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 in
147 * get_futex_key_refs(), through either ihold or atomic_inc, depending on the
150 * This yields the following case (where X:=waiters, Y:=futex):
158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
159 * the guarantee that we cannot both miss the futex variable change and the
162 * Note that a new waiter is accounted for in (a) even when it is possible that
163 * the wait call can return error, in which case we backtrack from it in (b).
164 * Refer to the comment in queue_lock().
166 * Similarly, in order to account for waiters being requeued on another
167 * address we always increment the waiters for the destination bucket before
168 * acquiring the lock. It then decrements them again after releasing it -
169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170 * will do the additional required waiter count housekeeping. This is done for
171 * double_lock_hb() and double_unlock_hb(), respectively.
174 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175 int __read_mostly futex_cmpxchg_enabled;
179 * Futex flags used to encode options to functions and preserve them across
182 #define FLAGS_SHARED 0x01
183 #define FLAGS_CLOCKRT 0x02
184 #define FLAGS_HAS_TIMEOUT 0x04
187 * Priority Inheritance state:
189 struct futex_pi_state {
191 * list of 'owned' pi_state instances - these have to be
192 * cleaned up in do_exit() if the task exits prematurely:
194 struct list_head list;
199 struct rt_mutex pi_mutex;
201 struct task_struct *owner;
208 * struct futex_q - The hashed futex queue entry, one per waiting task
209 * @list: priority-sorted list of tasks waiting on this futex
210 * @task: the task waiting on the futex
211 * @lock_ptr: the hash bucket lock
212 * @key: the key the futex is hashed on
213 * @pi_state: optional priority inheritance state
214 * @rt_waiter: rt_waiter storage for use with requeue_pi
215 * @requeue_pi_key: the requeue_pi target futex key
216 * @bitset: bitset for the optional bitmasked wakeup
218 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
219 * we can wake only the relevant ones (hashed queues may be shared).
221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223 * The order of wakeup is always to make the first condition true, then
226 * PI futexes are typically woken before they are removed from the hash list via
227 * the rt_mutex code. See unqueue_me_pi().
230 struct plist_node list;
232 struct task_struct *task;
233 spinlock_t *lock_ptr;
235 struct futex_pi_state *pi_state;
236 struct rt_mutex_waiter *rt_waiter;
237 union futex_key *requeue_pi_key;
241 static const struct futex_q futex_q_init = {
242 /* list gets initialized in queue_me()*/
243 .key = FUTEX_KEY_INIT,
244 .bitset = FUTEX_BITSET_MATCH_ANY
248 * Hash buckets are shared by all the futex_keys that hash to the same
249 * location. Each key may have multiple futex_q structures, one for each task
250 * waiting on a futex.
252 struct futex_hash_bucket {
255 struct plist_head chain;
256 } ____cacheline_aligned_in_smp;
258 static unsigned long __read_mostly futex_hashsize;
260 static struct futex_hash_bucket *futex_queues;
262 static inline void futex_get_mm(union futex_key *key)
264 atomic_inc(&key->private.mm->mm_count);
266 * Ensure futex_get_mm() implies a full barrier such that
267 * get_futex_key() implies a full barrier. This is relied upon
268 * as full barrier (B), see the ordering comment above.
270 smp_mb__after_atomic();
274 * Reflects a new waiter being added to the waitqueue.
276 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
279 atomic_inc(&hb->waiters);
281 * Full barrier (A), see the ordering comment above.
283 smp_mb__after_atomic();
288 * Reflects a waiter being removed from the waitqueue by wakeup
291 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
294 atomic_dec(&hb->waiters);
298 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
301 return atomic_read(&hb->waiters);
308 * We hash on the keys returned from get_futex_key (see below).
310 static struct futex_hash_bucket *hash_futex(union futex_key *key)
312 u32 hash = jhash2((u32*)&key->both.word,
313 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
315 return &futex_queues[hash & (futex_hashsize - 1)];
319 * Return 1 if two futex_keys are equal, 0 otherwise.
321 static inline int match_futex(union futex_key *key1, union futex_key *key2)
324 && key1->both.word == key2->both.word
325 && key1->both.ptr == key2->both.ptr
326 && key1->both.offset == key2->both.offset);
330 * Take a reference to the resource addressed by a key.
331 * Can be called while holding spinlocks.
334 static void get_futex_key_refs(union futex_key *key)
339 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
341 ihold(key->shared.inode); /* implies MB (B) */
343 case FUT_OFF_MMSHARED:
344 futex_get_mm(key); /* implies MB (B) */
350 * Drop a reference to the resource addressed by a key.
351 * The hash bucket spinlock must not be held.
353 static void drop_futex_key_refs(union futex_key *key)
355 if (!key->both.ptr) {
356 /* If we're here then we tried to put a key we failed to get */
361 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
363 iput(key->shared.inode);
365 case FUT_OFF_MMSHARED:
366 mmdrop(key->private.mm);
372 * get_futex_key() - Get parameters which are the keys for a futex
373 * @uaddr: virtual address of the futex
374 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
375 * @key: address where result is stored.
376 * @rw: mapping needs to be read/write (values: VERIFY_READ,
379 * Return: a negative error code or 0
381 * The key words are stored in *key on success.
383 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
384 * offset_within_page). For private mappings, it's (uaddr, current->mm).
385 * We can usually work out the index without swapping in the page.
387 * lock_page() might sleep, the caller should not hold a spinlock.
390 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
392 unsigned long address = (unsigned long)uaddr;
393 struct mm_struct *mm = current->mm;
394 struct page *page, *page_head;
398 * The futex address must be "naturally" aligned.
400 key->both.offset = address % PAGE_SIZE;
401 if (unlikely((address % sizeof(u32)) != 0))
403 address -= key->both.offset;
405 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
409 * PROCESS_PRIVATE futexes are fast.
410 * As the mm cannot disappear under us and the 'key' only needs
411 * virtual address, we dont even have to find the underlying vma.
412 * Note : We do have to check 'uaddr' is a valid user address,
413 * but access_ok() should be faster than find_vma()
416 key->private.mm = mm;
417 key->private.address = address;
418 get_futex_key_refs(key); /* implies MB (B) */
423 err = get_user_pages_fast(address, 1, 1, &page);
425 * If write access is not required (eg. FUTEX_WAIT), try
426 * and get read-only access.
428 if (err == -EFAULT && rw == VERIFY_READ) {
429 err = get_user_pages_fast(address, 1, 0, &page);
437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
439 if (unlikely(PageTail(page))) {
441 /* serialize against __split_huge_page_splitting() */
443 if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
444 page_head = compound_head(page);
446 * page_head is valid pointer but we must pin
447 * it before taking the PG_lock and/or
448 * PG_compound_lock. The moment we re-enable
449 * irqs __split_huge_page_splitting() can
450 * return and the head page can be freed from
451 * under us. We can't take the PG_lock and/or
452 * PG_compound_lock on a page that could be
453 * freed from under us.
455 if (page != page_head) {
466 page_head = compound_head(page);
467 if (page != page_head) {
473 lock_page(page_head);
476 * If page_head->mapping is NULL, then it cannot be a PageAnon
477 * page; but it might be the ZERO_PAGE or in the gate area or
478 * in a special mapping (all cases which we are happy to fail);
479 * or it may have been a good file page when get_user_pages_fast
480 * found it, but truncated or holepunched or subjected to
481 * invalidate_complete_page2 before we got the page lock (also
482 * cases which we are happy to fail). And we hold a reference,
483 * so refcount care in invalidate_complete_page's remove_mapping
484 * prevents drop_caches from setting mapping to NULL beneath us.
486 * The case we do have to guard against is when memory pressure made
487 * shmem_writepage move it from filecache to swapcache beneath us:
488 * an unlikely race, but we do need to retry for page_head->mapping.
490 if (!page_head->mapping) {
491 int shmem_swizzled = PageSwapCache(page_head);
492 unlock_page(page_head);
500 * Private mappings are handled in a simple way.
502 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
503 * it's a read-only handle, it's expected that futexes attach to
504 * the object not the particular process.
506 if (PageAnon(page_head)) {
508 * A RO anonymous page will never change and thus doesn't make
509 * sense for futex operations.
516 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
517 key->private.mm = mm;
518 key->private.address = address;
520 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
521 key->shared.inode = page_head->mapping->host;
522 key->shared.pgoff = basepage_index(page);
525 get_futex_key_refs(key); /* implies MB (B) */
528 unlock_page(page_head);
533 static inline void put_futex_key(union futex_key *key)
535 drop_futex_key_refs(key);
539 * fault_in_user_writeable() - Fault in user address and verify RW access
540 * @uaddr: pointer to faulting user space address
542 * Slow path to fixup the fault we just took in the atomic write
545 * We have no generic implementation of a non-destructive write to the
546 * user address. We know that we faulted in the atomic pagefault
547 * disabled section so we can as well avoid the #PF overhead by
548 * calling get_user_pages() right away.
550 static int fault_in_user_writeable(u32 __user *uaddr)
552 struct mm_struct *mm = current->mm;
555 down_read(&mm->mmap_sem);
556 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
558 up_read(&mm->mmap_sem);
560 return ret < 0 ? ret : 0;
564 * futex_top_waiter() - Return the highest priority waiter on a futex
565 * @hb: the hash bucket the futex_q's reside in
566 * @key: the futex key (to distinguish it from other futex futex_q's)
568 * Must be called with the hb lock held.
570 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
571 union futex_key *key)
573 struct futex_q *this;
575 plist_for_each_entry(this, &hb->chain, list) {
576 if (match_futex(&this->key, key))
582 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
583 u32 uval, u32 newval)
588 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
594 static int get_futex_value_locked(u32 *dest, u32 __user *from)
599 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
602 return ret ? -EFAULT : 0;
609 static int refill_pi_state_cache(void)
611 struct futex_pi_state *pi_state;
613 if (likely(current->pi_state_cache))
616 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
621 INIT_LIST_HEAD(&pi_state->list);
622 /* pi_mutex gets initialized later */
623 pi_state->owner = NULL;
624 atomic_set(&pi_state->refcount, 1);
625 pi_state->key = FUTEX_KEY_INIT;
627 current->pi_state_cache = pi_state;
632 static struct futex_pi_state * alloc_pi_state(void)
634 struct futex_pi_state *pi_state = current->pi_state_cache;
637 current->pi_state_cache = NULL;
642 static void free_pi_state(struct futex_pi_state *pi_state)
644 if (!atomic_dec_and_test(&pi_state->refcount))
648 * If pi_state->owner is NULL, the owner is most probably dying
649 * and has cleaned up the pi_state already
651 if (pi_state->owner) {
652 raw_spin_lock_irq(&pi_state->owner->pi_lock);
653 list_del_init(&pi_state->list);
654 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
656 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
659 if (current->pi_state_cache)
663 * pi_state->list is already empty.
664 * clear pi_state->owner.
665 * refcount is at 0 - put it back to 1.
667 pi_state->owner = NULL;
668 atomic_set(&pi_state->refcount, 1);
669 current->pi_state_cache = pi_state;
674 * Look up the task based on what TID userspace gave us.
677 static struct task_struct * futex_find_get_task(pid_t pid)
679 struct task_struct *p;
682 p = find_task_by_vpid(pid);
692 * This task is holding PI mutexes at exit time => bad.
693 * Kernel cleans up PI-state, but userspace is likely hosed.
694 * (Robust-futex cleanup is separate and might save the day for userspace.)
696 void exit_pi_state_list(struct task_struct *curr)
698 struct list_head *next, *head = &curr->pi_state_list;
699 struct futex_pi_state *pi_state;
700 struct futex_hash_bucket *hb;
701 union futex_key key = FUTEX_KEY_INIT;
703 if (!futex_cmpxchg_enabled)
706 * We are a ZOMBIE and nobody can enqueue itself on
707 * pi_state_list anymore, but we have to be careful
708 * versus waiters unqueueing themselves:
710 raw_spin_lock_irq(&curr->pi_lock);
711 while (!list_empty(head)) {
714 pi_state = list_entry(next, struct futex_pi_state, list);
716 hb = hash_futex(&key);
717 raw_spin_unlock_irq(&curr->pi_lock);
719 spin_lock(&hb->lock);
721 raw_spin_lock_irq(&curr->pi_lock);
723 * We dropped the pi-lock, so re-check whether this
724 * task still owns the PI-state:
726 if (head->next != next) {
727 spin_unlock(&hb->lock);
731 WARN_ON(pi_state->owner != curr);
732 WARN_ON(list_empty(&pi_state->list));
733 list_del_init(&pi_state->list);
734 pi_state->owner = NULL;
735 raw_spin_unlock_irq(&curr->pi_lock);
737 rt_mutex_unlock(&pi_state->pi_mutex);
739 spin_unlock(&hb->lock);
741 raw_spin_lock_irq(&curr->pi_lock);
743 raw_spin_unlock_irq(&curr->pi_lock);
747 * We need to check the following states:
749 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
751 * [1] NULL | --- | --- | 0 | 0/1 | Valid
752 * [2] NULL | --- | --- | >0 | 0/1 | Valid
754 * [3] Found | NULL | -- | Any | 0/1 | Invalid
756 * [4] Found | Found | NULL | 0 | 1 | Valid
757 * [5] Found | Found | NULL | >0 | 1 | Invalid
759 * [6] Found | Found | task | 0 | 1 | Valid
761 * [7] Found | Found | NULL | Any | 0 | Invalid
763 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
764 * [9] Found | Found | task | 0 | 0 | Invalid
765 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
767 * [1] Indicates that the kernel can acquire the futex atomically. We
768 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
770 * [2] Valid, if TID does not belong to a kernel thread. If no matching
771 * thread is found then it indicates that the owner TID has died.
773 * [3] Invalid. The waiter is queued on a non PI futex
775 * [4] Valid state after exit_robust_list(), which sets the user space
776 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
778 * [5] The user space value got manipulated between exit_robust_list()
779 * and exit_pi_state_list()
781 * [6] Valid state after exit_pi_state_list() which sets the new owner in
782 * the pi_state but cannot access the user space value.
784 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
786 * [8] Owner and user space value match
788 * [9] There is no transient state which sets the user space TID to 0
789 * except exit_robust_list(), but this is indicated by the
790 * FUTEX_OWNER_DIED bit. See [4]
792 * [10] There is no transient state which leaves owner and user space
797 * Validate that the existing waiter has a pi_state and sanity check
798 * the pi_state against the user space value. If correct, attach to
801 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
802 struct futex_pi_state **ps)
804 pid_t pid = uval & FUTEX_TID_MASK;
807 * Userspace might have messed up non-PI and PI futexes [3]
809 if (unlikely(!pi_state))
812 WARN_ON(!atomic_read(&pi_state->refcount));
815 * Handle the owner died case:
817 if (uval & FUTEX_OWNER_DIED) {
819 * exit_pi_state_list sets owner to NULL and wakes the
820 * topmost waiter. The task which acquires the
821 * pi_state->rt_mutex will fixup owner.
823 if (!pi_state->owner) {
825 * No pi state owner, but the user space TID
826 * is not 0. Inconsistent state. [5]
831 * Take a ref on the state and return success. [4]
837 * If TID is 0, then either the dying owner has not
838 * yet executed exit_pi_state_list() or some waiter
839 * acquired the rtmutex in the pi state, but did not
840 * yet fixup the TID in user space.
842 * Take a ref on the state and return success. [6]
848 * If the owner died bit is not set, then the pi_state
849 * must have an owner. [7]
851 if (!pi_state->owner)
856 * Bail out if user space manipulated the futex value. If pi
857 * state exists then the owner TID must be the same as the
858 * user space TID. [9/10]
860 if (pid != task_pid_vnr(pi_state->owner))
863 atomic_inc(&pi_state->refcount);
869 * Lookup the task for the TID provided from user space and attach to
870 * it after doing proper sanity checks.
872 static int attach_to_pi_owner(u32 uval, union futex_key *key,
873 struct futex_pi_state **ps)
875 pid_t pid = uval & FUTEX_TID_MASK;
876 struct futex_pi_state *pi_state;
877 struct task_struct *p;
880 * We are the first waiter - try to look up the real owner and attach
881 * the new pi_state to it, but bail out when TID = 0 [1]
885 p = futex_find_get_task(pid);
895 * We need to look at the task state flags to figure out,
896 * whether the task is exiting. To protect against the do_exit
897 * change of the task flags, we do this protected by
900 raw_spin_lock_irq(&p->pi_lock);
901 if (unlikely(p->flags & PF_EXITING)) {
903 * The task is on the way out. When PF_EXITPIDONE is
904 * set, we know that the task has finished the
907 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
909 raw_spin_unlock_irq(&p->pi_lock);
915 * No existing pi state. First waiter. [2]
917 pi_state = alloc_pi_state();
920 * Initialize the pi_mutex in locked state and make @p
923 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
925 /* Store the key for possible exit cleanups: */
926 pi_state->key = *key;
928 WARN_ON(!list_empty(&pi_state->list));
929 list_add(&pi_state->list, &p->pi_state_list);
931 raw_spin_unlock_irq(&p->pi_lock);
940 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
941 union futex_key *key, struct futex_pi_state **ps)
943 struct futex_q *match = futex_top_waiter(hb, key);
946 * If there is a waiter on that futex, validate it and
947 * attach to the pi_state when the validation succeeds.
950 return attach_to_pi_state(uval, match->pi_state, ps);
953 * We are the first waiter - try to look up the owner based on
954 * @uval and attach to it.
956 return attach_to_pi_owner(uval, key, ps);
959 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
961 u32 uninitialized_var(curval);
963 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
966 /*If user space value changed, let the caller retry */
967 return curval != uval ? -EAGAIN : 0;
971 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
972 * @uaddr: the pi futex user address
973 * @hb: the pi futex hash bucket
974 * @key: the futex key associated with uaddr and hb
975 * @ps: the pi_state pointer where we store the result of the
977 * @task: the task to perform the atomic lock work for. This will
978 * be "current" except in the case of requeue pi.
979 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
983 * 1 - acquired the lock;
986 * The hb->lock and futex_key refs shall be held by the caller.
988 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
989 union futex_key *key,
990 struct futex_pi_state **ps,
991 struct task_struct *task, int set_waiters)
993 u32 uval, newval, vpid = task_pid_vnr(task);
994 struct futex_q *match;
998 * Read the user space value first so we can validate a few
999 * things before proceeding further.
1001 if (get_futex_value_locked(&uval, uaddr))
1007 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1011 * Lookup existing state first. If it exists, try to attach to
1014 match = futex_top_waiter(hb, key);
1016 return attach_to_pi_state(uval, match->pi_state, ps);
1019 * No waiter and user TID is 0. We are here because the
1020 * waiters or the owner died bit is set or called from
1021 * requeue_cmp_pi or for whatever reason something took the
1024 if (!(uval & FUTEX_TID_MASK)) {
1026 * We take over the futex. No other waiters and the user space
1027 * TID is 0. We preserve the owner died bit.
1029 newval = uval & FUTEX_OWNER_DIED;
1032 /* The futex requeue_pi code can enforce the waiters bit */
1034 newval |= FUTEX_WAITERS;
1036 ret = lock_pi_update_atomic(uaddr, uval, newval);
1037 /* If the take over worked, return 1 */
1038 return ret < 0 ? ret : 1;
1042 * First waiter. Set the waiters bit before attaching ourself to
1043 * the owner. If owner tries to unlock, it will be forced into
1044 * the kernel and blocked on hb->lock.
1046 newval = uval | FUTEX_WAITERS;
1047 ret = lock_pi_update_atomic(uaddr, uval, newval);
1051 * If the update of the user space value succeeded, we try to
1052 * attach to the owner. If that fails, no harm done, we only
1053 * set the FUTEX_WAITERS bit in the user space variable.
1055 return attach_to_pi_owner(uval, key, ps);
1059 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1060 * @q: The futex_q to unqueue
1062 * The q->lock_ptr must not be NULL and must be held by the caller.
1064 static void __unqueue_futex(struct futex_q *q)
1066 struct futex_hash_bucket *hb;
1068 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1069 || WARN_ON(plist_node_empty(&q->list)))
1072 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1073 plist_del(&q->list, &hb->chain);
1078 * The hash bucket lock must be held when this is called.
1079 * Afterwards, the futex_q must not be accessed.
1081 static void wake_futex(struct futex_q *q)
1083 struct task_struct *p = q->task;
1085 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1089 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1090 * a non-futex wake up happens on another CPU then the task
1091 * might exit and p would dereference a non-existing task
1092 * struct. Prevent this by holding a reference on p across the
1099 * The waiting task can free the futex_q as soon as
1100 * q->lock_ptr = NULL is written, without taking any locks. A
1101 * memory barrier is required here to prevent the following
1102 * store to lock_ptr from getting ahead of the plist_del.
1107 wake_up_state(p, TASK_NORMAL);
1111 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1113 struct task_struct *new_owner;
1114 struct futex_pi_state *pi_state = this->pi_state;
1115 u32 uninitialized_var(curval), newval;
1122 * If current does not own the pi_state then the futex is
1123 * inconsistent and user space fiddled with the futex value.
1125 if (pi_state->owner != current)
1128 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1129 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1132 * It is possible that the next waiter (the one that brought
1133 * this owner to the kernel) timed out and is no longer
1134 * waiting on the lock.
1137 new_owner = this->task;
1140 * We pass it to the next owner. The WAITERS bit is always
1141 * kept enabled while there is PI state around. We cleanup the
1142 * owner died bit, because we are the owner.
1144 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1146 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1148 else if (curval != uval)
1151 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1155 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1156 WARN_ON(list_empty(&pi_state->list));
1157 list_del_init(&pi_state->list);
1158 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1160 raw_spin_lock_irq(&new_owner->pi_lock);
1161 WARN_ON(!list_empty(&pi_state->list));
1162 list_add(&pi_state->list, &new_owner->pi_state_list);
1163 pi_state->owner = new_owner;
1164 raw_spin_unlock_irq(&new_owner->pi_lock);
1166 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1167 rt_mutex_unlock(&pi_state->pi_mutex);
1173 * Express the locking dependencies for lockdep:
1176 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1179 spin_lock(&hb1->lock);
1181 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1182 } else { /* hb1 > hb2 */
1183 spin_lock(&hb2->lock);
1184 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1189 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1191 spin_unlock(&hb1->lock);
1193 spin_unlock(&hb2->lock);
1197 * Wake up waiters matching bitset queued on this futex (uaddr).
1200 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1202 struct futex_hash_bucket *hb;
1203 struct futex_q *this, *next;
1204 union futex_key key = FUTEX_KEY_INIT;
1210 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1211 if (unlikely(ret != 0))
1214 hb = hash_futex(&key);
1216 /* Make sure we really have tasks to wakeup */
1217 if (!hb_waiters_pending(hb))
1220 spin_lock(&hb->lock);
1222 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1223 if (match_futex (&this->key, &key)) {
1224 if (this->pi_state || this->rt_waiter) {
1229 /* Check if one of the bits is set in both bitsets */
1230 if (!(this->bitset & bitset))
1234 if (++ret >= nr_wake)
1239 spin_unlock(&hb->lock);
1241 put_futex_key(&key);
1247 * Wake up all waiters hashed on the physical page that is mapped
1248 * to this virtual address:
1251 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1252 int nr_wake, int nr_wake2, int op)
1254 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1255 struct futex_hash_bucket *hb1, *hb2;
1256 struct futex_q *this, *next;
1260 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1261 if (unlikely(ret != 0))
1263 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1264 if (unlikely(ret != 0))
1267 hb1 = hash_futex(&key1);
1268 hb2 = hash_futex(&key2);
1271 double_lock_hb(hb1, hb2);
1272 op_ret = futex_atomic_op_inuser(op, uaddr2);
1273 if (unlikely(op_ret < 0)) {
1275 double_unlock_hb(hb1, hb2);
1279 * we don't get EFAULT from MMU faults if we don't have an MMU,
1280 * but we might get them from range checking
1286 if (unlikely(op_ret != -EFAULT)) {
1291 ret = fault_in_user_writeable(uaddr2);
1295 if (!(flags & FLAGS_SHARED))
1298 put_futex_key(&key2);
1299 put_futex_key(&key1);
1303 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1304 if (match_futex (&this->key, &key1)) {
1305 if (this->pi_state || this->rt_waiter) {
1310 if (++ret >= nr_wake)
1317 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1318 if (match_futex (&this->key, &key2)) {
1319 if (this->pi_state || this->rt_waiter) {
1324 if (++op_ret >= nr_wake2)
1332 double_unlock_hb(hb1, hb2);
1334 put_futex_key(&key2);
1336 put_futex_key(&key1);
1342 * requeue_futex() - Requeue a futex_q from one hb to another
1343 * @q: the futex_q to requeue
1344 * @hb1: the source hash_bucket
1345 * @hb2: the target hash_bucket
1346 * @key2: the new key for the requeued futex_q
1349 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1350 struct futex_hash_bucket *hb2, union futex_key *key2)
1354 * If key1 and key2 hash to the same bucket, no need to
1357 if (likely(&hb1->chain != &hb2->chain)) {
1358 plist_del(&q->list, &hb1->chain);
1359 hb_waiters_dec(hb1);
1360 plist_add(&q->list, &hb2->chain);
1361 hb_waiters_inc(hb2);
1362 q->lock_ptr = &hb2->lock;
1364 get_futex_key_refs(key2);
1369 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1371 * @key: the key of the requeue target futex
1372 * @hb: the hash_bucket of the requeue target futex
1374 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1375 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1376 * to the requeue target futex so the waiter can detect the wakeup on the right
1377 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1378 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1379 * to protect access to the pi_state to fixup the owner later. Must be called
1380 * with both q->lock_ptr and hb->lock held.
1383 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1384 struct futex_hash_bucket *hb)
1386 get_futex_key_refs(key);
1391 WARN_ON(!q->rt_waiter);
1392 q->rt_waiter = NULL;
1394 q->lock_ptr = &hb->lock;
1396 wake_up_state(q->task, TASK_NORMAL);
1400 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1401 * @pifutex: the user address of the to futex
1402 * @hb1: the from futex hash bucket, must be locked by the caller
1403 * @hb2: the to futex hash bucket, must be locked by the caller
1404 * @key1: the from futex key
1405 * @key2: the to futex key
1406 * @ps: address to store the pi_state pointer
1407 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1409 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1410 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1411 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1412 * hb1 and hb2 must be held by the caller.
1415 * 0 - failed to acquire the lock atomically;
1416 * >0 - acquired the lock, return value is vpid of the top_waiter
1419 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1420 struct futex_hash_bucket *hb1,
1421 struct futex_hash_bucket *hb2,
1422 union futex_key *key1, union futex_key *key2,
1423 struct futex_pi_state **ps, int set_waiters)
1425 struct futex_q *top_waiter = NULL;
1429 if (get_futex_value_locked(&curval, pifutex))
1433 * Find the top_waiter and determine if there are additional waiters.
1434 * If the caller intends to requeue more than 1 waiter to pifutex,
1435 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1436 * as we have means to handle the possible fault. If not, don't set
1437 * the bit unecessarily as it will force the subsequent unlock to enter
1440 top_waiter = futex_top_waiter(hb1, key1);
1442 /* There are no waiters, nothing for us to do. */
1446 /* Ensure we requeue to the expected futex. */
1447 if (!match_futex(top_waiter->requeue_pi_key, key2))
1451 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1452 * the contended case or if set_waiters is 1. The pi_state is returned
1453 * in ps in contended cases.
1455 vpid = task_pid_vnr(top_waiter->task);
1456 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1459 requeue_pi_wake_futex(top_waiter, key2, hb2);
1466 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1467 * @uaddr1: source futex user address
1468 * @flags: futex flags (FLAGS_SHARED, etc.)
1469 * @uaddr2: target futex user address
1470 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1471 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1472 * @cmpval: @uaddr1 expected value (or %NULL)
1473 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1474 * pi futex (pi to pi requeue is not supported)
1476 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1477 * uaddr2 atomically on behalf of the top waiter.
1480 * >=0 - on success, the number of tasks requeued or woken;
1483 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1484 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1485 u32 *cmpval, int requeue_pi)
1487 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1488 int drop_count = 0, task_count = 0, ret;
1489 struct futex_pi_state *pi_state = NULL;
1490 struct futex_hash_bucket *hb1, *hb2;
1491 struct futex_q *this, *next;
1495 * Requeue PI only works on two distinct uaddrs. This
1496 * check is only valid for private futexes. See below.
1498 if (uaddr1 == uaddr2)
1502 * requeue_pi requires a pi_state, try to allocate it now
1503 * without any locks in case it fails.
1505 if (refill_pi_state_cache())
1508 * requeue_pi must wake as many tasks as it can, up to nr_wake
1509 * + nr_requeue, since it acquires the rt_mutex prior to
1510 * returning to userspace, so as to not leave the rt_mutex with
1511 * waiters and no owner. However, second and third wake-ups
1512 * cannot be predicted as they involve race conditions with the
1513 * first wake and a fault while looking up the pi_state. Both
1514 * pthread_cond_signal() and pthread_cond_broadcast() should
1522 if (pi_state != NULL) {
1524 * We will have to lookup the pi_state again, so free this one
1525 * to keep the accounting correct.
1527 free_pi_state(pi_state);
1531 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1532 if (unlikely(ret != 0))
1534 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1535 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1536 if (unlikely(ret != 0))
1540 * The check above which compares uaddrs is not sufficient for
1541 * shared futexes. We need to compare the keys:
1543 if (requeue_pi && match_futex(&key1, &key2)) {
1548 hb1 = hash_futex(&key1);
1549 hb2 = hash_futex(&key2);
1552 hb_waiters_inc(hb2);
1553 double_lock_hb(hb1, hb2);
1555 if (likely(cmpval != NULL)) {
1558 ret = get_futex_value_locked(&curval, uaddr1);
1560 if (unlikely(ret)) {
1561 double_unlock_hb(hb1, hb2);
1562 hb_waiters_dec(hb2);
1564 ret = get_user(curval, uaddr1);
1568 if (!(flags & FLAGS_SHARED))
1571 put_futex_key(&key2);
1572 put_futex_key(&key1);
1575 if (curval != *cmpval) {
1581 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1583 * Attempt to acquire uaddr2 and wake the top waiter. If we
1584 * intend to requeue waiters, force setting the FUTEX_WAITERS
1585 * bit. We force this here where we are able to easily handle
1586 * faults rather in the requeue loop below.
1588 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1589 &key2, &pi_state, nr_requeue);
1592 * At this point the top_waiter has either taken uaddr2 or is
1593 * waiting on it. If the former, then the pi_state will not
1594 * exist yet, look it up one more time to ensure we have a
1595 * reference to it. If the lock was taken, ret contains the
1596 * vpid of the top waiter task.
1603 * If we acquired the lock, then the user
1604 * space value of uaddr2 should be vpid. It
1605 * cannot be changed by the top waiter as it
1606 * is blocked on hb2 lock if it tries to do
1607 * so. If something fiddled with it behind our
1608 * back the pi state lookup might unearth
1609 * it. So we rather use the known value than
1610 * rereading and handing potential crap to
1613 ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1620 double_unlock_hb(hb1, hb2);
1621 hb_waiters_dec(hb2);
1622 put_futex_key(&key2);
1623 put_futex_key(&key1);
1624 ret = fault_in_user_writeable(uaddr2);
1630 * Two reasons for this:
1631 * - Owner is exiting and we just wait for the
1633 * - The user space value changed.
1635 double_unlock_hb(hb1, hb2);
1636 hb_waiters_dec(hb2);
1637 put_futex_key(&key2);
1638 put_futex_key(&key1);
1646 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1647 if (task_count - nr_wake >= nr_requeue)
1650 if (!match_futex(&this->key, &key1))
1654 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1655 * be paired with each other and no other futex ops.
1657 * We should never be requeueing a futex_q with a pi_state,
1658 * which is awaiting a futex_unlock_pi().
1660 if ((requeue_pi && !this->rt_waiter) ||
1661 (!requeue_pi && this->rt_waiter) ||
1668 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1669 * lock, we already woke the top_waiter. If not, it will be
1670 * woken by futex_unlock_pi().
1672 if (++task_count <= nr_wake && !requeue_pi) {
1677 /* Ensure we requeue to the expected futex for requeue_pi. */
1678 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1684 * Requeue nr_requeue waiters and possibly one more in the case
1685 * of requeue_pi if we couldn't acquire the lock atomically.
1688 /* Prepare the waiter to take the rt_mutex. */
1689 atomic_inc(&pi_state->refcount);
1690 this->pi_state = pi_state;
1691 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1695 /* We got the lock. */
1696 requeue_pi_wake_futex(this, &key2, hb2);
1701 this->pi_state = NULL;
1702 free_pi_state(pi_state);
1706 requeue_futex(this, hb1, hb2, &key2);
1711 double_unlock_hb(hb1, hb2);
1712 hb_waiters_dec(hb2);
1715 * drop_futex_key_refs() must be called outside the spinlocks. During
1716 * the requeue we moved futex_q's from the hash bucket at key1 to the
1717 * one at key2 and updated their key pointer. We no longer need to
1718 * hold the references to key1.
1720 while (--drop_count >= 0)
1721 drop_futex_key_refs(&key1);
1724 put_futex_key(&key2);
1726 put_futex_key(&key1);
1728 if (pi_state != NULL)
1729 free_pi_state(pi_state);
1730 return ret ? ret : task_count;
1733 /* The key must be already stored in q->key. */
1734 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1735 __acquires(&hb->lock)
1737 struct futex_hash_bucket *hb;
1739 hb = hash_futex(&q->key);
1742 * Increment the counter before taking the lock so that
1743 * a potential waker won't miss a to-be-slept task that is
1744 * waiting for the spinlock. This is safe as all queue_lock()
1745 * users end up calling queue_me(). Similarly, for housekeeping,
1746 * decrement the counter at queue_unlock() when some error has
1747 * occurred and we don't end up adding the task to the list.
1751 q->lock_ptr = &hb->lock;
1753 spin_lock(&hb->lock); /* implies MB (A) */
1758 queue_unlock(struct futex_hash_bucket *hb)
1759 __releases(&hb->lock)
1761 spin_unlock(&hb->lock);
1766 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1767 * @q: The futex_q to enqueue
1768 * @hb: The destination hash bucket
1770 * The hb->lock must be held by the caller, and is released here. A call to
1771 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1772 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1773 * or nothing if the unqueue is done as part of the wake process and the unqueue
1774 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1777 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1778 __releases(&hb->lock)
1783 * The priority used to register this element is
1784 * - either the real thread-priority for the real-time threads
1785 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1786 * - or MAX_RT_PRIO for non-RT threads.
1787 * Thus, all RT-threads are woken first in priority order, and
1788 * the others are woken last, in FIFO order.
1790 prio = min(current->normal_prio, MAX_RT_PRIO);
1792 plist_node_init(&q->list, prio);
1793 plist_add(&q->list, &hb->chain);
1795 spin_unlock(&hb->lock);
1799 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1800 * @q: The futex_q to unqueue
1802 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1803 * be paired with exactly one earlier call to queue_me().
1806 * 1 - if the futex_q was still queued (and we removed unqueued it);
1807 * 0 - if the futex_q was already removed by the waking thread
1809 static int unqueue_me(struct futex_q *q)
1811 spinlock_t *lock_ptr;
1814 /* In the common case we don't take the spinlock, which is nice. */
1816 lock_ptr = q->lock_ptr;
1818 if (lock_ptr != NULL) {
1819 spin_lock(lock_ptr);
1821 * q->lock_ptr can change between reading it and
1822 * spin_lock(), causing us to take the wrong lock. This
1823 * corrects the race condition.
1825 * Reasoning goes like this: if we have the wrong lock,
1826 * q->lock_ptr must have changed (maybe several times)
1827 * between reading it and the spin_lock(). It can
1828 * change again after the spin_lock() but only if it was
1829 * already changed before the spin_lock(). It cannot,
1830 * however, change back to the original value. Therefore
1831 * we can detect whether we acquired the correct lock.
1833 if (unlikely(lock_ptr != q->lock_ptr)) {
1834 spin_unlock(lock_ptr);
1839 BUG_ON(q->pi_state);
1841 spin_unlock(lock_ptr);
1845 drop_futex_key_refs(&q->key);
1850 * PI futexes can not be requeued and must remove themself from the
1851 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1854 static void unqueue_me_pi(struct futex_q *q)
1855 __releases(q->lock_ptr)
1859 BUG_ON(!q->pi_state);
1860 free_pi_state(q->pi_state);
1863 spin_unlock(q->lock_ptr);
1867 * Fixup the pi_state owner with the new owner.
1869 * Must be called with hash bucket lock held and mm->sem held for non
1872 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1873 struct task_struct *newowner)
1875 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1876 struct futex_pi_state *pi_state = q->pi_state;
1877 struct task_struct *oldowner = pi_state->owner;
1878 u32 uval, uninitialized_var(curval), newval;
1882 if (!pi_state->owner)
1883 newtid |= FUTEX_OWNER_DIED;
1886 * We are here either because we stole the rtmutex from the
1887 * previous highest priority waiter or we are the highest priority
1888 * waiter but failed to get the rtmutex the first time.
1889 * We have to replace the newowner TID in the user space variable.
1890 * This must be atomic as we have to preserve the owner died bit here.
1892 * Note: We write the user space value _before_ changing the pi_state
1893 * because we can fault here. Imagine swapped out pages or a fork
1894 * that marked all the anonymous memory readonly for cow.
1896 * Modifying pi_state _before_ the user space value would
1897 * leave the pi_state in an inconsistent state when we fault
1898 * here, because we need to drop the hash bucket lock to
1899 * handle the fault. This might be observed in the PID check
1900 * in lookup_pi_state.
1903 if (get_futex_value_locked(&uval, uaddr))
1907 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1909 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1917 * We fixed up user space. Now we need to fix the pi_state
1920 if (pi_state->owner != NULL) {
1921 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1922 WARN_ON(list_empty(&pi_state->list));
1923 list_del_init(&pi_state->list);
1924 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1927 pi_state->owner = newowner;
1929 raw_spin_lock_irq(&newowner->pi_lock);
1930 WARN_ON(!list_empty(&pi_state->list));
1931 list_add(&pi_state->list, &newowner->pi_state_list);
1932 raw_spin_unlock_irq(&newowner->pi_lock);
1936 * To handle the page fault we need to drop the hash bucket
1937 * lock here. That gives the other task (either the highest priority
1938 * waiter itself or the task which stole the rtmutex) the
1939 * chance to try the fixup of the pi_state. So once we are
1940 * back from handling the fault we need to check the pi_state
1941 * after reacquiring the hash bucket lock and before trying to
1942 * do another fixup. When the fixup has been done already we
1946 spin_unlock(q->lock_ptr);
1948 ret = fault_in_user_writeable(uaddr);
1950 spin_lock(q->lock_ptr);
1953 * Check if someone else fixed it for us:
1955 if (pi_state->owner != oldowner)
1964 static long futex_wait_restart(struct restart_block *restart);
1967 * fixup_owner() - Post lock pi_state and corner case management
1968 * @uaddr: user address of the futex
1969 * @q: futex_q (contains pi_state and access to the rt_mutex)
1970 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1972 * After attempting to lock an rt_mutex, this function is called to cleanup
1973 * the pi_state owner as well as handle race conditions that may allow us to
1974 * acquire the lock. Must be called with the hb lock held.
1977 * 1 - success, lock taken;
1978 * 0 - success, lock not taken;
1979 * <0 - on error (-EFAULT)
1981 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1983 struct task_struct *owner;
1988 * Got the lock. We might not be the anticipated owner if we
1989 * did a lock-steal - fix up the PI-state in that case:
1991 if (q->pi_state->owner != current)
1992 ret = fixup_pi_state_owner(uaddr, q, current);
1997 * Catch the rare case, where the lock was released when we were on the
1998 * way back before we locked the hash bucket.
2000 if (q->pi_state->owner == current) {
2002 * Try to get the rt_mutex now. This might fail as some other
2003 * task acquired the rt_mutex after we removed ourself from the
2004 * rt_mutex waiters list.
2006 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2012 * pi_state is incorrect, some other task did a lock steal and
2013 * we returned due to timeout or signal without taking the
2014 * rt_mutex. Too late.
2016 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2017 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2019 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2020 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2021 ret = fixup_pi_state_owner(uaddr, q, owner);
2026 * Paranoia check. If we did not take the lock, then we should not be
2027 * the owner of the rt_mutex.
2029 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2030 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2031 "pi-state %p\n", ret,
2032 q->pi_state->pi_mutex.owner,
2033 q->pi_state->owner);
2036 return ret ? ret : locked;
2040 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2041 * @hb: the futex hash bucket, must be locked by the caller
2042 * @q: the futex_q to queue up on
2043 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2045 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2046 struct hrtimer_sleeper *timeout)
2049 * The task state is guaranteed to be set before another task can
2050 * wake it. set_current_state() is implemented using set_mb() and
2051 * queue_me() calls spin_unlock() upon completion, both serializing
2052 * access to the hash list and forcing another memory barrier.
2054 set_current_state(TASK_INTERRUPTIBLE);
2059 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2060 if (!hrtimer_active(&timeout->timer))
2061 timeout->task = NULL;
2065 * If we have been removed from the hash list, then another task
2066 * has tried to wake us, and we can skip the call to schedule().
2068 if (likely(!plist_node_empty(&q->list))) {
2070 * If the timer has already expired, current will already be
2071 * flagged for rescheduling. Only call schedule if there
2072 * is no timeout, or if it has yet to expire.
2074 if (!timeout || timeout->task)
2075 freezable_schedule();
2077 __set_current_state(TASK_RUNNING);
2081 * futex_wait_setup() - Prepare to wait on a futex
2082 * @uaddr: the futex userspace address
2083 * @val: the expected value
2084 * @flags: futex flags (FLAGS_SHARED, etc.)
2085 * @q: the associated futex_q
2086 * @hb: storage for hash_bucket pointer to be returned to caller
2088 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2089 * compare it with the expected value. Handle atomic faults internally.
2090 * Return with the hb lock held and a q.key reference on success, and unlocked
2091 * with no q.key reference on failure.
2094 * 0 - uaddr contains val and hb has been locked;
2095 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2097 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2098 struct futex_q *q, struct futex_hash_bucket **hb)
2104 * Access the page AFTER the hash-bucket is locked.
2105 * Order is important:
2107 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2108 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2110 * The basic logical guarantee of a futex is that it blocks ONLY
2111 * if cond(var) is known to be true at the time of blocking, for
2112 * any cond. If we locked the hash-bucket after testing *uaddr, that
2113 * would open a race condition where we could block indefinitely with
2114 * cond(var) false, which would violate the guarantee.
2116 * On the other hand, we insert q and release the hash-bucket only
2117 * after testing *uaddr. This guarantees that futex_wait() will NOT
2118 * absorb a wakeup if *uaddr does not match the desired values
2119 * while the syscall executes.
2122 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2123 if (unlikely(ret != 0))
2127 *hb = queue_lock(q);
2129 ret = get_futex_value_locked(&uval, uaddr);
2134 ret = get_user(uval, uaddr);
2138 if (!(flags & FLAGS_SHARED))
2141 put_futex_key(&q->key);
2152 put_futex_key(&q->key);
2156 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2157 ktime_t *abs_time, u32 bitset)
2159 struct hrtimer_sleeper timeout, *to = NULL;
2160 struct restart_block *restart;
2161 struct futex_hash_bucket *hb;
2162 struct futex_q q = futex_q_init;
2172 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2173 CLOCK_REALTIME : CLOCK_MONOTONIC,
2175 hrtimer_init_sleeper(to, current);
2176 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2177 current->timer_slack_ns);
2182 * Prepare to wait on uaddr. On success, holds hb lock and increments
2185 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2189 /* queue_me and wait for wakeup, timeout, or a signal. */
2190 futex_wait_queue_me(hb, &q, to);
2192 /* If we were woken (and unqueued), we succeeded, whatever. */
2194 /* unqueue_me() drops q.key ref */
2195 if (!unqueue_me(&q))
2198 if (to && !to->task)
2202 * We expect signal_pending(current), but we might be the
2203 * victim of a spurious wakeup as well.
2205 if (!signal_pending(current))
2212 restart = ¤t_thread_info()->restart_block;
2213 restart->fn = futex_wait_restart;
2214 restart->futex.uaddr = uaddr;
2215 restart->futex.val = val;
2216 restart->futex.time = abs_time->tv64;
2217 restart->futex.bitset = bitset;
2218 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2220 ret = -ERESTART_RESTARTBLOCK;
2224 hrtimer_cancel(&to->timer);
2225 destroy_hrtimer_on_stack(&to->timer);
2231 static long futex_wait_restart(struct restart_block *restart)
2233 u32 __user *uaddr = restart->futex.uaddr;
2234 ktime_t t, *tp = NULL;
2236 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2237 t.tv64 = restart->futex.time;
2240 restart->fn = do_no_restart_syscall;
2242 return (long)futex_wait(uaddr, restart->futex.flags,
2243 restart->futex.val, tp, restart->futex.bitset);
2248 * Userspace tried a 0 -> TID atomic transition of the futex value
2249 * and failed. The kernel side here does the whole locking operation:
2250 * if there are waiters then it will block, it does PI, etc. (Due to
2251 * races the kernel might see a 0 value of the futex too.)
2253 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2254 ktime_t *time, int trylock)
2256 struct hrtimer_sleeper timeout, *to = NULL;
2257 struct futex_hash_bucket *hb;
2258 struct futex_q q = futex_q_init;
2261 if (refill_pi_state_cache())
2266 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2268 hrtimer_init_sleeper(to, current);
2269 hrtimer_set_expires(&to->timer, *time);
2273 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2274 if (unlikely(ret != 0))
2278 hb = queue_lock(&q);
2280 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2281 if (unlikely(ret)) {
2284 /* We got the lock. */
2286 goto out_unlock_put_key;
2291 * Two reasons for this:
2292 * - Task is exiting and we just wait for the
2294 * - The user space value changed.
2297 put_futex_key(&q.key);
2301 goto out_unlock_put_key;
2306 * Only actually queue now that the atomic ops are done:
2310 WARN_ON(!q.pi_state);
2312 * Block on the PI mutex:
2315 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2317 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2318 /* Fixup the trylock return value: */
2319 ret = ret ? 0 : -EWOULDBLOCK;
2322 spin_lock(q.lock_ptr);
2324 * Fixup the pi_state owner and possibly acquire the lock if we
2327 res = fixup_owner(uaddr, &q, !ret);
2329 * If fixup_owner() returned an error, proprogate that. If it acquired
2330 * the lock, clear our -ETIMEDOUT or -EINTR.
2333 ret = (res < 0) ? res : 0;
2336 * If fixup_owner() faulted and was unable to handle the fault, unlock
2337 * it and return the fault to userspace.
2339 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2340 rt_mutex_unlock(&q.pi_state->pi_mutex);
2342 /* Unqueue and drop the lock */
2351 put_futex_key(&q.key);
2354 destroy_hrtimer_on_stack(&to->timer);
2355 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2360 ret = fault_in_user_writeable(uaddr);
2364 if (!(flags & FLAGS_SHARED))
2367 put_futex_key(&q.key);
2372 * Userspace attempted a TID -> 0 atomic transition, and failed.
2373 * This is the in-kernel slowpath: we look up the PI state (if any),
2374 * and do the rt-mutex unlock.
2376 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2378 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2379 union futex_key key = FUTEX_KEY_INIT;
2380 struct futex_hash_bucket *hb;
2381 struct futex_q *match;
2385 if (get_user(uval, uaddr))
2388 * We release only a lock we actually own:
2390 if ((uval & FUTEX_TID_MASK) != vpid)
2393 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2397 hb = hash_futex(&key);
2398 spin_lock(&hb->lock);
2401 * Check waiters first. We do not trust user space values at
2402 * all and we at least want to know if user space fiddled
2403 * with the futex value instead of blindly unlocking.
2405 match = futex_top_waiter(hb, &key);
2407 ret = wake_futex_pi(uaddr, uval, match);
2409 * The atomic access to the futex value generated a
2410 * pagefault, so retry the user-access and the wakeup:
2418 * We have no kernel internal state, i.e. no waiters in the
2419 * kernel. Waiters which are about to queue themselves are stuck
2420 * on hb->lock. So we can safely ignore them. We do neither
2421 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2424 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2428 * If uval has changed, let user space handle it.
2430 ret = (curval == uval) ? 0 : -EAGAIN;
2433 spin_unlock(&hb->lock);
2434 put_futex_key(&key);
2438 spin_unlock(&hb->lock);
2439 put_futex_key(&key);
2441 ret = fault_in_user_writeable(uaddr);
2449 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2450 * @hb: the hash_bucket futex_q was original enqueued on
2451 * @q: the futex_q woken while waiting to be requeued
2452 * @key2: the futex_key of the requeue target futex
2453 * @timeout: the timeout associated with the wait (NULL if none)
2455 * Detect if the task was woken on the initial futex as opposed to the requeue
2456 * target futex. If so, determine if it was a timeout or a signal that caused
2457 * the wakeup and return the appropriate error code to the caller. Must be
2458 * called with the hb lock held.
2461 * 0 = no early wakeup detected;
2462 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2465 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2466 struct futex_q *q, union futex_key *key2,
2467 struct hrtimer_sleeper *timeout)
2472 * With the hb lock held, we avoid races while we process the wakeup.
2473 * We only need to hold hb (and not hb2) to ensure atomicity as the
2474 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2475 * It can't be requeued from uaddr2 to something else since we don't
2476 * support a PI aware source futex for requeue.
2478 if (!match_futex(&q->key, key2)) {
2479 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2481 * We were woken prior to requeue by a timeout or a signal.
2482 * Unqueue the futex_q and determine which it was.
2484 plist_del(&q->list, &hb->chain);
2487 /* Handle spurious wakeups gracefully */
2489 if (timeout && !timeout->task)
2491 else if (signal_pending(current))
2492 ret = -ERESTARTNOINTR;
2498 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2499 * @uaddr: the futex we initially wait on (non-pi)
2500 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2501 * the same type, no requeueing from private to shared, etc.
2502 * @val: the expected value of uaddr
2503 * @abs_time: absolute timeout
2504 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2505 * @uaddr2: the pi futex we will take prior to returning to user-space
2507 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2508 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2509 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2510 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2511 * without one, the pi logic would not know which task to boost/deboost, if
2512 * there was a need to.
2514 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2515 * via the following--
2516 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2517 * 2) wakeup on uaddr2 after a requeue
2521 * If 3, cleanup and return -ERESTARTNOINTR.
2523 * If 2, we may then block on trying to take the rt_mutex and return via:
2524 * 5) successful lock
2527 * 8) other lock acquisition failure
2529 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2531 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2537 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2538 u32 val, ktime_t *abs_time, u32 bitset,
2541 struct hrtimer_sleeper timeout, *to = NULL;
2542 struct rt_mutex_waiter rt_waiter;
2543 struct rt_mutex *pi_mutex = NULL;
2544 struct futex_hash_bucket *hb;
2545 union futex_key key2 = FUTEX_KEY_INIT;
2546 struct futex_q q = futex_q_init;
2549 if (uaddr == uaddr2)
2557 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2558 CLOCK_REALTIME : CLOCK_MONOTONIC,
2560 hrtimer_init_sleeper(to, current);
2561 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2562 current->timer_slack_ns);
2566 * The waiter is allocated on our stack, manipulated by the requeue
2567 * code while we sleep on uaddr.
2569 debug_rt_mutex_init_waiter(&rt_waiter);
2570 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2571 RB_CLEAR_NODE(&rt_waiter.tree_entry);
2572 rt_waiter.task = NULL;
2574 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2575 if (unlikely(ret != 0))
2579 q.rt_waiter = &rt_waiter;
2580 q.requeue_pi_key = &key2;
2583 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2586 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2591 * The check above which compares uaddrs is not sufficient for
2592 * shared futexes. We need to compare the keys:
2594 if (match_futex(&q.key, &key2)) {
2599 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2600 futex_wait_queue_me(hb, &q, to);
2602 spin_lock(&hb->lock);
2603 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2604 spin_unlock(&hb->lock);
2609 * In order for us to be here, we know our q.key == key2, and since
2610 * we took the hb->lock above, we also know that futex_requeue() has
2611 * completed and we no longer have to concern ourselves with a wakeup
2612 * race with the atomic proxy lock acquisition by the requeue code. The
2613 * futex_requeue dropped our key1 reference and incremented our key2
2617 /* Check if the requeue code acquired the second futex for us. */
2620 * Got the lock. We might not be the anticipated owner if we
2621 * did a lock-steal - fix up the PI-state in that case.
2623 if (q.pi_state && (q.pi_state->owner != current)) {
2624 spin_lock(q.lock_ptr);
2625 ret = fixup_pi_state_owner(uaddr2, &q, current);
2626 spin_unlock(q.lock_ptr);
2630 * We have been woken up by futex_unlock_pi(), a timeout, or a
2631 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2634 WARN_ON(!q.pi_state);
2635 pi_mutex = &q.pi_state->pi_mutex;
2636 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2637 debug_rt_mutex_free_waiter(&rt_waiter);
2639 spin_lock(q.lock_ptr);
2641 * Fixup the pi_state owner and possibly acquire the lock if we
2644 res = fixup_owner(uaddr2, &q, !ret);
2646 * If fixup_owner() returned an error, proprogate that. If it
2647 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2650 ret = (res < 0) ? res : 0;
2652 /* Unqueue and drop the lock. */
2657 * If fixup_pi_state_owner() faulted and was unable to handle the
2658 * fault, unlock the rt_mutex and return the fault to userspace.
2660 if (ret == -EFAULT) {
2661 if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2662 rt_mutex_unlock(pi_mutex);
2663 } else if (ret == -EINTR) {
2665 * We've already been requeued, but cannot restart by calling
2666 * futex_lock_pi() directly. We could restart this syscall, but
2667 * it would detect that the user space "val" changed and return
2668 * -EWOULDBLOCK. Save the overhead of the restart and return
2669 * -EWOULDBLOCK directly.
2675 put_futex_key(&q.key);
2677 put_futex_key(&key2);
2681 hrtimer_cancel(&to->timer);
2682 destroy_hrtimer_on_stack(&to->timer);
2688 * Support for robust futexes: the kernel cleans up held futexes at
2691 * Implementation: user-space maintains a per-thread list of locks it
2692 * is holding. Upon do_exit(), the kernel carefully walks this list,
2693 * and marks all locks that are owned by this thread with the
2694 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2695 * always manipulated with the lock held, so the list is private and
2696 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2697 * field, to allow the kernel to clean up if the thread dies after
2698 * acquiring the lock, but just before it could have added itself to
2699 * the list. There can only be one such pending lock.
2703 * sys_set_robust_list() - Set the robust-futex list head of a task
2704 * @head: pointer to the list-head
2705 * @len: length of the list-head, as userspace expects
2707 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2710 if (!futex_cmpxchg_enabled)
2713 * The kernel knows only one size for now:
2715 if (unlikely(len != sizeof(*head)))
2718 current->robust_list = head;
2724 * sys_get_robust_list() - Get the robust-futex list head of a task
2725 * @pid: pid of the process [zero for current task]
2726 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2727 * @len_ptr: pointer to a length field, the kernel fills in the header size
2729 SYSCALL_DEFINE3(get_robust_list, int, pid,
2730 struct robust_list_head __user * __user *, head_ptr,
2731 size_t __user *, len_ptr)
2733 struct robust_list_head __user *head;
2735 struct task_struct *p;
2737 if (!futex_cmpxchg_enabled)
2746 p = find_task_by_vpid(pid);
2752 if (!ptrace_may_access(p, PTRACE_MODE_READ))
2755 head = p->robust_list;
2758 if (put_user(sizeof(*head), len_ptr))
2760 return put_user(head, head_ptr);
2769 * Process a futex-list entry, check whether it's owned by the
2770 * dying task, and do notification if so:
2772 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2774 u32 uval, uninitialized_var(nval), mval;
2777 if (get_user(uval, uaddr))
2780 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2782 * Ok, this dying thread is truly holding a futex
2783 * of interest. Set the OWNER_DIED bit atomically
2784 * via cmpxchg, and if the value had FUTEX_WAITERS
2785 * set, wake up a waiter (if any). (We have to do a
2786 * futex_wake() even if OWNER_DIED is already set -
2787 * to handle the rare but possible case of recursive
2788 * thread-death.) The rest of the cleanup is done in
2791 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2793 * We are not holding a lock here, but we want to have
2794 * the pagefault_disable/enable() protection because
2795 * we want to handle the fault gracefully. If the
2796 * access fails we try to fault in the futex with R/W
2797 * verification via get_user_pages. get_user() above
2798 * does not guarantee R/W access. If that fails we
2799 * give up and leave the futex locked.
2801 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2802 if (fault_in_user_writeable(uaddr))
2810 * Wake robust non-PI futexes here. The wakeup of
2811 * PI futexes happens in exit_pi_state():
2813 if (!pi && (uval & FUTEX_WAITERS))
2814 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2820 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2822 static inline int fetch_robust_entry(struct robust_list __user **entry,
2823 struct robust_list __user * __user *head,
2826 unsigned long uentry;
2828 if (get_user(uentry, (unsigned long __user *)head))
2831 *entry = (void __user *)(uentry & ~1UL);
2838 * Walk curr->robust_list (very carefully, it's a userspace list!)
2839 * and mark any locks found there dead, and notify any waiters.
2841 * We silently return on any sign of list-walking problem.
2843 void exit_robust_list(struct task_struct *curr)
2845 struct robust_list_head __user *head = curr->robust_list;
2846 struct robust_list __user *entry, *next_entry, *pending;
2847 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2848 unsigned int uninitialized_var(next_pi);
2849 unsigned long futex_offset;
2852 if (!futex_cmpxchg_enabled)
2856 * Fetch the list head (which was registered earlier, via
2857 * sys_set_robust_list()):
2859 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2862 * Fetch the relative futex offset:
2864 if (get_user(futex_offset, &head->futex_offset))
2867 * Fetch any possibly pending lock-add first, and handle it
2870 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2873 next_entry = NULL; /* avoid warning with gcc */
2874 while (entry != &head->list) {
2876 * Fetch the next entry in the list before calling
2877 * handle_futex_death:
2879 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2881 * A pending lock might already be on the list, so
2882 * don't process it twice:
2884 if (entry != pending)
2885 if (handle_futex_death((void __user *)entry + futex_offset,
2893 * Avoid excessively long or circular lists:
2902 handle_futex_death((void __user *)pending + futex_offset,
2906 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2907 u32 __user *uaddr2, u32 val2, u32 val3)
2909 int cmd = op & FUTEX_CMD_MASK;
2910 unsigned int flags = 0;
2912 if (!(op & FUTEX_PRIVATE_FLAG))
2913 flags |= FLAGS_SHARED;
2915 if (op & FUTEX_CLOCK_REALTIME) {
2916 flags |= FLAGS_CLOCKRT;
2917 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2923 case FUTEX_UNLOCK_PI:
2924 case FUTEX_TRYLOCK_PI:
2925 case FUTEX_WAIT_REQUEUE_PI:
2926 case FUTEX_CMP_REQUEUE_PI:
2927 if (!futex_cmpxchg_enabled)
2933 val3 = FUTEX_BITSET_MATCH_ANY;
2934 case FUTEX_WAIT_BITSET:
2935 return futex_wait(uaddr, flags, val, timeout, val3);
2937 val3 = FUTEX_BITSET_MATCH_ANY;
2938 case FUTEX_WAKE_BITSET:
2939 return futex_wake(uaddr, flags, val, val3);
2941 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2942 case FUTEX_CMP_REQUEUE:
2943 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2945 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2947 return futex_lock_pi(uaddr, flags, val, timeout, 0);
2948 case FUTEX_UNLOCK_PI:
2949 return futex_unlock_pi(uaddr, flags);
2950 case FUTEX_TRYLOCK_PI:
2951 return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2952 case FUTEX_WAIT_REQUEUE_PI:
2953 val3 = FUTEX_BITSET_MATCH_ANY;
2954 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2956 case FUTEX_CMP_REQUEUE_PI:
2957 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2963 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2964 struct timespec __user *, utime, u32 __user *, uaddr2,
2968 ktime_t t, *tp = NULL;
2970 int cmd = op & FUTEX_CMD_MASK;
2972 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2973 cmd == FUTEX_WAIT_BITSET ||
2974 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2975 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2977 if (!timespec_valid(&ts))
2980 t = timespec_to_ktime(ts);
2981 if (cmd == FUTEX_WAIT)
2982 t = ktime_add_safe(ktime_get(), t);
2986 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2987 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2989 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2990 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2991 val2 = (u32) (unsigned long) utime;
2993 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2996 static void __init futex_detect_cmpxchg(void)
2998 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3002 * This will fail and we want it. Some arch implementations do
3003 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3004 * functionality. We want to know that before we call in any
3005 * of the complex code paths. Also we want to prevent
3006 * registration of robust lists in that case. NULL is
3007 * guaranteed to fault and we get -EFAULT on functional
3008 * implementation, the non-functional ones will return
3011 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3012 futex_cmpxchg_enabled = 1;
3016 static int __init futex_init(void)
3018 unsigned int futex_shift;
3021 #if CONFIG_BASE_SMALL
3022 futex_hashsize = 16;
3024 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3027 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3029 futex_hashsize < 256 ? HASH_SMALL : 0,
3031 futex_hashsize, futex_hashsize);
3032 futex_hashsize = 1UL << futex_shift;
3034 futex_detect_cmpxchg();
3036 for (i = 0; i < futex_hashsize; i++) {
3037 atomic_set(&futex_queues[i].waiters, 0);
3038 plist_head_init(&futex_queues[i].chain);
3039 spin_lock_init(&futex_queues[i].lock);
3044 __initcall(futex_init);