xref: /linux/kernel/futex/pi.c (revision 06d07429858317ded2db7986113a9e0129cd599b)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 
3 #include <linux/slab.h>
4 #include <linux/sched/rt.h>
5 #include <linux/sched/task.h>
6 
7 #include "futex.h"
8 #include "../locking/rtmutex_common.h"
9 
10 /*
11  * PI code:
12  */
refill_pi_state_cache(void)13 int refill_pi_state_cache(void)
14 {
15 	struct futex_pi_state *pi_state;
16 
17 	if (likely(current->pi_state_cache))
18 		return 0;
19 
20 	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
21 
22 	if (!pi_state)
23 		return -ENOMEM;
24 
25 	INIT_LIST_HEAD(&pi_state->list);
26 	/* pi_mutex gets initialized later */
27 	pi_state->owner = NULL;
28 	refcount_set(&pi_state->refcount, 1);
29 	pi_state->key = FUTEX_KEY_INIT;
30 
31 	current->pi_state_cache = pi_state;
32 
33 	return 0;
34 }
35 
alloc_pi_state(void)36 static struct futex_pi_state *alloc_pi_state(void)
37 {
38 	struct futex_pi_state *pi_state = current->pi_state_cache;
39 
40 	WARN_ON(!pi_state);
41 	current->pi_state_cache = NULL;
42 
43 	return pi_state;
44 }
45 
pi_state_update_owner(struct futex_pi_state * pi_state,struct task_struct * new_owner)46 static void pi_state_update_owner(struct futex_pi_state *pi_state,
47 				  struct task_struct *new_owner)
48 {
49 	struct task_struct *old_owner = pi_state->owner;
50 
51 	lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
52 
53 	if (old_owner) {
54 		raw_spin_lock(&old_owner->pi_lock);
55 		WARN_ON(list_empty(&pi_state->list));
56 		list_del_init(&pi_state->list);
57 		raw_spin_unlock(&old_owner->pi_lock);
58 	}
59 
60 	if (new_owner) {
61 		raw_spin_lock(&new_owner->pi_lock);
62 		WARN_ON(!list_empty(&pi_state->list));
63 		list_add(&pi_state->list, &new_owner->pi_state_list);
64 		pi_state->owner = new_owner;
65 		raw_spin_unlock(&new_owner->pi_lock);
66 	}
67 }
68 
get_pi_state(struct futex_pi_state * pi_state)69 void get_pi_state(struct futex_pi_state *pi_state)
70 {
71 	WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
72 }
73 
74 /*
75  * Drops a reference to the pi_state object and frees or caches it
76  * when the last reference is gone.
77  */
put_pi_state(struct futex_pi_state * pi_state)78 void put_pi_state(struct futex_pi_state *pi_state)
79 {
80 	if (!pi_state)
81 		return;
82 
83 	if (!refcount_dec_and_test(&pi_state->refcount))
84 		return;
85 
86 	/*
87 	 * If pi_state->owner is NULL, the owner is most probably dying
88 	 * and has cleaned up the pi_state already
89 	 */
90 	if (pi_state->owner) {
91 		unsigned long flags;
92 
93 		raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
94 		pi_state_update_owner(pi_state, NULL);
95 		rt_mutex_proxy_unlock(&pi_state->pi_mutex);
96 		raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
97 	}
98 
99 	if (current->pi_state_cache) {
100 		kfree(pi_state);
101 	} else {
102 		/*
103 		 * pi_state->list is already empty.
104 		 * clear pi_state->owner.
105 		 * refcount is at 0 - put it back to 1.
106 		 */
107 		pi_state->owner = NULL;
108 		refcount_set(&pi_state->refcount, 1);
109 		current->pi_state_cache = pi_state;
110 	}
111 }
112 
113 /*
114  * We need to check the following states:
115  *
116  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
117  *
118  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
119  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
120  *
121  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
122  *
123  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
124  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
125  *
126  * [6]  Found  | Found    | task      | 0         | 1      | Valid
127  *
128  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
129  *
130  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
131  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
132  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
133  *
134  * [1]	Indicates that the kernel can acquire the futex atomically. We
135  *	came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
136  *
137  * [2]	Valid, if TID does not belong to a kernel thread. If no matching
138  *      thread is found then it indicates that the owner TID has died.
139  *
140  * [3]	Invalid. The waiter is queued on a non PI futex
141  *
142  * [4]	Valid state after exit_robust_list(), which sets the user space
143  *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
144  *
145  * [5]	The user space value got manipulated between exit_robust_list()
146  *	and exit_pi_state_list()
147  *
148  * [6]	Valid state after exit_pi_state_list() which sets the new owner in
149  *	the pi_state but cannot access the user space value.
150  *
151  * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
152  *
153  * [8]	Owner and user space value match
154  *
155  * [9]	There is no transient state which sets the user space TID to 0
156  *	except exit_robust_list(), but this is indicated by the
157  *	FUTEX_OWNER_DIED bit. See [4]
158  *
159  * [10] There is no transient state which leaves owner and user space
160  *	TID out of sync. Except one error case where the kernel is denied
161  *	write access to the user address, see fixup_pi_state_owner().
162  *
163  *
164  * Serialization and lifetime rules:
165  *
166  * hb->lock:
167  *
168  *	hb -> futex_q, relation
169  *	futex_q -> pi_state, relation
170  *
171  *	(cannot be raw because hb can contain arbitrary amount
172  *	 of futex_q's)
173  *
174  * pi_mutex->wait_lock:
175  *
176  *	{uval, pi_state}
177  *
178  *	(and pi_mutex 'obviously')
179  *
180  * p->pi_lock:
181  *
182  *	p->pi_state_list -> pi_state->list, relation
183  *	pi_mutex->owner -> pi_state->owner, relation
184  *
185  * pi_state->refcount:
186  *
187  *	pi_state lifetime
188  *
189  *
190  * Lock order:
191  *
192  *   hb->lock
193  *     pi_mutex->wait_lock
194  *       p->pi_lock
195  *
196  */
197 
198 /*
199  * Validate that the existing waiter has a pi_state and sanity check
200  * the pi_state against the user space value. If correct, attach to
201  * it.
202  */
attach_to_pi_state(u32 __user * uaddr,u32 uval,struct futex_pi_state * pi_state,struct futex_pi_state ** ps)203 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
204 			      struct futex_pi_state *pi_state,
205 			      struct futex_pi_state **ps)
206 {
207 	pid_t pid = uval & FUTEX_TID_MASK;
208 	u32 uval2;
209 	int ret;
210 
211 	/*
212 	 * Userspace might have messed up non-PI and PI futexes [3]
213 	 */
214 	if (unlikely(!pi_state))
215 		return -EINVAL;
216 
217 	/*
218 	 * We get here with hb->lock held, and having found a
219 	 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
220 	 * has dropped the hb->lock in between futex_queue() and futex_unqueue_pi(),
221 	 * which in turn means that futex_lock_pi() still has a reference on
222 	 * our pi_state.
223 	 *
224 	 * The waiter holding a reference on @pi_state also protects against
225 	 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
226 	 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
227 	 * free pi_state before we can take a reference ourselves.
228 	 */
229 	WARN_ON(!refcount_read(&pi_state->refcount));
230 
231 	/*
232 	 * Now that we have a pi_state, we can acquire wait_lock
233 	 * and do the state validation.
234 	 */
235 	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
236 
237 	/*
238 	 * Since {uval, pi_state} is serialized by wait_lock, and our current
239 	 * uval was read without holding it, it can have changed. Verify it
240 	 * still is what we expect it to be, otherwise retry the entire
241 	 * operation.
242 	 */
243 	if (futex_get_value_locked(&uval2, uaddr))
244 		goto out_efault;
245 
246 	if (uval != uval2)
247 		goto out_eagain;
248 
249 	/*
250 	 * Handle the owner died case:
251 	 */
252 	if (uval & FUTEX_OWNER_DIED) {
253 		/*
254 		 * exit_pi_state_list sets owner to NULL and wakes the
255 		 * topmost waiter. The task which acquires the
256 		 * pi_state->rt_mutex will fixup owner.
257 		 */
258 		if (!pi_state->owner) {
259 			/*
260 			 * No pi state owner, but the user space TID
261 			 * is not 0. Inconsistent state. [5]
262 			 */
263 			if (pid)
264 				goto out_einval;
265 			/*
266 			 * Take a ref on the state and return success. [4]
267 			 */
268 			goto out_attach;
269 		}
270 
271 		/*
272 		 * If TID is 0, then either the dying owner has not
273 		 * yet executed exit_pi_state_list() or some waiter
274 		 * acquired the rtmutex in the pi state, but did not
275 		 * yet fixup the TID in user space.
276 		 *
277 		 * Take a ref on the state and return success. [6]
278 		 */
279 		if (!pid)
280 			goto out_attach;
281 	} else {
282 		/*
283 		 * If the owner died bit is not set, then the pi_state
284 		 * must have an owner. [7]
285 		 */
286 		if (!pi_state->owner)
287 			goto out_einval;
288 	}
289 
290 	/*
291 	 * Bail out if user space manipulated the futex value. If pi
292 	 * state exists then the owner TID must be the same as the
293 	 * user space TID. [9/10]
294 	 */
295 	if (pid != task_pid_vnr(pi_state->owner))
296 		goto out_einval;
297 
298 out_attach:
299 	get_pi_state(pi_state);
300 	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
301 	*ps = pi_state;
302 	return 0;
303 
304 out_einval:
305 	ret = -EINVAL;
306 	goto out_error;
307 
308 out_eagain:
309 	ret = -EAGAIN;
310 	goto out_error;
311 
312 out_efault:
313 	ret = -EFAULT;
314 	goto out_error;
315 
316 out_error:
317 	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
318 	return ret;
319 }
320 
handle_exit_race(u32 __user * uaddr,u32 uval,struct task_struct * tsk)321 static int handle_exit_race(u32 __user *uaddr, u32 uval,
322 			    struct task_struct *tsk)
323 {
324 	u32 uval2;
325 
326 	/*
327 	 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
328 	 * caller that the alleged owner is busy.
329 	 */
330 	if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
331 		return -EBUSY;
332 
333 	/*
334 	 * Reread the user space value to handle the following situation:
335 	 *
336 	 * CPU0				CPU1
337 	 *
338 	 * sys_exit()			sys_futex()
339 	 *  do_exit()			 futex_lock_pi()
340 	 *                                futex_lock_pi_atomic()
341 	 *   exit_signals(tsk)		    No waiters:
342 	 *    tsk->flags |= PF_EXITING;	    *uaddr == 0x00000PID
343 	 *  mm_release(tsk)		    Set waiter bit
344 	 *   exit_robust_list(tsk) {	    *uaddr = 0x80000PID;
345 	 *      Set owner died		    attach_to_pi_owner() {
346 	 *    *uaddr = 0xC0000000;	     tsk = get_task(PID);
347 	 *   }				     if (!tsk->flags & PF_EXITING) {
348 	 *  ...				       attach();
349 	 *  tsk->futex_state =               } else {
350 	 *	FUTEX_STATE_DEAD;              if (tsk->futex_state !=
351 	 *					  FUTEX_STATE_DEAD)
352 	 *				         return -EAGAIN;
353 	 *				       return -ESRCH; <--- FAIL
354 	 *				     }
355 	 *
356 	 * Returning ESRCH unconditionally is wrong here because the
357 	 * user space value has been changed by the exiting task.
358 	 *
359 	 * The same logic applies to the case where the exiting task is
360 	 * already gone.
361 	 */
362 	if (futex_get_value_locked(&uval2, uaddr))
363 		return -EFAULT;
364 
365 	/* If the user space value has changed, try again. */
366 	if (uval2 != uval)
367 		return -EAGAIN;
368 
369 	/*
370 	 * The exiting task did not have a robust list, the robust list was
371 	 * corrupted or the user space value in *uaddr is simply bogus.
372 	 * Give up and tell user space.
373 	 */
374 	return -ESRCH;
375 }
376 
__attach_to_pi_owner(struct task_struct * p,union futex_key * key,struct futex_pi_state ** ps)377 static void __attach_to_pi_owner(struct task_struct *p, union futex_key *key,
378 				 struct futex_pi_state **ps)
379 {
380 	/*
381 	 * No existing pi state. First waiter. [2]
382 	 *
383 	 * This creates pi_state, we have hb->lock held, this means nothing can
384 	 * observe this state, wait_lock is irrelevant.
385 	 */
386 	struct futex_pi_state *pi_state = alloc_pi_state();
387 
388 	/*
389 	 * Initialize the pi_mutex in locked state and make @p
390 	 * the owner of it:
391 	 */
392 	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
393 
394 	/* Store the key for possible exit cleanups: */
395 	pi_state->key = *key;
396 
397 	WARN_ON(!list_empty(&pi_state->list));
398 	list_add(&pi_state->list, &p->pi_state_list);
399 	/*
400 	 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
401 	 * because there is no concurrency as the object is not published yet.
402 	 */
403 	pi_state->owner = p;
404 
405 	*ps = pi_state;
406 }
407 /*
408  * Lookup the task for the TID provided from user space and attach to
409  * it after doing proper sanity checks.
410  */
attach_to_pi_owner(u32 __user * uaddr,u32 uval,union futex_key * key,struct futex_pi_state ** ps,struct task_struct ** exiting)411 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
412 			      struct futex_pi_state **ps,
413 			      struct task_struct **exiting)
414 {
415 	pid_t pid = uval & FUTEX_TID_MASK;
416 	struct task_struct *p;
417 
418 	/*
419 	 * We are the first waiter - try to look up the real owner and attach
420 	 * the new pi_state to it, but bail out when TID = 0 [1]
421 	 *
422 	 * The !pid check is paranoid. None of the call sites should end up
423 	 * with pid == 0, but better safe than sorry. Let the caller retry
424 	 */
425 	if (!pid)
426 		return -EAGAIN;
427 	p = find_get_task_by_vpid(pid);
428 	if (!p)
429 		return handle_exit_race(uaddr, uval, NULL);
430 
431 	if (unlikely(p->flags & PF_KTHREAD)) {
432 		put_task_struct(p);
433 		return -EPERM;
434 	}
435 
436 	/*
437 	 * We need to look at the task state to figure out, whether the
438 	 * task is exiting. To protect against the change of the task state
439 	 * in futex_exit_release(), we do this protected by p->pi_lock:
440 	 */
441 	raw_spin_lock_irq(&p->pi_lock);
442 	if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
443 		/*
444 		 * The task is on the way out. When the futex state is
445 		 * FUTEX_STATE_DEAD, we know that the task has finished
446 		 * the cleanup:
447 		 */
448 		int ret = handle_exit_race(uaddr, uval, p);
449 
450 		raw_spin_unlock_irq(&p->pi_lock);
451 		/*
452 		 * If the owner task is between FUTEX_STATE_EXITING and
453 		 * FUTEX_STATE_DEAD then store the task pointer and keep
454 		 * the reference on the task struct. The calling code will
455 		 * drop all locks, wait for the task to reach
456 		 * FUTEX_STATE_DEAD and then drop the refcount. This is
457 		 * required to prevent a live lock when the current task
458 		 * preempted the exiting task between the two states.
459 		 */
460 		if (ret == -EBUSY)
461 			*exiting = p;
462 		else
463 			put_task_struct(p);
464 		return ret;
465 	}
466 
467 	__attach_to_pi_owner(p, key, ps);
468 	raw_spin_unlock_irq(&p->pi_lock);
469 
470 	put_task_struct(p);
471 
472 	return 0;
473 }
474 
lock_pi_update_atomic(u32 __user * uaddr,u32 uval,u32 newval)475 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
476 {
477 	int err;
478 	u32 curval;
479 
480 	if (unlikely(should_fail_futex(true)))
481 		return -EFAULT;
482 
483 	err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
484 	if (unlikely(err))
485 		return err;
486 
487 	/* If user space value changed, let the caller retry */
488 	return curval != uval ? -EAGAIN : 0;
489 }
490 
491 /**
492  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
493  * @uaddr:		the pi futex user address
494  * @hb:			the pi futex hash bucket
495  * @key:		the futex key associated with uaddr and hb
496  * @ps:			the pi_state pointer where we store the result of the
497  *			lookup
498  * @task:		the task to perform the atomic lock work for.  This will
499  *			be "current" except in the case of requeue pi.
500  * @exiting:		Pointer to store the task pointer of the owner task
501  *			which is in the middle of exiting
502  * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
503  *
504  * Return:
505  *  -  0 - ready to wait;
506  *  -  1 - acquired the lock;
507  *  - <0 - error
508  *
509  * The hb->lock must be held by the caller.
510  *
511  * @exiting is only set when the return value is -EBUSY. If so, this holds
512  * a refcount on the exiting task on return and the caller needs to drop it
513  * after waiting for the exit to complete.
514  */
futex_lock_pi_atomic(u32 __user * uaddr,struct futex_hash_bucket * hb,union futex_key * key,struct futex_pi_state ** ps,struct task_struct * task,struct task_struct ** exiting,int set_waiters)515 int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
516 			 union futex_key *key,
517 			 struct futex_pi_state **ps,
518 			 struct task_struct *task,
519 			 struct task_struct **exiting,
520 			 int set_waiters)
521 {
522 	u32 uval, newval, vpid = task_pid_vnr(task);
523 	struct futex_q *top_waiter;
524 	int ret;
525 
526 	/*
527 	 * Read the user space value first so we can validate a few
528 	 * things before proceeding further.
529 	 */
530 	if (futex_get_value_locked(&uval, uaddr))
531 		return -EFAULT;
532 
533 	if (unlikely(should_fail_futex(true)))
534 		return -EFAULT;
535 
536 	/*
537 	 * Detect deadlocks.
538 	 */
539 	if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
540 		return -EDEADLK;
541 
542 	if ((unlikely(should_fail_futex(true))))
543 		return -EDEADLK;
544 
545 	/*
546 	 * Lookup existing state first. If it exists, try to attach to
547 	 * its pi_state.
548 	 */
549 	top_waiter = futex_top_waiter(hb, key);
550 	if (top_waiter)
551 		return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
552 
553 	/*
554 	 * No waiter and user TID is 0. We are here because the
555 	 * waiters or the owner died bit is set or called from
556 	 * requeue_cmp_pi or for whatever reason something took the
557 	 * syscall.
558 	 */
559 	if (!(uval & FUTEX_TID_MASK)) {
560 		/*
561 		 * We take over the futex. No other waiters and the user space
562 		 * TID is 0. We preserve the owner died bit.
563 		 */
564 		newval = uval & FUTEX_OWNER_DIED;
565 		newval |= vpid;
566 
567 		/* The futex requeue_pi code can enforce the waiters bit */
568 		if (set_waiters)
569 			newval |= FUTEX_WAITERS;
570 
571 		ret = lock_pi_update_atomic(uaddr, uval, newval);
572 		if (ret)
573 			return ret;
574 
575 		/*
576 		 * If the waiter bit was requested the caller also needs PI
577 		 * state attached to the new owner of the user space futex.
578 		 *
579 		 * @task is guaranteed to be alive and it cannot be exiting
580 		 * because it is either sleeping or waiting in
581 		 * futex_requeue_pi_wakeup_sync().
582 		 *
583 		 * No need to do the full attach_to_pi_owner() exercise
584 		 * because @task is known and valid.
585 		 */
586 		if (set_waiters) {
587 			raw_spin_lock_irq(&task->pi_lock);
588 			__attach_to_pi_owner(task, key, ps);
589 			raw_spin_unlock_irq(&task->pi_lock);
590 		}
591 		return 1;
592 	}
593 
594 	/*
595 	 * First waiter. Set the waiters bit before attaching ourself to
596 	 * the owner. If owner tries to unlock, it will be forced into
597 	 * the kernel and blocked on hb->lock.
598 	 */
599 	newval = uval | FUTEX_WAITERS;
600 	ret = lock_pi_update_atomic(uaddr, uval, newval);
601 	if (ret)
602 		return ret;
603 	/*
604 	 * If the update of the user space value succeeded, we try to
605 	 * attach to the owner. If that fails, no harm done, we only
606 	 * set the FUTEX_WAITERS bit in the user space variable.
607 	 */
608 	return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
609 }
610 
611 /*
612  * Caller must hold a reference on @pi_state.
613  */
wake_futex_pi(u32 __user * uaddr,u32 uval,struct futex_pi_state * pi_state,struct rt_mutex_waiter * top_waiter)614 static int wake_futex_pi(u32 __user *uaddr, u32 uval,
615 			 struct futex_pi_state *pi_state,
616 			 struct rt_mutex_waiter *top_waiter)
617 {
618 	struct task_struct *new_owner;
619 	bool postunlock = false;
620 	DEFINE_RT_WAKE_Q(wqh);
621 	u32 curval, newval;
622 	int ret = 0;
623 
624 	new_owner = top_waiter->task;
625 
626 	/*
627 	 * We pass it to the next owner. The WAITERS bit is always kept
628 	 * enabled while there is PI state around. We cleanup the owner
629 	 * died bit, because we are the owner.
630 	 */
631 	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
632 
633 	if (unlikely(should_fail_futex(true))) {
634 		ret = -EFAULT;
635 		goto out_unlock;
636 	}
637 
638 	ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
639 	if (!ret && (curval != uval)) {
640 		/*
641 		 * If a unconditional UNLOCK_PI operation (user space did not
642 		 * try the TID->0 transition) raced with a waiter setting the
643 		 * FUTEX_WAITERS flag between get_user() and locking the hash
644 		 * bucket lock, retry the operation.
645 		 */
646 		if ((FUTEX_TID_MASK & curval) == uval)
647 			ret = -EAGAIN;
648 		else
649 			ret = -EINVAL;
650 	}
651 
652 	if (!ret) {
653 		/*
654 		 * This is a point of no return; once we modified the uval
655 		 * there is no going back and subsequent operations must
656 		 * not fail.
657 		 */
658 		pi_state_update_owner(pi_state, new_owner);
659 		postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wqh);
660 	}
661 
662 out_unlock:
663 	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
664 
665 	if (postunlock)
666 		rt_mutex_postunlock(&wqh);
667 
668 	return ret;
669 }
670 
__fixup_pi_state_owner(u32 __user * uaddr,struct futex_q * q,struct task_struct * argowner)671 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
672 				  struct task_struct *argowner)
673 {
674 	struct futex_pi_state *pi_state = q->pi_state;
675 	struct task_struct *oldowner, *newowner;
676 	u32 uval, curval, newval, newtid;
677 	int err = 0;
678 
679 	oldowner = pi_state->owner;
680 
681 	/*
682 	 * We are here because either:
683 	 *
684 	 *  - we stole the lock and pi_state->owner needs updating to reflect
685 	 *    that (@argowner == current),
686 	 *
687 	 * or:
688 	 *
689 	 *  - someone stole our lock and we need to fix things to point to the
690 	 *    new owner (@argowner == NULL).
691 	 *
692 	 * Either way, we have to replace the TID in the user space variable.
693 	 * This must be atomic as we have to preserve the owner died bit here.
694 	 *
695 	 * Note: We write the user space value _before_ changing the pi_state
696 	 * because we can fault here. Imagine swapped out pages or a fork
697 	 * that marked all the anonymous memory readonly for cow.
698 	 *
699 	 * Modifying pi_state _before_ the user space value would leave the
700 	 * pi_state in an inconsistent state when we fault here, because we
701 	 * need to drop the locks to handle the fault. This might be observed
702 	 * in the PID checks when attaching to PI state .
703 	 */
704 retry:
705 	if (!argowner) {
706 		if (oldowner != current) {
707 			/*
708 			 * We raced against a concurrent self; things are
709 			 * already fixed up. Nothing to do.
710 			 */
711 			return 0;
712 		}
713 
714 		if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
715 			/* We got the lock. pi_state is correct. Tell caller. */
716 			return 1;
717 		}
718 
719 		/*
720 		 * The trylock just failed, so either there is an owner or
721 		 * there is a higher priority waiter than this one.
722 		 */
723 		newowner = rt_mutex_owner(&pi_state->pi_mutex);
724 		/*
725 		 * If the higher priority waiter has not yet taken over the
726 		 * rtmutex then newowner is NULL. We can't return here with
727 		 * that state because it's inconsistent vs. the user space
728 		 * state. So drop the locks and try again. It's a valid
729 		 * situation and not any different from the other retry
730 		 * conditions.
731 		 */
732 		if (unlikely(!newowner)) {
733 			err = -EAGAIN;
734 			goto handle_err;
735 		}
736 	} else {
737 		WARN_ON_ONCE(argowner != current);
738 		if (oldowner == current) {
739 			/*
740 			 * We raced against a concurrent self; things are
741 			 * already fixed up. Nothing to do.
742 			 */
743 			return 1;
744 		}
745 		newowner = argowner;
746 	}
747 
748 	newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
749 	/* Owner died? */
750 	if (!pi_state->owner)
751 		newtid |= FUTEX_OWNER_DIED;
752 
753 	err = futex_get_value_locked(&uval, uaddr);
754 	if (err)
755 		goto handle_err;
756 
757 	for (;;) {
758 		newval = (uval & FUTEX_OWNER_DIED) | newtid;
759 
760 		err = futex_cmpxchg_value_locked(&curval, uaddr, uval, newval);
761 		if (err)
762 			goto handle_err;
763 
764 		if (curval == uval)
765 			break;
766 		uval = curval;
767 	}
768 
769 	/*
770 	 * We fixed up user space. Now we need to fix the pi_state
771 	 * itself.
772 	 */
773 	pi_state_update_owner(pi_state, newowner);
774 
775 	return argowner == current;
776 
777 	/*
778 	 * In order to reschedule or handle a page fault, we need to drop the
779 	 * locks here. In the case of a fault, this gives the other task
780 	 * (either the highest priority waiter itself or the task which stole
781 	 * the rtmutex) the chance to try the fixup of the pi_state. So once we
782 	 * are back from handling the fault we need to check the pi_state after
783 	 * reacquiring the locks and before trying to do another fixup. When
784 	 * the fixup has been done already we simply return.
785 	 *
786 	 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
787 	 * drop hb->lock since the caller owns the hb -> futex_q relation.
788 	 * Dropping the pi_mutex->wait_lock requires the state revalidate.
789 	 */
790 handle_err:
791 	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
792 	spin_unlock(q->lock_ptr);
793 
794 	switch (err) {
795 	case -EFAULT:
796 		err = fault_in_user_writeable(uaddr);
797 		break;
798 
799 	case -EAGAIN:
800 		cond_resched();
801 		err = 0;
802 		break;
803 
804 	default:
805 		WARN_ON_ONCE(1);
806 		break;
807 	}
808 
809 	spin_lock(q->lock_ptr);
810 	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
811 
812 	/*
813 	 * Check if someone else fixed it for us:
814 	 */
815 	if (pi_state->owner != oldowner)
816 		return argowner == current;
817 
818 	/* Retry if err was -EAGAIN or the fault in succeeded */
819 	if (!err)
820 		goto retry;
821 
822 	/*
823 	 * fault_in_user_writeable() failed so user state is immutable. At
824 	 * best we can make the kernel state consistent but user state will
825 	 * be most likely hosed and any subsequent unlock operation will be
826 	 * rejected due to PI futex rule [10].
827 	 *
828 	 * Ensure that the rtmutex owner is also the pi_state owner despite
829 	 * the user space value claiming something different. There is no
830 	 * point in unlocking the rtmutex if current is the owner as it
831 	 * would need to wait until the next waiter has taken the rtmutex
832 	 * to guarantee consistent state. Keep it simple. Userspace asked
833 	 * for this wreckaged state.
834 	 *
835 	 * The rtmutex has an owner - either current or some other
836 	 * task. See the EAGAIN loop above.
837 	 */
838 	pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
839 
840 	return err;
841 }
842 
fixup_pi_state_owner(u32 __user * uaddr,struct futex_q * q,struct task_struct * argowner)843 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
844 				struct task_struct *argowner)
845 {
846 	struct futex_pi_state *pi_state = q->pi_state;
847 	int ret;
848 
849 	lockdep_assert_held(q->lock_ptr);
850 
851 	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
852 	ret = __fixup_pi_state_owner(uaddr, q, argowner);
853 	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
854 	return ret;
855 }
856 
857 /**
858  * fixup_pi_owner() - Post lock pi_state and corner case management
859  * @uaddr:	user address of the futex
860  * @q:		futex_q (contains pi_state and access to the rt_mutex)
861  * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
862  *
863  * After attempting to lock an rt_mutex, this function is called to cleanup
864  * the pi_state owner as well as handle race conditions that may allow us to
865  * acquire the lock. Must be called with the hb lock held.
866  *
867  * Return:
868  *  -  1 - success, lock taken;
869  *  -  0 - success, lock not taken;
870  *  - <0 - on error (-EFAULT)
871  */
fixup_pi_owner(u32 __user * uaddr,struct futex_q * q,int locked)872 int fixup_pi_owner(u32 __user *uaddr, struct futex_q *q, int locked)
873 {
874 	if (locked) {
875 		/*
876 		 * Got the lock. We might not be the anticipated owner if we
877 		 * did a lock-steal - fix up the PI-state in that case:
878 		 *
879 		 * Speculative pi_state->owner read (we don't hold wait_lock);
880 		 * since we own the lock pi_state->owner == current is the
881 		 * stable state, anything else needs more attention.
882 		 */
883 		if (q->pi_state->owner != current)
884 			return fixup_pi_state_owner(uaddr, q, current);
885 		return 1;
886 	}
887 
888 	/*
889 	 * If we didn't get the lock; check if anybody stole it from us. In
890 	 * that case, we need to fix up the uval to point to them instead of
891 	 * us, otherwise bad things happen. [10]
892 	 *
893 	 * Another speculative read; pi_state->owner == current is unstable
894 	 * but needs our attention.
895 	 */
896 	if (q->pi_state->owner == current)
897 		return fixup_pi_state_owner(uaddr, q, NULL);
898 
899 	/*
900 	 * Paranoia check. If we did not take the lock, then we should not be
901 	 * the owner of the rt_mutex. Warn and establish consistent state.
902 	 */
903 	if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
904 		return fixup_pi_state_owner(uaddr, q, current);
905 
906 	return 0;
907 }
908 
909 /*
910  * Userspace tried a 0 -> TID atomic transition of the futex value
911  * and failed. The kernel side here does the whole locking operation:
912  * if there are waiters then it will block as a consequence of relying
913  * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
914  * a 0 value of the futex too.).
915  *
916  * Also serves as futex trylock_pi()'ing, and due semantics.
917  */
futex_lock_pi(u32 __user * uaddr,unsigned int flags,ktime_t * time,int trylock)918 int futex_lock_pi(u32 __user *uaddr, unsigned int flags, ktime_t *time, int trylock)
919 {
920 	struct hrtimer_sleeper timeout, *to;
921 	struct task_struct *exiting = NULL;
922 	struct rt_mutex_waiter rt_waiter;
923 	struct futex_hash_bucket *hb;
924 	struct futex_q q = futex_q_init;
925 	int res, ret;
926 
927 	if (!IS_ENABLED(CONFIG_FUTEX_PI))
928 		return -ENOSYS;
929 
930 	if (refill_pi_state_cache())
931 		return -ENOMEM;
932 
933 	to = futex_setup_timer(time, &timeout, flags, 0);
934 
935 retry:
936 	ret = get_futex_key(uaddr, flags, &q.key, FUTEX_WRITE);
937 	if (unlikely(ret != 0))
938 		goto out;
939 
940 retry_private:
941 	hb = futex_q_lock(&q);
942 
943 	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
944 				   &exiting, 0);
945 	if (unlikely(ret)) {
946 		/*
947 		 * Atomic work succeeded and we got the lock,
948 		 * or failed. Either way, we do _not_ block.
949 		 */
950 		switch (ret) {
951 		case 1:
952 			/* We got the lock. */
953 			ret = 0;
954 			goto out_unlock_put_key;
955 		case -EFAULT:
956 			goto uaddr_faulted;
957 		case -EBUSY:
958 		case -EAGAIN:
959 			/*
960 			 * Two reasons for this:
961 			 * - EBUSY: Task is exiting and we just wait for the
962 			 *   exit to complete.
963 			 * - EAGAIN: The user space value changed.
964 			 */
965 			futex_q_unlock(hb);
966 			/*
967 			 * Handle the case where the owner is in the middle of
968 			 * exiting. Wait for the exit to complete otherwise
969 			 * this task might loop forever, aka. live lock.
970 			 */
971 			wait_for_owner_exiting(ret, exiting);
972 			cond_resched();
973 			goto retry;
974 		default:
975 			goto out_unlock_put_key;
976 		}
977 	}
978 
979 	WARN_ON(!q.pi_state);
980 
981 	/*
982 	 * Only actually queue now that the atomic ops are done:
983 	 */
984 	__futex_queue(&q, hb);
985 
986 	if (trylock) {
987 		ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
988 		/* Fixup the trylock return value: */
989 		ret = ret ? 0 : -EWOULDBLOCK;
990 		goto no_block;
991 	}
992 
993 	/*
994 	 * Must be done before we enqueue the waiter, here is unfortunately
995 	 * under the hb lock, but that *should* work because it does nothing.
996 	 */
997 	rt_mutex_pre_schedule();
998 
999 	rt_mutex_init_waiter(&rt_waiter);
1000 
1001 	/*
1002 	 * On PREEMPT_RT, when hb->lock becomes an rt_mutex, we must not
1003 	 * hold it while doing rt_mutex_start_proxy(), because then it will
1004 	 * include hb->lock in the blocking chain, even through we'll not in
1005 	 * fact hold it while blocking. This will lead it to report -EDEADLK
1006 	 * and BUG when futex_unlock_pi() interleaves with this.
1007 	 *
1008 	 * Therefore acquire wait_lock while holding hb->lock, but drop the
1009 	 * latter before calling __rt_mutex_start_proxy_lock(). This
1010 	 * interleaves with futex_unlock_pi() -- which does a similar lock
1011 	 * handoff -- such that the latter can observe the futex_q::pi_state
1012 	 * before __rt_mutex_start_proxy_lock() is done.
1013 	 */
1014 	raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
1015 	spin_unlock(q.lock_ptr);
1016 	/*
1017 	 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
1018 	 * such that futex_unlock_pi() is guaranteed to observe the waiter when
1019 	 * it sees the futex_q::pi_state.
1020 	 */
1021 	ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
1022 	raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
1023 
1024 	if (ret) {
1025 		if (ret == 1)
1026 			ret = 0;
1027 		goto cleanup;
1028 	}
1029 
1030 	if (unlikely(to))
1031 		hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
1032 
1033 	ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
1034 
1035 cleanup:
1036 	/*
1037 	 * If we failed to acquire the lock (deadlock/signal/timeout), we must
1038 	 * must unwind the above, however we canont lock hb->lock because
1039 	 * rt_mutex already has a waiter enqueued and hb->lock can itself try
1040 	 * and enqueue an rt_waiter through rtlock.
1041 	 *
1042 	 * Doing the cleanup without holding hb->lock can cause inconsistent
1043 	 * state between hb and pi_state, but only in the direction of not
1044 	 * seeing a waiter that is leaving.
1045 	 *
1046 	 * See futex_unlock_pi(), it deals with this inconsistency.
1047 	 *
1048 	 * There be dragons here, since we must deal with the inconsistency on
1049 	 * the way out (here), it is impossible to detect/warn about the race
1050 	 * the other way around (missing an incoming waiter).
1051 	 *
1052 	 * What could possibly go wrong...
1053 	 */
1054 	if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
1055 		ret = 0;
1056 
1057 	/*
1058 	 * Now that the rt_waiter has been dequeued, it is safe to use
1059 	 * spinlock/rtlock (which might enqueue its own rt_waiter) and fix up
1060 	 * the
1061 	 */
1062 	spin_lock(q.lock_ptr);
1063 	/*
1064 	 * Waiter is unqueued.
1065 	 */
1066 	rt_mutex_post_schedule();
1067 no_block:
1068 	/*
1069 	 * Fixup the pi_state owner and possibly acquire the lock if we
1070 	 * haven't already.
1071 	 */
1072 	res = fixup_pi_owner(uaddr, &q, !ret);
1073 	/*
1074 	 * If fixup_pi_owner() returned an error, propagate that.  If it acquired
1075 	 * the lock, clear our -ETIMEDOUT or -EINTR.
1076 	 */
1077 	if (res)
1078 		ret = (res < 0) ? res : 0;
1079 
1080 	futex_unqueue_pi(&q);
1081 	spin_unlock(q.lock_ptr);
1082 	goto out;
1083 
1084 out_unlock_put_key:
1085 	futex_q_unlock(hb);
1086 
1087 out:
1088 	if (to) {
1089 		hrtimer_cancel(&to->timer);
1090 		destroy_hrtimer_on_stack(&to->timer);
1091 	}
1092 	return ret != -EINTR ? ret : -ERESTARTNOINTR;
1093 
1094 uaddr_faulted:
1095 	futex_q_unlock(hb);
1096 
1097 	ret = fault_in_user_writeable(uaddr);
1098 	if (ret)
1099 		goto out;
1100 
1101 	if (!(flags & FLAGS_SHARED))
1102 		goto retry_private;
1103 
1104 	goto retry;
1105 }
1106 
1107 /*
1108  * Userspace attempted a TID -> 0 atomic transition, and failed.
1109  * This is the in-kernel slowpath: we look up the PI state (if any),
1110  * and do the rt-mutex unlock.
1111  */
futex_unlock_pi(u32 __user * uaddr,unsigned int flags)1112 int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
1113 {
1114 	u32 curval, uval, vpid = task_pid_vnr(current);
1115 	union futex_key key = FUTEX_KEY_INIT;
1116 	struct futex_hash_bucket *hb;
1117 	struct futex_q *top_waiter;
1118 	int ret;
1119 
1120 	if (!IS_ENABLED(CONFIG_FUTEX_PI))
1121 		return -ENOSYS;
1122 
1123 retry:
1124 	if (get_user(uval, uaddr))
1125 		return -EFAULT;
1126 	/*
1127 	 * We release only a lock we actually own:
1128 	 */
1129 	if ((uval & FUTEX_TID_MASK) != vpid)
1130 		return -EPERM;
1131 
1132 	ret = get_futex_key(uaddr, flags, &key, FUTEX_WRITE);
1133 	if (ret)
1134 		return ret;
1135 
1136 	hb = futex_hash(&key);
1137 	spin_lock(&hb->lock);
1138 retry_hb:
1139 
1140 	/*
1141 	 * Check waiters first. We do not trust user space values at
1142 	 * all and we at least want to know if user space fiddled
1143 	 * with the futex value instead of blindly unlocking.
1144 	 */
1145 	top_waiter = futex_top_waiter(hb, &key);
1146 	if (top_waiter) {
1147 		struct futex_pi_state *pi_state = top_waiter->pi_state;
1148 		struct rt_mutex_waiter *rt_waiter;
1149 
1150 		ret = -EINVAL;
1151 		if (!pi_state)
1152 			goto out_unlock;
1153 
1154 		/*
1155 		 * If current does not own the pi_state then the futex is
1156 		 * inconsistent and user space fiddled with the futex value.
1157 		 */
1158 		if (pi_state->owner != current)
1159 			goto out_unlock;
1160 
1161 		/*
1162 		 * By taking wait_lock while still holding hb->lock, we ensure
1163 		 * there is no point where we hold neither; and thereby
1164 		 * wake_futex_pi() must observe any new waiters.
1165 		 *
1166 		 * Since the cleanup: case in futex_lock_pi() removes the
1167 		 * rt_waiter without holding hb->lock, it is possible for
1168 		 * wake_futex_pi() to not find a waiter while the above does,
1169 		 * in this case the waiter is on the way out and it can be
1170 		 * ignored.
1171 		 *
1172 		 * In particular; this forces __rt_mutex_start_proxy() to
1173 		 * complete such that we're guaranteed to observe the
1174 		 * rt_waiter.
1175 		 */
1176 		raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1177 
1178 		/*
1179 		 * Futex vs rt_mutex waiter state -- if there are no rt_mutex
1180 		 * waiters even though futex thinks there are, then the waiter
1181 		 * is leaving. The entry needs to be removed from the list so a
1182 		 * new futex_lock_pi() is not using this stale PI-state while
1183 		 * the futex is available in user space again.
1184 		 * There can be more than one task on its way out so it needs
1185 		 * to retry.
1186 		 */
1187 		rt_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1188 		if (!rt_waiter) {
1189 			__futex_unqueue(top_waiter);
1190 			raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1191 			goto retry_hb;
1192 		}
1193 
1194 		get_pi_state(pi_state);
1195 		spin_unlock(&hb->lock);
1196 
1197 		/* drops pi_state->pi_mutex.wait_lock */
1198 		ret = wake_futex_pi(uaddr, uval, pi_state, rt_waiter);
1199 
1200 		put_pi_state(pi_state);
1201 
1202 		/*
1203 		 * Success, we're done! No tricky corner cases.
1204 		 */
1205 		if (!ret)
1206 			return ret;
1207 		/*
1208 		 * The atomic access to the futex value generated a
1209 		 * pagefault, so retry the user-access and the wakeup:
1210 		 */
1211 		if (ret == -EFAULT)
1212 			goto pi_faulted;
1213 		/*
1214 		 * A unconditional UNLOCK_PI op raced against a waiter
1215 		 * setting the FUTEX_WAITERS bit. Try again.
1216 		 */
1217 		if (ret == -EAGAIN)
1218 			goto pi_retry;
1219 		/*
1220 		 * wake_futex_pi has detected invalid state. Tell user
1221 		 * space.
1222 		 */
1223 		return ret;
1224 	}
1225 
1226 	/*
1227 	 * We have no kernel internal state, i.e. no waiters in the
1228 	 * kernel. Waiters which are about to queue themselves are stuck
1229 	 * on hb->lock. So we can safely ignore them. We do neither
1230 	 * preserve the WAITERS bit not the OWNER_DIED one. We are the
1231 	 * owner.
1232 	 */
1233 	if ((ret = futex_cmpxchg_value_locked(&curval, uaddr, uval, 0))) {
1234 		spin_unlock(&hb->lock);
1235 		switch (ret) {
1236 		case -EFAULT:
1237 			goto pi_faulted;
1238 
1239 		case -EAGAIN:
1240 			goto pi_retry;
1241 
1242 		default:
1243 			WARN_ON_ONCE(1);
1244 			return ret;
1245 		}
1246 	}
1247 
1248 	/*
1249 	 * If uval has changed, let user space handle it.
1250 	 */
1251 	ret = (curval == uval) ? 0 : -EAGAIN;
1252 
1253 out_unlock:
1254 	spin_unlock(&hb->lock);
1255 	return ret;
1256 
1257 pi_retry:
1258 	cond_resched();
1259 	goto retry;
1260 
1261 pi_faulted:
1262 
1263 	ret = fault_in_user_writeable(uaddr);
1264 	if (!ret)
1265 		goto retry;
1266 
1267 	return ret;
1268 }
1269 
1270