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;
258 static unsigned long __read_mostly futex_hashsize;
260 static struct futex_hash_bucket *futex_queues;
263 * Fault injections for futexes.
265 #ifdef CONFIG_FAIL_FUTEX
268 struct fault_attr attr;
272 .attr = FAULT_ATTR_INITIALIZER,
276 static int __init setup_fail_futex(char *str)
278 return setup_fault_attr(&fail_futex.attr, str);
280 __setup("fail_futex=", setup_fail_futex);
282 static bool should_fail_futex(bool fshared)
284 if (fail_futex.ignore_private && !fshared)
287 return should_fail(&fail_futex.attr, 1);
290 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
292 static int __init fail_futex_debugfs(void)
294 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
297 dir = fault_create_debugfs_attr("fail_futex", NULL,
302 if (!debugfs_create_bool("ignore-private", mode, dir,
303 &fail_futex.ignore_private)) {
304 debugfs_remove_recursive(dir);
311 late_initcall(fail_futex_debugfs);
313 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
316 static inline bool should_fail_futex(bool fshared)
320 #endif /* CONFIG_FAIL_FUTEX */
322 static inline void futex_get_mm(union futex_key *key)
324 atomic_inc(&key->private.mm->mm_count);
326 * Ensure futex_get_mm() implies a full barrier such that
327 * get_futex_key() implies a full barrier. This is relied upon
328 * as full barrier (B), see the ordering comment above.
330 smp_mb__after_atomic();
334 * Reflects a new waiter being added to the waitqueue.
336 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
339 atomic_inc(&hb->waiters);
341 * Full barrier (A), see the ordering comment above.
343 smp_mb__after_atomic();
348 * Reflects a waiter being removed from the waitqueue by wakeup
351 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
354 atomic_dec(&hb->waiters);
358 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
361 return atomic_read(&hb->waiters);
368 * We hash on the keys returned from get_futex_key (see below).
370 static struct futex_hash_bucket *hash_futex(union futex_key *key)
372 u32 hash = jhash2((u32*)&key->both.word,
373 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
375 return &futex_queues[hash & (futex_hashsize - 1)];
379 * Return 1 if two futex_keys are equal, 0 otherwise.
381 static inline int match_futex(union futex_key *key1, union futex_key *key2)
384 && key1->both.word == key2->both.word
385 && key1->both.ptr == key2->both.ptr
386 && key1->both.offset == key2->both.offset);
390 * Take a reference to the resource addressed by a key.
391 * Can be called while holding spinlocks.
394 static void get_futex_key_refs(union futex_key *key)
399 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
401 ihold(key->shared.inode); /* implies MB (B) */
403 case FUT_OFF_MMSHARED:
404 futex_get_mm(key); /* implies MB (B) */
408 * Private futexes do not hold reference on an inode or
409 * mm, therefore the only purpose of calling get_futex_key_refs
410 * is because we need the barrier for the lockless waiter check.
412 smp_mb(); /* explicit MB (B) */
417 * Drop a reference to the resource addressed by a key.
418 * The hash bucket spinlock must not be held. This is
419 * a no-op for private futexes, see comment in the get
422 static void drop_futex_key_refs(union futex_key *key)
424 if (!key->both.ptr) {
425 /* If we're here then we tried to put a key we failed to get */
430 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
432 iput(key->shared.inode);
434 case FUT_OFF_MMSHARED:
435 mmdrop(key->private.mm);
441 * get_futex_key() - Get parameters which are the keys for a futex
442 * @uaddr: virtual address of the futex
443 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
444 * @key: address where result is stored.
445 * @rw: mapping needs to be read/write (values: VERIFY_READ,
448 * Return: a negative error code or 0
450 * The key words are stored in *key on success.
452 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
453 * offset_within_page). For private mappings, it's (uaddr, current->mm).
454 * We can usually work out the index without swapping in the page.
456 * lock_page() might sleep, the caller should not hold a spinlock.
459 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
461 unsigned long address = (unsigned long)uaddr;
462 struct mm_struct *mm = current->mm;
467 * The futex address must be "naturally" aligned.
469 key->both.offset = address % PAGE_SIZE;
470 if (unlikely((address % sizeof(u32)) != 0))
472 address -= key->both.offset;
474 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
477 if (unlikely(should_fail_futex(fshared)))
481 * PROCESS_PRIVATE futexes are fast.
482 * As the mm cannot disappear under us and the 'key' only needs
483 * virtual address, we dont even have to find the underlying vma.
484 * Note : We do have to check 'uaddr' is a valid user address,
485 * but access_ok() should be faster than find_vma()
488 key->private.mm = mm;
489 key->private.address = address;
490 get_futex_key_refs(key); /* implies MB (B) */
495 /* Ignore any VERIFY_READ mapping (futex common case) */
496 if (unlikely(should_fail_futex(fshared)))
499 err = get_user_pages_fast(address, 1, 1, &page);
501 * If write access is not required (eg. FUTEX_WAIT), try
502 * and get read-only access.
504 if (err == -EFAULT && rw == VERIFY_READ) {
505 err = get_user_pages_fast(address, 1, 0, &page);
515 * If page->mapping is NULL, then it cannot be a PageAnon
516 * page; but it might be the ZERO_PAGE or in the gate area or
517 * in a special mapping (all cases which we are happy to fail);
518 * or it may have been a good file page when get_user_pages_fast
519 * found it, but truncated or holepunched or subjected to
520 * invalidate_complete_page2 before we got the page lock (also
521 * cases which we are happy to fail). And we hold a reference,
522 * so refcount care in invalidate_complete_page's remove_mapping
523 * prevents drop_caches from setting mapping to NULL beneath us.
525 * The case we do have to guard against is when memory pressure made
526 * shmem_writepage move it from filecache to swapcache beneath us:
527 * an unlikely race, but we do need to retry for page->mapping.
529 if (!page->mapping) {
530 int shmem_swizzled = PageSwapCache(page);
539 * Private mappings are handled in a simple way.
541 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
542 * it's a read-only handle, it's expected that futexes attach to
543 * the object not the particular process.
545 if (PageAnon(page)) {
547 * A RO anonymous page will never change and thus doesn't make
548 * sense for futex operations.
550 if (unlikely(should_fail_futex(fshared)) || ro) {
555 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
556 key->private.mm = mm;
557 key->private.address = address;
559 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
560 key->shared.inode = page->mapping->host;
561 key->shared.pgoff = basepage_index(page);
564 get_futex_key_refs(key); /* implies MB (B) */
572 static inline void put_futex_key(union futex_key *key)
574 drop_futex_key_refs(key);
578 * fault_in_user_writeable() - Fault in user address and verify RW access
579 * @uaddr: pointer to faulting user space address
581 * Slow path to fixup the fault we just took in the atomic write
584 * We have no generic implementation of a non-destructive write to the
585 * user address. We know that we faulted in the atomic pagefault
586 * disabled section so we can as well avoid the #PF overhead by
587 * calling get_user_pages() right away.
589 static int fault_in_user_writeable(u32 __user *uaddr)
591 struct mm_struct *mm = current->mm;
594 down_read(&mm->mmap_sem);
595 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
597 up_read(&mm->mmap_sem);
599 return ret < 0 ? ret : 0;
603 * futex_top_waiter() - Return the highest priority waiter on a futex
604 * @hb: the hash bucket the futex_q's reside in
605 * @key: the futex key (to distinguish it from other futex futex_q's)
607 * Must be called with the hb lock held.
609 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
610 union futex_key *key)
612 struct futex_q *this;
614 plist_for_each_entry(this, &hb->chain, list) {
615 if (match_futex(&this->key, key))
621 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
622 u32 uval, u32 newval)
627 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
633 static int get_futex_value_locked(u32 *dest, u32 __user *from)
638 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
641 return ret ? -EFAULT : 0;
648 static int refill_pi_state_cache(void)
650 struct futex_pi_state *pi_state;
652 if (likely(current->pi_state_cache))
655 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
660 INIT_LIST_HEAD(&pi_state->list);
661 /* pi_mutex gets initialized later */
662 pi_state->owner = NULL;
663 atomic_set(&pi_state->refcount, 1);
664 pi_state->key = FUTEX_KEY_INIT;
666 current->pi_state_cache = pi_state;
671 static struct futex_pi_state * alloc_pi_state(void)
673 struct futex_pi_state *pi_state = current->pi_state_cache;
676 current->pi_state_cache = NULL;
682 * Must be called with the hb lock held.
684 static void free_pi_state(struct futex_pi_state *pi_state)
689 if (!atomic_dec_and_test(&pi_state->refcount))
693 * If pi_state->owner is NULL, the owner is most probably dying
694 * and has cleaned up the pi_state already
696 if (pi_state->owner) {
697 raw_spin_lock_irq(&pi_state->owner->pi_lock);
698 list_del_init(&pi_state->list);
699 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
701 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
704 if (current->pi_state_cache)
708 * pi_state->list is already empty.
709 * clear pi_state->owner.
710 * refcount is at 0 - put it back to 1.
712 pi_state->owner = NULL;
713 atomic_set(&pi_state->refcount, 1);
714 current->pi_state_cache = pi_state;
719 * Look up the task based on what TID userspace gave us.
722 static struct task_struct * futex_find_get_task(pid_t pid)
724 struct task_struct *p;
727 p = find_task_by_vpid(pid);
737 * This task is holding PI mutexes at exit time => bad.
738 * Kernel cleans up PI-state, but userspace is likely hosed.
739 * (Robust-futex cleanup is separate and might save the day for userspace.)
741 void exit_pi_state_list(struct task_struct *curr)
743 struct list_head *next, *head = &curr->pi_state_list;
744 struct futex_pi_state *pi_state;
745 struct futex_hash_bucket *hb;
746 union futex_key key = FUTEX_KEY_INIT;
748 if (!futex_cmpxchg_enabled)
751 * We are a ZOMBIE and nobody can enqueue itself on
752 * pi_state_list anymore, but we have to be careful
753 * versus waiters unqueueing themselves:
755 raw_spin_lock_irq(&curr->pi_lock);
756 while (!list_empty(head)) {
759 pi_state = list_entry(next, struct futex_pi_state, list);
761 hb = hash_futex(&key);
762 raw_spin_unlock_irq(&curr->pi_lock);
764 spin_lock(&hb->lock);
766 raw_spin_lock_irq(&curr->pi_lock);
768 * We dropped the pi-lock, so re-check whether this
769 * task still owns the PI-state:
771 if (head->next != next) {
772 spin_unlock(&hb->lock);
776 WARN_ON(pi_state->owner != curr);
777 WARN_ON(list_empty(&pi_state->list));
778 list_del_init(&pi_state->list);
779 pi_state->owner = NULL;
780 raw_spin_unlock_irq(&curr->pi_lock);
782 rt_mutex_unlock(&pi_state->pi_mutex);
784 spin_unlock(&hb->lock);
786 raw_spin_lock_irq(&curr->pi_lock);
788 raw_spin_unlock_irq(&curr->pi_lock);
792 * We need to check the following states:
794 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
796 * [1] NULL | --- | --- | 0 | 0/1 | Valid
797 * [2] NULL | --- | --- | >0 | 0/1 | Valid
799 * [3] Found | NULL | -- | Any | 0/1 | Invalid
801 * [4] Found | Found | NULL | 0 | 1 | Valid
802 * [5] Found | Found | NULL | >0 | 1 | Invalid
804 * [6] Found | Found | task | 0 | 1 | Valid
806 * [7] Found | Found | NULL | Any | 0 | Invalid
808 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
809 * [9] Found | Found | task | 0 | 0 | Invalid
810 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
812 * [1] Indicates that the kernel can acquire the futex atomically. We
813 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
815 * [2] Valid, if TID does not belong to a kernel thread. If no matching
816 * thread is found then it indicates that the owner TID has died.
818 * [3] Invalid. The waiter is queued on a non PI futex
820 * [4] Valid state after exit_robust_list(), which sets the user space
821 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
823 * [5] The user space value got manipulated between exit_robust_list()
824 * and exit_pi_state_list()
826 * [6] Valid state after exit_pi_state_list() which sets the new owner in
827 * the pi_state but cannot access the user space value.
829 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
831 * [8] Owner and user space value match
833 * [9] There is no transient state which sets the user space TID to 0
834 * except exit_robust_list(), but this is indicated by the
835 * FUTEX_OWNER_DIED bit. See [4]
837 * [10] There is no transient state which leaves owner and user space
842 * Validate that the existing waiter has a pi_state and sanity check
843 * the pi_state against the user space value. If correct, attach to
846 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
847 struct futex_pi_state **ps)
849 pid_t pid = uval & FUTEX_TID_MASK;
852 * Userspace might have messed up non-PI and PI futexes [3]
854 if (unlikely(!pi_state))
857 WARN_ON(!atomic_read(&pi_state->refcount));
860 * Handle the owner died case:
862 if (uval & FUTEX_OWNER_DIED) {
864 * exit_pi_state_list sets owner to NULL and wakes the
865 * topmost waiter. The task which acquires the
866 * pi_state->rt_mutex will fixup owner.
868 if (!pi_state->owner) {
870 * No pi state owner, but the user space TID
871 * is not 0. Inconsistent state. [5]
876 * Take a ref on the state and return success. [4]
882 * If TID is 0, then either the dying owner has not
883 * yet executed exit_pi_state_list() or some waiter
884 * acquired the rtmutex in the pi state, but did not
885 * yet fixup the TID in user space.
887 * Take a ref on the state and return success. [6]
893 * If the owner died bit is not set, then the pi_state
894 * must have an owner. [7]
896 if (!pi_state->owner)
901 * Bail out if user space manipulated the futex value. If pi
902 * state exists then the owner TID must be the same as the
903 * user space TID. [9/10]
905 if (pid != task_pid_vnr(pi_state->owner))
908 atomic_inc(&pi_state->refcount);
914 * Lookup the task for the TID provided from user space and attach to
915 * it after doing proper sanity checks.
917 static int attach_to_pi_owner(u32 uval, union futex_key *key,
918 struct futex_pi_state **ps)
920 pid_t pid = uval & FUTEX_TID_MASK;
921 struct futex_pi_state *pi_state;
922 struct task_struct *p;
925 * We are the first waiter - try to look up the real owner and attach
926 * the new pi_state to it, but bail out when TID = 0 [1]
930 p = futex_find_get_task(pid);
934 if (unlikely(p->flags & PF_KTHREAD)) {
940 * We need to look at the task state flags to figure out,
941 * whether the task is exiting. To protect against the do_exit
942 * change of the task flags, we do this protected by
945 raw_spin_lock_irq(&p->pi_lock);
946 if (unlikely(p->flags & PF_EXITING)) {
948 * The task is on the way out. When PF_EXITPIDONE is
949 * set, we know that the task has finished the
952 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
954 raw_spin_unlock_irq(&p->pi_lock);
960 * No existing pi state. First waiter. [2]
962 pi_state = alloc_pi_state();
965 * Initialize the pi_mutex in locked state and make @p
968 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
970 /* Store the key for possible exit cleanups: */
971 pi_state->key = *key;
973 WARN_ON(!list_empty(&pi_state->list));
974 list_add(&pi_state->list, &p->pi_state_list);
976 raw_spin_unlock_irq(&p->pi_lock);
985 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
986 union futex_key *key, struct futex_pi_state **ps)
988 struct futex_q *match = futex_top_waiter(hb, key);
991 * If there is a waiter on that futex, validate it and
992 * attach to the pi_state when the validation succeeds.
995 return attach_to_pi_state(uval, match->pi_state, ps);
998 * We are the first waiter - try to look up the owner based on
999 * @uval and attach to it.
1001 return attach_to_pi_owner(uval, key, ps);
1004 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1006 u32 uninitialized_var(curval);
1008 if (unlikely(should_fail_futex(true)))
1011 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1014 /*If user space value changed, let the caller retry */
1015 return curval != uval ? -EAGAIN : 0;
1019 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1020 * @uaddr: the pi futex user address
1021 * @hb: the pi futex hash bucket
1022 * @key: the futex key associated with uaddr and hb
1023 * @ps: the pi_state pointer where we store the result of the
1025 * @task: the task to perform the atomic lock work for. This will
1026 * be "current" except in the case of requeue pi.
1027 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1030 * 0 - ready to wait;
1031 * 1 - acquired the lock;
1034 * The hb->lock and futex_key refs shall be held by the caller.
1036 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1037 union futex_key *key,
1038 struct futex_pi_state **ps,
1039 struct task_struct *task, int set_waiters)
1041 u32 uval, newval, vpid = task_pid_vnr(task);
1042 struct futex_q *match;
1046 * Read the user space value first so we can validate a few
1047 * things before proceeding further.
1049 if (get_futex_value_locked(&uval, uaddr))
1052 if (unlikely(should_fail_futex(true)))
1058 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1061 if ((unlikely(should_fail_futex(true))))
1065 * Lookup existing state first. If it exists, try to attach to
1068 match = futex_top_waiter(hb, key);
1070 return attach_to_pi_state(uval, match->pi_state, ps);
1073 * No waiter and user TID is 0. We are here because the
1074 * waiters or the owner died bit is set or called from
1075 * requeue_cmp_pi or for whatever reason something took the
1078 if (!(uval & FUTEX_TID_MASK)) {
1080 * We take over the futex. No other waiters and the user space
1081 * TID is 0. We preserve the owner died bit.
1083 newval = uval & FUTEX_OWNER_DIED;
1086 /* The futex requeue_pi code can enforce the waiters bit */
1088 newval |= FUTEX_WAITERS;
1090 ret = lock_pi_update_atomic(uaddr, uval, newval);
1091 /* If the take over worked, return 1 */
1092 return ret < 0 ? ret : 1;
1096 * First waiter. Set the waiters bit before attaching ourself to
1097 * the owner. If owner tries to unlock, it will be forced into
1098 * the kernel and blocked on hb->lock.
1100 newval = uval | FUTEX_WAITERS;
1101 ret = lock_pi_update_atomic(uaddr, uval, newval);
1105 * If the update of the user space value succeeded, we try to
1106 * attach to the owner. If that fails, no harm done, we only
1107 * set the FUTEX_WAITERS bit in the user space variable.
1109 return attach_to_pi_owner(uval, key, ps);
1113 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1114 * @q: The futex_q to unqueue
1116 * The q->lock_ptr must not be NULL and must be held by the caller.
1118 static void __unqueue_futex(struct futex_q *q)
1120 struct futex_hash_bucket *hb;
1122 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1123 || WARN_ON(plist_node_empty(&q->list)))
1126 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1127 plist_del(&q->list, &hb->chain);
1132 * The hash bucket lock must be held when this is called.
1133 * Afterwards, the futex_q must not be accessed. Callers
1134 * must ensure to later call wake_up_q() for the actual
1137 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1139 struct task_struct *p = q->task;
1141 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1145 * Queue the task for later wakeup for after we've released
1146 * the hb->lock. wake_q_add() grabs reference to p.
1148 wake_q_add(wake_q, p);
1151 * The waiting task can free the futex_q as soon as
1152 * q->lock_ptr = NULL is written, without taking any locks. A
1153 * memory barrier is required here to prevent the following
1154 * store to lock_ptr from getting ahead of the plist_del.
1160 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1161 struct futex_hash_bucket *hb)
1163 struct task_struct *new_owner;
1164 struct futex_pi_state *pi_state = this->pi_state;
1165 u32 uninitialized_var(curval), newval;
1174 * If current does not own the pi_state then the futex is
1175 * inconsistent and user space fiddled with the futex value.
1177 if (pi_state->owner != current)
1180 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1181 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1184 * It is possible that the next waiter (the one that brought
1185 * this owner to the kernel) timed out and is no longer
1186 * waiting on the lock.
1189 new_owner = this->task;
1192 * We pass it to the next owner. The WAITERS bit is always
1193 * kept enabled while there is PI state around. We cleanup the
1194 * owner died bit, because we are the owner.
1196 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1198 if (unlikely(should_fail_futex(true)))
1201 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1203 else if (curval != uval)
1206 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1210 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1211 WARN_ON(list_empty(&pi_state->list));
1212 list_del_init(&pi_state->list);
1213 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1215 raw_spin_lock_irq(&new_owner->pi_lock);
1216 WARN_ON(!list_empty(&pi_state->list));
1217 list_add(&pi_state->list, &new_owner->pi_state_list);
1218 pi_state->owner = new_owner;
1219 raw_spin_unlock_irq(&new_owner->pi_lock);
1221 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1223 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1226 * First unlock HB so the waiter does not spin on it once he got woken
1227 * up. Second wake up the waiter before the priority is adjusted. If we
1228 * deboost first (and lose our higher priority), then the task might get
1229 * scheduled away before the wake up can take place.
1231 spin_unlock(&hb->lock);
1234 rt_mutex_adjust_prio(current);
1240 * Express the locking dependencies for lockdep:
1243 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1246 spin_lock(&hb1->lock);
1248 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1249 } else { /* hb1 > hb2 */
1250 spin_lock(&hb2->lock);
1251 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1256 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1258 spin_unlock(&hb1->lock);
1260 spin_unlock(&hb2->lock);
1264 * Wake up waiters matching bitset queued on this futex (uaddr).
1267 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1269 struct futex_hash_bucket *hb;
1270 struct futex_q *this, *next;
1271 union futex_key key = FUTEX_KEY_INIT;
1278 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1279 if (unlikely(ret != 0))
1282 hb = hash_futex(&key);
1284 /* Make sure we really have tasks to wakeup */
1285 if (!hb_waiters_pending(hb))
1288 spin_lock(&hb->lock);
1290 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1291 if (match_futex (&this->key, &key)) {
1292 if (this->pi_state || this->rt_waiter) {
1297 /* Check if one of the bits is set in both bitsets */
1298 if (!(this->bitset & bitset))
1301 mark_wake_futex(&wake_q, this);
1302 if (++ret >= nr_wake)
1307 spin_unlock(&hb->lock);
1310 put_futex_key(&key);
1316 * Wake up all waiters hashed on the physical page that is mapped
1317 * to this virtual address:
1320 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1321 int nr_wake, int nr_wake2, int op)
1323 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1324 struct futex_hash_bucket *hb1, *hb2;
1325 struct futex_q *this, *next;
1330 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1331 if (unlikely(ret != 0))
1333 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1334 if (unlikely(ret != 0))
1337 hb1 = hash_futex(&key1);
1338 hb2 = hash_futex(&key2);
1341 double_lock_hb(hb1, hb2);
1342 op_ret = futex_atomic_op_inuser(op, uaddr2);
1343 if (unlikely(op_ret < 0)) {
1345 double_unlock_hb(hb1, hb2);
1349 * we don't get EFAULT from MMU faults if we don't have an MMU,
1350 * but we might get them from range checking
1356 if (unlikely(op_ret != -EFAULT)) {
1361 ret = fault_in_user_writeable(uaddr2);
1365 if (!(flags & FLAGS_SHARED))
1368 put_futex_key(&key2);
1369 put_futex_key(&key1);
1373 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1374 if (match_futex (&this->key, &key1)) {
1375 if (this->pi_state || this->rt_waiter) {
1379 mark_wake_futex(&wake_q, this);
1380 if (++ret >= nr_wake)
1387 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1388 if (match_futex (&this->key, &key2)) {
1389 if (this->pi_state || this->rt_waiter) {
1393 mark_wake_futex(&wake_q, this);
1394 if (++op_ret >= nr_wake2)
1402 double_unlock_hb(hb1, hb2);
1405 put_futex_key(&key2);
1407 put_futex_key(&key1);
1413 * requeue_futex() - Requeue a futex_q from one hb to another
1414 * @q: the futex_q to requeue
1415 * @hb1: the source hash_bucket
1416 * @hb2: the target hash_bucket
1417 * @key2: the new key for the requeued futex_q
1420 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1421 struct futex_hash_bucket *hb2, union futex_key *key2)
1425 * If key1 and key2 hash to the same bucket, no need to
1428 if (likely(&hb1->chain != &hb2->chain)) {
1429 plist_del(&q->list, &hb1->chain);
1430 hb_waiters_dec(hb1);
1431 plist_add(&q->list, &hb2->chain);
1432 hb_waiters_inc(hb2);
1433 q->lock_ptr = &hb2->lock;
1435 get_futex_key_refs(key2);
1440 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1442 * @key: the key of the requeue target futex
1443 * @hb: the hash_bucket of the requeue target futex
1445 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1446 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1447 * to the requeue target futex so the waiter can detect the wakeup on the right
1448 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1449 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1450 * to protect access to the pi_state to fixup the owner later. Must be called
1451 * with both q->lock_ptr and hb->lock held.
1454 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1455 struct futex_hash_bucket *hb)
1457 get_futex_key_refs(key);
1462 WARN_ON(!q->rt_waiter);
1463 q->rt_waiter = NULL;
1465 q->lock_ptr = &hb->lock;
1467 wake_up_state(q->task, TASK_NORMAL);
1471 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1472 * @pifutex: the user address of the to futex
1473 * @hb1: the from futex hash bucket, must be locked by the caller
1474 * @hb2: the to futex hash bucket, must be locked by the caller
1475 * @key1: the from futex key
1476 * @key2: the to futex key
1477 * @ps: address to store the pi_state pointer
1478 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1480 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1481 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1482 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1483 * hb1 and hb2 must be held by the caller.
1486 * 0 - failed to acquire the lock atomically;
1487 * >0 - acquired the lock, return value is vpid of the top_waiter
1490 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1491 struct futex_hash_bucket *hb1,
1492 struct futex_hash_bucket *hb2,
1493 union futex_key *key1, union futex_key *key2,
1494 struct futex_pi_state **ps, int set_waiters)
1496 struct futex_q *top_waiter = NULL;
1500 if (get_futex_value_locked(&curval, pifutex))
1503 if (unlikely(should_fail_futex(true)))
1507 * Find the top_waiter and determine if there are additional waiters.
1508 * If the caller intends to requeue more than 1 waiter to pifutex,
1509 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1510 * as we have means to handle the possible fault. If not, don't set
1511 * the bit unecessarily as it will force the subsequent unlock to enter
1514 top_waiter = futex_top_waiter(hb1, key1);
1516 /* There are no waiters, nothing for us to do. */
1520 /* Ensure we requeue to the expected futex. */
1521 if (!match_futex(top_waiter->requeue_pi_key, key2))
1525 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1526 * the contended case or if set_waiters is 1. The pi_state is returned
1527 * in ps in contended cases.
1529 vpid = task_pid_vnr(top_waiter->task);
1530 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1533 requeue_pi_wake_futex(top_waiter, key2, hb2);
1540 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1541 * @uaddr1: source futex user address
1542 * @flags: futex flags (FLAGS_SHARED, etc.)
1543 * @uaddr2: target futex user address
1544 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1545 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1546 * @cmpval: @uaddr1 expected value (or %NULL)
1547 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1548 * pi futex (pi to pi requeue is not supported)
1550 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1551 * uaddr2 atomically on behalf of the top waiter.
1554 * >=0 - on success, the number of tasks requeued or woken;
1557 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1558 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1559 u32 *cmpval, int requeue_pi)
1561 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1562 int drop_count = 0, task_count = 0, ret;
1563 struct futex_pi_state *pi_state = NULL;
1564 struct futex_hash_bucket *hb1, *hb2;
1565 struct futex_q *this, *next;
1570 * Requeue PI only works on two distinct uaddrs. This
1571 * check is only valid for private futexes. See below.
1573 if (uaddr1 == uaddr2)
1577 * requeue_pi requires a pi_state, try to allocate it now
1578 * without any locks in case it fails.
1580 if (refill_pi_state_cache())
1583 * requeue_pi must wake as many tasks as it can, up to nr_wake
1584 * + nr_requeue, since it acquires the rt_mutex prior to
1585 * returning to userspace, so as to not leave the rt_mutex with
1586 * waiters and no owner. However, second and third wake-ups
1587 * cannot be predicted as they involve race conditions with the
1588 * first wake and a fault while looking up the pi_state. Both
1589 * pthread_cond_signal() and pthread_cond_broadcast() should
1597 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1598 if (unlikely(ret != 0))
1600 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1601 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1602 if (unlikely(ret != 0))
1606 * The check above which compares uaddrs is not sufficient for
1607 * shared futexes. We need to compare the keys:
1609 if (requeue_pi && match_futex(&key1, &key2)) {
1614 hb1 = hash_futex(&key1);
1615 hb2 = hash_futex(&key2);
1618 hb_waiters_inc(hb2);
1619 double_lock_hb(hb1, hb2);
1621 if (likely(cmpval != NULL)) {
1624 ret = get_futex_value_locked(&curval, uaddr1);
1626 if (unlikely(ret)) {
1627 double_unlock_hb(hb1, hb2);
1628 hb_waiters_dec(hb2);
1630 ret = get_user(curval, uaddr1);
1634 if (!(flags & FLAGS_SHARED))
1637 put_futex_key(&key2);
1638 put_futex_key(&key1);
1641 if (curval != *cmpval) {
1647 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1649 * Attempt to acquire uaddr2 and wake the top waiter. If we
1650 * intend to requeue waiters, force setting the FUTEX_WAITERS
1651 * bit. We force this here where we are able to easily handle
1652 * faults rather in the requeue loop below.
1654 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1655 &key2, &pi_state, nr_requeue);
1658 * At this point the top_waiter has either taken uaddr2 or is
1659 * waiting on it. If the former, then the pi_state will not
1660 * exist yet, look it up one more time to ensure we have a
1661 * reference to it. If the lock was taken, ret contains the
1662 * vpid of the top waiter task.
1669 * If we acquired the lock, then the user
1670 * space value of uaddr2 should be vpid. It
1671 * cannot be changed by the top waiter as it
1672 * is blocked on hb2 lock if it tries to do
1673 * so. If something fiddled with it behind our
1674 * back the pi state lookup might unearth
1675 * it. So we rather use the known value than
1676 * rereading and handing potential crap to
1679 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1686 free_pi_state(pi_state);
1688 double_unlock_hb(hb1, hb2);
1689 hb_waiters_dec(hb2);
1690 put_futex_key(&key2);
1691 put_futex_key(&key1);
1692 ret = fault_in_user_writeable(uaddr2);
1698 * Two reasons for this:
1699 * - Owner is exiting and we just wait for the
1701 * - The user space value changed.
1703 free_pi_state(pi_state);
1705 double_unlock_hb(hb1, hb2);
1706 hb_waiters_dec(hb2);
1707 put_futex_key(&key2);
1708 put_futex_key(&key1);
1716 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1717 if (task_count - nr_wake >= nr_requeue)
1720 if (!match_futex(&this->key, &key1))
1724 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1725 * be paired with each other and no other futex ops.
1727 * We should never be requeueing a futex_q with a pi_state,
1728 * which is awaiting a futex_unlock_pi().
1730 if ((requeue_pi && !this->rt_waiter) ||
1731 (!requeue_pi && this->rt_waiter) ||
1738 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1739 * lock, we already woke the top_waiter. If not, it will be
1740 * woken by futex_unlock_pi().
1742 if (++task_count <= nr_wake && !requeue_pi) {
1743 mark_wake_futex(&wake_q, this);
1747 /* Ensure we requeue to the expected futex for requeue_pi. */
1748 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1754 * Requeue nr_requeue waiters and possibly one more in the case
1755 * of requeue_pi if we couldn't acquire the lock atomically.
1758 /* Prepare the waiter to take the rt_mutex. */
1759 atomic_inc(&pi_state->refcount);
1760 this->pi_state = pi_state;
1761 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1765 /* We got the lock. */
1766 requeue_pi_wake_futex(this, &key2, hb2);
1771 this->pi_state = NULL;
1772 free_pi_state(pi_state);
1776 requeue_futex(this, hb1, hb2, &key2);
1781 free_pi_state(pi_state);
1782 double_unlock_hb(hb1, hb2);
1784 hb_waiters_dec(hb2);
1787 * drop_futex_key_refs() must be called outside the spinlocks. During
1788 * the requeue we moved futex_q's from the hash bucket at key1 to the
1789 * one at key2 and updated their key pointer. We no longer need to
1790 * hold the references to key1.
1792 while (--drop_count >= 0)
1793 drop_futex_key_refs(&key1);
1796 put_futex_key(&key2);
1798 put_futex_key(&key1);
1800 return ret ? ret : task_count;
1803 /* The key must be already stored in q->key. */
1804 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1805 __acquires(&hb->lock)
1807 struct futex_hash_bucket *hb;
1809 hb = hash_futex(&q->key);
1812 * Increment the counter before taking the lock so that
1813 * a potential waker won't miss a to-be-slept task that is
1814 * waiting for the spinlock. This is safe as all queue_lock()
1815 * users end up calling queue_me(). Similarly, for housekeeping,
1816 * decrement the counter at queue_unlock() when some error has
1817 * occurred and we don't end up adding the task to the list.
1821 q->lock_ptr = &hb->lock;
1823 spin_lock(&hb->lock); /* implies MB (A) */
1828 queue_unlock(struct futex_hash_bucket *hb)
1829 __releases(&hb->lock)
1831 spin_unlock(&hb->lock);
1836 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1837 * @q: The futex_q to enqueue
1838 * @hb: The destination hash bucket
1840 * The hb->lock must be held by the caller, and is released here. A call to
1841 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1842 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1843 * or nothing if the unqueue is done as part of the wake process and the unqueue
1844 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1847 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1848 __releases(&hb->lock)
1853 * The priority used to register this element is
1854 * - either the real thread-priority for the real-time threads
1855 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1856 * - or MAX_RT_PRIO for non-RT threads.
1857 * Thus, all RT-threads are woken first in priority order, and
1858 * the others are woken last, in FIFO order.
1860 prio = min(current->normal_prio, MAX_RT_PRIO);
1862 plist_node_init(&q->list, prio);
1863 plist_add(&q->list, &hb->chain);
1865 spin_unlock(&hb->lock);
1869 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1870 * @q: The futex_q to unqueue
1872 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1873 * be paired with exactly one earlier call to queue_me().
1876 * 1 - if the futex_q was still queued (and we removed unqueued it);
1877 * 0 - if the futex_q was already removed by the waking thread
1879 static int unqueue_me(struct futex_q *q)
1881 spinlock_t *lock_ptr;
1884 /* In the common case we don't take the spinlock, which is nice. */
1886 lock_ptr = q->lock_ptr;
1888 if (lock_ptr != NULL) {
1889 spin_lock(lock_ptr);
1891 * q->lock_ptr can change between reading it and
1892 * spin_lock(), causing us to take the wrong lock. This
1893 * corrects the race condition.
1895 * Reasoning goes like this: if we have the wrong lock,
1896 * q->lock_ptr must have changed (maybe several times)
1897 * between reading it and the spin_lock(). It can
1898 * change again after the spin_lock() but only if it was
1899 * already changed before the spin_lock(). It cannot,
1900 * however, change back to the original value. Therefore
1901 * we can detect whether we acquired the correct lock.
1903 if (unlikely(lock_ptr != q->lock_ptr)) {
1904 spin_unlock(lock_ptr);
1909 BUG_ON(q->pi_state);
1911 spin_unlock(lock_ptr);
1915 drop_futex_key_refs(&q->key);
1920 * PI futexes can not be requeued and must remove themself from the
1921 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1924 static void unqueue_me_pi(struct futex_q *q)
1925 __releases(q->lock_ptr)
1929 BUG_ON(!q->pi_state);
1930 free_pi_state(q->pi_state);
1933 spin_unlock(q->lock_ptr);
1937 * Fixup the pi_state owner with the new owner.
1939 * Must be called with hash bucket lock held and mm->sem held for non
1942 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1943 struct task_struct *newowner)
1945 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1946 struct futex_pi_state *pi_state = q->pi_state;
1947 struct task_struct *oldowner = pi_state->owner;
1948 u32 uval, uninitialized_var(curval), newval;
1952 if (!pi_state->owner)
1953 newtid |= FUTEX_OWNER_DIED;
1956 * We are here either because we stole the rtmutex from the
1957 * previous highest priority waiter or we are the highest priority
1958 * waiter but failed to get the rtmutex the first time.
1959 * We have to replace the newowner TID in the user space variable.
1960 * This must be atomic as we have to preserve the owner died bit here.
1962 * Note: We write the user space value _before_ changing the pi_state
1963 * because we can fault here. Imagine swapped out pages or a fork
1964 * that marked all the anonymous memory readonly for cow.
1966 * Modifying pi_state _before_ the user space value would
1967 * leave the pi_state in an inconsistent state when we fault
1968 * here, because we need to drop the hash bucket lock to
1969 * handle the fault. This might be observed in the PID check
1970 * in lookup_pi_state.
1973 if (get_futex_value_locked(&uval, uaddr))
1977 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1979 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1987 * We fixed up user space. Now we need to fix the pi_state
1990 if (pi_state->owner != NULL) {
1991 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1992 WARN_ON(list_empty(&pi_state->list));
1993 list_del_init(&pi_state->list);
1994 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1997 pi_state->owner = newowner;
1999 raw_spin_lock_irq(&newowner->pi_lock);
2000 WARN_ON(!list_empty(&pi_state->list));
2001 list_add(&pi_state->list, &newowner->pi_state_list);
2002 raw_spin_unlock_irq(&newowner->pi_lock);
2006 * To handle the page fault we need to drop the hash bucket
2007 * lock here. That gives the other task (either the highest priority
2008 * waiter itself or the task which stole the rtmutex) the
2009 * chance to try the fixup of the pi_state. So once we are
2010 * back from handling the fault we need to check the pi_state
2011 * after reacquiring the hash bucket lock and before trying to
2012 * do another fixup. When the fixup has been done already we
2016 spin_unlock(q->lock_ptr);
2018 ret = fault_in_user_writeable(uaddr);
2020 spin_lock(q->lock_ptr);
2023 * Check if someone else fixed it for us:
2025 if (pi_state->owner != oldowner)
2034 static long futex_wait_restart(struct restart_block *restart);
2037 * fixup_owner() - Post lock pi_state and corner case management
2038 * @uaddr: user address of the futex
2039 * @q: futex_q (contains pi_state and access to the rt_mutex)
2040 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2042 * After attempting to lock an rt_mutex, this function is called to cleanup
2043 * the pi_state owner as well as handle race conditions that may allow us to
2044 * acquire the lock. Must be called with the hb lock held.
2047 * 1 - success, lock taken;
2048 * 0 - success, lock not taken;
2049 * <0 - on error (-EFAULT)
2051 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2053 struct task_struct *owner;
2058 * Got the lock. We might not be the anticipated owner if we
2059 * did a lock-steal - fix up the PI-state in that case:
2061 if (q->pi_state->owner != current)
2062 ret = fixup_pi_state_owner(uaddr, q, current);
2067 * Catch the rare case, where the lock was released when we were on the
2068 * way back before we locked the hash bucket.
2070 if (q->pi_state->owner == current) {
2072 * Try to get the rt_mutex now. This might fail as some other
2073 * task acquired the rt_mutex after we removed ourself from the
2074 * rt_mutex waiters list.
2076 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2082 * pi_state is incorrect, some other task did a lock steal and
2083 * we returned due to timeout or signal without taking the
2084 * rt_mutex. Too late.
2086 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2087 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2089 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2090 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2091 ret = fixup_pi_state_owner(uaddr, q, owner);
2096 * Paranoia check. If we did not take the lock, then we should not be
2097 * the owner of the rt_mutex.
2099 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2100 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2101 "pi-state %p\n", ret,
2102 q->pi_state->pi_mutex.owner,
2103 q->pi_state->owner);
2106 return ret ? ret : locked;
2110 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2111 * @hb: the futex hash bucket, must be locked by the caller
2112 * @q: the futex_q to queue up on
2113 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2115 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2116 struct hrtimer_sleeper *timeout)
2119 * The task state is guaranteed to be set before another task can
2120 * wake it. set_current_state() is implemented using smp_store_mb() and
2121 * queue_me() calls spin_unlock() upon completion, both serializing
2122 * access to the hash list and forcing another memory barrier.
2124 set_current_state(TASK_INTERRUPTIBLE);
2129 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2132 * If we have been removed from the hash list, then another task
2133 * has tried to wake us, and we can skip the call to schedule().
2135 if (likely(!plist_node_empty(&q->list))) {
2137 * If the timer has already expired, current will already be
2138 * flagged for rescheduling. Only call schedule if there
2139 * is no timeout, or if it has yet to expire.
2141 if (!timeout || timeout->task)
2142 freezable_schedule();
2144 __set_current_state(TASK_RUNNING);
2148 * futex_wait_setup() - Prepare to wait on a futex
2149 * @uaddr: the futex userspace address
2150 * @val: the expected value
2151 * @flags: futex flags (FLAGS_SHARED, etc.)
2152 * @q: the associated futex_q
2153 * @hb: storage for hash_bucket pointer to be returned to caller
2155 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2156 * compare it with the expected value. Handle atomic faults internally.
2157 * Return with the hb lock held and a q.key reference on success, and unlocked
2158 * with no q.key reference on failure.
2161 * 0 - uaddr contains val and hb has been locked;
2162 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2164 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2165 struct futex_q *q, struct futex_hash_bucket **hb)
2171 * Access the page AFTER the hash-bucket is locked.
2172 * Order is important:
2174 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2175 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2177 * The basic logical guarantee of a futex is that it blocks ONLY
2178 * if cond(var) is known to be true at the time of blocking, for
2179 * any cond. If we locked the hash-bucket after testing *uaddr, that
2180 * would open a race condition where we could block indefinitely with
2181 * cond(var) false, which would violate the guarantee.
2183 * On the other hand, we insert q and release the hash-bucket only
2184 * after testing *uaddr. This guarantees that futex_wait() will NOT
2185 * absorb a wakeup if *uaddr does not match the desired values
2186 * while the syscall executes.
2189 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2190 if (unlikely(ret != 0))
2194 *hb = queue_lock(q);
2196 ret = get_futex_value_locked(&uval, uaddr);
2201 ret = get_user(uval, uaddr);
2205 if (!(flags & FLAGS_SHARED))
2208 put_futex_key(&q->key);
2219 put_futex_key(&q->key);
2223 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2224 ktime_t *abs_time, u32 bitset)
2226 struct hrtimer_sleeper timeout, *to = NULL;
2227 struct restart_block *restart;
2228 struct futex_hash_bucket *hb;
2229 struct futex_q q = futex_q_init;
2239 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2240 CLOCK_REALTIME : CLOCK_MONOTONIC,
2242 hrtimer_init_sleeper(to, current);
2243 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2244 current->timer_slack_ns);
2249 * Prepare to wait on uaddr. On success, holds hb lock and increments
2252 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2256 /* queue_me and wait for wakeup, timeout, or a signal. */
2257 futex_wait_queue_me(hb, &q, to);
2259 /* If we were woken (and unqueued), we succeeded, whatever. */
2261 /* unqueue_me() drops q.key ref */
2262 if (!unqueue_me(&q))
2265 if (to && !to->task)
2269 * We expect signal_pending(current), but we might be the
2270 * victim of a spurious wakeup as well.
2272 if (!signal_pending(current))
2279 restart = ¤t->restart_block;
2280 restart->fn = futex_wait_restart;
2281 restart->futex.uaddr = uaddr;
2282 restart->futex.val = val;
2283 restart->futex.time = abs_time->tv64;
2284 restart->futex.bitset = bitset;
2285 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2287 ret = -ERESTART_RESTARTBLOCK;
2291 hrtimer_cancel(&to->timer);
2292 destroy_hrtimer_on_stack(&to->timer);
2298 static long futex_wait_restart(struct restart_block *restart)
2300 u32 __user *uaddr = restart->futex.uaddr;
2301 ktime_t t, *tp = NULL;
2303 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2304 t.tv64 = restart->futex.time;
2307 restart->fn = do_no_restart_syscall;
2309 return (long)futex_wait(uaddr, restart->futex.flags,
2310 restart->futex.val, tp, restart->futex.bitset);
2315 * Userspace tried a 0 -> TID atomic transition of the futex value
2316 * and failed. The kernel side here does the whole locking operation:
2317 * if there are waiters then it will block as a consequence of relying
2318 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2319 * a 0 value of the futex too.).
2321 * Also serves as futex trylock_pi()'ing, and due semantics.
2323 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2324 ktime_t *time, int trylock)
2326 struct hrtimer_sleeper timeout, *to = NULL;
2327 struct futex_hash_bucket *hb;
2328 struct futex_q q = futex_q_init;
2331 if (refill_pi_state_cache())
2336 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2338 hrtimer_init_sleeper(to, current);
2339 hrtimer_set_expires(&to->timer, *time);
2343 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2344 if (unlikely(ret != 0))
2348 hb = queue_lock(&q);
2350 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2351 if (unlikely(ret)) {
2353 * Atomic work succeeded and we got the lock,
2354 * or failed. Either way, we do _not_ block.
2358 /* We got the lock. */
2360 goto out_unlock_put_key;
2365 * Two reasons for this:
2366 * - Task is exiting and we just wait for the
2368 * - The user space value changed.
2371 put_futex_key(&q.key);
2375 goto out_unlock_put_key;
2380 * Only actually queue now that the atomic ops are done:
2384 WARN_ON(!q.pi_state);
2386 * Block on the PI mutex:
2389 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2391 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2392 /* Fixup the trylock return value: */
2393 ret = ret ? 0 : -EWOULDBLOCK;
2396 spin_lock(q.lock_ptr);
2398 * Fixup the pi_state owner and possibly acquire the lock if we
2401 res = fixup_owner(uaddr, &q, !ret);
2403 * If fixup_owner() returned an error, proprogate that. If it acquired
2404 * the lock, clear our -ETIMEDOUT or -EINTR.
2407 ret = (res < 0) ? res : 0;
2410 * If fixup_owner() faulted and was unable to handle the fault, unlock
2411 * it and return the fault to userspace.
2413 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2414 rt_mutex_unlock(&q.pi_state->pi_mutex);
2416 /* Unqueue and drop the lock */
2425 put_futex_key(&q.key);
2428 destroy_hrtimer_on_stack(&to->timer);
2429 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2434 ret = fault_in_user_writeable(uaddr);
2438 if (!(flags & FLAGS_SHARED))
2441 put_futex_key(&q.key);
2446 * Userspace attempted a TID -> 0 atomic transition, and failed.
2447 * This is the in-kernel slowpath: we look up the PI state (if any),
2448 * and do the rt-mutex unlock.
2450 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2452 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2453 union futex_key key = FUTEX_KEY_INIT;
2454 struct futex_hash_bucket *hb;
2455 struct futex_q *match;
2459 if (get_user(uval, uaddr))
2462 * We release only a lock we actually own:
2464 if ((uval & FUTEX_TID_MASK) != vpid)
2467 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2471 hb = hash_futex(&key);
2472 spin_lock(&hb->lock);
2475 * Check waiters first. We do not trust user space values at
2476 * all and we at least want to know if user space fiddled
2477 * with the futex value instead of blindly unlocking.
2479 match = futex_top_waiter(hb, &key);
2481 ret = wake_futex_pi(uaddr, uval, match, hb);
2483 * In case of success wake_futex_pi dropped the hash
2489 * The atomic access to the futex value generated a
2490 * pagefault, so retry the user-access and the wakeup:
2495 * wake_futex_pi has detected invalid state. Tell user
2502 * We have no kernel internal state, i.e. no waiters in the
2503 * kernel. Waiters which are about to queue themselves are stuck
2504 * on hb->lock. So we can safely ignore them. We do neither
2505 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2508 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2512 * If uval has changed, let user space handle it.
2514 ret = (curval == uval) ? 0 : -EAGAIN;
2517 spin_unlock(&hb->lock);
2519 put_futex_key(&key);
2523 spin_unlock(&hb->lock);
2524 put_futex_key(&key);
2526 ret = fault_in_user_writeable(uaddr);
2534 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2535 * @hb: the hash_bucket futex_q was original enqueued on
2536 * @q: the futex_q woken while waiting to be requeued
2537 * @key2: the futex_key of the requeue target futex
2538 * @timeout: the timeout associated with the wait (NULL if none)
2540 * Detect if the task was woken on the initial futex as opposed to the requeue
2541 * target futex. If so, determine if it was a timeout or a signal that caused
2542 * the wakeup and return the appropriate error code to the caller. Must be
2543 * called with the hb lock held.
2546 * 0 = no early wakeup detected;
2547 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2550 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2551 struct futex_q *q, union futex_key *key2,
2552 struct hrtimer_sleeper *timeout)
2557 * With the hb lock held, we avoid races while we process the wakeup.
2558 * We only need to hold hb (and not hb2) to ensure atomicity as the
2559 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2560 * It can't be requeued from uaddr2 to something else since we don't
2561 * support a PI aware source futex for requeue.
2563 if (!match_futex(&q->key, key2)) {
2564 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2566 * We were woken prior to requeue by a timeout or a signal.
2567 * Unqueue the futex_q and determine which it was.
2569 plist_del(&q->list, &hb->chain);
2572 /* Handle spurious wakeups gracefully */
2574 if (timeout && !timeout->task)
2576 else if (signal_pending(current))
2577 ret = -ERESTARTNOINTR;
2583 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2584 * @uaddr: the futex we initially wait on (non-pi)
2585 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2586 * the same type, no requeueing from private to shared, etc.
2587 * @val: the expected value of uaddr
2588 * @abs_time: absolute timeout
2589 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2590 * @uaddr2: the pi futex we will take prior to returning to user-space
2592 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2593 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2594 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2595 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2596 * without one, the pi logic would not know which task to boost/deboost, if
2597 * there was a need to.
2599 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2600 * via the following--
2601 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2602 * 2) wakeup on uaddr2 after a requeue
2606 * If 3, cleanup and return -ERESTARTNOINTR.
2608 * If 2, we may then block on trying to take the rt_mutex and return via:
2609 * 5) successful lock
2612 * 8) other lock acquisition failure
2614 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2616 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2622 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2623 u32 val, ktime_t *abs_time, u32 bitset,
2626 struct hrtimer_sleeper timeout, *to = NULL;
2627 struct rt_mutex_waiter rt_waiter;
2628 struct rt_mutex *pi_mutex = NULL;
2629 struct futex_hash_bucket *hb;
2630 union futex_key key2 = FUTEX_KEY_INIT;
2631 struct futex_q q = futex_q_init;
2634 if (uaddr == uaddr2)
2642 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2643 CLOCK_REALTIME : CLOCK_MONOTONIC,
2645 hrtimer_init_sleeper(to, current);
2646 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2647 current->timer_slack_ns);
2651 * The waiter is allocated on our stack, manipulated by the requeue
2652 * code while we sleep on uaddr.
2654 debug_rt_mutex_init_waiter(&rt_waiter);
2655 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2656 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2657 rt_waiter.task = NULL;
2659 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2660 if (unlikely(ret != 0))
2664 q.rt_waiter = &rt_waiter;
2665 q.requeue_pi_key = &key2;
2668 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2671 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2676 * The check above which compares uaddrs is not sufficient for
2677 * shared futexes. We need to compare the keys:
2679 if (match_futex(&q.key, &key2)) {
2685 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2686 futex_wait_queue_me(hb, &q, to);
2688 spin_lock(&hb->lock);
2689 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2690 spin_unlock(&hb->lock);
2695 * In order for us to be here, we know our q.key == key2, and since
2696 * we took the hb->lock above, we also know that futex_requeue() has
2697 * completed and we no longer have to concern ourselves with a wakeup
2698 * race with the atomic proxy lock acquisition by the requeue code. The
2699 * futex_requeue dropped our key1 reference and incremented our key2
2703 /* Check if the requeue code acquired the second futex for us. */
2706 * Got the lock. We might not be the anticipated owner if we
2707 * did a lock-steal - fix up the PI-state in that case.
2709 if (q.pi_state && (q.pi_state->owner != current)) {
2710 spin_lock(q.lock_ptr);
2711 ret = fixup_pi_state_owner(uaddr2, &q, current);
2712 spin_unlock(q.lock_ptr);
2716 * We have been woken up by futex_unlock_pi(), a timeout, or a
2717 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2720 WARN_ON(!q.pi_state);
2721 pi_mutex = &q.pi_state->pi_mutex;
2722 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2723 debug_rt_mutex_free_waiter(&rt_waiter);
2725 spin_lock(q.lock_ptr);
2727 * Fixup the pi_state owner and possibly acquire the lock if we
2730 res = fixup_owner(uaddr2, &q, !ret);
2732 * If fixup_owner() returned an error, proprogate that. If it
2733 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2736 ret = (res < 0) ? res : 0;
2738 /* Unqueue and drop the lock. */
2743 * If fixup_pi_state_owner() faulted and was unable to handle the
2744 * fault, unlock the rt_mutex and return the fault to userspace.
2746 if (ret == -EFAULT) {
2747 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2748 rt_mutex_unlock(pi_mutex);
2749 } else if (ret == -EINTR) {
2751 * We've already been requeued, but cannot restart by calling
2752 * futex_lock_pi() directly. We could restart this syscall, but
2753 * it would detect that the user space "val" changed and return
2754 * -EWOULDBLOCK. Save the overhead of the restart and return
2755 * -EWOULDBLOCK directly.
2761 put_futex_key(&q.key);
2763 put_futex_key(&key2);
2767 hrtimer_cancel(&to->timer);
2768 destroy_hrtimer_on_stack(&to->timer);
2774 * Support for robust futexes: the kernel cleans up held futexes at
2777 * Implementation: user-space maintains a per-thread list of locks it
2778 * is holding. Upon do_exit(), the kernel carefully walks this list,
2779 * and marks all locks that are owned by this thread with the
2780 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2781 * always manipulated with the lock held, so the list is private and
2782 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2783 * field, to allow the kernel to clean up if the thread dies after
2784 * acquiring the lock, but just before it could have added itself to
2785 * the list. There can only be one such pending lock.
2789 * sys_set_robust_list() - Set the robust-futex list head of a task
2790 * @head: pointer to the list-head
2791 * @len: length of the list-head, as userspace expects
2793 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2796 if (!futex_cmpxchg_enabled)
2799 * The kernel knows only one size for now:
2801 if (unlikely(len != sizeof(*head)))
2804 current->robust_list = head;
2810 * sys_get_robust_list() - Get the robust-futex list head of a task
2811 * @pid: pid of the process [zero for current task]
2812 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2813 * @len_ptr: pointer to a length field, the kernel fills in the header size
2815 SYSCALL_DEFINE3(get_robust_list, int, pid,
2816 struct robust_list_head __user * __user *, head_ptr,
2817 size_t __user *, len_ptr)
2819 struct robust_list_head __user *head;
2821 struct task_struct *p;
2823 if (!futex_cmpxchg_enabled)
2832 p = find_task_by_vpid(pid);
2838 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2841 head = p->robust_list;
2844 if (put_user(sizeof(*head), len_ptr))
2846 return put_user(head, head_ptr);
2855 * Process a futex-list entry, check whether it's owned by the
2856 * dying task, and do notification if so:
2858 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2860 u32 uval, uninitialized_var(nval), mval;
2863 if (get_user(uval, uaddr))
2866 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2868 * Ok, this dying thread is truly holding a futex
2869 * of interest. Set the OWNER_DIED bit atomically
2870 * via cmpxchg, and if the value had FUTEX_WAITERS
2871 * set, wake up a waiter (if any). (We have to do a
2872 * futex_wake() even if OWNER_DIED is already set -
2873 * to handle the rare but possible case of recursive
2874 * thread-death.) The rest of the cleanup is done in
2877 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2879 * We are not holding a lock here, but we want to have
2880 * the pagefault_disable/enable() protection because
2881 * we want to handle the fault gracefully. If the
2882 * access fails we try to fault in the futex with R/W
2883 * verification via get_user_pages. get_user() above
2884 * does not guarantee R/W access. If that fails we
2885 * give up and leave the futex locked.
2887 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2888 if (fault_in_user_writeable(uaddr))
2896 * Wake robust non-PI futexes here. The wakeup of
2897 * PI futexes happens in exit_pi_state():
2899 if (!pi && (uval & FUTEX_WAITERS))
2900 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2906 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2908 static inline int fetch_robust_entry(struct robust_list __user **entry,
2909 struct robust_list __user * __user *head,
2912 unsigned long uentry;
2914 if (get_user(uentry, (unsigned long __user *)head))
2917 *entry = (void __user *)(uentry & ~1UL);
2924 * Walk curr->robust_list (very carefully, it's a userspace list!)
2925 * and mark any locks found there dead, and notify any waiters.
2927 * We silently return on any sign of list-walking problem.
2929 void exit_robust_list(struct task_struct *curr)
2931 struct robust_list_head __user *head = curr->robust_list;
2932 struct robust_list __user *entry, *next_entry, *pending;
2933 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2934 unsigned int uninitialized_var(next_pi);
2935 unsigned long futex_offset;
2938 if (!futex_cmpxchg_enabled)
2942 * Fetch the list head (which was registered earlier, via
2943 * sys_set_robust_list()):
2945 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2948 * Fetch the relative futex offset:
2950 if (get_user(futex_offset, &head->futex_offset))
2953 * Fetch any possibly pending lock-add first, and handle it
2956 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2959 next_entry = NULL; /* avoid warning with gcc */
2960 while (entry != &head->list) {
2962 * Fetch the next entry in the list before calling
2963 * handle_futex_death:
2965 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2967 * A pending lock might already be on the list, so
2968 * don't process it twice:
2970 if (entry != pending)
2971 if (handle_futex_death((void __user *)entry + futex_offset,
2979 * Avoid excessively long or circular lists:
2988 handle_futex_death((void __user *)pending + futex_offset,
2992 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2993 u32 __user *uaddr2, u32 val2, u32 val3)
2995 int cmd = op & FUTEX_CMD_MASK;
2996 unsigned int flags = 0;
2998 if (!(op & FUTEX_PRIVATE_FLAG))
2999 flags |= FLAGS_SHARED;
3001 if (op & FUTEX_CLOCK_REALTIME) {
3002 flags |= FLAGS_CLOCKRT;
3003 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3009 case FUTEX_UNLOCK_PI:
3010 case FUTEX_TRYLOCK_PI:
3011 case FUTEX_WAIT_REQUEUE_PI:
3012 case FUTEX_CMP_REQUEUE_PI:
3013 if (!futex_cmpxchg_enabled)
3019 val3 = FUTEX_BITSET_MATCH_ANY;
3020 case FUTEX_WAIT_BITSET:
3021 return futex_wait(uaddr, flags, val, timeout, val3);
3023 val3 = FUTEX_BITSET_MATCH_ANY;
3024 case FUTEX_WAKE_BITSET:
3025 return futex_wake(uaddr, flags, val, val3);
3027 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3028 case FUTEX_CMP_REQUEUE:
3029 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3031 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3033 return futex_lock_pi(uaddr, flags, timeout, 0);
3034 case FUTEX_UNLOCK_PI:
3035 return futex_unlock_pi(uaddr, flags);
3036 case FUTEX_TRYLOCK_PI:
3037 return futex_lock_pi(uaddr, flags, NULL, 1);
3038 case FUTEX_WAIT_REQUEUE_PI:
3039 val3 = FUTEX_BITSET_MATCH_ANY;
3040 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3042 case FUTEX_CMP_REQUEUE_PI:
3043 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3049 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3050 struct timespec __user *, utime, u32 __user *, uaddr2,
3054 ktime_t t, *tp = NULL;
3056 int cmd = op & FUTEX_CMD_MASK;
3058 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3059 cmd == FUTEX_WAIT_BITSET ||
3060 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3061 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3063 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3065 if (!timespec_valid(&ts))
3068 t = timespec_to_ktime(ts);
3069 if (cmd == FUTEX_WAIT)
3070 t = ktime_add_safe(ktime_get(), t);
3074 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3075 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3077 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3078 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3079 val2 = (u32) (unsigned long) utime;
3081 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3084 static void __init futex_detect_cmpxchg(void)
3086 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3090 * This will fail and we want it. Some arch implementations do
3091 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3092 * functionality. We want to know that before we call in any
3093 * of the complex code paths. Also we want to prevent
3094 * registration of robust lists in that case. NULL is
3095 * guaranteed to fault and we get -EFAULT on functional
3096 * implementation, the non-functional ones will return
3099 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3100 futex_cmpxchg_enabled = 1;
3104 static int __init futex_init(void)
3106 unsigned int futex_shift;
3109 #if CONFIG_BASE_SMALL
3110 futex_hashsize = 16;
3112 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3115 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3117 futex_hashsize < 256 ? HASH_SMALL : 0,
3119 futex_hashsize, futex_hashsize);
3120 futex_hashsize = 1UL << futex_shift;
3122 futex_detect_cmpxchg();
3124 for (i = 0; i < futex_hashsize; i++) {
3125 atomic_set(&futex_queues[i].waiters, 0);
3126 plist_head_init(&futex_queues[i].chain);
3127 spin_lock_init(&futex_queues[i].lock);
3132 __initcall(futex_init);