1 // SPDX-License-Identifier: GPL-2.0-or-later
3 #include <linux/sched/task.h>
4 #include <linux/sched/signal.h>
5 #include <linux/freezer.h>
10 * READ this before attempting to hack on futexes!
12 * Basic futex operation and ordering guarantees
13 * =============================================
15 * The waiter reads the futex value in user space and calls
16 * futex_wait(). This function computes the hash bucket and acquires
17 * the hash bucket lock. After that it reads the futex user space value
18 * again and verifies that the data has not changed. If it has not changed
19 * it enqueues itself into the hash bucket, releases the hash bucket lock
22 * The waker side modifies the user space value of the futex and calls
23 * futex_wake(). This function computes the hash bucket and acquires the
24 * hash bucket lock. Then it looks for waiters on that futex in the hash
25 * bucket and wakes them.
27 * In futex wake up scenarios where no tasks are blocked on a futex, taking
28 * the hb spinlock can be avoided and simply return. In order for this
29 * optimization to work, ordering guarantees must exist so that the waiter
30 * being added to the list is acknowledged when the list is concurrently being
31 * checked by the waker, avoiding scenarios like the following:
35 * sys_futex(WAIT, futex, val);
36 * futex_wait(futex, val);
39 * sys_futex(WAKE, futex);
44 * lock(hash_bucket(futex));
46 * unlock(hash_bucket(futex));
49 * This would cause the waiter on CPU 0 to wait forever because it
50 * missed the transition of the user space value from val to newval
51 * and the waker did not find the waiter in the hash bucket queue.
53 * The correct serialization ensures that a waiter either observes
54 * the changed user space value before blocking or is woken by a
59 * sys_futex(WAIT, futex, val);
60 * futex_wait(futex, val);
63 * smp_mb(); (A) <-- paired with -.
65 * lock(hash_bucket(futex)); |
69 * | sys_futex(WAKE, futex);
70 * | futex_wake(futex);
72 * `--------> smp_mb(); (B)
75 * unlock(hash_bucket(futex));
76 * schedule(); if (waiters)
77 * lock(hash_bucket(futex));
78 * else wake_waiters(futex);
79 * waiters--; (b) unlock(hash_bucket(futex));
81 * Where (A) orders the waiters increment and the futex value read through
82 * atomic operations (see futex_hb_waiters_inc) and where (B) orders the write
83 * to futex and the waiters read (see futex_hb_waiters_pending()).
85 * This yields the following case (where X:=waiters, Y:=futex):
93 * Which guarantees that x==0 && y==0 is impossible; which translates back into
94 * the guarantee that we cannot both miss the futex variable change and the
97 * Note that a new waiter is accounted for in (a) even when it is possible that
98 * the wait call can return error, in which case we backtrack from it in (b).
99 * Refer to the comment in futex_q_lock().
101 * Similarly, in order to account for waiters being requeued on another
102 * address we always increment the waiters for the destination bucket before
103 * acquiring the lock. It then decrements them again after releasing it -
104 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
105 * will do the additional required waiter count housekeeping. This is done for
106 * double_lock_hb() and double_unlock_hb(), respectively.
109 bool __futex_wake_mark(struct futex_q *q)
111 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
116 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
117 * is written, without taking any locks. This is possible in the event
118 * of a spurious wakeup, for example. A memory barrier is required here
119 * to prevent the following store to lock_ptr from getting ahead of the
120 * plist_del in __futex_unqueue().
122 smp_store_release(&q->lock_ptr, NULL);
128 * The hash bucket lock must be held when this is called.
129 * Afterwards, the futex_q must not be accessed. Callers
130 * must ensure to later call wake_up_q() for the actual
133 void futex_wake_mark(struct wake_q_head *wake_q, struct futex_q *q)
135 struct task_struct *p = q->task;
139 if (!__futex_wake_mark(q)) {
145 * Queue the task for later wakeup for after we've released
148 wake_q_add_safe(wake_q, p);
152 * Wake up waiters matching bitset queued on this futex (uaddr).
154 int futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
156 struct futex_hash_bucket *hb;
157 struct futex_q *this, *next;
158 union futex_key key = FUTEX_KEY_INIT;
159 DEFINE_WAKE_Q(wake_q);
165 ret = get_futex_key(uaddr, flags, &key, FUTEX_READ);
166 if (unlikely(ret != 0))
169 if ((flags & FLAGS_STRICT) && !nr_wake)
172 hb = futex_hash(&key);
174 /* Make sure we really have tasks to wakeup */
175 if (!futex_hb_waiters_pending(hb))
178 spin_lock(&hb->lock);
180 plist_for_each_entry_safe(this, next, &hb->chain, list) {
181 if (futex_match (&this->key, &key)) {
182 if (this->pi_state || this->rt_waiter) {
187 /* Check if one of the bits is set in both bitsets */
188 if (!(this->bitset & bitset))
191 this->wake(&wake_q, this);
192 if (++ret >= nr_wake)
197 spin_unlock(&hb->lock);
202 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
204 unsigned int op = (encoded_op & 0x70000000) >> 28;
205 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
206 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
207 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
210 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
211 if (oparg < 0 || oparg > 31) {
212 char comm[sizeof(current->comm)];
214 * kill this print and return -EINVAL when userspace
217 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
218 get_task_comm(comm, current), oparg);
225 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
231 case FUTEX_OP_CMP_EQ:
232 return oldval == cmparg;
233 case FUTEX_OP_CMP_NE:
234 return oldval != cmparg;
235 case FUTEX_OP_CMP_LT:
236 return oldval < cmparg;
237 case FUTEX_OP_CMP_GE:
238 return oldval >= cmparg;
239 case FUTEX_OP_CMP_LE:
240 return oldval <= cmparg;
241 case FUTEX_OP_CMP_GT:
242 return oldval > cmparg;
249 * Wake up all waiters hashed on the physical page that is mapped
250 * to this virtual address:
252 int futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
253 int nr_wake, int nr_wake2, int op)
255 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
256 struct futex_hash_bucket *hb1, *hb2;
257 struct futex_q *this, *next;
259 DEFINE_WAKE_Q(wake_q);
262 ret = get_futex_key(uaddr1, flags, &key1, FUTEX_READ);
263 if (unlikely(ret != 0))
265 ret = get_futex_key(uaddr2, flags, &key2, FUTEX_WRITE);
266 if (unlikely(ret != 0))
269 hb1 = futex_hash(&key1);
270 hb2 = futex_hash(&key2);
273 double_lock_hb(hb1, hb2);
274 op_ret = futex_atomic_op_inuser(op, uaddr2);
275 if (unlikely(op_ret < 0)) {
276 double_unlock_hb(hb1, hb2);
278 if (!IS_ENABLED(CONFIG_MMU) ||
279 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
281 * we don't get EFAULT from MMU faults if we don't have
282 * an MMU, but we might get them from range checking
288 if (op_ret == -EFAULT) {
289 ret = fault_in_user_writeable(uaddr2);
295 if (!(flags & FLAGS_SHARED))
300 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
301 if (futex_match (&this->key, &key1)) {
302 if (this->pi_state || this->rt_waiter) {
306 this->wake(&wake_q, this);
307 if (++ret >= nr_wake)
314 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
315 if (futex_match (&this->key, &key2)) {
316 if (this->pi_state || this->rt_waiter) {
320 this->wake(&wake_q, this);
321 if (++op_ret >= nr_wake2)
329 double_unlock_hb(hb1, hb2);
334 static long futex_wait_restart(struct restart_block *restart);
337 * futex_wait_queue() - futex_queue() and wait for wakeup, timeout, or signal
338 * @hb: the futex hash bucket, must be locked by the caller
339 * @q: the futex_q to queue up on
340 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
342 void futex_wait_queue(struct futex_hash_bucket *hb, struct futex_q *q,
343 struct hrtimer_sleeper *timeout)
346 * The task state is guaranteed to be set before another task can
347 * wake it. set_current_state() is implemented using smp_store_mb() and
348 * futex_queue() calls spin_unlock() upon completion, both serializing
349 * access to the hash list and forcing another memory barrier.
351 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
356 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
359 * If we have been removed from the hash list, then another task
360 * has tried to wake us, and we can skip the call to schedule().
362 if (likely(!plist_node_empty(&q->list))) {
364 * If the timer has already expired, current will already be
365 * flagged for rescheduling. Only call schedule if there
366 * is no timeout, or if it has yet to expire.
368 if (!timeout || timeout->task)
371 __set_current_state(TASK_RUNNING);
375 * futex_unqueue_multiple - Remove various futexes from their hash bucket
376 * @v: The list of futexes to unqueue
377 * @count: Number of futexes in the list
379 * Helper to unqueue a list of futexes. This can't fail.
382 * - >=0 - Index of the last futex that was awoken;
383 * - -1 - No futex was awoken
385 int futex_unqueue_multiple(struct futex_vector *v, int count)
389 for (i = 0; i < count; i++) {
390 if (!futex_unqueue(&v[i].q))
398 * futex_wait_multiple_setup - Prepare to wait and enqueue multiple futexes
399 * @vs: The futex list to wait on
400 * @count: The size of the list
401 * @woken: Index of the last woken futex, if any. Used to notify the
402 * caller that it can return this index to userspace (return parameter)
404 * Prepare multiple futexes in a single step and enqueue them. This may fail if
405 * the futex list is invalid or if any futex was already awoken. On success the
406 * task is ready to interruptible sleep.
409 * - 1 - One of the futexes was woken by another thread
411 * - <0 - -EFAULT, -EWOULDBLOCK or -EINVAL
413 int futex_wait_multiple_setup(struct futex_vector *vs, int count, int *woken)
415 struct futex_hash_bucket *hb;
421 * Enqueuing multiple futexes is tricky, because we need to enqueue
422 * each futex on the list before dealing with the next one to avoid
423 * deadlocking on the hash bucket. But, before enqueuing, we need to
424 * make sure that current->state is TASK_INTERRUPTIBLE, so we don't
425 * lose any wake events, which cannot be done before the get_futex_key
426 * of the next key, because it calls get_user_pages, which can sleep.
427 * Thus, we fetch the list of futexes keys in two steps, by first
428 * pinning all the memory keys in the futex key, and only then we read
429 * each key and queue the corresponding futex.
431 * Private futexes doesn't need to recalculate hash in retry, so skip
432 * get_futex_key() when retrying.
435 for (i = 0; i < count; i++) {
436 if (!(vs[i].w.flags & FLAGS_SHARED) && retry)
439 ret = get_futex_key(u64_to_user_ptr(vs[i].w.uaddr),
441 &vs[i].q.key, FUTEX_READ);
447 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
449 for (i = 0; i < count; i++) {
450 u32 __user *uaddr = (u32 __user *)(unsigned long)vs[i].w.uaddr;
451 struct futex_q *q = &vs[i].q;
452 u32 val = vs[i].w.val;
454 hb = futex_q_lock(q);
455 ret = futex_get_value_locked(&uval, uaddr);
457 if (!ret && uval == val) {
459 * The bucket lock can't be held while dealing with the
460 * next futex. Queue each futex at this moment so hb can
468 __set_current_state(TASK_RUNNING);
471 * Even if something went wrong, if we find out that a futex
472 * was woken, we don't return error and return this index to
475 *woken = futex_unqueue_multiple(vs, i);
481 * If we need to handle a page fault, we need to do so
482 * without any lock and any enqueued futex (otherwise
483 * we could lose some wakeup). So we do it here, after
484 * undoing all the work done so far. In success, we
485 * retry all the work.
487 if (get_user(uval, uaddr))
502 * futex_sleep_multiple - Check sleeping conditions and sleep
503 * @vs: List of futexes to wait for
504 * @count: Length of vs
507 * Sleep if and only if the timeout hasn't expired and no futex on the list has
510 static void futex_sleep_multiple(struct futex_vector *vs, unsigned int count,
511 struct hrtimer_sleeper *to)
516 for (; count; count--, vs++) {
517 if (!READ_ONCE(vs->q.lock_ptr))
525 * futex_wait_multiple - Prepare to wait on and enqueue several futexes
526 * @vs: The list of futexes to wait on
527 * @count: The number of objects
528 * @to: Timeout before giving up and returning to userspace
530 * Entry point for the FUTEX_WAIT_MULTIPLE futex operation, this function
531 * sleeps on a group of futexes and returns on the first futex that is
532 * wake, or after the timeout has elapsed.
535 * - >=0 - Hint to the futex that was awoken
538 int futex_wait_multiple(struct futex_vector *vs, unsigned int count,
539 struct hrtimer_sleeper *to)
544 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
547 ret = futex_wait_multiple_setup(vs, count, &hint);
550 /* A futex was woken during setup */
556 futex_sleep_multiple(vs, count, to);
558 __set_current_state(TASK_RUNNING);
560 ret = futex_unqueue_multiple(vs, count);
566 else if (signal_pending(current))
569 * The final case is a spurious wakeup, for
576 * futex_wait_setup() - Prepare to wait on a futex
577 * @uaddr: the futex userspace address
578 * @val: the expected value
579 * @flags: futex flags (FLAGS_SHARED, etc.)
580 * @q: the associated futex_q
581 * @hb: storage for hash_bucket pointer to be returned to caller
583 * Setup the futex_q and locate the hash_bucket. Get the futex value and
584 * compare it with the expected value. Handle atomic faults internally.
585 * Return with the hb lock held on success, and unlocked on failure.
588 * - 0 - uaddr contains val and hb has been locked;
589 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
591 int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
592 struct futex_q *q, struct futex_hash_bucket **hb)
598 * Access the page AFTER the hash-bucket is locked.
599 * Order is important:
601 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
602 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
604 * The basic logical guarantee of a futex is that it blocks ONLY
605 * if cond(var) is known to be true at the time of blocking, for
606 * any cond. If we locked the hash-bucket after testing *uaddr, that
607 * would open a race condition where we could block indefinitely with
608 * cond(var) false, which would violate the guarantee.
610 * On the other hand, we insert q and release the hash-bucket only
611 * after testing *uaddr. This guarantees that futex_wait() will NOT
612 * absorb a wakeup if *uaddr does not match the desired values
613 * while the syscall executes.
616 ret = get_futex_key(uaddr, flags, &q->key, FUTEX_READ);
617 if (unlikely(ret != 0))
621 *hb = futex_q_lock(q);
623 ret = futex_get_value_locked(&uval, uaddr);
628 ret = get_user(uval, uaddr);
632 if (!(flags & FLAGS_SHARED))
646 int __futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
647 struct hrtimer_sleeper *to, u32 bitset)
649 struct futex_q q = futex_q_init;
650 struct futex_hash_bucket *hb;
660 * Prepare to wait on uaddr. On success, it holds hb->lock and q
663 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
667 /* futex_queue and wait for wakeup, timeout, or a signal. */
668 futex_wait_queue(hb, &q, to);
670 /* If we were woken (and unqueued), we succeeded, whatever. */
671 if (!futex_unqueue(&q))
678 * We expect signal_pending(current), but we might be the
679 * victim of a spurious wakeup as well.
681 if (!signal_pending(current))
687 int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val, ktime_t *abs_time, u32 bitset)
689 struct hrtimer_sleeper timeout, *to;
690 struct restart_block *restart;
693 to = futex_setup_timer(abs_time, &timeout, flags,
694 current->timer_slack_ns);
696 ret = __futex_wait(uaddr, flags, val, to, bitset);
698 /* No timeout, nothing to clean up. */
702 hrtimer_cancel(&to->timer);
703 destroy_hrtimer_on_stack(&to->timer);
705 if (ret == -ERESTARTSYS) {
706 restart = ¤t->restart_block;
707 restart->futex.uaddr = uaddr;
708 restart->futex.val = val;
709 restart->futex.time = *abs_time;
710 restart->futex.bitset = bitset;
711 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
713 return set_restart_fn(restart, futex_wait_restart);
719 static long futex_wait_restart(struct restart_block *restart)
721 u32 __user *uaddr = restart->futex.uaddr;
722 ktime_t t, *tp = NULL;
724 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
725 t = restart->futex.time;
728 restart->fn = do_no_restart_syscall;
730 return (long)futex_wait(uaddr, restart->futex.flags,
731 restart->futex.val, tp, restart->futex.bitset);