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 * Basic futex operation and ordering guarantees:
75 * The waiter reads the futex value in user space and calls
76 * futex_wait(). This function computes the hash bucket and acquires
77 * the hash bucket lock. After that it reads the futex user space value
78 * again and verifies that the data has not changed. If it has not changed
79 * it enqueues itself into the hash bucket, releases the hash bucket lock
82 * The waker side modifies the user space value of the futex and calls
83 * futex_wake(). This function computes the hash bucket and acquires the
84 * hash bucket lock. Then it looks for waiters on that futex in the hash
85 * bucket and wakes them.
87 * In futex wake up scenarios where no tasks are blocked on a futex, taking
88 * the hb spinlock can be avoided and simply return. In order for this
89 * optimization to work, ordering guarantees must exist so that the waiter
90 * being added to the list is acknowledged when the list is concurrently being
91 * checked by the waker, avoiding scenarios like the following:
95 * sys_futex(WAIT, futex, val);
96 * futex_wait(futex, val);
99 * sys_futex(WAKE, futex);
104 * lock(hash_bucket(futex));
106 * unlock(hash_bucket(futex));
109 * This would cause the waiter on CPU 0 to wait forever because it
110 * missed the transition of the user space value from val to newval
111 * and the waker did not find the waiter in the hash bucket queue.
113 * The correct serialization ensures that a waiter either observes
114 * the changed user space value before blocking or is woken by a
119 * sys_futex(WAIT, futex, val);
120 * futex_wait(futex, val);
123 * mb(); (A) <-- paired with -.
125 * lock(hash_bucket(futex)); |
129 * | sys_futex(WAKE, futex);
130 * | futex_wake(futex);
132 * `-------> mb(); (B)
135 * unlock(hash_bucket(futex));
136 * schedule(); if (waiters)
137 * lock(hash_bucket(futex));
138 * wake_waiters(futex);
139 * unlock(hash_bucket(futex));
141 * Where (A) orders the waiters increment and the futex value read -- this
142 * is guaranteed by the head counter in the hb spinlock; and where (B)
143 * orders the write to futex and the waiters read -- this is done by the
144 * barriers in get_futex_key_refs(), through either ihold or atomic_inc,
145 * depending on the futex type.
147 * This yields the following case (where X:=waiters, Y:=futex):
155 * Which guarantees that x==0 && y==0 is impossible; which translates back into
156 * the guarantee that we cannot both miss the futex variable change and the
160 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
161 int __read_mostly futex_cmpxchg_enabled;
165 * Futex flags used to encode options to functions and preserve them across
168 #define FLAGS_SHARED 0x01
169 #define FLAGS_CLOCKRT 0x02
170 #define FLAGS_HAS_TIMEOUT 0x04
173 * Priority Inheritance state:
175 struct futex_pi_state {
177 * list of 'owned' pi_state instances - these have to be
178 * cleaned up in do_exit() if the task exits prematurely:
180 struct list_head list;
185 struct rt_mutex pi_mutex;
187 struct task_struct *owner;
194 * struct futex_q - The hashed futex queue entry, one per waiting task
195 * @list: priority-sorted list of tasks waiting on this futex
196 * @task: the task waiting on the futex
197 * @lock_ptr: the hash bucket lock
198 * @key: the key the futex is hashed on
199 * @pi_state: optional priority inheritance state
200 * @rt_waiter: rt_waiter storage for use with requeue_pi
201 * @requeue_pi_key: the requeue_pi target futex key
202 * @bitset: bitset for the optional bitmasked wakeup
204 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
205 * we can wake only the relevant ones (hashed queues may be shared).
207 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
208 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
209 * The order of wakeup is always to make the first condition true, then
212 * PI futexes are typically woken before they are removed from the hash list via
213 * the rt_mutex code. See unqueue_me_pi().
216 struct plist_node list;
218 struct task_struct *task;
219 spinlock_t *lock_ptr;
221 struct futex_pi_state *pi_state;
222 struct rt_mutex_waiter *rt_waiter;
223 union futex_key *requeue_pi_key;
227 static const struct futex_q futex_q_init = {
228 /* list gets initialized in queue_me()*/
229 .key = FUTEX_KEY_INIT,
230 .bitset = FUTEX_BITSET_MATCH_ANY
234 * Hash buckets are shared by all the futex_keys that hash to the same
235 * location. Each key may have multiple futex_q structures, one for each task
236 * waiting on a futex.
238 struct futex_hash_bucket {
241 struct plist_head chain;
242 } ____cacheline_aligned_in_smp;
244 static unsigned long __read_mostly futex_hashsize;
246 static struct futex_hash_bucket *futex_queues;
248 static inline void futex_get_mm(union futex_key *key)
250 atomic_inc(&key->private.mm->mm_count);
252 * Ensure futex_get_mm() implies a full barrier such that
253 * get_futex_key() implies a full barrier. This is relied upon
254 * as full barrier (B), see the ordering comment above.
256 smp_mb__after_atomic_inc();
260 * Reflects a new waiter being added to the waitqueue.
262 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
265 atomic_inc(&hb->waiters);
267 * Full barrier (A), see the ordering comment above.
269 smp_mb__after_atomic_inc();
274 * Reflects a waiter being removed from the waitqueue by wakeup
277 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
280 atomic_dec(&hb->waiters);
284 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
287 return atomic_read(&hb->waiters);
294 * We hash on the keys returned from get_futex_key (see below).
296 static struct futex_hash_bucket *hash_futex(union futex_key *key)
298 u32 hash = jhash2((u32*)&key->both.word,
299 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
301 return &futex_queues[hash & (futex_hashsize - 1)];
305 * Return 1 if two futex_keys are equal, 0 otherwise.
307 static inline int match_futex(union futex_key *key1, union futex_key *key2)
310 && key1->both.word == key2->both.word
311 && key1->both.ptr == key2->both.ptr
312 && key1->both.offset == key2->both.offset);
316 * Take a reference to the resource addressed by a key.
317 * Can be called while holding spinlocks.
320 static void get_futex_key_refs(union futex_key *key)
325 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
327 ihold(key->shared.inode); /* implies MB (B) */
329 case FUT_OFF_MMSHARED:
330 futex_get_mm(key); /* implies MB (B) */
336 * Drop a reference to the resource addressed by a key.
337 * The hash bucket spinlock must not be held.
339 static void drop_futex_key_refs(union futex_key *key)
341 if (!key->both.ptr) {
342 /* If we're here then we tried to put a key we failed to get */
347 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
349 iput(key->shared.inode);
351 case FUT_OFF_MMSHARED:
352 mmdrop(key->private.mm);
358 * get_futex_key() - Get parameters which are the keys for a futex
359 * @uaddr: virtual address of the futex
360 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
361 * @key: address where result is stored.
362 * @rw: mapping needs to be read/write (values: VERIFY_READ,
365 * Return: a negative error code or 0
367 * The key words are stored in *key on success.
369 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
370 * offset_within_page). For private mappings, it's (uaddr, current->mm).
371 * We can usually work out the index without swapping in the page.
373 * lock_page() might sleep, the caller should not hold a spinlock.
376 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
378 unsigned long address = (unsigned long)uaddr;
379 struct mm_struct *mm = current->mm;
380 struct page *page, *page_head;
384 * The futex address must be "naturally" aligned.
386 key->both.offset = address % PAGE_SIZE;
387 if (unlikely((address % sizeof(u32)) != 0))
389 address -= key->both.offset;
391 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
395 * PROCESS_PRIVATE futexes are fast.
396 * As the mm cannot disappear under us and the 'key' only needs
397 * virtual address, we dont even have to find the underlying vma.
398 * Note : We do have to check 'uaddr' is a valid user address,
399 * but access_ok() should be faster than find_vma()
402 key->private.mm = mm;
403 key->private.address = address;
404 get_futex_key_refs(key); /* implies MB (B) */
409 err = get_user_pages_fast(address, 1, 1, &page);
411 * If write access is not required (eg. FUTEX_WAIT), try
412 * and get read-only access.
414 if (err == -EFAULT && rw == VERIFY_READ) {
415 err = get_user_pages_fast(address, 1, 0, &page);
423 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
425 if (unlikely(PageTail(page))) {
427 /* serialize against __split_huge_page_splitting() */
429 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
430 page_head = compound_head(page);
432 * page_head is valid pointer but we must pin
433 * it before taking the PG_lock and/or
434 * PG_compound_lock. The moment we re-enable
435 * irqs __split_huge_page_splitting() can
436 * return and the head page can be freed from
437 * under us. We can't take the PG_lock and/or
438 * PG_compound_lock on a page that could be
439 * freed from under us.
441 if (page != page_head) {
452 page_head = compound_head(page);
453 if (page != page_head) {
459 lock_page(page_head);
462 * If page_head->mapping is NULL, then it cannot be a PageAnon
463 * page; but it might be the ZERO_PAGE or in the gate area or
464 * in a special mapping (all cases which we are happy to fail);
465 * or it may have been a good file page when get_user_pages_fast
466 * found it, but truncated or holepunched or subjected to
467 * invalidate_complete_page2 before we got the page lock (also
468 * cases which we are happy to fail). And we hold a reference,
469 * so refcount care in invalidate_complete_page's remove_mapping
470 * prevents drop_caches from setting mapping to NULL beneath us.
472 * The case we do have to guard against is when memory pressure made
473 * shmem_writepage move it from filecache to swapcache beneath us:
474 * an unlikely race, but we do need to retry for page_head->mapping.
476 if (!page_head->mapping) {
477 int shmem_swizzled = PageSwapCache(page_head);
478 unlock_page(page_head);
486 * Private mappings are handled in a simple way.
488 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
489 * it's a read-only handle, it's expected that futexes attach to
490 * the object not the particular process.
492 if (PageAnon(page_head)) {
494 * A RO anonymous page will never change and thus doesn't make
495 * sense for futex operations.
502 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
503 key->private.mm = mm;
504 key->private.address = address;
506 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
507 key->shared.inode = page_head->mapping->host;
508 key->shared.pgoff = basepage_index(page);
511 get_futex_key_refs(key); /* implies MB (B) */
514 unlock_page(page_head);
519 static inline void put_futex_key(union futex_key *key)
521 drop_futex_key_refs(key);
525 * fault_in_user_writeable() - Fault in user address and verify RW access
526 * @uaddr: pointer to faulting user space address
528 * Slow path to fixup the fault we just took in the atomic write
531 * We have no generic implementation of a non-destructive write to the
532 * user address. We know that we faulted in the atomic pagefault
533 * disabled section so we can as well avoid the #PF overhead by
534 * calling get_user_pages() right away.
536 static int fault_in_user_writeable(u32 __user *uaddr)
538 struct mm_struct *mm = current->mm;
541 down_read(&mm->mmap_sem);
542 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
544 up_read(&mm->mmap_sem);
546 return ret < 0 ? ret : 0;
550 * futex_top_waiter() - Return the highest priority waiter on a futex
551 * @hb: the hash bucket the futex_q's reside in
552 * @key: the futex key (to distinguish it from other futex futex_q's)
554 * Must be called with the hb lock held.
556 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
557 union futex_key *key)
559 struct futex_q *this;
561 plist_for_each_entry(this, &hb->chain, list) {
562 if (match_futex(&this->key, key))
568 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
569 u32 uval, u32 newval)
574 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
580 static int get_futex_value_locked(u32 *dest, u32 __user *from)
585 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
588 return ret ? -EFAULT : 0;
595 static int refill_pi_state_cache(void)
597 struct futex_pi_state *pi_state;
599 if (likely(current->pi_state_cache))
602 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
607 INIT_LIST_HEAD(&pi_state->list);
608 /* pi_mutex gets initialized later */
609 pi_state->owner = NULL;
610 atomic_set(&pi_state->refcount, 1);
611 pi_state->key = FUTEX_KEY_INIT;
613 current->pi_state_cache = pi_state;
618 static struct futex_pi_state * alloc_pi_state(void)
620 struct futex_pi_state *pi_state = current->pi_state_cache;
623 current->pi_state_cache = NULL;
628 static void free_pi_state(struct futex_pi_state *pi_state)
630 if (!atomic_dec_and_test(&pi_state->refcount))
634 * If pi_state->owner is NULL, the owner is most probably dying
635 * and has cleaned up the pi_state already
637 if (pi_state->owner) {
638 raw_spin_lock_irq(&pi_state->owner->pi_lock);
639 list_del_init(&pi_state->list);
640 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
642 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
645 if (current->pi_state_cache)
649 * pi_state->list is already empty.
650 * clear pi_state->owner.
651 * refcount is at 0 - put it back to 1.
653 pi_state->owner = NULL;
654 atomic_set(&pi_state->refcount, 1);
655 current->pi_state_cache = pi_state;
660 * Look up the task based on what TID userspace gave us.
663 static struct task_struct * futex_find_get_task(pid_t pid)
665 struct task_struct *p;
668 p = find_task_by_vpid(pid);
678 * This task is holding PI mutexes at exit time => bad.
679 * Kernel cleans up PI-state, but userspace is likely hosed.
680 * (Robust-futex cleanup is separate and might save the day for userspace.)
682 void exit_pi_state_list(struct task_struct *curr)
684 struct list_head *next, *head = &curr->pi_state_list;
685 struct futex_pi_state *pi_state;
686 struct futex_hash_bucket *hb;
687 union futex_key key = FUTEX_KEY_INIT;
689 if (!futex_cmpxchg_enabled)
692 * We are a ZOMBIE and nobody can enqueue itself on
693 * pi_state_list anymore, but we have to be careful
694 * versus waiters unqueueing themselves:
696 raw_spin_lock_irq(&curr->pi_lock);
697 while (!list_empty(head)) {
700 pi_state = list_entry(next, struct futex_pi_state, list);
702 hb = hash_futex(&key);
703 raw_spin_unlock_irq(&curr->pi_lock);
705 spin_lock(&hb->lock);
707 raw_spin_lock_irq(&curr->pi_lock);
709 * We dropped the pi-lock, so re-check whether this
710 * task still owns the PI-state:
712 if (head->next != next) {
713 spin_unlock(&hb->lock);
717 WARN_ON(pi_state->owner != curr);
718 WARN_ON(list_empty(&pi_state->list));
719 list_del_init(&pi_state->list);
720 pi_state->owner = NULL;
721 raw_spin_unlock_irq(&curr->pi_lock);
723 rt_mutex_unlock(&pi_state->pi_mutex);
725 spin_unlock(&hb->lock);
727 raw_spin_lock_irq(&curr->pi_lock);
729 raw_spin_unlock_irq(&curr->pi_lock);
733 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
734 union futex_key *key, struct futex_pi_state **ps,
735 struct task_struct *task)
737 struct futex_pi_state *pi_state = NULL;
738 struct futex_q *this, *next;
739 struct task_struct *p;
740 pid_t pid = uval & FUTEX_TID_MASK;
742 plist_for_each_entry_safe(this, next, &hb->chain, list) {
743 if (match_futex(&this->key, key)) {
745 * Another waiter already exists - bump up
746 * the refcount and return its pi_state:
748 pi_state = this->pi_state;
750 * Userspace might have messed up non-PI and PI futexes
752 if (unlikely(!pi_state))
755 WARN_ON(!atomic_read(&pi_state->refcount));
758 * When pi_state->owner is NULL then the owner died
759 * and another waiter is on the fly. pi_state->owner
760 * is fixed up by the task which acquires
761 * pi_state->rt_mutex.
763 * We do not check for pid == 0 which can happen when
764 * the owner died and robust_list_exit() cleared the
767 if (pid && pi_state->owner) {
769 * Bail out if user space manipulated the
772 if (pid != task_pid_vnr(pi_state->owner))
777 * Protect against a corrupted uval. If uval
778 * is 0x80000000 then pid is 0 and the waiter
779 * bit is set. So the deadlock check in the
780 * calling code has failed and we did not fall
781 * into the check above due to !pid.
783 if (task && pi_state->owner == task)
786 atomic_inc(&pi_state->refcount);
794 * We are the first waiter - try to look up the real owner and attach
795 * the new pi_state to it, but bail out when TID = 0
799 p = futex_find_get_task(pid);
809 * We need to look at the task state flags to figure out,
810 * whether the task is exiting. To protect against the do_exit
811 * change of the task flags, we do this protected by
814 raw_spin_lock_irq(&p->pi_lock);
815 if (unlikely(p->flags & PF_EXITING)) {
817 * The task is on the way out. When PF_EXITPIDONE is
818 * set, we know that the task has finished the
821 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
823 raw_spin_unlock_irq(&p->pi_lock);
828 pi_state = alloc_pi_state();
831 * Initialize the pi_mutex in locked state and make 'p'
834 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
836 /* Store the key for possible exit cleanups: */
837 pi_state->key = *key;
839 WARN_ON(!list_empty(&pi_state->list));
840 list_add(&pi_state->list, &p->pi_state_list);
842 raw_spin_unlock_irq(&p->pi_lock);
852 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
853 * @uaddr: the pi futex user address
854 * @hb: the pi futex hash bucket
855 * @key: the futex key associated with uaddr and hb
856 * @ps: the pi_state pointer where we store the result of the
858 * @task: the task to perform the atomic lock work for. This will
859 * be "current" except in the case of requeue pi.
860 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
864 * 1 - acquired the lock;
867 * The hb->lock and futex_key refs shall be held by the caller.
869 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
870 union futex_key *key,
871 struct futex_pi_state **ps,
872 struct task_struct *task, int set_waiters)
874 int lock_taken, ret, force_take = 0;
875 u32 uval, newval, curval, vpid = task_pid_vnr(task);
878 ret = lock_taken = 0;
881 * To avoid races, we attempt to take the lock here again
882 * (by doing a 0 -> TID atomic cmpxchg), while holding all
883 * the locks. It will most likely not succeed.
887 newval |= FUTEX_WAITERS;
889 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
895 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
899 * Surprise - we got the lock. Just return to userspace:
901 if (unlikely(!curval))
907 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
908 * to wake at the next unlock.
910 newval = curval | FUTEX_WAITERS;
913 * Should we force take the futex? See below.
915 if (unlikely(force_take)) {
917 * Keep the OWNER_DIED and the WAITERS bit and set the
920 newval = (curval & ~FUTEX_TID_MASK) | vpid;
925 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
927 if (unlikely(curval != uval))
931 * We took the lock due to forced take over.
933 if (unlikely(lock_taken))
937 * We dont have the lock. Look up the PI state (or create it if
938 * we are the first waiter):
940 ret = lookup_pi_state(uval, hb, key, ps, task);
946 * We failed to find an owner for this
947 * futex. So we have no pi_state to block
948 * on. This can happen in two cases:
951 * 2) A stale FUTEX_WAITERS bit
953 * Re-read the futex value.
955 if (get_futex_value_locked(&curval, uaddr))
959 * If the owner died or we have a stale
960 * WAITERS bit the owner TID in the user space
963 if (!(curval & FUTEX_TID_MASK)) {
976 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
977 * @q: The futex_q to unqueue
979 * The q->lock_ptr must not be NULL and must be held by the caller.
981 static void __unqueue_futex(struct futex_q *q)
983 struct futex_hash_bucket *hb;
985 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
986 || WARN_ON(plist_node_empty(&q->list)))
989 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
990 plist_del(&q->list, &hb->chain);
995 * The hash bucket lock must be held when this is called.
996 * Afterwards, the futex_q must not be accessed.
998 static void wake_futex(struct futex_q *q)
1000 struct task_struct *p = q->task;
1002 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1006 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1007 * a non-futex wake up happens on another CPU then the task
1008 * might exit and p would dereference a non-existing task
1009 * struct. Prevent this by holding a reference on p across the
1016 * The waiting task can free the futex_q as soon as
1017 * q->lock_ptr = NULL is written, without taking any locks. A
1018 * memory barrier is required here to prevent the following
1019 * store to lock_ptr from getting ahead of the plist_del.
1024 wake_up_state(p, TASK_NORMAL);
1028 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1030 struct task_struct *new_owner;
1031 struct futex_pi_state *pi_state = this->pi_state;
1032 u32 uninitialized_var(curval), newval;
1038 * If current does not own the pi_state then the futex is
1039 * inconsistent and user space fiddled with the futex value.
1041 if (pi_state->owner != current)
1044 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1045 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1048 * It is possible that the next waiter (the one that brought
1049 * this owner to the kernel) timed out and is no longer
1050 * waiting on the lock.
1053 new_owner = this->task;
1056 * We pass it to the next owner. (The WAITERS bit is always
1057 * kept enabled while there is PI state around. We must also
1058 * preserve the owner died bit.)
1060 if (!(uval & FUTEX_OWNER_DIED)) {
1063 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1065 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1067 else if (curval != uval)
1070 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1075 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1076 WARN_ON(list_empty(&pi_state->list));
1077 list_del_init(&pi_state->list);
1078 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1080 raw_spin_lock_irq(&new_owner->pi_lock);
1081 WARN_ON(!list_empty(&pi_state->list));
1082 list_add(&pi_state->list, &new_owner->pi_state_list);
1083 pi_state->owner = new_owner;
1084 raw_spin_unlock_irq(&new_owner->pi_lock);
1086 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1087 rt_mutex_unlock(&pi_state->pi_mutex);
1092 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1094 u32 uninitialized_var(oldval);
1097 * There is no waiter, so we unlock the futex. The owner died
1098 * bit has not to be preserved here. We are the owner:
1100 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1109 * Express the locking dependencies for lockdep:
1112 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1115 spin_lock(&hb1->lock);
1117 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1118 } else { /* hb1 > hb2 */
1119 spin_lock(&hb2->lock);
1120 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1125 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1127 spin_unlock(&hb1->lock);
1129 spin_unlock(&hb2->lock);
1133 * Wake up waiters matching bitset queued on this futex (uaddr).
1136 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1138 struct futex_hash_bucket *hb;
1139 struct futex_q *this, *next;
1140 union futex_key key = FUTEX_KEY_INIT;
1146 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1147 if (unlikely(ret != 0))
1150 hb = hash_futex(&key);
1152 /* Make sure we really have tasks to wakeup */
1153 if (!hb_waiters_pending(hb))
1156 spin_lock(&hb->lock);
1158 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1159 if (match_futex (&this->key, &key)) {
1160 if (this->pi_state || this->rt_waiter) {
1165 /* Check if one of the bits is set in both bitsets */
1166 if (!(this->bitset & bitset))
1170 if (++ret >= nr_wake)
1175 spin_unlock(&hb->lock);
1177 put_futex_key(&key);
1183 * Wake up all waiters hashed on the physical page that is mapped
1184 * to this virtual address:
1187 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1188 int nr_wake, int nr_wake2, int op)
1190 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1191 struct futex_hash_bucket *hb1, *hb2;
1192 struct futex_q *this, *next;
1196 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1197 if (unlikely(ret != 0))
1199 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1200 if (unlikely(ret != 0))
1203 hb1 = hash_futex(&key1);
1204 hb2 = hash_futex(&key2);
1207 double_lock_hb(hb1, hb2);
1208 op_ret = futex_atomic_op_inuser(op, uaddr2);
1209 if (unlikely(op_ret < 0)) {
1211 double_unlock_hb(hb1, hb2);
1215 * we don't get EFAULT from MMU faults if we don't have an MMU,
1216 * but we might get them from range checking
1222 if (unlikely(op_ret != -EFAULT)) {
1227 ret = fault_in_user_writeable(uaddr2);
1231 if (!(flags & FLAGS_SHARED))
1234 put_futex_key(&key2);
1235 put_futex_key(&key1);
1239 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1240 if (match_futex (&this->key, &key1)) {
1241 if (this->pi_state || this->rt_waiter) {
1246 if (++ret >= nr_wake)
1253 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1254 if (match_futex (&this->key, &key2)) {
1255 if (this->pi_state || this->rt_waiter) {
1260 if (++op_ret >= nr_wake2)
1268 double_unlock_hb(hb1, hb2);
1270 put_futex_key(&key2);
1272 put_futex_key(&key1);
1278 * requeue_futex() - Requeue a futex_q from one hb to another
1279 * @q: the futex_q to requeue
1280 * @hb1: the source hash_bucket
1281 * @hb2: the target hash_bucket
1282 * @key2: the new key for the requeued futex_q
1285 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1286 struct futex_hash_bucket *hb2, union futex_key *key2)
1290 * If key1 and key2 hash to the same bucket, no need to
1293 if (likely(&hb1->chain != &hb2->chain)) {
1294 plist_del(&q->list, &hb1->chain);
1295 hb_waiters_dec(hb1);
1296 plist_add(&q->list, &hb2->chain);
1297 hb_waiters_inc(hb2);
1298 q->lock_ptr = &hb2->lock;
1300 get_futex_key_refs(key2);
1305 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1307 * @key: the key of the requeue target futex
1308 * @hb: the hash_bucket of the requeue target futex
1310 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1311 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1312 * to the requeue target futex so the waiter can detect the wakeup on the right
1313 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1314 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1315 * to protect access to the pi_state to fixup the owner later. Must be called
1316 * with both q->lock_ptr and hb->lock held.
1319 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1320 struct futex_hash_bucket *hb)
1322 get_futex_key_refs(key);
1327 WARN_ON(!q->rt_waiter);
1328 q->rt_waiter = NULL;
1330 q->lock_ptr = &hb->lock;
1332 wake_up_state(q->task, TASK_NORMAL);
1336 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1337 * @pifutex: the user address of the to futex
1338 * @hb1: the from futex hash bucket, must be locked by the caller
1339 * @hb2: the to futex hash bucket, must be locked by the caller
1340 * @key1: the from futex key
1341 * @key2: the to futex key
1342 * @ps: address to store the pi_state pointer
1343 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1345 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1346 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1347 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1348 * hb1 and hb2 must be held by the caller.
1351 * 0 - failed to acquire the lock atomically;
1352 * >0 - acquired the lock, return value is vpid of the top_waiter
1355 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1356 struct futex_hash_bucket *hb1,
1357 struct futex_hash_bucket *hb2,
1358 union futex_key *key1, union futex_key *key2,
1359 struct futex_pi_state **ps, int set_waiters)
1361 struct futex_q *top_waiter = NULL;
1365 if (get_futex_value_locked(&curval, pifutex))
1369 * Find the top_waiter and determine if there are additional waiters.
1370 * If the caller intends to requeue more than 1 waiter to pifutex,
1371 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1372 * as we have means to handle the possible fault. If not, don't set
1373 * the bit unecessarily as it will force the subsequent unlock to enter
1376 top_waiter = futex_top_waiter(hb1, key1);
1378 /* There are no waiters, nothing for us to do. */
1382 /* Ensure we requeue to the expected futex. */
1383 if (!match_futex(top_waiter->requeue_pi_key, key2))
1387 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1388 * the contended case or if set_waiters is 1. The pi_state is returned
1389 * in ps in contended cases.
1391 vpid = task_pid_vnr(top_waiter->task);
1392 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1395 requeue_pi_wake_futex(top_waiter, key2, hb2);
1402 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1403 * @uaddr1: source futex user address
1404 * @flags: futex flags (FLAGS_SHARED, etc.)
1405 * @uaddr2: target futex user address
1406 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1407 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1408 * @cmpval: @uaddr1 expected value (or %NULL)
1409 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1410 * pi futex (pi to pi requeue is not supported)
1412 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1413 * uaddr2 atomically on behalf of the top waiter.
1416 * >=0 - on success, the number of tasks requeued or woken;
1419 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1420 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1421 u32 *cmpval, int requeue_pi)
1423 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1424 int drop_count = 0, task_count = 0, ret;
1425 struct futex_pi_state *pi_state = NULL;
1426 struct futex_hash_bucket *hb1, *hb2;
1427 struct futex_q *this, *next;
1431 * Requeue PI only works on two distinct uaddrs. This
1432 * check is only valid for private futexes. See below.
1434 if (uaddr1 == uaddr2)
1438 * requeue_pi requires a pi_state, try to allocate it now
1439 * without any locks in case it fails.
1441 if (refill_pi_state_cache())
1444 * requeue_pi must wake as many tasks as it can, up to nr_wake
1445 * + nr_requeue, since it acquires the rt_mutex prior to
1446 * returning to userspace, so as to not leave the rt_mutex with
1447 * waiters and no owner. However, second and third wake-ups
1448 * cannot be predicted as they involve race conditions with the
1449 * first wake and a fault while looking up the pi_state. Both
1450 * pthread_cond_signal() and pthread_cond_broadcast() should
1458 if (pi_state != NULL) {
1460 * We will have to lookup the pi_state again, so free this one
1461 * to keep the accounting correct.
1463 free_pi_state(pi_state);
1467 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1468 if (unlikely(ret != 0))
1470 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1471 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1472 if (unlikely(ret != 0))
1476 * The check above which compares uaddrs is not sufficient for
1477 * shared futexes. We need to compare the keys:
1479 if (requeue_pi && match_futex(&key1, &key2)) {
1484 hb1 = hash_futex(&key1);
1485 hb2 = hash_futex(&key2);
1488 hb_waiters_inc(hb2);
1489 double_lock_hb(hb1, hb2);
1491 if (likely(cmpval != NULL)) {
1494 ret = get_futex_value_locked(&curval, uaddr1);
1496 if (unlikely(ret)) {
1497 double_unlock_hb(hb1, hb2);
1498 hb_waiters_dec(hb2);
1500 ret = get_user(curval, uaddr1);
1504 if (!(flags & FLAGS_SHARED))
1507 put_futex_key(&key2);
1508 put_futex_key(&key1);
1511 if (curval != *cmpval) {
1517 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1519 * Attempt to acquire uaddr2 and wake the top waiter. If we
1520 * intend to requeue waiters, force setting the FUTEX_WAITERS
1521 * bit. We force this here where we are able to easily handle
1522 * faults rather in the requeue loop below.
1524 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1525 &key2, &pi_state, nr_requeue);
1528 * At this point the top_waiter has either taken uaddr2 or is
1529 * waiting on it. If the former, then the pi_state will not
1530 * exist yet, look it up one more time to ensure we have a
1531 * reference to it. If the lock was taken, ret contains the
1532 * vpid of the top waiter task.
1539 * If we acquired the lock, then the user
1540 * space value of uaddr2 should be vpid. It
1541 * cannot be changed by the top waiter as it
1542 * is blocked on hb2 lock if it tries to do
1543 * so. If something fiddled with it behind our
1544 * back the pi state lookup might unearth
1545 * it. So we rather use the known value than
1546 * rereading and handing potential crap to
1549 ret = lookup_pi_state(ret, hb2, &key2, &pi_state, NULL);
1556 double_unlock_hb(hb1, hb2);
1557 hb_waiters_dec(hb2);
1558 put_futex_key(&key2);
1559 put_futex_key(&key1);
1560 ret = fault_in_user_writeable(uaddr2);
1565 /* The owner was exiting, try again. */
1566 double_unlock_hb(hb1, hb2);
1567 hb_waiters_dec(hb2);
1568 put_futex_key(&key2);
1569 put_futex_key(&key1);
1577 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1578 if (task_count - nr_wake >= nr_requeue)
1581 if (!match_futex(&this->key, &key1))
1585 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1586 * be paired with each other and no other futex ops.
1588 * We should never be requeueing a futex_q with a pi_state,
1589 * which is awaiting a futex_unlock_pi().
1591 if ((requeue_pi && !this->rt_waiter) ||
1592 (!requeue_pi && this->rt_waiter) ||
1599 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1600 * lock, we already woke the top_waiter. If not, it will be
1601 * woken by futex_unlock_pi().
1603 if (++task_count <= nr_wake && !requeue_pi) {
1608 /* Ensure we requeue to the expected futex for requeue_pi. */
1609 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1615 * Requeue nr_requeue waiters and possibly one more in the case
1616 * of requeue_pi if we couldn't acquire the lock atomically.
1619 /* Prepare the waiter to take the rt_mutex. */
1620 atomic_inc(&pi_state->refcount);
1621 this->pi_state = pi_state;
1622 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1626 /* We got the lock. */
1627 requeue_pi_wake_futex(this, &key2, hb2);
1632 this->pi_state = NULL;
1633 free_pi_state(pi_state);
1637 requeue_futex(this, hb1, hb2, &key2);
1642 double_unlock_hb(hb1, hb2);
1643 hb_waiters_dec(hb2);
1646 * drop_futex_key_refs() must be called outside the spinlocks. During
1647 * the requeue we moved futex_q's from the hash bucket at key1 to the
1648 * one at key2 and updated their key pointer. We no longer need to
1649 * hold the references to key1.
1651 while (--drop_count >= 0)
1652 drop_futex_key_refs(&key1);
1655 put_futex_key(&key2);
1657 put_futex_key(&key1);
1659 if (pi_state != NULL)
1660 free_pi_state(pi_state);
1661 return ret ? ret : task_count;
1664 /* The key must be already stored in q->key. */
1665 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1666 __acquires(&hb->lock)
1668 struct futex_hash_bucket *hb;
1670 hb = hash_futex(&q->key);
1673 * Increment the counter before taking the lock so that
1674 * a potential waker won't miss a to-be-slept task that is
1675 * waiting for the spinlock. This is safe as all queue_lock()
1676 * users end up calling queue_me(). Similarly, for housekeeping,
1677 * decrement the counter at queue_unlock() when some error has
1678 * occurred and we don't end up adding the task to the list.
1682 q->lock_ptr = &hb->lock;
1684 spin_lock(&hb->lock); /* implies MB (A) */
1689 queue_unlock(struct futex_hash_bucket *hb)
1690 __releases(&hb->lock)
1692 spin_unlock(&hb->lock);
1697 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1698 * @q: The futex_q to enqueue
1699 * @hb: The destination hash bucket
1701 * The hb->lock must be held by the caller, and is released here. A call to
1702 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1703 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1704 * or nothing if the unqueue is done as part of the wake process and the unqueue
1705 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1708 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1709 __releases(&hb->lock)
1714 * The priority used to register this element is
1715 * - either the real thread-priority for the real-time threads
1716 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1717 * - or MAX_RT_PRIO for non-RT threads.
1718 * Thus, all RT-threads are woken first in priority order, and
1719 * the others are woken last, in FIFO order.
1721 prio = min(current->normal_prio, MAX_RT_PRIO);
1723 plist_node_init(&q->list, prio);
1724 plist_add(&q->list, &hb->chain);
1726 spin_unlock(&hb->lock);
1730 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1731 * @q: The futex_q to unqueue
1733 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1734 * be paired with exactly one earlier call to queue_me().
1737 * 1 - if the futex_q was still queued (and we removed unqueued it);
1738 * 0 - if the futex_q was already removed by the waking thread
1740 static int unqueue_me(struct futex_q *q)
1742 spinlock_t *lock_ptr;
1745 /* In the common case we don't take the spinlock, which is nice. */
1747 lock_ptr = q->lock_ptr;
1749 if (lock_ptr != NULL) {
1750 spin_lock(lock_ptr);
1752 * q->lock_ptr can change between reading it and
1753 * spin_lock(), causing us to take the wrong lock. This
1754 * corrects the race condition.
1756 * Reasoning goes like this: if we have the wrong lock,
1757 * q->lock_ptr must have changed (maybe several times)
1758 * between reading it and the spin_lock(). It can
1759 * change again after the spin_lock() but only if it was
1760 * already changed before the spin_lock(). It cannot,
1761 * however, change back to the original value. Therefore
1762 * we can detect whether we acquired the correct lock.
1764 if (unlikely(lock_ptr != q->lock_ptr)) {
1765 spin_unlock(lock_ptr);
1770 BUG_ON(q->pi_state);
1772 spin_unlock(lock_ptr);
1776 drop_futex_key_refs(&q->key);
1781 * PI futexes can not be requeued and must remove themself from the
1782 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1785 static void unqueue_me_pi(struct futex_q *q)
1786 __releases(q->lock_ptr)
1790 BUG_ON(!q->pi_state);
1791 free_pi_state(q->pi_state);
1794 spin_unlock(q->lock_ptr);
1798 * Fixup the pi_state owner with the new owner.
1800 * Must be called with hash bucket lock held and mm->sem held for non
1803 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1804 struct task_struct *newowner)
1806 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1807 struct futex_pi_state *pi_state = q->pi_state;
1808 struct task_struct *oldowner = pi_state->owner;
1809 u32 uval, uninitialized_var(curval), newval;
1813 if (!pi_state->owner)
1814 newtid |= FUTEX_OWNER_DIED;
1817 * We are here either because we stole the rtmutex from the
1818 * previous highest priority waiter or we are the highest priority
1819 * waiter but failed to get the rtmutex the first time.
1820 * We have to replace the newowner TID in the user space variable.
1821 * This must be atomic as we have to preserve the owner died bit here.
1823 * Note: We write the user space value _before_ changing the pi_state
1824 * because we can fault here. Imagine swapped out pages or a fork
1825 * that marked all the anonymous memory readonly for cow.
1827 * Modifying pi_state _before_ the user space value would
1828 * leave the pi_state in an inconsistent state when we fault
1829 * here, because we need to drop the hash bucket lock to
1830 * handle the fault. This might be observed in the PID check
1831 * in lookup_pi_state.
1834 if (get_futex_value_locked(&uval, uaddr))
1838 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1840 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1848 * We fixed up user space. Now we need to fix the pi_state
1851 if (pi_state->owner != NULL) {
1852 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1853 WARN_ON(list_empty(&pi_state->list));
1854 list_del_init(&pi_state->list);
1855 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1858 pi_state->owner = newowner;
1860 raw_spin_lock_irq(&newowner->pi_lock);
1861 WARN_ON(!list_empty(&pi_state->list));
1862 list_add(&pi_state->list, &newowner->pi_state_list);
1863 raw_spin_unlock_irq(&newowner->pi_lock);
1867 * To handle the page fault we need to drop the hash bucket
1868 * lock here. That gives the other task (either the highest priority
1869 * waiter itself or the task which stole the rtmutex) the
1870 * chance to try the fixup of the pi_state. So once we are
1871 * back from handling the fault we need to check the pi_state
1872 * after reacquiring the hash bucket lock and before trying to
1873 * do another fixup. When the fixup has been done already we
1877 spin_unlock(q->lock_ptr);
1879 ret = fault_in_user_writeable(uaddr);
1881 spin_lock(q->lock_ptr);
1884 * Check if someone else fixed it for us:
1886 if (pi_state->owner != oldowner)
1895 static long futex_wait_restart(struct restart_block *restart);
1898 * fixup_owner() - Post lock pi_state and corner case management
1899 * @uaddr: user address of the futex
1900 * @q: futex_q (contains pi_state and access to the rt_mutex)
1901 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1903 * After attempting to lock an rt_mutex, this function is called to cleanup
1904 * the pi_state owner as well as handle race conditions that may allow us to
1905 * acquire the lock. Must be called with the hb lock held.
1908 * 1 - success, lock taken;
1909 * 0 - success, lock not taken;
1910 * <0 - on error (-EFAULT)
1912 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1914 struct task_struct *owner;
1919 * Got the lock. We might not be the anticipated owner if we
1920 * did a lock-steal - fix up the PI-state in that case:
1922 if (q->pi_state->owner != current)
1923 ret = fixup_pi_state_owner(uaddr, q, current);
1928 * Catch the rare case, where the lock was released when we were on the
1929 * way back before we locked the hash bucket.
1931 if (q->pi_state->owner == current) {
1933 * Try to get the rt_mutex now. This might fail as some other
1934 * task acquired the rt_mutex after we removed ourself from the
1935 * rt_mutex waiters list.
1937 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1943 * pi_state is incorrect, some other task did a lock steal and
1944 * we returned due to timeout or signal without taking the
1945 * rt_mutex. Too late.
1947 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1948 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1950 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1951 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1952 ret = fixup_pi_state_owner(uaddr, q, owner);
1957 * Paranoia check. If we did not take the lock, then we should not be
1958 * the owner of the rt_mutex.
1960 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1961 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1962 "pi-state %p\n", ret,
1963 q->pi_state->pi_mutex.owner,
1964 q->pi_state->owner);
1967 return ret ? ret : locked;
1971 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1972 * @hb: the futex hash bucket, must be locked by the caller
1973 * @q: the futex_q to queue up on
1974 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1976 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1977 struct hrtimer_sleeper *timeout)
1980 * The task state is guaranteed to be set before another task can
1981 * wake it. set_current_state() is implemented using set_mb() and
1982 * queue_me() calls spin_unlock() upon completion, both serializing
1983 * access to the hash list and forcing another memory barrier.
1985 set_current_state(TASK_INTERRUPTIBLE);
1990 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1991 if (!hrtimer_active(&timeout->timer))
1992 timeout->task = NULL;
1996 * If we have been removed from the hash list, then another task
1997 * has tried to wake us, and we can skip the call to schedule().
1999 if (likely(!plist_node_empty(&q->list))) {
2001 * If the timer has already expired, current will already be
2002 * flagged for rescheduling. Only call schedule if there
2003 * is no timeout, or if it has yet to expire.
2005 if (!timeout || timeout->task)
2006 freezable_schedule();
2008 __set_current_state(TASK_RUNNING);
2012 * futex_wait_setup() - Prepare to wait on a futex
2013 * @uaddr: the futex userspace address
2014 * @val: the expected value
2015 * @flags: futex flags (FLAGS_SHARED, etc.)
2016 * @q: the associated futex_q
2017 * @hb: storage for hash_bucket pointer to be returned to caller
2019 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2020 * compare it with the expected value. Handle atomic faults internally.
2021 * Return with the hb lock held and a q.key reference on success, and unlocked
2022 * with no q.key reference on failure.
2025 * 0 - uaddr contains val and hb has been locked;
2026 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2028 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2029 struct futex_q *q, struct futex_hash_bucket **hb)
2035 * Access the page AFTER the hash-bucket is locked.
2036 * Order is important:
2038 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2039 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2041 * The basic logical guarantee of a futex is that it blocks ONLY
2042 * if cond(var) is known to be true at the time of blocking, for
2043 * any cond. If we locked the hash-bucket after testing *uaddr, that
2044 * would open a race condition where we could block indefinitely with
2045 * cond(var) false, which would violate the guarantee.
2047 * On the other hand, we insert q and release the hash-bucket only
2048 * after testing *uaddr. This guarantees that futex_wait() will NOT
2049 * absorb a wakeup if *uaddr does not match the desired values
2050 * while the syscall executes.
2053 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2054 if (unlikely(ret != 0))
2058 *hb = queue_lock(q);
2060 ret = get_futex_value_locked(&uval, uaddr);
2065 ret = get_user(uval, uaddr);
2069 if (!(flags & FLAGS_SHARED))
2072 put_futex_key(&q->key);
2083 put_futex_key(&q->key);
2087 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2088 ktime_t *abs_time, u32 bitset)
2090 struct hrtimer_sleeper timeout, *to = NULL;
2091 struct restart_block *restart;
2092 struct futex_hash_bucket *hb;
2093 struct futex_q q = futex_q_init;
2103 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2104 CLOCK_REALTIME : CLOCK_MONOTONIC,
2106 hrtimer_init_sleeper(to, current);
2107 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2108 current->timer_slack_ns);
2113 * Prepare to wait on uaddr. On success, holds hb lock and increments
2116 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2120 /* queue_me and wait for wakeup, timeout, or a signal. */
2121 futex_wait_queue_me(hb, &q, to);
2123 /* If we were woken (and unqueued), we succeeded, whatever. */
2125 /* unqueue_me() drops q.key ref */
2126 if (!unqueue_me(&q))
2129 if (to && !to->task)
2133 * We expect signal_pending(current), but we might be the
2134 * victim of a spurious wakeup as well.
2136 if (!signal_pending(current))
2143 restart = ¤t_thread_info()->restart_block;
2144 restart->fn = futex_wait_restart;
2145 restart->futex.uaddr = uaddr;
2146 restart->futex.val = val;
2147 restart->futex.time = abs_time->tv64;
2148 restart->futex.bitset = bitset;
2149 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2151 ret = -ERESTART_RESTARTBLOCK;
2155 hrtimer_cancel(&to->timer);
2156 destroy_hrtimer_on_stack(&to->timer);
2162 static long futex_wait_restart(struct restart_block *restart)
2164 u32 __user *uaddr = restart->futex.uaddr;
2165 ktime_t t, *tp = NULL;
2167 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2168 t.tv64 = restart->futex.time;
2171 restart->fn = do_no_restart_syscall;
2173 return (long)futex_wait(uaddr, restart->futex.flags,
2174 restart->futex.val, tp, restart->futex.bitset);
2179 * Userspace tried a 0 -> TID atomic transition of the futex value
2180 * and failed. The kernel side here does the whole locking operation:
2181 * if there are waiters then it will block, it does PI, etc. (Due to
2182 * races the kernel might see a 0 value of the futex too.)
2184 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2185 ktime_t *time, int trylock)
2187 struct hrtimer_sleeper timeout, *to = NULL;
2188 struct futex_hash_bucket *hb;
2189 struct futex_q q = futex_q_init;
2192 if (refill_pi_state_cache())
2197 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2199 hrtimer_init_sleeper(to, current);
2200 hrtimer_set_expires(&to->timer, *time);
2204 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2205 if (unlikely(ret != 0))
2209 hb = queue_lock(&q);
2211 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2212 if (unlikely(ret)) {
2215 /* We got the lock. */
2217 goto out_unlock_put_key;
2222 * Task is exiting and we just wait for the
2226 put_futex_key(&q.key);
2230 goto out_unlock_put_key;
2235 * Only actually queue now that the atomic ops are done:
2239 WARN_ON(!q.pi_state);
2241 * Block on the PI mutex:
2244 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2246 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2247 /* Fixup the trylock return value: */
2248 ret = ret ? 0 : -EWOULDBLOCK;
2251 spin_lock(q.lock_ptr);
2253 * Fixup the pi_state owner and possibly acquire the lock if we
2256 res = fixup_owner(uaddr, &q, !ret);
2258 * If fixup_owner() returned an error, proprogate that. If it acquired
2259 * the lock, clear our -ETIMEDOUT or -EINTR.
2262 ret = (res < 0) ? res : 0;
2265 * If fixup_owner() faulted and was unable to handle the fault, unlock
2266 * it and return the fault to userspace.
2268 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2269 rt_mutex_unlock(&q.pi_state->pi_mutex);
2271 /* Unqueue and drop the lock */
2280 put_futex_key(&q.key);
2283 destroy_hrtimer_on_stack(&to->timer);
2284 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2289 ret = fault_in_user_writeable(uaddr);
2293 if (!(flags & FLAGS_SHARED))
2296 put_futex_key(&q.key);
2301 * Userspace attempted a TID -> 0 atomic transition, and failed.
2302 * This is the in-kernel slowpath: we look up the PI state (if any),
2303 * and do the rt-mutex unlock.
2305 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2307 struct futex_hash_bucket *hb;
2308 struct futex_q *this, *next;
2309 union futex_key key = FUTEX_KEY_INIT;
2310 u32 uval, vpid = task_pid_vnr(current);
2314 if (get_user(uval, uaddr))
2317 * We release only a lock we actually own:
2319 if ((uval & FUTEX_TID_MASK) != vpid)
2322 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2323 if (unlikely(ret != 0))
2326 hb = hash_futex(&key);
2327 spin_lock(&hb->lock);
2330 * To avoid races, try to do the TID -> 0 atomic transition
2331 * again. If it succeeds then we can return without waking
2334 if (!(uval & FUTEX_OWNER_DIED) &&
2335 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2338 * Rare case: we managed to release the lock atomically,
2339 * no need to wake anyone else up:
2341 if (unlikely(uval == vpid))
2345 * Ok, other tasks may need to be woken up - check waiters
2346 * and do the wakeup if necessary:
2348 plist_for_each_entry_safe(this, next, &hb->chain, list) {
2349 if (!match_futex (&this->key, &key))
2351 ret = wake_futex_pi(uaddr, uval, this);
2353 * The atomic access to the futex value
2354 * generated a pagefault, so retry the
2355 * user-access and the wakeup:
2362 * No waiters - kernel unlocks the futex:
2364 if (!(uval & FUTEX_OWNER_DIED)) {
2365 ret = unlock_futex_pi(uaddr, uval);
2371 spin_unlock(&hb->lock);
2372 put_futex_key(&key);
2378 spin_unlock(&hb->lock);
2379 put_futex_key(&key);
2381 ret = fault_in_user_writeable(uaddr);
2389 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2390 * @hb: the hash_bucket futex_q was original enqueued on
2391 * @q: the futex_q woken while waiting to be requeued
2392 * @key2: the futex_key of the requeue target futex
2393 * @timeout: the timeout associated with the wait (NULL if none)
2395 * Detect if the task was woken on the initial futex as opposed to the requeue
2396 * target futex. If so, determine if it was a timeout or a signal that caused
2397 * the wakeup and return the appropriate error code to the caller. Must be
2398 * called with the hb lock held.
2401 * 0 = no early wakeup detected;
2402 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2405 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2406 struct futex_q *q, union futex_key *key2,
2407 struct hrtimer_sleeper *timeout)
2412 * With the hb lock held, we avoid races while we process the wakeup.
2413 * We only need to hold hb (and not hb2) to ensure atomicity as the
2414 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2415 * It can't be requeued from uaddr2 to something else since we don't
2416 * support a PI aware source futex for requeue.
2418 if (!match_futex(&q->key, key2)) {
2419 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2421 * We were woken prior to requeue by a timeout or a signal.
2422 * Unqueue the futex_q and determine which it was.
2424 plist_del(&q->list, &hb->chain);
2427 /* Handle spurious wakeups gracefully */
2429 if (timeout && !timeout->task)
2431 else if (signal_pending(current))
2432 ret = -ERESTARTNOINTR;
2438 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2439 * @uaddr: the futex we initially wait on (non-pi)
2440 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2441 * the same type, no requeueing from private to shared, etc.
2442 * @val: the expected value of uaddr
2443 * @abs_time: absolute timeout
2444 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2445 * @uaddr2: the pi futex we will take prior to returning to user-space
2447 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2448 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2449 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2450 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2451 * without one, the pi logic would not know which task to boost/deboost, if
2452 * there was a need to.
2454 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2455 * via the following--
2456 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2457 * 2) wakeup on uaddr2 after a requeue
2461 * If 3, cleanup and return -ERESTARTNOINTR.
2463 * If 2, we may then block on trying to take the rt_mutex and return via:
2464 * 5) successful lock
2467 * 8) other lock acquisition failure
2469 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2471 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2477 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2478 u32 val, ktime_t *abs_time, u32 bitset,
2481 struct hrtimer_sleeper timeout, *to = NULL;
2482 struct rt_mutex_waiter rt_waiter;
2483 struct rt_mutex *pi_mutex = NULL;
2484 struct futex_hash_bucket *hb;
2485 union futex_key key2 = FUTEX_KEY_INIT;
2486 struct futex_q q = futex_q_init;
2489 if (uaddr == uaddr2)
2497 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2498 CLOCK_REALTIME : CLOCK_MONOTONIC,
2500 hrtimer_init_sleeper(to, current);
2501 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2502 current->timer_slack_ns);
2506 * The waiter is allocated on our stack, manipulated by the requeue
2507 * code while we sleep on uaddr.
2509 debug_rt_mutex_init_waiter(&rt_waiter);
2510 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2511 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2512 rt_waiter.task = NULL;
2514 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2515 if (unlikely(ret != 0))
2519 q.rt_waiter = &rt_waiter;
2520 q.requeue_pi_key = &key2;
2523 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2526 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2531 * The check above which compares uaddrs is not sufficient for
2532 * shared futexes. We need to compare the keys:
2534 if (match_futex(&q.key, &key2)) {
2539 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2540 futex_wait_queue_me(hb, &q, to);
2542 spin_lock(&hb->lock);
2543 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2544 spin_unlock(&hb->lock);
2549 * In order for us to be here, we know our q.key == key2, and since
2550 * we took the hb->lock above, we also know that futex_requeue() has
2551 * completed and we no longer have to concern ourselves with a wakeup
2552 * race with the atomic proxy lock acquisition by the requeue code. The
2553 * futex_requeue dropped our key1 reference and incremented our key2
2557 /* Check if the requeue code acquired the second futex for us. */
2560 * Got the lock. We might not be the anticipated owner if we
2561 * did a lock-steal - fix up the PI-state in that case.
2563 if (q.pi_state && (q.pi_state->owner != current)) {
2564 spin_lock(q.lock_ptr);
2565 ret = fixup_pi_state_owner(uaddr2, &q, current);
2566 spin_unlock(q.lock_ptr);
2570 * We have been woken up by futex_unlock_pi(), a timeout, or a
2571 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2574 WARN_ON(!q.pi_state);
2575 pi_mutex = &q.pi_state->pi_mutex;
2576 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2577 debug_rt_mutex_free_waiter(&rt_waiter);
2579 spin_lock(q.lock_ptr);
2581 * Fixup the pi_state owner and possibly acquire the lock if we
2584 res = fixup_owner(uaddr2, &q, !ret);
2586 * If fixup_owner() returned an error, proprogate that. If it
2587 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2590 ret = (res < 0) ? res : 0;
2592 /* Unqueue and drop the lock. */
2597 * If fixup_pi_state_owner() faulted and was unable to handle the
2598 * fault, unlock the rt_mutex and return the fault to userspace.
2600 if (ret == -EFAULT) {
2601 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2602 rt_mutex_unlock(pi_mutex);
2603 } else if (ret == -EINTR) {
2605 * We've already been requeued, but cannot restart by calling
2606 * futex_lock_pi() directly. We could restart this syscall, but
2607 * it would detect that the user space "val" changed and return
2608 * -EWOULDBLOCK. Save the overhead of the restart and return
2609 * -EWOULDBLOCK directly.
2615 put_futex_key(&q.key);
2617 put_futex_key(&key2);
2621 hrtimer_cancel(&to->timer);
2622 destroy_hrtimer_on_stack(&to->timer);
2628 * Support for robust futexes: the kernel cleans up held futexes at
2631 * Implementation: user-space maintains a per-thread list of locks it
2632 * is holding. Upon do_exit(), the kernel carefully walks this list,
2633 * and marks all locks that are owned by this thread with the
2634 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2635 * always manipulated with the lock held, so the list is private and
2636 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2637 * field, to allow the kernel to clean up if the thread dies after
2638 * acquiring the lock, but just before it could have added itself to
2639 * the list. There can only be one such pending lock.
2643 * sys_set_robust_list() - Set the robust-futex list head of a task
2644 * @head: pointer to the list-head
2645 * @len: length of the list-head, as userspace expects
2647 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2650 if (!futex_cmpxchg_enabled)
2653 * The kernel knows only one size for now:
2655 if (unlikely(len != sizeof(*head)))
2658 current->robust_list = head;
2664 * sys_get_robust_list() - Get the robust-futex list head of a task
2665 * @pid: pid of the process [zero for current task]
2666 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2667 * @len_ptr: pointer to a length field, the kernel fills in the header size
2669 SYSCALL_DEFINE3(get_robust_list, int, pid,
2670 struct robust_list_head __user * __user *, head_ptr,
2671 size_t __user *, len_ptr)
2673 struct robust_list_head __user *head;
2675 struct task_struct *p;
2677 if (!futex_cmpxchg_enabled)
2686 p = find_task_by_vpid(pid);
2692 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2695 head = p->robust_list;
2698 if (put_user(sizeof(*head), len_ptr))
2700 return put_user(head, head_ptr);
2709 * Process a futex-list entry, check whether it's owned by the
2710 * dying task, and do notification if so:
2712 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2714 u32 uval, uninitialized_var(nval), mval;
2717 if (get_user(uval, uaddr))
2720 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2722 * Ok, this dying thread is truly holding a futex
2723 * of interest. Set the OWNER_DIED bit atomically
2724 * via cmpxchg, and if the value had FUTEX_WAITERS
2725 * set, wake up a waiter (if any). (We have to do a
2726 * futex_wake() even if OWNER_DIED is already set -
2727 * to handle the rare but possible case of recursive
2728 * thread-death.) The rest of the cleanup is done in
2731 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2733 * We are not holding a lock here, but we want to have
2734 * the pagefault_disable/enable() protection because
2735 * we want to handle the fault gracefully. If the
2736 * access fails we try to fault in the futex with R/W
2737 * verification via get_user_pages. get_user() above
2738 * does not guarantee R/W access. If that fails we
2739 * give up and leave the futex locked.
2741 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2742 if (fault_in_user_writeable(uaddr))
2750 * Wake robust non-PI futexes here. The wakeup of
2751 * PI futexes happens in exit_pi_state():
2753 if (!pi && (uval & FUTEX_WAITERS))
2754 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2760 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2762 static inline int fetch_robust_entry(struct robust_list __user **entry,
2763 struct robust_list __user * __user *head,
2766 unsigned long uentry;
2768 if (get_user(uentry, (unsigned long __user *)head))
2771 *entry = (void __user *)(uentry & ~1UL);
2778 * Walk curr->robust_list (very carefully, it's a userspace list!)
2779 * and mark any locks found there dead, and notify any waiters.
2781 * We silently return on any sign of list-walking problem.
2783 void exit_robust_list(struct task_struct *curr)
2785 struct robust_list_head __user *head = curr->robust_list;
2786 struct robust_list __user *entry, *next_entry, *pending;
2787 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2788 unsigned int uninitialized_var(next_pi);
2789 unsigned long futex_offset;
2792 if (!futex_cmpxchg_enabled)
2796 * Fetch the list head (which was registered earlier, via
2797 * sys_set_robust_list()):
2799 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2802 * Fetch the relative futex offset:
2804 if (get_user(futex_offset, &head->futex_offset))
2807 * Fetch any possibly pending lock-add first, and handle it
2810 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2813 next_entry = NULL; /* avoid warning with gcc */
2814 while (entry != &head->list) {
2816 * Fetch the next entry in the list before calling
2817 * handle_futex_death:
2819 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2821 * A pending lock might already be on the list, so
2822 * don't process it twice:
2824 if (entry != pending)
2825 if (handle_futex_death((void __user *)entry + futex_offset,
2833 * Avoid excessively long or circular lists:
2842 handle_futex_death((void __user *)pending + futex_offset,
2846 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2847 u32 __user *uaddr2, u32 val2, u32 val3)
2849 int cmd = op & FUTEX_CMD_MASK;
2850 unsigned int flags = 0;
2852 if (!(op & FUTEX_PRIVATE_FLAG))
2853 flags |= FLAGS_SHARED;
2855 if (op & FUTEX_CLOCK_REALTIME) {
2856 flags |= FLAGS_CLOCKRT;
2857 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2863 case FUTEX_UNLOCK_PI:
2864 case FUTEX_TRYLOCK_PI:
2865 case FUTEX_WAIT_REQUEUE_PI:
2866 case FUTEX_CMP_REQUEUE_PI:
2867 if (!futex_cmpxchg_enabled)
2873 val3 = FUTEX_BITSET_MATCH_ANY;
2874 case FUTEX_WAIT_BITSET:
2875 return futex_wait(uaddr, flags, val, timeout, val3);
2877 val3 = FUTEX_BITSET_MATCH_ANY;
2878 case FUTEX_WAKE_BITSET:
2879 return futex_wake(uaddr, flags, val, val3);
2881 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2882 case FUTEX_CMP_REQUEUE:
2883 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2885 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2887 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2888 case FUTEX_UNLOCK_PI:
2889 return futex_unlock_pi(uaddr, flags);
2890 case FUTEX_TRYLOCK_PI:
2891 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2892 case FUTEX_WAIT_REQUEUE_PI:
2893 val3 = FUTEX_BITSET_MATCH_ANY;
2894 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2896 case FUTEX_CMP_REQUEUE_PI:
2897 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2903 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2904 struct timespec __user *, utime, u32 __user *, uaddr2,
2908 ktime_t t, *tp = NULL;
2910 int cmd = op & FUTEX_CMD_MASK;
2912 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2913 cmd == FUTEX_WAIT_BITSET ||
2914 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2915 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2917 if (!timespec_valid(&ts))
2920 t = timespec_to_ktime(ts);
2921 if (cmd == FUTEX_WAIT)
2922 t = ktime_add_safe(ktime_get(), t);
2926 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2927 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2929 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2930 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2931 val2 = (u32) (unsigned long) utime;
2933 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2936 static void __init futex_detect_cmpxchg(void)
2938 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
2942 * This will fail and we want it. Some arch implementations do
2943 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2944 * functionality. We want to know that before we call in any
2945 * of the complex code paths. Also we want to prevent
2946 * registration of robust lists in that case. NULL is
2947 * guaranteed to fault and we get -EFAULT on functional
2948 * implementation, the non-functional ones will return
2951 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2952 futex_cmpxchg_enabled = 1;
2956 static int __init futex_init(void)
2958 unsigned int futex_shift;
2961 #if CONFIG_BASE_SMALL
2962 futex_hashsize = 16;
2964 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
2967 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
2969 futex_hashsize < 256 ? HASH_SMALL : 0,
2971 futex_hashsize, futex_hashsize);
2972 futex_hashsize = 1UL << futex_shift;
2974 futex_detect_cmpxchg();
2976 for (i = 0; i < futex_hashsize; i++) {
2977 atomic_set(&futex_queues[i].waiters, 0);
2978 plist_head_init(&futex_queues[i].chain);
2979 spin_lock_init(&futex_queues[i].lock);
2984 __initcall(futex_init);