Re: Requesting for freeze exception for VT-d posted-interrupts

[Wei Liu <wei.liu2@citrix.com>]

On Tue, Jul 14, 2015 at 05:01:33PM +0100, Jan Beulich wrote:
> >>> On 14.07.15 at 17:02, <wei.liu2@citrix.com> wrote:
> > On Tue, Jul 14, 2015 at 03:46:46PM +0100, Jan Beulich wrote:
> >> >>> On 14.07.15 at 16:17, <wei.liu2@citrix.com> wrote:
> >> > On Tue, Jul 14, 2015 at 11:09:15AM +0100, Jan Beulich wrote:
> >> >> >>> On 14.07.15 at 11:21, <wei.liu2@citrix.com> wrote:
> >> >> > On Tue, Jul 14, 2015 at 05:51:02AM +0000, Wu, Feng wrote:
> >> >> >> Is it possible to get to 4.6 if making this feature default off?
> >> >> > 
> >> >> > Note that I'm not the only one who makes the decision and I can't speak
> >> >> > for maintainers. The first thing you ought to do is to convince
> >> >> > maintainers, not me.
> >> >> > 
> >> >> > If you ask for my opinion -- I don't see a point in releasing feature
> >> >> > with security flaw in design, even if it is off by default. 
> >> >> 
> >> >> It was actually me who suggested that by flagging this experimental
> >> >> and defaulting it to off, chances would increase for this to be allowed
> >> >> in without said issue fixed.
> >> > 
> >> > Are you satisfied with that?  Currently I only know from this email
> >> > there is concern with regard to security but I don't know what it is and
> >> > how big an impact it can possibly have.
> >> > 
> >> > I could maybe go dig up that series and try to understand what is the
> >> > security implication, but it would take a long time and I'm not sure I
> >> > have the right technical background to make the call.
> >> 
> >> The thing is that the way vCPU-s are being put on lists attached to
> >> pCPU-s, in a pathological case (which can be "helped" by a malicious
> >> tool stack) all vCPU-s could pile up on one such list. List traversal (in
> >> an interrupt handler) could then take (almost) arbitrarily long.
> > 
> > You mentioned "malicious toolstack", does that mean this feature, if on,
> > doesn't expose new attack vector to malicious guest?
> 
> I think getting a guest to affect this would be more involved, but
> I can't entirely exclude it.
> 
> > And what do you mean by "malicious toolstack"? I don't see patches
> > related to toolstack.
> 
> This is because the tool stack can control placement of vCPU-s on
> pCPU-s, not because new tool stack code is being added.
> 

I fished out the thread and tried my best to digest that.

It does seem that it requires tool stack help to pin too many vcpus to a
pcpu to trigger a problem.  Another possibility I can think of is that
the scheduler piling too many vcpus on one pcpu.

All in all, neither of the above two approaches can be used directly by
a malicious guest, so the concern with regard to security is not as big
as I thought.

Jan, I agree with you this feature should be marked as experimental if
we are to merge it for 4.6.

Wu, note the decision has not been made because the patch series is not
in shape yet, in theory you do have time to address all issues by
Friday if you want to try. I will revisit this on Friday.

Wei.

> Jan