Memory coherency issue with IO thread offloading?

[Michael Ellerman]

Daniel Black <daniel at mariadb.org> writes:
> Thanks Jens, Nick, Christophe and Michael for your work so far.
>
> Apologies for the out of thread email.
>
> Confirming MariabD-10.6+ is required( when we added liburing), and
> previous versions used libaio (which tested without incident as mpe
> retested).

Thanks! I was trying to follow the threads from the mariadb/server
github to find the CI but didn't have much luck. No .travis.yml these
days :)

> We were (we're now back on the old good kernel Jens indicated) getting
> failures like https://buildbot.mariadb.org/#/builders/231/builds/16857
> in a container (of various distro userspaces) on bare metal.
>
> bare metal end of /proc/cpuinfo
>
> processor    : 127
> cpu        : POWER9, altivec supported
> clock        : 3283.000000MHz
> revision    : 2.2 (pvr 004e 1202)
>
> timebase    : 512000000
> platform    : PowerNV
> model        : 9006-22P
> machine        : PowerNV 9006-22P
> firmware    : OPAL
> MMU        : Radix

Thanks, looks like that's a Boston, which at least has the same CPU as
the machine I'm testing on.

I have also reproduced it on bare metal now, but it seems much easier to
hit it in an LPAR (on P10). Not sure that's an actual data point, could
just be due to other activity on the box.

cheers