &rt_waiter;
2674 q.requeue_pi_key = &key2;
2677 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2680 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2685 * The check above which compares uaddrs is not sufficient for
2686 * shared futexes. We need to compare the keys:
2688 if (match_futex(&q.key, &key2)) {
2694 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2695 futex_wait_queue_me(hb, &q, to);
2697 spin_lock(&hb->lock);
2698 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2699 spin_unlock(&hb->lock);
2704 * In order for us to be here, we know our q.key == key2, and since
2705 * we took the hb->lock above, we also know that futex_requeue() has
2706 * completed and we no longer have to concern ourselves with a wakeup
2707 * race with the atomic proxy lock acquisition by the requeue code. The
2708 * futex_requeue dropped our key1 reference and incremented our key2
2712 /* Check if the requeue code acquired the second futex for us. */
2715 * Got the lock. We might not be the anticipated owner if we
2716 * did a lock-steal - fix up the PI-state in that case.
2718 if (q.pi_state && (q.pi_state->owner != current)) {
2719 spin_lock(q.lock_ptr);
2720 ret = fixup_pi_state_owner(uaddr2, &q, current);
2721 spin_unlock(q.lock_ptr);
2725 * We have been woken up by futex_unlock_pi(), a timeout, or a
2726 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2729 WARN_ON(!q.pi_state);
2730 pi_mutex = &q.pi_state->pi_mutex;
2731 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2732 debug_rt_mutex_free_waiter(&rt_waiter);
2734 spin_lock(q.lock_ptr);
2736 * Fixup the pi_state owner and possibly acquire the lock if we
2739 res = fixup_owner(uaddr2, &q, !ret);
2741 * If fixup_owner() returned an error, proprogate that. If it
2742 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2745 ret = (res < 0) ? res : 0;
2747 /* Unqueue and drop the lock. */
2752 * If fixup_pi_state_owner() faulted and was unable to handle the
2753 * fault, unlock the rt_mutex and return the fault to userspace.
2755 if (ret == -EFAULT) {
2756 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2757 rt_mutex_unlock(pi_mutex);
2758 } else if (ret == -EINTR) {
2760 * We've already been requeued, but cannot restart by calling
2761 * futex_lock_pi() directly. We could restart this syscall, but
2762 * it would detect that the user space "val" changed and return
2763 * -EWOULDBLOCK. Save the overhead of the restart and return
2764 * -EWOULDBLOCK directly.
2770 put_futex_key(&q.key);
2772 put_futex_key(&key2);
2776 hrtimer_cancel(&to->timer);
2777 destroy_hrtimer_on_stack(&to->timer);
2783 * Support for robust futexes: the kernel cleans up held futexes at
2786 * Implementation: user-space maintains a per-thread list of locks it
2787 * is holding. Upon do_exit(), the kernel carefully walks this list,
2788 * and marks all locks that are owned by this thread with the
2789 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2790 * always manipulated with the lock held, so the list is private and
2791 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2792 * field, to allow the kernel to clean up if the thread dies after
2793 * acquiring the lock, but just before it could have added itself to
2794 * the list. There can only be one such pending lock.
2798 * sys_set_robust_list() - Set the robust-futex list head of a task
2799 * @head: pointer to the list-head
2800 * @len: length of the list-head, as userspace expects
2802 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2805 if (!futex_cmpxchg_enabled)
2808 * The kernel knows only one size for now:
2810 if (unlikely(len != sizeof(*head)))
2813 current->robust_list = head;
2819 * sys_get_robust_list() - Get the robust-futex list head of a task
2820 * @pid: pid of the process [zero for current task]
2821 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2822 * @len_ptr: pointer to a length field, the kernel fills in the header size
2824 SYSCALL_DEFINE3(get_robust_list, int, pid,
2825 struct robust_list_head __user * __user *, head_ptr,
2826 size_t __user *, len_ptr)
2828 struct robust_list_head __user *head;
2830 struct task_struct *p;
2832 if (!futex_cmpxchg_enabled)
2841 p = find_task_by_vpid(pid);
2847 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2850 head = p->robust_list;
2853 if (put_user(sizeof(*head), len_ptr))
2855 return put_user(head, head_ptr);
2864 * Process a futex-list entry, check whether it's owned by the
2865 * dying task, and do notification if so:
2867 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2869 u32 uval, uninitialized_var(nval), mval;
2872 if (get_user(uval, uaddr))
2875 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2877 * Ok, this dying thread is truly holding a futex
2878 * of interest. Set the OWNER_DIED bit atomically
2879 * via cmpxchg, and if the value had FUTEX_WAITERS
2880 * set, wake up a waiter (if any). (We have to do a
2881 * futex_wake() even if OWNER_DIED is already set -
2882 * to handle the rare but possible case of recursive
2883 * thread-death.) The rest of the cleanup is done in
2886 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2888 * We are not holding a lock here, but we want to have
2889 * the pagefault_disable/enable() protection because
2890 * we want to handle the fault gracefully. If the
2891 * access fails we try to fault in the futex with R/W
2892 * verification via get_user_pages. get_user() above
2893 * does not guarantee R/W access. If that fails we
2894 * give up and leave the futex locked.
2896 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2897 if (fault_in_user_writeable(uaddr))
2905 * Wake robust non-PI futexes here. The wakeup of
2906 * PI futexes happens in exit_pi_state():
2908 if (!pi && (uval & FUTEX_WAITERS))
2909 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2915 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2917 static inline int fetch_robust_entry(struct robust_list __user **entry,
2918 struct robust_list __user * __user *head,
2921 unsigned long uentry;
2923 if (get_user(uentry, (unsigned long __user *)head))
2926 *entry = (void __user *)(uentry & ~1UL);
2933 * Walk curr->robust_list (very carefully, it's a userspace list!)
2934 * and mark any locks found there dead, and notify any waiters.
2936 * We silently return on any sign of list-walking problem.
2938 void exit_robust_list(struct task_struct *curr)
2940 struct robust_list_head __user *head = curr->robust_list;
2941 struct robust_list __user *entry, *next_entry, *pending;
2942 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2943 unsigned int uninitialized_var(next_pi);
2944 unsigned long futex_offset;
2947 if (!futex_cmpxchg_enabled)
2951 * Fetch the list head (which was registered earlier, via
2952 * sys_set_robust_list()):
2954 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2957 * Fetch the relative futex offset:
2959 if (get_user(futex_offset, &head->futex_offset))
2962 * Fetch any possibly pending lock-add first, and handle it
2965 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2968 next_entry = NULL; /* avoid warning with gcc */
2969 while (entry != &head->list) {
2971 * Fetch the next entry in the list before calling
2972 * handle_futex_death:
2974 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2976 * A pending lock might already be on the list, so
2977 * don't process it twice:
2979 if (entry != pending)
2980 if (handle_futex_death((void __user *)entry + futex_offset,
2988 * Avoid excessively long or circular lists:
2997 handle_futex_death((void __user *)pending + futex_offset,
3001 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3002 u32 __user *uaddr2, u32 val2, u32 val3)
3004 int cmd = op & FUTEX_CMD_MASK;
3005 unsigned int flags = 0;
3007 if (!(op & FUTEX_PRIVATE_FLAG))
3008 flags |= FLAGS_SHARED;
3010 if (op & FUTEX_CLOCK_REALTIME) {
3011 flags |= FLAGS_CLOCKRT;
3012 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3018 case FUTEX_UNLOCK_PI:
3019 case FUTEX_TRYLOCK_PI:
3020 case FUTEX_WAIT_REQUEUE_PI:
3021 case FUTEX_CMP_REQUEUE_PI:
3022 if (!futex_cmpxchg_enabled)
3028 val3 = FUTEX_BITSET_MATCH_ANY;
3029 case FUTEX_WAIT_BITSET:
3030 return futex_wait(uaddr, flags, val, timeout, val3);
3032 val3 = FUTEX_BITSET_MATCH_ANY;
3033 case FUTEX_WAKE_BITSET:
3034 return futex_wake(uaddr, flags, val, val3);
3036 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3037 case FUTEX_CMP_REQUEUE:
3038 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3040 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3042 return futex_lock_pi(uaddr, flags, timeout, 0);
3043 case FUTEX_UNLOCK_PI:
3044 return futex_unlock_pi(uaddr, flags);
3045 case FUTEX_TRYLOCK_PI:
3046 return futex_lock_pi(uaddr, flags, NULL, 1);
3047 case FUTEX_WAIT_REQUEUE_PI:
3048 val3 = FUTEX_BITSET_MATCH_ANY;
3049 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3051 case FUTEX_CMP_REQUEUE_PI:
3052 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3058 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3059 struct timespec __user *, utime, u32 __user *, uaddr2,
3063 ktime_t t, *tp = NULL;
3065 int cmd = op & FUTEX_CMD_MASK;
3067 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3068 cmd == FUTEX_WAIT_BITSET ||
3069 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3070 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3072 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3074 if (!timespec_valid(&ts))
3077 t = timespec_to_ktime(ts);
3078 if (cmd == FUTEX_WAIT)
3079 t = ktime_add_safe(ktime_get(), t);
3083 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3084 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3086 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3087 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3088 val2 = (u32) (unsigned long) utime;
3090 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3093 static void __init futex_detect_cmpxchg(void)
3095 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3099 * This will fail and we want it. Some arch implementations do
3100 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3101 * functionality. We want to know that before we call in any
3102 * of the complex code paths. Also we want to prevent
3103 * registration of robust lists in that case. NULL is
3104 * guaranteed to fault and we get -EFAULT on functional
3105 * implementation, the non-functional ones will return
3108 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3109 futex_cmpxchg_enabled = 1;
3113 static int __init futex_init(void)
3115 unsigned int futex_shift;
3118 #if CONFIG_BASE_SMALL
3119 futex_hashsize = 16;
3121 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3124 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3126 futex_hashsize < 256 ? HASH_SMALL : 0,
3128 futex_hashsize, futex_hashsize);
3129 futex_hashsize = 1UL << futex_shift;
3131 futex_detect_cmpxchg();
3133 for (i = 0; i < futex_hashsize; i++) {
3134 atomic_set(&futex_queues[i].waiters, 0);
3135 plist_head_init(&futex_queues[i].chain);
3136 spin_lock_init(&futex_queues[i].lock);
3141 __initcall(futex_init);