[PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Zqiang]

When RCU_BOOST is enabled, the boost kthreads will boosting readers
who are blocking a given grace period, if the current reader tasks
have a higher priority than boost kthreads(the boost kthreads priority
not always 1, if the kthread_prio is set), boosting is useless, skip
current task and select next task to boosting, reduce the time for a
given grace period.

Signed-off-by: Zqiang <qiang1.zhang@intel.com>
---
 kernel/rcu/tree_plugin.h | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
index c3d212bc5338..d35b6da66bbd 100644
--- a/kernel/rcu/tree_plugin.h
+++ b/kernel/rcu/tree_plugin.h
@@ -12,6 +12,7 @@
  */
 
 #include "../locking/rtmutex_common.h"
+#include <linux/sched/deadline.h>
 
 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
 {
@@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
 	 * section.
 	 */
 	t = container_of(tb, struct task_struct, rcu_node_entry);
+	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
+		tb = rcu_next_node_entry(t, rnp);
+		WRITE_ONCE(rnp->boost_tasks, tb);
+		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
+		goto end;
+	}
+
 	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 	/* Lock only for side effect: boosts task t's priority. */
 	rt_mutex_lock(&rnp->boost_mtx);
 	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
 	rnp->n_boosts++;
-
+end:
 	return READ_ONCE(rnp->exp_tasks) != NULL ||
 	       READ_ONCE(rnp->boost_tasks) != NULL;
 }
-- 
2.25.1

Re: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Neeraj Upadhyay]

On 3/4/2022 2:56 PM, Zqiang wrote:
> When RCU_BOOST is enabled, the boost kthreads will boosting readers
> who are blocking a given grace period, if the current reader tasks
> have a higher priority than boost kthreads(the boost kthreads priority
> not always 1, if the kthread_prio is set), boosting is useless, skip
> current task and select next task to boosting, reduce the time for a
> given grace period.
> 
> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> ---
>   kernel/rcu/tree_plugin.h | 10 +++++++++-
>   1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index c3d212bc5338..d35b6da66bbd 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -12,6 +12,7 @@
>    */
>   
>   #include "../locking/rtmutex_common.h"
> +#include <linux/sched/deadline.h>
>   
>   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
>   {
> @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
>   	 * section.
>   	 */
>   	t = container_of(tb, struct task_struct, rcu_node_entry);
> +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> +		tb = rcu_next_node_entry(t, rnp);
> +		WRITE_ONCE(rnp->boost_tasks, tb);
> +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> +		goto end;
> +	}
> +
>   	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
>   	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
>   	/* Lock only for side effect: boosts task t's priority. */
>   	rt_mutex_lock(&rnp->boost_mtx);
>   	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
>   	rnp->n_boosts++;
> -
> +end:

Nit: maybe rename the label to "skip_boost:" ?

Code looks fine; however, out of curiosity; given that the higher 
priority tasks, in general, would exit their read side critical section
quickly and boost the next blocking reader on exiting their read side 
section; do you see noticeable reduction in grace period timings with 
the change for certain type of workloads?


Thanks
Neeraj

>   	return READ_ONCE(rnp->exp_tasks) != NULL ||
>   	       READ_ONCE(rnp->boost_tasks) != NULL;
>   }

RE: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Zhang, Qiang1]

On 3/4/2022 2:56 PM, Zqiang wrote:
> When RCU_BOOST is enabled, the boost kthreads will boosting readers
> who are blocking a given grace period, if the current reader tasks
> have a higher priority than boost kthreads(the boost kthreads priority
> not always 1, if the kthread_prio is set), boosting is useless, skip
> current task and select next task to boosting, reduce the time for a
> given grace period.
> 
> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> ---
>   kernel/rcu/tree_plugin.h | 10 +++++++++-
>   1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index c3d212bc5338..d35b6da66bbd 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -12,6 +12,7 @@
>    */
>   
>   #include "../locking/rtmutex_common.h"
> +#include <linux/sched/deadline.h>
>   
>   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
>   {
> @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
>   	 * section.
>   	 */
>   	t = container_of(tb, struct task_struct, rcu_node_entry);
> +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> +		tb = rcu_next_node_entry(t, rnp);
> +		WRITE_ONCE(rnp->boost_tasks, tb);
> +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> +		goto end;
> +	}
> +
>   	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
>   	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
>   	/* Lock only for side effect: boosts task t's priority. */
>   	rt_mutex_lock(&rnp->boost_mtx);
>   	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
>   	rnp->n_boosts++;
> -
> +end:
>>
>>Nit: maybe rename the label to "skip_boost:" ?
>>
>>Code looks fine; however, out of curiosity; given that the higher 
>>priority tasks, in general, would exit their read side critical section
>>quickly and boost the next blocking reader on exiting their read side 
>>section; do you see noticeable reduction in grace period timings with 
>>the change for certain type of workloads?

Thanks for feedback ,  In preempt-RT systems, there will be many real-time threads (most
of them are created by users themselves ),  their priority is higher or lower than boost kthreads
(kthread_prio is set),  for rt tasks  with higher priority than boost kthreads, maybe it will exit
read side critical quickly,  maybe not, if it is preempted by a higher priority task,  If try to boost operation, 
this increases the boosts kthread waiting time, as a result, the next blkd tasks cannot be
boosted in time.  of course, I don't deny it,  there are also reasons that user priority setting is inappropriate.

Thanks
Zqiang

>>
>>
>>Thanks
>>Neeraj

>   	return READ_ONCE(rnp->exp_tasks) != NULL ||
>   	       READ_ONCE(rnp->boost_tasks) != NULL;
>   }

Re: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Paul E. McKenney]

On Mon, Mar 07, 2022 at 02:03:17AM +0000, Zhang, Qiang1 wrote:
> On 3/4/2022 2:56 PM, Zqiang wrote:
> > When RCU_BOOST is enabled, the boost kthreads will boosting readers
> > who are blocking a given grace period, if the current reader tasks
> > have a higher priority than boost kthreads(the boost kthreads priority
> > not always 1, if the kthread_prio is set), boosting is useless, skip
> > current task and select next task to boosting, reduce the time for a
> > given grace period.
> > 
> > Signed-off-by: Zqiang <qiang1.zhang@intel.com>

Adding to CC to get more eyes on this.  I am not necessarily opposed to
it, but I don't do that much RT work myself these days.

							Thanx, Paul

> > ---
> >   kernel/rcu/tree_plugin.h | 10 +++++++++-
> >   1 file changed, 9 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > index c3d212bc5338..d35b6da66bbd 100644
> > --- a/kernel/rcu/tree_plugin.h
> > +++ b/kernel/rcu/tree_plugin.h
> > @@ -12,6 +12,7 @@
> >    */
> >   
> >   #include "../locking/rtmutex_common.h"
> > +#include <linux/sched/deadline.h>
> >   
> >   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> >   {
> > @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
> >   	 * section.
> >   	 */
> >   	t = container_of(tb, struct task_struct, rcu_node_entry);
> > +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> > +		tb = rcu_next_node_entry(t, rnp);
> > +		WRITE_ONCE(rnp->boost_tasks, tb);
> > +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > +		goto end;
> > +	}
> > +
> >   	rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
> >   	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> >   	/* Lock only for side effect: boosts task t's priority. */
> >   	rt_mutex_lock(&rnp->boost_mtx);
> >   	rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
> >   	rnp->n_boosts++;
> > -
> > +end:
> >>
> >>Nit: maybe rename the label to "skip_boost:" ?
> >>
> >>Code looks fine; however, out of curiosity; given that the higher 
> >>priority tasks, in general, would exit their read side critical section
> >>quickly and boost the next blocking reader on exiting their read side 
> >>section; do you see noticeable reduction in grace period timings with 
> >>the change for certain type of workloads?
> 
> Thanks for feedback ,  In preempt-RT systems, there will be many real-time threads (most
> of them are created by users themselves ),  their priority is higher or lower than boost kthreads
> (kthread_prio is set),  for rt tasks  with higher priority than boost kthreads, maybe it will exit
> read side critical quickly,  maybe not, if it is preempted by a higher priority task,  If try to boost operation, 
> this increases the boosts kthread waiting time, as a result, the next blkd tasks cannot be
> boosted in time.  of course, I don't deny it,  there are also reasons that user priority setting is inappropriate.
> 
> Thanks
> Zqiang
> 
> >>
> >>
> >>Thanks
> >>Neeraj
> 
> >   	return READ_ONCE(rnp->exp_tasks) != NULL ||
> >   	       READ_ONCE(rnp->boost_tasks) != NULL;
> >   }

Re: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Uladzislau Rezki]

> On Mon, Mar 07, 2022 at 02:03:17AM +0000, Zhang, Qiang1 wrote:
> > On 3/4/2022 2:56 PM, Zqiang wrote:
> > > When RCU_BOOST is enabled, the boost kthreads will boosting readers
> > > who are blocking a given grace period, if the current reader tasks
> > > have a higher priority than boost kthreads(the boost kthreads priority
> > > not always 1, if the kthread_prio is set), boosting is useless, skip
> > > current task and select next task to boosting, reduce the time for a
> > > given grace period.
> > > 
> > > Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> 
> Adding to CC to get more eyes on this.  I am not necessarily opposed to
> it, but I don't do that much RT work myself these days.
> 
> 							Thanx, Paul
> 
> > > ---
> > >   kernel/rcu/tree_plugin.h | 10 +++++++++-
> > >   1 file changed, 9 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > index c3d212bc5338..d35b6da66bbd 100644
> > > --- a/kernel/rcu/tree_plugin.h
> > > +++ b/kernel/rcu/tree_plugin.h
> > > @@ -12,6 +12,7 @@
> > >    */
> > >   
> > >   #include "../locking/rtmutex_common.h"
> > > +#include <linux/sched/deadline.h>
> > >   
> > >   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> > >   {
> > > @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
> > >   	 * section.
> > >   	 */
> > >   	t = container_of(tb, struct task_struct, rcu_node_entry);
> > > +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> > > +		tb = rcu_next_node_entry(t, rnp);
> > > +		WRITE_ONCE(rnp->boost_tasks, tb);
> > > +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > +		goto end;
> > > +	}
> > > +
Why do you bypass the expedited grace period and boost any tasks anyway?
Same way the expedited gp can be blocked by higher prior tasks SCHED_DEADLINE
or SCHED_FIFO.

Thanks!

--
Vlad Rezki

Re: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Paul E. McKenney]

On Tue, Mar 08, 2022 at 07:04:21PM +0100, Uladzislau Rezki wrote:
> > On Mon, Mar 07, 2022 at 02:03:17AM +0000, Zhang, Qiang1 wrote:
> > > On 3/4/2022 2:56 PM, Zqiang wrote:
> > > > When RCU_BOOST is enabled, the boost kthreads will boosting readers
> > > > who are blocking a given grace period, if the current reader tasks
> > > > have a higher priority than boost kthreads(the boost kthreads priority
> > > > not always 1, if the kthread_prio is set), boosting is useless, skip
> > > > current task and select next task to boosting, reduce the time for a
> > > > given grace period.
> > > > 
> > > > Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> > 
> > Adding to CC to get more eyes on this.  I am not necessarily opposed to
> > it, but I don't do that much RT work myself these days.
> > 
> > 							Thanx, Paul
> > 
> > > > ---
> > > >   kernel/rcu/tree_plugin.h | 10 +++++++++-
> > > >   1 file changed, 9 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > index c3d212bc5338..d35b6da66bbd 100644
> > > > --- a/kernel/rcu/tree_plugin.h
> > > > +++ b/kernel/rcu/tree_plugin.h
> > > > @@ -12,6 +12,7 @@
> > > >    */
> > > >   
> > > >   #include "../locking/rtmutex_common.h"
> > > > +#include <linux/sched/deadline.h>
> > > >   
> > > >   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> > > >   {
> > > > @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
> > > >   	 * section.
> > > >   	 */
> > > >   	t = container_of(tb, struct task_struct, rcu_node_entry);
> > > > +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> > > > +		tb = rcu_next_node_entry(t, rnp);
> > > > +		WRITE_ONCE(rnp->boost_tasks, tb);
> > > > +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > > +		goto end;
> > > > +	}
> > > > +
> Why do you bypass the expedited grace period and boost any tasks anyway?
> Same way the expedited gp can be blocked by higher prior tasks SCHED_DEADLINE
> or SCHED_FIFO.

Just to make sure that I understand...

Are you pointing out that a SCHED_DEADLINE task might have exhausted
its budget, so that boosting might nonetheless be helpful?

Me, I honestly don't know what happens in that case, so I am just asking
the question.  And adding Juri on CC.  ;-)

							Thanx, Paul

Re: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Uladzislau Rezki]

On Tue, Mar 08, 2022 at 10:13:55AM -0800, Paul E. McKenney wrote:
> On Tue, Mar 08, 2022 at 07:04:21PM +0100, Uladzislau Rezki wrote:
> > > On Mon, Mar 07, 2022 at 02:03:17AM +0000, Zhang, Qiang1 wrote:
> > > > On 3/4/2022 2:56 PM, Zqiang wrote:
> > > > > When RCU_BOOST is enabled, the boost kthreads will boosting readers
> > > > > who are blocking a given grace period, if the current reader tasks
> > > > > have a higher priority than boost kthreads(the boost kthreads priority
> > > > > not always 1, if the kthread_prio is set), boosting is useless, skip
> > > > > current task and select next task to boosting, reduce the time for a
> > > > > given grace period.
> > > > > 
> > > > > Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> > > 
> > > Adding to CC to get more eyes on this.  I am not necessarily opposed to
> > > it, but I don't do that much RT work myself these days.
> > > 
> > > 							Thanx, Paul
> > > 
> > > > > ---
> > > > >   kernel/rcu/tree_plugin.h | 10 +++++++++-
> > > > >   1 file changed, 9 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> > > > > index c3d212bc5338..d35b6da66bbd 100644
> > > > > --- a/kernel/rcu/tree_plugin.h
> > > > > +++ b/kernel/rcu/tree_plugin.h
> > > > > @@ -12,6 +12,7 @@
> > > > >    */
> > > > >   
> > > > >   #include "../locking/rtmutex_common.h"
> > > > > +#include <linux/sched/deadline.h>
> > > > >   
> > > > >   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> > > > >   {
> > > > > @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
> > > > >   	 * section.
> > > > >   	 */
> > > > >   	t = container_of(tb, struct task_struct, rcu_node_entry);
> > > > > +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> > > > > +		tb = rcu_next_node_entry(t, rnp);
> > > > > +		WRITE_ONCE(rnp->boost_tasks, tb);
> > > > > +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > > > +		goto end;
> > > > > +	}
> > > > > +
> > Why do you bypass the expedited grace period and boost any tasks anyway?
> > Same way the expedited gp can be blocked by higher prior tasks SCHED_DEADLINE
> > or SCHED_FIFO.
> 
> Just to make sure that I understand...
> 
> Are you pointing out that a SCHED_DEADLINE task might have exhausted
> its budget, so that boosting might nonetheless be helpful?
> 
SCHED_DEADLINE we can not preempt nor stop it somehow(highest prio class),
it has some budget it makes use of. If it is in critical section then it
will leave asap, i do not take into account here IRQs and so on. I do not
see a reason to boost it.

>
> Me, I honestly don't know what happens in that case, so I am just asking
> the question.  And adding Juri on CC.  ;-)
> 
Juri should know more :)

--
Vlad Rezki

RE: [PATCH] rcu: Only boost rcu reader tasks with lower priority than boost kthreads

[Zhang, Qiang1]

> On Mon, Mar 07, 2022 at 02:03:17AM +0000, Zhang, Qiang1 wrote:
> > On 3/4/2022 2:56 PM, Zqiang wrote:
> > > When RCU_BOOST is enabled, the boost kthreads will boosting 
> > > readers who are blocking a given grace period, if the current 
> > > reader tasks have a higher priority than boost kthreads(the boost 
> > > kthreads priority not always 1, if the kthread_prio is set), 
> > > boosting is useless, skip current task and select next task to 
> > > boosting, reduce the time for a given grace period.
> > > 
> > > Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> 
> Adding to CC to get more eyes on this.  I am not necessarily opposed 
> to it, but I don't do that much RT work myself these days.
> 
> 							Thanx, Paul
> 
> > > ---
> > >   kernel/rcu/tree_plugin.h | 10 +++++++++-
> > >   1 file changed, 9 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h 
> > > index c3d212bc5338..d35b6da66bbd 100644
> > > --- a/kernel/rcu/tree_plugin.h
> > > +++ b/kernel/rcu/tree_plugin.h
> > > @@ -12,6 +12,7 @@
> > >    */
> > >   
> > >   #include "../locking/rtmutex_common.h"
> > > +#include <linux/sched/deadline.h>
> > >   
> > >   static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
> > >   {
> > > @@ -1065,13 +1066,20 @@ static int rcu_boost(struct rcu_node *rnp)
> > >   	 * section.
> > >   	 */
> > >   	t = container_of(tb, struct task_struct, rcu_node_entry);
> > > +	if (!rnp->exp_tasks && (dl_task(t) || t->prio <= current->prio)) {
> > > +		tb = rcu_next_node_entry(t, rnp);
> > > +		WRITE_ONCE(rnp->boost_tasks, tb);
> > > +		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > +		goto end;
> > > +	}
> > > +
>>>>Why do you bypass the expedited grace period and boost any tasks anyway?
>>>>Same way the expedited gp can be blocked by higher prior tasks SCHED_DEADLINE or SCHED_FIFO.

Thanks advice, it really ignores expedited grace period, may be
should generate 'boost_exp_tasks' pointer to point 'rnp->exp_tasks', do the same thing
as normal grace period.

Thanks
Zqiang

>>>>
>>>>Thanks!
>>>>
>>>>--
>>>>Vlad Rezki