[PATCH] mm/damon/core: skip needless update of next_{aggregation,ops_update}_sis

[PATCH] mm/damon/core: skip needless update of next_{aggregation,ops_update}_sis

[Bijan Tabatabai]

From: Bijan Tabatabai <bijantabatab@micron.com>

In damon_set_attrs(), ctx->next_{aggregation,ops_update}_sis would be
reset, even if the sample interval, aggregation interval, or ops update
interval were not changed. If damon_set_attrs() is called relatively
frequently, such as by frequent "commit" operations, aggregation and ops
update operations could be needlessly delayed.

This patch avoids this by only updating next_{aggregation,ops_update}_sis
if the relevant intervals were changed.

Cc: Bijan Tabatabai <bijan311@gmail.com>
Signed-off-by: Bijan Tabatabai <bijantabatab@micron.com>
---
This patch came from discussions in [1].

[1] https://lore.kernel.org/all/20250805162022.4920-1-bijan311@gmail.com/
---
 mm/damon/core.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

diff --git a/mm/damon/core.c b/mm/damon/core.c
index 6a2fe1f2c952..1c3d8b92257c 100644
--- a/mm/damon/core.c
+++ b/mm/damon/core.c
@@ -693,6 +693,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
 	unsigned long sample_interval = attrs->sample_interval ?
 		attrs->sample_interval : 1;
 	struct damos *s;
+	bool sample_interval_changed = ctx->attrs.sample_interval !=
+		attrs->sample_interval;
+	bool aggr_interval_changed = ctx->attrs.aggr_interval !=
+		attrs->aggr_interval;
+	bool ops_update_interval_changed = ctx->attrs.ops_update_interval !=
+		attrs->ops_update_interval;
 	bool aggregating = ctx->passed_sample_intervals <
 		ctx->next_aggregation_sis;
 
@@ -710,10 +716,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
 	if (!attrs->aggr_samples)
 		attrs->aggr_samples = attrs->aggr_interval / sample_interval;
 
-	ctx->next_aggregation_sis = ctx->passed_sample_intervals +
-		attrs->aggr_interval / sample_interval;
-	ctx->next_ops_update_sis = ctx->passed_sample_intervals +
-		attrs->ops_update_interval / sample_interval;
+	if (sample_interval_changed || aggr_interval_changed)
+		ctx->next_aggregation_sis = ctx->passed_sample_intervals +
+			attrs->aggr_interval / sample_interval;
+	if (sample_interval_changed || ops_update_interval_changed)
+		ctx->next_ops_update_sis = ctx->passed_sample_intervals +
+			attrs->ops_update_interval / sample_interval;
 
 	damon_update_monitoring_results(ctx, attrs, aggregating);
 	ctx->attrs = *attrs;
-- 
2.43.0

Re: [PATCH] mm/damon/core: skip needless update of next_{aggregation,ops_update}_sis

[SeongJae Park]

Hi Bijan,

On Wed,  6 Aug 2025 11:43:16 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:

> From: Bijan Tabatabai <bijantabatab@micron.com>
> 
> In damon_set_attrs(), ctx->next_{aggregation,ops_update}_sis would be
> reset, even if the sample interval, aggregation interval, or ops update
> interval were not changed. If damon_set_attrs() is called relatively
> frequently, such as by frequent "commit" operations, aggregation and ops
> update operations could be needlessly delayed.
> 
> This patch avoids this by only updating next_{aggregation,ops_update}_sis
> if the relevant intervals were changed.
> 
> Cc: Bijan Tabatabai <bijan311@gmail.com>
> Signed-off-by: Bijan Tabatabai <bijantabatab@micron.com>
> ---
> This patch came from discussions in [1].
> 
> [1] https://lore.kernel.org/all/20250805162022.4920-1-bijan311@gmail.com/

Thank you for sending this patch as we discussed on the thread!

> ---
>  mm/damon/core.c | 16 ++++++++++++----
>  1 file changed, 12 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/damon/core.c b/mm/damon/core.c
> index 6a2fe1f2c952..1c3d8b92257c 100644
> --- a/mm/damon/core.c
> +++ b/mm/damon/core.c
> @@ -693,6 +693,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
>  	unsigned long sample_interval = attrs->sample_interval ?
>  		attrs->sample_interval : 1;
>  	struct damos *s;
> +	bool sample_interval_changed = ctx->attrs.sample_interval !=
> +		attrs->sample_interval;
> +	bool aggr_interval_changed = ctx->attrs.aggr_interval !=
> +		attrs->aggr_interval;
> +	bool ops_update_interval_changed = ctx->attrs.ops_update_interval !=
> +		attrs->ops_update_interval;
>  	bool aggregating = ctx->passed_sample_intervals <
>  		ctx->next_aggregation_sis;
>  
> @@ -710,10 +716,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
>  	if (!attrs->aggr_samples)
>  		attrs->aggr_samples = attrs->aggr_interval / sample_interval;
>  
> -	ctx->next_aggregation_sis = ctx->passed_sample_intervals +
> -		attrs->aggr_interval / sample_interval;
> -	ctx->next_ops_update_sis = ctx->passed_sample_intervals +
> -		attrs->ops_update_interval / sample_interval;
> +	if (sample_interval_changed || aggr_interval_changed)
> +		ctx->next_aggregation_sis = ctx->passed_sample_intervals +
> +			attrs->aggr_interval / sample_interval;
> +	if (sample_interval_changed || ops_update_interval_changed)
> +		ctx->next_ops_update_sis = ctx->passed_sample_intervals +
> +			attrs->ops_update_interval / sample_interval;
>  
>  	damon_update_monitoring_results(ctx, attrs, aggregating);
>  	ctx->attrs = *attrs;

Long story short, this (original) code is bit complicated and hence I suggest
to make this change less optimum but simpler.

damon_update_monitoring_results() assumes it is called only just after
next_{aggr,ops_update}_sis are  updated.  And the assumption is important for
the pseudo-moving-sum access frequency maintenance.  As a result, this can make
the monitoring results temporarily corrupted, and splat warning once.  Please
refer to commit 591c4c78be063 ("mm/damon/core: warn and fix nr_accesses[_bp]
corruption") or the patch thread[1] if you want more details.

Also damon_set_attrs() is called not only from commit situation.  So I think
this maybe not an ideal part to modify.

What about modifying damon_commit_ctx() to check if new and old
damon_ctx->attrs are entirely same, and skip calling damon_set_attrs() in the
case?  Doing the entire damon_attrs comparison might be suboptimum, but would
make the change simpler.  I assume the suboptimum comparison is not a real
problem for your use case, so I think that could be a good tradeoff?

Also, I realize the moving sum access frequencies maintenance and
damon_set_attrs() are not well documented and organized.  I will try to make
those better cleaned up and documented.

[1] https://lkml.kernel.org/r/20250513002715.40126-2-sj@kernel.org


Thanks,
SJ

> -- 
> 2.43.0

Re: [PATCH] mm/damon/core: skip needless update of next_{aggregation,ops_update}_sis

[Bijan Tabatabai]

On Wed, Aug 6, 2025 at 2:09 PM SeongJae Park <sj@kernel.org> wrote:
>
> Hi Bijan,
>
> On Wed,  6 Aug 2025 11:43:16 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:
>
> > From: Bijan Tabatabai <bijantabatab@micron.com>
> >
> > In damon_set_attrs(), ctx->next_{aggregation,ops_update}_sis would be
> > reset, even if the sample interval, aggregation interval, or ops update
> > interval were not changed. If damon_set_attrs() is called relatively
> > frequently, such as by frequent "commit" operations, aggregation and ops
> > update operations could be needlessly delayed.
> >
> > This patch avoids this by only updating next_{aggregation,ops_update}_sis
> > if the relevant intervals were changed.
> >
> > Cc: Bijan Tabatabai <bijan311@gmail.com>
> > Signed-off-by: Bijan Tabatabai <bijantabatab@micron.com>
> > ---
> > This patch came from discussions in [1].
> >
> > [1] https://lore.kernel.org/all/20250805162022.4920-1-bijan311@gmail.com/
>
> Thank you for sending this patch as we discussed on the thread!
>
> > ---
> >  mm/damon/core.c | 16 ++++++++++++----
> >  1 file changed, 12 insertions(+), 4 deletions(-)
> >
> > diff --git a/mm/damon/core.c b/mm/damon/core.c
> > index 6a2fe1f2c952..1c3d8b92257c 100644
> > --- a/mm/damon/core.c
> > +++ b/mm/damon/core.c
> > @@ -693,6 +693,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
> >       unsigned long sample_interval = attrs->sample_interval ?
> >               attrs->sample_interval : 1;
> >       struct damos *s;
> > +     bool sample_interval_changed = ctx->attrs.sample_interval !=
> > +             attrs->sample_interval;
> > +     bool aggr_interval_changed = ctx->attrs.aggr_interval !=
> > +             attrs->aggr_interval;
> > +     bool ops_update_interval_changed = ctx->attrs.ops_update_interval !=
> > +             attrs->ops_update_interval;
> >       bool aggregating = ctx->passed_sample_intervals <
> >               ctx->next_aggregation_sis;
> >
> > @@ -710,10 +716,12 @@ int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
> >       if (!attrs->aggr_samples)
> >               attrs->aggr_samples = attrs->aggr_interval / sample_interval;
> >
> > -     ctx->next_aggregation_sis = ctx->passed_sample_intervals +
> > -             attrs->aggr_interval / sample_interval;
> > -     ctx->next_ops_update_sis = ctx->passed_sample_intervals +
> > -             attrs->ops_update_interval / sample_interval;
> > +     if (sample_interval_changed || aggr_interval_changed)
> > +             ctx->next_aggregation_sis = ctx->passed_sample_intervals +
> > +                     attrs->aggr_interval / sample_interval;
> > +     if (sample_interval_changed || ops_update_interval_changed)
> > +             ctx->next_ops_update_sis = ctx->passed_sample_intervals +
> > +                     attrs->ops_update_interval / sample_interval;
> >
> >       damon_update_monitoring_results(ctx, attrs, aggregating);
> >       ctx->attrs = *attrs;
>
> Long story short, this (original) code is bit complicated and hence I suggest
> to make this change less optimum but simpler.
>
> damon_update_monitoring_results() assumes it is called only just after
> next_{aggr,ops_update}_sis are  updated.  And the assumption is important for
> the pseudo-moving-sum access frequency maintenance.  As a result, this can make
> the monitoring results temporarily corrupted, and splat warning once.  Please
> refer to commit 591c4c78be063 ("mm/damon/core: warn and fix nr_accesses[_bp]
> corruption") or the patch thread[1] if you want more details.
>
> Also damon_set_attrs() is called not only from commit situation.  So I think
> this maybe not an ideal part to modify.
>
> What about modifying damon_commit_ctx() to check if new and old
> damon_ctx->attrs are entirely same, and skip calling damon_set_attrs() in the
> case?  Doing the entire damon_attrs comparison might be suboptimum, but would
> make the change simpler.  I assume the suboptimum comparison is not a real
> problem for your use case, so I think that could be a good tradeoff?

I can definitely do this. Checking a few extra fields is no big deal.

Silly question, but think it's best to get it out of the way before
sending another patch: do you think there's a more elegant way of just
having a dumb comparison function like

bool damon_attrs_equal(struct damon_attrs *a, struct damon_attrs *b)
{
        return a->sample_interval == b->sample_interval &&
            a->aggr_interval == b->aggr_interval &&
            ...
}

And I assume I shouldn't compare the aggr_samples field because it's
private, is that right?

[...]

Thanks for your patience,
Bijan

Re: [PATCH] mm/damon/core: skip needless update of next_{aggregation,ops_update}_sis

[SeongJae Park]

On Wed, 6 Aug 2025 14:49:21 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:

> On Wed, Aug 6, 2025 at 2:09 PM SeongJae Park <sj@kernel.org> wrote:
> >
> > Hi Bijan,
> >
> > On Wed,  6 Aug 2025 11:43:16 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:
> >
> > > From: Bijan Tabatabai <bijantabatab@micron.com>
> > >
> > > In damon_set_attrs(), ctx->next_{aggregation,ops_update}_sis would be
> > > reset, even if the sample interval, aggregation interval, or ops update
> > > interval were not changed. If damon_set_attrs() is called relatively
> > > frequently, such as by frequent "commit" operations, aggregation and ops
> > > update operations could be needlessly delayed.
> > >
> > > This patch avoids this by only updating next_{aggregation,ops_update}_sis
> > > if the relevant intervals were changed.
[...]
> > What about modifying damon_commit_ctx() to check if new and old
> > damon_ctx->attrs are entirely same, and skip calling damon_set_attrs() in the
> > case?  Doing the entire damon_attrs comparison might be suboptimum, but would
> > make the change simpler.  I assume the suboptimum comparison is not a real
> > problem for your use case, so I think that could be a good tradeoff?
> 
> I can definitely do this. Checking a few extra fields is no big deal.
> 
> Silly question, but think it's best to get it out of the way before
> sending another patch: do you think there's a more elegant way of just
> having a dumb comparison function like
> 
> bool damon_attrs_equal(struct damon_attrs *a, struct damon_attrs *b)
> {
>         return a->sample_interval == b->sample_interval &&
>             a->aggr_interval == b->aggr_interval &&
>             ...
> }
> 
> And I assume I shouldn't compare the aggr_samples field because it's
> private, is that right?

Ah, you're right, thank you for asking this!

Maybe we can copy src->attrs to a local damon_attrs variable, overwrite
aggr_samples of the copy and dst, and memcmp() the copy and dst?

I wouldn't mind the dumb comparison function, though, if you prefer.  I'll
defer the decision to you!


Thanks,
SJ

[...]