Re: [RFC PATCH v0] Docs/mm/damon/design: document data attributes report infra design (beyond page table accessed-bit milestone 2)

[SJ Park <sj@kernel.org>]

Hi Aveline,

On Wed, 11 Nov 2026 18:35:52 +1100 Aveline Noir <jm5905938@gmail.com> wrote:

> Hi SJ,
> 
> On Sat, 17 Oct 2026 12:57:45 -0700 SJ Park wrote:
> 
> > +The data attribute probe supports the reports.  The probes can be configured to
> > +hit for a given memory region depending on the existence of specific types of
> > +reports that associated with the memory region in given report buffers.
> 
> I'm wondering how this would interact with the region size.  If each
> page of a region is reported with probability q, a region of s pages
> would hit with probability 1 - (1 - q)^s, so a large region may hit
> more often than a small region of the same access density.

Thank you for sharing this finding.  Seems this theory assumes the probe will
hit for report of any byte of the region exists in the report buffers.  Under
the assumption, that makes sense.

But I'd say the assumed implementation is not realistic.  That's because DAMON
design does probing for only single sampling byte of the region, not the entire
bytes of the region.  In a real implementation that is aligned with the design,
the probe will hit if a report associated with the sampling byte exists.

> The page
> table accessed-bit check samples one page per region, so its hit rate
> doesn't depend on the region size in the same way.

I'd say it is not because it uses accessed-bit, but it respects the address
sampling design.

> 
> A quick replay of the crediting rule of Ravi's RFC v4 [1] on synthetic
> traces showed a similar trend.  The adaptive regions adjustment would
> mitigate this to some extent, but large cold-only regions still kept
> high hit rates after split and merge in the replay.  I'm not sure how
> much it matters on real workloads, though.
> 
> Would crediting the hits relative to the region size, e.g., reported
> bytes / region bytes, make sense to you?  It would need probe_hits
> wider than u8, so it is not free.  If you think it is worth exploring,
> I may run more experiments and share the results.

As long as the real implementation respects the address sampling design, I
think this will not be needed.

Thank you again for raising this question, and please let me know if I missed
something.


Thanks,
SJ

[...]