[Skiboot] (no subject)
Benjamin Herrenschmidt
benh at kernel.crashing.org
Sat Dec 2 14:39:15 AEDT 2017
On Sat, 2017-12-02 at 10:13 +1000, Nicholas Piggin wrote:
> On Fri, 01 Dec 2017 14:51:56 -0600
> Benjamin Herrenschmidt <benh at kernel.crashing.org> wrote:
>
> > On Fri, 2017-12-01 at 01:52 +1000, Nicholas Piggin wrote:
> > > Firstly, Linux should set up MMU registers like PIDR properly in
> > > its per-CPU mmu initialisation at boot, patch for that should not
> > > be controversial.
> >
> > So I was wondering how we got things into the PWC since we are
> > in real mode when we do the flush, but the above explains it.
> >
> > We come up with a stale PIDR and turn the MMU on. We only
> > execute/load/store from Q3 but prefetch/speculation can hit Q0 and thus
> > get crap into the PWC.
> >
> > So I think setting PIDR to 0 is the main fix. Cleaning up the rest also
> > makes sense of course.
> >
> > Or am I missing something else still ?
>
> Yes that's what happens. After realizing this, the bug no longer requires
> translations to be cached in real mode (does the architecture guarantee
> that?)
I don't think we can speculate accross contexts.
> I think LPID is safer to be zeroed as well because in theory we might
> get speculative access into quadrant 0/1.
Right.
Cheers,
Ben.
More information about the Skiboot
mailing list