xref: /linux/kernel/futex/requeue.c (revision 764e77d868a5b932c709e20ddb5993f9111a841c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 
3 #include <linux/plist.h>
4 #include <linux/sched/signal.h>
5 
6 #include "futex.h"
7 #include "../locking/rtmutex_common.h"
8 
9 /*
10  * On PREEMPT_RT, the hash bucket lock is a 'sleeping' spinlock with an
11  * underlying rtmutex. The task which is about to be requeued could have
12  * just woken up (timeout, signal). After the wake up the task has to
13  * acquire hash bucket lock, which is held by the requeue code.  As a task
14  * can only be blocked on _ONE_ rtmutex at a time, the proxy lock blocking
15  * and the hash bucket lock blocking would collide and corrupt state.
16  *
17  * On !PREEMPT_RT this is not a problem and everything could be serialized
18  * on hash bucket lock, but aside of having the benefit of common code,
19  * this allows to avoid doing the requeue when the task is already on the
20  * way out and taking the hash bucket lock of the original uaddr1 when the
21  * requeue has been completed.
22  *
23  * The following state transitions are valid:
24  *
25  * On the waiter side:
26  *   Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_IGNORE
27  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_WAIT
28  *
29  * On the requeue side:
30  *   Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_INPROGRESS
31  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_DONE/LOCKED
32  *   Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_NONE (requeue failed)
33  *   Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_DONE/LOCKED
34  *   Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_IGNORE (requeue failed)
35  *
36  * The requeue side ignores a waiter with state Q_REQUEUE_PI_IGNORE as this
37  * signals that the waiter is already on the way out. It also means that
38  * the waiter is still on the 'wait' futex, i.e. uaddr1.
39  *
40  * The waiter side signals early wakeup to the requeue side either through
41  * setting state to Q_REQUEUE_PI_IGNORE or to Q_REQUEUE_PI_WAIT depending
42  * on the current state. In case of Q_REQUEUE_PI_IGNORE it can immediately
43  * proceed to take the hash bucket lock of uaddr1. If it set state to WAIT,
44  * which means the wakeup is interleaving with a requeue in progress it has
45  * to wait for the requeue side to change the state. Either to DONE/LOCKED
46  * or to IGNORE. DONE/LOCKED means the waiter q is now on the uaddr2 futex
47  * and either blocked (DONE) or has acquired it (LOCKED). IGNORE is set by
48  * the requeue side when the requeue attempt failed via deadlock detection
49  * and therefore the waiter q is still on the uaddr1 futex.
50  */
51 enum {
52 	Q_REQUEUE_PI_NONE		=  0,
53 	Q_REQUEUE_PI_IGNORE,
54 	Q_REQUEUE_PI_IN_PROGRESS,
55 	Q_REQUEUE_PI_WAIT,
56 	Q_REQUEUE_PI_DONE,
57 	Q_REQUEUE_PI_LOCKED,
58 };
59 
60 const struct futex_q futex_q_init = {
61 	/* list gets initialized in futex_queue()*/
62 	.wake		= futex_wake_mark,
63 	.key		= FUTEX_KEY_INIT,
64 	.bitset		= FUTEX_BITSET_MATCH_ANY,
65 	.requeue_state	= ATOMIC_INIT(Q_REQUEUE_PI_NONE),
66 };
67 
68 /**
69  * requeue_futex() - Requeue a futex_q from one hb to another
70  * @q:		the futex_q to requeue
71  * @hb1:	the source hash_bucket
72  * @hb2:	the target hash_bucket
73  * @key2:	the new key for the requeued futex_q
74  */
75 static inline
76 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
77 		   struct futex_hash_bucket *hb2, union futex_key *key2)
78 {
79 
80 	/*
81 	 * If key1 and key2 hash to the same bucket, no need to
82 	 * requeue.
83 	 */
84 	if (likely(&hb1->chain != &hb2->chain)) {
85 		plist_del(&q->list, &hb1->chain);
86 		futex_hb_waiters_dec(hb1);
87 		futex_hb_waiters_inc(hb2);
88 		plist_add(&q->list, &hb2->chain);
89 		q->lock_ptr = &hb2->lock;
90 		/*
91 		 * hb1 and hb2 belong to the same futex_hash_bucket_private
92 		 * because if we managed get a reference on hb1 then it can't be
93 		 * replaced. Therefore we avoid put(hb1)+get(hb2) here.
94 		 */
95 	}
96 	q->key = *key2;
97 }
98 
99 static inline bool futex_requeue_pi_prepare(struct futex_q *q,
100 					    struct futex_pi_state *pi_state)
101 {
102 	int old, new;
103 
104 	/*
105 	 * Set state to Q_REQUEUE_PI_IN_PROGRESS unless an early wakeup has
106 	 * already set Q_REQUEUE_PI_IGNORE to signal that requeue should
107 	 * ignore the waiter.
108 	 */
109 	old = atomic_read_acquire(&q->requeue_state);
110 	do {
111 		if (old == Q_REQUEUE_PI_IGNORE)
112 			return false;
113 
114 		/*
115 		 * futex_proxy_trylock_atomic() might have set it to
116 		 * IN_PROGRESS and a interleaved early wake to WAIT.
117 		 *
118 		 * It was considered to have an extra state for that
119 		 * trylock, but that would just add more conditionals
120 		 * all over the place for a dubious value.
121 		 */
122 		if (old != Q_REQUEUE_PI_NONE)
123 			break;
124 
125 		new = Q_REQUEUE_PI_IN_PROGRESS;
126 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
127 
128 	q->pi_state = pi_state;
129 	return true;
130 }
131 
132 static inline void futex_requeue_pi_complete(struct futex_q *q, int locked)
133 {
134 	int old, new;
135 
136 	old = atomic_read_acquire(&q->requeue_state);
137 	do {
138 		if (old == Q_REQUEUE_PI_IGNORE)
139 			return;
140 
141 		if (locked >= 0) {
142 			/* Requeue succeeded. Set DONE or LOCKED */
143 			WARN_ON_ONCE(old != Q_REQUEUE_PI_IN_PROGRESS &&
144 				     old != Q_REQUEUE_PI_WAIT);
145 			new = Q_REQUEUE_PI_DONE + locked;
146 		} else if (old == Q_REQUEUE_PI_IN_PROGRESS) {
147 			/* Deadlock, no early wakeup interleave */
148 			new = Q_REQUEUE_PI_NONE;
149 		} else {
150 			/* Deadlock, early wakeup interleave. */
151 			WARN_ON_ONCE(old != Q_REQUEUE_PI_WAIT);
152 			new = Q_REQUEUE_PI_IGNORE;
153 		}
154 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
155 
156 #ifdef CONFIG_PREEMPT_RT
157 	/* If the waiter interleaved with the requeue let it know */
158 	if (unlikely(old == Q_REQUEUE_PI_WAIT))
159 		rcuwait_wake_up(&q->requeue_wait);
160 #endif
161 }
162 
163 static inline int futex_requeue_pi_wakeup_sync(struct futex_q *q)
164 {
165 	int old, new;
166 
167 	old = atomic_read_acquire(&q->requeue_state);
168 	do {
169 		/* Is requeue done already? */
170 		if (old >= Q_REQUEUE_PI_DONE)
171 			return old;
172 
173 		/*
174 		 * If not done, then tell the requeue code to either ignore
175 		 * the waiter or to wake it up once the requeue is done.
176 		 */
177 		new = Q_REQUEUE_PI_WAIT;
178 		if (old == Q_REQUEUE_PI_NONE)
179 			new = Q_REQUEUE_PI_IGNORE;
180 	} while (!atomic_try_cmpxchg(&q->requeue_state, &old, new));
181 
182 	/* If the requeue was in progress, wait for it to complete */
183 	if (old == Q_REQUEUE_PI_IN_PROGRESS) {
184 #ifdef CONFIG_PREEMPT_RT
185 		rcuwait_wait_event(&q->requeue_wait,
186 				   atomic_read(&q->requeue_state) != Q_REQUEUE_PI_WAIT,
187 				   TASK_UNINTERRUPTIBLE);
188 #else
189 		(void)atomic_cond_read_relaxed(&q->requeue_state, VAL != Q_REQUEUE_PI_WAIT);
190 #endif
191 	}
192 
193 	/*
194 	 * Requeue is now either prohibited or complete. Reread state
195 	 * because during the wait above it might have changed. Nothing
196 	 * will modify q->requeue_state after this point.
197 	 */
198 	return atomic_read(&q->requeue_state);
199 }
200 
201 /**
202  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
203  * @q:		the futex_q
204  * @key:	the key of the requeue target futex
205  * @hb:		the hash_bucket of the requeue target futex
206  *
207  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
208  * target futex if it is uncontended or via a lock steal.
209  *
210  * 1) Set @q::key to the requeue target futex key so the waiter can detect
211  *    the wakeup on the right futex.
212  *
213  * 2) Dequeue @q from the hash bucket.
214  *
215  * 3) Set @q::rt_waiter to NULL so the woken up task can detect atomic lock
216  *    acquisition.
217  *
218  * 4) Set the q->lock_ptr to the requeue target hb->lock for the case that
219  *    the waiter has to fixup the pi state.
220  *
221  * 5) Complete the requeue state so the waiter can make progress. After
222  *    this point the waiter task can return from the syscall immediately in
223  *    case that the pi state does not have to be fixed up.
224  *
225  * 6) Wake the waiter task.
226  *
227  * Must be called with both q->lock_ptr and hb->lock held.
228  */
229 static inline
230 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
231 			   struct futex_hash_bucket *hb)
232 {
233 	struct task_struct *task;
234 
235 	q->key = *key;
236 	__futex_unqueue(q);
237 
238 	WARN_ON(!q->rt_waiter);
239 	q->rt_waiter = NULL;
240 	/*
241 	 * Acquire a reference for the waiter to ensure valid
242 	 * futex_q::lock_ptr.
243 	 */
244 	if (futex_key_is_private(key))
245 		q->drop_fph = futex_private_hash(key->private.mm);
246 	q->lock_ptr = &hb->lock;
247 	task = READ_ONCE(q->task);
248 
249 	/* Signal locked state to the waiter */
250 	futex_requeue_pi_complete(q, 1);
251 	wake_up_state(task, TASK_NORMAL);
252 }
253 
254 /**
255  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
256  * @pifutex:		the user address of the to futex
257  * @hb1:		the from futex hash bucket, must be locked by the caller
258  * @hb2:		the to futex hash bucket, must be locked by the caller
259  * @key1:		the from futex key
260  * @key2:		the to futex key
261  * @ps:			address to store the pi_state pointer
262  * @exiting:		Pointer to store the task pointer of the owner task
263  *			which is in the middle of exiting
264  * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
265  *
266  * Try and get the lock on behalf of the top waiter if we can do it atomically.
267  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
268  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
269  * hb1 and hb2 must be held by the caller.
270  *
271  * @exiting is only set when the return value is -EBUSY. If so, this holds
272  * a refcount on the exiting task on return and the caller needs to drop it
273  * after waiting for the exit to complete.
274  *
275  * Return:
276  *  -  0 - failed to acquire the lock atomically;
277  *  - >0 - acquired the lock, return value is vpid of the top_waiter
278  *  - <0 - error
279  */
280 static int
281 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
282 			   struct futex_hash_bucket *hb2, union futex_key *key1,
283 			   union futex_key *key2, struct futex_pi_state **ps,
284 			   struct task_struct **exiting, int set_waiters)
285 {
286 	struct futex_q *top_waiter;
287 	u32 curval;
288 	int ret;
289 
290 	if (futex_get_value_locked(&curval, pifutex))
291 		return -EFAULT;
292 
293 	if (unlikely(should_fail_futex(true)))
294 		return -EFAULT;
295 
296 	/*
297 	 * Find the top_waiter and determine if there are additional waiters.
298 	 * If the caller intends to requeue more than 1 waiter to pifutex,
299 	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
300 	 * as we have means to handle the possible fault.  If not, don't set
301 	 * the bit unnecessarily as it will force the subsequent unlock to enter
302 	 * the kernel.
303 	 */
304 	top_waiter = futex_top_waiter(hb1, key1);
305 
306 	/* There are no waiters, nothing for us to do. */
307 	if (!top_waiter)
308 		return 0;
309 
310 	/*
311 	 * Ensure that this is a waiter sitting in futex_wait_requeue_pi()
312 	 * and waiting on the 'waitqueue' futex which is always !PI.
313 	 */
314 	if (!top_waiter->rt_waiter || top_waiter->pi_state)
315 		return -EINVAL;
316 
317 	/* Ensure we requeue to the expected futex. */
318 	if (!futex_match(top_waiter->requeue_pi_key, key2))
319 		return -EINVAL;
320 
321 	/* Ensure that this does not race against an early wakeup */
322 	if (!futex_requeue_pi_prepare(top_waiter, NULL)) {
323 		plist_del(&top_waiter->list, &hb1->chain);
324 		futex_hb_waiters_dec(hb1);
325 		return -EAGAIN;
326 	}
327 
328 	/*
329 	 * Try to take the lock for top_waiter and set the FUTEX_WAITERS bit
330 	 * in the contended case or if @set_waiters is true.
331 	 *
332 	 * In the contended case PI state is attached to the lock owner. If
333 	 * the user space lock can be acquired then PI state is attached to
334 	 * the new owner (@top_waiter->task) when @set_waiters is true.
335 	 */
336 	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
337 				   exiting, set_waiters);
338 	if (ret == 1) {
339 		/*
340 		 * Lock was acquired in user space and PI state was
341 		 * attached to @top_waiter->task. That means state is fully
342 		 * consistent and the waiter can return to user space
343 		 * immediately after the wakeup.
344 		 */
345 		requeue_pi_wake_futex(top_waiter, key2, hb2);
346 	} else if (ret < 0) {
347 		/* Rewind top_waiter::requeue_state */
348 		futex_requeue_pi_complete(top_waiter, ret);
349 	} else {
350 		/*
351 		 * futex_lock_pi_atomic() did not acquire the user space
352 		 * futex, but managed to establish the proxy lock and pi
353 		 * state. top_waiter::requeue_state cannot be fixed up here
354 		 * because the waiter is not enqueued on the rtmutex
355 		 * yet. This is handled at the callsite depending on the
356 		 * result of rt_mutex_start_proxy_lock() which is
357 		 * guaranteed to be reached with this function returning 0.
358 		 */
359 	}
360 	return ret;
361 }
362 
363 /**
364  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
365  * @uaddr1:	source futex user address
366  * @flags1:	futex flags (FLAGS_SHARED, etc.)
367  * @uaddr2:	target futex user address
368  * @flags2:	futex flags (FLAGS_SHARED, etc.)
369  * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
370  * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
371  * @cmpval:	@uaddr1 expected value (or %NULL)
372  * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
373  *		pi futex (pi to pi requeue is not supported)
374  *
375  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
376  * uaddr2 atomically on behalf of the top waiter.
377  *
378  * Return:
379  *  - >=0 - on success, the number of tasks requeued or woken;
380  *  -  <0 - on error
381  */
382 int futex_requeue(u32 __user *uaddr1, unsigned int flags1,
383 		  u32 __user *uaddr2, unsigned int flags2,
384 		  int nr_wake, int nr_requeue, u32 *cmpval, int requeue_pi)
385 {
386 	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
387 	int task_count = 0, ret;
388 	struct futex_pi_state *pi_state = NULL;
389 	struct futex_q *this, *next;
390 	DEFINE_WAKE_Q(wake_q);
391 
392 	if (nr_wake < 0 || nr_requeue < 0)
393 		return -EINVAL;
394 
395 	/*
396 	 * When PI not supported: return -ENOSYS if requeue_pi is true,
397 	 * consequently the compiler knows requeue_pi is always false past
398 	 * this point which will optimize away all the conditional code
399 	 * further down.
400 	 */
401 	if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
402 		return -ENOSYS;
403 
404 	if (requeue_pi) {
405 		/*
406 		 * Requeue PI only works on two distinct uaddrs. This
407 		 * check is only valid for private futexes. See below.
408 		 */
409 		if (uaddr1 == uaddr2)
410 			return -EINVAL;
411 
412 		/*
413 		 * futex_requeue() allows the caller to define the number
414 		 * of waiters to wake up via the @nr_wake argument. With
415 		 * REQUEUE_PI, waking up more than one waiter is creating
416 		 * more problems than it solves. Waking up a waiter makes
417 		 * only sense if the PI futex @uaddr2 is uncontended as
418 		 * this allows the requeue code to acquire the futex
419 		 * @uaddr2 before waking the waiter. The waiter can then
420 		 * return to user space without further action. A secondary
421 		 * wakeup would just make the futex_wait_requeue_pi()
422 		 * handling more complex, because that code would have to
423 		 * look up pi_state and do more or less all the handling
424 		 * which the requeue code has to do for the to be requeued
425 		 * waiters. So restrict the number of waiters to wake to
426 		 * one, and only wake it up when the PI futex is
427 		 * uncontended. Otherwise requeue it and let the unlock of
428 		 * the PI futex handle the wakeup.
429 		 *
430 		 * All REQUEUE_PI users, e.g. pthread_cond_signal() and
431 		 * pthread_cond_broadcast() must use nr_wake=1.
432 		 */
433 		if (nr_wake != 1)
434 			return -EINVAL;
435 
436 		/*
437 		 * requeue_pi requires a pi_state, try to allocate it now
438 		 * without any locks in case it fails.
439 		 */
440 		if (refill_pi_state_cache())
441 			return -ENOMEM;
442 	}
443 
444 retry:
445 	ret = get_futex_key(uaddr1, flags1, &key1, FUTEX_READ);
446 	if (unlikely(ret != 0))
447 		return ret;
448 	ret = get_futex_key(uaddr2, flags2, &key2,
449 			    requeue_pi ? FUTEX_WRITE : FUTEX_READ);
450 	if (unlikely(ret != 0))
451 		return ret;
452 
453 	/*
454 	 * The check above which compares uaddrs is not sufficient for
455 	 * shared futexes. We need to compare the keys:
456 	 */
457 	if (requeue_pi && futex_match(&key1, &key2))
458 		return -EINVAL;
459 
460 retry_private:
461 	if (1) {
462 		CLASS(hbr, hbr1)(&key1);
463 		CLASS(hbr, hbr2)(&key2);
464 		auto hb1 = hbr1.hb;
465 		auto hb2 = hbr2.hb;
466 
467 		futex_hb_waiters_inc(hb2);
468 		double_lock_hb(hb1, hb2);
469 
470 		if (likely(cmpval != NULL)) {
471 			u32 curval;
472 
473 			ret = futex_get_value_locked(&curval, uaddr1);
474 
475 			if (unlikely(ret)) {
476 				futex_hb_waiters_dec(hb2);
477 				double_unlock_hb(hb1, hb2);
478 
479 				ret = get_user(curval, uaddr1);
480 				if (ret)
481 					return ret;
482 
483 				if (!(flags1 & FLAGS_SHARED))
484 					goto retry_private;
485 
486 				goto retry;
487 			}
488 			if (curval != *cmpval) {
489 				ret = -EAGAIN;
490 				goto out_unlock;
491 			}
492 		}
493 
494 		if (requeue_pi) {
495 			struct task_struct *exiting = NULL;
496 
497 			/*
498 			 * Attempt to acquire uaddr2 and wake the top waiter. If we
499 			 * intend to requeue waiters, force setting the FUTEX_WAITERS
500 			 * bit.  We force this here where we are able to easily handle
501 			 * faults rather in the requeue loop below.
502 			 *
503 			 * Updates topwaiter::requeue_state if a top waiter exists.
504 			 */
505 			ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
506 							 &key2, &pi_state,
507 							 &exiting, nr_requeue);
508 
509 			/*
510 			 * At this point the top_waiter has either taken uaddr2 or
511 			 * is waiting on it. In both cases pi_state has been
512 			 * established and an initial refcount on it. In case of an
513 			 * error there's nothing.
514 			 *
515 			 * The top waiter's requeue_state is up to date:
516 			 *
517 			 *  - If the lock was acquired atomically (ret == 1), then
518 			 *    the state is Q_REQUEUE_PI_LOCKED.
519 			 *
520 			 *    The top waiter has been dequeued and woken up and can
521 			 *    return to user space immediately. The kernel/user
522 			 *    space state is consistent. In case that there must be
523 			 *    more waiters requeued the WAITERS bit in the user
524 			 *    space futex is set so the top waiter task has to go
525 			 *    into the syscall slowpath to unlock the futex. This
526 			 *    will block until this requeue operation has been
527 			 *    completed and the hash bucket locks have been
528 			 *    dropped.
529 			 *
530 			 *  - If the trylock failed with an error (ret < 0) then
531 			 *    the state is either Q_REQUEUE_PI_NONE, i.e. "nothing
532 			 *    happened", or Q_REQUEUE_PI_IGNORE when there was an
533 			 *    interleaved early wakeup.
534 			 *
535 			 *  - If the trylock did not succeed (ret == 0) then the
536 			 *    state is either Q_REQUEUE_PI_IN_PROGRESS or
537 			 *    Q_REQUEUE_PI_WAIT if an early wakeup interleaved.
538 			 *    This will be cleaned up in the loop below, which
539 			 *    cannot fail because futex_proxy_trylock_atomic() did
540 			 *    the same sanity checks for requeue_pi as the loop
541 			 *    below does.
542 			 */
543 			switch (ret) {
544 			case 0:
545 				/* We hold a reference on the pi state. */
546 				break;
547 
548 			case 1:
549 				/*
550 				 * futex_proxy_trylock_atomic() acquired the user space
551 				 * futex. Adjust task_count.
552 				 */
553 				task_count++;
554 				ret = 0;
555 				break;
556 
557 				/*
558 				 * If the above failed, then pi_state is NULL and
559 				 * waiter::requeue_state is correct.
560 				 */
561 			case -EFAULT:
562 				futex_hb_waiters_dec(hb2);
563 				double_unlock_hb(hb1, hb2);
564 				ret = fault_in_user_writeable(uaddr2);
565 				if (!ret)
566 					goto retry;
567 				return ret;
568 			case -EBUSY:
569 			case -EAGAIN:
570 				/*
571 				 * Two reasons for this:
572 				 * - EBUSY: Owner is exiting and we just wait for the
573 				 *   exit to complete.
574 				 * - EAGAIN: The user space value changed.
575 				 */
576 				futex_hb_waiters_dec(hb2);
577 				double_unlock_hb(hb1, hb2);
578 				/*
579 				 * Handle the case where the owner is in the middle of
580 				 * exiting. Wait for the exit to complete otherwise
581 				 * this task might loop forever, aka. live lock.
582 				 */
583 				wait_for_owner_exiting(ret, exiting);
584 				cond_resched();
585 				goto retry;
586 			default:
587 				goto out_unlock;
588 			}
589 		}
590 
591 		plist_for_each_entry_safe(this, next, &hb1->chain, list) {
592 			if (task_count - nr_wake >= nr_requeue)
593 				break;
594 
595 			if (!futex_match(&this->key, &key1))
596 				continue;
597 
598 			/*
599 			 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
600 			 * be paired with each other and no other futex ops.
601 			 *
602 			 * We should never be requeueing a futex_q with a pi_state,
603 			 * which is awaiting a futex_unlock_pi().
604 			 */
605 			if ((requeue_pi && !this->rt_waiter) ||
606 			    (!requeue_pi && this->rt_waiter) ||
607 			    this->pi_state) {
608 				ret = -EINVAL;
609 				break;
610 			}
611 
612 			/* Plain futexes just wake or requeue and are done */
613 			if (!requeue_pi) {
614 				if (++task_count <= nr_wake)
615 					this->wake(&wake_q, this);
616 				else
617 					requeue_futex(this, hb1, hb2, &key2);
618 				continue;
619 			}
620 
621 			/* Ensure we requeue to the expected futex for requeue_pi. */
622 			if (!futex_match(this->requeue_pi_key, &key2)) {
623 				ret = -EINVAL;
624 				break;
625 			}
626 
627 			/*
628 			 * Requeue nr_requeue waiters and possibly one more in the case
629 			 * of requeue_pi if we couldn't acquire the lock atomically.
630 			 *
631 			 * Prepare the waiter to take the rt_mutex. Take a refcount
632 			 * on the pi_state and store the pointer in the futex_q
633 			 * object of the waiter.
634 			 */
635 			get_pi_state(pi_state);
636 
637 			/* Don't requeue when the waiter is already on the way out. */
638 			if (!futex_requeue_pi_prepare(this, pi_state)) {
639 				/*
640 				 * Early woken waiter signaled that it is on the
641 				 * way out. Drop the pi_state reference and try the
642 				 * next waiter. @this->pi_state is still NULL.
643 				 */
644 				put_pi_state(pi_state);
645 				continue;
646 			}
647 
648 			/* Self-deadlock: non-top waiter already owns the PI futex. */
649 			if (rt_mutex_owner(&pi_state->pi_mutex) == this->task) {
650 				ret = -EDEADLK;
651 				break;
652 			}
653 
654 			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
655 							this->rt_waiter,
656 							this->task);
657 
658 			if (ret == 1) {
659 				/*
660 				 * We got the lock. We do neither drop the refcount
661 				 * on pi_state nor clear this->pi_state because the
662 				 * waiter needs the pi_state for cleaning up the
663 				 * user space value. It will drop the refcount
664 				 * after doing so. this::requeue_state is updated
665 				 * in the wakeup as well.
666 				 */
667 				requeue_pi_wake_futex(this, &key2, hb2);
668 				task_count++;
669 			} else if (!ret) {
670 				/* Waiter is queued, move it to hb2 */
671 				requeue_futex(this, hb1, hb2, &key2);
672 				futex_requeue_pi_complete(this, 0);
673 				task_count++;
674 			} else {
675 				/*
676 				 * rt_mutex_start_proxy_lock() detected a potential
677 				 * deadlock when we tried to queue that waiter.
678 				 * Drop the pi_state reference which we took above
679 				 * and remove the pointer to the state from the
680 				 * waiters futex_q object.
681 				 */
682 				this->pi_state = NULL;
683 				put_pi_state(pi_state);
684 				futex_requeue_pi_complete(this, ret);
685 				/*
686 				 * We stop queueing more waiters and let user space
687 				 * deal with the mess.
688 				 */
689 				break;
690 			}
691 		}
692 
693 		/*
694 		 * We took an extra initial reference to the pi_state in
695 		 * futex_proxy_trylock_atomic(). We need to drop it here again.
696 		 */
697 		put_pi_state(pi_state);
698 
699 out_unlock:
700 		futex_hb_waiters_dec(hb2);
701 		double_unlock_hb(hb1, hb2);
702 	}
703 	wake_up_q(&wake_q);
704 	return ret ? ret : task_count;
705 }
706 
707 /**
708  * handle_early_requeue_pi_wakeup() - Handle early wakeup on the initial futex
709  * @hb:		the hash_bucket futex_q was original enqueued on
710  * @q:		the futex_q woken while waiting to be requeued
711  * @timeout:	the timeout associated with the wait (NULL if none)
712  *
713  * Determine the cause for the early wakeup.
714  *
715  * Return:
716  *  -EWOULDBLOCK or -ETIMEDOUT or -ERESTARTNOINTR
717  */
718 static inline
719 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
720 				   struct futex_q *q,
721 				   struct hrtimer_sleeper *timeout)
722 {
723 	int ret;
724 
725 	/*
726 	 * With the hb lock held, we avoid races while we process the wakeup.
727 	 * We only need to hold hb (and not hb2) to ensure atomicity as the
728 	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
729 	 * It can't be requeued from uaddr2 to something else since we don't
730 	 * support a PI aware source futex for requeue.
731 	 */
732 	WARN_ON_ONCE(&hb->lock != q->lock_ptr);
733 
734 	/*
735 	 * We were woken prior to requeue by a timeout or a signal.
736 	 * Conditionally unqueue the futex_q and determine which it was.
737 	 */
738 	if (!plist_node_empty(&q->list)) {
739 		plist_del(&q->list, &hb->chain);
740 		futex_hb_waiters_dec(hb);
741 	}
742 
743 	/* Handle spurious wakeups gracefully */
744 	ret = -EWOULDBLOCK;
745 	if (timeout && !timeout->task)
746 		ret = -ETIMEDOUT;
747 	else if (signal_pending(current))
748 		ret = -ERESTARTNOINTR;
749 	return ret;
750 }
751 
752 /**
753  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
754  * @uaddr:	the futex we initially wait on (non-pi)
755  * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
756  *		the same type, no requeueing from private to shared, etc.
757  * @val:	the expected value of uaddr
758  * @abs_time:	absolute timeout
759  * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
760  * @uaddr2:	the pi futex we will take prior to returning to user-space
761  *
762  * The caller will wait on uaddr and will be requeued by futex_requeue() to
763  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
764  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
765  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
766  * without one, the pi logic would not know which task to boost/deboost, if
767  * there was a need to.
768  *
769  * We call schedule in futex_wait_queue() when we enqueue and return there
770  * via the following--
771  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
772  * 2) wakeup on uaddr2 after a requeue
773  * 3) signal
774  * 4) timeout
775  *
776  * If 3, cleanup and return -ERESTARTNOINTR.
777  *
778  * If 2, we may then block on trying to take the rt_mutex and return via:
779  * 5) successful lock
780  * 6) signal
781  * 7) timeout
782  * 8) other lock acquisition failure
783  *
784  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
785  *
786  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
787  *
788  * Return:
789  *  -  0 - On success;
790  *  - <0 - On error
791  */
792 int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
793 			  u32 val, ktime_t *abs_time, u32 bitset,
794 			  u32 __user *uaddr2)
795 {
796 	struct hrtimer_sleeper timeout, *to;
797 	struct rt_mutex_waiter rt_waiter;
798 	union futex_key key2 = FUTEX_KEY_INIT;
799 	struct futex_q q = futex_q_init;
800 	struct rt_mutex_base *pi_mutex;
801 	int res, ret;
802 
803 	if (!IS_ENABLED(CONFIG_FUTEX_PI))
804 		return -ENOSYS;
805 
806 	if (uaddr == uaddr2)
807 		return -EINVAL;
808 
809 	if (!bitset)
810 		return -EINVAL;
811 
812 	to = futex_setup_timer(abs_time, &timeout, flags,
813 			       current->timer_slack_ns);
814 
815 	/*
816 	 * The waiter is allocated on our stack, manipulated by the requeue
817 	 * code while we sleep on uaddr.
818 	 */
819 	rt_mutex_init_waiter(&rt_waiter);
820 
821 	ret = get_futex_key(uaddr2, flags, &key2, FUTEX_WRITE);
822 	if (unlikely(ret != 0))
823 		goto out;
824 
825 	q.bitset = bitset;
826 	q.rt_waiter = &rt_waiter;
827 	q.requeue_pi_key = &key2;
828 
829 	/*
830 	 * Prepare to wait on uaddr. On success, it holds hb->lock and q
831 	 * is initialized.
832 	 */
833 	ret = futex_wait_setup(uaddr, val, flags, &q, &key2, current);
834 	if (ret)
835 		goto out;
836 
837 	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
838 	futex_do_wait(&q, to);
839 
840 	switch (futex_requeue_pi_wakeup_sync(&q)) {
841 	case Q_REQUEUE_PI_IGNORE:
842 		{
843 			CLASS(hbr, hbr)(&q.key);
844 			auto hb = hbr.hb;
845 			/* The waiter is still on uaddr1 */
846 			spin_lock(&hb->lock);
847 			ret = handle_early_requeue_pi_wakeup(hb, &q, to);
848 			spin_unlock(&hb->lock);
849 		}
850 		break;
851 
852 	case Q_REQUEUE_PI_LOCKED:
853 		/* The requeue acquired the lock */
854 		if (q.pi_state && (q.pi_state->owner != current)) {
855 			futex_q_lockptr_lock(&q);
856 			ret = fixup_pi_owner(uaddr2, &q, true);
857 			/*
858 			 * Drop the reference to the pi state which the
859 			 * requeue_pi() code acquired for us.
860 			 */
861 			put_pi_state(q.pi_state);
862 			spin_unlock(q.lock_ptr);
863 			/*
864 			 * Adjust the return value. It's either -EFAULT or
865 			 * success (1) but the caller expects 0 for success.
866 			 */
867 			ret = ret < 0 ? ret : 0;
868 		}
869 		break;
870 
871 	case Q_REQUEUE_PI_DONE:
872 		/* Requeue completed. Current is 'pi_blocked_on' the rtmutex */
873 		pi_mutex = &q.pi_state->pi_mutex;
874 		ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
875 
876 		/*
877 		 * See futex_unlock_pi()'s cleanup: comment.
878 		 */
879 		if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
880 			ret = 0;
881 
882 		futex_q_lockptr_lock(&q);
883 		debug_rt_mutex_free_waiter(&rt_waiter);
884 		/*
885 		 * Fixup the pi_state owner and possibly acquire the lock if we
886 		 * haven't already.
887 		 */
888 		res = fixup_pi_owner(uaddr2, &q, !ret);
889 		/*
890 		 * If fixup_pi_owner() returned an error, propagate that.  If it
891 		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
892 		 */
893 		if (res)
894 			ret = (res < 0) ? res : 0;
895 
896 		futex_unqueue_pi(&q);
897 		spin_unlock(q.lock_ptr);
898 
899 		if (ret == -EINTR) {
900 			/*
901 			 * We've already been requeued, but cannot restart
902 			 * by calling futex_lock_pi() directly. We could
903 			 * restart this syscall, but it would detect that
904 			 * the user space "val" changed and return
905 			 * -EWOULDBLOCK.  Save the overhead of the restart
906 			 * and return -EWOULDBLOCK directly.
907 			 */
908 			ret = -EWOULDBLOCK;
909 		}
910 		break;
911 	default:
912 		BUG();
913 	}
914 	/* Additional reference from requeue_pi_wake_futex() */
915 	futex_private_hash_put(q.drop_fph);
916 
917 out:
918 	if (to) {
919 		hrtimer_cancel(&to->timer);
920 		destroy_hrtimer_on_stack(&to->timer);
921 	}
922 	return ret;
923 }
924 
925