Re: [PATCH v7 16/23] sched: Add deactivated (sleeping) owner handling to find_proxy_task()

[Metin Kaya <metin.kaya@arm.com>]

On 20/12/2023 12:18 am, John Stultz wrote:
> From: Peter Zijlstra <peterz@infradead.org>
> 
> If the blocked_on chain resolves to a sleeping owner, deactivate
> selected task, and enqueue it on the sleeping owner task.
> Then re-activate it later when the owner is woken up.
> 
> NOTE: This has been particularly challenging to get working
> properly, and some of the locking is particularly ackward. I'd

                                                     awkward?

> very much appreciate review and feedback for ways to simplify
> this.
> 
> Cc: Joel Fernandes <joelaf@google.com>
> Cc: Qais Yousef <qyousef@google.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
> Cc: Valentin Schneider <vschneid@redhat.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Ben Segall <bsegall@google.com>
> Cc: Zimuzo Ezeozue <zezeozue@google.com>
> Cc: Youssef Esmat <youssefesmat@google.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
> Cc: Will Deacon <will@kernel.org>
> Cc: Waiman Long <longman@redhat.com>
> Cc: Boqun Feng <boqun.feng@gmail.com>
> Cc: "Paul E. McKenney" <paulmck@kernel.org>
> Cc: Metin Kaya <Metin.Kaya@arm.com>
> Cc: Xuewen Yan <xuewen.yan94@gmail.com>
> Cc: K Prateek Nayak <kprateek.nayak@amd.com>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: kernel-team@android.com
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Signed-off-by: Juri Lelli <juri.lelli@redhat.com>
> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
> Signed-off-by: Connor O'Brien <connoro@google.com>
> [jstultz: This was broken out from the larger proxy() patch]
> Signed-off-by: John Stultz <jstultz@google.com>
> ---
> v5:
> * Split out from larger proxy patch
> v6:
> * Major rework, replacing the single list head per task with
>    per-task list head and nodes, creating a tree structure so
>    we only wake up decendents of the task woken.

                      descendants

> * Reworked the locking to take the task->pi_lock, so we can
>    avoid mid-chain wakeup races from try_to_wake_up() called by
>    the ww_mutex logic.
> v7:
> * Drop ununessary __nested lock annotation, as we already drop

          unnecessary

>    the lock prior.
> * Add comments on #else & #endif lines, and clearer function
>    names, and commit message tweaks as suggested by Metin Kaya
> * Move activate_blocked_entities() call from ttwu_queue to
>    try_to_wake_up() to simplify locking. Thanks to questions from
>    Metin Kaya
> * Fix irqsave/irqrestore usage now we call this outside where
>    the pi_lock is held
> * Fix activate_blocked_entitites not preserving wake_cpu
> * Fix for UP builds
> ---
>   include/linux/sched.h |   3 +
>   kernel/fork.c         |   4 +
>   kernel/sched/core.c   | 214 ++++++++++++++++++++++++++++++++++++++----
>   3 files changed, 202 insertions(+), 19 deletions(-)
> 
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 8020e224e057..6f982948a105 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1158,6 +1158,9 @@ struct task_struct {
>   	enum blocked_on_state		blocked_on_state;
>   	struct mutex			*blocked_on;	/* lock we're blocked on */
>   	struct task_struct		*blocked_donor;	/* task that is boosting this task */
> +	struct list_head		blocked_head;  /* tasks blocked on this task */
> +	struct list_head		blocked_node;  /* our entry on someone elses blocked_head */
> +	struct task_struct		*sleeping_owner; /* task our blocked_node is enqueued on */
>   	raw_spinlock_t			blocked_lock;

As we discussed off-list, there are now 7 fields introduced in 
task_struct because of Proxy Execution support. I recommended to move 
them into a container struct to save task_struct from a slight pollution 
(similarly initialization of these fields in copy_process() can be moved 
into a separate function), but you mentioned fair points about different 
locks for different fields.

I wonder what other folks think about this minor concern.

>   
>   #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
> diff --git a/kernel/fork.c b/kernel/fork.c
> index 138fc23cad43..56f5e19c268e 100644
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -2460,6 +2460,10 @@ __latent_entropy struct task_struct *copy_process(
>   	p->blocked_on_state = BO_RUNNABLE;
>   	p->blocked_on = NULL; /* not blocked yet */
>   	p->blocked_donor = NULL; /* nobody is boosting p yet */
> +
> +	INIT_LIST_HEAD(&p->blocked_head);
> +	INIT_LIST_HEAD(&p->blocked_node);
> +	p->sleeping_owner = NULL;
>   #ifdef CONFIG_BCACHE
>   	p->sequential_io	= 0;
>   	p->sequential_io_avg	= 0;
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index e0afa228bc9d..0cd63bd0bdcd 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3785,6 +3785,133 @@ static inline void ttwu_do_wakeup(struct task_struct *p)
>   	trace_sched_wakeup(p);
>   }
>   
> +#ifdef CONFIG_SCHED_PROXY_EXEC
> +static void do_activate_task(struct rq *rq, struct task_struct *p, int en_flags)
> +{
> +	lockdep_assert_rq_held(rq);
> +
> +	if (!sched_proxy_exec()) {
> +		activate_task(rq, p, en_flags);
> +		return;
> +	}
> +
> +	if (p->sleeping_owner) {
> +		struct task_struct *owner = p->sleeping_owner;
> +
> +		raw_spin_lock(&owner->blocked_lock);
> +		list_del_init(&p->blocked_node);
> +		p->sleeping_owner = NULL;
> +		raw_spin_unlock(&owner->blocked_lock);
> +	}
> +
> +	/*
> +	 * By calling activate_task with blocked_lock held, we
> +	 * order against the find_proxy_task() blocked_task case
> +	 * such that no more blocked tasks will be enqueued on p
> +	 * once we release p->blocked_lock.
> +	 */
> +	raw_spin_lock(&p->blocked_lock);
> +	WARN_ON(task_cpu(p) != cpu_of(rq));
> +	activate_task(rq, p, en_flags);
> +	raw_spin_unlock(&p->blocked_lock);
> +}
> +
> +#ifdef CONFIG_SMP
> +static inline void proxy_set_task_cpu(struct task_struct *p, int cpu)
> +{
> +	unsigned int wake_cpu;
> +
> +	/* Preserve wake_cpu */
> +	wake_cpu = p->wake_cpu;
> +	__set_task_cpu(p, cpu);
> +	p->wake_cpu = wake_cpu;
> +}
> +#else /* !CONFIG_SMP */
> +static inline void proxy_set_task_cpu(struct task_struct *p, int cpu)
> +{
> +	__set_task_cpu(p, cpu);
> +}
> +#endif /* CONFIG_SMP */
> +
> +static void activate_blocked_entities(struct rq *target_rq,
> +				      struct task_struct *owner,
> +				      int wake_flags)
> +{
> +	unsigned long flags;
> +	struct rq_flags rf;
> +	int target_cpu = cpu_of(target_rq);
> +	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
> +
> +	if (wake_flags & WF_MIGRATED)
> +		en_flags |= ENQUEUE_MIGRATED;
> +	/*
> +	 * A whole bunch of 'proxy' tasks back this blocked task, wake
> +	 * them all up to give this task its 'fair' share.
> +	 */
> +	raw_spin_lock_irqsave(&owner->blocked_lock, flags);
> +	while (!list_empty(&owner->blocked_head)) {
> +		struct task_struct *pp;
> +		unsigned int state;
> +
> +		pp = list_first_entry(&owner->blocked_head,
> +				      struct task_struct,
> +				      blocked_node);
> +		BUG_ON(pp == owner);
> +		list_del_init(&pp->blocked_node);
> +		WARN_ON(!pp->sleeping_owner);
> +		pp->sleeping_owner = NULL;
> +		raw_spin_unlock_irqrestore(&owner->blocked_lock, flags);
> +
> +		raw_spin_lock_irqsave(&pp->pi_lock, flags);
> +		state = READ_ONCE(pp->__state);
> +		/* Avoid racing with ttwu */
> +		if (state == TASK_WAKING) {
> +			raw_spin_unlock_irqrestore(&pp->pi_lock, flags);
> +			raw_spin_lock_irqsave(&owner->blocked_lock, flags);
> +			continue;
> +		}
> +		if (READ_ONCE(pp->on_rq)) {
> +			/*
> +			 * We raced with a non mutex handoff activation of pp.
> +			 * That activation will also take care of activating
> +			 * all of the tasks after pp in the blocked_entry list,
> +			 * so we're done here.
> +			 */
> +			raw_spin_unlock_irqrestore(&pp->pi_lock, flags);
> +			raw_spin_lock_irqsave(&owner->blocked_lock, flags);
> +			continue;
> +		}
> +
> +		proxy_set_task_cpu(pp, target_cpu);
> +
> +		rq_lock_irqsave(target_rq, &rf);
> +		update_rq_clock(target_rq);
> +		do_activate_task(target_rq, pp, en_flags);
> +		resched_curr(target_rq);
> +		rq_unlock_irqrestore(target_rq, &rf);
> +		raw_spin_unlock_irqrestore(&pp->pi_lock, flags);
> +
> +		/* recurse - XXX This needs to be reworked to avoid recursing */
> +		activate_blocked_entities(target_rq, pp, wake_flags);
> +
> +		raw_spin_lock_irqsave(&owner->blocked_lock, flags);
> +	}
> +	raw_spin_unlock_irqrestore(&owner->blocked_lock, flags);
> +}
> +#else /* !CONFIG_SCHED_PROXY_EXEC */
> +static inline void do_activate_task(struct rq *rq, struct task_struct *p,
> +				    int en_flags)
> +{
> +	activate_task(rq, p, en_flags);
> +}
> +
> +static inline void activate_blocked_entities(struct rq *target_rq,
> +					     struct task_struct *owner,
> +					     int wake_flags)
> +{
> +}
> +#endif /* CONFIG_SCHED_PROXY_EXEC */
> +
>   #ifdef CONFIG_SMP
>   static inline bool proxy_needs_return(struct rq *rq, struct task_struct *p)
>   {
> @@ -3839,7 +3966,7 @@ ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
>   		atomic_dec(&task_rq(p)->nr_iowait);
>   	}
>   
> -	activate_task(rq, p, en_flags);
> +	do_activate_task(rq, p, en_flags);
>   	wakeup_preempt(rq, p, wake_flags);
>   
>   	ttwu_do_wakeup(p);
> @@ -3936,13 +4063,19 @@ void sched_ttwu_pending(void *arg)
>   	update_rq_clock(rq);
>   
>   	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
> +		int wake_flags;
>   		if (WARN_ON_ONCE(p->on_cpu))
>   			smp_cond_load_acquire(&p->on_cpu, !VAL);
>   
>   		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
>   			set_task_cpu(p, cpu_of(rq));
>   
> -		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
> +		wake_flags = p->sched_remote_wakeup ? WF_MIGRATED : 0;
> +		ttwu_do_activate(rq, p, wake_flags, &rf);
> +		rq_unlock(rq, &rf);
> +		activate_blocked_entities(rq, p, wake_flags);

I'm unsure if it's a big deal, but IRQs are disabled here and 
activate_blocked_entities() disables them again.

> +		rq_lock(rq, &rf);
> +		update_rq_clock(rq);
>   	}
>   
>   	/*
> @@ -4423,6 +4556,7 @@ int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
>   	if (p->blocked_on_state == BO_WAKING)
>   		p->blocked_on_state = BO_RUNNABLE;
>   	raw_spin_unlock_irqrestore(&p->blocked_lock, flags);
> +	activate_blocked_entities(cpu_rq(cpu), p, wake_flags);
>   out:
>   	if (success)
>   		ttwu_stat(p, task_cpu(p), wake_flags);
> @@ -6663,19 +6797,6 @@ proxy_resched_idle(struct rq *rq, struct task_struct *next)
>   	return rq->idle;
>   }
>   
> -static bool proxy_deactivate(struct rq *rq, struct task_struct *next)
> -{
> -	unsigned long state = READ_ONCE(next->__state);
> -
> -	/* Don't deactivate if the state has been changed to TASK_RUNNING */
> -	if (state == TASK_RUNNING)
> -		return false;
> -	if (!try_to_deactivate_task(rq, next, state, true))

Now we can drop the last argument (deactivate_cond) of 
try_to_deactivate_task() since it can be determined via 
!task_is_blocked(p) I think. IOW:

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 30dfb6f14f2b..5b3b4ebca65d 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -6866,11 +6866,11 @@ pick_next_task(struct rq *rq, struct task_struct 
*prev, struct rq_flags *rf)
  #endif

  static bool try_to_deactivate_task(struct rq *rq, struct task_struct *p,
-				   unsigned long task_state, bool deactivate_cond)
+				   unsigned long task_state)
  {
  	if (signal_pending_state(task_state, p)) {
  		WRITE_ONCE(p->__state, TASK_RUNNING);
-	} else if (deactivate_cond) {
+	} else if (!task_is_blocked(p)) {
  		p->sched_contributes_to_load =
  			(task_state & TASK_UNINTERRUPTIBLE) &&
  			!(task_state & TASK_NOLOAD) &&
@@ -7329,8 +7329,7 @@ static void __sched notrace __schedule(unsigned 
int sched_mode)
  	 */
  	prev_state = READ_ONCE(prev->__state);
  	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
-		try_to_deactivate_task(rq, prev, prev_state,
-				       !task_is_blocked(prev));
+		try_to_deactivate_task(rq, prev, prev_state);
  		switch_count = &prev->nvcsw;
  	}


> -		return false;
> -	proxy_resched_idle(rq, next);
> -	return true;
> -}
> -
>   #ifdef CONFIG_SMP
>   /*
>    * If the blocked-on relationship crosses CPUs, migrate @p to the
> @@ -6761,6 +6882,31 @@ proxy_migrate_task(struct rq *rq, struct rq_flags *rf,
>   }
>   #endif /* CONFIG_SMP */
>   
> +static void proxy_enqueue_on_owner(struct rq *rq, struct task_struct *owner,
> +				   struct task_struct *next)
> +{
> +	/*
> +	 * ttwu_activate() will pick them up and place them on whatever rq
> +	 * @owner will run next.
> +	 */
> +	if (!owner->on_rq) {
> +		BUG_ON(!next->on_rq);
> +		deactivate_task(rq, next, DEQUEUE_SLEEP);
> +		if (task_current_selected(rq, next)) {
> +			put_prev_task(rq, next);
> +			rq_set_selected(rq, rq->idle);
> +		}
> +		/*
> +		 * ttwu_do_activate must not have a chance to activate p
> +		 * elsewhere before it's fully extricated from its old rq.
> +		 */
> +		WARN_ON(next->sleeping_owner);
> +		next->sleeping_owner = owner;
> +		smp_mb();
> +		list_add(&next->blocked_node, &owner->blocked_head);
> +	}
> +}
> +
>   /*
>    * Find who @next (currently blocked on a mutex) can proxy for.
>    *
> @@ -6866,10 +7012,40 @@ find_proxy_task(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
>   		}
>   
>   		if (!owner->on_rq) {
> -			/* XXX Don't handle blocked owners yet */
> -			if (!proxy_deactivate(rq, next))
> -				ret = next;
> -			goto out;
> +			/*
> +			 * rq->curr must not be added to the blocked_head list or else
> +			 * ttwu_do_activate could enqueue it elsewhere before it switches
> +			 * out here. The approach to avoid this is the same as in the
> +			 * migrate_task case.
> +			 */
> +			if (curr_in_chain) {
> +				raw_spin_unlock(&p->blocked_lock);
> +				raw_spin_unlock(&mutex->wait_lock);
> +				return proxy_resched_idle(rq, next);
> +			}
> +
> +			/*
> +			 * If !@owner->on_rq, holding @rq->lock will not pin the task,
> +			 * so we cannot drop @mutex->wait_lock until we're sure its a blocked
> +			 * task on this rq.
> +			 *
> +			 * We use @owner->blocked_lock to serialize against ttwu_activate().
> +			 * Either we see its new owner->on_rq or it will see our list_add().
> +			 */
> +			if (owner != p) {
> +				raw_spin_unlock(&p->blocked_lock);
> +				raw_spin_lock(&owner->blocked_lock);
> +			}
> +
> +			proxy_enqueue_on_owner(rq, owner, next);
> +
> +			if (task_current_selected(rq, next)) {
> +				put_prev_task(rq, next);
> +				rq_set_selected(rq, rq->idle);
> +			}
> +			raw_spin_unlock(&owner->blocked_lock);
> +			raw_spin_unlock(&mutex->wait_lock);
> +			return NULL; /* retry task selection */
>   		}
>   
>   		if (owner == p) {