Re: [PATCH v9 12/22] s390: vfio-ap: sysfs interfaces to configure control domains

From: Halil Pasic
Date: Wed Aug 22 2018 - 06:43:21 EST

On 08/22/2018 11:42 AM, Cornelia Huck wrote:
On Wed, 22 Aug 2018 01:18:20 +0200
Halil Pasic <pasic@xxxxxxxxxxxxx> wrote:

On 08/21/2018 07:07 PM, Tony Krowiak wrote:
This convention has been enforced by the kernel since v1. This is also
enforced by both the LPAR as well as in z/VM. The following is from the
PR/SM Planning Guide:

Control Domain
A logical partition's control domains are those cryptographic domains for which remote secure
administration functions can be established and administered from this logical partition. This
logical partitionâs control domains must include its usage domains. For each index selected in the
usage domain index list, you must select the same index in the control domain index list

That's interesting.

IMHO this quote is quite a half-full half-empty cup one:
* it mandates the set of usage domains is a subset of the set
of the control domains, but
* it speaks of independent controls, namely about the 'usage domain index'
and the 'control domain index list' and makes the enforcement of the rule
a job of the administrator (instead of codifying it in the controls).

I'm wondering if a configuration with a usage domain that is not also a
control domain is rejected outright? Anybody tried that? :)

I did not. This is my first exposure to the PR/SM Planning Guide. As I
stated previously the HMC interface enforces the convention by UI design:
in the activation profile you can either configure a domain as 'control' or
'control and usage' domain -- think radio button. I have no idea how
is this information feed into PR/SM and same goes for alternatives to
specify it. So I'm also very curious about this.

Another interesting question: On what level does z/VM and PR/SM
enforce the convention (i.e. on privilege level does the code
doing the enforcement run)?

Consequently, I'm going to opt for ensuring this is clearly documented. Based on the fact you've
requested clarification of many points described in this section of the doc, I
think I'll try putting my meager skills as a wordsmith to work to hopefully clarify things.
I'll run it by you when I complete that task to see if I've succeeded:)

I don't think just a doc update will do. Let me explain why.

What describe as "... note that the AQM and ADM masks configured for the
mediated matrix device will be logically OR'd together to create the ADM
stored in the CRYCB referenced from the guest's SIE state description."
is a gotcha at best. The member of struct ap_matrix and the member of the
respective apcb in the crycb are both called 'adm', but ap_matrix.adm is
not an ADM as we know it from the architecture, but rather ~ AQM & ADM.

I feel pretty strongly about this one. If we want to keep the enforcement
in the kernel, I guess, the assign_domain should set the bit corresponding
bit not only in ap_matrix.aqm but also in ap_matrix.adm. When the
ap_matrix is committed into the crycb no further manipulating the masks
should take place.

Would you be fine if the control domain interface stated that it is
used to configure _additional_ control domains and the usage domain
interface stated that it is used to define usage and implicitly also
control domains? (And make the usage domain interface also set the
equivalent bit in the control domain mask.)

I'm fine with the interface, otherwise I would not have r-b-ed the patch.
What I strongly dislike is the implementation is IMHO very confusing (along
the lines "surprise, surprise it is called adm but it ain't adm"). This
implementation detail however can be changed any time, so I did not want
to start a big discussion as we wanted to get this out ASAP. But since
it was brought up, I decided to put in my two cents.

I don't feel strongly about whether to enforce this convention about AQM
and ADM in the kernel or not. Frankly, I don't know what is behind the
rule. Since I can't tell if any problems are to be expected if this
convention is violated, I would feel more comfortable if the rule was
accommodated higher in the management stack.

I guess it depends:

- If this is a case of: "Don't configure control domains that are not
also usage domains. You are likely to go through
{code,firmware,hardware} paths that are generally not used.",
configure it in the kernel.
- If this rather is "Everybody is doing that, it's a general
convention.", configure it higher up in the stack (libvirt?)

My guess is that it's the second, but I really don't know. Maybe
somebody else will help us answer this question, and also tell
what is the rationale behind the rule.