Re: [RFC PATCH 1/5] mm/damon/core: fix nr_accesses_bp underflow in damon_moving_sum

[SeongJae Park <sj@kernel.org>]

Hello Ravi,

On Sat, 16 May 2026 14:03:53 -0700 Ravi Jonnalagadda <ravis.opensrc@gmail.com> wrote:

> Guard against unsigned integer underflow when nomvsum/len_window
> exceeds mvsum.

How could this happen?  mvsum is assumed to be same to nomvsum at the beginning
of the window.  Hence, even if there is only zero new_value, at the end of the
window, mvsum should be exactly zero.  Of course there could be a bug that
breaks the assumption.

> When that subtraction wraps, the moving sum returns a
> near-ULONG_MAX value and corrupts nr_accesses_bp.
> 
> If subtrahend > mvsum, return new_value: this clamps the moving-sum
> estimate to the current observation rather than wrapping.

I guess you saw this issue in real, and this change should fix the issue.  But
I think we should know why and how mvsum < nomvum / len_window can unexpectedly
happen, and fix that.

Could you share more details about when and how the situation happens?


Thanks,
SJ

[...]