From: Bijan Tabatabai <bijantabatab@micron.com>

This patchset adds two DAMON commands, commit_schemes_dests and
wait_for_schemes_apply, that I have found useful for using a migrate_hot
scheme with migration dests.

The commit_schemes_dests command, similar to the existing
commit_schemes_quota_goals, is used to commit only the dests fields of
schemes. This has a couple of benefits:
1) It is more efficient than recommitting all the DAMON data.
2) Doing a full commit resets the aggregation and ops_update intervals. If
a user sends the full commit command frequently (relatively to those
intervals) the aggregation and ops_update events will be prevented from
triggering. Having a separate commit command side steps this problem.

The wait_for_schemes_apply command causes the calling thread to wait until
all schemes have been applied. It does this by calling damos_walk() with a
NULL walk_fn. This can be useful, for example, if a user wants to know when
new scheme parameters they've committed have been applied. Another use case
could be if a user wants to record the system state every time a scheme is
applied for debuging purposes.

The functionality of wait_for_schemes_apply can be achieved with the
existing update_schemes_tried_bytes and update_schemes_tried_regions
commands. However, having a separate command avoids extra work and makes
user intent clearer when used in scripts.

The first two patches implement the commit_schemes_dests command.
The third patch implements the wait_for_schemes_apply command.
The fourth and fifth patches add documentation for these new commands.

Bijan Tabatabai (5):
  mm/damon/core: Add damos_destroy_dests()
  mm/damon/sysfs: Implement a command to only commit scheme dests
  mm/damon/sysfs: Implement a command to wait until schemes are applied
  Docs/ABI/damon: Document new DAMON commands
  Docs/admin-guide/mm/damon/usage: Document new DAMON commands

 .../ABI/testing/sysfs-kernel-mm-damon         | 21 ++++----
 Documentation/admin-guide/mm/damon/usage.rst  |  4 ++
 include/linux/damon.h                         |  2 +
 mm/damon/core.c                               | 12 +++--
 mm/damon/sysfs-common.h                       |  3 ++
 mm/damon/sysfs-schemes.c                      | 35 ++++++++++--
 mm/damon/sysfs.c                              | 54 +++++++++++++++++++
 7 files changed, 116 insertions(+), 15 deletions(-)

-- 
2.43.5

Hi Bijan,

Subjects of patches usually be converted into lowercase when be merged into mm
tree.  I'd suggest using lowercase even in patches stage, if those aim to be
merged into mm tree.

On Tue,  5 Aug 2025 11:20:17 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:

> From: Bijan Tabatabai <bijantabatab@micron.com>
> 
> This patchset adds two DAMON commands, commit_schemes_dests and
> wait_for_schemes_apply, that I have found useful for using a migrate_hot
> scheme with migration dests.

Thank you for this patchset!  I believe we shouldn't be afraid at adding
features, but should also be carefult at making good solutions for real
problems.  So, my main response to this patchset is, hopefully unsurprisingly,
I'd like to better understand the requirements and the problem you are
encountering.

> 
> The commit_schemes_dests command, similar to the existing
> commit_schemes_quota_goals, is used to commit only the dests fields of
> schemes. This has a couple of benefits:
> 1) It is more efficient than recommitting all the DAMON data.
> 2) Doing a full commit resets the aggregation and ops_update intervals. If
> a user sends the full commit command frequently (relatively to those
> intervals) the aggregation and ops_update events will be prevented from
> triggering. Having a separate commit command side steps this problem.

I agree the commit command of DAMON sysfs interface is inefficient, and could
make the infinite intervals delay problem.  But, I didn't expect there could be
use cases that use commit feature frequently enough to make the inefficiency
and the intervals delay be real problems.  Could you please let me know more
details about your use case and how severe problem DAMON is causing?

Depending on the real problem, I'm wondering if optimizing commit command can
be a solution.  For example, skipping the update of next_aggregation_sis and
next_ops_update_sis when the intervals are not changed might be able to solve
the intervals delay problem.

> 
> The wait_for_schemes_apply command causes the calling thread to wait until
> all schemes have been applied. It does this by calling damos_walk() with a
> NULL walk_fn. This can be useful, for example, if a user wants to know when
> new scheme parameters they've committed have been applied. Another use case
> could be if a user wants to record the system state every time a scheme is
> applied for debuging purposes.
> 
> The functionality of wait_for_schemes_apply can be achieved with the
> existing update_schemes_tried_bytes and update_schemes_tried_regions
> commands. However, having a separate command avoids extra work and makes
> user intent clearer when used in scripts.

I agree extra works are always better to be avoided.  But is the overhead large
enough to be a real problem for your use case?  I also agree it could make the
user script cleaner, but adding a kernel feature only for user scripts
readabilities sounds like too much, since the user script could have its own
abstract layers for its readability.

Also, even if the new command is implemented, since the DAMOS schemes continue
running, the system status will keep changing.  If you cannot do the recording
of the system state in a restricted time, the recorded information might not be
that reliable.  So I'm not sure if you really need this strict waiting in this
way.

Could you please share more details about what you want to do with the new
command, and how much problem you are seeing?  I'm particularly curious what
system state you want to record, and why you need to wait the exact time
interval.

Depending on the real use case and the problem, I think we might be able to
reuse the DAMOS scheme stats, or implement more strict and reliable time
control features, say, making kdamond or schemes pause and resume as users ask?


Thanks,
SJ

[...]

Hi SJ,


On Tue, Aug 5, 2025 at 7:40 PM SeongJae Park <sj@kernel.org> wrote:
>
> Hi Bijan,
>
> Subjects of patches usually be converted into lowercase when be merged into mm
> tree.  I'd suggest using lowercase even in patches stage, if those aim to be
> merged into mm tree.

Noted, thanks.

[...]
> > The commit_schemes_dests command, similar to the existing
> > commit_schemes_quota_goals, is used to commit only the dests fields of
> > schemes. This has a couple of benefits:
> > 1) It is more efficient than recommitting all the DAMON data.
> > 2) Doing a full commit resets the aggregation and ops_update intervals. If
> > a user sends the full commit command frequently (relatively to those
> > intervals) the aggregation and ops_update events will be prevented from
> > triggering. Having a separate commit command side steps this problem.
>
> I agree the commit command of DAMON sysfs interface is inefficient, and could
> make the infinite intervals delay problem.  But, I didn't expect there could be
> use cases that use commit feature frequently enough to make the inefficiency
> and the intervals delay be real problems.  Could you please let me know more
> details about your use case and how severe problem DAMON is causing?

In my use case, I am trying to optimize the interleave ratio of
applications to maximize their performance without prior knowledge of
their behavior. To do this, we take the steps of updating the ratio,
observing how the system reacts to the change in ratio, and update the
ratio again accordingly. Because we want to approach the ideal
interleave ratio quickly, we update the weights frequently, motivating
the commit_schemes_dests. Similarly, we want to observe how the system
reacts to the change only after the change has been applied,
motivating wait_for_schemes_apply.

The consequences are not very severe. The problem can be worked around
by either updating less frequently, at the cost of converging slower,
or decreasing the maximum aggregation period, which from what I
understand may affect the access monitoring behavior.

> Depending on the real problem, I'm wondering if optimizing commit command can
> be a solution.  For example, skipping the update of next_aggregation_sis and
> next_ops_update_sis when the intervals are not changed might be able to solve
> the intervals delay problem.

This would work for my use case. Another option might be to have a
more general commit_schemes command, which may be useful to other use
cases. I'll defer to your judgement on which would be better.

> >
> > The wait_for_schemes_apply command causes the calling thread to wait until
> > all schemes have been applied. It does this by calling damos_walk() with a
> > NULL walk_fn. This can be useful, for example, if a user wants to know when
> > new scheme parameters they've committed have been applied. Another use case
> > could be if a user wants to record the system state every time a scheme is
> > applied for debuging purposes.
> >
> > The functionality of wait_for_schemes_apply can be achieved with the
> > existing update_schemes_tried_bytes and update_schemes_tried_regions
> > commands. However, having a separate command avoids extra work and makes
> > user intent clearer when used in scripts.
>
> I agree extra works are always better to be avoided.  But is the overhead large
> enough to be a real problem for your use case?  I also agree it could make the
> user script cleaner, but adding a kernel feature only for user scripts
> readabilities sounds like too much, since the user script could have its own
> abstract layers for its readability.

Totally fair. I will drop wait_for_apply_schemes in any future versions.

> Also, even if the new command is implemented, since the DAMOS schemes continue
> running, the system status will keep changing.  If you cannot do the recording
> of the system state in a restricted time, the recorded information might not be
> that reliable.  So I'm not sure if you really need this strict waiting in this
> way.

Fair. That was not something I was personally using the command for,
just another possible use case I thought of. Regardless of the
usefulness of that, the existing commands using damos_walk would work
well enough.

> Could you please share more details about what you want to do with the new
> command, and how much problem you are seeing?  I'm particularly curious what
> system state you want to record, and why you need to wait the exact time
> interval.

I mentioned this above, but I am using this to wait for new migration
weights to be applied before monitoring how the change affects
applications, but again, this can be done with existing commands.

Thanks for your response,
Bijan

On Tue, 5 Aug 2025 20:27:52 -0500 Bijan Tabatabai <bijan311@gmail.com> wrote:

> Hi SJ,
> 
> 
> On Tue, Aug 5, 2025 at 7:40 PM SeongJae Park <sj@kernel.org> wrote:
> >
> > Hi Bijan,
> >
> > Subjects of patches usually be converted into lowercase when be merged into mm
> > tree.  I'd suggest using lowercase even in patches stage, if those aim to be
> > merged into mm tree.
> 
> Noted, thanks.
> 
> [...]
> > > The commit_schemes_dests command, similar to the existing
> > > commit_schemes_quota_goals, is used to commit only the dests fields of
> > > schemes. This has a couple of benefits:
> > > 1) It is more efficient than recommitting all the DAMON data.
> > > 2) Doing a full commit resets the aggregation and ops_update intervals. If
> > > a user sends the full commit command frequently (relatively to those
> > > intervals) the aggregation and ops_update events will be prevented from
> > > triggering. Having a separate commit command side steps this problem.
> >
> > I agree the commit command of DAMON sysfs interface is inefficient, and could
> > make the infinite intervals delay problem.  But, I didn't expect there could be
> > use cases that use commit feature frequently enough to make the inefficiency
> > and the intervals delay be real problems.  Could you please let me know more
> > details about your use case and how severe problem DAMON is causing?
> 
> In my use case, I am trying to optimize the interleave ratio of
> applications to maximize their performance without prior knowledge of
> their behavior. To do this, we take the steps of updating the ratio,
> observing how the system reacts to the change in ratio, and update the
> ratio again accordingly. Because we want to approach the ideal
> interleave ratio quickly, we update the weights frequently, motivating
> the commit_schemes_dests. Similarly, we want to observe how the system
> reacts to the change only after the change has been applied,
> motivating wait_for_schemes_apply.

Thank you for sharing these details!

It sounds like you are using DAMOS without any quota, and the target workload
has static memory mapping.  Hence all migrations for the new weights can be
completed after one DAMOS schemes apply interval, and no more migration will
happen until new weights are given.  And that's why you want
wait_for_schemes_apply, since when the command is finished is when all new
weights based interleaving is done.  Am I understanding correctly?

> 
> The consequences are not very severe. The problem can be worked around
> by either updating less frequently, at the cost of converging slower,
> or decreasing the maximum aggregation period, which from what I
> understand may affect the access monitoring behavior.

Sounds suboptimal work arounds for you...

> 
> > Depending on the real problem, I'm wondering if optimizing commit command can
> > be a solution.  For example, skipping the update of next_aggregation_sis and
> > next_ops_update_sis when the intervals are not changed might be able to solve
> > the intervals delay problem.
> 
> This would work for my use case.

Great to hear this, and I agree.  The commit operation internally uses
damon_call(), which takes up to one sampling interval.  Also the migration
operation that you will wait for would also take no small time, depending on
the amount of pages to migrate.  Compared to those, I think the commit speed
increase due to committing unnecessary paramters may relatively short.
deal.

Meanwhile, I was concerning the continuous next_{aggregation,ops_update_}sis
delay could be a real problem.  And this option should solve the real problem.

> Another option might be to have a
> more general commit_schemes command, which may be useful to other use
> cases. I'll defer to your judgement on which would be better.

If my above theory is not wrong, I'd suggest making the commit operation
optimization.  If it turns out to be not enough for your or others' use cases,
we can further consider commit_schemes.

> 
> > >
> > > The wait_for_schemes_apply command causes the calling thread to wait until
> > > all schemes have been applied. It does this by calling damos_walk() with a
> > > NULL walk_fn. This can be useful, for example, if a user wants to know when
> > > new scheme parameters they've committed have been applied. Another use case
> > > could be if a user wants to record the system state every time a scheme is
> > > applied for debuging purposes.
> > >
> > > The functionality of wait_for_schemes_apply can be achieved with the
> > > existing update_schemes_tried_bytes and update_schemes_tried_regions
> > > commands. However, having a separate command avoids extra work and makes
> > > user intent clearer when used in scripts.
> >
> > I agree extra works are always better to be avoided.  But is the overhead large
> > enough to be a real problem for your use case?  I also agree it could make the
> > user script cleaner, but adding a kernel feature only for user scripts
> > readabilities sounds like too much, since the user script could have its own
> > abstract layers for its readability.
> 
> Totally fair. I will drop wait_for_apply_schemes in any future versions.
> 
> > Also, even if the new command is implemented, since the DAMOS schemes continue
> > running, the system status will keep changing.  If you cannot do the recording
> > of the system state in a restricted time, the recorded information might not be
> > that reliable.  So I'm not sure if you really need this strict waiting in this
> > way.
> 
> Fair. That was not something I was personally using the command for,
> just another possible use case I thought of. Regardless of the
> usefulness of that, the existing commands using damos_walk would work
> well enough.

I'm glad to hear we found a way to go.  As you may know,
update_schemes_tried_bytes would be more efficient, so I would suggest that
more than update_schemes_tried_regions.

If the wait is not strictly need to be accurate, maybe monitoring the DAMOS
scheme stats in auto-update mode[1] until any change is made could also be a
way.  The stat update for a scheme will be done only after the scheme is
applied to all applicable regions for a round.

> 
> > Could you please share more details about what you want to do with the new
> > command, and how much problem you are seeing?  I'm particularly curious what
> > system state you want to record, and why you need to wait the exact time
> > interval.
> 
> I mentioned this above, but I am using this to wait for new migration
> weights to be applied before monitoring how the change affects
> applications, but again, this can be done with existing commands.

Thank you again for kindly sharing your use case and participate in this
constructive discussion :)

[1] https://lore.kernel.org/20250717055448.56976-1-sj@kernel.org


Thanks,
SJ

[...]