Re: [PATCH 6.6.y 0/2] jbd2: backport CVE-2026-89566, CVE-2026-89567

[Artem Dinaburg]

Hi Ted,

> Note that some tests can be flaky. So when you see a test failure, it
> can be useful to run a test multiple times

Thanks for the advice. We are still working through the testing and
wanted to send you a progress update.

We have completed three full iterations of
ext4/all,xfs/4k,btrfs/4k -g auto on the same unpatched 6.6.159-rc1
baseline (72bada21be7b), followed by one full iteration with the
submitted two-patch jbd2 series.

We used QEMU/KVM with fresh guest and test disks for each
configuration.

Following your suggestion, we also completed ten focused trials on
each kernel for 21 suspect configuration/test pairs: 420 trials in
total. Each trial used a separately booted guest and fresh disks.

In those ten runs on each kernel, we observed the following flaky
tests. Counts below are failures out of ten, unpatched / patched:

Flaky on both kernels:

  btrfs/4k btrfs/028: 3 / 3
  btrfs/4k btrfs/192: 7 / 6
  btrfs/4k generic/449: 4 / 7
  ext4/1k generic/051: 3 / 5
  ext4/1k generic/388: 2 / 3
  ext4/1k generic/475: 5 / 3
  ext4/adv generic/455: 5 / 5
  ext4/adv generic/482: 4 / 2
  ext4/ext3 generic/388: 8 / 9

Flaky on the unpatched kernel only:

  ext4/bigalloc_4k generic/320: 2 / 0

Flaky on the patched kernel only:

  ext4/4k generic/045: 0 / 1

We are now running 200 additional iterations of generic/045 on each
kernel in each of ext4/4k and ext4/1k. Ten runs leave wide confidence
intervals: a rare failure could easily have been missed on the
unpatched kernel. The additional runs will let us estimate failure
rates and their 95% confidence intervals to assess whether the test
fails more often on the patched kernel. We will finish all planned
runs before comparing the results.


Thanks,
Artem Dinaburg

On Sat, Oct 10, 2026 at 7:54 PM Theodore Tso <tytso@xxxxxxx> wrote:
>
> On Sat, Oct 10, 2026 at 06:12:10PM -0500, Artem Dinaburg wrote:
> >
> > Our initial unpatched/patched runs showed some differences in both
> > directions. We cannot yet attribute those to the series, so we have
> > paused the patched comparison and are running three complete
> > iterations on the same unpatched baseline to assess repeatability.
>
> Note that some tests can be flaky.  So when you see a test failure, it
> can be useful to run a test multiple times, so you can see if the test
> is a reliable hard failure, or one that only fails, say, 30% of the
> time.  So to run generic/363 10 times in the ext4/1k configuration,
> you might do soemthing like this:
>
>    kvm-xfstests -c ext4/1k generic/363 -C 10
>
>                                                 - Ted
>