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