Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   4 *
   5 * started by Ingo Molnar and Thomas Gleixner.
   6 *
   7 *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   8 *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   9 *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
  10 *  Copyright (C) 2006 Esben Nielsen
  11 * Adaptive Spinlocks:
  12 *  Copyright (C) 2008 Novell, Inc., Gregory Haskins, Sven Dietrich,
  13 *				     and Peter Morreale,
  14 * Adaptive Spinlocks simplification:
  15 *  Copyright (C) 2008 Red Hat, Inc., Steven Rostedt <srostedt@redhat.com>
  16 *
  17 *  See Documentation/locking/rt-mutex-design.rst for details.
  18 */
  19#include <linux/sched.h>
  20#include <linux/sched/debug.h>
  21#include <linux/sched/deadline.h>
  22#include <linux/sched/signal.h>
  23#include <linux/sched/rt.h>
 
  24#include <linux/sched/wake_q.h>
  25#include <linux/ww_mutex.h>
  26
  27#include <trace/events/lock.h>
  28
  29#include "rtmutex_common.h"
  30
  31#ifndef WW_RT
  32# define build_ww_mutex()	(false)
  33# define ww_container_of(rtm)	NULL
  34
  35static inline int __ww_mutex_add_waiter(struct rt_mutex_waiter *waiter,
  36					struct rt_mutex *lock,
  37					struct ww_acquire_ctx *ww_ctx,
  38					struct wake_q_head *wake_q)
  39{
  40	return 0;
  41}
  42
  43static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
  44					    struct ww_acquire_ctx *ww_ctx,
  45					    struct wake_q_head *wake_q)
  46{
  47}
  48
  49static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
  50					  struct ww_acquire_ctx *ww_ctx)
  51{
  52}
  53
  54static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
  55					struct rt_mutex_waiter *waiter,
  56					struct ww_acquire_ctx *ww_ctx)
  57{
  58	return 0;
  59}
  60
  61#else
  62# define build_ww_mutex()	(true)
  63# define ww_container_of(rtm)	container_of(rtm, struct ww_mutex, base)
  64# include "ww_mutex.h"
  65#endif
  66
  67/*
  68 * lock->owner state tracking:
  69 *
  70 * lock->owner holds the task_struct pointer of the owner. Bit 0
  71 * is used to keep track of the "lock has waiters" state.
  72 *
  73 * owner	bit0
  74 * NULL		0	lock is free (fast acquire possible)
  75 * NULL		1	lock is free and has waiters and the top waiter
  76 *				is going to take the lock*
  77 * taskpointer	0	lock is held (fast release possible)
  78 * taskpointer	1	lock is held and has waiters**
  79 *
  80 * The fast atomic compare exchange based acquire and release is only
  81 * possible when bit 0 of lock->owner is 0.
  82 *
  83 * (*) It also can be a transitional state when grabbing the lock
  84 * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  85 * we need to set the bit0 before looking at the lock, and the owner may be
  86 * NULL in this small time, hence this can be a transitional state.
  87 *
  88 * (**) There is a small time when bit 0 is set but there are no
  89 * waiters. This can happen when grabbing the lock in the slow path.
  90 * To prevent a cmpxchg of the owner releasing the lock, we need to
  91 * set this bit before looking at the lock.
  92 */
  93
  94static __always_inline struct task_struct *
  95rt_mutex_owner_encode(struct rt_mutex_base *lock, struct task_struct *owner)
  96{
  97	unsigned long val = (unsigned long)owner;
  98
  99	if (rt_mutex_has_waiters(lock))
 100		val |= RT_MUTEX_HAS_WAITERS;
 101
 102	return (struct task_struct *)val;
 103}
 104
 105static __always_inline void
 106rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
 107{
 108	/*
 109	 * lock->wait_lock is held but explicit acquire semantics are needed
 110	 * for a new lock owner so WRITE_ONCE is insufficient.
 111	 */
 112	xchg_acquire(&lock->owner, rt_mutex_owner_encode(lock, owner));
 113}
 114
 115static __always_inline void rt_mutex_clear_owner(struct rt_mutex_base *lock)
 116{
 117	/* lock->wait_lock is held so the unlock provides release semantics. */
 118	WRITE_ONCE(lock->owner, rt_mutex_owner_encode(lock, NULL));
 119}
 120
 121static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
 122{
 123	lock->owner = (struct task_struct *)
 124			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
 125}
 126
 127static __always_inline void
 128fixup_rt_mutex_waiters(struct rt_mutex_base *lock, bool acquire_lock)
 129{
 130	unsigned long owner, *p = (unsigned long *) &lock->owner;
 131
 132	if (rt_mutex_has_waiters(lock))
 133		return;
 134
 135	/*
 136	 * The rbtree has no waiters enqueued, now make sure that the
 137	 * lock->owner still has the waiters bit set, otherwise the
 138	 * following can happen:
 139	 *
 140	 * CPU 0	CPU 1		CPU2
 141	 * l->owner=T1
 142	 *		rt_mutex_lock(l)
 143	 *		lock(l->lock)
 144	 *		l->owner = T1 | HAS_WAITERS;
 145	 *		enqueue(T2)
 146	 *		boost()
 147	 *		  unlock(l->lock)
 148	 *		block()
 149	 *
 150	 *				rt_mutex_lock(l)
 151	 *				lock(l->lock)
 152	 *				l->owner = T1 | HAS_WAITERS;
 153	 *				enqueue(T3)
 154	 *				boost()
 155	 *				  unlock(l->lock)
 156	 *				block()
 157	 *		signal(->T2)	signal(->T3)
 158	 *		lock(l->lock)
 159	 *		dequeue(T2)
 160	 *		deboost()
 161	 *		  unlock(l->lock)
 162	 *				lock(l->lock)
 163	 *				dequeue(T3)
 164	 *				 ==> wait list is empty
 165	 *				deboost()
 166	 *				 unlock(l->lock)
 167	 *		lock(l->lock)
 168	 *		fixup_rt_mutex_waiters()
 169	 *		  if (wait_list_empty(l) {
 170	 *		    l->owner = owner
 171	 *		    owner = l->owner & ~HAS_WAITERS;
 172	 *		      ==> l->owner = T1
 173	 *		  }
 174	 *				lock(l->lock)
 175	 * rt_mutex_unlock(l)		fixup_rt_mutex_waiters()
 176	 *				  if (wait_list_empty(l) {
 177	 *				    owner = l->owner & ~HAS_WAITERS;
 178	 * cmpxchg(l->owner, T1, NULL)
 179	 *  ===> Success (l->owner = NULL)
 180	 *
 181	 *				    l->owner = owner
 182	 *				      ==> l->owner = T1
 183	 *				  }
 184	 *
 185	 * With the check for the waiter bit in place T3 on CPU2 will not
 186	 * overwrite. All tasks fiddling with the waiters bit are
 187	 * serialized by l->lock, so nothing else can modify the waiters
 188	 * bit. If the bit is set then nothing can change l->owner either
 189	 * so the simple RMW is safe. The cmpxchg() will simply fail if it
 190	 * happens in the middle of the RMW because the waiters bit is
 191	 * still set.
 192	 */
 193	owner = READ_ONCE(*p);
 194	if (owner & RT_MUTEX_HAS_WAITERS) {
 195		/*
 196		 * See rt_mutex_set_owner() and rt_mutex_clear_owner() on
 197		 * why xchg_acquire() is used for updating owner for
 198		 * locking and WRITE_ONCE() for unlocking.
 199		 *
 200		 * WRITE_ONCE() would work for the acquire case too, but
 201		 * in case that the lock acquisition failed it might
 202		 * force other lockers into the slow path unnecessarily.
 203		 */
 204		if (acquire_lock)
 205			xchg_acquire(p, owner & ~RT_MUTEX_HAS_WAITERS);
 206		else
 207			WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
 208	}
 209}
 210
 211/*
 212 * We can speed up the acquire/release, if there's no debugging state to be
 213 * set up.
 214 */
 215#ifndef CONFIG_DEBUG_RT_MUTEXES
 216static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
 217						     struct task_struct *old,
 218						     struct task_struct *new)
 219{
 220	return try_cmpxchg_acquire(&lock->owner, &old, new);
 221}
 222
 223static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
 224{
 225	return rt_mutex_cmpxchg_acquire(lock, NULL, current);
 226}
 227
 228static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
 229						     struct task_struct *old,
 230						     struct task_struct *new)
 231{
 232	return try_cmpxchg_release(&lock->owner, &old, new);
 233}
 234
 235/*
 236 * Callers must hold the ->wait_lock -- which is the whole purpose as we force
 237 * all future threads that attempt to [Rmw] the lock to the slowpath. As such
 238 * relaxed semantics suffice.
 239 */
 240static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 241{
 242	unsigned long *p = (unsigned long *) &lock->owner;
 243	unsigned long owner, new;
 244
 245	owner = READ_ONCE(*p);
 246	do {
 247		new = owner | RT_MUTEX_HAS_WAITERS;
 248	} while (!try_cmpxchg_relaxed(p, &owner, new));
 249
 250	/*
 251	 * The cmpxchg loop above is relaxed to avoid back-to-back ACQUIRE
 252	 * operations in the event of contention. Ensure the successful
 253	 * cmpxchg is visible.
 254	 */
 255	smp_mb__after_atomic();
 256}
 257
 258/*
 259 * Safe fastpath aware unlock:
 260 * 1) Clear the waiters bit
 261 * 2) Drop lock->wait_lock
 262 * 3) Try to unlock the lock with cmpxchg
 263 */
 264static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
 265						 unsigned long flags)
 266	__releases(lock->wait_lock)
 267{
 268	struct task_struct *owner = rt_mutex_owner(lock);
 269
 270	clear_rt_mutex_waiters(lock);
 271	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 272	/*
 273	 * If a new waiter comes in between the unlock and the cmpxchg
 274	 * we have two situations:
 275	 *
 276	 * unlock(wait_lock);
 277	 *					lock(wait_lock);
 278	 * cmpxchg(p, owner, 0) == owner
 279	 *					mark_rt_mutex_waiters(lock);
 280	 *					acquire(lock);
 281	 * or:
 282	 *
 283	 * unlock(wait_lock);
 284	 *					lock(wait_lock);
 285	 *					mark_rt_mutex_waiters(lock);
 286	 *
 287	 * cmpxchg(p, owner, 0) != owner
 288	 *					enqueue_waiter();
 289	 *					unlock(wait_lock);
 290	 * lock(wait_lock);
 291	 * wake waiter();
 292	 * unlock(wait_lock);
 293	 *					lock(wait_lock);
 294	 *					acquire(lock);
 295	 */
 296	return rt_mutex_cmpxchg_release(lock, owner, NULL);
 297}
 298
 299#else
 300static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
 301						     struct task_struct *old,
 302						     struct task_struct *new)
 303{
 304	return false;
 305
 306}
 307
 308static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock);
 309
 310static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
 311{
 312	/*
 313	 * With debug enabled rt_mutex_cmpxchg trylock() will always fail.
 314	 *
 315	 * Avoid unconditionally taking the slow path by using
 316	 * rt_mutex_slow_trylock() which is covered by the debug code and can
 317	 * acquire a non-contended rtmutex.
 318	 */
 319	return rt_mutex_slowtrylock(lock);
 320}
 321
 322static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
 323						     struct task_struct *old,
 324						     struct task_struct *new)
 325{
 326	return false;
 327}
 328
 329static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 330{
 331	lock->owner = (struct task_struct *)
 332			((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 333}
 334
 335/*
 336 * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 337 */
 338static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
 339						 unsigned long flags)
 340	__releases(lock->wait_lock)
 341{
 342	lock->owner = NULL;
 343	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 344	return true;
 345}
 346#endif
 347
 348static __always_inline int __waiter_prio(struct task_struct *task)
 349{
 350	int prio = task->prio;
 351
 352	if (!rt_or_dl_prio(prio))
 353		return DEFAULT_PRIO;
 354
 355	return prio;
 356}
 357
 358/*
 359 * Update the waiter->tree copy of the sort keys.
 360 */
 361static __always_inline void
 362waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
 363{
 364	lockdep_assert_held(&waiter->lock->wait_lock);
 365	lockdep_assert(RB_EMPTY_NODE(&waiter->tree.entry));
 366
 367	waiter->tree.prio = __waiter_prio(task);
 368	waiter->tree.deadline = task->dl.deadline;
 369}
 370
 371/*
 372 * Update the waiter->pi_tree copy of the sort keys (from the tree copy).
 373 */
 374static __always_inline void
 375waiter_clone_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
 376{
 377	lockdep_assert_held(&waiter->lock->wait_lock);
 378	lockdep_assert_held(&task->pi_lock);
 379	lockdep_assert(RB_EMPTY_NODE(&waiter->pi_tree.entry));
 380
 381	waiter->pi_tree.prio = waiter->tree.prio;
 382	waiter->pi_tree.deadline = waiter->tree.deadline;
 383}
 384
 385/*
 386 * Only use with rt_waiter_node_{less,equal}()
 387 */
 388#define task_to_waiter_node(p)	\
 389	&(struct rt_waiter_node){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
 390#define task_to_waiter(p)	\
 391	&(struct rt_mutex_waiter){ .tree = *task_to_waiter_node(p) }
 392
 393static __always_inline int rt_waiter_node_less(struct rt_waiter_node *left,
 394					       struct rt_waiter_node *right)
 395{
 396	if (left->prio < right->prio)
 397		return 1;
 398
 399	/*
 400	 * If both waiters have dl_prio(), we check the deadlines of the
 401	 * associated tasks.
 402	 * If left waiter has a dl_prio(), and we didn't return 1 above,
 403	 * then right waiter has a dl_prio() too.
 404	 */
 405	if (dl_prio(left->prio))
 406		return dl_time_before(left->deadline, right->deadline);
 407
 408	return 0;
 409}
 410
 411static __always_inline int rt_waiter_node_equal(struct rt_waiter_node *left,
 412						 struct rt_waiter_node *right)
 413{
 414	if (left->prio != right->prio)
 415		return 0;
 416
 417	/*
 418	 * If both waiters have dl_prio(), we check the deadlines of the
 419	 * associated tasks.
 420	 * If left waiter has a dl_prio(), and we didn't return 0 above,
 421	 * then right waiter has a dl_prio() too.
 422	 */
 423	if (dl_prio(left->prio))
 424		return left->deadline == right->deadline;
 425
 426	return 1;
 427}
 428
 429static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
 430				  struct rt_mutex_waiter *top_waiter)
 431{
 432	if (rt_waiter_node_less(&waiter->tree, &top_waiter->tree))
 433		return true;
 434
 435#ifdef RT_MUTEX_BUILD_SPINLOCKS
 436	/*
 437	 * Note that RT tasks are excluded from same priority (lateral)
 438	 * steals to prevent the introduction of an unbounded latency.
 439	 */
 440	if (rt_or_dl_prio(waiter->tree.prio))
 441		return false;
 442
 443	return rt_waiter_node_equal(&waiter->tree, &top_waiter->tree);
 444#else
 445	return false;
 446#endif
 447}
 448
 449#define __node_2_waiter(node) \
 450	rb_entry((node), struct rt_mutex_waiter, tree.entry)
 451
 452static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
 453{
 454	struct rt_mutex_waiter *aw = __node_2_waiter(a);
 455	struct rt_mutex_waiter *bw = __node_2_waiter(b);
 456
 457	if (rt_waiter_node_less(&aw->tree, &bw->tree))
 458		return 1;
 459
 460	if (!build_ww_mutex())
 461		return 0;
 462
 463	if (rt_waiter_node_less(&bw->tree, &aw->tree))
 464		return 0;
 465
 466	/* NOTE: relies on waiter->ww_ctx being set before insertion */
 467	if (aw->ww_ctx) {
 468		if (!bw->ww_ctx)
 469			return 1;
 470
 471		return (signed long)(aw->ww_ctx->stamp -
 472				     bw->ww_ctx->stamp) < 0;
 473	}
 474
 475	return 0;
 476}
 477
 478static __always_inline void
 479rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 480{
 481	lockdep_assert_held(&lock->wait_lock);
 482
 483	rb_add_cached(&waiter->tree.entry, &lock->waiters, __waiter_less);
 484}
 485
 486static __always_inline void
 487rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 488{
 489	lockdep_assert_held(&lock->wait_lock);
 490
 491	if (RB_EMPTY_NODE(&waiter->tree.entry))
 492		return;
 493
 494	rb_erase_cached(&waiter->tree.entry, &lock->waiters);
 495	RB_CLEAR_NODE(&waiter->tree.entry);
 496}
 497
 498#define __node_2_rt_node(node) \
 499	rb_entry((node), struct rt_waiter_node, entry)
 500
 501static __always_inline bool __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
 
 502{
 503	return rt_waiter_node_less(__node_2_rt_node(a), __node_2_rt_node(b));
 504}
 505
 506static __always_inline void
 507rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 508{
 509	lockdep_assert_held(&task->pi_lock);
 510
 511	rb_add_cached(&waiter->pi_tree.entry, &task->pi_waiters, __pi_waiter_less);
 512}
 513
 514static __always_inline void
 515rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 516{
 517	lockdep_assert_held(&task->pi_lock);
 518
 519	if (RB_EMPTY_NODE(&waiter->pi_tree.entry))
 520		return;
 521
 522	rb_erase_cached(&waiter->pi_tree.entry, &task->pi_waiters);
 523	RB_CLEAR_NODE(&waiter->pi_tree.entry);
 524}
 525
 526static __always_inline void rt_mutex_adjust_prio(struct rt_mutex_base *lock,
 527						 struct task_struct *p)
 528{
 529	struct task_struct *pi_task = NULL;
 530
 531	lockdep_assert_held(&lock->wait_lock);
 532	lockdep_assert(rt_mutex_owner(lock) == p);
 533	lockdep_assert_held(&p->pi_lock);
 534
 535	if (task_has_pi_waiters(p))
 536		pi_task = task_top_pi_waiter(p)->task;
 537
 538	rt_mutex_setprio(p, pi_task);
 539}
 540
 541/* RT mutex specific wake_q wrappers */
 542static __always_inline void rt_mutex_wake_q_add_task(struct rt_wake_q_head *wqh,
 543						     struct task_struct *task,
 544						     unsigned int wake_state)
 545{
 546	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wake_state == TASK_RTLOCK_WAIT) {
 547		if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 548			WARN_ON_ONCE(wqh->rtlock_task);
 549		get_task_struct(task);
 550		wqh->rtlock_task = task;
 551	} else {
 552		wake_q_add(&wqh->head, task);
 553	}
 554}
 555
 556static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
 557						struct rt_mutex_waiter *w)
 558{
 559	rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
 560}
 561
 562static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
 563{
 564	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
 565		wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
 566		put_task_struct(wqh->rtlock_task);
 567		wqh->rtlock_task = NULL;
 568	}
 569
 570	if (!wake_q_empty(&wqh->head))
 571		wake_up_q(&wqh->head);
 572
 573	/* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
 574	preempt_enable();
 575}
 576
 577/*
 578 * Deadlock detection is conditional:
 579 *
 580 * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 581 * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 582 *
 583 * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 584 * conducted independent of the detect argument.
 585 *
 586 * If the waiter argument is NULL this indicates the deboost path and
 587 * deadlock detection is disabled independent of the detect argument
 588 * and the config settings.
 589 */
 590static __always_inline bool
 591rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 592			      enum rtmutex_chainwalk chwalk)
 593{
 594	if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
 595		return waiter != NULL;
 596	return chwalk == RT_MUTEX_FULL_CHAINWALK;
 597}
 598
 599static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
 
 
 
 
 
 600{
 601	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 602}
 603
 604/*
 605 * Adjust the priority chain. Also used for deadlock detection.
 606 * Decreases task's usage by one - may thus free the task.
 607 *
 608 * @task:	the task owning the mutex (owner) for which a chain walk is
 609 *		probably needed
 610 * @chwalk:	do we have to carry out deadlock detection?
 611 * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
 612 *		things for a task that has just got its priority adjusted, and
 613 *		is waiting on a mutex)
 614 * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
 615 *		we dropped its pi_lock. Is never dereferenced, only used for
 616 *		comparison to detect lock chain changes.
 617 * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 618 *		its priority to the mutex owner (can be NULL in the case
 619 *		depicted above or if the top waiter is gone away and we are
 620 *		actually deboosting the owner)
 621 * @top_task:	the current top waiter
 622 *
 623 * Returns 0 or -EDEADLK.
 624 *
 625 * Chain walk basics and protection scope
 626 *
 627 * [R] refcount on task
 628 * [Pn] task->pi_lock held
 629 * [L] rtmutex->wait_lock held
 630 *
 631 * Normal locking order:
 632 *
 633 *   rtmutex->wait_lock
 634 *     task->pi_lock
 635 *
 636 * Step	Description				Protected by
 637 *	function arguments:
 638 *	@task					[R]
 639 *	@orig_lock if != NULL			@top_task is blocked on it
 640 *	@next_lock				Unprotected. Cannot be
 641 *						dereferenced. Only used for
 642 *						comparison.
 643 *	@orig_waiter if != NULL			@top_task is blocked on it
 644 *	@top_task				current, or in case of proxy
 645 *						locking protected by calling
 646 *						code
 647 *	again:
 648 *	  loop_sanity_check();
 649 *	retry:
 650 * [1]	  lock(task->pi_lock);			[R] acquire [P1]
 651 * [2]	  waiter = task->pi_blocked_on;		[P1]
 652 * [3]	  check_exit_conditions_1();		[P1]
 653 * [4]	  lock = waiter->lock;			[P1]
 654 * [5]	  if (!try_lock(lock->wait_lock)) {	[P1] try to acquire [L]
 655 *	    unlock(task->pi_lock);		release [P1]
 656 *	    goto retry;
 657 *	  }
 658 * [6]	  check_exit_conditions_2();		[P1] + [L]
 659 * [7]	  requeue_lock_waiter(lock, waiter);	[P1] + [L]
 660 * [8]	  unlock(task->pi_lock);		release [P1]
 661 *	  put_task_struct(task);		release [R]
 662 * [9]	  check_exit_conditions_3();		[L]
 663 * [10]	  task = owner(lock);			[L]
 664 *	  get_task_struct(task);		[L] acquire [R]
 665 *	  lock(task->pi_lock);			[L] acquire [P2]
 666 * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P2] + [L]
 667 * [12]	  check_exit_conditions_4();		[P2] + [L]
 668 * [13]	  unlock(task->pi_lock);		release [P2]
 669 *	  unlock(lock->wait_lock);		release [L]
 670 *	  goto again;
 671 *
 672 * Where P1 is the blocking task and P2 is the lock owner; going up one step
 673 * the owner becomes the next blocked task etc..
 674 *
 675*
 676 */
 677static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
 678					      enum rtmutex_chainwalk chwalk,
 679					      struct rt_mutex_base *orig_lock,
 680					      struct rt_mutex_base *next_lock,
 681					      struct rt_mutex_waiter *orig_waiter,
 682					      struct task_struct *top_task)
 683{
 684	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 685	struct rt_mutex_waiter *prerequeue_top_waiter;
 686	int ret = 0, depth = 0;
 687	struct rt_mutex_base *lock;
 688	bool detect_deadlock;
 689	bool requeue = true;
 690
 691	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 692
 693	/*
 694	 * The (de)boosting is a step by step approach with a lot of
 695	 * pitfalls. We want this to be preemptible and we want hold a
 696	 * maximum of two locks per step. So we have to check
 697	 * carefully whether things change under us.
 698	 */
 699 again:
 700	/*
 701	 * We limit the lock chain length for each invocation.
 702	 */
 703	if (++depth > max_lock_depth) {
 704		static int prev_max;
 705
 706		/*
 707		 * Print this only once. If the admin changes the limit,
 708		 * print a new message when reaching the limit again.
 709		 */
 710		if (prev_max != max_lock_depth) {
 711			prev_max = max_lock_depth;
 712			printk(KERN_WARNING "Maximum lock depth %d reached "
 713			       "task: %s (%d)\n", max_lock_depth,
 714			       top_task->comm, task_pid_nr(top_task));
 715		}
 716		put_task_struct(task);
 717
 718		return -EDEADLK;
 719	}
 720
 721	/*
 722	 * We are fully preemptible here and only hold the refcount on
 723	 * @task. So everything can have changed under us since the
 724	 * caller or our own code below (goto retry/again) dropped all
 725	 * locks.
 726	 */
 727 retry:
 728	/*
 729	 * [1] Task cannot go away as we did a get_task() before !
 730	 */
 731	raw_spin_lock_irq(&task->pi_lock);
 732
 733	/*
 734	 * [2] Get the waiter on which @task is blocked on.
 735	 */
 736	waiter = task->pi_blocked_on;
 737
 738	/*
 739	 * [3] check_exit_conditions_1() protected by task->pi_lock.
 740	 */
 741
 742	/*
 743	 * Check whether the end of the boosting chain has been
 744	 * reached or the state of the chain has changed while we
 745	 * dropped the locks.
 746	 */
 747	if (!waiter)
 748		goto out_unlock_pi;
 749
 750	/*
 751	 * Check the orig_waiter state. After we dropped the locks,
 752	 * the previous owner of the lock might have released the lock.
 753	 */
 754	if (orig_waiter && !rt_mutex_owner(orig_lock))
 755		goto out_unlock_pi;
 756
 757	/*
 758	 * We dropped all locks after taking a refcount on @task, so
 759	 * the task might have moved on in the lock chain or even left
 760	 * the chain completely and blocks now on an unrelated lock or
 761	 * on @orig_lock.
 762	 *
 763	 * We stored the lock on which @task was blocked in @next_lock,
 764	 * so we can detect the chain change.
 765	 */
 766	if (next_lock != waiter->lock)
 767		goto out_unlock_pi;
 768
 769	/*
 770	 * There could be 'spurious' loops in the lock graph due to ww_mutex,
 771	 * consider:
 772	 *
 773	 *   P1: A, ww_A, ww_B
 774	 *   P2: ww_B, ww_A
 775	 *   P3: A
 776	 *
 777	 * P3 should not return -EDEADLK because it gets trapped in the cycle
 778	 * created by P1 and P2 (which will resolve -- and runs into
 779	 * max_lock_depth above). Therefore disable detect_deadlock such that
 780	 * the below termination condition can trigger once all relevant tasks
 781	 * are boosted.
 782	 *
 783	 * Even when we start with ww_mutex we can disable deadlock detection,
 784	 * since we would supress a ww_mutex induced deadlock at [6] anyway.
 785	 * Supressing it here however is not sufficient since we might still
 786	 * hit [6] due to adjustment driven iteration.
 787	 *
 788	 * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
 789	 * utterly fail to report it; lockdep should.
 790	 */
 791	if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
 792		detect_deadlock = false;
 793
 794	/*
 795	 * Drop out, when the task has no waiters. Note,
 796	 * top_waiter can be NULL, when we are in the deboosting
 797	 * mode!
 798	 */
 799	if (top_waiter) {
 800		if (!task_has_pi_waiters(task))
 801			goto out_unlock_pi;
 802		/*
 803		 * If deadlock detection is off, we stop here if we
 804		 * are not the top pi waiter of the task. If deadlock
 805		 * detection is enabled we continue, but stop the
 806		 * requeueing in the chain walk.
 807		 */
 808		if (top_waiter != task_top_pi_waiter(task)) {
 809			if (!detect_deadlock)
 810				goto out_unlock_pi;
 811			else
 812				requeue = false;
 813		}
 814	}
 815
 816	/*
 817	 * If the waiter priority is the same as the task priority
 818	 * then there is no further priority adjustment necessary.  If
 819	 * deadlock detection is off, we stop the chain walk. If its
 820	 * enabled we continue, but stop the requeueing in the chain
 821	 * walk.
 822	 */
 823	if (rt_waiter_node_equal(&waiter->tree, task_to_waiter_node(task))) {
 824		if (!detect_deadlock)
 825			goto out_unlock_pi;
 826		else
 827			requeue = false;
 828	}
 829
 830	/*
 831	 * [4] Get the next lock; per holding task->pi_lock we can't unblock
 832	 * and guarantee @lock's existence.
 833	 */
 834	lock = waiter->lock;
 835	/*
 836	 * [5] We need to trylock here as we are holding task->pi_lock,
 837	 * which is the reverse lock order versus the other rtmutex
 838	 * operations.
 839	 *
 840	 * Per the above, holding task->pi_lock guarantees lock exists, so
 841	 * inverting this lock order is infeasible from a life-time
 842	 * perspective.
 843	 */
 844	if (!raw_spin_trylock(&lock->wait_lock)) {
 845		raw_spin_unlock_irq(&task->pi_lock);
 846		cpu_relax();
 847		goto retry;
 848	}
 849
 850	/*
 851	 * [6] check_exit_conditions_2() protected by task->pi_lock and
 852	 * lock->wait_lock.
 853	 *
 854	 * Deadlock detection. If the lock is the same as the original
 855	 * lock which caused us to walk the lock chain or if the
 856	 * current lock is owned by the task which initiated the chain
 857	 * walk, we detected a deadlock.
 858	 */
 859	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 860		ret = -EDEADLK;
 861
 862		/*
 863		 * When the deadlock is due to ww_mutex; also see above. Don't
 864		 * report the deadlock and instead let the ww_mutex wound/die
 865		 * logic pick which of the contending threads gets -EDEADLK.
 866		 *
 867		 * NOTE: assumes the cycle only contains a single ww_class; any
 868		 * other configuration and we fail to report; also, see
 869		 * lockdep.
 870		 */
 871		if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
 872			ret = 0;
 873
 874		raw_spin_unlock(&lock->wait_lock);
 
 875		goto out_unlock_pi;
 876	}
 877
 878	/*
 879	 * If we just follow the lock chain for deadlock detection, no
 880	 * need to do all the requeue operations. To avoid a truckload
 881	 * of conditionals around the various places below, just do the
 882	 * minimum chain walk checks.
 883	 */
 884	if (!requeue) {
 885		/*
 886		 * No requeue[7] here. Just release @task [8]
 887		 */
 888		raw_spin_unlock(&task->pi_lock);
 889		put_task_struct(task);
 890
 891		/*
 892		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
 893		 * If there is no owner of the lock, end of chain.
 894		 */
 895		if (!rt_mutex_owner(lock)) {
 896			raw_spin_unlock_irq(&lock->wait_lock);
 897			return 0;
 898		}
 899
 900		/* [10] Grab the next task, i.e. owner of @lock */
 901		task = get_task_struct(rt_mutex_owner(lock));
 902		raw_spin_lock(&task->pi_lock);
 903
 904		/*
 905		 * No requeue [11] here. We just do deadlock detection.
 906		 *
 907		 * [12] Store whether owner is blocked
 908		 * itself. Decision is made after dropping the locks
 909		 */
 910		next_lock = task_blocked_on_lock(task);
 911		/*
 912		 * Get the top waiter for the next iteration
 913		 */
 914		top_waiter = rt_mutex_top_waiter(lock);
 915
 916		/* [13] Drop locks */
 917		raw_spin_unlock(&task->pi_lock);
 918		raw_spin_unlock_irq(&lock->wait_lock);
 919
 920		/* If owner is not blocked, end of chain. */
 921		if (!next_lock)
 922			goto out_put_task;
 923		goto again;
 924	}
 925
 926	/*
 927	 * Store the current top waiter before doing the requeue
 928	 * operation on @lock. We need it for the boost/deboost
 929	 * decision below.
 930	 */
 931	prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 932
 933	/* [7] Requeue the waiter in the lock waiter tree. */
 934	rt_mutex_dequeue(lock, waiter);
 935
 936	/*
 937	 * Update the waiter prio fields now that we're dequeued.
 938	 *
 939	 * These values can have changed through either:
 940	 *
 941	 *   sys_sched_set_scheduler() / sys_sched_setattr()
 942	 *
 943	 * or
 944	 *
 945	 *   DL CBS enforcement advancing the effective deadline.
 
 
 
 
 
 946	 */
 947	waiter_update_prio(waiter, task);
 
 948
 949	rt_mutex_enqueue(lock, waiter);
 950
 951	/*
 952	 * [8] Release the (blocking) task in preparation for
 953	 * taking the owner task in [10].
 954	 *
 955	 * Since we hold lock->waiter_lock, task cannot unblock, even if we
 956	 * release task->pi_lock.
 957	 */
 958	raw_spin_unlock(&task->pi_lock);
 959	put_task_struct(task);
 960
 961	/*
 962	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
 963	 *
 964	 * We must abort the chain walk if there is no lock owner even
 965	 * in the dead lock detection case, as we have nothing to
 966	 * follow here. This is the end of the chain we are walking.
 967	 */
 968	if (!rt_mutex_owner(lock)) {
 969		/*
 970		 * If the requeue [7] above changed the top waiter,
 971		 * then we need to wake the new top waiter up to try
 972		 * to get the lock.
 973		 */
 974		top_waiter = rt_mutex_top_waiter(lock);
 975		if (prerequeue_top_waiter != top_waiter)
 976			wake_up_state(top_waiter->task, top_waiter->wake_state);
 977		raw_spin_unlock_irq(&lock->wait_lock);
 978		return 0;
 979	}
 980
 981	/*
 982	 * [10] Grab the next task, i.e. the owner of @lock
 983	 *
 984	 * Per holding lock->wait_lock and checking for !owner above, there
 985	 * must be an owner and it cannot go away.
 986	 */
 987	task = get_task_struct(rt_mutex_owner(lock));
 988	raw_spin_lock(&task->pi_lock);
 989
 990	/* [11] requeue the pi waiters if necessary */
 991	if (waiter == rt_mutex_top_waiter(lock)) {
 992		/*
 993		 * The waiter became the new top (highest priority)
 994		 * waiter on the lock. Replace the previous top waiter
 995		 * in the owner tasks pi waiters tree with this waiter
 996		 * and adjust the priority of the owner.
 997		 */
 998		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 999		waiter_clone_prio(waiter, task);
1000		rt_mutex_enqueue_pi(task, waiter);
1001		rt_mutex_adjust_prio(lock, task);
1002
1003	} else if (prerequeue_top_waiter == waiter) {
1004		/*
1005		 * The waiter was the top waiter on the lock, but is
1006		 * no longer the top priority waiter. Replace waiter in
1007		 * the owner tasks pi waiters tree with the new top
1008		 * (highest priority) waiter and adjust the priority
1009		 * of the owner.
1010		 * The new top waiter is stored in @waiter so that
1011		 * @waiter == @top_waiter evaluates to true below and
1012		 * we continue to deboost the rest of the chain.
1013		 */
1014		rt_mutex_dequeue_pi(task, waiter);
1015		waiter = rt_mutex_top_waiter(lock);
1016		waiter_clone_prio(waiter, task);
1017		rt_mutex_enqueue_pi(task, waiter);
1018		rt_mutex_adjust_prio(lock, task);
1019	} else {
1020		/*
1021		 * Nothing changed. No need to do any priority
1022		 * adjustment.
1023		 */
1024	}
1025
1026	/*
1027	 * [12] check_exit_conditions_4() protected by task->pi_lock
1028	 * and lock->wait_lock. The actual decisions are made after we
1029	 * dropped the locks.
1030	 *
1031	 * Check whether the task which owns the current lock is pi
1032	 * blocked itself. If yes we store a pointer to the lock for
1033	 * the lock chain change detection above. After we dropped
1034	 * task->pi_lock next_lock cannot be dereferenced anymore.
1035	 */
1036	next_lock = task_blocked_on_lock(task);
1037	/*
1038	 * Store the top waiter of @lock for the end of chain walk
1039	 * decision below.
1040	 */
1041	top_waiter = rt_mutex_top_waiter(lock);
1042
1043	/* [13] Drop the locks */
1044	raw_spin_unlock(&task->pi_lock);
1045	raw_spin_unlock_irq(&lock->wait_lock);
1046
1047	/*
1048	 * Make the actual exit decisions [12], based on the stored
1049	 * values.
1050	 *
1051	 * We reached the end of the lock chain. Stop right here. No
1052	 * point to go back just to figure that out.
1053	 */
1054	if (!next_lock)
1055		goto out_put_task;
1056
1057	/*
1058	 * If the current waiter is not the top waiter on the lock,
1059	 * then we can stop the chain walk here if we are not in full
1060	 * deadlock detection mode.
1061	 */
1062	if (!detect_deadlock && waiter != top_waiter)
1063		goto out_put_task;
1064
1065	goto again;
1066
1067 out_unlock_pi:
1068	raw_spin_unlock_irq(&task->pi_lock);
1069 out_put_task:
1070	put_task_struct(task);
1071
1072	return ret;
1073}
1074
1075/*
1076 * Try to take an rt-mutex
1077 *
1078 * Must be called with lock->wait_lock held and interrupts disabled
1079 *
1080 * @lock:   The lock to be acquired.
1081 * @task:   The task which wants to acquire the lock
1082 * @waiter: The waiter that is queued to the lock's wait tree if the
1083 *	    callsite called task_blocked_on_lock(), otherwise NULL
1084 */
1085static int __sched
1086try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
1087		     struct rt_mutex_waiter *waiter)
1088{
1089	lockdep_assert_held(&lock->wait_lock);
1090
1091	/*
1092	 * Before testing whether we can acquire @lock, we set the
1093	 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
1094	 * other tasks which try to modify @lock into the slow path
1095	 * and they serialize on @lock->wait_lock.
1096	 *
1097	 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
1098	 * as explained at the top of this file if and only if:
1099	 *
1100	 * - There is a lock owner. The caller must fixup the
1101	 *   transient state if it does a trylock or leaves the lock
1102	 *   function due to a signal or timeout.
1103	 *
1104	 * - @task acquires the lock and there are no other
1105	 *   waiters. This is undone in rt_mutex_set_owner(@task) at
1106	 *   the end of this function.
1107	 */
1108	mark_rt_mutex_waiters(lock);
1109
1110	/*
1111	 * If @lock has an owner, give up.
1112	 */
1113	if (rt_mutex_owner(lock))
1114		return 0;
1115
1116	/*
1117	 * If @waiter != NULL, @task has already enqueued the waiter
1118	 * into @lock waiter tree. If @waiter == NULL then this is a
1119	 * trylock attempt.
1120	 */
1121	if (waiter) {
1122		struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
1123
1124		/*
1125		 * If waiter is the highest priority waiter of @lock,
1126		 * or allowed to steal it, take it over.
1127		 */
1128		if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1129			/*
1130			 * We can acquire the lock. Remove the waiter from the
1131			 * lock waiters tree.
1132			 */
1133			rt_mutex_dequeue(lock, waiter);
1134		} else {
1135			return 0;
1136		}
 
 
 
 
 
 
1137	} else {
1138		/*
1139		 * If the lock has waiters already we check whether @task is
1140		 * eligible to take over the lock.
1141		 *
1142		 * If there are no other waiters, @task can acquire
1143		 * the lock.  @task->pi_blocked_on is NULL, so it does
1144		 * not need to be dequeued.
1145		 */
1146		if (rt_mutex_has_waiters(lock)) {
1147			/* Check whether the trylock can steal it. */
1148			if (!rt_mutex_steal(task_to_waiter(task),
1149					    rt_mutex_top_waiter(lock)))
 
 
 
 
1150				return 0;
1151
1152			/*
1153			 * The current top waiter stays enqueued. We
1154			 * don't have to change anything in the lock
1155			 * waiters order.
1156			 */
1157		} else {
1158			/*
1159			 * No waiters. Take the lock without the
1160			 * pi_lock dance.@task->pi_blocked_on is NULL
1161			 * and we have no waiters to enqueue in @task
1162			 * pi waiters tree.
1163			 */
1164			goto takeit;
1165		}
1166	}
1167
1168	/*
1169	 * Clear @task->pi_blocked_on. Requires protection by
1170	 * @task->pi_lock. Redundant operation for the @waiter == NULL
1171	 * case, but conditionals are more expensive than a redundant
1172	 * store.
1173	 */
1174	raw_spin_lock(&task->pi_lock);
1175	task->pi_blocked_on = NULL;
1176	/*
1177	 * Finish the lock acquisition. @task is the new owner. If
1178	 * other waiters exist we have to insert the highest priority
1179	 * waiter into @task->pi_waiters tree.
1180	 */
1181	if (rt_mutex_has_waiters(lock))
1182		rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
1183	raw_spin_unlock(&task->pi_lock);
1184
1185takeit:
1186	/*
1187	 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1188	 * are still waiters or clears it.
1189	 */
1190	rt_mutex_set_owner(lock, task);
1191
1192	return 1;
1193}
1194
1195/*
1196 * Task blocks on lock.
1197 *
1198 * Prepare waiter and propagate pi chain
1199 *
1200 * This must be called with lock->wait_lock held and interrupts disabled
1201 */
1202static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
1203					   struct rt_mutex_waiter *waiter,
1204					   struct task_struct *task,
1205					   struct ww_acquire_ctx *ww_ctx,
1206					   enum rtmutex_chainwalk chwalk,
1207					   struct wake_q_head *wake_q)
1208{
1209	struct task_struct *owner = rt_mutex_owner(lock);
1210	struct rt_mutex_waiter *top_waiter = waiter;
1211	struct rt_mutex_base *next_lock;
1212	int chain_walk = 0, res;
1213
1214	lockdep_assert_held(&lock->wait_lock);
1215
1216	/*
1217	 * Early deadlock detection. We really don't want the task to
1218	 * enqueue on itself just to untangle the mess later. It's not
1219	 * only an optimization. We drop the locks, so another waiter
1220	 * can come in before the chain walk detects the deadlock. So
1221	 * the other will detect the deadlock and return -EDEADLOCK,
1222	 * which is wrong, as the other waiter is not in a deadlock
1223	 * situation.
1224	 *
1225	 * Except for ww_mutex, in that case the chain walk must already deal
1226	 * with spurious cycles, see the comments at [3] and [6].
1227	 */
1228	if (owner == task && !(build_ww_mutex() && ww_ctx))
1229		return -EDEADLK;
1230
1231	raw_spin_lock(&task->pi_lock);
1232	waiter->task = task;
1233	waiter->lock = lock;
1234	waiter_update_prio(waiter, task);
1235	waiter_clone_prio(waiter, task);
1236
1237	/* Get the top priority waiter on the lock */
1238	if (rt_mutex_has_waiters(lock))
1239		top_waiter = rt_mutex_top_waiter(lock);
1240	rt_mutex_enqueue(lock, waiter);
1241
1242	task->pi_blocked_on = waiter;
1243
1244	raw_spin_unlock(&task->pi_lock);
1245
1246	if (build_ww_mutex() && ww_ctx) {
1247		struct rt_mutex *rtm;
1248
1249		/* Check whether the waiter should back out immediately */
1250		rtm = container_of(lock, struct rt_mutex, rtmutex);
1251		res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx, wake_q);
1252		if (res) {
1253			raw_spin_lock(&task->pi_lock);
1254			rt_mutex_dequeue(lock, waiter);
1255			task->pi_blocked_on = NULL;
1256			raw_spin_unlock(&task->pi_lock);
1257			return res;
1258		}
1259	}
1260
1261	if (!owner)
1262		return 0;
1263
1264	raw_spin_lock(&owner->pi_lock);
1265	if (waiter == rt_mutex_top_waiter(lock)) {
1266		rt_mutex_dequeue_pi(owner, top_waiter);
1267		rt_mutex_enqueue_pi(owner, waiter);
1268
1269		rt_mutex_adjust_prio(lock, owner);
1270		if (owner->pi_blocked_on)
1271			chain_walk = 1;
1272	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1273		chain_walk = 1;
1274	}
1275
1276	/* Store the lock on which owner is blocked or NULL */
1277	next_lock = task_blocked_on_lock(owner);
1278
1279	raw_spin_unlock(&owner->pi_lock);
1280	/*
1281	 * Even if full deadlock detection is on, if the owner is not
1282	 * blocked itself, we can avoid finding this out in the chain
1283	 * walk.
1284	 */
1285	if (!chain_walk || !next_lock)
1286		return 0;
1287
1288	/*
1289	 * The owner can't disappear while holding a lock,
1290	 * so the owner struct is protected by wait_lock.
1291	 * Gets dropped in rt_mutex_adjust_prio_chain()!
1292	 */
1293	get_task_struct(owner);
1294
1295	preempt_disable();
1296	raw_spin_unlock_irq(&lock->wait_lock);
1297	/* wake up any tasks on the wake_q before calling rt_mutex_adjust_prio_chain */
1298	wake_up_q(wake_q);
1299	wake_q_init(wake_q);
1300	preempt_enable();
1301
1302
1303	res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1304					 next_lock, waiter, task);
1305
1306	raw_spin_lock_irq(&lock->wait_lock);
1307
1308	return res;
1309}
1310
1311/*
1312 * Remove the top waiter from the current tasks pi waiter tree and
1313 * queue it up.
1314 *
1315 * Called with lock->wait_lock held and interrupts disabled.
1316 */
1317static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1318					    struct rt_mutex_base *lock)
1319{
1320	struct rt_mutex_waiter *waiter;
1321
1322	lockdep_assert_held(&lock->wait_lock);
1323
1324	raw_spin_lock(&current->pi_lock);
1325
1326	waiter = rt_mutex_top_waiter(lock);
1327
1328	/*
1329	 * Remove it from current->pi_waiters and deboost.
1330	 *
1331	 * We must in fact deboost here in order to ensure we call
1332	 * rt_mutex_setprio() to update p->pi_top_task before the
1333	 * task unblocks.
1334	 */
1335	rt_mutex_dequeue_pi(current, waiter);
1336	rt_mutex_adjust_prio(lock, current);
1337
1338	/*
1339	 * As we are waking up the top waiter, and the waiter stays
1340	 * queued on the lock until it gets the lock, this lock
1341	 * obviously has waiters. Just set the bit here and this has
1342	 * the added benefit of forcing all new tasks into the
1343	 * slow path making sure no task of lower priority than
1344	 * the top waiter can steal this lock.
1345	 */
1346	lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1347
1348	/*
1349	 * We deboosted before waking the top waiter task such that we don't
1350	 * run two tasks with the 'same' priority (and ensure the
1351	 * p->pi_top_task pointer points to a blocked task). This however can
1352	 * lead to priority inversion if we would get preempted after the
1353	 * deboost but before waking our donor task, hence the preempt_disable()
1354	 * before unlock.
1355	 *
1356	 * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1357	 */
1358	preempt_disable();
1359	rt_mutex_wake_q_add(wqh, waiter);
1360	raw_spin_unlock(&current->pi_lock);
1361}
1362
1363static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1364{
1365	int ret = try_to_take_rt_mutex(lock, current, NULL);
1366
1367	/*
1368	 * try_to_take_rt_mutex() sets the lock waiters bit
1369	 * unconditionally. Clean this up.
1370	 */
1371	fixup_rt_mutex_waiters(lock, true);
1372
1373	return ret;
1374}
1375
1376/*
1377 * Slow path try-lock function:
1378 */
1379static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1380{
1381	unsigned long flags;
1382	int ret;
1383
1384	/*
1385	 * If the lock already has an owner we fail to get the lock.
1386	 * This can be done without taking the @lock->wait_lock as
1387	 * it is only being read, and this is a trylock anyway.
1388	 */
1389	if (rt_mutex_owner(lock))
1390		return 0;
1391
1392	/*
1393	 * The mutex has currently no owner. Lock the wait lock and try to
1394	 * acquire the lock. We use irqsave here to support early boot calls.
1395	 */
1396	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1397
1398	ret = __rt_mutex_slowtrylock(lock);
1399
1400	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1401
1402	return ret;
1403}
1404
1405static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
 
 
 
1406{
1407	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1408		return 1;
1409
1410	return rt_mutex_slowtrylock(lock);
 
1411}
1412
1413/*
1414 * Slow path to release a rt-mutex.
 
 
1415 */
1416static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1417{
1418	DEFINE_RT_WAKE_Q(wqh);
1419	unsigned long flags;
1420
1421	/* irqsave required to support early boot calls */
1422	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1423
1424	debug_rt_mutex_unlock(lock);
1425
1426	/*
1427	 * We must be careful here if the fast path is enabled. If we
1428	 * have no waiters queued we cannot set owner to NULL here
1429	 * because of:
1430	 *
1431	 * foo->lock->owner = NULL;
1432	 *			rtmutex_lock(foo->lock);   <- fast path
1433	 *			free = atomic_dec_and_test(foo->refcnt);
1434	 *			rtmutex_unlock(foo->lock); <- fast path
1435	 *			if (free)
1436	 *				kfree(foo);
1437	 * raw_spin_unlock(foo->lock->wait_lock);
1438	 *
1439	 * So for the fastpath enabled kernel:
1440	 *
1441	 * Nothing can set the waiters bit as long as we hold
1442	 * lock->wait_lock. So we do the following sequence:
1443	 *
1444	 *	owner = rt_mutex_owner(lock);
1445	 *	clear_rt_mutex_waiters(lock);
1446	 *	raw_spin_unlock(&lock->wait_lock);
1447	 *	if (cmpxchg(&lock->owner, owner, 0) == owner)
1448	 *		return;
1449	 *	goto retry;
1450	 *
1451	 * The fastpath disabled variant is simple as all access to
1452	 * lock->owner is serialized by lock->wait_lock:
1453	 *
1454	 *	lock->owner = NULL;
1455	 *	raw_spin_unlock(&lock->wait_lock);
1456	 */
1457	while (!rt_mutex_has_waiters(lock)) {
1458		/* Drops lock->wait_lock ! */
1459		if (unlock_rt_mutex_safe(lock, flags) == true)
1460			return;
1461		/* Relock the rtmutex and try again */
1462		raw_spin_lock_irqsave(&lock->wait_lock, flags);
1463	}
1464
1465	/*
1466	 * The wakeup next waiter path does not suffer from the above
1467	 * race. See the comments there.
1468	 *
1469	 * Queue the next waiter for wakeup once we release the wait_lock.
1470	 */
1471	mark_wakeup_next_waiter(&wqh, lock);
1472	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1473
1474	rt_mutex_wake_up_q(&wqh);
1475}
1476
1477static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
 
 
 
 
 
 
 
1478{
1479	if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1480		return;
1481
1482	rt_mutex_slowunlock(lock);
 
 
 
 
 
 
 
 
 
1483}
1484
1485#ifdef CONFIG_SMP
1486static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1487				  struct rt_mutex_waiter *waiter,
1488				  struct task_struct *owner)
 
 
 
 
1489{
1490	bool res = true;
1491
1492	rcu_read_lock();
1493	for (;;) {
1494		/* If owner changed, trylock again. */
1495		if (owner != rt_mutex_owner(lock))
1496			break;
1497		/*
1498		 * Ensure that @owner is dereferenced after checking that
1499		 * the lock owner still matches @owner. If that fails,
1500		 * @owner might point to freed memory. If it still matches,
1501		 * the rcu_read_lock() ensures the memory stays valid.
1502		 */
1503		barrier();
1504		/*
1505		 * Stop spinning when:
1506		 *  - the lock owner has been scheduled out
1507		 *  - current is not longer the top waiter
1508		 *  - current is requested to reschedule (redundant
1509		 *    for CONFIG_PREEMPT_RCU=y)
1510		 *  - the VCPU on which owner runs is preempted
1511		 */
1512		if (!owner_on_cpu(owner) || need_resched() ||
1513		    !rt_mutex_waiter_is_top_waiter(lock, waiter)) {
1514			res = false;
1515			break;
1516		}
1517		cpu_relax();
1518	}
1519	rcu_read_unlock();
1520	return res;
1521}
1522#else
1523static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1524				  struct rt_mutex_waiter *waiter,
1525				  struct task_struct *owner)
 
 
 
 
 
 
1526{
1527	return false;
1528}
 
1529#endif
1530
1531#ifdef RT_MUTEX_BUILD_MUTEX
1532/*
1533 * Functions required for:
1534 *	- rtmutex, futex on all kernels
1535 *	- mutex and rwsem substitutions on RT kernels
 
 
 
1536 */
 
 
 
 
 
1537
1538/*
1539 * Remove a waiter from a lock and give up
1540 *
1541 * Must be called with lock->wait_lock held and interrupts disabled. It must
1542 * have just failed to try_to_take_rt_mutex().
 
 
 
 
 
 
1543 */
1544static void __sched remove_waiter(struct rt_mutex_base *lock,
1545				  struct rt_mutex_waiter *waiter)
1546{
1547	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1548	struct task_struct *owner = rt_mutex_owner(lock);
1549	struct rt_mutex_base *next_lock;
1550
1551	lockdep_assert_held(&lock->wait_lock);
1552
1553	raw_spin_lock(&current->pi_lock);
1554	rt_mutex_dequeue(lock, waiter);
1555	current->pi_blocked_on = NULL;
1556	raw_spin_unlock(&current->pi_lock);
1557
1558	/*
1559	 * Only update priority if the waiter was the highest priority
1560	 * waiter of the lock and there is an owner to update.
1561	 */
1562	if (!owner || !is_top_waiter)
1563		return;
1564
1565	raw_spin_lock(&owner->pi_lock);
1566
1567	rt_mutex_dequeue_pi(owner, waiter);
1568
1569	if (rt_mutex_has_waiters(lock))
1570		rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1571
1572	rt_mutex_adjust_prio(lock, owner);
1573
1574	/* Store the lock on which owner is blocked or NULL */
1575	next_lock = task_blocked_on_lock(owner);
 
1576
1577	raw_spin_unlock(&owner->pi_lock);
 
 
1578
1579	/*
1580	 * Don't walk the chain, if the owner task is not blocked
1581	 * itself.
1582	 */
1583	if (!next_lock)
 
 
 
 
1584		return;
1585
1586	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1587	get_task_struct(owner);
1588
1589	raw_spin_unlock_irq(&lock->wait_lock);
1590
1591	rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1592				   next_lock, NULL, current);
 
 
 
 
 
1593
1594	raw_spin_lock_irq(&lock->wait_lock);
 
 
1595}
1596
1597/**
1598 * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
1599 * @lock:		 the rt_mutex to take
1600 * @ww_ctx:		 WW mutex context pointer
1601 * @state:		 the state the task should block in (TASK_INTERRUPTIBLE
1602 *			 or TASK_UNINTERRUPTIBLE)
1603 * @timeout:		 the pre-initialized and started timer, or NULL for none
1604 * @waiter:		 the pre-initialized rt_mutex_waiter
1605 * @wake_q:		 wake_q of tasks to wake when we drop the lock->wait_lock
1606 *
1607 * Must be called with lock->wait_lock held and interrupts disabled
 
1608 */
1609static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1610					   struct ww_acquire_ctx *ww_ctx,
1611					   unsigned int state,
1612					   struct hrtimer_sleeper *timeout,
1613					   struct rt_mutex_waiter *waiter,
1614					   struct wake_q_head *wake_q)
1615	__releases(&lock->wait_lock) __acquires(&lock->wait_lock)
1616{
1617	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1618	struct task_struct *owner;
1619	int ret = 0;
1620
1621	for (;;) {
1622		/* Try to acquire the lock: */
1623		if (try_to_take_rt_mutex(lock, current, waiter))
1624			break;
1625
1626		if (timeout && !timeout->task) {
1627			ret = -ETIMEDOUT;
1628			break;
1629		}
1630		if (signal_pending_state(state, current)) {
1631			ret = -EINTR;
1632			break;
1633		}
1634
1635		if (build_ww_mutex() && ww_ctx) {
1636			ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1637			if (ret)
1638				break;
1639		}
 
 
1640
1641		if (waiter == rt_mutex_top_waiter(lock))
1642			owner = rt_mutex_owner(lock);
1643		else
1644			owner = NULL;
1645		preempt_disable();
1646		raw_spin_unlock_irq(&lock->wait_lock);
1647		if (wake_q) {
1648			wake_up_q(wake_q);
1649			wake_q_init(wake_q);
1650		}
1651		preempt_enable();
1652
1653		if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1654			rt_mutex_schedule();
 
 
 
1655
1656		raw_spin_lock_irq(&lock->wait_lock);
1657		set_current_state(state);
1658	}
1659
1660	__set_current_state(TASK_RUNNING);
1661	return ret;
1662}
1663
1664static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1665					     struct rt_mutex_base *lock,
1666					     struct rt_mutex_waiter *w)
 
 
 
 
 
 
 
 
 
 
1667{
1668	/*
1669	 * If the result is not -EDEADLOCK or the caller requested
1670	 * deadlock detection, nothing to do here.
1671	 */
1672	if (res != -EDEADLOCK || detect_deadlock)
1673		return;
1674
1675	if (build_ww_mutex() && w->ww_ctx)
1676		return;
1677
1678	raw_spin_unlock_irq(&lock->wait_lock);
 
 
1679
1680	WARN(1, "rtmutex deadlock detected\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1681
1682	while (1) {
1683		set_current_state(TASK_INTERRUPTIBLE);
1684		rt_mutex_schedule();
1685	}
 
 
 
 
 
 
 
 
 
 
 
 
1686}
1687
1688/**
1689 * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
1690 * @lock:	The rtmutex to block lock
1691 * @ww_ctx:	WW mutex context pointer
1692 * @state:	The task state for sleeping
1693 * @chwalk:	Indicator whether full or partial chainwalk is requested
1694 * @waiter:	Initializer waiter for blocking
1695 * @wake_q:	The wake_q to wake tasks after we release the wait_lock
1696 */
1697static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1698				       struct ww_acquire_ctx *ww_ctx,
1699				       unsigned int state,
1700				       enum rtmutex_chainwalk chwalk,
1701				       struct rt_mutex_waiter *waiter,
1702				       struct wake_q_head *wake_q)
 
 
 
 
 
 
 
1703{
1704	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1705	struct ww_mutex *ww = ww_container_of(rtm);
1706	int ret;
1707
1708	lockdep_assert_held(&lock->wait_lock);
1709
1710	/* Try to acquire the lock again: */
1711	if (try_to_take_rt_mutex(lock, current, NULL)) {
1712		if (build_ww_mutex() && ww_ctx) {
1713			__ww_mutex_check_waiters(rtm, ww_ctx, wake_q);
1714			ww_mutex_lock_acquired(ww, ww_ctx);
1715		}
1716		return 0;
1717	}
1718
1719	set_current_state(state);
1720
1721	trace_contention_begin(lock, LCB_F_RT);
1722
1723	ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk, wake_q);
1724	if (likely(!ret))
1725		ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter, wake_q);
1726
1727	if (likely(!ret)) {
1728		/* acquired the lock */
1729		if (build_ww_mutex() && ww_ctx) {
1730			if (!ww_ctx->is_wait_die)
1731				__ww_mutex_check_waiters(rtm, ww_ctx, wake_q);
1732			ww_mutex_lock_acquired(ww, ww_ctx);
1733		}
1734	} else {
1735		__set_current_state(TASK_RUNNING);
1736		remove_waiter(lock, waiter);
1737		rt_mutex_handle_deadlock(ret, chwalk, lock, waiter);
1738	}
1739
1740	/*
1741	 * try_to_take_rt_mutex() sets the waiter bit
1742	 * unconditionally. We might have to fix that up.
1743	 */
1744	fixup_rt_mutex_waiters(lock, true);
1745
1746	trace_contention_end(lock, ret);
1747
1748	return ret;
1749}
1750
1751static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1752					     struct ww_acquire_ctx *ww_ctx,
1753					     unsigned int state,
1754					     struct wake_q_head *wake_q)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1755{
1756	struct rt_mutex_waiter waiter;
1757	int ret;
1758
1759	rt_mutex_init_waiter(&waiter);
1760	waiter.ww_ctx = ww_ctx;
1761
1762	ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1763				  &waiter, wake_q);
1764
1765	debug_rt_mutex_free_waiter(&waiter);
1766	return ret;
1767}
1768
1769/*
1770 * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
1771 * @lock:	The rtmutex to block lock
1772 * @ww_ctx:	WW mutex context pointer
1773 * @state:	The task state for sleeping
1774 */
1775static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1776				     struct ww_acquire_ctx *ww_ctx,
1777				     unsigned int state)
 
 
 
 
 
 
 
 
 
 
 
1778{
1779	DEFINE_WAKE_Q(wake_q);
1780	unsigned long flags;
1781	int ret;
1782
 
 
 
 
1783	/*
1784	 * Do all pre-schedule work here, before we queue a waiter and invoke
1785	 * PI -- any such work that trips on rtlock (PREEMPT_RT spinlock) would
1786	 * otherwise recurse back into task_blocks_on_rt_mutex() through
1787	 * rtlock_slowlock() and will then enqueue a second waiter for this
1788	 * same task and things get really confusing real fast.
1789	 */
1790	rt_mutex_pre_schedule();
1791
1792	/*
1793	 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1794	 * be called in early boot if the cmpxchg() fast path is disabled
1795	 * (debug, no architecture support). In this case we will acquire the
1796	 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1797	 * enable interrupts in that early boot case. So we need to use the
1798	 * irqsave/restore variants.
1799	 */
1800	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1801	ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state, &wake_q);
1802	preempt_disable();
1803	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1804	wake_up_q(&wake_q);
1805	preempt_enable();
1806	rt_mutex_post_schedule();
1807
1808	return ret;
1809}
1810
1811static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1812					   unsigned int state)
1813{
1814	lockdep_assert(!current->pi_blocked_on);
1815
1816	if (likely(rt_mutex_try_acquire(lock)))
1817		return 0;
1818
1819	return rt_mutex_slowlock(lock, NULL, state);
1820}
1821#endif /* RT_MUTEX_BUILD_MUTEX */
1822
1823#ifdef RT_MUTEX_BUILD_SPINLOCKS
1824/*
1825 * Functions required for spin/rw_lock substitution on RT kernels
1826 */
1827
1828/**
1829 * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1830 * @lock:	The underlying RT mutex
1831 * @wake_q:	The wake_q to wake tasks after we release the wait_lock
1832 */
1833static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock,
1834					   struct wake_q_head *wake_q)
1835	__releases(&lock->wait_lock) __acquires(&lock->wait_lock)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1836{
1837	struct rt_mutex_waiter waiter;
1838	struct task_struct *owner;
1839
1840	lockdep_assert_held(&lock->wait_lock);
1841
1842	if (try_to_take_rt_mutex(lock, current, NULL))
1843		return;
1844
1845	rt_mutex_init_rtlock_waiter(&waiter);
1846
1847	/* Save current state and set state to TASK_RTLOCK_WAIT */
1848	current_save_and_set_rtlock_wait_state();
1849
1850	trace_contention_begin(lock, LCB_F_RT);
1851
1852	task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK, wake_q);
1853
1854	for (;;) {
1855		/* Try to acquire the lock again */
1856		if (try_to_take_rt_mutex(lock, current, &waiter))
1857			break;
1858
1859		if (&waiter == rt_mutex_top_waiter(lock))
1860			owner = rt_mutex_owner(lock);
1861		else
1862			owner = NULL;
1863		preempt_disable();
1864		raw_spin_unlock_irq(&lock->wait_lock);
1865		wake_up_q(wake_q);
1866		wake_q_init(wake_q);
1867		preempt_enable();
1868
1869		if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1870			schedule_rtlock();
1871
1872		raw_spin_lock_irq(&lock->wait_lock);
1873		set_current_state(TASK_RTLOCK_WAIT);
1874	}
1875
1876	/* Restore the task state */
1877	current_restore_rtlock_saved_state();
1878
1879	/*
1880	 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1881	 * We might have to fix that up:
1882	 */
1883	fixup_rt_mutex_waiters(lock, true);
1884	debug_rt_mutex_free_waiter(&waiter);
 
1885
1886	trace_contention_end(lock, 0);
1887}
1888
1889static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
 
1890{
1891	unsigned long flags;
1892	DEFINE_WAKE_Q(wake_q);
1893
1894	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1895	rtlock_slowlock_locked(lock, &wake_q);
1896	preempt_disable();
1897	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1898	wake_up_q(&wake_q);
1899	preempt_enable();
1900}
1901
1902#endif /* RT_MUTEX_BUILD_SPINLOCKS */
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   4 *
   5 * started by Ingo Molnar and Thomas Gleixner.
   6 *
   7 *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   8 *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   9 *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
  10 *  Copyright (C) 2006 Esben Nielsen
 
 
 
 
 
  11 *
  12 *  See Documentation/locking/rt-mutex-design.rst for details.
  13 */
  14#include <linux/spinlock.h>
  15#include <linux/export.h>
 
  16#include <linux/sched/signal.h>
  17#include <linux/sched/rt.h>
  18#include <linux/sched/deadline.h>
  19#include <linux/sched/wake_q.h>
  20#include <linux/sched/debug.h>
  21#include <linux/timer.h>
 
  22
  23#include "rtmutex_common.h"
  24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  25/*
  26 * lock->owner state tracking:
  27 *
  28 * lock->owner holds the task_struct pointer of the owner. Bit 0
  29 * is used to keep track of the "lock has waiters" state.
  30 *
  31 * owner	bit0
  32 * NULL		0	lock is free (fast acquire possible)
  33 * NULL		1	lock is free and has waiters and the top waiter
  34 *				is going to take the lock*
  35 * taskpointer	0	lock is held (fast release possible)
  36 * taskpointer	1	lock is held and has waiters**
  37 *
  38 * The fast atomic compare exchange based acquire and release is only
  39 * possible when bit 0 of lock->owner is 0.
  40 *
  41 * (*) It also can be a transitional state when grabbing the lock
  42 * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  43 * we need to set the bit0 before looking at the lock, and the owner may be
  44 * NULL in this small time, hence this can be a transitional state.
  45 *
  46 * (**) There is a small time when bit 0 is set but there are no
  47 * waiters. This can happen when grabbing the lock in the slow path.
  48 * To prevent a cmpxchg of the owner releasing the lock, we need to
  49 * set this bit before looking at the lock.
  50 */
  51
  52static __always_inline void
  53rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  54{
  55	unsigned long val = (unsigned long)owner;
  56
  57	if (rt_mutex_has_waiters(lock))
  58		val |= RT_MUTEX_HAS_WAITERS;
  59
  60	WRITE_ONCE(lock->owner, (struct task_struct *)val);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  61}
  62
  63static __always_inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  64{
  65	lock->owner = (struct task_struct *)
  66			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  67}
  68
  69static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex *lock)
 
  70{
  71	unsigned long owner, *p = (unsigned long *) &lock->owner;
  72
  73	if (rt_mutex_has_waiters(lock))
  74		return;
  75
  76	/*
  77	 * The rbtree has no waiters enqueued, now make sure that the
  78	 * lock->owner still has the waiters bit set, otherwise the
  79	 * following can happen:
  80	 *
  81	 * CPU 0	CPU 1		CPU2
  82	 * l->owner=T1
  83	 *		rt_mutex_lock(l)
  84	 *		lock(l->lock)
  85	 *		l->owner = T1 | HAS_WAITERS;
  86	 *		enqueue(T2)
  87	 *		boost()
  88	 *		  unlock(l->lock)
  89	 *		block()
  90	 *
  91	 *				rt_mutex_lock(l)
  92	 *				lock(l->lock)
  93	 *				l->owner = T1 | HAS_WAITERS;
  94	 *				enqueue(T3)
  95	 *				boost()
  96	 *				  unlock(l->lock)
  97	 *				block()
  98	 *		signal(->T2)	signal(->T3)
  99	 *		lock(l->lock)
 100	 *		dequeue(T2)
 101	 *		deboost()
 102	 *		  unlock(l->lock)
 103	 *				lock(l->lock)
 104	 *				dequeue(T3)
 105	 *				 ==> wait list is empty
 106	 *				deboost()
 107	 *				 unlock(l->lock)
 108	 *		lock(l->lock)
 109	 *		fixup_rt_mutex_waiters()
 110	 *		  if (wait_list_empty(l) {
 111	 *		    l->owner = owner
 112	 *		    owner = l->owner & ~HAS_WAITERS;
 113	 *		      ==> l->owner = T1
 114	 *		  }
 115	 *				lock(l->lock)
 116	 * rt_mutex_unlock(l)		fixup_rt_mutex_waiters()
 117	 *				  if (wait_list_empty(l) {
 118	 *				    owner = l->owner & ~HAS_WAITERS;
 119	 * cmpxchg(l->owner, T1, NULL)
 120	 *  ===> Success (l->owner = NULL)
 121	 *
 122	 *				    l->owner = owner
 123	 *				      ==> l->owner = T1
 124	 *				  }
 125	 *
 126	 * With the check for the waiter bit in place T3 on CPU2 will not
 127	 * overwrite. All tasks fiddling with the waiters bit are
 128	 * serialized by l->lock, so nothing else can modify the waiters
 129	 * bit. If the bit is set then nothing can change l->owner either
 130	 * so the simple RMW is safe. The cmpxchg() will simply fail if it
 131	 * happens in the middle of the RMW because the waiters bit is
 132	 * still set.
 133	 */
 134	owner = READ_ONCE(*p);
 135	if (owner & RT_MUTEX_HAS_WAITERS)
 136		WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
 
 
 
 
 
 
 
 
 
 
 
 
 
 137}
 138
 139/*
 140 * We can speed up the acquire/release, if there's no debugging state to be
 141 * set up.
 142 */
 143#ifndef CONFIG_DEBUG_RT_MUTEXES
 144# define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
 145# define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 146
 147/*
 148 * Callers must hold the ->wait_lock -- which is the whole purpose as we force
 149 * all future threads that attempt to [Rmw] the lock to the slowpath. As such
 150 * relaxed semantics suffice.
 151 */
 152static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 153{
 154	unsigned long owner, *p = (unsigned long *) &lock->owner;
 
 155
 
 156	do {
 157		owner = *p;
 158	} while (cmpxchg_relaxed(p, owner,
 159				 owner | RT_MUTEX_HAS_WAITERS) != owner);
 
 
 
 
 
 
 160}
 161
 162/*
 163 * Safe fastpath aware unlock:
 164 * 1) Clear the waiters bit
 165 * 2) Drop lock->wait_lock
 166 * 3) Try to unlock the lock with cmpxchg
 167 */
 168static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 169						 unsigned long flags)
 170	__releases(lock->wait_lock)
 171{
 172	struct task_struct *owner = rt_mutex_owner(lock);
 173
 174	clear_rt_mutex_waiters(lock);
 175	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 176	/*
 177	 * If a new waiter comes in between the unlock and the cmpxchg
 178	 * we have two situations:
 179	 *
 180	 * unlock(wait_lock);
 181	 *					lock(wait_lock);
 182	 * cmpxchg(p, owner, 0) == owner
 183	 *					mark_rt_mutex_waiters(lock);
 184	 *					acquire(lock);
 185	 * or:
 186	 *
 187	 * unlock(wait_lock);
 188	 *					lock(wait_lock);
 189	 *					mark_rt_mutex_waiters(lock);
 190	 *
 191	 * cmpxchg(p, owner, 0) != owner
 192	 *					enqueue_waiter();
 193	 *					unlock(wait_lock);
 194	 * lock(wait_lock);
 195	 * wake waiter();
 196	 * unlock(wait_lock);
 197	 *					lock(wait_lock);
 198	 *					acquire(lock);
 199	 */
 200	return rt_mutex_cmpxchg_release(lock, owner, NULL);
 201}
 202
 203#else
 204# define rt_mutex_cmpxchg_acquire(l,c,n)	(0)
 205# define rt_mutex_cmpxchg_release(l,c,n)	(0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 206
 207static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 208{
 209	lock->owner = (struct task_struct *)
 210			((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 211}
 212
 213/*
 214 * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 215 */
 216static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 217						 unsigned long flags)
 218	__releases(lock->wait_lock)
 219{
 220	lock->owner = NULL;
 221	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 222	return true;
 223}
 224#endif
 225
 
 
 
 
 
 
 
 
 
 
 226/*
 227 * Only use with rt_mutex_waiter_{less,equal}()
 228 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 229#define task_to_waiter(p)	\
 230	&(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
 231
 232static __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 233						struct rt_mutex_waiter *right)
 234{
 235	if (left->prio < right->prio)
 236		return 1;
 237
 238	/*
 239	 * If both waiters have dl_prio(), we check the deadlines of the
 240	 * associated tasks.
 241	 * If left waiter has a dl_prio(), and we didn't return 1 above,
 242	 * then right waiter has a dl_prio() too.
 243	 */
 244	if (dl_prio(left->prio))
 245		return dl_time_before(left->deadline, right->deadline);
 246
 247	return 0;
 248}
 249
 250static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
 251						 struct rt_mutex_waiter *right)
 252{
 253	if (left->prio != right->prio)
 254		return 0;
 255
 256	/*
 257	 * If both waiters have dl_prio(), we check the deadlines of the
 258	 * associated tasks.
 259	 * If left waiter has a dl_prio(), and we didn't return 0 above,
 260	 * then right waiter has a dl_prio() too.
 261	 */
 262	if (dl_prio(left->prio))
 263		return left->deadline == right->deadline;
 264
 265	return 1;
 266}
 267
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 268#define __node_2_waiter(node) \
 269	rb_entry((node), struct rt_mutex_waiter, tree_entry)
 270
 271static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
 272{
 273	return rt_mutex_waiter_less(__node_2_waiter(a), __node_2_waiter(b));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 274}
 275
 276static __always_inline void
 277rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 278{
 279	rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
 
 
 280}
 281
 282static __always_inline void
 283rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 284{
 285	if (RB_EMPTY_NODE(&waiter->tree_entry))
 
 
 286		return;
 287
 288	rb_erase_cached(&waiter->tree_entry, &lock->waiters);
 289	RB_CLEAR_NODE(&waiter->tree_entry);
 290}
 291
 292#define __node_2_pi_waiter(node) \
 293	rb_entry((node), struct rt_mutex_waiter, pi_tree_entry)
 294
 295static __always_inline bool
 296__pi_waiter_less(struct rb_node *a, const struct rb_node *b)
 297{
 298	return rt_mutex_waiter_less(__node_2_pi_waiter(a), __node_2_pi_waiter(b));
 299}
 300
 301static __always_inline void
 302rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 303{
 304	rb_add_cached(&waiter->pi_tree_entry, &task->pi_waiters, __pi_waiter_less);
 
 
 305}
 306
 307static __always_inline void
 308rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 309{
 310	if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 
 
 311		return;
 312
 313	rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
 314	RB_CLEAR_NODE(&waiter->pi_tree_entry);
 315}
 316
 317static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
 
 318{
 319	struct task_struct *pi_task = NULL;
 320
 
 
 321	lockdep_assert_held(&p->pi_lock);
 322
 323	if (task_has_pi_waiters(p))
 324		pi_task = task_top_pi_waiter(p)->task;
 325
 326	rt_mutex_setprio(p, pi_task);
 327}
 328
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 329/*
 330 * Deadlock detection is conditional:
 331 *
 332 * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 333 * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 334 *
 335 * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 336 * conducted independent of the detect argument.
 337 *
 338 * If the waiter argument is NULL this indicates the deboost path and
 339 * deadlock detection is disabled independent of the detect argument
 340 * and the config settings.
 341 */
 342static __always_inline bool
 343rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 344			      enum rtmutex_chainwalk chwalk)
 345{
 346	if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
 347		return waiter != NULL;
 348	return chwalk == RT_MUTEX_FULL_CHAINWALK;
 349}
 350
 351/*
 352 * Max number of times we'll walk the boosting chain:
 353 */
 354int max_lock_depth = 1024;
 355
 356static __always_inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 357{
 358	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 359}
 360
 361/*
 362 * Adjust the priority chain. Also used for deadlock detection.
 363 * Decreases task's usage by one - may thus free the task.
 364 *
 365 * @task:	the task owning the mutex (owner) for which a chain walk is
 366 *		probably needed
 367 * @chwalk:	do we have to carry out deadlock detection?
 368 * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
 369 *		things for a task that has just got its priority adjusted, and
 370 *		is waiting on a mutex)
 371 * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
 372 *		we dropped its pi_lock. Is never dereferenced, only used for
 373 *		comparison to detect lock chain changes.
 374 * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 375 *		its priority to the mutex owner (can be NULL in the case
 376 *		depicted above or if the top waiter is gone away and we are
 377 *		actually deboosting the owner)
 378 * @top_task:	the current top waiter
 379 *
 380 * Returns 0 or -EDEADLK.
 381 *
 382 * Chain walk basics and protection scope
 383 *
 384 * [R] refcount on task
 385 * [P] task->pi_lock held
 386 * [L] rtmutex->wait_lock held
 387 *
 
 
 
 
 
 388 * Step	Description				Protected by
 389 *	function arguments:
 390 *	@task					[R]
 391 *	@orig_lock if != NULL			@top_task is blocked on it
 392 *	@next_lock				Unprotected. Cannot be
 393 *						dereferenced. Only used for
 394 *						comparison.
 395 *	@orig_waiter if != NULL			@top_task is blocked on it
 396 *	@top_task				current, or in case of proxy
 397 *						locking protected by calling
 398 *						code
 399 *	again:
 400 *	  loop_sanity_check();
 401 *	retry:
 402 * [1]	  lock(task->pi_lock);			[R] acquire [P]
 403 * [2]	  waiter = task->pi_blocked_on;		[P]
 404 * [3]	  check_exit_conditions_1();		[P]
 405 * [4]	  lock = waiter->lock;			[P]
 406 * [5]	  if (!try_lock(lock->wait_lock)) {	[P] try to acquire [L]
 407 *	    unlock(task->pi_lock);		release [P]
 408 *	    goto retry;
 409 *	  }
 410 * [6]	  check_exit_conditions_2();		[P] + [L]
 411 * [7]	  requeue_lock_waiter(lock, waiter);	[P] + [L]
 412 * [8]	  unlock(task->pi_lock);		release [P]
 413 *	  put_task_struct(task);		release [R]
 414 * [9]	  check_exit_conditions_3();		[L]
 415 * [10]	  task = owner(lock);			[L]
 416 *	  get_task_struct(task);		[L] acquire [R]
 417 *	  lock(task->pi_lock);			[L] acquire [P]
 418 * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 419 * [12]	  check_exit_conditions_4();		[P] + [L]
 420 * [13]	  unlock(task->pi_lock);		release [P]
 421 *	  unlock(lock->wait_lock);		release [L]
 422 *	  goto again;
 
 
 
 
 
 423 */
 424static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
 425					      enum rtmutex_chainwalk chwalk,
 426					      struct rt_mutex *orig_lock,
 427					      struct rt_mutex *next_lock,
 428					      struct rt_mutex_waiter *orig_waiter,
 429					      struct task_struct *top_task)
 430{
 431	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 432	struct rt_mutex_waiter *prerequeue_top_waiter;
 433	int ret = 0, depth = 0;
 434	struct rt_mutex *lock;
 435	bool detect_deadlock;
 436	bool requeue = true;
 437
 438	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 439
 440	/*
 441	 * The (de)boosting is a step by step approach with a lot of
 442	 * pitfalls. We want this to be preemptible and we want hold a
 443	 * maximum of two locks per step. So we have to check
 444	 * carefully whether things change under us.
 445	 */
 446 again:
 447	/*
 448	 * We limit the lock chain length for each invocation.
 449	 */
 450	if (++depth > max_lock_depth) {
 451		static int prev_max;
 452
 453		/*
 454		 * Print this only once. If the admin changes the limit,
 455		 * print a new message when reaching the limit again.
 456		 */
 457		if (prev_max != max_lock_depth) {
 458			prev_max = max_lock_depth;
 459			printk(KERN_WARNING "Maximum lock depth %d reached "
 460			       "task: %s (%d)\n", max_lock_depth,
 461			       top_task->comm, task_pid_nr(top_task));
 462		}
 463		put_task_struct(task);
 464
 465		return -EDEADLK;
 466	}
 467
 468	/*
 469	 * We are fully preemptible here and only hold the refcount on
 470	 * @task. So everything can have changed under us since the
 471	 * caller or our own code below (goto retry/again) dropped all
 472	 * locks.
 473	 */
 474 retry:
 475	/*
 476	 * [1] Task cannot go away as we did a get_task() before !
 477	 */
 478	raw_spin_lock_irq(&task->pi_lock);
 479
 480	/*
 481	 * [2] Get the waiter on which @task is blocked on.
 482	 */
 483	waiter = task->pi_blocked_on;
 484
 485	/*
 486	 * [3] check_exit_conditions_1() protected by task->pi_lock.
 487	 */
 488
 489	/*
 490	 * Check whether the end of the boosting chain has been
 491	 * reached or the state of the chain has changed while we
 492	 * dropped the locks.
 493	 */
 494	if (!waiter)
 495		goto out_unlock_pi;
 496
 497	/*
 498	 * Check the orig_waiter state. After we dropped the locks,
 499	 * the previous owner of the lock might have released the lock.
 500	 */
 501	if (orig_waiter && !rt_mutex_owner(orig_lock))
 502		goto out_unlock_pi;
 503
 504	/*
 505	 * We dropped all locks after taking a refcount on @task, so
 506	 * the task might have moved on in the lock chain or even left
 507	 * the chain completely and blocks now on an unrelated lock or
 508	 * on @orig_lock.
 509	 *
 510	 * We stored the lock on which @task was blocked in @next_lock,
 511	 * so we can detect the chain change.
 512	 */
 513	if (next_lock != waiter->lock)
 514		goto out_unlock_pi;
 515
 516	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 517	 * Drop out, when the task has no waiters. Note,
 518	 * top_waiter can be NULL, when we are in the deboosting
 519	 * mode!
 520	 */
 521	if (top_waiter) {
 522		if (!task_has_pi_waiters(task))
 523			goto out_unlock_pi;
 524		/*
 525		 * If deadlock detection is off, we stop here if we
 526		 * are not the top pi waiter of the task. If deadlock
 527		 * detection is enabled we continue, but stop the
 528		 * requeueing in the chain walk.
 529		 */
 530		if (top_waiter != task_top_pi_waiter(task)) {
 531			if (!detect_deadlock)
 532				goto out_unlock_pi;
 533			else
 534				requeue = false;
 535		}
 536	}
 537
 538	/*
 539	 * If the waiter priority is the same as the task priority
 540	 * then there is no further priority adjustment necessary.  If
 541	 * deadlock detection is off, we stop the chain walk. If its
 542	 * enabled we continue, but stop the requeueing in the chain
 543	 * walk.
 544	 */
 545	if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
 546		if (!detect_deadlock)
 547			goto out_unlock_pi;
 548		else
 549			requeue = false;
 550	}
 551
 552	/*
 553	 * [4] Get the next lock
 
 554	 */
 555	lock = waiter->lock;
 556	/*
 557	 * [5] We need to trylock here as we are holding task->pi_lock,
 558	 * which is the reverse lock order versus the other rtmutex
 559	 * operations.
 
 
 
 
 560	 */
 561	if (!raw_spin_trylock(&lock->wait_lock)) {
 562		raw_spin_unlock_irq(&task->pi_lock);
 563		cpu_relax();
 564		goto retry;
 565	}
 566
 567	/*
 568	 * [6] check_exit_conditions_2() protected by task->pi_lock and
 569	 * lock->wait_lock.
 570	 *
 571	 * Deadlock detection. If the lock is the same as the original
 572	 * lock which caused us to walk the lock chain or if the
 573	 * current lock is owned by the task which initiated the chain
 574	 * walk, we detected a deadlock.
 575	 */
 576	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 577		raw_spin_unlock(&lock->wait_lock);
 578		ret = -EDEADLK;
 579		goto out_unlock_pi;
 580	}
 581
 582	/*
 583	 * If we just follow the lock chain for deadlock detection, no
 584	 * need to do all the requeue operations. To avoid a truckload
 585	 * of conditionals around the various places below, just do the
 586	 * minimum chain walk checks.
 587	 */
 588	if (!requeue) {
 589		/*
 590		 * No requeue[7] here. Just release @task [8]
 591		 */
 592		raw_spin_unlock(&task->pi_lock);
 593		put_task_struct(task);
 594
 595		/*
 596		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
 597		 * If there is no owner of the lock, end of chain.
 598		 */
 599		if (!rt_mutex_owner(lock)) {
 600			raw_spin_unlock_irq(&lock->wait_lock);
 601			return 0;
 602		}
 603
 604		/* [10] Grab the next task, i.e. owner of @lock */
 605		task = get_task_struct(rt_mutex_owner(lock));
 606		raw_spin_lock(&task->pi_lock);
 607
 608		/*
 609		 * No requeue [11] here. We just do deadlock detection.
 610		 *
 611		 * [12] Store whether owner is blocked
 612		 * itself. Decision is made after dropping the locks
 613		 */
 614		next_lock = task_blocked_on_lock(task);
 615		/*
 616		 * Get the top waiter for the next iteration
 617		 */
 618		top_waiter = rt_mutex_top_waiter(lock);
 619
 620		/* [13] Drop locks */
 621		raw_spin_unlock(&task->pi_lock);
 622		raw_spin_unlock_irq(&lock->wait_lock);
 623
 624		/* If owner is not blocked, end of chain. */
 625		if (!next_lock)
 626			goto out_put_task;
 627		goto again;
 628	}
 629
 630	/*
 631	 * Store the current top waiter before doing the requeue
 632	 * operation on @lock. We need it for the boost/deboost
 633	 * decision below.
 634	 */
 635	prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 636
 637	/* [7] Requeue the waiter in the lock waiter tree. */
 638	rt_mutex_dequeue(lock, waiter);
 639
 640	/*
 641	 * Update the waiter prio fields now that we're dequeued.
 642	 *
 643	 * These values can have changed through either:
 644	 *
 645	 *   sys_sched_set_scheduler() / sys_sched_setattr()
 646	 *
 647	 * or
 648	 *
 649	 *   DL CBS enforcement advancing the effective deadline.
 650	 *
 651	 * Even though pi_waiters also uses these fields, and that tree is only
 652	 * updated in [11], we can do this here, since we hold [L], which
 653	 * serializes all pi_waiters access and rb_erase() does not care about
 654	 * the values of the node being removed.
 655	 */
 656	waiter->prio = task->prio;
 657	waiter->deadline = task->dl.deadline;
 658
 659	rt_mutex_enqueue(lock, waiter);
 660
 661	/* [8] Release the task */
 
 
 
 
 
 
 662	raw_spin_unlock(&task->pi_lock);
 663	put_task_struct(task);
 664
 665	/*
 666	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
 667	 *
 668	 * We must abort the chain walk if there is no lock owner even
 669	 * in the dead lock detection case, as we have nothing to
 670	 * follow here. This is the end of the chain we are walking.
 671	 */
 672	if (!rt_mutex_owner(lock)) {
 673		/*
 674		 * If the requeue [7] above changed the top waiter,
 675		 * then we need to wake the new top waiter up to try
 676		 * to get the lock.
 677		 */
 678		if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 679			wake_up_process(rt_mutex_top_waiter(lock)->task);
 
 680		raw_spin_unlock_irq(&lock->wait_lock);
 681		return 0;
 682	}
 683
 684	/* [10] Grab the next task, i.e. the owner of @lock */
 
 
 
 
 
 685	task = get_task_struct(rt_mutex_owner(lock));
 686	raw_spin_lock(&task->pi_lock);
 687
 688	/* [11] requeue the pi waiters if necessary */
 689	if (waiter == rt_mutex_top_waiter(lock)) {
 690		/*
 691		 * The waiter became the new top (highest priority)
 692		 * waiter on the lock. Replace the previous top waiter
 693		 * in the owner tasks pi waiters tree with this waiter
 694		 * and adjust the priority of the owner.
 695		 */
 696		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 
 697		rt_mutex_enqueue_pi(task, waiter);
 698		rt_mutex_adjust_prio(task);
 699
 700	} else if (prerequeue_top_waiter == waiter) {
 701		/*
 702		 * The waiter was the top waiter on the lock, but is
 703		 * no longer the top priority waiter. Replace waiter in
 704		 * the owner tasks pi waiters tree with the new top
 705		 * (highest priority) waiter and adjust the priority
 706		 * of the owner.
 707		 * The new top waiter is stored in @waiter so that
 708		 * @waiter == @top_waiter evaluates to true below and
 709		 * we continue to deboost the rest of the chain.
 710		 */
 711		rt_mutex_dequeue_pi(task, waiter);
 712		waiter = rt_mutex_top_waiter(lock);
 
 713		rt_mutex_enqueue_pi(task, waiter);
 714		rt_mutex_adjust_prio(task);
 715	} else {
 716		/*
 717		 * Nothing changed. No need to do any priority
 718		 * adjustment.
 719		 */
 720	}
 721
 722	/*
 723	 * [12] check_exit_conditions_4() protected by task->pi_lock
 724	 * and lock->wait_lock. The actual decisions are made after we
 725	 * dropped the locks.
 726	 *
 727	 * Check whether the task which owns the current lock is pi
 728	 * blocked itself. If yes we store a pointer to the lock for
 729	 * the lock chain change detection above. After we dropped
 730	 * task->pi_lock next_lock cannot be dereferenced anymore.
 731	 */
 732	next_lock = task_blocked_on_lock(task);
 733	/*
 734	 * Store the top waiter of @lock for the end of chain walk
 735	 * decision below.
 736	 */
 737	top_waiter = rt_mutex_top_waiter(lock);
 738
 739	/* [13] Drop the locks */
 740	raw_spin_unlock(&task->pi_lock);
 741	raw_spin_unlock_irq(&lock->wait_lock);
 742
 743	/*
 744	 * Make the actual exit decisions [12], based on the stored
 745	 * values.
 746	 *
 747	 * We reached the end of the lock chain. Stop right here. No
 748	 * point to go back just to figure that out.
 749	 */
 750	if (!next_lock)
 751		goto out_put_task;
 752
 753	/*
 754	 * If the current waiter is not the top waiter on the lock,
 755	 * then we can stop the chain walk here if we are not in full
 756	 * deadlock detection mode.
 757	 */
 758	if (!detect_deadlock && waiter != top_waiter)
 759		goto out_put_task;
 760
 761	goto again;
 762
 763 out_unlock_pi:
 764	raw_spin_unlock_irq(&task->pi_lock);
 765 out_put_task:
 766	put_task_struct(task);
 767
 768	return ret;
 769}
 770
 771/*
 772 * Try to take an rt-mutex
 773 *
 774 * Must be called with lock->wait_lock held and interrupts disabled
 775 *
 776 * @lock:   The lock to be acquired.
 777 * @task:   The task which wants to acquire the lock
 778 * @waiter: The waiter that is queued to the lock's wait tree if the
 779 *	    callsite called task_blocked_on_lock(), otherwise NULL
 780 */
 781static int __sched
 782try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 783		     struct rt_mutex_waiter *waiter)
 784{
 785	lockdep_assert_held(&lock->wait_lock);
 786
 787	/*
 788	 * Before testing whether we can acquire @lock, we set the
 789	 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 790	 * other tasks which try to modify @lock into the slow path
 791	 * and they serialize on @lock->wait_lock.
 792	 *
 793	 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 794	 * as explained at the top of this file if and only if:
 795	 *
 796	 * - There is a lock owner. The caller must fixup the
 797	 *   transient state if it does a trylock or leaves the lock
 798	 *   function due to a signal or timeout.
 799	 *
 800	 * - @task acquires the lock and there are no other
 801	 *   waiters. This is undone in rt_mutex_set_owner(@task) at
 802	 *   the end of this function.
 803	 */
 804	mark_rt_mutex_waiters(lock);
 805
 806	/*
 807	 * If @lock has an owner, give up.
 808	 */
 809	if (rt_mutex_owner(lock))
 810		return 0;
 811
 812	/*
 813	 * If @waiter != NULL, @task has already enqueued the waiter
 814	 * into @lock waiter tree. If @waiter == NULL then this is a
 815	 * trylock attempt.
 816	 */
 817	if (waiter) {
 
 
 818		/*
 819		 * If waiter is not the highest priority waiter of
 820		 * @lock, give up.
 821		 */
 822		if (waiter != rt_mutex_top_waiter(lock))
 
 
 
 
 
 
 823			return 0;
 824
 825		/*
 826		 * We can acquire the lock. Remove the waiter from the
 827		 * lock waiters tree.
 828		 */
 829		rt_mutex_dequeue(lock, waiter);
 830
 831	} else {
 832		/*
 833		 * If the lock has waiters already we check whether @task is
 834		 * eligible to take over the lock.
 835		 *
 836		 * If there are no other waiters, @task can acquire
 837		 * the lock.  @task->pi_blocked_on is NULL, so it does
 838		 * not need to be dequeued.
 839		 */
 840		if (rt_mutex_has_waiters(lock)) {
 841			/*
 842			 * If @task->prio is greater than or equal to
 843			 * the top waiter priority (kernel view),
 844			 * @task lost.
 845			 */
 846			if (!rt_mutex_waiter_less(task_to_waiter(task),
 847						  rt_mutex_top_waiter(lock)))
 848				return 0;
 849
 850			/*
 851			 * The current top waiter stays enqueued. We
 852			 * don't have to change anything in the lock
 853			 * waiters order.
 854			 */
 855		} else {
 856			/*
 857			 * No waiters. Take the lock without the
 858			 * pi_lock dance.@task->pi_blocked_on is NULL
 859			 * and we have no waiters to enqueue in @task
 860			 * pi waiters tree.
 861			 */
 862			goto takeit;
 863		}
 864	}
 865
 866	/*
 867	 * Clear @task->pi_blocked_on. Requires protection by
 868	 * @task->pi_lock. Redundant operation for the @waiter == NULL
 869	 * case, but conditionals are more expensive than a redundant
 870	 * store.
 871	 */
 872	raw_spin_lock(&task->pi_lock);
 873	task->pi_blocked_on = NULL;
 874	/*
 875	 * Finish the lock acquisition. @task is the new owner. If
 876	 * other waiters exist we have to insert the highest priority
 877	 * waiter into @task->pi_waiters tree.
 878	 */
 879	if (rt_mutex_has_waiters(lock))
 880		rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
 881	raw_spin_unlock(&task->pi_lock);
 882
 883takeit:
 884	/*
 885	 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
 886	 * are still waiters or clears it.
 887	 */
 888	rt_mutex_set_owner(lock, task);
 889
 890	return 1;
 891}
 892
 893/*
 894 * Task blocks on lock.
 895 *
 896 * Prepare waiter and propagate pi chain
 897 *
 898 * This must be called with lock->wait_lock held and interrupts disabled
 899 */
 900static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
 901					   struct rt_mutex_waiter *waiter,
 902					   struct task_struct *task,
 903					   enum rtmutex_chainwalk chwalk)
 
 
 904{
 905	struct task_struct *owner = rt_mutex_owner(lock);
 906	struct rt_mutex_waiter *top_waiter = waiter;
 907	struct rt_mutex *next_lock;
 908	int chain_walk = 0, res;
 909
 910	lockdep_assert_held(&lock->wait_lock);
 911
 912	/*
 913	 * Early deadlock detection. We really don't want the task to
 914	 * enqueue on itself just to untangle the mess later. It's not
 915	 * only an optimization. We drop the locks, so another waiter
 916	 * can come in before the chain walk detects the deadlock. So
 917	 * the other will detect the deadlock and return -EDEADLOCK,
 918	 * which is wrong, as the other waiter is not in a deadlock
 919	 * situation.
 
 
 
 920	 */
 921	if (owner == task)
 922		return -EDEADLK;
 923
 924	raw_spin_lock(&task->pi_lock);
 925	waiter->task = task;
 926	waiter->lock = lock;
 927	waiter->prio = task->prio;
 928	waiter->deadline = task->dl.deadline;
 929
 930	/* Get the top priority waiter on the lock */
 931	if (rt_mutex_has_waiters(lock))
 932		top_waiter = rt_mutex_top_waiter(lock);
 933	rt_mutex_enqueue(lock, waiter);
 934
 935	task->pi_blocked_on = waiter;
 936
 937	raw_spin_unlock(&task->pi_lock);
 938
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 939	if (!owner)
 940		return 0;
 941
 942	raw_spin_lock(&owner->pi_lock);
 943	if (waiter == rt_mutex_top_waiter(lock)) {
 944		rt_mutex_dequeue_pi(owner, top_waiter);
 945		rt_mutex_enqueue_pi(owner, waiter);
 946
 947		rt_mutex_adjust_prio(owner);
 948		if (owner->pi_blocked_on)
 949			chain_walk = 1;
 950	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 951		chain_walk = 1;
 952	}
 953
 954	/* Store the lock on which owner is blocked or NULL */
 955	next_lock = task_blocked_on_lock(owner);
 956
 957	raw_spin_unlock(&owner->pi_lock);
 958	/*
 959	 * Even if full deadlock detection is on, if the owner is not
 960	 * blocked itself, we can avoid finding this out in the chain
 961	 * walk.
 962	 */
 963	if (!chain_walk || !next_lock)
 964		return 0;
 965
 966	/*
 967	 * The owner can't disappear while holding a lock,
 968	 * so the owner struct is protected by wait_lock.
 969	 * Gets dropped in rt_mutex_adjust_prio_chain()!
 970	 */
 971	get_task_struct(owner);
 972
 
 973	raw_spin_unlock_irq(&lock->wait_lock);
 
 
 
 
 
 974
 975	res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
 976					 next_lock, waiter, task);
 977
 978	raw_spin_lock_irq(&lock->wait_lock);
 979
 980	return res;
 981}
 982
 983/*
 984 * Remove the top waiter from the current tasks pi waiter tree and
 985 * queue it up.
 986 *
 987 * Called with lock->wait_lock held and interrupts disabled.
 988 */
 989static void __sched mark_wakeup_next_waiter(struct wake_q_head *wake_q,
 990					    struct rt_mutex *lock)
 991{
 992	struct rt_mutex_waiter *waiter;
 993
 
 
 994	raw_spin_lock(&current->pi_lock);
 995
 996	waiter = rt_mutex_top_waiter(lock);
 997
 998	/*
 999	 * Remove it from current->pi_waiters and deboost.
1000	 *
1001	 * We must in fact deboost here in order to ensure we call
1002	 * rt_mutex_setprio() to update p->pi_top_task before the
1003	 * task unblocks.
1004	 */
1005	rt_mutex_dequeue_pi(current, waiter);
1006	rt_mutex_adjust_prio(current);
1007
1008	/*
1009	 * As we are waking up the top waiter, and the waiter stays
1010	 * queued on the lock until it gets the lock, this lock
1011	 * obviously has waiters. Just set the bit here and this has
1012	 * the added benefit of forcing all new tasks into the
1013	 * slow path making sure no task of lower priority than
1014	 * the top waiter can steal this lock.
1015	 */
1016	lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1017
1018	/*
1019	 * We deboosted before waking the top waiter task such that we don't
1020	 * run two tasks with the 'same' priority (and ensure the
1021	 * p->pi_top_task pointer points to a blocked task). This however can
1022	 * lead to priority inversion if we would get preempted after the
1023	 * deboost but before waking our donor task, hence the preempt_disable()
1024	 * before unlock.
1025	 *
1026	 * Pairs with preempt_enable() in rt_mutex_postunlock();
1027	 */
1028	preempt_disable();
1029	wake_q_add(wake_q, waiter->task);
1030	raw_spin_unlock(&current->pi_lock);
1031}
1032
1033/*
1034 * Remove a waiter from a lock and give up
1035 *
1036 * Must be called with lock->wait_lock held and interrupts disabled. I must
1037 * have just failed to try_to_take_rt_mutex().
1038 */
1039static void __sched remove_waiter(struct rt_mutex *lock,
1040				  struct rt_mutex_waiter *waiter)
1041{
1042	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1043	struct task_struct *owner = rt_mutex_owner(lock);
1044	struct rt_mutex *next_lock;
1045
1046	lockdep_assert_held(&lock->wait_lock);
1047
1048	raw_spin_lock(&current->pi_lock);
1049	rt_mutex_dequeue(lock, waiter);
1050	current->pi_blocked_on = NULL;
1051	raw_spin_unlock(&current->pi_lock);
1052
1053	/*
1054	 * Only update priority if the waiter was the highest priority
1055	 * waiter of the lock and there is an owner to update.
1056	 */
1057	if (!owner || !is_top_waiter)
1058		return;
1059
1060	raw_spin_lock(&owner->pi_lock);
1061
1062	rt_mutex_dequeue_pi(owner, waiter);
1063
1064	if (rt_mutex_has_waiters(lock))
1065		rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1066
1067	rt_mutex_adjust_prio(owner);
1068
1069	/* Store the lock on which owner is blocked or NULL */
1070	next_lock = task_blocked_on_lock(owner);
1071
1072	raw_spin_unlock(&owner->pi_lock);
1073
1074	/*
1075	 * Don't walk the chain, if the owner task is not blocked
1076	 * itself.
1077	 */
1078	if (!next_lock)
1079		return;
1080
1081	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1082	get_task_struct(owner);
1083
1084	raw_spin_unlock_irq(&lock->wait_lock);
1085
1086	rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1087				   next_lock, NULL, current);
1088
1089	raw_spin_lock_irq(&lock->wait_lock);
1090}
1091
1092/*
1093 * Recheck the pi chain, in case we got a priority setting
1094 *
1095 * Called from sched_setscheduler
1096 */
1097void __sched rt_mutex_adjust_pi(struct task_struct *task)
1098{
1099	struct rt_mutex_waiter *waiter;
1100	struct rt_mutex *next_lock;
1101	unsigned long flags;
1102
1103	raw_spin_lock_irqsave(&task->pi_lock, flags);
1104
1105	waiter = task->pi_blocked_on;
1106	if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1107		raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1108		return;
1109	}
1110	next_lock = waiter->lock;
1111	raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1112
1113	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1114	get_task_struct(task);
1115
1116	rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1117				   next_lock, NULL, task);
1118}
1119
1120void __sched rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1121{
1122	debug_rt_mutex_init_waiter(waiter);
1123	RB_CLEAR_NODE(&waiter->pi_tree_entry);
1124	RB_CLEAR_NODE(&waiter->tree_entry);
1125	waiter->task = NULL;
1126}
1127
1128/**
1129 * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1130 * @lock:		 the rt_mutex to take
1131 * @state:		 the state the task should block in (TASK_INTERRUPTIBLE
1132 *			 or TASK_UNINTERRUPTIBLE)
1133 * @timeout:		 the pre-initialized and started timer, or NULL for none
1134 * @waiter:		 the pre-initialized rt_mutex_waiter
1135 *
1136 * Must be called with lock->wait_lock held and interrupts disabled
1137 */
1138static int __sched __rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
1139				       struct hrtimer_sleeper *timeout,
1140				       struct rt_mutex_waiter *waiter)
1141{
1142	int ret = 0;
1143
1144	for (;;) {
1145		/* Try to acquire the lock: */
1146		if (try_to_take_rt_mutex(lock, current, waiter))
1147			break;
1148
1149		if (timeout && !timeout->task) {
1150			ret = -ETIMEDOUT;
1151			break;
1152		}
1153		if (signal_pending_state(state, current)) {
1154			ret = -EINTR;
1155			break;
1156		}
1157
1158		raw_spin_unlock_irq(&lock->wait_lock);
1159
1160		schedule();
1161
1162		raw_spin_lock_irq(&lock->wait_lock);
1163		set_current_state(state);
1164	}
1165
1166	__set_current_state(TASK_RUNNING);
1167	return ret;
1168}
1169
1170static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1171					     struct rt_mutex_waiter *w)
1172{
1173	/*
1174	 * If the result is not -EDEADLOCK or the caller requested
1175	 * deadlock detection, nothing to do here.
1176	 */
1177	if (res != -EDEADLOCK || detect_deadlock)
1178		return;
1179
1180	/*
1181	 * Yell loudly and stop the task right here.
1182	 */
1183	WARN(1, "rtmutex deadlock detected\n");
1184	while (1) {
1185		set_current_state(TASK_INTERRUPTIBLE);
1186		schedule();
1187	}
1188}
1189
1190/*
1191 * Slow path lock function:
1192 */
1193static int __sched rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
1194				     struct hrtimer_sleeper *timeout,
1195				     enum rtmutex_chainwalk chwalk)
1196{
1197	struct rt_mutex_waiter waiter;
1198	unsigned long flags;
1199	int ret = 0;
1200
1201	rt_mutex_init_waiter(&waiter);
1202
1203	/*
1204	 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1205	 * be called in early boot if the cmpxchg() fast path is disabled
1206	 * (debug, no architecture support). In this case we will acquire the
1207	 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1208	 * enable interrupts in that early boot case. So we need to use the
1209	 * irqsave/restore variants.
1210	 */
1211	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1212
1213	/* Try to acquire the lock again: */
1214	if (try_to_take_rt_mutex(lock, current, NULL)) {
1215		raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1216		return 0;
1217	}
1218
1219	set_current_state(state);
1220
1221	/* Setup the timer, when timeout != NULL */
1222	if (unlikely(timeout))
1223		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1224
1225	ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1226
1227	if (likely(!ret))
1228		/* sleep on the mutex */
1229		ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1230
1231	if (unlikely(ret)) {
1232		__set_current_state(TASK_RUNNING);
1233		remove_waiter(lock, &waiter);
1234		rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1235	}
1236
1237	/*
1238	 * try_to_take_rt_mutex() sets the waiter bit
1239	 * unconditionally. We might have to fix that up.
1240	 */
1241	fixup_rt_mutex_waiters(lock);
1242
1243	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1244
1245	/* Remove pending timer: */
1246	if (unlikely(timeout))
1247		hrtimer_cancel(&timeout->timer);
1248
1249	debug_rt_mutex_free_waiter(&waiter);
1250
1251	return ret;
1252}
1253
1254static int __sched __rt_mutex_slowtrylock(struct rt_mutex *lock)
1255{
1256	int ret = try_to_take_rt_mutex(lock, current, NULL);
1257
1258	/*
1259	 * try_to_take_rt_mutex() sets the lock waiters bit
1260	 * unconditionally. Clean this up.
1261	 */
1262	fixup_rt_mutex_waiters(lock);
1263
1264	return ret;
1265}
1266
1267/*
1268 * Slow path try-lock function:
1269 */
1270static int __sched rt_mutex_slowtrylock(struct rt_mutex *lock)
1271{
1272	unsigned long flags;
1273	int ret;
1274
1275	/*
1276	 * If the lock already has an owner we fail to get the lock.
1277	 * This can be done without taking the @lock->wait_lock as
1278	 * it is only being read, and this is a trylock anyway.
1279	 */
1280	if (rt_mutex_owner(lock))
1281		return 0;
1282
1283	/*
1284	 * The mutex has currently no owner. Lock the wait lock and try to
1285	 * acquire the lock. We use irqsave here to support early boot calls.
1286	 */
1287	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1288
1289	ret = __rt_mutex_slowtrylock(lock);
1290
1291	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1292
1293	return ret;
1294}
1295
1296/*
1297 * Performs the wakeup of the top-waiter and re-enables preemption.
1298 */
1299void __sched rt_mutex_postunlock(struct wake_q_head *wake_q)
1300{
1301	wake_up_q(wake_q);
 
1302
1303	/* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
1304	preempt_enable();
1305}
1306
1307/*
1308 * Slow path to release a rt-mutex.
1309 *
1310 * Return whether the current task needs to call rt_mutex_postunlock().
1311 */
1312static void __sched rt_mutex_slowunlock(struct rt_mutex *lock)
1313{
1314	DEFINE_WAKE_Q(wake_q);
1315	unsigned long flags;
1316
1317	/* irqsave required to support early boot calls */
1318	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1319
1320	debug_rt_mutex_unlock(lock);
1321
1322	/*
1323	 * We must be careful here if the fast path is enabled. If we
1324	 * have no waiters queued we cannot set owner to NULL here
1325	 * because of:
1326	 *
1327	 * foo->lock->owner = NULL;
1328	 *			rtmutex_lock(foo->lock);   <- fast path
1329	 *			free = atomic_dec_and_test(foo->refcnt);
1330	 *			rtmutex_unlock(foo->lock); <- fast path
1331	 *			if (free)
1332	 *				kfree(foo);
1333	 * raw_spin_unlock(foo->lock->wait_lock);
1334	 *
1335	 * So for the fastpath enabled kernel:
1336	 *
1337	 * Nothing can set the waiters bit as long as we hold
1338	 * lock->wait_lock. So we do the following sequence:
1339	 *
1340	 *	owner = rt_mutex_owner(lock);
1341	 *	clear_rt_mutex_waiters(lock);
1342	 *	raw_spin_unlock(&lock->wait_lock);
1343	 *	if (cmpxchg(&lock->owner, owner, 0) == owner)
1344	 *		return;
1345	 *	goto retry;
1346	 *
1347	 * The fastpath disabled variant is simple as all access to
1348	 * lock->owner is serialized by lock->wait_lock:
1349	 *
1350	 *	lock->owner = NULL;
1351	 *	raw_spin_unlock(&lock->wait_lock);
1352	 */
1353	while (!rt_mutex_has_waiters(lock)) {
1354		/* Drops lock->wait_lock ! */
1355		if (unlock_rt_mutex_safe(lock, flags) == true)
1356			return;
1357		/* Relock the rtmutex and try again */
1358		raw_spin_lock_irqsave(&lock->wait_lock, flags);
1359	}
1360
1361	/*
1362	 * The wakeup next waiter path does not suffer from the above
1363	 * race. See the comments there.
1364	 *
1365	 * Queue the next waiter for wakeup once we release the wait_lock.
1366	 */
1367	mark_wakeup_next_waiter(&wake_q, lock);
1368	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1369
1370	rt_mutex_postunlock(&wake_q);
1371}
1372
1373/*
1374 * debug aware fast / slowpath lock,trylock,unlock
1375 *
1376 * The atomic acquire/release ops are compiled away, when either the
1377 * architecture does not support cmpxchg or when debugging is enabled.
1378 */
1379static __always_inline int __rt_mutex_lock(struct rt_mutex *lock, long state,
1380					   unsigned int subclass)
1381{
1382	int ret;
 
1383
1384	might_sleep();
1385	mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
1386
1387	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1388		return 0;
1389
1390	ret = rt_mutex_slowlock(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1391	if (ret)
1392		mutex_release(&lock->dep_map, _RET_IP_);
1393	return ret;
1394}
1395
1396#ifdef CONFIG_DEBUG_LOCK_ALLOC
1397/**
1398 * rt_mutex_lock_nested - lock a rt_mutex
1399 *
1400 * @lock: the rt_mutex to be locked
1401 * @subclass: the lockdep subclass
1402 */
1403void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
1404{
1405	__rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, subclass);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1406}
1407EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
1408
1409#else /* !CONFIG_DEBUG_LOCK_ALLOC */
1410
1411/**
1412 * rt_mutex_lock - lock a rt_mutex
1413 *
1414 * @lock: the rt_mutex to be locked
1415 */
1416void __sched rt_mutex_lock(struct rt_mutex *lock)
1417{
1418	__rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0);
1419}
1420EXPORT_SYMBOL_GPL(rt_mutex_lock);
1421#endif
1422
1423/**
1424 * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1425 *
1426 * @lock:		the rt_mutex to be locked
1427 *
1428 * Returns:
1429 *  0		on success
1430 * -EINTR	when interrupted by a signal
1431 */
1432int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1433{
1434	return __rt_mutex_lock(lock, TASK_INTERRUPTIBLE, 0);
1435}
1436EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1437
1438/**
1439 * rt_mutex_trylock - try to lock a rt_mutex
1440 *
1441 * @lock:	the rt_mutex to be locked
1442 *
1443 * This function can only be called in thread context. It's safe to call it
1444 * from atomic regions, but not from hard or soft interrupt context.
1445 *
1446 * Returns:
1447 *  1 on success
1448 *  0 on contention
1449 */
1450int __sched rt_mutex_trylock(struct rt_mutex *lock)
 
1451{
1452	int ret;
 
 
 
 
1453
1454	if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES) && WARN_ON_ONCE(!in_task()))
1455		return 0;
 
 
1456
1457	/*
1458	 * No lockdep annotation required because lockdep disables the fast
1459	 * path.
1460	 */
1461	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1462		return 1;
 
 
 
 
 
 
 
 
 
1463
1464	ret = rt_mutex_slowtrylock(lock);
1465	if (ret)
1466		mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1467
1468	return ret;
1469}
1470EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1471
1472/**
1473 * rt_mutex_unlock - unlock a rt_mutex
1474 *
1475 * @lock: the rt_mutex to be unlocked
1476 */
1477void __sched rt_mutex_unlock(struct rt_mutex *lock)
1478{
1479	mutex_release(&lock->dep_map, _RET_IP_);
1480	if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1481		return;
1482
1483	rt_mutex_slowunlock(lock);
1484}
1485EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 
1486
1487/*
1488 * Futex variants, must not use fastpath.
1489 */
1490int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1491{
1492	return rt_mutex_slowtrylock(lock);
1493}
1494
1495int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1496{
1497	return __rt_mutex_slowtrylock(lock);
1498}
1499
1500/**
1501 * __rt_mutex_futex_unlock - Futex variant, that since futex variants
1502 * do not use the fast-path, can be simple and will not need to retry.
 
 
 
 
 
 
1503 *
1504 * @lock:	The rt_mutex to be unlocked
1505 * @wake_q:	The wake queue head from which to get the next lock waiter
1506 */
1507bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1508				     struct wake_q_head *wake_q)
 
 
 
 
 
1509{
1510	lockdep_assert_held(&lock->wait_lock);
 
 
1511
1512	debug_rt_mutex_unlock(lock);
 
 
 
1513
1514	if (!rt_mutex_has_waiters(lock)) {
1515		lock->owner = NULL;
1516		return false; /* done */
1517	}
 
 
 
 
1518
1519	/*
1520	 * We've already deboosted, mark_wakeup_next_waiter() will
1521	 * retain preempt_disabled when we drop the wait_lock, to
1522	 * avoid inversion prior to the wakeup.  preempt_disable()
1523	 * therein pairs with rt_mutex_postunlock().
1524	 */
1525	mark_wakeup_next_waiter(wake_q, lock);
1526
1527	return true; /* call postunlock() */
1528}
 
 
 
 
 
 
 
 
 
1529
1530void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1531{
1532	DEFINE_WAKE_Q(wake_q);
1533	unsigned long flags;
1534	bool postunlock;
1535
1536	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1537	postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1538	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1539
1540	if (postunlock)
1541		rt_mutex_postunlock(&wake_q);
1542}
1543
1544/**
1545 * __rt_mutex_init - initialize the rt_mutex
1546 *
1547 * @lock:	The rt_mutex to be initialized
1548 * @name:	The lock name used for debugging
1549 * @key:	The lock class key used for debugging
1550 *
1551 * Initialize the rt_mutex to unlocked state.
1552 *
1553 * Initializing of a locked rt_mutex is not allowed
1554 */
1555void __sched __rt_mutex_init(struct rt_mutex *lock, const char *name,
1556		     struct lock_class_key *key)
1557{
1558	debug_check_no_locks_freed((void *)lock, sizeof(*lock));
1559	lockdep_init_map_wait(&lock->dep_map, name, key, 0, LD_WAIT_SLEEP);
 
 
 
 
 
 
 
1560
1561	__rt_mutex_basic_init(lock);
1562}
1563EXPORT_SYMBOL_GPL(__rt_mutex_init);
1564
1565/**
1566 * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1567 *				proxy owner
1568 *
1569 * @lock:	the rt_mutex to be locked
1570 * @proxy_owner:the task to set as owner
1571 *
1572 * No locking. Caller has to do serializing itself
1573 *
1574 * Special API call for PI-futex support. This initializes the rtmutex and
1575 * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1576 * possible at this point because the pi_state which contains the rtmutex
1577 * is not yet visible to other tasks.
1578 */
1579void __sched rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1580					struct task_struct *proxy_owner)
1581{
1582	__rt_mutex_basic_init(lock);
1583	rt_mutex_set_owner(lock, proxy_owner);
1584}
1585
1586/**
1587 * rt_mutex_proxy_unlock - release a lock on behalf of owner
1588 *
1589 * @lock:	the rt_mutex to be locked
1590 *
1591 * No locking. Caller has to do serializing itself
1592 *
1593 * Special API call for PI-futex support. This merrily cleans up the rtmutex
1594 * (debugging) state. Concurrent operations on this rt_mutex are not
1595 * possible because it belongs to the pi_state which is about to be freed
1596 * and it is not longer visible to other tasks.
1597 */
1598void __sched rt_mutex_proxy_unlock(struct rt_mutex *lock)
1599{
1600	debug_rt_mutex_proxy_unlock(lock);
1601	rt_mutex_set_owner(lock, NULL);
1602}
1603
1604/**
1605 * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1606 * @lock:		the rt_mutex to take
1607 * @waiter:		the pre-initialized rt_mutex_waiter
1608 * @task:		the task to prepare
1609 *
1610 * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1611 * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1612 *
1613 * NOTE: does _NOT_ remove the @waiter on failure; must either call
1614 * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
1615 *
1616 * Returns:
1617 *  0 - task blocked on lock
1618 *  1 - acquired the lock for task, caller should wake it up
1619 * <0 - error
1620 *
1621 * Special API call for PI-futex support.
1622 */
1623int __sched __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1624					struct rt_mutex_waiter *waiter,
1625					struct task_struct *task)
1626{
 
 
1627	int ret;
1628
1629	lockdep_assert_held(&lock->wait_lock);
1630
1631	if (try_to_take_rt_mutex(lock, task, NULL))
1632		return 1;
 
 
 
 
 
 
 
 
 
 
1633
1634	/* We enforce deadlock detection for futexes */
1635	ret = task_blocks_on_rt_mutex(lock, waiter, task,
1636				      RT_MUTEX_FULL_CHAINWALK);
1637
1638	if (ret && !rt_mutex_owner(lock)) {
1639		/*
1640		 * Reset the return value. We might have
1641		 * returned with -EDEADLK and the owner
1642		 * released the lock while we were walking the
1643		 * pi chain.  Let the waiter sort it out.
1644		 */
1645		ret = 0;
 
 
 
1646	}
1647
 
 
 
 
 
 
 
 
1648	return ret;
1649}
1650
1651/**
1652 * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1653 * @lock:		the rt_mutex to take
1654 * @waiter:		the pre-initialized rt_mutex_waiter
1655 * @task:		the task to prepare
1656 *
1657 * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1658 * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1659 *
1660 * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
1661 * on failure.
1662 *
1663 * Returns:
1664 *  0 - task blocked on lock
1665 *  1 - acquired the lock for task, caller should wake it up
1666 * <0 - error
1667 *
1668 * Special API call for PI-futex support.
1669 */
1670int __sched rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1671				      struct rt_mutex_waiter *waiter,
1672				      struct task_struct *task)
1673{
 
1674	int ret;
1675
1676	raw_spin_lock_irq(&lock->wait_lock);
1677	ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1678	if (unlikely(ret))
1679		remove_waiter(lock, waiter);
1680	raw_spin_unlock_irq(&lock->wait_lock);
1681
 
1682	return ret;
1683}
1684
1685/**
1686 * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1687 * @lock:		the rt_mutex we were woken on
1688 * @to:			the timeout, null if none. hrtimer should already have
1689 *			been started.
1690 * @waiter:		the pre-initialized rt_mutex_waiter
1691 *
1692 * Wait for the lock acquisition started on our behalf by
1693 * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1694 * rt_mutex_cleanup_proxy_lock().
1695 *
1696 * Returns:
1697 *  0 - success
1698 * <0 - error, one of -EINTR, -ETIMEDOUT
1699 *
1700 * Special API call for PI-futex support
1701 */
1702int __sched rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1703				     struct hrtimer_sleeper *to,
1704				     struct rt_mutex_waiter *waiter)
1705{
 
 
1706	int ret;
1707
1708	raw_spin_lock_irq(&lock->wait_lock);
1709	/* sleep on the mutex */
1710	set_current_state(TASK_INTERRUPTIBLE);
1711	ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1712	/*
1713	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1714	 * have to fix that up.
 
 
 
 
 
 
 
 
 
 
 
 
 
1715	 */
1716	fixup_rt_mutex_waiters(lock);
1717	raw_spin_unlock_irq(&lock->wait_lock);
 
 
 
 
 
1718
1719	return ret;
1720}
1721
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1722/**
1723 * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1724 * @lock:		the rt_mutex we were woken on
1725 * @waiter:		the pre-initialized rt_mutex_waiter
1726 *
1727 * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
1728 * rt_mutex_wait_proxy_lock().
1729 *
1730 * Unless we acquired the lock; we're still enqueued on the wait-list and can
1731 * in fact still be granted ownership until we're removed. Therefore we can
1732 * find we are in fact the owner and must disregard the
1733 * rt_mutex_wait_proxy_lock() failure.
1734 *
1735 * Returns:
1736 *  true  - did the cleanup, we done.
1737 *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1738 *          caller should disregards its return value.
1739 *
1740 * Special API call for PI-futex support
1741 */
1742bool __sched rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1743					 struct rt_mutex_waiter *waiter)
1744{
1745	bool cleanup = false;
 
 
 
 
 
 
 
 
 
 
 
1746
1747	raw_spin_lock_irq(&lock->wait_lock);
1748	/*
1749	 * Do an unconditional try-lock, this deals with the lock stealing
1750	 * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1751	 * sets a NULL owner.
1752	 *
1753	 * We're not interested in the return value, because the subsequent
1754	 * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1755	 * we will own the lock and it will have removed the waiter. If we
1756	 * failed the trylock, we're still not owner and we need to remove
1757	 * ourselves.
1758	 */
1759	try_to_take_rt_mutex(lock, current, waiter);
1760	/*
1761	 * Unless we're the owner; we're still enqueued on the wait_list.
1762	 * So check if we became owner, if not, take us off the wait_list.
1763	 */
1764	if (rt_mutex_owner(lock) != current) {
1765		remove_waiter(lock, waiter);
1766		cleanup = true;
 
 
 
 
1767	}
 
 
 
 
1768	/*
1769	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1770	 * have to fix that up.
1771	 */
1772	fixup_rt_mutex_waiters(lock);
1773
1774	raw_spin_unlock_irq(&lock->wait_lock);
1775
1776	return cleanup;
1777}
1778
1779#ifdef CONFIG_DEBUG_RT_MUTEXES
1780void rt_mutex_debug_task_free(struct task_struct *task)
1781{
1782	DEBUG_LOCKS_WARN_ON(!RB_EMPTY_ROOT(&task->pi_waiters.rb_root));
1783	DEBUG_LOCKS_WARN_ON(task->pi_blocked_on);
 
 
 
 
 
 
 
1784}
1785#endif