Re: [RFC PATCH 2/3] x86/mm: make sure LAM is up-to-date during context switching

[Yosry Ahmed <yosryahmed@google.com>]

On Mon, Mar 11, 2024 at 5:42 AM Kirill A. Shutemov
<kirill.shutemov@linux.intel.com> wrote:
>
> On Sat, Mar 09, 2024 at 01:37:06PM -0800, Yosry Ahmed wrote:
> > On Sat, Mar 9, 2024 at 8:34 AM Kirill A. Shutemov
> > <kirill.shutemov@linux.intel.com> wrote:
> > >
> > > On Sat, Mar 09, 2024 at 02:19:19AM +0000, Yosry Ahmed wrote:
> > > > I don't see how skipping set_tlbstate_lam_mode() for kthreads fixes this
> > > > problem. Do you mind elaborating?
> > >
> > > Define what problem is.
> > >
> > > Yes, in this scenario kthread gets more permissive LAM mode than it needs.
> > > But nothing breaks.
> >
> >
> > The problem here is not how the kthread runs at all. It is the fact
> > that if that kthread context switches into the user process that has
> > enabled LAM, it may not update CR3 because the mm doesn't change.
> > switch_mm_irqs_off() will only update CR3 in this case if there is a
> > pending TLB flush. Otherwise, we just return, even if the LAM for this
> > mm has changed.
> >
> > This can cause the process that has enabled LAM to run with LAM
> > disabled and fault on tagged addresses, right? Did I miss something?
>
> You are right. I think IPI is the way to go.
>
> Will you prepare a patch?

I am working on it.