Date:   Mon, 25 Sep 2023 15:53:02 +0000
From:   SeongJae Park <sj@...nel.org>
To:     杨欢 <link@...o.com>
Cc:     "SeongJae Park" <sj@...nel.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        open list <linux-kernel@...r.kernel.org>,
        "open list:DATA ACCESS MONITOR" <damon@...ts.linux.dev>,
        "open list:DATA ACCESS MONITOR" <linux-mm@...ck.org>,
        "opensource . kernel" <opensource.kernel@...o.com>
Subject: Re: [RFC 0/2] Damos filter cleanup code and implement workingset

Hi Huan,

On Mon, 25 Sep 2023 02:10:58 +0000 杨欢 <link@...o.com> wrote:

> HI SJ,
> > Hi Huan,
> >
> > On Fri, 22 Sep 2023 02:54:58 +0000 杨欢 <link@...o.com> wrote:
> >
> >> HI SJ,
> >>
> >> Thanks for your patient response.
> > Happy to hear this kind acknowledgement :)
> >
> >>> [Some people who received this message don't often get email from sj@...nel.org. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> >>>
> >>> Hi Huan,
> >>>
> >>>
> >>> First of all, thank you for this patch.  I have some high level comments and
> >>> questions as below.
> >>>
> >>> On Tue, 19 Sep 2023 17:52:33 +0800 Huan Yang <link@...o.com> wrote:
> >>>
> >>>> Now damos support filter contains two type.
> >>>> The first is scheme filter which will invoke each scheme apply,
> >>>> second is scheme action filter, which will filter out unwanted action.
> >>> IMHO, that's not clear definition of the types.  Conceptually, all the filters
> >>> are same.  Nonetheless, they are having different behaviors because they are
> >>> handled in different layer depending on which layer is more efficient to handle
> >> Thanks for these instructions to help me understand the design idea of
> >> damos filter.
> >>> those[1].
> >>>
> >>> I agree this is complicated and a bit verbose explanation, but I don't feel
> >>> your scheme vs action definition is making it easier to understand.
> >>>
> >>>> But this implement is scattered in different implementations
> >>> Indeed the implementation is scattered in core layer and the ops layer
> >>> depending on the filter types.  But I think this is needed for efficient
> >>> handling of those.
> >> IMO, some simple filter can have a common implement, like anon filter? And,
> >> for some special type, each layer offer their own?
> > Do you mean there could be two filter types that better to be handled in
> > different layer for efficiency, but share common implementation?  Could you
> > please give me a more specific example?
> 
> It's hard to offer a specific example due to current ops implement is so
> great to cover
> 
> much situation. Maybe Heterogeneous memory may add a new ops(just
> examples of
> 
> imprudence, like intel's or other network memory?) . And offer a common
> ops filter code
> 
> can may some simple type be reused.
> 
> >
> >>>> and hard to reuse or extend.
> >>>   From your first patch, which extending the filter, the essential change for the
> >>> extension is adding another enum to 'enum damos_filter_type' (1 line) and
> >>> addition of switch case in '__damos_pa_filter_out()' (3 lines).  I don't think
> >>> it looks too complicated.  If you're saying it was hard to understand which
> >> Yes, indeed.
> >>> part really need to be modifed, I think that makes sense.  But in the case,
> >>> we might need more comments rather than structural change.
> >>>
> >>>> This patchset clean up those filter code, move into filter.c and add header
> >>>> to expose filter func.(patch 2)
> >>> Again, I agree more code cleanup is needed.  But I'm not sure if this is the
> >>> right way.  Especially, I'm unsure if it really need to have a dedicated file.
> >> I think the filter code scatter into each layer may make code hard to
> >> reuse, other ops
> >>
> >> may need anon filter or something. So, make code into a dedicated file
> >> may good?
> >>
> >> (just my opinion.)
> > Again, I'm not super confident about my understanding.  But sounds like you're
> > partly worrying about duplication of some code in each ops implementation
> > (modules in same layer).  Please correct me if I'm wrong, with specific,
> > detailed and realistic examples.
> >
> > If my guess is not that wrong, I can agree to the concern.  Nevertheless, we
> > already have a dedicated source file for such common code between ops
> > implementaions, namely ops-common.c.
> Yes, no need to split a single file to drive filter.
> >
> > That said, we don't have duplicated DAMOS filters implementation code in
> > multipe ops implementation at the moment.  It could have such duplication in
> > the future, but I think it's not too late to make such cleanup after or just
> > before such duplication heppen.  IOW, I'd suggest to not make changes for
> > something that _might_ happen in future.
> 
> Yes, indeed. We needn't to make this right now.(That's the why RFC,
> meanwhile, my code is
> 
> not good.)
> 
> >
> >>> To my humble eyes, it doesn't look like making code clearly easier to read
> >>> compared to the current version.  This is probably because I don't feel your
> >>> scheme vs action definition is clear to understand.  Neither it is
> >> Yes, indeed, current code not clean, the idea didn't take shape.
> >>> deduplicating code.  The patch adds lines more that deleted ones, according to
> >>> its diffstat.  For the reason, I don't see clear benefit of this change.
> >> In this code, maybe just a little benefit when other ops need to filter
> >> anon page? :)
> > As mentioned above, it's unclear when, how, and for what benefit we will
> > support anon pages filter from vaddr.  I'd again suggest to not make change
> > only for future change that not clear if we really want to make for now.
> >
> >>>     I
> >>> might be biased, having NIH (Not Invented Here) sindrome, or missing something.
> >>> Please let me know if so.
> >>>
> >>>> And add a new filter "workingset" to protect workingset page.
> >>> Can you explain expected use cases of this new filter type in more detail?
> >>> Specifically, for what scheme action and under what situation this new filter
> >> IMO, page if marked as workingset, mean page in task's core
> >> workload(maybe have
> >>
> >> seen the refault, and trigger refault protect). So, for lru sort or reclaim,
> >>
> >> we'd better not touch it?(If any wrong, please let me know.)
> > Still unclear to me.  Could I ask you more specific and detailed case?
> 
> I may have misunderstood, I suggest you just look:
> 
> Page workingset flag will mark to page when page need to deactive or
> refault and
> 
> find it already marked. In some memory path(migrate, reclaim or else.),
> touch
> 
> workingset page require special attention.(Enter psi mm stall or else)
> 
> So, I think help filter out this is helpful.(Of course, just thought
> experiment, no helpful data).
> 
> As, above and below. This RFC patch. :). I will share when get valuable
> data.
> 
> >
> >>> type will be needed?  If you have some test data for the use case and could
> >>> share it, it would be very helpful for me to understand why it is needed.
> >> Sorry, this type just from my knowledge of MM, have no test data.
> >>
> >> For futher learn of DAMON, I'll try it.
> > Yes, that will be very helpful.
> >
> > And from this point, I'm getting an impression that the purpose of this RFC is
> > not for making a real change for specific use case that assumed to make real
> > benefits, but just for getting opinions about some imaginable changes that
> > _might_ be helpful, and _might_ be made in future.  If so, making the point
> Yes, I just learn DAMON a little time, and offer some thinking for this.
> > more clear would be helpful for me to give you opinion earlier.  If that's the
> > case, my opinion is this.  I agree DAMON code has many rooms of improvement in
> > terms of readability, so cleanup patches are welcome.  Nevertheless, I'd prefer
> > to make changes only when it is reasonable to expect it's providing _real_
> > improvement just after be applied, or at least very near and specific future.
> Yes, keep this and change when we need indeed. :)

Ok, makes sense.  Let's do further discussion after some more data and
realistic detailed example be available. :)  Looking forward to!


Thanks,
SJ

> >
> >
> > Thanks,
> > SJ
> 
> Thans,
> 
> Huan
> 
> >>>> Do we need this and cleanup it?
> >>> I think I cannot answer for now, and your further clarification and patient
> >>> explanation could be helpful.
> >>>
> >>> [1] https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/tree/Documentation/mm/damon/design.rst?h=mm-everything-2023-09-20-19-38#n400
> >>>
> >>>
> >>> Thanks,
> >>> SJ
> >> Thanks,
> >>
> >> Huan
> >>
> >>>> Huan Yang (2):
> >>>>     mm/damos/filter: Add workingset page filter
> >>>>     mm/damos/filter: Damos filter cleanup
> >>>>
> >>>>    include/linux/damon.h    |  62 +-----------------
> >>>>    mm/damon/Makefile        |   2 +-
> >>>>    mm/damon/core-test.h     |   7 ++
> >>>>    mm/damon/core.c          |  93 ++++-----------------------
> >>>>    mm/damon/filter.c        | 135 +++++++++++++++++++++++++++++++++++++++
> >>>>    mm/damon/filter.h        | 119 ++++++++++++++++++++++++++++++++++
> >>>>    mm/damon/paddr.c         |  29 +++------
> >>>>    mm/damon/reclaim.c       |  48 +++++++++++---
> >>>>    mm/damon/sysfs-schemes.c |   1 +
> >>>>    9 files changed, 325 insertions(+), 171 deletions(-)
> >>>>    create mode 100644 mm/damon/filter.c
> >>>>    create mode 100644 mm/damon/filter.h
> >>>>
> >>>> --
> >>>> 2.34.1
> >>>>
> >>>>

Powered by blists - more mailing lists