これはLinux Advent Calendar 2019の17日目が空いてしまったので代打で投稿する記事です。

一言でまとめると

Linuxのrtmutexは、

mutexとほぼ同じように使える
ロック待ちが優先度付きになる
SCHED_DEADLINE、SCHED_FIFO, SCHED_RRなどのときだけ優先度付きになる
タスク間で優先度の継承をする
「これを使えばすべて解決」するような万能機能ではない

はじめに

LinuxにはCONFIG_RT_MUTEXESが有効のときに使えるrtmutexがある。これは、mutexロック待ちをするタスクがSCHED_DEADLINE, SCHED_FIFO, SCHED_RRなどのリアルタイムスケジューリングのときに意味を持ち、ロック待ちが優先度付きになり、リアルタイムスケジューリングのタスクが先にロックを取れるようになる。また、リアルタイムのタスクがロック待ちをすると、ロックを持つタスクが優先度の継承を受けて優先度がブーストする。

対比されるのは普通のmutexで、こっちはロック待ちはFirst In First Outの先に来たもの順になる。

下記ではLinux-5.4くらいとglibc-2.30くらいを見ています。

Documentation

普通のmutexについてはここ、

kernel/Documentation/locking/mutex-design.rst
rtmutexの基本的なコンセプトについてはここ、
kernel/Documentation/locking/rt-mutex.rst
rtmutexの詳細実装についてはここ、
kernel/Documentation/locking/rt-mutex-design.rst
rtmutexをfutexへ適用する話についてはここ、
kernel/Documentation/pi-futex.txt
またLWN.netにまとまった記事があるので同様に紹介しておく。
Priority inheritance in the kernel [LWN.net]

良い子のみんなは、以下の私の文章は当てにせずに、ちゃんと上記ドキュメントを読みましょうね。

コードを確認してみる

本当に優先度継承しているのかをコードを見て確認しておく。...

コードのコピペが長いので念のため折りたたんでおく

**rt_mutex_waiter_less()**あたりを見てみると、

kernel/kernel/locking/rtmutex.c

static inline int
rt_mutex_waiter_less(struct rt_mutex_waiter *left,
		     struct rt_mutex_waiter *right)
{
	if (left->prio < right->prio)
		return 1;

	/*
	 * If both waiters have dl_prio(), we check the deadlines of the
	 * associated tasks.
	 * If left waiter has a dl_prio(), and we didn't return 1 above,
	 * then right waiter has a dl_prio() too.
	 */
	if (dl_prio(left->prio))
		return dl_time_before(left->deadline, right->deadline);

	return 0;
}

**rt_mutex_waiter_equal()**あたりは、

kernel/kernel/locking/rtmutex.c

static inline int
rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
		      struct rt_mutex_waiter *right)
{
	if (left->prio != right->prio)
		return 0;

	/*
	 * If both waiters have dl_prio(), we check the deadlines of the
	 * associated tasks.
	 * If left waiter has a dl_prio(), and we didn't return 0 above,
	 * then right waiter has a dl_prio() too.
	 */
	if (dl_prio(left->prio))
		return left->deadline == right->deadline;

	return 1;
}

となっていて、リアルタイムタスクの優先度で並び替えをしている様子がわかる。詳細は省略するけど、これらの比較関数を使ってrbtreeで優先度順にタスクを並び替えるように実装している。そいやちょうどrbtreeの説明記事がったので紹介しておこう。

赤黒木の本質 - Qiita

ドキュメントにも登場しているrt_mutex_adjust_prio_chain()あたりは、長いのを承知で全部載せると、

kernel/kernel/locking/rtmutex.c

/*
 * Adjust the priority chain. Also used for deadlock detection.
 * Decreases task's usage by one - may thus free the task.
 *
 * @task:	the task owning the mutex (owner) for which a chain walk is
 *		probably needed
 * @chwalk:	do we have to carry out deadlock detection?
 * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
 *		things for a task that has just got its priority adjusted, and
 *		is waiting on a mutex)
 * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
 *		we dropped its pi_lock. Is never dereferenced, only used for
 *		comparison to detect lock chain changes.
 * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 *		its priority to the mutex owner (can be NULL in the case
 *		depicted above or if the top waiter is gone away and we are
 *		actually deboosting the owner)
 * @top_task:	the current top waiter
 *
 * Returns 0 or -EDEADLK.
 *
 * Chain walk basics and protection scope
 *
 * [R] refcount on task
 * [P] task->pi_lock held
 * [L] rtmutex->wait_lock held
 *
 * Step	Description				Protected by
 *	function arguments:
 *	@task					[R]
 *	@orig_lock if != NULL			@top_task is blocked on it
 *	@next_lock				Unprotected. Cannot be
 *						dereferenced. Only used for
 *						comparison.
 *	@orig_waiter if != NULL			@top_task is blocked on it
 *	@top_task				current, or in case of proxy
 *						locking protected by calling
 *						code
 *	again:
 *	  loop_sanity_check();
 *	retry:
 * [1]	  lock(task->pi_lock);			[R] acquire [P]
 * [2]	  waiter = task->pi_blocked_on;		[P]
 * [3]	  check_exit_conditions_1();		[P]
 * [4]	  lock = waiter->lock;			[P]
 * [5]	  if (!try_lock(lock->wait_lock)) {	[P] try to acquire [L]
 *	    unlock(task->pi_lock);		release [P]
 *	    goto retry;
 *	  }
 * [6]	  check_exit_conditions_2();		[P] + [L]
 * [7]	  requeue_lock_waiter(lock, waiter);	[P] + [L]
 * [8]	  unlock(task->pi_lock);		release [P]
 *	  put_task_struct(task);		release [R]
 * [9]	  check_exit_conditions_3();		[L]
 * [10]	  task = owner(lock);			[L]
 *	  get_task_struct(task);		[L] acquire [R]
 *	  lock(task->pi_lock);			[L] acquire [P]
 * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 * [12]	  check_exit_conditions_4();		[P] + [L]
 * [13]	  unlock(task->pi_lock);		release [P]
 *	  unlock(lock->wait_lock);		release [L]
 *	  goto again;
 */
static int rt_mutex_adjust_prio_chain(struct task_struct *task,
				      enum rtmutex_chainwalk chwalk,
				      struct rt_mutex *orig_lock,
				      struct rt_mutex *next_lock,
				      struct rt_mutex_waiter *orig_waiter,
				      struct task_struct *top_task)
{
	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
	struct rt_mutex_waiter *prerequeue_top_waiter;
	int ret = 0, depth = 0;
	struct rt_mutex *lock;
	bool detect_deadlock;
	bool requeue = true;

	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);

	/*
	 * The (de)boosting is a step by step approach with a lot of
	 * pitfalls. We want this to be preemptible and we want hold a
	 * maximum of two locks per step. So we have to check
	 * carefully whether things change under us.
	 */
 again:
	/*
	 * We limit the lock chain length for each invocation.
	 */
	if (++depth > max_lock_depth) {
		static int prev_max;

		/*
		 * Print this only once. If the admin changes the limit,
		 * print a new message when reaching the limit again.
		 */
		if (prev_max != max_lock_depth) {
			prev_max = max_lock_depth;
			printk(KERN_WARNING "Maximum lock depth %d reached "
			       "task: %s (%d)\n", max_lock_depth,
			       top_task->comm, task_pid_nr(top_task));
		}
		put_task_struct(task);

		return -EDEADLK;
	}

	/*
	 * We are fully preemptible here and only hold the refcount on
	 * @task. So everything can have changed under us since the
	 * caller or our own code below (goto retry/again) dropped all
	 * locks.
	 */
 retry:
	/*
	 * [1] Task cannot go away as we did a get_task() before !
	 */
	raw_spin_lock_irq(&task->pi_lock);

	/*
	 * [2] Get the waiter on which @task is blocked on.
	 */
	waiter = task->pi_blocked_on;

	/*
	 * [3] check_exit_conditions_1() protected by task->pi_lock.
	 */

	/*
	 * Check whether the end of the boosting chain has been
	 * reached or the state of the chain has changed while we
	 * dropped the locks.
	 */
	if (!waiter)
		goto out_unlock_pi;

	/*
	 * Check the orig_waiter state. After we dropped the locks,
	 * the previous owner of the lock might have released the lock.
	 */
	if (orig_waiter && !rt_mutex_owner(orig_lock))
		goto out_unlock_pi;

	/*
	 * We dropped all locks after taking a refcount on @task, so
	 * the task might have moved on in the lock chain or even left
	 * the chain completely and blocks now on an unrelated lock or
	 * on @orig_lock.
	 *
	 * We stored the lock on which @task was blocked in @next_lock,
	 * so we can detect the chain change.
	 */
	if (next_lock != waiter->lock)
		goto out_unlock_pi;

	/*
	 * Drop out, when the task has no waiters. Note,
	 * top_waiter can be NULL, when we are in the deboosting
	 * mode!
	 */
	if (top_waiter) {
		if (!task_has_pi_waiters(task))
			goto out_unlock_pi;
		/*
		 * If deadlock detection is off, we stop here if we
		 * are not the top pi waiter of the task. If deadlock
		 * detection is enabled we continue, but stop the
		 * requeueing in the chain walk.
		 */
		if (top_waiter != task_top_pi_waiter(task)) {
			if (!detect_deadlock)
				goto out_unlock_pi;
			else
				requeue = false;
		}
	}

	/*
	 * If the waiter priority is the same as the task priority
	 * then there is no further priority adjustment necessary.  If
	 * deadlock detection is off, we stop the chain walk. If its
	 * enabled we continue, but stop the requeueing in the chain
	 * walk.
	 */
	if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
		if (!detect_deadlock)
			goto out_unlock_pi;
		else
			requeue = false;
	}

	/*
	 * [4] Get the next lock
	 */
	lock = waiter->lock;
	/*
	 * [5] We need to trylock here as we are holding task->pi_lock,
	 * which is the reverse lock order versus the other rtmutex
	 * operations.
	 */
	if (!raw_spin_trylock(&lock->wait_lock)) {
		raw_spin_unlock_irq(&task->pi_lock);
		cpu_relax();
		goto retry;
	}

	/*
	 * [6] check_exit_conditions_2() protected by task->pi_lock and
	 * lock->wait_lock.
	 *
	 * Deadlock detection. If the lock is the same as the original
	 * lock which caused us to walk the lock chain or if the
	 * current lock is owned by the task which initiated the chain
	 * walk, we detected a deadlock.
	 */
	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
		debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
		raw_spin_unlock(&lock->wait_lock);
		ret = -EDEADLK;
		goto out_unlock_pi;
	}

	/*
	 * If we just follow the lock chain for deadlock detection, no
	 * need to do all the requeue operations. To avoid a truckload
	 * of conditionals around the various places below, just do the
	 * minimum chain walk checks.
	 */
	if (!requeue) {
		/*
		 * No requeue[7] here. Just release @task [8]
		 */
		raw_spin_unlock(&task->pi_lock);
		put_task_struct(task);

		/*
		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
		 * If there is no owner of the lock, end of chain.
		 */
		if (!rt_mutex_owner(lock)) {
			raw_spin_unlock_irq(&lock->wait_lock);
			return 0;
		}

		/* [10] Grab the next task, i.e. owner of @lock */
		task = get_task_struct(rt_mutex_owner(lock));
		raw_spin_lock(&task->pi_lock);

		/*
		 * No requeue [11] here. We just do deadlock detection.
		 *
		 * [12] Store whether owner is blocked
		 * itself. Decision is made after dropping the locks
		 */
		next_lock = task_blocked_on_lock(task);
		/*
		 * Get the top waiter for the next iteration
		 */
		top_waiter = rt_mutex_top_waiter(lock);

		/* [13] Drop locks */
		raw_spin_unlock(&task->pi_lock);
		raw_spin_unlock_irq(&lock->wait_lock);

		/* If owner is not blocked, end of chain. */
		if (!next_lock)
			goto out_put_task;
		goto again;
	}

	/*
	 * Store the current top waiter before doing the requeue
	 * operation on @lock. We need it for the boost/deboost
	 * decision below.
	 */
	prerequeue_top_waiter = rt_mutex_top_waiter(lock);

	/* [7] Requeue the waiter in the lock waiter tree. */
	rt_mutex_dequeue(lock, waiter);

	/*
	 * Update the waiter prio fields now that we're dequeued.
	 *
	 * These values can have changed through either:
	 *
	 *   sys_sched_set_scheduler() / sys_sched_setattr()
	 *
	 * or
	 *
	 *   DL CBS enforcement advancing the effective deadline.
	 *
	 * Even though pi_waiters also uses these fields, and that tree is only
	 * updated in [11], we can do this here, since we hold [L], which
	 * serializes all pi_waiters access and rb_erase() does not care about
	 * the values of the node being removed.
	 */
	waiter->prio = task->prio;
	waiter->deadline = task->dl.deadline;

	rt_mutex_enqueue(lock, waiter);

	/* [8] Release the task */
	raw_spin_unlock(&task->pi_lock);
	put_task_struct(task);

	/*
	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
	 *
	 * We must abort the chain walk if there is no lock owner even
	 * in the dead lock detection case, as we have nothing to
	 * follow here. This is the end of the chain we are walking.
	 */
	if (!rt_mutex_owner(lock)) {
		/*
		 * If the requeue [7] above changed the top waiter,
		 * then we need to wake the new top waiter up to try
		 * to get the lock.
		 */
		if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
			wake_up_process(rt_mutex_top_waiter(lock)->task);
		raw_spin_unlock_irq(&lock->wait_lock);
		return 0;
	}

	/* [10] Grab the next task, i.e. the owner of @lock */
	task = get_task_struct(rt_mutex_owner(lock));
	raw_spin_lock(&task->pi_lock);

	/* [11] requeue the pi waiters if necessary */
	if (waiter == rt_mutex_top_waiter(lock)) {
		/*
		 * The waiter became the new top (highest priority)
		 * waiter on the lock. Replace the previous top waiter
		 * in the owner tasks pi waiters tree with this waiter
		 * and adjust the priority of the owner.
		 */
		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
		rt_mutex_enqueue_pi(task, waiter);
		rt_mutex_adjust_prio(task);

	} else if (prerequeue_top_waiter == waiter) {
		/*
		 * The waiter was the top waiter on the lock, but is
		 * no longer the top prority waiter. Replace waiter in
		 * the owner tasks pi waiters tree with the new top
		 * (highest priority) waiter and adjust the priority
		 * of the owner.
		 * The new top waiter is stored in @waiter so that
		 * @waiter == @top_waiter evaluates to true below and
		 * we continue to deboost the rest of the chain.
		 */
		rt_mutex_dequeue_pi(task, waiter);
		waiter = rt_mutex_top_waiter(lock);
		rt_mutex_enqueue_pi(task, waiter);
		rt_mutex_adjust_prio(task);
	} else {
		/*
		 * Nothing changed. No need to do any priority
		 * adjustment.
		 */
	}

	/*
	 * [12] check_exit_conditions_4() protected by task->pi_lock
	 * and lock->wait_lock. The actual decisions are made after we
	 * dropped the locks.
	 *
	 * Check whether the task which owns the current lock is pi
	 * blocked itself. If yes we store a pointer to the lock for
	 * the lock chain change detection above. After we dropped
	 * task->pi_lock next_lock cannot be dereferenced anymore.
	 */
	next_lock = task_blocked_on_lock(task);
	/*
	 * Store the top waiter of @lock for the end of chain walk
	 * decision below.
	 */
	top_waiter = rt_mutex_top_waiter(lock);

	/* [13] Drop the locks */
	raw_spin_unlock(&task->pi_lock);
	raw_spin_unlock_irq(&lock->wait_lock);

	/*
	 * Make the actual exit decisions [12], based on the stored
	 * values.
	 *
	 * We reached the end of the lock chain. Stop right here. No
	 * point to go back just to figure that out.
	 */
	if (!next_lock)
		goto out_put_task;

	/*
	 * If the current waiter is not the top waiter on the lock,
	 * then we can stop the chain walk here if we are not in full
	 * deadlock detection mode.
	 */
	if (!detect_deadlock && waiter != top_waiter)
		goto out_put_task;

	goto again;

 out_unlock_pi:
	raw_spin_unlock_irq(&task->pi_lock);
 out_put_task:
	put_task_struct(task);

	return ret;
}

となっていて、優先度の継承をいろんな条件を網羅しながら苦労して実装している様子がわかる。またrt_mutex_setprio()は、CONFIG_RT_MUTEXESが有効な場合に実装が分かれていて、これも長いけど同様に全部載せると、

kernel/sched/core.c

# ifdef CONFIG_RT_MUTEXES

static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
{
	if (pi_task)
		prio = min(prio, pi_task->prio);

	return prio;
}

static inline int rt_effective_prio(struct task_struct *p, int prio)
{
	struct task_struct *pi_task = rt_mutex_get_top_task(p);

	return __rt_effective_prio(pi_task, prio);
}

/*
 * rt_mutex_setprio - set the current priority of a task
 * @p: task to boost
 * @pi_task: donor task
 *
 * This function changes the 'effective' priority of a task. It does
 * not touch ->normal_prio like __setscheduler().
 *
 * Used by the rt_mutex code to implement priority inheritance
 * logic. Call site only calls if the priority of the task changed.
 */
void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
{
	int prio, oldprio, queued, running, queue_flag =
		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
	const struct sched_class *prev_class;
	struct rq_flags rf;
	struct rq *rq;

	/* XXX used to be waiter->prio, not waiter->task->prio */
	prio = __rt_effective_prio(pi_task, p->normal_prio);

	/*
	 * If nothing changed; bail early.
	 */
	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
		return;

	rq = __task_rq_lock(p, &rf);
	update_rq_clock(rq);
	/*
	 * Set under pi_lock && rq->lock, such that the value can be used under
	 * either lock.
	 *
	 * Note that there is loads of tricky to make this pointer cache work
	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
	 * ensure a task is de-boosted (pi_task is set to NULL) before the
	 * task is allowed to run again (and can exit). This ensures the pointer
	 * points to a blocked task -- which guaratees the task is present.
	 */
	p->pi_top_task = pi_task;

	/*
	 * For FIFO/RR we only need to set prio, if that matches we're done.
	 */
	if (prio == p->prio && !dl_prio(prio))
		goto out_unlock;

	/*
	 * Idle task boosting is a nono in general. There is one
	 * exception, when PREEMPT_RT and NOHZ is active:
	 *
	 * The idle task calls get_next_timer_interrupt() and holds
	 * the timer wheel base->lock on the CPU and another CPU wants
	 * to access the timer (probably to cancel it). We can safely
	 * ignore the boosting request, as the idle CPU runs this code
	 * with interrupts disabled and will complete the lock
	 * protected section without being interrupted. So there is no
	 * real need to boost.
	 */
	if (unlikely(p == rq->idle)) {
		WARN_ON(p != rq->curr);
		WARN_ON(p->pi_blocked_on);
		goto out_unlock;
	}

	trace_sched_pi_setprio(p, pi_task);
	oldprio = p->prio;

	if (oldprio == prio)
		queue_flag &= ~DEQUEUE_MOVE;

	prev_class = p->sched_class;
	queued = task_on_rq_queued(p);
	running = task_current(rq, p);
	if (queued)
		dequeue_task(rq, p, queue_flag);
	if (running)
		put_prev_task(rq, p);

	/*
	 * Boosting condition are:
	 * 1. -rt task is running and holds mutex A
	 *      --> -dl task blocks on mutex A
	 *
	 * 2. -dl task is running and holds mutex A
	 *      --> -dl task blocks on mutex A and could preempt the
	 *          running task
	 */
	if (dl_prio(prio)) {
		if (!dl_prio(p->normal_prio) ||
		    (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
			p->dl.dl_boosted = 1;
			queue_flag |= ENQUEUE_REPLENISH;
		} else
			p->dl.dl_boosted = 0;
		p->sched_class = &dl_sched_class;
	} else if (rt_prio(prio)) {
		if (dl_prio(oldprio))
			p->dl.dl_boosted = 0;
		if (oldprio < prio)
			queue_flag |= ENQUEUE_HEAD;
		p->sched_class = &rt_sched_class;
	} else {
		if (dl_prio(oldprio))
			p->dl.dl_boosted = 0;
		if (rt_prio(oldprio))
			p->rt.timeout = 0;
		p->sched_class = &fair_sched_class;
	}

	p->prio = prio;

	if (queued)
		enqueue_task(rq, p, queue_flag);
	if (running)
		set_next_task(rq, p);

	check_class_changed(rq, p, prev_class, oldprio);
out_unlock:
	/* Avoid rq from going away on us: */
	preempt_disable();
	__task_rq_unlock(rq, &rf);

	balance_callback(rq);
	preempt_enable();
}。

となっていて、実際にタスクの優先度を変更してリスケしている様子がわかる。

FUTEX_PI

Linux-v4.14からCONFIG_FUTEX_PIというのが登場していて、優先度継承のユーザランドへの開放は4.14から...と思っていたらそうではなく、これはこのcommitによりLinux-2.6.18くらいからすでに入っているようだった。CONFIG_FUTEX_PIは、すでに実装された機能をdisableにできるようにするため導入されたようだ。

もう少し具体的にコードを見ていく。kernel/kernel/futex.cのdo_futex()より

kernel/kernel/futex.c

	switch (cmd) {
	case FUTEX_WAIT:
		val3 = FUTEX_BITSET_MATCH_ANY;
		/* fall through */
	case FUTEX_WAIT_BITSET:
		return futex_wait(uaddr, flags, val, timeout, val3);
	case FUTEX_WAKE:
		val3 = FUTEX_BITSET_MATCH_ANY;
		/* fall through */
	case FUTEX_WAKE_BITSET:
		return futex_wake(uaddr, flags, val, val3);
	case FUTEX_REQUEUE:
		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
	case FUTEX_CMP_REQUEUE:
		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
	case FUTEX_WAKE_OP:
		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
	case FUTEX_LOCK_PI:
		return futex_lock_pi(uaddr, flags, timeout, 0);
	case FUTEX_UNLOCK_PI:
		return futex_unlock_pi(uaddr, flags);
	case FUTEX_TRYLOCK_PI:
		return futex_lock_pi(uaddr, flags, NULL, 1);
	case FUTEX_WAIT_REQUEUE_PI:
		val3 = FUTEX_BITSET_MATCH_ANY;
		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
					     uaddr2);
	case FUTEX_CMP_REQUEUE_PI:
		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
	}

となっていて、HOGEHOGE_PI(例えばFUTEX_LOCK_PI)をつけてfutexシステムコールするとfutex_piなコードへつながる。細かくは省略するけど、hogehoge_pi()な関数は中でrt_mutex()を使っている。

肝心のFUTEX_LOCK_PIを使っているユーザランドのコードはというと、glibc/nptl/pthread_mutex_lock.cの**__pthread_mutex_lock_full()**より、

glibc/nptl/pthread_mutex_lock.c

    case PTHREAD_MUTEX_PI_RECURSIVE_NP:
    case PTHREAD_MUTEX_PI_ERRORCHECK_NP:
    case PTHREAD_MUTEX_PI_NORMAL_NP:
    case PTHREAD_MUTEX_PI_ADAPTIVE_NP:
    case PTHREAD_MUTEX_PI_ROBUST_RECURSIVE_NP:
    case PTHREAD_MUTEX_PI_ROBUST_ERRORCHECK_NP:
    case PTHREAD_MUTEX_PI_ROBUST_NORMAL_NP:
    case PTHREAD_MUTEX_PI_ROBUST_ADAPTIVE_NP:

glibc/nptl/pthread_mutex_lock.c

    if (oldval != 0)
    {
        /* The mutex is locked.  The kernel will now take care of
        everything.  */
        int private = (robust
                       ? PTHREAD_ROBUST_MUTEX_PSHARED (mutex)
                       : PTHREAD_MUTEX_PSHARED (mutex));                                                      
        INTERNAL_SYSCALL_DECL (__err);                                                                        
        int e = INTERNAL_SYSCALL (futex, __err, 4, &mutex->__data.__lock,
                                 __lll_private_flag (FUTEX_LOCK_PI,
                                                     private), 1, 0);

となっていて、PTHREAD_MUTEX_PI_NORMAL_NPなどのタイプの時に使われる。PTHREAD_MUTEX_PI_NORMAL_NPは、glibc/nptl/pthreadP.hより、

glibc/nptl/pthreadP.h

enum
{
  PTHREAD_MUTEX_KIND_MASK_NP = 3,

  PTHREAD_MUTEX_ELISION_NP    = 256,
  PTHREAD_MUTEX_NO_ELISION_NP = 512,

  PTHREAD_MUTEX_ROBUST_NORMAL_NP = 16,
  PTHREAD_MUTEX_ROBUST_RECURSIVE_NP
  = PTHREAD_MUTEX_ROBUST_NORMAL_NP | PTHREAD_MUTEX_RECURSIVE_NP,
  PTHREAD_MUTEX_ROBUST_ERRORCHECK_NP
  = PTHREAD_MUTEX_ROBUST_NORMAL_NP | PTHREAD_MUTEX_ERRORCHECK_NP,
  PTHREAD_MUTEX_ROBUST_ADAPTIVE_NP
  = PTHREAD_MUTEX_ROBUST_NORMAL_NP | PTHREAD_MUTEX_ADAPTIVE_NP,
  PTHREAD_MUTEX_PRIO_INHERIT_NP = 32,
  PTHREAD_MUTEX_PI_NORMAL_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_NORMAL,
  PTHREAD_MUTEX_PI_RECURSIVE_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_RECURSIVE_NP,
  PTHREAD_MUTEX_PI_ERRORCHECK_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ERRORCHECK_NP,
  PTHREAD_MUTEX_PI_ADAPTIVE_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ADAPTIVE_NP,
  PTHREAD_MUTEX_PI_ROBUST_NORMAL_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ROBUST_NORMAL_NP,
  PTHREAD_MUTEX_PI_ROBUST_RECURSIVE_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ROBUST_RECURSIVE_NP,
  PTHREAD_MUTEX_PI_ROBUST_ERRORCHECK_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ROBUST_ERRORCHECK_NP,
  PTHREAD_MUTEX_PI_ROBUST_ADAPTIVE_NP
  = PTHREAD_MUTEX_PRIO_INHERIT_NP | PTHREAD_MUTEX_ROBUST_ADAPTIVE_NP,
  PTHREAD_MUTEX_PRIO_PROTECT_NP = 64,
  PTHREAD_MUTEX_PP_NORMAL_NP
  = PTHREAD_MUTEX_PRIO_PROTECT_NP | PTHREAD_MUTEX_NORMAL,
  PTHREAD_MUTEX_PP_RECURSIVE_NP
  = PTHREAD_MUTEX_PRIO_PROTECT_NP | PTHREAD_MUTEX_RECURSIVE_NP,
  PTHREAD_MUTEX_PP_ERRORCHECK_NP
  = PTHREAD_MUTEX_PRIO_PROTECT_NP | PTHREAD_MUTEX_ERRORCHECK_NP,
  PTHREAD_MUTEX_PP_ADAPTIVE_NP
  = PTHREAD_MUTEX_PRIO_PROTECT_NP | PTHREAD_MUTEX_ADAPTIVE_NP,
  PTHREAD_MUTEX_ELISION_FLAGS_NP
  = PTHREAD_MUTEX_ELISION_NP | PTHREAD_MUTEX_NO_ELISION_NP,

  PTHREAD_MUTEX_TIMED_ELISION_NP =
  PTHREAD_MUTEX_TIMED_NP | PTHREAD_MUTEX_ELISION_NP,
  PTHREAD_MUTEX_TIMED_NO_ELISION_NP =
  PTHREAD_MUTEX_TIMED_NP | PTHREAD_MUTEX_NO_ELISION_NP,
};                                                                                                    
# define PTHREAD_MUTEX_PSHARED_BIT 128

となっていて、ちょっとわかりにくいけど、端的には、PTHREAD_MUTEX_PRIO_INHERIT_NPつきのmutexかどうかで決まる。PTHREAD_MUTEX_PRIO_INHERIT_NPについては、man pthread_mutexattr_setprotocol(3)に詳しく書かれている。要点だけを書くと、

PTHREAD_PRIO_NONEは、なにもしない(デフォルトで選ばれる)
PTHREAD_PRIO_INHERITは、優先度継承する。
PTHREAD_PRIO_PROTECTは、あらかじめシーリングした値までlock取るタイミングで優先度が上がる。

シーリング値の設定はpthread_mutexattr_setprioceiling(3)で行う。

というわけで、pthread_mutexattr_tを適切に使うことで、Linuxのrtmutexによる優先度継承をユーザランドでも利用できることがわかった。

PIは何の略称？(追記)

(2020/01/01追記)
「PI」という略称が登場するけど、これは、Documentationに記載のある通り、Priority Inheritance(優先度継承)の略称です。

ポエム: Linuxにおけるリアルタイムシステムとは(その2)

まずは2017年4月頃に書いた以前のポエムをどうぞ。

Linuxのkernel内ではrtmutexではないふつうのmutexが非常に多いので、原則論でいくなら、リアルタイムのタスクがkernelを動き回るだけでアウトとなる。とはいえ、一部のmutexを除いて、それほど長時間保持されるものでもないので、CPU使用率常に高い状態とかでもない限りは、そこまで足を引っ張るわけではないんじゃないかと思う。

CONFIG_IRQ_FORCED_THREADINGが有効の場合、cmdlineにthreadirqsを足すことで、割り込み登録したエントリ関数がSCHED_FIFOの50の割り込みスレッドで駆動するようになる。割り込みより優先したいタスクがある場合にprio順で扱えるようになる。ただ、オーバヘッドは増えるので、これを使わなくても済むような設計がよいのは言うまでもない。

Linux-4.4からMLOCK_ONFAULTが登場し、mlock2(2)やmlockall(2)がちょっと便利になった。依存するライブラリや関数をコントロールしきれない場合の足しになるかもしれない。ただ、不必要なライブラリや関数を減らして、mlock()すべき領域を精査したほうが良いのは言うまでもない。

今回はrtmutexを紹介しているわけだけど、mutexの待ち時間や待ち行列が気になる時点でふつうはリアルタイム処理がすでに破綻していることがほとんどだと思う。rtmutexに頼ろうとしている時点でもはや設計しなおしが必要なほど手遅れなんだと思っておいたほうが良い。

私が関わり始めた＊年前は「Linuxなんかでリアルタイム処理なんてできるわけがない」などとよく言われたものだけど、そこそこのレベルでよければ、Linuxの機能強化とともに、上記のような対応をすることでリアルタイム志向のシステム構築がしやすくなってきたかと思う。

一方で、ALSAやgstreamerなど、本来ならリアルタイム処理が必要なものが、重層なライブラリとたくさんのスレッドのもとで富豪的に動かされているというのが現在の実装の現実だったりもする。「最低256MBはほしい」みたいな話もあり、大規模組み込みLinuxをやってきた立場からは違和感みたいなのを感じつつも、ラズパイみたいなSoCも最近は安いし、下手するとDRAM/eMMCのほうが高かったりするし、安いハードを使って富豪的プログラミングしラピッド開発するというのが一つの時代の流れなのかもしれない。

あとがき

調べるきっかけになったとある案件では、結果的にrtmutexで救われるようなケースだったので助かったけど、一般的には「rtmutexを使えば解決！」というケースは少ないんじゃないかと思う。独自ドライバで独自タスク管理を行うようなケースくらいじゃないだろうか。また、FUTEX_PIを使えばユーザランドのpthread_mutexでも優先度継承ができるとはいえ、やはりそれに頼るというのもなにか間違っている気がする。

最近はじっくり設計するような機会がめっきり減ったので、こういう小手先で問題をごまかすようなやり方に手を染めつつあって、違和感を覚えつつもやもやするおっさんがお届けしました。

参考サイト