On Tue, Jul 26, 2022 at 11:21 AM Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>
> Darrick,
>
> This backport series contains mostly fixes from v5.14 release along
> with three deferred patches from the joint 5.10/5.15 series [1].
>
> I ran the auto group 10 times on baseline (v5.10.131) and this series
> with no observed regressions.
>
> I ran the recoveryloop group 100 times with no observed regressions.
> The soak group run is in progress (10+) with no observed regressions
> so far.
>
> I am somewhat disappointed from not seeing any improvement in the
> results of the recoveryloop tests comapred to baseline.
>
> This is the summary of the recoveryloop test results on both baseline
> and backport branch:
>
> generic,455, generic/457, generic/646: pass
> generic/019, generic/475, generic/648: failing often in all config
> generic/388: failing often with reflink_1024
> generic/388: failing at ~1/50 rate for any config
> generic/482: failing often on V4 configs
> generic/482: failing at ~1/100 rate for V5 configs
> xfs/057: failing at ~1/200 rate for any config
>
> I observed no failures in soak group so far neither on baseline nor
> on backport branch. I will update when I have more results.
>
Some more results after 1.5 days of spinning:
1. soak group reached 100 runs (x5 configs) with no failures
2. Ran all the tests also on debian/testing with xfsprogs 5.18 and
observed a very similar fail/pass pattern as with xfsprogs 5.10
3. Started to run the 3 passing recoveryloop tests 1000 times and
an interesting pattern emerged -
generic/455 failed 3 times on baseline (out of 250 runs x 5 configs),
but if has not failed on backport branch yet (700 runs x 5 configs).
And it's not just failures, it's proper data corruptions, e.g.
"testfile2.mark1 md5sum mismatched" (and not always on mark1)
I will keep this loop spinning, but I am cautiously optimistic about
this being an actual proof of bug fix.
If these results don't change, I would be happy to get an ACK for the
series so I can post it after the long soaking.
Thanks,
Amir.
