[BUG] DAMON_STAT: idle percentiles out of order on 32-bit

[BUG] DAMON_STAT: idle percentiles out of order on 32-bit

[Aveline Noir]

Hi,

In an i386 QEMU guest built from mm-new commit ad9f2d463666,
DAMON_STAT reported these memory_idle_ms_percentiles values after a
statistics update:

  guest uptime:       22.51 s
  aggr_interval_us:   3200000
  percentile 0:      19200 ms
  percentile 50:         0 ms
  percentile 100:        0 ms

The 0th percentile is greater than the 50th and 100th.  The guest had
256 MiB of RAM (MemTotal: 254216 kB), a running kdamond,
CONFIG_X86_32=y, CONFIG_DAMON_STAT=y,
CONFIG_DAMON_STAT_ENABLED_DEFAULT=y, and CONFIG_COMPAT_32BIT_TIME=y.
aggr_interval_us changed from 0 to 1600000 and then 3200000, showing
that statistics updates ran before the final read.

In damon_stat_set_idletime_percentiles(), the loop compares

  next_percentile <= accounted_bytes * 100 / total_sz

where accounted_bytes and total_sz are unsigned long.  On 32-bit, when
total_sz exceeds 42,949,672 bytes, the wrapped product cannot yield
100%, even after all bytes are accounted for.  Upper entries in
memory_idle_ms_percentiles then retain an earlier value or remain zero
on first use.  DAMON_STAT monitors the system RAM address range, which
exceeds that threshold in this guest.  I did not capture a per-region
trace in this run.

The guest init read the module parameters after 22 seconds of uptime;
it did not write to them.  The i386 config, guest init, serial log, and
build/run steps:

  https://github.com/jm5905938/kernel-mail/tree/195d7017352a3d0f4a0475b98f3b03650f5c14ea/cases/2026/damon/2026-10-15-bug-reports/03-stat-percentiles-32bit-overflow

Thanks,
Aveline Noir

Re: [BUG] DAMON_STAT: idle percentiles out of order on 32-bit

[SJ Park]

Hi Avelin,

On Thu, 15 Oct 2026 04:24:12 -0700 Aveline Noir <jm5905938@gmail.com> wrote:

> Hi,
> 
> In an i386 QEMU guest built from mm-new commit ad9f2d463666,
> DAMON_STAT reported these memory_idle_ms_percentiles values after a
> statistics update:
> 
>   guest uptime:       22.51 s
>   aggr_interval_us:   3200000
>   percentile 0:      19200 ms
>   percentile 50:         0 ms
>   percentile 100:        0 ms
> 
> The 0th percentile is greater than the 50th and 100th.  The guest had
> 256 MiB of RAM (MemTotal: 254216 kB), a running kdamond,
> CONFIG_X86_32=y, CONFIG_DAMON_STAT=y,
> CONFIG_DAMON_STAT_ENABLED_DEFAULT=y, and CONFIG_COMPAT_32BIT_TIME=y.
> aggr_interval_us changed from 0 to 1600000 and then 3200000, showing
> that statistics updates ran before the final read.
> 
> In damon_stat_set_idletime_percentiles(), the loop compares
> 
>   next_percentile <= accounted_bytes * 100 / total_sz
> 
> where accounted_bytes and total_sz are unsigned long.  On 32-bit, when
> total_sz exceeds 42,949,672 bytes, the wrapped product cannot yield
> 100%, even after all bytes are accounted for.  Upper entries in
> memory_idle_ms_percentiles then retain an earlier value or remain zero
> on first use.  DAMON_STAT monitors the system RAM address range, which
> exceeds that threshold in this guest.  I did not capture a per-region
> trace in this run.
> 
> The guest init read the module parameters after 22 seconds of uptime;
> it did not write to them.  The i386 config, guest init, serial log, and
> build/run steps:

Could you share more details about your real use case that require the setup
and how this becomes a real problem?


Thanks,
SJ

[...]

Re: [BUG] DAMON_STAT: idle percentiles out of order on 32-bit

[Aveline Noir]

Hi SJ,

I haven't seen this affect a real workload. I found it while testing
DAMON_STAT on 32-bit.

The upper percentiles can stay at zero or keep old values. If someone
uses them to judge how much memory is cold or to choose a reclaim
threshold, they could get the wrong picture.

I'll send the fix and test results in a separate mail.

Thanks,
Aveline Noir