Re: [PATCH v2.1] mm/damon/core: fix false positive in damos_quota_is_full() when esz is zero

[Liew Rui Yan <aethernet65535@gmail.com>]

On Fri, 04 Sep 2026 07:05:35 -0700 SJ Park <sj@kernel.org> wrote:

> On Fri,  4 Sep 2026 16:07:41 +0800 Liew Rui Yan <aethernet65535@gmail.com> wrote:
> 
> > First, I'd like to clarify that this isn't a problem encountered by a
> > real user, it's just a scenario I came up with.
> 
> Thank you for clarifying this.
> 
> > 
> > 1. Users sample qt_exceeds periodically (e.g., every 10 minutes).
> 
> What's the purpose of this sampling?
> 
> > 
> > 2. Within this 10 minute sampling interval, the counter aggregates both
> >    the real quota exhaustions and the increments caused by esz==0.
> > 
> > 3. When users notice a high qt_exceeds value, they eventually realize
> >    (perhaps by reading the code or documentation) that it includes the
> >    counts from the esz==0 state.
> > 
> > 4. To get the actual quota exhaustion statistics, the user is now forced
> >    to perform additional testing and implement external filtering to
> >    separate the esz==0 increments from the real exceeds.
> > 
> > Even if we explicitly state in the documentation that qt_exceeds
> > includes the esz==0 counts, it still burdens the user.  The user still
> > has to figure out how to filter out the esz==0 increments externally to
> > get the signal they actually care about.
> 
> Users set the temporal goal.  They can know when the goal is achieved since
> most of the goal metrics are already exposed to user space.  Users can also
> show the current effective quotas.  I agree that can be cumbersome, but how
> problematic it is?  Also, as I asked above, why they want to do this after all?
> 
> > 
> > Honestly, I struggle to imagine any valid use case where a user would
> > actually rely on the qt_exceeds increments caused by esz==0 to make
> > decisions.
> > 
> > If the only purpose of qt_exceeds is to let users "easily notice" if the
> > quota is too small,
> 
> I agree it could be a signal to show if the quota is too small.  But the real
> purpose of qt_exceeds is, in my opinion, letting users understand how DAMOS is
> internally working now.  After all, how much quota means if it is too small or
> not?  That all depends on the real use case and complicated things including
> their SLO etc.

Thank you for your clarify.

> 
> If documentation is saying the purpose of qt_exceeds is to show if the quota is
> too small, that is what need to be updated.

I completely agree your perspective.

This is the current documentation of qt_exceeds:

    - ``qt_exceeds``: Total number of times the quota of the scheme has exceeded.

Although it state the purpose of this statistic, I think adding a
note to clarify that this stat also increase when the quota is zero (but
not unlimited) would be helpful for users.  For example:

    Usually, a quota of zero means the DAMOS scheme has an unlimited
    quota, so qt_exceeds will not increase.  However, if user sets a
    temporal quota goal, the quota is set to zero once the goal is
    [over]-achieved.  In this situation, qt_exceeds will still increase.

I can prepare a formal documentation patch based on this if you agree.

[...]

That said, it's not important for me to add explanations to the
document, but may I know why commit [2] changed the behavior which
introduced by commit [1]?

Commit [1] Behavior:

    if (quota->esz && quota->changed_sz >= quota->esz)
        s->stat.qt_exceeds++;

Commit [2] Behavior:

    if (damos_quota_is_full(quota, c->min_region_sz))
        s->stat.qt_exceeds++;

Before commit [2], qt_exceeds will only increase when quota->esz is not
zero, but after commit [2], qt_exceeds also increase even when
quota->esz is zero.  I'd love to understand the rationale behind this
change to better grasp the design evolution.

[1] 6268eac34ca30 ("mm/damon/schemes: account how many times quota limit has exceeded")
    (Fri Jan 14 14:10:20 2022 -0800)
[2] c7ec7d5f6b3d1 ("mm/damon/core: handle <min_region_sz remaining quota as empty")
    (Mon Apr 27 18:33:50 2026 -0700)

Best regards,
Rui Yan