Linux Virtualization - Jun 2020 - Should SEV-ES #VC use IST? (Re: [PATCH] Allow RDTSC and RDTSCP from userspace)

On Tue, Jun 23, 2020 at 04:39:26PM +0100, Andrew Cooper
wrote:
> On 23/06/2020 16:23, Peter Zijlstra wrote:
> > On Tue, Jun 23, 2020 at 04:59:14PM +0200, Joerg Roedel wrote:
> >> Yes, this is a start, it doesn't cover the case where the NMI
stack is
> >> in-between, so I think you need to walk down regs->sp too.
> > That shouldn't be possible with the current code, I think.
> 
> NMI; #MC; Anything which IRET but isn't fatal - #DB, or #BP from
> patching, #GP from *_safe(), etc; NMI
> 
> Sure its a corner case, but did you hear that IST is evil?
Isn't current #MC unconditionally fatal from kernel? But yes, I was
sorta aware people want that changed.

And yes, NMI can recurse, mostly on #BP and #PF. Like I wrote, its
broken vs #MC.

But Joerg was talking about IST recursion with NMI in the middle,
something like: #DB, NMI, #DB, and not already being fatal. This one in
particular is ruled out by #DB itself clearing DR7 (but NMI would also
do that).
> P.S. did you also hear that with Rowhammer, userspace has a nonzero
> quantity of control over generating #MC, depending on how ECC is
> configured on the platform.
Yes, excellent stuff.

On 6/23/20 8:52 AM, Peter Zijlstra wrote:
> Isn't current #MC unconditionally fatal from kernel? But yes, I was
> sorta aware people want that changed.
Not unconditionally.  copy_to_iter_mcsafe() is a good example of one
thing we _can_ handle.

On Tue, Jun 23, 2020 at 09:03:56AM -0700, Dave Hansen
wrote:
> On 6/23/20 8:52 AM, Peter Zijlstra wrote:
> > Isn't current #MC unconditionally fatal from kernel? But yes, I
was
> > sorta aware people want that changed.
> 
> Not unconditionally.  copy_to_iter_mcsafe() is a good example of one
> thing we _can_ handle.
Urgh, I thought that stuff was still pending.

Anyway, the important thing is that it is fatal if we hit early NMI.
Which I think still holds.