On Wed, Sep 22, 2021 at 03:50:25PM +0200, Jan Beulich wrote:
> On 22.09.2021 15:31, Andrew Cooper wrote:
> > On 21/09/2021 08:17, Jan Beulich wrote:
> >> @@ -933,7 +934,18 @@ int __init dom0_construct_pv(struct doma
> >> #ifdef CONFIG_SHADOW_PAGING
> >> if ( opt_dom0_shadow )
> >> {
> >> + bool preempted;
> >> +
> >> printk("Switching dom0 to using shadow paging\n");
> >> +
> >> + nr_pt_pages = dom0_paging_pages(d, nr_pages);
> >> +
> >> + do {
> >> + preempted = false;
> >> + shadow_set_allocation(d, nr_pt_pages, &preempted);
> >> + process_pending_softirqs();
> >> + } while ( preempted );
> >
> > This is still broken.
> >
> > The loop setting the shadow allocation needs to be outside of this
> > conditional, because it is not related to early activation of the l1tf
> > tasklet.
>
> Well, I'm not sure what to say. On v1 you already said so. But then you
> didn't care to reply to me responding:
>
> "Are you suggesting to set up a (perhaps large) shadow pool just in
> case we need to enable shadow mode on Dom0? And all of this memory
> to then remain unused in the majority of cases?
>
> Plus even if so, I'd view this as a 2nd, independent step, largely
> orthogonal to the handling of "dom0=shadow". If somebody really
> wanted that, I think this should be driven by an explicit setting
> of the shadow pool size, indicating the admin is willing to waste
> the memory.
Maybe an acceptable compromise would be to allocate the pool if
opt_dom0_shadow || opt_pv_l1tf_hwdom?
opt_pv_l1tf_hwdom is not enabled by default, so an admin opting to
enable it should also be willing to reserve the memory it requires in
case it needs activating during runtime.
Roger.
