Re: [PATCH] hw/xen: Set suppress-vmdesc for Xen machines

[Anthony PERARD <anthony.perard@xxxxxxxxxx>]

On Fri, Oct 16, 2020 at 12:01:47PM -0400, Jason Andryuk wrote:
> On Fri, Oct 16, 2020 at 11:38 AM Anthony PERARD
> <anthony.perard@xxxxxxxxxx> wrote:
> >
> > On Tue, Oct 13, 2020 at 03:05:06PM -0400, Jason Andryuk wrote:
> > > xen-save-devices-state doesn't currently generate a vmdesc, so restore
> > > always triggers "Expected vmdescription section, but got 0".  This is
> > > not a problem when restore comes from a file.  However, when QEMU runs
> > > in a linux stubdom and comes over a console, EOF is not received.  This
> > > causes a delay restoring - though it does restore.
> > >
> > > Setting suppress-vmdesc skips looking for the vmdesc during restore and
> > > avoids the wait.
> >
> > suppress-vmdesc is only used during restore, right? So starting a guest
> > without it, saving the guest and restoring the guest with
> > suppress-vmdesc=on added will work as intended? (I'm checking that migration
> > across update of QEMU will work.)
> 
> vmdesc is a json description of the migration stream that comes after
> the QEMU migration stream.  For our purposes, <migration
> stream><vmdesc json blob>.  Normal QEMU savevm will generate it,
> unless suppress-vmdesc is set.  QEMU restore will read it because:
> "Try to read in the VMDESC section as well, so that dumping tools that
> intercept our migration stream have the chance to see it."
> 
> Xen save does not go through savevm, but instead
> xen-save-devices-state, which is a subset of the QEMU savevm.  It
> skips RAM since that is read out through Xen interfaces.  Xen uses
> xen-load-devices-state to restore device state.  That goes through the
> common qemu_loadvm_state which tries to read the vmdesc stream.
> 
> For Xen, yes, suppress-vmdesc only matters for the restore case, and
> it suppresses the attempt to read the vmdesc.  I think every Xen
> restore currently has "Expected vmdescription section, but got -1" in
> the -dm.log since the vmdesc is missing.  I have not tested restoring
> across this change, but since it just controls reading and discarding
> the vmdesc stream, I don't think it will break migration across
> update.

Thanks for the explanation.

Acked-by: Anthony PERARD <anthony.perard@xxxxxxxxxx>

Do you think you could send a patch for libxl as well? Since libxl in
some cases may use the "pc machine instead of "xenfv". I can send the
patch otherwise.

Cheers,

-- 
Anthony PERARD