kernel/locking/rtmutex_api.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * rtmutex API
   4  */
   5 #include <linux/spinlock.h>
   6 #include <linux/export.h>
   7
   8 #define RT_MUTEX_BUILD_MUTEX
   9 #include "rtmutex.c"
  10
  11 /*
  12  * Max number of times we'll walk the boosting chain:
  13  */
  14 int max_lock_depth = 1024;
  15
  16 /*
  17  * Debug aware fast / slowpath lock,trylock,unlock
  18  *
  19  * The atomic acquire/release ops are compiled away, when either the
  20  * architecture does not support cmpxchg or when debugging is enabled.
  21  */
  22 static __always_inline int __rt_mutex_lock_common(struct rt_mutex *lock,
  23                                                   unsigned int state,
  24                                                   unsigned int subclass)
  25 {
  26         int ret;
  27
  28         might_sleep();
  29         mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
  30         ret = __rt_mutex_lock(&lock->rtmutex, state);
  31         if (ret)
  32                 mutex_release(&lock->dep_map, _RET_IP_);
  33         return ret;
  34 }
  35
  36 void rt_mutex_base_init(struct rt_mutex_base *rtb)
  37 {
  38         __rt_mutex_base_init(rtb);
  39 }
  40 EXPORT_SYMBOL(rt_mutex_base_init);
  41
  42 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  43 /**
  44  * rt_mutex_lock_nested - lock a rt_mutex
  45  *
  46  * @lock: the rt_mutex to be locked
  47  * @subclass: the lockdep subclass
  48  */
  49 void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
  50 {
  51         __rt_mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, subclass);
  52 }
  53 EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
  54
  55 #else /* !CONFIG_DEBUG_LOCK_ALLOC */
  56
  57 /**
  58  * rt_mutex_lock - lock a rt_mutex
  59  *
  60  * @lock: the rt_mutex to be locked
  61  */
  62 void __sched rt_mutex_lock(struct rt_mutex *lock)
  63 {
  64         __rt_mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0);
  65 }
  66 EXPORT_SYMBOL_GPL(rt_mutex_lock);
  67 #endif
  68
  69 /**
  70  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
  71  *
  72  * @lock:               the rt_mutex to be locked
  73  *
  74  * Returns:
  75  *  0           on success
  76  * -EINTR       when interrupted by a signal
  77  */
  78 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
  79 {
  80         return __rt_mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0);
  81 }
  82 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
  83
  84 /**
  85  * rt_mutex_trylock - try to lock a rt_mutex
  86  *
  87  * @lock:       the rt_mutex to be locked
  88  *
  89  * This function can only be called in thread context. It's safe to call it
  90  * from atomic regions, but not from hard or soft interrupt context.
  91  *
  92  * Returns:
  93  *  1 on success
  94  *  0 on contention
  95  */
  96 int __sched rt_mutex_trylock(struct rt_mutex *lock)
  97 {
  98         int ret;
  99
 100         if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES) && WARN_ON_ONCE(!in_task()))
 101                 return 0;
 102
 103         ret = __rt_mutex_trylock(&lock->rtmutex);
 104         if (ret)
 105                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
 106
 107         return ret;
 108 }
 109 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 110
 111 /**
 112  * rt_mutex_unlock - unlock a rt_mutex
 113  *
 114  * @lock: the rt_mutex to be unlocked
 115  */
 116 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 117 {
 118         mutex_release(&lock->dep_map, _RET_IP_);
 119         __rt_mutex_unlock(&lock->rtmutex);
 120 }
 121 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 122
 123 /*
 124  * Futex variants, must not use fastpath.
 125  */
 126 int __sched rt_mutex_futex_trylock(struct rt_mutex_base *lock)
 127 {
 128         return rt_mutex_slowtrylock(lock);
 129 }
 130
 131 int __sched __rt_mutex_futex_trylock(struct rt_mutex_base *lock)
 132 {
 133         return __rt_mutex_slowtrylock(lock);
 134 }
 135
 136 /**
 137  * __rt_mutex_futex_unlock - Futex variant, that since futex variants
 138  * do not use the fast-path, can be simple and will not need to retry.
 139  *
 140  * @lock:       The rt_mutex to be unlocked
 141  * @wqh:        The wake queue head from which to get the next lock waiter
 142  */
 143 bool __sched __rt_mutex_futex_unlock(struct rt_mutex_base *lock,
 144                                      struct rt_wake_q_head *wqh)
 145 {
 146         lockdep_assert_held(&lock->wait_lock);
 147
 148         debug_rt_mutex_unlock(lock);
 149
 150         if (!rt_mutex_has_waiters(lock)) {
 151                 lock->owner = NULL;
 152                 return false; /* done */
 153         }
 154
 155         /*
 156          * We've already deboosted, mark_wakeup_next_waiter() will
 157          * retain preempt_disabled when we drop the wait_lock, to
 158          * avoid inversion prior to the wakeup.  preempt_disable()
 159          * therein pairs with rt_mutex_postunlock().
 160          */
 161         mark_wakeup_next_waiter(wqh, lock);
 162
 163         return true; /* call postunlock() */
 164 }
 165
 166 void __sched rt_mutex_futex_unlock(struct rt_mutex_base *lock)
 167 {
 168         DEFINE_RT_WAKE_Q(wqh);
 169         unsigned long flags;
 170         bool postunlock;
 171
 172         raw_spin_lock_irqsave(&lock->wait_lock, flags);
 173         postunlock = __rt_mutex_futex_unlock(lock, &wqh);
 174         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 175
 176         if (postunlock)
 177                 rt_mutex_postunlock(&wqh);
 178 }
 179
 180 /**
 181  * __rt_mutex_init - initialize the rt_mutex
 182  *
 183  * @lock:       The rt_mutex to be initialized
 184  * @name:       The lock name used for debugging
 185  * @key:        The lock class key used for debugging
 186  *
 187  * Initialize the rt_mutex to unlocked state.
 188  *
 189  * Initializing of a locked rt_mutex is not allowed
 190  */
 191 void __sched __rt_mutex_init(struct rt_mutex *lock, const char *name,
 192                              struct lock_class_key *key)
 193 {
 194         debug_check_no_locks_freed((void *)lock, sizeof(*lock));
 195         __rt_mutex_base_init(&lock->rtmutex);
 196         lockdep_init_map_wait(&lock->dep_map, name, key, 0, LD_WAIT_SLEEP);
 197 }
 198 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 199
 200 /**
 201  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 202  *                              proxy owner
 203  *
 204  * @lock:       the rt_mutex to be locked
 205  * @proxy_owner:the task to set as owner
 206  *
 207  * No locking. Caller has to do serializing itself
 208  *
 209  * Special API call for PI-futex support. This initializes the rtmutex and
 210  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
 211  * possible at this point because the pi_state which contains the rtmutex
 212  * is not yet visible to other tasks.
 213  */
 214 void __sched rt_mutex_init_proxy_locked(struct rt_mutex_base *lock,
 215                                         struct task_struct *proxy_owner)
 216 {
 217         static struct lock_class_key pi_futex_key;
 218
 219         __rt_mutex_base_init(lock);
 220         /*
 221          * On PREEMPT_RT the futex hashbucket spinlock becomes 'sleeping'
 222          * and rtmutex based. That causes a lockdep false positive, because
 223          * some of the futex functions invoke spin_unlock(&hb->lock) with
 224          * the wait_lock of the rtmutex associated to the pi_futex held.
 225          * spin_unlock() in turn takes wait_lock of the rtmutex on which
 226          * the spinlock is based, which makes lockdep notice a lock
 227          * recursion. Give the futex/rtmutex wait_lock a separate key.
 228          */
 229         lockdep_set_class(&lock->wait_lock, &pi_futex_key);
 230         rt_mutex_set_owner(lock, proxy_owner);
 231 }
 232
 233 /**
 234  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 235  *
 236  * @lock:       the rt_mutex to be locked
 237  *
 238  * No locking. Caller has to do serializing itself
 239  *
 240  * Special API call for PI-futex support. This just cleans up the rtmutex
 241  * (debugging) state. Concurrent operations on this rt_mutex are not
 242  * possible because it belongs to the pi_state which is about to be freed
 243  * and it is not longer visible to other tasks.
 244  */
 245 void __sched rt_mutex_proxy_unlock(struct rt_mutex_base *lock)
 246 {
 247         debug_rt_mutex_proxy_unlock(lock);
 248         rt_mutex_set_owner(lock, NULL);
 249 }
 250
 251 /**
 252  * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
 253  * @lock:               the rt_mutex to take
 254  * @waiter:             the pre-initialized rt_mutex_waiter
 255  * @task:               the task to prepare
 256  *
 257  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
 258  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
 259  *
 260  * NOTE: does _NOT_ remove the @waiter on failure; must either call
 261  * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
 262  *
 263  * Returns:
 264  *  0 - task blocked on lock
 265  *  1 - acquired the lock for task, caller should wake it up
 266  * <0 - error
 267  *
 268  * Special API call for PI-futex support.
 269  */
 270 int __sched __rt_mutex_start_proxy_lock(struct rt_mutex_base *lock,
 271                                         struct rt_mutex_waiter *waiter,
 272                                         struct task_struct *task)
 273 {
 274         int ret;
 275
 276         lockdep_assert_held(&lock->wait_lock);
 277
 278         if (try_to_take_rt_mutex(lock, task, NULL))
 279                 return 1;
 280
 281         /* We enforce deadlock detection for futexes */
 282         ret = task_blocks_on_rt_mutex(lock, waiter, task, NULL,
 283                                       RT_MUTEX_FULL_CHAINWALK);
 284
 285         if (ret && !rt_mutex_owner(lock)) {
 286                 /*
 287                  * Reset the return value. We might have
 288                  * returned with -EDEADLK and the owner
 289                  * released the lock while we were walking the
 290                  * pi chain.  Let the waiter sort it out.
 291                  */
 292                 ret = 0;
 293         }
 294
 295         return ret;
 296 }
 297
 298 /**
 299  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
 300  * @lock:               the rt_mutex to take
 301  * @waiter:             the pre-initialized rt_mutex_waiter
 302  * @task:               the task to prepare
 303  *
 304  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
 305  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
 306  *
 307  * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
 308  * on failure.
 309  *
 310  * Returns:
 311  *  0 - task blocked on lock
 312  *  1 - acquired the lock for task, caller should wake it up
 313  * <0 - error
 314  *
 315  * Special API call for PI-futex support.
 316  */
 317 int __sched rt_mutex_start_proxy_lock(struct rt_mutex_base *lock,
 318                                       struct rt_mutex_waiter *waiter,
 319                                       struct task_struct *task)
 320 {
 321         int ret;
 322
 323         raw_spin_lock_irq(&lock->wait_lock);
 324         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
 325         if (unlikely(ret))
 326                 remove_waiter(lock, waiter);
 327         raw_spin_unlock_irq(&lock->wait_lock);
 328
 329         return ret;
 330 }
 331
 332 /**
 333  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
 334  * @lock:               the rt_mutex we were woken on
 335  * @to:                 the timeout, null if none. hrtimer should already have
 336  *                      been started.
 337  * @waiter:             the pre-initialized rt_mutex_waiter
 338  *
 339  * Wait for the lock acquisition started on our behalf by
 340  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
 341  * rt_mutex_cleanup_proxy_lock().
 342  *
 343  * Returns:
 344  *  0 - success
 345  * <0 - error, one of -EINTR, -ETIMEDOUT
 346  *
 347  * Special API call for PI-futex support
 348  */
 349 int __sched rt_mutex_wait_proxy_lock(struct rt_mutex_base *lock,
 350                                      struct hrtimer_sleeper *to,
 351                                      struct rt_mutex_waiter *waiter)
 352 {
 353         int ret;
 354
 355         raw_spin_lock_irq(&lock->wait_lock);
 356         /* sleep on the mutex */
 357         set_current_state(TASK_INTERRUPTIBLE);
 358         ret = rt_mutex_slowlock_block(lock, NULL, TASK_INTERRUPTIBLE, to, waiter);
 359         /*
 360          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
 361          * have to fix that up.
 362          */
 363         fixup_rt_mutex_waiters(lock);
 364         raw_spin_unlock_irq(&lock->wait_lock);
 365
 366         return ret;
 367 }
 368
 369 /**
 370  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
 371  * @lock:               the rt_mutex we were woken on
 372  * @waiter:             the pre-initialized rt_mutex_waiter
 373  *
 374  * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
 375  * rt_mutex_wait_proxy_lock().
 376  *
 377  * Unless we acquired the lock; we're still enqueued on the wait-list and can
 378  * in fact still be granted ownership until we're removed. Therefore we can
 379  * find we are in fact the owner and must disregard the
 380  * rt_mutex_wait_proxy_lock() failure.
 381  *
 382  * Returns:
 383  *  true  - did the cleanup, we done.
 384  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
 385  *          caller should disregards its return value.
 386  *
 387  * Special API call for PI-futex support
 388  */
 389 bool __sched rt_mutex_cleanup_proxy_lock(struct rt_mutex_base *lock,
 390                                          struct rt_mutex_waiter *waiter)
 391 {
 392         bool cleanup = false;
 393
 394         raw_spin_lock_irq(&lock->wait_lock);
 395         /*
 396          * Do an unconditional try-lock, this deals with the lock stealing
 397          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
 398          * sets a NULL owner.
 399          *
 400          * We're not interested in the return value, because the subsequent
 401          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
 402          * we will own the lock and it will have removed the waiter. If we
 403          * failed the trylock, we're still not owner and we need to remove
 404          * ourselves.
 405          */
 406         try_to_take_rt_mutex(lock, current, waiter);
 407         /*
 408          * Unless we're the owner; we're still enqueued on the wait_list.
 409          * So check if we became owner, if not, take us off the wait_list.
 410          */
 411         if (rt_mutex_owner(lock) != current) {
 412                 remove_waiter(lock, waiter);
 413                 cleanup = true;
 414         }
 415         /*
 416          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
 417          * have to fix that up.
 418          */
 419         fixup_rt_mutex_waiters(lock);
 420
 421         raw_spin_unlock_irq(&lock->wait_lock);
 422
 423         return cleanup;
 424 }
 425
 426 /*
 427  * Recheck the pi chain, in case we got a priority setting
 428  *
 429  * Called from sched_setscheduler
 430  */
 431 void __sched rt_mutex_adjust_pi(struct task_struct *task)
 432 {
 433         struct rt_mutex_waiter *waiter;
 434         struct rt_mutex_base *next_lock;
 435         unsigned long flags;
 436
 437         raw_spin_lock_irqsave(&task->pi_lock, flags);
 438
 439         waiter = task->pi_blocked_on;
 440         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
 441                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 442                 return;
 443         }
 444         next_lock = waiter->lock;
 445         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 446
 447         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 448         get_task_struct(task);
 449
 450         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
 451                                    next_lock, NULL, task);
 452 }
 453
 454 /*
 455  * Performs the wakeup of the top-waiter and re-enables preemption.
 456  */
 457 void __sched rt_mutex_postunlock(struct rt_wake_q_head *wqh)
 458 {
 459         rt_mutex_wake_up_q(wqh);
 460 }
 461
 462 #ifdef CONFIG_DEBUG_RT_MUTEXES
 463 void rt_mutex_debug_task_free(struct task_struct *task)
 464 {
 465         DEBUG_LOCKS_WARN_ON(!RB_EMPTY_ROOT(&task->pi_waiters.rb_root));
 466         DEBUG_LOCKS_WARN_ON(task->pi_blocked_on);
 467 }
 468 #endif
 469
 470 #ifdef CONFIG_PREEMPT_RT
 471 /* Mutexes */
 472 void __mutex_rt_init(struct mutex *mutex, const char *name,
 473                      struct lock_class_key *key)
 474 {
 475         debug_check_no_locks_freed((void *)mutex, sizeof(*mutex));
 476         lockdep_init_map_wait(&mutex->dep_map, name, key, 0, LD_WAIT_SLEEP);
 477 }
 478 EXPORT_SYMBOL(__mutex_rt_init);
 479
 480 static __always_inline int __mutex_lock_common(struct mutex *lock,
 481                                                unsigned int state,
 482                                                unsigned int subclass,
 483                                                struct lockdep_map *nest_lock,
 484                                                unsigned long ip)
 485 {
 486         int ret;
 487
 488         might_sleep();
 489         mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
 490         ret = __rt_mutex_lock(&lock->rtmutex, state);
 491         if (ret)
 492                 mutex_release(&lock->dep_map, ip);
 493         else
 494                 lock_acquired(&lock->dep_map, ip);
 495         return ret;
 496 }
 497
 498 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 499 void __sched mutex_lock_nested(struct mutex *lock, unsigned int subclass)
 500 {
 501         __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, subclass, NULL, _RET_IP_);
 502 }
 503 EXPORT_SYMBOL_GPL(mutex_lock_nested);
 504
 505 void __sched _mutex_lock_nest_lock(struct mutex *lock,
 506                                    struct lockdep_map *nest_lock)
 507 {
 508         __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0, nest_lock, _RET_IP_);
 509 }
 510 EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
 511
 512 int __sched mutex_lock_interruptible_nested(struct mutex *lock,
 513                                             unsigned int subclass)
 514 {
 515         return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, subclass, NULL, _RET_IP_);
 516 }
 517 EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
 518
 519 int __sched mutex_lock_killable_nested(struct mutex *lock,
 520                                             unsigned int subclass)
 521 {
 522         return __mutex_lock_common(lock, TASK_KILLABLE, subclass, NULL, _RET_IP_);
 523 }
 524 EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
 525
 526 void __sched mutex_lock_io_nested(struct mutex *lock, unsigned int subclass)
 527 {
 528         int token;
 529
 530         might_sleep();
 531
 532         token = io_schedule_prepare();
 533         __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, subclass, NULL, _RET_IP_);
 534         io_schedule_finish(token);
 535 }
 536 EXPORT_SYMBOL_GPL(mutex_lock_io_nested);
 537
 538 #else /* CONFIG_DEBUG_LOCK_ALLOC */
 539
 540 void __sched mutex_lock(struct mutex *lock)
 541 {
 542         __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0, NULL, _RET_IP_);
 543 }
 544 EXPORT_SYMBOL(mutex_lock);
 545
 546 int __sched mutex_lock_interruptible(struct mutex *lock)
 547 {
 548         return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0, NULL, _RET_IP_);
 549 }
 550 EXPORT_SYMBOL(mutex_lock_interruptible);
 551
 552 int __sched mutex_lock_killable(struct mutex *lock)
 553 {
 554         return __mutex_lock_common(lock, TASK_KILLABLE, 0, NULL, _RET_IP_);
 555 }
 556 EXPORT_SYMBOL(mutex_lock_killable);
 557
 558 void __sched mutex_lock_io(struct mutex *lock)
 559 {
 560         int token = io_schedule_prepare();
 561
 562         __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0, NULL, _RET_IP_);
 563         io_schedule_finish(token);
 564 }
 565 EXPORT_SYMBOL(mutex_lock_io);
 566 #endif /* !CONFIG_DEBUG_LOCK_ALLOC */
 567
 568 int __sched mutex_trylock(struct mutex *lock)
 569 {
 570         int ret;
 571
 572         if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES) && WARN_ON_ONCE(!in_task()))
 573                 return 0;
 574
 575         ret = __rt_mutex_trylock(&lock->rtmutex);
 576         if (ret)
 577                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
 578
 579         return ret;
 580 }
 581 EXPORT_SYMBOL(mutex_trylock);
 582
 583 void __sched mutex_unlock(struct mutex *lock)
 584 {
 585         mutex_release(&lock->dep_map, _RET_IP_);
 586         __rt_mutex_unlock(&lock->rtmutex);
 587 }
 588 EXPORT_SYMBOL(mutex_unlock);
 589
 590 #endif /* CONFIG_PREEMPT_RT */