Re: [RFC 00/19] KVM: s390/crypto/vfio: guest dedicated crypto adapters

[Cornelia Huck <cohuck@redhat.com>]

On Fri, 13 Oct 2017 13:38:45 -0400
Tony Krowiak <akrowiak@linux.vnet.ibm.com> wrote:

> Tony Krowiak (19):
>   KVM: s390: SIE considerations for AP Queue virtualization
>   KVM: s390: refactor crypto initialization
>   s390/zcrypt: new AP matrix bus
>   s390/zcrypt: create an AP matrix device on the AP matrix bus
>   s390/zcrypt: base implementation of AP matrix device driver
>   s390/zcrypt: register matrix device with VFIO mediated device
>     framework
>   KVM: s390: introduce AP matrix configuration interface
>   s390/zcrypt: support for assigning adapters to matrix mdev
>   s390/zcrypt: validate adapter assignment
>   s390/zcrypt: sysfs interfaces supporting AP domain assignment
>   s390/zcrypt: validate domain assignment
>   s390/zcrypt: sysfs support for control domain assignment
>   s390/zcrypt: validate control domain assignment
>   KVM: s390: Connect the AP mediated matrix device to KVM
>   s390/zcrypt: introduce ioctl access to VFIO AP Matrix driver
>   KVM: s390: interface to configure KVM guest's AP matrix
>   KVM: s390: validate input to AP matrix config interface
>   KVM: s390: New ioctl to configure KVM guest's AP matrix
>   s390/facilities: enable AP facilities needed by guest

I'll try to summarize all of this in my own words, both to make sure I
understand the design correctly and to give others a different view on
this.

[I'm completely disregarding control domains here.]

On s390, we have cryptographic coprocessor cards, which are modeled on
Linux as devices on the AP bus. There's also a concept called domains,
which means an individual queue of a crypto device is basically a
(card,domain) tuple. We model this something like the following
(assuming we have access to cards 3 and 4 and domains 1 and 2):

AP -> card3 -> queue (3,1)
            -> queue (3,2)
   -> card4 -> queue (4,1)
            -> queue (4,2)

(The AP bus is a bit different for backwards compat.)

If we want to virtualize this, we can use a feature provided by the
hardware. We basically attach a satellite control block to our main
hardware virtualization control block and the hardware takes care of
(mostly) everything.

For this control block, we don't specify explicit tuples, but a list of
cards and a list of domains. The guest will get access to the cross
product.

Because of this, we need to take care that the lists provided to
different guests don't overlap; i.e., we need to enforce sane
configurations. Otherwise, one guest may get access to things like
secret keys for another guest.

The idea of this patch set is to introduce a new device, the matrix
device. This matrix device hangs off a different root and acts as the
node where mdev devices hang off.

If you now want to give the tuples (4,1) and (4,2), you need to do the
following:

- Unbind the (4,1) and (4,2) tuples from their ap bus driver.
- Bind the (4,1) and (4,2) tuples to the ap matrix driver.
- Create the mediated device.
- Assign card 4 and domains 1 and 2.

QEMU will now simply consume the mediated device and things should work.