Re: [PATCH v7 03/15] s390/zcrypt: driver callback to indicate resource in use

From: Tony Krowiak
Date: Tue Apr 28 2020 - 10:37:10 EST




On 4/28/20 7:07 AM, Harald Freudenberger wrote:
On 28.04.20 00:24, Tony Krowiak wrote:

On 4/27/20 4:20 AM, Pierre Morel wrote:

On 2020-04-07 21:20, Tony Krowiak wrote:
Introduces a new driver callback to prevent a root user from unbinding
an AP queue from its device driver if the queue is in use. The intent of
this callback is to provide a driver with the means to prevent a root user
from inadvertently taking a queue away from a guest and giving it to the
host while the guest is still using it.
How can we know, at this point if the guest uses or not the queue?
The struct ap_matrix_mdev has a field, struct kvm *kvm, which holds a pointer to KVM when
the matrix mdev is in use by a guest. This patch series also introduces a shadow_crycb (soon to
be shadow_apcb) which holds the AP configuration for the guest. Between those two things,
the driver can detect when a queue is in use by a guest.

Do you want to say that this prevents to take away a queue when it is currently assigned to a VFIO device?
and with a guest currently using this VFIO device?
No, I do not. The intent here is to enforce the proper procedure for giving up a queue so it is done
deliberately. Before taking a queue away from the matrix mdev, its APQN should be unassigned
from the matrix mdev. That is not to say that if there are major objections to this that we can't
base in_use upon the queue being in use by a guest at the time. Maybe that is preferable to
the community. I'll leave it to them to state their case.

The callback will
be invoked whenever a change to the AP bus's sysfs apmask or aqmask
attributes would result in one or more AP queues being removed from its
driver. If the callback responds in the affirmative for any driver
queried, the change to the apmask or aqmask will be rejected with a device
in use error.
AFAIU you mean that Linux's driver's binding and unbinding mechanism is not sufficient to avoid this issue because unbind can not be refused by the driver.
Correct!


The reason why we do not want a single queue to be removed from the VFIO driver is because the VFIO drivers works on a matrix, not on queues, and for the matrix to be consistent it needs to acquire all queues defined by the cross product of all APID and AQID assigned to the matrix.
Not correct. The reason why is because we do not want a queue to be surreptitiously removed
without the guest administrator being aware of its removal.

This functionality is valid for the host as for the guests and is handled automatically by the firmware with the CRYCB.
The AP bus uses QCI to retrieve the host CRYCB and build the hosts AP queues.

If instead to mix VFIO CRYCB matrix handling and queues at the same level inside the AP bus we separate these different firmware entities in two different software entities.

If we make the AP bus sit above a CRYCB/Matrix bus, and in the way virtualize the QCI and test AP queue instructions:
- we can directly pass a matrix device to the guest though a VFIO matrix device
- the consistence will be automatic
- the VFIO device and parent device will be of the same kind which would make the design much more clearer.
- there will be no need for these callback because the consistence of the matrix will be guaranteed by firmware
As stated in my response above, the issue here is not consistency. While the design you describe
may be reasonable, it is a major departure from what is out in the field. In other words, that ship
has sailed.


For this patch, only non-default drivers will be queried. Currently,
there is only one non-default driver, the vfio_ap device driver.
You mean that the admin may take queues away from the "default driver", while the queue is in use, to give it to an other driver?
Why is it to avoid in one way and not in the other way?
Because the default drivers have direct control over the queues and can ensure they are empty
and reset before giving up control. The vfio driver does not have direct control over the queues
because they have been passed through to the guest.
No, that's not true. The 'default' drivers have no change to do anything with an APQN when it is removed
from the driver. They get the very same notification which is the remove() callback as the vfio dd gets
and have the very same change to do something here. The more interesting thing here is, that the remove()
callback invocation is usually because a hardware HAS BEEN GONE AWAY. Neither the 'default' drivers
nor the vfio dd can do a reset on a not-any-more existing APQN.
And it is also not true that the vfio dd has no direct control over the queue because they have been passed
through to the guest. It's the job of the vfio dd to modify the guest's APM, AQM, ADM masks to disable
the guest's access to the APQN and then the vfio can (try to) do a reset.

The context here is when a sysadmin deliberately takes one or more queues away from a
guest by changing the apmask or aqmask; we are not talking about the the case where an
adapter is deconfigured or disappears. The idea here is to prevent a sysadmin for the host
from taking a queue away from a KVM guest that is using it. IMHO, control
over that queue should belong to the guest until such time as the guest gives it up or the
guest is terminated. Since the zcrypt drivers are directly responsible for their AP queues,
it is not necessary to implement this callback, although there is nothing precluding that.

The
vfio_ap device driver manages AP queues passed through to one or more
guests
I read this as if a queue may be passed to several guest...
please, rephrase or explain.
AP queues is plural, so it is true that AP queues can be passed through
to more than one guest. I see your point, however, so I'll reword that
to be more clear.

and we don't want to unexpectedly take AP resources away from
guests which are most likely independently administered.
When you say "independently administered", you mean as a second admin inside the host, don't you?
I mean that a guest can be administered by a different person than the host administrator.
Again, I'll try to clarify this.


Regards,
Pierre