Re: PROBLEM: Probabilistic segfault on AMD hardware with INVLPGB

From: Henrik Böving

Date: Tue Jun 23 2026 - 08:29:35 EST


Thanks for the quick response!

> Upgrading the firmware on the host solved those
> segfaults for us.

Our hardware is currently on microcode version 0x0B002162, while the
advisory you linked names 0x0B00211E as the patched version for our
model. So I think we should already be sufficiently up to date for this
patch, but I don't know (and couldn't find anything on the web) about
the details of AMD microcode versioning.

Henrik

On Mon, Jun 22, 2026 at 9:00 PM Rik van Riel <riel@xxxxxxxxxxx> wrote:
>
> On Fri, 2026-06-19 at 15:17 +0200, Henrik Böving wrote:
> > [1.] One line summary of the problem
> > Probabilistic segfault on AMD hardware with INVLPGB
> >
> > [2.] Full description of the problem/report:
> >
> > Since upgrading from kernel 6.14 to 7.1, we've been observing
> > probabilistic segfaults on our build machines. We were able to
> > reproduce
> > these segfaults on two copies of the same hardware, equipped with an
> > AMD EPYC 9455 48-Core Processor. Subsequently, we managed to bisect
> > the
> > segfault down to the introduction of INVLPGB-based page invalidation
> > by
> > enabling/disabling BROADCAST_TLB_FLUSH on kernel 6.15. The segfaults
> > also do not reproduce on hardware that does not have the INVLPGB
> > instruction.
> >
> We have also seen increased segfault rates when we
> first deployed Turin CPUs. This turned out to be
> AMD erratum AMD-SB-3029.
>
> Upgrading the firmware on the host solved those
> segfaults for us.
>
> Doing a bulk data analysis across the Meta fleet
> shows that Turin segfault rates are in the same
> range as those of other CPU microarchitectures,
> of several vendors.
>
> https://www.amd.com/en/resources/product-security/bulletin/amd-sb-3029.html
>
> Now, it is possible that INVLPGB has increased
> the likelihood of hitting some race condition, and
> it also possible that the kernel has a TLB
> flushing bug somewhere, but it may make sense
> to start with a firmware upgrade.
>
> --
> All Rights Reversed.