Re: [Xen-devel] [PATCH] xen/domctl: lower loglevel of XEN_DOMCTL_memory_mapping

[Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>]

On Thu, Sep 10, 2015 at 02:59:17AM -0600, Jan Beulich wrote:
> >>> On 10.09.15 at 10:55, <tiejun.chen@xxxxxxxxx> wrote:
> >>  Sort of (the patch has the intended effect, but for its size very
> >> many rough edges).
> >>
> > 
> > I guess we need to amend the original parameter, once_mapping_mfns, like 
> > this,
> > 
> > /* xen_once_mapping_mfns: memory mapping mfn bumbers once. */
> > unsigned int xen_once_mapping_mfns;
> > size_param("once_mapping_mfns", xen_once_mapping_mfns);
> > 
> > static void xen_once_mapping_mfns_setup(void)
> > {
> >      if ( once_mapping_mfns < 64 )
> >          xen_once_mapping_mfns = 64;
> >      else if ( once_mapping_mfns > 1024 )
> >          xen_once_mapping_mfns = 1024;
> >      else
> >          xen_once_mapping_mfns = once_mapping_mfns;
> > }
> 
> Right, that's one of the things that would need taking care of.
> (Whether enforcing an upper limit is actually needed I'm not
> sure - we generally allow the admin to shoot himself in the foot
> if he wants to. And whether the lower limit should be 64 instead
> of just ensuring the limit is not zero is another question.)

64 was semi-arbitrary  - it ended up giving good latency on
highly scalar machines (8 socket). Higher numbers ended up
affecting the latency.

But higher numbers on small socket machines were OK.
(As they do not have 8 IOMMU VT-d chipsets all potentially
flodding the QPI with serialized cache flushes).
> 
> Jan
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel