Re: [RFC PATCH v3 07/28] mm/damon/core: do data attributes monitoring

[SeongJae Park <sj@kernel.org>]

On Sat, 16 May 2026 19:00:10 +0000 sashiko-bot@kernel.org wrote:

> Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:
> - [High] The `probe_hits` array counters can silently overflow, yielding corrupted monitoring results.
> - [High] DAMON dynamically merges regions based strictly on `nr_accesses`, which corrupts the spatial granularity of `probe_hits`.
> - [High] If DAMON is run without `check_accesses`, all regions will be aggressively merged up to `sz_limit`, neutralizing data attribute monitoring.
> - [High] Size-weighted average calculations for `probe_hits` overflow `unsigned long` on 32-bit systems.
> --
> 
> commit 55dcc5294b4ee96c9d7e5a24e9dcdb0d798c401d
> Author: SeongJae Park <sj@kernel.org>
> 
> mm/damon/core: do data attributes monitoring
> 
> Implement the data attributes monitoring execution.  Update kdamond to
> invoke the probes application callback, and reset the aggregated number
> of per-region per-probe positive samples for every aggregation interval.
> 
> > diff --git a/mm/damon/core.c b/mm/damon/core.c
> > index e5667660b4cd9..dde3c8d8fef89 100644
> > --- a/mm/damon/core.c
> > +++ b/mm/damon/core.c
> > @@ -1886,10 +1886,14 @@ static void kdamond_reset_aggregated(struct damon_ctx *c)
> >  		struct damon_region *r;
> >  
> >  		damon_for_each_region(r, t) {
> > +			int i;
> > +
> >  			trace_damon_aggregated(ti, r, damon_nr_regions(t));
> >  			damon_warn_fix_nr_accesses_corruption(r);
> >  			r->last_nr_accesses = r->nr_accesses;
> >  			r->nr_accesses = 0;
> > +			for (i = 0; i < DAMON_MAX_PROBES; i++)
> > +				r->probe_hits[i] = 0;
> 
> Can probe_hits overflow here?
> 
> The probe_hits array is defined as unsigned char, which has a maximum value
> of 255.
> 
> The number of samples per interval is user-configurable and can legitimately
> exceed 255 (e.g., a 1000ms aggregation with a 1ms sample interval yields 1000
> samples).
> 
> If a probe matches continuously under such configurations, could the counter
> silently overflow and wrap around, corrupting the monitoring results?

No blocker, as I previously explained [1].

> 
> >  			damon_verify_reset_aggregated(r, c);
> >  		}
> >  		ti++;
> > @@ -3179,6 +3183,8 @@ static int kdamond_fn(void *data)
> >  
> >  		if (ctx->ops.check_accesses)
> >  			max_nr_accesses = ctx->ops.check_accesses(ctx);
> > +		if (ctx->ops.apply_probes)
> > +			ctx->ops.apply_probes(ctx);
> 
> If DAMON is configured strictly for data attribute monitoring where
> check_accesses is NULL but apply_probes is set, max_nr_accesses and all
> regional nr_accesses persistently evaluate to 0.
> 
> Will this cause all adjacent regions to have a calculated difference
> threshold of 0, resulting in all regions merging indiscriminately up to
> sz_limit during the subsequent kdamond_merge_regions() call?
> 
> Additionally, damon_merge_regions_of() merges adjacent regions by evaluating
> only nr_accesses:
> 
> abs(prev->nr_accesses - r->nr_accesses) <= thres
> 
> Since this merge condition ignores probe_hits, could two contiguous regions
> with identical access frequencies but vastly different data attributes be
> incorrectly merged?

Correct, and that's the intended behavior.  Data attributes are just additional
information.  Data access pattern, specifically nr_accesses, is the primary
information that drives region splits and merges.  I'm planning to let users
set the primary information in future, though.

> 
> Does calculating the size-weighted average for probe_hits risk overflowing on
> 32-bit systems?
> 
> When damon_merge_two_regions() calculates averages using:
> 
> (metric * sz_l + metric * sz_r) / (sz_l + sz_r)
> 
> Both sz_l and sz_r are evaluated in bytes as unsigned long. On 32-bit
> systems where ULONG_MAX is roughly 4 GB, a 200 MB region with 25 probe_hits
> yields an intermediate product of 5,000,000,000.
> 
> Does this silently overflow the 32-bit unsigned long and corrupt the region
> metrics?

I previously answered [1] this.  32-bit arch users could use addr_unit to avoid
this issue.

> 
> >  
> >  		if (time_after_eq(ctx->passed_sample_intervals,
> >  					next_aggregation_sis)) {
> 
> -- 
> Sashiko AI review · https://sashiko.dev/#/patchset/20260516183712.81393-1-sj@kernel.org?part=7

[1] https://lore.kernel.org/20260516003559.141476-1-sj@kernel.org


Thanks,
SJ