1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/memblock.h>
38 #include <linux/fault-inject.h>
39 #include <linux/slab.h>
42 #include "../locking/rtmutex_common.h"
45 * The base of the bucket array and its size are always used together
46 * (after initialization only in futex_hash()), so ensure that they
47 * reside in the same cacheline.
50 struct futex_hash_bucket *queues;
51 unsigned long hashsize;
52 } __futex_data __read_mostly __aligned(2*sizeof(long));
53 #define futex_queues (__futex_data.queues)
54 #define futex_hashsize (__futex_data.hashsize)
58 * Fault injections for futexes.
60 #ifdef CONFIG_FAIL_FUTEX
63 struct fault_attr attr;
67 .attr = FAULT_ATTR_INITIALIZER,
68 .ignore_private = false,
71 static int __init setup_fail_futex(char *str)
73 return setup_fault_attr(&fail_futex.attr, str);
75 __setup("fail_futex=", setup_fail_futex);
77 bool should_fail_futex(bool fshared)
79 if (fail_futex.ignore_private && !fshared)
82 return should_fail(&fail_futex.attr, 1);
85 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
87 static int __init fail_futex_debugfs(void)
89 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
92 dir = fault_create_debugfs_attr("fail_futex", NULL,
97 debugfs_create_bool("ignore-private", mode, dir,
98 &fail_futex.ignore_private);
102 late_initcall(fail_futex_debugfs);
104 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
106 #endif /* CONFIG_FAIL_FUTEX */
109 * futex_hash - Return the hash bucket in the global hash
110 * @key: Pointer to the futex key for which the hash is calculated
112 * We hash on the keys returned from get_futex_key (see below) and return the
113 * corresponding hash bucket in the global hash.
115 struct futex_hash_bucket *futex_hash(union futex_key *key)
117 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
120 return &futex_queues[hash & (futex_hashsize - 1)];
125 * futex_setup_timer - set up the sleeping hrtimer.
126 * @time: ptr to the given timeout value
127 * @timeout: the hrtimer_sleeper structure to be set up
128 * @flags: futex flags
129 * @range_ns: optional range in ns
131 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
134 struct hrtimer_sleeper *
135 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
136 int flags, u64 range_ns)
141 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
142 CLOCK_REALTIME : CLOCK_MONOTONIC,
145 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
146 * effectively the same as calling hrtimer_set_expires().
148 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
154 * Generate a machine wide unique identifier for this inode.
156 * This relies on u64 not wrapping in the life-time of the machine; which with
157 * 1ns resolution means almost 585 years.
159 * This further relies on the fact that a well formed program will not unmap
160 * the file while it has a (shared) futex waiting on it. This mapping will have
161 * a file reference which pins the mount and inode.
163 * If for some reason an inode gets evicted and read back in again, it will get
164 * a new sequence number and will _NOT_ match, even though it is the exact same
167 * It is important that futex_match() will never have a false-positive, esp.
168 * for PI futexes that can mess up the state. The above argues that false-negatives
169 * are only possible for malformed programs.
171 static u64 get_inode_sequence_number(struct inode *inode)
173 static atomic64_t i_seq;
176 /* Does the inode already have a sequence number? */
177 old = atomic64_read(&inode->i_sequence);
182 u64 new = atomic64_add_return(1, &i_seq);
183 if (WARN_ON_ONCE(!new))
186 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
194 * get_futex_key() - Get parameters which are the keys for a futex
195 * @uaddr: virtual address of the futex
196 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
197 * @key: address where result is stored.
198 * @rw: mapping needs to be read/write (values: FUTEX_READ,
201 * Return: a negative error code or 0
203 * The key words are stored in @key on success.
205 * For shared mappings (when @fshared), the key is:
207 * ( inode->i_sequence, page->index, offset_within_page )
209 * [ also see get_inode_sequence_number() ]
211 * For private mappings (or when !@fshared), the key is:
213 * ( current->mm, address, 0 )
215 * This allows (cross process, where applicable) identification of the futex
216 * without keeping the page pinned for the duration of the FUTEX_WAIT.
218 * lock_page() might sleep, the caller should not hold a spinlock.
220 int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
221 enum futex_access rw)
223 unsigned long address = (unsigned long)uaddr;
224 struct mm_struct *mm = current->mm;
225 struct page *page, *tail;
226 struct address_space *mapping;
230 * The futex address must be "naturally" aligned.
232 key->both.offset = address % PAGE_SIZE;
233 if (unlikely((address % sizeof(u32)) != 0))
235 address -= key->both.offset;
237 if (unlikely(!access_ok(uaddr, sizeof(u32))))
240 if (unlikely(should_fail_futex(fshared)))
244 * PROCESS_PRIVATE futexes are fast.
245 * As the mm cannot disappear under us and the 'key' only needs
246 * virtual address, we dont even have to find the underlying vma.
247 * Note : We do have to check 'uaddr' is a valid user address,
248 * but access_ok() should be faster than find_vma()
252 * On no-MMU, shared futexes are treated as private, therefore
253 * we must not include the current process in the key. Since
254 * there is only one address space, the address is a unique key
257 if (IS_ENABLED(CONFIG_MMU))
258 key->private.mm = mm;
260 key->private.mm = NULL;
262 key->private.address = address;
267 /* Ignore any VERIFY_READ mapping (futex common case) */
268 if (unlikely(should_fail_futex(true)))
271 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
273 * If write access is not required (eg. FUTEX_WAIT), try
274 * and get read-only access.
276 if (err == -EFAULT && rw == FUTEX_READ) {
277 err = get_user_pages_fast(address, 1, 0, &page);
286 * The treatment of mapping from this point on is critical. The page
287 * lock protects many things but in this context the page lock
288 * stabilizes mapping, prevents inode freeing in the shared
289 * file-backed region case and guards against movement to swap cache.
291 * Strictly speaking the page lock is not needed in all cases being
292 * considered here and page lock forces unnecessarily serialization
293 * From this point on, mapping will be re-verified if necessary and
294 * page lock will be acquired only if it is unavoidable
296 * Mapping checks require the head page for any compound page so the
297 * head page and mapping is looked up now. For anonymous pages, it
298 * does not matter if the page splits in the future as the key is
299 * based on the address. For filesystem-backed pages, the tail is
300 * required as the index of the page determines the key. For
301 * base pages, there is no tail page and tail == page.
304 page = compound_head(page);
305 mapping = READ_ONCE(page->mapping);
308 * If page->mapping is NULL, then it cannot be a PageAnon
309 * page; but it might be the ZERO_PAGE or in the gate area or
310 * in a special mapping (all cases which we are happy to fail);
311 * or it may have been a good file page when get_user_pages_fast
312 * found it, but truncated or holepunched or subjected to
313 * invalidate_complete_page2 before we got the page lock (also
314 * cases which we are happy to fail). And we hold a reference,
315 * so refcount care in invalidate_inode_page's remove_mapping
316 * prevents drop_caches from setting mapping to NULL beneath us.
318 * The case we do have to guard against is when memory pressure made
319 * shmem_writepage move it from filecache to swapcache beneath us:
320 * an unlikely race, but we do need to retry for page->mapping.
322 if (unlikely(!mapping)) {
326 * Page lock is required to identify which special case above
327 * applies. If this is really a shmem page then the page lock
328 * will prevent unexpected transitions.
331 shmem_swizzled = PageSwapCache(page) || page->mapping;
342 * Private mappings are handled in a simple way.
344 * If the futex key is stored on an anonymous page, then the associated
345 * object is the mm which is implicitly pinned by the calling process.
347 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
348 * it's a read-only handle, it's expected that futexes attach to
349 * the object not the particular process.
351 if (PageAnon(page)) {
353 * A RO anonymous page will never change and thus doesn't make
354 * sense for futex operations.
356 if (unlikely(should_fail_futex(true)) || ro) {
361 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
362 key->private.mm = mm;
363 key->private.address = address;
369 * The associated futex object in this case is the inode and
370 * the page->mapping must be traversed. Ordinarily this should
371 * be stabilised under page lock but it's not strictly
372 * necessary in this case as we just want to pin the inode, not
373 * update the radix tree or anything like that.
375 * The RCU read lock is taken as the inode is finally freed
376 * under RCU. If the mapping still matches expectations then the
377 * mapping->host can be safely accessed as being a valid inode.
381 if (READ_ONCE(page->mapping) != mapping) {
388 inode = READ_ONCE(mapping->host);
396 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
397 key->shared.i_seq = get_inode_sequence_number(inode);
398 key->shared.pgoff = page_to_pgoff(tail);
408 * fault_in_user_writeable() - Fault in user address and verify RW access
409 * @uaddr: pointer to faulting user space address
411 * Slow path to fixup the fault we just took in the atomic write
414 * We have no generic implementation of a non-destructive write to the
415 * user address. We know that we faulted in the atomic pagefault
416 * disabled section so we can as well avoid the #PF overhead by
417 * calling get_user_pages() right away.
419 int fault_in_user_writeable(u32 __user *uaddr)
421 struct mm_struct *mm = current->mm;
425 ret = fixup_user_fault(mm, (unsigned long)uaddr,
426 FAULT_FLAG_WRITE, NULL);
427 mmap_read_unlock(mm);
429 return ret < 0 ? ret : 0;
433 * futex_top_waiter() - Return the highest priority waiter on a futex
434 * @hb: the hash bucket the futex_q's reside in
435 * @key: the futex key (to distinguish it from other futex futex_q's)
437 * Must be called with the hb lock held.
439 struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, union futex_key *key)
441 struct futex_q *this;
443 plist_for_each_entry(this, &hb->chain, list) {
444 if (futex_match(&this->key, key))
450 int futex_cmpxchg_value_locked(u32 *curval, u32 __user *uaddr, u32 uval, u32 newval)
455 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
461 int futex_get_value_locked(u32 *dest, u32 __user *from)
466 ret = __get_user(*dest, from);
469 return ret ? -EFAULT : 0;
473 * wait_for_owner_exiting - Block until the owner has exited
474 * @ret: owner's current futex lock status
475 * @exiting: Pointer to the exiting task
477 * Caller must hold a refcount on @exiting.
479 void wait_for_owner_exiting(int ret, struct task_struct *exiting)
482 WARN_ON_ONCE(exiting);
486 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
489 mutex_lock(&exiting->futex_exit_mutex);
491 * No point in doing state checking here. If the waiter got here
492 * while the task was in exec()->exec_futex_release() then it can
493 * have any FUTEX_STATE_* value when the waiter has acquired the
494 * mutex. OK, if running, EXITING or DEAD if it reached exit()
495 * already. Highly unlikely and not a problem. Just one more round
496 * through the futex maze.
498 mutex_unlock(&exiting->futex_exit_mutex);
500 put_task_struct(exiting);
504 * __futex_unqueue() - Remove the futex_q from its futex_hash_bucket
505 * @q: The futex_q to unqueue
507 * The q->lock_ptr must not be NULL and must be held by the caller.
509 void __futex_unqueue(struct futex_q *q)
511 struct futex_hash_bucket *hb;
513 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
515 lockdep_assert_held(q->lock_ptr);
517 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
518 plist_del(&q->list, &hb->chain);
519 futex_hb_waiters_dec(hb);
522 /* The key must be already stored in q->key. */
523 struct futex_hash_bucket *futex_q_lock(struct futex_q *q)
524 __acquires(&hb->lock)
526 struct futex_hash_bucket *hb;
528 hb = futex_hash(&q->key);
531 * Increment the counter before taking the lock so that
532 * a potential waker won't miss a to-be-slept task that is
533 * waiting for the spinlock. This is safe as all futex_q_lock()
534 * users end up calling futex_queue(). Similarly, for housekeeping,
535 * decrement the counter at futex_q_unlock() when some error has
536 * occurred and we don't end up adding the task to the list.
538 futex_hb_waiters_inc(hb); /* implies smp_mb(); (A) */
540 q->lock_ptr = &hb->lock;
542 spin_lock(&hb->lock);
546 void futex_q_unlock(struct futex_hash_bucket *hb)
547 __releases(&hb->lock)
549 spin_unlock(&hb->lock);
550 futex_hb_waiters_dec(hb);
553 void __futex_queue(struct futex_q *q, struct futex_hash_bucket *hb)
558 * The priority used to register this element is
559 * - either the real thread-priority for the real-time threads
560 * (i.e. threads with a priority lower than MAX_RT_PRIO)
561 * - or MAX_RT_PRIO for non-RT threads.
562 * Thus, all RT-threads are woken first in priority order, and
563 * the others are woken last, in FIFO order.
565 prio = min(current->normal_prio, MAX_RT_PRIO);
567 plist_node_init(&q->list, prio);
568 plist_add(&q->list, &hb->chain);
573 * futex_unqueue() - Remove the futex_q from its futex_hash_bucket
574 * @q: The futex_q to unqueue
576 * The q->lock_ptr must not be held by the caller. A call to futex_unqueue() must
577 * be paired with exactly one earlier call to futex_queue().
580 * - 1 - if the futex_q was still queued (and we removed unqueued it);
581 * - 0 - if the futex_q was already removed by the waking thread
583 int futex_unqueue(struct futex_q *q)
585 spinlock_t *lock_ptr;
588 /* In the common case we don't take the spinlock, which is nice. */
591 * q->lock_ptr can change between this read and the following spin_lock.
592 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
593 * optimizing lock_ptr out of the logic below.
595 lock_ptr = READ_ONCE(q->lock_ptr);
596 if (lock_ptr != NULL) {
599 * q->lock_ptr can change between reading it and
600 * spin_lock(), causing us to take the wrong lock. This
601 * corrects the race condition.
603 * Reasoning goes like this: if we have the wrong lock,
604 * q->lock_ptr must have changed (maybe several times)
605 * between reading it and the spin_lock(). It can
606 * change again after the spin_lock() but only if it was
607 * already changed before the spin_lock(). It cannot,
608 * however, change back to the original value. Therefore
609 * we can detect whether we acquired the correct lock.
611 if (unlikely(lock_ptr != q->lock_ptr)) {
612 spin_unlock(lock_ptr);
619 spin_unlock(lock_ptr);
627 * PI futexes can not be requeued and must remove themselves from the
628 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
630 void futex_unqueue_pi(struct futex_q *q)
634 BUG_ON(!q->pi_state);
635 put_pi_state(q->pi_state);
639 /* Constants for the pending_op argument of handle_futex_death */
640 #define HANDLE_DEATH_PENDING true
641 #define HANDLE_DEATH_LIST false
644 * Process a futex-list entry, check whether it's owned by the
645 * dying task, and do notification if so:
647 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
648 bool pi, bool pending_op)
650 u32 uval, nval, mval;
654 /* Futex address must be 32bit aligned */
655 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
659 if (get_user(uval, uaddr))
663 * Special case for regular (non PI) futexes. The unlock path in
664 * user space has two race scenarios:
666 * 1. The unlock path releases the user space futex value and
667 * before it can execute the futex() syscall to wake up
668 * waiters it is killed.
670 * 2. A woken up waiter is killed before it can acquire the
671 * futex in user space.
673 * In the second case, the wake up notification could be generated
674 * by the unlock path in user space after setting the futex value
675 * to zero or by the kernel after setting the OWNER_DIED bit below.
677 * In both cases the TID validation below prevents a wakeup of
678 * potential waiters which can cause these waiters to block
681 * In both cases the following conditions are met:
683 * 1) task->robust_list->list_op_pending != NULL
684 * @pending_op == true
685 * 2) The owner part of user space futex value == 0
686 * 3) Regular futex: @pi == false
688 * If these conditions are met, it is safe to attempt waking up a
689 * potential waiter without touching the user space futex value and
690 * trying to set the OWNER_DIED bit. If the futex value is zero,
691 * the rest of the user space mutex state is consistent, so a woken
692 * waiter will just take over the uncontended futex. Setting the
693 * OWNER_DIED bit would create inconsistent state and malfunction
694 * of the user space owner died handling. Otherwise, the OWNER_DIED
695 * bit is already set, and the woken waiter is expected to deal with
698 owner = uval & FUTEX_TID_MASK;
700 if (pending_op && !pi && !owner) {
701 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
705 if (owner != task_pid_vnr(curr))
709 * Ok, this dying thread is truly holding a futex
710 * of interest. Set the OWNER_DIED bit atomically
711 * via cmpxchg, and if the value had FUTEX_WAITERS
712 * set, wake up a waiter (if any). (We have to do a
713 * futex_wake() even if OWNER_DIED is already set -
714 * to handle the rare but possible case of recursive
715 * thread-death.) The rest of the cleanup is done in
718 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
721 * We are not holding a lock here, but we want to have
722 * the pagefault_disable/enable() protection because
723 * we want to handle the fault gracefully. If the
724 * access fails we try to fault in the futex with R/W
725 * verification via get_user_pages. get_user() above
726 * does not guarantee R/W access. If that fails we
727 * give up and leave the futex locked.
729 if ((err = futex_cmpxchg_value_locked(&nval, uaddr, uval, mval))) {
732 if (fault_in_user_writeable(uaddr))
750 * Wake robust non-PI futexes here. The wakeup of
751 * PI futexes happens in exit_pi_state():
753 if (!pi && (uval & FUTEX_WAITERS))
754 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
760 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
762 static inline int fetch_robust_entry(struct robust_list __user **entry,
763 struct robust_list __user * __user *head,
766 unsigned long uentry;
768 if (get_user(uentry, (unsigned long __user *)head))
771 *entry = (void __user *)(uentry & ~1UL);
778 * Walk curr->robust_list (very carefully, it's a userspace list!)
779 * and mark any locks found there dead, and notify any waiters.
781 * We silently return on any sign of list-walking problem.
783 static void exit_robust_list(struct task_struct *curr)
785 struct robust_list_head __user *head = curr->robust_list;
786 struct robust_list __user *entry, *next_entry, *pending;
787 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
788 unsigned int next_pi;
789 unsigned long futex_offset;
793 * Fetch the list head (which was registered earlier, via
794 * sys_set_robust_list()):
796 if (fetch_robust_entry(&entry, &head->list.next, &pi))
799 * Fetch the relative futex offset:
801 if (get_user(futex_offset, &head->futex_offset))
804 * Fetch any possibly pending lock-add first, and handle it
807 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
810 next_entry = NULL; /* avoid warning with gcc */
811 while (entry != &head->list) {
813 * Fetch the next entry in the list before calling
814 * handle_futex_death:
816 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
818 * A pending lock might already be on the list, so
819 * don't process it twice:
821 if (entry != pending) {
822 if (handle_futex_death((void __user *)entry + futex_offset,
823 curr, pi, HANDLE_DEATH_LIST))
831 * Avoid excessively long or circular lists:
840 handle_futex_death((void __user *)pending + futex_offset,
841 curr, pip, HANDLE_DEATH_PENDING);
846 static void __user *futex_uaddr(struct robust_list __user *entry,
847 compat_long_t futex_offset)
849 compat_uptr_t base = ptr_to_compat(entry);
850 void __user *uaddr = compat_ptr(base + futex_offset);
856 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
859 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
860 compat_uptr_t __user *head, unsigned int *pi)
862 if (get_user(*uentry, head))
865 *entry = compat_ptr((*uentry) & ~1);
866 *pi = (unsigned int)(*uentry) & 1;
872 * Walk curr->robust_list (very carefully, it's a userspace list!)
873 * and mark any locks found there dead, and notify any waiters.
875 * We silently return on any sign of list-walking problem.
877 static void compat_exit_robust_list(struct task_struct *curr)
879 struct compat_robust_list_head __user *head = curr->compat_robust_list;
880 struct robust_list __user *entry, *next_entry, *pending;
881 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
882 unsigned int next_pi;
883 compat_uptr_t uentry, next_uentry, upending;
884 compat_long_t futex_offset;
888 * Fetch the list head (which was registered earlier, via
889 * sys_set_robust_list()):
891 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
894 * Fetch the relative futex offset:
896 if (get_user(futex_offset, &head->futex_offset))
899 * Fetch any possibly pending lock-add first, and handle it
902 if (compat_fetch_robust_entry(&upending, &pending,
903 &head->list_op_pending, &pip))
906 next_entry = NULL; /* avoid warning with gcc */
907 while (entry != (struct robust_list __user *) &head->list) {
909 * Fetch the next entry in the list before calling
910 * handle_futex_death:
912 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
913 (compat_uptr_t __user *)&entry->next, &next_pi);
915 * A pending lock might already be on the list, so
916 * dont process it twice:
918 if (entry != pending) {
919 void __user *uaddr = futex_uaddr(entry, futex_offset);
921 if (handle_futex_death(uaddr, curr, pi,
927 uentry = next_uentry;
931 * Avoid excessively long or circular lists:
939 void __user *uaddr = futex_uaddr(pending, futex_offset);
941 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
946 #ifdef CONFIG_FUTEX_PI
949 * This task is holding PI mutexes at exit time => bad.
950 * Kernel cleans up PI-state, but userspace is likely hosed.
951 * (Robust-futex cleanup is separate and might save the day for userspace.)
953 static void exit_pi_state_list(struct task_struct *curr)
955 struct list_head *next, *head = &curr->pi_state_list;
956 struct futex_pi_state *pi_state;
957 struct futex_hash_bucket *hb;
958 union futex_key key = FUTEX_KEY_INIT;
961 * We are a ZOMBIE and nobody can enqueue itself on
962 * pi_state_list anymore, but we have to be careful
963 * versus waiters unqueueing themselves:
965 raw_spin_lock_irq(&curr->pi_lock);
966 while (!list_empty(head)) {
968 pi_state = list_entry(next, struct futex_pi_state, list);
970 hb = futex_hash(&key);
973 * We can race against put_pi_state() removing itself from the
974 * list (a waiter going away). put_pi_state() will first
975 * decrement the reference count and then modify the list, so
976 * its possible to see the list entry but fail this reference
979 * In that case; drop the locks to let put_pi_state() make
980 * progress and retry the loop.
982 if (!refcount_inc_not_zero(&pi_state->refcount)) {
983 raw_spin_unlock_irq(&curr->pi_lock);
985 raw_spin_lock_irq(&curr->pi_lock);
988 raw_spin_unlock_irq(&curr->pi_lock);
990 spin_lock(&hb->lock);
991 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
992 raw_spin_lock(&curr->pi_lock);
994 * We dropped the pi-lock, so re-check whether this
995 * task still owns the PI-state:
997 if (head->next != next) {
998 /* retain curr->pi_lock for the loop invariant */
999 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1000 spin_unlock(&hb->lock);
1001 put_pi_state(pi_state);
1005 WARN_ON(pi_state->owner != curr);
1006 WARN_ON(list_empty(&pi_state->list));
1007 list_del_init(&pi_state->list);
1008 pi_state->owner = NULL;
1010 raw_spin_unlock(&curr->pi_lock);
1011 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1012 spin_unlock(&hb->lock);
1014 rt_mutex_futex_unlock(&pi_state->pi_mutex);
1015 put_pi_state(pi_state);
1017 raw_spin_lock_irq(&curr->pi_lock);
1019 raw_spin_unlock_irq(&curr->pi_lock);
1022 static inline void exit_pi_state_list(struct task_struct *curr) { }
1025 static void futex_cleanup(struct task_struct *tsk)
1027 if (unlikely(tsk->robust_list)) {
1028 exit_robust_list(tsk);
1029 tsk->robust_list = NULL;
1032 #ifdef CONFIG_COMPAT
1033 if (unlikely(tsk->compat_robust_list)) {
1034 compat_exit_robust_list(tsk);
1035 tsk->compat_robust_list = NULL;
1039 if (unlikely(!list_empty(&tsk->pi_state_list)))
1040 exit_pi_state_list(tsk);
1044 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
1045 * @tsk: task to set the state on
1047 * Set the futex exit state of the task lockless. The futex waiter code
1048 * observes that state when a task is exiting and loops until the task has
1049 * actually finished the futex cleanup. The worst case for this is that the
1050 * waiter runs through the wait loop until the state becomes visible.
1052 * This is called from the recursive fault handling path in make_task_dead().
1054 * This is best effort. Either the futex exit code has run already or
1055 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
1056 * take it over. If not, the problem is pushed back to user space. If the
1057 * futex exit code did not run yet, then an already queued waiter might
1058 * block forever, but there is nothing which can be done about that.
1060 void futex_exit_recursive(struct task_struct *tsk)
1062 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
1063 if (tsk->futex_state == FUTEX_STATE_EXITING)
1064 mutex_unlock(&tsk->futex_exit_mutex);
1065 tsk->futex_state = FUTEX_STATE_DEAD;
1068 static void futex_cleanup_begin(struct task_struct *tsk)
1071 * Prevent various race issues against a concurrent incoming waiter
1072 * including live locks by forcing the waiter to block on
1073 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
1074 * attach_to_pi_owner().
1076 mutex_lock(&tsk->futex_exit_mutex);
1079 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
1081 * This ensures that all subsequent checks of tsk->futex_state in
1082 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
1083 * tsk->pi_lock held.
1085 * It guarantees also that a pi_state which was queued right before
1086 * the state change under tsk->pi_lock by a concurrent waiter must
1087 * be observed in exit_pi_state_list().
1089 raw_spin_lock_irq(&tsk->pi_lock);
1090 tsk->futex_state = FUTEX_STATE_EXITING;
1091 raw_spin_unlock_irq(&tsk->pi_lock);
1094 static void futex_cleanup_end(struct task_struct *tsk, int state)
1097 * Lockless store. The only side effect is that an observer might
1098 * take another loop until it becomes visible.
1100 tsk->futex_state = state;
1102 * Drop the exit protection. This unblocks waiters which observed
1103 * FUTEX_STATE_EXITING to reevaluate the state.
1105 mutex_unlock(&tsk->futex_exit_mutex);
1108 void futex_exec_release(struct task_struct *tsk)
1111 * The state handling is done for consistency, but in the case of
1112 * exec() there is no way to prevent further damage as the PID stays
1113 * the same. But for the unlikely and arguably buggy case that a
1114 * futex is held on exec(), this provides at least as much state
1115 * consistency protection which is possible.
1117 futex_cleanup_begin(tsk);
1120 * Reset the state to FUTEX_STATE_OK. The task is alive and about
1121 * exec a new binary.
1123 futex_cleanup_end(tsk, FUTEX_STATE_OK);
1126 void futex_exit_release(struct task_struct *tsk)
1128 futex_cleanup_begin(tsk);
1130 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
1133 static int __init futex_init(void)
1135 unsigned int futex_shift;
1138 #if CONFIG_BASE_SMALL
1139 futex_hashsize = 16;
1141 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
1144 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
1145 futex_hashsize, 0, 0,
1147 futex_hashsize, futex_hashsize);
1148 futex_hashsize = 1UL << futex_shift;
1150 for (i = 0; i < futex_hashsize; i++) {
1151 atomic_set(&futex_queues[i].waiters, 0);
1152 plist_head_init(&futex_queues[i].chain);
1153 spin_lock_init(&futex_queues[i].lock);
1158 core_initcall(futex_init);