Use dentry_path() to create full path to inode object
[pohmelfs.git] / kernel / futex.c
blob1614be20173dcb19ef45e65fa25d07f8d3c65194
1 /*
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>
49 #include <linux/fs.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>
63 #include <asm/futex.h>
65 #include "rtmutex_common.h"
67 int __read_mostly futex_cmpxchg_enabled;
69 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
72 * Futex flags used to encode options to functions and preserve them across
73 * restarts.
75 #define FLAGS_SHARED 0x01
76 #define FLAGS_CLOCKRT 0x02
77 #define FLAGS_HAS_TIMEOUT 0x04
80 * Priority Inheritance state:
82 struct futex_pi_state {
84 * list of 'owned' pi_state instances - these have to be
85 * cleaned up in do_exit() if the task exits prematurely:
87 struct list_head list;
90 * The PI object:
92 struct rt_mutex pi_mutex;
94 struct task_struct *owner;
95 atomic_t refcount;
97 union futex_key key;
101 * struct futex_q - The hashed futex queue entry, one per waiting task
102 * @list: priority-sorted list of tasks waiting on this futex
103 * @task: the task waiting on the futex
104 * @lock_ptr: the hash bucket lock
105 * @key: the key the futex is hashed on
106 * @pi_state: optional priority inheritance state
107 * @rt_waiter: rt_waiter storage for use with requeue_pi
108 * @requeue_pi_key: the requeue_pi target futex key
109 * @bitset: bitset for the optional bitmasked wakeup
111 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
112 * we can wake only the relevant ones (hashed queues may be shared).
114 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
115 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
116 * The order of wakeup is always to make the first condition true, then
117 * the second.
119 * PI futexes are typically woken before they are removed from the hash list via
120 * the rt_mutex code. See unqueue_me_pi().
122 struct futex_q {
123 struct plist_node list;
125 struct task_struct *task;
126 spinlock_t *lock_ptr;
127 union futex_key key;
128 struct futex_pi_state *pi_state;
129 struct rt_mutex_waiter *rt_waiter;
130 union futex_key *requeue_pi_key;
131 u32 bitset;
134 static const struct futex_q futex_q_init = {
135 /* list gets initialized in queue_me()*/
136 .key = FUTEX_KEY_INIT,
137 .bitset = FUTEX_BITSET_MATCH_ANY
141 * Hash buckets are shared by all the futex_keys that hash to the same
142 * location. Each key may have multiple futex_q structures, one for each task
143 * waiting on a futex.
145 struct futex_hash_bucket {
146 spinlock_t lock;
147 struct plist_head chain;
150 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
153 * We hash on the keys returned from get_futex_key (see below).
155 static struct futex_hash_bucket *hash_futex(union futex_key *key)
157 u32 hash = jhash2((u32*)&key->both.word,
158 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
159 key->both.offset);
160 return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
164 * Return 1 if two futex_keys are equal, 0 otherwise.
166 static inline int match_futex(union futex_key *key1, union futex_key *key2)
168 return (key1 && key2
169 && key1->both.word == key2->both.word
170 && key1->both.ptr == key2->both.ptr
171 && key1->both.offset == key2->both.offset);
175 * Take a reference to the resource addressed by a key.
176 * Can be called while holding spinlocks.
179 static void get_futex_key_refs(union futex_key *key)
181 if (!key->both.ptr)
182 return;
184 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
185 case FUT_OFF_INODE:
186 ihold(key->shared.inode);
187 break;
188 case FUT_OFF_MMSHARED:
189 atomic_inc(&key->private.mm->mm_count);
190 break;
195 * Drop a reference to the resource addressed by a key.
196 * The hash bucket spinlock must not be held.
198 static void drop_futex_key_refs(union futex_key *key)
200 if (!key->both.ptr) {
201 /* If we're here then we tried to put a key we failed to get */
202 WARN_ON_ONCE(1);
203 return;
206 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
207 case FUT_OFF_INODE:
208 iput(key->shared.inode);
209 break;
210 case FUT_OFF_MMSHARED:
211 mmdrop(key->private.mm);
212 break;
217 * get_futex_key() - Get parameters which are the keys for a futex
218 * @uaddr: virtual address of the futex
219 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
220 * @key: address where result is stored.
221 * @rw: mapping needs to be read/write (values: VERIFY_READ,
222 * VERIFY_WRITE)
224 * Returns a negative error code or 0
225 * The key words are stored in *key on success.
227 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
228 * offset_within_page). For private mappings, it's (uaddr, current->mm).
229 * We can usually work out the index without swapping in the page.
231 * lock_page() might sleep, the caller should not hold a spinlock.
233 static int
234 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
236 unsigned long address = (unsigned long)uaddr;
237 struct mm_struct *mm = current->mm;
238 struct page *page, *page_head;
239 int err, ro = 0;
242 * The futex address must be "naturally" aligned.
244 key->both.offset = address % PAGE_SIZE;
245 if (unlikely((address % sizeof(u32)) != 0))
246 return -EINVAL;
247 address -= key->both.offset;
250 * PROCESS_PRIVATE futexes are fast.
251 * As the mm cannot disappear under us and the 'key' only needs
252 * virtual address, we dont even have to find the underlying vma.
253 * Note : We do have to check 'uaddr' is a valid user address,
254 * but access_ok() should be faster than find_vma()
256 if (!fshared) {
257 if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
258 return -EFAULT;
259 key->private.mm = mm;
260 key->private.address = address;
261 get_futex_key_refs(key);
262 return 0;
265 again:
266 err = get_user_pages_fast(address, 1, 1, &page);
268 * If write access is not required (eg. FUTEX_WAIT), try
269 * and get read-only access.
271 if (err == -EFAULT && rw == VERIFY_READ) {
272 err = get_user_pages_fast(address, 1, 0, &page);
273 ro = 1;
275 if (err < 0)
276 return err;
277 else
278 err = 0;
280 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
281 page_head = page;
282 if (unlikely(PageTail(page))) {
283 put_page(page);
284 /* serialize against __split_huge_page_splitting() */
285 local_irq_disable();
286 if (likely(__get_user_pages_fast(address, 1, 1, &page) == 1)) {
287 page_head = compound_head(page);
289 * page_head is valid pointer but we must pin
290 * it before taking the PG_lock and/or
291 * PG_compound_lock. The moment we re-enable
292 * irqs __split_huge_page_splitting() can
293 * return and the head page can be freed from
294 * under us. We can't take the PG_lock and/or
295 * PG_compound_lock on a page that could be
296 * freed from under us.
298 if (page != page_head) {
299 get_page(page_head);
300 put_page(page);
302 local_irq_enable();
303 } else {
304 local_irq_enable();
305 goto again;
308 #else
309 page_head = compound_head(page);
310 if (page != page_head) {
311 get_page(page_head);
312 put_page(page);
314 #endif
316 lock_page(page_head);
319 * If page_head->mapping is NULL, then it cannot be a PageAnon
320 * page; but it might be the ZERO_PAGE or in the gate area or
321 * in a special mapping (all cases which we are happy to fail);
322 * or it may have been a good file page when get_user_pages_fast
323 * found it, but truncated or holepunched or subjected to
324 * invalidate_complete_page2 before we got the page lock (also
325 * cases which we are happy to fail). And we hold a reference,
326 * so refcount care in invalidate_complete_page's remove_mapping
327 * prevents drop_caches from setting mapping to NULL beneath us.
329 * The case we do have to guard against is when memory pressure made
330 * shmem_writepage move it from filecache to swapcache beneath us:
331 * an unlikely race, but we do need to retry for page_head->mapping.
333 if (!page_head->mapping) {
334 int shmem_swizzled = PageSwapCache(page_head);
335 unlock_page(page_head);
336 put_page(page_head);
337 if (shmem_swizzled)
338 goto again;
339 return -EFAULT;
343 * Private mappings are handled in a simple way.
345 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
346 * it's a read-only handle, it's expected that futexes attach to
347 * the object not the particular process.
349 if (PageAnon(page_head)) {
351 * A RO anonymous page will never change and thus doesn't make
352 * sense for futex operations.
354 if (ro) {
355 err = -EFAULT;
356 goto out;
359 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
360 key->private.mm = mm;
361 key->private.address = address;
362 } else {
363 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
364 key->shared.inode = page_head->mapping->host;
365 key->shared.pgoff = page_head->index;
368 get_futex_key_refs(key);
370 out:
371 unlock_page(page_head);
372 put_page(page_head);
373 return err;
376 static inline void put_futex_key(union futex_key *key)
378 drop_futex_key_refs(key);
382 * fault_in_user_writeable() - Fault in user address and verify RW access
383 * @uaddr: pointer to faulting user space address
385 * Slow path to fixup the fault we just took in the atomic write
386 * access to @uaddr.
388 * We have no generic implementation of a non-destructive write to the
389 * user address. We know that we faulted in the atomic pagefault
390 * disabled section so we can as well avoid the #PF overhead by
391 * calling get_user_pages() right away.
393 static int fault_in_user_writeable(u32 __user *uaddr)
395 struct mm_struct *mm = current->mm;
396 int ret;
398 down_read(&mm->mmap_sem);
399 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
400 FAULT_FLAG_WRITE);
401 up_read(&mm->mmap_sem);
403 return ret < 0 ? ret : 0;
407 * futex_top_waiter() - Return the highest priority waiter on a futex
408 * @hb: the hash bucket the futex_q's reside in
409 * @key: the futex key (to distinguish it from other futex futex_q's)
411 * Must be called with the hb lock held.
413 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
414 union futex_key *key)
416 struct futex_q *this;
418 plist_for_each_entry(this, &hb->chain, list) {
419 if (match_futex(&this->key, key))
420 return this;
422 return NULL;
425 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
426 u32 uval, u32 newval)
428 int ret;
430 pagefault_disable();
431 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
432 pagefault_enable();
434 return ret;
437 static int get_futex_value_locked(u32 *dest, u32 __user *from)
439 int ret;
441 pagefault_disable();
442 ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
443 pagefault_enable();
445 return ret ? -EFAULT : 0;
450 * PI code:
452 static int refill_pi_state_cache(void)
454 struct futex_pi_state *pi_state;
456 if (likely(current->pi_state_cache))
457 return 0;
459 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
461 if (!pi_state)
462 return -ENOMEM;
464 INIT_LIST_HEAD(&pi_state->list);
465 /* pi_mutex gets initialized later */
466 pi_state->owner = NULL;
467 atomic_set(&pi_state->refcount, 1);
468 pi_state->key = FUTEX_KEY_INIT;
470 current->pi_state_cache = pi_state;
472 return 0;
475 static struct futex_pi_state * alloc_pi_state(void)
477 struct futex_pi_state *pi_state = current->pi_state_cache;
479 WARN_ON(!pi_state);
480 current->pi_state_cache = NULL;
482 return pi_state;
485 static void free_pi_state(struct futex_pi_state *pi_state)
487 if (!atomic_dec_and_test(&pi_state->refcount))
488 return;
491 * If pi_state->owner is NULL, the owner is most probably dying
492 * and has cleaned up the pi_state already
494 if (pi_state->owner) {
495 raw_spin_lock_irq(&pi_state->owner->pi_lock);
496 list_del_init(&pi_state->list);
497 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
499 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
502 if (current->pi_state_cache)
503 kfree(pi_state);
504 else {
506 * pi_state->list is already empty.
507 * clear pi_state->owner.
508 * refcount is at 0 - put it back to 1.
510 pi_state->owner = NULL;
511 atomic_set(&pi_state->refcount, 1);
512 current->pi_state_cache = pi_state;
517 * Look up the task based on what TID userspace gave us.
518 * We dont trust it.
520 static struct task_struct * futex_find_get_task(pid_t pid)
522 struct task_struct *p;
524 rcu_read_lock();
525 p = find_task_by_vpid(pid);
526 if (p)
527 get_task_struct(p);
529 rcu_read_unlock();
531 return p;
535 * This task is holding PI mutexes at exit time => bad.
536 * Kernel cleans up PI-state, but userspace is likely hosed.
537 * (Robust-futex cleanup is separate and might save the day for userspace.)
539 void exit_pi_state_list(struct task_struct *curr)
541 struct list_head *next, *head = &curr->pi_state_list;
542 struct futex_pi_state *pi_state;
543 struct futex_hash_bucket *hb;
544 union futex_key key = FUTEX_KEY_INIT;
546 if (!futex_cmpxchg_enabled)
547 return;
549 * We are a ZOMBIE and nobody can enqueue itself on
550 * pi_state_list anymore, but we have to be careful
551 * versus waiters unqueueing themselves:
553 raw_spin_lock_irq(&curr->pi_lock);
554 while (!list_empty(head)) {
556 next = head->next;
557 pi_state = list_entry(next, struct futex_pi_state, list);
558 key = pi_state->key;
559 hb = hash_futex(&key);
560 raw_spin_unlock_irq(&curr->pi_lock);
562 spin_lock(&hb->lock);
564 raw_spin_lock_irq(&curr->pi_lock);
566 * We dropped the pi-lock, so re-check whether this
567 * task still owns the PI-state:
569 if (head->next != next) {
570 spin_unlock(&hb->lock);
571 continue;
574 WARN_ON(pi_state->owner != curr);
575 WARN_ON(list_empty(&pi_state->list));
576 list_del_init(&pi_state->list);
577 pi_state->owner = NULL;
578 raw_spin_unlock_irq(&curr->pi_lock);
580 rt_mutex_unlock(&pi_state->pi_mutex);
582 spin_unlock(&hb->lock);
584 raw_spin_lock_irq(&curr->pi_lock);
586 raw_spin_unlock_irq(&curr->pi_lock);
589 static int
590 lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
591 union futex_key *key, struct futex_pi_state **ps)
593 struct futex_pi_state *pi_state = NULL;
594 struct futex_q *this, *next;
595 struct plist_head *head;
596 struct task_struct *p;
597 pid_t pid = uval & FUTEX_TID_MASK;
599 head = &hb->chain;
601 plist_for_each_entry_safe(this, next, head, list) {
602 if (match_futex(&this->key, key)) {
604 * Another waiter already exists - bump up
605 * the refcount and return its pi_state:
607 pi_state = this->pi_state;
609 * Userspace might have messed up non-PI and PI futexes
611 if (unlikely(!pi_state))
612 return -EINVAL;
614 WARN_ON(!atomic_read(&pi_state->refcount));
617 * When pi_state->owner is NULL then the owner died
618 * and another waiter is on the fly. pi_state->owner
619 * is fixed up by the task which acquires
620 * pi_state->rt_mutex.
622 * We do not check for pid == 0 which can happen when
623 * the owner died and robust_list_exit() cleared the
624 * TID.
626 if (pid && pi_state->owner) {
628 * Bail out if user space manipulated the
629 * futex value.
631 if (pid != task_pid_vnr(pi_state->owner))
632 return -EINVAL;
635 atomic_inc(&pi_state->refcount);
636 *ps = pi_state;
638 return 0;
643 * We are the first waiter - try to look up the real owner and attach
644 * the new pi_state to it, but bail out when TID = 0
646 if (!pid)
647 return -ESRCH;
648 p = futex_find_get_task(pid);
649 if (!p)
650 return -ESRCH;
653 * We need to look at the task state flags to figure out,
654 * whether the task is exiting. To protect against the do_exit
655 * change of the task flags, we do this protected by
656 * p->pi_lock:
658 raw_spin_lock_irq(&p->pi_lock);
659 if (unlikely(p->flags & PF_EXITING)) {
661 * The task is on the way out. When PF_EXITPIDONE is
662 * set, we know that the task has finished the
663 * cleanup:
665 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
667 raw_spin_unlock_irq(&p->pi_lock);
668 put_task_struct(p);
669 return ret;
672 pi_state = alloc_pi_state();
675 * Initialize the pi_mutex in locked state and make 'p'
676 * the owner of it:
678 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
680 /* Store the key for possible exit cleanups: */
681 pi_state->key = *key;
683 WARN_ON(!list_empty(&pi_state->list));
684 list_add(&pi_state->list, &p->pi_state_list);
685 pi_state->owner = p;
686 raw_spin_unlock_irq(&p->pi_lock);
688 put_task_struct(p);
690 *ps = pi_state;
692 return 0;
696 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
697 * @uaddr: the pi futex user address
698 * @hb: the pi futex hash bucket
699 * @key: the futex key associated with uaddr and hb
700 * @ps: the pi_state pointer where we store the result of the
701 * lookup
702 * @task: the task to perform the atomic lock work for. This will
703 * be "current" except in the case of requeue pi.
704 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
706 * Returns:
707 * 0 - ready to wait
708 * 1 - acquired the lock
709 * <0 - error
711 * The hb->lock and futex_key refs shall be held by the caller.
713 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
714 union futex_key *key,
715 struct futex_pi_state **ps,
716 struct task_struct *task, int set_waiters)
718 int lock_taken, ret, ownerdied = 0;
719 u32 uval, newval, curval, vpid = task_pid_vnr(task);
721 retry:
722 ret = lock_taken = 0;
725 * To avoid races, we attempt to take the lock here again
726 * (by doing a 0 -> TID atomic cmpxchg), while holding all
727 * the locks. It will most likely not succeed.
729 newval = vpid;
730 if (set_waiters)
731 newval |= FUTEX_WAITERS;
733 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
734 return -EFAULT;
737 * Detect deadlocks.
739 if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
740 return -EDEADLK;
743 * Surprise - we got the lock. Just return to userspace:
745 if (unlikely(!curval))
746 return 1;
748 uval = curval;
751 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
752 * to wake at the next unlock.
754 newval = curval | FUTEX_WAITERS;
757 * There are two cases, where a futex might have no owner (the
758 * owner TID is 0): OWNER_DIED. We take over the futex in this
759 * case. We also do an unconditional take over, when the owner
760 * of the futex died.
762 * This is safe as we are protected by the hash bucket lock !
764 if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
765 /* Keep the OWNER_DIED bit */
766 newval = (curval & ~FUTEX_TID_MASK) | vpid;
767 ownerdied = 0;
768 lock_taken = 1;
771 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
772 return -EFAULT;
773 if (unlikely(curval != uval))
774 goto retry;
777 * We took the lock due to owner died take over.
779 if (unlikely(lock_taken))
780 return 1;
783 * We dont have the lock. Look up the PI state (or create it if
784 * we are the first waiter):
786 ret = lookup_pi_state(uval, hb, key, ps);
788 if (unlikely(ret)) {
789 switch (ret) {
790 case -ESRCH:
792 * No owner found for this futex. Check if the
793 * OWNER_DIED bit is set to figure out whether
794 * this is a robust futex or not.
796 if (get_futex_value_locked(&curval, uaddr))
797 return -EFAULT;
800 * We simply start over in case of a robust
801 * futex. The code above will take the futex
802 * and return happy.
804 if (curval & FUTEX_OWNER_DIED) {
805 ownerdied = 1;
806 goto retry;
808 default:
809 break;
813 return ret;
817 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
818 * @q: The futex_q to unqueue
820 * The q->lock_ptr must not be NULL and must be held by the caller.
822 static void __unqueue_futex(struct futex_q *q)
824 struct futex_hash_bucket *hb;
826 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
827 || WARN_ON(plist_node_empty(&q->list)))
828 return;
830 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
831 plist_del(&q->list, &hb->chain);
835 * The hash bucket lock must be held when this is called.
836 * Afterwards, the futex_q must not be accessed.
838 static void wake_futex(struct futex_q *q)
840 struct task_struct *p = q->task;
843 * We set q->lock_ptr = NULL _before_ we wake up the task. If
844 * a non-futex wake up happens on another CPU then the task
845 * might exit and p would dereference a non-existing task
846 * struct. Prevent this by holding a reference on p across the
847 * wake up.
849 get_task_struct(p);
851 __unqueue_futex(q);
853 * The waiting task can free the futex_q as soon as
854 * q->lock_ptr = NULL is written, without taking any locks. A
855 * memory barrier is required here to prevent the following
856 * store to lock_ptr from getting ahead of the plist_del.
858 smp_wmb();
859 q->lock_ptr = NULL;
861 wake_up_state(p, TASK_NORMAL);
862 put_task_struct(p);
865 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
867 struct task_struct *new_owner;
868 struct futex_pi_state *pi_state = this->pi_state;
869 u32 uninitialized_var(curval), newval;
871 if (!pi_state)
872 return -EINVAL;
875 * If current does not own the pi_state then the futex is
876 * inconsistent and user space fiddled with the futex value.
878 if (pi_state->owner != current)
879 return -EINVAL;
881 raw_spin_lock(&pi_state->pi_mutex.wait_lock);
882 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
885 * It is possible that the next waiter (the one that brought
886 * this owner to the kernel) timed out and is no longer
887 * waiting on the lock.
889 if (!new_owner)
890 new_owner = this->task;
893 * We pass it to the next owner. (The WAITERS bit is always
894 * kept enabled while there is PI state around. We must also
895 * preserve the owner died bit.)
897 if (!(uval & FUTEX_OWNER_DIED)) {
898 int ret = 0;
900 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
902 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
903 ret = -EFAULT;
904 else if (curval != uval)
905 ret = -EINVAL;
906 if (ret) {
907 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
908 return ret;
912 raw_spin_lock_irq(&pi_state->owner->pi_lock);
913 WARN_ON(list_empty(&pi_state->list));
914 list_del_init(&pi_state->list);
915 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
917 raw_spin_lock_irq(&new_owner->pi_lock);
918 WARN_ON(!list_empty(&pi_state->list));
919 list_add(&pi_state->list, &new_owner->pi_state_list);
920 pi_state->owner = new_owner;
921 raw_spin_unlock_irq(&new_owner->pi_lock);
923 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
924 rt_mutex_unlock(&pi_state->pi_mutex);
926 return 0;
929 static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
931 u32 uninitialized_var(oldval);
934 * There is no waiter, so we unlock the futex. The owner died
935 * bit has not to be preserved here. We are the owner:
937 if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
938 return -EFAULT;
939 if (oldval != uval)
940 return -EAGAIN;
942 return 0;
946 * Express the locking dependencies for lockdep:
948 static inline void
949 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
951 if (hb1 <= hb2) {
952 spin_lock(&hb1->lock);
953 if (hb1 < hb2)
954 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
955 } else { /* hb1 > hb2 */
956 spin_lock(&hb2->lock);
957 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
961 static inline void
962 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
964 spin_unlock(&hb1->lock);
965 if (hb1 != hb2)
966 spin_unlock(&hb2->lock);
970 * Wake up waiters matching bitset queued on this futex (uaddr).
972 static int
973 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
975 struct futex_hash_bucket *hb;
976 struct futex_q *this, *next;
977 struct plist_head *head;
978 union futex_key key = FUTEX_KEY_INIT;
979 int ret;
981 if (!bitset)
982 return -EINVAL;
984 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
985 if (unlikely(ret != 0))
986 goto out;
988 hb = hash_futex(&key);
989 spin_lock(&hb->lock);
990 head = &hb->chain;
992 plist_for_each_entry_safe(this, next, head, list) {
993 if (match_futex (&this->key, &key)) {
994 if (this->pi_state || this->rt_waiter) {
995 ret = -EINVAL;
996 break;
999 /* Check if one of the bits is set in both bitsets */
1000 if (!(this->bitset & bitset))
1001 continue;
1003 wake_futex(this);
1004 if (++ret >= nr_wake)
1005 break;
1009 spin_unlock(&hb->lock);
1010 put_futex_key(&key);
1011 out:
1012 return ret;
1016 * Wake up all waiters hashed on the physical page that is mapped
1017 * to this virtual address:
1019 static int
1020 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1021 int nr_wake, int nr_wake2, int op)
1023 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1024 struct futex_hash_bucket *hb1, *hb2;
1025 struct plist_head *head;
1026 struct futex_q *this, *next;
1027 int ret, op_ret;
1029 retry:
1030 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1031 if (unlikely(ret != 0))
1032 goto out;
1033 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1034 if (unlikely(ret != 0))
1035 goto out_put_key1;
1037 hb1 = hash_futex(&key1);
1038 hb2 = hash_futex(&key2);
1040 retry_private:
1041 double_lock_hb(hb1, hb2);
1042 op_ret = futex_atomic_op_inuser(op, uaddr2);
1043 if (unlikely(op_ret < 0)) {
1045 double_unlock_hb(hb1, hb2);
1047 #ifndef CONFIG_MMU
1049 * we don't get EFAULT from MMU faults if we don't have an MMU,
1050 * but we might get them from range checking
1052 ret = op_ret;
1053 goto out_put_keys;
1054 #endif
1056 if (unlikely(op_ret != -EFAULT)) {
1057 ret = op_ret;
1058 goto out_put_keys;
1061 ret = fault_in_user_writeable(uaddr2);
1062 if (ret)
1063 goto out_put_keys;
1065 if (!(flags & FLAGS_SHARED))
1066 goto retry_private;
1068 put_futex_key(&key2);
1069 put_futex_key(&key1);
1070 goto retry;
1073 head = &hb1->chain;
1075 plist_for_each_entry_safe(this, next, head, list) {
1076 if (match_futex (&this->key, &key1)) {
1077 wake_futex(this);
1078 if (++ret >= nr_wake)
1079 break;
1083 if (op_ret > 0) {
1084 head = &hb2->chain;
1086 op_ret = 0;
1087 plist_for_each_entry_safe(this, next, head, list) {
1088 if (match_futex (&this->key, &key2)) {
1089 wake_futex(this);
1090 if (++op_ret >= nr_wake2)
1091 break;
1094 ret += op_ret;
1097 double_unlock_hb(hb1, hb2);
1098 out_put_keys:
1099 put_futex_key(&key2);
1100 out_put_key1:
1101 put_futex_key(&key1);
1102 out:
1103 return ret;
1107 * requeue_futex() - Requeue a futex_q from one hb to another
1108 * @q: the futex_q to requeue
1109 * @hb1: the source hash_bucket
1110 * @hb2: the target hash_bucket
1111 * @key2: the new key for the requeued futex_q
1113 static inline
1114 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1115 struct futex_hash_bucket *hb2, union futex_key *key2)
1119 * If key1 and key2 hash to the same bucket, no need to
1120 * requeue.
1122 if (likely(&hb1->chain != &hb2->chain)) {
1123 plist_del(&q->list, &hb1->chain);
1124 plist_add(&q->list, &hb2->chain);
1125 q->lock_ptr = &hb2->lock;
1127 get_futex_key_refs(key2);
1128 q->key = *key2;
1132 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1133 * @q: the futex_q
1134 * @key: the key of the requeue target futex
1135 * @hb: the hash_bucket of the requeue target futex
1137 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1138 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1139 * to the requeue target futex so the waiter can detect the wakeup on the right
1140 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1141 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1142 * to protect access to the pi_state to fixup the owner later. Must be called
1143 * with both q->lock_ptr and hb->lock held.
1145 static inline
1146 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1147 struct futex_hash_bucket *hb)
1149 get_futex_key_refs(key);
1150 q->key = *key;
1152 __unqueue_futex(q);
1154 WARN_ON(!q->rt_waiter);
1155 q->rt_waiter = NULL;
1157 q->lock_ptr = &hb->lock;
1159 wake_up_state(q->task, TASK_NORMAL);
1163 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1164 * @pifutex: the user address of the to futex
1165 * @hb1: the from futex hash bucket, must be locked by the caller
1166 * @hb2: the to futex hash bucket, must be locked by the caller
1167 * @key1: the from futex key
1168 * @key2: the to futex key
1169 * @ps: address to store the pi_state pointer
1170 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1172 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1173 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1174 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1175 * hb1 and hb2 must be held by the caller.
1177 * Returns:
1178 * 0 - failed to acquire the lock atomicly
1179 * 1 - acquired the lock
1180 * <0 - error
1182 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1183 struct futex_hash_bucket *hb1,
1184 struct futex_hash_bucket *hb2,
1185 union futex_key *key1, union futex_key *key2,
1186 struct futex_pi_state **ps, int set_waiters)
1188 struct futex_q *top_waiter = NULL;
1189 u32 curval;
1190 int ret;
1192 if (get_futex_value_locked(&curval, pifutex))
1193 return -EFAULT;
1196 * Find the top_waiter and determine if there are additional waiters.
1197 * If the caller intends to requeue more than 1 waiter to pifutex,
1198 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1199 * as we have means to handle the possible fault. If not, don't set
1200 * the bit unecessarily as it will force the subsequent unlock to enter
1201 * the kernel.
1203 top_waiter = futex_top_waiter(hb1, key1);
1205 /* There are no waiters, nothing for us to do. */
1206 if (!top_waiter)
1207 return 0;
1209 /* Ensure we requeue to the expected futex. */
1210 if (!match_futex(top_waiter->requeue_pi_key, key2))
1211 return -EINVAL;
1214 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1215 * the contended case or if set_waiters is 1. The pi_state is returned
1216 * in ps in contended cases.
1218 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1219 set_waiters);
1220 if (ret == 1)
1221 requeue_pi_wake_futex(top_waiter, key2, hb2);
1223 return ret;
1227 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1228 * @uaddr1: source futex user address
1229 * @flags: futex flags (FLAGS_SHARED, etc.)
1230 * @uaddr2: target futex user address
1231 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1232 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1233 * @cmpval: @uaddr1 expected value (or %NULL)
1234 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1235 * pi futex (pi to pi requeue is not supported)
1237 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1238 * uaddr2 atomically on behalf of the top waiter.
1240 * Returns:
1241 * >=0 - on success, the number of tasks requeued or woken
1242 * <0 - on error
1244 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1245 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1246 u32 *cmpval, int requeue_pi)
1248 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1249 int drop_count = 0, task_count = 0, ret;
1250 struct futex_pi_state *pi_state = NULL;
1251 struct futex_hash_bucket *hb1, *hb2;
1252 struct plist_head *head1;
1253 struct futex_q *this, *next;
1254 u32 curval2;
1256 if (requeue_pi) {
1258 * requeue_pi requires a pi_state, try to allocate it now
1259 * without any locks in case it fails.
1261 if (refill_pi_state_cache())
1262 return -ENOMEM;
1264 * requeue_pi must wake as many tasks as it can, up to nr_wake
1265 * + nr_requeue, since it acquires the rt_mutex prior to
1266 * returning to userspace, so as to not leave the rt_mutex with
1267 * waiters and no owner. However, second and third wake-ups
1268 * cannot be predicted as they involve race conditions with the
1269 * first wake and a fault while looking up the pi_state. Both
1270 * pthread_cond_signal() and pthread_cond_broadcast() should
1271 * use nr_wake=1.
1273 if (nr_wake != 1)
1274 return -EINVAL;
1277 retry:
1278 if (pi_state != NULL) {
1280 * We will have to lookup the pi_state again, so free this one
1281 * to keep the accounting correct.
1283 free_pi_state(pi_state);
1284 pi_state = NULL;
1287 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1288 if (unlikely(ret != 0))
1289 goto out;
1290 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1291 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1292 if (unlikely(ret != 0))
1293 goto out_put_key1;
1295 hb1 = hash_futex(&key1);
1296 hb2 = hash_futex(&key2);
1298 retry_private:
1299 double_lock_hb(hb1, hb2);
1301 if (likely(cmpval != NULL)) {
1302 u32 curval;
1304 ret = get_futex_value_locked(&curval, uaddr1);
1306 if (unlikely(ret)) {
1307 double_unlock_hb(hb1, hb2);
1309 ret = get_user(curval, uaddr1);
1310 if (ret)
1311 goto out_put_keys;
1313 if (!(flags & FLAGS_SHARED))
1314 goto retry_private;
1316 put_futex_key(&key2);
1317 put_futex_key(&key1);
1318 goto retry;
1320 if (curval != *cmpval) {
1321 ret = -EAGAIN;
1322 goto out_unlock;
1326 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1328 * Attempt to acquire uaddr2 and wake the top waiter. If we
1329 * intend to requeue waiters, force setting the FUTEX_WAITERS
1330 * bit. We force this here where we are able to easily handle
1331 * faults rather in the requeue loop below.
1333 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1334 &key2, &pi_state, nr_requeue);
1337 * At this point the top_waiter has either taken uaddr2 or is
1338 * waiting on it. If the former, then the pi_state will not
1339 * exist yet, look it up one more time to ensure we have a
1340 * reference to it.
1342 if (ret == 1) {
1343 WARN_ON(pi_state);
1344 drop_count++;
1345 task_count++;
1346 ret = get_futex_value_locked(&curval2, uaddr2);
1347 if (!ret)
1348 ret = lookup_pi_state(curval2, hb2, &key2,
1349 &pi_state);
1352 switch (ret) {
1353 case 0:
1354 break;
1355 case -EFAULT:
1356 double_unlock_hb(hb1, hb2);
1357 put_futex_key(&key2);
1358 put_futex_key(&key1);
1359 ret = fault_in_user_writeable(uaddr2);
1360 if (!ret)
1361 goto retry;
1362 goto out;
1363 case -EAGAIN:
1364 /* The owner was exiting, try again. */
1365 double_unlock_hb(hb1, hb2);
1366 put_futex_key(&key2);
1367 put_futex_key(&key1);
1368 cond_resched();
1369 goto retry;
1370 default:
1371 goto out_unlock;
1375 head1 = &hb1->chain;
1376 plist_for_each_entry_safe(this, next, head1, list) {
1377 if (task_count - nr_wake >= nr_requeue)
1378 break;
1380 if (!match_futex(&this->key, &key1))
1381 continue;
1384 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1385 * be paired with each other and no other futex ops.
1387 if ((requeue_pi && !this->rt_waiter) ||
1388 (!requeue_pi && this->rt_waiter)) {
1389 ret = -EINVAL;
1390 break;
1394 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1395 * lock, we already woke the top_waiter. If not, it will be
1396 * woken by futex_unlock_pi().
1398 if (++task_count <= nr_wake && !requeue_pi) {
1399 wake_futex(this);
1400 continue;
1403 /* Ensure we requeue to the expected futex for requeue_pi. */
1404 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1405 ret = -EINVAL;
1406 break;
1410 * Requeue nr_requeue waiters and possibly one more in the case
1411 * of requeue_pi if we couldn't acquire the lock atomically.
1413 if (requeue_pi) {
1414 /* Prepare the waiter to take the rt_mutex. */
1415 atomic_inc(&pi_state->refcount);
1416 this->pi_state = pi_state;
1417 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1418 this->rt_waiter,
1419 this->task, 1);
1420 if (ret == 1) {
1421 /* We got the lock. */
1422 requeue_pi_wake_futex(this, &key2, hb2);
1423 drop_count++;
1424 continue;
1425 } else if (ret) {
1426 /* -EDEADLK */
1427 this->pi_state = NULL;
1428 free_pi_state(pi_state);
1429 goto out_unlock;
1432 requeue_futex(this, hb1, hb2, &key2);
1433 drop_count++;
1436 out_unlock:
1437 double_unlock_hb(hb1, hb2);
1440 * drop_futex_key_refs() must be called outside the spinlocks. During
1441 * the requeue we moved futex_q's from the hash bucket at key1 to the
1442 * one at key2 and updated their key pointer. We no longer need to
1443 * hold the references to key1.
1445 while (--drop_count >= 0)
1446 drop_futex_key_refs(&key1);
1448 out_put_keys:
1449 put_futex_key(&key2);
1450 out_put_key1:
1451 put_futex_key(&key1);
1452 out:
1453 if (pi_state != NULL)
1454 free_pi_state(pi_state);
1455 return ret ? ret : task_count;
1458 /* The key must be already stored in q->key. */
1459 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1460 __acquires(&hb->lock)
1462 struct futex_hash_bucket *hb;
1464 hb = hash_futex(&q->key);
1465 q->lock_ptr = &hb->lock;
1467 spin_lock(&hb->lock);
1468 return hb;
1471 static inline void
1472 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1473 __releases(&hb->lock)
1475 spin_unlock(&hb->lock);
1479 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1480 * @q: The futex_q to enqueue
1481 * @hb: The destination hash bucket
1483 * The hb->lock must be held by the caller, and is released here. A call to
1484 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1485 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1486 * or nothing if the unqueue is done as part of the wake process and the unqueue
1487 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1488 * an example).
1490 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1491 __releases(&hb->lock)
1493 int prio;
1496 * The priority used to register this element is
1497 * - either the real thread-priority for the real-time threads
1498 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1499 * - or MAX_RT_PRIO for non-RT threads.
1500 * Thus, all RT-threads are woken first in priority order, and
1501 * the others are woken last, in FIFO order.
1503 prio = min(current->normal_prio, MAX_RT_PRIO);
1505 plist_node_init(&q->list, prio);
1506 plist_add(&q->list, &hb->chain);
1507 q->task = current;
1508 spin_unlock(&hb->lock);
1512 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1513 * @q: The futex_q to unqueue
1515 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1516 * be paired with exactly one earlier call to queue_me().
1518 * Returns:
1519 * 1 - if the futex_q was still queued (and we removed unqueued it)
1520 * 0 - if the futex_q was already removed by the waking thread
1522 static int unqueue_me(struct futex_q *q)
1524 spinlock_t *lock_ptr;
1525 int ret = 0;
1527 /* In the common case we don't take the spinlock, which is nice. */
1528 retry:
1529 lock_ptr = q->lock_ptr;
1530 barrier();
1531 if (lock_ptr != NULL) {
1532 spin_lock(lock_ptr);
1534 * q->lock_ptr can change between reading it and
1535 * spin_lock(), causing us to take the wrong lock. This
1536 * corrects the race condition.
1538 * Reasoning goes like this: if we have the wrong lock,
1539 * q->lock_ptr must have changed (maybe several times)
1540 * between reading it and the spin_lock(). It can
1541 * change again after the spin_lock() but only if it was
1542 * already changed before the spin_lock(). It cannot,
1543 * however, change back to the original value. Therefore
1544 * we can detect whether we acquired the correct lock.
1546 if (unlikely(lock_ptr != q->lock_ptr)) {
1547 spin_unlock(lock_ptr);
1548 goto retry;
1550 __unqueue_futex(q);
1552 BUG_ON(q->pi_state);
1554 spin_unlock(lock_ptr);
1555 ret = 1;
1558 drop_futex_key_refs(&q->key);
1559 return ret;
1563 * PI futexes can not be requeued and must remove themself from the
1564 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1565 * and dropped here.
1567 static void unqueue_me_pi(struct futex_q *q)
1568 __releases(q->lock_ptr)
1570 __unqueue_futex(q);
1572 BUG_ON(!q->pi_state);
1573 free_pi_state(q->pi_state);
1574 q->pi_state = NULL;
1576 spin_unlock(q->lock_ptr);
1580 * Fixup the pi_state owner with the new owner.
1582 * Must be called with hash bucket lock held and mm->sem held for non
1583 * private futexes.
1585 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1586 struct task_struct *newowner)
1588 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1589 struct futex_pi_state *pi_state = q->pi_state;
1590 struct task_struct *oldowner = pi_state->owner;
1591 u32 uval, uninitialized_var(curval), newval;
1592 int ret;
1594 /* Owner died? */
1595 if (!pi_state->owner)
1596 newtid |= FUTEX_OWNER_DIED;
1599 * We are here either because we stole the rtmutex from the
1600 * previous highest priority waiter or we are the highest priority
1601 * waiter but failed to get the rtmutex the first time.
1602 * We have to replace the newowner TID in the user space variable.
1603 * This must be atomic as we have to preserve the owner died bit here.
1605 * Note: We write the user space value _before_ changing the pi_state
1606 * because we can fault here. Imagine swapped out pages or a fork
1607 * that marked all the anonymous memory readonly for cow.
1609 * Modifying pi_state _before_ the user space value would
1610 * leave the pi_state in an inconsistent state when we fault
1611 * here, because we need to drop the hash bucket lock to
1612 * handle the fault. This might be observed in the PID check
1613 * in lookup_pi_state.
1615 retry:
1616 if (get_futex_value_locked(&uval, uaddr))
1617 goto handle_fault;
1619 while (1) {
1620 newval = (uval & FUTEX_OWNER_DIED) | newtid;
1622 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1623 goto handle_fault;
1624 if (curval == uval)
1625 break;
1626 uval = curval;
1630 * We fixed up user space. Now we need to fix the pi_state
1631 * itself.
1633 if (pi_state->owner != NULL) {
1634 raw_spin_lock_irq(&pi_state->owner->pi_lock);
1635 WARN_ON(list_empty(&pi_state->list));
1636 list_del_init(&pi_state->list);
1637 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1640 pi_state->owner = newowner;
1642 raw_spin_lock_irq(&newowner->pi_lock);
1643 WARN_ON(!list_empty(&pi_state->list));
1644 list_add(&pi_state->list, &newowner->pi_state_list);
1645 raw_spin_unlock_irq(&newowner->pi_lock);
1646 return 0;
1649 * To handle the page fault we need to drop the hash bucket
1650 * lock here. That gives the other task (either the highest priority
1651 * waiter itself or the task which stole the rtmutex) the
1652 * chance to try the fixup of the pi_state. So once we are
1653 * back from handling the fault we need to check the pi_state
1654 * after reacquiring the hash bucket lock and before trying to
1655 * do another fixup. When the fixup has been done already we
1656 * simply return.
1658 handle_fault:
1659 spin_unlock(q->lock_ptr);
1661 ret = fault_in_user_writeable(uaddr);
1663 spin_lock(q->lock_ptr);
1666 * Check if someone else fixed it for us:
1668 if (pi_state->owner != oldowner)
1669 return 0;
1671 if (ret)
1672 return ret;
1674 goto retry;
1677 static long futex_wait_restart(struct restart_block *restart);
1680 * fixup_owner() - Post lock pi_state and corner case management
1681 * @uaddr: user address of the futex
1682 * @q: futex_q (contains pi_state and access to the rt_mutex)
1683 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1685 * After attempting to lock an rt_mutex, this function is called to cleanup
1686 * the pi_state owner as well as handle race conditions that may allow us to
1687 * acquire the lock. Must be called with the hb lock held.
1689 * Returns:
1690 * 1 - success, lock taken
1691 * 0 - success, lock not taken
1692 * <0 - on error (-EFAULT)
1694 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
1696 struct task_struct *owner;
1697 int ret = 0;
1699 if (locked) {
1701 * Got the lock. We might not be the anticipated owner if we
1702 * did a lock-steal - fix up the PI-state in that case:
1704 if (q->pi_state->owner != current)
1705 ret = fixup_pi_state_owner(uaddr, q, current);
1706 goto out;
1710 * Catch the rare case, where the lock was released when we were on the
1711 * way back before we locked the hash bucket.
1713 if (q->pi_state->owner == current) {
1715 * Try to get the rt_mutex now. This might fail as some other
1716 * task acquired the rt_mutex after we removed ourself from the
1717 * rt_mutex waiters list.
1719 if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1720 locked = 1;
1721 goto out;
1725 * pi_state is incorrect, some other task did a lock steal and
1726 * we returned due to timeout or signal without taking the
1727 * rt_mutex. Too late.
1729 raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
1730 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1731 if (!owner)
1732 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
1733 raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
1734 ret = fixup_pi_state_owner(uaddr, q, owner);
1735 goto out;
1739 * Paranoia check. If we did not take the lock, then we should not be
1740 * the owner of the rt_mutex.
1742 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1743 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1744 "pi-state %p\n", ret,
1745 q->pi_state->pi_mutex.owner,
1746 q->pi_state->owner);
1748 out:
1749 return ret ? ret : locked;
1753 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1754 * @hb: the futex hash bucket, must be locked by the caller
1755 * @q: the futex_q to queue up on
1756 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1758 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1759 struct hrtimer_sleeper *timeout)
1762 * The task state is guaranteed to be set before another task can
1763 * wake it. set_current_state() is implemented using set_mb() and
1764 * queue_me() calls spin_unlock() upon completion, both serializing
1765 * access to the hash list and forcing another memory barrier.
1767 set_current_state(TASK_INTERRUPTIBLE);
1768 queue_me(q, hb);
1770 /* Arm the timer */
1771 if (timeout) {
1772 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1773 if (!hrtimer_active(&timeout->timer))
1774 timeout->task = NULL;
1778 * If we have been removed from the hash list, then another task
1779 * has tried to wake us, and we can skip the call to schedule().
1781 if (likely(!plist_node_empty(&q->list))) {
1783 * If the timer has already expired, current will already be
1784 * flagged for rescheduling. Only call schedule if there
1785 * is no timeout, or if it has yet to expire.
1787 if (!timeout || timeout->task)
1788 schedule();
1790 __set_current_state(TASK_RUNNING);
1794 * futex_wait_setup() - Prepare to wait on a futex
1795 * @uaddr: the futex userspace address
1796 * @val: the expected value
1797 * @flags: futex flags (FLAGS_SHARED, etc.)
1798 * @q: the associated futex_q
1799 * @hb: storage for hash_bucket pointer to be returned to caller
1801 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1802 * compare it with the expected value. Handle atomic faults internally.
1803 * Return with the hb lock held and a q.key reference on success, and unlocked
1804 * with no q.key reference on failure.
1806 * Returns:
1807 * 0 - uaddr contains val and hb has been locked
1808 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
1810 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
1811 struct futex_q *q, struct futex_hash_bucket **hb)
1813 u32 uval;
1814 int ret;
1817 * Access the page AFTER the hash-bucket is locked.
1818 * Order is important:
1820 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1821 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1823 * The basic logical guarantee of a futex is that it blocks ONLY
1824 * if cond(var) is known to be true at the time of blocking, for
1825 * any cond. If we locked the hash-bucket after testing *uaddr, that
1826 * would open a race condition where we could block indefinitely with
1827 * cond(var) false, which would violate the guarantee.
1829 * On the other hand, we insert q and release the hash-bucket only
1830 * after testing *uaddr. This guarantees that futex_wait() will NOT
1831 * absorb a wakeup if *uaddr does not match the desired values
1832 * while the syscall executes.
1834 retry:
1835 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
1836 if (unlikely(ret != 0))
1837 return ret;
1839 retry_private:
1840 *hb = queue_lock(q);
1842 ret = get_futex_value_locked(&uval, uaddr);
1844 if (ret) {
1845 queue_unlock(q, *hb);
1847 ret = get_user(uval, uaddr);
1848 if (ret)
1849 goto out;
1851 if (!(flags & FLAGS_SHARED))
1852 goto retry_private;
1854 put_futex_key(&q->key);
1855 goto retry;
1858 if (uval != val) {
1859 queue_unlock(q, *hb);
1860 ret = -EWOULDBLOCK;
1863 out:
1864 if (ret)
1865 put_futex_key(&q->key);
1866 return ret;
1869 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
1870 ktime_t *abs_time, u32 bitset)
1872 struct hrtimer_sleeper timeout, *to = NULL;
1873 struct restart_block *restart;
1874 struct futex_hash_bucket *hb;
1875 struct futex_q q = futex_q_init;
1876 int ret;
1878 if (!bitset)
1879 return -EINVAL;
1880 q.bitset = bitset;
1882 if (abs_time) {
1883 to = &timeout;
1885 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
1886 CLOCK_REALTIME : CLOCK_MONOTONIC,
1887 HRTIMER_MODE_ABS);
1888 hrtimer_init_sleeper(to, current);
1889 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1890 current->timer_slack_ns);
1893 retry:
1895 * Prepare to wait on uaddr. On success, holds hb lock and increments
1896 * q.key refs.
1898 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
1899 if (ret)
1900 goto out;
1902 /* queue_me and wait for wakeup, timeout, or a signal. */
1903 futex_wait_queue_me(hb, &q, to);
1905 /* If we were woken (and unqueued), we succeeded, whatever. */
1906 ret = 0;
1907 /* unqueue_me() drops q.key ref */
1908 if (!unqueue_me(&q))
1909 goto out;
1910 ret = -ETIMEDOUT;
1911 if (to && !to->task)
1912 goto out;
1915 * We expect signal_pending(current), but we might be the
1916 * victim of a spurious wakeup as well.
1918 if (!signal_pending(current))
1919 goto retry;
1921 ret = -ERESTARTSYS;
1922 if (!abs_time)
1923 goto out;
1925 restart = &current_thread_info()->restart_block;
1926 restart->fn = futex_wait_restart;
1927 restart->futex.uaddr = uaddr;
1928 restart->futex.val = val;
1929 restart->futex.time = abs_time->tv64;
1930 restart->futex.bitset = bitset;
1931 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
1933 ret = -ERESTART_RESTARTBLOCK;
1935 out:
1936 if (to) {
1937 hrtimer_cancel(&to->timer);
1938 destroy_hrtimer_on_stack(&to->timer);
1940 return ret;
1944 static long futex_wait_restart(struct restart_block *restart)
1946 u32 __user *uaddr = restart->futex.uaddr;
1947 ktime_t t, *tp = NULL;
1949 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1950 t.tv64 = restart->futex.time;
1951 tp = &t;
1953 restart->fn = do_no_restart_syscall;
1955 return (long)futex_wait(uaddr, restart->futex.flags,
1956 restart->futex.val, tp, restart->futex.bitset);
1961 * Userspace tried a 0 -> TID atomic transition of the futex value
1962 * and failed. The kernel side here does the whole locking operation:
1963 * if there are waiters then it will block, it does PI, etc. (Due to
1964 * races the kernel might see a 0 value of the futex too.)
1966 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
1967 ktime_t *time, int trylock)
1969 struct hrtimer_sleeper timeout, *to = NULL;
1970 struct futex_hash_bucket *hb;
1971 struct futex_q q = futex_q_init;
1972 int res, ret;
1974 if (refill_pi_state_cache())
1975 return -ENOMEM;
1977 if (time) {
1978 to = &timeout;
1979 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1980 HRTIMER_MODE_ABS);
1981 hrtimer_init_sleeper(to, current);
1982 hrtimer_set_expires(&to->timer, *time);
1985 retry:
1986 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
1987 if (unlikely(ret != 0))
1988 goto out;
1990 retry_private:
1991 hb = queue_lock(&q);
1993 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1994 if (unlikely(ret)) {
1995 switch (ret) {
1996 case 1:
1997 /* We got the lock. */
1998 ret = 0;
1999 goto out_unlock_put_key;
2000 case -EFAULT:
2001 goto uaddr_faulted;
2002 case -EAGAIN:
2004 * Task is exiting and we just wait for the
2005 * exit to complete.
2007 queue_unlock(&q, hb);
2008 put_futex_key(&q.key);
2009 cond_resched();
2010 goto retry;
2011 default:
2012 goto out_unlock_put_key;
2017 * Only actually queue now that the atomic ops are done:
2019 queue_me(&q, hb);
2021 WARN_ON(!q.pi_state);
2023 * Block on the PI mutex:
2025 if (!trylock)
2026 ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2027 else {
2028 ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2029 /* Fixup the trylock return value: */
2030 ret = ret ? 0 : -EWOULDBLOCK;
2033 spin_lock(q.lock_ptr);
2035 * Fixup the pi_state owner and possibly acquire the lock if we
2036 * haven't already.
2038 res = fixup_owner(uaddr, &q, !ret);
2040 * If fixup_owner() returned an error, proprogate that. If it acquired
2041 * the lock, clear our -ETIMEDOUT or -EINTR.
2043 if (res)
2044 ret = (res < 0) ? res : 0;
2047 * If fixup_owner() faulted and was unable to handle the fault, unlock
2048 * it and return the fault to userspace.
2050 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2051 rt_mutex_unlock(&q.pi_state->pi_mutex);
2053 /* Unqueue and drop the lock */
2054 unqueue_me_pi(&q);
2056 goto out_put_key;
2058 out_unlock_put_key:
2059 queue_unlock(&q, hb);
2061 out_put_key:
2062 put_futex_key(&q.key);
2063 out:
2064 if (to)
2065 destroy_hrtimer_on_stack(&to->timer);
2066 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2068 uaddr_faulted:
2069 queue_unlock(&q, hb);
2071 ret = fault_in_user_writeable(uaddr);
2072 if (ret)
2073 goto out_put_key;
2075 if (!(flags & FLAGS_SHARED))
2076 goto retry_private;
2078 put_futex_key(&q.key);
2079 goto retry;
2083 * Userspace attempted a TID -> 0 atomic transition, and failed.
2084 * This is the in-kernel slowpath: we look up the PI state (if any),
2085 * and do the rt-mutex unlock.
2087 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2089 struct futex_hash_bucket *hb;
2090 struct futex_q *this, *next;
2091 struct plist_head *head;
2092 union futex_key key = FUTEX_KEY_INIT;
2093 u32 uval, vpid = task_pid_vnr(current);
2094 int ret;
2096 retry:
2097 if (get_user(uval, uaddr))
2098 return -EFAULT;
2100 * We release only a lock we actually own:
2102 if ((uval & FUTEX_TID_MASK) != vpid)
2103 return -EPERM;
2105 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2106 if (unlikely(ret != 0))
2107 goto out;
2109 hb = hash_futex(&key);
2110 spin_lock(&hb->lock);
2113 * To avoid races, try to do the TID -> 0 atomic transition
2114 * again. If it succeeds then we can return without waking
2115 * anyone else up:
2117 if (!(uval & FUTEX_OWNER_DIED) &&
2118 cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2119 goto pi_faulted;
2121 * Rare case: we managed to release the lock atomically,
2122 * no need to wake anyone else up:
2124 if (unlikely(uval == vpid))
2125 goto out_unlock;
2128 * Ok, other tasks may need to be woken up - check waiters
2129 * and do the wakeup if necessary:
2131 head = &hb->chain;
2133 plist_for_each_entry_safe(this, next, head, list) {
2134 if (!match_futex (&this->key, &key))
2135 continue;
2136 ret = wake_futex_pi(uaddr, uval, this);
2138 * The atomic access to the futex value
2139 * generated a pagefault, so retry the
2140 * user-access and the wakeup:
2142 if (ret == -EFAULT)
2143 goto pi_faulted;
2144 goto out_unlock;
2147 * No waiters - kernel unlocks the futex:
2149 if (!(uval & FUTEX_OWNER_DIED)) {
2150 ret = unlock_futex_pi(uaddr, uval);
2151 if (ret == -EFAULT)
2152 goto pi_faulted;
2155 out_unlock:
2156 spin_unlock(&hb->lock);
2157 put_futex_key(&key);
2159 out:
2160 return ret;
2162 pi_faulted:
2163 spin_unlock(&hb->lock);
2164 put_futex_key(&key);
2166 ret = fault_in_user_writeable(uaddr);
2167 if (!ret)
2168 goto retry;
2170 return ret;
2174 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2175 * @hb: the hash_bucket futex_q was original enqueued on
2176 * @q: the futex_q woken while waiting to be requeued
2177 * @key2: the futex_key of the requeue target futex
2178 * @timeout: the timeout associated with the wait (NULL if none)
2180 * Detect if the task was woken on the initial futex as opposed to the requeue
2181 * target futex. If so, determine if it was a timeout or a signal that caused
2182 * the wakeup and return the appropriate error code to the caller. Must be
2183 * called with the hb lock held.
2185 * Returns
2186 * 0 - no early wakeup detected
2187 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2189 static inline
2190 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2191 struct futex_q *q, union futex_key *key2,
2192 struct hrtimer_sleeper *timeout)
2194 int ret = 0;
2197 * With the hb lock held, we avoid races while we process the wakeup.
2198 * We only need to hold hb (and not hb2) to ensure atomicity as the
2199 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2200 * It can't be requeued from uaddr2 to something else since we don't
2201 * support a PI aware source futex for requeue.
2203 if (!match_futex(&q->key, key2)) {
2204 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2206 * We were woken prior to requeue by a timeout or a signal.
2207 * Unqueue the futex_q and determine which it was.
2209 plist_del(&q->list, &hb->chain);
2211 /* Handle spurious wakeups gracefully */
2212 ret = -EWOULDBLOCK;
2213 if (timeout && !timeout->task)
2214 ret = -ETIMEDOUT;
2215 else if (signal_pending(current))
2216 ret = -ERESTARTNOINTR;
2218 return ret;
2222 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2223 * @uaddr: the futex we initially wait on (non-pi)
2224 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2225 * the same type, no requeueing from private to shared, etc.
2226 * @val: the expected value of uaddr
2227 * @abs_time: absolute timeout
2228 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2229 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2230 * @uaddr2: the pi futex we will take prior to returning to user-space
2232 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2233 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2234 * complete the acquisition of the rt_mutex prior to returning to userspace.
2235 * This ensures the rt_mutex maintains an owner when it has waiters; without
2236 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2237 * need to.
2239 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2240 * via the following:
2241 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2242 * 2) wakeup on uaddr2 after a requeue
2243 * 3) signal
2244 * 4) timeout
2246 * If 3, cleanup and return -ERESTARTNOINTR.
2248 * If 2, we may then block on trying to take the rt_mutex and return via:
2249 * 5) successful lock
2250 * 6) signal
2251 * 7) timeout
2252 * 8) other lock acquisition failure
2254 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2256 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2258 * Returns:
2259 * 0 - On success
2260 * <0 - On error
2262 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2263 u32 val, ktime_t *abs_time, u32 bitset,
2264 u32 __user *uaddr2)
2266 struct hrtimer_sleeper timeout, *to = NULL;
2267 struct rt_mutex_waiter rt_waiter;
2268 struct rt_mutex *pi_mutex = NULL;
2269 struct futex_hash_bucket *hb;
2270 union futex_key key2 = FUTEX_KEY_INIT;
2271 struct futex_q q = futex_q_init;
2272 int res, ret;
2274 if (!bitset)
2275 return -EINVAL;
2277 if (abs_time) {
2278 to = &timeout;
2279 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2280 CLOCK_REALTIME : CLOCK_MONOTONIC,
2281 HRTIMER_MODE_ABS);
2282 hrtimer_init_sleeper(to, current);
2283 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2284 current->timer_slack_ns);
2288 * The waiter is allocated on our stack, manipulated by the requeue
2289 * code while we sleep on uaddr.
2291 debug_rt_mutex_init_waiter(&rt_waiter);
2292 rt_waiter.task = NULL;
2294 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2295 if (unlikely(ret != 0))
2296 goto out;
2298 q.bitset = bitset;
2299 q.rt_waiter = &rt_waiter;
2300 q.requeue_pi_key = &key2;
2303 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2304 * count.
2306 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2307 if (ret)
2308 goto out_key2;
2310 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2311 futex_wait_queue_me(hb, &q, to);
2313 spin_lock(&hb->lock);
2314 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2315 spin_unlock(&hb->lock);
2316 if (ret)
2317 goto out_put_keys;
2320 * In order for us to be here, we know our q.key == key2, and since
2321 * we took the hb->lock above, we also know that futex_requeue() has
2322 * completed and we no longer have to concern ourselves with a wakeup
2323 * race with the atomic proxy lock acquisition by the requeue code. The
2324 * futex_requeue dropped our key1 reference and incremented our key2
2325 * reference count.
2328 /* Check if the requeue code acquired the second futex for us. */
2329 if (!q.rt_waiter) {
2331 * Got the lock. We might not be the anticipated owner if we
2332 * did a lock-steal - fix up the PI-state in that case.
2334 if (q.pi_state && (q.pi_state->owner != current)) {
2335 spin_lock(q.lock_ptr);
2336 ret = fixup_pi_state_owner(uaddr2, &q, current);
2337 spin_unlock(q.lock_ptr);
2339 } else {
2341 * We have been woken up by futex_unlock_pi(), a timeout, or a
2342 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2343 * the pi_state.
2345 WARN_ON(!&q.pi_state);
2346 pi_mutex = &q.pi_state->pi_mutex;
2347 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2348 debug_rt_mutex_free_waiter(&rt_waiter);
2350 spin_lock(q.lock_ptr);
2352 * Fixup the pi_state owner and possibly acquire the lock if we
2353 * haven't already.
2355 res = fixup_owner(uaddr2, &q, !ret);
2357 * If fixup_owner() returned an error, proprogate that. If it
2358 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2360 if (res)
2361 ret = (res < 0) ? res : 0;
2363 /* Unqueue and drop the lock. */
2364 unqueue_me_pi(&q);
2368 * If fixup_pi_state_owner() faulted and was unable to handle the
2369 * fault, unlock the rt_mutex and return the fault to userspace.
2371 if (ret == -EFAULT) {
2372 if (rt_mutex_owner(pi_mutex) == current)
2373 rt_mutex_unlock(pi_mutex);
2374 } else if (ret == -EINTR) {
2376 * We've already been requeued, but cannot restart by calling
2377 * futex_lock_pi() directly. We could restart this syscall, but
2378 * it would detect that the user space "val" changed and return
2379 * -EWOULDBLOCK. Save the overhead of the restart and return
2380 * -EWOULDBLOCK directly.
2382 ret = -EWOULDBLOCK;
2385 out_put_keys:
2386 put_futex_key(&q.key);
2387 out_key2:
2388 put_futex_key(&key2);
2390 out:
2391 if (to) {
2392 hrtimer_cancel(&to->timer);
2393 destroy_hrtimer_on_stack(&to->timer);
2395 return ret;
2399 * Support for robust futexes: the kernel cleans up held futexes at
2400 * thread exit time.
2402 * Implementation: user-space maintains a per-thread list of locks it
2403 * is holding. Upon do_exit(), the kernel carefully walks this list,
2404 * and marks all locks that are owned by this thread with the
2405 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2406 * always manipulated with the lock held, so the list is private and
2407 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2408 * field, to allow the kernel to clean up if the thread dies after
2409 * acquiring the lock, but just before it could have added itself to
2410 * the list. There can only be one such pending lock.
2414 * sys_set_robust_list() - Set the robust-futex list head of a task
2415 * @head: pointer to the list-head
2416 * @len: length of the list-head, as userspace expects
2418 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2419 size_t, len)
2421 if (!futex_cmpxchg_enabled)
2422 return -ENOSYS;
2424 * The kernel knows only one size for now:
2426 if (unlikely(len != sizeof(*head)))
2427 return -EINVAL;
2429 current->robust_list = head;
2431 return 0;
2435 * sys_get_robust_list() - Get the robust-futex list head of a task
2436 * @pid: pid of the process [zero for current task]
2437 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2438 * @len_ptr: pointer to a length field, the kernel fills in the header size
2440 SYSCALL_DEFINE3(get_robust_list, int, pid,
2441 struct robust_list_head __user * __user *, head_ptr,
2442 size_t __user *, len_ptr)
2444 struct robust_list_head __user *head;
2445 unsigned long ret;
2446 const struct cred *cred = current_cred(), *pcred;
2448 if (!futex_cmpxchg_enabled)
2449 return -ENOSYS;
2451 if (!pid)
2452 head = current->robust_list;
2453 else {
2454 struct task_struct *p;
2456 ret = -ESRCH;
2457 rcu_read_lock();
2458 p = find_task_by_vpid(pid);
2459 if (!p)
2460 goto err_unlock;
2461 ret = -EPERM;
2462 pcred = __task_cred(p);
2463 /* If victim is in different user_ns, then uids are not
2464 comparable, so we must have CAP_SYS_PTRACE */
2465 if (cred->user->user_ns != pcred->user->user_ns) {
2466 if (!ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2467 goto err_unlock;
2468 goto ok;
2470 /* If victim is in same user_ns, then uids are comparable */
2471 if (cred->euid != pcred->euid &&
2472 cred->euid != pcred->uid &&
2473 !ns_capable(pcred->user->user_ns, CAP_SYS_PTRACE))
2474 goto err_unlock;
2476 head = p->robust_list;
2477 rcu_read_unlock();
2480 if (put_user(sizeof(*head), len_ptr))
2481 return -EFAULT;
2482 return put_user(head, head_ptr);
2484 err_unlock:
2485 rcu_read_unlock();
2487 return ret;
2491 * Process a futex-list entry, check whether it's owned by the
2492 * dying task, and do notification if so:
2494 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2496 u32 uval, uninitialized_var(nval), mval;
2498 retry:
2499 if (get_user(uval, uaddr))
2500 return -1;
2502 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2504 * Ok, this dying thread is truly holding a futex
2505 * of interest. Set the OWNER_DIED bit atomically
2506 * via cmpxchg, and if the value had FUTEX_WAITERS
2507 * set, wake up a waiter (if any). (We have to do a
2508 * futex_wake() even if OWNER_DIED is already set -
2509 * to handle the rare but possible case of recursive
2510 * thread-death.) The rest of the cleanup is done in
2511 * userspace.
2513 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2515 * We are not holding a lock here, but we want to have
2516 * the pagefault_disable/enable() protection because
2517 * we want to handle the fault gracefully. If the
2518 * access fails we try to fault in the futex with R/W
2519 * verification via get_user_pages. get_user() above
2520 * does not guarantee R/W access. If that fails we
2521 * give up and leave the futex locked.
2523 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2524 if (fault_in_user_writeable(uaddr))
2525 return -1;
2526 goto retry;
2528 if (nval != uval)
2529 goto retry;
2532 * Wake robust non-PI futexes here. The wakeup of
2533 * PI futexes happens in exit_pi_state():
2535 if (!pi && (uval & FUTEX_WAITERS))
2536 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2538 return 0;
2542 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2544 static inline int fetch_robust_entry(struct robust_list __user **entry,
2545 struct robust_list __user * __user *head,
2546 unsigned int *pi)
2548 unsigned long uentry;
2550 if (get_user(uentry, (unsigned long __user *)head))
2551 return -EFAULT;
2553 *entry = (void __user *)(uentry & ~1UL);
2554 *pi = uentry & 1;
2556 return 0;
2560 * Walk curr->robust_list (very carefully, it's a userspace list!)
2561 * and mark any locks found there dead, and notify any waiters.
2563 * We silently return on any sign of list-walking problem.
2565 void exit_robust_list(struct task_struct *curr)
2567 struct robust_list_head __user *head = curr->robust_list;
2568 struct robust_list __user *entry, *next_entry, *pending;
2569 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2570 unsigned int uninitialized_var(next_pi);
2571 unsigned long futex_offset;
2572 int rc;
2574 if (!futex_cmpxchg_enabled)
2575 return;
2578 * Fetch the list head (which was registered earlier, via
2579 * sys_set_robust_list()):
2581 if (fetch_robust_entry(&entry, &head->list.next, &pi))
2582 return;
2584 * Fetch the relative futex offset:
2586 if (get_user(futex_offset, &head->futex_offset))
2587 return;
2589 * Fetch any possibly pending lock-add first, and handle it
2590 * if it exists:
2592 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2593 return;
2595 next_entry = NULL; /* avoid warning with gcc */
2596 while (entry != &head->list) {
2598 * Fetch the next entry in the list before calling
2599 * handle_futex_death:
2601 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2603 * A pending lock might already be on the list, so
2604 * don't process it twice:
2606 if (entry != pending)
2607 if (handle_futex_death((void __user *)entry + futex_offset,
2608 curr, pi))
2609 return;
2610 if (rc)
2611 return;
2612 entry = next_entry;
2613 pi = next_pi;
2615 * Avoid excessively long or circular lists:
2617 if (!--limit)
2618 break;
2620 cond_resched();
2623 if (pending)
2624 handle_futex_death((void __user *)pending + futex_offset,
2625 curr, pip);
2628 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2629 u32 __user *uaddr2, u32 val2, u32 val3)
2631 int ret = -ENOSYS, cmd = op & FUTEX_CMD_MASK;
2632 unsigned int flags = 0;
2634 if (!(op & FUTEX_PRIVATE_FLAG))
2635 flags |= FLAGS_SHARED;
2637 if (op & FUTEX_CLOCK_REALTIME) {
2638 flags |= FLAGS_CLOCKRT;
2639 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2640 return -ENOSYS;
2643 switch (cmd) {
2644 case FUTEX_WAIT:
2645 val3 = FUTEX_BITSET_MATCH_ANY;
2646 case FUTEX_WAIT_BITSET:
2647 ret = futex_wait(uaddr, flags, val, timeout, val3);
2648 break;
2649 case FUTEX_WAKE:
2650 val3 = FUTEX_BITSET_MATCH_ANY;
2651 case FUTEX_WAKE_BITSET:
2652 ret = futex_wake(uaddr, flags, val, val3);
2653 break;
2654 case FUTEX_REQUEUE:
2655 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2656 break;
2657 case FUTEX_CMP_REQUEUE:
2658 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2659 break;
2660 case FUTEX_WAKE_OP:
2661 ret = futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2662 break;
2663 case FUTEX_LOCK_PI:
2664 if (futex_cmpxchg_enabled)
2665 ret = futex_lock_pi(uaddr, flags, val, timeout, 0);
2666 break;
2667 case FUTEX_UNLOCK_PI:
2668 if (futex_cmpxchg_enabled)
2669 ret = futex_unlock_pi(uaddr, flags);
2670 break;
2671 case FUTEX_TRYLOCK_PI:
2672 if (futex_cmpxchg_enabled)
2673 ret = futex_lock_pi(uaddr, flags, 0, timeout, 1);
2674 break;
2675 case FUTEX_WAIT_REQUEUE_PI:
2676 val3 = FUTEX_BITSET_MATCH_ANY;
2677 ret = futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2678 uaddr2);
2679 break;
2680 case FUTEX_CMP_REQUEUE_PI:
2681 ret = futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2682 break;
2683 default:
2684 ret = -ENOSYS;
2686 return ret;
2690 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2691 struct timespec __user *, utime, u32 __user *, uaddr2,
2692 u32, val3)
2694 struct timespec ts;
2695 ktime_t t, *tp = NULL;
2696 u32 val2 = 0;
2697 int cmd = op & FUTEX_CMD_MASK;
2699 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2700 cmd == FUTEX_WAIT_BITSET ||
2701 cmd == FUTEX_WAIT_REQUEUE_PI)) {
2702 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2703 return -EFAULT;
2704 if (!timespec_valid(&ts))
2705 return -EINVAL;
2707 t = timespec_to_ktime(ts);
2708 if (cmd == FUTEX_WAIT)
2709 t = ktime_add_safe(ktime_get(), t);
2710 tp = &t;
2713 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2714 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2716 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2717 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2718 val2 = (u32) (unsigned long) utime;
2720 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2723 static int __init futex_init(void)
2725 u32 curval;
2726 int i;
2729 * This will fail and we want it. Some arch implementations do
2730 * runtime detection of the futex_atomic_cmpxchg_inatomic()
2731 * functionality. We want to know that before we call in any
2732 * of the complex code paths. Also we want to prevent
2733 * registration of robust lists in that case. NULL is
2734 * guaranteed to fault and we get -EFAULT on functional
2735 * implementation, the non-functional ones will return
2736 * -ENOSYS.
2738 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
2739 futex_cmpxchg_enabled = 1;
2741 for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2742 plist_head_init(&futex_queues[i].chain);
2743 spin_lock_init(&futex_queues[i].lock);
2746 return 0;
2748 __initcall(futex_init);