xref: /linux/kernel/futex/requeue.c (revision 79997eda0d31bc68203c95ecb978773ee6ce7a1f)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 
3 #include <linux/sched/signal.h>
4 
5 #include "futex.h"
6 #include "../locking/rtmutex_common.h"
7 
8 /*
9  * On PREEMPT_RT, the hash bucket lock is a 'sleeping' spinlock with an
10  * underlying rtmutex. The task which is about to be requeued could have
11  * just woken up (timeout, signal). After the wake up the task has to
12  * acquire hash bucket lock, which is held by the requeue code.  As a task
13  * can only be blocked on _ONE_ rtmutex at a time, the proxy lock blocking
14  * and the hash bucket lock blocking would collide and corrupt state.
15  *
16  * On !PREEMPT_RT this is not a problem and everything could be serialized
17  * on hash bucket lock, but aside of having the benefit of common code,
18  * this allows to avoid doing the requeue when the task is already on the
19  * way out and taking the hash bucket lock of the original uaddr1 when the
20  * requeue has been completed.
21  *
22  * The following state transitions are valid:
23  *
24  * On the waiter side:
25  *   Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_IGNORE
26  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_WAIT
27  *
28  * On the requeue side:
29  *   Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_INPROGRESS
30  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_DONE/LOCKED
31  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_NONE (requeue failed)
32  *   Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_DONE/LOCKED
33  *   Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_IGNORE (requeue failed)
34  *
35  * The requeue side ignores a waiter with state Q_REQUEUE_PI_IGNORE as this
36  * signals that the waiter is already on the way out. It also means that
37  * the waiter is still on the 'wait' futex, i.e. uaddr1.
38  *
39  * The waiter side signals early wakeup to the requeue side either through
40  * setting state to Q_REQUEUE_PI_IGNORE or to Q_REQUEUE_PI_WAIT depending
41  * on the current state. In case of Q_REQUEUE_PI_IGNORE it can immediately
42  * proceed to take the hash bucket lock of uaddr1. If it set state to WAIT,
43  * which means the wakeup is interleaving with a requeue in progress it has
44  * to wait for the requeue side to change the state. Either to DONE/LOCKED
45  * or to IGNORE. DONE/LOCKED means the waiter q is now on the uaddr2 futex
46  * and either blocked (DONE) or has acquired it (LOCKED). IGNORE is set by
47  * the requeue side when the requeue attempt failed via deadlock detection
48  * and therefore the waiter q is still on the uaddr1 futex.
49  */
50 enum {
51 	Q_REQUEUE_PI_NONE		=  0,
52 	Q_REQUEUE_PI_IGNORE,
53 	Q_REQUEUE_PI_IN_PROGRESS,
54 	Q_REQUEUE_PI_WAIT,
55 	Q_REQUEUE_PI_DONE,
56 	Q_REQUEUE_PI_LOCKED,
57 };
58 
59 const struct futex_q futex_q_init = {
60 	/* list gets initialized in futex_queue()*/
61 	.wake		= futex_wake_mark,
62 	.key		= FUTEX_KEY_INIT,
63 	.bitset		= FUTEX_BITSET_MATCH_ANY,
64 	.requeue_state	= ATOMIC_INIT(Q_REQUEUE_PI_NONE),
65 };
66 
67 /**
68  * requeue_futex() - Requeue a futex_q from one hb to another
69  * @q:		the futex_q to requeue
70  * @hb1:	the source hash_bucket
71  * @hb2:	the target hash_bucket
72  * @key2:	the new key for the requeued futex_q
73  */
74 static inline
75 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
76 		   struct futex_hash_bucket *hb2, union futex_key *key2)
77 {
78 
79 	/*
80 	 * If key1 and key2 hash to the same bucket, no need to
81 	 * requeue.
82 	 */
83 	if (likely(&hb1->chain != &hb2->chain)) {
84 		plist_del(&q->list, &hb1->chain);
85 		futex_hb_waiters_dec(hb1);
86 		futex_hb_waiters_inc(hb2);
87 		plist_add(&q->list, &hb2->chain);
88 		q->lock_ptr = &hb2->lock;
89 	}
90 	q->key = *key2;
91 }
92 
93 static inline bool futex_requeue_pi_prepare(struct futex_q *q,
94 					    struct futex_pi_state *pi_state)
95 {
96 	int old, new;
97 
98 	/*
99 	 * Set state to Q_REQUEUE_PI_IN_PROGRESS unless an early wakeup has
100 	 * already set Q_REQUEUE_PI_IGNORE to signal that requeue should
101 	 * ignore the waiter.
102 	 */
103 	old = atomic_read_acquire(&q->requeue_state);
104 	do {
105 		if (old == Q_REQUEUE_PI_IGNORE)
106 			return false;
107 
108 		/*
109 		 * futex_proxy_trylock_atomic() might have set it to
110 		 * IN_PROGRESS and a interleaved early wake to WAIT.
111 		 *
112 		 * It was considered to have an extra state for that
113 		 * trylock, but that would just add more conditionals
114 		 * all over the place for a dubious value.
115 		 */
116 		if (old != Q_REQUEUE_PI_NONE)
117 			break;
118 
119 		new = Q_REQUEUE_PI_IN_PROGRESS;
120 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
121 
122 	q->pi_state = pi_state;
123 	return true;
124 }
125 
126 static inline void futex_requeue_pi_complete(struct futex_q *q, int locked)
127 {
128 	int old, new;
129 
130 	old = atomic_read_acquire(&q->requeue_state);
131 	do {
132 		if (old == Q_REQUEUE_PI_IGNORE)
133 			return;
134 
135 		if (locked >= 0) {
136 			/* Requeue succeeded. Set DONE or LOCKED */
137 			WARN_ON_ONCE(old != Q_REQUEUE_PI_IN_PROGRESS &&
138 				     old != Q_REQUEUE_PI_WAIT);
139 			new = Q_REQUEUE_PI_DONE + locked;
140 		} else if (old == Q_REQUEUE_PI_IN_PROGRESS) {
141 			/* Deadlock, no early wakeup interleave */
142 			new = Q_REQUEUE_PI_NONE;
143 		} else {
144 			/* Deadlock, early wakeup interleave. */
145 			WARN_ON_ONCE(old != Q_REQUEUE_PI_WAIT);
146 			new = Q_REQUEUE_PI_IGNORE;
147 		}
148 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
149 
150 #ifdef CONFIG_PREEMPT_RT
151 	/* If the waiter interleaved with the requeue let it know */
152 	if (unlikely(old == Q_REQUEUE_PI_WAIT))
153 		rcuwait_wake_up(&q->requeue_wait);
154 #endif
155 }
156 
157 static inline int futex_requeue_pi_wakeup_sync(struct futex_q *q)
158 {
159 	int old, new;
160 
161 	old = atomic_read_acquire(&q->requeue_state);
162 	do {
163 		/* Is requeue done already? */
164 		if (old >= Q_REQUEUE_PI_DONE)
165 			return old;
166 
167 		/*
168 		 * If not done, then tell the requeue code to either ignore
169 		 * the waiter or to wake it up once the requeue is done.
170 		 */
171 		new = Q_REQUEUE_PI_WAIT;
172 		if (old == Q_REQUEUE_PI_NONE)
173 			new = Q_REQUEUE_PI_IGNORE;
174 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
175 
176 	/* If the requeue was in progress, wait for it to complete */
177 	if (old == Q_REQUEUE_PI_IN_PROGRESS) {
178 #ifdef CONFIG_PREEMPT_RT
179 		rcuwait_wait_event(&q->requeue_wait,
180 				   atomic_read(&q->requeue_state) != Q_REQUEUE_PI_WAIT,
181 				   TASK_UNINTERRUPTIBLE);
182 #else
183 		(void)atomic_cond_read_relaxed(&q->requeue_state, VAL != Q_REQUEUE_PI_WAIT);
184 #endif
185 	}
186 
187 	/*
188 	 * Requeue is now either prohibited or complete. Reread state
189 	 * because during the wait above it might have changed. Nothing
190 	 * will modify q->requeue_state after this point.
191 	 */
192 	return atomic_read(&q->requeue_state);
193 }
194 
195 /**
196  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
197  * @q:		the futex_q
198  * @key:	the key of the requeue target futex
199  * @hb:		the hash_bucket of the requeue target futex
200  *
201  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
202  * target futex if it is uncontended or via a lock steal.
203  *
204  * 1) Set @q::key to the requeue target futex key so the waiter can detect
205  *    the wakeup on the right futex.
206  *
207  * 2) Dequeue @q from the hash bucket.
208  *
209  * 3) Set @q::rt_waiter to NULL so the woken up task can detect atomic lock
210  *    acquisition.
211  *
212  * 4) Set the q->lock_ptr to the requeue target hb->lock for the case that
213  *    the waiter has to fixup the pi state.
214  *
215  * 5) Complete the requeue state so the waiter can make progress. After
216  *    this point the waiter task can return from the syscall immediately in
217  *    case that the pi state does not have to be fixed up.
218  *
219  * 6) Wake the waiter task.
220  *
221  * Must be called with both q->lock_ptr and hb->lock held.
222  */
223 static inline
224 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
225 			   struct futex_hash_bucket *hb)
226 {
227 	q->key = *key;
228 
229 	__futex_unqueue(q);
230 
231 	WARN_ON(!q->rt_waiter);
232 	q->rt_waiter = NULL;
233 
234 	q->lock_ptr = &hb->lock;
235 
236 	/* Signal locked state to the waiter */
237 	futex_requeue_pi_complete(q, 1);
238 	wake_up_state(q->task, TASK_NORMAL);
239 }
240 
241 /**
242  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
243  * @pifutex:		the user address of the to futex
244  * @hb1:		the from futex hash bucket, must be locked by the caller
245  * @hb2:		the to futex hash bucket, must be locked by the caller
246  * @key1:		the from futex key
247  * @key2:		the to futex key
248  * @ps:			address to store the pi_state pointer
249  * @exiting:		Pointer to store the task pointer of the owner task
250  *			which is in the middle of exiting
251  * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
252  *
253  * Try and get the lock on behalf of the top waiter if we can do it atomically.
254  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
255  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
256  * hb1 and hb2 must be held by the caller.
257  *
258  * @exiting is only set when the return value is -EBUSY. If so, this holds
259  * a refcount on the exiting task on return and the caller needs to drop it
260  * after waiting for the exit to complete.
261  *
262  * Return:
263  *  -  0 - failed to acquire the lock atomically;
264  *  - >0 - acquired the lock, return value is vpid of the top_waiter
265  *  - <0 - error
266  */
267 static int
268 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
269 			   struct futex_hash_bucket *hb2, union futex_key *key1,
270 			   union futex_key *key2, struct futex_pi_state **ps,
271 			   struct task_struct **exiting, int set_waiters)
272 {
273 	struct futex_q *top_waiter;
274 	u32 curval;
275 	int ret;
276 
277 	if (futex_get_value_locked(&curval, pifutex))
278 		return -EFAULT;
279 
280 	if (unlikely(should_fail_futex(true)))
281 		return -EFAULT;
282 
283 	/*
284 	 * Find the top_waiter and determine if there are additional waiters.
285 	 * If the caller intends to requeue more than 1 waiter to pifutex,
286 	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
287 	 * as we have means to handle the possible fault.  If not, don't set
288 	 * the bit unnecessarily as it will force the subsequent unlock to enter
289 	 * the kernel.
290 	 */
291 	top_waiter = futex_top_waiter(hb1, key1);
292 
293 	/* There are no waiters, nothing for us to do. */
294 	if (!top_waiter)
295 		return 0;
296 
297 	/*
298 	 * Ensure that this is a waiter sitting in futex_wait_requeue_pi()
299 	 * and waiting on the 'waitqueue' futex which is always !PI.
300 	 */
301 	if (!top_waiter->rt_waiter || top_waiter->pi_state)
302 		return -EINVAL;
303 
304 	/* Ensure we requeue to the expected futex. */
305 	if (!futex_match(top_waiter->requeue_pi_key, key2))
306 		return -EINVAL;
307 
308 	/* Ensure that this does not race against an early wakeup */
309 	if (!futex_requeue_pi_prepare(top_waiter, NULL))
310 		return -EAGAIN;
311 
312 	/*
313 	 * Try to take the lock for top_waiter and set the FUTEX_WAITERS bit
314 	 * in the contended case or if @set_waiters is true.
315 	 *
316 	 * In the contended case PI state is attached to the lock owner. If
317 	 * the user space lock can be acquired then PI state is attached to
318 	 * the new owner (@top_waiter->task) when @set_waiters is true.
319 	 */
320 	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
321 				   exiting, set_waiters);
322 	if (ret == 1) {
323 		/*
324 		 * Lock was acquired in user space and PI state was
325 		 * attached to @top_waiter->task. That means state is fully
326 		 * consistent and the waiter can return to user space
327 		 * immediately after the wakeup.
328 		 */
329 		requeue_pi_wake_futex(top_waiter, key2, hb2);
330 	} else if (ret < 0) {
331 		/* Rewind top_waiter::requeue_state */
332 		futex_requeue_pi_complete(top_waiter, ret);
333 	} else {
334 		/*
335 		 * futex_lock_pi_atomic() did not acquire the user space
336 		 * futex, but managed to establish the proxy lock and pi
337 		 * state. top_waiter::requeue_state cannot be fixed up here
338 		 * because the waiter is not enqueued on the rtmutex
339 		 * yet. This is handled at the callsite depending on the
340 		 * result of rt_mutex_start_proxy_lock() which is
341 		 * guaranteed to be reached with this function returning 0.
342 		 */
343 	}
344 	return ret;
345 }
346 
347 /**
348  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
349  * @uaddr1:	source futex user address
350  * @flags1:	futex flags (FLAGS_SHARED, etc.)
351  * @uaddr2:	target futex user address
352  * @flags2:	futex flags (FLAGS_SHARED, etc.)
353  * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
354  * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
355  * @cmpval:	@uaddr1 expected value (or %NULL)
356  * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
357  *		pi futex (pi to pi requeue is not supported)
358  *
359  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
360  * uaddr2 atomically on behalf of the top waiter.
361  *
362  * Return:
363  *  - >=0 - on success, the number of tasks requeued or woken;
364  *  -  <0 - on error
365  */
366 int futex_requeue(u32 __user *uaddr1, unsigned int flags1,
367 		  u32 __user *uaddr2, unsigned int flags2,
368 		  int nr_wake, int nr_requeue, u32 *cmpval, int requeue_pi)
369 {
370 	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
371 	int task_count = 0, ret;
372 	struct futex_pi_state *pi_state = NULL;
373 	struct futex_hash_bucket *hb1, *hb2;
374 	struct futex_q *this, *next;
375 	DEFINE_WAKE_Q(wake_q);
376 
377 	if (nr_wake < 0 || nr_requeue < 0)
378 		return -EINVAL;
379 
380 	/*
381 	 * When PI not supported: return -ENOSYS if requeue_pi is true,
382 	 * consequently the compiler knows requeue_pi is always false past
383 	 * this point which will optimize away all the conditional code
384 	 * further down.
385 	 */
386 	if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
387 		return -ENOSYS;
388 
389 	if (requeue_pi) {
390 		/*
391 		 * Requeue PI only works on two distinct uaddrs. This
392 		 * check is only valid for private futexes. See below.
393 		 */
394 		if (uaddr1 == uaddr2)
395 			return -EINVAL;
396 
397 		/*
398 		 * futex_requeue() allows the caller to define the number
399 		 * of waiters to wake up via the @nr_wake argument. With
400 		 * REQUEUE_PI, waking up more than one waiter is creating
401 		 * more problems than it solves. Waking up a waiter makes
402 		 * only sense if the PI futex @uaddr2 is uncontended as
403 		 * this allows the requeue code to acquire the futex
404 		 * @uaddr2 before waking the waiter. The waiter can then
405 		 * return to user space without further action. A secondary
406 		 * wakeup would just make the futex_wait_requeue_pi()
407 		 * handling more complex, because that code would have to
408 		 * look up pi_state and do more or less all the handling
409 		 * which the requeue code has to do for the to be requeued
410 		 * waiters. So restrict the number of waiters to wake to
411 		 * one, and only wake it up when the PI futex is
412 		 * uncontended. Otherwise requeue it and let the unlock of
413 		 * the PI futex handle the wakeup.
414 		 *
415 		 * All REQUEUE_PI users, e.g. pthread_cond_signal() and
416 		 * pthread_cond_broadcast() must use nr_wake=1.
417 		 */
418 		if (nr_wake != 1)
419 			return -EINVAL;
420 
421 		/*
422 		 * requeue_pi requires a pi_state, try to allocate it now
423 		 * without any locks in case it fails.
424 		 */
425 		if (refill_pi_state_cache())
426 			return -ENOMEM;
427 	}
428 
429 retry:
430 	ret = get_futex_key(uaddr1, flags1, &key1, FUTEX_READ);
431 	if (unlikely(ret != 0))
432 		return ret;
433 	ret = get_futex_key(uaddr2, flags2, &key2,
434 			    requeue_pi ? FUTEX_WRITE : FUTEX_READ);
435 	if (unlikely(ret != 0))
436 		return ret;
437 
438 	/*
439 	 * The check above which compares uaddrs is not sufficient for
440 	 * shared futexes. We need to compare the keys:
441 	 */
442 	if (requeue_pi && futex_match(&key1, &key2))
443 		return -EINVAL;
444 
445 	hb1 = futex_hash(&key1);
446 	hb2 = futex_hash(&key2);
447 
448 retry_private:
449 	futex_hb_waiters_inc(hb2);
450 	double_lock_hb(hb1, hb2);
451 
452 	if (likely(cmpval != NULL)) {
453 		u32 curval;
454 
455 		ret = futex_get_value_locked(&curval, uaddr1);
456 
457 		if (unlikely(ret)) {
458 			double_unlock_hb(hb1, hb2);
459 			futex_hb_waiters_dec(hb2);
460 
461 			ret = get_user(curval, uaddr1);
462 			if (ret)
463 				return ret;
464 
465 			if (!(flags1 & FLAGS_SHARED))
466 				goto retry_private;
467 
468 			goto retry;
469 		}
470 		if (curval != *cmpval) {
471 			ret = -EAGAIN;
472 			goto out_unlock;
473 		}
474 	}
475 
476 	if (requeue_pi) {
477 		struct task_struct *exiting = NULL;
478 
479 		/*
480 		 * Attempt to acquire uaddr2 and wake the top waiter. If we
481 		 * intend to requeue waiters, force setting the FUTEX_WAITERS
482 		 * bit.  We force this here where we are able to easily handle
483 		 * faults rather in the requeue loop below.
484 		 *
485 		 * Updates topwaiter::requeue_state if a top waiter exists.
486 		 */
487 		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
488 						 &key2, &pi_state,
489 						 &exiting, nr_requeue);
490 
491 		/*
492 		 * At this point the top_waiter has either taken uaddr2 or
493 		 * is waiting on it. In both cases pi_state has been
494 		 * established and an initial refcount on it. In case of an
495 		 * error there's nothing.
496 		 *
497 		 * The top waiter's requeue_state is up to date:
498 		 *
499 		 *  - If the lock was acquired atomically (ret == 1), then
500 		 *    the state is Q_REQUEUE_PI_LOCKED.
501 		 *
502 		 *    The top waiter has been dequeued and woken up and can
503 		 *    return to user space immediately. The kernel/user
504 		 *    space state is consistent. In case that there must be
505 		 *    more waiters requeued the WAITERS bit in the user
506 		 *    space futex is set so the top waiter task has to go
507 		 *    into the syscall slowpath to unlock the futex. This
508 		 *    will block until this requeue operation has been
509 		 *    completed and the hash bucket locks have been
510 		 *    dropped.
511 		 *
512 		 *  - If the trylock failed with an error (ret < 0) then
513 		 *    the state is either Q_REQUEUE_PI_NONE, i.e. "nothing
514 		 *    happened", or Q_REQUEUE_PI_IGNORE when there was an
515 		 *    interleaved early wakeup.
516 		 *
517 		 *  - If the trylock did not succeed (ret == 0) then the
518 		 *    state is either Q_REQUEUE_PI_IN_PROGRESS or
519 		 *    Q_REQUEUE_PI_WAIT if an early wakeup interleaved.
520 		 *    This will be cleaned up in the loop below, which
521 		 *    cannot fail because futex_proxy_trylock_atomic() did
522 		 *    the same sanity checks for requeue_pi as the loop
523 		 *    below does.
524 		 */
525 		switch (ret) {
526 		case 0:
527 			/* We hold a reference on the pi state. */
528 			break;
529 
530 		case 1:
531 			/*
532 			 * futex_proxy_trylock_atomic() acquired the user space
533 			 * futex. Adjust task_count.
534 			 */
535 			task_count++;
536 			ret = 0;
537 			break;
538 
539 		/*
540 		 * If the above failed, then pi_state is NULL and
541 		 * waiter::requeue_state is correct.
542 		 */
543 		case -EFAULT:
544 			double_unlock_hb(hb1, hb2);
545 			futex_hb_waiters_dec(hb2);
546 			ret = fault_in_user_writeable(uaddr2);
547 			if (!ret)
548 				goto retry;
549 			return ret;
550 		case -EBUSY:
551 		case -EAGAIN:
552 			/*
553 			 * Two reasons for this:
554 			 * - EBUSY: Owner is exiting and we just wait for the
555 			 *   exit to complete.
556 			 * - EAGAIN: The user space value changed.
557 			 */
558 			double_unlock_hb(hb1, hb2);
559 			futex_hb_waiters_dec(hb2);
560 			/*
561 			 * Handle the case where the owner is in the middle of
562 			 * exiting. Wait for the exit to complete otherwise
563 			 * this task might loop forever, aka. live lock.
564 			 */
565 			wait_for_owner_exiting(ret, exiting);
566 			cond_resched();
567 			goto retry;
568 		default:
569 			goto out_unlock;
570 		}
571 	}
572 
573 	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
574 		if (task_count - nr_wake >= nr_requeue)
575 			break;
576 
577 		if (!futex_match(&this->key, &key1))
578 			continue;
579 
580 		/*
581 		 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
582 		 * be paired with each other and no other futex ops.
583 		 *
584 		 * We should never be requeueing a futex_q with a pi_state,
585 		 * which is awaiting a futex_unlock_pi().
586 		 */
587 		if ((requeue_pi && !this->rt_waiter) ||
588 		    (!requeue_pi && this->rt_waiter) ||
589 		    this->pi_state) {
590 			ret = -EINVAL;
591 			break;
592 		}
593 
594 		/* Plain futexes just wake or requeue and are done */
595 		if (!requeue_pi) {
596 			if (++task_count <= nr_wake)
597 				this->wake(&wake_q, this);
598 			else
599 				requeue_futex(this, hb1, hb2, &key2);
600 			continue;
601 		}
602 
603 		/* Ensure we requeue to the expected futex for requeue_pi. */
604 		if (!futex_match(this->requeue_pi_key, &key2)) {
605 			ret = -EINVAL;
606 			break;
607 		}
608 
609 		/*
610 		 * Requeue nr_requeue waiters and possibly one more in the case
611 		 * of requeue_pi if we couldn't acquire the lock atomically.
612 		 *
613 		 * Prepare the waiter to take the rt_mutex. Take a refcount
614 		 * on the pi_state and store the pointer in the futex_q
615 		 * object of the waiter.
616 		 */
617 		get_pi_state(pi_state);
618 
619 		/* Don't requeue when the waiter is already on the way out. */
620 		if (!futex_requeue_pi_prepare(this, pi_state)) {
621 			/*
622 			 * Early woken waiter signaled that it is on the
623 			 * way out. Drop the pi_state reference and try the
624 			 * next waiter. @this->pi_state is still NULL.
625 			 */
626 			put_pi_state(pi_state);
627 			continue;
628 		}
629 
630 		ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
631 						this->rt_waiter,
632 						this->task);
633 
634 		if (ret == 1) {
635 			/*
636 			 * We got the lock. We do neither drop the refcount
637 			 * on pi_state nor clear this->pi_state because the
638 			 * waiter needs the pi_state for cleaning up the
639 			 * user space value. It will drop the refcount
640 			 * after doing so. this::requeue_state is updated
641 			 * in the wakeup as well.
642 			 */
643 			requeue_pi_wake_futex(this, &key2, hb2);
644 			task_count++;
645 		} else if (!ret) {
646 			/* Waiter is queued, move it to hb2 */
647 			requeue_futex(this, hb1, hb2, &key2);
648 			futex_requeue_pi_complete(this, 0);
649 			task_count++;
650 		} else {
651 			/*
652 			 * rt_mutex_start_proxy_lock() detected a potential
653 			 * deadlock when we tried to queue that waiter.
654 			 * Drop the pi_state reference which we took above
655 			 * and remove the pointer to the state from the
656 			 * waiters futex_q object.
657 			 */
658 			this->pi_state = NULL;
659 			put_pi_state(pi_state);
660 			futex_requeue_pi_complete(this, ret);
661 			/*
662 			 * We stop queueing more waiters and let user space
663 			 * deal with the mess.
664 			 */
665 			break;
666 		}
667 	}
668 
669 	/*
670 	 * We took an extra initial reference to the pi_state in
671 	 * futex_proxy_trylock_atomic(). We need to drop it here again.
672 	 */
673 	put_pi_state(pi_state);
674 
675 out_unlock:
676 	double_unlock_hb(hb1, hb2);
677 	wake_up_q(&wake_q);
678 	futex_hb_waiters_dec(hb2);
679 	return ret ? ret : task_count;
680 }
681 
682 /**
683  * handle_early_requeue_pi_wakeup() - Handle early wakeup on the initial futex
684  * @hb:		the hash_bucket futex_q was original enqueued on
685  * @q:		the futex_q woken while waiting to be requeued
686  * @timeout:	the timeout associated with the wait (NULL if none)
687  *
688  * Determine the cause for the early wakeup.
689  *
690  * Return:
691  *  -EWOULDBLOCK or -ETIMEDOUT or -ERESTARTNOINTR
692  */
693 static inline
694 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
695 				   struct futex_q *q,
696 				   struct hrtimer_sleeper *timeout)
697 {
698 	int ret;
699 
700 	/*
701 	 * With the hb lock held, we avoid races while we process the wakeup.
702 	 * We only need to hold hb (and not hb2) to ensure atomicity as the
703 	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
704 	 * It can't be requeued from uaddr2 to something else since we don't
705 	 * support a PI aware source futex for requeue.
706 	 */
707 	WARN_ON_ONCE(&hb->lock != q->lock_ptr);
708 
709 	/*
710 	 * We were woken prior to requeue by a timeout or a signal.
711 	 * Unqueue the futex_q and determine which it was.
712 	 */
713 	plist_del(&q->list, &hb->chain);
714 	futex_hb_waiters_dec(hb);
715 
716 	/* Handle spurious wakeups gracefully */
717 	ret = -EWOULDBLOCK;
718 	if (timeout && !timeout->task)
719 		ret = -ETIMEDOUT;
720 	else if (signal_pending(current))
721 		ret = -ERESTARTNOINTR;
722 	return ret;
723 }
724 
725 /**
726  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
727  * @uaddr:	the futex we initially wait on (non-pi)
728  * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
729  *		the same type, no requeueing from private to shared, etc.
730  * @val:	the expected value of uaddr
731  * @abs_time:	absolute timeout
732  * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
733  * @uaddr2:	the pi futex we will take prior to returning to user-space
734  *
735  * The caller will wait on uaddr and will be requeued by futex_requeue() to
736  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
737  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
738  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
739  * without one, the pi logic would not know which task to boost/deboost, if
740  * there was a need to.
741  *
742  * We call schedule in futex_wait_queue() when we enqueue and return there
743  * via the following--
744  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
745  * 2) wakeup on uaddr2 after a requeue
746  * 3) signal
747  * 4) timeout
748  *
749  * If 3, cleanup and return -ERESTARTNOINTR.
750  *
751  * If 2, we may then block on trying to take the rt_mutex and return via:
752  * 5) successful lock
753  * 6) signal
754  * 7) timeout
755  * 8) other lock acquisition failure
756  *
757  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
758  *
759  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
760  *
761  * Return:
762  *  -  0 - On success;
763  *  - <0 - On error
764  */
765 int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
766 			  u32 val, ktime_t *abs_time, u32 bitset,
767 			  u32 __user *uaddr2)
768 {
769 	struct hrtimer_sleeper timeout, *to;
770 	struct rt_mutex_waiter rt_waiter;
771 	struct futex_hash_bucket *hb;
772 	union futex_key key2 = FUTEX_KEY_INIT;
773 	struct futex_q q = futex_q_init;
774 	struct rt_mutex_base *pi_mutex;
775 	int res, ret;
776 
777 	if (!IS_ENABLED(CONFIG_FUTEX_PI))
778 		return -ENOSYS;
779 
780 	if (uaddr == uaddr2)
781 		return -EINVAL;
782 
783 	if (!bitset)
784 		return -EINVAL;
785 
786 	to = futex_setup_timer(abs_time, &timeout, flags,
787 			       current->timer_slack_ns);
788 
789 	/*
790 	 * The waiter is allocated on our stack, manipulated by the requeue
791 	 * code while we sleep on uaddr.
792 	 */
793 	rt_mutex_init_waiter(&rt_waiter);
794 
795 	ret = get_futex_key(uaddr2, flags, &key2, FUTEX_WRITE);
796 	if (unlikely(ret != 0))
797 		goto out;
798 
799 	q.bitset = bitset;
800 	q.rt_waiter = &rt_waiter;
801 	q.requeue_pi_key = &key2;
802 
803 	/*
804 	 * Prepare to wait on uaddr. On success, it holds hb->lock and q
805 	 * is initialized.
806 	 */
807 	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
808 	if (ret)
809 		goto out;
810 
811 	/*
812 	 * The check above which compares uaddrs is not sufficient for
813 	 * shared futexes. We need to compare the keys:
814 	 */
815 	if (futex_match(&q.key, &key2)) {
816 		futex_q_unlock(hb);
817 		ret = -EINVAL;
818 		goto out;
819 	}
820 
821 	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
822 	futex_wait_queue(hb, &q, to);
823 
824 	switch (futex_requeue_pi_wakeup_sync(&q)) {
825 	case Q_REQUEUE_PI_IGNORE:
826 		/* The waiter is still on uaddr1 */
827 		spin_lock(&hb->lock);
828 		ret = handle_early_requeue_pi_wakeup(hb, &q, to);
829 		spin_unlock(&hb->lock);
830 		break;
831 
832 	case Q_REQUEUE_PI_LOCKED:
833 		/* The requeue acquired the lock */
834 		if (q.pi_state && (q.pi_state->owner != current)) {
835 			spin_lock(q.lock_ptr);
836 			ret = fixup_pi_owner(uaddr2, &q, true);
837 			/*
838 			 * Drop the reference to the pi state which the
839 			 * requeue_pi() code acquired for us.
840 			 */
841 			put_pi_state(q.pi_state);
842 			spin_unlock(q.lock_ptr);
843 			/*
844 			 * Adjust the return value. It's either -EFAULT or
845 			 * success (1) but the caller expects 0 for success.
846 			 */
847 			ret = ret < 0 ? ret : 0;
848 		}
849 		break;
850 
851 	case Q_REQUEUE_PI_DONE:
852 		/* Requeue completed. Current is 'pi_blocked_on' the rtmutex */
853 		pi_mutex = &q.pi_state->pi_mutex;
854 		ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
855 
856 		/*
857 		 * See futex_unlock_pi()'s cleanup: comment.
858 		 */
859 		if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
860 			ret = 0;
861 
862 		spin_lock(q.lock_ptr);
863 		debug_rt_mutex_free_waiter(&rt_waiter);
864 		/*
865 		 * Fixup the pi_state owner and possibly acquire the lock if we
866 		 * haven't already.
867 		 */
868 		res = fixup_pi_owner(uaddr2, &q, !ret);
869 		/*
870 		 * If fixup_pi_owner() returned an error, propagate that.  If it
871 		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
872 		 */
873 		if (res)
874 			ret = (res < 0) ? res : 0;
875 
876 		futex_unqueue_pi(&q);
877 		spin_unlock(q.lock_ptr);
878 
879 		if (ret == -EINTR) {
880 			/*
881 			 * We've already been requeued, but cannot restart
882 			 * by calling futex_lock_pi() directly. We could
883 			 * restart this syscall, but it would detect that
884 			 * the user space "val" changed and return
885 			 * -EWOULDBLOCK.  Save the overhead of the restart
886 			 * and return -EWOULDBLOCK directly.
887 			 */
888 			ret = -EWOULDBLOCK;
889 		}
890 		break;
891 	default:
892 		BUG();
893 	}
894 
895 out:
896 	if (to) {
897 		hrtimer_cancel(&to->timer);
898 		destroy_hrtimer_on_stack(&to->timer);
899 	}
900 	return ret;
901 }
902 
903