Re: [PATCH v6 21/21] s390: doc: detailed specifications for AP virtualization

From: Tony Krowiak
Date: Thu Jul 05 2018 - 09:30:02 EST

On 07/04/2018 12:31 PM, Boris Fiuczynski wrote:
On 07/03/2018 06:36 PM, Tony Krowiak wrote:
On 07/02/2018 07:10 PM, Halil Pasic wrote:

On 06/29/2018 11:11 PM, Tony Krowiak wrote:
This patch provides documentation describing the AP architecture and
design concepts behind the virtualization of AP devices. It also
includes an example of how to configure AP devices for exclusive
use of KVM guests.

Signed-off-by: Tony Krowiak <akrowiak@xxxxxxxxxxxxx>

I don't like the design of external interfaces except for:
* cpu model features, and
* reset handling.

In particular:

1) The architecture is such that authorizing access (via APM, AQM and ADM)
to an AP queue that is currently not configured (e.g. the card not physically
plugged, or just configured off). That seems to be a perfectly normal use

Your assign operations however enforce that the resource is bound to your
driver, and thus the existence of the resource in the host.

It is clear: we need to avoid passing trough resources to guests that are not
dedicated for this purpose (e.g. a queue utilized by zcrypt). But IMHO
we need a different mechanism.

Interesting that you wait until v6 to bring this up. I agree, this is a normal
use case, but there is currently no mechanism in the AP bus for drivers to
reserve devices that are not yet configured. There is proposed solution in the
works, but until such time that is available the only choice is to disallow
assignment of AP queues to a guest that are not bound to the vfio_ap device driver.

2) I see no benefit in deferring the exclusivity check to vfio_ap_mdev_open().
The downside is however pretty obvious: management software is notified about
a 'bad configuration' only at an attempted guest start-up. And your current QEMU
patches are not very helpful in conveying this piece of information.

It only becomes a 'bad configuration' if the two guests are started concurrently.
Is there value in being able to configure two mediated devices with the same
queue if the intent is to never run two guests using those mediated devices
simultaneously? If so, then the only time the exclusivity check can be done
is when the guest opens the mediated device. If not, then we can certainly
prevent multiple mediated devices from being assigned the same queue.

In my view, while a mediated device is used by a guest, it is not a guest and
can be configured any way an administrator prefers. If we get concurrence
that doing an exclusivity check when an adapter or domain is assigned to
the mediated device, I'll make that change.

I've talked with Boris, and AFAIR he said this is not acceptable to him (@Boris
can you confirm).

Then I suggest Boris participate in the review and explain why.

[To make things a bit easier I am not going to address the aspect of not-currently-exiting host resources.]
Your current implementation does provide active configurations that work with existing host resources. These need to be bound to the vfio_ap driver.
Libvirt allows to define objects (e.g. domains or networks). These are just definitions and do NOT bind any resources. The defined resources are bound once the definition is started.
Currently I am assuming that an ap matrix device is defined in libvirt outside of a libvirt domain (an ap definition). The mediated device of the ap matrix device is used in a libvirt domain by referencing it via its UID.
When a libvirt domain is started the mediated device should exist and be configured correctly as every other host resource.
Therefore there needs to be something new in libvirt that allows one to define, start, stop and undefine an ap matrix device. After a define the ap definition for an ap matrix device would exist in libvirt only.
Once you start the ap definition the result should be a well configured ready to be used mediated device representing the ap definition which can be used configuration-error free by a libvirt domain. Please not that the start of an ap definition is independent from the start of a libvirt domain using the ap definition.
Can you explain to me how that can be accomplished?

I can make a similar case for the mediated devices. Mediated devices play no role in guest configuration until a vfio-ap
device is specified on the QEMU command line when starting a guest. In other words, a mediated device configuration is
independent from the start of a guest using the mediated device. To answer your question then, if there are two or more
mediated devices with the same APQN(s) assigned, then only start one libvirt domain that uses one of these mediated
devices. This begs the question: Does libvirt preclude one from defining a domain that uses a host device (of any kind)
that must be dedicated to a single guest? If not, then isn't it incumbent upon the administrator to ensure he doesn't
start two guests with the same dedicated host device? Wouldn't that same logic apply to AP devices?

Having said that, I have no problem disallowing assignment of an AP queue to more than one mediated device, however; suppose
an administrator - for whatever reason - wants to create multiple mediated devices with the same APQN(s) assigned, but
never intends to run more than one guest using one of those mediated devices concurrently. The question is - as I have
asked in another response - is there a use case for allowing an administrator to configure multiple mediated devices with
the same APQN assigned?

3) We indicate the reason for failure due to a configuration problem (exclusivity
or resource allocation) via pr_err() that is via kernel messages. I don't think
this is very tooling/management software friendly, and I hope we don't expect admins
to work with the sysfs interface long term. I mean the effects of the admin actions
are not very persistent. Thus if the interface is a painful one, we are talking
about potentially frequent pain.

We have multiple layers of software, each with its own logging facilities. Figuring
out what went wrong when a guest fails to start is always a painful process IMHO.
Typically, one has to view the log for each component in the stack to figure out
what went wrong and often times, still can't figure it out. Of course, we can help
out here by having QEMU put out a better message when this problem occurs. But the
bottom line is, does the community think that allowing an administrator to configure
multiple mediated devices with the same queues have value? In other words, are
there potential use cases that would required this?

4) If I were to act out the role of the administrator, I would prefer to think of
specifying or changing the access controls of a guest in respect to AP (that is
setting the AP matrix) as a single atomic operation -- which either succeeds or fails.

I don't understand what you are describing here. How would this be done? Are you
suggesting the admin somehow provides the masks en masse?

The operation should succeed for any valid configuration, and fail for any invalid

The current piecemeal approach seems even less fitting if we consider changing the
access controls of a running guest. AFAIK changing access controls for a running
guest is possible, and I don't see a reason why should we artificially prohibit this.

Setting and clearing bits in the APM/AQM/ADM of a guest's CRYCB is certainly possible,
but there is a lot more to it than merely setting and clearing bits. What you seem
to be describing here is hot plug/unplug which I stated in the cover letter is
forthcoming. It is currently prohibited for good reason.

I think the current sysfs interface for manipulating the matrix is good for
manual playing around, but I would prefer having an interface that is better
suited for programs (e.g. ioctl).

That wouldn't be a problem, but do we have a use case for it?