Re: [xen-unstable test] 57852: regressions - FAIL

[Ian Campbell <ian.campbell@citrix.com>]

On Thu, 2015-06-11 at 09:45 +0100, Ian Campbell wrote:
> On Thu, 2015-06-11 at 08:02 +0100, Jan Beulich wrote:
> > >>> On 10.06.15 at 16:08, <ian.campbell@citrix.com> wrote:
> > > On Wed, 2015-06-10 at 14:45 +0100, Jan Beulich wrote:
> > >> So if we're going to approach Intel with this - will you or should I?
> > > 
> > > I think it'd be best coming from you.
> > 
> > Just have sent it off; in putting together the technical details it
> > became clear that elbling* indeed are at a newer microcode level,
> > so I think this at least slightly raises the chances of an update to
> > help fiano* (if so I of course wonder why the vendor hasn't made
> > a suitable BIOS update available yet).
> 
> It's possible that there is one which we've not applied...

I've now run a bunch of adhoc runs with the microcode update in place
(from 0x416 to 0x428 on these particular machines):

58468	fiano0	guest-stop
58479	fiano0	guest-stop
58485	fiano0	windows-install
58494	fiano0	guest-stop
58499	fiano1	guest-stop
58509	fiano1	windows-install
58516	fiano1	guest-stop
58527*	fiano0	guest-stop
58531	fiano0	guest-stop
58534	fiano0	guest-stop
58537	fiano0	guest-stop
58538	fiano1	guest-stop
58544	fiano1	guest-stop
58547	fiano1	guest-stop
58550	fiano0	guest-stop
58555	fiano0	guest-stop
58557	fiano0	guest-stop
58560	fiano1	guest-stop
58563	fiano1	guest-stop
58565	fiano1	windows-install

(*) rebuilt binaries because previous build was gc'd, same versions as
before.

So 3/20 = 15% failure rate (fiano0: 1/11=9%; fiano1: 2/9=22%). Which is
better than the ~50% seen at the start of this thread, so it is worth
applying the ucode update I think (and it would have been regardless the
right thing to do), 

I do think a 15-20% failure rate might be worthy of further
investigation by Intel too, since the failure rate with no-apicv was
1/13 = 7% (fiano0: 1/7=14%, fiano1: 0/6=0%), although those numbers are
less significant due to fewer runs.

Ian.