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>
68 #include <asm/futex.h>
70 #include "locking/rtmutex_common.h"
73 * READ this before attempting to hack on futexes!
75 * Basic futex operation and ordering guarantees
76 * =============================================
78 * The waiter reads the futex value in user space and calls
79 * futex_wait(). This function computes the hash bucket and acquires
80 * the hash bucket lock. After that it reads the futex user space value
81 * again and verifies that the data has not changed. If it has not changed
82 * it enqueues itself into the hash bucket, releases the hash bucket lock
85 * The waker side modifies the user space value of the futex and calls
86 * futex_wake(). This function computes the hash bucket and acquires the
87 * hash bucket lock. Then it looks for waiters on that futex in the hash
88 * bucket and wakes them.
90 * In futex wake up scenarios where no tasks are blocked on a futex, taking
91 * the hb spinlock can be avoided and simply return. In order for this
92 * optimization to work, ordering guarantees must exist so that the waiter
93 * being added to the list is acknowledged when the list is concurrently being
94 * checked by the waker, avoiding scenarios like the following:
98 * sys_futex(WAIT, futex, val);
99 * futex_wait(futex, val);
102 * sys_futex(WAKE, futex);
107 * lock(hash_bucket(futex));
109 * unlock(hash_bucket(futex));
112 * This would cause the waiter on CPU 0 to wait forever because it
113 * missed the transition of the user space value from val to newval
114 * and the waker did not find the waiter in the hash bucket queue.
116 * The correct serialization ensures that a waiter either observes
117 * the changed user space value before blocking or is woken by a
122 * sys_futex(WAIT, futex, val);
123 * futex_wait(futex, val);
126 * mb(); (A) <-- paired with -.
128 * lock(hash_bucket(futex)); |
132 * | sys_futex(WAKE, futex);
133 * | futex_wake(futex);
135 * `-------> mb(); (B)
138 * unlock(hash_bucket(futex));
139 * schedule(); if (waiters)
140 * lock(hash_bucket(futex));
141 * else wake_waiters(futex);
142 * waiters--; (b) unlock(hash_bucket(futex));
144 * Where (A) orders the waiters increment and the futex value read through
145 * atomic operations (see hb_waiters_inc) and where (B) orders the write
146 * to futex and the waiters read -- this is done by the barriers for both
147 * shared and private futexes in get_futex_key_refs().
149 * This yields the following case (where X:=waiters, Y:=futex):
157 * Which guarantees that x==0 && y==0 is impossible; which translates back into
158 * the guarantee that we cannot both miss the futex variable change and the
161 * Note that a new waiter is accounted for in (a) even when it is possible that
162 * the wait call can return error, in which case we backtrack from it in (b).
163 * Refer to the comment in queue_lock().
165 * Similarly, in order to account for waiters being requeued on another
166 * address we always increment the waiters for the destination bucket before
167 * acquiring the lock. It then decrements them again after releasing it -
168 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
169 * will do the additional required waiter count housekeeping. This is done for
170 * double_lock_hb() and double_unlock_hb(), respectively.
173 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
174 int __read_mostly futex_cmpxchg_enabled;
178 * Futex flags used to encode options to functions and preserve them across
181 #define FLAGS_SHARED 0x01
182 #define FLAGS_CLOCKRT 0x02
183 #define FLAGS_HAS_TIMEOUT 0x04
186 * Priority Inheritance state:
188 struct futex_pi_state {
190 * list of 'owned' pi_state instances - these have to be
191 * cleaned up in do_exit() if the task exits prematurely:
193 struct list_head list;
198 struct rt_mutex pi_mutex;
200 struct task_struct *owner;
207 * struct futex_q - The hashed futex queue entry, one per waiting task
208 * @list: priority-sorted list of tasks waiting on this futex
209 * @task: the task waiting on the futex
210 * @lock_ptr: the hash bucket lock
211 * @key: the key the futex is hashed on
212 * @pi_state: optional priority inheritance state
213 * @rt_waiter: rt_waiter storage for use with requeue_pi
214 * @requeue_pi_key: the requeue_pi target futex key
215 * @bitset: bitset for the optional bitmasked wakeup
217 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
218 * we can wake only the relevant ones (hashed queues may be shared).
220 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
221 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
222 * The order of wakeup is always to make the first condition true, then
225 * PI futexes are typically woken before they are removed from the hash list via
226 * the rt_mutex code. See unqueue_me_pi().
229 struct plist_node list;
231 struct task_struct *task;
232 spinlock_t *lock_ptr;
234 struct futex_pi_state *pi_state;
235 struct rt_mutex_waiter *rt_waiter;
236 union futex_key *requeue_pi_key;
240 static const struct futex_q futex_q_init = {
241 /* list gets initialized in queue_me()*/
242 .key = FUTEX_KEY_INIT,
243 .bitset = FUTEX_BITSET_MATCH_ANY
247 * Hash buckets are shared by all the futex_keys that hash to the same
248 * location. Each key may have multiple futex_q structures, one for each task
249 * waiting on a futex.
251 struct futex_hash_bucket {
254 struct plist_head chain;
255 } ____cacheline_aligned_in_smp;
257 static unsigned long __read_mostly futex_hashsize;
259 static struct futex_hash_bucket *futex_queues;
261 static inline void futex_get_mm(union futex_key *key)
263 atomic_inc(&key->private.mm->mm_count);
265 * Ensure futex_get_mm() implies a full barrier such that
266 * get_futex_key() implies a full barrier. This is relied upon
267 * as full barrier (B), see the ordering comment above.
269 smp_mb__after_atomic();
273 * Reflects a new waiter being added to the waitqueue.
275 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
278 atomic_inc(&hb->waiters);
280 * Full barrier (A), see the ordering comment above.
282 smp_mb__after_atomic();
287 * Reflects a waiter being removed from the waitqueue by wakeup
290 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
293 atomic_dec(&hb->waiters);
297 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
300 return atomic_read(&hb->waiters);
307 * We hash on the keys returned from get_futex_key (see below).
309 static struct futex_hash_bucket *hash_futex(union futex_key *key)
311 u32 hash = jhash2((u32*)&key->both.word,
312 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
314 return &futex_queues[hash & (futex_hashsize - 1)];
318 * Return 1 if two futex_keys are equal, 0 otherwise.
320 static inline int match_futex(union futex_key *key1, union futex_key *key2)
323 && key1->both.word == key2->both.word
324 && key1->both.ptr == key2->both.ptr
325 && key1->both.offset == key2->both.offset);
329 * Take a reference to the resource addressed by a key.
330 * Can be called while holding spinlocks.
333 static void get_futex_key_refs(union futex_key *key)
338 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
340 ihold(key->shared.inode); /* implies MB (B) */
342 case FUT_OFF_MMSHARED:
343 futex_get_mm(key); /* implies MB (B) */
347 * Private futexes do not hold reference on an inode or
348 * mm, therefore the only purpose of calling get_futex_key_refs
349 * is because we need the barrier for the lockless waiter check.
351 smp_mb(); /* explicit MB (B) */
356 * Drop a reference to the resource addressed by a key.
357 * The hash bucket spinlock must not be held. This is
358 * a no-op for private futexes, see comment in the get
361 static void drop_futex_key_refs(union futex_key *key)
363 if (!key->both.ptr) {
364 /* If we're here then we tried to put a key we failed to get */
369 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
371 iput(key->shared.inode);
373 case FUT_OFF_MMSHARED:
374 mmdrop(key->private.mm);
380 * get_futex_key() - Get parameters which are the keys for a futex
381 * @uaddr: virtual address of the futex
382 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
383 * @key: address where result is stored.
384 * @rw: mapping needs to be read/write (values: VERIFY_READ,
387 * Return: a negative error code or 0
389 * The key words are stored in *key on success.
391 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
392 * offset_within_page). For private mappings, it's (uaddr, current->mm).
393 * We can usually work out the index without swapping in the page.
395 * lock_page() might sleep, the caller should not hold a spinlock.
398 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
400 unsigned long address = (unsigned long)uaddr;
401 struct mm_struct *mm = current->mm;
402 struct page *page, *page_head;
406 * The futex address must be "naturally" aligned.
408 key->both.offset = address % PAGE_SIZE;
409 if (unlikely((address % sizeof(u32)) != 0))
411 address -= key->both.offset;
413 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
417 * PROCESS_PRIVATE futexes are fast.
418 * As the mm cannot disappear under us and the 'key' only needs
419 * virtual address, we dont even have to find the underlying vma.
420 * Note : We do have to check 'uaddr' is a valid user address,
421 * but access_ok() should be faster than find_vma()
424 key->private.mm = mm;
425 key->private.address = address;
426 get_futex_key_refs(key); /* implies MB (B) */
431 err = get_user_pages_fast(address, 1, 1, &page);
433 * If write access is not required (eg. FUTEX_WAIT), try
434 * and get read-only access.
436 if (err == -EFAULT && rw == VERIFY_READ) {
437 err = get_user_pages_fast(address, 1, 0, &page);
445 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
447 if (unlikely(PageTail(page))) {
449 /* serialize against __split_huge_page_splitting() */
451 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
452 page_head = compound_head(page);
454 * page_head is valid pointer but we must pin
455 * it before taking the PG_lock and/or
456 * PG_compound_lock. The moment we re-enable
457 * irqs __split_huge_page_splitting() can
458 * return and the head page can be freed from
459 * under us. We can't take the PG_lock and/or
460 * PG_compound_lock on a page that could be
461 * freed from under us.
463 if (page != page_head) {
474 page_head = compound_head(page);
475 if (page != page_head) {
481 lock_page(page_head);
484 * If page_head->mapping is NULL, then it cannot be a PageAnon
485 * page; but it might be the ZERO_PAGE or in the gate area or
486 * in a special mapping (all cases which we are happy to fail);
487 * or it may have been a good file page when get_user_pages_fast
488 * found it, but truncated or holepunched or subjected to
489 * invalidate_complete_page2 before we got the page lock (also
490 * cases which we are happy to fail). And we hold a reference,
491 * so refcount care in invalidate_complete_page's remove_mapping
492 * prevents drop_caches from setting mapping to NULL beneath us.
494 * The case we do have to guard against is when memory pressure made
495 * shmem_writepage move it from filecache to swapcache beneath us:
496 * an unlikely race, but we do need to retry for page_head->mapping.
498 if (!page_head->mapping) {
499 int shmem_swizzled = PageSwapCache(page_head);
500 unlock_page(page_head);
508 * Private mappings are handled in a simple way.
510 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
511 * it's a read-only handle, it's expected that futexes attach to
512 * the object not the particular process.
514 if (PageAnon(page_head)) {
516 * A RO anonymous page will never change and thus doesn't make
517 * sense for futex operations.
524 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
525 key->private.mm = mm;
526 key->private.address = address;
528 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
529 key->shared.inode = page_head->mapping->host;
530 key->shared.pgoff = basepage_index(page);
533 get_futex_key_refs(key); /* implies MB (B) */
536 unlock_page(page_head);
541 static inline void put_futex_key(union futex_key *key)
543 drop_futex_key_refs(key);
547 * fault_in_user_writeable() - Fault in user address and verify RW access
548 * @uaddr: pointer to faulting user space address
550 * Slow path to fixup the fault we just took in the atomic write
553 * We have no generic implementation of a non-destructive write to the
554 * user address. We know that we faulted in the atomic pagefault
555 * disabled section so we can as well avoid the #PF overhead by
556 * calling get_user_pages() right away.
558 static int fault_in_user_writeable(u32 __user *uaddr)
560 struct mm_struct *mm = current->mm;
563 down_read(&mm->mmap_sem);
564 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
566 up_read(&mm->mmap_sem);
568 return ret < 0 ? ret : 0;
572 * futex_top_waiter() - Return the highest priority waiter on a futex
573 * @hb: the hash bucket the futex_q's reside in
574 * @key: the futex key (to distinguish it from other futex futex_q's)
576 * Must be called with the hb lock held.
578 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
579 union futex_key *key)
581 struct futex_q *this;
583 plist_for_each_entry(this, &hb->chain, list) {
584 if (match_futex(&this->key, key))
590 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
591 u32 uval, u32 newval)
596 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
602 static int get_futex_value_locked(u32 *dest, u32 __user *from)
607 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
610 return ret ? -EFAULT : 0;
617 static int refill_pi_state_cache(void)
619 struct futex_pi_state *pi_state;
621 if (likely(current->pi_state_cache))
624 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
629 INIT_LIST_HEAD(&pi_state->list);
630 /* pi_mutex gets initialized later */
631 pi_state->owner = NULL;
632 atomic_set(&pi_state->refcount, 1);
633 pi_state->key = FUTEX_KEY_INIT;
635 current->pi_state_cache = pi_state;
640 static struct futex_pi_state * alloc_pi_state(void)
642 struct futex_pi_state *pi_state = current->pi_state_cache;
645 current->pi_state_cache = NULL;
651 * Must be called with the hb lock held.
653 static void free_pi_state(struct futex_pi_state *pi_state)
658 if (!atomic_dec_and_test(&pi_state->refcount))
662 * If pi_state->owner is NULL, the owner is most probably dying
663 * and has cleaned up the pi_state already
665 if (pi_state->owner) {
666 raw_spin_lock_irq(&pi_state->owner->pi_lock);
667 list_del_init(&pi_state->list);
668 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
670 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
673 if (current->pi_state_cache)
677 * pi_state->list is already empty.
678 * clear pi_state->owner.
679 * refcount is at 0 - put it back to 1.
681 pi_state->owner = NULL;
682 atomic_set(&pi_state->refcount, 1);
683 current->pi_state_cache = pi_state;
688 * Look up the task based on what TID userspace gave us.
691 static struct task_struct * futex_find_get_task(pid_t pid)
693 struct task_struct *p;
696 p = find_task_by_vpid(pid);
706 * This task is holding PI mutexes at exit time => bad.
707 * Kernel cleans up PI-state, but userspace is likely hosed.
708 * (Robust-futex cleanup is separate and might save the day for userspace.)
710 void exit_pi_state_list(struct task_struct *curr)
712 struct list_head *next, *head = &curr->pi_state_list;
713 struct futex_pi_state *pi_state;
714 struct futex_hash_bucket *hb;
715 union futex_key key = FUTEX_KEY_INIT;
717 if (!futex_cmpxchg_enabled)
720 * We are a ZOMBIE and nobody can enqueue itself on
721 * pi_state_list anymore, but we have to be careful
722 * versus waiters unqueueing themselves:
724 raw_spin_lock_irq(&curr->pi_lock);
725 while (!list_empty(head)) {
728 pi_state = list_entry(next, struct futex_pi_state, list);
730 hb = hash_futex(&key);
731 raw_spin_unlock_irq(&curr->pi_lock);
733 spin_lock(&hb->lock);
735 raw_spin_lock_irq(&curr->pi_lock);
737 * We dropped the pi-lock, so re-check whether this
738 * task still owns the PI-state:
740 if (head->next != next) {
741 raw_spin_unlock_irq(&curr->pi_lock);
742 spin_unlock(&hb->lock);
743 raw_spin_lock_irq(&curr->pi_lock);
747 WARN_ON(pi_state->owner != curr);
748 WARN_ON(list_empty(&pi_state->list));
749 list_del_init(&pi_state->list);
750 pi_state->owner = NULL;
751 raw_spin_unlock_irq(&curr->pi_lock);
753 rt_mutex_unlock(&pi_state->pi_mutex);
755 spin_unlock(&hb->lock);
757 raw_spin_lock_irq(&curr->pi_lock);
759 raw_spin_unlock_irq(&curr->pi_lock);
763 * We need to check the following states:
765 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
767 * [1] NULL | --- | --- | 0 | 0/1 | Valid
768 * [2] NULL | --- | --- | >0 | 0/1 | Valid
770 * [3] Found | NULL | -- | Any | 0/1 | Invalid
772 * [4] Found | Found | NULL | 0 | 1 | Valid
773 * [5] Found | Found | NULL | >0 | 1 | Invalid
775 * [6] Found | Found | task | 0 | 1 | Valid
777 * [7] Found | Found | NULL | Any | 0 | Invalid
779 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
780 * [9] Found | Found | task | 0 | 0 | Invalid
781 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
783 * [1] Indicates that the kernel can acquire the futex atomically. We
784 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
786 * [2] Valid, if TID does not belong to a kernel thread. If no matching
787 * thread is found then it indicates that the owner TID has died.
789 * [3] Invalid. The waiter is queued on a non PI futex
791 * [4] Valid state after exit_robust_list(), which sets the user space
792 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
794 * [5] The user space value got manipulated between exit_robust_list()
795 * and exit_pi_state_list()
797 * [6] Valid state after exit_pi_state_list() which sets the new owner in
798 * the pi_state but cannot access the user space value.
800 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
802 * [8] Owner and user space value match
804 * [9] There is no transient state which sets the user space TID to 0
805 * except exit_robust_list(), but this is indicated by the
806 * FUTEX_OWNER_DIED bit. See [4]
808 * [10] There is no transient state which leaves owner and user space
813 * Validate that the existing waiter has a pi_state and sanity check
814 * the pi_state against the user space value. If correct, attach to
817 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
818 struct futex_pi_state **ps)
820 pid_t pid = uval & FUTEX_TID_MASK;
823 * Userspace might have messed up non-PI and PI futexes [3]
825 if (unlikely(!pi_state))
828 WARN_ON(!atomic_read(&pi_state->refcount));
831 * Handle the owner died case:
833 if (uval & FUTEX_OWNER_DIED) {
835 * exit_pi_state_list sets owner to NULL and wakes the
836 * topmost waiter. The task which acquires the
837 * pi_state->rt_mutex will fixup owner.
839 if (!pi_state->owner) {
841 * No pi state owner, but the user space TID
842 * is not 0. Inconsistent state. [5]
847 * Take a ref on the state and return success. [4]
853 * If TID is 0, then either the dying owner has not
854 * yet executed exit_pi_state_list() or some waiter
855 * acquired the rtmutex in the pi state, but did not
856 * yet fixup the TID in user space.
858 * Take a ref on the state and return success. [6]
864 * If the owner died bit is not set, then the pi_state
865 * must have an owner. [7]
867 if (!pi_state->owner)
872 * Bail out if user space manipulated the futex value. If pi
873 * state exists then the owner TID must be the same as the
874 * user space TID. [9/10]
876 if (pid != task_pid_vnr(pi_state->owner))
879 atomic_inc(&pi_state->refcount);
885 * Lookup the task for the TID provided from user space and attach to
886 * it after doing proper sanity checks.
888 static int attach_to_pi_owner(u32 uval, union futex_key *key,
889 struct futex_pi_state **ps)
891 pid_t pid = uval & FUTEX_TID_MASK;
892 struct futex_pi_state *pi_state;
893 struct task_struct *p;
896 * We are the first waiter - try to look up the real owner and attach
897 * the new pi_state to it, but bail out when TID = 0 [1]
901 p = futex_find_get_task(pid);
905 if (unlikely(p->flags & PF_KTHREAD)) {
911 * We need to look at the task state flags to figure out,
912 * whether the task is exiting. To protect against the do_exit
913 * change of the task flags, we do this protected by
916 raw_spin_lock_irq(&p->pi_lock);
917 if (unlikely(p->flags & PF_EXITING)) {
919 * The task is on the way out. When PF_EXITPIDONE is
920 * set, we know that the task has finished the
923 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
925 raw_spin_unlock_irq(&p->pi_lock);
931 * No existing pi state. First waiter. [2]
933 pi_state = alloc_pi_state();
936 * Initialize the pi_mutex in locked state and make @p
939 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
941 /* Store the key for possible exit cleanups: */
942 pi_state->key = *key;
944 WARN_ON(!list_empty(&pi_state->list));
945 list_add(&pi_state->list, &p->pi_state_list);
947 raw_spin_unlock_irq(&p->pi_lock);
956 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
957 union futex_key *key, struct futex_pi_state **ps)
959 struct futex_q *match = futex_top_waiter(hb, key);
962 * If there is a waiter on that futex, validate it and
963 * attach to the pi_state when the validation succeeds.
966 return attach_to_pi_state(uval, match->pi_state, ps);
969 * We are the first waiter - try to look up the owner based on
970 * @uval and attach to it.
972 return attach_to_pi_owner(uval, key, ps);
975 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
977 u32 uninitialized_var(curval);
979 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
982 /*If user space value changed, let the caller retry */
983 return curval != uval ? -EAGAIN : 0;
987 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
988 * @uaddr: the pi futex user address
989 * @hb: the pi futex hash bucket
990 * @key: the futex key associated with uaddr and hb
991 * @ps: the pi_state pointer where we store the result of the
993 * @task: the task to perform the atomic lock work for. This will
994 * be "current" except in the case of requeue pi.
995 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
999 * 1 - acquired the lock;
1002 * The hb->lock and futex_key refs shall be held by the caller.
1004 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1005 union futex_key *key,
1006 struct futex_pi_state **ps,
1007 struct task_struct *task, int set_waiters)
1009 u32 uval, newval, vpid = task_pid_vnr(task);
1010 struct futex_q *match;
1014 * Read the user space value first so we can validate a few
1015 * things before proceeding further.
1017 if (get_futex_value_locked(&uval, uaddr))
1023 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1027 * Lookup existing state first. If it exists, try to attach to
1030 match = futex_top_waiter(hb, key);
1032 return attach_to_pi_state(uval, match->pi_state, ps);
1035 * No waiter and user TID is 0. We are here because the
1036 * waiters or the owner died bit is set or called from
1037 * requeue_cmp_pi or for whatever reason something took the
1040 if (!(uval & FUTEX_TID_MASK)) {
1042 * We take over the futex. No other waiters and the user space
1043 * TID is 0. We preserve the owner died bit.
1045 newval = uval & FUTEX_OWNER_DIED;
1048 /* The futex requeue_pi code can enforce the waiters bit */
1050 newval |= FUTEX_WAITERS;
1052 ret = lock_pi_update_atomic(uaddr, uval, newval);
1053 /* If the take over worked, return 1 */
1054 return ret < 0 ? ret : 1;
1058 * First waiter. Set the waiters bit before attaching ourself to
1059 * the owner. If owner tries to unlock, it will be forced into
1060 * the kernel and blocked on hb->lock.
1062 newval = uval | FUTEX_WAITERS;
1063 ret = lock_pi_update_atomic(uaddr, uval, newval);
1067 * If the update of the user space value succeeded, we try to
1068 * attach to the owner. If that fails, no harm done, we only
1069 * set the FUTEX_WAITERS bit in the user space variable.
1071 return attach_to_pi_owner(uval, key, ps);
1075 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1076 * @q: The futex_q to unqueue
1078 * The q->lock_ptr must not be NULL and must be held by the caller.
1080 static void __unqueue_futex(struct futex_q *q)
1082 struct futex_hash_bucket *hb;
1084 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1085 || WARN_ON(plist_node_empty(&q->list)))
1088 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1089 plist_del(&q->list, &hb->chain);
1094 * The hash bucket lock must be held when this is called.
1095 * Afterwards, the futex_q must not be accessed. Callers
1096 * must ensure to later call wake_up_q() for the actual
1099 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1101 struct task_struct *p = q->task;
1103 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1107 * Queue the task for later wakeup for after we've released
1108 * the hb->lock. wake_q_add() grabs reference to p.
1110 wake_q_add(wake_q, p);
1113 * The waiting task can free the futex_q as soon as
1114 * q->lock_ptr = NULL is written, without taking any locks. A
1115 * memory barrier is required here to prevent the following
1116 * store to lock_ptr from getting ahead of the plist_del.
1122 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1123 struct futex_hash_bucket *hb)
1125 struct task_struct *new_owner;
1126 struct futex_pi_state *pi_state = this->pi_state;
1127 u32 uninitialized_var(curval), newval;
1135 * If current does not own the pi_state then the futex is
1136 * inconsistent and user space fiddled with the futex value.
1138 if (pi_state->owner != current)
1141 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1142 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1145 * It is possible that the next waiter (the one that brought
1146 * this owner to the kernel) timed out and is no longer
1147 * waiting on the lock.
1150 new_owner = this->task;
1153 * We pass it to the next owner. The WAITERS bit is always
1154 * kept enabled while there is PI state around. We cleanup the
1155 * owner died bit, because we are the owner.
1157 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1159 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1161 else if (curval != uval)
1164 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1168 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1169 WARN_ON(list_empty(&pi_state->list));
1170 list_del_init(&pi_state->list);
1171 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1173 raw_spin_lock_irq(&new_owner->pi_lock);
1174 WARN_ON(!list_empty(&pi_state->list));
1175 list_add(&pi_state->list, &new_owner->pi_state_list);
1176 pi_state->owner = new_owner;
1177 raw_spin_unlock_irq(&new_owner->pi_lock);
1179 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1181 deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex);
1184 * We deboost after dropping hb->lock. That prevents a double
1187 spin_unlock(&hb->lock);
1190 rt_mutex_adjust_prio(current);
1196 * Express the locking dependencies for lockdep:
1199 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1202 spin_lock(&hb1->lock);
1204 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1205 } else { /* hb1 > hb2 */
1206 spin_lock(&hb2->lock);
1207 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1212 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1214 spin_unlock(&hb1->lock);
1216 spin_unlock(&hb2->lock);
1220 * Wake up waiters matching bitset queued on this futex (uaddr).
1223 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1225 struct futex_hash_bucket *hb;
1226 struct futex_q *this, *next;
1227 union futex_key key = FUTEX_KEY_INIT;
1234 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1235 if (unlikely(ret != 0))
1238 hb = hash_futex(&key);
1240 /* Make sure we really have tasks to wakeup */
1241 if (!hb_waiters_pending(hb))
1244 spin_lock(&hb->lock);
1246 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1247 if (match_futex (&this->key, &key)) {
1248 if (this->pi_state || this->rt_waiter) {
1253 /* Check if one of the bits is set in both bitsets */
1254 if (!(this->bitset & bitset))
1257 mark_wake_futex(&wake_q, this);
1258 if (++ret >= nr_wake)
1263 spin_unlock(&hb->lock);
1266 put_futex_key(&key);
1272 * Wake up all waiters hashed on the physical page that is mapped
1273 * to this virtual address:
1276 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1277 int nr_wake, int nr_wake2, int op)
1279 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1280 struct futex_hash_bucket *hb1, *hb2;
1281 struct futex_q *this, *next;
1286 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1287 if (unlikely(ret != 0))
1289 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1290 if (unlikely(ret != 0))
1293 hb1 = hash_futex(&key1);
1294 hb2 = hash_futex(&key2);
1297 double_lock_hb(hb1, hb2);
1298 op_ret = futex_atomic_op_inuser(op, uaddr2);
1299 if (unlikely(op_ret < 0)) {
1301 double_unlock_hb(hb1, hb2);
1305 * we don't get EFAULT from MMU faults if we don't have an MMU,
1306 * but we might get them from range checking
1312 if (unlikely(op_ret != -EFAULT)) {
1317 ret = fault_in_user_writeable(uaddr2);
1321 if (!(flags & FLAGS_SHARED))
1324 put_futex_key(&key2);
1325 put_futex_key(&key1);
1329 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1330 if (match_futex (&this->key, &key1)) {
1331 if (this->pi_state || this->rt_waiter) {
1335 mark_wake_futex(&wake_q, this);
1336 if (++ret >= nr_wake)
1343 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1344 if (match_futex (&this->key, &key2)) {
1345 if (this->pi_state || this->rt_waiter) {
1349 mark_wake_futex(&wake_q, this);
1350 if (++op_ret >= nr_wake2)
1358 double_unlock_hb(hb1, hb2);
1361 put_futex_key(&key2);
1363 put_futex_key(&key1);
1369 * requeue_futex() - Requeue a futex_q from one hb to another
1370 * @q: the futex_q to requeue
1371 * @hb1: the source hash_bucket
1372 * @hb2: the target hash_bucket
1373 * @key2: the new key for the requeued futex_q
1376 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1377 struct futex_hash_bucket *hb2, union futex_key *key2)
1381 * If key1 and key2 hash to the same bucket, no need to
1384 if (likely(&hb1->chain != &hb2->chain)) {
1385 plist_del(&q->list, &hb1->chain);
1386 hb_waiters_dec(hb1);
1387 plist_add(&q->list, &hb2->chain);
1388 hb_waiters_inc(hb2);
1389 q->lock_ptr = &hb2->lock;
1391 get_futex_key_refs(key2);
1396 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1398 * @key: the key of the requeue target futex
1399 * @hb: the hash_bucket of the requeue target futex
1401 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1402 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1403 * to the requeue target futex so the waiter can detect the wakeup on the right
1404 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1405 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1406 * to protect access to the pi_state to fixup the owner later. Must be called
1407 * with both q->lock_ptr and hb->lock held.
1410 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1411 struct futex_hash_bucket *hb)
1413 get_futex_key_refs(key);
1418 WARN_ON(!q->rt_waiter);
1419 q->rt_waiter = NULL;
1421 q->lock_ptr = &hb->lock;
1423 wake_up_state(q->task, TASK_NORMAL);
1427 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1428 * @pifutex: the user address of the to futex
1429 * @hb1: the from futex hash bucket, must be locked by the caller
1430 * @hb2: the to futex hash bucket, must be locked by the caller
1431 * @key1: the from futex key
1432 * @key2: the to futex key
1433 * @ps: address to store the pi_state pointer
1434 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1436 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1437 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1438 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1439 * hb1 and hb2 must be held by the caller.
1442 * 0 - failed to acquire the lock atomically;
1443 * >0 - acquired the lock, return value is vpid of the top_waiter
1446 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1447 struct futex_hash_bucket *hb1,
1448 struct futex_hash_bucket *hb2,
1449 union futex_key *key1, union futex_key *key2,
1450 struct futex_pi_state **ps, int set_waiters)
1452 struct futex_q *top_waiter = NULL;
1456 if (get_futex_value_locked(&curval, pifutex))
1460 * Find the top_waiter and determine if there are additional waiters.
1461 * If the caller intends to requeue more than 1 waiter to pifutex,
1462 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1463 * as we have means to handle the possible fault. If not, don't set
1464 * the bit unecessarily as it will force the subsequent unlock to enter
1467 top_waiter = futex_top_waiter(hb1, key1);
1469 /* There are no waiters, nothing for us to do. */
1473 /* Ensure we requeue to the expected futex. */
1474 if (!match_futex(top_waiter->requeue_pi_key, key2))
1478 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1479 * the contended case or if set_waiters is 1. The pi_state is returned
1480 * in ps in contended cases.
1482 vpid = task_pid_vnr(top_waiter->task);
1483 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1486 requeue_pi_wake_futex(top_waiter, key2, hb2);
1493 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1494 * @uaddr1: source futex user address
1495 * @flags: futex flags (FLAGS_SHARED, etc.)
1496 * @uaddr2: target futex user address
1497 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1498 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1499 * @cmpval: @uaddr1 expected value (or %NULL)
1500 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1501 * pi futex (pi to pi requeue is not supported)
1503 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1504 * uaddr2 atomically on behalf of the top waiter.
1507 * >=0 - on success, the number of tasks requeued or woken;
1510 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1511 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1512 u32 *cmpval, int requeue_pi)
1514 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1515 int drop_count = 0, task_count = 0, ret;
1516 struct futex_pi_state *pi_state = NULL;
1517 struct futex_hash_bucket *hb1, *hb2;
1518 struct futex_q *this, *next;
1523 * Requeue PI only works on two distinct uaddrs. This
1524 * check is only valid for private futexes. See below.
1526 if (uaddr1 == uaddr2)
1530 * requeue_pi requires a pi_state, try to allocate it now
1531 * without any locks in case it fails.
1533 if (refill_pi_state_cache())
1536 * requeue_pi must wake as many tasks as it can, up to nr_wake
1537 * + nr_requeue, since it acquires the rt_mutex prior to
1538 * returning to userspace, so as to not leave the rt_mutex with
1539 * waiters and no owner. However, second and third wake-ups
1540 * cannot be predicted as they involve race conditions with the
1541 * first wake and a fault while looking up the pi_state. Both
1542 * pthread_cond_signal() and pthread_cond_broadcast() should
1550 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1551 if (unlikely(ret != 0))
1553 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1554 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1555 if (unlikely(ret != 0))
1559 * The check above which compares uaddrs is not sufficient for
1560 * shared futexes. We need to compare the keys:
1562 if (requeue_pi && match_futex(&key1, &key2)) {
1567 hb1 = hash_futex(&key1);
1568 hb2 = hash_futex(&key2);
1571 hb_waiters_inc(hb2);
1572 double_lock_hb(hb1, hb2);
1574 if (likely(cmpval != NULL)) {
1577 ret = get_futex_value_locked(&curval, uaddr1);
1579 if (unlikely(ret)) {
1580 double_unlock_hb(hb1, hb2);
1581 hb_waiters_dec(hb2);
1583 ret = get_user(curval, uaddr1);
1587 if (!(flags & FLAGS_SHARED))
1590 put_futex_key(&key2);
1591 put_futex_key(&key1);
1594 if (curval != *cmpval) {
1600 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1602 * Attempt to acquire uaddr2 and wake the top waiter. If we
1603 * intend to requeue waiters, force setting the FUTEX_WAITERS
1604 * bit. We force this here where we are able to easily handle
1605 * faults rather in the requeue loop below.
1607 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1608 &key2, &pi_state, nr_requeue);
1611 * At this point the top_waiter has either taken uaddr2 or is
1612 * waiting on it. If the former, then the pi_state will not
1613 * exist yet, look it up one more time to ensure we have a
1614 * reference to it. If the lock was taken, ret contains the
1615 * vpid of the top waiter task.
1622 * If we acquired the lock, then the user
1623 * space value of uaddr2 should be vpid. It
1624 * cannot be changed by the top waiter as it
1625 * is blocked on hb2 lock if it tries to do
1626 * so. If something fiddled with it behind our
1627 * back the pi state lookup might unearth
1628 * it. So we rather use the known value than
1629 * rereading and handing potential crap to
1632 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1639 free_pi_state(pi_state);
1641 double_unlock_hb(hb1, hb2);
1642 hb_waiters_dec(hb2);
1643 put_futex_key(&key2);
1644 put_futex_key(&key1);
1645 ret = fault_in_user_writeable(uaddr2);
1651 * Two reasons for this:
1652 * - Owner is exiting and we just wait for the
1654 * - The user space value changed.
1656 free_pi_state(pi_state);
1658 double_unlock_hb(hb1, hb2);
1659 hb_waiters_dec(hb2);
1660 put_futex_key(&key2);
1661 put_futex_key(&key1);
1669 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1670 if (task_count - nr_wake >= nr_requeue)
1673 if (!match_futex(&this->key, &key1))
1677 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1678 * be paired with each other and no other futex ops.
1680 * We should never be requeueing a futex_q with a pi_state,
1681 * which is awaiting a futex_unlock_pi().
1683 if ((requeue_pi && !this->rt_waiter) ||
1684 (!requeue_pi && this->rt_waiter) ||
1691 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1692 * lock, we already woke the top_waiter. If not, it will be
1693 * woken by futex_unlock_pi().
1695 if (++task_count <= nr_wake && !requeue_pi) {
1696 mark_wake_futex(&wake_q, this);
1700 /* Ensure we requeue to the expected futex for requeue_pi. */
1701 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1707 * Requeue nr_requeue waiters and possibly one more in the case
1708 * of requeue_pi if we couldn't acquire the lock atomically.
1711 /* Prepare the waiter to take the rt_mutex. */
1712 atomic_inc(&pi_state->refcount);
1713 this->pi_state = pi_state;
1714 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1718 /* We got the lock. */
1719 requeue_pi_wake_futex(this, &key2, hb2);
1722 } else if (ret == -EAGAIN) {
1724 * Waiter was woken by timeout or
1725 * signal and has set pi_blocked_on to
1726 * PI_WAKEUP_INPROGRESS before we
1727 * tried to enqueue it on the rtmutex.
1729 this->pi_state = NULL;
1730 free_pi_state(pi_state);
1734 this->pi_state = NULL;
1735 free_pi_state(pi_state);
1739 requeue_futex(this, hb1, hb2, &key2);
1744 free_pi_state(pi_state);
1745 double_unlock_hb(hb1, hb2);
1747 hb_waiters_dec(hb2);
1750 * drop_futex_key_refs() must be called outside the spinlocks. During
1751 * the requeue we moved futex_q's from the hash bucket at key1 to the
1752 * one at key2 and updated their key pointer. We no longer need to
1753 * hold the references to key1.
1755 while (--drop_count >= 0)
1756 drop_futex_key_refs(&key1);
1759 put_futex_key(&key2);
1761 put_futex_key(&key1);
1763 return ret ? ret : task_count;
1766 /* The key must be already stored in q->key. */
1767 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1768 __acquires(&hb->lock)
1770 struct futex_hash_bucket *hb;
1772 hb = hash_futex(&q->key);
1775 * Increment the counter before taking the lock so that
1776 * a potential waker won't miss a to-be-slept task that is
1777 * waiting for the spinlock. This is safe as all queue_lock()
1778 * users end up calling queue_me(). Similarly, for housekeeping,
1779 * decrement the counter at queue_unlock() when some error has
1780 * occurred and we don't end up adding the task to the list.
1784 q->lock_ptr = &hb->lock;
1786 spin_lock(&hb->lock); /* implies MB (A) */
1791 queue_unlock(struct futex_hash_bucket *hb)
1792 __releases(&hb->lock)
1794 spin_unlock(&hb->lock);
1799 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1800 * @q: The futex_q to enqueue
1801 * @hb: The destination hash bucket
1803 * The hb->lock must be held by the caller, and is released here. A call to
1804 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1805 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1806 * or nothing if the unqueue is done as part of the wake process and the unqueue
1807 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1810 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1811 __releases(&hb->lock)
1816 * The priority used to register this element is
1817 * - either the real thread-priority for the real-time threads
1818 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1819 * - or MAX_RT_PRIO for non-RT threads.
1820 * Thus, all RT-threads are woken first in priority order, and
1821 * the others are woken last, in FIFO order.
1823 prio = min(current->normal_prio, MAX_RT_PRIO);
1825 plist_node_init(&q->list, prio);
1826 plist_add(&q->list, &hb->chain);
1828 spin_unlock(&hb->lock);
1832 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1833 * @q: The futex_q to unqueue
1835 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1836 * be paired with exactly one earlier call to queue_me().
1839 * 1 - if the futex_q was still queued (and we removed unqueued it);
1840 * 0 - if the futex_q was already removed by the waking thread
1842 static int unqueue_me(struct futex_q *q)
1844 spinlock_t *lock_ptr;
1847 /* In the common case we don't take the spinlock, which is nice. */
1849 lock_ptr = q->lock_ptr;
1851 if (lock_ptr != NULL) {
1852 spin_lock(lock_ptr);
1854 * q->lock_ptr can change between reading it and
1855 * spin_lock(), causing us to take the wrong lock. This
1856 * corrects the race condition.
1858 * Reasoning goes like this: if we have the wrong lock,
1859 * q->lock_ptr must have changed (maybe several times)
1860 * between reading it and the spin_lock(). It can
1861 * change again after the spin_lock() but only if it was
1862 * already changed before the spin_lock(). It cannot,
1863 * however, change back to the original value. Therefore
1864 * we can detect whether we acquired the correct lock.
1866 if (unlikely(lock_ptr != q->lock_ptr)) {
1867 spin_unlock(lock_ptr);
1872 BUG_ON(q->pi_state);
1874 spin_unlock(lock_ptr);
1878 drop_futex_key_refs(&q->key);
1883 * PI futexes can not be requeued and must remove themself from the
1884 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1887 static void unqueue_me_pi(struct futex_q *q)
1888 __releases(q->lock_ptr)
1892 BUG_ON(!q->pi_state);
1893 free_pi_state(q->pi_state);
1896 spin_unlock(q->lock_ptr);
1900 * Fixup the pi_state owner with the new owner.
1902 * Must be called with hash bucket lock held and mm->sem held for non
1905 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1906 struct task_struct *newowner)
1908 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1909 struct futex_pi_state *pi_state = q->pi_state;
1910 struct task_struct *oldowner = pi_state->owner;
1911 u32 uval, uninitialized_var(curval), newval;
1915 if (!pi_state->owner)
1916 newtid |= FUTEX_OWNER_DIED;
1919 * We are here either because we stole the rtmutex from the
1920 * previous highest priority waiter or we are the highest priority
1921 * waiter but failed to get the rtmutex the first time.
1922 * We have to replace the newowner TID in the user space variable.
1923 * This must be atomic as we have to preserve the owner died bit here.
1925 * Note: We write the user space value _before_ changing the pi_state
1926 * because we can fault here. Imagine swapped out pages or a fork
1927 * that marked all the anonymous memory readonly for cow.
1929 * Modifying pi_state _before_ the user space value would
1930 * leave the pi_state in an inconsistent state when we fault
1931 * here, because we need to drop the hash bucket lock to
1932 * handle the fault. This might be observed in the PID check
1933 * in lookup_pi_state.
1936 if (get_futex_value_locked(&uval, uaddr))
1940 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1942 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1950 * We fixed up user space. Now we need to fix the pi_state
1953 if (pi_state->owner != NULL) {
1954 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1955 WARN_ON(list_empty(&pi_state->list));
1956 list_del_init(&pi_state->list);
1957 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1960 pi_state->owner = newowner;
1962 raw_spin_lock_irq(&newowner->pi_lock);
1963 WARN_ON(!list_empty(&pi_state->list));
1964 list_add(&pi_state->list, &newowner->pi_state_list);
1965 raw_spin_unlock_irq(&newowner->pi_lock);
1969 * To handle the page fault we need to drop the hash bucket
1970 * lock here. That gives the other task (either the highest priority
1971 * waiter itself or the task which stole the rtmutex) the
1972 * chance to try the fixup of the pi_state. So once we are
1973 * back from handling the fault we need to check the pi_state
1974 * after reacquiring the hash bucket lock and before trying to
1975 * do another fixup. When the fixup has been done already we
1979 spin_unlock(q->lock_ptr);
1981 ret = fault_in_user_writeable(uaddr);
1983 spin_lock(q->lock_ptr);
1986 * Check if someone else fixed it for us:
1988 if (pi_state->owner != oldowner)
1997 static long futex_wait_restart(struct restart_block *restart);
2000 * fixup_owner() - Post lock pi_state and corner case management
2001 * @uaddr: user address of the futex
2002 * @q: futex_q (contains pi_state and access to the rt_mutex)
2003 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2005 * After attempting to lock an rt_mutex, this function is called to cleanup
2006 * the pi_state owner as well as handle race conditions that may allow us to
2007 * acquire the lock. Must be called with the hb lock held.
2010 * 1 - success, lock taken;
2011 * 0 - success, lock not taken;
2012 * <0 - on error (-EFAULT)
2014 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2016 struct task_struct *owner;
2021 * Got the lock. We might not be the anticipated owner if we
2022 * did a lock-steal - fix up the PI-state in that case:
2024 if (q->pi_state->owner != current)
2025 ret = fixup_pi_state_owner(uaddr, q, current);
2030 * Catch the rare case, where the lock was released when we were on the
2031 * way back before we locked the hash bucket.
2033 if (q->pi_state->owner == current) {
2035 * Try to get the rt_mutex now. This might fail as some other
2036 * task acquired the rt_mutex after we removed ourself from the
2037 * rt_mutex waiters list.
2039 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2045 * pi_state is incorrect, some other task did a lock steal and
2046 * we returned due to timeout or signal without taking the
2047 * rt_mutex. Too late.
2049 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2050 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2052 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2053 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2054 ret = fixup_pi_state_owner(uaddr, q, owner);
2059 * Paranoia check. If we did not take the lock, then we should not be
2060 * the owner of the rt_mutex.
2062 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2063 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2064 "pi-state %p\n", ret,
2065 q->pi_state->pi_mutex.owner,
2066 q->pi_state->owner);
2069 return ret ? ret : locked;
2073 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2074 * @hb: the futex hash bucket, must be locked by the caller
2075 * @q: the futex_q to queue up on
2076 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2078 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2079 struct hrtimer_sleeper *timeout)
2082 * The task state is guaranteed to be set before another task can
2083 * wake it. set_current_state() is implemented using set_mb() and
2084 * queue_me() calls spin_unlock() upon completion, both serializing
2085 * access to the hash list and forcing another memory barrier.
2087 set_current_state(TASK_INTERRUPTIBLE);
2092 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2093 if (!hrtimer_active(&timeout->timer))
2094 timeout->task = NULL;
2098 * If we have been removed from the hash list, then another task
2099 * has tried to wake us, and we can skip the call to schedule().
2101 if (likely(!plist_node_empty(&q->list))) {
2103 * If the timer has already expired, current will already be
2104 * flagged for rescheduling. Only call schedule if there
2105 * is no timeout, or if it has yet to expire.
2107 if (!timeout || timeout->task)
2108 freezable_schedule();
2110 __set_current_state(TASK_RUNNING);
2114 * futex_wait_setup() - Prepare to wait on a futex
2115 * @uaddr: the futex userspace address
2116 * @val: the expected value
2117 * @flags: futex flags (FLAGS_SHARED, etc.)
2118 * @q: the associated futex_q
2119 * @hb: storage for hash_bucket pointer to be returned to caller
2121 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2122 * compare it with the expected value. Handle atomic faults internally.
2123 * Return with the hb lock held and a q.key reference on success, and unlocked
2124 * with no q.key reference on failure.
2127 * 0 - uaddr contains val and hb has been locked;
2128 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2130 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2131 struct futex_q *q, struct futex_hash_bucket **hb)
2137 * Access the page AFTER the hash-bucket is locked.
2138 * Order is important:
2140 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2141 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2143 * The basic logical guarantee of a futex is that it blocks ONLY
2144 * if cond(var) is known to be true at the time of blocking, for
2145 * any cond. If we locked the hash-bucket after testing *uaddr, that
2146 * would open a race condition where we could block indefinitely with
2147 * cond(var) false, which would violate the guarantee.
2149 * On the other hand, we insert q and release the hash-bucket only
2150 * after testing *uaddr. This guarantees that futex_wait() will NOT
2151 * absorb a wakeup if *uaddr does not match the desired values
2152 * while the syscall executes.
2155 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2156 if (unlikely(ret != 0))
2160 *hb = queue_lock(q);
2162 ret = get_futex_value_locked(&uval, uaddr);
2167 ret = get_user(uval, uaddr);
2171 if (!(flags & FLAGS_SHARED))
2174 put_futex_key(&q->key);
2185 put_futex_key(&q->key);
2189 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2190 ktime_t *abs_time, u32 bitset)
2192 struct hrtimer_sleeper timeout, *to = NULL;
2193 struct restart_block *restart;
2194 struct futex_hash_bucket *hb;
2195 struct futex_q q = futex_q_init;
2205 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2206 CLOCK_REALTIME : CLOCK_MONOTONIC,
2208 hrtimer_init_sleeper(to, current);
2209 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2210 current->timer_slack_ns);
2215 * Prepare to wait on uaddr. On success, holds hb lock and increments
2218 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2222 /* queue_me and wait for wakeup, timeout, or a signal. */
2223 futex_wait_queue_me(hb, &q, to);
2225 /* If we were woken (and unqueued), we succeeded, whatever. */
2227 /* unqueue_me() drops q.key ref */
2228 if (!unqueue_me(&q))
2231 if (to && !to->task)
2235 * We expect signal_pending(current), but we might be the
2236 * victim of a spurious wakeup as well.
2238 if (!signal_pending(current))
2245 restart = ¤t->restart_block;
2246 restart->fn = futex_wait_restart;
2247 restart->futex.uaddr = uaddr;
2248 restart->futex.val = val;
2249 restart->futex.time = abs_time->tv64;
2250 restart->futex.bitset = bitset;
2251 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2253 ret = -ERESTART_RESTARTBLOCK;
2257 hrtimer_cancel(&to->timer);
2258 destroy_hrtimer_on_stack(&to->timer);
2264 static long futex_wait_restart(struct restart_block *restart)
2266 u32 __user *uaddr = restart->futex.uaddr;
2267 ktime_t t, *tp = NULL;
2269 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2270 t.tv64 = restart->futex.time;
2273 restart->fn = do_no_restart_syscall;
2275 return (long)futex_wait(uaddr, restart->futex.flags,
2276 restart->futex.val, tp, restart->futex.bitset);
2281 * Userspace tried a 0 -> TID atomic transition of the futex value
2282 * and failed. The kernel side here does the whole locking operation:
2283 * if there are waiters then it will block, it does PI, etc. (Due to
2284 * races the kernel might see a 0 value of the futex too.)
2286 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2287 ktime_t *time, int trylock)
2289 struct hrtimer_sleeper timeout, *to = NULL;
2290 struct futex_hash_bucket *hb;
2291 struct futex_q q = futex_q_init;
2294 if (refill_pi_state_cache())
2299 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2301 hrtimer_init_sleeper(to, current);
2302 hrtimer_set_expires(&to->timer, *time);
2306 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2307 if (unlikely(ret != 0))
2311 hb = queue_lock(&q);
2313 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2314 if (unlikely(ret)) {
2317 /* We got the lock. */
2319 goto out_unlock_put_key;
2324 * Two reasons for this:
2325 * - Task is exiting and we just wait for the
2327 * - The user space value changed.
2330 put_futex_key(&q.key);
2334 goto out_unlock_put_key;
2339 * Only actually queue now that the atomic ops are done:
2343 WARN_ON(!q.pi_state);
2345 * Block on the PI mutex:
2348 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2350 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2351 /* Fixup the trylock return value: */
2352 ret = ret ? 0 : -EWOULDBLOCK;
2355 spin_lock(q.lock_ptr);
2357 * Fixup the pi_state owner and possibly acquire the lock if we
2360 res = fixup_owner(uaddr, &q, !ret);
2362 * If fixup_owner() returned an error, proprogate that. If it acquired
2363 * the lock, clear our -ETIMEDOUT or -EINTR.
2366 ret = (res < 0) ? res : 0;
2369 * If fixup_owner() faulted and was unable to handle the fault, unlock
2370 * it and return the fault to userspace.
2372 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2373 rt_mutex_unlock(&q.pi_state->pi_mutex);
2375 /* Unqueue and drop the lock */
2384 put_futex_key(&q.key);
2387 destroy_hrtimer_on_stack(&to->timer);
2388 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2393 ret = fault_in_user_writeable(uaddr);
2397 if (!(flags & FLAGS_SHARED))
2400 put_futex_key(&q.key);
2405 * Userspace attempted a TID -> 0 atomic transition, and failed.
2406 * This is the in-kernel slowpath: we look up the PI state (if any),
2407 * and do the rt-mutex unlock.
2409 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2411 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2412 union futex_key key = FUTEX_KEY_INIT;
2413 struct futex_hash_bucket *hb;
2414 struct futex_q *match;
2418 if (get_user(uval, uaddr))
2421 * We release only a lock we actually own:
2423 if ((uval & FUTEX_TID_MASK) != vpid)
2426 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2430 hb = hash_futex(&key);
2431 spin_lock(&hb->lock);
2434 * Check waiters first. We do not trust user space values at
2435 * all and we at least want to know if user space fiddled
2436 * with the futex value instead of blindly unlocking.
2438 match = futex_top_waiter(hb, &key);
2440 ret = wake_futex_pi(uaddr, uval, match, hb);
2443 * In case of success wake_futex_pi dropped the hash
2450 * The atomic access to the futex value generated a
2451 * pagefault, so retry the user-access and the wakeup:
2457 * wake_futex_pi has detected invalid state. Tell user
2464 * We have no kernel internal state, i.e. no waiters in the
2465 * kernel. Waiters which are about to queue themselves are stuck
2466 * on hb->lock. So we can safely ignore them. We do neither
2467 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2470 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2474 * If uval has changed, let user space handle it.
2476 ret = (curval == uval) ? 0 : -EAGAIN;
2479 spin_unlock(&hb->lock);
2481 put_futex_key(&key);
2485 spin_unlock(&hb->lock);
2486 put_futex_key(&key);
2488 ret = fault_in_user_writeable(uaddr);
2496 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2497 * @hb: the hash_bucket futex_q was original enqueued on
2498 * @q: the futex_q woken while waiting to be requeued
2499 * @key2: the futex_key of the requeue target futex
2500 * @timeout: the timeout associated with the wait (NULL if none)
2502 * Detect if the task was woken on the initial futex as opposed to the requeue
2503 * target futex. If so, determine if it was a timeout or a signal that caused
2504 * the wakeup and return the appropriate error code to the caller. Must be
2505 * called with the hb lock held.
2508 * 0 = no early wakeup detected;
2509 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2512 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2513 struct futex_q *q, union futex_key *key2,
2514 struct hrtimer_sleeper *timeout)
2519 * With the hb lock held, we avoid races while we process the wakeup.
2520 * We only need to hold hb (and not hb2) to ensure atomicity as the
2521 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2522 * It can't be requeued from uaddr2 to something else since we don't
2523 * support a PI aware source futex for requeue.
2525 if (!match_futex(&q->key, key2)) {
2526 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2528 * We were woken prior to requeue by a timeout or a signal.
2529 * Unqueue the futex_q and determine which it was.
2531 plist_del(&q->list, &hb->chain);
2534 /* Handle spurious wakeups gracefully */
2536 if (timeout && !timeout->task)
2538 else if (signal_pending(current))
2539 ret = -ERESTARTNOINTR;
2545 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2546 * @uaddr: the futex we initially wait on (non-pi)
2547 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2548 * the same type, no requeueing from private to shared, etc.
2549 * @val: the expected value of uaddr
2550 * @abs_time: absolute timeout
2551 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2552 * @uaddr2: the pi futex we will take prior to returning to user-space
2554 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2555 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2556 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2557 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2558 * without one, the pi logic would not know which task to boost/deboost, if
2559 * there was a need to.
2561 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2562 * via the following--
2563 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2564 * 2) wakeup on uaddr2 after a requeue
2568 * If 3, cleanup and return -ERESTARTNOINTR.
2570 * If 2, we may then block on trying to take the rt_mutex and return via:
2571 * 5) successful lock
2574 * 8) other lock acquisition failure
2576 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2578 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2584 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2585 u32 val, ktime_t *abs_time, u32 bitset,
2588 struct hrtimer_sleeper timeout, *to = NULL;
2589 struct rt_mutex_waiter rt_waiter;
2590 struct rt_mutex *pi_mutex = NULL;
2591 struct futex_hash_bucket *hb, *hb2;
2592 union futex_key key2 = FUTEX_KEY_INIT;
2593 struct futex_q q = futex_q_init;
2596 if (uaddr == uaddr2)
2604 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2605 CLOCK_REALTIME : CLOCK_MONOTONIC,
2607 hrtimer_init_sleeper(to, current);
2608 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2609 current->timer_slack_ns);
2613 * The waiter is allocated on our stack, manipulated by the requeue
2614 * code while we sleep on uaddr.
2616 rt_mutex_init_waiter(&rt_waiter, false);
2618 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2619 if (unlikely(ret != 0))
2623 q.rt_waiter = &rt_waiter;
2624 q.requeue_pi_key = &key2;
2627 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2630 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2635 * The check above which compares uaddrs is not sufficient for
2636 * shared futexes. We need to compare the keys:
2638 if (match_futex(&q.key, &key2)) {
2644 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2645 futex_wait_queue_me(hb, &q, to);
2648 * On RT we must avoid races with requeue and trying to block
2649 * on two mutexes (hb->lock and uaddr2's rtmutex) by
2650 * serializing access to pi_blocked_on with pi_lock.
2652 raw_spin_lock_irq(¤t->pi_lock);
2653 if (current->pi_blocked_on) {
2655 * We have been requeued or are in the process of
2658 raw_spin_unlock_irq(¤t->pi_lock);
2661 * Setting pi_blocked_on to PI_WAKEUP_INPROGRESS
2662 * prevents a concurrent requeue from moving us to the
2663 * uaddr2 rtmutex. After that we can safely acquire
2664 * (and possibly block on) hb->lock.
2666 current->pi_blocked_on = PI_WAKEUP_INPROGRESS;
2667 raw_spin_unlock_irq(¤t->pi_lock);
2669 spin_lock(&hb->lock);
2672 * Clean up pi_blocked_on. We might leak it otherwise
2673 * when we succeeded with the hb->lock in the fast
2676 raw_spin_lock_irq(¤t->pi_lock);
2677 current->pi_blocked_on = NULL;
2678 raw_spin_unlock_irq(¤t->pi_lock);
2680 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2681 spin_unlock(&hb->lock);
2687 * In order to be here, we have either been requeued, are in
2688 * the process of being requeued, or requeue successfully
2689 * acquired uaddr2 on our behalf. If pi_blocked_on was
2690 * non-null above, we may be racing with a requeue. Do not
2691 * rely on q->lock_ptr to be hb2->lock until after blocking on
2692 * hb->lock or hb2->lock. The futex_requeue dropped our key1
2693 * reference and incremented our key2 reference count.
2695 hb2 = hash_futex(&key2);
2697 /* Check if the requeue code acquired the second futex for us. */
2700 * Got the lock. We might not be the anticipated owner if we
2701 * did a lock-steal - fix up the PI-state in that case.
2703 if (q.pi_state && (q.pi_state->owner != current)) {
2704 spin_lock(&hb2->lock);
2705 BUG_ON(&hb2->lock != q.lock_ptr);
2706 ret = fixup_pi_state_owner(uaddr2, &q, current);
2707 spin_unlock(&hb2->lock);
2711 * We have been woken up by futex_unlock_pi(), a timeout, or a
2712 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2715 WARN_ON(!q.pi_state);
2716 pi_mutex = &q.pi_state->pi_mutex;
2717 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2718 debug_rt_mutex_free_waiter(&rt_waiter);
2720 spin_lock(&hb2->lock);
2721 BUG_ON(&hb2->lock != q.lock_ptr);
2723 * Fixup the pi_state owner and possibly acquire the lock if we
2726 res = fixup_owner(uaddr2, &q, !ret);
2728 * If fixup_owner() returned an error, proprogate that. If it
2729 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2732 ret = (res < 0) ? res : 0;
2734 /* Unqueue and drop the lock. */
2739 * If fixup_pi_state_owner() faulted and was unable to handle the
2740 * fault, unlock the rt_mutex and return the fault to userspace.
2742 if (ret == -EFAULT) {
2743 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2744 rt_mutex_unlock(pi_mutex);
2745 } else if (ret == -EINTR) {
2747 * We've already been requeued, but cannot restart by calling
2748 * futex_lock_pi() directly. We could restart this syscall, but
2749 * it would detect that the user space "val" changed and return
2750 * -EWOULDBLOCK. Save the overhead of the restart and return
2751 * -EWOULDBLOCK directly.
2757 put_futex_key(&q.key);
2759 put_futex_key(&key2);
2763 hrtimer_cancel(&to->timer);
2764 destroy_hrtimer_on_stack(&to->timer);
2770 * Support for robust futexes: the kernel cleans up held futexes at
2773 * Implementation: user-space maintains a per-thread list of locks it
2774 * is holding. Upon do_exit(), the kernel carefully walks this list,
2775 * and marks all locks that are owned by this thread with the
2776 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2777 * always manipulated with the lock held, so the list is private and
2778 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2779 * field, to allow the kernel to clean up if the thread dies after
2780 * acquiring the lock, but just before it could have added itself to
2781 * the list. There can only be one such pending lock.
2785 * sys_set_robust_list() - Set the robust-futex list head of a task
2786 * @head: pointer to the list-head
2787 * @len: length of the list-head, as userspace expects
2789 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2792 if (!futex_cmpxchg_enabled)
2795 * The kernel knows only one size for now:
2797 if (unlikely(len != sizeof(*head)))
2800 current->robust_list = head;
2806 * sys_get_robust_list() - Get the robust-futex list head of a task
2807 * @pid: pid of the process [zero for current task]
2808 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2809 * @len_ptr: pointer to a length field, the kernel fills in the header size
2811 SYSCALL_DEFINE3(get_robust_list, int, pid,
2812 struct robust_list_head __user * __user *, head_ptr,
2813 size_t __user *, len_ptr)
2815 struct robust_list_head __user *head;
2817 struct task_struct *p;
2819 if (!futex_cmpxchg_enabled)
2828 p = find_task_by_vpid(pid);
2834 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2837 head = p->robust_list;
2840 if (put_user(sizeof(*head), len_ptr))
2842 return put_user(head, head_ptr);
2851 * Process a futex-list entry, check whether it's owned by the
2852 * dying task, and do notification if so:
2854 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2856 u32 uval, uninitialized_var(nval), mval;
2859 if (get_user(uval, uaddr))
2862 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2864 * Ok, this dying thread is truly holding a futex
2865 * of interest. Set the OWNER_DIED bit atomically
2866 * via cmpxchg, and if the value had FUTEX_WAITERS
2867 * set, wake up a waiter (if any). (We have to do a
2868 * futex_wake() even if OWNER_DIED is already set -
2869 * to handle the rare but possible case of recursive
2870 * thread-death.) The rest of the cleanup is done in
2873 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2875 * We are not holding a lock here, but we want to have
2876 * the pagefault_disable/enable() protection because
2877 * we want to handle the fault gracefully. If the
2878 * access fails we try to fault in the futex with R/W
2879 * verification via get_user_pages. get_user() above
2880 * does not guarantee R/W access. If that fails we
2881 * give up and leave the futex locked.
2883 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2884 if (fault_in_user_writeable(uaddr))
2892 * Wake robust non-PI futexes here. The wakeup of
2893 * PI futexes happens in exit_pi_state():
2895 if (!pi && (uval & FUTEX_WAITERS))
2896 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2902 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2904 static inline int fetch_robust_entry(struct robust_list __user **entry,
2905 struct robust_list __user * __user *head,
2908 unsigned long uentry;
2910 if (get_user(uentry, (unsigned long __user *)head))
2913 *entry = (void __user *)(uentry & ~1UL);
2920 * Walk curr->robust_list (very carefully, it's a userspace list!)
2921 * and mark any locks found there dead, and notify any waiters.
2923 * We silently return on any sign of list-walking problem.
2925 void exit_robust_list(struct task_struct *curr)
2927 struct robust_list_head __user *head = curr->robust_list;
2928 struct robust_list __user *entry, *next_entry, *pending;
2929 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2930 unsigned int uninitialized_var(next_pi);
2931 unsigned long futex_offset;
2934 if (!futex_cmpxchg_enabled)
2938 * Fetch the list head (which was registered earlier, via
2939 * sys_set_robust_list()):
2941 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2944 * Fetch the relative futex offset:
2946 if (get_user(futex_offset, &head->futex_offset))
2949 * Fetch any possibly pending lock-add first, and handle it
2952 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2955 next_entry = NULL; /* avoid warning with gcc */
2956 while (entry != &head->list) {
2958 * Fetch the next entry in the list before calling
2959 * handle_futex_death:
2961 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2963 * A pending lock might already be on the list, so
2964 * don't process it twice:
2966 if (entry != pending)
2967 if (handle_futex_death((void __user *)entry + futex_offset,
2975 * Avoid excessively long or circular lists:
2984 handle_futex_death((void __user *)pending + futex_offset,
2988 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2989 u32 __user *uaddr2, u32 val2, u32 val3)
2991 int cmd = op & FUTEX_CMD_MASK;
2992 unsigned int flags = 0;
2994 if (!(op & FUTEX_PRIVATE_FLAG))
2995 flags |= FLAGS_SHARED;
2997 if (op & FUTEX_CLOCK_REALTIME) {
2998 flags |= FLAGS_CLOCKRT;
2999 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3005 case FUTEX_UNLOCK_PI:
3006 case FUTEX_TRYLOCK_PI:
3007 case FUTEX_WAIT_REQUEUE_PI:
3008 case FUTEX_CMP_REQUEUE_PI:
3009 if (!futex_cmpxchg_enabled)
3015 val3 = FUTEX_BITSET_MATCH_ANY;
3016 case FUTEX_WAIT_BITSET:
3017 return futex_wait(uaddr, flags, val, timeout, val3);
3019 val3 = FUTEX_BITSET_MATCH_ANY;
3020 case FUTEX_WAKE_BITSET:
3021 return futex_wake(uaddr, flags, val, val3);
3023 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3024 case FUTEX_CMP_REQUEUE:
3025 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3027 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3029 return futex_lock_pi(uaddr, flags, timeout, 0);
3030 case FUTEX_UNLOCK_PI:
3031 return futex_unlock_pi(uaddr, flags);
3032 case FUTEX_TRYLOCK_PI:
3033 return futex_lock_pi(uaddr, flags, NULL, 1);
3034 case FUTEX_WAIT_REQUEUE_PI:
3035 val3 = FUTEX_BITSET_MATCH_ANY;
3036 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3038 case FUTEX_CMP_REQUEUE_PI:
3039 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3045 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3046 struct timespec __user *, utime, u32 __user *, uaddr2,
3050 ktime_t t, *tp = NULL;
3052 int cmd = op & FUTEX_CMD_MASK;
3054 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3055 cmd == FUTEX_WAIT_BITSET ||
3056 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3057 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3059 if (!timespec_valid(&ts))
3062 t = timespec_to_ktime(ts);
3063 if (cmd == FUTEX_WAIT)
3064 t = ktime_add_safe(ktime_get(), t);
3068 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3069 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3071 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3072 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3073 val2 = (u32) (unsigned long) utime;
3075 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3078 static void __init futex_detect_cmpxchg(void)
3080 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3084 * This will fail and we want it. Some arch implementations do
3085 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3086 * functionality. We want to know that before we call in any
3087 * of the complex code paths. Also we want to prevent
3088 * registration of robust lists in that case. NULL is
3089 * guaranteed to fault and we get -EFAULT on functional
3090 * implementation, the non-functional ones will return
3093 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3094 futex_cmpxchg_enabled = 1;
3098 static int __init futex_init(void)
3100 unsigned int futex_shift;
3103 #if CONFIG_BASE_SMALL
3104 futex_hashsize = 16;
3106 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3109 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3111 futex_hashsize < 256 ? HASH_SMALL : 0,
3113 futex_hashsize, futex_hashsize);
3114 futex_hashsize = 1UL << futex_shift;
3116 futex_detect_cmpxchg();
3118 for (i = 0; i < futex_hashsize; i++) {
3119 atomic_set(&futex_queues[i].waiters, 0);
3120 plist_head_init(&futex_queues[i].chain);
3121 spin_lock_init(&futex_queues[i].lock);
3126 __initcall(futex_init);